1 /* 2 * Copyright (c) 1998, 2013, 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, &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 Node* adr = array_addressing(elem_type, 1); 61 if (stopped()) return; // guaranteed null or range check 62 Node* val = pop(); 63 dec_sp(2); // Pop array and index 64 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(elem_type); 65 store_to_memory(control(), adr, val, elem_type, adr_type, StoreNode::release_if_reference(elem_type)); 66 } 67 68 69 //------------------------------array_addressing------------------------------- 70 // Pull array and index from the stack. Compute pointer-to-element. 71 Node* Parse::array_addressing(BasicType type, int vals, const Type* *result2) { 72 Node *idx = peek(0+vals); // Get from stack without popping 73 Node *ary = peek(1+vals); // in case of exception 74 75 // Null check the array base, with correct stack contents 76 ary = null_check(ary, T_ARRAY); 77 // Compile-time detect of null-exception? 78 if (stopped()) return top(); 79 80 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr(); 81 const TypeInt* sizetype = arytype->size(); 82 const Type* elemtype = arytype->elem(); 83 84 if (UseUniqueSubclasses && result2 != NULL) { 85 const Type* el = elemtype->make_ptr(); 86 if (el && el->isa_instptr()) { 87 const TypeInstPtr* toop = el->is_instptr(); 88 if (toop->klass()->as_instance_klass()->unique_concrete_subklass()) { 89 // If we load from "AbstractClass[]" we must see "ConcreteSubClass". 90 const Type* subklass = Type::get_const_type(toop->klass()); 91 elemtype = subklass->join_speculative(el); 92 } 93 } 94 } 95 96 // Check for big class initializers with all constant offsets 97 // feeding into a known-size array. 98 const TypeInt* idxtype = _gvn.type(idx)->is_int(); 99 // See if the highest idx value is less than the lowest array bound, 100 // and if the idx value cannot be negative: 101 bool need_range_check = true; 102 if (idxtype->_hi < sizetype->_lo && idxtype->_lo >= 0) { 103 need_range_check = false; 104 if (C->log() != NULL) C->log()->elem("observe that='!need_range_check'"); 105 } 106 107 ciKlass * arytype_klass = arytype->klass(); 108 if ((arytype_klass != NULL) && (!arytype_klass->is_loaded())) { 109 // Only fails for some -Xcomp runs 110 // The class is unloaded. We have to run this bytecode in the interpreter. 111 uncommon_trap(Deoptimization::Reason_unloaded, 112 Deoptimization::Action_reinterpret, 113 arytype->klass(), "!loaded array"); 114 return top(); 115 } 116 117 // Do the range check 118 if (GenerateRangeChecks && need_range_check) { 119 Node* tst; 120 if (sizetype->_hi <= 0) { 121 // The greatest array bound is negative, so we can conclude that we're 122 // compiling unreachable code, but the unsigned compare trick used below 123 // only works with non-negative lengths. Instead, hack "tst" to be zero so 124 // the uncommon_trap path will always be taken. 125 tst = _gvn.intcon(0); 126 } else { 127 // Range is constant in array-oop, so we can use the original state of mem 128 Node* len = load_array_length(ary); 129 130 // Test length vs index (standard trick using unsigned compare) 131 Node* chk = _gvn.transform( new (C) CmpUNode(idx, len) ); 132 BoolTest::mask btest = BoolTest::lt; 133 tst = _gvn.transform( new (C) BoolNode(chk, btest) ); 134 } 135 // Branch to failure if out of bounds 136 { BuildCutout unless(this, tst, PROB_MAX); 137 if (C->allow_range_check_smearing()) { 138 // Do not use builtin_throw, since range checks are sometimes 139 // made more stringent by an optimistic transformation. 140 // This creates "tentative" range checks at this point, 141 // which are not guaranteed to throw exceptions. 142 // See IfNode::Ideal, is_range_check, adjust_check. 143 uncommon_trap(Deoptimization::Reason_range_check, 144 Deoptimization::Action_make_not_entrant, 145 NULL, "range_check"); 146 } else { 147 // If we have already recompiled with the range-check-widening 148 // heroic optimization turned off, then we must really be throwing 149 // range check exceptions. 150 builtin_throw(Deoptimization::Reason_range_check, idx); 151 } 152 } 153 } 154 // Check for always knowing you are throwing a range-check exception 155 if (stopped()) return top(); 156 157 Node* ptr = array_element_address(ary, idx, type, sizetype); 158 159 if (result2 != NULL) *result2 = elemtype; 160 161 assert(ptr != top(), "top should go hand-in-hand with stopped"); 162 163 return ptr; 164 } 165 166 167 // returns IfNode 168 IfNode* Parse::jump_if_fork_int(Node* a, Node* b, BoolTest::mask mask) { 169 Node *cmp = _gvn.transform( new (C) CmpINode( a, b)); // two cases: shiftcount > 32 and shiftcount <= 32 170 Node *tst = _gvn.transform( new (C) BoolNode( cmp, mask)); 171 IfNode *iff = create_and_map_if( control(), tst, ((mask == BoolTest::eq) ? PROB_STATIC_INFREQUENT : PROB_FAIR), COUNT_UNKNOWN ); 172 return iff; 173 } 174 175 // return Region node 176 Node* Parse::jump_if_join(Node* iffalse, Node* iftrue) { 177 Node *region = new (C) RegionNode(3); // 2 results 178 record_for_igvn(region); 179 region->init_req(1, iffalse); 180 region->init_req(2, iftrue ); 181 _gvn.set_type(region, Type::CONTROL); 182 region = _gvn.transform(region); 183 set_control (region); 184 return region; 185 } 186 187 188 //------------------------------helper for tableswitch------------------------- 189 void Parse::jump_if_true_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index) { 190 // True branch, use existing map info 191 { PreserveJVMState pjvms(this); 192 Node *iftrue = _gvn.transform( new (C) IfTrueNode (iff) ); 193 set_control( iftrue ); 194 profile_switch_case(prof_table_index); 195 merge_new_path(dest_bci_if_true); 196 } 197 198 // False branch 199 Node *iffalse = _gvn.transform( new (C) IfFalseNode(iff) ); 200 set_control( iffalse ); 201 } 202 203 void Parse::jump_if_false_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index) { 204 // True branch, use existing map info 205 { PreserveJVMState pjvms(this); 206 Node *iffalse = _gvn.transform( new (C) IfFalseNode (iff) ); 207 set_control( iffalse ); 208 profile_switch_case(prof_table_index); 209 merge_new_path(dest_bci_if_true); 210 } 211 212 // False branch 213 Node *iftrue = _gvn.transform( new (C) IfTrueNode(iff) ); 214 set_control( iftrue ); 215 } 216 217 void Parse::jump_if_always_fork(int dest_bci, int prof_table_index) { 218 // False branch, use existing map and control() 219 profile_switch_case(prof_table_index); 220 merge_new_path(dest_bci); 221 } 222 223 224 extern "C" { 225 static int jint_cmp(const void *i, const void *j) { 226 int a = *(jint *)i; 227 int b = *(jint *)j; 228 return a > b ? 1 : a < b ? -1 : 0; 229 } 230 } 231 232 233 // Default value for methodData switch indexing. Must be a negative value to avoid 234 // conflict with any legal switch index. 235 #define NullTableIndex -1 236 237 class SwitchRange : public StackObj { 238 // a range of integers coupled with a bci destination 239 jint _lo; // inclusive lower limit 240 jint _hi; // inclusive upper limit 241 int _dest; 242 int _table_index; // index into method data table 243 244 public: 245 jint lo() const { return _lo; } 246 jint hi() const { return _hi; } 247 int dest() const { return _dest; } 248 int table_index() const { return _table_index; } 249 bool is_singleton() const { return _lo == _hi; } 250 251 void setRange(jint lo, jint hi, int dest, int table_index) { 252 assert(lo <= hi, "must be a non-empty range"); 253 _lo = lo, _hi = hi; _dest = dest; _table_index = table_index; 254 } 255 bool adjoinRange(jint lo, jint hi, int dest, int table_index) { 256 assert(lo <= hi, "must be a non-empty range"); 257 if (lo == _hi+1 && dest == _dest && table_index == _table_index) { 258 _hi = hi; 259 return true; 260 } 261 return false; 262 } 263 264 void set (jint value, int dest, int table_index) { 265 setRange(value, value, dest, table_index); 266 } 267 bool adjoin(jint value, int dest, int table_index) { 268 return adjoinRange(value, value, dest, table_index); 269 } 270 271 void print() { 272 if (is_singleton()) 273 tty->print(" {%d}=>%d", lo(), dest()); 274 else if (lo() == min_jint) 275 tty->print(" {..%d}=>%d", hi(), dest()); 276 else if (hi() == max_jint) 277 tty->print(" {%d..}=>%d", lo(), dest()); 278 else 279 tty->print(" {%d..%d}=>%d", lo(), hi(), dest()); 280 } 281 }; 282 283 284 //-------------------------------do_tableswitch-------------------------------- 285 void Parse::do_tableswitch() { 286 Node* lookup = pop(); 287 288 // Get information about tableswitch 289 int default_dest = iter().get_dest_table(0); 290 int lo_index = iter().get_int_table(1); 291 int hi_index = iter().get_int_table(2); 292 int len = hi_index - lo_index + 1; 293 294 if (len < 1) { 295 // If this is a backward branch, add safepoint 296 maybe_add_safepoint(default_dest); 297 merge(default_dest); 298 return; 299 } 300 301 // generate decision tree, using trichotomy when possible 302 int rnum = len+2; 303 bool makes_backward_branch = false; 304 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum); 305 int rp = -1; 306 if (lo_index != min_jint) { 307 ranges[++rp].setRange(min_jint, lo_index-1, default_dest, NullTableIndex); 308 } 309 for (int j = 0; j < len; j++) { 310 jint match_int = lo_index+j; 311 int dest = iter().get_dest_table(j+3); 312 makes_backward_branch |= (dest <= bci()); 313 int table_index = method_data_update() ? j : NullTableIndex; 314 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index)) { 315 ranges[++rp].set(match_int, dest, table_index); 316 } 317 } 318 jint highest = lo_index+(len-1); 319 assert(ranges[rp].hi() == highest, ""); 320 if (highest != max_jint 321 && !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex)) { 322 ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex); 323 } 324 assert(rp < len+2, "not too many ranges"); 325 326 // Safepoint in case if backward branch observed 327 if( makes_backward_branch && UseLoopSafepoints ) 328 add_safepoint(); 329 330 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]); 331 } 332 333 334 //------------------------------do_lookupswitch-------------------------------- 335 void Parse::do_lookupswitch() { 336 Node *lookup = pop(); // lookup value 337 // Get information about lookupswitch 338 int default_dest = iter().get_dest_table(0); 339 int len = iter().get_int_table(1); 340 341 if (len < 1) { // If this is a backward branch, add safepoint 342 maybe_add_safepoint(default_dest); 343 merge(default_dest); 344 return; 345 } 346 347 // generate decision tree, using trichotomy when possible 348 jint* table = NEW_RESOURCE_ARRAY(jint, len*2); 349 { 350 for( int j = 0; j < len; j++ ) { 351 table[j+j+0] = iter().get_int_table(2+j+j); 352 table[j+j+1] = iter().get_dest_table(2+j+j+1); 353 } 354 qsort( table, len, 2*sizeof(table[0]), jint_cmp ); 355 } 356 357 int rnum = len*2+1; 358 bool makes_backward_branch = false; 359 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum); 360 int rp = -1; 361 for( int j = 0; j < len; j++ ) { 362 jint match_int = table[j+j+0]; 363 int dest = table[j+j+1]; 364 int next_lo = rp < 0 ? min_jint : ranges[rp].hi()+1; 365 int table_index = method_data_update() ? j : NullTableIndex; 366 makes_backward_branch |= (dest <= bci()); 367 if( match_int != next_lo ) { 368 ranges[++rp].setRange(next_lo, match_int-1, default_dest, NullTableIndex); 369 } 370 if( rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index) ) { 371 ranges[++rp].set(match_int, dest, table_index); 372 } 373 } 374 jint highest = table[2*(len-1)]; 375 assert(ranges[rp].hi() == highest, ""); 376 if( highest != max_jint 377 && !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex) ) { 378 ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex); 379 } 380 assert(rp < rnum, "not too many ranges"); 381 382 // Safepoint in case backward branch observed 383 if( makes_backward_branch && UseLoopSafepoints ) 384 add_safepoint(); 385 386 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]); 387 } 388 389 //----------------------------create_jump_tables------------------------------- 390 bool Parse::create_jump_tables(Node* key_val, SwitchRange* lo, SwitchRange* hi) { 391 // Are jumptables enabled 392 if (!UseJumpTables) return false; 393 394 // Are jumptables supported 395 if (!Matcher::has_match_rule(Op_Jump)) return false; 396 397 // Don't make jump table if profiling 398 if (method_data_update()) return false; 399 400 // Decide if a guard is needed to lop off big ranges at either (or 401 // both) end(s) of the input set. We'll call this the default target 402 // even though we can't be sure that it is the true "default". 403 404 bool needs_guard = false; 405 int default_dest; 406 int64 total_outlier_size = 0; 407 int64 hi_size = ((int64)hi->hi()) - ((int64)hi->lo()) + 1; 408 int64 lo_size = ((int64)lo->hi()) - ((int64)lo->lo()) + 1; 409 410 if (lo->dest() == hi->dest()) { 411 total_outlier_size = hi_size + lo_size; 412 default_dest = lo->dest(); 413 } else if (lo_size > hi_size) { 414 total_outlier_size = lo_size; 415 default_dest = lo->dest(); 416 } else { 417 total_outlier_size = hi_size; 418 default_dest = hi->dest(); 419 } 420 421 // If a guard test will eliminate very sparse end ranges, then 422 // it is worth the cost of an extra jump. 423 if (total_outlier_size > (MaxJumpTableSparseness * 4)) { 424 needs_guard = true; 425 if (default_dest == lo->dest()) lo++; 426 if (default_dest == hi->dest()) hi--; 427 } 428 429 // Find the total number of cases and ranges 430 int64 num_cases = ((int64)hi->hi()) - ((int64)lo->lo()) + 1; 431 int num_range = hi - lo + 1; 432 433 // Don't create table if: too large, too small, or too sparse. 434 if (num_cases < MinJumpTableSize || num_cases > MaxJumpTableSize) 435 return false; 436 if (num_cases > (MaxJumpTableSparseness * num_range)) 437 return false; 438 439 // Normalize table lookups to zero 440 int lowval = lo->lo(); 441 key_val = _gvn.transform( new (C) SubINode(key_val, _gvn.intcon(lowval)) ); 442 443 // Generate a guard to protect against input keyvals that aren't 444 // in the switch domain. 445 if (needs_guard) { 446 Node* size = _gvn.intcon(num_cases); 447 Node* cmp = _gvn.transform( new (C) CmpUNode(key_val, size) ); 448 Node* tst = _gvn.transform( new (C) BoolNode(cmp, BoolTest::ge) ); 449 IfNode* iff = create_and_map_if( control(), tst, PROB_FAIR, COUNT_UNKNOWN); 450 jump_if_true_fork(iff, default_dest, NullTableIndex); 451 } 452 453 // Create an ideal node JumpTable that has projections 454 // of all possible ranges for a switch statement 455 // The key_val input must be converted to a pointer offset and scaled. 456 // Compare Parse::array_addressing above. 457 #ifdef _LP64 458 // Clean the 32-bit int into a real 64-bit offset. 459 // Otherwise, the jint value 0 might turn into an offset of 0x0800000000. 460 const TypeLong* lkeytype = TypeLong::make(CONST64(0), num_cases-1, Type::WidenMin); 461 key_val = _gvn.transform( new (C) ConvI2LNode(key_val, lkeytype) ); 462 #endif 463 // Shift the value by wordsize so we have an index into the table, rather 464 // than a switch value 465 Node *shiftWord = _gvn.MakeConX(wordSize); 466 key_val = _gvn.transform( new (C) MulXNode( key_val, shiftWord)); 467 468 // Create the JumpNode 469 Node* jtn = _gvn.transform( new (C) JumpNode(control(), key_val, num_cases) ); 470 471 // These are the switch destinations hanging off the jumpnode 472 int i = 0; 473 for (SwitchRange* r = lo; r <= hi; r++) { 474 for (int64 j = r->lo(); j <= r->hi(); j++, i++) { 475 Node* input = _gvn.transform(new (C) JumpProjNode(jtn, i, r->dest(), (int)(j - lowval))); 476 { 477 PreserveJVMState pjvms(this); 478 set_control(input); 479 jump_if_always_fork(r->dest(), r->table_index()); 480 } 481 } 482 } 483 assert(i == num_cases, "miscount of cases"); 484 stop_and_kill_map(); // no more uses for this JVMS 485 return true; 486 } 487 488 //----------------------------jump_switch_ranges------------------------------- 489 void Parse::jump_switch_ranges(Node* key_val, SwitchRange *lo, SwitchRange *hi, int switch_depth) { 490 Block* switch_block = block(); 491 492 if (switch_depth == 0) { 493 // Do special processing for the top-level call. 494 assert(lo->lo() == min_jint, "initial range must exhaust Type::INT"); 495 assert(hi->hi() == max_jint, "initial range must exhaust Type::INT"); 496 497 // Decrement pred-numbers for the unique set of nodes. 498 #ifdef ASSERT 499 // Ensure that the block's successors are a (duplicate-free) set. 500 int successors_counted = 0; // block occurrences in [hi..lo] 501 int unique_successors = switch_block->num_successors(); 502 for (int i = 0; i < unique_successors; i++) { 503 Block* target = switch_block->successor_at(i); 504 505 // Check that the set of successors is the same in both places. 506 int successors_found = 0; 507 for (SwitchRange* p = lo; p <= hi; p++) { 508 if (p->dest() == target->start()) successors_found++; 509 } 510 assert(successors_found > 0, "successor must be known"); 511 successors_counted += successors_found; 512 } 513 assert(successors_counted == (hi-lo)+1, "no unexpected successors"); 514 #endif 515 516 // Maybe prune the inputs, based on the type of key_val. 517 jint min_val = min_jint; 518 jint max_val = max_jint; 519 const TypeInt* ti = key_val->bottom_type()->isa_int(); 520 if (ti != NULL) { 521 min_val = ti->_lo; 522 max_val = ti->_hi; 523 assert(min_val <= max_val, "invalid int type"); 524 } 525 while (lo->hi() < min_val) lo++; 526 if (lo->lo() < min_val) lo->setRange(min_val, lo->hi(), lo->dest(), lo->table_index()); 527 while (hi->lo() > max_val) hi--; 528 if (hi->hi() > max_val) hi->setRange(hi->lo(), max_val, hi->dest(), hi->table_index()); 529 } 530 531 #ifndef PRODUCT 532 if (switch_depth == 0) { 533 _max_switch_depth = 0; 534 _est_switch_depth = log2_intptr((hi-lo+1)-1)+1; 535 } 536 #endif 537 538 assert(lo <= hi, "must be a non-empty set of ranges"); 539 if (lo == hi) { 540 jump_if_always_fork(lo->dest(), lo->table_index()); 541 } else { 542 assert(lo->hi() == (lo+1)->lo()-1, "contiguous ranges"); 543 assert(hi->lo() == (hi-1)->hi()+1, "contiguous ranges"); 544 545 if (create_jump_tables(key_val, lo, hi)) return; 546 547 int nr = hi - lo + 1; 548 549 SwitchRange* mid = lo + nr/2; 550 // if there is an easy choice, pivot at a singleton: 551 if (nr > 3 && !mid->is_singleton() && (mid-1)->is_singleton()) mid--; 552 553 assert(lo < mid && mid <= hi, "good pivot choice"); 554 assert(nr != 2 || mid == hi, "should pick higher of 2"); 555 assert(nr != 3 || mid == hi-1, "should pick middle of 3"); 556 557 Node *test_val = _gvn.intcon(mid->lo()); 558 559 if (mid->is_singleton()) { 560 IfNode *iff_ne = jump_if_fork_int(key_val, test_val, BoolTest::ne); 561 jump_if_false_fork(iff_ne, mid->dest(), mid->table_index()); 562 563 // Special Case: If there are exactly three ranges, and the high 564 // and low range each go to the same place, omit the "gt" test, 565 // since it will not discriminate anything. 566 bool eq_test_only = (hi == lo+2 && hi->dest() == lo->dest()); 567 if (eq_test_only) { 568 assert(mid == hi-1, ""); 569 } 570 571 // if there is a higher range, test for it and process it: 572 if (mid < hi && !eq_test_only) { 573 // two comparisons of same values--should enable 1 test for 2 branches 574 // Use BoolTest::le instead of BoolTest::gt 575 IfNode *iff_le = jump_if_fork_int(key_val, test_val, BoolTest::le); 576 Node *iftrue = _gvn.transform( new (C) IfTrueNode(iff_le) ); 577 Node *iffalse = _gvn.transform( new (C) IfFalseNode(iff_le) ); 578 { PreserveJVMState pjvms(this); 579 set_control(iffalse); 580 jump_switch_ranges(key_val, mid+1, hi, switch_depth+1); 581 } 582 set_control(iftrue); 583 } 584 585 } else { 586 // mid is a range, not a singleton, so treat mid..hi as a unit 587 IfNode *iff_ge = jump_if_fork_int(key_val, test_val, BoolTest::ge); 588 589 // if there is a higher range, test for it and process it: 590 if (mid == hi) { 591 jump_if_true_fork(iff_ge, mid->dest(), mid->table_index()); 592 } else { 593 Node *iftrue = _gvn.transform( new (C) IfTrueNode(iff_ge) ); 594 Node *iffalse = _gvn.transform( new (C) IfFalseNode(iff_ge) ); 595 { PreserveJVMState pjvms(this); 596 set_control(iftrue); 597 jump_switch_ranges(key_val, mid, hi, switch_depth+1); 598 } 599 set_control(iffalse); 600 } 601 } 602 603 // in any case, process the lower range 604 jump_switch_ranges(key_val, lo, mid-1, switch_depth+1); 605 } 606 607 // Decrease pred_count for each successor after all is done. 608 if (switch_depth == 0) { 609 int unique_successors = switch_block->num_successors(); 610 for (int i = 0; i < unique_successors; i++) { 611 Block* target = switch_block->successor_at(i); 612 // Throw away the pre-allocated path for each unique successor. 613 target->next_path_num(); 614 } 615 } 616 617 #ifndef PRODUCT 618 _max_switch_depth = MAX2(switch_depth, _max_switch_depth); 619 if (TraceOptoParse && Verbose && WizardMode && switch_depth == 0) { 620 SwitchRange* r; 621 int nsing = 0; 622 for( r = lo; r <= hi; r++ ) { 623 if( r->is_singleton() ) nsing++; 624 } 625 tty->print(">>> "); 626 _method->print_short_name(); 627 tty->print_cr(" switch decision tree"); 628 tty->print_cr(" %d ranges (%d singletons), max_depth=%d, est_depth=%d", 629 hi-lo+1, nsing, _max_switch_depth, _est_switch_depth); 630 if (_max_switch_depth > _est_switch_depth) { 631 tty->print_cr("******** BAD SWITCH DEPTH ********"); 632 } 633 tty->print(" "); 634 for( r = lo; r <= hi; r++ ) { 635 r->print(); 636 } 637 tty->print_cr(""); 638 } 639 #endif 640 } 641 642 void Parse::modf() { 643 Node *f2 = pop(); 644 Node *f1 = pop(); 645 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::modf_Type(), 646 CAST_FROM_FN_PTR(address, SharedRuntime::frem), 647 "frem", NULL, //no memory effects 648 f1, f2); 649 Node* res = _gvn.transform(new (C) ProjNode(c, TypeFunc::Parms + 0)); 650 651 push(res); 652 } 653 654 void Parse::modd() { 655 Node *d2 = pop_pair(); 656 Node *d1 = pop_pair(); 657 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::Math_DD_D_Type(), 658 CAST_FROM_FN_PTR(address, SharedRuntime::drem), 659 "drem", NULL, //no memory effects 660 d1, top(), d2, top()); 661 Node* res_d = _gvn.transform(new (C) ProjNode(c, TypeFunc::Parms + 0)); 662 663 #ifdef ASSERT 664 Node* res_top = _gvn.transform(new (C) ProjNode(c, TypeFunc::Parms + 1)); 665 assert(res_top == top(), "second value must be top"); 666 #endif 667 668 push_pair(res_d); 669 } 670 671 void Parse::l2f() { 672 Node* f2 = pop(); 673 Node* f1 = pop(); 674 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::l2f_Type(), 675 CAST_FROM_FN_PTR(address, SharedRuntime::l2f), 676 "l2f", NULL, //no memory effects 677 f1, f2); 678 Node* res = _gvn.transform(new (C) ProjNode(c, TypeFunc::Parms + 0)); 679 680 push(res); 681 } 682 683 void Parse::do_irem() { 684 // Must keep both values on the expression-stack during null-check 685 zero_check_int(peek()); 686 // Compile-time detect of null-exception? 687 if (stopped()) return; 688 689 Node* b = pop(); 690 Node* a = pop(); 691 692 const Type *t = _gvn.type(b); 693 if (t != Type::TOP) { 694 const TypeInt *ti = t->is_int(); 695 if (ti->is_con()) { 696 int divisor = ti->get_con(); 697 // check for positive power of 2 698 if (divisor > 0 && 699 (divisor & ~(divisor-1)) == divisor) { 700 // yes ! 701 Node *mask = _gvn.intcon((divisor - 1)); 702 // Sigh, must handle negative dividends 703 Node *zero = _gvn.intcon(0); 704 IfNode *ifff = jump_if_fork_int(a, zero, BoolTest::lt); 705 Node *iff = _gvn.transform( new (C) IfFalseNode(ifff) ); 706 Node *ift = _gvn.transform( new (C) IfTrueNode (ifff) ); 707 Node *reg = jump_if_join(ift, iff); 708 Node *phi = PhiNode::make(reg, NULL, TypeInt::INT); 709 // Negative path; negate/and/negate 710 Node *neg = _gvn.transform( new (C) SubINode(zero, a) ); 711 Node *andn= _gvn.transform( new (C) AndINode(neg, mask) ); 712 Node *negn= _gvn.transform( new (C) SubINode(zero, andn) ); 713 phi->init_req(1, negn); 714 // Fast positive case 715 Node *andx = _gvn.transform( new (C) AndINode(a, mask) ); 716 phi->init_req(2, andx); 717 // Push the merge 718 push( _gvn.transform(phi) ); 719 return; 720 } 721 } 722 } 723 // Default case 724 push( _gvn.transform( new (C) ModINode(control(),a,b) ) ); 725 } 726 727 // Handle jsr and jsr_w bytecode 728 void Parse::do_jsr() { 729 assert(bc() == Bytecodes::_jsr || bc() == Bytecodes::_jsr_w, "wrong bytecode"); 730 731 // Store information about current state, tagged with new _jsr_bci 732 int return_bci = iter().next_bci(); 733 int jsr_bci = (bc() == Bytecodes::_jsr) ? iter().get_dest() : iter().get_far_dest(); 734 735 // Update method data 736 profile_taken_branch(jsr_bci); 737 738 // The way we do things now, there is only one successor block 739 // for the jsr, because the target code is cloned by ciTypeFlow. 740 Block* target = successor_for_bci(jsr_bci); 741 742 // What got pushed? 743 const Type* ret_addr = target->peek(); 744 assert(ret_addr->singleton(), "must be a constant (cloned jsr body)"); 745 746 // Effect on jsr on stack 747 push(_gvn.makecon(ret_addr)); 748 749 // Flow to the jsr. 750 merge(jsr_bci); 751 } 752 753 // Handle ret bytecode 754 void Parse::do_ret() { 755 // Find to whom we return. 756 assert(block()->num_successors() == 1, "a ret can only go one place now"); 757 Block* target = block()->successor_at(0); 758 assert(!target->is_ready(), "our arrival must be expected"); 759 profile_ret(target->flow()->start()); 760 int pnum = target->next_path_num(); 761 merge_common(target, pnum); 762 } 763 764 //--------------------------dynamic_branch_prediction-------------------------- 765 // Try to gather dynamic branch prediction behavior. Return a probability 766 // of the branch being taken and set the "cnt" field. Returns a -1.0 767 // if we need to use static prediction for some reason. 768 float Parse::dynamic_branch_prediction(float &cnt) { 769 ResourceMark rm; 770 771 cnt = COUNT_UNKNOWN; 772 773 // Use MethodData information if it is available 774 // FIXME: free the ProfileData structure 775 ciMethodData* methodData = method()->method_data(); 776 if (!methodData->is_mature()) return PROB_UNKNOWN; 777 ciProfileData* data = methodData->bci_to_data(bci()); 778 if (!data->is_JumpData()) return PROB_UNKNOWN; 779 780 // get taken and not taken values 781 int taken = data->as_JumpData()->taken(); 782 int not_taken = 0; 783 if (data->is_BranchData()) { 784 not_taken = data->as_BranchData()->not_taken(); 785 } 786 787 // scale the counts to be commensurate with invocation counts: 788 taken = method()->scale_count(taken); 789 not_taken = method()->scale_count(not_taken); 790 791 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful. 792 // We also check that individual counters are positive first, overwise the sum can become positive. 793 if (taken < 0 || not_taken < 0 || taken + not_taken < 40) { 794 if (C->log() != NULL) { 795 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d'", iter().get_dest(), taken, not_taken); 796 } 797 return PROB_UNKNOWN; 798 } 799 800 // Compute frequency that we arrive here 801 float sum = taken + not_taken; 802 // Adjust, if this block is a cloned private block but the 803 // Jump counts are shared. Taken the private counts for 804 // just this path instead of the shared counts. 805 if( block()->count() > 0 ) 806 sum = block()->count(); 807 cnt = sum / FreqCountInvocations; 808 809 // Pin probability to sane limits 810 float prob; 811 if( !taken ) 812 prob = (0+PROB_MIN) / 2; 813 else if( !not_taken ) 814 prob = (1+PROB_MAX) / 2; 815 else { // Compute probability of true path 816 prob = (float)taken / (float)(taken + not_taken); 817 if (prob > PROB_MAX) prob = PROB_MAX; 818 if (prob < PROB_MIN) prob = PROB_MIN; 819 } 820 821 assert((cnt > 0.0f) && (prob > 0.0f), 822 "Bad frequency assignment in if"); 823 824 if (C->log() != NULL) { 825 const char* prob_str = NULL; 826 if (prob >= PROB_MAX) prob_str = (prob == PROB_MAX) ? "max" : "always"; 827 if (prob <= PROB_MIN) prob_str = (prob == PROB_MIN) ? "min" : "never"; 828 char prob_str_buf[30]; 829 if (prob_str == NULL) { 830 sprintf(prob_str_buf, "%g", prob); 831 prob_str = prob_str_buf; 832 } 833 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d' cnt='%g' prob='%s'", 834 iter().get_dest(), taken, not_taken, cnt, prob_str); 835 } 836 return prob; 837 } 838 839 //-----------------------------branch_prediction------------------------------- 840 float Parse::branch_prediction(float& cnt, 841 BoolTest::mask btest, 842 int target_bci) { 843 float prob = dynamic_branch_prediction(cnt); 844 // If prob is unknown, switch to static prediction 845 if (prob != PROB_UNKNOWN) return prob; 846 847 prob = PROB_FAIR; // Set default value 848 if (btest == BoolTest::eq) // Exactly equal test? 849 prob = PROB_STATIC_INFREQUENT; // Assume its relatively infrequent 850 else if (btest == BoolTest::ne) 851 prob = PROB_STATIC_FREQUENT; // Assume its relatively frequent 852 853 // If this is a conditional test guarding a backwards branch, 854 // assume its a loop-back edge. Make it a likely taken branch. 855 if (target_bci < bci()) { 856 if (is_osr_parse()) { // Could be a hot OSR'd loop; force deopt 857 // Since it's an OSR, we probably have profile data, but since 858 // branch_prediction returned PROB_UNKNOWN, the counts are too small. 859 // Let's make a special check here for completely zero counts. 860 ciMethodData* methodData = method()->method_data(); 861 if (!methodData->is_empty()) { 862 ciProfileData* data = methodData->bci_to_data(bci()); 863 // Only stop for truly zero counts, which mean an unknown part 864 // of the OSR-ed method, and we want to deopt to gather more stats. 865 // If you have ANY counts, then this loop is simply 'cold' relative 866 // to the OSR loop. 867 if (data->as_BranchData()->taken() + 868 data->as_BranchData()->not_taken() == 0 ) { 869 // This is the only way to return PROB_UNKNOWN: 870 return PROB_UNKNOWN; 871 } 872 } 873 } 874 prob = PROB_STATIC_FREQUENT; // Likely to take backwards branch 875 } 876 877 assert(prob != PROB_UNKNOWN, "must have some guess at this point"); 878 return prob; 879 } 880 881 // The magic constants are chosen so as to match the output of 882 // branch_prediction() when the profile reports a zero taken count. 883 // It is important to distinguish zero counts unambiguously, because 884 // some branches (e.g., _213_javac.Assembler.eliminate) validly produce 885 // very small but nonzero probabilities, which if confused with zero 886 // counts would keep the program recompiling indefinitely. 887 bool Parse::seems_never_taken(float prob) { 888 return prob < PROB_MIN; 889 } 890 891 // True if the comparison seems to be the kind that will not change its 892 // statistics from true to false. See comments in adjust_map_after_if. 893 // This question is only asked along paths which are already 894 // classifed as untaken (by seems_never_taken), so really, 895 // if a path is never taken, its controlling comparison is 896 // already acting in a stable fashion. If the comparison 897 // seems stable, we will put an expensive uncommon trap 898 // on the untaken path. To be conservative, and to allow 899 // partially executed counted loops to be compiled fully, 900 // we will plant uncommon traps only after pointer comparisons. 901 bool Parse::seems_stable_comparison(BoolTest::mask btest, Node* cmp) { 902 for (int depth = 4; depth > 0; depth--) { 903 // The following switch can find CmpP here over half the time for 904 // dynamic language code rich with type tests. 905 // Code using counted loops or array manipulations (typical 906 // of benchmarks) will have many (>80%) CmpI instructions. 907 switch (cmp->Opcode()) { 908 case Op_CmpP: 909 // A never-taken null check looks like CmpP/BoolTest::eq. 910 // These certainly should be closed off as uncommon traps. 911 if (btest == BoolTest::eq) 912 return true; 913 // A never-failed type check looks like CmpP/BoolTest::ne. 914 // Let's put traps on those, too, so that we don't have to compile 915 // unused paths with indeterminate dynamic type information. 916 if (ProfileDynamicTypes) 917 return true; 918 return false; 919 920 case Op_CmpI: 921 // A small minority (< 10%) of CmpP are masked as CmpI, 922 // as if by boolean conversion ((p == q? 1: 0) != 0). 923 // Detect that here, even if it hasn't optimized away yet. 924 // Specifically, this covers the 'instanceof' operator. 925 if (btest == BoolTest::ne || btest == BoolTest::eq) { 926 if (_gvn.type(cmp->in(2))->singleton() && 927 cmp->in(1)->is_Phi()) { 928 PhiNode* phi = cmp->in(1)->as_Phi(); 929 int true_path = phi->is_diamond_phi(); 930 if (true_path > 0 && 931 _gvn.type(phi->in(1))->singleton() && 932 _gvn.type(phi->in(2))->singleton()) { 933 // phi->region->if_proj->ifnode->bool->cmp 934 BoolNode* bol = phi->in(0)->in(1)->in(0)->in(1)->as_Bool(); 935 btest = bol->_test._test; 936 cmp = bol->in(1); 937 continue; 938 } 939 } 940 } 941 return false; 942 } 943 } 944 return false; 945 } 946 947 //-------------------------------repush_if_args-------------------------------- 948 // Push arguments of an "if" bytecode back onto the stack by adjusting _sp. 949 inline int Parse::repush_if_args() { 950 #ifndef PRODUCT 951 if (PrintOpto && WizardMode) { 952 tty->print("defending against excessive implicit null exceptions on %s @%d in ", 953 Bytecodes::name(iter().cur_bc()), iter().cur_bci()); 954 method()->print_name(); tty->cr(); 955 } 956 #endif 957 int bc_depth = - Bytecodes::depth(iter().cur_bc()); 958 assert(bc_depth == 1 || bc_depth == 2, "only two kinds of branches"); 959 DEBUG_ONLY(sync_jvms()); // argument(n) requires a synced jvms 960 assert(argument(0) != NULL, "must exist"); 961 assert(bc_depth == 1 || argument(1) != NULL, "two must exist"); 962 inc_sp(bc_depth); 963 return bc_depth; 964 } 965 966 //----------------------------------do_ifnull---------------------------------- 967 void Parse::do_ifnull(BoolTest::mask btest, Node *c) { 968 int target_bci = iter().get_dest(); 969 970 Block* branch_block = successor_for_bci(target_bci); 971 Block* next_block = successor_for_bci(iter().next_bci()); 972 973 float cnt; 974 float prob = branch_prediction(cnt, btest, target_bci); 975 if (prob == PROB_UNKNOWN) { 976 // (An earlier version of do_ifnull omitted this trap for OSR methods.) 977 #ifndef PRODUCT 978 if (PrintOpto && Verbose) 979 tty->print_cr("Never-taken edge stops compilation at bci %d",bci()); 980 #endif 981 repush_if_args(); // to gather stats on loop 982 // We need to mark this branch as taken so that if we recompile we will 983 // see that it is possible. In the tiered system the interpreter doesn't 984 // do profiling and by the time we get to the lower tier from the interpreter 985 // the path may be cold again. Make sure it doesn't look untaken 986 profile_taken_branch(target_bci, !ProfileInterpreter); 987 uncommon_trap(Deoptimization::Reason_unreached, 988 Deoptimization::Action_reinterpret, 989 NULL, "cold"); 990 if (C->eliminate_boxing()) { 991 // Mark the successor blocks as parsed 992 branch_block->next_path_num(); 993 next_block->next_path_num(); 994 } 995 return; 996 } 997 998 explicit_null_checks_inserted++; 999 1000 // Generate real control flow 1001 Node *tst = _gvn.transform( new (C) BoolNode( c, btest ) ); 1002 1003 // Sanity check the probability value 1004 assert(prob > 0.0f,"Bad probability in Parser"); 1005 // Need xform to put node in hash table 1006 IfNode *iff = create_and_xform_if( control(), tst, prob, cnt ); 1007 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser"); 1008 // True branch 1009 { PreserveJVMState pjvms(this); 1010 Node* iftrue = _gvn.transform( new (C) IfTrueNode (iff) ); 1011 set_control(iftrue); 1012 1013 if (stopped()) { // Path is dead? 1014 explicit_null_checks_elided++; 1015 if (C->eliminate_boxing()) { 1016 // Mark the successor block as parsed 1017 branch_block->next_path_num(); 1018 } 1019 } else { // Path is live. 1020 // Update method data 1021 profile_taken_branch(target_bci); 1022 adjust_map_after_if(btest, c, prob, branch_block, next_block); 1023 if (!stopped()) { 1024 merge(target_bci); 1025 } 1026 } 1027 } 1028 1029 // False branch 1030 Node* iffalse = _gvn.transform( new (C) IfFalseNode(iff) ); 1031 set_control(iffalse); 1032 1033 if (stopped()) { // Path is dead? 1034 explicit_null_checks_elided++; 1035 if (C->eliminate_boxing()) { 1036 // Mark the successor block as parsed 1037 next_block->next_path_num(); 1038 } 1039 } else { // Path is live. 1040 // Update method data 1041 profile_not_taken_branch(); 1042 adjust_map_after_if(BoolTest(btest).negate(), c, 1.0-prob, 1043 next_block, branch_block); 1044 } 1045 } 1046 1047 //------------------------------------do_if------------------------------------ 1048 void Parse::do_if(BoolTest::mask btest, Node* c) { 1049 int target_bci = iter().get_dest(); 1050 1051 Block* branch_block = successor_for_bci(target_bci); 1052 Block* next_block = successor_for_bci(iter().next_bci()); 1053 1054 float cnt; 1055 float prob = branch_prediction(cnt, btest, target_bci); 1056 float untaken_prob = 1.0 - prob; 1057 1058 if (prob == PROB_UNKNOWN) { 1059 #ifndef PRODUCT 1060 if (PrintOpto && Verbose) 1061 tty->print_cr("Never-taken edge stops compilation at bci %d",bci()); 1062 #endif 1063 repush_if_args(); // to gather stats on loop 1064 // We need to mark this branch as taken so that if we recompile we will 1065 // see that it is possible. In the tiered system the interpreter doesn't 1066 // do profiling and by the time we get to the lower tier from the interpreter 1067 // the path may be cold again. Make sure it doesn't look untaken 1068 profile_taken_branch(target_bci, !ProfileInterpreter); 1069 uncommon_trap(Deoptimization::Reason_unreached, 1070 Deoptimization::Action_reinterpret, 1071 NULL, "cold"); 1072 if (C->eliminate_boxing()) { 1073 // Mark the successor blocks as parsed 1074 branch_block->next_path_num(); 1075 next_block->next_path_num(); 1076 } 1077 return; 1078 } 1079 1080 // Sanity check the probability value 1081 assert(0.0f < prob && prob < 1.0f,"Bad probability in Parser"); 1082 1083 bool taken_if_true = true; 1084 // Convert BoolTest to canonical form: 1085 if (!BoolTest(btest).is_canonical()) { 1086 btest = BoolTest(btest).negate(); 1087 taken_if_true = false; 1088 // prob is NOT updated here; it remains the probability of the taken 1089 // path (as opposed to the prob of the path guarded by an 'IfTrueNode'). 1090 } 1091 assert(btest != BoolTest::eq, "!= is the only canonical exact test"); 1092 1093 Node* tst0 = new (C) BoolNode(c, btest); 1094 Node* tst = _gvn.transform(tst0); 1095 BoolTest::mask taken_btest = BoolTest::illegal; 1096 BoolTest::mask untaken_btest = BoolTest::illegal; 1097 1098 if (tst->is_Bool()) { 1099 // Refresh c from the transformed bool node, since it may be 1100 // simpler than the original c. Also re-canonicalize btest. 1101 // This wins when (Bool ne (Conv2B p) 0) => (Bool ne (CmpP p NULL)). 1102 // That can arise from statements like: if (x instanceof C) ... 1103 if (tst != tst0) { 1104 // Canonicalize one more time since transform can change it. 1105 btest = tst->as_Bool()->_test._test; 1106 if (!BoolTest(btest).is_canonical()) { 1107 // Reverse edges one more time... 1108 tst = _gvn.transform( tst->as_Bool()->negate(&_gvn) ); 1109 btest = tst->as_Bool()->_test._test; 1110 assert(BoolTest(btest).is_canonical(), "sanity"); 1111 taken_if_true = !taken_if_true; 1112 } 1113 c = tst->in(1); 1114 } 1115 BoolTest::mask neg_btest = BoolTest(btest).negate(); 1116 taken_btest = taken_if_true ? btest : neg_btest; 1117 untaken_btest = taken_if_true ? neg_btest : btest; 1118 } 1119 1120 // Generate real control flow 1121 float true_prob = (taken_if_true ? prob : untaken_prob); 1122 IfNode* iff = create_and_map_if(control(), tst, true_prob, cnt); 1123 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser"); 1124 Node* taken_branch = new (C) IfTrueNode(iff); 1125 Node* untaken_branch = new (C) IfFalseNode(iff); 1126 if (!taken_if_true) { // Finish conversion to canonical form 1127 Node* tmp = taken_branch; 1128 taken_branch = untaken_branch; 1129 untaken_branch = tmp; 1130 } 1131 1132 // Branch is taken: 1133 { PreserveJVMState pjvms(this); 1134 taken_branch = _gvn.transform(taken_branch); 1135 set_control(taken_branch); 1136 1137 if (stopped()) { 1138 if (C->eliminate_boxing()) { 1139 // Mark the successor block as parsed 1140 branch_block->next_path_num(); 1141 } 1142 } else { 1143 // Update method data 1144 profile_taken_branch(target_bci); 1145 adjust_map_after_if(taken_btest, c, prob, branch_block, next_block); 1146 if (!stopped()) { 1147 merge(target_bci); 1148 } 1149 } 1150 } 1151 1152 untaken_branch = _gvn.transform(untaken_branch); 1153 set_control(untaken_branch); 1154 1155 // Branch not taken. 1156 if (stopped()) { 1157 if (C->eliminate_boxing()) { 1158 // Mark the successor block as parsed 1159 next_block->next_path_num(); 1160 } 1161 } else { 1162 // Update method data 1163 profile_not_taken_branch(); 1164 adjust_map_after_if(untaken_btest, c, untaken_prob, 1165 next_block, branch_block); 1166 } 1167 } 1168 1169 //----------------------------adjust_map_after_if------------------------------ 1170 // Adjust the JVM state to reflect the result of taking this path. 1171 // Basically, it means inspecting the CmpNode controlling this 1172 // branch, seeing how it constrains a tested value, and then 1173 // deciding if it's worth our while to encode this constraint 1174 // as graph nodes in the current abstract interpretation map. 1175 void Parse::adjust_map_after_if(BoolTest::mask btest, Node* c, float prob, 1176 Block* path, Block* other_path) { 1177 if (stopped() || !c->is_Cmp() || btest == BoolTest::illegal) 1178 return; // nothing to do 1179 1180 bool is_fallthrough = (path == successor_for_bci(iter().next_bci())); 1181 1182 if (seems_never_taken(prob) && seems_stable_comparison(btest, c)) { 1183 // If this might possibly turn into an implicit null check, 1184 // and the null has never yet been seen, we need to generate 1185 // an uncommon trap, so as to recompile instead of suffering 1186 // with very slow branches. (We'll get the slow branches if 1187 // the program ever changes phase and starts seeing nulls here.) 1188 // 1189 // We do not inspect for a null constant, since a node may 1190 // optimize to 'null' later on. 1191 // 1192 // Null checks, and other tests which expect inequality, 1193 // show btest == BoolTest::eq along the non-taken branch. 1194 // On the other hand, type tests, must-be-null tests, 1195 // and other tests which expect pointer equality, 1196 // show btest == BoolTest::ne along the non-taken branch. 1197 // We prune both types of branches if they look unused. 1198 repush_if_args(); 1199 // We need to mark this branch as taken so that if we recompile we will 1200 // see that it is possible. In the tiered system the interpreter doesn't 1201 // do profiling and by the time we get to the lower tier from the interpreter 1202 // the path may be cold again. Make sure it doesn't look untaken 1203 if (is_fallthrough) { 1204 profile_not_taken_branch(!ProfileInterpreter); 1205 } else { 1206 profile_taken_branch(iter().get_dest(), !ProfileInterpreter); 1207 } 1208 uncommon_trap(Deoptimization::Reason_unreached, 1209 Deoptimization::Action_reinterpret, 1210 NULL, 1211 (is_fallthrough ? "taken always" : "taken never")); 1212 return; 1213 } 1214 1215 Node* val = c->in(1); 1216 Node* con = c->in(2); 1217 const Type* tcon = _gvn.type(con); 1218 const Type* tval = _gvn.type(val); 1219 bool have_con = tcon->singleton(); 1220 if (tval->singleton()) { 1221 if (!have_con) { 1222 // Swap, so constant is in con. 1223 con = val; 1224 tcon = tval; 1225 val = c->in(2); 1226 tval = _gvn.type(val); 1227 btest = BoolTest(btest).commute(); 1228 have_con = true; 1229 } else { 1230 // Do we have two constants? Then leave well enough alone. 1231 have_con = false; 1232 } 1233 } 1234 if (!have_con) // remaining adjustments need a con 1235 return; 1236 1237 sharpen_type_after_if(btest, con, tcon, val, tval); 1238 } 1239 1240 1241 static Node* extract_obj_from_klass_load(PhaseGVN* gvn, Node* n) { 1242 Node* ldk; 1243 if (n->is_DecodeNKlass()) { 1244 if (n->in(1)->Opcode() != Op_LoadNKlass) { 1245 return NULL; 1246 } else { 1247 ldk = n->in(1); 1248 } 1249 } else if (n->Opcode() != Op_LoadKlass) { 1250 return NULL; 1251 } else { 1252 ldk = n; 1253 } 1254 assert(ldk != NULL && ldk->is_Load(), "should have found a LoadKlass or LoadNKlass node"); 1255 1256 Node* adr = ldk->in(MemNode::Address); 1257 intptr_t off = 0; 1258 Node* obj = AddPNode::Ideal_base_and_offset(adr, gvn, off); 1259 if (obj == NULL || off != oopDesc::klass_offset_in_bytes()) // loading oopDesc::_klass? 1260 return NULL; 1261 const TypePtr* tp = gvn->type(obj)->is_ptr(); 1262 if (tp == NULL || !(tp->isa_instptr() || tp->isa_aryptr())) // is obj a Java object ptr? 1263 return NULL; 1264 1265 return obj; 1266 } 1267 1268 void Parse::sharpen_type_after_if(BoolTest::mask btest, 1269 Node* con, const Type* tcon, 1270 Node* val, const Type* tval) { 1271 // Look for opportunities to sharpen the type of a node 1272 // whose klass is compared with a constant klass. 1273 if (btest == BoolTest::eq && tcon->isa_klassptr()) { 1274 Node* obj = extract_obj_from_klass_load(&_gvn, val); 1275 const TypeOopPtr* con_type = tcon->isa_klassptr()->as_instance_type(); 1276 if (obj != NULL && (con_type->isa_instptr() || con_type->isa_aryptr())) { 1277 // Found: 1278 // Bool(CmpP(LoadKlass(obj._klass), ConP(Foo.klass)), [eq]) 1279 // or the narrowOop equivalent. 1280 const Type* obj_type = _gvn.type(obj); 1281 const TypeOopPtr* tboth = obj_type->join_speculative(con_type)->isa_oopptr(); 1282 if (tboth != NULL && tboth->klass_is_exact() && tboth != obj_type && 1283 tboth->higher_equal(obj_type)) { 1284 // obj has to be of the exact type Foo if the CmpP succeeds. 1285 int obj_in_map = map()->find_edge(obj); 1286 JVMState* jvms = this->jvms(); 1287 if (obj_in_map >= 0 && 1288 (jvms->is_loc(obj_in_map) || jvms->is_stk(obj_in_map))) { 1289 TypeNode* ccast = new (C) CheckCastPPNode(control(), obj, tboth); 1290 const Type* tcc = ccast->as_Type()->type(); 1291 assert(tcc != obj_type && tcc->higher_equal(obj_type), "must improve"); 1292 // Delay transform() call to allow recovery of pre-cast value 1293 // at the control merge. 1294 _gvn.set_type_bottom(ccast); 1295 record_for_igvn(ccast); 1296 // Here's the payoff. 1297 replace_in_map(obj, ccast); 1298 } 1299 } 1300 } 1301 } 1302 1303 int val_in_map = map()->find_edge(val); 1304 if (val_in_map < 0) return; // replace_in_map would be useless 1305 { 1306 JVMState* jvms = this->jvms(); 1307 if (!(jvms->is_loc(val_in_map) || 1308 jvms->is_stk(val_in_map))) 1309 return; // again, it would be useless 1310 } 1311 1312 // Check for a comparison to a constant, and "know" that the compared 1313 // value is constrained on this path. 1314 assert(tcon->singleton(), ""); 1315 ConstraintCastNode* ccast = NULL; 1316 Node* cast = NULL; 1317 1318 switch (btest) { 1319 case BoolTest::eq: // Constant test? 1320 { 1321 const Type* tboth = tcon->join_speculative(tval); 1322 if (tboth == tval) break; // Nothing to gain. 1323 if (tcon->isa_int()) { 1324 ccast = new (C) CastIINode(val, tboth); 1325 } else if (tcon == TypePtr::NULL_PTR) { 1326 // Cast to null, but keep the pointer identity temporarily live. 1327 ccast = new (C) CastPPNode(val, tboth); 1328 } else { 1329 const TypeF* tf = tcon->isa_float_constant(); 1330 const TypeD* td = tcon->isa_double_constant(); 1331 // Exclude tests vs float/double 0 as these could be 1332 // either +0 or -0. Just because you are equal to +0 1333 // doesn't mean you ARE +0! 1334 // Note, following code also replaces Long and Oop values. 1335 if ((!tf || tf->_f != 0.0) && 1336 (!td || td->_d != 0.0)) 1337 cast = con; // Replace non-constant val by con. 1338 } 1339 } 1340 break; 1341 1342 case BoolTest::ne: 1343 if (tcon == TypePtr::NULL_PTR) { 1344 cast = cast_not_null(val, false); 1345 } 1346 break; 1347 1348 default: 1349 // (At this point we could record int range types with CastII.) 1350 break; 1351 } 1352 1353 if (ccast != NULL) { 1354 const Type* tcc = ccast->as_Type()->type(); 1355 assert(tcc != tval && tcc->higher_equal(tval), "must improve"); 1356 // Delay transform() call to allow recovery of pre-cast value 1357 // at the control merge. 1358 ccast->set_req(0, control()); 1359 _gvn.set_type_bottom(ccast); 1360 record_for_igvn(ccast); 1361 cast = ccast; 1362 } 1363 1364 if (cast != NULL) { // Here's the payoff. 1365 replace_in_map(val, cast); 1366 } 1367 } 1368 1369 /** 1370 * Use speculative type to optimize CmpP node: if comparison is 1371 * against the low level class, cast the object to the speculative 1372 * type if any. CmpP should then go away. 1373 * 1374 * @param c expected CmpP node 1375 * @return result of CmpP on object casted to speculative type 1376 * 1377 */ 1378 Node* Parse::optimize_cmp_with_klass(Node* c) { 1379 // If this is transformed by the _gvn to a comparison with the low 1380 // level klass then we may be able to use speculation 1381 if (c->Opcode() == Op_CmpP && 1382 (c->in(1)->Opcode() == Op_LoadKlass || c->in(1)->Opcode() == Op_DecodeNKlass) && 1383 c->in(2)->is_Con()) { 1384 Node* load_klass = NULL; 1385 Node* decode = NULL; 1386 if (c->in(1)->Opcode() == Op_DecodeNKlass) { 1387 decode = c->in(1); 1388 load_klass = c->in(1)->in(1); 1389 } else { 1390 load_klass = c->in(1); 1391 } 1392 if (load_klass->in(2)->is_AddP()) { 1393 Node* addp = load_klass->in(2); 1394 Node* obj = addp->in(AddPNode::Address); 1395 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 1396 if (obj_type->speculative_type_not_null() != NULL) { 1397 ciKlass* k = obj_type->speculative_type(); 1398 inc_sp(2); 1399 obj = maybe_cast_profiled_obj(obj, k); 1400 dec_sp(2); 1401 // Make the CmpP use the casted obj 1402 addp = basic_plus_adr(obj, addp->in(AddPNode::Offset)); 1403 load_klass = load_klass->clone(); 1404 load_klass->set_req(2, addp); 1405 load_klass = _gvn.transform(load_klass); 1406 if (decode != NULL) { 1407 decode = decode->clone(); 1408 decode->set_req(1, load_klass); 1409 load_klass = _gvn.transform(decode); 1410 } 1411 c = c->clone(); 1412 c->set_req(1, load_klass); 1413 c = _gvn.transform(c); 1414 } 1415 } 1416 } 1417 return c; 1418 } 1419 1420 //------------------------------do_one_bytecode-------------------------------- 1421 // Parse this bytecode, and alter the Parsers JVM->Node mapping 1422 void Parse::do_one_bytecode() { 1423 Node *a, *b, *c, *d; // Handy temps 1424 BoolTest::mask btest; 1425 int i; 1426 1427 assert(!has_exceptions(), "bytecode entry state must be clear of throws"); 1428 1429 if (C->check_node_count(NodeLimitFudgeFactor * 5, 1430 "out of nodes parsing method")) { 1431 return; 1432 } 1433 1434 #ifdef ASSERT 1435 // for setting breakpoints 1436 if (TraceOptoParse) { 1437 tty->print(" @"); 1438 dump_bci(bci()); 1439 tty->cr(); 1440 } 1441 #endif 1442 1443 switch (bc()) { 1444 case Bytecodes::_nop: 1445 // do nothing 1446 break; 1447 case Bytecodes::_lconst_0: 1448 push_pair(longcon(0)); 1449 break; 1450 1451 case Bytecodes::_lconst_1: 1452 push_pair(longcon(1)); 1453 break; 1454 1455 case Bytecodes::_fconst_0: 1456 push(zerocon(T_FLOAT)); 1457 break; 1458 1459 case Bytecodes::_fconst_1: 1460 push(makecon(TypeF::ONE)); 1461 break; 1462 1463 case Bytecodes::_fconst_2: 1464 push(makecon(TypeF::make(2.0f))); 1465 break; 1466 1467 case Bytecodes::_dconst_0: 1468 push_pair(zerocon(T_DOUBLE)); 1469 break; 1470 1471 case Bytecodes::_dconst_1: 1472 push_pair(makecon(TypeD::ONE)); 1473 break; 1474 1475 case Bytecodes::_iconst_m1:push(intcon(-1)); break; 1476 case Bytecodes::_iconst_0: push(intcon( 0)); break; 1477 case Bytecodes::_iconst_1: push(intcon( 1)); break; 1478 case Bytecodes::_iconst_2: push(intcon( 2)); break; 1479 case Bytecodes::_iconst_3: push(intcon( 3)); break; 1480 case Bytecodes::_iconst_4: push(intcon( 4)); break; 1481 case Bytecodes::_iconst_5: push(intcon( 5)); break; 1482 case Bytecodes::_bipush: push(intcon(iter().get_constant_u1())); break; 1483 case Bytecodes::_sipush: push(intcon(iter().get_constant_u2())); break; 1484 case Bytecodes::_aconst_null: push(null()); break; 1485 case Bytecodes::_ldc: 1486 case Bytecodes::_ldc_w: 1487 case Bytecodes::_ldc2_w: 1488 // If the constant is unresolved, run this BC once in the interpreter. 1489 { 1490 ciConstant constant = iter().get_constant(); 1491 if (constant.basic_type() == T_OBJECT && 1492 !constant.as_object()->is_loaded()) { 1493 int index = iter().get_constant_pool_index(); 1494 constantTag tag = iter().get_constant_pool_tag(index); 1495 uncommon_trap(Deoptimization::make_trap_request 1496 (Deoptimization::Reason_unloaded, 1497 Deoptimization::Action_reinterpret, 1498 index), 1499 NULL, tag.internal_name()); 1500 break; 1501 } 1502 assert(constant.basic_type() != T_OBJECT || constant.as_object()->is_instance(), 1503 "must be java_mirror of klass"); 1504 bool pushed = push_constant(constant, true); 1505 guarantee(pushed, "must be possible to push this constant"); 1506 } 1507 1508 break; 1509 1510 case Bytecodes::_aload_0: 1511 push( local(0) ); 1512 break; 1513 case Bytecodes::_aload_1: 1514 push( local(1) ); 1515 break; 1516 case Bytecodes::_aload_2: 1517 push( local(2) ); 1518 break; 1519 case Bytecodes::_aload_3: 1520 push( local(3) ); 1521 break; 1522 case Bytecodes::_aload: 1523 push( local(iter().get_index()) ); 1524 break; 1525 1526 case Bytecodes::_fload_0: 1527 case Bytecodes::_iload_0: 1528 push( local(0) ); 1529 break; 1530 case Bytecodes::_fload_1: 1531 case Bytecodes::_iload_1: 1532 push( local(1) ); 1533 break; 1534 case Bytecodes::_fload_2: 1535 case Bytecodes::_iload_2: 1536 push( local(2) ); 1537 break; 1538 case Bytecodes::_fload_3: 1539 case Bytecodes::_iload_3: 1540 push( local(3) ); 1541 break; 1542 case Bytecodes::_fload: 1543 case Bytecodes::_iload: 1544 push( local(iter().get_index()) ); 1545 break; 1546 case Bytecodes::_lload_0: 1547 push_pair_local( 0 ); 1548 break; 1549 case Bytecodes::_lload_1: 1550 push_pair_local( 1 ); 1551 break; 1552 case Bytecodes::_lload_2: 1553 push_pair_local( 2 ); 1554 break; 1555 case Bytecodes::_lload_3: 1556 push_pair_local( 3 ); 1557 break; 1558 case Bytecodes::_lload: 1559 push_pair_local( iter().get_index() ); 1560 break; 1561 1562 case Bytecodes::_dload_0: 1563 push_pair_local(0); 1564 break; 1565 case Bytecodes::_dload_1: 1566 push_pair_local(1); 1567 break; 1568 case Bytecodes::_dload_2: 1569 push_pair_local(2); 1570 break; 1571 case Bytecodes::_dload_3: 1572 push_pair_local(3); 1573 break; 1574 case Bytecodes::_dload: 1575 push_pair_local(iter().get_index()); 1576 break; 1577 case Bytecodes::_fstore_0: 1578 case Bytecodes::_istore_0: 1579 case Bytecodes::_astore_0: 1580 set_local( 0, pop() ); 1581 break; 1582 case Bytecodes::_fstore_1: 1583 case Bytecodes::_istore_1: 1584 case Bytecodes::_astore_1: 1585 set_local( 1, pop() ); 1586 break; 1587 case Bytecodes::_fstore_2: 1588 case Bytecodes::_istore_2: 1589 case Bytecodes::_astore_2: 1590 set_local( 2, pop() ); 1591 break; 1592 case Bytecodes::_fstore_3: 1593 case Bytecodes::_istore_3: 1594 case Bytecodes::_astore_3: 1595 set_local( 3, pop() ); 1596 break; 1597 case Bytecodes::_fstore: 1598 case Bytecodes::_istore: 1599 case Bytecodes::_astore: 1600 set_local( iter().get_index(), pop() ); 1601 break; 1602 // long stores 1603 case Bytecodes::_lstore_0: 1604 set_pair_local( 0, pop_pair() ); 1605 break; 1606 case Bytecodes::_lstore_1: 1607 set_pair_local( 1, pop_pair() ); 1608 break; 1609 case Bytecodes::_lstore_2: 1610 set_pair_local( 2, pop_pair() ); 1611 break; 1612 case Bytecodes::_lstore_3: 1613 set_pair_local( 3, pop_pair() ); 1614 break; 1615 case Bytecodes::_lstore: 1616 set_pair_local( iter().get_index(), pop_pair() ); 1617 break; 1618 1619 // double stores 1620 case Bytecodes::_dstore_0: 1621 set_pair_local( 0, dstore_rounding(pop_pair()) ); 1622 break; 1623 case Bytecodes::_dstore_1: 1624 set_pair_local( 1, dstore_rounding(pop_pair()) ); 1625 break; 1626 case Bytecodes::_dstore_2: 1627 set_pair_local( 2, dstore_rounding(pop_pair()) ); 1628 break; 1629 case Bytecodes::_dstore_3: 1630 set_pair_local( 3, dstore_rounding(pop_pair()) ); 1631 break; 1632 case Bytecodes::_dstore: 1633 set_pair_local( iter().get_index(), dstore_rounding(pop_pair()) ); 1634 break; 1635 1636 case Bytecodes::_pop: dec_sp(1); break; 1637 case Bytecodes::_pop2: dec_sp(2); break; 1638 case Bytecodes::_swap: 1639 a = pop(); 1640 b = pop(); 1641 push(a); 1642 push(b); 1643 break; 1644 case Bytecodes::_dup: 1645 a = pop(); 1646 push(a); 1647 push(a); 1648 break; 1649 case Bytecodes::_dup_x1: 1650 a = pop(); 1651 b = pop(); 1652 push( a ); 1653 push( b ); 1654 push( a ); 1655 break; 1656 case Bytecodes::_dup_x2: 1657 a = pop(); 1658 b = pop(); 1659 c = pop(); 1660 push( a ); 1661 push( c ); 1662 push( b ); 1663 push( a ); 1664 break; 1665 case Bytecodes::_dup2: 1666 a = pop(); 1667 b = pop(); 1668 push( b ); 1669 push( a ); 1670 push( b ); 1671 push( a ); 1672 break; 1673 1674 case Bytecodes::_dup2_x1: 1675 // before: .. c, b, a 1676 // after: .. b, a, c, b, a 1677 // not tested 1678 a = pop(); 1679 b = pop(); 1680 c = pop(); 1681 push( b ); 1682 push( a ); 1683 push( c ); 1684 push( b ); 1685 push( a ); 1686 break; 1687 case Bytecodes::_dup2_x2: 1688 // before: .. d, c, b, a 1689 // after: .. b, a, d, c, b, a 1690 // not tested 1691 a = pop(); 1692 b = pop(); 1693 c = pop(); 1694 d = pop(); 1695 push( b ); 1696 push( a ); 1697 push( d ); 1698 push( c ); 1699 push( b ); 1700 push( a ); 1701 break; 1702 1703 case Bytecodes::_arraylength: { 1704 // Must do null-check with value on expression stack 1705 Node *ary = null_check(peek(), T_ARRAY); 1706 // Compile-time detect of null-exception? 1707 if (stopped()) return; 1708 a = pop(); 1709 push(load_array_length(a)); 1710 break; 1711 } 1712 1713 case Bytecodes::_baload: array_load(T_BYTE); break; 1714 case Bytecodes::_caload: array_load(T_CHAR); break; 1715 case Bytecodes::_iaload: array_load(T_INT); break; 1716 case Bytecodes::_saload: array_load(T_SHORT); break; 1717 case Bytecodes::_faload: array_load(T_FLOAT); break; 1718 case Bytecodes::_aaload: array_load(T_OBJECT); break; 1719 case Bytecodes::_laload: { 1720 a = array_addressing(T_LONG, 0); 1721 if (stopped()) return; // guaranteed null or range check 1722 dec_sp(2); // Pop array and index 1723 push_pair(make_load(control(), a, TypeLong::LONG, T_LONG, TypeAryPtr::LONGS, MemNode::unordered)); 1724 break; 1725 } 1726 case Bytecodes::_daload: { 1727 a = array_addressing(T_DOUBLE, 0); 1728 if (stopped()) return; // guaranteed null or range check 1729 dec_sp(2); // Pop array and index 1730 push_pair(make_load(control(), a, Type::DOUBLE, T_DOUBLE, TypeAryPtr::DOUBLES, MemNode::unordered)); 1731 break; 1732 } 1733 case Bytecodes::_bastore: array_store(T_BYTE); break; 1734 case Bytecodes::_castore: array_store(T_CHAR); break; 1735 case Bytecodes::_iastore: array_store(T_INT); break; 1736 case Bytecodes::_sastore: array_store(T_SHORT); break; 1737 case Bytecodes::_fastore: array_store(T_FLOAT); break; 1738 case Bytecodes::_aastore: { 1739 d = array_addressing(T_OBJECT, 1); 1740 if (stopped()) return; // guaranteed null or range check 1741 array_store_check(); 1742 c = pop(); // Oop to store 1743 b = pop(); // index (already used) 1744 a = pop(); // the array itself 1745 const TypeOopPtr* elemtype = _gvn.type(a)->is_aryptr()->elem()->make_oopptr(); 1746 const TypeAryPtr* adr_type = TypeAryPtr::OOPS; 1747 Node* store = store_oop_to_array(control(), a, d, adr_type, c, elemtype, T_OBJECT, MemNode::release); 1748 break; 1749 } 1750 case Bytecodes::_lastore: { 1751 a = array_addressing(T_LONG, 2); 1752 if (stopped()) return; // guaranteed null or range check 1753 c = pop_pair(); 1754 dec_sp(2); // Pop array and index 1755 store_to_memory(control(), a, c, T_LONG, TypeAryPtr::LONGS, MemNode::unordered); 1756 break; 1757 } 1758 case Bytecodes::_dastore: { 1759 a = array_addressing(T_DOUBLE, 2); 1760 if (stopped()) return; // guaranteed null or range check 1761 c = pop_pair(); 1762 dec_sp(2); // Pop array and index 1763 c = dstore_rounding(c); 1764 store_to_memory(control(), a, c, T_DOUBLE, TypeAryPtr::DOUBLES, MemNode::unordered); 1765 break; 1766 } 1767 case Bytecodes::_getfield: 1768 do_getfield(); 1769 break; 1770 1771 case Bytecodes::_getstatic: 1772 do_getstatic(); 1773 break; 1774 1775 case Bytecodes::_putfield: 1776 do_putfield(); 1777 break; 1778 1779 case Bytecodes::_putstatic: 1780 do_putstatic(); 1781 break; 1782 1783 case Bytecodes::_irem: 1784 do_irem(); 1785 break; 1786 case Bytecodes::_idiv: 1787 // Must keep both values on the expression-stack during null-check 1788 zero_check_int(peek()); 1789 // Compile-time detect of null-exception? 1790 if (stopped()) return; 1791 b = pop(); 1792 a = pop(); 1793 push( _gvn.transform( new (C) DivINode(control(),a,b) ) ); 1794 break; 1795 case Bytecodes::_imul: 1796 b = pop(); a = pop(); 1797 push( _gvn.transform( new (C) MulINode(a,b) ) ); 1798 break; 1799 case Bytecodes::_iadd: 1800 b = pop(); a = pop(); 1801 push( _gvn.transform( new (C) AddINode(a,b) ) ); 1802 break; 1803 case Bytecodes::_ineg: 1804 a = pop(); 1805 push( _gvn.transform( new (C) SubINode(_gvn.intcon(0),a)) ); 1806 break; 1807 case Bytecodes::_isub: 1808 b = pop(); a = pop(); 1809 push( _gvn.transform( new (C) SubINode(a,b) ) ); 1810 break; 1811 case Bytecodes::_iand: 1812 b = pop(); a = pop(); 1813 push( _gvn.transform( new (C) AndINode(a,b) ) ); 1814 break; 1815 case Bytecodes::_ior: 1816 b = pop(); a = pop(); 1817 push( _gvn.transform( new (C) OrINode(a,b) ) ); 1818 break; 1819 case Bytecodes::_ixor: 1820 b = pop(); a = pop(); 1821 push( _gvn.transform( new (C) XorINode(a,b) ) ); 1822 break; 1823 case Bytecodes::_ishl: 1824 b = pop(); a = pop(); 1825 push( _gvn.transform( new (C) LShiftINode(a,b) ) ); 1826 break; 1827 case Bytecodes::_ishr: 1828 b = pop(); a = pop(); 1829 push( _gvn.transform( new (C) RShiftINode(a,b) ) ); 1830 break; 1831 case Bytecodes::_iushr: 1832 b = pop(); a = pop(); 1833 push( _gvn.transform( new (C) URShiftINode(a,b) ) ); 1834 break; 1835 1836 case Bytecodes::_fneg: 1837 a = pop(); 1838 b = _gvn.transform(new (C) NegFNode (a)); 1839 push(b); 1840 break; 1841 1842 case Bytecodes::_fsub: 1843 b = pop(); 1844 a = pop(); 1845 c = _gvn.transform( new (C) SubFNode(a,b) ); 1846 d = precision_rounding(c); 1847 push( d ); 1848 break; 1849 1850 case Bytecodes::_fadd: 1851 b = pop(); 1852 a = pop(); 1853 c = _gvn.transform( new (C) AddFNode(a,b) ); 1854 d = precision_rounding(c); 1855 push( d ); 1856 break; 1857 1858 case Bytecodes::_fmul: 1859 b = pop(); 1860 a = pop(); 1861 c = _gvn.transform( new (C) MulFNode(a,b) ); 1862 d = precision_rounding(c); 1863 push( d ); 1864 break; 1865 1866 case Bytecodes::_fdiv: 1867 b = pop(); 1868 a = pop(); 1869 c = _gvn.transform( new (C) DivFNode(0,a,b) ); 1870 d = precision_rounding(c); 1871 push( d ); 1872 break; 1873 1874 case Bytecodes::_frem: 1875 if (Matcher::has_match_rule(Op_ModF)) { 1876 // Generate a ModF node. 1877 b = pop(); 1878 a = pop(); 1879 c = _gvn.transform( new (C) ModFNode(0,a,b) ); 1880 d = precision_rounding(c); 1881 push( d ); 1882 } 1883 else { 1884 // Generate a call. 1885 modf(); 1886 } 1887 break; 1888 1889 case Bytecodes::_fcmpl: 1890 b = pop(); 1891 a = pop(); 1892 c = _gvn.transform( new (C) CmpF3Node( a, b)); 1893 push(c); 1894 break; 1895 case Bytecodes::_fcmpg: 1896 b = pop(); 1897 a = pop(); 1898 1899 // Same as fcmpl but need to flip the unordered case. Swap the inputs, 1900 // which negates the result sign except for unordered. Flip the unordered 1901 // as well by using CmpF3 which implements unordered-lesser instead of 1902 // unordered-greater semantics. Finally, commute the result bits. Result 1903 // is same as using a CmpF3Greater except we did it with CmpF3 alone. 1904 c = _gvn.transform( new (C) CmpF3Node( b, a)); 1905 c = _gvn.transform( new (C) SubINode(_gvn.intcon(0),c) ); 1906 push(c); 1907 break; 1908 1909 case Bytecodes::_f2i: 1910 a = pop(); 1911 push(_gvn.transform(new (C) ConvF2INode(a))); 1912 break; 1913 1914 case Bytecodes::_d2i: 1915 a = pop_pair(); 1916 b = _gvn.transform(new (C) ConvD2INode(a)); 1917 push( b ); 1918 break; 1919 1920 case Bytecodes::_f2d: 1921 a = pop(); 1922 b = _gvn.transform( new (C) ConvF2DNode(a)); 1923 push_pair( b ); 1924 break; 1925 1926 case Bytecodes::_d2f: 1927 a = pop_pair(); 1928 b = _gvn.transform( new (C) ConvD2FNode(a)); 1929 // This breaks _227_mtrt (speed & correctness) and _222_mpegaudio (speed) 1930 //b = _gvn.transform(new (C) RoundFloatNode(0, b) ); 1931 push( b ); 1932 break; 1933 1934 case Bytecodes::_l2f: 1935 if (Matcher::convL2FSupported()) { 1936 a = pop_pair(); 1937 b = _gvn.transform( new (C) ConvL2FNode(a)); 1938 // For i486.ad, FILD doesn't restrict precision to 24 or 53 bits. 1939 // Rather than storing the result into an FP register then pushing 1940 // out to memory to round, the machine instruction that implements 1941 // ConvL2D is responsible for rounding. 1942 // c = precision_rounding(b); 1943 c = _gvn.transform(b); 1944 push(c); 1945 } else { 1946 l2f(); 1947 } 1948 break; 1949 1950 case Bytecodes::_l2d: 1951 a = pop_pair(); 1952 b = _gvn.transform( new (C) ConvL2DNode(a)); 1953 // For i486.ad, rounding is always necessary (see _l2f above). 1954 // c = dprecision_rounding(b); 1955 c = _gvn.transform(b); 1956 push_pair(c); 1957 break; 1958 1959 case Bytecodes::_f2l: 1960 a = pop(); 1961 b = _gvn.transform( new (C) ConvF2LNode(a)); 1962 push_pair(b); 1963 break; 1964 1965 case Bytecodes::_d2l: 1966 a = pop_pair(); 1967 b = _gvn.transform( new (C) ConvD2LNode(a)); 1968 push_pair(b); 1969 break; 1970 1971 case Bytecodes::_dsub: 1972 b = pop_pair(); 1973 a = pop_pair(); 1974 c = _gvn.transform( new (C) SubDNode(a,b) ); 1975 d = dprecision_rounding(c); 1976 push_pair( d ); 1977 break; 1978 1979 case Bytecodes::_dadd: 1980 b = pop_pair(); 1981 a = pop_pair(); 1982 c = _gvn.transform( new (C) AddDNode(a,b) ); 1983 d = dprecision_rounding(c); 1984 push_pair( d ); 1985 break; 1986 1987 case Bytecodes::_dmul: 1988 b = pop_pair(); 1989 a = pop_pair(); 1990 c = _gvn.transform( new (C) MulDNode(a,b) ); 1991 d = dprecision_rounding(c); 1992 push_pair( d ); 1993 break; 1994 1995 case Bytecodes::_ddiv: 1996 b = pop_pair(); 1997 a = pop_pair(); 1998 c = _gvn.transform( new (C) DivDNode(0,a,b) ); 1999 d = dprecision_rounding(c); 2000 push_pair( d ); 2001 break; 2002 2003 case Bytecodes::_dneg: 2004 a = pop_pair(); 2005 b = _gvn.transform(new (C) NegDNode (a)); 2006 push_pair(b); 2007 break; 2008 2009 case Bytecodes::_drem: 2010 if (Matcher::has_match_rule(Op_ModD)) { 2011 // Generate a ModD node. 2012 b = pop_pair(); 2013 a = pop_pair(); 2014 // a % b 2015 2016 c = _gvn.transform( new (C) ModDNode(0,a,b) ); 2017 d = dprecision_rounding(c); 2018 push_pair( d ); 2019 } 2020 else { 2021 // Generate a call. 2022 modd(); 2023 } 2024 break; 2025 2026 case Bytecodes::_dcmpl: 2027 b = pop_pair(); 2028 a = pop_pair(); 2029 c = _gvn.transform( new (C) CmpD3Node( a, b)); 2030 push(c); 2031 break; 2032 2033 case Bytecodes::_dcmpg: 2034 b = pop_pair(); 2035 a = pop_pair(); 2036 // Same as dcmpl but need to flip the unordered case. 2037 // Commute the inputs, which negates the result sign except for unordered. 2038 // Flip the unordered as well by using CmpD3 which implements 2039 // unordered-lesser instead of unordered-greater semantics. 2040 // Finally, negate the result bits. Result is same as using a 2041 // CmpD3Greater except we did it with CmpD3 alone. 2042 c = _gvn.transform( new (C) CmpD3Node( b, a)); 2043 c = _gvn.transform( new (C) SubINode(_gvn.intcon(0),c) ); 2044 push(c); 2045 break; 2046 2047 2048 // Note for longs -> lo word is on TOS, hi word is on TOS - 1 2049 case Bytecodes::_land: 2050 b = pop_pair(); 2051 a = pop_pair(); 2052 c = _gvn.transform( new (C) AndLNode(a,b) ); 2053 push_pair(c); 2054 break; 2055 case Bytecodes::_lor: 2056 b = pop_pair(); 2057 a = pop_pair(); 2058 c = _gvn.transform( new (C) OrLNode(a,b) ); 2059 push_pair(c); 2060 break; 2061 case Bytecodes::_lxor: 2062 b = pop_pair(); 2063 a = pop_pair(); 2064 c = _gvn.transform( new (C) XorLNode(a,b) ); 2065 push_pair(c); 2066 break; 2067 2068 case Bytecodes::_lshl: 2069 b = pop(); // the shift count 2070 a = pop_pair(); // value to be shifted 2071 c = _gvn.transform( new (C) LShiftLNode(a,b) ); 2072 push_pair(c); 2073 break; 2074 case Bytecodes::_lshr: 2075 b = pop(); // the shift count 2076 a = pop_pair(); // value to be shifted 2077 c = _gvn.transform( new (C) RShiftLNode(a,b) ); 2078 push_pair(c); 2079 break; 2080 case Bytecodes::_lushr: 2081 b = pop(); // the shift count 2082 a = pop_pair(); // value to be shifted 2083 c = _gvn.transform( new (C) URShiftLNode(a,b) ); 2084 push_pair(c); 2085 break; 2086 case Bytecodes::_lmul: 2087 b = pop_pair(); 2088 a = pop_pair(); 2089 c = _gvn.transform( new (C) MulLNode(a,b) ); 2090 push_pair(c); 2091 break; 2092 2093 case Bytecodes::_lrem: 2094 // Must keep both values on the expression-stack during null-check 2095 assert(peek(0) == top(), "long word order"); 2096 zero_check_long(peek(1)); 2097 // Compile-time detect of null-exception? 2098 if (stopped()) return; 2099 b = pop_pair(); 2100 a = pop_pair(); 2101 c = _gvn.transform( new (C) ModLNode(control(),a,b) ); 2102 push_pair(c); 2103 break; 2104 2105 case Bytecodes::_ldiv: 2106 // Must keep both values on the expression-stack during null-check 2107 assert(peek(0) == top(), "long word order"); 2108 zero_check_long(peek(1)); 2109 // Compile-time detect of null-exception? 2110 if (stopped()) return; 2111 b = pop_pair(); 2112 a = pop_pair(); 2113 c = _gvn.transform( new (C) DivLNode(control(),a,b) ); 2114 push_pair(c); 2115 break; 2116 2117 case Bytecodes::_ladd: 2118 b = pop_pair(); 2119 a = pop_pair(); 2120 c = _gvn.transform( new (C) AddLNode(a,b) ); 2121 push_pair(c); 2122 break; 2123 case Bytecodes::_lsub: 2124 b = pop_pair(); 2125 a = pop_pair(); 2126 c = _gvn.transform( new (C) SubLNode(a,b) ); 2127 push_pair(c); 2128 break; 2129 case Bytecodes::_lcmp: 2130 // Safepoints are now inserted _before_ branches. The long-compare 2131 // bytecode painfully produces a 3-way value (-1,0,+1) which requires a 2132 // slew of control flow. These are usually followed by a CmpI vs zero and 2133 // a branch; this pattern then optimizes to the obvious long-compare and 2134 // branch. However, if the branch is backwards there's a Safepoint 2135 // inserted. The inserted Safepoint captures the JVM state at the 2136 // pre-branch point, i.e. it captures the 3-way value. Thus if a 2137 // long-compare is used to control a loop the debug info will force 2138 // computation of the 3-way value, even though the generated code uses a 2139 // long-compare and branch. We try to rectify the situation by inserting 2140 // a SafePoint here and have it dominate and kill the safepoint added at a 2141 // following backwards branch. At this point the JVM state merely holds 2 2142 // longs but not the 3-way value. 2143 if( UseLoopSafepoints ) { 2144 switch( iter().next_bc() ) { 2145 case Bytecodes::_ifgt: 2146 case Bytecodes::_iflt: 2147 case Bytecodes::_ifge: 2148 case Bytecodes::_ifle: 2149 case Bytecodes::_ifne: 2150 case Bytecodes::_ifeq: 2151 // If this is a backwards branch in the bytecodes, add Safepoint 2152 maybe_add_safepoint(iter().next_get_dest()); 2153 } 2154 } 2155 b = pop_pair(); 2156 a = pop_pair(); 2157 c = _gvn.transform( new (C) CmpL3Node( a, b )); 2158 push(c); 2159 break; 2160 2161 case Bytecodes::_lneg: 2162 a = pop_pair(); 2163 b = _gvn.transform( new (C) SubLNode(longcon(0),a)); 2164 push_pair(b); 2165 break; 2166 case Bytecodes::_l2i: 2167 a = pop_pair(); 2168 push( _gvn.transform( new (C) ConvL2INode(a))); 2169 break; 2170 case Bytecodes::_i2l: 2171 a = pop(); 2172 b = _gvn.transform( new (C) ConvI2LNode(a)); 2173 push_pair(b); 2174 break; 2175 case Bytecodes::_i2b: 2176 // Sign extend 2177 a = pop(); 2178 a = _gvn.transform( new (C) LShiftINode(a,_gvn.intcon(24)) ); 2179 a = _gvn.transform( new (C) RShiftINode(a,_gvn.intcon(24)) ); 2180 push( a ); 2181 break; 2182 case Bytecodes::_i2s: 2183 a = pop(); 2184 a = _gvn.transform( new (C) LShiftINode(a,_gvn.intcon(16)) ); 2185 a = _gvn.transform( new (C) RShiftINode(a,_gvn.intcon(16)) ); 2186 push( a ); 2187 break; 2188 case Bytecodes::_i2c: 2189 a = pop(); 2190 push( _gvn.transform( new (C) AndINode(a,_gvn.intcon(0xFFFF)) ) ); 2191 break; 2192 2193 case Bytecodes::_i2f: 2194 a = pop(); 2195 b = _gvn.transform( new (C) ConvI2FNode(a) ) ; 2196 c = precision_rounding(b); 2197 push (b); 2198 break; 2199 2200 case Bytecodes::_i2d: 2201 a = pop(); 2202 b = _gvn.transform( new (C) ConvI2DNode(a)); 2203 push_pair(b); 2204 break; 2205 2206 case Bytecodes::_iinc: // Increment local 2207 i = iter().get_index(); // Get local index 2208 set_local( i, _gvn.transform( new (C) AddINode( _gvn.intcon(iter().get_iinc_con()), local(i) ) ) ); 2209 break; 2210 2211 // Exit points of synchronized methods must have an unlock node 2212 case Bytecodes::_return: 2213 return_current(NULL); 2214 break; 2215 2216 case Bytecodes::_ireturn: 2217 case Bytecodes::_areturn: 2218 case Bytecodes::_freturn: 2219 return_current(pop()); 2220 break; 2221 case Bytecodes::_lreturn: 2222 return_current(pop_pair()); 2223 break; 2224 case Bytecodes::_dreturn: 2225 return_current(pop_pair()); 2226 break; 2227 2228 case Bytecodes::_athrow: 2229 // null exception oop throws NULL pointer exception 2230 null_check(peek()); 2231 if (stopped()) return; 2232 // Hook the thrown exception directly to subsequent handlers. 2233 if (BailoutToInterpreterForThrows) { 2234 // Keep method interpreted from now on. 2235 uncommon_trap(Deoptimization::Reason_unhandled, 2236 Deoptimization::Action_make_not_compilable); 2237 return; 2238 } 2239 if (env()->jvmti_can_post_on_exceptions()) { 2240 // check if we must post exception events, take uncommon trap if so (with must_throw = false) 2241 uncommon_trap_if_should_post_on_exceptions(Deoptimization::Reason_unhandled, false); 2242 } 2243 // Here if either can_post_on_exceptions or should_post_on_exceptions is false 2244 add_exception_state(make_exception_state(peek())); 2245 break; 2246 2247 case Bytecodes::_goto: // fall through 2248 case Bytecodes::_goto_w: { 2249 int target_bci = (bc() == Bytecodes::_goto) ? iter().get_dest() : iter().get_far_dest(); 2250 2251 // If this is a backwards branch in the bytecodes, add Safepoint 2252 maybe_add_safepoint(target_bci); 2253 2254 // Update method data 2255 profile_taken_branch(target_bci); 2256 2257 // Merge the current control into the target basic block 2258 merge(target_bci); 2259 2260 // See if we can get some profile data and hand it off to the next block 2261 Block *target_block = block()->successor_for_bci(target_bci); 2262 if (target_block->pred_count() != 1) break; 2263 ciMethodData* methodData = method()->method_data(); 2264 if (!methodData->is_mature()) break; 2265 ciProfileData* data = methodData->bci_to_data(bci()); 2266 assert( data->is_JumpData(), "" ); 2267 int taken = ((ciJumpData*)data)->taken(); 2268 taken = method()->scale_count(taken); 2269 target_block->set_count(taken); 2270 break; 2271 } 2272 2273 case Bytecodes::_ifnull: btest = BoolTest::eq; goto handle_if_null; 2274 case Bytecodes::_ifnonnull: btest = BoolTest::ne; goto handle_if_null; 2275 handle_if_null: 2276 // If this is a backwards branch in the bytecodes, add Safepoint 2277 maybe_add_safepoint(iter().get_dest()); 2278 a = null(); 2279 b = pop(); 2280 if (!_gvn.type(b)->speculative_maybe_null() && 2281 !too_many_traps(Deoptimization::Reason_speculate_null_check)) { 2282 inc_sp(1); 2283 Node* null_ctl = top(); 2284 b = null_check_oop(b, &null_ctl, true, true, true); 2285 assert(null_ctl->is_top(), "no null control here"); 2286 dec_sp(1); 2287 } 2288 c = _gvn.transform( new (C) CmpPNode(b, a) ); 2289 do_ifnull(btest, c); 2290 break; 2291 2292 case Bytecodes::_if_acmpeq: btest = BoolTest::eq; goto handle_if_acmp; 2293 case Bytecodes::_if_acmpne: btest = BoolTest::ne; goto handle_if_acmp; 2294 handle_if_acmp: 2295 // If this is a backwards branch in the bytecodes, add Safepoint 2296 maybe_add_safepoint(iter().get_dest()); 2297 a = pop(); 2298 b = pop(); 2299 c = _gvn.transform( new (C) CmpPNode(b, a) ); 2300 c = optimize_cmp_with_klass(c); 2301 do_if(btest, c); 2302 break; 2303 2304 case Bytecodes::_ifeq: btest = BoolTest::eq; goto handle_ifxx; 2305 case Bytecodes::_ifne: btest = BoolTest::ne; goto handle_ifxx; 2306 case Bytecodes::_iflt: btest = BoolTest::lt; goto handle_ifxx; 2307 case Bytecodes::_ifle: btest = BoolTest::le; goto handle_ifxx; 2308 case Bytecodes::_ifgt: btest = BoolTest::gt; goto handle_ifxx; 2309 case Bytecodes::_ifge: btest = BoolTest::ge; goto handle_ifxx; 2310 handle_ifxx: 2311 // If this is a backwards branch in the bytecodes, add Safepoint 2312 maybe_add_safepoint(iter().get_dest()); 2313 a = _gvn.intcon(0); 2314 b = pop(); 2315 c = _gvn.transform( new (C) CmpINode(b, a) ); 2316 do_if(btest, c); 2317 break; 2318 2319 case Bytecodes::_if_icmpeq: btest = BoolTest::eq; goto handle_if_icmp; 2320 case Bytecodes::_if_icmpne: btest = BoolTest::ne; goto handle_if_icmp; 2321 case Bytecodes::_if_icmplt: btest = BoolTest::lt; goto handle_if_icmp; 2322 case Bytecodes::_if_icmple: btest = BoolTest::le; goto handle_if_icmp; 2323 case Bytecodes::_if_icmpgt: btest = BoolTest::gt; goto handle_if_icmp; 2324 case Bytecodes::_if_icmpge: btest = BoolTest::ge; goto handle_if_icmp; 2325 handle_if_icmp: 2326 // If this is a backwards branch in the bytecodes, add Safepoint 2327 maybe_add_safepoint(iter().get_dest()); 2328 a = pop(); 2329 b = pop(); 2330 c = _gvn.transform( new (C) CmpINode( b, a ) ); 2331 do_if(btest, c); 2332 break; 2333 2334 case Bytecodes::_tableswitch: 2335 do_tableswitch(); 2336 break; 2337 2338 case Bytecodes::_lookupswitch: 2339 do_lookupswitch(); 2340 break; 2341 2342 case Bytecodes::_invokestatic: 2343 case Bytecodes::_invokedynamic: 2344 case Bytecodes::_invokespecial: 2345 case Bytecodes::_invokevirtual: 2346 case Bytecodes::_invokeinterface: 2347 do_call(); 2348 break; 2349 case Bytecodes::_checkcast: 2350 do_checkcast(); 2351 break; 2352 case Bytecodes::_instanceof: 2353 do_instanceof(); 2354 break; 2355 case Bytecodes::_anewarray: 2356 do_anewarray(); 2357 break; 2358 case Bytecodes::_newarray: 2359 do_newarray((BasicType)iter().get_index()); 2360 break; 2361 case Bytecodes::_multianewarray: 2362 do_multianewarray(); 2363 break; 2364 case Bytecodes::_new: 2365 do_new(); 2366 break; 2367 2368 case Bytecodes::_jsr: 2369 case Bytecodes::_jsr_w: 2370 do_jsr(); 2371 break; 2372 2373 case Bytecodes::_ret: 2374 do_ret(); 2375 break; 2376 2377 2378 case Bytecodes::_monitorenter: 2379 do_monitor_enter(); 2380 break; 2381 2382 case Bytecodes::_monitorexit: 2383 do_monitor_exit(); 2384 break; 2385 2386 case Bytecodes::_breakpoint: 2387 // Breakpoint set concurrently to compile 2388 // %%% use an uncommon trap? 2389 C->record_failure("breakpoint in method"); 2390 return; 2391 2392 default: 2393 #ifndef PRODUCT 2394 map()->dump(99); 2395 #endif 2396 tty->print("\nUnhandled bytecode %s\n", Bytecodes::name(bc()) ); 2397 ShouldNotReachHere(); 2398 } 2399 2400 #ifndef PRODUCT 2401 IdealGraphPrinter *printer = IdealGraphPrinter::printer(); 2402 if(printer) { 2403 char buffer[256]; 2404 sprintf(buffer, "Bytecode %d: %s", bci(), Bytecodes::name(bc())); 2405 bool old = printer->traverse_outs(); 2406 printer->set_traverse_outs(true); 2407 printer->print_method(C, buffer, 4); 2408 printer->set_traverse_outs(old); 2409 } 2410 #endif 2411 }