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(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() == 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 == NULL) { 830 return PROB_UNKNOWN; 831 } 832 if (!data->is_JumpData()) return PROB_UNKNOWN; 833 834 // get taken and not taken values 835 taken = data->as_JumpData()->taken(); 836 not_taken = 0; 837 if (data->is_BranchData()) { 838 not_taken = data->as_BranchData()->not_taken(); 839 } 840 841 // scale the counts to be commensurate with invocation counts: 842 taken = method()->scale_count(taken); 843 not_taken = method()->scale_count(not_taken); 844 } 845 846 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful. 847 // We also check that individual counters are positive first, otherwise the sum can become positive. 848 if (taken < 0 || not_taken < 0 || taken + not_taken < 40) { 849 if (C->log() != NULL) { 850 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d'", iter().get_dest(), taken, not_taken); 851 } 852 return PROB_UNKNOWN; 853 } 854 855 // Compute frequency that we arrive here 856 float sum = taken + not_taken; 857 // Adjust, if this block is a cloned private block but the 858 // Jump counts are shared. Taken the private counts for 859 // just this path instead of the shared counts. 860 if( block()->count() > 0 ) 861 sum = block()->count(); 862 cnt = sum / FreqCountInvocations; 863 864 // Pin probability to sane limits 865 float prob; 866 if( !taken ) 867 prob = (0+PROB_MIN) / 2; 868 else if( !not_taken ) 869 prob = (1+PROB_MAX) / 2; 870 else { // Compute probability of true path 871 prob = (float)taken / (float)(taken + not_taken); 872 if (prob > PROB_MAX) prob = PROB_MAX; 873 if (prob < PROB_MIN) prob = PROB_MIN; 874 } 875 876 assert((cnt > 0.0f) && (prob > 0.0f), 877 "Bad frequency assignment in if"); 878 879 if (C->log() != NULL) { 880 const char* prob_str = NULL; 881 if (prob >= PROB_MAX) prob_str = (prob == PROB_MAX) ? "max" : "always"; 882 if (prob <= PROB_MIN) prob_str = (prob == PROB_MIN) ? "min" : "never"; 883 char prob_str_buf[30]; 884 if (prob_str == NULL) { 885 sprintf(prob_str_buf, "%g", prob); 886 prob_str = prob_str_buf; 887 } 888 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d' cnt='%f' prob='%s'", 889 iter().get_dest(), taken, not_taken, cnt, prob_str); 890 } 891 return prob; 892 } 893 894 //-----------------------------branch_prediction------------------------------- 895 float Parse::branch_prediction(float& cnt, 896 BoolTest::mask btest, 897 int target_bci, 898 Node* test) { 899 float prob = dynamic_branch_prediction(cnt, btest, test); 900 // If prob is unknown, switch to static prediction 901 if (prob != PROB_UNKNOWN) return prob; 902 903 prob = PROB_FAIR; // Set default value 904 if (btest == BoolTest::eq) // Exactly equal test? 905 prob = PROB_STATIC_INFREQUENT; // Assume its relatively infrequent 906 else if (btest == BoolTest::ne) 907 prob = PROB_STATIC_FREQUENT; // Assume its relatively frequent 908 909 // If this is a conditional test guarding a backwards branch, 910 // assume its a loop-back edge. Make it a likely taken branch. 911 if (target_bci < bci()) { 912 if (is_osr_parse()) { // Could be a hot OSR'd loop; force deopt 913 // Since it's an OSR, we probably have profile data, but since 914 // branch_prediction returned PROB_UNKNOWN, the counts are too small. 915 // Let's make a special check here for completely zero counts. 916 ciMethodData* methodData = method()->method_data(); 917 if (!methodData->is_empty()) { 918 ciProfileData* data = methodData->bci_to_data(bci()); 919 // Only stop for truly zero counts, which mean an unknown part 920 // of the OSR-ed method, and we want to deopt to gather more stats. 921 // If you have ANY counts, then this loop is simply 'cold' relative 922 // to the OSR loop. 923 if (data == NULL || 924 (data->as_BranchData()->taken() + data->as_BranchData()->not_taken() == 0)) { 925 // This is the only way to return PROB_UNKNOWN: 926 return PROB_UNKNOWN; 927 } 928 } 929 } 930 prob = PROB_STATIC_FREQUENT; // Likely to take backwards branch 931 } 932 933 assert(prob != PROB_UNKNOWN, "must have some guess at this point"); 934 return prob; 935 } 936 937 // The magic constants are chosen so as to match the output of 938 // branch_prediction() when the profile reports a zero taken count. 939 // It is important to distinguish zero counts unambiguously, because 940 // some branches (e.g., _213_javac.Assembler.eliminate) validly produce 941 // very small but nonzero probabilities, which if confused with zero 942 // counts would keep the program recompiling indefinitely. 943 bool Parse::seems_never_taken(float prob) const { 944 return prob < PROB_MIN; 945 } 946 947 // True if the comparison seems to be the kind that will not change its 948 // statistics from true to false. See comments in adjust_map_after_if. 949 // This question is only asked along paths which are already 950 // classifed as untaken (by seems_never_taken), so really, 951 // if a path is never taken, its controlling comparison is 952 // already acting in a stable fashion. If the comparison 953 // seems stable, we will put an expensive uncommon trap 954 // on the untaken path. 955 bool Parse::seems_stable_comparison() const { 956 if (C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if)) { 957 return false; 958 } 959 return true; 960 } 961 962 //-------------------------------repush_if_args-------------------------------- 963 // Push arguments of an "if" bytecode back onto the stack by adjusting _sp. 964 inline int Parse::repush_if_args() { 965 if (PrintOpto && WizardMode) { 966 tty->print("defending against excessive implicit null exceptions on %s @%d in ", 967 Bytecodes::name(iter().cur_bc()), iter().cur_bci()); 968 method()->print_name(); tty->cr(); 969 } 970 int bc_depth = - Bytecodes::depth(iter().cur_bc()); 971 assert(bc_depth == 1 || bc_depth == 2, "only two kinds of branches"); 972 DEBUG_ONLY(sync_jvms()); // argument(n) requires a synced jvms 973 assert(argument(0) != NULL, "must exist"); 974 assert(bc_depth == 1 || argument(1) != NULL, "two must exist"); 975 inc_sp(bc_depth); 976 return bc_depth; 977 } 978 979 //----------------------------------do_ifnull---------------------------------- 980 void Parse::do_ifnull(BoolTest::mask btest, Node *c) { 981 int target_bci = iter().get_dest(); 982 983 Block* branch_block = successor_for_bci(target_bci); 984 Block* next_block = successor_for_bci(iter().next_bci()); 985 986 float cnt; 987 float prob = branch_prediction(cnt, btest, target_bci, c); 988 if (prob == PROB_UNKNOWN) { 989 // (An earlier version of do_ifnull omitted this trap for OSR methods.) 990 if (PrintOpto && Verbose) { 991 tty->print_cr("Never-taken edge stops compilation at bci %d", bci()); 992 } 993 repush_if_args(); // to gather stats on loop 994 // We need to mark this branch as taken so that if we recompile we will 995 // see that it is possible. In the tiered system the interpreter doesn't 996 // do profiling and by the time we get to the lower tier from the interpreter 997 // the path may be cold again. Make sure it doesn't look untaken 998 profile_taken_branch(target_bci, !ProfileInterpreter); 999 uncommon_trap(Deoptimization::Reason_unreached, 1000 Deoptimization::Action_reinterpret, 1001 NULL, "cold"); 1002 if (C->eliminate_boxing()) { 1003 // Mark the successor blocks as parsed 1004 branch_block->next_path_num(); 1005 next_block->next_path_num(); 1006 } 1007 return; 1008 } 1009 1010 NOT_PRODUCT(explicit_null_checks_inserted++); 1011 1012 // Generate real control flow 1013 Node *tst = _gvn.transform( new BoolNode( c, btest ) ); 1014 1015 // Sanity check the probability value 1016 assert(prob > 0.0f,"Bad probability in Parser"); 1017 // Need xform to put node in hash table 1018 IfNode *iff = create_and_xform_if( control(), tst, prob, cnt ); 1019 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser"); 1020 // True branch 1021 { PreserveJVMState pjvms(this); 1022 Node* iftrue = _gvn.transform( new IfTrueNode (iff) ); 1023 set_control(iftrue); 1024 1025 if (stopped()) { // Path is dead? 1026 NOT_PRODUCT(explicit_null_checks_elided++); 1027 if (C->eliminate_boxing()) { 1028 // Mark the successor block as parsed 1029 branch_block->next_path_num(); 1030 } 1031 } else { // Path is live. 1032 // Update method data 1033 profile_taken_branch(target_bci); 1034 adjust_map_after_if(btest, c, prob, branch_block, next_block); 1035 if (!stopped()) { 1036 merge(target_bci); 1037 } 1038 } 1039 } 1040 1041 // False branch 1042 Node* iffalse = _gvn.transform( new IfFalseNode(iff) ); 1043 set_control(iffalse); 1044 1045 if (stopped()) { // Path is dead? 1046 NOT_PRODUCT(explicit_null_checks_elided++); 1047 if (C->eliminate_boxing()) { 1048 // Mark the successor block as parsed 1049 next_block->next_path_num(); 1050 } 1051 } else { // Path is live. 1052 // Update method data 1053 profile_not_taken_branch(); 1054 adjust_map_after_if(BoolTest(btest).negate(), c, 1.0-prob, 1055 next_block, branch_block); 1056 } 1057 } 1058 1059 //------------------------------------do_if------------------------------------ 1060 void Parse::do_if(BoolTest::mask btest, Node* c) { 1061 int target_bci = iter().get_dest(); 1062 1063 Block* branch_block = successor_for_bci(target_bci); 1064 Block* next_block = successor_for_bci(iter().next_bci()); 1065 1066 float cnt; 1067 float prob = branch_prediction(cnt, btest, target_bci, c); 1068 float untaken_prob = 1.0 - prob; 1069 1070 if (prob == PROB_UNKNOWN) { 1071 if (PrintOpto && Verbose) { 1072 tty->print_cr("Never-taken edge stops compilation at bci %d", bci()); 1073 } 1074 repush_if_args(); // to gather stats on loop 1075 // We need to mark this branch as taken so that if we recompile we will 1076 // see that it is possible. In the tiered system the interpreter doesn't 1077 // do profiling and by the time we get to the lower tier from the interpreter 1078 // the path may be cold again. Make sure it doesn't look untaken 1079 profile_taken_branch(target_bci, !ProfileInterpreter); 1080 uncommon_trap(Deoptimization::Reason_unreached, 1081 Deoptimization::Action_reinterpret, 1082 NULL, "cold"); 1083 if (C->eliminate_boxing()) { 1084 // Mark the successor blocks as parsed 1085 branch_block->next_path_num(); 1086 next_block->next_path_num(); 1087 } 1088 return; 1089 } 1090 1091 // Sanity check the probability value 1092 assert(0.0f < prob && prob < 1.0f,"Bad probability in Parser"); 1093 1094 bool taken_if_true = true; 1095 // Convert BoolTest to canonical form: 1096 if (!BoolTest(btest).is_canonical()) { 1097 btest = BoolTest(btest).negate(); 1098 taken_if_true = false; 1099 // prob is NOT updated here; it remains the probability of the taken 1100 // path (as opposed to the prob of the path guarded by an 'IfTrueNode'). 1101 } 1102 assert(btest != BoolTest::eq, "!= is the only canonical exact test"); 1103 1104 Node* tst0 = new BoolNode(c, btest); 1105 Node* tst = _gvn.transform(tst0); 1106 BoolTest::mask taken_btest = BoolTest::illegal; 1107 BoolTest::mask untaken_btest = BoolTest::illegal; 1108 1109 if (tst->is_Bool()) { 1110 // Refresh c from the transformed bool node, since it may be 1111 // simpler than the original c. Also re-canonicalize btest. 1112 // This wins when (Bool ne (Conv2B p) 0) => (Bool ne (CmpP p NULL)). 1113 // That can arise from statements like: if (x instanceof C) ... 1114 if (tst != tst0) { 1115 // Canonicalize one more time since transform can change it. 1116 btest = tst->as_Bool()->_test._test; 1117 if (!BoolTest(btest).is_canonical()) { 1118 // Reverse edges one more time... 1119 tst = _gvn.transform( tst->as_Bool()->negate(&_gvn) ); 1120 btest = tst->as_Bool()->_test._test; 1121 assert(BoolTest(btest).is_canonical(), "sanity"); 1122 taken_if_true = !taken_if_true; 1123 } 1124 c = tst->in(1); 1125 } 1126 BoolTest::mask neg_btest = BoolTest(btest).negate(); 1127 taken_btest = taken_if_true ? btest : neg_btest; 1128 untaken_btest = taken_if_true ? neg_btest : btest; 1129 } 1130 1131 // Generate real control flow 1132 float true_prob = (taken_if_true ? prob : untaken_prob); 1133 IfNode* iff = create_and_map_if(control(), tst, true_prob, cnt); 1134 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser"); 1135 Node* taken_branch = new IfTrueNode(iff); 1136 Node* untaken_branch = new IfFalseNode(iff); 1137 if (!taken_if_true) { // Finish conversion to canonical form 1138 Node* tmp = taken_branch; 1139 taken_branch = untaken_branch; 1140 untaken_branch = tmp; 1141 } 1142 1143 // Branch is taken: 1144 { PreserveJVMState pjvms(this); 1145 taken_branch = _gvn.transform(taken_branch); 1146 set_control(taken_branch); 1147 1148 if (stopped()) { 1149 if (C->eliminate_boxing()) { 1150 // Mark the successor block as parsed 1151 branch_block->next_path_num(); 1152 } 1153 } else { 1154 // Update method data 1155 profile_taken_branch(target_bci); 1156 adjust_map_after_if(taken_btest, c, prob, branch_block, next_block); 1157 if (!stopped()) { 1158 merge(target_bci); 1159 } 1160 } 1161 } 1162 1163 untaken_branch = _gvn.transform(untaken_branch); 1164 set_control(untaken_branch); 1165 1166 // Branch not taken. 1167 if (stopped()) { 1168 if (C->eliminate_boxing()) { 1169 // Mark the successor block as parsed 1170 next_block->next_path_num(); 1171 } 1172 } else { 1173 // Update method data 1174 profile_not_taken_branch(); 1175 adjust_map_after_if(untaken_btest, c, untaken_prob, 1176 next_block, branch_block); 1177 } 1178 } 1179 1180 bool Parse::path_is_suitable_for_uncommon_trap(float prob) const { 1181 // Don't want to speculate on uncommon traps when running with -Xcomp 1182 if (!UseInterpreter) { 1183 return false; 1184 } 1185 return (seems_never_taken(prob) && seems_stable_comparison()); 1186 } 1187 1188 //----------------------------adjust_map_after_if------------------------------ 1189 // Adjust the JVM state to reflect the result of taking this path. 1190 // Basically, it means inspecting the CmpNode controlling this 1191 // branch, seeing how it constrains a tested value, and then 1192 // deciding if it's worth our while to encode this constraint 1193 // as graph nodes in the current abstract interpretation map. 1194 void Parse::adjust_map_after_if(BoolTest::mask btest, Node* c, float prob, 1195 Block* path, Block* other_path) { 1196 if (stopped() || !c->is_Cmp() || btest == BoolTest::illegal) 1197 return; // nothing to do 1198 1199 bool is_fallthrough = (path == successor_for_bci(iter().next_bci())); 1200 1201 if (path_is_suitable_for_uncommon_trap(prob)) { 1202 repush_if_args(); 1203 uncommon_trap(Deoptimization::Reason_unstable_if, 1204 Deoptimization::Action_reinterpret, 1205 NULL, 1206 (is_fallthrough ? "taken always" : "taken never")); 1207 return; 1208 } 1209 1210 Node* val = c->in(1); 1211 Node* con = c->in(2); 1212 const Type* tcon = _gvn.type(con); 1213 const Type* tval = _gvn.type(val); 1214 bool have_con = tcon->singleton(); 1215 if (tval->singleton()) { 1216 if (!have_con) { 1217 // Swap, so constant is in con. 1218 con = val; 1219 tcon = tval; 1220 val = c->in(2); 1221 tval = _gvn.type(val); 1222 btest = BoolTest(btest).commute(); 1223 have_con = true; 1224 } else { 1225 // Do we have two constants? Then leave well enough alone. 1226 have_con = false; 1227 } 1228 } 1229 if (!have_con) // remaining adjustments need a con 1230 return; 1231 1232 sharpen_type_after_if(btest, con, tcon, val, tval); 1233 } 1234 1235 1236 static Node* extract_obj_from_klass_load(PhaseGVN* gvn, Node* n) { 1237 Node* ldk; 1238 if (n->is_DecodeNKlass()) { 1239 if (n->in(1)->Opcode() != Op_LoadNKlass) { 1240 return NULL; 1241 } else { 1242 ldk = n->in(1); 1243 } 1244 } else if (n->Opcode() != Op_LoadKlass) { 1245 return NULL; 1246 } else { 1247 ldk = n; 1248 } 1249 assert(ldk != NULL && ldk->is_Load(), "should have found a LoadKlass or LoadNKlass node"); 1250 1251 Node* adr = ldk->in(MemNode::Address); 1252 intptr_t off = 0; 1253 Node* obj = AddPNode::Ideal_base_and_offset(adr, gvn, off); 1254 if (obj == NULL || off != oopDesc::klass_offset_in_bytes()) // loading oopDesc::_klass? 1255 return NULL; 1256 const TypePtr* tp = gvn->type(obj)->is_ptr(); 1257 if (tp == NULL || !(tp->isa_instptr() || tp->isa_aryptr())) // is obj a Java object ptr? 1258 return NULL; 1259 1260 return obj; 1261 } 1262 1263 void Parse::sharpen_type_after_if(BoolTest::mask btest, 1264 Node* con, const Type* tcon, 1265 Node* val, const Type* tval) { 1266 // Look for opportunities to sharpen the type of a node 1267 // whose klass is compared with a constant klass. 1268 if (btest == BoolTest::eq && tcon->isa_klassptr()) { 1269 Node* obj = extract_obj_from_klass_load(&_gvn, val); 1270 const TypeOopPtr* con_type = tcon->isa_klassptr()->as_instance_type(); 1271 if (obj != NULL && (con_type->isa_instptr() || con_type->isa_aryptr())) { 1272 // Found: 1273 // Bool(CmpP(LoadKlass(obj._klass), ConP(Foo.klass)), [eq]) 1274 // or the narrowOop equivalent. 1275 const Type* obj_type = _gvn.type(obj); 1276 const TypeOopPtr* tboth = obj_type->join_speculative(con_type)->isa_oopptr(); 1277 if (tboth != NULL && tboth->klass_is_exact() && tboth != obj_type && 1278 tboth->higher_equal(obj_type)) { 1279 // obj has to be of the exact type Foo if the CmpP succeeds. 1280 int obj_in_map = map()->find_edge(obj); 1281 JVMState* jvms = this->jvms(); 1282 if (obj_in_map >= 0 && 1283 (jvms->is_loc(obj_in_map) || jvms->is_stk(obj_in_map))) { 1284 TypeNode* ccast = new CheckCastPPNode(control(), obj, tboth); 1285 const Type* tcc = ccast->as_Type()->type(); 1286 assert(tcc != obj_type && tcc->higher_equal(obj_type), "must improve"); 1287 // Delay transform() call to allow recovery of pre-cast value 1288 // at the control merge. 1289 _gvn.set_type_bottom(ccast); 1290 record_for_igvn(ccast); 1291 // Here's the payoff. 1292 replace_in_map(obj, ccast); 1293 } 1294 } 1295 } 1296 } 1297 1298 int val_in_map = map()->find_edge(val); 1299 if (val_in_map < 0) return; // replace_in_map would be useless 1300 { 1301 JVMState* jvms = this->jvms(); 1302 if (!(jvms->is_loc(val_in_map) || 1303 jvms->is_stk(val_in_map))) 1304 return; // again, it would be useless 1305 } 1306 1307 // Check for a comparison to a constant, and "know" that the compared 1308 // value is constrained on this path. 1309 assert(tcon->singleton(), ""); 1310 ConstraintCastNode* ccast = NULL; 1311 Node* cast = NULL; 1312 1313 switch (btest) { 1314 case BoolTest::eq: // Constant test? 1315 { 1316 const Type* tboth = tcon->join_speculative(tval); 1317 if (tboth == tval) break; // Nothing to gain. 1318 if (tcon->isa_int()) { 1319 ccast = new CastIINode(val, tboth); 1320 } else if (tcon == TypePtr::NULL_PTR) { 1321 // Cast to null, but keep the pointer identity temporarily live. 1322 ccast = new CastPPNode(val, tboth); 1323 } else { 1324 const TypeF* tf = tcon->isa_float_constant(); 1325 const TypeD* td = tcon->isa_double_constant(); 1326 // Exclude tests vs float/double 0 as these could be 1327 // either +0 or -0. Just because you are equal to +0 1328 // doesn't mean you ARE +0! 1329 // Note, following code also replaces Long and Oop values. 1330 if ((!tf || tf->_f != 0.0) && 1331 (!td || td->_d != 0.0)) 1332 cast = con; // Replace non-constant val by con. 1333 } 1334 } 1335 break; 1336 1337 case BoolTest::ne: 1338 if (tcon == TypePtr::NULL_PTR) { 1339 cast = cast_not_null(val, false); 1340 } 1341 break; 1342 1343 default: 1344 // (At this point we could record int range types with CastII.) 1345 break; 1346 } 1347 1348 if (ccast != NULL) { 1349 const Type* tcc = ccast->as_Type()->type(); 1350 assert(tcc != tval && tcc->higher_equal(tval), "must improve"); 1351 // Delay transform() call to allow recovery of pre-cast value 1352 // at the control merge. 1353 ccast->set_req(0, control()); 1354 _gvn.set_type_bottom(ccast); 1355 record_for_igvn(ccast); 1356 cast = ccast; 1357 } 1358 1359 if (cast != NULL) { // Here's the payoff. 1360 replace_in_map(val, cast); 1361 } 1362 } 1363 1364 /** 1365 * Use speculative type to optimize CmpP node: if comparison is 1366 * against the low level class, cast the object to the speculative 1367 * type if any. CmpP should then go away. 1368 * 1369 * @param c expected CmpP node 1370 * @return result of CmpP on object casted to speculative type 1371 * 1372 */ 1373 Node* Parse::optimize_cmp_with_klass(Node* c) { 1374 // If this is transformed by the _gvn to a comparison with the low 1375 // level klass then we may be able to use speculation 1376 if (c->Opcode() == Op_CmpP && 1377 (c->in(1)->Opcode() == Op_LoadKlass || c->in(1)->Opcode() == Op_DecodeNKlass) && 1378 c->in(2)->is_Con()) { 1379 Node* load_klass = NULL; 1380 Node* decode = NULL; 1381 if (c->in(1)->Opcode() == Op_DecodeNKlass) { 1382 decode = c->in(1); 1383 load_klass = c->in(1)->in(1); 1384 } else { 1385 load_klass = c->in(1); 1386 } 1387 if (load_klass->in(2)->is_AddP()) { 1388 Node* addp = load_klass->in(2); 1389 Node* obj = addp->in(AddPNode::Address); 1390 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 1391 if (obj_type->speculative_type_not_null() != NULL) { 1392 ciKlass* k = obj_type->speculative_type(); 1393 inc_sp(2); 1394 obj = maybe_cast_profiled_obj(obj, k); 1395 dec_sp(2); 1396 // Make the CmpP use the casted obj 1397 addp = basic_plus_adr(obj, addp->in(AddPNode::Offset)); 1398 load_klass = load_klass->clone(); 1399 load_klass->set_req(2, addp); 1400 load_klass = _gvn.transform(load_klass); 1401 if (decode != NULL) { 1402 decode = decode->clone(); 1403 decode->set_req(1, load_klass); 1404 load_klass = _gvn.transform(decode); 1405 } 1406 c = c->clone(); 1407 c->set_req(1, load_klass); 1408 c = _gvn.transform(c); 1409 } 1410 } 1411 } 1412 return c; 1413 } 1414 1415 //------------------------------do_one_bytecode-------------------------------- 1416 // Parse this bytecode, and alter the Parsers JVM->Node mapping 1417 void Parse::do_one_bytecode() { 1418 Node *a, *b, *c, *d; // Handy temps 1419 BoolTest::mask btest; 1420 int i; 1421 1422 assert(!has_exceptions(), "bytecode entry state must be clear of throws"); 1423 1424 if (C->check_node_count(NodeLimitFudgeFactor * 5, 1425 "out of nodes parsing method")) { 1426 return; 1427 } 1428 1429 #ifdef ASSERT 1430 // for setting breakpoints 1431 if (TraceOptoParse) { 1432 tty->print(" @"); 1433 dump_bci(bci()); 1434 tty->cr(); 1435 } 1436 #endif 1437 1438 switch (bc()) { 1439 case Bytecodes::_nop: 1440 // do nothing 1441 break; 1442 case Bytecodes::_lconst_0: 1443 push_pair(longcon(0)); 1444 break; 1445 1446 case Bytecodes::_lconst_1: 1447 push_pair(longcon(1)); 1448 break; 1449 1450 case Bytecodes::_fconst_0: 1451 push(zerocon(T_FLOAT)); 1452 break; 1453 1454 case Bytecodes::_fconst_1: 1455 push(makecon(TypeF::ONE)); 1456 break; 1457 1458 case Bytecodes::_fconst_2: 1459 push(makecon(TypeF::make(2.0f))); 1460 break; 1461 1462 case Bytecodes::_dconst_0: 1463 push_pair(zerocon(T_DOUBLE)); 1464 break; 1465 1466 case Bytecodes::_dconst_1: 1467 push_pair(makecon(TypeD::ONE)); 1468 break; 1469 1470 case Bytecodes::_iconst_m1:push(intcon(-1)); break; 1471 case Bytecodes::_iconst_0: push(intcon( 0)); break; 1472 case Bytecodes::_iconst_1: push(intcon( 1)); break; 1473 case Bytecodes::_iconst_2: push(intcon( 2)); break; 1474 case Bytecodes::_iconst_3: push(intcon( 3)); break; 1475 case Bytecodes::_iconst_4: push(intcon( 4)); break; 1476 case Bytecodes::_iconst_5: push(intcon( 5)); break; 1477 case Bytecodes::_bipush: push(intcon(iter().get_constant_u1())); break; 1478 case Bytecodes::_sipush: push(intcon(iter().get_constant_u2())); break; 1479 case Bytecodes::_aconst_null: push(null()); break; 1480 case Bytecodes::_ldc: 1481 case Bytecodes::_ldc_w: 1482 case Bytecodes::_ldc2_w: 1483 // If the constant is unresolved, run this BC once in the interpreter. 1484 { 1485 ciConstant constant = iter().get_constant(); 1486 if (constant.basic_type() == T_OBJECT && 1487 !constant.as_object()->is_loaded()) { 1488 int index = iter().get_constant_pool_index(); 1489 constantTag tag = iter().get_constant_pool_tag(index); 1490 uncommon_trap(Deoptimization::make_trap_request 1491 (Deoptimization::Reason_unloaded, 1492 Deoptimization::Action_reinterpret, 1493 index), 1494 NULL, tag.internal_name()); 1495 break; 1496 } 1497 assert(constant.basic_type() != T_OBJECT || constant.as_object()->is_instance(), 1498 "must be java_mirror of klass"); 1499 const Type* con_type = Type::make_from_constant(constant); 1500 if (con_type != NULL) { 1501 push_node(con_type->basic_type(), makecon(con_type)); 1502 } 1503 } 1504 1505 break; 1506 1507 case Bytecodes::_aload_0: 1508 push( local(0) ); 1509 break; 1510 case Bytecodes::_aload_1: 1511 push( local(1) ); 1512 break; 1513 case Bytecodes::_aload_2: 1514 push( local(2) ); 1515 break; 1516 case Bytecodes::_aload_3: 1517 push( local(3) ); 1518 break; 1519 case Bytecodes::_aload: 1520 push( local(iter().get_index()) ); 1521 break; 1522 1523 case Bytecodes::_fload_0: 1524 case Bytecodes::_iload_0: 1525 push( local(0) ); 1526 break; 1527 case Bytecodes::_fload_1: 1528 case Bytecodes::_iload_1: 1529 push( local(1) ); 1530 break; 1531 case Bytecodes::_fload_2: 1532 case Bytecodes::_iload_2: 1533 push( local(2) ); 1534 break; 1535 case Bytecodes::_fload_3: 1536 case Bytecodes::_iload_3: 1537 push( local(3) ); 1538 break; 1539 case Bytecodes::_fload: 1540 case Bytecodes::_iload: 1541 push( local(iter().get_index()) ); 1542 break; 1543 case Bytecodes::_lload_0: 1544 push_pair_local( 0 ); 1545 break; 1546 case Bytecodes::_lload_1: 1547 push_pair_local( 1 ); 1548 break; 1549 case Bytecodes::_lload_2: 1550 push_pair_local( 2 ); 1551 break; 1552 case Bytecodes::_lload_3: 1553 push_pair_local( 3 ); 1554 break; 1555 case Bytecodes::_lload: 1556 push_pair_local( iter().get_index() ); 1557 break; 1558 1559 case Bytecodes::_dload_0: 1560 push_pair_local(0); 1561 break; 1562 case Bytecodes::_dload_1: 1563 push_pair_local(1); 1564 break; 1565 case Bytecodes::_dload_2: 1566 push_pair_local(2); 1567 break; 1568 case Bytecodes::_dload_3: 1569 push_pair_local(3); 1570 break; 1571 case Bytecodes::_dload: 1572 push_pair_local(iter().get_index()); 1573 break; 1574 case Bytecodes::_fstore_0: 1575 case Bytecodes::_istore_0: 1576 case Bytecodes::_astore_0: 1577 set_local( 0, pop() ); 1578 break; 1579 case Bytecodes::_fstore_1: 1580 case Bytecodes::_istore_1: 1581 case Bytecodes::_astore_1: 1582 set_local( 1, pop() ); 1583 break; 1584 case Bytecodes::_fstore_2: 1585 case Bytecodes::_istore_2: 1586 case Bytecodes::_astore_2: 1587 set_local( 2, pop() ); 1588 break; 1589 case Bytecodes::_fstore_3: 1590 case Bytecodes::_istore_3: 1591 case Bytecodes::_astore_3: 1592 set_local( 3, pop() ); 1593 break; 1594 case Bytecodes::_fstore: 1595 case Bytecodes::_istore: 1596 case Bytecodes::_astore: 1597 set_local( iter().get_index(), pop() ); 1598 break; 1599 // long stores 1600 case Bytecodes::_lstore_0: 1601 set_pair_local( 0, pop_pair() ); 1602 break; 1603 case Bytecodes::_lstore_1: 1604 set_pair_local( 1, pop_pair() ); 1605 break; 1606 case Bytecodes::_lstore_2: 1607 set_pair_local( 2, pop_pair() ); 1608 break; 1609 case Bytecodes::_lstore_3: 1610 set_pair_local( 3, pop_pair() ); 1611 break; 1612 case Bytecodes::_lstore: 1613 set_pair_local( iter().get_index(), pop_pair() ); 1614 break; 1615 1616 // double stores 1617 case Bytecodes::_dstore_0: 1618 set_pair_local( 0, dstore_rounding(pop_pair()) ); 1619 break; 1620 case Bytecodes::_dstore_1: 1621 set_pair_local( 1, dstore_rounding(pop_pair()) ); 1622 break; 1623 case Bytecodes::_dstore_2: 1624 set_pair_local( 2, dstore_rounding(pop_pair()) ); 1625 break; 1626 case Bytecodes::_dstore_3: 1627 set_pair_local( 3, dstore_rounding(pop_pair()) ); 1628 break; 1629 case Bytecodes::_dstore: 1630 set_pair_local( iter().get_index(), dstore_rounding(pop_pair()) ); 1631 break; 1632 1633 case Bytecodes::_pop: dec_sp(1); break; 1634 case Bytecodes::_pop2: dec_sp(2); break; 1635 case Bytecodes::_swap: 1636 a = pop(); 1637 b = pop(); 1638 push(a); 1639 push(b); 1640 break; 1641 case Bytecodes::_dup: 1642 a = pop(); 1643 push(a); 1644 push(a); 1645 break; 1646 case Bytecodes::_dup_x1: 1647 a = pop(); 1648 b = pop(); 1649 push( a ); 1650 push( b ); 1651 push( a ); 1652 break; 1653 case Bytecodes::_dup_x2: 1654 a = pop(); 1655 b = pop(); 1656 c = pop(); 1657 push( a ); 1658 push( c ); 1659 push( b ); 1660 push( a ); 1661 break; 1662 case Bytecodes::_dup2: 1663 a = pop(); 1664 b = pop(); 1665 push( b ); 1666 push( a ); 1667 push( b ); 1668 push( a ); 1669 break; 1670 1671 case Bytecodes::_dup2_x1: 1672 // before: .. c, b, a 1673 // after: .. b, a, c, b, a 1674 // not tested 1675 a = pop(); 1676 b = pop(); 1677 c = pop(); 1678 push( b ); 1679 push( a ); 1680 push( c ); 1681 push( b ); 1682 push( a ); 1683 break; 1684 case Bytecodes::_dup2_x2: 1685 // before: .. d, c, b, a 1686 // after: .. b, a, d, c, b, a 1687 // not tested 1688 a = pop(); 1689 b = pop(); 1690 c = pop(); 1691 d = pop(); 1692 push( b ); 1693 push( a ); 1694 push( d ); 1695 push( c ); 1696 push( b ); 1697 push( a ); 1698 break; 1699 1700 case Bytecodes::_arraylength: { 1701 // Must do null-check with value on expression stack 1702 Node *ary = null_check(peek(), T_ARRAY); 1703 // Compile-time detect of null-exception? 1704 if (stopped()) return; 1705 a = pop(); 1706 push(load_array_length(a)); 1707 break; 1708 } 1709 1710 case Bytecodes::_baload: array_load(T_BYTE); break; 1711 case Bytecodes::_caload: array_load(T_CHAR); break; 1712 case Bytecodes::_iaload: array_load(T_INT); break; 1713 case Bytecodes::_saload: array_load(T_SHORT); break; 1714 case Bytecodes::_faload: array_load(T_FLOAT); break; 1715 case Bytecodes::_aaload: array_load(T_OBJECT); break; 1716 case Bytecodes::_laload: { 1717 a = array_addressing(T_LONG, 0); 1718 if (stopped()) return; // guaranteed null or range check 1719 dec_sp(2); // Pop array and index 1720 push_pair(make_load(control(), a, TypeLong::LONG, T_LONG, TypeAryPtr::LONGS, MemNode::unordered)); 1721 break; 1722 } 1723 case Bytecodes::_daload: { 1724 a = array_addressing(T_DOUBLE, 0); 1725 if (stopped()) return; // guaranteed null or range check 1726 dec_sp(2); // Pop array and index 1727 push_pair(make_load(control(), a, Type::DOUBLE, T_DOUBLE, TypeAryPtr::DOUBLES, MemNode::unordered)); 1728 break; 1729 } 1730 case Bytecodes::_bastore: array_store(T_BYTE); break; 1731 case Bytecodes::_castore: array_store(T_CHAR); break; 1732 case Bytecodes::_iastore: array_store(T_INT); break; 1733 case Bytecodes::_sastore: array_store(T_SHORT); break; 1734 case Bytecodes::_fastore: array_store(T_FLOAT); break; 1735 case Bytecodes::_aastore: { 1736 d = array_addressing(T_OBJECT, 1); 1737 if (stopped()) return; // guaranteed null or range check 1738 array_store_check(); 1739 c = pop(); // Oop to store 1740 b = pop(); // index (already used) 1741 a = pop(); // the array itself 1742 const TypeOopPtr* elemtype = _gvn.type(a)->is_aryptr()->elem()->make_oopptr(); 1743 const TypeAryPtr* adr_type = TypeAryPtr::OOPS; 1744 Node* store = store_oop_to_array(control(), a, d, adr_type, c, elemtype, T_OBJECT, 1745 StoreNode::release_if_reference(T_OBJECT)); 1746 break; 1747 } 1748 case Bytecodes::_lastore: { 1749 a = array_addressing(T_LONG, 2); 1750 if (stopped()) return; // guaranteed null or range check 1751 c = pop_pair(); 1752 dec_sp(2); // Pop array and index 1753 store_to_memory(control(), a, c, T_LONG, TypeAryPtr::LONGS, MemNode::unordered); 1754 break; 1755 } 1756 case Bytecodes::_dastore: { 1757 a = array_addressing(T_DOUBLE, 2); 1758 if (stopped()) return; // guaranteed null or range check 1759 c = pop_pair(); 1760 dec_sp(2); // Pop array and index 1761 c = dstore_rounding(c); 1762 store_to_memory(control(), a, c, T_DOUBLE, TypeAryPtr::DOUBLES, MemNode::unordered); 1763 break; 1764 } 1765 case Bytecodes::_getfield: 1766 do_getfield(); 1767 break; 1768 1769 case Bytecodes::_getstatic: 1770 do_getstatic(); 1771 break; 1772 1773 case Bytecodes::_putfield: 1774 do_putfield(); 1775 break; 1776 1777 case Bytecodes::_putstatic: 1778 do_putstatic(); 1779 break; 1780 1781 case Bytecodes::_irem: 1782 do_irem(); 1783 break; 1784 case Bytecodes::_idiv: 1785 // Must keep both values on the expression-stack during null-check 1786 zero_check_int(peek()); 1787 // Compile-time detect of null-exception? 1788 if (stopped()) return; 1789 b = pop(); 1790 a = pop(); 1791 push( _gvn.transform( new DivINode(control(),a,b) ) ); 1792 break; 1793 case Bytecodes::_imul: 1794 b = pop(); a = pop(); 1795 push( _gvn.transform( new MulINode(a,b) ) ); 1796 break; 1797 case Bytecodes::_iadd: 1798 b = pop(); a = pop(); 1799 push( _gvn.transform( new AddINode(a,b) ) ); 1800 break; 1801 case Bytecodes::_ineg: 1802 a = pop(); 1803 push( _gvn.transform( new SubINode(_gvn.intcon(0),a)) ); 1804 break; 1805 case Bytecodes::_isub: 1806 b = pop(); a = pop(); 1807 push( _gvn.transform( new SubINode(a,b) ) ); 1808 break; 1809 case Bytecodes::_iand: 1810 b = pop(); a = pop(); 1811 push( _gvn.transform( new AndINode(a,b) ) ); 1812 break; 1813 case Bytecodes::_ior: 1814 b = pop(); a = pop(); 1815 push( _gvn.transform( new OrINode(a,b) ) ); 1816 break; 1817 case Bytecodes::_ixor: 1818 b = pop(); a = pop(); 1819 push( _gvn.transform( new XorINode(a,b) ) ); 1820 break; 1821 case Bytecodes::_ishl: 1822 b = pop(); a = pop(); 1823 push( _gvn.transform( new LShiftINode(a,b) ) ); 1824 break; 1825 case Bytecodes::_ishr: 1826 b = pop(); a = pop(); 1827 push( _gvn.transform( new RShiftINode(a,b) ) ); 1828 break; 1829 case Bytecodes::_iushr: 1830 b = pop(); a = pop(); 1831 push( _gvn.transform( new URShiftINode(a,b) ) ); 1832 break; 1833 1834 case Bytecodes::_fneg: 1835 a = pop(); 1836 b = _gvn.transform(new NegFNode (a)); 1837 push(b); 1838 break; 1839 1840 case Bytecodes::_fsub: 1841 b = pop(); 1842 a = pop(); 1843 c = _gvn.transform( new SubFNode(a,b) ); 1844 d = precision_rounding(c); 1845 push( d ); 1846 break; 1847 1848 case Bytecodes::_fadd: 1849 b = pop(); 1850 a = pop(); 1851 c = _gvn.transform( new AddFNode(a,b) ); 1852 d = precision_rounding(c); 1853 push( d ); 1854 break; 1855 1856 case Bytecodes::_fmul: 1857 b = pop(); 1858 a = pop(); 1859 c = _gvn.transform( new MulFNode(a,b) ); 1860 d = precision_rounding(c); 1861 push( d ); 1862 break; 1863 1864 case Bytecodes::_fdiv: 1865 b = pop(); 1866 a = pop(); 1867 c = _gvn.transform( new DivFNode(0,a,b) ); 1868 d = precision_rounding(c); 1869 push( d ); 1870 break; 1871 1872 case Bytecodes::_frem: 1873 if (Matcher::has_match_rule(Op_ModF)) { 1874 // Generate a ModF node. 1875 b = pop(); 1876 a = pop(); 1877 c = _gvn.transform( new ModFNode(0,a,b) ); 1878 d = precision_rounding(c); 1879 push( d ); 1880 } 1881 else { 1882 // Generate a call. 1883 modf(); 1884 } 1885 break; 1886 1887 case Bytecodes::_fcmpl: 1888 b = pop(); 1889 a = pop(); 1890 c = _gvn.transform( new CmpF3Node( a, b)); 1891 push(c); 1892 break; 1893 case Bytecodes::_fcmpg: 1894 b = pop(); 1895 a = pop(); 1896 1897 // Same as fcmpl but need to flip the unordered case. Swap the inputs, 1898 // which negates the result sign except for unordered. Flip the unordered 1899 // as well by using CmpF3 which implements unordered-lesser instead of 1900 // unordered-greater semantics. Finally, commute the result bits. Result 1901 // is same as using a CmpF3Greater except we did it with CmpF3 alone. 1902 c = _gvn.transform( new CmpF3Node( b, a)); 1903 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) ); 1904 push(c); 1905 break; 1906 1907 case Bytecodes::_f2i: 1908 a = pop(); 1909 push(_gvn.transform(new ConvF2INode(a))); 1910 break; 1911 1912 case Bytecodes::_d2i: 1913 a = pop_pair(); 1914 b = _gvn.transform(new ConvD2INode(a)); 1915 push( b ); 1916 break; 1917 1918 case Bytecodes::_f2d: 1919 a = pop(); 1920 b = _gvn.transform( new ConvF2DNode(a)); 1921 push_pair( b ); 1922 break; 1923 1924 case Bytecodes::_d2f: 1925 a = pop_pair(); 1926 b = _gvn.transform( new ConvD2FNode(a)); 1927 // This breaks _227_mtrt (speed & correctness) and _222_mpegaudio (speed) 1928 //b = _gvn.transform(new RoundFloatNode(0, b) ); 1929 push( b ); 1930 break; 1931 1932 case Bytecodes::_l2f: 1933 if (Matcher::convL2FSupported()) { 1934 a = pop_pair(); 1935 b = _gvn.transform( new ConvL2FNode(a)); 1936 // For i486.ad, FILD doesn't restrict precision to 24 or 53 bits. 1937 // Rather than storing the result into an FP register then pushing 1938 // out to memory to round, the machine instruction that implements 1939 // ConvL2D is responsible for rounding. 1940 // c = precision_rounding(b); 1941 c = _gvn.transform(b); 1942 push(c); 1943 } else { 1944 l2f(); 1945 } 1946 break; 1947 1948 case Bytecodes::_l2d: 1949 a = pop_pair(); 1950 b = _gvn.transform( new ConvL2DNode(a)); 1951 // For i486.ad, rounding is always necessary (see _l2f above). 1952 // c = dprecision_rounding(b); 1953 c = _gvn.transform(b); 1954 push_pair(c); 1955 break; 1956 1957 case Bytecodes::_f2l: 1958 a = pop(); 1959 b = _gvn.transform( new ConvF2LNode(a)); 1960 push_pair(b); 1961 break; 1962 1963 case Bytecodes::_d2l: 1964 a = pop_pair(); 1965 b = _gvn.transform( new ConvD2LNode(a)); 1966 push_pair(b); 1967 break; 1968 1969 case Bytecodes::_dsub: 1970 b = pop_pair(); 1971 a = pop_pair(); 1972 c = _gvn.transform( new SubDNode(a,b) ); 1973 d = dprecision_rounding(c); 1974 push_pair( d ); 1975 break; 1976 1977 case Bytecodes::_dadd: 1978 b = pop_pair(); 1979 a = pop_pair(); 1980 c = _gvn.transform( new AddDNode(a,b) ); 1981 d = dprecision_rounding(c); 1982 push_pair( d ); 1983 break; 1984 1985 case Bytecodes::_dmul: 1986 b = pop_pair(); 1987 a = pop_pair(); 1988 c = _gvn.transform( new MulDNode(a,b) ); 1989 d = dprecision_rounding(c); 1990 push_pair( d ); 1991 break; 1992 1993 case Bytecodes::_ddiv: 1994 b = pop_pair(); 1995 a = pop_pair(); 1996 c = _gvn.transform( new DivDNode(0,a,b) ); 1997 d = dprecision_rounding(c); 1998 push_pair( d ); 1999 break; 2000 2001 case Bytecodes::_dneg: 2002 a = pop_pair(); 2003 b = _gvn.transform(new NegDNode (a)); 2004 push_pair(b); 2005 break; 2006 2007 case Bytecodes::_drem: 2008 if (Matcher::has_match_rule(Op_ModD)) { 2009 // Generate a ModD node. 2010 b = pop_pair(); 2011 a = pop_pair(); 2012 // a % b 2013 2014 c = _gvn.transform( new ModDNode(0,a,b) ); 2015 d = dprecision_rounding(c); 2016 push_pair( d ); 2017 } 2018 else { 2019 // Generate a call. 2020 modd(); 2021 } 2022 break; 2023 2024 case Bytecodes::_dcmpl: 2025 b = pop_pair(); 2026 a = pop_pair(); 2027 c = _gvn.transform( new CmpD3Node( a, b)); 2028 push(c); 2029 break; 2030 2031 case Bytecodes::_dcmpg: 2032 b = pop_pair(); 2033 a = pop_pair(); 2034 // Same as dcmpl but need to flip the unordered case. 2035 // Commute the inputs, which negates the result sign except for unordered. 2036 // Flip the unordered as well by using CmpD3 which implements 2037 // unordered-lesser instead of unordered-greater semantics. 2038 // Finally, negate the result bits. Result is same as using a 2039 // CmpD3Greater except we did it with CmpD3 alone. 2040 c = _gvn.transform( new CmpD3Node( b, a)); 2041 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) ); 2042 push(c); 2043 break; 2044 2045 2046 // Note for longs -> lo word is on TOS, hi word is on TOS - 1 2047 case Bytecodes::_land: 2048 b = pop_pair(); 2049 a = pop_pair(); 2050 c = _gvn.transform( new AndLNode(a,b) ); 2051 push_pair(c); 2052 break; 2053 case Bytecodes::_lor: 2054 b = pop_pair(); 2055 a = pop_pair(); 2056 c = _gvn.transform( new OrLNode(a,b) ); 2057 push_pair(c); 2058 break; 2059 case Bytecodes::_lxor: 2060 b = pop_pair(); 2061 a = pop_pair(); 2062 c = _gvn.transform( new XorLNode(a,b) ); 2063 push_pair(c); 2064 break; 2065 2066 case Bytecodes::_lshl: 2067 b = pop(); // the shift count 2068 a = pop_pair(); // value to be shifted 2069 c = _gvn.transform( new LShiftLNode(a,b) ); 2070 push_pair(c); 2071 break; 2072 case Bytecodes::_lshr: 2073 b = pop(); // the shift count 2074 a = pop_pair(); // value to be shifted 2075 c = _gvn.transform( new RShiftLNode(a,b) ); 2076 push_pair(c); 2077 break; 2078 case Bytecodes::_lushr: 2079 b = pop(); // the shift count 2080 a = pop_pair(); // value to be shifted 2081 c = _gvn.transform( new URShiftLNode(a,b) ); 2082 push_pair(c); 2083 break; 2084 case Bytecodes::_lmul: 2085 b = pop_pair(); 2086 a = pop_pair(); 2087 c = _gvn.transform( new MulLNode(a,b) ); 2088 push_pair(c); 2089 break; 2090 2091 case Bytecodes::_lrem: 2092 // Must keep both values on the expression-stack during null-check 2093 assert(peek(0) == top(), "long word order"); 2094 zero_check_long(peek(1)); 2095 // Compile-time detect of null-exception? 2096 if (stopped()) return; 2097 b = pop_pair(); 2098 a = pop_pair(); 2099 c = _gvn.transform( new ModLNode(control(),a,b) ); 2100 push_pair(c); 2101 break; 2102 2103 case Bytecodes::_ldiv: 2104 // Must keep both values on the expression-stack during null-check 2105 assert(peek(0) == top(), "long word order"); 2106 zero_check_long(peek(1)); 2107 // Compile-time detect of null-exception? 2108 if (stopped()) return; 2109 b = pop_pair(); 2110 a = pop_pair(); 2111 c = _gvn.transform( new DivLNode(control(),a,b) ); 2112 push_pair(c); 2113 break; 2114 2115 case Bytecodes::_ladd: 2116 b = pop_pair(); 2117 a = pop_pair(); 2118 c = _gvn.transform( new AddLNode(a,b) ); 2119 push_pair(c); 2120 break; 2121 case Bytecodes::_lsub: 2122 b = pop_pair(); 2123 a = pop_pair(); 2124 c = _gvn.transform( new SubLNode(a,b) ); 2125 push_pair(c); 2126 break; 2127 case Bytecodes::_lcmp: 2128 // Safepoints are now inserted _before_ branches. The long-compare 2129 // bytecode painfully produces a 3-way value (-1,0,+1) which requires a 2130 // slew of control flow. These are usually followed by a CmpI vs zero and 2131 // a branch; this pattern then optimizes to the obvious long-compare and 2132 // branch. However, if the branch is backwards there's a Safepoint 2133 // inserted. The inserted Safepoint captures the JVM state at the 2134 // pre-branch point, i.e. it captures the 3-way value. Thus if a 2135 // long-compare is used to control a loop the debug info will force 2136 // computation of the 3-way value, even though the generated code uses a 2137 // long-compare and branch. We try to rectify the situation by inserting 2138 // a SafePoint here and have it dominate and kill the safepoint added at a 2139 // following backwards branch. At this point the JVM state merely holds 2 2140 // longs but not the 3-way value. 2141 if( UseLoopSafepoints ) { 2142 switch( iter().next_bc() ) { 2143 case Bytecodes::_ifgt: 2144 case Bytecodes::_iflt: 2145 case Bytecodes::_ifge: 2146 case Bytecodes::_ifle: 2147 case Bytecodes::_ifne: 2148 case Bytecodes::_ifeq: 2149 // If this is a backwards branch in the bytecodes, add Safepoint 2150 maybe_add_safepoint(iter().next_get_dest()); 2151 } 2152 } 2153 b = pop_pair(); 2154 a = pop_pair(); 2155 c = _gvn.transform( new CmpL3Node( a, b )); 2156 push(c); 2157 break; 2158 2159 case Bytecodes::_lneg: 2160 a = pop_pair(); 2161 b = _gvn.transform( new SubLNode(longcon(0),a)); 2162 push_pair(b); 2163 break; 2164 case Bytecodes::_l2i: 2165 a = pop_pair(); 2166 push( _gvn.transform( new ConvL2INode(a))); 2167 break; 2168 case Bytecodes::_i2l: 2169 a = pop(); 2170 b = _gvn.transform( new ConvI2LNode(a)); 2171 push_pair(b); 2172 break; 2173 case Bytecodes::_i2b: 2174 // Sign extend 2175 a = pop(); 2176 a = _gvn.transform( new LShiftINode(a,_gvn.intcon(24)) ); 2177 a = _gvn.transform( new RShiftINode(a,_gvn.intcon(24)) ); 2178 push( a ); 2179 break; 2180 case Bytecodes::_i2s: 2181 a = pop(); 2182 a = _gvn.transform( new LShiftINode(a,_gvn.intcon(16)) ); 2183 a = _gvn.transform( new RShiftINode(a,_gvn.intcon(16)) ); 2184 push( a ); 2185 break; 2186 case Bytecodes::_i2c: 2187 a = pop(); 2188 push( _gvn.transform( new AndINode(a,_gvn.intcon(0xFFFF)) ) ); 2189 break; 2190 2191 case Bytecodes::_i2f: 2192 a = pop(); 2193 b = _gvn.transform( new ConvI2FNode(a) ) ; 2194 c = precision_rounding(b); 2195 push (b); 2196 break; 2197 2198 case Bytecodes::_i2d: 2199 a = pop(); 2200 b = _gvn.transform( new ConvI2DNode(a)); 2201 push_pair(b); 2202 break; 2203 2204 case Bytecodes::_iinc: // Increment local 2205 i = iter().get_index(); // Get local index 2206 set_local( i, _gvn.transform( new AddINode( _gvn.intcon(iter().get_iinc_con()), local(i) ) ) ); 2207 break; 2208 2209 // Exit points of synchronized methods must have an unlock node 2210 case Bytecodes::_return: 2211 return_current(NULL); 2212 break; 2213 2214 case Bytecodes::_ireturn: 2215 case Bytecodes::_areturn: 2216 case Bytecodes::_freturn: 2217 return_current(pop()); 2218 break; 2219 case Bytecodes::_lreturn: 2220 return_current(pop_pair()); 2221 break; 2222 case Bytecodes::_dreturn: 2223 return_current(pop_pair()); 2224 break; 2225 2226 case Bytecodes::_athrow: 2227 // null exception oop throws NULL pointer exception 2228 null_check(peek()); 2229 if (stopped()) return; 2230 // Hook the thrown exception directly to subsequent handlers. 2231 if (BailoutToInterpreterForThrows) { 2232 // Keep method interpreted from now on. 2233 uncommon_trap(Deoptimization::Reason_unhandled, 2234 Deoptimization::Action_make_not_compilable); 2235 return; 2236 } 2237 if (env()->jvmti_can_post_on_exceptions()) { 2238 // check if we must post exception events, take uncommon trap if so (with must_throw = false) 2239 uncommon_trap_if_should_post_on_exceptions(Deoptimization::Reason_unhandled, false); 2240 } 2241 // Here if either can_post_on_exceptions or should_post_on_exceptions is false 2242 add_exception_state(make_exception_state(peek())); 2243 break; 2244 2245 case Bytecodes::_goto: // fall through 2246 case Bytecodes::_goto_w: { 2247 int target_bci = (bc() == Bytecodes::_goto) ? iter().get_dest() : iter().get_far_dest(); 2248 2249 // If this is a backwards branch in the bytecodes, add Safepoint 2250 maybe_add_safepoint(target_bci); 2251 2252 // Update method data 2253 profile_taken_branch(target_bci); 2254 2255 // Merge the current control into the target basic block 2256 merge(target_bci); 2257 2258 // See if we can get some profile data and hand it off to the next block 2259 Block *target_block = block()->successor_for_bci(target_bci); 2260 if (target_block->pred_count() != 1) break; 2261 ciMethodData* methodData = method()->method_data(); 2262 if (!methodData->is_mature()) break; 2263 ciProfileData* data = methodData->bci_to_data(bci()); 2264 assert( data->is_JumpData(), "" ); 2265 int taken = ((ciJumpData*)data)->taken(); 2266 taken = method()->scale_count(taken); 2267 target_block->set_count(taken); 2268 break; 2269 } 2270 2271 case Bytecodes::_ifnull: btest = BoolTest::eq; goto handle_if_null; 2272 case Bytecodes::_ifnonnull: btest = BoolTest::ne; goto handle_if_null; 2273 handle_if_null: 2274 // If this is a backwards branch in the bytecodes, add Safepoint 2275 maybe_add_safepoint(iter().get_dest()); 2276 a = null(); 2277 b = pop(); 2278 if (!_gvn.type(b)->speculative_maybe_null() && 2279 !too_many_traps(Deoptimization::Reason_speculate_null_check)) { 2280 inc_sp(1); 2281 Node* null_ctl = top(); 2282 b = null_check_oop(b, &null_ctl, true, true, true); 2283 assert(null_ctl->is_top(), "no null control here"); 2284 dec_sp(1); 2285 } 2286 c = _gvn.transform( new CmpPNode(b, a) ); 2287 do_ifnull(btest, c); 2288 break; 2289 2290 case Bytecodes::_if_acmpeq: btest = BoolTest::eq; goto handle_if_acmp; 2291 case Bytecodes::_if_acmpne: btest = BoolTest::ne; goto handle_if_acmp; 2292 handle_if_acmp: 2293 // If this is a backwards branch in the bytecodes, add Safepoint 2294 maybe_add_safepoint(iter().get_dest()); 2295 a = pop(); 2296 b = pop(); 2297 c = _gvn.transform( new CmpPNode(b, a) ); 2298 c = optimize_cmp_with_klass(c); 2299 do_if(btest, c); 2300 break; 2301 2302 case Bytecodes::_ifeq: btest = BoolTest::eq; goto handle_ifxx; 2303 case Bytecodes::_ifne: btest = BoolTest::ne; goto handle_ifxx; 2304 case Bytecodes::_iflt: btest = BoolTest::lt; goto handle_ifxx; 2305 case Bytecodes::_ifle: btest = BoolTest::le; goto handle_ifxx; 2306 case Bytecodes::_ifgt: btest = BoolTest::gt; goto handle_ifxx; 2307 case Bytecodes::_ifge: btest = BoolTest::ge; goto handle_ifxx; 2308 handle_ifxx: 2309 // If this is a backwards branch in the bytecodes, add Safepoint 2310 maybe_add_safepoint(iter().get_dest()); 2311 a = _gvn.intcon(0); 2312 b = pop(); 2313 c = _gvn.transform( new CmpINode(b, a) ); 2314 do_if(btest, c); 2315 break; 2316 2317 case Bytecodes::_if_icmpeq: btest = BoolTest::eq; goto handle_if_icmp; 2318 case Bytecodes::_if_icmpne: btest = BoolTest::ne; goto handle_if_icmp; 2319 case Bytecodes::_if_icmplt: btest = BoolTest::lt; goto handle_if_icmp; 2320 case Bytecodes::_if_icmple: btest = BoolTest::le; goto handle_if_icmp; 2321 case Bytecodes::_if_icmpgt: btest = BoolTest::gt; goto handle_if_icmp; 2322 case Bytecodes::_if_icmpge: btest = BoolTest::ge; goto handle_if_icmp; 2323 handle_if_icmp: 2324 // If this is a backwards branch in the bytecodes, add Safepoint 2325 maybe_add_safepoint(iter().get_dest()); 2326 a = pop(); 2327 b = pop(); 2328 c = _gvn.transform( new CmpINode( b, a ) ); 2329 do_if(btest, c); 2330 break; 2331 2332 case Bytecodes::_tableswitch: 2333 do_tableswitch(); 2334 break; 2335 2336 case Bytecodes::_lookupswitch: 2337 do_lookupswitch(); 2338 break; 2339 2340 case Bytecodes::_invokestatic: 2341 case Bytecodes::_invokedynamic: 2342 case Bytecodes::_invokespecial: 2343 case Bytecodes::_invokevirtual: 2344 case Bytecodes::_invokeinterface: 2345 do_call(); 2346 break; 2347 case Bytecodes::_checkcast: 2348 do_checkcast(); 2349 break; 2350 case Bytecodes::_instanceof: 2351 do_instanceof(); 2352 break; 2353 case Bytecodes::_anewarray: 2354 do_anewarray(); 2355 break; 2356 case Bytecodes::_newarray: 2357 do_newarray((BasicType)iter().get_index()); 2358 break; 2359 case Bytecodes::_multianewarray: 2360 do_multianewarray(); 2361 break; 2362 case Bytecodes::_new: 2363 do_new(); 2364 break; 2365 2366 case Bytecodes::_jsr: 2367 case Bytecodes::_jsr_w: 2368 do_jsr(); 2369 break; 2370 2371 case Bytecodes::_ret: 2372 do_ret(); 2373 break; 2374 2375 2376 case Bytecodes::_monitorenter: 2377 do_monitor_enter(); 2378 break; 2379 2380 case Bytecodes::_monitorexit: 2381 do_monitor_exit(); 2382 break; 2383 2384 case Bytecodes::_breakpoint: 2385 // Breakpoint set concurrently to compile 2386 // %%% use an uncommon trap? 2387 C->record_failure("breakpoint in method"); 2388 return; 2389 2390 default: 2391 #ifndef PRODUCT 2392 map()->dump(99); 2393 #endif 2394 tty->print("\nUnhandled bytecode %s\n", Bytecodes::name(bc()) ); 2395 ShouldNotReachHere(); 2396 } 2397 2398 #ifndef PRODUCT 2399 IdealGraphPrinter *printer = C->printer(); 2400 if (printer && printer->should_print(1)) { 2401 char buffer[256]; 2402 sprintf(buffer, "Bytecode %d: %s", bci(), Bytecodes::name(bc())); 2403 bool old = printer->traverse_outs(); 2404 printer->set_traverse_outs(true); 2405 printer->print_method(buffer, 4); 2406 printer->set_traverse_outs(old); 2407 } 2408 #endif 2409 }