1 /* 2 * Copyright (c) 1998, 2016, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "ci/ciMethodData.hpp" 27 #include "classfile/systemDictionary.hpp" 28 #include "classfile/vmSymbols.hpp" 29 #include "compiler/compileLog.hpp" 30 #include "interpreter/linkResolver.hpp" 31 #include "memory/universe.inline.hpp" 32 #include "opto/addnode.hpp" 33 #include "opto/divnode.hpp" 34 #include "opto/idealGraphPrinter.hpp" 35 #include "opto/matcher.hpp" 36 #include "opto/memnode.hpp" 37 #include "opto/mulnode.hpp" 38 #include "opto/parse.hpp" 39 #include "opto/runtime.hpp" 40 #include "runtime/deoptimization.hpp" 41 #include "runtime/sharedRuntime.hpp" 42 43 extern int explicit_null_checks_inserted, 44 explicit_null_checks_elided; 45 46 //---------------------------------array_load---------------------------------- 47 void Parse::array_load(BasicType elem_type) { 48 const Type* elem = Type::TOP; 49 Node* adr = array_addressing(elem_type, 0, false, &elem); 50 if (stopped()) return; // guaranteed null or range check 51 dec_sp(2); // Pop array and index 52 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(elem_type); 53 Node* ld = make_load(control(), adr, elem, elem_type, adr_type, MemNode::unordered); 54 push(ld); 55 } 56 57 58 //--------------------------------array_store---------------------------------- 59 void Parse::array_store(BasicType elem_type) { 60 const Type* elem = Type::TOP; 61 Node* adr = array_addressing(elem_type, 1, true, &elem); 62 if (stopped()) return; // guaranteed null or range check 63 Node* val = pop(); 64 dec_sp(2); // Pop array and index 65 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(elem_type); 66 if (elem == TypeInt::BOOL) { 67 elem_type = T_BOOLEAN; 68 } 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 (C) CmpUNode(idx, len) ); 136 BoolTest::mask btest = BoolTest::lt; 137 tst = _gvn.transform( new (C) 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 // Make array address computation control dependent to prevent it 168 // from floating above the range check during loop optimizations. 169 Node* ptr = array_element_address(ary, idx, type, sizetype, control()); 170 171 if (result2 != NULL) *result2 = elemtype; 172 173 assert(ptr != top(), "top should go hand-in-hand with stopped"); 174 175 return ptr; 176 } 177 178 179 // returns IfNode 180 IfNode* Parse::jump_if_fork_int(Node* a, Node* b, BoolTest::mask mask) { 181 Node *cmp = _gvn.transform( new (C) CmpINode( a, b)); // two cases: shiftcount > 32 and shiftcount <= 32 182 Node *tst = _gvn.transform( new (C) BoolNode( cmp, mask)); 183 IfNode *iff = create_and_map_if( control(), tst, ((mask == BoolTest::eq) ? PROB_STATIC_INFREQUENT : PROB_FAIR), COUNT_UNKNOWN ); 184 return iff; 185 } 186 187 // return Region node 188 Node* Parse::jump_if_join(Node* iffalse, Node* iftrue) { 189 Node *region = new (C) RegionNode(3); // 2 results 190 record_for_igvn(region); 191 region->init_req(1, iffalse); 192 region->init_req(2, iftrue ); 193 _gvn.set_type(region, Type::CONTROL); 194 region = _gvn.transform(region); 195 set_control (region); 196 return region; 197 } 198 199 200 //------------------------------helper for tableswitch------------------------- 201 void Parse::jump_if_true_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index) { 202 // True branch, use existing map info 203 { PreserveJVMState pjvms(this); 204 Node *iftrue = _gvn.transform( new (C) IfTrueNode (iff) ); 205 set_control( iftrue ); 206 profile_switch_case(prof_table_index); 207 merge_new_path(dest_bci_if_true); 208 } 209 210 // False branch 211 Node *iffalse = _gvn.transform( new (C) IfFalseNode(iff) ); 212 set_control( iffalse ); 213 } 214 215 void Parse::jump_if_false_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index) { 216 // True branch, use existing map info 217 { PreserveJVMState pjvms(this); 218 Node *iffalse = _gvn.transform( new (C) IfFalseNode (iff) ); 219 set_control( iffalse ); 220 profile_switch_case(prof_table_index); 221 merge_new_path(dest_bci_if_true); 222 } 223 224 // False branch 225 Node *iftrue = _gvn.transform( new (C) IfTrueNode(iff) ); 226 set_control( iftrue ); 227 } 228 229 void Parse::jump_if_always_fork(int dest_bci, int prof_table_index) { 230 // False branch, use existing map and control() 231 profile_switch_case(prof_table_index); 232 merge_new_path(dest_bci); 233 } 234 235 236 extern "C" { 237 static int jint_cmp(const void *i, const void *j) { 238 int a = *(jint *)i; 239 int b = *(jint *)j; 240 return a > b ? 1 : a < b ? -1 : 0; 241 } 242 } 243 244 245 // Default value for methodData switch indexing. Must be a negative value to avoid 246 // conflict with any legal switch index. 247 #define NullTableIndex -1 248 249 class SwitchRange : public StackObj { 250 // a range of integers coupled with a bci destination 251 jint _lo; // inclusive lower limit 252 jint _hi; // inclusive upper limit 253 int _dest; 254 int _table_index; // index into method data table 255 256 public: 257 jint lo() const { return _lo; } 258 jint hi() const { return _hi; } 259 int dest() const { return _dest; } 260 int table_index() const { return _table_index; } 261 bool is_singleton() const { return _lo == _hi; } 262 263 void setRange(jint lo, jint hi, int dest, int table_index) { 264 assert(lo <= hi, "must be a non-empty range"); 265 _lo = lo, _hi = hi; _dest = dest; _table_index = table_index; 266 } 267 bool adjoinRange(jint lo, jint hi, int dest, int table_index) { 268 assert(lo <= hi, "must be a non-empty range"); 269 if (lo == _hi+1 && dest == _dest && table_index == _table_index) { 270 _hi = hi; 271 return true; 272 } 273 return false; 274 } 275 276 void set (jint value, int dest, int table_index) { 277 setRange(value, value, dest, table_index); 278 } 279 bool adjoin(jint value, int dest, int table_index) { 280 return adjoinRange(value, value, dest, table_index); 281 } 282 283 void print() { 284 if (is_singleton()) 285 tty->print(" {%d}=>%d", lo(), dest()); 286 else if (lo() == min_jint) 287 tty->print(" {..%d}=>%d", hi(), dest()); 288 else if (hi() == max_jint) 289 tty->print(" {%d..}=>%d", lo(), dest()); 290 else 291 tty->print(" {%d..%d}=>%d", lo(), hi(), dest()); 292 } 293 }; 294 295 296 //-------------------------------do_tableswitch-------------------------------- 297 void Parse::do_tableswitch() { 298 Node* lookup = pop(); 299 300 // Get information about tableswitch 301 int default_dest = iter().get_dest_table(0); 302 int lo_index = iter().get_int_table(1); 303 int hi_index = iter().get_int_table(2); 304 int len = hi_index - lo_index + 1; 305 306 if (len < 1) { 307 // If this is a backward branch, add safepoint 308 maybe_add_safepoint(default_dest); 309 merge(default_dest); 310 return; 311 } 312 313 // generate decision tree, using trichotomy when possible 314 int rnum = len+2; 315 bool makes_backward_branch = false; 316 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum); 317 int rp = -1; 318 if (lo_index != min_jint) { 319 ranges[++rp].setRange(min_jint, lo_index-1, default_dest, NullTableIndex); 320 } 321 for (int j = 0; j < len; j++) { 322 jint match_int = lo_index+j; 323 int dest = iter().get_dest_table(j+3); 324 makes_backward_branch |= (dest <= bci()); 325 int table_index = method_data_update() ? j : NullTableIndex; 326 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index)) { 327 ranges[++rp].set(match_int, dest, table_index); 328 } 329 } 330 jint highest = lo_index+(len-1); 331 assert(ranges[rp].hi() == highest, ""); 332 if (highest != max_jint 333 && !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex)) { 334 ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex); 335 } 336 assert(rp < len+2, "not too many ranges"); 337 338 // Safepoint in case if backward branch observed 339 if( makes_backward_branch && UseLoopSafepoints ) 340 add_safepoint(); 341 342 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]); 343 } 344 345 346 //------------------------------do_lookupswitch-------------------------------- 347 void Parse::do_lookupswitch() { 348 Node *lookup = pop(); // lookup value 349 // Get information about lookupswitch 350 int default_dest = iter().get_dest_table(0); 351 int len = iter().get_int_table(1); 352 353 if (len < 1) { // If this is a backward branch, add safepoint 354 maybe_add_safepoint(default_dest); 355 merge(default_dest); 356 return; 357 } 358 359 // generate decision tree, using trichotomy when possible 360 jint* table = NEW_RESOURCE_ARRAY(jint, len*2); 361 { 362 for( int j = 0; j < len; j++ ) { 363 table[j+j+0] = iter().get_int_table(2+j+j); 364 table[j+j+1] = iter().get_dest_table(2+j+j+1); 365 } 366 qsort( table, len, 2*sizeof(table[0]), jint_cmp ); 367 } 368 369 int rnum = len*2+1; 370 bool makes_backward_branch = false; 371 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum); 372 int rp = -1; 373 for( int j = 0; j < len; j++ ) { 374 jint match_int = table[j+j+0]; 375 int dest = table[j+j+1]; 376 int next_lo = rp < 0 ? min_jint : ranges[rp].hi()+1; 377 int table_index = method_data_update() ? j : NullTableIndex; 378 makes_backward_branch |= (dest <= bci()); 379 if( match_int != next_lo ) { 380 ranges[++rp].setRange(next_lo, match_int-1, default_dest, NullTableIndex); 381 } 382 if( rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index) ) { 383 ranges[++rp].set(match_int, dest, table_index); 384 } 385 } 386 jint highest = table[2*(len-1)]; 387 assert(ranges[rp].hi() == highest, ""); 388 if( highest != max_jint 389 && !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex) ) { 390 ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex); 391 } 392 assert(rp < rnum, "not too many ranges"); 393 394 // Safepoint in case backward branch observed 395 if( makes_backward_branch && UseLoopSafepoints ) 396 add_safepoint(); 397 398 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]); 399 } 400 401 //----------------------------create_jump_tables------------------------------- 402 bool Parse::create_jump_tables(Node* key_val, SwitchRange* lo, SwitchRange* hi) { 403 // Are jumptables enabled 404 if (!UseJumpTables) return false; 405 406 // Are jumptables supported 407 if (!Matcher::has_match_rule(Op_Jump)) return false; 408 409 // Don't make jump table if profiling 410 if (method_data_update()) return false; 411 412 // Decide if a guard is needed to lop off big ranges at either (or 413 // both) end(s) of the input set. We'll call this the default target 414 // even though we can't be sure that it is the true "default". 415 416 bool needs_guard = false; 417 int default_dest; 418 int64 total_outlier_size = 0; 419 int64 hi_size = ((int64)hi->hi()) - ((int64)hi->lo()) + 1; 420 int64 lo_size = ((int64)lo->hi()) - ((int64)lo->lo()) + 1; 421 422 if (lo->dest() == hi->dest()) { 423 total_outlier_size = hi_size + lo_size; 424 default_dest = lo->dest(); 425 } else if (lo_size > hi_size) { 426 total_outlier_size = lo_size; 427 default_dest = lo->dest(); 428 } else { 429 total_outlier_size = hi_size; 430 default_dest = hi->dest(); 431 } 432 433 // If a guard test will eliminate very sparse end ranges, then 434 // it is worth the cost of an extra jump. 435 if (total_outlier_size > (MaxJumpTableSparseness * 4)) { 436 needs_guard = true; 437 if (default_dest == lo->dest()) lo++; 438 if (default_dest == hi->dest()) hi--; 439 } 440 441 // Find the total number of cases and ranges 442 int64 num_cases = ((int64)hi->hi()) - ((int64)lo->lo()) + 1; 443 int num_range = hi - lo + 1; 444 445 // Don't create table if: too large, too small, or too sparse. 446 if (num_cases < MinJumpTableSize || num_cases > MaxJumpTableSize) 447 return false; 448 if (num_cases > (MaxJumpTableSparseness * num_range)) 449 return false; 450 451 // Normalize table lookups to zero 452 int lowval = lo->lo(); 453 key_val = _gvn.transform( new (C) SubINode(key_val, _gvn.intcon(lowval)) ); 454 455 // Generate a guard to protect against input keyvals that aren't 456 // in the switch domain. 457 if (needs_guard) { 458 Node* size = _gvn.intcon(num_cases); 459 Node* cmp = _gvn.transform( new (C) CmpUNode(key_val, size) ); 460 Node* tst = _gvn.transform( new (C) BoolNode(cmp, BoolTest::ge) ); 461 IfNode* iff = create_and_map_if( control(), tst, PROB_FAIR, COUNT_UNKNOWN); 462 jump_if_true_fork(iff, default_dest, NullTableIndex); 463 } 464 465 // Create an ideal node JumpTable that has projections 466 // of all possible ranges for a switch statement 467 // The key_val input must be converted to a pointer offset and scaled. 468 // Compare Parse::array_addressing above. 469 #ifdef _LP64 470 // Clean the 32-bit int into a real 64-bit offset. 471 // Otherwise, the jint value 0 might turn into an offset of 0x0800000000. 472 const TypeInt* ikeytype = TypeInt::make(0, num_cases-1, Type::WidenMin); 473 // Make I2L conversion control dependent to prevent it from 474 // floating above the range check during loop optimizations. 475 key_val = C->constrained_convI2L(&_gvn, key_val, ikeytype, control()); 476 #endif 477 478 // Shift the value by wordsize so we have an index into the table, rather 479 // than a switch value 480 Node *shiftWord = _gvn.MakeConX(wordSize); 481 key_val = _gvn.transform( new (C) MulXNode( key_val, shiftWord)); 482 483 // Create the JumpNode 484 Node* jtn = _gvn.transform( new (C) JumpNode(control(), key_val, num_cases) ); 485 486 // These are the switch destinations hanging off the jumpnode 487 int i = 0; 488 for (SwitchRange* r = lo; r <= hi; r++) { 489 for (int64 j = r->lo(); j <= r->hi(); j++, i++) { 490 Node* input = _gvn.transform(new (C) JumpProjNode(jtn, i, r->dest(), (int)(j - lowval))); 491 { 492 PreserveJVMState pjvms(this); 493 set_control(input); 494 jump_if_always_fork(r->dest(), r->table_index()); 495 } 496 } 497 } 498 assert(i == num_cases, "miscount of cases"); 499 stop_and_kill_map(); // no more uses for this JVMS 500 return true; 501 } 502 503 //----------------------------jump_switch_ranges------------------------------- 504 void Parse::jump_switch_ranges(Node* key_val, SwitchRange *lo, SwitchRange *hi, int switch_depth) { 505 Block* switch_block = block(); 506 507 if (switch_depth == 0) { 508 // Do special processing for the top-level call. 509 assert(lo->lo() == min_jint, "initial range must exhaust Type::INT"); 510 assert(hi->hi() == max_jint, "initial range must exhaust Type::INT"); 511 512 // Decrement pred-numbers for the unique set of nodes. 513 #ifdef ASSERT 514 // Ensure that the block's successors are a (duplicate-free) set. 515 int successors_counted = 0; // block occurrences in [hi..lo] 516 int unique_successors = switch_block->num_successors(); 517 for (int i = 0; i < unique_successors; i++) { 518 Block* target = switch_block->successor_at(i); 519 520 // Check that the set of successors is the same in both places. 521 int successors_found = 0; 522 for (SwitchRange* p = lo; p <= hi; p++) { 523 if (p->dest() == target->start()) successors_found++; 524 } 525 assert(successors_found > 0, "successor must be known"); 526 successors_counted += successors_found; 527 } 528 assert(successors_counted == (hi-lo)+1, "no unexpected successors"); 529 #endif 530 531 // Maybe prune the inputs, based on the type of key_val. 532 jint min_val = min_jint; 533 jint max_val = max_jint; 534 const TypeInt* ti = key_val->bottom_type()->isa_int(); 535 if (ti != NULL) { 536 min_val = ti->_lo; 537 max_val = ti->_hi; 538 assert(min_val <= max_val, "invalid int type"); 539 } 540 while (lo->hi() < min_val) lo++; 541 if (lo->lo() < min_val) lo->setRange(min_val, lo->hi(), lo->dest(), lo->table_index()); 542 while (hi->lo() > max_val) hi--; 543 if (hi->hi() > max_val) hi->setRange(hi->lo(), max_val, hi->dest(), hi->table_index()); 544 } 545 546 #ifndef PRODUCT 547 if (switch_depth == 0) { 548 _max_switch_depth = 0; 549 _est_switch_depth = log2_intptr((hi-lo+1)-1)+1; 550 } 551 #endif 552 553 assert(lo <= hi, "must be a non-empty set of ranges"); 554 if (lo == hi) { 555 jump_if_always_fork(lo->dest(), lo->table_index()); 556 } else { 557 assert(lo->hi() == (lo+1)->lo()-1, "contiguous ranges"); 558 assert(hi->lo() == (hi-1)->hi()+1, "contiguous ranges"); 559 560 if (create_jump_tables(key_val, lo, hi)) return; 561 562 int nr = hi - lo + 1; 563 564 SwitchRange* mid = lo + nr/2; 565 // if there is an easy choice, pivot at a singleton: 566 if (nr > 3 && !mid->is_singleton() && (mid-1)->is_singleton()) mid--; 567 568 assert(lo < mid && mid <= hi, "good pivot choice"); 569 assert(nr != 2 || mid == hi, "should pick higher of 2"); 570 assert(nr != 3 || mid == hi-1, "should pick middle of 3"); 571 572 Node *test_val = _gvn.intcon(mid->lo()); 573 574 if (mid->is_singleton()) { 575 IfNode *iff_ne = jump_if_fork_int(key_val, test_val, BoolTest::ne); 576 jump_if_false_fork(iff_ne, mid->dest(), mid->table_index()); 577 578 // Special Case: If there are exactly three ranges, and the high 579 // and low range each go to the same place, omit the "gt" test, 580 // since it will not discriminate anything. 581 bool eq_test_only = (hi == lo+2 && hi->dest() == lo->dest()); 582 if (eq_test_only) { 583 assert(mid == hi-1, ""); 584 } 585 586 // if there is a higher range, test for it and process it: 587 if (mid < hi && !eq_test_only) { 588 // two comparisons of same values--should enable 1 test for 2 branches 589 // Use BoolTest::le instead of BoolTest::gt 590 IfNode *iff_le = jump_if_fork_int(key_val, test_val, BoolTest::le); 591 Node *iftrue = _gvn.transform( new (C) IfTrueNode(iff_le) ); 592 Node *iffalse = _gvn.transform( new (C) IfFalseNode(iff_le) ); 593 { PreserveJVMState pjvms(this); 594 set_control(iffalse); 595 jump_switch_ranges(key_val, mid+1, hi, switch_depth+1); 596 } 597 set_control(iftrue); 598 } 599 600 } else { 601 // mid is a range, not a singleton, so treat mid..hi as a unit 602 IfNode *iff_ge = jump_if_fork_int(key_val, test_val, BoolTest::ge); 603 604 // if there is a higher range, test for it and process it: 605 if (mid == hi) { 606 jump_if_true_fork(iff_ge, mid->dest(), mid->table_index()); 607 } else { 608 Node *iftrue = _gvn.transform( new (C) IfTrueNode(iff_ge) ); 609 Node *iffalse = _gvn.transform( new (C) IfFalseNode(iff_ge) ); 610 { PreserveJVMState pjvms(this); 611 set_control(iftrue); 612 jump_switch_ranges(key_val, mid, hi, switch_depth+1); 613 } 614 set_control(iffalse); 615 } 616 } 617 618 // in any case, process the lower range 619 jump_switch_ranges(key_val, lo, mid-1, switch_depth+1); 620 } 621 622 // Decrease pred_count for each successor after all is done. 623 if (switch_depth == 0) { 624 int unique_successors = switch_block->num_successors(); 625 for (int i = 0; i < unique_successors; i++) { 626 Block* target = switch_block->successor_at(i); 627 // Throw away the pre-allocated path for each unique successor. 628 target->next_path_num(); 629 } 630 } 631 632 #ifndef PRODUCT 633 _max_switch_depth = MAX2(switch_depth, _max_switch_depth); 634 if (TraceOptoParse && Verbose && WizardMode && switch_depth == 0) { 635 SwitchRange* r; 636 int nsing = 0; 637 for( r = lo; r <= hi; r++ ) { 638 if( r->is_singleton() ) nsing++; 639 } 640 tty->print(">>> "); 641 _method->print_short_name(); 642 tty->print_cr(" switch decision tree"); 643 tty->print_cr(" %d ranges (%d singletons), max_depth=%d, est_depth=%d", 644 (int) (hi-lo+1), nsing, _max_switch_depth, _est_switch_depth); 645 if (_max_switch_depth > _est_switch_depth) { 646 tty->print_cr("******** BAD SWITCH DEPTH ********"); 647 } 648 tty->print(" "); 649 for( r = lo; r <= hi; r++ ) { 650 r->print(); 651 } 652 tty->cr(); 653 } 654 #endif 655 } 656 657 void Parse::modf() { 658 Node *f2 = pop(); 659 Node *f1 = pop(); 660 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::modf_Type(), 661 CAST_FROM_FN_PTR(address, SharedRuntime::frem), 662 "frem", NULL, //no memory effects 663 f1, f2); 664 Node* res = _gvn.transform(new (C) ProjNode(c, TypeFunc::Parms + 0)); 665 666 push(res); 667 } 668 669 void Parse::modd() { 670 Node *d2 = pop_pair(); 671 Node *d1 = pop_pair(); 672 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::Math_DD_D_Type(), 673 CAST_FROM_FN_PTR(address, SharedRuntime::drem), 674 "drem", NULL, //no memory effects 675 d1, top(), d2, top()); 676 Node* res_d = _gvn.transform(new (C) ProjNode(c, TypeFunc::Parms + 0)); 677 678 #ifdef ASSERT 679 Node* res_top = _gvn.transform(new (C) ProjNode(c, TypeFunc::Parms + 1)); 680 assert(res_top == top(), "second value must be top"); 681 #endif 682 683 push_pair(res_d); 684 } 685 686 void Parse::l2f() { 687 Node* f2 = pop(); 688 Node* f1 = pop(); 689 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::l2f_Type(), 690 CAST_FROM_FN_PTR(address, SharedRuntime::l2f), 691 "l2f", NULL, //no memory effects 692 f1, f2); 693 Node* res = _gvn.transform(new (C) ProjNode(c, TypeFunc::Parms + 0)); 694 695 push(res); 696 } 697 698 void Parse::do_irem() { 699 // Must keep both values on the expression-stack during null-check 700 zero_check_int(peek()); 701 // Compile-time detect of null-exception? 702 if (stopped()) return; 703 704 Node* b = pop(); 705 Node* a = pop(); 706 707 const Type *t = _gvn.type(b); 708 if (t != Type::TOP) { 709 const TypeInt *ti = t->is_int(); 710 if (ti->is_con()) { 711 int divisor = ti->get_con(); 712 // check for positive power of 2 713 if (divisor > 0 && 714 (divisor & ~(divisor-1)) == divisor) { 715 // yes ! 716 Node *mask = _gvn.intcon((divisor - 1)); 717 // Sigh, must handle negative dividends 718 Node *zero = _gvn.intcon(0); 719 IfNode *ifff = jump_if_fork_int(a, zero, BoolTest::lt); 720 Node *iff = _gvn.transform( new (C) IfFalseNode(ifff) ); 721 Node *ift = _gvn.transform( new (C) IfTrueNode (ifff) ); 722 Node *reg = jump_if_join(ift, iff); 723 Node *phi = PhiNode::make(reg, NULL, TypeInt::INT); 724 // Negative path; negate/and/negate 725 Node *neg = _gvn.transform( new (C) SubINode(zero, a) ); 726 Node *andn= _gvn.transform( new (C) AndINode(neg, mask) ); 727 Node *negn= _gvn.transform( new (C) SubINode(zero, andn) ); 728 phi->init_req(1, negn); 729 // Fast positive case 730 Node *andx = _gvn.transform( new (C) AndINode(a, mask) ); 731 phi->init_req(2, andx); 732 // Push the merge 733 push( _gvn.transform(phi) ); 734 return; 735 } 736 } 737 } 738 // Default case 739 push( _gvn.transform( new (C) ModINode(control(),a,b) ) ); 740 } 741 742 // Handle jsr and jsr_w bytecode 743 void Parse::do_jsr() { 744 assert(bc() == Bytecodes::_jsr || bc() == Bytecodes::_jsr_w, "wrong bytecode"); 745 746 // Store information about current state, tagged with new _jsr_bci 747 int return_bci = iter().next_bci(); 748 int jsr_bci = (bc() == Bytecodes::_jsr) ? iter().get_dest() : iter().get_far_dest(); 749 750 // Update method data 751 profile_taken_branch(jsr_bci); 752 753 // The way we do things now, there is only one successor block 754 // for the jsr, because the target code is cloned by ciTypeFlow. 755 Block* target = successor_for_bci(jsr_bci); 756 757 // What got pushed? 758 const Type* ret_addr = target->peek(); 759 assert(ret_addr->singleton(), "must be a constant (cloned jsr body)"); 760 761 // Effect on jsr on stack 762 push(_gvn.makecon(ret_addr)); 763 764 // Flow to the jsr. 765 merge(jsr_bci); 766 } 767 768 // Handle ret bytecode 769 void Parse::do_ret() { 770 // Find to whom we return. 771 assert(block()->num_successors() == 1, "a ret can only go one place now"); 772 Block* target = block()->successor_at(0); 773 assert(!target->is_ready(), "our arrival must be expected"); 774 profile_ret(target->flow()->start()); 775 int pnum = target->next_path_num(); 776 merge_common(target, pnum); 777 } 778 779 static bool has_injected_profile(BoolTest::mask btest, Node* test, int& taken, int& not_taken) { 780 if (btest != BoolTest::eq && btest != BoolTest::ne) { 781 // Only ::eq and ::ne are supported for profile injection. 782 return false; 783 } 784 if (test->is_Cmp() && 785 test->in(1)->Opcode() == Op_ProfileBoolean) { 786 ProfileBooleanNode* profile = (ProfileBooleanNode*)test->in(1); 787 int false_cnt = profile->false_count(); 788 int true_cnt = profile->true_count(); 789 790 // Counts matching depends on the actual test operation (::eq or ::ne). 791 // No need to scale the counts because profile injection was designed 792 // to feed exact counts into VM. 793 taken = (btest == BoolTest::eq) ? false_cnt : true_cnt; 794 not_taken = (btest == BoolTest::eq) ? true_cnt : false_cnt; 795 796 profile->consume(); 797 return true; 798 } 799 return false; 800 } 801 //--------------------------dynamic_branch_prediction-------------------------- 802 // Try to gather dynamic branch prediction behavior. Return a probability 803 // of the branch being taken and set the "cnt" field. Returns a -1.0 804 // if we need to use static prediction for some reason. 805 float Parse::dynamic_branch_prediction(float &cnt, BoolTest::mask btest, Node* test) { 806 ResourceMark rm; 807 808 cnt = COUNT_UNKNOWN; 809 810 int taken = 0; 811 int not_taken = 0; 812 813 bool use_mdo = !has_injected_profile(btest, test, taken, not_taken); 814 815 if (use_mdo) { 816 // Use MethodData information if it is available 817 // FIXME: free the ProfileData structure 818 ciMethodData* methodData = method()->method_data(); 819 if (!methodData->is_mature()) return PROB_UNKNOWN; 820 ciProfileData* data = methodData->bci_to_data(bci()); 821 if (data == NULL) { 822 return PROB_UNKNOWN; 823 } 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='%g' 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 == NULL || 916 (data->as_BranchData()->taken() + 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 #ifndef PRODUCT 958 if (PrintOpto && WizardMode) { 959 tty->print("defending against excessive implicit null exceptions on %s @%d in ", 960 Bytecodes::name(iter().cur_bc()), iter().cur_bci()); 961 method()->print_name(); tty->cr(); 962 } 963 #endif 964 int bc_depth = - Bytecodes::depth(iter().cur_bc()); 965 assert(bc_depth == 1 || bc_depth == 2, "only two kinds of branches"); 966 DEBUG_ONLY(sync_jvms()); // argument(n) requires a synced jvms 967 assert(argument(0) != NULL, "must exist"); 968 assert(bc_depth == 1 || argument(1) != NULL, "two must exist"); 969 inc_sp(bc_depth); 970 return bc_depth; 971 } 972 973 //----------------------------------do_ifnull---------------------------------- 974 void Parse::do_ifnull(BoolTest::mask btest, Node *c) { 975 int target_bci = iter().get_dest(); 976 977 Block* branch_block = successor_for_bci(target_bci); 978 Block* next_block = successor_for_bci(iter().next_bci()); 979 980 float cnt; 981 float prob = branch_prediction(cnt, btest, target_bci, c); 982 if (prob == PROB_UNKNOWN) { 983 // (An earlier version of do_ifnull omitted this trap for OSR methods.) 984 #ifndef PRODUCT 985 if (PrintOpto && Verbose) 986 tty->print_cr("Never-taken edge stops compilation at bci %d",bci()); 987 #endif 988 repush_if_args(); // to gather stats on loop 989 // We need to mark this branch as taken so that if we recompile we will 990 // see that it is possible. In the tiered system the interpreter doesn't 991 // do profiling and by the time we get to the lower tier from the interpreter 992 // the path may be cold again. Make sure it doesn't look untaken 993 profile_taken_branch(target_bci, !ProfileInterpreter); 994 uncommon_trap(Deoptimization::Reason_unreached, 995 Deoptimization::Action_reinterpret, 996 NULL, "cold"); 997 if (C->eliminate_boxing()) { 998 // Mark the successor blocks as parsed 999 branch_block->next_path_num(); 1000 next_block->next_path_num(); 1001 } 1002 return; 1003 } 1004 1005 explicit_null_checks_inserted++; 1006 1007 // Generate real control flow 1008 Node *tst = _gvn.transform( new (C) BoolNode( c, btest ) ); 1009 1010 // Sanity check the probability value 1011 assert(prob > 0.0f,"Bad probability in Parser"); 1012 // Need xform to put node in hash table 1013 IfNode *iff = create_and_xform_if( control(), tst, prob, cnt ); 1014 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser"); 1015 // True branch 1016 { PreserveJVMState pjvms(this); 1017 Node* iftrue = _gvn.transform( new (C) IfTrueNode (iff) ); 1018 set_control(iftrue); 1019 1020 if (stopped()) { // Path is dead? 1021 explicit_null_checks_elided++; 1022 if (C->eliminate_boxing()) { 1023 // Mark the successor block as parsed 1024 branch_block->next_path_num(); 1025 } 1026 } else { // Path is live. 1027 // Update method data 1028 profile_taken_branch(target_bci); 1029 adjust_map_after_if(btest, c, prob, branch_block, next_block); 1030 if (!stopped()) { 1031 merge(target_bci); 1032 } 1033 } 1034 } 1035 1036 // False branch 1037 Node* iffalse = _gvn.transform( new (C) IfFalseNode(iff) ); 1038 set_control(iffalse); 1039 1040 if (stopped()) { // Path is dead? 1041 explicit_null_checks_elided++; 1042 if (C->eliminate_boxing()) { 1043 // Mark the successor block as parsed 1044 next_block->next_path_num(); 1045 } 1046 } else { // Path is live. 1047 // Update method data 1048 profile_not_taken_branch(); 1049 adjust_map_after_if(BoolTest(btest).negate(), c, 1.0-prob, 1050 next_block, branch_block); 1051 } 1052 } 1053 1054 //------------------------------------do_if------------------------------------ 1055 void Parse::do_if(BoolTest::mask btest, Node* c) { 1056 int target_bci = iter().get_dest(); 1057 1058 Block* branch_block = successor_for_bci(target_bci); 1059 Block* next_block = successor_for_bci(iter().next_bci()); 1060 1061 float cnt; 1062 float prob = branch_prediction(cnt, btest, target_bci, c); 1063 float untaken_prob = 1.0 - prob; 1064 1065 if (prob == PROB_UNKNOWN) { 1066 #ifndef PRODUCT 1067 if (PrintOpto && Verbose) 1068 tty->print_cr("Never-taken edge stops compilation at bci %d",bci()); 1069 #endif 1070 repush_if_args(); // to gather stats on loop 1071 // We need to mark this branch as taken so that if we recompile we will 1072 // see that it is possible. In the tiered system the interpreter doesn't 1073 // do profiling and by the time we get to the lower tier from the interpreter 1074 // the path may be cold again. Make sure it doesn't look untaken 1075 profile_taken_branch(target_bci, !ProfileInterpreter); 1076 uncommon_trap(Deoptimization::Reason_unreached, 1077 Deoptimization::Action_reinterpret, 1078 NULL, "cold"); 1079 if (C->eliminate_boxing()) { 1080 // Mark the successor blocks as parsed 1081 branch_block->next_path_num(); 1082 next_block->next_path_num(); 1083 } 1084 return; 1085 } 1086 1087 // Sanity check the probability value 1088 assert(0.0f < prob && prob < 1.0f,"Bad probability in Parser"); 1089 1090 bool taken_if_true = true; 1091 // Convert BoolTest to canonical form: 1092 if (!BoolTest(btest).is_canonical()) { 1093 btest = BoolTest(btest).negate(); 1094 taken_if_true = false; 1095 // prob is NOT updated here; it remains the probability of the taken 1096 // path (as opposed to the prob of the path guarded by an 'IfTrueNode'). 1097 } 1098 assert(btest != BoolTest::eq, "!= is the only canonical exact test"); 1099 1100 Node* tst0 = new (C) BoolNode(c, btest); 1101 Node* tst = _gvn.transform(tst0); 1102 BoolTest::mask taken_btest = BoolTest::illegal; 1103 BoolTest::mask untaken_btest = BoolTest::illegal; 1104 1105 if (tst->is_Bool()) { 1106 // Refresh c from the transformed bool node, since it may be 1107 // simpler than the original c. Also re-canonicalize btest. 1108 // This wins when (Bool ne (Conv2B p) 0) => (Bool ne (CmpP p NULL)). 1109 // That can arise from statements like: if (x instanceof C) ... 1110 if (tst != tst0) { 1111 // Canonicalize one more time since transform can change it. 1112 btest = tst->as_Bool()->_test._test; 1113 if (!BoolTest(btest).is_canonical()) { 1114 // Reverse edges one more time... 1115 tst = _gvn.transform( tst->as_Bool()->negate(&_gvn) ); 1116 btest = tst->as_Bool()->_test._test; 1117 assert(BoolTest(btest).is_canonical(), "sanity"); 1118 taken_if_true = !taken_if_true; 1119 } 1120 c = tst->in(1); 1121 } 1122 BoolTest::mask neg_btest = BoolTest(btest).negate(); 1123 taken_btest = taken_if_true ? btest : neg_btest; 1124 untaken_btest = taken_if_true ? neg_btest : btest; 1125 } 1126 1127 // Generate real control flow 1128 float true_prob = (taken_if_true ? prob : untaken_prob); 1129 IfNode* iff = create_and_map_if(control(), tst, true_prob, cnt); 1130 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser"); 1131 Node* taken_branch = new (C) IfTrueNode(iff); 1132 Node* untaken_branch = new (C) IfFalseNode(iff); 1133 if (!taken_if_true) { // Finish conversion to canonical form 1134 Node* tmp = taken_branch; 1135 taken_branch = untaken_branch; 1136 untaken_branch = tmp; 1137 } 1138 1139 // Branch is taken: 1140 { PreserveJVMState pjvms(this); 1141 taken_branch = _gvn.transform(taken_branch); 1142 set_control(taken_branch); 1143 1144 if (stopped()) { 1145 if (C->eliminate_boxing()) { 1146 // Mark the successor block as parsed 1147 branch_block->next_path_num(); 1148 } 1149 } else { 1150 // Update method data 1151 profile_taken_branch(target_bci); 1152 adjust_map_after_if(taken_btest, c, prob, branch_block, next_block); 1153 if (!stopped()) { 1154 merge(target_bci); 1155 } 1156 } 1157 } 1158 1159 untaken_branch = _gvn.transform(untaken_branch); 1160 set_control(untaken_branch); 1161 1162 // Branch not taken. 1163 if (stopped()) { 1164 if (C->eliminate_boxing()) { 1165 // Mark the successor block as parsed 1166 next_block->next_path_num(); 1167 } 1168 } else { 1169 // Update method data 1170 profile_not_taken_branch(); 1171 adjust_map_after_if(untaken_btest, c, untaken_prob, 1172 next_block, branch_block); 1173 } 1174 } 1175 1176 bool Parse::path_is_suitable_for_uncommon_trap(float prob) const { 1177 // Don't want to speculate on uncommon traps when running with -Xcomp 1178 if (!UseInterpreter) { 1179 return false; 1180 } 1181 return (seems_never_taken(prob) && seems_stable_comparison()); 1182 } 1183 1184 //----------------------------adjust_map_after_if------------------------------ 1185 // Adjust the JVM state to reflect the result of taking this path. 1186 // Basically, it means inspecting the CmpNode controlling this 1187 // branch, seeing how it constrains a tested value, and then 1188 // deciding if it's worth our while to encode this constraint 1189 // as graph nodes in the current abstract interpretation map. 1190 void Parse::adjust_map_after_if(BoolTest::mask btest, Node* c, float prob, 1191 Block* path, Block* other_path) { 1192 if (stopped() || !c->is_Cmp() || btest == BoolTest::illegal) 1193 return; // nothing to do 1194 1195 bool is_fallthrough = (path == successor_for_bci(iter().next_bci())); 1196 1197 if (path_is_suitable_for_uncommon_trap(prob)) { 1198 repush_if_args(); 1199 uncommon_trap(Deoptimization::Reason_unstable_if, 1200 Deoptimization::Action_reinterpret, 1201 NULL, 1202 (is_fallthrough ? "taken always" : "taken never")); 1203 return; 1204 } 1205 1206 Node* val = c->in(1); 1207 Node* con = c->in(2); 1208 const Type* tcon = _gvn.type(con); 1209 const Type* tval = _gvn.type(val); 1210 bool have_con = tcon->singleton(); 1211 if (tval->singleton()) { 1212 if (!have_con) { 1213 // Swap, so constant is in con. 1214 con = val; 1215 tcon = tval; 1216 val = c->in(2); 1217 tval = _gvn.type(val); 1218 btest = BoolTest(btest).commute(); 1219 have_con = true; 1220 } else { 1221 // Do we have two constants? Then leave well enough alone. 1222 have_con = false; 1223 } 1224 } 1225 if (!have_con) // remaining adjustments need a con 1226 return; 1227 1228 sharpen_type_after_if(btest, con, tcon, val, tval); 1229 } 1230 1231 1232 static Node* extract_obj_from_klass_load(PhaseGVN* gvn, Node* n) { 1233 Node* ldk; 1234 if (n->is_DecodeNKlass()) { 1235 if (n->in(1)->Opcode() != Op_LoadNKlass) { 1236 return NULL; 1237 } else { 1238 ldk = n->in(1); 1239 } 1240 } else if (n->Opcode() != Op_LoadKlass) { 1241 return NULL; 1242 } else { 1243 ldk = n; 1244 } 1245 assert(ldk != NULL && ldk->is_Load(), "should have found a LoadKlass or LoadNKlass node"); 1246 1247 Node* adr = ldk->in(MemNode::Address); 1248 intptr_t off = 0; 1249 Node* obj = AddPNode::Ideal_base_and_offset(adr, gvn, off); 1250 if (obj == NULL || off != oopDesc::klass_offset_in_bytes()) // loading oopDesc::_klass? 1251 return NULL; 1252 const TypePtr* tp = gvn->type(obj)->is_ptr(); 1253 if (tp == NULL || !(tp->isa_instptr() || tp->isa_aryptr())) // is obj a Java object ptr? 1254 return NULL; 1255 1256 return obj; 1257 } 1258 1259 void Parse::sharpen_type_after_if(BoolTest::mask btest, 1260 Node* con, const Type* tcon, 1261 Node* val, const Type* tval) { 1262 // Look for opportunities to sharpen the type of a node 1263 // whose klass is compared with a constant klass. 1264 if (btest == BoolTest::eq && tcon->isa_klassptr()) { 1265 Node* obj = extract_obj_from_klass_load(&_gvn, val); 1266 const TypeOopPtr* con_type = tcon->isa_klassptr()->as_instance_type(); 1267 if (obj != NULL && (con_type->isa_instptr() || con_type->isa_aryptr())) { 1268 // Found: 1269 // Bool(CmpP(LoadKlass(obj._klass), ConP(Foo.klass)), [eq]) 1270 // or the narrowOop equivalent. 1271 const Type* obj_type = _gvn.type(obj); 1272 const TypeOopPtr* tboth = obj_type->join_speculative(con_type)->isa_oopptr(); 1273 if (tboth != NULL && tboth->klass_is_exact() && tboth != obj_type && 1274 tboth->higher_equal(obj_type)) { 1275 // obj has to be of the exact type Foo if the CmpP succeeds. 1276 int obj_in_map = map()->find_edge(obj); 1277 JVMState* jvms = this->jvms(); 1278 if (obj_in_map >= 0 && 1279 (jvms->is_loc(obj_in_map) || jvms->is_stk(obj_in_map))) { 1280 TypeNode* ccast = new (C) CheckCastPPNode(control(), obj, tboth); 1281 const Type* tcc = ccast->as_Type()->type(); 1282 assert(tcc != obj_type && tcc->higher_equal_speculative(obj_type), "must improve"); 1283 // Delay transform() call to allow recovery of pre-cast value 1284 // at the control merge. 1285 _gvn.set_type_bottom(ccast); 1286 record_for_igvn(ccast); 1287 // Here's the payoff. 1288 replace_in_map(obj, ccast); 1289 } 1290 } 1291 } 1292 } 1293 1294 int val_in_map = map()->find_edge(val); 1295 if (val_in_map < 0) return; // replace_in_map would be useless 1296 { 1297 JVMState* jvms = this->jvms(); 1298 if (!(jvms->is_loc(val_in_map) || 1299 jvms->is_stk(val_in_map))) 1300 return; // again, it would be useless 1301 } 1302 1303 // Check for a comparison to a constant, and "know" that the compared 1304 // value is constrained on this path. 1305 assert(tcon->singleton(), ""); 1306 ConstraintCastNode* ccast = NULL; 1307 Node* cast = NULL; 1308 1309 switch (btest) { 1310 case BoolTest::eq: // Constant test? 1311 { 1312 const Type* tboth = tcon->join_speculative(tval); 1313 if (tboth == tval) break; // Nothing to gain. 1314 if (tcon->isa_int()) { 1315 ccast = new (C) CastIINode(val, tboth); 1316 } else if (tcon == TypePtr::NULL_PTR) { 1317 // Cast to null, but keep the pointer identity temporarily live. 1318 ccast = new (C) CastPPNode(val, tboth); 1319 } else { 1320 const TypeF* tf = tcon->isa_float_constant(); 1321 const TypeD* td = tcon->isa_double_constant(); 1322 // Exclude tests vs float/double 0 as these could be 1323 // either +0 or -0. Just because you are equal to +0 1324 // doesn't mean you ARE +0! 1325 // Note, following code also replaces Long and Oop values. 1326 if ((!tf || tf->_f != 0.0) && 1327 (!td || td->_d != 0.0)) 1328 cast = con; // Replace non-constant val by con. 1329 } 1330 } 1331 break; 1332 1333 case BoolTest::ne: 1334 if (tcon == TypePtr::NULL_PTR) { 1335 cast = cast_not_null(val, false); 1336 } 1337 break; 1338 1339 default: 1340 // (At this point we could record int range types with CastII.) 1341 break; 1342 } 1343 1344 if (ccast != NULL) { 1345 const Type* tcc = ccast->as_Type()->type(); 1346 assert(tcc != tval && tcc->higher_equal_speculative(tval), "must improve"); 1347 // Delay transform() call to allow recovery of pre-cast value 1348 // at the control merge. 1349 ccast->set_req(0, control()); 1350 _gvn.set_type_bottom(ccast); 1351 record_for_igvn(ccast); 1352 cast = ccast; 1353 } 1354 1355 if (cast != NULL) { // Here's the payoff. 1356 replace_in_map(val, cast); 1357 } 1358 } 1359 1360 /** 1361 * Use speculative type to optimize CmpP node: if comparison is 1362 * against the low level class, cast the object to the speculative 1363 * type if any. CmpP should then go away. 1364 * 1365 * @param c expected CmpP node 1366 * @return result of CmpP on object casted to speculative type 1367 * 1368 */ 1369 Node* Parse::optimize_cmp_with_klass(Node* c) { 1370 // If this is transformed by the _gvn to a comparison with the low 1371 // level klass then we may be able to use speculation 1372 if (c->Opcode() == Op_CmpP && 1373 (c->in(1)->Opcode() == Op_LoadKlass || c->in(1)->Opcode() == Op_DecodeNKlass) && 1374 c->in(2)->is_Con()) { 1375 Node* load_klass = NULL; 1376 Node* decode = NULL; 1377 if (c->in(1)->Opcode() == Op_DecodeNKlass) { 1378 decode = c->in(1); 1379 load_klass = c->in(1)->in(1); 1380 } else { 1381 load_klass = c->in(1); 1382 } 1383 if (load_klass->in(2)->is_AddP()) { 1384 Node* addp = load_klass->in(2); 1385 Node* obj = addp->in(AddPNode::Address); 1386 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 1387 if (obj_type->speculative_type() != NULL) { 1388 ciKlass* k = obj_type->speculative_type(); 1389 inc_sp(2); 1390 obj = maybe_cast_profiled_obj(obj, k); 1391 dec_sp(2); 1392 // Make the CmpP use the casted obj 1393 addp = basic_plus_adr(obj, addp->in(AddPNode::Offset)); 1394 load_klass = load_klass->clone(); 1395 load_klass->set_req(2, addp); 1396 load_klass = _gvn.transform(load_klass); 1397 if (decode != NULL) { 1398 decode = decode->clone(); 1399 decode->set_req(1, load_klass); 1400 load_klass = _gvn.transform(decode); 1401 } 1402 c = c->clone(); 1403 c->set_req(1, load_klass); 1404 c = _gvn.transform(c); 1405 } 1406 } 1407 } 1408 return c; 1409 } 1410 1411 //------------------------------do_one_bytecode-------------------------------- 1412 // Parse this bytecode, and alter the Parsers JVM->Node mapping 1413 void Parse::do_one_bytecode() { 1414 Node *a, *b, *c, *d; // Handy temps 1415 BoolTest::mask btest; 1416 int i; 1417 1418 assert(!has_exceptions(), "bytecode entry state must be clear of throws"); 1419 1420 if (C->check_node_count(NodeLimitFudgeFactor * 5, 1421 "out of nodes parsing method")) { 1422 return; 1423 } 1424 1425 #ifdef ASSERT 1426 // for setting breakpoints 1427 if (TraceOptoParse) { 1428 tty->print(" @"); 1429 dump_bci(bci()); 1430 tty->cr(); 1431 } 1432 #endif 1433 1434 switch (bc()) { 1435 case Bytecodes::_nop: 1436 // do nothing 1437 break; 1438 case Bytecodes::_lconst_0: 1439 push_pair(longcon(0)); 1440 break; 1441 1442 case Bytecodes::_lconst_1: 1443 push_pair(longcon(1)); 1444 break; 1445 1446 case Bytecodes::_fconst_0: 1447 push(zerocon(T_FLOAT)); 1448 break; 1449 1450 case Bytecodes::_fconst_1: 1451 push(makecon(TypeF::ONE)); 1452 break; 1453 1454 case Bytecodes::_fconst_2: 1455 push(makecon(TypeF::make(2.0f))); 1456 break; 1457 1458 case Bytecodes::_dconst_0: 1459 push_pair(zerocon(T_DOUBLE)); 1460 break; 1461 1462 case Bytecodes::_dconst_1: 1463 push_pair(makecon(TypeD::ONE)); 1464 break; 1465 1466 case Bytecodes::_iconst_m1:push(intcon(-1)); break; 1467 case Bytecodes::_iconst_0: push(intcon( 0)); break; 1468 case Bytecodes::_iconst_1: push(intcon( 1)); break; 1469 case Bytecodes::_iconst_2: push(intcon( 2)); break; 1470 case Bytecodes::_iconst_3: push(intcon( 3)); break; 1471 case Bytecodes::_iconst_4: push(intcon( 4)); break; 1472 case Bytecodes::_iconst_5: push(intcon( 5)); break; 1473 case Bytecodes::_bipush: push(intcon(iter().get_constant_u1())); break; 1474 case Bytecodes::_sipush: push(intcon(iter().get_constant_u2())); break; 1475 case Bytecodes::_aconst_null: push(null()); break; 1476 case Bytecodes::_ldc: 1477 case Bytecodes::_ldc_w: 1478 case Bytecodes::_ldc2_w: 1479 // If the constant is unresolved, run this BC once in the interpreter. 1480 { 1481 ciConstant constant = iter().get_constant(); 1482 if (constant.basic_type() == T_OBJECT && 1483 !constant.as_object()->is_loaded()) { 1484 int index = iter().get_constant_pool_index(); 1485 constantTag tag = iter().get_constant_pool_tag(index); 1486 uncommon_trap(Deoptimization::make_trap_request 1487 (Deoptimization::Reason_unloaded, 1488 Deoptimization::Action_reinterpret, 1489 index), 1490 NULL, tag.internal_name()); 1491 break; 1492 } 1493 assert(constant.basic_type() != T_OBJECT || constant.as_object()->is_instance(), 1494 "must be java_mirror of klass"); 1495 bool pushed = push_constant(constant, true); 1496 guarantee(pushed, "must be possible to push this constant"); 1497 } 1498 1499 break; 1500 1501 case Bytecodes::_aload_0: 1502 push( local(0) ); 1503 break; 1504 case Bytecodes::_aload_1: 1505 push( local(1) ); 1506 break; 1507 case Bytecodes::_aload_2: 1508 push( local(2) ); 1509 break; 1510 case Bytecodes::_aload_3: 1511 push( local(3) ); 1512 break; 1513 case Bytecodes::_aload: 1514 push( local(iter().get_index()) ); 1515 break; 1516 1517 case Bytecodes::_fload_0: 1518 case Bytecodes::_iload_0: 1519 push( local(0) ); 1520 break; 1521 case Bytecodes::_fload_1: 1522 case Bytecodes::_iload_1: 1523 push( local(1) ); 1524 break; 1525 case Bytecodes::_fload_2: 1526 case Bytecodes::_iload_2: 1527 push( local(2) ); 1528 break; 1529 case Bytecodes::_fload_3: 1530 case Bytecodes::_iload_3: 1531 push( local(3) ); 1532 break; 1533 case Bytecodes::_fload: 1534 case Bytecodes::_iload: 1535 push( local(iter().get_index()) ); 1536 break; 1537 case Bytecodes::_lload_0: 1538 push_pair_local( 0 ); 1539 break; 1540 case Bytecodes::_lload_1: 1541 push_pair_local( 1 ); 1542 break; 1543 case Bytecodes::_lload_2: 1544 push_pair_local( 2 ); 1545 break; 1546 case Bytecodes::_lload_3: 1547 push_pair_local( 3 ); 1548 break; 1549 case Bytecodes::_lload: 1550 push_pair_local( iter().get_index() ); 1551 break; 1552 1553 case Bytecodes::_dload_0: 1554 push_pair_local(0); 1555 break; 1556 case Bytecodes::_dload_1: 1557 push_pair_local(1); 1558 break; 1559 case Bytecodes::_dload_2: 1560 push_pair_local(2); 1561 break; 1562 case Bytecodes::_dload_3: 1563 push_pair_local(3); 1564 break; 1565 case Bytecodes::_dload: 1566 push_pair_local(iter().get_index()); 1567 break; 1568 case Bytecodes::_fstore_0: 1569 case Bytecodes::_istore_0: 1570 case Bytecodes::_astore_0: 1571 set_local( 0, pop() ); 1572 break; 1573 case Bytecodes::_fstore_1: 1574 case Bytecodes::_istore_1: 1575 case Bytecodes::_astore_1: 1576 set_local( 1, pop() ); 1577 break; 1578 case Bytecodes::_fstore_2: 1579 case Bytecodes::_istore_2: 1580 case Bytecodes::_astore_2: 1581 set_local( 2, pop() ); 1582 break; 1583 case Bytecodes::_fstore_3: 1584 case Bytecodes::_istore_3: 1585 case Bytecodes::_astore_3: 1586 set_local( 3, pop() ); 1587 break; 1588 case Bytecodes::_fstore: 1589 case Bytecodes::_istore: 1590 case Bytecodes::_astore: 1591 set_local( iter().get_index(), pop() ); 1592 break; 1593 // long stores 1594 case Bytecodes::_lstore_0: 1595 set_pair_local( 0, pop_pair() ); 1596 break; 1597 case Bytecodes::_lstore_1: 1598 set_pair_local( 1, pop_pair() ); 1599 break; 1600 case Bytecodes::_lstore_2: 1601 set_pair_local( 2, pop_pair() ); 1602 break; 1603 case Bytecodes::_lstore_3: 1604 set_pair_local( 3, pop_pair() ); 1605 break; 1606 case Bytecodes::_lstore: 1607 set_pair_local( iter().get_index(), pop_pair() ); 1608 break; 1609 1610 // double stores 1611 case Bytecodes::_dstore_0: 1612 set_pair_local( 0, dstore_rounding(pop_pair()) ); 1613 break; 1614 case Bytecodes::_dstore_1: 1615 set_pair_local( 1, dstore_rounding(pop_pair()) ); 1616 break; 1617 case Bytecodes::_dstore_2: 1618 set_pair_local( 2, dstore_rounding(pop_pair()) ); 1619 break; 1620 case Bytecodes::_dstore_3: 1621 set_pair_local( 3, dstore_rounding(pop_pair()) ); 1622 break; 1623 case Bytecodes::_dstore: 1624 set_pair_local( iter().get_index(), dstore_rounding(pop_pair()) ); 1625 break; 1626 1627 case Bytecodes::_pop: dec_sp(1); break; 1628 case Bytecodes::_pop2: dec_sp(2); break; 1629 case Bytecodes::_swap: 1630 a = pop(); 1631 b = pop(); 1632 push(a); 1633 push(b); 1634 break; 1635 case Bytecodes::_dup: 1636 a = pop(); 1637 push(a); 1638 push(a); 1639 break; 1640 case Bytecodes::_dup_x1: 1641 a = pop(); 1642 b = pop(); 1643 push( a ); 1644 push( b ); 1645 push( a ); 1646 break; 1647 case Bytecodes::_dup_x2: 1648 a = pop(); 1649 b = pop(); 1650 c = pop(); 1651 push( a ); 1652 push( c ); 1653 push( b ); 1654 push( a ); 1655 break; 1656 case Bytecodes::_dup2: 1657 a = pop(); 1658 b = pop(); 1659 push( b ); 1660 push( a ); 1661 push( b ); 1662 push( a ); 1663 break; 1664 1665 case Bytecodes::_dup2_x1: 1666 // before: .. c, b, a 1667 // after: .. b, a, c, b, a 1668 // not tested 1669 a = pop(); 1670 b = pop(); 1671 c = pop(); 1672 push( b ); 1673 push( a ); 1674 push( c ); 1675 push( b ); 1676 push( a ); 1677 break; 1678 case Bytecodes::_dup2_x2: 1679 // before: .. d, c, b, a 1680 // after: .. b, a, d, c, b, a 1681 // not tested 1682 a = pop(); 1683 b = pop(); 1684 c = pop(); 1685 d = pop(); 1686 push( b ); 1687 push( a ); 1688 push( d ); 1689 push( c ); 1690 push( b ); 1691 push( a ); 1692 break; 1693 1694 case Bytecodes::_arraylength: { 1695 // Must do null-check with value on expression stack 1696 Node *ary = null_check(peek(), T_ARRAY); 1697 // Compile-time detect of null-exception? 1698 if (stopped()) return; 1699 a = pop(); 1700 push(load_array_length(a)); 1701 break; 1702 } 1703 1704 case Bytecodes::_baload: array_load(T_BYTE); break; 1705 case Bytecodes::_caload: array_load(T_CHAR); break; 1706 case Bytecodes::_iaload: array_load(T_INT); break; 1707 case Bytecodes::_saload: array_load(T_SHORT); break; 1708 case Bytecodes::_faload: array_load(T_FLOAT); break; 1709 case Bytecodes::_aaload: array_load(T_OBJECT); break; 1710 case Bytecodes::_laload: { 1711 a = array_addressing(T_LONG, 0, false); 1712 if (stopped()) return; // guaranteed null or range check 1713 dec_sp(2); // Pop array and index 1714 push_pair(make_load(control(), a, TypeLong::LONG, T_LONG, TypeAryPtr::LONGS, MemNode::unordered)); 1715 break; 1716 } 1717 case Bytecodes::_daload: { 1718 a = array_addressing(T_DOUBLE, 0, false); 1719 if (stopped()) return; // guaranteed null or range check 1720 dec_sp(2); // Pop array and index 1721 push_pair(make_load(control(), a, Type::DOUBLE, T_DOUBLE, TypeAryPtr::DOUBLES, MemNode::unordered)); 1722 break; 1723 } 1724 case Bytecodes::_bastore: array_store(T_BYTE); break; 1725 case Bytecodes::_castore: array_store(T_CHAR); break; 1726 case Bytecodes::_iastore: array_store(T_INT); break; 1727 case Bytecodes::_sastore: array_store(T_SHORT); break; 1728 case Bytecodes::_fastore: array_store(T_FLOAT); break; 1729 case Bytecodes::_aastore: { 1730 d = array_addressing(T_OBJECT, 1, true); 1731 if (stopped()) return; // guaranteed null or range check 1732 array_store_check(); 1733 c = pop(); // Oop to store 1734 b = pop(); // index (already used) 1735 a = pop(); // the array itself 1736 const TypeOopPtr* elemtype = _gvn.type(a)->is_aryptr()->elem()->make_oopptr(); 1737 const TypeAryPtr* adr_type = TypeAryPtr::OOPS; 1738 // Note: We don't need a write barrier for Shenandoah on a here, because 1739 // a is not used except for an assert. The address d already has the 1740 // write barrier. Adding a barrier on a only results in additional code 1741 // being generated. 1742 c = shenandoah_read_barrier_storeval(c); 1743 Node* store = store_oop_to_array(control(), a, d, adr_type, c, elemtype, T_OBJECT, MemNode::release); 1744 break; 1745 } 1746 case Bytecodes::_lastore: { 1747 a = array_addressing(T_LONG, 2, true); 1748 if (stopped()) return; // guaranteed null or range check 1749 c = pop_pair(); 1750 dec_sp(2); // Pop array and index 1751 store_to_memory(control(), a, c, T_LONG, TypeAryPtr::LONGS, MemNode::unordered); 1752 break; 1753 } 1754 case Bytecodes::_dastore: { 1755 a = array_addressing(T_DOUBLE, 2, true); 1756 if (stopped()) return; // guaranteed null or range check 1757 c = pop_pair(); 1758 dec_sp(2); // Pop array and index 1759 c = dstore_rounding(c); 1760 store_to_memory(control(), a, c, T_DOUBLE, TypeAryPtr::DOUBLES, MemNode::unordered); 1761 break; 1762 } 1763 case Bytecodes::_getfield: 1764 do_getfield(); 1765 break; 1766 1767 case Bytecodes::_getstatic: 1768 do_getstatic(); 1769 break; 1770 1771 case Bytecodes::_putfield: 1772 do_putfield(); 1773 break; 1774 1775 case Bytecodes::_putstatic: 1776 do_putstatic(); 1777 break; 1778 1779 case Bytecodes::_irem: 1780 do_irem(); 1781 break; 1782 case Bytecodes::_idiv: 1783 // Must keep both values on the expression-stack during null-check 1784 zero_check_int(peek()); 1785 // Compile-time detect of null-exception? 1786 if (stopped()) return; 1787 b = pop(); 1788 a = pop(); 1789 push( _gvn.transform( new (C) DivINode(control(),a,b) ) ); 1790 break; 1791 case Bytecodes::_imul: 1792 b = pop(); a = pop(); 1793 push( _gvn.transform( new (C) MulINode(a,b) ) ); 1794 break; 1795 case Bytecodes::_iadd: 1796 b = pop(); a = pop(); 1797 push( _gvn.transform( new (C) AddINode(a,b) ) ); 1798 break; 1799 case Bytecodes::_ineg: 1800 a = pop(); 1801 push( _gvn.transform( new (C) SubINode(_gvn.intcon(0),a)) ); 1802 break; 1803 case Bytecodes::_isub: 1804 b = pop(); a = pop(); 1805 push( _gvn.transform( new (C) SubINode(a,b) ) ); 1806 break; 1807 case Bytecodes::_iand: 1808 b = pop(); a = pop(); 1809 push( _gvn.transform( new (C) AndINode(a,b) ) ); 1810 break; 1811 case Bytecodes::_ior: 1812 b = pop(); a = pop(); 1813 push( _gvn.transform( new (C) OrINode(a,b) ) ); 1814 break; 1815 case Bytecodes::_ixor: 1816 b = pop(); a = pop(); 1817 push( _gvn.transform( new (C) XorINode(a,b) ) ); 1818 break; 1819 case Bytecodes::_ishl: 1820 b = pop(); a = pop(); 1821 push( _gvn.transform( new (C) LShiftINode(a,b) ) ); 1822 break; 1823 case Bytecodes::_ishr: 1824 b = pop(); a = pop(); 1825 push( _gvn.transform( new (C) RShiftINode(a,b) ) ); 1826 break; 1827 case Bytecodes::_iushr: 1828 b = pop(); a = pop(); 1829 push( _gvn.transform( new (C) URShiftINode(a,b) ) ); 1830 break; 1831 1832 case Bytecodes::_fneg: 1833 a = pop(); 1834 b = _gvn.transform(new (C) NegFNode (a)); 1835 push(b); 1836 break; 1837 1838 case Bytecodes::_fsub: 1839 b = pop(); 1840 a = pop(); 1841 c = _gvn.transform( new (C) SubFNode(a,b) ); 1842 d = precision_rounding(c); 1843 push( d ); 1844 break; 1845 1846 case Bytecodes::_fadd: 1847 b = pop(); 1848 a = pop(); 1849 c = _gvn.transform( new (C) AddFNode(a,b) ); 1850 d = precision_rounding(c); 1851 push( d ); 1852 break; 1853 1854 case Bytecodes::_fmul: 1855 b = pop(); 1856 a = pop(); 1857 c = _gvn.transform( new (C) MulFNode(a,b) ); 1858 d = precision_rounding(c); 1859 push( d ); 1860 break; 1861 1862 case Bytecodes::_fdiv: 1863 b = pop(); 1864 a = pop(); 1865 c = _gvn.transform( new (C) DivFNode(0,a,b) ); 1866 d = precision_rounding(c); 1867 push( d ); 1868 break; 1869 1870 case Bytecodes::_frem: 1871 if (Matcher::has_match_rule(Op_ModF)) { 1872 // Generate a ModF node. 1873 b = pop(); 1874 a = pop(); 1875 c = _gvn.transform( new (C) ModFNode(0,a,b) ); 1876 d = precision_rounding(c); 1877 push( d ); 1878 } 1879 else { 1880 // Generate a call. 1881 modf(); 1882 } 1883 break; 1884 1885 case Bytecodes::_fcmpl: 1886 b = pop(); 1887 a = pop(); 1888 c = _gvn.transform( new (C) CmpF3Node( a, b)); 1889 push(c); 1890 break; 1891 case Bytecodes::_fcmpg: 1892 b = pop(); 1893 a = pop(); 1894 1895 // Same as fcmpl but need to flip the unordered case. Swap the inputs, 1896 // which negates the result sign except for unordered. Flip the unordered 1897 // as well by using CmpF3 which implements unordered-lesser instead of 1898 // unordered-greater semantics. Finally, commute the result bits. Result 1899 // is same as using a CmpF3Greater except we did it with CmpF3 alone. 1900 c = _gvn.transform( new (C) CmpF3Node( b, a)); 1901 c = _gvn.transform( new (C) SubINode(_gvn.intcon(0),c) ); 1902 push(c); 1903 break; 1904 1905 case Bytecodes::_f2i: 1906 a = pop(); 1907 push(_gvn.transform(new (C) ConvF2INode(a))); 1908 break; 1909 1910 case Bytecodes::_d2i: 1911 a = pop_pair(); 1912 b = _gvn.transform(new (C) ConvD2INode(a)); 1913 push( b ); 1914 break; 1915 1916 case Bytecodes::_f2d: 1917 a = pop(); 1918 b = _gvn.transform( new (C) ConvF2DNode(a)); 1919 push_pair( b ); 1920 break; 1921 1922 case Bytecodes::_d2f: 1923 a = pop_pair(); 1924 b = _gvn.transform( new (C) ConvD2FNode(a)); 1925 // This breaks _227_mtrt (speed & correctness) and _222_mpegaudio (speed) 1926 //b = _gvn.transform(new (C) RoundFloatNode(0, b) ); 1927 push( b ); 1928 break; 1929 1930 case Bytecodes::_l2f: 1931 if (Matcher::convL2FSupported()) { 1932 a = pop_pair(); 1933 b = _gvn.transform( new (C) ConvL2FNode(a)); 1934 // For i486.ad, FILD doesn't restrict precision to 24 or 53 bits. 1935 // Rather than storing the result into an FP register then pushing 1936 // out to memory to round, the machine instruction that implements 1937 // ConvL2D is responsible for rounding. 1938 // c = precision_rounding(b); 1939 c = _gvn.transform(b); 1940 push(c); 1941 } else { 1942 l2f(); 1943 } 1944 break; 1945 1946 case Bytecodes::_l2d: 1947 a = pop_pair(); 1948 b = _gvn.transform( new (C) ConvL2DNode(a)); 1949 // For i486.ad, rounding is always necessary (see _l2f above). 1950 // c = dprecision_rounding(b); 1951 c = _gvn.transform(b); 1952 push_pair(c); 1953 break; 1954 1955 case Bytecodes::_f2l: 1956 a = pop(); 1957 b = _gvn.transform( new (C) ConvF2LNode(a)); 1958 push_pair(b); 1959 break; 1960 1961 case Bytecodes::_d2l: 1962 a = pop_pair(); 1963 b = _gvn.transform( new (C) ConvD2LNode(a)); 1964 push_pair(b); 1965 break; 1966 1967 case Bytecodes::_dsub: 1968 b = pop_pair(); 1969 a = pop_pair(); 1970 c = _gvn.transform( new (C) SubDNode(a,b) ); 1971 d = dprecision_rounding(c); 1972 push_pair( d ); 1973 break; 1974 1975 case Bytecodes::_dadd: 1976 b = pop_pair(); 1977 a = pop_pair(); 1978 c = _gvn.transform( new (C) AddDNode(a,b) ); 1979 d = dprecision_rounding(c); 1980 push_pair( d ); 1981 break; 1982 1983 case Bytecodes::_dmul: 1984 b = pop_pair(); 1985 a = pop_pair(); 1986 c = _gvn.transform( new (C) MulDNode(a,b) ); 1987 d = dprecision_rounding(c); 1988 push_pair( d ); 1989 break; 1990 1991 case Bytecodes::_ddiv: 1992 b = pop_pair(); 1993 a = pop_pair(); 1994 c = _gvn.transform( new (C) DivDNode(0,a,b) ); 1995 d = dprecision_rounding(c); 1996 push_pair( d ); 1997 break; 1998 1999 case Bytecodes::_dneg: 2000 a = pop_pair(); 2001 b = _gvn.transform(new (C) NegDNode (a)); 2002 push_pair(b); 2003 break; 2004 2005 case Bytecodes::_drem: 2006 if (Matcher::has_match_rule(Op_ModD)) { 2007 // Generate a ModD node. 2008 b = pop_pair(); 2009 a = pop_pair(); 2010 // a % b 2011 2012 c = _gvn.transform( new (C) ModDNode(0,a,b) ); 2013 d = dprecision_rounding(c); 2014 push_pair( d ); 2015 } 2016 else { 2017 // Generate a call. 2018 modd(); 2019 } 2020 break; 2021 2022 case Bytecodes::_dcmpl: 2023 b = pop_pair(); 2024 a = pop_pair(); 2025 c = _gvn.transform( new (C) CmpD3Node( a, b)); 2026 push(c); 2027 break; 2028 2029 case Bytecodes::_dcmpg: 2030 b = pop_pair(); 2031 a = pop_pair(); 2032 // Same as dcmpl but need to flip the unordered case. 2033 // Commute the inputs, which negates the result sign except for unordered. 2034 // Flip the unordered as well by using CmpD3 which implements 2035 // unordered-lesser instead of unordered-greater semantics. 2036 // Finally, negate the result bits. Result is same as using a 2037 // CmpD3Greater except we did it with CmpD3 alone. 2038 c = _gvn.transform( new (C) CmpD3Node( b, a)); 2039 c = _gvn.transform( new (C) SubINode(_gvn.intcon(0),c) ); 2040 push(c); 2041 break; 2042 2043 2044 // Note for longs -> lo word is on TOS, hi word is on TOS - 1 2045 case Bytecodes::_land: 2046 b = pop_pair(); 2047 a = pop_pair(); 2048 c = _gvn.transform( new (C) AndLNode(a,b) ); 2049 push_pair(c); 2050 break; 2051 case Bytecodes::_lor: 2052 b = pop_pair(); 2053 a = pop_pair(); 2054 c = _gvn.transform( new (C) OrLNode(a,b) ); 2055 push_pair(c); 2056 break; 2057 case Bytecodes::_lxor: 2058 b = pop_pair(); 2059 a = pop_pair(); 2060 c = _gvn.transform( new (C) XorLNode(a,b) ); 2061 push_pair(c); 2062 break; 2063 2064 case Bytecodes::_lshl: 2065 b = pop(); // the shift count 2066 a = pop_pair(); // value to be shifted 2067 c = _gvn.transform( new (C) LShiftLNode(a,b) ); 2068 push_pair(c); 2069 break; 2070 case Bytecodes::_lshr: 2071 b = pop(); // the shift count 2072 a = pop_pair(); // value to be shifted 2073 c = _gvn.transform( new (C) RShiftLNode(a,b) ); 2074 push_pair(c); 2075 break; 2076 case Bytecodes::_lushr: 2077 b = pop(); // the shift count 2078 a = pop_pair(); // value to be shifted 2079 c = _gvn.transform( new (C) URShiftLNode(a,b) ); 2080 push_pair(c); 2081 break; 2082 case Bytecodes::_lmul: 2083 b = pop_pair(); 2084 a = pop_pair(); 2085 c = _gvn.transform( new (C) MulLNode(a,b) ); 2086 push_pair(c); 2087 break; 2088 2089 case Bytecodes::_lrem: 2090 // Must keep both values on the expression-stack during null-check 2091 assert(peek(0) == top(), "long word order"); 2092 zero_check_long(peek(1)); 2093 // Compile-time detect of null-exception? 2094 if (stopped()) return; 2095 b = pop_pair(); 2096 a = pop_pair(); 2097 c = _gvn.transform( new (C) ModLNode(control(),a,b) ); 2098 push_pair(c); 2099 break; 2100 2101 case Bytecodes::_ldiv: 2102 // Must keep both values on the expression-stack during null-check 2103 assert(peek(0) == top(), "long word order"); 2104 zero_check_long(peek(1)); 2105 // Compile-time detect of null-exception? 2106 if (stopped()) return; 2107 b = pop_pair(); 2108 a = pop_pair(); 2109 c = _gvn.transform( new (C) DivLNode(control(),a,b) ); 2110 push_pair(c); 2111 break; 2112 2113 case Bytecodes::_ladd: 2114 b = pop_pair(); 2115 a = pop_pair(); 2116 c = _gvn.transform( new (C) AddLNode(a,b) ); 2117 push_pair(c); 2118 break; 2119 case Bytecodes::_lsub: 2120 b = pop_pair(); 2121 a = pop_pair(); 2122 c = _gvn.transform( new (C) SubLNode(a,b) ); 2123 push_pair(c); 2124 break; 2125 case Bytecodes::_lcmp: 2126 // Safepoints are now inserted _before_ branches. The long-compare 2127 // bytecode painfully produces a 3-way value (-1,0,+1) which requires a 2128 // slew of control flow. These are usually followed by a CmpI vs zero and 2129 // a branch; this pattern then optimizes to the obvious long-compare and 2130 // branch. However, if the branch is backwards there's a Safepoint 2131 // inserted. The inserted Safepoint captures the JVM state at the 2132 // pre-branch point, i.e. it captures the 3-way value. Thus if a 2133 // long-compare is used to control a loop the debug info will force 2134 // computation of the 3-way value, even though the generated code uses a 2135 // long-compare and branch. We try to rectify the situation by inserting 2136 // a SafePoint here and have it dominate and kill the safepoint added at a 2137 // following backwards branch. At this point the JVM state merely holds 2 2138 // longs but not the 3-way value. 2139 if( UseLoopSafepoints ) { 2140 switch( iter().next_bc() ) { 2141 case Bytecodes::_ifgt: 2142 case Bytecodes::_iflt: 2143 case Bytecodes::_ifge: 2144 case Bytecodes::_ifle: 2145 case Bytecodes::_ifne: 2146 case Bytecodes::_ifeq: 2147 // If this is a backwards branch in the bytecodes, add Safepoint 2148 maybe_add_safepoint(iter().next_get_dest()); 2149 } 2150 } 2151 b = pop_pair(); 2152 a = pop_pair(); 2153 c = _gvn.transform( new (C) CmpL3Node( a, b )); 2154 push(c); 2155 break; 2156 2157 case Bytecodes::_lneg: 2158 a = pop_pair(); 2159 b = _gvn.transform( new (C) SubLNode(longcon(0),a)); 2160 push_pair(b); 2161 break; 2162 case Bytecodes::_l2i: 2163 a = pop_pair(); 2164 push( _gvn.transform( new (C) ConvL2INode(a))); 2165 break; 2166 case Bytecodes::_i2l: 2167 a = pop(); 2168 b = _gvn.transform( new (C) ConvI2LNode(a)); 2169 push_pair(b); 2170 break; 2171 case Bytecodes::_i2b: 2172 // Sign extend 2173 a = pop(); 2174 a = _gvn.transform( new (C) LShiftINode(a,_gvn.intcon(24)) ); 2175 a = _gvn.transform( new (C) RShiftINode(a,_gvn.intcon(24)) ); 2176 push( a ); 2177 break; 2178 case Bytecodes::_i2s: 2179 a = pop(); 2180 a = _gvn.transform( new (C) LShiftINode(a,_gvn.intcon(16)) ); 2181 a = _gvn.transform( new (C) RShiftINode(a,_gvn.intcon(16)) ); 2182 push( a ); 2183 break; 2184 case Bytecodes::_i2c: 2185 a = pop(); 2186 push( _gvn.transform( new (C) AndINode(a,_gvn.intcon(0xFFFF)) ) ); 2187 break; 2188 2189 case Bytecodes::_i2f: 2190 a = pop(); 2191 b = _gvn.transform( new (C) ConvI2FNode(a) ) ; 2192 c = precision_rounding(b); 2193 push (b); 2194 break; 2195 2196 case Bytecodes::_i2d: 2197 a = pop(); 2198 b = _gvn.transform( new (C) ConvI2DNode(a)); 2199 push_pair(b); 2200 break; 2201 2202 case Bytecodes::_iinc: // Increment local 2203 i = iter().get_index(); // Get local index 2204 set_local( i, _gvn.transform( new (C) AddINode( _gvn.intcon(iter().get_iinc_con()), local(i) ) ) ); 2205 break; 2206 2207 // Exit points of synchronized methods must have an unlock node 2208 case Bytecodes::_return: 2209 return_current(NULL); 2210 break; 2211 2212 case Bytecodes::_ireturn: 2213 case Bytecodes::_areturn: 2214 case Bytecodes::_freturn: 2215 return_current(pop()); 2216 break; 2217 case Bytecodes::_lreturn: 2218 return_current(pop_pair()); 2219 break; 2220 case Bytecodes::_dreturn: 2221 return_current(pop_pair()); 2222 break; 2223 2224 case Bytecodes::_athrow: 2225 // null exception oop throws NULL pointer exception 2226 null_check(peek()); 2227 if (stopped()) return; 2228 // Hook the thrown exception directly to subsequent handlers. 2229 if (BailoutToInterpreterForThrows) { 2230 // Keep method interpreted from now on. 2231 uncommon_trap(Deoptimization::Reason_unhandled, 2232 Deoptimization::Action_make_not_compilable); 2233 return; 2234 } 2235 if (env()->jvmti_can_post_on_exceptions()) { 2236 // check if we must post exception events, take uncommon trap if so (with must_throw = false) 2237 uncommon_trap_if_should_post_on_exceptions(Deoptimization::Reason_unhandled, false); 2238 } 2239 // Here if either can_post_on_exceptions or should_post_on_exceptions is false 2240 add_exception_state(make_exception_state(peek())); 2241 break; 2242 2243 case Bytecodes::_goto: // fall through 2244 case Bytecodes::_goto_w: { 2245 int target_bci = (bc() == Bytecodes::_goto) ? iter().get_dest() : iter().get_far_dest(); 2246 2247 // If this is a backwards branch in the bytecodes, add Safepoint 2248 maybe_add_safepoint(target_bci); 2249 2250 // Update method data 2251 profile_taken_branch(target_bci); 2252 2253 // Merge the current control into the target basic block 2254 merge(target_bci); 2255 2256 // See if we can get some profile data and hand it off to the next block 2257 Block *target_block = block()->successor_for_bci(target_bci); 2258 if (target_block->pred_count() != 1) break; 2259 ciMethodData* methodData = method()->method_data(); 2260 if (!methodData->is_mature()) break; 2261 ciProfileData* data = methodData->bci_to_data(bci()); 2262 assert( data->is_JumpData(), "" ); 2263 int taken = ((ciJumpData*)data)->taken(); 2264 taken = method()->scale_count(taken); 2265 target_block->set_count(taken); 2266 break; 2267 } 2268 2269 case Bytecodes::_ifnull: btest = BoolTest::eq; goto handle_if_null; 2270 case Bytecodes::_ifnonnull: btest = BoolTest::ne; goto handle_if_null; 2271 handle_if_null: 2272 // If this is a backwards branch in the bytecodes, add Safepoint 2273 maybe_add_safepoint(iter().get_dest()); 2274 a = null(); 2275 b = pop(); 2276 c = _gvn.transform( new (C) CmpPNode(b, a) ); 2277 do_ifnull(btest, c); 2278 break; 2279 2280 case Bytecodes::_if_acmpeq: btest = BoolTest::eq; goto handle_if_acmp; 2281 case Bytecodes::_if_acmpne: btest = BoolTest::ne; goto handle_if_acmp; 2282 handle_if_acmp: 2283 // If this is a backwards branch in the bytecodes, add Safepoint 2284 maybe_add_safepoint(iter().get_dest()); 2285 a = pop(); 2286 b = pop(); 2287 if (UseShenandoahGC && ShenandoahAcmpBarrier) { 2288 a = shenandoah_write_barrier(a); 2289 b = shenandoah_write_barrier(b); 2290 } 2291 c = _gvn.transform(new (C) 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 (C) 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 (C) 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) { 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 }