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 extern int explicit_null_checks_inserted, 48 explicit_null_checks_elided; 49 50 //---------------------------------array_load---------------------------------- 51 void Parse::array_load(BasicType elem_type) { 52 const Type* elem = Type::TOP; 53 Node* adr = array_addressing(elem_type, 0, false, &elem); 54 if (stopped()) return; // guaranteed null or range check 55 dec_sp(2); // Pop array and index 56 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(elem_type); 57 Node* ld = make_load(control(), adr, elem, elem_type, adr_type, MemNode::unordered); 58 push(ld); 59 } 60 61 62 //--------------------------------array_store---------------------------------- 63 void Parse::array_store(BasicType elem_type) { 64 Node* adr = array_addressing(elem_type, 1, true); 65 if (stopped()) return; // guaranteed null or range check 66 Node* val = pop(); 67 dec_sp(2); // Pop array and index 68 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(elem_type); 69 store_to_memory(control(), adr, val, elem_type, adr_type, StoreNode::release_if_reference(elem_type)); 70 } 71 72 73 //------------------------------array_addressing------------------------------- 74 // Pull array and index from the stack. Compute pointer-to-element. 75 Node* Parse::array_addressing(BasicType type, int vals, bool is_store, const Type* *result2) { 76 Node *idx = peek(0+vals); // Get from stack without popping 77 Node *ary = peek(1+vals); // in case of exception 78 79 // Null check the array base, with correct stack contents 80 ary = null_check(ary, T_ARRAY); 81 // Compile-time detect of null-exception? 82 if (stopped()) return top(); 83 84 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr(); 85 const TypeInt* sizetype = arytype->size(); 86 const Type* elemtype = arytype->elem(); 87 88 if (UseUniqueSubclasses && result2 != NULL) { 89 const Type* el = elemtype->make_ptr(); 90 if (el && el->isa_instptr()) { 91 const TypeInstPtr* toop = el->is_instptr(); 92 if (toop->klass()->as_instance_klass()->unique_concrete_subklass()) { 93 // If we load from "AbstractClass[]" we must see "ConcreteSubClass". 94 const Type* subklass = Type::get_const_type(toop->klass()); 95 elemtype = subklass->join_speculative(el); 96 } 97 } 98 } 99 100 // Check for big class initializers with all constant offsets 101 // feeding into a known-size array. 102 const TypeInt* idxtype = _gvn.type(idx)->is_int(); 103 // See if the highest idx value is less than the lowest array bound, 104 // and if the idx value cannot be negative: 105 bool need_range_check = true; 106 if (idxtype->_hi < sizetype->_lo && idxtype->_lo >= 0) { 107 need_range_check = false; 108 if (C->log() != NULL) C->log()->elem("observe that='!need_range_check'"); 109 } 110 111 ciKlass * arytype_klass = arytype->klass(); 112 if ((arytype_klass != NULL) && (!arytype_klass->is_loaded())) { 113 // Only fails for some -Xcomp runs 114 // The class is unloaded. We have to run this bytecode in the interpreter. 115 uncommon_trap(Deoptimization::Reason_unloaded, 116 Deoptimization::Action_reinterpret, 117 arytype->klass(), "!loaded array"); 118 return top(); 119 } 120 121 // Do the range check 122 if (GenerateRangeChecks && need_range_check) { 123 Node* tst; 124 if (sizetype->_hi <= 0) { 125 // The greatest array bound is negative, so we can conclude that we're 126 // compiling unreachable code, but the unsigned compare trick used below 127 // only works with non-negative lengths. Instead, hack "tst" to be zero so 128 // the uncommon_trap path will always be taken. 129 tst = _gvn.intcon(0); 130 } else { 131 // Range is constant in array-oop, so we can use the original state of mem 132 Node* len = load_array_length(ary); 133 134 // Test length vs index (standard trick using unsigned compare) 135 Node* chk = _gvn.transform( new CmpUNode(idx, len) ); 136 BoolTest::mask btest = BoolTest::lt; 137 tst = _gvn.transform( new BoolNode(chk, btest) ); 138 } 139 // Branch to failure if out of bounds 140 { BuildCutout unless(this, tst, PROB_MAX); 141 if (C->allow_range_check_smearing()) { 142 // Do not use builtin_throw, since range checks are sometimes 143 // made more stringent by an optimistic transformation. 144 // This creates "tentative" range checks at this point, 145 // which are not guaranteed to throw exceptions. 146 // See IfNode::Ideal, is_range_check, adjust_check. 147 uncommon_trap(Deoptimization::Reason_range_check, 148 Deoptimization::Action_make_not_entrant, 149 NULL, "range_check"); 150 } else { 151 // If we have already recompiled with the range-check-widening 152 // heroic optimization turned off, then we must really be throwing 153 // range check exceptions. 154 builtin_throw(Deoptimization::Reason_range_check, idx); 155 } 156 } 157 } 158 // Check for always knowing you are throwing a range-check exception 159 if (stopped()) return top(); 160 161 if (is_store) { 162 ary = shenandoah_write_barrier(ary); 163 } else { 164 ary = shenandoah_read_barrier(ary); 165 } 166 167 Node* ptr = array_element_address(ary, idx, type, sizetype); 168 169 if (result2 != NULL) *result2 = elemtype; 170 171 assert(ptr != top(), "top should go hand-in-hand with stopped"); 172 173 return ptr; 174 } 175 176 177 // returns IfNode 178 IfNode* Parse::jump_if_fork_int(Node* a, Node* b, BoolTest::mask mask) { 179 Node *cmp = _gvn.transform( new CmpINode( a, b)); // two cases: shiftcount > 32 and shiftcount <= 32 180 Node *tst = _gvn.transform( new BoolNode( cmp, mask)); 181 IfNode *iff = create_and_map_if( control(), tst, ((mask == BoolTest::eq) ? PROB_STATIC_INFREQUENT : PROB_FAIR), COUNT_UNKNOWN ); 182 return iff; 183 } 184 185 // return Region node 186 Node* Parse::jump_if_join(Node* iffalse, Node* iftrue) { 187 Node *region = new RegionNode(3); // 2 results 188 record_for_igvn(region); 189 region->init_req(1, iffalse); 190 region->init_req(2, iftrue ); 191 _gvn.set_type(region, Type::CONTROL); 192 region = _gvn.transform(region); 193 set_control (region); 194 return region; 195 } 196 197 198 //------------------------------helper for tableswitch------------------------- 199 void Parse::jump_if_true_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index) { 200 // True branch, use existing map info 201 { PreserveJVMState pjvms(this); 202 Node *iftrue = _gvn.transform( new IfTrueNode (iff) ); 203 set_control( iftrue ); 204 profile_switch_case(prof_table_index); 205 merge_new_path(dest_bci_if_true); 206 } 207 208 // False branch 209 Node *iffalse = _gvn.transform( new IfFalseNode(iff) ); 210 set_control( iffalse ); 211 } 212 213 void Parse::jump_if_false_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index) { 214 // True branch, use existing map info 215 { PreserveJVMState pjvms(this); 216 Node *iffalse = _gvn.transform( new IfFalseNode (iff) ); 217 set_control( iffalse ); 218 profile_switch_case(prof_table_index); 219 merge_new_path(dest_bci_if_true); 220 } 221 222 // False branch 223 Node *iftrue = _gvn.transform( new IfTrueNode(iff) ); 224 set_control( iftrue ); 225 } 226 227 void Parse::jump_if_always_fork(int dest_bci, int prof_table_index) { 228 // False branch, use existing map and control() 229 profile_switch_case(prof_table_index); 230 merge_new_path(dest_bci); 231 } 232 233 234 extern "C" { 235 static int jint_cmp(const void *i, const void *j) { 236 int a = *(jint *)i; 237 int b = *(jint *)j; 238 return a > b ? 1 : a < b ? -1 : 0; 239 } 240 } 241 242 243 // Default value for methodData switch indexing. Must be a negative value to avoid 244 // conflict with any legal switch index. 245 #define NullTableIndex -1 246 247 class SwitchRange : public StackObj { 248 // a range of integers coupled with a bci destination 249 jint _lo; // inclusive lower limit 250 jint _hi; // inclusive upper limit 251 int _dest; 252 int _table_index; // index into method data table 253 254 public: 255 jint lo() const { return _lo; } 256 jint hi() const { return _hi; } 257 int dest() const { return _dest; } 258 int table_index() const { return _table_index; } 259 bool is_singleton() const { return _lo == _hi; } 260 261 void setRange(jint lo, jint hi, int dest, int table_index) { 262 assert(lo <= hi, "must be a non-empty range"); 263 _lo = lo, _hi = hi; _dest = dest; _table_index = table_index; 264 } 265 bool adjoinRange(jint lo, jint hi, int dest, int table_index) { 266 assert(lo <= hi, "must be a non-empty range"); 267 if (lo == _hi+1 && dest == _dest && table_index == _table_index) { 268 _hi = hi; 269 return true; 270 } 271 return false; 272 } 273 274 void set (jint value, int dest, int table_index) { 275 setRange(value, value, dest, table_index); 276 } 277 bool adjoin(jint value, int dest, int table_index) { 278 return adjoinRange(value, value, dest, table_index); 279 } 280 281 void print() { 282 if (is_singleton()) 283 tty->print(" {%d}=>%d", lo(), dest()); 284 else if (lo() == min_jint) 285 tty->print(" {..%d}=>%d", hi(), dest()); 286 else if (hi() == max_jint) 287 tty->print(" {%d..}=>%d", lo(), dest()); 288 else 289 tty->print(" {%d..%d}=>%d", lo(), hi(), dest()); 290 } 291 }; 292 293 294 //-------------------------------do_tableswitch-------------------------------- 295 void Parse::do_tableswitch() { 296 Node* lookup = pop(); 297 298 // Get information about tableswitch 299 int default_dest = iter().get_dest_table(0); 300 int lo_index = iter().get_int_table(1); 301 int hi_index = iter().get_int_table(2); 302 int len = hi_index - lo_index + 1; 303 304 if (len < 1) { 305 // If this is a backward branch, add safepoint 306 maybe_add_safepoint(default_dest); 307 merge(default_dest); 308 return; 309 } 310 311 // generate decision tree, using trichotomy when possible 312 int rnum = len+2; 313 bool makes_backward_branch = false; 314 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum); 315 int rp = -1; 316 if (lo_index != min_jint) { 317 ranges[++rp].setRange(min_jint, lo_index-1, default_dest, NullTableIndex); 318 } 319 for (int j = 0; j < len; j++) { 320 jint match_int = lo_index+j; 321 int dest = iter().get_dest_table(j+3); 322 makes_backward_branch |= (dest <= bci()); 323 int table_index = method_data_update() ? j : NullTableIndex; 324 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index)) { 325 ranges[++rp].set(match_int, dest, table_index); 326 } 327 } 328 jint highest = lo_index+(len-1); 329 assert(ranges[rp].hi() == highest, ""); 330 if (highest != max_jint 331 && !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex)) { 332 ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex); 333 } 334 assert(rp < len+2, "not too many ranges"); 335 336 // Safepoint in case if backward branch observed 337 if( makes_backward_branch && UseLoopSafepoints ) 338 add_safepoint(); 339 340 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]); 341 } 342 343 344 //------------------------------do_lookupswitch-------------------------------- 345 void Parse::do_lookupswitch() { 346 Node *lookup = pop(); // lookup value 347 // Get information about lookupswitch 348 int default_dest = iter().get_dest_table(0); 349 int len = iter().get_int_table(1); 350 351 if (len < 1) { // If this is a backward branch, add safepoint 352 maybe_add_safepoint(default_dest); 353 merge(default_dest); 354 return; 355 } 356 357 // generate decision tree, using trichotomy when possible 358 jint* table = NEW_RESOURCE_ARRAY(jint, len*2); 359 { 360 for( int j = 0; j < len; j++ ) { 361 table[j+j+0] = iter().get_int_table(2+j+j); 362 table[j+j+1] = iter().get_dest_table(2+j+j+1); 363 } 364 qsort( table, len, 2*sizeof(table[0]), jint_cmp ); 365 } 366 367 int rnum = len*2+1; 368 bool makes_backward_branch = false; 369 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum); 370 int rp = -1; 371 for( int j = 0; j < len; j++ ) { 372 jint match_int = table[j+j+0]; 373 int dest = table[j+j+1]; 374 int next_lo = rp < 0 ? min_jint : ranges[rp].hi()+1; 375 int table_index = method_data_update() ? j : NullTableIndex; 376 makes_backward_branch |= (dest <= bci()); 377 if( match_int != next_lo ) { 378 ranges[++rp].setRange(next_lo, match_int-1, default_dest, NullTableIndex); 379 } 380 if( rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index) ) { 381 ranges[++rp].set(match_int, dest, table_index); 382 } 383 } 384 jint highest = table[2*(len-1)]; 385 assert(ranges[rp].hi() == highest, ""); 386 if( highest != max_jint 387 && !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex) ) { 388 ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex); 389 } 390 assert(rp < rnum, "not too many ranges"); 391 392 // Safepoint in case backward branch observed 393 if( makes_backward_branch && UseLoopSafepoints ) 394 add_safepoint(); 395 396 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]); 397 } 398 399 //----------------------------create_jump_tables------------------------------- 400 bool Parse::create_jump_tables(Node* key_val, SwitchRange* lo, SwitchRange* hi) { 401 // Are jumptables enabled 402 if (!UseJumpTables) return false; 403 404 // Are jumptables supported 405 if (!Matcher::has_match_rule(Op_Jump)) return false; 406 407 // Don't make jump table if profiling 408 if (method_data_update()) return false; 409 410 // Decide if a guard is needed to lop off big ranges at either (or 411 // both) end(s) of the input set. We'll call this the default target 412 // even though we can't be sure that it is the true "default". 413 414 bool needs_guard = false; 415 int default_dest; 416 int64_t total_outlier_size = 0; 417 int64_t hi_size = ((int64_t)hi->hi()) - ((int64_t)hi->lo()) + 1; 418 int64_t lo_size = ((int64_t)lo->hi()) - ((int64_t)lo->lo()) + 1; 419 420 if (lo->dest() == hi->dest()) { 421 total_outlier_size = hi_size + lo_size; 422 default_dest = lo->dest(); 423 } else if (lo_size > hi_size) { 424 total_outlier_size = lo_size; 425 default_dest = lo->dest(); 426 } else { 427 total_outlier_size = hi_size; 428 default_dest = hi->dest(); 429 } 430 431 // If a guard test will eliminate very sparse end ranges, then 432 // it is worth the cost of an extra jump. 433 if (total_outlier_size > (MaxJumpTableSparseness * 4)) { 434 needs_guard = true; 435 if (default_dest == lo->dest()) lo++; 436 if (default_dest == hi->dest()) hi--; 437 } 438 439 // Find the total number of cases and ranges 440 int64_t num_cases = ((int64_t)hi->hi()) - ((int64_t)lo->lo()) + 1; 441 int num_range = hi - lo + 1; 442 443 // Don't create table if: too large, too small, or too sparse. 444 if (num_cases < MinJumpTableSize || num_cases > MaxJumpTableSize) 445 return false; 446 if (num_cases > (MaxJumpTableSparseness * num_range)) 447 return false; 448 449 // Normalize table lookups to zero 450 int lowval = lo->lo(); 451 key_val = _gvn.transform( new SubINode(key_val, _gvn.intcon(lowval)) ); 452 453 // Generate a guard to protect against input keyvals that aren't 454 // in the switch domain. 455 if (needs_guard) { 456 Node* size = _gvn.intcon(num_cases); 457 Node* cmp = _gvn.transform( new CmpUNode(key_val, size) ); 458 Node* tst = _gvn.transform( new BoolNode(cmp, BoolTest::ge) ); 459 IfNode* iff = create_and_map_if( control(), tst, PROB_FAIR, COUNT_UNKNOWN); 460 jump_if_true_fork(iff, default_dest, NullTableIndex); 461 } 462 463 // Create an ideal node JumpTable that has projections 464 // of all possible ranges for a switch statement 465 // The key_val input must be converted to a pointer offset and scaled. 466 // Compare Parse::array_addressing above. 467 #ifdef _LP64 468 // Clean the 32-bit int into a real 64-bit offset. 469 // Otherwise, the jint value 0 might turn into an offset of 0x0800000000. 470 const TypeLong* lkeytype = TypeLong::make(CONST64(0), num_cases-1, Type::WidenMin); 471 key_val = _gvn.transform( new ConvI2LNode(key_val, lkeytype) ); 472 #endif 473 // Shift the value by wordsize so we have an index into the table, rather 474 // than a switch value 475 Node *shiftWord = _gvn.MakeConX(wordSize); 476 key_val = _gvn.transform( new MulXNode( key_val, shiftWord)); 477 478 // Create the JumpNode 479 Node* jtn = _gvn.transform( new JumpNode(control(), key_val, num_cases) ); 480 481 // These are the switch destinations hanging off the jumpnode 482 int i = 0; 483 for (SwitchRange* r = lo; r <= hi; r++) { 484 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) { 485 Node* input = _gvn.transform(new JumpProjNode(jtn, i, r->dest(), (int)(j - lowval))); 486 { 487 PreserveJVMState pjvms(this); 488 set_control(input); 489 jump_if_always_fork(r->dest(), r->table_index()); 490 } 491 } 492 } 493 assert(i == num_cases, "miscount of cases"); 494 stop_and_kill_map(); // no more uses for this JVMS 495 return true; 496 } 497 498 //----------------------------jump_switch_ranges------------------------------- 499 void Parse::jump_switch_ranges(Node* key_val, SwitchRange *lo, SwitchRange *hi, int switch_depth) { 500 Block* switch_block = block(); 501 502 if (switch_depth == 0) { 503 // Do special processing for the top-level call. 504 assert(lo->lo() == min_jint, "initial range must exhaust Type::INT"); 505 assert(hi->hi() == max_jint, "initial range must exhaust Type::INT"); 506 507 // Decrement pred-numbers for the unique set of nodes. 508 #ifdef ASSERT 509 // Ensure that the block's successors are a (duplicate-free) set. 510 int successors_counted = 0; // block occurrences in [hi..lo] 511 int unique_successors = switch_block->num_successors(); 512 for (int i = 0; i < unique_successors; i++) { 513 Block* target = switch_block->successor_at(i); 514 515 // Check that the set of successors is the same in both places. 516 int successors_found = 0; 517 for (SwitchRange* p = lo; p <= hi; p++) { 518 if (p->dest() == target->start()) successors_found++; 519 } 520 assert(successors_found > 0, "successor must be known"); 521 successors_counted += successors_found; 522 } 523 assert(successors_counted == (hi-lo)+1, "no unexpected successors"); 524 #endif 525 526 // Maybe prune the inputs, based on the type of key_val. 527 jint min_val = min_jint; 528 jint max_val = max_jint; 529 const TypeInt* ti = key_val->bottom_type()->isa_int(); 530 if (ti != NULL) { 531 min_val = ti->_lo; 532 max_val = ti->_hi; 533 assert(min_val <= max_val, "invalid int type"); 534 } 535 while (lo->hi() < min_val) lo++; 536 if (lo->lo() < min_val) lo->setRange(min_val, lo->hi(), lo->dest(), lo->table_index()); 537 while (hi->lo() > max_val) hi--; 538 if (hi->hi() > max_val) hi->setRange(hi->lo(), max_val, hi->dest(), hi->table_index()); 539 } 540 541 #ifndef PRODUCT 542 if (switch_depth == 0) { 543 _max_switch_depth = 0; 544 _est_switch_depth = log2_intptr((hi-lo+1)-1)+1; 545 } 546 #endif 547 548 assert(lo <= hi, "must be a non-empty set of ranges"); 549 if (lo == hi) { 550 jump_if_always_fork(lo->dest(), lo->table_index()); 551 } else { 552 assert(lo->hi() == (lo+1)->lo()-1, "contiguous ranges"); 553 assert(hi->lo() == (hi-1)->hi()+1, "contiguous ranges"); 554 555 if (create_jump_tables(key_val, lo, hi)) return; 556 557 int nr = hi - lo + 1; 558 559 SwitchRange* mid = lo + nr/2; 560 // if there is an easy choice, pivot at a singleton: 561 if (nr > 3 && !mid->is_singleton() && (mid-1)->is_singleton()) mid--; 562 563 assert(lo < mid && mid <= hi, "good pivot choice"); 564 assert(nr != 2 || mid == hi, "should pick higher of 2"); 565 assert(nr != 3 || mid == hi-1, "should pick middle of 3"); 566 567 Node *test_val = _gvn.intcon(mid->lo()); 568 569 if (mid->is_singleton()) { 570 IfNode *iff_ne = jump_if_fork_int(key_val, test_val, BoolTest::ne); 571 jump_if_false_fork(iff_ne, mid->dest(), mid->table_index()); 572 573 // Special Case: If there are exactly three ranges, and the high 574 // and low range each go to the same place, omit the "gt" test, 575 // since it will not discriminate anything. 576 bool eq_test_only = (hi == lo+2 && hi->dest() == lo->dest()); 577 if (eq_test_only) { 578 assert(mid == hi-1, ""); 579 } 580 581 // if there is a higher range, test for it and process it: 582 if (mid < hi && !eq_test_only) { 583 // two comparisons of same values--should enable 1 test for 2 branches 584 // Use BoolTest::le instead of BoolTest::gt 585 IfNode *iff_le = jump_if_fork_int(key_val, test_val, BoolTest::le); 586 Node *iftrue = _gvn.transform( new IfTrueNode(iff_le) ); 587 Node *iffalse = _gvn.transform( new IfFalseNode(iff_le) ); 588 { PreserveJVMState pjvms(this); 589 set_control(iffalse); 590 jump_switch_ranges(key_val, mid+1, hi, switch_depth+1); 591 } 592 set_control(iftrue); 593 } 594 595 } else { 596 // mid is a range, not a singleton, so treat mid..hi as a unit 597 IfNode *iff_ge = jump_if_fork_int(key_val, test_val, BoolTest::ge); 598 599 // if there is a higher range, test for it and process it: 600 if (mid == hi) { 601 jump_if_true_fork(iff_ge, mid->dest(), mid->table_index()); 602 } else { 603 Node *iftrue = _gvn.transform( new IfTrueNode(iff_ge) ); 604 Node *iffalse = _gvn.transform( new IfFalseNode(iff_ge) ); 605 { PreserveJVMState pjvms(this); 606 set_control(iftrue); 607 jump_switch_ranges(key_val, mid, hi, switch_depth+1); 608 } 609 set_control(iffalse); 610 } 611 } 612 613 // in any case, process the lower range 614 jump_switch_ranges(key_val, lo, mid-1, switch_depth+1); 615 } 616 617 // Decrease pred_count for each successor after all is done. 618 if (switch_depth == 0) { 619 int unique_successors = switch_block->num_successors(); 620 for (int i = 0; i < unique_successors; i++) { 621 Block* target = switch_block->successor_at(i); 622 // Throw away the pre-allocated path for each unique successor. 623 target->next_path_num(); 624 } 625 } 626 627 #ifndef PRODUCT 628 _max_switch_depth = MAX2(switch_depth, _max_switch_depth); 629 if (TraceOptoParse && Verbose && WizardMode && switch_depth == 0) { 630 SwitchRange* r; 631 int nsing = 0; 632 for( r = lo; r <= hi; r++ ) { 633 if( r->is_singleton() ) nsing++; 634 } 635 tty->print(">>> "); 636 _method->print_short_name(); 637 tty->print_cr(" switch decision tree"); 638 tty->print_cr(" %d ranges (%d singletons), max_depth=%d, est_depth=%d", 639 (int) (hi-lo+1), nsing, _max_switch_depth, _est_switch_depth); 640 if (_max_switch_depth > _est_switch_depth) { 641 tty->print_cr("******** BAD SWITCH DEPTH ********"); 642 } 643 tty->print(" "); 644 for( r = lo; r <= hi; r++ ) { 645 r->print(); 646 } 647 tty->cr(); 648 } 649 #endif 650 } 651 652 void Parse::modf() { 653 Node *f2 = pop(); 654 Node *f1 = pop(); 655 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::modf_Type(), 656 CAST_FROM_FN_PTR(address, SharedRuntime::frem), 657 "frem", NULL, //no memory effects 658 f1, f2); 659 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0)); 660 661 push(res); 662 } 663 664 void Parse::modd() { 665 Node *d2 = pop_pair(); 666 Node *d1 = pop_pair(); 667 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::Math_DD_D_Type(), 668 CAST_FROM_FN_PTR(address, SharedRuntime::drem), 669 "drem", NULL, //no memory effects 670 d1, top(), d2, top()); 671 Node* res_d = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0)); 672 673 #ifdef ASSERT 674 Node* res_top = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 1)); 675 assert(res_top == top(), "second value must be top"); 676 #endif 677 678 push_pair(res_d); 679 } 680 681 void Parse::l2f() { 682 Node* f2 = pop(); 683 Node* f1 = pop(); 684 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::l2f_Type(), 685 CAST_FROM_FN_PTR(address, SharedRuntime::l2f), 686 "l2f", NULL, //no memory effects 687 f1, f2); 688 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0)); 689 690 push(res); 691 } 692 693 void Parse::do_irem() { 694 // Must keep both values on the expression-stack during null-check 695 zero_check_int(peek()); 696 // Compile-time detect of null-exception? 697 if (stopped()) return; 698 699 Node* b = pop(); 700 Node* a = pop(); 701 702 const Type *t = _gvn.type(b); 703 if (t != Type::TOP) { 704 const TypeInt *ti = t->is_int(); 705 if (ti->is_con()) { 706 int divisor = ti->get_con(); 707 // check for positive power of 2 708 if (divisor > 0 && 709 (divisor & ~(divisor-1)) == divisor) { 710 // yes ! 711 Node *mask = _gvn.intcon((divisor - 1)); 712 // Sigh, must handle negative dividends 713 Node *zero = _gvn.intcon(0); 714 IfNode *ifff = jump_if_fork_int(a, zero, BoolTest::lt); 715 Node *iff = _gvn.transform( new IfFalseNode(ifff) ); 716 Node *ift = _gvn.transform( new IfTrueNode (ifff) ); 717 Node *reg = jump_if_join(ift, iff); 718 Node *phi = PhiNode::make(reg, NULL, TypeInt::INT); 719 // Negative path; negate/and/negate 720 Node *neg = _gvn.transform( new SubINode(zero, a) ); 721 Node *andn= _gvn.transform( new AndINode(neg, mask) ); 722 Node *negn= _gvn.transform( new SubINode(zero, andn) ); 723 phi->init_req(1, negn); 724 // Fast positive case 725 Node *andx = _gvn.transform( new AndINode(a, mask) ); 726 phi->init_req(2, andx); 727 // Push the merge 728 push( _gvn.transform(phi) ); 729 return; 730 } 731 } 732 } 733 // Default case 734 push( _gvn.transform( new ModINode(control(),a,b) ) ); 735 } 736 737 // Handle jsr and jsr_w bytecode 738 void Parse::do_jsr() { 739 assert(bc() == Bytecodes::_jsr || bc() == Bytecodes::_jsr_w, "wrong bytecode"); 740 741 // Store information about current state, tagged with new _jsr_bci 742 int return_bci = iter().next_bci(); 743 int jsr_bci = (bc() == Bytecodes::_jsr) ? iter().get_dest() : iter().get_far_dest(); 744 745 // Update method data 746 profile_taken_branch(jsr_bci); 747 748 // The way we do things now, there is only one successor block 749 // for the jsr, because the target code is cloned by ciTypeFlow. 750 Block* target = successor_for_bci(jsr_bci); 751 752 // What got pushed? 753 const Type* ret_addr = target->peek(); 754 assert(ret_addr->singleton(), "must be a constant (cloned jsr body)"); 755 756 // Effect on jsr on stack 757 push(_gvn.makecon(ret_addr)); 758 759 // Flow to the jsr. 760 merge(jsr_bci); 761 } 762 763 // Handle ret bytecode 764 void Parse::do_ret() { 765 // Find to whom we return. 766 assert(block()->num_successors() == 1, "a ret can only go one place now"); 767 Block* target = block()->successor_at(0); 768 assert(!target->is_ready(), "our arrival must be expected"); 769 profile_ret(target->flow()->start()); 770 int pnum = target->next_path_num(); 771 merge_common(target, pnum); 772 } 773 774 static bool has_injected_profile(BoolTest::mask btest, Node* test, int& taken, int& not_taken) { 775 if (btest != BoolTest::eq && btest != BoolTest::ne) { 776 // Only ::eq and ::ne are supported for profile injection. 777 return false; 778 } 779 if (test->is_Cmp() && 780 test->in(1)->Opcode() == Op_ProfileBoolean) { 781 ProfileBooleanNode* profile = (ProfileBooleanNode*)test->in(1); 782 int false_cnt = profile->false_count(); 783 int true_cnt = profile->true_count(); 784 785 // Counts matching depends on the actual test operation (::eq or ::ne). 786 // No need to scale the counts because profile injection was designed 787 // to feed exact counts into VM. 788 taken = (btest == BoolTest::eq) ? false_cnt : true_cnt; 789 not_taken = (btest == BoolTest::eq) ? true_cnt : false_cnt; 790 791 profile->consume(); 792 return true; 793 } 794 return false; 795 } 796 //--------------------------dynamic_branch_prediction-------------------------- 797 // Try to gather dynamic branch prediction behavior. Return a probability 798 // of the branch being taken and set the "cnt" field. Returns a -1.0 799 // if we need to use static prediction for some reason. 800 float Parse::dynamic_branch_prediction(float &cnt, BoolTest::mask btest, Node* test) { 801 ResourceMark rm; 802 803 cnt = COUNT_UNKNOWN; 804 805 int taken = 0; 806 int not_taken = 0; 807 808 bool use_mdo = !has_injected_profile(btest, test, taken, not_taken); 809 810 if (use_mdo) { 811 // Use MethodData information if it is available 812 // FIXME: free the ProfileData structure 813 ciMethodData* methodData = method()->method_data(); 814 if (!methodData->is_mature()) return PROB_UNKNOWN; 815 ciProfileData* data = methodData->bci_to_data(bci()); 816 if (!data->is_JumpData()) return PROB_UNKNOWN; 817 818 // get taken and not taken values 819 taken = data->as_JumpData()->taken(); 820 not_taken = 0; 821 if (data->is_BranchData()) { 822 not_taken = data->as_BranchData()->not_taken(); 823 } 824 825 // scale the counts to be commensurate with invocation counts: 826 taken = method()->scale_count(taken); 827 not_taken = method()->scale_count(not_taken); 828 } 829 830 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful. 831 // We also check that individual counters are positive first, otherwise the sum can become positive. 832 if (taken < 0 || not_taken < 0 || taken + not_taken < 40) { 833 if (C->log() != NULL) { 834 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d'", iter().get_dest(), taken, not_taken); 835 } 836 return PROB_UNKNOWN; 837 } 838 839 // Compute frequency that we arrive here 840 float sum = taken + not_taken; 841 // Adjust, if this block is a cloned private block but the 842 // Jump counts are shared. Taken the private counts for 843 // just this path instead of the shared counts. 844 if( block()->count() > 0 ) 845 sum = block()->count(); 846 cnt = sum / FreqCountInvocations; 847 848 // Pin probability to sane limits 849 float prob; 850 if( !taken ) 851 prob = (0+PROB_MIN) / 2; 852 else if( !not_taken ) 853 prob = (1+PROB_MAX) / 2; 854 else { // Compute probability of true path 855 prob = (float)taken / (float)(taken + not_taken); 856 if (prob > PROB_MAX) prob = PROB_MAX; 857 if (prob < PROB_MIN) prob = PROB_MIN; 858 } 859 860 assert((cnt > 0.0f) && (prob > 0.0f), 861 "Bad frequency assignment in if"); 862 863 if (C->log() != NULL) { 864 const char* prob_str = NULL; 865 if (prob >= PROB_MAX) prob_str = (prob == PROB_MAX) ? "max" : "always"; 866 if (prob <= PROB_MIN) prob_str = (prob == PROB_MIN) ? "min" : "never"; 867 char prob_str_buf[30]; 868 if (prob_str == NULL) { 869 sprintf(prob_str_buf, "%g", prob); 870 prob_str = prob_str_buf; 871 } 872 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d' cnt='%f' prob='%s'", 873 iter().get_dest(), taken, not_taken, cnt, prob_str); 874 } 875 return prob; 876 } 877 878 //-----------------------------branch_prediction------------------------------- 879 float Parse::branch_prediction(float& cnt, 880 BoolTest::mask btest, 881 int target_bci, 882 Node* test) { 883 float prob = dynamic_branch_prediction(cnt, btest, test); 884 // If prob is unknown, switch to static prediction 885 if (prob != PROB_UNKNOWN) return prob; 886 887 prob = PROB_FAIR; // Set default value 888 if (btest == BoolTest::eq) // Exactly equal test? 889 prob = PROB_STATIC_INFREQUENT; // Assume its relatively infrequent 890 else if (btest == BoolTest::ne) 891 prob = PROB_STATIC_FREQUENT; // Assume its relatively frequent 892 893 // If this is a conditional test guarding a backwards branch, 894 // assume its a loop-back edge. Make it a likely taken branch. 895 if (target_bci < bci()) { 896 if (is_osr_parse()) { // Could be a hot OSR'd loop; force deopt 897 // Since it's an OSR, we probably have profile data, but since 898 // branch_prediction returned PROB_UNKNOWN, the counts are too small. 899 // Let's make a special check here for completely zero counts. 900 ciMethodData* methodData = method()->method_data(); 901 if (!methodData->is_empty()) { 902 ciProfileData* data = methodData->bci_to_data(bci()); 903 // Only stop for truly zero counts, which mean an unknown part 904 // of the OSR-ed method, and we want to deopt to gather more stats. 905 // If you have ANY counts, then this loop is simply 'cold' relative 906 // to the OSR loop. 907 if (data->as_BranchData()->taken() + 908 data->as_BranchData()->not_taken() == 0 ) { 909 // This is the only way to return PROB_UNKNOWN: 910 return PROB_UNKNOWN; 911 } 912 } 913 } 914 prob = PROB_STATIC_FREQUENT; // Likely to take backwards branch 915 } 916 917 assert(prob != PROB_UNKNOWN, "must have some guess at this point"); 918 return prob; 919 } 920 921 // The magic constants are chosen so as to match the output of 922 // branch_prediction() when the profile reports a zero taken count. 923 // It is important to distinguish zero counts unambiguously, because 924 // some branches (e.g., _213_javac.Assembler.eliminate) validly produce 925 // very small but nonzero probabilities, which if confused with zero 926 // counts would keep the program recompiling indefinitely. 927 bool Parse::seems_never_taken(float prob) const { 928 return prob < PROB_MIN; 929 } 930 931 // True if the comparison seems to be the kind that will not change its 932 // statistics from true to false. See comments in adjust_map_after_if. 933 // This question is only asked along paths which are already 934 // classifed as untaken (by seems_never_taken), so really, 935 // if a path is never taken, its controlling comparison is 936 // already acting in a stable fashion. If the comparison 937 // seems stable, we will put an expensive uncommon trap 938 // on the untaken path. 939 bool Parse::seems_stable_comparison() const { 940 if (C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if)) { 941 return false; 942 } 943 return true; 944 } 945 946 //-------------------------------repush_if_args-------------------------------- 947 // Push arguments of an "if" bytecode back onto the stack by adjusting _sp. 948 inline int Parse::repush_if_args() { 949 #ifndef PRODUCT 950 if (PrintOpto && WizardMode) { 951 tty->print("defending against excessive implicit null exceptions on %s @%d in ", 952 Bytecodes::name(iter().cur_bc()), iter().cur_bci()); 953 method()->print_name(); tty->cr(); 954 } 955 #endif 956 int bc_depth = - Bytecodes::depth(iter().cur_bc()); 957 assert(bc_depth == 1 || bc_depth == 2, "only two kinds of branches"); 958 DEBUG_ONLY(sync_jvms()); // argument(n) requires a synced jvms 959 assert(argument(0) != NULL, "must exist"); 960 assert(bc_depth == 1 || argument(1) != NULL, "two must exist"); 961 inc_sp(bc_depth); 962 return bc_depth; 963 } 964 965 //----------------------------------do_ifnull---------------------------------- 966 void Parse::do_ifnull(BoolTest::mask btest, Node *c) { 967 int target_bci = iter().get_dest(); 968 969 Block* branch_block = successor_for_bci(target_bci); 970 Block* next_block = successor_for_bci(iter().next_bci()); 971 972 float cnt; 973 float prob = branch_prediction(cnt, btest, target_bci, c); 974 if (prob == PROB_UNKNOWN) { 975 // (An earlier version of do_ifnull omitted this trap for OSR methods.) 976 #ifndef PRODUCT 977 if (PrintOpto && Verbose) 978 tty->print_cr("Never-taken edge stops compilation at bci %d",bci()); 979 #endif 980 repush_if_args(); // to gather stats on loop 981 // We need to mark this branch as taken so that if we recompile we will 982 // see that it is possible. In the tiered system the interpreter doesn't 983 // do profiling and by the time we get to the lower tier from the interpreter 984 // the path may be cold again. Make sure it doesn't look untaken 985 profile_taken_branch(target_bci, !ProfileInterpreter); 986 uncommon_trap(Deoptimization::Reason_unreached, 987 Deoptimization::Action_reinterpret, 988 NULL, "cold"); 989 if (C->eliminate_boxing()) { 990 // Mark the successor blocks as parsed 991 branch_block->next_path_num(); 992 next_block->next_path_num(); 993 } 994 return; 995 } 996 997 explicit_null_checks_inserted++; 998 999 // Generate real control flow 1000 Node *tst = _gvn.transform( new BoolNode( c, btest ) ); 1001 1002 // Sanity check the probability value 1003 assert(prob > 0.0f,"Bad probability in Parser"); 1004 // Need xform to put node in hash table 1005 IfNode *iff = create_and_xform_if( control(), tst, prob, cnt ); 1006 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser"); 1007 // True branch 1008 { PreserveJVMState pjvms(this); 1009 Node* iftrue = _gvn.transform( new IfTrueNode (iff) ); 1010 set_control(iftrue); 1011 1012 if (stopped()) { // Path is dead? 1013 explicit_null_checks_elided++; 1014 if (C->eliminate_boxing()) { 1015 // Mark the successor block as parsed 1016 branch_block->next_path_num(); 1017 } 1018 } else { // Path is live. 1019 // Update method data 1020 profile_taken_branch(target_bci); 1021 adjust_map_after_if(btest, c, prob, branch_block, next_block); 1022 if (!stopped()) { 1023 merge(target_bci); 1024 } 1025 } 1026 } 1027 1028 // False branch 1029 Node* iffalse = _gvn.transform( new IfFalseNode(iff) ); 1030 set_control(iffalse); 1031 1032 if (stopped()) { // Path is dead? 1033 explicit_null_checks_elided++; 1034 if (C->eliminate_boxing()) { 1035 // Mark the successor block as parsed 1036 next_block->next_path_num(); 1037 } 1038 } else { // Path is live. 1039 // Update method data 1040 profile_not_taken_branch(); 1041 adjust_map_after_if(BoolTest(btest).negate(), c, 1.0-prob, 1042 next_block, branch_block); 1043 } 1044 } 1045 1046 //------------------------------------do_if------------------------------------ 1047 void Parse::do_if(BoolTest::mask btest, Node* c) { 1048 int target_bci = iter().get_dest(); 1049 1050 Block* branch_block = successor_for_bci(target_bci); 1051 Block* next_block = successor_for_bci(iter().next_bci()); 1052 1053 float cnt; 1054 float prob = branch_prediction(cnt, btest, target_bci, c); 1055 float untaken_prob = 1.0 - prob; 1056 1057 if (prob == PROB_UNKNOWN) { 1058 #ifndef PRODUCT 1059 if (PrintOpto && Verbose) 1060 tty->print_cr("Never-taken edge stops compilation at bci %d",bci()); 1061 #endif 1062 repush_if_args(); // to gather stats on loop 1063 // We need to mark this branch as taken so that if we recompile we will 1064 // see that it is possible. In the tiered system the interpreter doesn't 1065 // do profiling and by the time we get to the lower tier from the interpreter 1066 // the path may be cold again. Make sure it doesn't look untaken 1067 profile_taken_branch(target_bci, !ProfileInterpreter); 1068 uncommon_trap(Deoptimization::Reason_unreached, 1069 Deoptimization::Action_reinterpret, 1070 NULL, "cold"); 1071 if (C->eliminate_boxing()) { 1072 // Mark the successor blocks as parsed 1073 branch_block->next_path_num(); 1074 next_block->next_path_num(); 1075 } 1076 return; 1077 } 1078 1079 // Sanity check the probability value 1080 assert(0.0f < prob && prob < 1.0f,"Bad probability in Parser"); 1081 1082 bool taken_if_true = true; 1083 // Convert BoolTest to canonical form: 1084 if (!BoolTest(btest).is_canonical()) { 1085 btest = BoolTest(btest).negate(); 1086 taken_if_true = false; 1087 // prob is NOT updated here; it remains the probability of the taken 1088 // path (as opposed to the prob of the path guarded by an 'IfTrueNode'). 1089 } 1090 assert(btest != BoolTest::eq, "!= is the only canonical exact test"); 1091 1092 Node* tst0 = new BoolNode(c, btest); 1093 Node* tst = _gvn.transform(tst0); 1094 BoolTest::mask taken_btest = BoolTest::illegal; 1095 BoolTest::mask untaken_btest = BoolTest::illegal; 1096 1097 if (tst->is_Bool()) { 1098 // Refresh c from the transformed bool node, since it may be 1099 // simpler than the original c. Also re-canonicalize btest. 1100 // This wins when (Bool ne (Conv2B p) 0) => (Bool ne (CmpP p NULL)). 1101 // That can arise from statements like: if (x instanceof C) ... 1102 if (tst != tst0) { 1103 // Canonicalize one more time since transform can change it. 1104 btest = tst->as_Bool()->_test._test; 1105 if (!BoolTest(btest).is_canonical()) { 1106 // Reverse edges one more time... 1107 tst = _gvn.transform( tst->as_Bool()->negate(&_gvn) ); 1108 btest = tst->as_Bool()->_test._test; 1109 assert(BoolTest(btest).is_canonical(), "sanity"); 1110 taken_if_true = !taken_if_true; 1111 } 1112 c = tst->in(1); 1113 } 1114 BoolTest::mask neg_btest = BoolTest(btest).negate(); 1115 taken_btest = taken_if_true ? btest : neg_btest; 1116 untaken_btest = taken_if_true ? neg_btest : btest; 1117 } 1118 1119 // Generate real control flow 1120 float true_prob = (taken_if_true ? prob : untaken_prob); 1121 IfNode* iff = create_and_map_if(control(), tst, true_prob, cnt); 1122 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser"); 1123 Node* taken_branch = new IfTrueNode(iff); 1124 Node* untaken_branch = new IfFalseNode(iff); 1125 if (!taken_if_true) { // Finish conversion to canonical form 1126 Node* tmp = taken_branch; 1127 taken_branch = untaken_branch; 1128 untaken_branch = tmp; 1129 } 1130 1131 // Branch is taken: 1132 { PreserveJVMState pjvms(this); 1133 taken_branch = _gvn.transform(taken_branch); 1134 set_control(taken_branch); 1135 1136 if (stopped()) { 1137 if (C->eliminate_boxing()) { 1138 // Mark the successor block as parsed 1139 branch_block->next_path_num(); 1140 } 1141 } else { 1142 // Update method data 1143 profile_taken_branch(target_bci); 1144 adjust_map_after_if(taken_btest, c, prob, branch_block, next_block); 1145 if (!stopped()) { 1146 merge(target_bci); 1147 } 1148 } 1149 } 1150 1151 untaken_branch = _gvn.transform(untaken_branch); 1152 set_control(untaken_branch); 1153 1154 // Branch not taken. 1155 if (stopped()) { 1156 if (C->eliminate_boxing()) { 1157 // Mark the successor block as parsed 1158 next_block->next_path_num(); 1159 } 1160 } else { 1161 // Update method data 1162 profile_not_taken_branch(); 1163 adjust_map_after_if(untaken_btest, c, untaken_prob, 1164 next_block, branch_block); 1165 } 1166 } 1167 1168 bool Parse::path_is_suitable_for_uncommon_trap(float prob) const { 1169 // Don't want to speculate on uncommon traps when running with -Xcomp 1170 if (!UseInterpreter) { 1171 return false; 1172 } 1173 return (seems_never_taken(prob) && seems_stable_comparison()); 1174 } 1175 1176 //----------------------------adjust_map_after_if------------------------------ 1177 // Adjust the JVM state to reflect the result of taking this path. 1178 // Basically, it means inspecting the CmpNode controlling this 1179 // branch, seeing how it constrains a tested value, and then 1180 // deciding if it's worth our while to encode this constraint 1181 // as graph nodes in the current abstract interpretation map. 1182 void Parse::adjust_map_after_if(BoolTest::mask btest, Node* c, float prob, 1183 Block* path, Block* other_path) { 1184 if (stopped() || !c->is_Cmp() || btest == BoolTest::illegal) 1185 return; // nothing to do 1186 1187 bool is_fallthrough = (path == successor_for_bci(iter().next_bci())); 1188 1189 if (path_is_suitable_for_uncommon_trap(prob)) { 1190 repush_if_args(); 1191 uncommon_trap(Deoptimization::Reason_unstable_if, 1192 Deoptimization::Action_reinterpret, 1193 NULL, 1194 (is_fallthrough ? "taken always" : "taken never")); 1195 return; 1196 } 1197 1198 Node* val = c->in(1); 1199 Node* con = c->in(2); 1200 const Type* tcon = _gvn.type(con); 1201 const Type* tval = _gvn.type(val); 1202 bool have_con = tcon->singleton(); 1203 if (tval->singleton()) { 1204 if (!have_con) { 1205 // Swap, so constant is in con. 1206 con = val; 1207 tcon = tval; 1208 val = c->in(2); 1209 tval = _gvn.type(val); 1210 btest = BoolTest(btest).commute(); 1211 have_con = true; 1212 } else { 1213 // Do we have two constants? Then leave well enough alone. 1214 have_con = false; 1215 } 1216 } 1217 if (!have_con) // remaining adjustments need a con 1218 return; 1219 1220 sharpen_type_after_if(btest, con, tcon, val, tval); 1221 } 1222 1223 1224 static Node* extract_obj_from_klass_load(PhaseGVN* gvn, Node* n) { 1225 Node* ldk; 1226 if (n->is_DecodeNKlass()) { 1227 if (n->in(1)->Opcode() != Op_LoadNKlass) { 1228 return NULL; 1229 } else { 1230 ldk = n->in(1); 1231 } 1232 } else if (n->Opcode() != Op_LoadKlass) { 1233 return NULL; 1234 } else { 1235 ldk = n; 1236 } 1237 assert(ldk != NULL && ldk->is_Load(), "should have found a LoadKlass or LoadNKlass node"); 1238 1239 Node* adr = ldk->in(MemNode::Address); 1240 intptr_t off = 0; 1241 Node* obj = AddPNode::Ideal_base_and_offset(adr, gvn, off); 1242 if (obj == NULL || off != oopDesc::klass_offset_in_bytes()) // loading oopDesc::_klass? 1243 return NULL; 1244 const TypePtr* tp = gvn->type(obj)->is_ptr(); 1245 if (tp == NULL || !(tp->isa_instptr() || tp->isa_aryptr())) // is obj a Java object ptr? 1246 return NULL; 1247 1248 return obj; 1249 } 1250 1251 void Parse::sharpen_type_after_if(BoolTest::mask btest, 1252 Node* con, const Type* tcon, 1253 Node* val, const Type* tval) { 1254 // Look for opportunities to sharpen the type of a node 1255 // whose klass is compared with a constant klass. 1256 if (btest == BoolTest::eq && tcon->isa_klassptr()) { 1257 Node* obj = extract_obj_from_klass_load(&_gvn, val); 1258 const TypeOopPtr* con_type = tcon->isa_klassptr()->as_instance_type(); 1259 if (obj != NULL && (con_type->isa_instptr() || con_type->isa_aryptr())) { 1260 // Found: 1261 // Bool(CmpP(LoadKlass(obj._klass), ConP(Foo.klass)), [eq]) 1262 // or the narrowOop equivalent. 1263 const Type* obj_type = _gvn.type(obj); 1264 const TypeOopPtr* tboth = obj_type->join_speculative(con_type)->isa_oopptr(); 1265 if (tboth != NULL && tboth->klass_is_exact() && tboth != obj_type && 1266 tboth->higher_equal(obj_type)) { 1267 // obj has to be of the exact type Foo if the CmpP succeeds. 1268 int obj_in_map = map()->find_edge(obj); 1269 JVMState* jvms = this->jvms(); 1270 if (obj_in_map >= 0 && 1271 (jvms->is_loc(obj_in_map) || jvms->is_stk(obj_in_map))) { 1272 TypeNode* ccast = new CheckCastPPNode(control(), obj, tboth); 1273 const Type* tcc = ccast->as_Type()->type(); 1274 assert(tcc != obj_type && tcc->higher_equal(obj_type), "must improve"); 1275 // Delay transform() call to allow recovery of pre-cast value 1276 // at the control merge. 1277 _gvn.set_type_bottom(ccast); 1278 record_for_igvn(ccast); 1279 // Here's the payoff. 1280 replace_in_map(obj, ccast); 1281 } 1282 } 1283 } 1284 } 1285 1286 int val_in_map = map()->find_edge(val); 1287 if (val_in_map < 0) return; // replace_in_map would be useless 1288 { 1289 JVMState* jvms = this->jvms(); 1290 if (!(jvms->is_loc(val_in_map) || 1291 jvms->is_stk(val_in_map))) 1292 return; // again, it would be useless 1293 } 1294 1295 // Check for a comparison to a constant, and "know" that the compared 1296 // value is constrained on this path. 1297 assert(tcon->singleton(), ""); 1298 ConstraintCastNode* ccast = NULL; 1299 Node* cast = NULL; 1300 1301 switch (btest) { 1302 case BoolTest::eq: // Constant test? 1303 { 1304 const Type* tboth = tcon->join_speculative(tval); 1305 if (tboth == tval) break; // Nothing to gain. 1306 if (tcon->isa_int()) { 1307 ccast = new CastIINode(val, tboth); 1308 } else if (tcon == TypePtr::NULL_PTR) { 1309 // Cast to null, but keep the pointer identity temporarily live. 1310 ccast = new CastPPNode(val, tboth); 1311 } else { 1312 const TypeF* tf = tcon->isa_float_constant(); 1313 const TypeD* td = tcon->isa_double_constant(); 1314 // Exclude tests vs float/double 0 as these could be 1315 // either +0 or -0. Just because you are equal to +0 1316 // doesn't mean you ARE +0! 1317 // Note, following code also replaces Long and Oop values. 1318 if ((!tf || tf->_f != 0.0) && 1319 (!td || td->_d != 0.0)) 1320 cast = con; // Replace non-constant val by con. 1321 } 1322 } 1323 break; 1324 1325 case BoolTest::ne: 1326 if (tcon == TypePtr::NULL_PTR) { 1327 cast = cast_not_null(val, false); 1328 } 1329 break; 1330 1331 default: 1332 // (At this point we could record int range types with CastII.) 1333 break; 1334 } 1335 1336 if (ccast != NULL) { 1337 const Type* tcc = ccast->as_Type()->type(); 1338 assert(tcc != tval && tcc->higher_equal(tval), "must improve"); 1339 // Delay transform() call to allow recovery of pre-cast value 1340 // at the control merge. 1341 ccast->set_req(0, control()); 1342 _gvn.set_type_bottom(ccast); 1343 record_for_igvn(ccast); 1344 cast = ccast; 1345 } 1346 1347 if (cast != NULL) { // Here's the payoff. 1348 replace_in_map(val, cast); 1349 } 1350 } 1351 1352 /** 1353 * Use speculative type to optimize CmpP node: if comparison is 1354 * against the low level class, cast the object to the speculative 1355 * type if any. CmpP should then go away. 1356 * 1357 * @param c expected CmpP node 1358 * @return result of CmpP on object casted to speculative type 1359 * 1360 */ 1361 Node* Parse::optimize_cmp_with_klass(Node* c) { 1362 // If this is transformed by the _gvn to a comparison with the low 1363 // level klass then we may be able to use speculation 1364 if (c->Opcode() == Op_CmpP && 1365 (c->in(1)->Opcode() == Op_LoadKlass || c->in(1)->Opcode() == Op_DecodeNKlass) && 1366 c->in(2)->is_Con()) { 1367 Node* load_klass = NULL; 1368 Node* decode = NULL; 1369 if (c->in(1)->Opcode() == Op_DecodeNKlass) { 1370 decode = c->in(1); 1371 load_klass = c->in(1)->in(1); 1372 } else { 1373 load_klass = c->in(1); 1374 } 1375 if (load_klass->in(2)->is_AddP()) { 1376 Node* addp = load_klass->in(2); 1377 Node* obj = addp->in(AddPNode::Address); 1378 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 1379 if (obj_type->speculative_type_not_null() != NULL) { 1380 ciKlass* k = obj_type->speculative_type(); 1381 inc_sp(2); 1382 obj = maybe_cast_profiled_obj(obj, k); 1383 dec_sp(2); 1384 // Make the CmpP use the casted obj 1385 addp = basic_plus_adr(obj, addp->in(AddPNode::Offset)); 1386 load_klass = load_klass->clone(); 1387 load_klass->set_req(2, addp); 1388 load_klass = _gvn.transform(load_klass); 1389 if (decode != NULL) { 1390 decode = decode->clone(); 1391 decode->set_req(1, load_klass); 1392 load_klass = _gvn.transform(decode); 1393 } 1394 c = c->clone(); 1395 c->set_req(1, load_klass); 1396 c = _gvn.transform(c); 1397 } 1398 } 1399 } 1400 return c; 1401 } 1402 1403 //------------------------------do_one_bytecode-------------------------------- 1404 // Parse this bytecode, and alter the Parsers JVM->Node mapping 1405 void Parse::do_one_bytecode() { 1406 Node *a, *b, *c, *d; // Handy temps 1407 BoolTest::mask btest; 1408 int i; 1409 1410 assert(!has_exceptions(), "bytecode entry state must be clear of throws"); 1411 1412 if (C->check_node_count(NodeLimitFudgeFactor * 5, 1413 "out of nodes parsing method")) { 1414 return; 1415 } 1416 1417 #ifdef ASSERT 1418 // for setting breakpoints 1419 if (TraceOptoParse) { 1420 tty->print(" @"); 1421 dump_bci(bci()); 1422 tty->cr(); 1423 } 1424 #endif 1425 1426 switch (bc()) { 1427 case Bytecodes::_nop: 1428 // do nothing 1429 break; 1430 case Bytecodes::_lconst_0: 1431 push_pair(longcon(0)); 1432 break; 1433 1434 case Bytecodes::_lconst_1: 1435 push_pair(longcon(1)); 1436 break; 1437 1438 case Bytecodes::_fconst_0: 1439 push(zerocon(T_FLOAT)); 1440 break; 1441 1442 case Bytecodes::_fconst_1: 1443 push(makecon(TypeF::ONE)); 1444 break; 1445 1446 case Bytecodes::_fconst_2: 1447 push(makecon(TypeF::make(2.0f))); 1448 break; 1449 1450 case Bytecodes::_dconst_0: 1451 push_pair(zerocon(T_DOUBLE)); 1452 break; 1453 1454 case Bytecodes::_dconst_1: 1455 push_pair(makecon(TypeD::ONE)); 1456 break; 1457 1458 case Bytecodes::_iconst_m1:push(intcon(-1)); break; 1459 case Bytecodes::_iconst_0: push(intcon( 0)); break; 1460 case Bytecodes::_iconst_1: push(intcon( 1)); break; 1461 case Bytecodes::_iconst_2: push(intcon( 2)); break; 1462 case Bytecodes::_iconst_3: push(intcon( 3)); break; 1463 case Bytecodes::_iconst_4: push(intcon( 4)); break; 1464 case Bytecodes::_iconst_5: push(intcon( 5)); break; 1465 case Bytecodes::_bipush: push(intcon(iter().get_constant_u1())); break; 1466 case Bytecodes::_sipush: push(intcon(iter().get_constant_u2())); break; 1467 case Bytecodes::_aconst_null: push(null()); break; 1468 case Bytecodes::_ldc: 1469 case Bytecodes::_ldc_w: 1470 case Bytecodes::_ldc2_w: 1471 // If the constant is unresolved, run this BC once in the interpreter. 1472 { 1473 ciConstant constant = iter().get_constant(); 1474 if (constant.basic_type() == T_OBJECT && 1475 !constant.as_object()->is_loaded()) { 1476 int index = iter().get_constant_pool_index(); 1477 constantTag tag = iter().get_constant_pool_tag(index); 1478 uncommon_trap(Deoptimization::make_trap_request 1479 (Deoptimization::Reason_unloaded, 1480 Deoptimization::Action_reinterpret, 1481 index), 1482 NULL, tag.internal_name()); 1483 break; 1484 } 1485 assert(constant.basic_type() != T_OBJECT || constant.as_object()->is_instance(), 1486 "must be java_mirror of klass"); 1487 const Type* con_type = Type::make_from_constant(constant); 1488 if (con_type != NULL) { 1489 push_node(con_type->basic_type(), makecon(con_type)); 1490 } 1491 } 1492 1493 break; 1494 1495 case Bytecodes::_aload_0: 1496 push( local(0) ); 1497 break; 1498 case Bytecodes::_aload_1: 1499 push( local(1) ); 1500 break; 1501 case Bytecodes::_aload_2: 1502 push( local(2) ); 1503 break; 1504 case Bytecodes::_aload_3: 1505 push( local(3) ); 1506 break; 1507 case Bytecodes::_aload: 1508 push( local(iter().get_index()) ); 1509 break; 1510 1511 case Bytecodes::_fload_0: 1512 case Bytecodes::_iload_0: 1513 push( local(0) ); 1514 break; 1515 case Bytecodes::_fload_1: 1516 case Bytecodes::_iload_1: 1517 push( local(1) ); 1518 break; 1519 case Bytecodes::_fload_2: 1520 case Bytecodes::_iload_2: 1521 push( local(2) ); 1522 break; 1523 case Bytecodes::_fload_3: 1524 case Bytecodes::_iload_3: 1525 push( local(3) ); 1526 break; 1527 case Bytecodes::_fload: 1528 case Bytecodes::_iload: 1529 push( local(iter().get_index()) ); 1530 break; 1531 case Bytecodes::_lload_0: 1532 push_pair_local( 0 ); 1533 break; 1534 case Bytecodes::_lload_1: 1535 push_pair_local( 1 ); 1536 break; 1537 case Bytecodes::_lload_2: 1538 push_pair_local( 2 ); 1539 break; 1540 case Bytecodes::_lload_3: 1541 push_pair_local( 3 ); 1542 break; 1543 case Bytecodes::_lload: 1544 push_pair_local( iter().get_index() ); 1545 break; 1546 1547 case Bytecodes::_dload_0: 1548 push_pair_local(0); 1549 break; 1550 case Bytecodes::_dload_1: 1551 push_pair_local(1); 1552 break; 1553 case Bytecodes::_dload_2: 1554 push_pair_local(2); 1555 break; 1556 case Bytecodes::_dload_3: 1557 push_pair_local(3); 1558 break; 1559 case Bytecodes::_dload: 1560 push_pair_local(iter().get_index()); 1561 break; 1562 case Bytecodes::_fstore_0: 1563 case Bytecodes::_istore_0: 1564 case Bytecodes::_astore_0: 1565 set_local( 0, pop() ); 1566 break; 1567 case Bytecodes::_fstore_1: 1568 case Bytecodes::_istore_1: 1569 case Bytecodes::_astore_1: 1570 set_local( 1, pop() ); 1571 break; 1572 case Bytecodes::_fstore_2: 1573 case Bytecodes::_istore_2: 1574 case Bytecodes::_astore_2: 1575 set_local( 2, pop() ); 1576 break; 1577 case Bytecodes::_fstore_3: 1578 case Bytecodes::_istore_3: 1579 case Bytecodes::_astore_3: 1580 set_local( 3, pop() ); 1581 break; 1582 case Bytecodes::_fstore: 1583 case Bytecodes::_istore: 1584 case Bytecodes::_astore: 1585 set_local( iter().get_index(), pop() ); 1586 break; 1587 // long stores 1588 case Bytecodes::_lstore_0: 1589 set_pair_local( 0, pop_pair() ); 1590 break; 1591 case Bytecodes::_lstore_1: 1592 set_pair_local( 1, pop_pair() ); 1593 break; 1594 case Bytecodes::_lstore_2: 1595 set_pair_local( 2, pop_pair() ); 1596 break; 1597 case Bytecodes::_lstore_3: 1598 set_pair_local( 3, pop_pair() ); 1599 break; 1600 case Bytecodes::_lstore: 1601 set_pair_local( iter().get_index(), pop_pair() ); 1602 break; 1603 1604 // double stores 1605 case Bytecodes::_dstore_0: 1606 set_pair_local( 0, dstore_rounding(pop_pair()) ); 1607 break; 1608 case Bytecodes::_dstore_1: 1609 set_pair_local( 1, dstore_rounding(pop_pair()) ); 1610 break; 1611 case Bytecodes::_dstore_2: 1612 set_pair_local( 2, dstore_rounding(pop_pair()) ); 1613 break; 1614 case Bytecodes::_dstore_3: 1615 set_pair_local( 3, dstore_rounding(pop_pair()) ); 1616 break; 1617 case Bytecodes::_dstore: 1618 set_pair_local( iter().get_index(), dstore_rounding(pop_pair()) ); 1619 break; 1620 1621 case Bytecodes::_pop: dec_sp(1); break; 1622 case Bytecodes::_pop2: dec_sp(2); break; 1623 case Bytecodes::_swap: 1624 a = pop(); 1625 b = pop(); 1626 push(a); 1627 push(b); 1628 break; 1629 case Bytecodes::_dup: 1630 a = pop(); 1631 push(a); 1632 push(a); 1633 break; 1634 case Bytecodes::_dup_x1: 1635 a = pop(); 1636 b = pop(); 1637 push( a ); 1638 push( b ); 1639 push( a ); 1640 break; 1641 case Bytecodes::_dup_x2: 1642 a = pop(); 1643 b = pop(); 1644 c = pop(); 1645 push( a ); 1646 push( c ); 1647 push( b ); 1648 push( a ); 1649 break; 1650 case Bytecodes::_dup2: 1651 a = pop(); 1652 b = pop(); 1653 push( b ); 1654 push( a ); 1655 push( b ); 1656 push( a ); 1657 break; 1658 1659 case Bytecodes::_dup2_x1: 1660 // before: .. c, b, a 1661 // after: .. b, a, c, b, a 1662 // not tested 1663 a = pop(); 1664 b = pop(); 1665 c = pop(); 1666 push( b ); 1667 push( a ); 1668 push( c ); 1669 push( b ); 1670 push( a ); 1671 break; 1672 case Bytecodes::_dup2_x2: 1673 // before: .. d, c, b, a 1674 // after: .. b, a, d, c, b, a 1675 // not tested 1676 a = pop(); 1677 b = pop(); 1678 c = pop(); 1679 d = pop(); 1680 push( b ); 1681 push( a ); 1682 push( d ); 1683 push( c ); 1684 push( b ); 1685 push( a ); 1686 break; 1687 1688 case Bytecodes::_arraylength: { 1689 // Must do null-check with value on expression stack 1690 Node *ary = null_check(peek(), T_ARRAY); 1691 // Compile-time detect of null-exception? 1692 if (stopped()) return; 1693 a = pop(); 1694 push(load_array_length(a)); 1695 break; 1696 } 1697 1698 case Bytecodes::_baload: array_load(T_BYTE); break; 1699 case Bytecodes::_caload: array_load(T_CHAR); break; 1700 case Bytecodes::_iaload: array_load(T_INT); break; 1701 case Bytecodes::_saload: array_load(T_SHORT); break; 1702 case Bytecodes::_faload: array_load(T_FLOAT); break; 1703 case Bytecodes::_aaload: array_load(T_OBJECT); break; 1704 case Bytecodes::_laload: { 1705 a = array_addressing(T_LONG, 0, false); 1706 if (stopped()) return; // guaranteed null or range check 1707 dec_sp(2); // Pop array and index 1708 push_pair(make_load(control(), a, TypeLong::LONG, T_LONG, TypeAryPtr::LONGS, MemNode::unordered)); 1709 break; 1710 } 1711 case Bytecodes::_daload: { 1712 a = array_addressing(T_DOUBLE, 0, false); 1713 if (stopped()) return; // guaranteed null or range check 1714 dec_sp(2); // Pop array and index 1715 push_pair(make_load(control(), a, Type::DOUBLE, T_DOUBLE, TypeAryPtr::DOUBLES, MemNode::unordered)); 1716 break; 1717 } 1718 case Bytecodes::_bastore: array_store(T_BYTE); break; 1719 case Bytecodes::_castore: array_store(T_CHAR); break; 1720 case Bytecodes::_iastore: array_store(T_INT); break; 1721 case Bytecodes::_sastore: array_store(T_SHORT); break; 1722 case Bytecodes::_fastore: array_store(T_FLOAT); break; 1723 case Bytecodes::_aastore: { 1724 d = array_addressing(T_OBJECT, 1, true); 1725 if (stopped()) return; // guaranteed null or range check 1726 array_store_check(); 1727 c = pop(); // Oop to store 1728 b = pop(); // index (already used) 1729 a = pop(); // the array itself 1730 const TypeOopPtr* elemtype = _gvn.type(a)->is_aryptr()->elem()->make_oopptr(); 1731 const TypeAryPtr* adr_type = TypeAryPtr::OOPS; 1732 // Note: We don't need a write barrier for Shenandoah on a here, because 1733 // a is not used except for an assert. The address d already has the 1734 // write barrier. Adding a barrier on a only results in additional code 1735 // being generated. 1736 c = shenandoah_read_barrier_nomem(c); 1737 Node* store = store_oop_to_array(control(), a, d, adr_type, c, elemtype, T_OBJECT, 1738 StoreNode::release_if_reference(T_OBJECT)); 1739 break; 1740 } 1741 case Bytecodes::_lastore: { 1742 a = array_addressing(T_LONG, 2, true); 1743 if (stopped()) return; // guaranteed null or range check 1744 c = pop_pair(); 1745 dec_sp(2); // Pop array and index 1746 store_to_memory(control(), a, c, T_LONG, TypeAryPtr::LONGS, MemNode::unordered); 1747 break; 1748 } 1749 case Bytecodes::_dastore: { 1750 a = array_addressing(T_DOUBLE, 2, true); 1751 if (stopped()) return; // guaranteed null or range check 1752 c = pop_pair(); 1753 dec_sp(2); // Pop array and index 1754 c = dstore_rounding(c); 1755 store_to_memory(control(), a, c, T_DOUBLE, TypeAryPtr::DOUBLES, MemNode::unordered); 1756 break; 1757 } 1758 case Bytecodes::_getfield: 1759 do_getfield(); 1760 break; 1761 1762 case Bytecodes::_getstatic: 1763 do_getstatic(); 1764 break; 1765 1766 case Bytecodes::_putfield: 1767 do_putfield(); 1768 break; 1769 1770 case Bytecodes::_putstatic: 1771 do_putstatic(); 1772 break; 1773 1774 case Bytecodes::_irem: 1775 do_irem(); 1776 break; 1777 case Bytecodes::_idiv: 1778 // Must keep both values on the expression-stack during null-check 1779 zero_check_int(peek()); 1780 // Compile-time detect of null-exception? 1781 if (stopped()) return; 1782 b = pop(); 1783 a = pop(); 1784 push( _gvn.transform( new DivINode(control(),a,b) ) ); 1785 break; 1786 case Bytecodes::_imul: 1787 b = pop(); a = pop(); 1788 push( _gvn.transform( new MulINode(a,b) ) ); 1789 break; 1790 case Bytecodes::_iadd: 1791 b = pop(); a = pop(); 1792 push( _gvn.transform( new AddINode(a,b) ) ); 1793 break; 1794 case Bytecodes::_ineg: 1795 a = pop(); 1796 push( _gvn.transform( new SubINode(_gvn.intcon(0),a)) ); 1797 break; 1798 case Bytecodes::_isub: 1799 b = pop(); a = pop(); 1800 push( _gvn.transform( new SubINode(a,b) ) ); 1801 break; 1802 case Bytecodes::_iand: 1803 b = pop(); a = pop(); 1804 push( _gvn.transform( new AndINode(a,b) ) ); 1805 break; 1806 case Bytecodes::_ior: 1807 b = pop(); a = pop(); 1808 push( _gvn.transform( new OrINode(a,b) ) ); 1809 break; 1810 case Bytecodes::_ixor: 1811 b = pop(); a = pop(); 1812 push( _gvn.transform( new XorINode(a,b) ) ); 1813 break; 1814 case Bytecodes::_ishl: 1815 b = pop(); a = pop(); 1816 push( _gvn.transform( new LShiftINode(a,b) ) ); 1817 break; 1818 case Bytecodes::_ishr: 1819 b = pop(); a = pop(); 1820 push( _gvn.transform( new RShiftINode(a,b) ) ); 1821 break; 1822 case Bytecodes::_iushr: 1823 b = pop(); a = pop(); 1824 push( _gvn.transform( new URShiftINode(a,b) ) ); 1825 break; 1826 1827 case Bytecodes::_fneg: 1828 a = pop(); 1829 b = _gvn.transform(new NegFNode (a)); 1830 push(b); 1831 break; 1832 1833 case Bytecodes::_fsub: 1834 b = pop(); 1835 a = pop(); 1836 c = _gvn.transform( new SubFNode(a,b) ); 1837 d = precision_rounding(c); 1838 push( d ); 1839 break; 1840 1841 case Bytecodes::_fadd: 1842 b = pop(); 1843 a = pop(); 1844 c = _gvn.transform( new AddFNode(a,b) ); 1845 d = precision_rounding(c); 1846 push( d ); 1847 break; 1848 1849 case Bytecodes::_fmul: 1850 b = pop(); 1851 a = pop(); 1852 c = _gvn.transform( new MulFNode(a,b) ); 1853 d = precision_rounding(c); 1854 push( d ); 1855 break; 1856 1857 case Bytecodes::_fdiv: 1858 b = pop(); 1859 a = pop(); 1860 c = _gvn.transform( new DivFNode(0,a,b) ); 1861 d = precision_rounding(c); 1862 push( d ); 1863 break; 1864 1865 case Bytecodes::_frem: 1866 if (Matcher::has_match_rule(Op_ModF)) { 1867 // Generate a ModF node. 1868 b = pop(); 1869 a = pop(); 1870 c = _gvn.transform( new ModFNode(0,a,b) ); 1871 d = precision_rounding(c); 1872 push( d ); 1873 } 1874 else { 1875 // Generate a call. 1876 modf(); 1877 } 1878 break; 1879 1880 case Bytecodes::_fcmpl: 1881 b = pop(); 1882 a = pop(); 1883 c = _gvn.transform( new CmpF3Node( a, b)); 1884 push(c); 1885 break; 1886 case Bytecodes::_fcmpg: 1887 b = pop(); 1888 a = pop(); 1889 1890 // Same as fcmpl but need to flip the unordered case. Swap the inputs, 1891 // which negates the result sign except for unordered. Flip the unordered 1892 // as well by using CmpF3 which implements unordered-lesser instead of 1893 // unordered-greater semantics. Finally, commute the result bits. Result 1894 // is same as using a CmpF3Greater except we did it with CmpF3 alone. 1895 c = _gvn.transform( new CmpF3Node( b, a)); 1896 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) ); 1897 push(c); 1898 break; 1899 1900 case Bytecodes::_f2i: 1901 a = pop(); 1902 push(_gvn.transform(new ConvF2INode(a))); 1903 break; 1904 1905 case Bytecodes::_d2i: 1906 a = pop_pair(); 1907 b = _gvn.transform(new ConvD2INode(a)); 1908 push( b ); 1909 break; 1910 1911 case Bytecodes::_f2d: 1912 a = pop(); 1913 b = _gvn.transform( new ConvF2DNode(a)); 1914 push_pair( b ); 1915 break; 1916 1917 case Bytecodes::_d2f: 1918 a = pop_pair(); 1919 b = _gvn.transform( new ConvD2FNode(a)); 1920 // This breaks _227_mtrt (speed & correctness) and _222_mpegaudio (speed) 1921 //b = _gvn.transform(new RoundFloatNode(0, b) ); 1922 push( b ); 1923 break; 1924 1925 case Bytecodes::_l2f: 1926 if (Matcher::convL2FSupported()) { 1927 a = pop_pair(); 1928 b = _gvn.transform( new ConvL2FNode(a)); 1929 // For i486.ad, FILD doesn't restrict precision to 24 or 53 bits. 1930 // Rather than storing the result into an FP register then pushing 1931 // out to memory to round, the machine instruction that implements 1932 // ConvL2D is responsible for rounding. 1933 // c = precision_rounding(b); 1934 c = _gvn.transform(b); 1935 push(c); 1936 } else { 1937 l2f(); 1938 } 1939 break; 1940 1941 case Bytecodes::_l2d: 1942 a = pop_pair(); 1943 b = _gvn.transform( new ConvL2DNode(a)); 1944 // For i486.ad, rounding is always necessary (see _l2f above). 1945 // c = dprecision_rounding(b); 1946 c = _gvn.transform(b); 1947 push_pair(c); 1948 break; 1949 1950 case Bytecodes::_f2l: 1951 a = pop(); 1952 b = _gvn.transform( new ConvF2LNode(a)); 1953 push_pair(b); 1954 break; 1955 1956 case Bytecodes::_d2l: 1957 a = pop_pair(); 1958 b = _gvn.transform( new ConvD2LNode(a)); 1959 push_pair(b); 1960 break; 1961 1962 case Bytecodes::_dsub: 1963 b = pop_pair(); 1964 a = pop_pair(); 1965 c = _gvn.transform( new SubDNode(a,b) ); 1966 d = dprecision_rounding(c); 1967 push_pair( d ); 1968 break; 1969 1970 case Bytecodes::_dadd: 1971 b = pop_pair(); 1972 a = pop_pair(); 1973 c = _gvn.transform( new AddDNode(a,b) ); 1974 d = dprecision_rounding(c); 1975 push_pair( d ); 1976 break; 1977 1978 case Bytecodes::_dmul: 1979 b = pop_pair(); 1980 a = pop_pair(); 1981 c = _gvn.transform( new MulDNode(a,b) ); 1982 d = dprecision_rounding(c); 1983 push_pair( d ); 1984 break; 1985 1986 case Bytecodes::_ddiv: 1987 b = pop_pair(); 1988 a = pop_pair(); 1989 c = _gvn.transform( new DivDNode(0,a,b) ); 1990 d = dprecision_rounding(c); 1991 push_pair( d ); 1992 break; 1993 1994 case Bytecodes::_dneg: 1995 a = pop_pair(); 1996 b = _gvn.transform(new NegDNode (a)); 1997 push_pair(b); 1998 break; 1999 2000 case Bytecodes::_drem: 2001 if (Matcher::has_match_rule(Op_ModD)) { 2002 // Generate a ModD node. 2003 b = pop_pair(); 2004 a = pop_pair(); 2005 // a % b 2006 2007 c = _gvn.transform( new ModDNode(0,a,b) ); 2008 d = dprecision_rounding(c); 2009 push_pair( d ); 2010 } 2011 else { 2012 // Generate a call. 2013 modd(); 2014 } 2015 break; 2016 2017 case Bytecodes::_dcmpl: 2018 b = pop_pair(); 2019 a = pop_pair(); 2020 c = _gvn.transform( new CmpD3Node( a, b)); 2021 push(c); 2022 break; 2023 2024 case Bytecodes::_dcmpg: 2025 b = pop_pair(); 2026 a = pop_pair(); 2027 // Same as dcmpl but need to flip the unordered case. 2028 // Commute the inputs, which negates the result sign except for unordered. 2029 // Flip the unordered as well by using CmpD3 which implements 2030 // unordered-lesser instead of unordered-greater semantics. 2031 // Finally, negate the result bits. Result is same as using a 2032 // CmpD3Greater except we did it with CmpD3 alone. 2033 c = _gvn.transform( new CmpD3Node( b, a)); 2034 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) ); 2035 push(c); 2036 break; 2037 2038 2039 // Note for longs -> lo word is on TOS, hi word is on TOS - 1 2040 case Bytecodes::_land: 2041 b = pop_pair(); 2042 a = pop_pair(); 2043 c = _gvn.transform( new AndLNode(a,b) ); 2044 push_pair(c); 2045 break; 2046 case Bytecodes::_lor: 2047 b = pop_pair(); 2048 a = pop_pair(); 2049 c = _gvn.transform( new OrLNode(a,b) ); 2050 push_pair(c); 2051 break; 2052 case Bytecodes::_lxor: 2053 b = pop_pair(); 2054 a = pop_pair(); 2055 c = _gvn.transform( new XorLNode(a,b) ); 2056 push_pair(c); 2057 break; 2058 2059 case Bytecodes::_lshl: 2060 b = pop(); // the shift count 2061 a = pop_pair(); // value to be shifted 2062 c = _gvn.transform( new LShiftLNode(a,b) ); 2063 push_pair(c); 2064 break; 2065 case Bytecodes::_lshr: 2066 b = pop(); // the shift count 2067 a = pop_pair(); // value to be shifted 2068 c = _gvn.transform( new RShiftLNode(a,b) ); 2069 push_pair(c); 2070 break; 2071 case Bytecodes::_lushr: 2072 b = pop(); // the shift count 2073 a = pop_pair(); // value to be shifted 2074 c = _gvn.transform( new URShiftLNode(a,b) ); 2075 push_pair(c); 2076 break; 2077 case Bytecodes::_lmul: 2078 b = pop_pair(); 2079 a = pop_pair(); 2080 c = _gvn.transform( new MulLNode(a,b) ); 2081 push_pair(c); 2082 break; 2083 2084 case Bytecodes::_lrem: 2085 // Must keep both values on the expression-stack during null-check 2086 assert(peek(0) == top(), "long word order"); 2087 zero_check_long(peek(1)); 2088 // Compile-time detect of null-exception? 2089 if (stopped()) return; 2090 b = pop_pair(); 2091 a = pop_pair(); 2092 c = _gvn.transform( new ModLNode(control(),a,b) ); 2093 push_pair(c); 2094 break; 2095 2096 case Bytecodes::_ldiv: 2097 // Must keep both values on the expression-stack during null-check 2098 assert(peek(0) == top(), "long word order"); 2099 zero_check_long(peek(1)); 2100 // Compile-time detect of null-exception? 2101 if (stopped()) return; 2102 b = pop_pair(); 2103 a = pop_pair(); 2104 c = _gvn.transform( new DivLNode(control(),a,b) ); 2105 push_pair(c); 2106 break; 2107 2108 case Bytecodes::_ladd: 2109 b = pop_pair(); 2110 a = pop_pair(); 2111 c = _gvn.transform( new AddLNode(a,b) ); 2112 push_pair(c); 2113 break; 2114 case Bytecodes::_lsub: 2115 b = pop_pair(); 2116 a = pop_pair(); 2117 c = _gvn.transform( new SubLNode(a,b) ); 2118 push_pair(c); 2119 break; 2120 case Bytecodes::_lcmp: 2121 // Safepoints are now inserted _before_ branches. The long-compare 2122 // bytecode painfully produces a 3-way value (-1,0,+1) which requires a 2123 // slew of control flow. These are usually followed by a CmpI vs zero and 2124 // a branch; this pattern then optimizes to the obvious long-compare and 2125 // branch. However, if the branch is backwards there's a Safepoint 2126 // inserted. The inserted Safepoint captures the JVM state at the 2127 // pre-branch point, i.e. it captures the 3-way value. Thus if a 2128 // long-compare is used to control a loop the debug info will force 2129 // computation of the 3-way value, even though the generated code uses a 2130 // long-compare and branch. We try to rectify the situation by inserting 2131 // a SafePoint here and have it dominate and kill the safepoint added at a 2132 // following backwards branch. At this point the JVM state merely holds 2 2133 // longs but not the 3-way value. 2134 if( UseLoopSafepoints ) { 2135 switch( iter().next_bc() ) { 2136 case Bytecodes::_ifgt: 2137 case Bytecodes::_iflt: 2138 case Bytecodes::_ifge: 2139 case Bytecodes::_ifle: 2140 case Bytecodes::_ifne: 2141 case Bytecodes::_ifeq: 2142 // If this is a backwards branch in the bytecodes, add Safepoint 2143 maybe_add_safepoint(iter().next_get_dest()); 2144 } 2145 } 2146 b = pop_pair(); 2147 a = pop_pair(); 2148 c = _gvn.transform( new CmpL3Node( a, b )); 2149 push(c); 2150 break; 2151 2152 case Bytecodes::_lneg: 2153 a = pop_pair(); 2154 b = _gvn.transform( new SubLNode(longcon(0),a)); 2155 push_pair(b); 2156 break; 2157 case Bytecodes::_l2i: 2158 a = pop_pair(); 2159 push( _gvn.transform( new ConvL2INode(a))); 2160 break; 2161 case Bytecodes::_i2l: 2162 a = pop(); 2163 b = _gvn.transform( new ConvI2LNode(a)); 2164 push_pair(b); 2165 break; 2166 case Bytecodes::_i2b: 2167 // Sign extend 2168 a = pop(); 2169 a = _gvn.transform( new LShiftINode(a,_gvn.intcon(24)) ); 2170 a = _gvn.transform( new RShiftINode(a,_gvn.intcon(24)) ); 2171 push( a ); 2172 break; 2173 case Bytecodes::_i2s: 2174 a = pop(); 2175 a = _gvn.transform( new LShiftINode(a,_gvn.intcon(16)) ); 2176 a = _gvn.transform( new RShiftINode(a,_gvn.intcon(16)) ); 2177 push( a ); 2178 break; 2179 case Bytecodes::_i2c: 2180 a = pop(); 2181 push( _gvn.transform( new AndINode(a,_gvn.intcon(0xFFFF)) ) ); 2182 break; 2183 2184 case Bytecodes::_i2f: 2185 a = pop(); 2186 b = _gvn.transform( new ConvI2FNode(a) ) ; 2187 c = precision_rounding(b); 2188 push (b); 2189 break; 2190 2191 case Bytecodes::_i2d: 2192 a = pop(); 2193 b = _gvn.transform( new ConvI2DNode(a)); 2194 push_pair(b); 2195 break; 2196 2197 case Bytecodes::_iinc: // Increment local 2198 i = iter().get_index(); // Get local index 2199 set_local( i, _gvn.transform( new AddINode( _gvn.intcon(iter().get_iinc_con()), local(i) ) ) ); 2200 break; 2201 2202 // Exit points of synchronized methods must have an unlock node 2203 case Bytecodes::_return: 2204 return_current(NULL); 2205 break; 2206 2207 case Bytecodes::_ireturn: 2208 case Bytecodes::_areturn: 2209 case Bytecodes::_freturn: 2210 return_current(pop()); 2211 break; 2212 case Bytecodes::_lreturn: 2213 return_current(pop_pair()); 2214 break; 2215 case Bytecodes::_dreturn: 2216 return_current(pop_pair()); 2217 break; 2218 2219 case Bytecodes::_athrow: 2220 // null exception oop throws NULL pointer exception 2221 null_check(peek()); 2222 if (stopped()) return; 2223 // Hook the thrown exception directly to subsequent handlers. 2224 if (BailoutToInterpreterForThrows) { 2225 // Keep method interpreted from now on. 2226 uncommon_trap(Deoptimization::Reason_unhandled, 2227 Deoptimization::Action_make_not_compilable); 2228 return; 2229 } 2230 if (env()->jvmti_can_post_on_exceptions()) { 2231 // check if we must post exception events, take uncommon trap if so (with must_throw = false) 2232 uncommon_trap_if_should_post_on_exceptions(Deoptimization::Reason_unhandled, false); 2233 } 2234 // Here if either can_post_on_exceptions or should_post_on_exceptions is false 2235 add_exception_state(make_exception_state(peek())); 2236 break; 2237 2238 case Bytecodes::_goto: // fall through 2239 case Bytecodes::_goto_w: { 2240 int target_bci = (bc() == Bytecodes::_goto) ? iter().get_dest() : iter().get_far_dest(); 2241 2242 // If this is a backwards branch in the bytecodes, add Safepoint 2243 maybe_add_safepoint(target_bci); 2244 2245 // Update method data 2246 profile_taken_branch(target_bci); 2247 2248 // Merge the current control into the target basic block 2249 merge(target_bci); 2250 2251 // See if we can get some profile data and hand it off to the next block 2252 Block *target_block = block()->successor_for_bci(target_bci); 2253 if (target_block->pred_count() != 1) break; 2254 ciMethodData* methodData = method()->method_data(); 2255 if (!methodData->is_mature()) break; 2256 ciProfileData* data = methodData->bci_to_data(bci()); 2257 assert( data->is_JumpData(), "" ); 2258 int taken = ((ciJumpData*)data)->taken(); 2259 taken = method()->scale_count(taken); 2260 target_block->set_count(taken); 2261 break; 2262 } 2263 2264 case Bytecodes::_ifnull: btest = BoolTest::eq; goto handle_if_null; 2265 case Bytecodes::_ifnonnull: btest = BoolTest::ne; goto handle_if_null; 2266 handle_if_null: 2267 // If this is a backwards branch in the bytecodes, add Safepoint 2268 maybe_add_safepoint(iter().get_dest()); 2269 a = null(); 2270 b = pop(); 2271 if (!_gvn.type(b)->speculative_maybe_null() && 2272 !too_many_traps(Deoptimization::Reason_speculate_null_check)) { 2273 inc_sp(1); 2274 Node* null_ctl = top(); 2275 b = null_check_oop(b, &null_ctl, true, true, true); 2276 assert(null_ctl->is_top(), "no null control here"); 2277 dec_sp(1); 2278 } 2279 c = _gvn.transform( new CmpPNode(b, a) ); 2280 do_ifnull(btest, c); 2281 break; 2282 2283 case Bytecodes::_if_acmpeq: btest = BoolTest::eq; goto handle_if_acmp; 2284 case Bytecodes::_if_acmpne: btest = BoolTest::ne; goto handle_if_acmp; 2285 handle_if_acmp: 2286 // If this is a backwards branch in the bytecodes, add Safepoint 2287 maybe_add_safepoint(iter().get_dest()); 2288 a = pop(); 2289 b = pop(); 2290 shenandoah_acmp_barrier(a, b); 2291 c = _gvn.transform( new CmpPNode(b, a) ); 2292 c = optimize_cmp_with_klass(c); 2293 do_if(btest, c); 2294 break; 2295 2296 case Bytecodes::_ifeq: btest = BoolTest::eq; goto handle_ifxx; 2297 case Bytecodes::_ifne: btest = BoolTest::ne; goto handle_ifxx; 2298 case Bytecodes::_iflt: btest = BoolTest::lt; goto handle_ifxx; 2299 case Bytecodes::_ifle: btest = BoolTest::le; goto handle_ifxx; 2300 case Bytecodes::_ifgt: btest = BoolTest::gt; goto handle_ifxx; 2301 case Bytecodes::_ifge: btest = BoolTest::ge; goto handle_ifxx; 2302 handle_ifxx: 2303 // If this is a backwards branch in the bytecodes, add Safepoint 2304 maybe_add_safepoint(iter().get_dest()); 2305 a = _gvn.intcon(0); 2306 b = pop(); 2307 c = _gvn.transform( new CmpINode(b, a) ); 2308 do_if(btest, c); 2309 break; 2310 2311 case Bytecodes::_if_icmpeq: btest = BoolTest::eq; goto handle_if_icmp; 2312 case Bytecodes::_if_icmpne: btest = BoolTest::ne; goto handle_if_icmp; 2313 case Bytecodes::_if_icmplt: btest = BoolTest::lt; goto handle_if_icmp; 2314 case Bytecodes::_if_icmple: btest = BoolTest::le; goto handle_if_icmp; 2315 case Bytecodes::_if_icmpgt: btest = BoolTest::gt; goto handle_if_icmp; 2316 case Bytecodes::_if_icmpge: btest = BoolTest::ge; goto handle_if_icmp; 2317 handle_if_icmp: 2318 // If this is a backwards branch in the bytecodes, add Safepoint 2319 maybe_add_safepoint(iter().get_dest()); 2320 a = pop(); 2321 b = pop(); 2322 c = _gvn.transform( new CmpINode( b, a ) ); 2323 do_if(btest, c); 2324 break; 2325 2326 case Bytecodes::_tableswitch: 2327 do_tableswitch(); 2328 break; 2329 2330 case Bytecodes::_lookupswitch: 2331 do_lookupswitch(); 2332 break; 2333 2334 case Bytecodes::_invokestatic: 2335 case Bytecodes::_invokedynamic: 2336 case Bytecodes::_invokespecial: 2337 case Bytecodes::_invokevirtual: 2338 case Bytecodes::_invokeinterface: 2339 do_call(); 2340 break; 2341 case Bytecodes::_checkcast: 2342 do_checkcast(); 2343 break; 2344 case Bytecodes::_instanceof: 2345 do_instanceof(); 2346 break; 2347 case Bytecodes::_anewarray: 2348 do_anewarray(); 2349 break; 2350 case Bytecodes::_newarray: 2351 do_newarray((BasicType)iter().get_index()); 2352 break; 2353 case Bytecodes::_multianewarray: 2354 do_multianewarray(); 2355 break; 2356 case Bytecodes::_new: 2357 do_new(); 2358 break; 2359 2360 case Bytecodes::_jsr: 2361 case Bytecodes::_jsr_w: 2362 do_jsr(); 2363 break; 2364 2365 case Bytecodes::_ret: 2366 do_ret(); 2367 break; 2368 2369 2370 case Bytecodes::_monitorenter: 2371 do_monitor_enter(); 2372 break; 2373 2374 case Bytecodes::_monitorexit: 2375 do_monitor_exit(); 2376 break; 2377 2378 case Bytecodes::_breakpoint: 2379 // Breakpoint set concurrently to compile 2380 // %%% use an uncommon trap? 2381 C->record_failure("breakpoint in method"); 2382 return; 2383 2384 default: 2385 #ifndef PRODUCT 2386 map()->dump(99); 2387 #endif 2388 tty->print("\nUnhandled bytecode %s\n", Bytecodes::name(bc()) ); 2389 ShouldNotReachHere(); 2390 } 2391 2392 #ifndef PRODUCT 2393 IdealGraphPrinter *printer = IdealGraphPrinter::printer(); 2394 if (printer && printer->should_print(_method)) { 2395 char buffer[256]; 2396 sprintf(buffer, "Bytecode %d: %s", bci(), Bytecodes::name(bc())); 2397 bool old = printer->traverse_outs(); 2398 printer->set_traverse_outs(true); 2399 printer->print_method(C, buffer, 4); 2400 printer->set_traverse_outs(old); 2401 } 2402 #endif 2403 }