compile.cpp revision 5945:d2907f74462e
1/* 2 * Copyright (c) 1997, 2012, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25#include "precompiled.hpp" 26#include "asm/macroAssembler.hpp" 27#include "asm/macroAssembler.inline.hpp" 28#include "classfile/systemDictionary.hpp" 29#include "code/exceptionHandlerTable.hpp" 30#include "code/nmethod.hpp" 31#include "compiler/compileLog.hpp" 32#include "compiler/disassembler.hpp" 33#include "compiler/oopMap.hpp" 34#include "opto/addnode.hpp" 35#include "opto/block.hpp" 36#include "opto/c2compiler.hpp" 37#include "opto/callGenerator.hpp" 38#include "opto/callnode.hpp" 39#include "opto/cfgnode.hpp" 40#include "opto/chaitin.hpp" 41#include "opto/compile.hpp" 42#include "opto/connode.hpp" 43#include "opto/divnode.hpp" 44#include "opto/escape.hpp" 45#include "opto/idealGraphPrinter.hpp" 46#include "opto/loopnode.hpp" 47#include "opto/machnode.hpp" 48#include "opto/macro.hpp" 49#include "opto/matcher.hpp" 50#include "opto/memnode.hpp" 51#include "opto/mulnode.hpp" 52#include "opto/node.hpp" 53#include "opto/opcodes.hpp" 54#include "opto/output.hpp" 55#include "opto/parse.hpp" 56#include "opto/phaseX.hpp" 57#include "opto/rootnode.hpp" 58#include "opto/runtime.hpp" 59#include "opto/stringopts.hpp" 60#include "opto/type.hpp" 61#include "opto/vectornode.hpp" 62#include "runtime/arguments.hpp" 63#include "runtime/signature.hpp" 64#include "runtime/stubRoutines.hpp" 65#include "runtime/timer.hpp" 66#include "utilities/copy.hpp" 67#ifdef TARGET_ARCH_MODEL_x86_32 68# include "adfiles/ad_x86_32.hpp" 69#endif 70#ifdef TARGET_ARCH_MODEL_x86_64 71# include "adfiles/ad_x86_64.hpp" 72#endif 73#ifdef TARGET_ARCH_MODEL_sparc 74# include "adfiles/ad_sparc.hpp" 75#endif 76#ifdef TARGET_ARCH_MODEL_zero 77# include "adfiles/ad_zero.hpp" 78#endif 79#ifdef TARGET_ARCH_MODEL_arm 80# include "adfiles/ad_arm.hpp" 81#endif 82#ifdef TARGET_ARCH_MODEL_ppc_32 83# include "adfiles/ad_ppc_32.hpp" 84#endif 85#ifdef TARGET_ARCH_MODEL_ppc_64 86# include "adfiles/ad_ppc_64.hpp" 87#endif 88 89 90// -------------------- Compile::mach_constant_base_node ----------------------- 91// Constant table base node singleton. 92MachConstantBaseNode* Compile::mach_constant_base_node() { 93 if (_mach_constant_base_node == NULL) { 94 _mach_constant_base_node = new (C) MachConstantBaseNode(); 95 _mach_constant_base_node->add_req(C->root()); 96 } 97 return _mach_constant_base_node; 98} 99 100 101/// Support for intrinsics. 102 103// Return the index at which m must be inserted (or already exists). 104// The sort order is by the address of the ciMethod, with is_virtual as minor key. 105int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) { 106#ifdef ASSERT 107 for (int i = 1; i < _intrinsics->length(); i++) { 108 CallGenerator* cg1 = _intrinsics->at(i-1); 109 CallGenerator* cg2 = _intrinsics->at(i); 110 assert(cg1->method() != cg2->method() 111 ? cg1->method() < cg2->method() 112 : cg1->is_virtual() < cg2->is_virtual(), 113 "compiler intrinsics list must stay sorted"); 114 } 115#endif 116 // Binary search sorted list, in decreasing intervals [lo, hi]. 117 int lo = 0, hi = _intrinsics->length()-1; 118 while (lo <= hi) { 119 int mid = (uint)(hi + lo) / 2; 120 ciMethod* mid_m = _intrinsics->at(mid)->method(); 121 if (m < mid_m) { 122 hi = mid-1; 123 } else if (m > mid_m) { 124 lo = mid+1; 125 } else { 126 // look at minor sort key 127 bool mid_virt = _intrinsics->at(mid)->is_virtual(); 128 if (is_virtual < mid_virt) { 129 hi = mid-1; 130 } else if (is_virtual > mid_virt) { 131 lo = mid+1; 132 } else { 133 return mid; // exact match 134 } 135 } 136 } 137 return lo; // inexact match 138} 139 140void Compile::register_intrinsic(CallGenerator* cg) { 141 if (_intrinsics == NULL) { 142 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 143 } 144 // This code is stolen from ciObjectFactory::insert. 145 // Really, GrowableArray should have methods for 146 // insert_at, remove_at, and binary_search. 147 int len = _intrinsics->length(); 148 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual()); 149 if (index == len) { 150 _intrinsics->append(cg); 151 } else { 152#ifdef ASSERT 153 CallGenerator* oldcg = _intrinsics->at(index); 154 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice"); 155#endif 156 _intrinsics->append(_intrinsics->at(len-1)); 157 int pos; 158 for (pos = len-2; pos >= index; pos--) { 159 _intrinsics->at_put(pos+1,_intrinsics->at(pos)); 160 } 161 _intrinsics->at_put(index, cg); 162 } 163 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 164} 165 166CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 167 assert(m->is_loaded(), "don't try this on unloaded methods"); 168 if (_intrinsics != NULL) { 169 int index = intrinsic_insertion_index(m, is_virtual); 170 if (index < _intrinsics->length() 171 && _intrinsics->at(index)->method() == m 172 && _intrinsics->at(index)->is_virtual() == is_virtual) { 173 return _intrinsics->at(index); 174 } 175 } 176 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 177 if (m->intrinsic_id() != vmIntrinsics::_none && 178 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 179 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 180 if (cg != NULL) { 181 // Save it for next time: 182 register_intrinsic(cg); 183 return cg; 184 } else { 185 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 186 } 187 } 188 return NULL; 189} 190 191// Compile:: register_library_intrinsics and make_vm_intrinsic are defined 192// in library_call.cpp. 193 194 195#ifndef PRODUCT 196// statistics gathering... 197 198juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 199jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 200 201bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 202 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 203 int oflags = _intrinsic_hist_flags[id]; 204 assert(flags != 0, "what happened?"); 205 if (is_virtual) { 206 flags |= _intrinsic_virtual; 207 } 208 bool changed = (flags != oflags); 209 if ((flags & _intrinsic_worked) != 0) { 210 juint count = (_intrinsic_hist_count[id] += 1); 211 if (count == 1) { 212 changed = true; // first time 213 } 214 // increment the overall count also: 215 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 216 } 217 if (changed) { 218 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 219 // Something changed about the intrinsic's virtuality. 220 if ((flags & _intrinsic_virtual) != 0) { 221 // This is the first use of this intrinsic as a virtual call. 222 if (oflags != 0) { 223 // We already saw it as a non-virtual, so note both cases. 224 flags |= _intrinsic_both; 225 } 226 } else if ((oflags & _intrinsic_both) == 0) { 227 // This is the first use of this intrinsic as a non-virtual 228 flags |= _intrinsic_both; 229 } 230 } 231 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 232 } 233 // update the overall flags also: 234 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 235 return changed; 236} 237 238static char* format_flags(int flags, char* buf) { 239 buf[0] = 0; 240 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 241 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 242 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 243 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 244 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 245 if (buf[0] == 0) strcat(buf, ","); 246 assert(buf[0] == ',', "must be"); 247 return &buf[1]; 248} 249 250void Compile::print_intrinsic_statistics() { 251 char flagsbuf[100]; 252 ttyLocker ttyl; 253 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 254 tty->print_cr("Compiler intrinsic usage:"); 255 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 256 if (total == 0) total = 1; // avoid div0 in case of no successes 257 #define PRINT_STAT_LINE(name, c, f) \ 258 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 259 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 260 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 261 int flags = _intrinsic_hist_flags[id]; 262 juint count = _intrinsic_hist_count[id]; 263 if ((flags | count) != 0) { 264 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 265 } 266 } 267 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 268 if (xtty != NULL) xtty->tail("statistics"); 269} 270 271void Compile::print_statistics() { 272 { ttyLocker ttyl; 273 if (xtty != NULL) xtty->head("statistics type='opto'"); 274 Parse::print_statistics(); 275 PhaseCCP::print_statistics(); 276 PhaseRegAlloc::print_statistics(); 277 Scheduling::print_statistics(); 278 PhasePeephole::print_statistics(); 279 PhaseIdealLoop::print_statistics(); 280 if (xtty != NULL) xtty->tail("statistics"); 281 } 282 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 283 // put this under its own <statistics> element. 284 print_intrinsic_statistics(); 285 } 286} 287#endif //PRODUCT 288 289// Support for bundling info 290Bundle* Compile::node_bundling(const Node *n) { 291 assert(valid_bundle_info(n), "oob"); 292 return &_node_bundling_base[n->_idx]; 293} 294 295bool Compile::valid_bundle_info(const Node *n) { 296 return (_node_bundling_limit > n->_idx); 297} 298 299 300void Compile::gvn_replace_by(Node* n, Node* nn) { 301 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 302 Node* use = n->last_out(i); 303 bool is_in_table = initial_gvn()->hash_delete(use); 304 uint uses_found = 0; 305 for (uint j = 0; j < use->len(); j++) { 306 if (use->in(j) == n) { 307 if (j < use->req()) 308 use->set_req(j, nn); 309 else 310 use->set_prec(j, nn); 311 uses_found++; 312 } 313 } 314 if (is_in_table) { 315 // reinsert into table 316 initial_gvn()->hash_find_insert(use); 317 } 318 record_for_igvn(use); 319 i -= uses_found; // we deleted 1 or more copies of this edge 320 } 321} 322 323 324static inline bool not_a_node(const Node* n) { 325 if (n == NULL) return true; 326 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 327 if (*(address*)n == badAddress) return true; // kill by Node::destruct 328 return false; 329} 330 331// Identify all nodes that are reachable from below, useful. 332// Use breadth-first pass that records state in a Unique_Node_List, 333// recursive traversal is slower. 334void Compile::identify_useful_nodes(Unique_Node_List &useful) { 335 int estimated_worklist_size = unique(); 336 useful.map( estimated_worklist_size, NULL ); // preallocate space 337 338 // Initialize worklist 339 if (root() != NULL) { useful.push(root()); } 340 // If 'top' is cached, declare it useful to preserve cached node 341 if( cached_top_node() ) { useful.push(cached_top_node()); } 342 343 // Push all useful nodes onto the list, breadthfirst 344 for( uint next = 0; next < useful.size(); ++next ) { 345 assert( next < unique(), "Unique useful nodes < total nodes"); 346 Node *n = useful.at(next); 347 uint max = n->len(); 348 for( uint i = 0; i < max; ++i ) { 349 Node *m = n->in(i); 350 if (not_a_node(m)) continue; 351 useful.push(m); 352 } 353 } 354} 355 356// Update dead_node_list with any missing dead nodes using useful 357// list. Consider all non-useful nodes to be useless i.e., dead nodes. 358void Compile::update_dead_node_list(Unique_Node_List &useful) { 359 uint max_idx = unique(); 360 VectorSet& useful_node_set = useful.member_set(); 361 362 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 363 // If node with index node_idx is not in useful set, 364 // mark it as dead in dead node list. 365 if (! useful_node_set.test(node_idx) ) { 366 record_dead_node(node_idx); 367 } 368 } 369} 370 371void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 372 int shift = 0; 373 for (int i = 0; i < inlines->length(); i++) { 374 CallGenerator* cg = inlines->at(i); 375 CallNode* call = cg->call_node(); 376 if (shift > 0) { 377 inlines->at_put(i-shift, cg); 378 } 379 if (!useful.member(call)) { 380 shift++; 381 } 382 } 383 inlines->trunc_to(inlines->length()-shift); 384} 385 386// Disconnect all useless nodes by disconnecting those at the boundary. 387void Compile::remove_useless_nodes(Unique_Node_List &useful) { 388 uint next = 0; 389 while (next < useful.size()) { 390 Node *n = useful.at(next++); 391 // Use raw traversal of out edges since this code removes out edges 392 int max = n->outcnt(); 393 for (int j = 0; j < max; ++j) { 394 Node* child = n->raw_out(j); 395 if (! useful.member(child)) { 396 assert(!child->is_top() || child != top(), 397 "If top is cached in Compile object it is in useful list"); 398 // Only need to remove this out-edge to the useless node 399 n->raw_del_out(j); 400 --j; 401 --max; 402 } 403 } 404 if (n->outcnt() == 1 && n->has_special_unique_user()) { 405 record_for_igvn(n->unique_out()); 406 } 407 } 408 // Remove useless macro and predicate opaq nodes 409 for (int i = C->macro_count()-1; i >= 0; i--) { 410 Node* n = C->macro_node(i); 411 if (!useful.member(n)) { 412 remove_macro_node(n); 413 } 414 } 415 // Remove useless expensive node 416 for (int i = C->expensive_count()-1; i >= 0; i--) { 417 Node* n = C->expensive_node(i); 418 if (!useful.member(n)) { 419 remove_expensive_node(n); 420 } 421 } 422 // clean up the late inline lists 423 remove_useless_late_inlines(&_string_late_inlines, useful); 424 remove_useless_late_inlines(&_boxing_late_inlines, useful); 425 remove_useless_late_inlines(&_late_inlines, useful); 426 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 427} 428 429//------------------------------frame_size_in_words----------------------------- 430// frame_slots in units of words 431int Compile::frame_size_in_words() const { 432 // shift is 0 in LP32 and 1 in LP64 433 const int shift = (LogBytesPerWord - LogBytesPerInt); 434 int words = _frame_slots >> shift; 435 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 436 return words; 437} 438 439// ============================================================================ 440//------------------------------CompileWrapper--------------------------------- 441class CompileWrapper : public StackObj { 442 Compile *const _compile; 443 public: 444 CompileWrapper(Compile* compile); 445 446 ~CompileWrapper(); 447}; 448 449CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 450 // the Compile* pointer is stored in the current ciEnv: 451 ciEnv* env = compile->env(); 452 assert(env == ciEnv::current(), "must already be a ciEnv active"); 453 assert(env->compiler_data() == NULL, "compile already active?"); 454 env->set_compiler_data(compile); 455 assert(compile == Compile::current(), "sanity"); 456 457 compile->set_type_dict(NULL); 458 compile->set_type_hwm(NULL); 459 compile->set_type_last_size(0); 460 compile->set_last_tf(NULL, NULL); 461 compile->set_indexSet_arena(NULL); 462 compile->set_indexSet_free_block_list(NULL); 463 compile->init_type_arena(); 464 Type::Initialize(compile); 465 _compile->set_scratch_buffer_blob(NULL); 466 _compile->begin_method(); 467} 468CompileWrapper::~CompileWrapper() { 469 _compile->end_method(); 470 if (_compile->scratch_buffer_blob() != NULL) 471 BufferBlob::free(_compile->scratch_buffer_blob()); 472 _compile->env()->set_compiler_data(NULL); 473} 474 475 476//----------------------------print_compile_messages--------------------------- 477void Compile::print_compile_messages() { 478#ifndef PRODUCT 479 // Check if recompiling 480 if (_subsume_loads == false && PrintOpto) { 481 // Recompiling without allowing machine instructions to subsume loads 482 tty->print_cr("*********************************************************"); 483 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 484 tty->print_cr("*********************************************************"); 485 } 486 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 487 // Recompiling without escape analysis 488 tty->print_cr("*********************************************************"); 489 tty->print_cr("** Bailout: Recompile without escape analysis **"); 490 tty->print_cr("*********************************************************"); 491 } 492 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 493 // Recompiling without boxing elimination 494 tty->print_cr("*********************************************************"); 495 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 496 tty->print_cr("*********************************************************"); 497 } 498 if (env()->break_at_compile()) { 499 // Open the debugger when compiling this method. 500 tty->print("### Breaking when compiling: "); 501 method()->print_short_name(); 502 tty->cr(); 503 BREAKPOINT; 504 } 505 506 if( PrintOpto ) { 507 if (is_osr_compilation()) { 508 tty->print("[OSR]%3d", _compile_id); 509 } else { 510 tty->print("%3d", _compile_id); 511 } 512 } 513#endif 514} 515 516 517//-----------------------init_scratch_buffer_blob------------------------------ 518// Construct a temporary BufferBlob and cache it for this compile. 519void Compile::init_scratch_buffer_blob(int const_size) { 520 // If there is already a scratch buffer blob allocated and the 521 // constant section is big enough, use it. Otherwise free the 522 // current and allocate a new one. 523 BufferBlob* blob = scratch_buffer_blob(); 524 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 525 // Use the current blob. 526 } else { 527 if (blob != NULL) { 528 BufferBlob::free(blob); 529 } 530 531 ResourceMark rm; 532 _scratch_const_size = const_size; 533 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size); 534 blob = BufferBlob::create("Compile::scratch_buffer", size); 535 // Record the buffer blob for next time. 536 set_scratch_buffer_blob(blob); 537 // Have we run out of code space? 538 if (scratch_buffer_blob() == NULL) { 539 // Let CompilerBroker disable further compilations. 540 record_failure("Not enough space for scratch buffer in CodeCache"); 541 return; 542 } 543 } 544 545 // Initialize the relocation buffers 546 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 547 set_scratch_locs_memory(locs_buf); 548} 549 550 551//-----------------------scratch_emit_size------------------------------------- 552// Helper function that computes size by emitting code 553uint Compile::scratch_emit_size(const Node* n) { 554 // Start scratch_emit_size section. 555 set_in_scratch_emit_size(true); 556 557 // Emit into a trash buffer and count bytes emitted. 558 // This is a pretty expensive way to compute a size, 559 // but it works well enough if seldom used. 560 // All common fixed-size instructions are given a size 561 // method by the AD file. 562 // Note that the scratch buffer blob and locs memory are 563 // allocated at the beginning of the compile task, and 564 // may be shared by several calls to scratch_emit_size. 565 // The allocation of the scratch buffer blob is particularly 566 // expensive, since it has to grab the code cache lock. 567 BufferBlob* blob = this->scratch_buffer_blob(); 568 assert(blob != NULL, "Initialize BufferBlob at start"); 569 assert(blob->size() > MAX_inst_size, "sanity"); 570 relocInfo* locs_buf = scratch_locs_memory(); 571 address blob_begin = blob->content_begin(); 572 address blob_end = (address)locs_buf; 573 assert(blob->content_contains(blob_end), "sanity"); 574 CodeBuffer buf(blob_begin, blob_end - blob_begin); 575 buf.initialize_consts_size(_scratch_const_size); 576 buf.initialize_stubs_size(MAX_stubs_size); 577 assert(locs_buf != NULL, "sanity"); 578 int lsize = MAX_locs_size / 3; 579 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 580 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 581 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 582 583 // Do the emission. 584 585 Label fakeL; // Fake label for branch instructions. 586 Label* saveL = NULL; 587 uint save_bnum = 0; 588 bool is_branch = n->is_MachBranch(); 589 if (is_branch) { 590 MacroAssembler masm(&buf); 591 masm.bind(fakeL); 592 n->as_MachBranch()->save_label(&saveL, &save_bnum); 593 n->as_MachBranch()->label_set(&fakeL, 0); 594 } 595 n->emit(buf, this->regalloc()); 596 if (is_branch) // Restore label. 597 n->as_MachBranch()->label_set(saveL, save_bnum); 598 599 // End scratch_emit_size section. 600 set_in_scratch_emit_size(false); 601 602 return buf.insts_size(); 603} 604 605 606// ============================================================================ 607//------------------------------Compile standard------------------------------- 608debug_only( int Compile::_debug_idx = 100000; ) 609 610// Compile a method. entry_bci is -1 for normal compilations and indicates 611// the continuation bci for on stack replacement. 612 613 614Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 615 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing ) 616 : Phase(Compiler), 617 _env(ci_env), 618 _log(ci_env->log()), 619 _compile_id(ci_env->compile_id()), 620 _save_argument_registers(false), 621 _stub_name(NULL), 622 _stub_function(NULL), 623 _stub_entry_point(NULL), 624 _method(target), 625 _entry_bci(osr_bci), 626 _initial_gvn(NULL), 627 _for_igvn(NULL), 628 _warm_calls(NULL), 629 _subsume_loads(subsume_loads), 630 _do_escape_analysis(do_escape_analysis), 631 _eliminate_boxing(eliminate_boxing), 632 _failure_reason(NULL), 633 _code_buffer("Compile::Fill_buffer"), 634 _orig_pc_slot(0), 635 _orig_pc_slot_offset_in_bytes(0), 636 _has_method_handle_invokes(false), 637 _mach_constant_base_node(NULL), 638 _node_bundling_limit(0), 639 _node_bundling_base(NULL), 640 _java_calls(0), 641 _inner_loops(0), 642 _scratch_const_size(-1), 643 _in_scratch_emit_size(false), 644 _dead_node_list(comp_arena()), 645 _dead_node_count(0), 646#ifndef PRODUCT 647 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), 648 _printer(IdealGraphPrinter::printer()), 649#endif 650 _congraph(NULL), 651 _late_inlines(comp_arena(), 2, 0, NULL), 652 _string_late_inlines(comp_arena(), 2, 0, NULL), 653 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 654 _late_inlines_pos(0), 655 _number_of_mh_late_inlines(0), 656 _inlining_progress(false), 657 _inlining_incrementally(false), 658 _print_inlining_list(NULL), 659 _print_inlining(0) { 660 C = this; 661 662 CompileWrapper cw(this); 663#ifndef PRODUCT 664 if (TimeCompiler2) { 665 tty->print(" "); 666 target->holder()->name()->print(); 667 tty->print("."); 668 target->print_short_name(); 669 tty->print(" "); 670 } 671 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2); 672 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false); 673 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); 674 if (!print_opto_assembly) { 675 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); 676 if (print_assembly && !Disassembler::can_decode()) { 677 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 678 print_opto_assembly = true; 679 } 680 } 681 set_print_assembly(print_opto_assembly); 682 set_parsed_irreducible_loop(false); 683#endif 684 685 if (ProfileTraps) { 686 // Make sure the method being compiled gets its own MDO, 687 // so we can at least track the decompile_count(). 688 method()->ensure_method_data(); 689 } 690 691 Init(::AliasLevel); 692 693 694 print_compile_messages(); 695 696 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) ) 697 _ilt = InlineTree::build_inline_tree_root(); 698 else 699 _ilt = NULL; 700 701 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 702 assert(num_alias_types() >= AliasIdxRaw, ""); 703 704#define MINIMUM_NODE_HASH 1023 705 // Node list that Iterative GVN will start with 706 Unique_Node_List for_igvn(comp_arena()); 707 set_for_igvn(&for_igvn); 708 709 // GVN that will be run immediately on new nodes 710 uint estimated_size = method()->code_size()*4+64; 711 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 712 PhaseGVN gvn(node_arena(), estimated_size); 713 set_initial_gvn(&gvn); 714 715 if (PrintInlining || PrintIntrinsics NOT_PRODUCT( || PrintOptoInlining)) { 716 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 717 } 718 { // Scope for timing the parser 719 TracePhase t3("parse", &_t_parser, true); 720 721 // Put top into the hash table ASAP. 722 initial_gvn()->transform_no_reclaim(top()); 723 724 // Set up tf(), start(), and find a CallGenerator. 725 CallGenerator* cg = NULL; 726 if (is_osr_compilation()) { 727 const TypeTuple *domain = StartOSRNode::osr_domain(); 728 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 729 init_tf(TypeFunc::make(domain, range)); 730 StartNode* s = new (this) StartOSRNode(root(), domain); 731 initial_gvn()->set_type_bottom(s); 732 init_start(s); 733 cg = CallGenerator::for_osr(method(), entry_bci()); 734 } else { 735 // Normal case. 736 init_tf(TypeFunc::make(method())); 737 StartNode* s = new (this) StartNode(root(), tf()->domain()); 738 initial_gvn()->set_type_bottom(s); 739 init_start(s); 740 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) { 741 // With java.lang.ref.reference.get() we must go through the 742 // intrinsic when G1 is enabled - even when get() is the root 743 // method of the compile - so that, if necessary, the value in 744 // the referent field of the reference object gets recorded by 745 // the pre-barrier code. 746 // Specifically, if G1 is enabled, the value in the referent 747 // field is recorded by the G1 SATB pre barrier. This will 748 // result in the referent being marked live and the reference 749 // object removed from the list of discovered references during 750 // reference processing. 751 cg = find_intrinsic(method(), false); 752 } 753 if (cg == NULL) { 754 float past_uses = method()->interpreter_invocation_count(); 755 float expected_uses = past_uses; 756 cg = CallGenerator::for_inline(method(), expected_uses); 757 } 758 } 759 if (failing()) return; 760 if (cg == NULL) { 761 record_method_not_compilable_all_tiers("cannot parse method"); 762 return; 763 } 764 JVMState* jvms = build_start_state(start(), tf()); 765 if ((jvms = cg->generate(jvms)) == NULL) { 766 record_method_not_compilable("method parse failed"); 767 return; 768 } 769 GraphKit kit(jvms); 770 771 if (!kit.stopped()) { 772 // Accept return values, and transfer control we know not where. 773 // This is done by a special, unique ReturnNode bound to root. 774 return_values(kit.jvms()); 775 } 776 777 if (kit.has_exceptions()) { 778 // Any exceptions that escape from this call must be rethrown 779 // to whatever caller is dynamically above us on the stack. 780 // This is done by a special, unique RethrowNode bound to root. 781 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 782 } 783 784 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 785 786 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 787 inline_string_calls(true); 788 } 789 790 if (failing()) return; 791 792 print_method("Before RemoveUseless", 3); 793 794 // Remove clutter produced by parsing. 795 if (!failing()) { 796 ResourceMark rm; 797 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 798 } 799 } 800 801 // Note: Large methods are capped off in do_one_bytecode(). 802 if (failing()) return; 803 804 // After parsing, node notes are no longer automagic. 805 // They must be propagated by register_new_node_with_optimizer(), 806 // clone(), or the like. 807 set_default_node_notes(NULL); 808 809 for (;;) { 810 int successes = Inline_Warm(); 811 if (failing()) return; 812 if (successes == 0) break; 813 } 814 815 // Drain the list. 816 Finish_Warm(); 817#ifndef PRODUCT 818 if (_printer) { 819 _printer->print_inlining(this); 820 } 821#endif 822 823 if (failing()) return; 824 NOT_PRODUCT( verify_graph_edges(); ) 825 826 // Now optimize 827 Optimize(); 828 if (failing()) return; 829 NOT_PRODUCT( verify_graph_edges(); ) 830 831#ifndef PRODUCT 832 if (PrintIdeal) { 833 ttyLocker ttyl; // keep the following output all in one block 834 // This output goes directly to the tty, not the compiler log. 835 // To enable tools to match it up with the compilation activity, 836 // be sure to tag this tty output with the compile ID. 837 if (xtty != NULL) { 838 xtty->head("ideal compile_id='%d'%s", compile_id(), 839 is_osr_compilation() ? " compile_kind='osr'" : 840 ""); 841 } 842 root()->dump(9999); 843 if (xtty != NULL) { 844 xtty->tail("ideal"); 845 } 846 } 847#endif 848 849 // Now that we know the size of all the monitors we can add a fixed slot 850 // for the original deopt pc. 851 852 _orig_pc_slot = fixed_slots(); 853 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 854 set_fixed_slots(next_slot); 855 856 // Now generate code 857 Code_Gen(); 858 if (failing()) return; 859 860 // Check if we want to skip execution of all compiled code. 861 { 862#ifndef PRODUCT 863 if (OptoNoExecute) { 864 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 865 return; 866 } 867 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler); 868#endif 869 870 if (is_osr_compilation()) { 871 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 872 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 873 } else { 874 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 875 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 876 } 877 878 env()->register_method(_method, _entry_bci, 879 &_code_offsets, 880 _orig_pc_slot_offset_in_bytes, 881 code_buffer(), 882 frame_size_in_words(), _oop_map_set, 883 &_handler_table, &_inc_table, 884 compiler, 885 env()->comp_level(), 886 has_unsafe_access(), 887 SharedRuntime::is_wide_vector(max_vector_size()) 888 ); 889 890 if (log() != NULL) // Print code cache state into compiler log 891 log()->code_cache_state(); 892 } 893} 894 895//------------------------------Compile---------------------------------------- 896// Compile a runtime stub 897Compile::Compile( ciEnv* ci_env, 898 TypeFunc_generator generator, 899 address stub_function, 900 const char *stub_name, 901 int is_fancy_jump, 902 bool pass_tls, 903 bool save_arg_registers, 904 bool return_pc ) 905 : Phase(Compiler), 906 _env(ci_env), 907 _log(ci_env->log()), 908 _compile_id(0), 909 _save_argument_registers(save_arg_registers), 910 _method(NULL), 911 _stub_name(stub_name), 912 _stub_function(stub_function), 913 _stub_entry_point(NULL), 914 _entry_bci(InvocationEntryBci), 915 _initial_gvn(NULL), 916 _for_igvn(NULL), 917 _warm_calls(NULL), 918 _orig_pc_slot(0), 919 _orig_pc_slot_offset_in_bytes(0), 920 _subsume_loads(true), 921 _do_escape_analysis(false), 922 _eliminate_boxing(false), 923 _failure_reason(NULL), 924 _code_buffer("Compile::Fill_buffer"), 925 _has_method_handle_invokes(false), 926 _mach_constant_base_node(NULL), 927 _node_bundling_limit(0), 928 _node_bundling_base(NULL), 929 _java_calls(0), 930 _inner_loops(0), 931#ifndef PRODUCT 932 _trace_opto_output(TraceOptoOutput), 933 _printer(NULL), 934#endif 935 _dead_node_list(comp_arena()), 936 _dead_node_count(0), 937 _congraph(NULL), 938 _number_of_mh_late_inlines(0), 939 _inlining_progress(false), 940 _inlining_incrementally(false), 941 _print_inlining_list(NULL), 942 _print_inlining(0) { 943 C = this; 944 945#ifndef PRODUCT 946 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false); 947 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false); 948 set_print_assembly(PrintFrameConverterAssembly); 949 set_parsed_irreducible_loop(false); 950#endif 951 CompileWrapper cw(this); 952 Init(/*AliasLevel=*/ 0); 953 init_tf((*generator)()); 954 955 { 956 // The following is a dummy for the sake of GraphKit::gen_stub 957 Unique_Node_List for_igvn(comp_arena()); 958 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 959 PhaseGVN gvn(Thread::current()->resource_area(),255); 960 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 961 gvn.transform_no_reclaim(top()); 962 963 GraphKit kit; 964 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 965 } 966 967 NOT_PRODUCT( verify_graph_edges(); ) 968 Code_Gen(); 969 if (failing()) return; 970 971 972 // Entry point will be accessed using compile->stub_entry_point(); 973 if (code_buffer() == NULL) { 974 Matcher::soft_match_failure(); 975 } else { 976 if (PrintAssembly && (WizardMode || Verbose)) 977 tty->print_cr("### Stub::%s", stub_name); 978 979 if (!failing()) { 980 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 981 982 // Make the NMethod 983 // For now we mark the frame as never safe for profile stackwalking 984 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 985 code_buffer(), 986 CodeOffsets::frame_never_safe, 987 // _code_offsets.value(CodeOffsets::Frame_Complete), 988 frame_size_in_words(), 989 _oop_map_set, 990 save_arg_registers); 991 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 992 993 _stub_entry_point = rs->entry_point(); 994 } 995 } 996} 997 998//------------------------------Init------------------------------------------- 999// Prepare for a single compilation 1000void Compile::Init(int aliaslevel) { 1001 _unique = 0; 1002 _regalloc = NULL; 1003 1004 _tf = NULL; // filled in later 1005 _top = NULL; // cached later 1006 _matcher = NULL; // filled in later 1007 _cfg = NULL; // filled in later 1008 1009 set_24_bit_selection_and_mode(Use24BitFP, false); 1010 1011 _node_note_array = NULL; 1012 _default_node_notes = NULL; 1013 1014 _immutable_memory = NULL; // filled in at first inquiry 1015 1016 // Globally visible Nodes 1017 // First set TOP to NULL to give safe behavior during creation of RootNode 1018 set_cached_top_node(NULL); 1019 set_root(new (this) RootNode()); 1020 // Now that you have a Root to point to, create the real TOP 1021 set_cached_top_node( new (this) ConNode(Type::TOP) ); 1022 set_recent_alloc(NULL, NULL); 1023 1024 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1025 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1026 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1027 env()->set_dependencies(new Dependencies(env())); 1028 1029 _fixed_slots = 0; 1030 set_has_split_ifs(false); 1031 set_has_loops(has_method() && method()->has_loops()); // first approximation 1032 set_has_stringbuilder(false); 1033 set_has_boxed_value(false); 1034 _trap_can_recompile = false; // no traps emitted yet 1035 _major_progress = true; // start out assuming good things will happen 1036 set_has_unsafe_access(false); 1037 set_max_vector_size(0); 1038 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1039 set_decompile_count(0); 1040 1041 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); 1042 set_num_loop_opts(LoopOptsCount); 1043 set_do_inlining(Inline); 1044 set_max_inline_size(MaxInlineSize); 1045 set_freq_inline_size(FreqInlineSize); 1046 set_do_scheduling(OptoScheduling); 1047 set_do_count_invocations(false); 1048 set_do_method_data_update(false); 1049 1050 if (debug_info()->recording_non_safepoints()) { 1051 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1052 (comp_arena(), 8, 0, NULL)); 1053 set_default_node_notes(Node_Notes::make(this)); 1054 } 1055 1056 // // -- Initialize types before each compile -- 1057 // // Update cached type information 1058 // if( _method && _method->constants() ) 1059 // Type::update_loaded_types(_method, _method->constants()); 1060 1061 // Init alias_type map. 1062 if (!_do_escape_analysis && aliaslevel == 3) 1063 aliaslevel = 2; // No unique types without escape analysis 1064 _AliasLevel = aliaslevel; 1065 const int grow_ats = 16; 1066 _max_alias_types = grow_ats; 1067 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1068 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1069 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1070 { 1071 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1072 } 1073 // Initialize the first few types. 1074 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1075 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1076 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1077 _num_alias_types = AliasIdxRaw+1; 1078 // Zero out the alias type cache. 1079 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1080 // A NULL adr_type hits in the cache right away. Preload the right answer. 1081 probe_alias_cache(NULL)->_index = AliasIdxTop; 1082 1083 _intrinsics = NULL; 1084 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1085 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1086 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1087 register_library_intrinsics(); 1088} 1089 1090//---------------------------init_start---------------------------------------- 1091// Install the StartNode on this compile object. 1092void Compile::init_start(StartNode* s) { 1093 if (failing()) 1094 return; // already failing 1095 assert(s == start(), ""); 1096} 1097 1098StartNode* Compile::start() const { 1099 assert(!failing(), ""); 1100 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1101 Node* start = root()->fast_out(i); 1102 if( start->is_Start() ) 1103 return start->as_Start(); 1104 } 1105 ShouldNotReachHere(); 1106 return NULL; 1107} 1108 1109//-------------------------------immutable_memory------------------------------------- 1110// Access immutable memory 1111Node* Compile::immutable_memory() { 1112 if (_immutable_memory != NULL) { 1113 return _immutable_memory; 1114 } 1115 StartNode* s = start(); 1116 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1117 Node *p = s->fast_out(i); 1118 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1119 _immutable_memory = p; 1120 return _immutable_memory; 1121 } 1122 } 1123 ShouldNotReachHere(); 1124 return NULL; 1125} 1126 1127//----------------------set_cached_top_node------------------------------------ 1128// Install the cached top node, and make sure Node::is_top works correctly. 1129void Compile::set_cached_top_node(Node* tn) { 1130 if (tn != NULL) verify_top(tn); 1131 Node* old_top = _top; 1132 _top = tn; 1133 // Calling Node::setup_is_top allows the nodes the chance to adjust 1134 // their _out arrays. 1135 if (_top != NULL) _top->setup_is_top(); 1136 if (old_top != NULL) old_top->setup_is_top(); 1137 assert(_top == NULL || top()->is_top(), ""); 1138} 1139 1140#ifdef ASSERT 1141uint Compile::count_live_nodes_by_graph_walk() { 1142 Unique_Node_List useful(comp_arena()); 1143 // Get useful node list by walking the graph. 1144 identify_useful_nodes(useful); 1145 return useful.size(); 1146} 1147 1148void Compile::print_missing_nodes() { 1149 1150 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1151 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1152 return; 1153 } 1154 1155 // This is an expensive function. It is executed only when the user 1156 // specifies VerifyIdealNodeCount option or otherwise knows the 1157 // additional work that needs to be done to identify reachable nodes 1158 // by walking the flow graph and find the missing ones using 1159 // _dead_node_list. 1160 1161 Unique_Node_List useful(comp_arena()); 1162 // Get useful node list by walking the graph. 1163 identify_useful_nodes(useful); 1164 1165 uint l_nodes = C->live_nodes(); 1166 uint l_nodes_by_walk = useful.size(); 1167 1168 if (l_nodes != l_nodes_by_walk) { 1169 if (_log != NULL) { 1170 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1171 _log->stamp(); 1172 _log->end_head(); 1173 } 1174 VectorSet& useful_member_set = useful.member_set(); 1175 int last_idx = l_nodes_by_walk; 1176 for (int i = 0; i < last_idx; i++) { 1177 if (useful_member_set.test(i)) { 1178 if (_dead_node_list.test(i)) { 1179 if (_log != NULL) { 1180 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1181 } 1182 if (PrintIdealNodeCount) { 1183 // Print the log message to tty 1184 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1185 useful.at(i)->dump(); 1186 } 1187 } 1188 } 1189 else if (! _dead_node_list.test(i)) { 1190 if (_log != NULL) { 1191 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1192 } 1193 if (PrintIdealNodeCount) { 1194 // Print the log message to tty 1195 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1196 } 1197 } 1198 } 1199 if (_log != NULL) { 1200 _log->tail("mismatched_nodes"); 1201 } 1202 } 1203} 1204#endif 1205 1206#ifndef PRODUCT 1207void Compile::verify_top(Node* tn) const { 1208 if (tn != NULL) { 1209 assert(tn->is_Con(), "top node must be a constant"); 1210 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1211 assert(tn->in(0) != NULL, "must have live top node"); 1212 } 1213} 1214#endif 1215 1216 1217///-------------------Managing Per-Node Debug & Profile Info------------------- 1218 1219void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1220 guarantee(arr != NULL, ""); 1221 int num_blocks = arr->length(); 1222 if (grow_by < num_blocks) grow_by = num_blocks; 1223 int num_notes = grow_by * _node_notes_block_size; 1224 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1225 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1226 while (num_notes > 0) { 1227 arr->append(notes); 1228 notes += _node_notes_block_size; 1229 num_notes -= _node_notes_block_size; 1230 } 1231 assert(num_notes == 0, "exact multiple, please"); 1232} 1233 1234bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1235 if (source == NULL || dest == NULL) return false; 1236 1237 if (dest->is_Con()) 1238 return false; // Do not push debug info onto constants. 1239 1240#ifdef ASSERT 1241 // Leave a bread crumb trail pointing to the original node: 1242 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1243 dest->set_debug_orig(source); 1244 } 1245#endif 1246 1247 if (node_note_array() == NULL) 1248 return false; // Not collecting any notes now. 1249 1250 // This is a copy onto a pre-existing node, which may already have notes. 1251 // If both nodes have notes, do not overwrite any pre-existing notes. 1252 Node_Notes* source_notes = node_notes_at(source->_idx); 1253 if (source_notes == NULL || source_notes->is_clear()) return false; 1254 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1255 if (dest_notes == NULL || dest_notes->is_clear()) { 1256 return set_node_notes_at(dest->_idx, source_notes); 1257 } 1258 1259 Node_Notes merged_notes = (*source_notes); 1260 // The order of operations here ensures that dest notes will win... 1261 merged_notes.update_from(dest_notes); 1262 return set_node_notes_at(dest->_idx, &merged_notes); 1263} 1264 1265 1266//--------------------------allow_range_check_smearing------------------------- 1267// Gating condition for coalescing similar range checks. 1268// Sometimes we try 'speculatively' replacing a series of a range checks by a 1269// single covering check that is at least as strong as any of them. 1270// If the optimization succeeds, the simplified (strengthened) range check 1271// will always succeed. If it fails, we will deopt, and then give up 1272// on the optimization. 1273bool Compile::allow_range_check_smearing() const { 1274 // If this method has already thrown a range-check, 1275 // assume it was because we already tried range smearing 1276 // and it failed. 1277 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1278 return !already_trapped; 1279} 1280 1281 1282//------------------------------flatten_alias_type----------------------------- 1283const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1284 int offset = tj->offset(); 1285 TypePtr::PTR ptr = tj->ptr(); 1286 1287 // Known instance (scalarizable allocation) alias only with itself. 1288 bool is_known_inst = tj->isa_oopptr() != NULL && 1289 tj->is_oopptr()->is_known_instance(); 1290 1291 // Process weird unsafe references. 1292 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1293 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1294 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1295 tj = TypeOopPtr::BOTTOM; 1296 ptr = tj->ptr(); 1297 offset = tj->offset(); 1298 } 1299 1300 // Array pointers need some flattening 1301 const TypeAryPtr *ta = tj->isa_aryptr(); 1302 if( ta && is_known_inst ) { 1303 if ( offset != Type::OffsetBot && 1304 offset > arrayOopDesc::length_offset_in_bytes() ) { 1305 offset = Type::OffsetBot; // Flatten constant access into array body only 1306 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1307 } 1308 } else if( ta && _AliasLevel >= 2 ) { 1309 // For arrays indexed by constant indices, we flatten the alias 1310 // space to include all of the array body. Only the header, klass 1311 // and array length can be accessed un-aliased. 1312 if( offset != Type::OffsetBot ) { 1313 if( ta->const_oop() ) { // MethodData* or Method* 1314 offset = Type::OffsetBot; // Flatten constant access into array body 1315 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1316 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1317 // range is OK as-is. 1318 tj = ta = TypeAryPtr::RANGE; 1319 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1320 tj = TypeInstPtr::KLASS; // all klass loads look alike 1321 ta = TypeAryPtr::RANGE; // generic ignored junk 1322 ptr = TypePtr::BotPTR; 1323 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1324 tj = TypeInstPtr::MARK; 1325 ta = TypeAryPtr::RANGE; // generic ignored junk 1326 ptr = TypePtr::BotPTR; 1327 } else { // Random constant offset into array body 1328 offset = Type::OffsetBot; // Flatten constant access into array body 1329 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1330 } 1331 } 1332 // Arrays of fixed size alias with arrays of unknown size. 1333 if (ta->size() != TypeInt::POS) { 1334 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1335 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1336 } 1337 // Arrays of known objects become arrays of unknown objects. 1338 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1339 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1340 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1341 } 1342 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1343 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1344 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1345 } 1346 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1347 // cannot be distinguished by bytecode alone. 1348 if (ta->elem() == TypeInt::BOOL) { 1349 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1350 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1351 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1352 } 1353 // During the 2nd round of IterGVN, NotNull castings are removed. 1354 // Make sure the Bottom and NotNull variants alias the same. 1355 // Also, make sure exact and non-exact variants alias the same. 1356 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) { 1357 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1358 } 1359 } 1360 1361 // Oop pointers need some flattening 1362 const TypeInstPtr *to = tj->isa_instptr(); 1363 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1364 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1365 if( ptr == TypePtr::Constant ) { 1366 if (to->klass() != ciEnv::current()->Class_klass() || 1367 offset < k->size_helper() * wordSize) { 1368 // No constant oop pointers (such as Strings); they alias with 1369 // unknown strings. 1370 assert(!is_known_inst, "not scalarizable allocation"); 1371 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1372 } 1373 } else if( is_known_inst ) { 1374 tj = to; // Keep NotNull and klass_is_exact for instance type 1375 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1376 // During the 2nd round of IterGVN, NotNull castings are removed. 1377 // Make sure the Bottom and NotNull variants alias the same. 1378 // Also, make sure exact and non-exact variants alias the same. 1379 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1380 } 1381 // Canonicalize the holder of this field 1382 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1383 // First handle header references such as a LoadKlassNode, even if the 1384 // object's klass is unloaded at compile time (4965979). 1385 if (!is_known_inst) { // Do it only for non-instance types 1386 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1387 } 1388 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1389 // Static fields are in the space above the normal instance 1390 // fields in the java.lang.Class instance. 1391 if (to->klass() != ciEnv::current()->Class_klass()) { 1392 to = NULL; 1393 tj = TypeOopPtr::BOTTOM; 1394 offset = tj->offset(); 1395 } 1396 } else { 1397 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1398 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1399 if( is_known_inst ) { 1400 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1401 } else { 1402 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1403 } 1404 } 1405 } 1406 } 1407 1408 // Klass pointers to object array klasses need some flattening 1409 const TypeKlassPtr *tk = tj->isa_klassptr(); 1410 if( tk ) { 1411 // If we are referencing a field within a Klass, we need 1412 // to assume the worst case of an Object. Both exact and 1413 // inexact types must flatten to the same alias class so 1414 // use NotNull as the PTR. 1415 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1416 1417 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1418 TypeKlassPtr::OBJECT->klass(), 1419 offset); 1420 } 1421 1422 ciKlass* klass = tk->klass(); 1423 if( klass->is_obj_array_klass() ) { 1424 ciKlass* k = TypeAryPtr::OOPS->klass(); 1425 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1426 k = TypeInstPtr::BOTTOM->klass(); 1427 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1428 } 1429 1430 // Check for precise loads from the primary supertype array and force them 1431 // to the supertype cache alias index. Check for generic array loads from 1432 // the primary supertype array and also force them to the supertype cache 1433 // alias index. Since the same load can reach both, we need to merge 1434 // these 2 disparate memories into the same alias class. Since the 1435 // primary supertype array is read-only, there's no chance of confusion 1436 // where we bypass an array load and an array store. 1437 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1438 if (offset == Type::OffsetBot || 1439 (offset >= primary_supers_offset && 1440 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1441 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1442 offset = in_bytes(Klass::secondary_super_cache_offset()); 1443 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1444 } 1445 } 1446 1447 // Flatten all Raw pointers together. 1448 if (tj->base() == Type::RawPtr) 1449 tj = TypeRawPtr::BOTTOM; 1450 1451 if (tj->base() == Type::AnyPtr) 1452 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1453 1454 // Flatten all to bottom for now 1455 switch( _AliasLevel ) { 1456 case 0: 1457 tj = TypePtr::BOTTOM; 1458 break; 1459 case 1: // Flatten to: oop, static, field or array 1460 switch (tj->base()) { 1461 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1462 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1463 case Type::AryPtr: // do not distinguish arrays at all 1464 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1465 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1466 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1467 default: ShouldNotReachHere(); 1468 } 1469 break; 1470 case 2: // No collapsing at level 2; keep all splits 1471 case 3: // No collapsing at level 3; keep all splits 1472 break; 1473 default: 1474 Unimplemented(); 1475 } 1476 1477 offset = tj->offset(); 1478 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1479 1480 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1481 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1482 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1483 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1484 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1485 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1486 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1487 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1488 assert( tj->ptr() != TypePtr::TopPTR && 1489 tj->ptr() != TypePtr::AnyNull && 1490 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1491// assert( tj->ptr() != TypePtr::Constant || 1492// tj->base() == Type::RawPtr || 1493// tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1494 1495 return tj; 1496} 1497 1498void Compile::AliasType::Init(int i, const TypePtr* at) { 1499 _index = i; 1500 _adr_type = at; 1501 _field = NULL; 1502 _is_rewritable = true; // default 1503 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1504 if (atoop != NULL && atoop->is_known_instance()) { 1505 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1506 _general_index = Compile::current()->get_alias_index(gt); 1507 } else { 1508 _general_index = 0; 1509 } 1510} 1511 1512//---------------------------------print_on------------------------------------ 1513#ifndef PRODUCT 1514void Compile::AliasType::print_on(outputStream* st) { 1515 if (index() < 10) 1516 st->print("@ <%d> ", index()); 1517 else st->print("@ <%d>", index()); 1518 st->print(is_rewritable() ? " " : " RO"); 1519 int offset = adr_type()->offset(); 1520 if (offset == Type::OffsetBot) 1521 st->print(" +any"); 1522 else st->print(" +%-3d", offset); 1523 st->print(" in "); 1524 adr_type()->dump_on(st); 1525 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1526 if (field() != NULL && tjp) { 1527 if (tjp->klass() != field()->holder() || 1528 tjp->offset() != field()->offset_in_bytes()) { 1529 st->print(" != "); 1530 field()->print(); 1531 st->print(" ***"); 1532 } 1533 } 1534} 1535 1536void print_alias_types() { 1537 Compile* C = Compile::current(); 1538 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1539 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1540 C->alias_type(idx)->print_on(tty); 1541 tty->cr(); 1542 } 1543} 1544#endif 1545 1546 1547//----------------------------probe_alias_cache-------------------------------- 1548Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1549 intptr_t key = (intptr_t) adr_type; 1550 key ^= key >> logAliasCacheSize; 1551 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1552} 1553 1554 1555//-----------------------------grow_alias_types-------------------------------- 1556void Compile::grow_alias_types() { 1557 const int old_ats = _max_alias_types; // how many before? 1558 const int new_ats = old_ats; // how many more? 1559 const int grow_ats = old_ats+new_ats; // how many now? 1560 _max_alias_types = grow_ats; 1561 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1562 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1563 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1564 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1565} 1566 1567 1568//--------------------------------find_alias_type------------------------------ 1569Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1570 if (_AliasLevel == 0) 1571 return alias_type(AliasIdxBot); 1572 1573 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1574 if (ace->_adr_type == adr_type) { 1575 return alias_type(ace->_index); 1576 } 1577 1578 // Handle special cases. 1579 if (adr_type == NULL) return alias_type(AliasIdxTop); 1580 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1581 1582 // Do it the slow way. 1583 const TypePtr* flat = flatten_alias_type(adr_type); 1584 1585#ifdef ASSERT 1586 assert(flat == flatten_alias_type(flat), "idempotent"); 1587 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1588 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1589 const TypeOopPtr* foop = flat->is_oopptr(); 1590 // Scalarizable allocations have exact klass always. 1591 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1592 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1593 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1594 } 1595 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1596#endif 1597 1598 int idx = AliasIdxTop; 1599 for (int i = 0; i < num_alias_types(); i++) { 1600 if (alias_type(i)->adr_type() == flat) { 1601 idx = i; 1602 break; 1603 } 1604 } 1605 1606 if (idx == AliasIdxTop) { 1607 if (no_create) return NULL; 1608 // Grow the array if necessary. 1609 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1610 // Add a new alias type. 1611 idx = _num_alias_types++; 1612 _alias_types[idx]->Init(idx, flat); 1613 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1614 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1615 if (flat->isa_instptr()) { 1616 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1617 && flat->is_instptr()->klass() == env()->Class_klass()) 1618 alias_type(idx)->set_rewritable(false); 1619 } 1620 if (flat->isa_klassptr()) { 1621 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1622 alias_type(idx)->set_rewritable(false); 1623 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1624 alias_type(idx)->set_rewritable(false); 1625 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1626 alias_type(idx)->set_rewritable(false); 1627 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1628 alias_type(idx)->set_rewritable(false); 1629 } 1630 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1631 // but the base pointer type is not distinctive enough to identify 1632 // references into JavaThread.) 1633 1634 // Check for final fields. 1635 const TypeInstPtr* tinst = flat->isa_instptr(); 1636 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1637 ciField* field; 1638 if (tinst->const_oop() != NULL && 1639 tinst->klass() == ciEnv::current()->Class_klass() && 1640 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1641 // static field 1642 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1643 field = k->get_field_by_offset(tinst->offset(), true); 1644 } else { 1645 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1646 field = k->get_field_by_offset(tinst->offset(), false); 1647 } 1648 assert(field == NULL || 1649 original_field == NULL || 1650 (field->holder() == original_field->holder() && 1651 field->offset() == original_field->offset() && 1652 field->is_static() == original_field->is_static()), "wrong field?"); 1653 // Set field() and is_rewritable() attributes. 1654 if (field != NULL) alias_type(idx)->set_field(field); 1655 } 1656 } 1657 1658 // Fill the cache for next time. 1659 ace->_adr_type = adr_type; 1660 ace->_index = idx; 1661 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1662 1663 // Might as well try to fill the cache for the flattened version, too. 1664 AliasCacheEntry* face = probe_alias_cache(flat); 1665 if (face->_adr_type == NULL) { 1666 face->_adr_type = flat; 1667 face->_index = idx; 1668 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1669 } 1670 1671 return alias_type(idx); 1672} 1673 1674 1675Compile::AliasType* Compile::alias_type(ciField* field) { 1676 const TypeOopPtr* t; 1677 if (field->is_static()) 1678 t = TypeInstPtr::make(field->holder()->java_mirror()); 1679 else 1680 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1681 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1682 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct"); 1683 return atp; 1684} 1685 1686 1687//------------------------------have_alias_type-------------------------------- 1688bool Compile::have_alias_type(const TypePtr* adr_type) { 1689 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1690 if (ace->_adr_type == adr_type) { 1691 return true; 1692 } 1693 1694 // Handle special cases. 1695 if (adr_type == NULL) return true; 1696 if (adr_type == TypePtr::BOTTOM) return true; 1697 1698 return find_alias_type(adr_type, true, NULL) != NULL; 1699} 1700 1701//-----------------------------must_alias-------------------------------------- 1702// True if all values of the given address type are in the given alias category. 1703bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1704 if (alias_idx == AliasIdxBot) return true; // the universal category 1705 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1706 if (alias_idx == AliasIdxTop) return false; // the empty category 1707 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1708 1709 // the only remaining possible overlap is identity 1710 int adr_idx = get_alias_index(adr_type); 1711 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1712 assert(adr_idx == alias_idx || 1713 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1714 && adr_type != TypeOopPtr::BOTTOM), 1715 "should not be testing for overlap with an unsafe pointer"); 1716 return adr_idx == alias_idx; 1717} 1718 1719//------------------------------can_alias-------------------------------------- 1720// True if any values of the given address type are in the given alias category. 1721bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1722 if (alias_idx == AliasIdxTop) return false; // the empty category 1723 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1724 if (alias_idx == AliasIdxBot) return true; // the universal category 1725 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1726 1727 // the only remaining possible overlap is identity 1728 int adr_idx = get_alias_index(adr_type); 1729 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1730 return adr_idx == alias_idx; 1731} 1732 1733 1734 1735//---------------------------pop_warm_call------------------------------------- 1736WarmCallInfo* Compile::pop_warm_call() { 1737 WarmCallInfo* wci = _warm_calls; 1738 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1739 return wci; 1740} 1741 1742//----------------------------Inline_Warm-------------------------------------- 1743int Compile::Inline_Warm() { 1744 // If there is room, try to inline some more warm call sites. 1745 // %%% Do a graph index compaction pass when we think we're out of space? 1746 if (!InlineWarmCalls) return 0; 1747 1748 int calls_made_hot = 0; 1749 int room_to_grow = NodeCountInliningCutoff - unique(); 1750 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1751 int amount_grown = 0; 1752 WarmCallInfo* call; 1753 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1754 int est_size = (int)call->size(); 1755 if (est_size > (room_to_grow - amount_grown)) { 1756 // This one won't fit anyway. Get rid of it. 1757 call->make_cold(); 1758 continue; 1759 } 1760 call->make_hot(); 1761 calls_made_hot++; 1762 amount_grown += est_size; 1763 amount_to_grow -= est_size; 1764 } 1765 1766 if (calls_made_hot > 0) set_major_progress(); 1767 return calls_made_hot; 1768} 1769 1770 1771//----------------------------Finish_Warm-------------------------------------- 1772void Compile::Finish_Warm() { 1773 if (!InlineWarmCalls) return; 1774 if (failing()) return; 1775 if (warm_calls() == NULL) return; 1776 1777 // Clean up loose ends, if we are out of space for inlining. 1778 WarmCallInfo* call; 1779 while ((call = pop_warm_call()) != NULL) { 1780 call->make_cold(); 1781 } 1782} 1783 1784//---------------------cleanup_loop_predicates----------------------- 1785// Remove the opaque nodes that protect the predicates so that all unused 1786// checks and uncommon_traps will be eliminated from the ideal graph 1787void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1788 if (predicate_count()==0) return; 1789 for (int i = predicate_count(); i > 0; i--) { 1790 Node * n = predicate_opaque1_node(i-1); 1791 assert(n->Opcode() == Op_Opaque1, "must be"); 1792 igvn.replace_node(n, n->in(1)); 1793 } 1794 assert(predicate_count()==0, "should be clean!"); 1795} 1796 1797// StringOpts and late inlining of string methods 1798void Compile::inline_string_calls(bool parse_time) { 1799 { 1800 // remove useless nodes to make the usage analysis simpler 1801 ResourceMark rm; 1802 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1803 } 1804 1805 { 1806 ResourceMark rm; 1807 print_method("Before StringOpts", 3); 1808 PhaseStringOpts pso(initial_gvn(), for_igvn()); 1809 print_method("After StringOpts", 3); 1810 } 1811 1812 // now inline anything that we skipped the first time around 1813 if (!parse_time) { 1814 _late_inlines_pos = _late_inlines.length(); 1815 } 1816 1817 while (_string_late_inlines.length() > 0) { 1818 CallGenerator* cg = _string_late_inlines.pop(); 1819 cg->do_late_inline(); 1820 if (failing()) return; 1821 } 1822 _string_late_inlines.trunc_to(0); 1823} 1824 1825// Late inlining of boxing methods 1826void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 1827 if (_boxing_late_inlines.length() > 0) { 1828 assert(has_boxed_value(), "inconsistent"); 1829 1830 PhaseGVN* gvn = initial_gvn(); 1831 set_inlining_incrementally(true); 1832 1833 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1834 for_igvn()->clear(); 1835 gvn->replace_with(&igvn); 1836 1837 while (_boxing_late_inlines.length() > 0) { 1838 CallGenerator* cg = _boxing_late_inlines.pop(); 1839 cg->do_late_inline(); 1840 if (failing()) return; 1841 } 1842 _boxing_late_inlines.trunc_to(0); 1843 1844 { 1845 ResourceMark rm; 1846 PhaseRemoveUseless pru(gvn, for_igvn()); 1847 } 1848 1849 igvn = PhaseIterGVN(gvn); 1850 igvn.optimize(); 1851 1852 set_inlining_progress(false); 1853 set_inlining_incrementally(false); 1854 } 1855} 1856 1857void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { 1858 assert(IncrementalInline, "incremental inlining should be on"); 1859 PhaseGVN* gvn = initial_gvn(); 1860 1861 set_inlining_progress(false); 1862 for_igvn()->clear(); 1863 gvn->replace_with(&igvn); 1864 1865 int i = 0; 1866 1867 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 1868 CallGenerator* cg = _late_inlines.at(i); 1869 _late_inlines_pos = i+1; 1870 cg->do_late_inline(); 1871 if (failing()) return; 1872 } 1873 int j = 0; 1874 for (; i < _late_inlines.length(); i++, j++) { 1875 _late_inlines.at_put(j, _late_inlines.at(i)); 1876 } 1877 _late_inlines.trunc_to(j); 1878 1879 { 1880 ResourceMark rm; 1881 PhaseRemoveUseless pru(gvn, for_igvn()); 1882 } 1883 1884 igvn = PhaseIterGVN(gvn); 1885} 1886 1887// Perform incremental inlining until bound on number of live nodes is reached 1888void Compile::inline_incrementally(PhaseIterGVN& igvn) { 1889 PhaseGVN* gvn = initial_gvn(); 1890 1891 set_inlining_incrementally(true); 1892 set_inlining_progress(true); 1893 uint low_live_nodes = 0; 1894 1895 while(inlining_progress() && _late_inlines.length() > 0) { 1896 1897 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 1898 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 1899 // PhaseIdealLoop is expensive so we only try it once we are 1900 // out of loop and we only try it again if the previous helped 1901 // got the number of nodes down significantly 1902 PhaseIdealLoop ideal_loop( igvn, false, true ); 1903 if (failing()) return; 1904 low_live_nodes = live_nodes(); 1905 _major_progress = true; 1906 } 1907 1908 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 1909 break; 1910 } 1911 } 1912 1913 inline_incrementally_one(igvn); 1914 1915 if (failing()) return; 1916 1917 igvn.optimize(); 1918 1919 if (failing()) return; 1920 } 1921 1922 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1923 1924 if (_string_late_inlines.length() > 0) { 1925 assert(has_stringbuilder(), "inconsistent"); 1926 for_igvn()->clear(); 1927 initial_gvn()->replace_with(&igvn); 1928 1929 inline_string_calls(false); 1930 1931 if (failing()) return; 1932 1933 { 1934 ResourceMark rm; 1935 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1936 } 1937 1938 igvn = PhaseIterGVN(gvn); 1939 1940 igvn.optimize(); 1941 } 1942 1943 set_inlining_incrementally(false); 1944} 1945 1946 1947//------------------------------Optimize--------------------------------------- 1948// Given a graph, optimize it. 1949void Compile::Optimize() { 1950 TracePhase t1("optimizer", &_t_optimizer, true); 1951 1952#ifndef PRODUCT 1953 if (env()->break_at_compile()) { 1954 BREAKPOINT; 1955 } 1956 1957#endif 1958 1959 ResourceMark rm; 1960 int loop_opts_cnt; 1961 1962 NOT_PRODUCT( verify_graph_edges(); ) 1963 1964 print_method("After Parsing"); 1965 1966 { 1967 // Iterative Global Value Numbering, including ideal transforms 1968 // Initialize IterGVN with types and values from parse-time GVN 1969 PhaseIterGVN igvn(initial_gvn()); 1970 { 1971 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); ) 1972 igvn.optimize(); 1973 } 1974 1975 print_method("Iter GVN 1", 2); 1976 1977 if (failing()) return; 1978 1979 { 1980 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); ) 1981 inline_incrementally(igvn); 1982 } 1983 1984 print_method("Incremental Inline", 2); 1985 1986 if (failing()) return; 1987 1988 if (eliminate_boxing()) { 1989 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); ) 1990 // Inline valueOf() methods now. 1991 inline_boxing_calls(igvn); 1992 1993 print_method("Incremental Boxing Inline", 2); 1994 1995 if (failing()) return; 1996 } 1997 1998 // No more new expensive nodes will be added to the list from here 1999 // so keep only the actual candidates for optimizations. 2000 cleanup_expensive_nodes(igvn); 2001 2002 // Perform escape analysis 2003 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2004 if (has_loops()) { 2005 // Cleanup graph (remove dead nodes). 2006 TracePhase t2("idealLoop", &_t_idealLoop, true); 2007 PhaseIdealLoop ideal_loop( igvn, false, true ); 2008 if (major_progress()) print_method("PhaseIdealLoop before EA", 2); 2009 if (failing()) return; 2010 } 2011 ConnectionGraph::do_analysis(this, &igvn); 2012 2013 if (failing()) return; 2014 2015 // Optimize out fields loads from scalar replaceable allocations. 2016 igvn.optimize(); 2017 print_method("Iter GVN after EA", 2); 2018 2019 if (failing()) return; 2020 2021 if (congraph() != NULL && macro_count() > 0) { 2022 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); ) 2023 PhaseMacroExpand mexp(igvn); 2024 mexp.eliminate_macro_nodes(); 2025 igvn.set_delay_transform(false); 2026 2027 igvn.optimize(); 2028 print_method("Iter GVN after eliminating allocations and locks", 2); 2029 2030 if (failing()) return; 2031 } 2032 } 2033 2034 // Loop transforms on the ideal graph. Range Check Elimination, 2035 // peeling, unrolling, etc. 2036 2037 // Set loop opts counter 2038 loop_opts_cnt = num_loop_opts(); 2039 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2040 { 2041 TracePhase t2("idealLoop", &_t_idealLoop, true); 2042 PhaseIdealLoop ideal_loop( igvn, true ); 2043 loop_opts_cnt--; 2044 if (major_progress()) print_method("PhaseIdealLoop 1", 2); 2045 if (failing()) return; 2046 } 2047 // Loop opts pass if partial peeling occurred in previous pass 2048 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 2049 TracePhase t3("idealLoop", &_t_idealLoop, true); 2050 PhaseIdealLoop ideal_loop( igvn, false ); 2051 loop_opts_cnt--; 2052 if (major_progress()) print_method("PhaseIdealLoop 2", 2); 2053 if (failing()) return; 2054 } 2055 // Loop opts pass for loop-unrolling before CCP 2056 if(major_progress() && (loop_opts_cnt > 0)) { 2057 TracePhase t4("idealLoop", &_t_idealLoop, true); 2058 PhaseIdealLoop ideal_loop( igvn, false ); 2059 loop_opts_cnt--; 2060 if (major_progress()) print_method("PhaseIdealLoop 3", 2); 2061 } 2062 if (!failing()) { 2063 // Verify that last round of loop opts produced a valid graph 2064 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 2065 PhaseIdealLoop::verify(igvn); 2066 } 2067 } 2068 if (failing()) return; 2069 2070 // Conditional Constant Propagation; 2071 PhaseCCP ccp( &igvn ); 2072 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2073 { 2074 TracePhase t2("ccp", &_t_ccp, true); 2075 ccp.do_transform(); 2076 } 2077 print_method("PhaseCPP 1", 2); 2078 2079 assert( true, "Break here to ccp.dump_old2new_map()"); 2080 2081 // Iterative Global Value Numbering, including ideal transforms 2082 { 2083 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); ) 2084 igvn = ccp; 2085 igvn.optimize(); 2086 } 2087 2088 print_method("Iter GVN 2", 2); 2089 2090 if (failing()) return; 2091 2092 // Loop transforms on the ideal graph. Range Check Elimination, 2093 // peeling, unrolling, etc. 2094 if(loop_opts_cnt > 0) { 2095 debug_only( int cnt = 0; ); 2096 while(major_progress() && (loop_opts_cnt > 0)) { 2097 TracePhase t2("idealLoop", &_t_idealLoop, true); 2098 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2099 PhaseIdealLoop ideal_loop( igvn, true); 2100 loop_opts_cnt--; 2101 if (major_progress()) print_method("PhaseIdealLoop iterations", 2); 2102 if (failing()) return; 2103 } 2104 } 2105 2106 { 2107 // Verify that all previous optimizations produced a valid graph 2108 // at least to this point, even if no loop optimizations were done. 2109 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 2110 PhaseIdealLoop::verify(igvn); 2111 } 2112 2113 { 2114 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); ) 2115 PhaseMacroExpand mex(igvn); 2116 if (mex.expand_macro_nodes()) { 2117 assert(failing(), "must bail out w/ explicit message"); 2118 return; 2119 } 2120 } 2121 2122 } // (End scope of igvn; run destructor if necessary for asserts.) 2123 2124 dump_inlining(); 2125 // A method with only infinite loops has no edges entering loops from root 2126 { 2127 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); ) 2128 if (final_graph_reshaping()) { 2129 assert(failing(), "must bail out w/ explicit message"); 2130 return; 2131 } 2132 } 2133 2134 print_method("Optimize finished", 2); 2135} 2136 2137 2138//------------------------------Code_Gen--------------------------------------- 2139// Given a graph, generate code for it 2140void Compile::Code_Gen() { 2141 if (failing()) return; 2142 2143 // Perform instruction selection. You might think we could reclaim Matcher 2144 // memory PDQ, but actually the Matcher is used in generating spill code. 2145 // Internals of the Matcher (including some VectorSets) must remain live 2146 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2147 // set a bit in reclaimed memory. 2148 2149 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2150 // nodes. Mapping is only valid at the root of each matched subtree. 2151 NOT_PRODUCT( verify_graph_edges(); ) 2152 2153 Node_List proj_list; 2154 Matcher m(proj_list); 2155 _matcher = &m; 2156 { 2157 TracePhase t2("matcher", &_t_matcher, true); 2158 m.match(); 2159 } 2160 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2161 // nodes. Mapping is only valid at the root of each matched subtree. 2162 NOT_PRODUCT( verify_graph_edges(); ) 2163 2164 // If you have too many nodes, or if matching has failed, bail out 2165 check_node_count(0, "out of nodes matching instructions"); 2166 if (failing()) return; 2167 2168 // Build a proper-looking CFG 2169 PhaseCFG cfg(node_arena(), root(), m); 2170 _cfg = &cfg; 2171 { 2172 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); ) 2173 cfg.Dominators(); 2174 if (failing()) return; 2175 2176 NOT_PRODUCT( verify_graph_edges(); ) 2177 2178 cfg.Estimate_Block_Frequency(); 2179 cfg.GlobalCodeMotion(m,unique(),proj_list); 2180 if (failing()) return; 2181 2182 print_method("Global code motion", 2); 2183 2184 NOT_PRODUCT( verify_graph_edges(); ) 2185 2186 debug_only( cfg.verify(); ) 2187 } 2188 NOT_PRODUCT( verify_graph_edges(); ) 2189 2190 PhaseChaitin regalloc(unique(), cfg, m); 2191 _regalloc = ®alloc; 2192 { 2193 TracePhase t2("regalloc", &_t_registerAllocation, true); 2194 // Perform register allocation. After Chaitin, use-def chains are 2195 // no longer accurate (at spill code) and so must be ignored. 2196 // Node->LRG->reg mappings are still accurate. 2197 _regalloc->Register_Allocate(); 2198 2199 // Bail out if the allocator builds too many nodes 2200 if (failing()) { 2201 return; 2202 } 2203 } 2204 2205 // Prior to register allocation we kept empty basic blocks in case the 2206 // the allocator needed a place to spill. After register allocation we 2207 // are not adding any new instructions. If any basic block is empty, we 2208 // can now safely remove it. 2209 { 2210 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); ) 2211 cfg.remove_empty(); 2212 if (do_freq_based_layout()) { 2213 PhaseBlockLayout layout(cfg); 2214 } else { 2215 cfg.set_loop_alignment(); 2216 } 2217 cfg.fixup_flow(); 2218 } 2219 2220 // Apply peephole optimizations 2221 if( OptoPeephole ) { 2222 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); ) 2223 PhasePeephole peep( _regalloc, cfg); 2224 peep.do_transform(); 2225 } 2226 2227 // Convert Nodes to instruction bits in a buffer 2228 { 2229 // %%%% workspace merge brought two timers together for one job 2230 TracePhase t2a("output", &_t_output, true); 2231 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); ) 2232 Output(); 2233 } 2234 2235 print_method("Final Code"); 2236 2237 // He's dead, Jim. 2238 _cfg = (PhaseCFG*)0xdeadbeef; 2239 _regalloc = (PhaseChaitin*)0xdeadbeef; 2240} 2241 2242 2243//------------------------------dump_asm--------------------------------------- 2244// Dump formatted assembly 2245#ifndef PRODUCT 2246void Compile::dump_asm(int *pcs, uint pc_limit) { 2247 bool cut_short = false; 2248 tty->print_cr("#"); 2249 tty->print("# "); _tf->dump(); tty->cr(); 2250 tty->print_cr("#"); 2251 2252 // For all blocks 2253 int pc = 0x0; // Program counter 2254 char starts_bundle = ' '; 2255 _regalloc->dump_frame(); 2256 2257 Node *n = NULL; 2258 for( uint i=0; i<_cfg->_num_blocks; i++ ) { 2259 if (VMThread::should_terminate()) { cut_short = true; break; } 2260 Block *b = _cfg->_blocks[i]; 2261 if (b->is_connector() && !Verbose) continue; 2262 n = b->_nodes[0]; 2263 if (pcs && n->_idx < pc_limit) 2264 tty->print("%3.3x ", pcs[n->_idx]); 2265 else 2266 tty->print(" "); 2267 b->dump_head( &_cfg->_bbs ); 2268 if (b->is_connector()) { 2269 tty->print_cr(" # Empty connector block"); 2270 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2271 tty->print_cr(" # Block is sole successor of call"); 2272 } 2273 2274 // For all instructions 2275 Node *delay = NULL; 2276 for( uint j = 0; j<b->_nodes.size(); j++ ) { 2277 if (VMThread::should_terminate()) { cut_short = true; break; } 2278 n = b->_nodes[j]; 2279 if (valid_bundle_info(n)) { 2280 Bundle *bundle = node_bundling(n); 2281 if (bundle->used_in_unconditional_delay()) { 2282 delay = n; 2283 continue; 2284 } 2285 if (bundle->starts_bundle()) 2286 starts_bundle = '+'; 2287 } 2288 2289 if (WizardMode) n->dump(); 2290 2291 if( !n->is_Region() && // Dont print in the Assembly 2292 !n->is_Phi() && // a few noisely useless nodes 2293 !n->is_Proj() && 2294 !n->is_MachTemp() && 2295 !n->is_SafePointScalarObject() && 2296 !n->is_Catch() && // Would be nice to print exception table targets 2297 !n->is_MergeMem() && // Not very interesting 2298 !n->is_top() && // Debug info table constants 2299 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2300 ) { 2301 if (pcs && n->_idx < pc_limit) 2302 tty->print("%3.3x", pcs[n->_idx]); 2303 else 2304 tty->print(" "); 2305 tty->print(" %c ", starts_bundle); 2306 starts_bundle = ' '; 2307 tty->print("\t"); 2308 n->format(_regalloc, tty); 2309 tty->cr(); 2310 } 2311 2312 // If we have an instruction with a delay slot, and have seen a delay, 2313 // then back up and print it 2314 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2315 assert(delay != NULL, "no unconditional delay instruction"); 2316 if (WizardMode) delay->dump(); 2317 2318 if (node_bundling(delay)->starts_bundle()) 2319 starts_bundle = '+'; 2320 if (pcs && n->_idx < pc_limit) 2321 tty->print("%3.3x", pcs[n->_idx]); 2322 else 2323 tty->print(" "); 2324 tty->print(" %c ", starts_bundle); 2325 starts_bundle = ' '; 2326 tty->print("\t"); 2327 delay->format(_regalloc, tty); 2328 tty->print_cr(""); 2329 delay = NULL; 2330 } 2331 2332 // Dump the exception table as well 2333 if( n->is_Catch() && (Verbose || WizardMode) ) { 2334 // Print the exception table for this offset 2335 _handler_table.print_subtable_for(pc); 2336 } 2337 } 2338 2339 if (pcs && n->_idx < pc_limit) 2340 tty->print_cr("%3.3x", pcs[n->_idx]); 2341 else 2342 tty->print_cr(""); 2343 2344 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2345 2346 } // End of per-block dump 2347 tty->print_cr(""); 2348 2349 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2350} 2351#endif 2352 2353//------------------------------Final_Reshape_Counts--------------------------- 2354// This class defines counters to help identify when a method 2355// may/must be executed using hardware with only 24-bit precision. 2356struct Final_Reshape_Counts : public StackObj { 2357 int _call_count; // count non-inlined 'common' calls 2358 int _float_count; // count float ops requiring 24-bit precision 2359 int _double_count; // count double ops requiring more precision 2360 int _java_call_count; // count non-inlined 'java' calls 2361 int _inner_loop_count; // count loops which need alignment 2362 VectorSet _visited; // Visitation flags 2363 Node_List _tests; // Set of IfNodes & PCTableNodes 2364 2365 Final_Reshape_Counts() : 2366 _call_count(0), _float_count(0), _double_count(0), 2367 _java_call_count(0), _inner_loop_count(0), 2368 _visited( Thread::current()->resource_area() ) { } 2369 2370 void inc_call_count () { _call_count ++; } 2371 void inc_float_count () { _float_count ++; } 2372 void inc_double_count() { _double_count++; } 2373 void inc_java_call_count() { _java_call_count++; } 2374 void inc_inner_loop_count() { _inner_loop_count++; } 2375 2376 int get_call_count () const { return _call_count ; } 2377 int get_float_count () const { return _float_count ; } 2378 int get_double_count() const { return _double_count; } 2379 int get_java_call_count() const { return _java_call_count; } 2380 int get_inner_loop_count() const { return _inner_loop_count; } 2381}; 2382 2383#ifdef ASSERT 2384static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2385 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2386 // Make sure the offset goes inside the instance layout. 2387 return k->contains_field_offset(tp->offset()); 2388 // Note that OffsetBot and OffsetTop are very negative. 2389} 2390#endif 2391 2392// Eliminate trivially redundant StoreCMs and accumulate their 2393// precedence edges. 2394void Compile::eliminate_redundant_card_marks(Node* n) { 2395 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2396 if (n->in(MemNode::Address)->outcnt() > 1) { 2397 // There are multiple users of the same address so it might be 2398 // possible to eliminate some of the StoreCMs 2399 Node* mem = n->in(MemNode::Memory); 2400 Node* adr = n->in(MemNode::Address); 2401 Node* val = n->in(MemNode::ValueIn); 2402 Node* prev = n; 2403 bool done = false; 2404 // Walk the chain of StoreCMs eliminating ones that match. As 2405 // long as it's a chain of single users then the optimization is 2406 // safe. Eliminating partially redundant StoreCMs would require 2407 // cloning copies down the other paths. 2408 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2409 if (adr == mem->in(MemNode::Address) && 2410 val == mem->in(MemNode::ValueIn)) { 2411 // redundant StoreCM 2412 if (mem->req() > MemNode::OopStore) { 2413 // Hasn't been processed by this code yet. 2414 n->add_prec(mem->in(MemNode::OopStore)); 2415 } else { 2416 // Already converted to precedence edge 2417 for (uint i = mem->req(); i < mem->len(); i++) { 2418 // Accumulate any precedence edges 2419 if (mem->in(i) != NULL) { 2420 n->add_prec(mem->in(i)); 2421 } 2422 } 2423 // Everything above this point has been processed. 2424 done = true; 2425 } 2426 // Eliminate the previous StoreCM 2427 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2428 assert(mem->outcnt() == 0, "should be dead"); 2429 mem->disconnect_inputs(NULL, this); 2430 } else { 2431 prev = mem; 2432 } 2433 mem = prev->in(MemNode::Memory); 2434 } 2435 } 2436} 2437 2438//------------------------------final_graph_reshaping_impl---------------------- 2439// Implement items 1-5 from final_graph_reshaping below. 2440void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2441 2442 if ( n->outcnt() == 0 ) return; // dead node 2443 uint nop = n->Opcode(); 2444 2445 // Check for 2-input instruction with "last use" on right input. 2446 // Swap to left input. Implements item (2). 2447 if( n->req() == 3 && // two-input instruction 2448 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2449 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2450 n->in(2)->outcnt() == 1 &&// right use IS a last use 2451 !n->in(2)->is_Con() ) { // right use is not a constant 2452 // Check for commutative opcode 2453 switch( nop ) { 2454 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2455 case Op_MaxI: case Op_MinI: 2456 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2457 case Op_AndL: case Op_XorL: case Op_OrL: 2458 case Op_AndI: case Op_XorI: case Op_OrI: { 2459 // Move "last use" input to left by swapping inputs 2460 n->swap_edges(1, 2); 2461 break; 2462 } 2463 default: 2464 break; 2465 } 2466 } 2467 2468#ifdef ASSERT 2469 if( n->is_Mem() ) { 2470 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2471 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2472 // oop will be recorded in oop map if load crosses safepoint 2473 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2474 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2475 "raw memory operations should have control edge"); 2476 } 2477#endif 2478 // Count FPU ops and common calls, implements item (3) 2479 switch( nop ) { 2480 // Count all float operations that may use FPU 2481 case Op_AddF: 2482 case Op_SubF: 2483 case Op_MulF: 2484 case Op_DivF: 2485 case Op_NegF: 2486 case Op_ModF: 2487 case Op_ConvI2F: 2488 case Op_ConF: 2489 case Op_CmpF: 2490 case Op_CmpF3: 2491 // case Op_ConvL2F: // longs are split into 32-bit halves 2492 frc.inc_float_count(); 2493 break; 2494 2495 case Op_ConvF2D: 2496 case Op_ConvD2F: 2497 frc.inc_float_count(); 2498 frc.inc_double_count(); 2499 break; 2500 2501 // Count all double operations that may use FPU 2502 case Op_AddD: 2503 case Op_SubD: 2504 case Op_MulD: 2505 case Op_DivD: 2506 case Op_NegD: 2507 case Op_ModD: 2508 case Op_ConvI2D: 2509 case Op_ConvD2I: 2510 // case Op_ConvL2D: // handled by leaf call 2511 // case Op_ConvD2L: // handled by leaf call 2512 case Op_ConD: 2513 case Op_CmpD: 2514 case Op_CmpD3: 2515 frc.inc_double_count(); 2516 break; 2517 case Op_Opaque1: // Remove Opaque Nodes before matching 2518 case Op_Opaque2: // Remove Opaque Nodes before matching 2519 n->subsume_by(n->in(1), this); 2520 break; 2521 case Op_CallStaticJava: 2522 case Op_CallJava: 2523 case Op_CallDynamicJava: 2524 frc.inc_java_call_count(); // Count java call site; 2525 case Op_CallRuntime: 2526 case Op_CallLeaf: 2527 case Op_CallLeafNoFP: { 2528 assert( n->is_Call(), "" ); 2529 CallNode *call = n->as_Call(); 2530 // Count call sites where the FP mode bit would have to be flipped. 2531 // Do not count uncommon runtime calls: 2532 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2533 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2534 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2535 frc.inc_call_count(); // Count the call site 2536 } else { // See if uncommon argument is shared 2537 Node *n = call->in(TypeFunc::Parms); 2538 int nop = n->Opcode(); 2539 // Clone shared simple arguments to uncommon calls, item (1). 2540 if( n->outcnt() > 1 && 2541 !n->is_Proj() && 2542 nop != Op_CreateEx && 2543 nop != Op_CheckCastPP && 2544 nop != Op_DecodeN && 2545 nop != Op_DecodeNKlass && 2546 !n->is_Mem() ) { 2547 Node *x = n->clone(); 2548 call->set_req( TypeFunc::Parms, x ); 2549 } 2550 } 2551 break; 2552 } 2553 2554 case Op_StoreD: 2555 case Op_LoadD: 2556 case Op_LoadD_unaligned: 2557 frc.inc_double_count(); 2558 goto handle_mem; 2559 case Op_StoreF: 2560 case Op_LoadF: 2561 frc.inc_float_count(); 2562 goto handle_mem; 2563 2564 case Op_StoreCM: 2565 { 2566 // Convert OopStore dependence into precedence edge 2567 Node* prec = n->in(MemNode::OopStore); 2568 n->del_req(MemNode::OopStore); 2569 n->add_prec(prec); 2570 eliminate_redundant_card_marks(n); 2571 } 2572 2573 // fall through 2574 2575 case Op_StoreB: 2576 case Op_StoreC: 2577 case Op_StorePConditional: 2578 case Op_StoreI: 2579 case Op_StoreL: 2580 case Op_StoreIConditional: 2581 case Op_StoreLConditional: 2582 case Op_CompareAndSwapI: 2583 case Op_CompareAndSwapL: 2584 case Op_CompareAndSwapP: 2585 case Op_CompareAndSwapN: 2586 case Op_GetAndAddI: 2587 case Op_GetAndAddL: 2588 case Op_GetAndSetI: 2589 case Op_GetAndSetL: 2590 case Op_GetAndSetP: 2591 case Op_GetAndSetN: 2592 case Op_StoreP: 2593 case Op_StoreN: 2594 case Op_StoreNKlass: 2595 case Op_LoadB: 2596 case Op_LoadUB: 2597 case Op_LoadUS: 2598 case Op_LoadI: 2599 case Op_LoadKlass: 2600 case Op_LoadNKlass: 2601 case Op_LoadL: 2602 case Op_LoadL_unaligned: 2603 case Op_LoadPLocked: 2604 case Op_LoadP: 2605 case Op_LoadN: 2606 case Op_LoadRange: 2607 case Op_LoadS: { 2608 handle_mem: 2609#ifdef ASSERT 2610 if( VerifyOptoOopOffsets ) { 2611 assert( n->is_Mem(), "" ); 2612 MemNode *mem = (MemNode*)n; 2613 // Check to see if address types have grounded out somehow. 2614 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2615 assert( !tp || oop_offset_is_sane(tp), "" ); 2616 } 2617#endif 2618 break; 2619 } 2620 2621 case Op_AddP: { // Assert sane base pointers 2622 Node *addp = n->in(AddPNode::Address); 2623 assert( !addp->is_AddP() || 2624 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2625 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2626 "Base pointers must match" ); 2627#ifdef _LP64 2628 if ((UseCompressedOops || UseCompressedKlassPointers) && 2629 addp->Opcode() == Op_ConP && 2630 addp == n->in(AddPNode::Base) && 2631 n->in(AddPNode::Offset)->is_Con()) { 2632 // Use addressing with narrow klass to load with offset on x86. 2633 // On sparc loading 32-bits constant and decoding it have less 2634 // instructions (4) then load 64-bits constant (7). 2635 // Do this transformation here since IGVN will convert ConN back to ConP. 2636 const Type* t = addp->bottom_type(); 2637 if (t->isa_oopptr() || t->isa_klassptr()) { 2638 Node* nn = NULL; 2639 2640 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass; 2641 2642 // Look for existing ConN node of the same exact type. 2643 Node* r = root(); 2644 uint cnt = r->outcnt(); 2645 for (uint i = 0; i < cnt; i++) { 2646 Node* m = r->raw_out(i); 2647 if (m!= NULL && m->Opcode() == op && 2648 m->bottom_type()->make_ptr() == t) { 2649 nn = m; 2650 break; 2651 } 2652 } 2653 if (nn != NULL) { 2654 // Decode a narrow oop to match address 2655 // [R12 + narrow_oop_reg<<3 + offset] 2656 if (t->isa_oopptr()) { 2657 nn = new (this) DecodeNNode(nn, t); 2658 } else { 2659 nn = new (this) DecodeNKlassNode(nn, t); 2660 } 2661 n->set_req(AddPNode::Base, nn); 2662 n->set_req(AddPNode::Address, nn); 2663 if (addp->outcnt() == 0) { 2664 addp->disconnect_inputs(NULL, this); 2665 } 2666 } 2667 } 2668 } 2669#endif 2670 break; 2671 } 2672 2673#ifdef _LP64 2674 case Op_CastPP: 2675 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 2676 Node* in1 = n->in(1); 2677 const Type* t = n->bottom_type(); 2678 Node* new_in1 = in1->clone(); 2679 new_in1->as_DecodeN()->set_type(t); 2680 2681 if (!Matcher::narrow_oop_use_complex_address()) { 2682 // 2683 // x86, ARM and friends can handle 2 adds in addressing mode 2684 // and Matcher can fold a DecodeN node into address by using 2685 // a narrow oop directly and do implicit NULL check in address: 2686 // 2687 // [R12 + narrow_oop_reg<<3 + offset] 2688 // NullCheck narrow_oop_reg 2689 // 2690 // On other platforms (Sparc) we have to keep new DecodeN node and 2691 // use it to do implicit NULL check in address: 2692 // 2693 // decode_not_null narrow_oop_reg, base_reg 2694 // [base_reg + offset] 2695 // NullCheck base_reg 2696 // 2697 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2698 // to keep the information to which NULL check the new DecodeN node 2699 // corresponds to use it as value in implicit_null_check(). 2700 // 2701 new_in1->set_req(0, n->in(0)); 2702 } 2703 2704 n->subsume_by(new_in1, this); 2705 if (in1->outcnt() == 0) { 2706 in1->disconnect_inputs(NULL, this); 2707 } 2708 } 2709 break; 2710 2711 case Op_CmpP: 2712 // Do this transformation here to preserve CmpPNode::sub() and 2713 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2714 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 2715 Node* in1 = n->in(1); 2716 Node* in2 = n->in(2); 2717 if (!in1->is_DecodeNarrowPtr()) { 2718 in2 = in1; 2719 in1 = n->in(2); 2720 } 2721 assert(in1->is_DecodeNarrowPtr(), "sanity"); 2722 2723 Node* new_in2 = NULL; 2724 if (in2->is_DecodeNarrowPtr()) { 2725 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 2726 new_in2 = in2->in(1); 2727 } else if (in2->Opcode() == Op_ConP) { 2728 const Type* t = in2->bottom_type(); 2729 if (t == TypePtr::NULL_PTR) { 2730 assert(in1->is_DecodeN(), "compare klass to null?"); 2731 // Don't convert CmpP null check into CmpN if compressed 2732 // oops implicit null check is not generated. 2733 // This will allow to generate normal oop implicit null check. 2734 if (Matcher::gen_narrow_oop_implicit_null_checks()) 2735 new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR); 2736 // 2737 // This transformation together with CastPP transformation above 2738 // will generated code for implicit NULL checks for compressed oops. 2739 // 2740 // The original code after Optimize() 2741 // 2742 // LoadN memory, narrow_oop_reg 2743 // decode narrow_oop_reg, base_reg 2744 // CmpP base_reg, NULL 2745 // CastPP base_reg // NotNull 2746 // Load [base_reg + offset], val_reg 2747 // 2748 // after these transformations will be 2749 // 2750 // LoadN memory, narrow_oop_reg 2751 // CmpN narrow_oop_reg, NULL 2752 // decode_not_null narrow_oop_reg, base_reg 2753 // Load [base_reg + offset], val_reg 2754 // 2755 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2756 // since narrow oops can be used in debug info now (see the code in 2757 // final_graph_reshaping_walk()). 2758 // 2759 // At the end the code will be matched to 2760 // on x86: 2761 // 2762 // Load_narrow_oop memory, narrow_oop_reg 2763 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2764 // NullCheck narrow_oop_reg 2765 // 2766 // and on sparc: 2767 // 2768 // Load_narrow_oop memory, narrow_oop_reg 2769 // decode_not_null narrow_oop_reg, base_reg 2770 // Load [base_reg + offset], val_reg 2771 // NullCheck base_reg 2772 // 2773 } else if (t->isa_oopptr()) { 2774 new_in2 = ConNode::make(this, t->make_narrowoop()); 2775 } else if (t->isa_klassptr()) { 2776 new_in2 = ConNode::make(this, t->make_narrowklass()); 2777 } 2778 } 2779 if (new_in2 != NULL) { 2780 Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2); 2781 n->subsume_by(cmpN, this); 2782 if (in1->outcnt() == 0) { 2783 in1->disconnect_inputs(NULL, this); 2784 } 2785 if (in2->outcnt() == 0) { 2786 in2->disconnect_inputs(NULL, this); 2787 } 2788 } 2789 } 2790 break; 2791 2792 case Op_DecodeN: 2793 case Op_DecodeNKlass: 2794 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 2795 // DecodeN could be pinned when it can't be fold into 2796 // an address expression, see the code for Op_CastPP above. 2797 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 2798 break; 2799 2800 case Op_EncodeP: 2801 case Op_EncodePKlass: { 2802 Node* in1 = n->in(1); 2803 if (in1->is_DecodeNarrowPtr()) { 2804 n->subsume_by(in1->in(1), this); 2805 } else if (in1->Opcode() == Op_ConP) { 2806 const Type* t = in1->bottom_type(); 2807 if (t == TypePtr::NULL_PTR) { 2808 assert(t->isa_oopptr(), "null klass?"); 2809 n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this); 2810 } else if (t->isa_oopptr()) { 2811 n->subsume_by(ConNode::make(this, t->make_narrowoop()), this); 2812 } else if (t->isa_klassptr()) { 2813 n->subsume_by(ConNode::make(this, t->make_narrowklass()), this); 2814 } 2815 } 2816 if (in1->outcnt() == 0) { 2817 in1->disconnect_inputs(NULL, this); 2818 } 2819 break; 2820 } 2821 2822 case Op_Proj: { 2823 if (OptimizeStringConcat) { 2824 ProjNode* p = n->as_Proj(); 2825 if (p->_is_io_use) { 2826 // Separate projections were used for the exception path which 2827 // are normally removed by a late inline. If it wasn't inlined 2828 // then they will hang around and should just be replaced with 2829 // the original one. 2830 Node* proj = NULL; 2831 // Replace with just one 2832 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 2833 Node *use = i.get(); 2834 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 2835 proj = use; 2836 break; 2837 } 2838 } 2839 assert(proj != NULL, "must be found"); 2840 p->subsume_by(proj, this); 2841 } 2842 } 2843 break; 2844 } 2845 2846 case Op_Phi: 2847 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 2848 // The EncodeP optimization may create Phi with the same edges 2849 // for all paths. It is not handled well by Register Allocator. 2850 Node* unique_in = n->in(1); 2851 assert(unique_in != NULL, ""); 2852 uint cnt = n->req(); 2853 for (uint i = 2; i < cnt; i++) { 2854 Node* m = n->in(i); 2855 assert(m != NULL, ""); 2856 if (unique_in != m) 2857 unique_in = NULL; 2858 } 2859 if (unique_in != NULL) { 2860 n->subsume_by(unique_in, this); 2861 } 2862 } 2863 break; 2864 2865#endif 2866 2867 case Op_ModI: 2868 if (UseDivMod) { 2869 // Check if a%b and a/b both exist 2870 Node* d = n->find_similar(Op_DivI); 2871 if (d) { 2872 // Replace them with a fused divmod if supported 2873 if (Matcher::has_match_rule(Op_DivModI)) { 2874 DivModINode* divmod = DivModINode::make(this, n); 2875 d->subsume_by(divmod->div_proj(), this); 2876 n->subsume_by(divmod->mod_proj(), this); 2877 } else { 2878 // replace a%b with a-((a/b)*b) 2879 Node* mult = new (this) MulINode(d, d->in(2)); 2880 Node* sub = new (this) SubINode(d->in(1), mult); 2881 n->subsume_by(sub, this); 2882 } 2883 } 2884 } 2885 break; 2886 2887 case Op_ModL: 2888 if (UseDivMod) { 2889 // Check if a%b and a/b both exist 2890 Node* d = n->find_similar(Op_DivL); 2891 if (d) { 2892 // Replace them with a fused divmod if supported 2893 if (Matcher::has_match_rule(Op_DivModL)) { 2894 DivModLNode* divmod = DivModLNode::make(this, n); 2895 d->subsume_by(divmod->div_proj(), this); 2896 n->subsume_by(divmod->mod_proj(), this); 2897 } else { 2898 // replace a%b with a-((a/b)*b) 2899 Node* mult = new (this) MulLNode(d, d->in(2)); 2900 Node* sub = new (this) SubLNode(d->in(1), mult); 2901 n->subsume_by(sub, this); 2902 } 2903 } 2904 } 2905 break; 2906 2907 case Op_LoadVector: 2908 case Op_StoreVector: 2909 break; 2910 2911 case Op_PackB: 2912 case Op_PackS: 2913 case Op_PackI: 2914 case Op_PackF: 2915 case Op_PackL: 2916 case Op_PackD: 2917 if (n->req()-1 > 2) { 2918 // Replace many operand PackNodes with a binary tree for matching 2919 PackNode* p = (PackNode*) n; 2920 Node* btp = p->binary_tree_pack(this, 1, n->req()); 2921 n->subsume_by(btp, this); 2922 } 2923 break; 2924 case Op_Loop: 2925 case Op_CountedLoop: 2926 if (n->as_Loop()->is_inner_loop()) { 2927 frc.inc_inner_loop_count(); 2928 } 2929 break; 2930 case Op_LShiftI: 2931 case Op_RShiftI: 2932 case Op_URShiftI: 2933 case Op_LShiftL: 2934 case Op_RShiftL: 2935 case Op_URShiftL: 2936 if (Matcher::need_masked_shift_count) { 2937 // The cpu's shift instructions don't restrict the count to the 2938 // lower 5/6 bits. We need to do the masking ourselves. 2939 Node* in2 = n->in(2); 2940 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 2941 const TypeInt* t = in2->find_int_type(); 2942 if (t != NULL && t->is_con()) { 2943 juint shift = t->get_con(); 2944 if (shift > mask) { // Unsigned cmp 2945 n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask))); 2946 } 2947 } else { 2948 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 2949 Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask))); 2950 n->set_req(2, shift); 2951 } 2952 } 2953 if (in2->outcnt() == 0) { // Remove dead node 2954 in2->disconnect_inputs(NULL, this); 2955 } 2956 } 2957 break; 2958 case Op_MemBarStoreStore: 2959 case Op_MemBarRelease: 2960 // Break the link with AllocateNode: it is no longer useful and 2961 // confuses register allocation. 2962 if (n->req() > MemBarNode::Precedent) { 2963 n->set_req(MemBarNode::Precedent, top()); 2964 } 2965 break; 2966 default: 2967 assert( !n->is_Call(), "" ); 2968 assert( !n->is_Mem(), "" ); 2969 break; 2970 } 2971 2972 // Collect CFG split points 2973 if (n->is_MultiBranch()) 2974 frc._tests.push(n); 2975} 2976 2977//------------------------------final_graph_reshaping_walk--------------------- 2978// Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 2979// requires that the walk visits a node's inputs before visiting the node. 2980void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 2981 ResourceArea *area = Thread::current()->resource_area(); 2982 Unique_Node_List sfpt(area); 2983 2984 frc._visited.set(root->_idx); // first, mark node as visited 2985 uint cnt = root->req(); 2986 Node *n = root; 2987 uint i = 0; 2988 while (true) { 2989 if (i < cnt) { 2990 // Place all non-visited non-null inputs onto stack 2991 Node* m = n->in(i); 2992 ++i; 2993 if (m != NULL && !frc._visited.test_set(m->_idx)) { 2994 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) 2995 sfpt.push(m); 2996 cnt = m->req(); 2997 nstack.push(n, i); // put on stack parent and next input's index 2998 n = m; 2999 i = 0; 3000 } 3001 } else { 3002 // Now do post-visit work 3003 final_graph_reshaping_impl( n, frc ); 3004 if (nstack.is_empty()) 3005 break; // finished 3006 n = nstack.node(); // Get node from stack 3007 cnt = n->req(); 3008 i = nstack.index(); 3009 nstack.pop(); // Shift to the next node on stack 3010 } 3011 } 3012 3013 // Skip next transformation if compressed oops are not used. 3014 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3015 (!UseCompressedOops && !UseCompressedKlassPointers)) 3016 return; 3017 3018 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3019 // It could be done for an uncommon traps or any safepoints/calls 3020 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3021 while (sfpt.size() > 0) { 3022 n = sfpt.pop(); 3023 JVMState *jvms = n->as_SafePoint()->jvms(); 3024 assert(jvms != NULL, "sanity"); 3025 int start = jvms->debug_start(); 3026 int end = n->req(); 3027 bool is_uncommon = (n->is_CallStaticJava() && 3028 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3029 for (int j = start; j < end; j++) { 3030 Node* in = n->in(j); 3031 if (in->is_DecodeNarrowPtr()) { 3032 bool safe_to_skip = true; 3033 if (!is_uncommon ) { 3034 // Is it safe to skip? 3035 for (uint i = 0; i < in->outcnt(); i++) { 3036 Node* u = in->raw_out(i); 3037 if (!u->is_SafePoint() || 3038 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 3039 safe_to_skip = false; 3040 } 3041 } 3042 } 3043 if (safe_to_skip) { 3044 n->set_req(j, in->in(1)); 3045 } 3046 if (in->outcnt() == 0) { 3047 in->disconnect_inputs(NULL, this); 3048 } 3049 } 3050 } 3051 } 3052} 3053 3054//------------------------------final_graph_reshaping-------------------------- 3055// Final Graph Reshaping. 3056// 3057// (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3058// and not commoned up and forced early. Must come after regular 3059// optimizations to avoid GVN undoing the cloning. Clone constant 3060// inputs to Loop Phis; these will be split by the allocator anyways. 3061// Remove Opaque nodes. 3062// (2) Move last-uses by commutative operations to the left input to encourage 3063// Intel update-in-place two-address operations and better register usage 3064// on RISCs. Must come after regular optimizations to avoid GVN Ideal 3065// calls canonicalizing them back. 3066// (3) Count the number of double-precision FP ops, single-precision FP ops 3067// and call sites. On Intel, we can get correct rounding either by 3068// forcing singles to memory (requires extra stores and loads after each 3069// FP bytecode) or we can set a rounding mode bit (requires setting and 3070// clearing the mode bit around call sites). The mode bit is only used 3071// if the relative frequency of single FP ops to calls is low enough. 3072// This is a key transform for SPEC mpeg_audio. 3073// (4) Detect infinite loops; blobs of code reachable from above but not 3074// below. Several of the Code_Gen algorithms fail on such code shapes, 3075// so we simply bail out. Happens a lot in ZKM.jar, but also happens 3076// from time to time in other codes (such as -Xcomp finalizer loops, etc). 3077// Detection is by looking for IfNodes where only 1 projection is 3078// reachable from below or CatchNodes missing some targets. 3079// (5) Assert for insane oop offsets in debug mode. 3080 3081bool Compile::final_graph_reshaping() { 3082 // an infinite loop may have been eliminated by the optimizer, 3083 // in which case the graph will be empty. 3084 if (root()->req() == 1) { 3085 record_method_not_compilable("trivial infinite loop"); 3086 return true; 3087 } 3088 3089 // Expensive nodes have their control input set to prevent the GVN 3090 // from freely commoning them. There's no GVN beyond this point so 3091 // no need to keep the control input. We want the expensive nodes to 3092 // be freely moved to the least frequent code path by gcm. 3093 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3094 for (int i = 0; i < expensive_count(); i++) { 3095 _expensive_nodes->at(i)->set_req(0, NULL); 3096 } 3097 3098 Final_Reshape_Counts frc; 3099 3100 // Visit everybody reachable! 3101 // Allocate stack of size C->unique()/2 to avoid frequent realloc 3102 Node_Stack nstack(unique() >> 1); 3103 final_graph_reshaping_walk(nstack, root(), frc); 3104 3105 // Check for unreachable (from below) code (i.e., infinite loops). 3106 for( uint i = 0; i < frc._tests.size(); i++ ) { 3107 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3108 // Get number of CFG targets. 3109 // Note that PCTables include exception targets after calls. 3110 uint required_outcnt = n->required_outcnt(); 3111 if (n->outcnt() != required_outcnt) { 3112 // Check for a few special cases. Rethrow Nodes never take the 3113 // 'fall-thru' path, so expected kids is 1 less. 3114 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3115 if (n->in(0)->in(0)->is_Call()) { 3116 CallNode *call = n->in(0)->in(0)->as_Call(); 3117 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3118 required_outcnt--; // Rethrow always has 1 less kid 3119 } else if (call->req() > TypeFunc::Parms && 3120 call->is_CallDynamicJava()) { 3121 // Check for null receiver. In such case, the optimizer has 3122 // detected that the virtual call will always result in a null 3123 // pointer exception. The fall-through projection of this CatchNode 3124 // will not be populated. 3125 Node *arg0 = call->in(TypeFunc::Parms); 3126 if (arg0->is_Type() && 3127 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3128 required_outcnt--; 3129 } 3130 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3131 call->req() > TypeFunc::Parms+1 && 3132 call->is_CallStaticJava()) { 3133 // Check for negative array length. In such case, the optimizer has 3134 // detected that the allocation attempt will always result in an 3135 // exception. There is no fall-through projection of this CatchNode . 3136 Node *arg1 = call->in(TypeFunc::Parms+1); 3137 if (arg1->is_Type() && 3138 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3139 required_outcnt--; 3140 } 3141 } 3142 } 3143 } 3144 // Recheck with a better notion of 'required_outcnt' 3145 if (n->outcnt() != required_outcnt) { 3146 record_method_not_compilable("malformed control flow"); 3147 return true; // Not all targets reachable! 3148 } 3149 } 3150 // Check that I actually visited all kids. Unreached kids 3151 // must be infinite loops. 3152 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3153 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3154 record_method_not_compilable("infinite loop"); 3155 return true; // Found unvisited kid; must be unreach 3156 } 3157 } 3158 3159 // If original bytecodes contained a mixture of floats and doubles 3160 // check if the optimizer has made it homogenous, item (3). 3161 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3162 frc.get_float_count() > 32 && 3163 frc.get_double_count() == 0 && 3164 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3165 set_24_bit_selection_and_mode( false, true ); 3166 } 3167 3168 set_java_calls(frc.get_java_call_count()); 3169 set_inner_loops(frc.get_inner_loop_count()); 3170 3171 // No infinite loops, no reason to bail out. 3172 return false; 3173} 3174 3175//-----------------------------too_many_traps---------------------------------- 3176// Report if there are too many traps at the current method and bci. 3177// Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3178bool Compile::too_many_traps(ciMethod* method, 3179 int bci, 3180 Deoptimization::DeoptReason reason) { 3181 ciMethodData* md = method->method_data(); 3182 if (md->is_empty()) { 3183 // Assume the trap has not occurred, or that it occurred only 3184 // because of a transient condition during start-up in the interpreter. 3185 return false; 3186 } 3187 if (md->has_trap_at(bci, reason) != 0) { 3188 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3189 // Also, if there are multiple reasons, or if there is no per-BCI record, 3190 // assume the worst. 3191 if (log()) 3192 log()->elem("observe trap='%s' count='%d'", 3193 Deoptimization::trap_reason_name(reason), 3194 md->trap_count(reason)); 3195 return true; 3196 } else { 3197 // Ignore method/bci and see if there have been too many globally. 3198 return too_many_traps(reason, md); 3199 } 3200} 3201 3202// Less-accurate variant which does not require a method and bci. 3203bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3204 ciMethodData* logmd) { 3205 if (trap_count(reason) >= (uint)PerMethodTrapLimit) { 3206 // Too many traps globally. 3207 // Note that we use cumulative trap_count, not just md->trap_count. 3208 if (log()) { 3209 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3210 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3211 Deoptimization::trap_reason_name(reason), 3212 mcount, trap_count(reason)); 3213 } 3214 return true; 3215 } else { 3216 // The coast is clear. 3217 return false; 3218 } 3219} 3220 3221//--------------------------too_many_recompiles-------------------------------- 3222// Report if there are too many recompiles at the current method and bci. 3223// Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3224// Is not eager to return true, since this will cause the compiler to use 3225// Action_none for a trap point, to avoid too many recompilations. 3226bool Compile::too_many_recompiles(ciMethod* method, 3227 int bci, 3228 Deoptimization::DeoptReason reason) { 3229 ciMethodData* md = method->method_data(); 3230 if (md->is_empty()) { 3231 // Assume the trap has not occurred, or that it occurred only 3232 // because of a transient condition during start-up in the interpreter. 3233 return false; 3234 } 3235 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3236 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3237 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3238 Deoptimization::DeoptReason per_bc_reason 3239 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3240 if ((per_bc_reason == Deoptimization::Reason_none 3241 || md->has_trap_at(bci, reason) != 0) 3242 // The trap frequency measure we care about is the recompile count: 3243 && md->trap_recompiled_at(bci) 3244 && md->overflow_recompile_count() >= bc_cutoff) { 3245 // Do not emit a trap here if it has already caused recompilations. 3246 // Also, if there are multiple reasons, or if there is no per-BCI record, 3247 // assume the worst. 3248 if (log()) 3249 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3250 Deoptimization::trap_reason_name(reason), 3251 md->trap_count(reason), 3252 md->overflow_recompile_count()); 3253 return true; 3254 } else if (trap_count(reason) != 0 3255 && decompile_count() >= m_cutoff) { 3256 // Too many recompiles globally, and we have seen this sort of trap. 3257 // Use cumulative decompile_count, not just md->decompile_count. 3258 if (log()) 3259 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3260 Deoptimization::trap_reason_name(reason), 3261 md->trap_count(reason), trap_count(reason), 3262 md->decompile_count(), decompile_count()); 3263 return true; 3264 } else { 3265 // The coast is clear. 3266 return false; 3267 } 3268} 3269 3270 3271#ifndef PRODUCT 3272//------------------------------verify_graph_edges--------------------------- 3273// Walk the Graph and verify that there is a one-to-one correspondence 3274// between Use-Def edges and Def-Use edges in the graph. 3275void Compile::verify_graph_edges(bool no_dead_code) { 3276 if (VerifyGraphEdges) { 3277 ResourceArea *area = Thread::current()->resource_area(); 3278 Unique_Node_List visited(area); 3279 // Call recursive graph walk to check edges 3280 _root->verify_edges(visited); 3281 if (no_dead_code) { 3282 // Now make sure that no visited node is used by an unvisited node. 3283 bool dead_nodes = 0; 3284 Unique_Node_List checked(area); 3285 while (visited.size() > 0) { 3286 Node* n = visited.pop(); 3287 checked.push(n); 3288 for (uint i = 0; i < n->outcnt(); i++) { 3289 Node* use = n->raw_out(i); 3290 if (checked.member(use)) continue; // already checked 3291 if (visited.member(use)) continue; // already in the graph 3292 if (use->is_Con()) continue; // a dead ConNode is OK 3293 // At this point, we have found a dead node which is DU-reachable. 3294 if (dead_nodes++ == 0) 3295 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3296 use->dump(2); 3297 tty->print_cr("---"); 3298 checked.push(use); // No repeats; pretend it is now checked. 3299 } 3300 } 3301 assert(dead_nodes == 0, "using nodes must be reachable from root"); 3302 } 3303 } 3304} 3305#endif 3306 3307// The Compile object keeps track of failure reasons separately from the ciEnv. 3308// This is required because there is not quite a 1-1 relation between the 3309// ciEnv and its compilation task and the Compile object. Note that one 3310// ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3311// to backtrack and retry without subsuming loads. Other than this backtracking 3312// behavior, the Compile's failure reason is quietly copied up to the ciEnv 3313// by the logic in C2Compiler. 3314void Compile::record_failure(const char* reason) { 3315 if (log() != NULL) { 3316 log()->elem("failure reason='%s' phase='compile'", reason); 3317 } 3318 if (_failure_reason == NULL) { 3319 // Record the first failure reason. 3320 _failure_reason = reason; 3321 } 3322 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3323 C->print_method(_failure_reason); 3324 } 3325 _root = NULL; // flush the graph, too 3326} 3327 3328Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog) 3329 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false), 3330 _phase_name(name), _dolog(dolog) 3331{ 3332 if (dolog) { 3333 C = Compile::current(); 3334 _log = C->log(); 3335 } else { 3336 C = NULL; 3337 _log = NULL; 3338 } 3339 if (_log != NULL) { 3340 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3341 _log->stamp(); 3342 _log->end_head(); 3343 } 3344} 3345 3346Compile::TracePhase::~TracePhase() { 3347 3348 C = Compile::current(); 3349 if (_dolog) { 3350 _log = C->log(); 3351 } else { 3352 _log = NULL; 3353 } 3354 3355#ifdef ASSERT 3356 if (PrintIdealNodeCount) { 3357 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3358 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3359 } 3360 3361 if (VerifyIdealNodeCount) { 3362 Compile::current()->print_missing_nodes(); 3363 } 3364#endif 3365 3366 if (_log != NULL) { 3367 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3368 } 3369} 3370 3371//============================================================================= 3372// Two Constant's are equal when the type and the value are equal. 3373bool Compile::Constant::operator==(const Constant& other) { 3374 if (type() != other.type() ) return false; 3375 if (can_be_reused() != other.can_be_reused()) return false; 3376 // For floating point values we compare the bit pattern. 3377 switch (type()) { 3378 case T_FLOAT: return (_v._value.i == other._v._value.i); 3379 case T_LONG: 3380 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3381 case T_OBJECT: 3382 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3383 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3384 case T_METADATA: return (_v._metadata == other._v._metadata); 3385 default: ShouldNotReachHere(); 3386 } 3387 return false; 3388} 3389 3390static int type_to_size_in_bytes(BasicType t) { 3391 switch (t) { 3392 case T_LONG: return sizeof(jlong ); 3393 case T_FLOAT: return sizeof(jfloat ); 3394 case T_DOUBLE: return sizeof(jdouble); 3395 case T_METADATA: return sizeof(Metadata*); 3396 // We use T_VOID as marker for jump-table entries (labels) which 3397 // need an internal word relocation. 3398 case T_VOID: 3399 case T_ADDRESS: 3400 case T_OBJECT: return sizeof(jobject); 3401 } 3402 3403 ShouldNotReachHere(); 3404 return -1; 3405} 3406 3407int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 3408 // sort descending 3409 if (a->freq() > b->freq()) return -1; 3410 if (a->freq() < b->freq()) return 1; 3411 return 0; 3412} 3413 3414void Compile::ConstantTable::calculate_offsets_and_size() { 3415 // First, sort the array by frequencies. 3416 _constants.sort(qsort_comparator); 3417 3418#ifdef ASSERT 3419 // Make sure all jump-table entries were sorted to the end of the 3420 // array (they have a negative frequency). 3421 bool found_void = false; 3422 for (int i = 0; i < _constants.length(); i++) { 3423 Constant con = _constants.at(i); 3424 if (con.type() == T_VOID) 3425 found_void = true; // jump-tables 3426 else 3427 assert(!found_void, "wrong sorting"); 3428 } 3429#endif 3430 3431 int offset = 0; 3432 for (int i = 0; i < _constants.length(); i++) { 3433 Constant* con = _constants.adr_at(i); 3434 3435 // Align offset for type. 3436 int typesize = type_to_size_in_bytes(con->type()); 3437 offset = align_size_up(offset, typesize); 3438 con->set_offset(offset); // set constant's offset 3439 3440 if (con->type() == T_VOID) { 3441 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 3442 offset = offset + typesize * n->outcnt(); // expand jump-table 3443 } else { 3444 offset = offset + typesize; 3445 } 3446 } 3447 3448 // Align size up to the next section start (which is insts; see 3449 // CodeBuffer::align_at_start). 3450 assert(_size == -1, "already set?"); 3451 _size = align_size_up(offset, CodeEntryAlignment); 3452} 3453 3454void Compile::ConstantTable::emit(CodeBuffer& cb) { 3455 MacroAssembler _masm(&cb); 3456 for (int i = 0; i < _constants.length(); i++) { 3457 Constant con = _constants.at(i); 3458 address constant_addr; 3459 switch (con.type()) { 3460 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 3461 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 3462 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 3463 case T_OBJECT: { 3464 jobject obj = con.get_jobject(); 3465 int oop_index = _masm.oop_recorder()->find_index(obj); 3466 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 3467 break; 3468 } 3469 case T_ADDRESS: { 3470 address addr = (address) con.get_jobject(); 3471 constant_addr = _masm.address_constant(addr); 3472 break; 3473 } 3474 // We use T_VOID as marker for jump-table entries (labels) which 3475 // need an internal word relocation. 3476 case T_VOID: { 3477 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 3478 // Fill the jump-table with a dummy word. The real value is 3479 // filled in later in fill_jump_table. 3480 address dummy = (address) n; 3481 constant_addr = _masm.address_constant(dummy); 3482 // Expand jump-table 3483 for (uint i = 1; i < n->outcnt(); i++) { 3484 address temp_addr = _masm.address_constant(dummy + i); 3485 assert(temp_addr, "consts section too small"); 3486 } 3487 break; 3488 } 3489 case T_METADATA: { 3490 Metadata* obj = con.get_metadata(); 3491 int metadata_index = _masm.oop_recorder()->find_index(obj); 3492 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 3493 break; 3494 } 3495 default: ShouldNotReachHere(); 3496 } 3497 assert(constant_addr, "consts section too small"); 3498 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), err_msg_res("must be: %d == %d", constant_addr - _masm.code()->consts()->start(), con.offset())); 3499 } 3500} 3501 3502int Compile::ConstantTable::find_offset(Constant& con) const { 3503 int idx = _constants.find(con); 3504 assert(idx != -1, "constant must be in constant table"); 3505 int offset = _constants.at(idx).offset(); 3506 assert(offset != -1, "constant table not emitted yet?"); 3507 return offset; 3508} 3509 3510void Compile::ConstantTable::add(Constant& con) { 3511 if (con.can_be_reused()) { 3512 int idx = _constants.find(con); 3513 if (idx != -1 && _constants.at(idx).can_be_reused()) { 3514 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 3515 return; 3516 } 3517 } 3518 (void) _constants.append(con); 3519} 3520 3521Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 3522 Block* b = Compile::current()->cfg()->_bbs[n->_idx]; 3523 Constant con(type, value, b->_freq); 3524 add(con); 3525 return con; 3526} 3527 3528Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 3529 Constant con(metadata); 3530 add(con); 3531 return con; 3532} 3533 3534Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 3535 jvalue value; 3536 BasicType type = oper->type()->basic_type(); 3537 switch (type) { 3538 case T_LONG: value.j = oper->constantL(); break; 3539 case T_FLOAT: value.f = oper->constantF(); break; 3540 case T_DOUBLE: value.d = oper->constantD(); break; 3541 case T_OBJECT: 3542 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 3543 case T_METADATA: return add((Metadata*)oper->constant()); break; 3544 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type))); 3545 } 3546 return add(n, type, value); 3547} 3548 3549Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 3550 jvalue value; 3551 // We can use the node pointer here to identify the right jump-table 3552 // as this method is called from Compile::Fill_buffer right before 3553 // the MachNodes are emitted and the jump-table is filled (means the 3554 // MachNode pointers do not change anymore). 3555 value.l = (jobject) n; 3556 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 3557 add(con); 3558 return con; 3559} 3560 3561void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 3562 // If called from Compile::scratch_emit_size do nothing. 3563 if (Compile::current()->in_scratch_emit_size()) return; 3564 3565 assert(labels.is_nonempty(), "must be"); 3566 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt())); 3567 3568 // Since MachConstantNode::constant_offset() also contains 3569 // table_base_offset() we need to subtract the table_base_offset() 3570 // to get the plain offset into the constant table. 3571 int offset = n->constant_offset() - table_base_offset(); 3572 3573 MacroAssembler _masm(&cb); 3574 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 3575 3576 for (uint i = 0; i < n->outcnt(); i++) { 3577 address* constant_addr = &jump_table_base[i]; 3578 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))); 3579 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 3580 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 3581 } 3582} 3583 3584void Compile::dump_inlining() { 3585 if (PrintInlining || PrintIntrinsics NOT_PRODUCT( || PrintOptoInlining)) { 3586 // Print inlining message for candidates that we couldn't inline 3587 // for lack of space or non constant receiver 3588 for (int i = 0; i < _late_inlines.length(); i++) { 3589 CallGenerator* cg = _late_inlines.at(i); 3590 cg->print_inlining_late("live nodes > LiveNodeCountInliningCutoff"); 3591 } 3592 Unique_Node_List useful; 3593 useful.push(root()); 3594 for (uint next = 0; next < useful.size(); ++next) { 3595 Node* n = useful.at(next); 3596 if (n->is_Call() && n->as_Call()->generator() != NULL && n->as_Call()->generator()->call_node() == n) { 3597 CallNode* call = n->as_Call(); 3598 CallGenerator* cg = call->generator(); 3599 cg->print_inlining_late("receiver not constant"); 3600 } 3601 uint max = n->len(); 3602 for ( uint i = 0; i < max; ++i ) { 3603 Node *m = n->in(i); 3604 if ( m == NULL ) continue; 3605 useful.push(m); 3606 } 3607 } 3608 for (int i = 0; i < _print_inlining_list->length(); i++) { 3609 tty->print(_print_inlining_list->at(i).ss()->as_string()); 3610 } 3611 } 3612} 3613 3614int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 3615 if (n1->Opcode() < n2->Opcode()) return -1; 3616 else if (n1->Opcode() > n2->Opcode()) return 1; 3617 3618 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())); 3619 for (uint i = 1; i < n1->req(); i++) { 3620 if (n1->in(i) < n2->in(i)) return -1; 3621 else if (n1->in(i) > n2->in(i)) return 1; 3622 } 3623 3624 return 0; 3625} 3626 3627int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 3628 Node* n1 = *n1p; 3629 Node* n2 = *n2p; 3630 3631 return cmp_expensive_nodes(n1, n2); 3632} 3633 3634void Compile::sort_expensive_nodes() { 3635 if (!expensive_nodes_sorted()) { 3636 _expensive_nodes->sort(cmp_expensive_nodes); 3637 } 3638} 3639 3640bool Compile::expensive_nodes_sorted() const { 3641 for (int i = 1; i < _expensive_nodes->length(); i++) { 3642 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 3643 return false; 3644 } 3645 } 3646 return true; 3647} 3648 3649bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 3650 if (_expensive_nodes->length() == 0) { 3651 return false; 3652 } 3653 3654 assert(OptimizeExpensiveOps, "optimization off?"); 3655 3656 // Take this opportunity to remove dead nodes from the list 3657 int j = 0; 3658 for (int i = 0; i < _expensive_nodes->length(); i++) { 3659 Node* n = _expensive_nodes->at(i); 3660 if (!n->is_unreachable(igvn)) { 3661 assert(n->is_expensive(), "should be expensive"); 3662 _expensive_nodes->at_put(j, n); 3663 j++; 3664 } 3665 } 3666 _expensive_nodes->trunc_to(j); 3667 3668 // Then sort the list so that similar nodes are next to each other 3669 // and check for at least two nodes of identical kind with same data 3670 // inputs. 3671 sort_expensive_nodes(); 3672 3673 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 3674 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 3675 return true; 3676 } 3677 } 3678 3679 return false; 3680} 3681 3682void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 3683 if (_expensive_nodes->length() == 0) { 3684 return; 3685 } 3686 3687 assert(OptimizeExpensiveOps, "optimization off?"); 3688 3689 // Sort to bring similar nodes next to each other and clear the 3690 // control input of nodes for which there's only a single copy. 3691 sort_expensive_nodes(); 3692 3693 int j = 0; 3694 int identical = 0; 3695 int i = 0; 3696 for (; i < _expensive_nodes->length()-1; i++) { 3697 assert(j <= i, "can't write beyond current index"); 3698 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 3699 identical++; 3700 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 3701 continue; 3702 } 3703 if (identical > 0) { 3704 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 3705 identical = 0; 3706 } else { 3707 Node* n = _expensive_nodes->at(i); 3708 igvn.hash_delete(n); 3709 n->set_req(0, NULL); 3710 igvn.hash_insert(n); 3711 } 3712 } 3713 if (identical > 0) { 3714 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 3715 } else if (_expensive_nodes->length() >= 1) { 3716 Node* n = _expensive_nodes->at(i); 3717 igvn.hash_delete(n); 3718 n->set_req(0, NULL); 3719 igvn.hash_insert(n); 3720 } 3721 _expensive_nodes->trunc_to(j); 3722} 3723 3724void Compile::add_expensive_node(Node * n) { 3725 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 3726 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 3727 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 3728 if (OptimizeExpensiveOps) { 3729 _expensive_nodes->append(n); 3730 } else { 3731 // Clear control input and let IGVN optimize expensive nodes if 3732 // OptimizeExpensiveOps is off. 3733 n->set_req(0, NULL); 3734 } 3735} 3736 3737// Auxiliary method to support randomized stressing/fuzzing. 3738// 3739// This method can be called the arbitrary number of times, with current count 3740// as the argument. The logic allows selecting a single candidate from the 3741// running list of candidates as follows: 3742// int count = 0; 3743// Cand* selected = null; 3744// while(cand = cand->next()) { 3745// if (randomized_select(++count)) { 3746// selected = cand; 3747// } 3748// } 3749// 3750// Including count equalizes the chances any candidate is "selected". 3751// This is useful when we don't have the complete list of candidates to choose 3752// from uniformly. In this case, we need to adjust the randomicity of the 3753// selection, or else we will end up biasing the selection towards the latter 3754// candidates. 3755// 3756// Quick back-envelope calculation shows that for the list of n candidates 3757// the equal probability for the candidate to persist as "best" can be 3758// achieved by replacing it with "next" k-th candidate with the probability 3759// of 1/k. It can be easily shown that by the end of the run, the 3760// probability for any candidate is converged to 1/n, thus giving the 3761// uniform distribution among all the candidates. 3762// 3763// We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 3764#define RANDOMIZED_DOMAIN_POW 29 3765#define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 3766#define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 3767bool Compile::randomized_select(int count) { 3768 assert(count > 0, "only positive"); 3769 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 3770} 3771