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