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