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 Node* idx = pop(); // Get from stack without popping 59 Node* ary = pop(); // in case of exception 60 //dec_sp(2); // Pop array and index 61 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr(); 62 if (arytype->klass()->is_value_array_klass()) { 63 ciValueArrayKlass* vak = arytype->klass()->as_value_array_klass(); 64 Node* vt = ValueTypeNode::make(gvn(), vak->element_klass()->as_value_klass(), map()->memory(), ary, adr); 65 push(vt); 66 return; 67 } 68 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(elem_type); 69 Node* ld = make_load(control(), adr, elem, elem_type, adr_type, MemNode::unordered); 70 push(ld); 71 } 72 73 74 //--------------------------------array_store---------------------------------- 75 void Parse::array_store(BasicType elem_type) { 76 Node* adr = array_addressing(elem_type, 1); 77 if (stopped()) return; // guaranteed null or range check 78 Node* val = pop(); 79 dec_sp(2); // Pop array and index 80 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(elem_type); 81 store_to_memory(control(), adr, val, elem_type, adr_type, StoreNode::release_if_reference(elem_type)); 82 } 83 84 85 //------------------------------array_addressing------------------------------- 86 // Pull array and index from the stack. Compute pointer-to-element. 87 Node* Parse::array_addressing(BasicType type, int vals, const Type* *result2) { 88 Node *idx = peek(0+vals); // Get from stack without popping 89 Node *ary = peek(1+vals); // in case of exception 90 91 // Null check the array base, with correct stack contents 92 ary = null_check(ary, T_ARRAY); 93 // Compile-time detect of null-exception? 94 if (stopped()) return top(); 95 96 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr(); 97 const TypeInt* sizetype = arytype->size(); 98 const Type* elemtype = arytype->elem(); 99 100 if (UseUniqueSubclasses && result2 != NULL) { 101 const Type* el = elemtype->make_ptr(); 102 if (el && el->isa_instptr()) { 103 const TypeInstPtr* toop = el->is_instptr(); 104 if (toop->klass()->as_instance_klass()->unique_concrete_subklass()) { 105 // If we load from "AbstractClass[]" we must see "ConcreteSubClass". 106 const Type* subklass = Type::get_const_type(toop->klass()); 107 elemtype = subklass->join_speculative(el); 108 } 109 } 110 } 111 112 // Check for big class initializers with all constant offsets 113 // feeding into a known-size array. 114 const TypeInt* idxtype = _gvn.type(idx)->is_int(); 115 // See if the highest idx value is less than the lowest array bound, 116 // and if the idx value cannot be negative: 117 bool need_range_check = true; 118 if (idxtype->_hi < sizetype->_lo && idxtype->_lo >= 0) { 119 need_range_check = false; 120 if (C->log() != NULL) C->log()->elem("observe that='!need_range_check'"); 121 } 122 123 ciKlass * arytype_klass = arytype->klass(); 124 if ((arytype_klass != NULL) && (!arytype_klass->is_loaded())) { 125 // Only fails for some -Xcomp runs 126 // The class is unloaded. We have to run this bytecode in the interpreter. 127 uncommon_trap(Deoptimization::Reason_unloaded, 128 Deoptimization::Action_reinterpret, 129 arytype->klass(), "!loaded array"); 130 return top(); 131 } 132 133 // Do the range check 134 if (GenerateRangeChecks && need_range_check) { 135 Node* tst; 136 if (sizetype->_hi <= 0) { 137 // The greatest array bound is negative, so we can conclude that we're 138 // compiling unreachable code, but the unsigned compare trick used below 139 // only works with non-negative lengths. Instead, hack "tst" to be zero so 140 // the uncommon_trap path will always be taken. 141 tst = _gvn.intcon(0); 142 } else { 143 // Range is constant in array-oop, so we can use the original state of mem 144 Node* len = load_array_length(ary); 145 146 // Test length vs index (standard trick using unsigned compare) 147 Node* chk = _gvn.transform( new CmpUNode(idx, len) ); 148 BoolTest::mask btest = BoolTest::lt; 149 tst = _gvn.transform( new BoolNode(chk, btest) ); 150 } 151 RangeCheckNode* rc = new RangeCheckNode(control(), tst, PROB_MAX, COUNT_UNKNOWN); 152 _gvn.set_type(rc, rc->Value(&_gvn)); 153 if (!tst->is_Con()) { 154 record_for_igvn(rc); 155 } 156 set_control(_gvn.transform(new IfTrueNode(rc))); 157 // Branch to failure if out of bounds 158 { 159 PreserveJVMState pjvms(this); 160 set_control(_gvn.transform(new IfFalseNode(rc))); 161 if (C->allow_range_check_smearing()) { 162 // Do not use builtin_throw, since range checks are sometimes 163 // made more stringent by an optimistic transformation. 164 // This creates "tentative" range checks at this point, 165 // which are not guaranteed to throw exceptions. 166 // See IfNode::Ideal, is_range_check, adjust_check. 167 uncommon_trap(Deoptimization::Reason_range_check, 168 Deoptimization::Action_make_not_entrant, 169 NULL, "range_check"); 170 } else { 171 // If we have already recompiled with the range-check-widening 172 // heroic optimization turned off, then we must really be throwing 173 // range check exceptions. 174 builtin_throw(Deoptimization::Reason_range_check, idx); 175 } 176 } 177 } 178 // Check for always knowing you are throwing a range-check exception 179 if (stopped()) return top(); 180 181 // Make array address computation control dependent to prevent it 182 // from floating above the range check during loop optimizations. 183 Node* ptr = array_element_address(ary, idx, type, sizetype, control()); 184 185 if (result2 != NULL) *result2 = elemtype; 186 187 assert(ptr != top(), "top should go hand-in-hand with stopped"); 188 189 return ptr; 190 } 191 192 193 // returns IfNode 194 IfNode* Parse::jump_if_fork_int(Node* a, Node* b, BoolTest::mask mask) { 195 Node *cmp = _gvn.transform( new CmpINode( a, b)); // two cases: shiftcount > 32 and shiftcount <= 32 196 Node *tst = _gvn.transform( new BoolNode( cmp, mask)); 197 IfNode *iff = create_and_map_if( control(), tst, ((mask == BoolTest::eq) ? PROB_STATIC_INFREQUENT : PROB_FAIR), COUNT_UNKNOWN ); 198 return iff; 199 } 200 201 // return Region node 202 Node* Parse::jump_if_join(Node* iffalse, Node* iftrue) { 203 Node *region = new RegionNode(3); // 2 results 204 record_for_igvn(region); 205 region->init_req(1, iffalse); 206 region->init_req(2, iftrue ); 207 _gvn.set_type(region, Type::CONTROL); 208 region = _gvn.transform(region); 209 set_control (region); 210 return region; 211 } 212 213 214 //------------------------------helper for tableswitch------------------------- 215 void Parse::jump_if_true_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index) { 216 // True branch, use existing map info 217 { PreserveJVMState pjvms(this); 218 Node *iftrue = _gvn.transform( new IfTrueNode (iff) ); 219 set_control( iftrue ); 220 profile_switch_case(prof_table_index); 221 merge_new_path(dest_bci_if_true); 222 } 223 224 // False branch 225 Node *iffalse = _gvn.transform( new IfFalseNode(iff) ); 226 set_control( iffalse ); 227 } 228 229 void Parse::jump_if_false_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index) { 230 // True branch, use existing map info 231 { PreserveJVMState pjvms(this); 232 Node *iffalse = _gvn.transform( new IfFalseNode (iff) ); 233 set_control( iffalse ); 234 profile_switch_case(prof_table_index); 235 merge_new_path(dest_bci_if_true); 236 } 237 238 // False branch 239 Node *iftrue = _gvn.transform( new IfTrueNode(iff) ); 240 set_control( iftrue ); 241 } 242 243 void Parse::jump_if_always_fork(int dest_bci, int prof_table_index) { 244 // False branch, use existing map and control() 245 profile_switch_case(prof_table_index); 246 merge_new_path(dest_bci); 247 } 248 249 250 extern "C" { 251 static int jint_cmp(const void *i, const void *j) { 252 int a = *(jint *)i; 253 int b = *(jint *)j; 254 return a > b ? 1 : a < b ? -1 : 0; 255 } 256 } 257 258 259 // Default value for methodData switch indexing. Must be a negative value to avoid 260 // conflict with any legal switch index. 261 #define NullTableIndex -1 262 263 class SwitchRange : public StackObj { 264 // a range of integers coupled with a bci destination 265 jint _lo; // inclusive lower limit 266 jint _hi; // inclusive upper limit 267 int _dest; 268 int _table_index; // index into method data table 269 270 public: 271 jint lo() const { return _lo; } 272 jint hi() const { return _hi; } 273 int dest() const { return _dest; } 274 int table_index() const { return _table_index; } 275 bool is_singleton() const { return _lo == _hi; } 276 277 void setRange(jint lo, jint hi, int dest, int table_index) { 278 assert(lo <= hi, "must be a non-empty range"); 279 _lo = lo, _hi = hi; _dest = dest; _table_index = table_index; 280 } 281 bool adjoinRange(jint lo, jint hi, int dest, int table_index) { 282 assert(lo <= hi, "must be a non-empty range"); 283 if (lo == _hi+1 && dest == _dest && table_index == _table_index) { 284 _hi = hi; 285 return true; 286 } 287 return false; 288 } 289 290 void set (jint value, int dest, int table_index) { 291 setRange(value, value, dest, table_index); 292 } 293 bool adjoin(jint value, int dest, int table_index) { 294 return adjoinRange(value, value, dest, table_index); 295 } 296 297 void print() { 298 if (is_singleton()) 299 tty->print(" {%d}=>%d", lo(), dest()); 300 else if (lo() == min_jint) 301 tty->print(" {..%d}=>%d", hi(), dest()); 302 else if (hi() == max_jint) 303 tty->print(" {%d..}=>%d", lo(), dest()); 304 else 305 tty->print(" {%d..%d}=>%d", lo(), hi(), dest()); 306 } 307 }; 308 309 310 //-------------------------------do_tableswitch-------------------------------- 311 void Parse::do_tableswitch() { 312 Node* lookup = pop(); 313 314 // Get information about tableswitch 315 int default_dest = iter().get_dest_table(0); 316 int lo_index = iter().get_int_table(1); 317 int hi_index = iter().get_int_table(2); 318 int len = hi_index - lo_index + 1; 319 320 if (len < 1) { 321 // If this is a backward branch, add safepoint 322 maybe_add_safepoint(default_dest); 323 merge(default_dest); 324 return; 325 } 326 327 // generate decision tree, using trichotomy when possible 328 int rnum = len+2; 329 bool makes_backward_branch = false; 330 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum); 331 int rp = -1; 332 if (lo_index != min_jint) { 333 ranges[++rp].setRange(min_jint, lo_index-1, default_dest, NullTableIndex); 334 } 335 for (int j = 0; j < len; j++) { 336 jint match_int = lo_index+j; 337 int dest = iter().get_dest_table(j+3); 338 makes_backward_branch |= (dest <= bci()); 339 int table_index = method_data_update() ? j : NullTableIndex; 340 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index)) { 341 ranges[++rp].set(match_int, dest, table_index); 342 } 343 } 344 jint highest = lo_index+(len-1); 345 assert(ranges[rp].hi() == highest, ""); 346 if (highest != max_jint 347 && !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex)) { 348 ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex); 349 } 350 assert(rp < len+2, "not too many ranges"); 351 352 // Safepoint in case if backward branch observed 353 if( makes_backward_branch && UseLoopSafepoints ) 354 add_safepoint(); 355 356 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]); 357 } 358 359 360 //------------------------------do_lookupswitch-------------------------------- 361 void Parse::do_lookupswitch() { 362 Node *lookup = pop(); // lookup value 363 // Get information about lookupswitch 364 int default_dest = iter().get_dest_table(0); 365 int len = iter().get_int_table(1); 366 367 if (len < 1) { // If this is a backward branch, add safepoint 368 maybe_add_safepoint(default_dest); 369 merge(default_dest); 370 return; 371 } 372 373 // generate decision tree, using trichotomy when possible 374 jint* table = NEW_RESOURCE_ARRAY(jint, len*2); 375 { 376 for( int j = 0; j < len; j++ ) { 377 table[j+j+0] = iter().get_int_table(2+j+j); 378 table[j+j+1] = iter().get_dest_table(2+j+j+1); 379 } 380 qsort( table, len, 2*sizeof(table[0]), jint_cmp ); 381 } 382 383 int rnum = len*2+1; 384 bool makes_backward_branch = false; 385 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum); 386 int rp = -1; 387 for( int j = 0; j < len; j++ ) { 388 jint match_int = table[j+j+0]; 389 int dest = table[j+j+1]; 390 int next_lo = rp < 0 ? min_jint : ranges[rp].hi()+1; 391 int table_index = method_data_update() ? j : NullTableIndex; 392 makes_backward_branch |= (dest <= bci()); 393 if( match_int != next_lo ) { 394 ranges[++rp].setRange(next_lo, match_int-1, default_dest, NullTableIndex); 395 } 396 if( rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index) ) { 397 ranges[++rp].set(match_int, dest, table_index); 398 } 399 } 400 jint highest = table[2*(len-1)]; 401 assert(ranges[rp].hi() == highest, ""); 402 if( highest != max_jint 403 && !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex) ) { 404 ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex); 405 } 406 assert(rp < rnum, "not too many ranges"); 407 408 // Safepoint in case backward branch observed 409 if( makes_backward_branch && UseLoopSafepoints ) 410 add_safepoint(); 411 412 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]); 413 } 414 415 //----------------------------create_jump_tables------------------------------- 416 bool Parse::create_jump_tables(Node* key_val, SwitchRange* lo, SwitchRange* hi) { 417 // Are jumptables enabled 418 if (!UseJumpTables) return false; 419 420 // Are jumptables supported 421 if (!Matcher::has_match_rule(Op_Jump)) return false; 422 423 // Don't make jump table if profiling 424 if (method_data_update()) return false; 425 426 // Decide if a guard is needed to lop off big ranges at either (or 427 // both) end(s) of the input set. We'll call this the default target 428 // even though we can't be sure that it is the true "default". 429 430 bool needs_guard = false; 431 int default_dest; 432 int64_t total_outlier_size = 0; 433 int64_t hi_size = ((int64_t)hi->hi()) - ((int64_t)hi->lo()) + 1; 434 int64_t lo_size = ((int64_t)lo->hi()) - ((int64_t)lo->lo()) + 1; 435 436 if (lo->dest() == hi->dest()) { 437 total_outlier_size = hi_size + lo_size; 438 default_dest = lo->dest(); 439 } else if (lo_size > hi_size) { 440 total_outlier_size = lo_size; 441 default_dest = lo->dest(); 442 } else { 443 total_outlier_size = hi_size; 444 default_dest = hi->dest(); 445 } 446 447 // If a guard test will eliminate very sparse end ranges, then 448 // it is worth the cost of an extra jump. 449 if (total_outlier_size > (MaxJumpTableSparseness * 4)) { 450 needs_guard = true; 451 if (default_dest == lo->dest()) lo++; 452 if (default_dest == hi->dest()) hi--; 453 } 454 455 // Find the total number of cases and ranges 456 int64_t num_cases = ((int64_t)hi->hi()) - ((int64_t)lo->lo()) + 1; 457 int num_range = hi - lo + 1; 458 459 // Don't create table if: too large, too small, or too sparse. 460 if (num_cases < MinJumpTableSize || num_cases > MaxJumpTableSize) 461 return false; 462 if (num_cases > (MaxJumpTableSparseness * num_range)) 463 return false; 464 465 // Normalize table lookups to zero 466 int lowval = lo->lo(); 467 key_val = _gvn.transform( new SubINode(key_val, _gvn.intcon(lowval)) ); 468 469 // Generate a guard to protect against input keyvals that aren't 470 // in the switch domain. 471 if (needs_guard) { 472 Node* size = _gvn.intcon(num_cases); 473 Node* cmp = _gvn.transform( new CmpUNode(key_val, size) ); 474 Node* tst = _gvn.transform( new BoolNode(cmp, BoolTest::ge) ); 475 IfNode* iff = create_and_map_if( control(), tst, PROB_FAIR, COUNT_UNKNOWN); 476 jump_if_true_fork(iff, default_dest, NullTableIndex); 477 } 478 479 // Create an ideal node JumpTable that has projections 480 // of all possible ranges for a switch statement 481 // The key_val input must be converted to a pointer offset and scaled. 482 // Compare Parse::array_addressing above. 483 484 // Clean the 32-bit int into a real 64-bit offset. 485 // Otherwise, the jint value 0 might turn into an offset of 0x0800000000. 486 const TypeInt* ikeytype = TypeInt::make(0, num_cases, Type::WidenMin); 487 // Make I2L conversion control dependent to prevent it from 488 // floating above the range check during loop optimizations. 489 key_val = C->conv_I2X_index(&_gvn, key_val, ikeytype, control()); 490 491 // Shift the value by wordsize so we have an index into the table, rather 492 // than a switch value 493 Node *shiftWord = _gvn.MakeConX(wordSize); 494 key_val = _gvn.transform( new MulXNode( key_val, shiftWord)); 495 496 // Create the JumpNode 497 Node* jtn = _gvn.transform( new JumpNode(control(), key_val, num_cases) ); 498 499 // These are the switch destinations hanging off the jumpnode 500 int i = 0; 501 for (SwitchRange* r = lo; r <= hi; r++) { 502 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) { 503 Node* input = _gvn.transform(new JumpProjNode(jtn, i, r->dest(), (int)(j - lowval))); 504 { 505 PreserveJVMState pjvms(this); 506 set_control(input); 507 jump_if_always_fork(r->dest(), r->table_index()); 508 } 509 } 510 } 511 assert(i == num_cases, "miscount of cases"); 512 stop_and_kill_map(); // no more uses for this JVMS 513 return true; 514 } 515 516 //----------------------------jump_switch_ranges------------------------------- 517 void Parse::jump_switch_ranges(Node* key_val, SwitchRange *lo, SwitchRange *hi, int switch_depth) { 518 Block* switch_block = block(); 519 520 if (switch_depth == 0) { 521 // Do special processing for the top-level call. 522 assert(lo->lo() == min_jint, "initial range must exhaust Type::INT"); 523 assert(hi->hi() == max_jint, "initial range must exhaust Type::INT"); 524 525 // Decrement pred-numbers for the unique set of nodes. 526 #ifdef ASSERT 527 // Ensure that the block's successors are a (duplicate-free) set. 528 int successors_counted = 0; // block occurrences in [hi..lo] 529 int unique_successors = switch_block->num_successors(); 530 for (int i = 0; i < unique_successors; i++) { 531 Block* target = switch_block->successor_at(i); 532 533 // Check that the set of successors is the same in both places. 534 int successors_found = 0; 535 for (SwitchRange* p = lo; p <= hi; p++) { 536 if (p->dest() == target->start()) successors_found++; 537 } 538 assert(successors_found > 0, "successor must be known"); 539 successors_counted += successors_found; 540 } 541 assert(successors_counted == (hi-lo)+1, "no unexpected successors"); 542 #endif 543 544 // Maybe prune the inputs, based on the type of key_val. 545 jint min_val = min_jint; 546 jint max_val = max_jint; 547 const TypeInt* ti = key_val->bottom_type()->isa_int(); 548 if (ti != NULL) { 549 min_val = ti->_lo; 550 max_val = ti->_hi; 551 assert(min_val <= max_val, "invalid int type"); 552 } 553 while (lo->hi() < min_val) lo++; 554 if (lo->lo() < min_val) lo->setRange(min_val, lo->hi(), lo->dest(), lo->table_index()); 555 while (hi->lo() > max_val) hi--; 556 if (hi->hi() > max_val) hi->setRange(hi->lo(), max_val, hi->dest(), hi->table_index()); 557 } 558 559 #ifndef PRODUCT 560 if (switch_depth == 0) { 561 _max_switch_depth = 0; 562 _est_switch_depth = log2_intptr((hi-lo+1)-1)+1; 563 } 564 #endif 565 566 assert(lo <= hi, "must be a non-empty set of ranges"); 567 if (lo == hi) { 568 jump_if_always_fork(lo->dest(), lo->table_index()); 569 } else { 570 assert(lo->hi() == (lo+1)->lo()-1, "contiguous ranges"); 571 assert(hi->lo() == (hi-1)->hi()+1, "contiguous ranges"); 572 573 if (create_jump_tables(key_val, lo, hi)) return; 574 575 int nr = hi - lo + 1; 576 577 SwitchRange* mid = lo + nr/2; 578 // if there is an easy choice, pivot at a singleton: 579 if (nr > 3 && !mid->is_singleton() && (mid-1)->is_singleton()) mid--; 580 581 assert(lo < mid && mid <= hi, "good pivot choice"); 582 assert(nr != 2 || mid == hi, "should pick higher of 2"); 583 assert(nr != 3 || mid == hi-1, "should pick middle of 3"); 584 585 Node *test_val = _gvn.intcon(mid->lo()); 586 587 if (mid->is_singleton()) { 588 IfNode *iff_ne = jump_if_fork_int(key_val, test_val, BoolTest::ne); 589 jump_if_false_fork(iff_ne, mid->dest(), mid->table_index()); 590 591 // Special Case: If there are exactly three ranges, and the high 592 // and low range each go to the same place, omit the "gt" test, 593 // since it will not discriminate anything. 594 bool eq_test_only = (hi == lo+2 && hi->dest() == lo->dest()); 595 if (eq_test_only) { 596 assert(mid == hi-1, ""); 597 } 598 599 // if there is a higher range, test for it and process it: 600 if (mid < hi && !eq_test_only) { 601 // two comparisons of same values--should enable 1 test for 2 branches 602 // Use BoolTest::le instead of BoolTest::gt 603 IfNode *iff_le = jump_if_fork_int(key_val, test_val, BoolTest::le); 604 Node *iftrue = _gvn.transform( new IfTrueNode(iff_le) ); 605 Node *iffalse = _gvn.transform( new IfFalseNode(iff_le) ); 606 { PreserveJVMState pjvms(this); 607 set_control(iffalse); 608 jump_switch_ranges(key_val, mid+1, hi, switch_depth+1); 609 } 610 set_control(iftrue); 611 } 612 613 } else { 614 // mid is a range, not a singleton, so treat mid..hi as a unit 615 IfNode *iff_ge = jump_if_fork_int(key_val, test_val, BoolTest::ge); 616 617 // if there is a higher range, test for it and process it: 618 if (mid == hi) { 619 jump_if_true_fork(iff_ge, mid->dest(), mid->table_index()); 620 } else { 621 Node *iftrue = _gvn.transform( new IfTrueNode(iff_ge) ); 622 Node *iffalse = _gvn.transform( new IfFalseNode(iff_ge) ); 623 { PreserveJVMState pjvms(this); 624 set_control(iftrue); 625 jump_switch_ranges(key_val, mid, hi, switch_depth+1); 626 } 627 set_control(iffalse); 628 } 629 } 630 631 // in any case, process the lower range 632 jump_switch_ranges(key_val, lo, mid-1, switch_depth+1); 633 } 634 635 // Decrease pred_count for each successor after all is done. 636 if (switch_depth == 0) { 637 int unique_successors = switch_block->num_successors(); 638 for (int i = 0; i < unique_successors; i++) { 639 Block* target = switch_block->successor_at(i); 640 // Throw away the pre-allocated path for each unique successor. 641 target->next_path_num(); 642 } 643 } 644 645 #ifndef PRODUCT 646 _max_switch_depth = MAX2(switch_depth, _max_switch_depth); 647 if (TraceOptoParse && Verbose && WizardMode && switch_depth == 0) { 648 SwitchRange* r; 649 int nsing = 0; 650 for( r = lo; r <= hi; r++ ) { 651 if( r->is_singleton() ) nsing++; 652 } 653 tty->print(">>> "); 654 _method->print_short_name(); 655 tty->print_cr(" switch decision tree"); 656 tty->print_cr(" %d ranges (%d singletons), max_depth=%d, est_depth=%d", 657 (int) (hi-lo+1), nsing, _max_switch_depth, _est_switch_depth); 658 if (_max_switch_depth > _est_switch_depth) { 659 tty->print_cr("******** BAD SWITCH DEPTH ********"); 660 } 661 tty->print(" "); 662 for( r = lo; r <= hi; r++ ) { 663 r->print(); 664 } 665 tty->cr(); 666 } 667 #endif 668 } 669 670 void Parse::modf() { 671 Node *f2 = pop(); 672 Node *f1 = pop(); 673 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::modf_Type(), 674 CAST_FROM_FN_PTR(address, SharedRuntime::frem), 675 "frem", NULL, //no memory effects 676 f1, f2); 677 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0)); 678 679 push(res); 680 } 681 682 void Parse::modd() { 683 Node *d2 = pop_pair(); 684 Node *d1 = pop_pair(); 685 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::Math_DD_D_Type(), 686 CAST_FROM_FN_PTR(address, SharedRuntime::drem), 687 "drem", NULL, //no memory effects 688 d1, top(), d2, top()); 689 Node* res_d = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0)); 690 691 #ifdef ASSERT 692 Node* res_top = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 1)); 693 assert(res_top == top(), "second value must be top"); 694 #endif 695 696 push_pair(res_d); 697 } 698 699 void Parse::l2f() { 700 Node* f2 = pop(); 701 Node* f1 = pop(); 702 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::l2f_Type(), 703 CAST_FROM_FN_PTR(address, SharedRuntime::l2f), 704 "l2f", NULL, //no memory effects 705 f1, f2); 706 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0)); 707 708 push(res); 709 } 710 711 void Parse::do_irem() { 712 // Must keep both values on the expression-stack during null-check 713 zero_check_int(peek()); 714 // Compile-time detect of null-exception? 715 if (stopped()) return; 716 717 Node* b = pop(); 718 Node* a = pop(); 719 720 const Type *t = _gvn.type(b); 721 if (t != Type::TOP) { 722 const TypeInt *ti = t->is_int(); 723 if (ti->is_con()) { 724 int divisor = ti->get_con(); 725 // check for positive power of 2 726 if (divisor > 0 && 727 (divisor & ~(divisor-1)) == divisor) { 728 // yes ! 729 Node *mask = _gvn.intcon((divisor - 1)); 730 // Sigh, must handle negative dividends 731 Node *zero = _gvn.intcon(0); 732 IfNode *ifff = jump_if_fork_int(a, zero, BoolTest::lt); 733 Node *iff = _gvn.transform( new IfFalseNode(ifff) ); 734 Node *ift = _gvn.transform( new IfTrueNode (ifff) ); 735 Node *reg = jump_if_join(ift, iff); 736 Node *phi = PhiNode::make(reg, NULL, TypeInt::INT); 737 // Negative path; negate/and/negate 738 Node *neg = _gvn.transform( new SubINode(zero, a) ); 739 Node *andn= _gvn.transform( new AndINode(neg, mask) ); 740 Node *negn= _gvn.transform( new SubINode(zero, andn) ); 741 phi->init_req(1, negn); 742 // Fast positive case 743 Node *andx = _gvn.transform( new AndINode(a, mask) ); 744 phi->init_req(2, andx); 745 // Push the merge 746 push( _gvn.transform(phi) ); 747 return; 748 } 749 } 750 } 751 // Default case 752 push( _gvn.transform( new ModINode(control(),a,b) ) ); 753 } 754 755 // Handle jsr and jsr_w bytecode 756 void Parse::do_jsr() { 757 assert(bc() == Bytecodes::_jsr || bc() == Bytecodes::_jsr_w, "wrong bytecode"); 758 759 // Store information about current state, tagged with new _jsr_bci 760 int return_bci = iter().next_bci(); 761 int jsr_bci = (bc() == Bytecodes::_jsr) ? iter().get_dest() : iter().get_far_dest(); 762 763 // Update method data 764 profile_taken_branch(jsr_bci); 765 766 // The way we do things now, there is only one successor block 767 // for the jsr, because the target code is cloned by ciTypeFlow. 768 Block* target = successor_for_bci(jsr_bci); 769 770 // What got pushed? 771 const Type* ret_addr = target->peek(); 772 assert(ret_addr->singleton(), "must be a constant (cloned jsr body)"); 773 774 // Effect on jsr on stack 775 push(_gvn.makecon(ret_addr)); 776 777 // Flow to the jsr. 778 merge(jsr_bci); 779 } 780 781 // Handle ret bytecode 782 void Parse::do_ret() { 783 // Find to whom we return. 784 assert(block()->num_successors() == 1, "a ret can only go one place now"); 785 Block* target = block()->successor_at(0); 786 assert(!target->is_ready(), "our arrival must be expected"); 787 profile_ret(target->flow()->start()); 788 int pnum = target->next_path_num(); 789 merge_common(target, pnum); 790 } 791 792 static bool has_injected_profile(BoolTest::mask btest, Node* test, int& taken, int& not_taken) { 793 if (btest != BoolTest::eq && btest != BoolTest::ne) { 794 // Only ::eq and ::ne are supported for profile injection. 795 return false; 796 } 797 if (test->is_Cmp() && 798 test->in(1)->Opcode() == Op_ProfileBoolean) { 799 ProfileBooleanNode* profile = (ProfileBooleanNode*)test->in(1); 800 int false_cnt = profile->false_count(); 801 int true_cnt = profile->true_count(); 802 803 // Counts matching depends on the actual test operation (::eq or ::ne). 804 // No need to scale the counts because profile injection was designed 805 // to feed exact counts into VM. 806 taken = (btest == BoolTest::eq) ? false_cnt : true_cnt; 807 not_taken = (btest == BoolTest::eq) ? true_cnt : false_cnt; 808 809 profile->consume(); 810 return true; 811 } 812 return false; 813 } 814 //--------------------------dynamic_branch_prediction-------------------------- 815 // Try to gather dynamic branch prediction behavior. Return a probability 816 // of the branch being taken and set the "cnt" field. Returns a -1.0 817 // if we need to use static prediction for some reason. 818 float Parse::dynamic_branch_prediction(float &cnt, BoolTest::mask btest, Node* test) { 819 ResourceMark rm; 820 821 cnt = COUNT_UNKNOWN; 822 823 int taken = 0; 824 int not_taken = 0; 825 826 bool use_mdo = !has_injected_profile(btest, test, taken, not_taken); 827 828 if (use_mdo) { 829 // Use MethodData information if it is available 830 // FIXME: free the ProfileData structure 831 ciMethodData* methodData = method()->method_data(); 832 if (!methodData->is_mature()) return PROB_UNKNOWN; 833 ciProfileData* data = methodData->bci_to_data(bci()); 834 if (!data->is_JumpData()) return PROB_UNKNOWN; 835 836 // get taken and not taken values 837 taken = data->as_JumpData()->taken(); 838 not_taken = 0; 839 if (data->is_BranchData()) { 840 not_taken = data->as_BranchData()->not_taken(); 841 } 842 843 // scale the counts to be commensurate with invocation counts: 844 taken = method()->scale_count(taken); 845 not_taken = method()->scale_count(not_taken); 846 } 847 848 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful. 849 // We also check that individual counters are positive first, otherwise the sum can become positive. 850 if (taken < 0 || not_taken < 0 || taken + not_taken < 40) { 851 if (C->log() != NULL) { 852 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d'", iter().get_dest(), taken, not_taken); 853 } 854 return PROB_UNKNOWN; 855 } 856 857 // Compute frequency that we arrive here 858 float sum = taken + not_taken; 859 // Adjust, if this block is a cloned private block but the 860 // Jump counts are shared. Taken the private counts for 861 // just this path instead of the shared counts. 862 if( block()->count() > 0 ) 863 sum = block()->count(); 864 cnt = sum / FreqCountInvocations; 865 866 // Pin probability to sane limits 867 float prob; 868 if( !taken ) 869 prob = (0+PROB_MIN) / 2; 870 else if( !not_taken ) 871 prob = (1+PROB_MAX) / 2; 872 else { // Compute probability of true path 873 prob = (float)taken / (float)(taken + not_taken); 874 if (prob > PROB_MAX) prob = PROB_MAX; 875 if (prob < PROB_MIN) prob = PROB_MIN; 876 } 877 878 assert((cnt > 0.0f) && (prob > 0.0f), 879 "Bad frequency assignment in if"); 880 881 if (C->log() != NULL) { 882 const char* prob_str = NULL; 883 if (prob >= PROB_MAX) prob_str = (prob == PROB_MAX) ? "max" : "always"; 884 if (prob <= PROB_MIN) prob_str = (prob == PROB_MIN) ? "min" : "never"; 885 char prob_str_buf[30]; 886 if (prob_str == NULL) { 887 sprintf(prob_str_buf, "%g", prob); 888 prob_str = prob_str_buf; 889 } 890 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d' cnt='%f' prob='%s'", 891 iter().get_dest(), taken, not_taken, cnt, prob_str); 892 } 893 return prob; 894 } 895 896 //-----------------------------branch_prediction------------------------------- 897 float Parse::branch_prediction(float& cnt, 898 BoolTest::mask btest, 899 int target_bci, 900 Node* test) { 901 float prob = dynamic_branch_prediction(cnt, btest, test); 902 // If prob is unknown, switch to static prediction 903 if (prob != PROB_UNKNOWN) return prob; 904 905 prob = PROB_FAIR; // Set default value 906 if (btest == BoolTest::eq) // Exactly equal test? 907 prob = PROB_STATIC_INFREQUENT; // Assume its relatively infrequent 908 else if (btest == BoolTest::ne) 909 prob = PROB_STATIC_FREQUENT; // Assume its relatively frequent 910 911 // If this is a conditional test guarding a backwards branch, 912 // assume its a loop-back edge. Make it a likely taken branch. 913 if (target_bci < bci()) { 914 if (is_osr_parse()) { // Could be a hot OSR'd loop; force deopt 915 // Since it's an OSR, we probably have profile data, but since 916 // branch_prediction returned PROB_UNKNOWN, the counts are too small. 917 // Let's make a special check here for completely zero counts. 918 ciMethodData* methodData = method()->method_data(); 919 if (!methodData->is_empty()) { 920 ciProfileData* data = methodData->bci_to_data(bci()); 921 // Only stop for truly zero counts, which mean an unknown part 922 // of the OSR-ed method, and we want to deopt to gather more stats. 923 // If you have ANY counts, then this loop is simply 'cold' relative 924 // to the OSR loop. 925 if (data->as_BranchData()->taken() + 926 data->as_BranchData()->not_taken() == 0 ) { 927 // This is the only way to return PROB_UNKNOWN: 928 return PROB_UNKNOWN; 929 } 930 } 931 } 932 prob = PROB_STATIC_FREQUENT; // Likely to take backwards branch 933 } 934 935 assert(prob != PROB_UNKNOWN, "must have some guess at this point"); 936 return prob; 937 } 938 939 // The magic constants are chosen so as to match the output of 940 // branch_prediction() when the profile reports a zero taken count. 941 // It is important to distinguish zero counts unambiguously, because 942 // some branches (e.g., _213_javac.Assembler.eliminate) validly produce 943 // very small but nonzero probabilities, which if confused with zero 944 // counts would keep the program recompiling indefinitely. 945 bool Parse::seems_never_taken(float prob) const { 946 return prob < PROB_MIN; 947 } 948 949 // True if the comparison seems to be the kind that will not change its 950 // statistics from true to false. See comments in adjust_map_after_if. 951 // This question is only asked along paths which are already 952 // classifed as untaken (by seems_never_taken), so really, 953 // if a path is never taken, its controlling comparison is 954 // already acting in a stable fashion. If the comparison 955 // seems stable, we will put an expensive uncommon trap 956 // on the untaken path. 957 bool Parse::seems_stable_comparison() const { 958 if (C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if)) { 959 return false; 960 } 961 return true; 962 } 963 964 //-------------------------------repush_if_args-------------------------------- 965 // Push arguments of an "if" bytecode back onto the stack by adjusting _sp. 966 inline int Parse::repush_if_args() { 967 if (PrintOpto && WizardMode) { 968 tty->print("defending against excessive implicit null exceptions on %s @%d in ", 969 Bytecodes::name(iter().cur_bc()), iter().cur_bci()); 970 method()->print_name(); tty->cr(); 971 } 972 int bc_depth = - Bytecodes::depth(iter().cur_bc()); 973 assert(bc_depth == 1 || bc_depth == 2, "only two kinds of branches"); 974 DEBUG_ONLY(sync_jvms()); // argument(n) requires a synced jvms 975 assert(argument(0) != NULL, "must exist"); 976 assert(bc_depth == 1 || argument(1) != NULL, "two must exist"); 977 inc_sp(bc_depth); 978 return bc_depth; 979 } 980 981 //----------------------------------do_ifnull---------------------------------- 982 void Parse::do_ifnull(BoolTest::mask btest, Node *c) { 983 int target_bci = iter().get_dest(); 984 985 Block* branch_block = successor_for_bci(target_bci); 986 Block* next_block = successor_for_bci(iter().next_bci()); 987 988 float cnt; 989 float prob = branch_prediction(cnt, btest, target_bci, c); 990 if (prob == PROB_UNKNOWN) { 991 // (An earlier version of do_ifnull omitted this trap for OSR methods.) 992 if (PrintOpto && Verbose) { 993 tty->print_cr("Never-taken edge stops compilation at bci %d", bci()); 994 } 995 repush_if_args(); // to gather stats on loop 996 // We need to mark this branch as taken so that if we recompile we will 997 // see that it is possible. In the tiered system the interpreter doesn't 998 // do profiling and by the time we get to the lower tier from the interpreter 999 // the path may be cold again. Make sure it doesn't look untaken 1000 profile_taken_branch(target_bci, !ProfileInterpreter); 1001 uncommon_trap(Deoptimization::Reason_unreached, 1002 Deoptimization::Action_reinterpret, 1003 NULL, "cold"); 1004 if (C->eliminate_boxing()) { 1005 // Mark the successor blocks as parsed 1006 branch_block->next_path_num(); 1007 next_block->next_path_num(); 1008 } 1009 return; 1010 } 1011 1012 NOT_PRODUCT(explicit_null_checks_inserted++); 1013 1014 // Generate real control flow 1015 Node *tst = _gvn.transform( new BoolNode( c, btest ) ); 1016 1017 // Sanity check the probability value 1018 assert(prob > 0.0f,"Bad probability in Parser"); 1019 // Need xform to put node in hash table 1020 IfNode *iff = create_and_xform_if( control(), tst, prob, cnt ); 1021 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser"); 1022 // True branch 1023 { PreserveJVMState pjvms(this); 1024 Node* iftrue = _gvn.transform( new IfTrueNode (iff) ); 1025 set_control(iftrue); 1026 1027 if (stopped()) { // Path is dead? 1028 NOT_PRODUCT(explicit_null_checks_elided++); 1029 if (C->eliminate_boxing()) { 1030 // Mark the successor block as parsed 1031 branch_block->next_path_num(); 1032 } 1033 } else { // Path is live. 1034 // Update method data 1035 profile_taken_branch(target_bci); 1036 adjust_map_after_if(btest, c, prob, branch_block, next_block); 1037 if (!stopped()) { 1038 merge(target_bci); 1039 } 1040 } 1041 } 1042 1043 // False branch 1044 Node* iffalse = _gvn.transform( new IfFalseNode(iff) ); 1045 set_control(iffalse); 1046 1047 if (stopped()) { // Path is dead? 1048 NOT_PRODUCT(explicit_null_checks_elided++); 1049 if (C->eliminate_boxing()) { 1050 // Mark the successor block as parsed 1051 next_block->next_path_num(); 1052 } 1053 } else { // Path is live. 1054 // Update method data 1055 profile_not_taken_branch(); 1056 adjust_map_after_if(BoolTest(btest).negate(), c, 1.0-prob, 1057 next_block, branch_block); 1058 } 1059 } 1060 1061 //------------------------------------do_if------------------------------------ 1062 void Parse::do_if(BoolTest::mask btest, Node* c) { 1063 int target_bci = iter().get_dest(); 1064 1065 Block* branch_block = successor_for_bci(target_bci); 1066 Block* next_block = successor_for_bci(iter().next_bci()); 1067 1068 float cnt; 1069 float prob = branch_prediction(cnt, btest, target_bci, c); 1070 float untaken_prob = 1.0 - prob; 1071 1072 if (prob == PROB_UNKNOWN) { 1073 if (PrintOpto && Verbose) { 1074 tty->print_cr("Never-taken edge stops compilation at bci %d", bci()); 1075 } 1076 repush_if_args(); // to gather stats on loop 1077 // We need to mark this branch as taken so that if we recompile we will 1078 // see that it is possible. In the tiered system the interpreter doesn't 1079 // do profiling and by the time we get to the lower tier from the interpreter 1080 // the path may be cold again. Make sure it doesn't look untaken 1081 profile_taken_branch(target_bci, !ProfileInterpreter); 1082 uncommon_trap(Deoptimization::Reason_unreached, 1083 Deoptimization::Action_reinterpret, 1084 NULL, "cold"); 1085 if (C->eliminate_boxing()) { 1086 // Mark the successor blocks as parsed 1087 branch_block->next_path_num(); 1088 next_block->next_path_num(); 1089 } 1090 return; 1091 } 1092 1093 // Sanity check the probability value 1094 assert(0.0f < prob && prob < 1.0f,"Bad probability in Parser"); 1095 1096 bool taken_if_true = true; 1097 // Convert BoolTest to canonical form: 1098 if (!BoolTest(btest).is_canonical()) { 1099 btest = BoolTest(btest).negate(); 1100 taken_if_true = false; 1101 // prob is NOT updated here; it remains the probability of the taken 1102 // path (as opposed to the prob of the path guarded by an 'IfTrueNode'). 1103 } 1104 assert(btest != BoolTest::eq, "!= is the only canonical exact test"); 1105 1106 Node* tst0 = new BoolNode(c, btest); 1107 Node* tst = _gvn.transform(tst0); 1108 BoolTest::mask taken_btest = BoolTest::illegal; 1109 BoolTest::mask untaken_btest = BoolTest::illegal; 1110 1111 if (tst->is_Bool()) { 1112 // Refresh c from the transformed bool node, since it may be 1113 // simpler than the original c. Also re-canonicalize btest. 1114 // This wins when (Bool ne (Conv2B p) 0) => (Bool ne (CmpP p NULL)). 1115 // That can arise from statements like: if (x instanceof C) ... 1116 if (tst != tst0) { 1117 // Canonicalize one more time since transform can change it. 1118 btest = tst->as_Bool()->_test._test; 1119 if (!BoolTest(btest).is_canonical()) { 1120 // Reverse edges one more time... 1121 tst = _gvn.transform( tst->as_Bool()->negate(&_gvn) ); 1122 btest = tst->as_Bool()->_test._test; 1123 assert(BoolTest(btest).is_canonical(), "sanity"); 1124 taken_if_true = !taken_if_true; 1125 } 1126 c = tst->in(1); 1127 } 1128 BoolTest::mask neg_btest = BoolTest(btest).negate(); 1129 taken_btest = taken_if_true ? btest : neg_btest; 1130 untaken_btest = taken_if_true ? neg_btest : btest; 1131 } 1132 1133 // Generate real control flow 1134 float true_prob = (taken_if_true ? prob : untaken_prob); 1135 IfNode* iff = create_and_map_if(control(), tst, true_prob, cnt); 1136 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser"); 1137 Node* taken_branch = new IfTrueNode(iff); 1138 Node* untaken_branch = new IfFalseNode(iff); 1139 if (!taken_if_true) { // Finish conversion to canonical form 1140 Node* tmp = taken_branch; 1141 taken_branch = untaken_branch; 1142 untaken_branch = tmp; 1143 } 1144 1145 // Branch is taken: 1146 { PreserveJVMState pjvms(this); 1147 taken_branch = _gvn.transform(taken_branch); 1148 set_control(taken_branch); 1149 1150 if (stopped()) { 1151 if (C->eliminate_boxing()) { 1152 // Mark the successor block as parsed 1153 branch_block->next_path_num(); 1154 } 1155 } else { 1156 // Update method data 1157 profile_taken_branch(target_bci); 1158 adjust_map_after_if(taken_btest, c, prob, branch_block, next_block); 1159 if (!stopped()) { 1160 merge(target_bci); 1161 } 1162 } 1163 } 1164 1165 untaken_branch = _gvn.transform(untaken_branch); 1166 set_control(untaken_branch); 1167 1168 // Branch not taken. 1169 if (stopped()) { 1170 if (C->eliminate_boxing()) { 1171 // Mark the successor block as parsed 1172 next_block->next_path_num(); 1173 } 1174 } else { 1175 // Update method data 1176 profile_not_taken_branch(); 1177 adjust_map_after_if(untaken_btest, c, untaken_prob, 1178 next_block, branch_block); 1179 } 1180 } 1181 1182 bool Parse::path_is_suitable_for_uncommon_trap(float prob) const { 1183 // Don't want to speculate on uncommon traps when running with -Xcomp 1184 if (!UseInterpreter) { 1185 return false; 1186 } 1187 return (seems_never_taken(prob) && seems_stable_comparison()); 1188 } 1189 1190 //----------------------------adjust_map_after_if------------------------------ 1191 // Adjust the JVM state to reflect the result of taking this path. 1192 // Basically, it means inspecting the CmpNode controlling this 1193 // branch, seeing how it constrains a tested value, and then 1194 // deciding if it's worth our while to encode this constraint 1195 // as graph nodes in the current abstract interpretation map. 1196 void Parse::adjust_map_after_if(BoolTest::mask btest, Node* c, float prob, 1197 Block* path, Block* other_path) { 1198 if (stopped() || !c->is_Cmp() || btest == BoolTest::illegal) 1199 return; // nothing to do 1200 1201 bool is_fallthrough = (path == successor_for_bci(iter().next_bci())); 1202 1203 if (path_is_suitable_for_uncommon_trap(prob)) { 1204 repush_if_args(); 1205 uncommon_trap(Deoptimization::Reason_unstable_if, 1206 Deoptimization::Action_reinterpret, 1207 NULL, 1208 (is_fallthrough ? "taken always" : "taken never")); 1209 return; 1210 } 1211 1212 Node* val = c->in(1); 1213 Node* con = c->in(2); 1214 const Type* tcon = _gvn.type(con); 1215 const Type* tval = _gvn.type(val); 1216 bool have_con = tcon->singleton(); 1217 if (tval->singleton()) { 1218 if (!have_con) { 1219 // Swap, so constant is in con. 1220 con = val; 1221 tcon = tval; 1222 val = c->in(2); 1223 tval = _gvn.type(val); 1224 btest = BoolTest(btest).commute(); 1225 have_con = true; 1226 } else { 1227 // Do we have two constants? Then leave well enough alone. 1228 have_con = false; 1229 } 1230 } 1231 if (!have_con) // remaining adjustments need a con 1232 return; 1233 1234 sharpen_type_after_if(btest, con, tcon, val, tval); 1235 } 1236 1237 1238 static Node* extract_obj_from_klass_load(PhaseGVN* gvn, Node* n) { 1239 Node* ldk; 1240 if (n->is_DecodeNKlass()) { 1241 if (n->in(1)->Opcode() != Op_LoadNKlass) { 1242 return NULL; 1243 } else { 1244 ldk = n->in(1); 1245 } 1246 } else if (n->Opcode() != Op_LoadKlass) { 1247 return NULL; 1248 } else { 1249 ldk = n; 1250 } 1251 assert(ldk != NULL && ldk->is_Load(), "should have found a LoadKlass or LoadNKlass node"); 1252 1253 Node* adr = ldk->in(MemNode::Address); 1254 intptr_t off = 0; 1255 Node* obj = AddPNode::Ideal_base_and_offset(adr, gvn, off); 1256 if (obj == NULL || off != oopDesc::klass_offset_in_bytes()) // loading oopDesc::_klass? 1257 return NULL; 1258 const TypePtr* tp = gvn->type(obj)->is_ptr(); 1259 if (tp == NULL || !(tp->isa_instptr() || tp->isa_aryptr())) // is obj a Java object ptr? 1260 return NULL; 1261 1262 return obj; 1263 } 1264 1265 void Parse::sharpen_type_after_if(BoolTest::mask btest, 1266 Node* con, const Type* tcon, 1267 Node* val, const Type* tval) { 1268 // Look for opportunities to sharpen the type of a node 1269 // whose klass is compared with a constant klass. 1270 if (btest == BoolTest::eq && tcon->isa_klassptr()) { 1271 Node* obj = extract_obj_from_klass_load(&_gvn, val); 1272 const TypeOopPtr* con_type = tcon->isa_klassptr()->as_instance_type(); 1273 if (obj != NULL && (con_type->isa_instptr() || con_type->isa_aryptr())) { 1274 // Found: 1275 // Bool(CmpP(LoadKlass(obj._klass), ConP(Foo.klass)), [eq]) 1276 // or the narrowOop equivalent. 1277 const Type* obj_type = _gvn.type(obj); 1278 const TypeOopPtr* tboth = obj_type->join_speculative(con_type)->isa_oopptr(); 1279 if (tboth != NULL && tboth->klass_is_exact() && tboth != obj_type && 1280 tboth->higher_equal(obj_type)) { 1281 // obj has to be of the exact type Foo if the CmpP succeeds. 1282 int obj_in_map = map()->find_edge(obj); 1283 JVMState* jvms = this->jvms(); 1284 if (obj_in_map >= 0 && 1285 (jvms->is_loc(obj_in_map) || jvms->is_stk(obj_in_map))) { 1286 TypeNode* ccast = new CheckCastPPNode(control(), obj, tboth); 1287 const Type* tcc = ccast->as_Type()->type(); 1288 assert(tcc != obj_type && tcc->higher_equal(obj_type), "must improve"); 1289 // Delay transform() call to allow recovery of pre-cast value 1290 // at the control merge. 1291 _gvn.set_type_bottom(ccast); 1292 record_for_igvn(ccast); 1293 // Here's the payoff. 1294 replace_in_map(obj, ccast); 1295 } 1296 } 1297 } 1298 } 1299 1300 int val_in_map = map()->find_edge(val); 1301 if (val_in_map < 0) return; // replace_in_map would be useless 1302 { 1303 JVMState* jvms = this->jvms(); 1304 if (!(jvms->is_loc(val_in_map) || 1305 jvms->is_stk(val_in_map))) 1306 return; // again, it would be useless 1307 } 1308 1309 // Check for a comparison to a constant, and "know" that the compared 1310 // value is constrained on this path. 1311 assert(tcon->singleton(), ""); 1312 ConstraintCastNode* ccast = NULL; 1313 Node* cast = NULL; 1314 1315 switch (btest) { 1316 case BoolTest::eq: // Constant test? 1317 { 1318 const Type* tboth = tcon->join_speculative(tval); 1319 if (tboth == tval) break; // Nothing to gain. 1320 if (tcon->isa_int()) { 1321 ccast = new CastIINode(val, tboth); 1322 } else if (tcon == TypePtr::NULL_PTR) { 1323 // Cast to null, but keep the pointer identity temporarily live. 1324 ccast = new CastPPNode(val, tboth); 1325 } else { 1326 const TypeF* tf = tcon->isa_float_constant(); 1327 const TypeD* td = tcon->isa_double_constant(); 1328 // Exclude tests vs float/double 0 as these could be 1329 // either +0 or -0. Just because you are equal to +0 1330 // doesn't mean you ARE +0! 1331 // Note, following code also replaces Long and Oop values. 1332 if ((!tf || tf->_f != 0.0) && 1333 (!td || td->_d != 0.0)) 1334 cast = con; // Replace non-constant val by con. 1335 } 1336 } 1337 break; 1338 1339 case BoolTest::ne: 1340 if (tcon == TypePtr::NULL_PTR) { 1341 cast = cast_not_null(val, false); 1342 } 1343 break; 1344 1345 default: 1346 // (At this point we could record int range types with CastII.) 1347 break; 1348 } 1349 1350 if (ccast != NULL) { 1351 const Type* tcc = ccast->as_Type()->type(); 1352 assert(tcc != tval && tcc->higher_equal(tval), "must improve"); 1353 // Delay transform() call to allow recovery of pre-cast value 1354 // at the control merge. 1355 ccast->set_req(0, control()); 1356 _gvn.set_type_bottom(ccast); 1357 record_for_igvn(ccast); 1358 cast = ccast; 1359 } 1360 1361 if (cast != NULL) { // Here's the payoff. 1362 replace_in_map(val, cast); 1363 } 1364 } 1365 1366 /** 1367 * Use speculative type to optimize CmpP node: if comparison is 1368 * against the low level class, cast the object to the speculative 1369 * type if any. CmpP should then go away. 1370 * 1371 * @param c expected CmpP node 1372 * @return result of CmpP on object casted to speculative type 1373 * 1374 */ 1375 Node* Parse::optimize_cmp_with_klass(Node* c) { 1376 // If this is transformed by the _gvn to a comparison with the low 1377 // level klass then we may be able to use speculation 1378 if (c->Opcode() == Op_CmpP && 1379 (c->in(1)->Opcode() == Op_LoadKlass || c->in(1)->Opcode() == Op_DecodeNKlass) && 1380 c->in(2)->is_Con()) { 1381 Node* load_klass = NULL; 1382 Node* decode = NULL; 1383 if (c->in(1)->Opcode() == Op_DecodeNKlass) { 1384 decode = c->in(1); 1385 load_klass = c->in(1)->in(1); 1386 } else { 1387 load_klass = c->in(1); 1388 } 1389 if (load_klass->in(2)->is_AddP()) { 1390 Node* addp = load_klass->in(2); 1391 Node* obj = addp->in(AddPNode::Address); 1392 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 1393 if (obj_type->speculative_type_not_null() != NULL) { 1394 ciKlass* k = obj_type->speculative_type(); 1395 inc_sp(2); 1396 obj = maybe_cast_profiled_obj(obj, k); 1397 dec_sp(2); 1398 // Make the CmpP use the casted obj 1399 addp = basic_plus_adr(obj, addp->in(AddPNode::Offset)); 1400 load_klass = load_klass->clone(); 1401 load_klass->set_req(2, addp); 1402 load_klass = _gvn.transform(load_klass); 1403 if (decode != NULL) { 1404 decode = decode->clone(); 1405 decode->set_req(1, load_klass); 1406 load_klass = _gvn.transform(decode); 1407 } 1408 c = c->clone(); 1409 c->set_req(1, load_klass); 1410 c = _gvn.transform(c); 1411 } 1412 } 1413 } 1414 return c; 1415 } 1416 1417 //------------------------------do_one_bytecode-------------------------------- 1418 // Parse this bytecode, and alter the Parsers JVM->Node mapping 1419 void Parse::do_one_bytecode() { 1420 Node *a, *b, *c, *d; // Handy temps 1421 BoolTest::mask btest; 1422 int i; 1423 1424 assert(!has_exceptions(), "bytecode entry state must be clear of throws"); 1425 1426 if (C->check_node_count(NodeLimitFudgeFactor * 5, 1427 "out of nodes parsing method")) { 1428 return; 1429 } 1430 1431 #ifdef ASSERT 1432 // for setting breakpoints 1433 if (TraceOptoParse) { 1434 tty->print(" @"); 1435 dump_bci(bci()); 1436 tty->cr(); 1437 } 1438 #endif 1439 1440 switch (bc()) { 1441 case Bytecodes::_nop: 1442 // do nothing 1443 break; 1444 case Bytecodes::_lconst_0: 1445 push_pair(longcon(0)); 1446 break; 1447 1448 case Bytecodes::_lconst_1: 1449 push_pair(longcon(1)); 1450 break; 1451 1452 case Bytecodes::_fconst_0: 1453 push(zerocon(T_FLOAT)); 1454 break; 1455 1456 case Bytecodes::_fconst_1: 1457 push(makecon(TypeF::ONE)); 1458 break; 1459 1460 case Bytecodes::_fconst_2: 1461 push(makecon(TypeF::make(2.0f))); 1462 break; 1463 1464 case Bytecodes::_dconst_0: 1465 push_pair(zerocon(T_DOUBLE)); 1466 break; 1467 1468 case Bytecodes::_dconst_1: 1469 push_pair(makecon(TypeD::ONE)); 1470 break; 1471 1472 case Bytecodes::_iconst_m1:push(intcon(-1)); break; 1473 case Bytecodes::_iconst_0: push(intcon( 0)); break; 1474 case Bytecodes::_iconst_1: push(intcon( 1)); break; 1475 case Bytecodes::_iconst_2: push(intcon( 2)); break; 1476 case Bytecodes::_iconst_3: push(intcon( 3)); break; 1477 case Bytecodes::_iconst_4: push(intcon( 4)); break; 1478 case Bytecodes::_iconst_5: push(intcon( 5)); break; 1479 case Bytecodes::_bipush: push(intcon(iter().get_constant_u1())); break; 1480 case Bytecodes::_sipush: push(intcon(iter().get_constant_u2())); break; 1481 case Bytecodes::_aconst_null: push(null()); break; 1482 case Bytecodes::_ldc: 1483 case Bytecodes::_ldc_w: 1484 case Bytecodes::_ldc2_w: 1485 // If the constant is unresolved, run this BC once in the interpreter. 1486 { 1487 ciConstant constant = iter().get_constant(); 1488 if (constant.basic_type() == T_OBJECT && 1489 !constant.as_object()->is_loaded()) { 1490 int index = iter().get_constant_pool_index(); 1491 constantTag tag = iter().get_constant_pool_tag(index); 1492 uncommon_trap(Deoptimization::make_trap_request 1493 (Deoptimization::Reason_unloaded, 1494 Deoptimization::Action_reinterpret, 1495 index), 1496 NULL, tag.internal_name()); 1497 break; 1498 } 1499 assert(constant.basic_type() != T_OBJECT || constant.as_object()->is_instance(), 1500 "must be java_mirror of klass"); 1501 const Type* con_type = Type::make_from_constant(constant); 1502 if (con_type != NULL) { 1503 push_node(con_type->basic_type(), makecon(con_type)); 1504 } 1505 } 1506 1507 break; 1508 1509 case Bytecodes::_aload_0: 1510 push( local(0) ); 1511 break; 1512 case Bytecodes::_aload_1: 1513 push( local(1) ); 1514 break; 1515 case Bytecodes::_aload_2: 1516 push( local(2) ); 1517 break; 1518 case Bytecodes::_aload_3: 1519 push( local(3) ); 1520 break; 1521 case Bytecodes::_aload: 1522 case Bytecodes::_vload: 1523 push( local(iter().get_index()) ); 1524 break; 1525 1526 case Bytecodes::_fload_0: 1527 case Bytecodes::_iload_0: 1528 push( local(0) ); 1529 break; 1530 case Bytecodes::_fload_1: 1531 case Bytecodes::_iload_1: 1532 push( local(1) ); 1533 break; 1534 case Bytecodes::_fload_2: 1535 case Bytecodes::_iload_2: 1536 push( local(2) ); 1537 break; 1538 case Bytecodes::_fload_3: 1539 case Bytecodes::_iload_3: 1540 push( local(3) ); 1541 break; 1542 case Bytecodes::_fload: 1543 case Bytecodes::_iload: 1544 push( local(iter().get_index()) ); 1545 break; 1546 case Bytecodes::_lload_0: 1547 push_pair_local( 0 ); 1548 break; 1549 case Bytecodes::_lload_1: 1550 push_pair_local( 1 ); 1551 break; 1552 case Bytecodes::_lload_2: 1553 push_pair_local( 2 ); 1554 break; 1555 case Bytecodes::_lload_3: 1556 push_pair_local( 3 ); 1557 break; 1558 case Bytecodes::_lload: 1559 push_pair_local( iter().get_index() ); 1560 break; 1561 1562 case Bytecodes::_dload_0: 1563 push_pair_local(0); 1564 break; 1565 case Bytecodes::_dload_1: 1566 push_pair_local(1); 1567 break; 1568 case Bytecodes::_dload_2: 1569 push_pair_local(2); 1570 break; 1571 case Bytecodes::_dload_3: 1572 push_pair_local(3); 1573 break; 1574 case Bytecodes::_dload: 1575 push_pair_local(iter().get_index()); 1576 break; 1577 case Bytecodes::_fstore_0: 1578 case Bytecodes::_istore_0: 1579 case Bytecodes::_astore_0: 1580 set_local( 0, pop() ); 1581 break; 1582 case Bytecodes::_fstore_1: 1583 case Bytecodes::_istore_1: 1584 case Bytecodes::_astore_1: 1585 set_local( 1, pop() ); 1586 break; 1587 case Bytecodes::_fstore_2: 1588 case Bytecodes::_istore_2: 1589 case Bytecodes::_astore_2: 1590 set_local( 2, pop() ); 1591 break; 1592 case Bytecodes::_fstore_3: 1593 case Bytecodes::_istore_3: 1594 case Bytecodes::_astore_3: 1595 set_local( 3, pop() ); 1596 break; 1597 case Bytecodes::_fstore: 1598 case Bytecodes::_istore: 1599 case Bytecodes::_astore: 1600 case Bytecodes::_vstore: 1601 set_local( iter().get_index(), pop() ); 1602 break; 1603 // long stores 1604 case Bytecodes::_lstore_0: 1605 set_pair_local( 0, pop_pair() ); 1606 break; 1607 case Bytecodes::_lstore_1: 1608 set_pair_local( 1, pop_pair() ); 1609 break; 1610 case Bytecodes::_lstore_2: 1611 set_pair_local( 2, pop_pair() ); 1612 break; 1613 case Bytecodes::_lstore_3: 1614 set_pair_local( 3, pop_pair() ); 1615 break; 1616 case Bytecodes::_lstore: 1617 set_pair_local( iter().get_index(), pop_pair() ); 1618 break; 1619 1620 // double stores 1621 case Bytecodes::_dstore_0: 1622 set_pair_local( 0, dstore_rounding(pop_pair()) ); 1623 break; 1624 case Bytecodes::_dstore_1: 1625 set_pair_local( 1, dstore_rounding(pop_pair()) ); 1626 break; 1627 case Bytecodes::_dstore_2: 1628 set_pair_local( 2, dstore_rounding(pop_pair()) ); 1629 break; 1630 case Bytecodes::_dstore_3: 1631 set_pair_local( 3, dstore_rounding(pop_pair()) ); 1632 break; 1633 case Bytecodes::_dstore: 1634 set_pair_local( iter().get_index(), dstore_rounding(pop_pair()) ); 1635 break; 1636 1637 case Bytecodes::_pop: dec_sp(1); break; 1638 case Bytecodes::_pop2: dec_sp(2); break; 1639 case Bytecodes::_swap: 1640 a = pop(); 1641 b = pop(); 1642 push(a); 1643 push(b); 1644 break; 1645 case Bytecodes::_dup: 1646 a = pop(); 1647 push(a); 1648 push(a); 1649 break; 1650 case Bytecodes::_dup_x1: 1651 a = pop(); 1652 b = pop(); 1653 push( a ); 1654 push( b ); 1655 push( a ); 1656 break; 1657 case Bytecodes::_dup_x2: 1658 a = pop(); 1659 b = pop(); 1660 c = pop(); 1661 push( a ); 1662 push( c ); 1663 push( b ); 1664 push( a ); 1665 break; 1666 case Bytecodes::_dup2: 1667 a = pop(); 1668 b = pop(); 1669 push( b ); 1670 push( a ); 1671 push( b ); 1672 push( a ); 1673 break; 1674 1675 case Bytecodes::_dup2_x1: 1676 // before: .. c, b, a 1677 // after: .. b, a, c, b, a 1678 // not tested 1679 a = pop(); 1680 b = pop(); 1681 c = pop(); 1682 push( b ); 1683 push( a ); 1684 push( c ); 1685 push( b ); 1686 push( a ); 1687 break; 1688 case Bytecodes::_dup2_x2: 1689 // before: .. d, c, b, a 1690 // after: .. b, a, d, c, b, a 1691 // not tested 1692 a = pop(); 1693 b = pop(); 1694 c = pop(); 1695 d = pop(); 1696 push( b ); 1697 push( a ); 1698 push( d ); 1699 push( c ); 1700 push( b ); 1701 push( a ); 1702 break; 1703 1704 case Bytecodes::_arraylength: { 1705 // Must do null-check with value on expression stack 1706 Node *ary = null_check(peek(), T_ARRAY); 1707 // Compile-time detect of null-exception? 1708 if (stopped()) return; 1709 a = pop(); 1710 push(load_array_length(a)); 1711 break; 1712 } 1713 1714 case Bytecodes::_baload: array_load(T_BYTE); break; 1715 case Bytecodes::_caload: array_load(T_CHAR); break; 1716 case Bytecodes::_iaload: array_load(T_INT); break; 1717 case Bytecodes::_saload: array_load(T_SHORT); break; 1718 case Bytecodes::_faload: array_load(T_FLOAT); break; 1719 case Bytecodes::_vaload: array_load(T_VALUETYPE); break; 1720 case Bytecodes::_aaload: array_load(T_OBJECT); break; 1721 case Bytecodes::_laload: { 1722 a = array_addressing(T_LONG, 0); 1723 if (stopped()) return; // guaranteed null or range check 1724 dec_sp(2); // Pop array and index 1725 push_pair(make_load(control(), a, TypeLong::LONG, T_LONG, TypeAryPtr::LONGS, MemNode::unordered)); 1726 break; 1727 } 1728 case Bytecodes::_daload: { 1729 a = array_addressing(T_DOUBLE, 0); 1730 if (stopped()) return; // guaranteed null or range check 1731 dec_sp(2); // Pop array and index 1732 push_pair(make_load(control(), a, Type::DOUBLE, T_DOUBLE, TypeAryPtr::DOUBLES, MemNode::unordered)); 1733 break; 1734 } 1735 case Bytecodes::_bastore: array_store(T_BYTE); break; 1736 case Bytecodes::_castore: array_store(T_CHAR); break; 1737 case Bytecodes::_iastore: array_store(T_INT); break; 1738 case Bytecodes::_sastore: array_store(T_SHORT); break; 1739 case Bytecodes::_fastore: array_store(T_FLOAT); break; 1740 case Bytecodes::_vastore: { 1741 d = array_addressing(T_OBJECT, 1); 1742 if (stopped()) return; // guaranteed null or range check 1743 // TODO fix this 1744 // array_store_check(); 1745 c = pop(); // Oop to store 1746 b = pop(); // index (already used) 1747 a = pop(); // the array itself 1748 const TypeAryPtr* arytype = _gvn.type(a)->is_aryptr(); 1749 const Type* elemtype = arytype->elem(); 1750 1751 if (elemtype->isa_valuetype()) { 1752 c->as_ValueType()->store(this, a, d); 1753 break; 1754 } 1755 1756 const TypeAryPtr* adr_type = TypeAryPtr::OOPS; 1757 Node* oop = c->as_ValueType()->store_to_memory(this); 1758 Node* store = store_oop_to_array(control(), a, d, adr_type, oop, elemtype->make_oopptr(), T_OBJECT, 1759 StoreNode::release_if_reference(T_OBJECT)); 1760 break; 1761 } 1762 case Bytecodes::_aastore: { 1763 d = array_addressing(T_OBJECT, 1); 1764 if (stopped()) return; // guaranteed null or range check 1765 array_store_check(); 1766 c = pop(); // Oop to store 1767 b = pop(); // index (already used) 1768 a = pop(); // the array itself 1769 const TypeOopPtr* elemtype = _gvn.type(a)->is_aryptr()->elem()->make_oopptr(); 1770 const TypeAryPtr* adr_type = TypeAryPtr::OOPS; 1771 Node* store = store_oop_to_array(control(), a, d, adr_type, c, elemtype, T_OBJECT, 1772 StoreNode::release_if_reference(T_OBJECT)); 1773 break; 1774 } 1775 case Bytecodes::_lastore: { 1776 a = array_addressing(T_LONG, 2); 1777 if (stopped()) return; // guaranteed null or range check 1778 c = pop_pair(); 1779 dec_sp(2); // Pop array and index 1780 store_to_memory(control(), a, c, T_LONG, TypeAryPtr::LONGS, MemNode::unordered); 1781 break; 1782 } 1783 case Bytecodes::_dastore: { 1784 a = array_addressing(T_DOUBLE, 2); 1785 if (stopped()) return; // guaranteed null or range check 1786 c = pop_pair(); 1787 dec_sp(2); // Pop array and index 1788 c = dstore_rounding(c); 1789 store_to_memory(control(), a, c, T_DOUBLE, TypeAryPtr::DOUBLES, MemNode::unordered); 1790 break; 1791 } 1792 1793 case Bytecodes::_getfield: 1794 do_getfield(); 1795 break; 1796 1797 case Bytecodes::_vgetfield: 1798 do_vgetfield(); 1799 break; 1800 1801 case Bytecodes::_getstatic: 1802 do_getstatic(); 1803 break; 1804 1805 case Bytecodes::_putfield: 1806 do_putfield(); 1807 break; 1808 1809 case Bytecodes::_putstatic: 1810 do_putstatic(); 1811 break; 1812 1813 case Bytecodes::_irem: 1814 do_irem(); 1815 break; 1816 case Bytecodes::_idiv: 1817 // Must keep both values on the expression-stack during null-check 1818 zero_check_int(peek()); 1819 // Compile-time detect of null-exception? 1820 if (stopped()) return; 1821 b = pop(); 1822 a = pop(); 1823 push( _gvn.transform( new DivINode(control(),a,b) ) ); 1824 break; 1825 case Bytecodes::_imul: 1826 b = pop(); a = pop(); 1827 push( _gvn.transform( new MulINode(a,b) ) ); 1828 break; 1829 case Bytecodes::_iadd: 1830 b = pop(); a = pop(); 1831 push( _gvn.transform( new AddINode(a,b) ) ); 1832 break; 1833 case Bytecodes::_ineg: 1834 a = pop(); 1835 push( _gvn.transform( new SubINode(_gvn.intcon(0),a)) ); 1836 break; 1837 case Bytecodes::_isub: 1838 b = pop(); a = pop(); 1839 push( _gvn.transform( new SubINode(a,b) ) ); 1840 break; 1841 case Bytecodes::_iand: 1842 b = pop(); a = pop(); 1843 push( _gvn.transform( new AndINode(a,b) ) ); 1844 break; 1845 case Bytecodes::_ior: 1846 b = pop(); a = pop(); 1847 push( _gvn.transform( new OrINode(a,b) ) ); 1848 break; 1849 case Bytecodes::_ixor: 1850 b = pop(); a = pop(); 1851 push( _gvn.transform( new XorINode(a,b) ) ); 1852 break; 1853 case Bytecodes::_ishl: 1854 b = pop(); a = pop(); 1855 push( _gvn.transform( new LShiftINode(a,b) ) ); 1856 break; 1857 case Bytecodes::_ishr: 1858 b = pop(); a = pop(); 1859 push( _gvn.transform( new RShiftINode(a,b) ) ); 1860 break; 1861 case Bytecodes::_iushr: 1862 b = pop(); a = pop(); 1863 push( _gvn.transform( new URShiftINode(a,b) ) ); 1864 break; 1865 1866 case Bytecodes::_fneg: 1867 a = pop(); 1868 b = _gvn.transform(new NegFNode (a)); 1869 push(b); 1870 break; 1871 1872 case Bytecodes::_fsub: 1873 b = pop(); 1874 a = pop(); 1875 c = _gvn.transform( new SubFNode(a,b) ); 1876 d = precision_rounding(c); 1877 push( d ); 1878 break; 1879 1880 case Bytecodes::_fadd: 1881 b = pop(); 1882 a = pop(); 1883 c = _gvn.transform( new AddFNode(a,b) ); 1884 d = precision_rounding(c); 1885 push( d ); 1886 break; 1887 1888 case Bytecodes::_fmul: 1889 b = pop(); 1890 a = pop(); 1891 c = _gvn.transform( new MulFNode(a,b) ); 1892 d = precision_rounding(c); 1893 push( d ); 1894 break; 1895 1896 case Bytecodes::_fdiv: 1897 b = pop(); 1898 a = pop(); 1899 c = _gvn.transform( new DivFNode(0,a,b) ); 1900 d = precision_rounding(c); 1901 push( d ); 1902 break; 1903 1904 case Bytecodes::_frem: 1905 if (Matcher::has_match_rule(Op_ModF)) { 1906 // Generate a ModF node. 1907 b = pop(); 1908 a = pop(); 1909 c = _gvn.transform( new ModFNode(0,a,b) ); 1910 d = precision_rounding(c); 1911 push( d ); 1912 } 1913 else { 1914 // Generate a call. 1915 modf(); 1916 } 1917 break; 1918 1919 case Bytecodes::_fcmpl: 1920 b = pop(); 1921 a = pop(); 1922 c = _gvn.transform( new CmpF3Node( a, b)); 1923 push(c); 1924 break; 1925 case Bytecodes::_fcmpg: 1926 b = pop(); 1927 a = pop(); 1928 1929 // Same as fcmpl but need to flip the unordered case. Swap the inputs, 1930 // which negates the result sign except for unordered. Flip the unordered 1931 // as well by using CmpF3 which implements unordered-lesser instead of 1932 // unordered-greater semantics. Finally, commute the result bits. Result 1933 // is same as using a CmpF3Greater except we did it with CmpF3 alone. 1934 c = _gvn.transform( new CmpF3Node( b, a)); 1935 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) ); 1936 push(c); 1937 break; 1938 1939 case Bytecodes::_f2i: 1940 a = pop(); 1941 push(_gvn.transform(new ConvF2INode(a))); 1942 break; 1943 1944 case Bytecodes::_d2i: 1945 a = pop_pair(); 1946 b = _gvn.transform(new ConvD2INode(a)); 1947 push( b ); 1948 break; 1949 1950 case Bytecodes::_f2d: 1951 a = pop(); 1952 b = _gvn.transform( new ConvF2DNode(a)); 1953 push_pair( b ); 1954 break; 1955 1956 case Bytecodes::_d2f: 1957 a = pop_pair(); 1958 b = _gvn.transform( new ConvD2FNode(a)); 1959 // This breaks _227_mtrt (speed & correctness) and _222_mpegaudio (speed) 1960 //b = _gvn.transform(new RoundFloatNode(0, b) ); 1961 push( b ); 1962 break; 1963 1964 case Bytecodes::_l2f: 1965 if (Matcher::convL2FSupported()) { 1966 a = pop_pair(); 1967 b = _gvn.transform( new ConvL2FNode(a)); 1968 // For i486.ad, FILD doesn't restrict precision to 24 or 53 bits. 1969 // Rather than storing the result into an FP register then pushing 1970 // out to memory to round, the machine instruction that implements 1971 // ConvL2D is responsible for rounding. 1972 // c = precision_rounding(b); 1973 c = _gvn.transform(b); 1974 push(c); 1975 } else { 1976 l2f(); 1977 } 1978 break; 1979 1980 case Bytecodes::_l2d: 1981 a = pop_pair(); 1982 b = _gvn.transform( new ConvL2DNode(a)); 1983 // For i486.ad, rounding is always necessary (see _l2f above). 1984 // c = dprecision_rounding(b); 1985 c = _gvn.transform(b); 1986 push_pair(c); 1987 break; 1988 1989 case Bytecodes::_f2l: 1990 a = pop(); 1991 b = _gvn.transform( new ConvF2LNode(a)); 1992 push_pair(b); 1993 break; 1994 1995 case Bytecodes::_d2l: 1996 a = pop_pair(); 1997 b = _gvn.transform( new ConvD2LNode(a)); 1998 push_pair(b); 1999 break; 2000 2001 case Bytecodes::_dsub: 2002 b = pop_pair(); 2003 a = pop_pair(); 2004 c = _gvn.transform( new SubDNode(a,b) ); 2005 d = dprecision_rounding(c); 2006 push_pair( d ); 2007 break; 2008 2009 case Bytecodes::_dadd: 2010 b = pop_pair(); 2011 a = pop_pair(); 2012 c = _gvn.transform( new AddDNode(a,b) ); 2013 d = dprecision_rounding(c); 2014 push_pair( d ); 2015 break; 2016 2017 case Bytecodes::_dmul: 2018 b = pop_pair(); 2019 a = pop_pair(); 2020 c = _gvn.transform( new MulDNode(a,b) ); 2021 d = dprecision_rounding(c); 2022 push_pair( d ); 2023 break; 2024 2025 case Bytecodes::_ddiv: 2026 b = pop_pair(); 2027 a = pop_pair(); 2028 c = _gvn.transform( new DivDNode(0,a,b) ); 2029 d = dprecision_rounding(c); 2030 push_pair( d ); 2031 break; 2032 2033 case Bytecodes::_dneg: 2034 a = pop_pair(); 2035 b = _gvn.transform(new NegDNode (a)); 2036 push_pair(b); 2037 break; 2038 2039 case Bytecodes::_drem: 2040 if (Matcher::has_match_rule(Op_ModD)) { 2041 // Generate a ModD node. 2042 b = pop_pair(); 2043 a = pop_pair(); 2044 // a % b 2045 2046 c = _gvn.transform( new ModDNode(0,a,b) ); 2047 d = dprecision_rounding(c); 2048 push_pair( d ); 2049 } 2050 else { 2051 // Generate a call. 2052 modd(); 2053 } 2054 break; 2055 2056 case Bytecodes::_dcmpl: 2057 b = pop_pair(); 2058 a = pop_pair(); 2059 c = _gvn.transform( new CmpD3Node( a, b)); 2060 push(c); 2061 break; 2062 2063 case Bytecodes::_dcmpg: 2064 b = pop_pair(); 2065 a = pop_pair(); 2066 // Same as dcmpl but need to flip the unordered case. 2067 // Commute the inputs, which negates the result sign except for unordered. 2068 // Flip the unordered as well by using CmpD3 which implements 2069 // unordered-lesser instead of unordered-greater semantics. 2070 // Finally, negate the result bits. Result is same as using a 2071 // CmpD3Greater except we did it with CmpD3 alone. 2072 c = _gvn.transform( new CmpD3Node( b, a)); 2073 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) ); 2074 push(c); 2075 break; 2076 2077 2078 // Note for longs -> lo word is on TOS, hi word is on TOS - 1 2079 case Bytecodes::_land: 2080 b = pop_pair(); 2081 a = pop_pair(); 2082 c = _gvn.transform( new AndLNode(a,b) ); 2083 push_pair(c); 2084 break; 2085 case Bytecodes::_lor: 2086 b = pop_pair(); 2087 a = pop_pair(); 2088 c = _gvn.transform( new OrLNode(a,b) ); 2089 push_pair(c); 2090 break; 2091 case Bytecodes::_lxor: 2092 b = pop_pair(); 2093 a = pop_pair(); 2094 c = _gvn.transform( new XorLNode(a,b) ); 2095 push_pair(c); 2096 break; 2097 2098 case Bytecodes::_lshl: 2099 b = pop(); // the shift count 2100 a = pop_pair(); // value to be shifted 2101 c = _gvn.transform( new LShiftLNode(a,b) ); 2102 push_pair(c); 2103 break; 2104 case Bytecodes::_lshr: 2105 b = pop(); // the shift count 2106 a = pop_pair(); // value to be shifted 2107 c = _gvn.transform( new RShiftLNode(a,b) ); 2108 push_pair(c); 2109 break; 2110 case Bytecodes::_lushr: 2111 b = pop(); // the shift count 2112 a = pop_pair(); // value to be shifted 2113 c = _gvn.transform( new URShiftLNode(a,b) ); 2114 push_pair(c); 2115 break; 2116 case Bytecodes::_lmul: 2117 b = pop_pair(); 2118 a = pop_pair(); 2119 c = _gvn.transform( new MulLNode(a,b) ); 2120 push_pair(c); 2121 break; 2122 2123 case Bytecodes::_lrem: 2124 // Must keep both values on the expression-stack during null-check 2125 assert(peek(0) == top(), "long word order"); 2126 zero_check_long(peek(1)); 2127 // Compile-time detect of null-exception? 2128 if (stopped()) return; 2129 b = pop_pair(); 2130 a = pop_pair(); 2131 c = _gvn.transform( new ModLNode(control(),a,b) ); 2132 push_pair(c); 2133 break; 2134 2135 case Bytecodes::_ldiv: 2136 // Must keep both values on the expression-stack during null-check 2137 assert(peek(0) == top(), "long word order"); 2138 zero_check_long(peek(1)); 2139 // Compile-time detect of null-exception? 2140 if (stopped()) return; 2141 b = pop_pair(); 2142 a = pop_pair(); 2143 c = _gvn.transform( new DivLNode(control(),a,b) ); 2144 push_pair(c); 2145 break; 2146 2147 case Bytecodes::_ladd: 2148 b = pop_pair(); 2149 a = pop_pair(); 2150 c = _gvn.transform( new AddLNode(a,b) ); 2151 push_pair(c); 2152 break; 2153 case Bytecodes::_lsub: 2154 b = pop_pair(); 2155 a = pop_pair(); 2156 c = _gvn.transform( new SubLNode(a,b) ); 2157 push_pair(c); 2158 break; 2159 case Bytecodes::_lcmp: 2160 // Safepoints are now inserted _before_ branches. The long-compare 2161 // bytecode painfully produces a 3-way value (-1,0,+1) which requires a 2162 // slew of control flow. These are usually followed by a CmpI vs zero and 2163 // a branch; this pattern then optimizes to the obvious long-compare and 2164 // branch. However, if the branch is backwards there's a Safepoint 2165 // inserted. The inserted Safepoint captures the JVM state at the 2166 // pre-branch point, i.e. it captures the 3-way value. Thus if a 2167 // long-compare is used to control a loop the debug info will force 2168 // computation of the 3-way value, even though the generated code uses a 2169 // long-compare and branch. We try to rectify the situation by inserting 2170 // a SafePoint here and have it dominate and kill the safepoint added at a 2171 // following backwards branch. At this point the JVM state merely holds 2 2172 // longs but not the 3-way value. 2173 if( UseLoopSafepoints ) { 2174 switch( iter().next_bc() ) { 2175 case Bytecodes::_ifgt: 2176 case Bytecodes::_iflt: 2177 case Bytecodes::_ifge: 2178 case Bytecodes::_ifle: 2179 case Bytecodes::_ifne: 2180 case Bytecodes::_ifeq: 2181 // If this is a backwards branch in the bytecodes, add Safepoint 2182 maybe_add_safepoint(iter().next_get_dest()); 2183 } 2184 } 2185 b = pop_pair(); 2186 a = pop_pair(); 2187 c = _gvn.transform( new CmpL3Node( a, b )); 2188 push(c); 2189 break; 2190 2191 case Bytecodes::_lneg: 2192 a = pop_pair(); 2193 b = _gvn.transform( new SubLNode(longcon(0),a)); 2194 push_pair(b); 2195 break; 2196 case Bytecodes::_l2i: 2197 a = pop_pair(); 2198 push( _gvn.transform( new ConvL2INode(a))); 2199 break; 2200 case Bytecodes::_i2l: 2201 a = pop(); 2202 b = _gvn.transform( new ConvI2LNode(a)); 2203 push_pair(b); 2204 break; 2205 case Bytecodes::_i2b: 2206 // Sign extend 2207 a = pop(); 2208 a = _gvn.transform( new LShiftINode(a,_gvn.intcon(24)) ); 2209 a = _gvn.transform( new RShiftINode(a,_gvn.intcon(24)) ); 2210 push( a ); 2211 break; 2212 case Bytecodes::_i2s: 2213 a = pop(); 2214 a = _gvn.transform( new LShiftINode(a,_gvn.intcon(16)) ); 2215 a = _gvn.transform( new RShiftINode(a,_gvn.intcon(16)) ); 2216 push( a ); 2217 break; 2218 case Bytecodes::_i2c: 2219 a = pop(); 2220 push( _gvn.transform( new AndINode(a,_gvn.intcon(0xFFFF)) ) ); 2221 break; 2222 2223 case Bytecodes::_i2f: 2224 a = pop(); 2225 b = _gvn.transform( new ConvI2FNode(a) ) ; 2226 c = precision_rounding(b); 2227 push (b); 2228 break; 2229 2230 case Bytecodes::_i2d: 2231 a = pop(); 2232 b = _gvn.transform( new ConvI2DNode(a)); 2233 push_pair(b); 2234 break; 2235 2236 case Bytecodes::_iinc: // Increment local 2237 i = iter().get_index(); // Get local index 2238 set_local( i, _gvn.transform( new AddINode( _gvn.intcon(iter().get_iinc_con()), local(i) ) ) ); 2239 break; 2240 2241 // Exit points of synchronized methods must have an unlock node 2242 case Bytecodes::_return: 2243 return_current(NULL); 2244 break; 2245 2246 case Bytecodes::_ireturn: 2247 case Bytecodes::_areturn: 2248 case Bytecodes::_vreturn: 2249 case Bytecodes::_freturn: 2250 return_current(pop()); 2251 break; 2252 case Bytecodes::_lreturn: 2253 return_current(pop_pair()); 2254 break; 2255 case Bytecodes::_dreturn: 2256 return_current(pop_pair()); 2257 break; 2258 2259 case Bytecodes::_athrow: 2260 // null exception oop throws NULL pointer exception 2261 null_check(peek()); 2262 if (stopped()) return; 2263 // Hook the thrown exception directly to subsequent handlers. 2264 if (BailoutToInterpreterForThrows) { 2265 // Keep method interpreted from now on. 2266 uncommon_trap(Deoptimization::Reason_unhandled, 2267 Deoptimization::Action_make_not_compilable); 2268 return; 2269 } 2270 if (env()->jvmti_can_post_on_exceptions()) { 2271 // check if we must post exception events, take uncommon trap if so (with must_throw = false) 2272 uncommon_trap_if_should_post_on_exceptions(Deoptimization::Reason_unhandled, false); 2273 } 2274 // Here if either can_post_on_exceptions or should_post_on_exceptions is false 2275 add_exception_state(make_exception_state(peek())); 2276 break; 2277 2278 case Bytecodes::_goto: // fall through 2279 case Bytecodes::_goto_w: { 2280 int target_bci = (bc() == Bytecodes::_goto) ? iter().get_dest() : iter().get_far_dest(); 2281 2282 // If this is a backwards branch in the bytecodes, add Safepoint 2283 maybe_add_safepoint(target_bci); 2284 2285 // Update method data 2286 profile_taken_branch(target_bci); 2287 2288 // Merge the current control into the target basic block 2289 merge(target_bci); 2290 2291 // See if we can get some profile data and hand it off to the next block 2292 Block *target_block = block()->successor_for_bci(target_bci); 2293 if (target_block->pred_count() != 1) break; 2294 ciMethodData* methodData = method()->method_data(); 2295 if (!methodData->is_mature()) break; 2296 ciProfileData* data = methodData->bci_to_data(bci()); 2297 assert( data->is_JumpData(), "" ); 2298 int taken = ((ciJumpData*)data)->taken(); 2299 taken = method()->scale_count(taken); 2300 target_block->set_count(taken); 2301 break; 2302 } 2303 2304 case Bytecodes::_ifnull: btest = BoolTest::eq; goto handle_if_null; 2305 case Bytecodes::_ifnonnull: btest = BoolTest::ne; goto handle_if_null; 2306 handle_if_null: 2307 // If this is a backwards branch in the bytecodes, add Safepoint 2308 maybe_add_safepoint(iter().get_dest()); 2309 a = null(); 2310 b = pop(); 2311 if (!_gvn.type(b)->speculative_maybe_null() && 2312 !too_many_traps(Deoptimization::Reason_speculate_null_check)) { 2313 inc_sp(1); 2314 Node* null_ctl = top(); 2315 b = null_check_oop(b, &null_ctl, true, true, true); 2316 assert(null_ctl->is_top(), "no null control here"); 2317 dec_sp(1); 2318 } 2319 c = _gvn.transform( new CmpPNode(b, a) ); 2320 do_ifnull(btest, c); 2321 break; 2322 2323 case Bytecodes::_if_acmpeq: btest = BoolTest::eq; goto handle_if_acmp; 2324 case Bytecodes::_if_acmpne: btest = BoolTest::ne; goto handle_if_acmp; 2325 handle_if_acmp: 2326 // If this is a backwards branch in the bytecodes, add Safepoint 2327 maybe_add_safepoint(iter().get_dest()); 2328 a = pop(); 2329 b = pop(); 2330 c = _gvn.transform( new CmpPNode(b, a) ); 2331 c = optimize_cmp_with_klass(c); 2332 do_if(btest, c); 2333 break; 2334 2335 case Bytecodes::_ifeq: btest = BoolTest::eq; goto handle_ifxx; 2336 case Bytecodes::_ifne: btest = BoolTest::ne; goto handle_ifxx; 2337 case Bytecodes::_iflt: btest = BoolTest::lt; goto handle_ifxx; 2338 case Bytecodes::_ifle: btest = BoolTest::le; goto handle_ifxx; 2339 case Bytecodes::_ifgt: btest = BoolTest::gt; goto handle_ifxx; 2340 case Bytecodes::_ifge: btest = BoolTest::ge; goto handle_ifxx; 2341 handle_ifxx: 2342 // If this is a backwards branch in the bytecodes, add Safepoint 2343 maybe_add_safepoint(iter().get_dest()); 2344 a = _gvn.intcon(0); 2345 b = pop(); 2346 c = _gvn.transform( new CmpINode(b, a) ); 2347 do_if(btest, c); 2348 break; 2349 2350 case Bytecodes::_if_icmpeq: btest = BoolTest::eq; goto handle_if_icmp; 2351 case Bytecodes::_if_icmpne: btest = BoolTest::ne; goto handle_if_icmp; 2352 case Bytecodes::_if_icmplt: btest = BoolTest::lt; goto handle_if_icmp; 2353 case Bytecodes::_if_icmple: btest = BoolTest::le; goto handle_if_icmp; 2354 case Bytecodes::_if_icmpgt: btest = BoolTest::gt; goto handle_if_icmp; 2355 case Bytecodes::_if_icmpge: btest = BoolTest::ge; goto handle_if_icmp; 2356 handle_if_icmp: 2357 // If this is a backwards branch in the bytecodes, add Safepoint 2358 maybe_add_safepoint(iter().get_dest()); 2359 a = pop(); 2360 b = pop(); 2361 c = _gvn.transform( new CmpINode( b, a ) ); 2362 do_if(btest, c); 2363 break; 2364 2365 case Bytecodes::_tableswitch: 2366 do_tableswitch(); 2367 break; 2368 2369 case Bytecodes::_lookupswitch: 2370 do_lookupswitch(); 2371 break; 2372 2373 case Bytecodes::_invokestatic: 2374 case Bytecodes::_invokedynamic: 2375 case Bytecodes::_invokespecial: 2376 case Bytecodes::_invokevirtual: 2377 case Bytecodes::_invokedirect: 2378 case Bytecodes::_invokeinterface: 2379 do_call(); 2380 break; 2381 case Bytecodes::_checkcast: 2382 do_checkcast(); 2383 break; 2384 case Bytecodes::_instanceof: 2385 do_instanceof(); 2386 break; 2387 case Bytecodes::_anewarray: 2388 do_newarray(); 2389 break; 2390 case Bytecodes::_newarray: 2391 do_newarray((BasicType)iter().get_index()); 2392 break; 2393 case Bytecodes::_multianewarray: 2394 do_multianewarray(); 2395 break; 2396 case Bytecodes::_new: 2397 do_new(); 2398 break; 2399 case Bytecodes::_vnew: 2400 do_vnew(); 2401 break; 2402 case Bytecodes::_vdefault: 2403 do_vdefault(); 2404 break; 2405 case Bytecodes::_vwithfield: 2406 do_vwithfield(); 2407 break; 2408 2409 case Bytecodes::_jsr: 2410 case Bytecodes::_jsr_w: 2411 do_jsr(); 2412 break; 2413 2414 case Bytecodes::_ret: 2415 do_ret(); 2416 break; 2417 2418 2419 case Bytecodes::_monitorenter: 2420 do_monitor_enter(); 2421 break; 2422 2423 case Bytecodes::_monitorexit: 2424 do_monitor_exit(); 2425 break; 2426 2427 case Bytecodes::_vunbox: 2428 do_vunbox(); 2429 break; 2430 2431 case Bytecodes::_vbox: 2432 do_vbox(); 2433 break; 2434 2435 case Bytecodes::_breakpoint: 2436 // Breakpoint set concurrently to compile 2437 // %%% use an uncommon trap? 2438 C->record_failure("breakpoint in method"); 2439 return; 2440 2441 default: 2442 #ifndef PRODUCT 2443 map()->dump(99); 2444 #endif 2445 tty->print("\nUnhandled bytecode %s\n", Bytecodes::name(bc()) ); 2446 ShouldNotReachHere(); 2447 } 2448 2449 #ifndef PRODUCT 2450 IdealGraphPrinter *printer = C->printer(); 2451 if (printer && printer->should_print(1)) { 2452 char buffer[256]; 2453 sprintf(buffer, "Bytecode %d: %s", bci(), Bytecodes::name(bc())); 2454 bool old = printer->traverse_outs(); 2455 printer->set_traverse_outs(true); 2456 printer->print_method(buffer, 4); 2457 printer->set_traverse_outs(old); 2458 } 2459 #endif 2460 }