1 /* 2 * Copyright (c) 1998, 2015, 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 "ci/ciMethodData.hpp" 27 #include "classfile/systemDictionary.hpp" 28 #include "classfile/vmSymbols.hpp" 29 #include "compiler/compileLog.hpp" 30 #include "interpreter/linkResolver.hpp" 31 #include "memory/universe.inline.hpp" 32 #include "oops/oop.inline.hpp" 33 #include "opto/addnode.hpp" 34 #include "opto/castnode.hpp" 35 #include "opto/convertnode.hpp" 36 #include "opto/divnode.hpp" 37 #include "opto/idealGraphPrinter.hpp" 38 #include "opto/matcher.hpp" 39 #include "opto/memnode.hpp" 40 #include "opto/mulnode.hpp" 41 #include "opto/opaquenode.hpp" 42 #include "opto/parse.hpp" 43 #include "opto/runtime.hpp" 44 #include "opto/valuetypenode.hpp" 45 #include "runtime/deoptimization.hpp" 46 #include "runtime/sharedRuntime.hpp" 47 48 #ifndef PRODUCT 49 extern int explicit_null_checks_inserted, 50 explicit_null_checks_elided; 51 #endif 52 53 //---------------------------------array_load---------------------------------- 54 void Parse::array_load(BasicType elem_type) { 55 const Type* elem = Type::TOP; 56 Node* adr = array_addressing(elem_type, 0, &elem); 57 if (stopped()) return; // guaranteed null or range check 58 dec_sp(2); // Pop array and index 59 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(elem_type); 60 Node* ld = make_load(control(), adr, elem, elem_type, adr_type, MemNode::unordered); 61 push(ld); 62 } 63 64 65 //--------------------------------array_store---------------------------------- 66 void Parse::array_store(BasicType elem_type) { 67 Node* adr = array_addressing(elem_type, 1); 68 if (stopped()) return; // guaranteed null or range check 69 Node* val = pop(); 70 dec_sp(2); // Pop array and index 71 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(elem_type); 72 store_to_memory(control(), adr, val, elem_type, adr_type, StoreNode::release_if_reference(elem_type)); 73 } 74 75 76 //------------------------------array_addressing------------------------------- 77 // Pull array and index from the stack. Compute pointer-to-element. 78 Node* Parse::array_addressing(BasicType type, int vals, const Type* *result2) { 79 Node *idx = peek(0+vals); // Get from stack without popping 80 Node *ary = peek(1+vals); // in case of exception 81 82 // Null check the array base, with correct stack contents 83 ary = null_check(ary, T_ARRAY); 84 // Compile-time detect of null-exception? 85 if (stopped()) return top(); 86 87 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr(); 88 const TypeInt* sizetype = arytype->size(); 89 const Type* elemtype = arytype->elem(); 90 91 if (UseUniqueSubclasses && result2 != NULL) { 92 const Type* el = elemtype->make_ptr(); 93 if (el && el->isa_instptr()) { 94 const TypeInstPtr* toop = el->is_instptr(); 95 if (toop->klass()->as_instance_klass()->unique_concrete_subklass()) { 96 // If we load from "AbstractClass[]" we must see "ConcreteSubClass". 97 const Type* subklass = Type::get_const_type(toop->klass()); 98 elemtype = subklass->join_speculative(el); 99 } 100 } 101 } 102 103 // Check for big class initializers with all constant offsets 104 // feeding into a known-size array. 105 const TypeInt* idxtype = _gvn.type(idx)->is_int(); 106 // See if the highest idx value is less than the lowest array bound, 107 // and if the idx value cannot be negative: 108 bool need_range_check = true; 109 if (idxtype->_hi < sizetype->_lo && idxtype->_lo >= 0) { 110 need_range_check = false; 111 if (C->log() != NULL) C->log()->elem("observe that='!need_range_check'"); 112 } 113 114 ciKlass * arytype_klass = arytype->klass(); 115 if ((arytype_klass != NULL) && (!arytype_klass->is_loaded())) { 116 // Only fails for some -Xcomp runs 117 // The class is unloaded. We have to run this bytecode in the interpreter. 118 uncommon_trap(Deoptimization::Reason_unloaded, 119 Deoptimization::Action_reinterpret, 120 arytype->klass(), "!loaded array"); 121 return top(); 122 } 123 124 // Do the range check 125 if (GenerateRangeChecks && need_range_check) { 126 Node* tst; 127 if (sizetype->_hi <= 0) { 128 // The greatest array bound is negative, so we can conclude that we're 129 // compiling unreachable code, but the unsigned compare trick used below 130 // only works with non-negative lengths. Instead, hack "tst" to be zero so 131 // the uncommon_trap path will always be taken. 132 tst = _gvn.intcon(0); 133 } else { 134 // Range is constant in array-oop, so we can use the original state of mem 135 Node* len = load_array_length(ary); 136 137 // Test length vs index (standard trick using unsigned compare) 138 Node* chk = _gvn.transform( new CmpUNode(idx, len) ); 139 BoolTest::mask btest = BoolTest::lt; 140 tst = _gvn.transform( new BoolNode(chk, btest) ); 141 } 142 RangeCheckNode* rc = new RangeCheckNode(control(), tst, PROB_MAX, COUNT_UNKNOWN); 143 _gvn.set_type(rc, rc->Value(&_gvn)); 144 if (!tst->is_Con()) { 145 record_for_igvn(rc); 146 } 147 set_control(_gvn.transform(new IfTrueNode(rc))); 148 // Branch to failure if out of bounds 149 { 150 PreserveJVMState pjvms(this); 151 set_control(_gvn.transform(new IfFalseNode(rc))); 152 if (C->allow_range_check_smearing()) { 153 // Do not use builtin_throw, since range checks are sometimes 154 // made more stringent by an optimistic transformation. 155 // This creates "tentative" range checks at this point, 156 // which are not guaranteed to throw exceptions. 157 // See IfNode::Ideal, is_range_check, adjust_check. 158 uncommon_trap(Deoptimization::Reason_range_check, 159 Deoptimization::Action_make_not_entrant, 160 NULL, "range_check"); 161 } else { 162 // If we have already recompiled with the range-check-widening 163 // heroic optimization turned off, then we must really be throwing 164 // range check exceptions. 165 builtin_throw(Deoptimization::Reason_range_check, idx); 166 } 167 } 168 } 169 // Check for always knowing you are throwing a range-check exception 170 if (stopped()) return top(); 171 172 // Make array address computation control dependent to prevent it 173 // from floating above the range check during loop optimizations. 174 Node* ptr = array_element_address(ary, idx, type, sizetype, control()); 175 176 if (result2 != NULL) *result2 = elemtype; 177 178 assert(ptr != top(), "top should go hand-in-hand with stopped"); 179 180 return ptr; 181 } 182 183 184 // returns IfNode 185 IfNode* Parse::jump_if_fork_int(Node* a, Node* b, BoolTest::mask mask) { 186 Node *cmp = _gvn.transform( new CmpINode( a, b)); // two cases: shiftcount > 32 and shiftcount <= 32 187 Node *tst = _gvn.transform( new BoolNode( cmp, mask)); 188 IfNode *iff = create_and_map_if( control(), tst, ((mask == BoolTest::eq) ? PROB_STATIC_INFREQUENT : PROB_FAIR), COUNT_UNKNOWN ); 189 return iff; 190 } 191 192 // return Region node 193 Node* Parse::jump_if_join(Node* iffalse, Node* iftrue) { 194 Node *region = new RegionNode(3); // 2 results 195 record_for_igvn(region); 196 region->init_req(1, iffalse); 197 region->init_req(2, iftrue ); 198 _gvn.set_type(region, Type::CONTROL); 199 region = _gvn.transform(region); 200 set_control (region); 201 return region; 202 } 203 204 205 //------------------------------helper for tableswitch------------------------- 206 void Parse::jump_if_true_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index) { 207 // True branch, use existing map info 208 { PreserveJVMState pjvms(this); 209 Node *iftrue = _gvn.transform( new IfTrueNode (iff) ); 210 set_control( iftrue ); 211 profile_switch_case(prof_table_index); 212 merge_new_path(dest_bci_if_true); 213 } 214 215 // False branch 216 Node *iffalse = _gvn.transform( new IfFalseNode(iff) ); 217 set_control( iffalse ); 218 } 219 220 void Parse::jump_if_false_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index) { 221 // True branch, use existing map info 222 { PreserveJVMState pjvms(this); 223 Node *iffalse = _gvn.transform( new IfFalseNode (iff) ); 224 set_control( iffalse ); 225 profile_switch_case(prof_table_index); 226 merge_new_path(dest_bci_if_true); 227 } 228 229 // False branch 230 Node *iftrue = _gvn.transform( new IfTrueNode(iff) ); 231 set_control( iftrue ); 232 } 233 234 void Parse::jump_if_always_fork(int dest_bci, int prof_table_index) { 235 // False branch, use existing map and control() 236 profile_switch_case(prof_table_index); 237 merge_new_path(dest_bci); 238 } 239 240 241 extern "C" { 242 static int jint_cmp(const void *i, const void *j) { 243 int a = *(jint *)i; 244 int b = *(jint *)j; 245 return a > b ? 1 : a < b ? -1 : 0; 246 } 247 } 248 249 250 // Default value for methodData switch indexing. Must be a negative value to avoid 251 // conflict with any legal switch index. 252 #define NullTableIndex -1 253 254 class SwitchRange : public StackObj { 255 // a range of integers coupled with a bci destination 256 jint _lo; // inclusive lower limit 257 jint _hi; // inclusive upper limit 258 int _dest; 259 int _table_index; // index into method data table 260 261 public: 262 jint lo() const { return _lo; } 263 jint hi() const { return _hi; } 264 int dest() const { return _dest; } 265 int table_index() const { return _table_index; } 266 bool is_singleton() const { return _lo == _hi; } 267 268 void setRange(jint lo, jint hi, int dest, int table_index) { 269 assert(lo <= hi, "must be a non-empty range"); 270 _lo = lo, _hi = hi; _dest = dest; _table_index = table_index; 271 } 272 bool adjoinRange(jint lo, jint hi, int dest, int table_index) { 273 assert(lo <= hi, "must be a non-empty range"); 274 if (lo == _hi+1 && dest == _dest && table_index == _table_index) { 275 _hi = hi; 276 return true; 277 } 278 return false; 279 } 280 281 void set (jint value, int dest, int table_index) { 282 setRange(value, value, dest, table_index); 283 } 284 bool adjoin(jint value, int dest, int table_index) { 285 return adjoinRange(value, value, dest, table_index); 286 } 287 288 void print() { 289 if (is_singleton()) 290 tty->print(" {%d}=>%d", lo(), dest()); 291 else if (lo() == min_jint) 292 tty->print(" {..%d}=>%d", hi(), dest()); 293 else if (hi() == max_jint) 294 tty->print(" {%d..}=>%d", lo(), dest()); 295 else 296 tty->print(" {%d..%d}=>%d", lo(), hi(), dest()); 297 } 298 }; 299 300 301 //-------------------------------do_tableswitch-------------------------------- 302 void Parse::do_tableswitch() { 303 Node* lookup = pop(); 304 305 // Get information about tableswitch 306 int default_dest = iter().get_dest_table(0); 307 int lo_index = iter().get_int_table(1); 308 int hi_index = iter().get_int_table(2); 309 int len = hi_index - lo_index + 1; 310 311 if (len < 1) { 312 // If this is a backward branch, add safepoint 313 maybe_add_safepoint(default_dest); 314 merge(default_dest); 315 return; 316 } 317 318 // generate decision tree, using trichotomy when possible 319 int rnum = len+2; 320 bool makes_backward_branch = false; 321 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum); 322 int rp = -1; 323 if (lo_index != min_jint) { 324 ranges[++rp].setRange(min_jint, lo_index-1, default_dest, NullTableIndex); 325 } 326 for (int j = 0; j < len; j++) { 327 jint match_int = lo_index+j; 328 int dest = iter().get_dest_table(j+3); 329 makes_backward_branch |= (dest <= bci()); 330 int table_index = method_data_update() ? j : NullTableIndex; 331 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index)) { 332 ranges[++rp].set(match_int, dest, table_index); 333 } 334 } 335 jint highest = lo_index+(len-1); 336 assert(ranges[rp].hi() == highest, ""); 337 if (highest != max_jint 338 && !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex)) { 339 ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex); 340 } 341 assert(rp < len+2, "not too many ranges"); 342 343 // Safepoint in case if backward branch observed 344 if( makes_backward_branch && UseLoopSafepoints ) 345 add_safepoint(); 346 347 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]); 348 } 349 350 351 //------------------------------do_lookupswitch-------------------------------- 352 void Parse::do_lookupswitch() { 353 Node *lookup = pop(); // lookup value 354 // Get information about lookupswitch 355 int default_dest = iter().get_dest_table(0); 356 int len = iter().get_int_table(1); 357 358 if (len < 1) { // If this is a backward branch, add safepoint 359 maybe_add_safepoint(default_dest); 360 merge(default_dest); 361 return; 362 } 363 364 // generate decision tree, using trichotomy when possible 365 jint* table = NEW_RESOURCE_ARRAY(jint, len*2); 366 { 367 for( int j = 0; j < len; j++ ) { 368 table[j+j+0] = iter().get_int_table(2+j+j); 369 table[j+j+1] = iter().get_dest_table(2+j+j+1); 370 } 371 qsort( table, len, 2*sizeof(table[0]), jint_cmp ); 372 } 373 374 int rnum = len*2+1; 375 bool makes_backward_branch = false; 376 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum); 377 int rp = -1; 378 for( int j = 0; j < len; j++ ) { 379 jint match_int = table[j+j+0]; 380 int dest = table[j+j+1]; 381 int next_lo = rp < 0 ? min_jint : ranges[rp].hi()+1; 382 int table_index = method_data_update() ? j : NullTableIndex; 383 makes_backward_branch |= (dest <= bci()); 384 if( match_int != next_lo ) { 385 ranges[++rp].setRange(next_lo, match_int-1, default_dest, NullTableIndex); 386 } 387 if( rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index) ) { 388 ranges[++rp].set(match_int, dest, table_index); 389 } 390 } 391 jint highest = table[2*(len-1)]; 392 assert(ranges[rp].hi() == highest, ""); 393 if( highest != max_jint 394 && !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex) ) { 395 ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex); 396 } 397 assert(rp < rnum, "not too many ranges"); 398 399 // Safepoint in case backward branch observed 400 if( makes_backward_branch && UseLoopSafepoints ) 401 add_safepoint(); 402 403 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]); 404 } 405 406 //----------------------------create_jump_tables------------------------------- 407 bool Parse::create_jump_tables(Node* key_val, SwitchRange* lo, SwitchRange* hi) { 408 // Are jumptables enabled 409 if (!UseJumpTables) return false; 410 411 // Are jumptables supported 412 if (!Matcher::has_match_rule(Op_Jump)) return false; 413 414 // Don't make jump table if profiling 415 if (method_data_update()) return false; 416 417 // Decide if a guard is needed to lop off big ranges at either (or 418 // both) end(s) of the input set. We'll call this the default target 419 // even though we can't be sure that it is the true "default". 420 421 bool needs_guard = false; 422 int default_dest; 423 int64_t total_outlier_size = 0; 424 int64_t hi_size = ((int64_t)hi->hi()) - ((int64_t)hi->lo()) + 1; 425 int64_t lo_size = ((int64_t)lo->hi()) - ((int64_t)lo->lo()) + 1; 426 427 if (lo->dest() == hi->dest()) { 428 total_outlier_size = hi_size + lo_size; 429 default_dest = lo->dest(); 430 } else if (lo_size > hi_size) { 431 total_outlier_size = lo_size; 432 default_dest = lo->dest(); 433 } else { 434 total_outlier_size = hi_size; 435 default_dest = hi->dest(); 436 } 437 438 // If a guard test will eliminate very sparse end ranges, then 439 // it is worth the cost of an extra jump. 440 if (total_outlier_size > (MaxJumpTableSparseness * 4)) { 441 needs_guard = true; 442 if (default_dest == lo->dest()) lo++; 443 if (default_dest == hi->dest()) hi--; 444 } 445 446 // Find the total number of cases and ranges 447 int64_t num_cases = ((int64_t)hi->hi()) - ((int64_t)lo->lo()) + 1; 448 int num_range = hi - lo + 1; 449 450 // Don't create table if: too large, too small, or too sparse. 451 if (num_cases < MinJumpTableSize || num_cases > MaxJumpTableSize) 452 return false; 453 if (num_cases > (MaxJumpTableSparseness * num_range)) 454 return false; 455 456 // Normalize table lookups to zero 457 int lowval = lo->lo(); 458 key_val = _gvn.transform( new SubINode(key_val, _gvn.intcon(lowval)) ); 459 460 // Generate a guard to protect against input keyvals that aren't 461 // in the switch domain. 462 if (needs_guard) { 463 Node* size = _gvn.intcon(num_cases); 464 Node* cmp = _gvn.transform( new CmpUNode(key_val, size) ); 465 Node* tst = _gvn.transform( new BoolNode(cmp, BoolTest::ge) ); 466 IfNode* iff = create_and_map_if( control(), tst, PROB_FAIR, COUNT_UNKNOWN); 467 jump_if_true_fork(iff, default_dest, NullTableIndex); 468 } 469 470 // Create an ideal node JumpTable that has projections 471 // of all possible ranges for a switch statement 472 // The key_val input must be converted to a pointer offset and scaled. 473 // Compare Parse::array_addressing above. 474 475 // Clean the 32-bit int into a real 64-bit offset. 476 // Otherwise, the jint value 0 might turn into an offset of 0x0800000000. 477 const TypeInt* ikeytype = TypeInt::make(0, num_cases, Type::WidenMin); 478 // Make I2L conversion control dependent to prevent it from 479 // floating above the range check during loop optimizations. 480 key_val = C->conv_I2X_index(&_gvn, key_val, ikeytype, control()); 481 482 // Shift the value by wordsize so we have an index into the table, rather 483 // than a switch value 484 Node *shiftWord = _gvn.MakeConX(wordSize); 485 key_val = _gvn.transform( new MulXNode( key_val, shiftWord)); 486 487 // Create the JumpNode 488 Node* jtn = _gvn.transform( new JumpNode(control(), key_val, num_cases) ); 489 490 // These are the switch destinations hanging off the jumpnode 491 int i = 0; 492 for (SwitchRange* r = lo; r <= hi; r++) { 493 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) { 494 Node* input = _gvn.transform(new JumpProjNode(jtn, i, r->dest(), (int)(j - lowval))); 495 { 496 PreserveJVMState pjvms(this); 497 set_control(input); 498 jump_if_always_fork(r->dest(), r->table_index()); 499 } 500 } 501 } 502 assert(i == num_cases, "miscount of cases"); 503 stop_and_kill_map(); // no more uses for this JVMS 504 return true; 505 } 506 507 //----------------------------jump_switch_ranges------------------------------- 508 void Parse::jump_switch_ranges(Node* key_val, SwitchRange *lo, SwitchRange *hi, int switch_depth) { 509 Block* switch_block = block(); 510 511 if (switch_depth == 0) { 512 // Do special processing for the top-level call. 513 assert(lo->lo() == min_jint, "initial range must exhaust Type::INT"); 514 assert(hi->hi() == max_jint, "initial range must exhaust Type::INT"); 515 516 // Decrement pred-numbers for the unique set of nodes. 517 #ifdef ASSERT 518 // Ensure that the block's successors are a (duplicate-free) set. 519 int successors_counted = 0; // block occurrences in [hi..lo] 520 int unique_successors = switch_block->num_successors(); 521 for (int i = 0; i < unique_successors; i++) { 522 Block* target = switch_block->successor_at(i); 523 524 // Check that the set of successors is the same in both places. 525 int successors_found = 0; 526 for (SwitchRange* p = lo; p <= hi; p++) { 527 if (p->dest() == target->start()) successors_found++; 528 } 529 assert(successors_found > 0, "successor must be known"); 530 successors_counted += successors_found; 531 } 532 assert(successors_counted == (hi-lo)+1, "no unexpected successors"); 533 #endif 534 535 // Maybe prune the inputs, based on the type of key_val. 536 jint min_val = min_jint; 537 jint max_val = max_jint; 538 const TypeInt* ti = key_val->bottom_type()->isa_int(); 539 if (ti != NULL) { 540 min_val = ti->_lo; 541 max_val = ti->_hi; 542 assert(min_val <= max_val, "invalid int type"); 543 } 544 while (lo->hi() < min_val) lo++; 545 if (lo->lo() < min_val) lo->setRange(min_val, lo->hi(), lo->dest(), lo->table_index()); 546 while (hi->lo() > max_val) hi--; 547 if (hi->hi() > max_val) hi->setRange(hi->lo(), max_val, hi->dest(), hi->table_index()); 548 } 549 550 #ifndef PRODUCT 551 if (switch_depth == 0) { 552 _max_switch_depth = 0; 553 _est_switch_depth = log2_intptr((hi-lo+1)-1)+1; 554 } 555 #endif 556 557 assert(lo <= hi, "must be a non-empty set of ranges"); 558 if (lo == hi) { 559 jump_if_always_fork(lo->dest(), lo->table_index()); 560 } else { 561 assert(lo->hi() == (lo+1)->lo()-1, "contiguous ranges"); 562 assert(hi->lo() == (hi-1)->hi()+1, "contiguous ranges"); 563 564 if (create_jump_tables(key_val, lo, hi)) return; 565 566 int nr = hi - lo + 1; 567 568 SwitchRange* mid = lo + nr/2; 569 // if there is an easy choice, pivot at a singleton: 570 if (nr > 3 && !mid->is_singleton() && (mid-1)->is_singleton()) mid--; 571 572 assert(lo < mid && mid <= hi, "good pivot choice"); 573 assert(nr != 2 || mid == hi, "should pick higher of 2"); 574 assert(nr != 3 || mid == hi-1, "should pick middle of 3"); 575 576 Node *test_val = _gvn.intcon(mid->lo()); 577 578 if (mid->is_singleton()) { 579 IfNode *iff_ne = jump_if_fork_int(key_val, test_val, BoolTest::ne); 580 jump_if_false_fork(iff_ne, mid->dest(), mid->table_index()); 581 582 // Special Case: If there are exactly three ranges, and the high 583 // and low range each go to the same place, omit the "gt" test, 584 // since it will not discriminate anything. 585 bool eq_test_only = (hi == lo+2 && hi->dest() == lo->dest()); 586 if (eq_test_only) { 587 assert(mid == hi-1, ""); 588 } 589 590 // if there is a higher range, test for it and process it: 591 if (mid < hi && !eq_test_only) { 592 // two comparisons of same values--should enable 1 test for 2 branches 593 // Use BoolTest::le instead of BoolTest::gt 594 IfNode *iff_le = jump_if_fork_int(key_val, test_val, BoolTest::le); 595 Node *iftrue = _gvn.transform( new IfTrueNode(iff_le) ); 596 Node *iffalse = _gvn.transform( new IfFalseNode(iff_le) ); 597 { PreserveJVMState pjvms(this); 598 set_control(iffalse); 599 jump_switch_ranges(key_val, mid+1, hi, switch_depth+1); 600 } 601 set_control(iftrue); 602 } 603 604 } else { 605 // mid is a range, not a singleton, so treat mid..hi as a unit 606 IfNode *iff_ge = jump_if_fork_int(key_val, test_val, BoolTest::ge); 607 608 // if there is a higher range, test for it and process it: 609 if (mid == hi) { 610 jump_if_true_fork(iff_ge, mid->dest(), mid->table_index()); 611 } else { 612 Node *iftrue = _gvn.transform( new IfTrueNode(iff_ge) ); 613 Node *iffalse = _gvn.transform( new IfFalseNode(iff_ge) ); 614 { PreserveJVMState pjvms(this); 615 set_control(iftrue); 616 jump_switch_ranges(key_val, mid, hi, switch_depth+1); 617 } 618 set_control(iffalse); 619 } 620 } 621 622 // in any case, process the lower range 623 jump_switch_ranges(key_val, lo, mid-1, switch_depth+1); 624 } 625 626 // Decrease pred_count for each successor after all is done. 627 if (switch_depth == 0) { 628 int unique_successors = switch_block->num_successors(); 629 for (int i = 0; i < unique_successors; i++) { 630 Block* target = switch_block->successor_at(i); 631 // Throw away the pre-allocated path for each unique successor. 632 target->next_path_num(); 633 } 634 } 635 636 #ifndef PRODUCT 637 _max_switch_depth = MAX2(switch_depth, _max_switch_depth); 638 if (TraceOptoParse && Verbose && WizardMode && switch_depth == 0) { 639 SwitchRange* r; 640 int nsing = 0; 641 for( r = lo; r <= hi; r++ ) { 642 if( r->is_singleton() ) nsing++; 643 } 644 tty->print(">>> "); 645 _method->print_short_name(); 646 tty->print_cr(" switch decision tree"); 647 tty->print_cr(" %d ranges (%d singletons), max_depth=%d, est_depth=%d", 648 (int) (hi-lo+1), nsing, _max_switch_depth, _est_switch_depth); 649 if (_max_switch_depth > _est_switch_depth) { 650 tty->print_cr("******** BAD SWITCH DEPTH ********"); 651 } 652 tty->print(" "); 653 for( r = lo; r <= hi; r++ ) { 654 r->print(); 655 } 656 tty->cr(); 657 } 658 #endif 659 } 660 661 void Parse::modf() { 662 Node *f2 = pop(); 663 Node *f1 = pop(); 664 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::modf_Type(), 665 CAST_FROM_FN_PTR(address, SharedRuntime::frem), 666 "frem", NULL, //no memory effects 667 f1, f2); 668 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0)); 669 670 push(res); 671 } 672 673 void Parse::modd() { 674 Node *d2 = pop_pair(); 675 Node *d1 = pop_pair(); 676 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::Math_DD_D_Type(), 677 CAST_FROM_FN_PTR(address, SharedRuntime::drem), 678 "drem", NULL, //no memory effects 679 d1, top(), d2, top()); 680 Node* res_d = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0)); 681 682 #ifdef ASSERT 683 Node* res_top = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 1)); 684 assert(res_top == top(), "second value must be top"); 685 #endif 686 687 push_pair(res_d); 688 } 689 690 void Parse::l2f() { 691 Node* f2 = pop(); 692 Node* f1 = pop(); 693 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::l2f_Type(), 694 CAST_FROM_FN_PTR(address, SharedRuntime::l2f), 695 "l2f", NULL, //no memory effects 696 f1, f2); 697 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0)); 698 699 push(res); 700 } 701 702 void Parse::do_irem() { 703 // Must keep both values on the expression-stack during null-check 704 zero_check_int(peek()); 705 // Compile-time detect of null-exception? 706 if (stopped()) return; 707 708 Node* b = pop(); 709 Node* a = pop(); 710 711 const Type *t = _gvn.type(b); 712 if (t != Type::TOP) { 713 const TypeInt *ti = t->is_int(); 714 if (ti->is_con()) { 715 int divisor = ti->get_con(); 716 // check for positive power of 2 717 if (divisor > 0 && 718 (divisor & ~(divisor-1)) == divisor) { 719 // yes ! 720 Node *mask = _gvn.intcon((divisor - 1)); 721 // Sigh, must handle negative dividends 722 Node *zero = _gvn.intcon(0); 723 IfNode *ifff = jump_if_fork_int(a, zero, BoolTest::lt); 724 Node *iff = _gvn.transform( new IfFalseNode(ifff) ); 725 Node *ift = _gvn.transform( new IfTrueNode (ifff) ); 726 Node *reg = jump_if_join(ift, iff); 727 Node *phi = PhiNode::make(reg, NULL, TypeInt::INT); 728 // Negative path; negate/and/negate 729 Node *neg = _gvn.transform( new SubINode(zero, a) ); 730 Node *andn= _gvn.transform( new AndINode(neg, mask) ); 731 Node *negn= _gvn.transform( new SubINode(zero, andn) ); 732 phi->init_req(1, negn); 733 // Fast positive case 734 Node *andx = _gvn.transform( new AndINode(a, mask) ); 735 phi->init_req(2, andx); 736 // Push the merge 737 push( _gvn.transform(phi) ); 738 return; 739 } 740 } 741 } 742 // Default case 743 push( _gvn.transform( new ModINode(control(),a,b) ) ); 744 } 745 746 // Handle jsr and jsr_w bytecode 747 void Parse::do_jsr() { 748 assert(bc() == Bytecodes::_jsr || bc() == Bytecodes::_jsr_w, "wrong bytecode"); 749 750 // Store information about current state, tagged with new _jsr_bci 751 int return_bci = iter().next_bci(); 752 int jsr_bci = (bc() == Bytecodes::_jsr) ? iter().get_dest() : iter().get_far_dest(); 753 754 // Update method data 755 profile_taken_branch(jsr_bci); 756 757 // The way we do things now, there is only one successor block 758 // for the jsr, because the target code is cloned by ciTypeFlow. 759 Block* target = successor_for_bci(jsr_bci); 760 761 // What got pushed? 762 const Type* ret_addr = target->peek(); 763 assert(ret_addr->singleton(), "must be a constant (cloned jsr body)"); 764 765 // Effect on jsr on stack 766 push(_gvn.makecon(ret_addr)); 767 768 // Flow to the jsr. 769 merge(jsr_bci); 770 } 771 772 // Handle ret bytecode 773 void Parse::do_ret() { 774 // Find to whom we return. 775 assert(block()->num_successors() == 1, "a ret can only go one place now"); 776 Block* target = block()->successor_at(0); 777 assert(!target->is_ready(), "our arrival must be expected"); 778 profile_ret(target->flow()->start()); 779 int pnum = target->next_path_num(); 780 merge_common(target, pnum); 781 } 782 783 static bool has_injected_profile(BoolTest::mask btest, Node* test, int& taken, int& not_taken) { 784 if (btest != BoolTest::eq && btest != BoolTest::ne) { 785 // Only ::eq and ::ne are supported for profile injection. 786 return false; 787 } 788 if (test->is_Cmp() && 789 test->in(1)->Opcode() == Op_ProfileBoolean) { 790 ProfileBooleanNode* profile = (ProfileBooleanNode*)test->in(1); 791 int false_cnt = profile->false_count(); 792 int true_cnt = profile->true_count(); 793 794 // Counts matching depends on the actual test operation (::eq or ::ne). 795 // No need to scale the counts because profile injection was designed 796 // to feed exact counts into VM. 797 taken = (btest == BoolTest::eq) ? false_cnt : true_cnt; 798 not_taken = (btest == BoolTest::eq) ? true_cnt : false_cnt; 799 800 profile->consume(); 801 return true; 802 } 803 return false; 804 } 805 //--------------------------dynamic_branch_prediction-------------------------- 806 // Try to gather dynamic branch prediction behavior. Return a probability 807 // of the branch being taken and set the "cnt" field. Returns a -1.0 808 // if we need to use static prediction for some reason. 809 float Parse::dynamic_branch_prediction(float &cnt, BoolTest::mask btest, Node* test) { 810 ResourceMark rm; 811 812 cnt = COUNT_UNKNOWN; 813 814 int taken = 0; 815 int not_taken = 0; 816 817 bool use_mdo = !has_injected_profile(btest, test, taken, not_taken); 818 819 if (use_mdo) { 820 // Use MethodData information if it is available 821 // FIXME: free the ProfileData structure 822 ciMethodData* methodData = method()->method_data(); 823 if (!methodData->is_mature()) return PROB_UNKNOWN; 824 ciProfileData* data = methodData->bci_to_data(bci()); 825 if (!data->is_JumpData()) return PROB_UNKNOWN; 826 827 // get taken and not taken values 828 taken = data->as_JumpData()->taken(); 829 not_taken = 0; 830 if (data->is_BranchData()) { 831 not_taken = data->as_BranchData()->not_taken(); 832 } 833 834 // scale the counts to be commensurate with invocation counts: 835 taken = method()->scale_count(taken); 836 not_taken = method()->scale_count(not_taken); 837 } 838 839 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful. 840 // We also check that individual counters are positive first, otherwise the sum can become positive. 841 if (taken < 0 || not_taken < 0 || taken + not_taken < 40) { 842 if (C->log() != NULL) { 843 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d'", iter().get_dest(), taken, not_taken); 844 } 845 return PROB_UNKNOWN; 846 } 847 848 // Compute frequency that we arrive here 849 float sum = taken + not_taken; 850 // Adjust, if this block is a cloned private block but the 851 // Jump counts are shared. Taken the private counts for 852 // just this path instead of the shared counts. 853 if( block()->count() > 0 ) 854 sum = block()->count(); 855 cnt = sum / FreqCountInvocations; 856 857 // Pin probability to sane limits 858 float prob; 859 if( !taken ) 860 prob = (0+PROB_MIN) / 2; 861 else if( !not_taken ) 862 prob = (1+PROB_MAX) / 2; 863 else { // Compute probability of true path 864 prob = (float)taken / (float)(taken + not_taken); 865 if (prob > PROB_MAX) prob = PROB_MAX; 866 if (prob < PROB_MIN) prob = PROB_MIN; 867 } 868 869 assert((cnt > 0.0f) && (prob > 0.0f), 870 "Bad frequency assignment in if"); 871 872 if (C->log() != NULL) { 873 const char* prob_str = NULL; 874 if (prob >= PROB_MAX) prob_str = (prob == PROB_MAX) ? "max" : "always"; 875 if (prob <= PROB_MIN) prob_str = (prob == PROB_MIN) ? "min" : "never"; 876 char prob_str_buf[30]; 877 if (prob_str == NULL) { 878 sprintf(prob_str_buf, "%g", prob); 879 prob_str = prob_str_buf; 880 } 881 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d' cnt='%f' prob='%s'", 882 iter().get_dest(), taken, not_taken, cnt, prob_str); 883 } 884 return prob; 885 } 886 887 //-----------------------------branch_prediction------------------------------- 888 float Parse::branch_prediction(float& cnt, 889 BoolTest::mask btest, 890 int target_bci, 891 Node* test) { 892 float prob = dynamic_branch_prediction(cnt, btest, test); 893 // If prob is unknown, switch to static prediction 894 if (prob != PROB_UNKNOWN) return prob; 895 896 prob = PROB_FAIR; // Set default value 897 if (btest == BoolTest::eq) // Exactly equal test? 898 prob = PROB_STATIC_INFREQUENT; // Assume its relatively infrequent 899 else if (btest == BoolTest::ne) 900 prob = PROB_STATIC_FREQUENT; // Assume its relatively frequent 901 902 // If this is a conditional test guarding a backwards branch, 903 // assume its a loop-back edge. Make it a likely taken branch. 904 if (target_bci < bci()) { 905 if (is_osr_parse()) { // Could be a hot OSR'd loop; force deopt 906 // Since it's an OSR, we probably have profile data, but since 907 // branch_prediction returned PROB_UNKNOWN, the counts are too small. 908 // Let's make a special check here for completely zero counts. 909 ciMethodData* methodData = method()->method_data(); 910 if (!methodData->is_empty()) { 911 ciProfileData* data = methodData->bci_to_data(bci()); 912 // Only stop for truly zero counts, which mean an unknown part 913 // of the OSR-ed method, and we want to deopt to gather more stats. 914 // If you have ANY counts, then this loop is simply 'cold' relative 915 // to the OSR loop. 916 if (data->as_BranchData()->taken() + 917 data->as_BranchData()->not_taken() == 0 ) { 918 // This is the only way to return PROB_UNKNOWN: 919 return PROB_UNKNOWN; 920 } 921 } 922 } 923 prob = PROB_STATIC_FREQUENT; // Likely to take backwards branch 924 } 925 926 assert(prob != PROB_UNKNOWN, "must have some guess at this point"); 927 return prob; 928 } 929 930 // The magic constants are chosen so as to match the output of 931 // branch_prediction() when the profile reports a zero taken count. 932 // It is important to distinguish zero counts unambiguously, because 933 // some branches (e.g., _213_javac.Assembler.eliminate) validly produce 934 // very small but nonzero probabilities, which if confused with zero 935 // counts would keep the program recompiling indefinitely. 936 bool Parse::seems_never_taken(float prob) const { 937 return prob < PROB_MIN; 938 } 939 940 // True if the comparison seems to be the kind that will not change its 941 // statistics from true to false. See comments in adjust_map_after_if. 942 // This question is only asked along paths which are already 943 // classifed as untaken (by seems_never_taken), so really, 944 // if a path is never taken, its controlling comparison is 945 // already acting in a stable fashion. If the comparison 946 // seems stable, we will put an expensive uncommon trap 947 // on the untaken path. 948 bool Parse::seems_stable_comparison() const { 949 if (C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if)) { 950 return false; 951 } 952 return true; 953 } 954 955 //-------------------------------repush_if_args-------------------------------- 956 // Push arguments of an "if" bytecode back onto the stack by adjusting _sp. 957 inline int Parse::repush_if_args() { 958 if (PrintOpto && WizardMode) { 959 tty->print("defending against excessive implicit null exceptions on %s @%d in ", 960 Bytecodes::name(iter().cur_bc()), iter().cur_bci()); 961 method()->print_name(); tty->cr(); 962 } 963 int bc_depth = - Bytecodes::depth(iter().cur_bc()); 964 assert(bc_depth == 1 || bc_depth == 2, "only two kinds of branches"); 965 DEBUG_ONLY(sync_jvms()); // argument(n) requires a synced jvms 966 assert(argument(0) != NULL, "must exist"); 967 assert(bc_depth == 1 || argument(1) != NULL, "two must exist"); 968 inc_sp(bc_depth); 969 return bc_depth; 970 } 971 972 //----------------------------------do_ifnull---------------------------------- 973 void Parse::do_ifnull(BoolTest::mask btest, Node *c) { 974 int target_bci = iter().get_dest(); 975 976 Block* branch_block = successor_for_bci(target_bci); 977 Block* next_block = successor_for_bci(iter().next_bci()); 978 979 float cnt; 980 float prob = branch_prediction(cnt, btest, target_bci, c); 981 if (prob == PROB_UNKNOWN) { 982 // (An earlier version of do_ifnull omitted this trap for OSR methods.) 983 if (PrintOpto && Verbose) { 984 tty->print_cr("Never-taken edge stops compilation at bci %d", bci()); 985 } 986 repush_if_args(); // to gather stats on loop 987 // We need to mark this branch as taken so that if we recompile we will 988 // see that it is possible. In the tiered system the interpreter doesn't 989 // do profiling and by the time we get to the lower tier from the interpreter 990 // the path may be cold again. Make sure it doesn't look untaken 991 profile_taken_branch(target_bci, !ProfileInterpreter); 992 uncommon_trap(Deoptimization::Reason_unreached, 993 Deoptimization::Action_reinterpret, 994 NULL, "cold"); 995 if (C->eliminate_boxing()) { 996 // Mark the successor blocks as parsed 997 branch_block->next_path_num(); 998 next_block->next_path_num(); 999 } 1000 return; 1001 } 1002 1003 NOT_PRODUCT(explicit_null_checks_inserted++); 1004 1005 // Generate real control flow 1006 Node *tst = _gvn.transform( new BoolNode( c, btest ) ); 1007 1008 // Sanity check the probability value 1009 assert(prob > 0.0f,"Bad probability in Parser"); 1010 // Need xform to put node in hash table 1011 IfNode *iff = create_and_xform_if( control(), tst, prob, cnt ); 1012 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser"); 1013 // True branch 1014 { PreserveJVMState pjvms(this); 1015 Node* iftrue = _gvn.transform( new IfTrueNode (iff) ); 1016 set_control(iftrue); 1017 1018 if (stopped()) { // Path is dead? 1019 NOT_PRODUCT(explicit_null_checks_elided++); 1020 if (C->eliminate_boxing()) { 1021 // Mark the successor block as parsed 1022 branch_block->next_path_num(); 1023 } 1024 } else { // Path is live. 1025 // Update method data 1026 profile_taken_branch(target_bci); 1027 adjust_map_after_if(btest, c, prob, branch_block, next_block); 1028 if (!stopped()) { 1029 merge(target_bci); 1030 } 1031 } 1032 } 1033 1034 // False branch 1035 Node* iffalse = _gvn.transform( new IfFalseNode(iff) ); 1036 set_control(iffalse); 1037 1038 if (stopped()) { // Path is dead? 1039 NOT_PRODUCT(explicit_null_checks_elided++); 1040 if (C->eliminate_boxing()) { 1041 // Mark the successor block as parsed 1042 next_block->next_path_num(); 1043 } 1044 } else { // Path is live. 1045 // Update method data 1046 profile_not_taken_branch(); 1047 adjust_map_after_if(BoolTest(btest).negate(), c, 1.0-prob, 1048 next_block, branch_block); 1049 } 1050 } 1051 1052 //------------------------------------do_if------------------------------------ 1053 void Parse::do_if(BoolTest::mask btest, Node* c) { 1054 int target_bci = iter().get_dest(); 1055 1056 Block* branch_block = successor_for_bci(target_bci); 1057 Block* next_block = successor_for_bci(iter().next_bci()); 1058 1059 float cnt; 1060 float prob = branch_prediction(cnt, btest, target_bci, c); 1061 float untaken_prob = 1.0 - prob; 1062 1063 if (prob == PROB_UNKNOWN) { 1064 if (PrintOpto && Verbose) { 1065 tty->print_cr("Never-taken edge stops compilation at bci %d", bci()); 1066 } 1067 repush_if_args(); // to gather stats on loop 1068 // We need to mark this branch as taken so that if we recompile we will 1069 // see that it is possible. In the tiered system the interpreter doesn't 1070 // do profiling and by the time we get to the lower tier from the interpreter 1071 // the path may be cold again. Make sure it doesn't look untaken 1072 profile_taken_branch(target_bci, !ProfileInterpreter); 1073 uncommon_trap(Deoptimization::Reason_unreached, 1074 Deoptimization::Action_reinterpret, 1075 NULL, "cold"); 1076 if (C->eliminate_boxing()) { 1077 // Mark the successor blocks as parsed 1078 branch_block->next_path_num(); 1079 next_block->next_path_num(); 1080 } 1081 return; 1082 } 1083 1084 // Sanity check the probability value 1085 assert(0.0f < prob && prob < 1.0f,"Bad probability in Parser"); 1086 1087 bool taken_if_true = true; 1088 // Convert BoolTest to canonical form: 1089 if (!BoolTest(btest).is_canonical()) { 1090 btest = BoolTest(btest).negate(); 1091 taken_if_true = false; 1092 // prob is NOT updated here; it remains the probability of the taken 1093 // path (as opposed to the prob of the path guarded by an 'IfTrueNode'). 1094 } 1095 assert(btest != BoolTest::eq, "!= is the only canonical exact test"); 1096 1097 Node* tst0 = new BoolNode(c, btest); 1098 Node* tst = _gvn.transform(tst0); 1099 BoolTest::mask taken_btest = BoolTest::illegal; 1100 BoolTest::mask untaken_btest = BoolTest::illegal; 1101 1102 if (tst->is_Bool()) { 1103 // Refresh c from the transformed bool node, since it may be 1104 // simpler than the original c. Also re-canonicalize btest. 1105 // This wins when (Bool ne (Conv2B p) 0) => (Bool ne (CmpP p NULL)). 1106 // That can arise from statements like: if (x instanceof C) ... 1107 if (tst != tst0) { 1108 // Canonicalize one more time since transform can change it. 1109 btest = tst->as_Bool()->_test._test; 1110 if (!BoolTest(btest).is_canonical()) { 1111 // Reverse edges one more time... 1112 tst = _gvn.transform( tst->as_Bool()->negate(&_gvn) ); 1113 btest = tst->as_Bool()->_test._test; 1114 assert(BoolTest(btest).is_canonical(), "sanity"); 1115 taken_if_true = !taken_if_true; 1116 } 1117 c = tst->in(1); 1118 } 1119 BoolTest::mask neg_btest = BoolTest(btest).negate(); 1120 taken_btest = taken_if_true ? btest : neg_btest; 1121 untaken_btest = taken_if_true ? neg_btest : btest; 1122 } 1123 1124 // Generate real control flow 1125 float true_prob = (taken_if_true ? prob : untaken_prob); 1126 IfNode* iff = create_and_map_if(control(), tst, true_prob, cnt); 1127 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser"); 1128 Node* taken_branch = new IfTrueNode(iff); 1129 Node* untaken_branch = new IfFalseNode(iff); 1130 if (!taken_if_true) { // Finish conversion to canonical form 1131 Node* tmp = taken_branch; 1132 taken_branch = untaken_branch; 1133 untaken_branch = tmp; 1134 } 1135 1136 // Branch is taken: 1137 { PreserveJVMState pjvms(this); 1138 taken_branch = _gvn.transform(taken_branch); 1139 set_control(taken_branch); 1140 1141 if (stopped()) { 1142 if (C->eliminate_boxing()) { 1143 // Mark the successor block as parsed 1144 branch_block->next_path_num(); 1145 } 1146 } else { 1147 // Update method data 1148 profile_taken_branch(target_bci); 1149 adjust_map_after_if(taken_btest, c, prob, branch_block, next_block); 1150 if (!stopped()) { 1151 merge(target_bci); 1152 } 1153 } 1154 } 1155 1156 untaken_branch = _gvn.transform(untaken_branch); 1157 set_control(untaken_branch); 1158 1159 // Branch not taken. 1160 if (stopped()) { 1161 if (C->eliminate_boxing()) { 1162 // Mark the successor block as parsed 1163 next_block->next_path_num(); 1164 } 1165 } else { 1166 // Update method data 1167 profile_not_taken_branch(); 1168 adjust_map_after_if(untaken_btest, c, untaken_prob, 1169 next_block, branch_block); 1170 } 1171 } 1172 1173 bool Parse::path_is_suitable_for_uncommon_trap(float prob) const { 1174 // Don't want to speculate on uncommon traps when running with -Xcomp 1175 if (!UseInterpreter) { 1176 return false; 1177 } 1178 return (seems_never_taken(prob) && seems_stable_comparison()); 1179 } 1180 1181 //----------------------------adjust_map_after_if------------------------------ 1182 // Adjust the JVM state to reflect the result of taking this path. 1183 // Basically, it means inspecting the CmpNode controlling this 1184 // branch, seeing how it constrains a tested value, and then 1185 // deciding if it's worth our while to encode this constraint 1186 // as graph nodes in the current abstract interpretation map. 1187 void Parse::adjust_map_after_if(BoolTest::mask btest, Node* c, float prob, 1188 Block* path, Block* other_path) { 1189 if (stopped() || !c->is_Cmp() || btest == BoolTest::illegal) 1190 return; // nothing to do 1191 1192 bool is_fallthrough = (path == successor_for_bci(iter().next_bci())); 1193 1194 if (path_is_suitable_for_uncommon_trap(prob)) { 1195 repush_if_args(); 1196 uncommon_trap(Deoptimization::Reason_unstable_if, 1197 Deoptimization::Action_reinterpret, 1198 NULL, 1199 (is_fallthrough ? "taken always" : "taken never")); 1200 return; 1201 } 1202 1203 Node* val = c->in(1); 1204 Node* con = c->in(2); 1205 const Type* tcon = _gvn.type(con); 1206 const Type* tval = _gvn.type(val); 1207 bool have_con = tcon->singleton(); 1208 if (tval->singleton()) { 1209 if (!have_con) { 1210 // Swap, so constant is in con. 1211 con = val; 1212 tcon = tval; 1213 val = c->in(2); 1214 tval = _gvn.type(val); 1215 btest = BoolTest(btest).commute(); 1216 have_con = true; 1217 } else { 1218 // Do we have two constants? Then leave well enough alone. 1219 have_con = false; 1220 } 1221 } 1222 if (!have_con) // remaining adjustments need a con 1223 return; 1224 1225 sharpen_type_after_if(btest, con, tcon, val, tval); 1226 } 1227 1228 1229 static Node* extract_obj_from_klass_load(PhaseGVN* gvn, Node* n) { 1230 Node* ldk; 1231 if (n->is_DecodeNKlass()) { 1232 if (n->in(1)->Opcode() != Op_LoadNKlass) { 1233 return NULL; 1234 } else { 1235 ldk = n->in(1); 1236 } 1237 } else if (n->Opcode() != Op_LoadKlass) { 1238 return NULL; 1239 } else { 1240 ldk = n; 1241 } 1242 assert(ldk != NULL && ldk->is_Load(), "should have found a LoadKlass or LoadNKlass node"); 1243 1244 Node* adr = ldk->in(MemNode::Address); 1245 intptr_t off = 0; 1246 Node* obj = AddPNode::Ideal_base_and_offset(adr, gvn, off); 1247 if (obj == NULL || off != oopDesc::klass_offset_in_bytes()) // loading oopDesc::_klass? 1248 return NULL; 1249 const TypePtr* tp = gvn->type(obj)->is_ptr(); 1250 if (tp == NULL || !(tp->isa_instptr() || tp->isa_aryptr())) // is obj a Java object ptr? 1251 return NULL; 1252 1253 return obj; 1254 } 1255 1256 void Parse::sharpen_type_after_if(BoolTest::mask btest, 1257 Node* con, const Type* tcon, 1258 Node* val, const Type* tval) { 1259 // Look for opportunities to sharpen the type of a node 1260 // whose klass is compared with a constant klass. 1261 if (btest == BoolTest::eq && tcon->isa_klassptr()) { 1262 Node* obj = extract_obj_from_klass_load(&_gvn, val); 1263 const TypeOopPtr* con_type = tcon->isa_klassptr()->as_instance_type(); 1264 if (obj != NULL && (con_type->isa_instptr() || con_type->isa_aryptr())) { 1265 // Found: 1266 // Bool(CmpP(LoadKlass(obj._klass), ConP(Foo.klass)), [eq]) 1267 // or the narrowOop equivalent. 1268 const Type* obj_type = _gvn.type(obj); 1269 const TypeOopPtr* tboth = obj_type->join_speculative(con_type)->isa_oopptr(); 1270 if (tboth != NULL && tboth->klass_is_exact() && tboth != obj_type && 1271 tboth->higher_equal(obj_type)) { 1272 // obj has to be of the exact type Foo if the CmpP succeeds. 1273 int obj_in_map = map()->find_edge(obj); 1274 JVMState* jvms = this->jvms(); 1275 if (obj_in_map >= 0 && 1276 (jvms->is_loc(obj_in_map) || jvms->is_stk(obj_in_map))) { 1277 TypeNode* ccast = new CheckCastPPNode(control(), obj, tboth); 1278 const Type* tcc = ccast->as_Type()->type(); 1279 assert(tcc != obj_type && tcc->higher_equal(obj_type), "must improve"); 1280 // Delay transform() call to allow recovery of pre-cast value 1281 // at the control merge. 1282 _gvn.set_type_bottom(ccast); 1283 record_for_igvn(ccast); 1284 // Here's the payoff. 1285 replace_in_map(obj, ccast); 1286 } 1287 } 1288 } 1289 } 1290 1291 int val_in_map = map()->find_edge(val); 1292 if (val_in_map < 0) return; // replace_in_map would be useless 1293 { 1294 JVMState* jvms = this->jvms(); 1295 if (!(jvms->is_loc(val_in_map) || 1296 jvms->is_stk(val_in_map))) 1297 return; // again, it would be useless 1298 } 1299 1300 // Check for a comparison to a constant, and "know" that the compared 1301 // value is constrained on this path. 1302 assert(tcon->singleton(), ""); 1303 ConstraintCastNode* ccast = NULL; 1304 Node* cast = NULL; 1305 1306 switch (btest) { 1307 case BoolTest::eq: // Constant test? 1308 { 1309 const Type* tboth = tcon->join_speculative(tval); 1310 if (tboth == tval) break; // Nothing to gain. 1311 if (tcon->isa_int()) { 1312 ccast = new CastIINode(val, tboth); 1313 } else if (tcon == TypePtr::NULL_PTR) { 1314 // Cast to null, but keep the pointer identity temporarily live. 1315 ccast = new CastPPNode(val, tboth); 1316 } else { 1317 const TypeF* tf = tcon->isa_float_constant(); 1318 const TypeD* td = tcon->isa_double_constant(); 1319 // Exclude tests vs float/double 0 as these could be 1320 // either +0 or -0. Just because you are equal to +0 1321 // doesn't mean you ARE +0! 1322 // Note, following code also replaces Long and Oop values. 1323 if ((!tf || tf->_f != 0.0) && 1324 (!td || td->_d != 0.0)) 1325 cast = con; // Replace non-constant val by con. 1326 } 1327 } 1328 break; 1329 1330 case BoolTest::ne: 1331 if (tcon == TypePtr::NULL_PTR) { 1332 cast = cast_not_null(val, false); 1333 } 1334 break; 1335 1336 default: 1337 // (At this point we could record int range types with CastII.) 1338 break; 1339 } 1340 1341 if (ccast != NULL) { 1342 const Type* tcc = ccast->as_Type()->type(); 1343 assert(tcc != tval && tcc->higher_equal(tval), "must improve"); 1344 // Delay transform() call to allow recovery of pre-cast value 1345 // at the control merge. 1346 ccast->set_req(0, control()); 1347 _gvn.set_type_bottom(ccast); 1348 record_for_igvn(ccast); 1349 cast = ccast; 1350 } 1351 1352 if (cast != NULL) { // Here's the payoff. 1353 replace_in_map(val, cast); 1354 } 1355 } 1356 1357 /** 1358 * Use speculative type to optimize CmpP node: if comparison is 1359 * against the low level class, cast the object to the speculative 1360 * type if any. CmpP should then go away. 1361 * 1362 * @param c expected CmpP node 1363 * @return result of CmpP on object casted to speculative type 1364 * 1365 */ 1366 Node* Parse::optimize_cmp_with_klass(Node* c) { 1367 // If this is transformed by the _gvn to a comparison with the low 1368 // level klass then we may be able to use speculation 1369 if (c->Opcode() == Op_CmpP && 1370 (c->in(1)->Opcode() == Op_LoadKlass || c->in(1)->Opcode() == Op_DecodeNKlass) && 1371 c->in(2)->is_Con()) { 1372 Node* load_klass = NULL; 1373 Node* decode = NULL; 1374 if (c->in(1)->Opcode() == Op_DecodeNKlass) { 1375 decode = c->in(1); 1376 load_klass = c->in(1)->in(1); 1377 } else { 1378 load_klass = c->in(1); 1379 } 1380 if (load_klass->in(2)->is_AddP()) { 1381 Node* addp = load_klass->in(2); 1382 Node* obj = addp->in(AddPNode::Address); 1383 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 1384 if (obj_type->speculative_type_not_null() != NULL) { 1385 ciKlass* k = obj_type->speculative_type(); 1386 inc_sp(2); 1387 obj = maybe_cast_profiled_obj(obj, k); 1388 dec_sp(2); 1389 // Make the CmpP use the casted obj 1390 addp = basic_plus_adr(obj, addp->in(AddPNode::Offset)); 1391 load_klass = load_klass->clone(); 1392 load_klass->set_req(2, addp); 1393 load_klass = _gvn.transform(load_klass); 1394 if (decode != NULL) { 1395 decode = decode->clone(); 1396 decode->set_req(1, load_klass); 1397 load_klass = _gvn.transform(decode); 1398 } 1399 c = c->clone(); 1400 c->set_req(1, load_klass); 1401 c = _gvn.transform(c); 1402 } 1403 } 1404 } 1405 return c; 1406 } 1407 1408 //------------------------------do_one_bytecode-------------------------------- 1409 // Parse this bytecode, and alter the Parsers JVM->Node mapping 1410 void Parse::do_one_bytecode() { 1411 Node *a, *b, *c, *d; // Handy temps 1412 BoolTest::mask btest; 1413 int i; 1414 1415 assert(!has_exceptions(), "bytecode entry state must be clear of throws"); 1416 1417 if (C->check_node_count(NodeLimitFudgeFactor * 5, 1418 "out of nodes parsing method")) { 1419 return; 1420 } 1421 1422 #ifdef ASSERT 1423 // for setting breakpoints 1424 if (TraceOptoParse) { 1425 tty->print(" @"); 1426 dump_bci(bci()); 1427 tty->cr(); 1428 } 1429 #endif 1430 1431 switch (bc()) { 1432 case Bytecodes::_nop: 1433 // do nothing 1434 break; 1435 case Bytecodes::_lconst_0: 1436 push_pair(longcon(0)); 1437 break; 1438 1439 case Bytecodes::_lconst_1: 1440 push_pair(longcon(1)); 1441 break; 1442 1443 case Bytecodes::_fconst_0: 1444 push(zerocon(T_FLOAT)); 1445 break; 1446 1447 case Bytecodes::_fconst_1: 1448 push(makecon(TypeF::ONE)); 1449 break; 1450 1451 case Bytecodes::_fconst_2: 1452 push(makecon(TypeF::make(2.0f))); 1453 break; 1454 1455 case Bytecodes::_dconst_0: 1456 push_pair(zerocon(T_DOUBLE)); 1457 break; 1458 1459 case Bytecodes::_dconst_1: 1460 push_pair(makecon(TypeD::ONE)); 1461 break; 1462 1463 case Bytecodes::_iconst_m1:push(intcon(-1)); break; 1464 case Bytecodes::_iconst_0: push(intcon( 0)); break; 1465 case Bytecodes::_iconst_1: push(intcon( 1)); break; 1466 case Bytecodes::_iconst_2: push(intcon( 2)); break; 1467 case Bytecodes::_iconst_3: push(intcon( 3)); break; 1468 case Bytecodes::_iconst_4: push(intcon( 4)); break; 1469 case Bytecodes::_iconst_5: push(intcon( 5)); break; 1470 case Bytecodes::_bipush: push(intcon(iter().get_constant_u1())); break; 1471 case Bytecodes::_sipush: push(intcon(iter().get_constant_u2())); break; 1472 case Bytecodes::_aconst_null: push(null()); break; 1473 case Bytecodes::_ldc: 1474 case Bytecodes::_ldc_w: 1475 case Bytecodes::_ldc2_w: 1476 // If the constant is unresolved, run this BC once in the interpreter. 1477 { 1478 ciConstant constant = iter().get_constant(); 1479 if (constant.basic_type() == T_OBJECT && 1480 !constant.as_object()->is_loaded()) { 1481 int index = iter().get_constant_pool_index(); 1482 constantTag tag = iter().get_constant_pool_tag(index); 1483 uncommon_trap(Deoptimization::make_trap_request 1484 (Deoptimization::Reason_unloaded, 1485 Deoptimization::Action_reinterpret, 1486 index), 1487 NULL, tag.internal_name()); 1488 break; 1489 } 1490 assert(constant.basic_type() != T_OBJECT || constant.as_object()->is_instance(), 1491 "must be java_mirror of klass"); 1492 const Type* con_type = Type::make_from_constant(constant); 1493 if (con_type != NULL) { 1494 push_node(con_type->basic_type(), makecon(con_type)); 1495 } 1496 } 1497 1498 break; 1499 1500 case Bytecodes::_aload_0: 1501 push( local(0) ); 1502 break; 1503 case Bytecodes::_aload_1: 1504 push( local(1) ); 1505 break; 1506 case Bytecodes::_aload_2: 1507 push( local(2) ); 1508 break; 1509 case Bytecodes::_aload_3: 1510 push( local(3) ); 1511 break; 1512 case Bytecodes::_aload: 1513 case Bytecodes::_vload: 1514 push( local(iter().get_index()) ); 1515 break; 1516 1517 case Bytecodes::_fload_0: 1518 case Bytecodes::_iload_0: 1519 push( local(0) ); 1520 break; 1521 case Bytecodes::_fload_1: 1522 case Bytecodes::_iload_1: 1523 push( local(1) ); 1524 break; 1525 case Bytecodes::_fload_2: 1526 case Bytecodes::_iload_2: 1527 push( local(2) ); 1528 break; 1529 case Bytecodes::_fload_3: 1530 case Bytecodes::_iload_3: 1531 push( local(3) ); 1532 break; 1533 case Bytecodes::_fload: 1534 case Bytecodes::_iload: 1535 push( local(iter().get_index()) ); 1536 break; 1537 case Bytecodes::_lload_0: 1538 push_pair_local( 0 ); 1539 break; 1540 case Bytecodes::_lload_1: 1541 push_pair_local( 1 ); 1542 break; 1543 case Bytecodes::_lload_2: 1544 push_pair_local( 2 ); 1545 break; 1546 case Bytecodes::_lload_3: 1547 push_pair_local( 3 ); 1548 break; 1549 case Bytecodes::_lload: 1550 push_pair_local( iter().get_index() ); 1551 break; 1552 1553 case Bytecodes::_dload_0: 1554 push_pair_local(0); 1555 break; 1556 case Bytecodes::_dload_1: 1557 push_pair_local(1); 1558 break; 1559 case Bytecodes::_dload_2: 1560 push_pair_local(2); 1561 break; 1562 case Bytecodes::_dload_3: 1563 push_pair_local(3); 1564 break; 1565 case Bytecodes::_dload: 1566 push_pair_local(iter().get_index()); 1567 break; 1568 case Bytecodes::_fstore_0: 1569 case Bytecodes::_istore_0: 1570 case Bytecodes::_astore_0: 1571 set_local( 0, pop() ); 1572 break; 1573 case Bytecodes::_fstore_1: 1574 case Bytecodes::_istore_1: 1575 case Bytecodes::_astore_1: 1576 set_local( 1, pop() ); 1577 break; 1578 case Bytecodes::_fstore_2: 1579 case Bytecodes::_istore_2: 1580 case Bytecodes::_astore_2: 1581 set_local( 2, pop() ); 1582 break; 1583 case Bytecodes::_fstore_3: 1584 case Bytecodes::_istore_3: 1585 case Bytecodes::_astore_3: 1586 set_local( 3, pop() ); 1587 break; 1588 case Bytecodes::_fstore: 1589 case Bytecodes::_istore: 1590 case Bytecodes::_astore: 1591 case Bytecodes::_vstore: 1592 set_local( iter().get_index(), pop() ); 1593 break; 1594 // long stores 1595 case Bytecodes::_lstore_0: 1596 set_pair_local( 0, pop_pair() ); 1597 break; 1598 case Bytecodes::_lstore_1: 1599 set_pair_local( 1, pop_pair() ); 1600 break; 1601 case Bytecodes::_lstore_2: 1602 set_pair_local( 2, pop_pair() ); 1603 break; 1604 case Bytecodes::_lstore_3: 1605 set_pair_local( 3, pop_pair() ); 1606 break; 1607 case Bytecodes::_lstore: 1608 set_pair_local( iter().get_index(), pop_pair() ); 1609 break; 1610 1611 // double stores 1612 case Bytecodes::_dstore_0: 1613 set_pair_local( 0, dstore_rounding(pop_pair()) ); 1614 break; 1615 case Bytecodes::_dstore_1: 1616 set_pair_local( 1, dstore_rounding(pop_pair()) ); 1617 break; 1618 case Bytecodes::_dstore_2: 1619 set_pair_local( 2, dstore_rounding(pop_pair()) ); 1620 break; 1621 case Bytecodes::_dstore_3: 1622 set_pair_local( 3, dstore_rounding(pop_pair()) ); 1623 break; 1624 case Bytecodes::_dstore: 1625 set_pair_local( iter().get_index(), dstore_rounding(pop_pair()) ); 1626 break; 1627 1628 case Bytecodes::_pop: dec_sp(1); break; 1629 case Bytecodes::_pop2: dec_sp(2); break; 1630 case Bytecodes::_swap: 1631 a = pop(); 1632 b = pop(); 1633 push(a); 1634 push(b); 1635 break; 1636 case Bytecodes::_dup: 1637 a = pop(); 1638 push(a); 1639 push(a); 1640 break; 1641 case Bytecodes::_dup_x1: 1642 a = pop(); 1643 b = pop(); 1644 push( a ); 1645 push( b ); 1646 push( a ); 1647 break; 1648 case Bytecodes::_dup_x2: 1649 a = pop(); 1650 b = pop(); 1651 c = pop(); 1652 push( a ); 1653 push( c ); 1654 push( b ); 1655 push( a ); 1656 break; 1657 case Bytecodes::_dup2: 1658 a = pop(); 1659 b = pop(); 1660 push( b ); 1661 push( a ); 1662 push( b ); 1663 push( a ); 1664 break; 1665 1666 case Bytecodes::_dup2_x1: 1667 // before: .. c, b, a 1668 // after: .. b, a, c, b, a 1669 // not tested 1670 a = pop(); 1671 b = pop(); 1672 c = pop(); 1673 push( b ); 1674 push( a ); 1675 push( c ); 1676 push( b ); 1677 push( a ); 1678 break; 1679 case Bytecodes::_dup2_x2: 1680 // before: .. d, c, b, a 1681 // after: .. b, a, d, c, b, a 1682 // not tested 1683 a = pop(); 1684 b = pop(); 1685 c = pop(); 1686 d = pop(); 1687 push( b ); 1688 push( a ); 1689 push( d ); 1690 push( c ); 1691 push( b ); 1692 push( a ); 1693 break; 1694 1695 case Bytecodes::_arraylength: { 1696 // Must do null-check with value on expression stack 1697 Node *ary = null_check(peek(), T_ARRAY); 1698 // Compile-time detect of null-exception? 1699 if (stopped()) return; 1700 a = pop(); 1701 push(load_array_length(a)); 1702 break; 1703 } 1704 1705 case Bytecodes::_baload: array_load(T_BYTE); break; 1706 case Bytecodes::_caload: array_load(T_CHAR); break; 1707 case Bytecodes::_iaload: array_load(T_INT); break; 1708 case Bytecodes::_saload: array_load(T_SHORT); break; 1709 case Bytecodes::_faload: array_load(T_FLOAT); break; 1710 case Bytecodes::_aaload: array_load(T_OBJECT); break; 1711 case Bytecodes::_laload: { 1712 a = array_addressing(T_LONG, 0); 1713 if (stopped()) return; // guaranteed null or range check 1714 dec_sp(2); // Pop array and index 1715 push_pair(make_load(control(), a, TypeLong::LONG, T_LONG, TypeAryPtr::LONGS, MemNode::unordered)); 1716 break; 1717 } 1718 case Bytecodes::_daload: { 1719 a = array_addressing(T_DOUBLE, 0); 1720 if (stopped()) return; // guaranteed null or range check 1721 dec_sp(2); // Pop array and index 1722 push_pair(make_load(control(), a, Type::DOUBLE, T_DOUBLE, TypeAryPtr::DOUBLES, MemNode::unordered)); 1723 break; 1724 } 1725 case Bytecodes::_bastore: array_store(T_BYTE); break; 1726 case Bytecodes::_castore: array_store(T_CHAR); break; 1727 case Bytecodes::_iastore: array_store(T_INT); break; 1728 case Bytecodes::_sastore: array_store(T_SHORT); break; 1729 case Bytecodes::_fastore: array_store(T_FLOAT); break; 1730 case Bytecodes::_aastore: { 1731 d = array_addressing(T_OBJECT, 1); 1732 if (stopped()) return; // guaranteed null or range check 1733 array_store_check(); 1734 c = pop(); // Oop to store 1735 b = pop(); // index (already used) 1736 a = pop(); // the array itself 1737 const TypeOopPtr* elemtype = _gvn.type(a)->is_aryptr()->elem()->make_oopptr(); 1738 const TypeAryPtr* adr_type = TypeAryPtr::OOPS; 1739 Node* store = store_oop_to_array(control(), a, d, adr_type, c, elemtype, T_OBJECT, 1740 StoreNode::release_if_reference(T_OBJECT)); 1741 break; 1742 } 1743 case Bytecodes::_lastore: { 1744 a = array_addressing(T_LONG, 2); 1745 if (stopped()) return; // guaranteed null or range check 1746 c = pop_pair(); 1747 dec_sp(2); // Pop array and index 1748 store_to_memory(control(), a, c, T_LONG, TypeAryPtr::LONGS, MemNode::unordered); 1749 break; 1750 } 1751 case Bytecodes::_dastore: { 1752 a = array_addressing(T_DOUBLE, 2); 1753 if (stopped()) return; // guaranteed null or range check 1754 c = pop_pair(); 1755 dec_sp(2); // Pop array and index 1756 c = dstore_rounding(c); 1757 store_to_memory(control(), a, c, T_DOUBLE, TypeAryPtr::DOUBLES, MemNode::unordered); 1758 break; 1759 } 1760 1761 case Bytecodes::_getfield: 1762 do_getfield(); 1763 break; 1764 1765 case Bytecodes::_vgetfield: 1766 do_vgetfield(); 1767 break; 1768 1769 case Bytecodes::_getstatic: 1770 do_getstatic(); 1771 break; 1772 1773 case Bytecodes::_putfield: 1774 do_putfield(); 1775 break; 1776 1777 case Bytecodes::_putstatic: 1778 do_putstatic(); 1779 break; 1780 1781 case Bytecodes::_irem: 1782 do_irem(); 1783 break; 1784 case Bytecodes::_idiv: 1785 // Must keep both values on the expression-stack during null-check 1786 zero_check_int(peek()); 1787 // Compile-time detect of null-exception? 1788 if (stopped()) return; 1789 b = pop(); 1790 a = pop(); 1791 push( _gvn.transform( new DivINode(control(),a,b) ) ); 1792 break; 1793 case Bytecodes::_imul: 1794 b = pop(); a = pop(); 1795 push( _gvn.transform( new MulINode(a,b) ) ); 1796 break; 1797 case Bytecodes::_iadd: 1798 b = pop(); a = pop(); 1799 push( _gvn.transform( new AddINode(a,b) ) ); 1800 break; 1801 case Bytecodes::_ineg: 1802 a = pop(); 1803 push( _gvn.transform( new SubINode(_gvn.intcon(0),a)) ); 1804 break; 1805 case Bytecodes::_isub: 1806 b = pop(); a = pop(); 1807 push( _gvn.transform( new SubINode(a,b) ) ); 1808 break; 1809 case Bytecodes::_iand: 1810 b = pop(); a = pop(); 1811 push( _gvn.transform( new AndINode(a,b) ) ); 1812 break; 1813 case Bytecodes::_ior: 1814 b = pop(); a = pop(); 1815 push( _gvn.transform( new OrINode(a,b) ) ); 1816 break; 1817 case Bytecodes::_ixor: 1818 b = pop(); a = pop(); 1819 push( _gvn.transform( new XorINode(a,b) ) ); 1820 break; 1821 case Bytecodes::_ishl: 1822 b = pop(); a = pop(); 1823 push( _gvn.transform( new LShiftINode(a,b) ) ); 1824 break; 1825 case Bytecodes::_ishr: 1826 b = pop(); a = pop(); 1827 push( _gvn.transform( new RShiftINode(a,b) ) ); 1828 break; 1829 case Bytecodes::_iushr: 1830 b = pop(); a = pop(); 1831 push( _gvn.transform( new URShiftINode(a,b) ) ); 1832 break; 1833 1834 case Bytecodes::_fneg: 1835 a = pop(); 1836 b = _gvn.transform(new NegFNode (a)); 1837 push(b); 1838 break; 1839 1840 case Bytecodes::_fsub: 1841 b = pop(); 1842 a = pop(); 1843 c = _gvn.transform( new SubFNode(a,b) ); 1844 d = precision_rounding(c); 1845 push( d ); 1846 break; 1847 1848 case Bytecodes::_fadd: 1849 b = pop(); 1850 a = pop(); 1851 c = _gvn.transform( new AddFNode(a,b) ); 1852 d = precision_rounding(c); 1853 push( d ); 1854 break; 1855 1856 case Bytecodes::_fmul: 1857 b = pop(); 1858 a = pop(); 1859 c = _gvn.transform( new MulFNode(a,b) ); 1860 d = precision_rounding(c); 1861 push( d ); 1862 break; 1863 1864 case Bytecodes::_fdiv: 1865 b = pop(); 1866 a = pop(); 1867 c = _gvn.transform( new DivFNode(0,a,b) ); 1868 d = precision_rounding(c); 1869 push( d ); 1870 break; 1871 1872 case Bytecodes::_frem: 1873 if (Matcher::has_match_rule(Op_ModF)) { 1874 // Generate a ModF node. 1875 b = pop(); 1876 a = pop(); 1877 c = _gvn.transform( new ModFNode(0,a,b) ); 1878 d = precision_rounding(c); 1879 push( d ); 1880 } 1881 else { 1882 // Generate a call. 1883 modf(); 1884 } 1885 break; 1886 1887 case Bytecodes::_fcmpl: 1888 b = pop(); 1889 a = pop(); 1890 c = _gvn.transform( new CmpF3Node( a, b)); 1891 push(c); 1892 break; 1893 case Bytecodes::_fcmpg: 1894 b = pop(); 1895 a = pop(); 1896 1897 // Same as fcmpl but need to flip the unordered case. Swap the inputs, 1898 // which negates the result sign except for unordered. Flip the unordered 1899 // as well by using CmpF3 which implements unordered-lesser instead of 1900 // unordered-greater semantics. Finally, commute the result bits. Result 1901 // is same as using a CmpF3Greater except we did it with CmpF3 alone. 1902 c = _gvn.transform( new CmpF3Node( b, a)); 1903 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) ); 1904 push(c); 1905 break; 1906 1907 case Bytecodes::_f2i: 1908 a = pop(); 1909 push(_gvn.transform(new ConvF2INode(a))); 1910 break; 1911 1912 case Bytecodes::_d2i: 1913 a = pop_pair(); 1914 b = _gvn.transform(new ConvD2INode(a)); 1915 push( b ); 1916 break; 1917 1918 case Bytecodes::_f2d: 1919 a = pop(); 1920 b = _gvn.transform( new ConvF2DNode(a)); 1921 push_pair( b ); 1922 break; 1923 1924 case Bytecodes::_d2f: 1925 a = pop_pair(); 1926 b = _gvn.transform( new ConvD2FNode(a)); 1927 // This breaks _227_mtrt (speed & correctness) and _222_mpegaudio (speed) 1928 //b = _gvn.transform(new RoundFloatNode(0, b) ); 1929 push( b ); 1930 break; 1931 1932 case Bytecodes::_l2f: 1933 if (Matcher::convL2FSupported()) { 1934 a = pop_pair(); 1935 b = _gvn.transform( new ConvL2FNode(a)); 1936 // For i486.ad, FILD doesn't restrict precision to 24 or 53 bits. 1937 // Rather than storing the result into an FP register then pushing 1938 // out to memory to round, the machine instruction that implements 1939 // ConvL2D is responsible for rounding. 1940 // c = precision_rounding(b); 1941 c = _gvn.transform(b); 1942 push(c); 1943 } else { 1944 l2f(); 1945 } 1946 break; 1947 1948 case Bytecodes::_l2d: 1949 a = pop_pair(); 1950 b = _gvn.transform( new ConvL2DNode(a)); 1951 // For i486.ad, rounding is always necessary (see _l2f above). 1952 // c = dprecision_rounding(b); 1953 c = _gvn.transform(b); 1954 push_pair(c); 1955 break; 1956 1957 case Bytecodes::_f2l: 1958 a = pop(); 1959 b = _gvn.transform( new ConvF2LNode(a)); 1960 push_pair(b); 1961 break; 1962 1963 case Bytecodes::_d2l: 1964 a = pop_pair(); 1965 b = _gvn.transform( new ConvD2LNode(a)); 1966 push_pair(b); 1967 break; 1968 1969 case Bytecodes::_dsub: 1970 b = pop_pair(); 1971 a = pop_pair(); 1972 c = _gvn.transform( new SubDNode(a,b) ); 1973 d = dprecision_rounding(c); 1974 push_pair( d ); 1975 break; 1976 1977 case Bytecodes::_dadd: 1978 b = pop_pair(); 1979 a = pop_pair(); 1980 c = _gvn.transform( new AddDNode(a,b) ); 1981 d = dprecision_rounding(c); 1982 push_pair( d ); 1983 break; 1984 1985 case Bytecodes::_dmul: 1986 b = pop_pair(); 1987 a = pop_pair(); 1988 c = _gvn.transform( new MulDNode(a,b) ); 1989 d = dprecision_rounding(c); 1990 push_pair( d ); 1991 break; 1992 1993 case Bytecodes::_ddiv: 1994 b = pop_pair(); 1995 a = pop_pair(); 1996 c = _gvn.transform( new DivDNode(0,a,b) ); 1997 d = dprecision_rounding(c); 1998 push_pair( d ); 1999 break; 2000 2001 case Bytecodes::_dneg: 2002 a = pop_pair(); 2003 b = _gvn.transform(new NegDNode (a)); 2004 push_pair(b); 2005 break; 2006 2007 case Bytecodes::_drem: 2008 if (Matcher::has_match_rule(Op_ModD)) { 2009 // Generate a ModD node. 2010 b = pop_pair(); 2011 a = pop_pair(); 2012 // a % b 2013 2014 c = _gvn.transform( new ModDNode(0,a,b) ); 2015 d = dprecision_rounding(c); 2016 push_pair( d ); 2017 } 2018 else { 2019 // Generate a call. 2020 modd(); 2021 } 2022 break; 2023 2024 case Bytecodes::_dcmpl: 2025 b = pop_pair(); 2026 a = pop_pair(); 2027 c = _gvn.transform( new CmpD3Node( a, b)); 2028 push(c); 2029 break; 2030 2031 case Bytecodes::_dcmpg: 2032 b = pop_pair(); 2033 a = pop_pair(); 2034 // Same as dcmpl but need to flip the unordered case. 2035 // Commute the inputs, which negates the result sign except for unordered. 2036 // Flip the unordered as well by using CmpD3 which implements 2037 // unordered-lesser instead of unordered-greater semantics. 2038 // Finally, negate the result bits. Result is same as using a 2039 // CmpD3Greater except we did it with CmpD3 alone. 2040 c = _gvn.transform( new CmpD3Node( b, a)); 2041 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) ); 2042 push(c); 2043 break; 2044 2045 2046 // Note for longs -> lo word is on TOS, hi word is on TOS - 1 2047 case Bytecodes::_land: 2048 b = pop_pair(); 2049 a = pop_pair(); 2050 c = _gvn.transform( new AndLNode(a,b) ); 2051 push_pair(c); 2052 break; 2053 case Bytecodes::_lor: 2054 b = pop_pair(); 2055 a = pop_pair(); 2056 c = _gvn.transform( new OrLNode(a,b) ); 2057 push_pair(c); 2058 break; 2059 case Bytecodes::_lxor: 2060 b = pop_pair(); 2061 a = pop_pair(); 2062 c = _gvn.transform( new XorLNode(a,b) ); 2063 push_pair(c); 2064 break; 2065 2066 case Bytecodes::_lshl: 2067 b = pop(); // the shift count 2068 a = pop_pair(); // value to be shifted 2069 c = _gvn.transform( new LShiftLNode(a,b) ); 2070 push_pair(c); 2071 break; 2072 case Bytecodes::_lshr: 2073 b = pop(); // the shift count 2074 a = pop_pair(); // value to be shifted 2075 c = _gvn.transform( new RShiftLNode(a,b) ); 2076 push_pair(c); 2077 break; 2078 case Bytecodes::_lushr: 2079 b = pop(); // the shift count 2080 a = pop_pair(); // value to be shifted 2081 c = _gvn.transform( new URShiftLNode(a,b) ); 2082 push_pair(c); 2083 break; 2084 case Bytecodes::_lmul: 2085 b = pop_pair(); 2086 a = pop_pair(); 2087 c = _gvn.transform( new MulLNode(a,b) ); 2088 push_pair(c); 2089 break; 2090 2091 case Bytecodes::_lrem: 2092 // Must keep both values on the expression-stack during null-check 2093 assert(peek(0) == top(), "long word order"); 2094 zero_check_long(peek(1)); 2095 // Compile-time detect of null-exception? 2096 if (stopped()) return; 2097 b = pop_pair(); 2098 a = pop_pair(); 2099 c = _gvn.transform( new ModLNode(control(),a,b) ); 2100 push_pair(c); 2101 break; 2102 2103 case Bytecodes::_ldiv: 2104 // Must keep both values on the expression-stack during null-check 2105 assert(peek(0) == top(), "long word order"); 2106 zero_check_long(peek(1)); 2107 // Compile-time detect of null-exception? 2108 if (stopped()) return; 2109 b = pop_pair(); 2110 a = pop_pair(); 2111 c = _gvn.transform( new DivLNode(control(),a,b) ); 2112 push_pair(c); 2113 break; 2114 2115 case Bytecodes::_ladd: 2116 b = pop_pair(); 2117 a = pop_pair(); 2118 c = _gvn.transform( new AddLNode(a,b) ); 2119 push_pair(c); 2120 break; 2121 case Bytecodes::_lsub: 2122 b = pop_pair(); 2123 a = pop_pair(); 2124 c = _gvn.transform( new SubLNode(a,b) ); 2125 push_pair(c); 2126 break; 2127 case Bytecodes::_lcmp: 2128 // Safepoints are now inserted _before_ branches. The long-compare 2129 // bytecode painfully produces a 3-way value (-1,0,+1) which requires a 2130 // slew of control flow. These are usually followed by a CmpI vs zero and 2131 // a branch; this pattern then optimizes to the obvious long-compare and 2132 // branch. However, if the branch is backwards there's a Safepoint 2133 // inserted. The inserted Safepoint captures the JVM state at the 2134 // pre-branch point, i.e. it captures the 3-way value. Thus if a 2135 // long-compare is used to control a loop the debug info will force 2136 // computation of the 3-way value, even though the generated code uses a 2137 // long-compare and branch. We try to rectify the situation by inserting 2138 // a SafePoint here and have it dominate and kill the safepoint added at a 2139 // following backwards branch. At this point the JVM state merely holds 2 2140 // longs but not the 3-way value. 2141 if( UseLoopSafepoints ) { 2142 switch( iter().next_bc() ) { 2143 case Bytecodes::_ifgt: 2144 case Bytecodes::_iflt: 2145 case Bytecodes::_ifge: 2146 case Bytecodes::_ifle: 2147 case Bytecodes::_ifne: 2148 case Bytecodes::_ifeq: 2149 // If this is a backwards branch in the bytecodes, add Safepoint 2150 maybe_add_safepoint(iter().next_get_dest()); 2151 } 2152 } 2153 b = pop_pair(); 2154 a = pop_pair(); 2155 c = _gvn.transform( new CmpL3Node( a, b )); 2156 push(c); 2157 break; 2158 2159 case Bytecodes::_lneg: 2160 a = pop_pair(); 2161 b = _gvn.transform( new SubLNode(longcon(0),a)); 2162 push_pair(b); 2163 break; 2164 case Bytecodes::_l2i: 2165 a = pop_pair(); 2166 push( _gvn.transform( new ConvL2INode(a))); 2167 break; 2168 case Bytecodes::_i2l: 2169 a = pop(); 2170 b = _gvn.transform( new ConvI2LNode(a)); 2171 push_pair(b); 2172 break; 2173 case Bytecodes::_i2b: 2174 // Sign extend 2175 a = pop(); 2176 a = _gvn.transform( new LShiftINode(a,_gvn.intcon(24)) ); 2177 a = _gvn.transform( new RShiftINode(a,_gvn.intcon(24)) ); 2178 push( a ); 2179 break; 2180 case Bytecodes::_i2s: 2181 a = pop(); 2182 a = _gvn.transform( new LShiftINode(a,_gvn.intcon(16)) ); 2183 a = _gvn.transform( new RShiftINode(a,_gvn.intcon(16)) ); 2184 push( a ); 2185 break; 2186 case Bytecodes::_i2c: 2187 a = pop(); 2188 push( _gvn.transform( new AndINode(a,_gvn.intcon(0xFFFF)) ) ); 2189 break; 2190 2191 case Bytecodes::_i2f: 2192 a = pop(); 2193 b = _gvn.transform( new ConvI2FNode(a) ) ; 2194 c = precision_rounding(b); 2195 push (b); 2196 break; 2197 2198 case Bytecodes::_i2d: 2199 a = pop(); 2200 b = _gvn.transform( new ConvI2DNode(a)); 2201 push_pair(b); 2202 break; 2203 2204 case Bytecodes::_iinc: // Increment local 2205 i = iter().get_index(); // Get local index 2206 set_local( i, _gvn.transform( new AddINode( _gvn.intcon(iter().get_iinc_con()), local(i) ) ) ); 2207 break; 2208 2209 // Exit points of synchronized methods must have an unlock node 2210 case Bytecodes::_return: 2211 return_current(NULL); 2212 break; 2213 2214 case Bytecodes::_ireturn: 2215 case Bytecodes::_areturn: 2216 case Bytecodes::_vreturn: 2217 case Bytecodes::_freturn: 2218 return_current(pop()); 2219 break; 2220 case Bytecodes::_lreturn: 2221 return_current(pop_pair()); 2222 break; 2223 case Bytecodes::_dreturn: 2224 return_current(pop_pair()); 2225 break; 2226 2227 case Bytecodes::_athrow: 2228 // null exception oop throws NULL pointer exception 2229 null_check(peek()); 2230 if (stopped()) return; 2231 // Hook the thrown exception directly to subsequent handlers. 2232 if (BailoutToInterpreterForThrows) { 2233 // Keep method interpreted from now on. 2234 uncommon_trap(Deoptimization::Reason_unhandled, 2235 Deoptimization::Action_make_not_compilable); 2236 return; 2237 } 2238 if (env()->jvmti_can_post_on_exceptions()) { 2239 // check if we must post exception events, take uncommon trap if so (with must_throw = false) 2240 uncommon_trap_if_should_post_on_exceptions(Deoptimization::Reason_unhandled, false); 2241 } 2242 // Here if either can_post_on_exceptions or should_post_on_exceptions is false 2243 add_exception_state(make_exception_state(peek())); 2244 break; 2245 2246 case Bytecodes::_goto: // fall through 2247 case Bytecodes::_goto_w: { 2248 int target_bci = (bc() == Bytecodes::_goto) ? iter().get_dest() : iter().get_far_dest(); 2249 2250 // If this is a backwards branch in the bytecodes, add Safepoint 2251 maybe_add_safepoint(target_bci); 2252 2253 // Update method data 2254 profile_taken_branch(target_bci); 2255 2256 // Merge the current control into the target basic block 2257 merge(target_bci); 2258 2259 // See if we can get some profile data and hand it off to the next block 2260 Block *target_block = block()->successor_for_bci(target_bci); 2261 if (target_block->pred_count() != 1) break; 2262 ciMethodData* methodData = method()->method_data(); 2263 if (!methodData->is_mature()) break; 2264 ciProfileData* data = methodData->bci_to_data(bci()); 2265 assert( data->is_JumpData(), "" ); 2266 int taken = ((ciJumpData*)data)->taken(); 2267 taken = method()->scale_count(taken); 2268 target_block->set_count(taken); 2269 break; 2270 } 2271 2272 case Bytecodes::_ifnull: btest = BoolTest::eq; goto handle_if_null; 2273 case Bytecodes::_ifnonnull: btest = BoolTest::ne; goto handle_if_null; 2274 handle_if_null: 2275 // If this is a backwards branch in the bytecodes, add Safepoint 2276 maybe_add_safepoint(iter().get_dest()); 2277 a = null(); 2278 b = pop(); 2279 if (!_gvn.type(b)->speculative_maybe_null() && 2280 !too_many_traps(Deoptimization::Reason_speculate_null_check)) { 2281 inc_sp(1); 2282 Node* null_ctl = top(); 2283 b = null_check_oop(b, &null_ctl, true, true, true); 2284 assert(null_ctl->is_top(), "no null control here"); 2285 dec_sp(1); 2286 } 2287 c = _gvn.transform( new CmpPNode(b, a) ); 2288 do_ifnull(btest, c); 2289 break; 2290 2291 case Bytecodes::_if_acmpeq: btest = BoolTest::eq; goto handle_if_acmp; 2292 case Bytecodes::_if_acmpne: btest = BoolTest::ne; goto handle_if_acmp; 2293 handle_if_acmp: 2294 // If this is a backwards branch in the bytecodes, add Safepoint 2295 maybe_add_safepoint(iter().get_dest()); 2296 a = pop(); 2297 b = pop(); 2298 c = _gvn.transform( new CmpPNode(b, a) ); 2299 c = optimize_cmp_with_klass(c); 2300 do_if(btest, c); 2301 break; 2302 2303 case Bytecodes::_ifeq: btest = BoolTest::eq; goto handle_ifxx; 2304 case Bytecodes::_ifne: btest = BoolTest::ne; goto handle_ifxx; 2305 case Bytecodes::_iflt: btest = BoolTest::lt; goto handle_ifxx; 2306 case Bytecodes::_ifle: btest = BoolTest::le; goto handle_ifxx; 2307 case Bytecodes::_ifgt: btest = BoolTest::gt; goto handle_ifxx; 2308 case Bytecodes::_ifge: btest = BoolTest::ge; goto handle_ifxx; 2309 handle_ifxx: 2310 // If this is a backwards branch in the bytecodes, add Safepoint 2311 maybe_add_safepoint(iter().get_dest()); 2312 a = _gvn.intcon(0); 2313 b = pop(); 2314 c = _gvn.transform( new CmpINode(b, a) ); 2315 do_if(btest, c); 2316 break; 2317 2318 case Bytecodes::_if_icmpeq: btest = BoolTest::eq; goto handle_if_icmp; 2319 case Bytecodes::_if_icmpne: btest = BoolTest::ne; goto handle_if_icmp; 2320 case Bytecodes::_if_icmplt: btest = BoolTest::lt; goto handle_if_icmp; 2321 case Bytecodes::_if_icmple: btest = BoolTest::le; goto handle_if_icmp; 2322 case Bytecodes::_if_icmpgt: btest = BoolTest::gt; goto handle_if_icmp; 2323 case Bytecodes::_if_icmpge: btest = BoolTest::ge; goto handle_if_icmp; 2324 handle_if_icmp: 2325 // If this is a backwards branch in the bytecodes, add Safepoint 2326 maybe_add_safepoint(iter().get_dest()); 2327 a = pop(); 2328 b = pop(); 2329 c = _gvn.transform( new CmpINode( b, a ) ); 2330 do_if(btest, c); 2331 break; 2332 2333 case Bytecodes::_tableswitch: 2334 do_tableswitch(); 2335 break; 2336 2337 case Bytecodes::_lookupswitch: 2338 do_lookupswitch(); 2339 break; 2340 2341 case Bytecodes::_invokestatic: 2342 case Bytecodes::_invokedynamic: 2343 case Bytecodes::_invokespecial: 2344 case Bytecodes::_invokevirtual: 2345 case Bytecodes::_invokedirect: 2346 case Bytecodes::_invokeinterface: 2347 do_call(); 2348 break; 2349 case Bytecodes::_checkcast: 2350 do_checkcast(); 2351 break; 2352 case Bytecodes::_instanceof: 2353 do_instanceof(); 2354 break; 2355 case Bytecodes::_anewarray: 2356 do_anewarray(); 2357 break; 2358 case Bytecodes::_newarray: 2359 do_newarray((BasicType)iter().get_index()); 2360 break; 2361 case Bytecodes::_multianewarray: 2362 do_multianewarray(); 2363 break; 2364 case Bytecodes::_new: 2365 do_new(); 2366 break; 2367 case Bytecodes::_vnew: 2368 do_vnew(); 2369 break; 2370 2371 case Bytecodes::_jsr: 2372 case Bytecodes::_jsr_w: 2373 do_jsr(); 2374 break; 2375 2376 case Bytecodes::_ret: 2377 do_ret(); 2378 break; 2379 2380 2381 case Bytecodes::_monitorenter: 2382 do_monitor_enter(); 2383 break; 2384 2385 case Bytecodes::_monitorexit: 2386 do_monitor_exit(); 2387 break; 2388 2389 case Bytecodes::_breakpoint: 2390 // Breakpoint set concurrently to compile 2391 // %%% use an uncommon trap? 2392 C->record_failure("breakpoint in method"); 2393 return; 2394 2395 default: 2396 #ifndef PRODUCT 2397 map()->dump(99); 2398 #endif 2399 tty->print("\nUnhandled bytecode %s\n", Bytecodes::name(bc()) ); 2400 ShouldNotReachHere(); 2401 } 2402 2403 #ifndef PRODUCT 2404 IdealGraphPrinter *printer = C->printer(); 2405 if (printer && printer->should_print(1)) { 2406 char buffer[256]; 2407 sprintf(buffer, "Bytecode %d: %s", bci(), Bytecodes::name(bc())); 2408 bool old = printer->traverse_outs(); 2409 printer->set_traverse_outs(true); 2410 printer->print_method(buffer, 4); 2411 printer->set_traverse_outs(old); 2412 } 2413 #endif 2414 }