1 /* 2 * Copyright (c) 1998, 2018, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "ci/ciMethodData.hpp" 27 #include "classfile/systemDictionary.hpp" 28 #include "classfile/vmSymbols.hpp" 29 #include "compiler/compileLog.hpp" 30 #include "interpreter/linkResolver.hpp" 31 #include "memory/resourceArea.hpp" 32 #include "memory/universe.hpp" 33 #include "oops/oop.inline.hpp" 34 #include "opto/addnode.hpp" 35 #include "opto/castnode.hpp" 36 #include "opto/convertnode.hpp" 37 #include "opto/divnode.hpp" 38 #include "opto/idealGraphPrinter.hpp" 39 #include "opto/matcher.hpp" 40 #include "opto/memnode.hpp" 41 #include "opto/mulnode.hpp" 42 #include "opto/opaquenode.hpp" 43 #include "opto/parse.hpp" 44 #include "opto/runtime.hpp" 45 #include "runtime/deoptimization.hpp" 46 #include "runtime/sharedRuntime.hpp" 47 48 #ifndef PRODUCT 49 extern int explicit_null_checks_inserted, 50 explicit_null_checks_elided; 51 #endif 52 53 //---------------------------------array_load---------------------------------- 54 void Parse::array_load(BasicType elem_type) { 55 const Type* elem = Type::TOP; 56 Node* adr = array_addressing(elem_type, 0, &elem); 57 if (stopped()) return; // guaranteed null or range check 58 dec_sp(2); // Pop array and index 59 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(elem_type); 60 Node* ld = make_load(control(), adr, elem, elem_type, adr_type, MemNode::unordered); 61 push(ld); 62 } 63 64 65 //--------------------------------array_store---------------------------------- 66 void Parse::array_store(BasicType elem_type) { 67 const Type* elem = Type::TOP; 68 Node* adr = array_addressing(elem_type, 1, &elem); 69 if (stopped()) return; // guaranteed null or range check 70 Node* val = pop(); 71 dec_sp(2); // Pop array and index 72 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(elem_type); 73 if (elem == TypeInt::BOOL) { 74 elem_type = T_BOOLEAN; 75 } 76 store_to_memory(control(), adr, val, elem_type, adr_type, StoreNode::release_if_reference(elem_type)); 77 } 78 79 80 //------------------------------array_addressing------------------------------- 81 // Pull array and index from the stack. Compute pointer-to-element. 82 Node* Parse::array_addressing(BasicType type, int vals, const Type* *result2) { 83 Node *idx = peek(0+vals); // Get from stack without popping 84 Node *ary = peek(1+vals); // in case of exception 85 86 // Null check the array base, with correct stack contents 87 ary = null_check(ary, T_ARRAY); 88 // Compile-time detect of null-exception? 89 if (stopped()) return top(); 90 91 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr(); 92 const TypeInt* sizetype = arytype->size(); 93 const Type* elemtype = arytype->elem(); 94 95 if (UseUniqueSubclasses && result2 != NULL) { 96 const Type* el = elemtype->make_ptr(); 97 if (el && el->isa_instptr()) { 98 const TypeInstPtr* toop = el->is_instptr(); 99 if (toop->klass()->as_instance_klass()->unique_concrete_subklass()) { 100 // If we load from "AbstractClass[]" we must see "ConcreteSubClass". 101 const Type* subklass = Type::get_const_type(toop->klass()); 102 elemtype = subklass->join_speculative(el); 103 } 104 } 105 } 106 107 // Check for big class initializers with all constant offsets 108 // feeding into a known-size array. 109 const TypeInt* idxtype = _gvn.type(idx)->is_int(); 110 // See if the highest idx value is less than the lowest array bound, 111 // and if the idx value cannot be negative: 112 bool need_range_check = true; 113 if (idxtype->_hi < sizetype->_lo && idxtype->_lo >= 0) { 114 need_range_check = false; 115 if (C->log() != NULL) C->log()->elem("observe that='!need_range_check'"); 116 } 117 118 ciKlass * arytype_klass = arytype->klass(); 119 if ((arytype_klass != NULL) && (!arytype_klass->is_loaded())) { 120 // Only fails for some -Xcomp runs 121 // The class is unloaded. We have to run this bytecode in the interpreter. 122 uncommon_trap(Deoptimization::Reason_unloaded, 123 Deoptimization::Action_reinterpret, 124 arytype->klass(), "!loaded array"); 125 return top(); 126 } 127 128 // Do the range check 129 if (GenerateRangeChecks && need_range_check) { 130 Node* tst; 131 if (sizetype->_hi <= 0) { 132 // The greatest array bound is negative, so we can conclude that we're 133 // compiling unreachable code, but the unsigned compare trick used below 134 // only works with non-negative lengths. Instead, hack "tst" to be zero so 135 // the uncommon_trap path will always be taken. 136 tst = _gvn.intcon(0); 137 } else { 138 // Range is constant in array-oop, so we can use the original state of mem 139 Node* len = load_array_length(ary); 140 141 // Test length vs index (standard trick using unsigned compare) 142 Node* chk = _gvn.transform( new CmpUNode(idx, len) ); 143 BoolTest::mask btest = BoolTest::lt; 144 tst = _gvn.transform( new BoolNode(chk, btest) ); 145 } 146 RangeCheckNode* rc = new RangeCheckNode(control(), tst, PROB_MAX, COUNT_UNKNOWN); 147 _gvn.set_type(rc, rc->Value(&_gvn)); 148 if (!tst->is_Con()) { 149 record_for_igvn(rc); 150 } 151 set_control(_gvn.transform(new IfTrueNode(rc))); 152 // Branch to failure if out of bounds 153 { 154 PreserveJVMState pjvms(this); 155 set_control(_gvn.transform(new IfFalseNode(rc))); 156 if (C->allow_range_check_smearing()) { 157 // Do not use builtin_throw, since range checks are sometimes 158 // made more stringent by an optimistic transformation. 159 // This creates "tentative" range checks at this point, 160 // which are not guaranteed to throw exceptions. 161 // See IfNode::Ideal, is_range_check, adjust_check. 162 uncommon_trap(Deoptimization::Reason_range_check, 163 Deoptimization::Action_make_not_entrant, 164 NULL, "range_check"); 165 } else { 166 // If we have already recompiled with the range-check-widening 167 // heroic optimization turned off, then we must really be throwing 168 // range check exceptions. 169 builtin_throw(Deoptimization::Reason_range_check, idx); 170 } 171 } 172 } 173 // Check for always knowing you are throwing a range-check exception 174 if (stopped()) return top(); 175 176 // Make array address computation control dependent to prevent it 177 // from floating above the range check during loop optimizations. 178 Node* ptr = array_element_address(ary, idx, type, sizetype, control()); 179 180 if (result2 != NULL) *result2 = elemtype; 181 182 assert(ptr != top(), "top should go hand-in-hand with stopped"); 183 184 return ptr; 185 } 186 187 188 // returns IfNode 189 IfNode* Parse::jump_if_fork_int(Node* a, Node* b, BoolTest::mask mask, float prob, float cnt) { 190 Node *cmp = _gvn.transform(new CmpINode(a, b)); // two cases: shiftcount > 32 and shiftcount <= 32 191 Node *tst = _gvn.transform(new BoolNode(cmp, mask)); 192 IfNode *iff = create_and_map_if(control(), tst, prob, cnt); 193 return iff; 194 } 195 196 // return Region node 197 Node* Parse::jump_if_join(Node* iffalse, Node* iftrue) { 198 Node *region = new RegionNode(3); // 2 results 199 record_for_igvn(region); 200 region->init_req(1, iffalse); 201 region->init_req(2, iftrue ); 202 _gvn.set_type(region, Type::CONTROL); 203 region = _gvn.transform(region); 204 set_control (region); 205 return region; 206 } 207 208 // sentinel value for the target bci to mark never taken branches 209 // (according to profiling) 210 static const int never_reached = INT_MAX; 211 212 //------------------------------helper for tableswitch------------------------- 213 void Parse::jump_if_true_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index, bool unc) { 214 // True branch, use existing map info 215 { PreserveJVMState pjvms(this); 216 Node *iftrue = _gvn.transform( new IfTrueNode (iff) ); 217 set_control( iftrue ); 218 if (unc) { 219 repush_if_args(); 220 uncommon_trap(Deoptimization::Reason_unstable_if, 221 Deoptimization::Action_reinterpret, 222 NULL, 223 "taken always"); 224 } else { 225 assert(dest_bci_if_true != never_reached, "inconsistent dest"); 226 profile_switch_case(prof_table_index); 227 merge_new_path(dest_bci_if_true); 228 } 229 } 230 231 // False branch 232 Node *iffalse = _gvn.transform( new IfFalseNode(iff) ); 233 set_control( iffalse ); 234 } 235 236 void Parse::jump_if_false_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index, bool unc) { 237 // True branch, use existing map info 238 { PreserveJVMState pjvms(this); 239 Node *iffalse = _gvn.transform( new IfFalseNode (iff) ); 240 set_control( iffalse ); 241 if (unc) { 242 repush_if_args(); 243 uncommon_trap(Deoptimization::Reason_unstable_if, 244 Deoptimization::Action_reinterpret, 245 NULL, 246 "taken never"); 247 } else { 248 assert(dest_bci_if_true != never_reached, "inconsistent dest"); 249 profile_switch_case(prof_table_index); 250 merge_new_path(dest_bci_if_true); 251 } 252 } 253 254 // False branch 255 Node *iftrue = _gvn.transform( new IfTrueNode(iff) ); 256 set_control( iftrue ); 257 } 258 259 void Parse::jump_if_always_fork(int dest_bci, int prof_table_index, bool unc) { 260 // False branch, use existing map and control() 261 if (unc) { 262 repush_if_args(); 263 uncommon_trap(Deoptimization::Reason_unstable_if, 264 Deoptimization::Action_reinterpret, 265 NULL, 266 "taken never"); 267 } else { 268 assert(dest_bci != never_reached, "inconsistent dest"); 269 profile_switch_case(prof_table_index); 270 merge_new_path(dest_bci); 271 } 272 } 273 274 275 extern "C" { 276 static int jint_cmp(const void *i, const void *j) { 277 int a = *(jint *)i; 278 int b = *(jint *)j; 279 return a > b ? 1 : a < b ? -1 : 0; 280 } 281 } 282 283 284 // Default value for methodData switch indexing. Must be a negative value to avoid 285 // conflict with any legal switch index. 286 #define NullTableIndex -1 287 288 class SwitchRange : public StackObj { 289 // a range of integers coupled with a bci destination 290 jint _lo; // inclusive lower limit 291 jint _hi; // inclusive upper limit 292 int _dest; 293 int _table_index; // index into method data table 294 float _cnt; // how many times this range was hit according to profiling 295 296 public: 297 jint lo() const { return _lo; } 298 jint hi() const { return _hi; } 299 int dest() const { return _dest; } 300 int table_index() const { return _table_index; } 301 bool is_singleton() const { return _lo == _hi; } 302 float cnt() const { return _cnt; } 303 304 void setRange(jint lo, jint hi, int dest, int table_index, float cnt) { 305 assert(lo <= hi, "must be a non-empty range"); 306 _lo = lo, _hi = hi; _dest = dest; _table_index = table_index; _cnt = cnt; 307 assert(_cnt >= 0, ""); 308 } 309 bool adjoinRange(jint lo, jint hi, int dest, int table_index, float cnt, bool trim_ranges) { 310 assert(lo <= hi, "must be a non-empty range"); 311 if (lo == _hi+1 && table_index == _table_index) { 312 // see merge_ranges() comment below 313 if (trim_ranges) { 314 if (cnt == 0) { 315 if (_cnt != 0) { 316 return false; 317 } 318 if (dest != _dest) { 319 _dest = never_reached; 320 } 321 } else { 322 if (_cnt == 0) { 323 return false; 324 } 325 if (dest != _dest) { 326 return false; 327 } 328 } 329 } else { 330 if (dest != _dest) { 331 return false; 332 } 333 } 334 _hi = hi; 335 _cnt += cnt; 336 return true; 337 } 338 return false; 339 } 340 341 void set (jint value, int dest, int table_index, float cnt) { 342 setRange(value, value, dest, table_index, cnt); 343 } 344 bool adjoin(jint value, int dest, int table_index, float cnt, bool trim_ranges) { 345 return adjoinRange(value, value, dest, table_index, cnt, trim_ranges); 346 } 347 bool adjoin(SwitchRange& other) { 348 return adjoinRange(other._lo, other._hi, other._dest, other._table_index, other._cnt, false); 349 } 350 351 void print() { 352 if (is_singleton()) 353 tty->print(" {%d}=>%d (cnt=%f)", lo(), dest(), cnt()); 354 else if (lo() == min_jint) 355 tty->print(" {..%d}=>%d (cnt=%f)", hi(), dest(), cnt()); 356 else if (hi() == max_jint) 357 tty->print(" {%d..}=>%d (cnt=%f)", lo(), dest(), cnt()); 358 else 359 tty->print(" {%d..%d}=>%d (cnt=%f)", lo(), hi(), dest(), cnt()); 360 } 361 }; 362 363 // We try to minimize the number of ranges and the size of the taken 364 // ones using profiling data. When ranges are created, 365 // SwitchRange::adjoinRange() only allows 2 adjoining ranges to merge 366 // if both were never hit or both were hit to build longer unreached 367 // ranges. Here, we now merge adjoining ranges with the same 368 // destination and finally set destination of unreached ranges to the 369 // special value never_reached because it can help minimize the number 370 // of tests that are necessary. 371 // 372 // For instance: 373 // [0, 1] to target1 sometimes taken 374 // [1, 2] to target1 never taken 375 // [2, 3] to target2 never taken 376 // would lead to: 377 // [0, 1] to target1 sometimes taken 378 // [1, 3] never taken 379 // 380 // (first 2 ranges to target1 are not merged) 381 static void merge_ranges(SwitchRange* ranges, int& rp) { 382 if (rp == 0) { 383 return; 384 } 385 int shift = 0; 386 for (int j = 0; j < rp; j++) { 387 SwitchRange& r1 = ranges[j-shift]; 388 SwitchRange& r2 = ranges[j+1]; 389 if (r1.adjoin(r2)) { 390 shift++; 391 } else if (shift > 0) { 392 ranges[j+1-shift] = r2; 393 } 394 } 395 rp -= shift; 396 for (int j = 0; j <= rp; j++) { 397 SwitchRange& r = ranges[j]; 398 if (r.cnt() == 0 && r.dest() != never_reached) { 399 r.setRange(r.lo(), r.hi(), never_reached, r.table_index(), r.cnt()); 400 } 401 } 402 } 403 404 //-------------------------------do_tableswitch-------------------------------- 405 void Parse::do_tableswitch() { 406 Node* lookup = pop(); 407 // Get information about tableswitch 408 int default_dest = iter().get_dest_table(0); 409 int lo_index = iter().get_int_table(1); 410 int hi_index = iter().get_int_table(2); 411 int len = hi_index - lo_index + 1; 412 413 if (len < 1) { 414 // If this is a backward branch, add safepoint 415 maybe_add_safepoint(default_dest); 416 merge(default_dest); 417 return; 418 } 419 420 ciMethodData* methodData = method()->method_data(); 421 ciMultiBranchData* profile = NULL; 422 if (methodData->is_mature() && UseSwitchProfiling) { 423 ciProfileData* data = methodData->bci_to_data(bci()); 424 if (data != NULL && data->is_MultiBranchData()) { 425 profile = (ciMultiBranchData*)data; 426 } 427 } 428 bool trim_ranges = !method_data_update() && !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if); 429 430 // generate decision tree, using trichotomy when possible 431 int rnum = len+2; 432 bool makes_backward_branch = false; 433 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum); 434 int rp = -1; 435 if (lo_index != min_jint) { 436 uint cnt = 1; 437 if (profile != NULL) { 438 cnt = profile->default_count() / (hi_index != max_jint ? 2 : 1); 439 } 440 ranges[++rp].setRange(min_jint, lo_index-1, default_dest, NullTableIndex, cnt); 441 } 442 for (int j = 0; j < len; j++) { 443 jint match_int = lo_index+j; 444 int dest = iter().get_dest_table(j+3); 445 makes_backward_branch |= (dest <= bci()); 446 int table_index = method_data_update() ? j : NullTableIndex; 447 uint cnt = 1; 448 if (profile != NULL) { 449 cnt = profile->count_at(j); 450 } 451 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index, cnt, trim_ranges)) { 452 ranges[++rp].set(match_int, dest, table_index, cnt); 453 } 454 } 455 jint highest = lo_index+(len-1); 456 assert(ranges[rp].hi() == highest, ""); 457 if (highest != max_jint) { 458 uint cnt = 1; 459 if (profile != NULL) { 460 cnt = profile->default_count() / (lo_index != min_jint ? 2 : 1); 461 } 462 if (!ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex, cnt, trim_ranges)) { 463 ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex, cnt); 464 } 465 } 466 assert(rp < len+2, "not too many ranges"); 467 468 if (trim_ranges) { 469 merge_ranges(ranges, rp); 470 } 471 472 // Safepoint in case if backward branch observed 473 if( makes_backward_branch && UseLoopSafepoints ) 474 add_safepoint(); 475 476 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]); 477 } 478 479 480 //------------------------------do_lookupswitch-------------------------------- 481 void Parse::do_lookupswitch() { 482 Node *lookup = pop(); // lookup value 483 // Get information about lookupswitch 484 int default_dest = iter().get_dest_table(0); 485 int len = iter().get_int_table(1); 486 487 if (len < 1) { // If this is a backward branch, add safepoint 488 maybe_add_safepoint(default_dest); 489 merge(default_dest); 490 return; 491 } 492 493 ciMethodData* methodData = method()->method_data(); 494 ciMultiBranchData* profile = NULL; 495 if (methodData->is_mature() && UseSwitchProfiling) { 496 ciProfileData* data = methodData->bci_to_data(bci()); 497 if (data != NULL && data->is_MultiBranchData()) { 498 profile = (ciMultiBranchData*)data; 499 } 500 } 501 bool trim_ranges = !method_data_update() && !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if); 502 503 // generate decision tree, using trichotomy when possible 504 jint* table = NEW_RESOURCE_ARRAY(jint, len*3); 505 { 506 for (int j = 0; j < len; j++) { 507 table[3*j+0] = iter().get_int_table(2+2*j); 508 table[3*j+1] = iter().get_dest_table(2+2*j+1); 509 table[3*j+2] = profile == NULL ? 1 : profile->count_at(j); 510 } 511 qsort(table, len, 3*sizeof(table[0]), jint_cmp); 512 } 513 514 float defaults = 0; 515 jint prev = min_jint; 516 for (int j = 0; j < len; j++) { 517 jint match_int = table[3*j+0]; 518 if (match_int != prev) { 519 defaults += (float)match_int - prev; 520 } 521 prev = match_int+1; 522 } 523 if (prev-1 != max_jint) { 524 defaults += (float)max_jint - prev + 1; 525 } 526 float default_cnt = 1; 527 if (profile != NULL) { 528 default_cnt = profile->default_count()/defaults; 529 } 530 531 int rnum = len*2+1; 532 bool makes_backward_branch = false; 533 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum); 534 int rp = -1; 535 for (int j = 0; j < len; j++) { 536 jint match_int = table[3*j+0]; 537 int dest = table[3*j+1]; 538 int cnt = table[3*j+2]; 539 int next_lo = rp < 0 ? min_jint : ranges[rp].hi()+1; 540 int table_index = method_data_update() ? j : NullTableIndex; 541 makes_backward_branch |= (dest <= bci()); 542 float c = default_cnt * ((float)match_int - next_lo); 543 if (match_int != next_lo && (rp < 0 || !ranges[rp].adjoinRange(next_lo, match_int-1, default_dest, NullTableIndex, c, trim_ranges))) { 544 assert(default_dest != never_reached, "sentinel value for dead destinations"); 545 ranges[++rp].setRange(next_lo, match_int-1, default_dest, NullTableIndex, c); 546 } 547 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index, cnt, trim_ranges)) { 548 assert(dest != never_reached, "sentinel value for dead destinations"); 549 ranges[++rp].set(match_int, dest, table_index, cnt); 550 } 551 } 552 jint highest = table[3*(len-1)]; 553 assert(ranges[rp].hi() == highest, ""); 554 if (highest != max_jint && 555 !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex, default_cnt * ((float)max_jint - highest), trim_ranges)) { 556 ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex, default_cnt * ((float)max_jint - highest)); 557 } 558 assert(rp < rnum, "not too many ranges"); 559 560 if (trim_ranges) { 561 merge_ranges(ranges, rp); 562 } 563 564 // Safepoint in case backward branch observed 565 if (makes_backward_branch && UseLoopSafepoints) 566 add_safepoint(); 567 568 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]); 569 } 570 571 static float if_prob(float taken_cnt, float total_cnt) { 572 assert(taken_cnt <= total_cnt, ""); 573 if (total_cnt == 0) { 574 return PROB_FAIR; 575 } 576 float p = taken_cnt / total_cnt; 577 return MIN2(MAX2(p, PROB_MIN), PROB_MAX); 578 } 579 580 static float if_cnt(float cnt) { 581 if (cnt == 0) { 582 return COUNT_UNKNOWN; 583 } 584 return cnt; 585 } 586 587 static float sum_of_cnts(SwitchRange *lo, SwitchRange *hi) { 588 float total_cnt = 0; 589 for (SwitchRange* sr = lo; sr <= hi; sr++) { 590 total_cnt += sr->cnt(); 591 } 592 return total_cnt; 593 } 594 595 class SwitchRanges : public ResourceObj { 596 public: 597 SwitchRange* _lo; 598 SwitchRange* _hi; 599 SwitchRange* _mid; 600 float _cost; 601 602 enum { 603 Start, 604 LeftDone, 605 RightDone, 606 Done 607 } _state; 608 609 SwitchRanges(SwitchRange *lo, SwitchRange *hi) 610 : _lo(lo), _hi(hi), _mid(NULL), 611 _cost(0), _state(Start) { 612 } 613 614 SwitchRanges() 615 : _lo(NULL), _hi(NULL), _mid(NULL), 616 _cost(0), _state(Start) {} 617 }; 618 619 // Estimate cost of performing a binary search on lo..hi 620 static float compute_tree_cost(SwitchRange *lo, SwitchRange *hi, float total_cnt) { 621 GrowableArray<SwitchRanges> tree; 622 SwitchRanges root(lo, hi); 623 tree.push(root); 624 625 float cost = 0; 626 do { 627 SwitchRanges& r = *tree.adr_at(tree.length()-1); 628 if (r._hi != r._lo) { 629 if (r._mid == NULL) { 630 float r_cnt = sum_of_cnts(r._lo, r._hi); 631 632 if (r_cnt == 0) { 633 tree.pop(); 634 cost = 0; 635 continue; 636 } 637 638 SwitchRange* mid = NULL; 639 mid = r._lo; 640 for (float cnt = 0; ; ) { 641 assert(mid <= r._hi, "out of bounds"); 642 cnt += mid->cnt(); 643 if (cnt > r_cnt / 2) { 644 break; 645 } 646 mid++; 647 } 648 assert(mid <= r._hi, "out of bounds"); 649 r._mid = mid; 650 r._cost = r_cnt / total_cnt; 651 } 652 r._cost += cost; 653 if (r._state < SwitchRanges::LeftDone && r._mid > r._lo) { 654 cost = 0; 655 r._state = SwitchRanges::LeftDone; 656 tree.push(SwitchRanges(r._lo, r._mid-1)); 657 } else if (r._state < SwitchRanges::RightDone) { 658 cost = 0; 659 r._state = SwitchRanges::RightDone; 660 tree.push(SwitchRanges(r._mid == r._lo ? r._mid+1 : r._mid, r._hi)); 661 } else { 662 tree.pop(); 663 cost = r._cost; 664 } 665 } else { 666 tree.pop(); 667 cost = r._cost; 668 } 669 } while (tree.length() > 0); 670 671 672 return cost; 673 } 674 675 // It sometimes pays off to test most common ranges before the binary search 676 void Parse::linear_search_switch_ranges(Node* key_val, SwitchRange*& lo, SwitchRange*& hi) { 677 uint nr = hi - lo + 1; 678 float total_cnt = sum_of_cnts(lo, hi); 679 680 float min = compute_tree_cost(lo, hi, total_cnt); 681 float extra = 1; 682 float sub = 0; 683 684 SwitchRange* array1 = lo; 685 SwitchRange* array2 = NEW_RESOURCE_ARRAY(SwitchRange, nr); 686 687 SwitchRange* ranges = NULL; 688 689 while (nr >= 2) { 690 assert(lo == array1 || lo == array2, "one the 2 already allocated arrays"); 691 ranges = (lo == array1) ? array2 : array1; 692 693 // Find highest frequency range 694 SwitchRange* candidate = lo; 695 for (SwitchRange* sr = lo+1; sr <= hi; sr++) { 696 if (sr->cnt() > candidate->cnt()) { 697 candidate = sr; 698 } 699 } 700 SwitchRange most_freq = *candidate; 701 if (most_freq.cnt() == 0) { 702 break; 703 } 704 705 // Copy remaining ranges into another array 706 int shift = 0; 707 for (uint i = 0; i < nr; i++) { 708 SwitchRange* sr = &lo[i]; 709 if (sr != candidate) { 710 ranges[i-shift] = *sr; 711 } else { 712 shift++; 713 if (i > 0 && i < nr-1) { 714 SwitchRange prev = lo[i-1]; 715 prev.setRange(prev.lo(), sr->hi(), prev.dest(), prev.table_index(), prev.cnt()); 716 if (prev.adjoin(lo[i+1])) { 717 shift++; 718 i++; 719 } 720 ranges[i-shift] = prev; 721 } 722 } 723 } 724 nr -= shift; 725 726 // Evaluate cost of testing the most common range and performing a 727 // binary search on the other ranges 728 float cost = extra + compute_tree_cost(&ranges[0], &ranges[nr-1], total_cnt); 729 if (cost >= min) { 730 break; 731 } 732 // swap arrays 733 lo = &ranges[0]; 734 hi = &ranges[nr-1]; 735 736 // It pays off: emit the test for the most common range 737 assert(most_freq.cnt() > 0, "must be taken"); 738 Node* val = _gvn.transform(new SubINode(key_val, _gvn.intcon(most_freq.lo()))); 739 Node* cmp = _gvn.transform(new CmpUNode(val, _gvn.intcon(most_freq.hi() - most_freq.lo()))); 740 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::le)); 741 IfNode* iff = create_and_map_if(control(), tst, if_prob(most_freq.cnt(), total_cnt), if_cnt(most_freq.cnt())); 742 jump_if_true_fork(iff, most_freq.dest(), most_freq.table_index(), false); 743 744 sub += most_freq.cnt() / total_cnt; 745 extra += 1 - sub; 746 min = cost; 747 } 748 } 749 750 //----------------------------create_jump_tables------------------------------- 751 bool Parse::create_jump_tables(Node* key_val, SwitchRange* lo, SwitchRange* hi) { 752 // Are jumptables enabled 753 if (!UseJumpTables) return false; 754 755 // Are jumptables supported 756 if (!Matcher::has_match_rule(Op_Jump)) return false; 757 758 // Don't make jump table if profiling 759 if (method_data_update()) return false; 760 761 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if); 762 763 // Decide if a guard is needed to lop off big ranges at either (or 764 // both) end(s) of the input set. We'll call this the default target 765 // even though we can't be sure that it is the true "default". 766 767 bool needs_guard = false; 768 int default_dest; 769 int64_t total_outlier_size = 0; 770 int64_t hi_size = ((int64_t)hi->hi()) - ((int64_t)hi->lo()) + 1; 771 int64_t lo_size = ((int64_t)lo->hi()) - ((int64_t)lo->lo()) + 1; 772 773 if (lo->dest() == hi->dest()) { 774 total_outlier_size = hi_size + lo_size; 775 default_dest = lo->dest(); 776 } else if (lo_size > hi_size) { 777 total_outlier_size = lo_size; 778 default_dest = lo->dest(); 779 } else { 780 total_outlier_size = hi_size; 781 default_dest = hi->dest(); 782 } 783 784 float total = sum_of_cnts(lo, hi); 785 float cost = compute_tree_cost(lo, hi, total); 786 787 // If a guard test will eliminate very sparse end ranges, then 788 // it is worth the cost of an extra jump. 789 float trimmed_cnt = 0; 790 if (total_outlier_size > (MaxJumpTableSparseness * 4)) { 791 needs_guard = true; 792 if (default_dest == lo->dest()) { 793 trimmed_cnt += lo->cnt(); 794 lo++; 795 } 796 if (default_dest == hi->dest()) { 797 trimmed_cnt += hi->cnt(); 798 hi--; 799 } 800 } 801 802 // Find the total number of cases and ranges 803 int64_t num_cases = ((int64_t)hi->hi()) - ((int64_t)lo->lo()) + 1; 804 int num_range = hi - lo + 1; 805 806 // Don't create table if: too large, too small, or too sparse. 807 if (num_cases > MaxJumpTableSize) 808 return false; 809 if (UseSwitchProfiling) { 810 // MinJumpTableSize is set so with a well balanced binary tree, 811 // when the number of ranges is MinJumpTableSize, it's cheaper to 812 // go through a JumpNode that a tree of IfNodes. Average cost of a 813 // tree of IfNodes with MinJumpTableSize is 814 // log2f(MinJumpTableSize) comparisons. So if the cost computed 815 // from profile data is less than log2f(MinJumpTableSize) then 816 // going with the binary search is cheaper. 817 if (cost < log2f(MinJumpTableSize)) { 818 return false; 819 } 820 } else { 821 if (num_cases < MinJumpTableSize) 822 return false; 823 } 824 if (num_cases > (MaxJumpTableSparseness * num_range)) 825 return false; 826 827 // Normalize table lookups to zero 828 int lowval = lo->lo(); 829 key_val = _gvn.transform( new SubINode(key_val, _gvn.intcon(lowval)) ); 830 831 // Generate a guard to protect against input keyvals that aren't 832 // in the switch domain. 833 if (needs_guard) { 834 Node* size = _gvn.intcon(num_cases); 835 Node* cmp = _gvn.transform(new CmpUNode(key_val, size)); 836 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::ge)); 837 IfNode* iff = create_and_map_if(control(), tst, if_prob(trimmed_cnt, total), if_cnt(trimmed_cnt)); 838 jump_if_true_fork(iff, default_dest, NullTableIndex, trim_ranges && trimmed_cnt == 0); 839 840 total -= trimmed_cnt; 841 } 842 843 // Create an ideal node JumpTable that has projections 844 // of all possible ranges for a switch statement 845 // The key_val input must be converted to a pointer offset and scaled. 846 // Compare Parse::array_addressing above. 847 848 // Clean the 32-bit int into a real 64-bit offset. 849 // Otherwise, the jint value 0 might turn into an offset of 0x0800000000. 850 const TypeInt* ikeytype = TypeInt::make(0, num_cases, Type::WidenMin); 851 // Make I2L conversion control dependent to prevent it from 852 // floating above the range check during loop optimizations. 853 key_val = C->conv_I2X_index(&_gvn, key_val, ikeytype, control()); 854 855 // Shift the value by wordsize so we have an index into the table, rather 856 // than a switch value 857 Node *shiftWord = _gvn.MakeConX(wordSize); 858 key_val = _gvn.transform( new MulXNode( key_val, shiftWord)); 859 860 // Create the JumpNode 861 Arena* arena = C->comp_arena(); 862 float* probs = (float*)arena->Amalloc(sizeof(float)*num_cases); 863 int i = 0; 864 if (total == 0) { 865 for (SwitchRange* r = lo; r <= hi; r++) { 866 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) { 867 probs[i] = 1.0F / num_cases; 868 } 869 } 870 } else { 871 for (SwitchRange* r = lo; r <= hi; r++) { 872 float prob = r->cnt()/total; 873 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) { 874 probs[i] = prob / (r->hi() - r->lo() + 1); 875 } 876 } 877 } 878 879 ciMethodData* methodData = method()->method_data(); 880 ciMultiBranchData* profile = NULL; 881 if (methodData->is_mature()) { 882 ciProfileData* data = methodData->bci_to_data(bci()); 883 if (data != NULL && data->is_MultiBranchData()) { 884 profile = (ciMultiBranchData*)data; 885 } 886 } 887 888 Node* jtn = _gvn.transform(new JumpNode(control(), key_val, num_cases, probs, profile == NULL ? COUNT_UNKNOWN : total)); 889 890 // These are the switch destinations hanging off the jumpnode 891 i = 0; 892 for (SwitchRange* r = lo; r <= hi; r++) { 893 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) { 894 Node* input = _gvn.transform(new JumpProjNode(jtn, i, r->dest(), (int)(j - lowval))); 895 { 896 PreserveJVMState pjvms(this); 897 set_control(input); 898 jump_if_always_fork(r->dest(), r->table_index(), trim_ranges && r->cnt() == 0); 899 } 900 } 901 } 902 assert(i == num_cases, "miscount of cases"); 903 stop_and_kill_map(); // no more uses for this JVMS 904 return true; 905 } 906 907 //----------------------------jump_switch_ranges------------------------------- 908 void Parse::jump_switch_ranges(Node* key_val, SwitchRange *lo, SwitchRange *hi, int switch_depth) { 909 Block* switch_block = block(); 910 bool trim_ranges = !method_data_update() && !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if); 911 912 if (switch_depth == 0) { 913 // Do special processing for the top-level call. 914 assert(lo->lo() == min_jint, "initial range must exhaust Type::INT"); 915 assert(hi->hi() == max_jint, "initial range must exhaust Type::INT"); 916 917 // Decrement pred-numbers for the unique set of nodes. 918 #ifdef ASSERT 919 if (!trim_ranges) { 920 // Ensure that the block's successors are a (duplicate-free) set. 921 int successors_counted = 0; // block occurrences in [hi..lo] 922 int unique_successors = switch_block->num_successors(); 923 for (int i = 0; i < unique_successors; i++) { 924 Block* target = switch_block->successor_at(i); 925 926 // Check that the set of successors is the same in both places. 927 int successors_found = 0; 928 for (SwitchRange* p = lo; p <= hi; p++) { 929 if (p->dest() == target->start()) successors_found++; 930 } 931 assert(successors_found > 0, "successor must be known"); 932 successors_counted += successors_found; 933 } 934 assert(successors_counted == (hi-lo)+1, "no unexpected successors"); 935 } 936 #endif 937 938 // Maybe prune the inputs, based on the type of key_val. 939 jint min_val = min_jint; 940 jint max_val = max_jint; 941 const TypeInt* ti = key_val->bottom_type()->isa_int(); 942 if (ti != NULL) { 943 min_val = ti->_lo; 944 max_val = ti->_hi; 945 assert(min_val <= max_val, "invalid int type"); 946 } 947 while (lo->hi() < min_val) { 948 lo++; 949 } 950 if (lo->lo() < min_val) { 951 lo->setRange(min_val, lo->hi(), lo->dest(), lo->table_index(), lo->cnt()); 952 } 953 while (hi->lo() > max_val) { 954 hi--; 955 } 956 if (hi->hi() > max_val) { 957 hi->setRange(hi->lo(), max_val, hi->dest(), hi->table_index(), hi->cnt()); 958 } 959 960 linear_search_switch_ranges(key_val, lo, hi); 961 } 962 963 #ifndef PRODUCT 964 if (switch_depth == 0) { 965 _max_switch_depth = 0; 966 _est_switch_depth = log2_intptr((hi-lo+1)-1)+1; 967 } 968 #endif 969 970 assert(lo <= hi, "must be a non-empty set of ranges"); 971 if (lo == hi) { 972 jump_if_always_fork(lo->dest(), lo->table_index(), trim_ranges && lo->cnt() == 0); 973 } else { 974 assert(lo->hi() == (lo+1)->lo()-1, "contiguous ranges"); 975 assert(hi->lo() == (hi-1)->hi()+1, "contiguous ranges"); 976 977 if (create_jump_tables(key_val, lo, hi)) return; 978 979 SwitchRange* mid = NULL; 980 float total_cnt = sum_of_cnts(lo, hi); 981 982 int nr = hi - lo + 1; 983 if (UseSwitchProfiling) { 984 // Don't keep the binary search tree balanced: pick up mid point 985 // that split frequencies in half. 986 float cnt = 0; 987 for (SwitchRange* sr = lo; sr <= hi; sr++) { 988 cnt += sr->cnt(); 989 if (cnt >= total_cnt / 2) { 990 mid = sr; 991 break; 992 } 993 } 994 } else { 995 mid = lo + nr/2; 996 997 // if there is an easy choice, pivot at a singleton: 998 if (nr > 3 && !mid->is_singleton() && (mid-1)->is_singleton()) mid--; 999 1000 assert(lo < mid && mid <= hi, "good pivot choice"); 1001 assert(nr != 2 || mid == hi, "should pick higher of 2"); 1002 assert(nr != 3 || mid == hi-1, "should pick middle of 3"); 1003 } 1004 1005 1006 Node *test_val = _gvn.intcon(mid == lo ? mid->hi() : mid->lo()); 1007 1008 if (mid->is_singleton()) { 1009 IfNode *iff_ne = jump_if_fork_int(key_val, test_val, BoolTest::ne, 1-if_prob(mid->cnt(), total_cnt), if_cnt(mid->cnt())); 1010 jump_if_false_fork(iff_ne, mid->dest(), mid->table_index(), trim_ranges && mid->cnt() == 0); 1011 1012 // Special Case: If there are exactly three ranges, and the high 1013 // and low range each go to the same place, omit the "gt" test, 1014 // since it will not discriminate anything. 1015 bool eq_test_only = (hi == lo+2 && hi->dest() == lo->dest() && mid == hi-1) || mid == lo; 1016 1017 // if there is a higher range, test for it and process it: 1018 if (mid < hi && !eq_test_only) { 1019 // two comparisons of same values--should enable 1 test for 2 branches 1020 // Use BoolTest::le instead of BoolTest::gt 1021 float cnt = sum_of_cnts(lo, mid-1); 1022 IfNode *iff_le = jump_if_fork_int(key_val, test_val, BoolTest::le, if_prob(cnt, total_cnt), if_cnt(cnt)); 1023 Node *iftrue = _gvn.transform( new IfTrueNode(iff_le) ); 1024 Node *iffalse = _gvn.transform( new IfFalseNode(iff_le) ); 1025 { PreserveJVMState pjvms(this); 1026 set_control(iffalse); 1027 jump_switch_ranges(key_val, mid+1, hi, switch_depth+1); 1028 } 1029 set_control(iftrue); 1030 } 1031 1032 } else { 1033 // mid is a range, not a singleton, so treat mid..hi as a unit 1034 float cnt = sum_of_cnts(mid == lo ? mid+1 : mid, hi); 1035 IfNode *iff_ge = jump_if_fork_int(key_val, test_val, mid == lo ? BoolTest::gt : BoolTest::ge, if_prob(cnt, total_cnt), if_cnt(cnt)); 1036 1037 // if there is a higher range, test for it and process it: 1038 if (mid == hi) { 1039 jump_if_true_fork(iff_ge, mid->dest(), mid->table_index(), trim_ranges && cnt == 0); 1040 } else { 1041 Node *iftrue = _gvn.transform( new IfTrueNode(iff_ge) ); 1042 Node *iffalse = _gvn.transform( new IfFalseNode(iff_ge) ); 1043 { PreserveJVMState pjvms(this); 1044 set_control(iftrue); 1045 jump_switch_ranges(key_val, mid == lo ? mid+1 : mid, hi, switch_depth+1); 1046 } 1047 set_control(iffalse); 1048 } 1049 } 1050 1051 // in any case, process the lower range 1052 if (mid == lo) { 1053 if (mid->is_singleton()) { 1054 jump_switch_ranges(key_val, lo+1, hi, switch_depth+1); 1055 } else { 1056 jump_if_always_fork(lo->dest(), lo->table_index(), trim_ranges && lo->cnt() == 0); 1057 } 1058 } else { 1059 jump_switch_ranges(key_val, lo, mid-1, switch_depth+1); 1060 } 1061 } 1062 1063 // Decrease pred_count for each successor after all is done. 1064 if (switch_depth == 0) { 1065 int unique_successors = switch_block->num_successors(); 1066 for (int i = 0; i < unique_successors; i++) { 1067 Block* target = switch_block->successor_at(i); 1068 // Throw away the pre-allocated path for each unique successor. 1069 target->next_path_num(); 1070 } 1071 } 1072 1073 #ifndef PRODUCT 1074 _max_switch_depth = MAX2(switch_depth, _max_switch_depth); 1075 if (TraceOptoParse && Verbose && WizardMode && switch_depth == 0) { 1076 SwitchRange* r; 1077 int nsing = 0; 1078 for( r = lo; r <= hi; r++ ) { 1079 if( r->is_singleton() ) nsing++; 1080 } 1081 tty->print(">>> "); 1082 _method->print_short_name(); 1083 tty->print_cr(" switch decision tree"); 1084 tty->print_cr(" %d ranges (%d singletons), max_depth=%d, est_depth=%d", 1085 (int) (hi-lo+1), nsing, _max_switch_depth, _est_switch_depth); 1086 if (_max_switch_depth > _est_switch_depth) { 1087 tty->print_cr("******** BAD SWITCH DEPTH ********"); 1088 } 1089 tty->print(" "); 1090 for( r = lo; r <= hi; r++ ) { 1091 r->print(); 1092 } 1093 tty->cr(); 1094 } 1095 #endif 1096 } 1097 1098 void Parse::modf() { 1099 Node *f2 = pop(); 1100 Node *f1 = pop(); 1101 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::modf_Type(), 1102 CAST_FROM_FN_PTR(address, SharedRuntime::frem), 1103 "frem", NULL, //no memory effects 1104 f1, f2); 1105 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0)); 1106 1107 push(res); 1108 } 1109 1110 void Parse::modd() { 1111 Node *d2 = pop_pair(); 1112 Node *d1 = pop_pair(); 1113 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::Math_DD_D_Type(), 1114 CAST_FROM_FN_PTR(address, SharedRuntime::drem), 1115 "drem", NULL, //no memory effects 1116 d1, top(), d2, top()); 1117 Node* res_d = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0)); 1118 1119 #ifdef ASSERT 1120 Node* res_top = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 1)); 1121 assert(res_top == top(), "second value must be top"); 1122 #endif 1123 1124 push_pair(res_d); 1125 } 1126 1127 void Parse::l2f() { 1128 Node* f2 = pop(); 1129 Node* f1 = pop(); 1130 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::l2f_Type(), 1131 CAST_FROM_FN_PTR(address, SharedRuntime::l2f), 1132 "l2f", NULL, //no memory effects 1133 f1, f2); 1134 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0)); 1135 1136 push(res); 1137 } 1138 1139 void Parse::do_irem() { 1140 // Must keep both values on the expression-stack during null-check 1141 zero_check_int(peek()); 1142 // Compile-time detect of null-exception? 1143 if (stopped()) return; 1144 1145 Node* b = pop(); 1146 Node* a = pop(); 1147 1148 const Type *t = _gvn.type(b); 1149 if (t != Type::TOP) { 1150 const TypeInt *ti = t->is_int(); 1151 if (ti->is_con()) { 1152 int divisor = ti->get_con(); 1153 // check for positive power of 2 1154 if (divisor > 0 && 1155 (divisor & ~(divisor-1)) == divisor) { 1156 // yes ! 1157 Node *mask = _gvn.intcon((divisor - 1)); 1158 // Sigh, must handle negative dividends 1159 Node *zero = _gvn.intcon(0); 1160 IfNode *ifff = jump_if_fork_int(a, zero, BoolTest::lt, PROB_FAIR, COUNT_UNKNOWN); 1161 Node *iff = _gvn.transform( new IfFalseNode(ifff) ); 1162 Node *ift = _gvn.transform( new IfTrueNode (ifff) ); 1163 Node *reg = jump_if_join(ift, iff); 1164 Node *phi = PhiNode::make(reg, NULL, TypeInt::INT); 1165 // Negative path; negate/and/negate 1166 Node *neg = _gvn.transform( new SubINode(zero, a) ); 1167 Node *andn= _gvn.transform( new AndINode(neg, mask) ); 1168 Node *negn= _gvn.transform( new SubINode(zero, andn) ); 1169 phi->init_req(1, negn); 1170 // Fast positive case 1171 Node *andx = _gvn.transform( new AndINode(a, mask) ); 1172 phi->init_req(2, andx); 1173 // Push the merge 1174 push( _gvn.transform(phi) ); 1175 return; 1176 } 1177 } 1178 } 1179 // Default case 1180 push( _gvn.transform( new ModINode(control(),a,b) ) ); 1181 } 1182 1183 // Handle jsr and jsr_w bytecode 1184 void Parse::do_jsr() { 1185 assert(bc() == Bytecodes::_jsr || bc() == Bytecodes::_jsr_w, "wrong bytecode"); 1186 1187 // Store information about current state, tagged with new _jsr_bci 1188 int return_bci = iter().next_bci(); 1189 int jsr_bci = (bc() == Bytecodes::_jsr) ? iter().get_dest() : iter().get_far_dest(); 1190 1191 // Update method data 1192 profile_taken_branch(jsr_bci); 1193 1194 // The way we do things now, there is only one successor block 1195 // for the jsr, because the target code is cloned by ciTypeFlow. 1196 Block* target = successor_for_bci(jsr_bci); 1197 1198 // What got pushed? 1199 const Type* ret_addr = target->peek(); 1200 assert(ret_addr->singleton(), "must be a constant (cloned jsr body)"); 1201 1202 // Effect on jsr on stack 1203 push(_gvn.makecon(ret_addr)); 1204 1205 // Flow to the jsr. 1206 merge(jsr_bci); 1207 } 1208 1209 // Handle ret bytecode 1210 void Parse::do_ret() { 1211 // Find to whom we return. 1212 assert(block()->num_successors() == 1, "a ret can only go one place now"); 1213 Block* target = block()->successor_at(0); 1214 assert(!target->is_ready(), "our arrival must be expected"); 1215 profile_ret(target->flow()->start()); 1216 int pnum = target->next_path_num(); 1217 merge_common(target, pnum); 1218 } 1219 1220 static bool has_injected_profile(BoolTest::mask btest, Node* test, int& taken, int& not_taken) { 1221 if (btest != BoolTest::eq && btest != BoolTest::ne) { 1222 // Only ::eq and ::ne are supported for profile injection. 1223 return false; 1224 } 1225 if (test->is_Cmp() && 1226 test->in(1)->Opcode() == Op_ProfileBoolean) { 1227 ProfileBooleanNode* profile = (ProfileBooleanNode*)test->in(1); 1228 int false_cnt = profile->false_count(); 1229 int true_cnt = profile->true_count(); 1230 1231 // Counts matching depends on the actual test operation (::eq or ::ne). 1232 // No need to scale the counts because profile injection was designed 1233 // to feed exact counts into VM. 1234 taken = (btest == BoolTest::eq) ? false_cnt : true_cnt; 1235 not_taken = (btest == BoolTest::eq) ? true_cnt : false_cnt; 1236 1237 profile->consume(); 1238 return true; 1239 } 1240 return false; 1241 } 1242 //--------------------------dynamic_branch_prediction-------------------------- 1243 // Try to gather dynamic branch prediction behavior. Return a probability 1244 // of the branch being taken and set the "cnt" field. Returns a -1.0 1245 // if we need to use static prediction for some reason. 1246 float Parse::dynamic_branch_prediction(float &cnt, BoolTest::mask btest, Node* test) { 1247 ResourceMark rm; 1248 1249 cnt = COUNT_UNKNOWN; 1250 1251 int taken = 0; 1252 int not_taken = 0; 1253 1254 bool use_mdo = !has_injected_profile(btest, test, taken, not_taken); 1255 1256 if (use_mdo) { 1257 // Use MethodData information if it is available 1258 // FIXME: free the ProfileData structure 1259 ciMethodData* methodData = method()->method_data(); 1260 if (!methodData->is_mature()) return PROB_UNKNOWN; 1261 ciProfileData* data = methodData->bci_to_data(bci()); 1262 if (data == NULL) { 1263 return PROB_UNKNOWN; 1264 } 1265 if (!data->is_JumpData()) return PROB_UNKNOWN; 1266 1267 // get taken and not taken values 1268 taken = data->as_JumpData()->taken(); 1269 not_taken = 0; 1270 if (data->is_BranchData()) { 1271 not_taken = data->as_BranchData()->not_taken(); 1272 } 1273 1274 // scale the counts to be commensurate with invocation counts: 1275 taken = method()->scale_count(taken); 1276 not_taken = method()->scale_count(not_taken); 1277 } 1278 1279 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful. 1280 // We also check that individual counters are positive first, otherwise the sum can become positive. 1281 if (taken < 0 || not_taken < 0 || taken + not_taken < 40) { 1282 if (C->log() != NULL) { 1283 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d'", iter().get_dest(), taken, not_taken); 1284 } 1285 return PROB_UNKNOWN; 1286 } 1287 1288 // Compute frequency that we arrive here 1289 float sum = taken + not_taken; 1290 // Adjust, if this block is a cloned private block but the 1291 // Jump counts are shared. Taken the private counts for 1292 // just this path instead of the shared counts. 1293 if( block()->count() > 0 ) 1294 sum = block()->count(); 1295 cnt = sum / FreqCountInvocations; 1296 1297 // Pin probability to sane limits 1298 float prob; 1299 if( !taken ) 1300 prob = (0+PROB_MIN) / 2; 1301 else if( !not_taken ) 1302 prob = (1+PROB_MAX) / 2; 1303 else { // Compute probability of true path 1304 prob = (float)taken / (float)(taken + not_taken); 1305 if (prob > PROB_MAX) prob = PROB_MAX; 1306 if (prob < PROB_MIN) prob = PROB_MIN; 1307 } 1308 1309 assert((cnt > 0.0f) && (prob > 0.0f), 1310 "Bad frequency assignment in if"); 1311 1312 if (C->log() != NULL) { 1313 const char* prob_str = NULL; 1314 if (prob >= PROB_MAX) prob_str = (prob == PROB_MAX) ? "max" : "always"; 1315 if (prob <= PROB_MIN) prob_str = (prob == PROB_MIN) ? "min" : "never"; 1316 char prob_str_buf[30]; 1317 if (prob_str == NULL) { 1318 sprintf(prob_str_buf, "%g", prob); 1319 prob_str = prob_str_buf; 1320 } 1321 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d' cnt='%f' prob='%s'", 1322 iter().get_dest(), taken, not_taken, cnt, prob_str); 1323 } 1324 return prob; 1325 } 1326 1327 //-----------------------------branch_prediction------------------------------- 1328 float Parse::branch_prediction(float& cnt, 1329 BoolTest::mask btest, 1330 int target_bci, 1331 Node* test) { 1332 float prob = dynamic_branch_prediction(cnt, btest, test); 1333 // If prob is unknown, switch to static prediction 1334 if (prob != PROB_UNKNOWN) return prob; 1335 1336 prob = PROB_FAIR; // Set default value 1337 if (btest == BoolTest::eq) // Exactly equal test? 1338 prob = PROB_STATIC_INFREQUENT; // Assume its relatively infrequent 1339 else if (btest == BoolTest::ne) 1340 prob = PROB_STATIC_FREQUENT; // Assume its relatively frequent 1341 1342 // If this is a conditional test guarding a backwards branch, 1343 // assume its a loop-back edge. Make it a likely taken branch. 1344 if (target_bci < bci()) { 1345 if (is_osr_parse()) { // Could be a hot OSR'd loop; force deopt 1346 // Since it's an OSR, we probably have profile data, but since 1347 // branch_prediction returned PROB_UNKNOWN, the counts are too small. 1348 // Let's make a special check here for completely zero counts. 1349 ciMethodData* methodData = method()->method_data(); 1350 if (!methodData->is_empty()) { 1351 ciProfileData* data = methodData->bci_to_data(bci()); 1352 // Only stop for truly zero counts, which mean an unknown part 1353 // of the OSR-ed method, and we want to deopt to gather more stats. 1354 // If you have ANY counts, then this loop is simply 'cold' relative 1355 // to the OSR loop. 1356 if (data == NULL || 1357 (data->as_BranchData()->taken() + data->as_BranchData()->not_taken() == 0)) { 1358 // This is the only way to return PROB_UNKNOWN: 1359 return PROB_UNKNOWN; 1360 } 1361 } 1362 } 1363 prob = PROB_STATIC_FREQUENT; // Likely to take backwards branch 1364 } 1365 1366 assert(prob != PROB_UNKNOWN, "must have some guess at this point"); 1367 return prob; 1368 } 1369 1370 // The magic constants are chosen so as to match the output of 1371 // branch_prediction() when the profile reports a zero taken count. 1372 // It is important to distinguish zero counts unambiguously, because 1373 // some branches (e.g., _213_javac.Assembler.eliminate) validly produce 1374 // very small but nonzero probabilities, which if confused with zero 1375 // counts would keep the program recompiling indefinitely. 1376 bool Parse::seems_never_taken(float prob) const { 1377 return prob < PROB_MIN; 1378 } 1379 1380 // True if the comparison seems to be the kind that will not change its 1381 // statistics from true to false. See comments in adjust_map_after_if. 1382 // This question is only asked along paths which are already 1383 // classifed as untaken (by seems_never_taken), so really, 1384 // if a path is never taken, its controlling comparison is 1385 // already acting in a stable fashion. If the comparison 1386 // seems stable, we will put an expensive uncommon trap 1387 // on the untaken path. 1388 bool Parse::seems_stable_comparison() const { 1389 if (C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if)) { 1390 return false; 1391 } 1392 return true; 1393 } 1394 1395 //-------------------------------repush_if_args-------------------------------- 1396 // Push arguments of an "if" bytecode back onto the stack by adjusting _sp. 1397 inline int Parse::repush_if_args() { 1398 if (PrintOpto && WizardMode) { 1399 tty->print("defending against excessive implicit null exceptions on %s @%d in ", 1400 Bytecodes::name(iter().cur_bc()), iter().cur_bci()); 1401 method()->print_name(); tty->cr(); 1402 } 1403 int bc_depth = - Bytecodes::depth(iter().cur_bc()); 1404 assert(bc_depth == 1 || bc_depth == 2, "only two kinds of branches"); 1405 DEBUG_ONLY(sync_jvms()); // argument(n) requires a synced jvms 1406 assert(argument(0) != NULL, "must exist"); 1407 assert(bc_depth == 1 || argument(1) != NULL, "two must exist"); 1408 inc_sp(bc_depth); 1409 return bc_depth; 1410 } 1411 1412 //----------------------------------do_ifnull---------------------------------- 1413 void Parse::do_ifnull(BoolTest::mask btest, Node *c) { 1414 int target_bci = iter().get_dest(); 1415 1416 Block* branch_block = successor_for_bci(target_bci); 1417 Block* next_block = successor_for_bci(iter().next_bci()); 1418 1419 float cnt; 1420 float prob = branch_prediction(cnt, btest, target_bci, c); 1421 if (prob == PROB_UNKNOWN) { 1422 // (An earlier version of do_ifnull omitted this trap for OSR methods.) 1423 if (PrintOpto && Verbose) { 1424 tty->print_cr("Never-taken edge stops compilation at bci %d", bci()); 1425 } 1426 repush_if_args(); // to gather stats on loop 1427 // We need to mark this branch as taken so that if we recompile we will 1428 // see that it is possible. In the tiered system the interpreter doesn't 1429 // do profiling and by the time we get to the lower tier from the interpreter 1430 // the path may be cold again. Make sure it doesn't look untaken 1431 profile_taken_branch(target_bci, !ProfileInterpreter); 1432 uncommon_trap(Deoptimization::Reason_unreached, 1433 Deoptimization::Action_reinterpret, 1434 NULL, "cold"); 1435 if (C->eliminate_boxing()) { 1436 // Mark the successor blocks as parsed 1437 branch_block->next_path_num(); 1438 next_block->next_path_num(); 1439 } 1440 return; 1441 } 1442 1443 NOT_PRODUCT(explicit_null_checks_inserted++); 1444 1445 // Generate real control flow 1446 Node *tst = _gvn.transform( new BoolNode( c, btest ) ); 1447 1448 // Sanity check the probability value 1449 assert(prob > 0.0f,"Bad probability in Parser"); 1450 // Need xform to put node in hash table 1451 IfNode *iff = create_and_xform_if( control(), tst, prob, cnt ); 1452 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser"); 1453 // True branch 1454 { PreserveJVMState pjvms(this); 1455 Node* iftrue = _gvn.transform( new IfTrueNode (iff) ); 1456 set_control(iftrue); 1457 1458 if (stopped()) { // Path is dead? 1459 NOT_PRODUCT(explicit_null_checks_elided++); 1460 if (C->eliminate_boxing()) { 1461 // Mark the successor block as parsed 1462 branch_block->next_path_num(); 1463 } 1464 } else { // Path is live. 1465 // Update method data 1466 profile_taken_branch(target_bci); 1467 adjust_map_after_if(btest, c, prob, branch_block, next_block); 1468 if (!stopped()) { 1469 merge(target_bci); 1470 } 1471 } 1472 } 1473 1474 // False branch 1475 Node* iffalse = _gvn.transform( new IfFalseNode(iff) ); 1476 set_control(iffalse); 1477 1478 if (stopped()) { // Path is dead? 1479 NOT_PRODUCT(explicit_null_checks_elided++); 1480 if (C->eliminate_boxing()) { 1481 // Mark the successor block as parsed 1482 next_block->next_path_num(); 1483 } 1484 } else { // Path is live. 1485 // Update method data 1486 profile_not_taken_branch(); 1487 adjust_map_after_if(BoolTest(btest).negate(), c, 1.0-prob, 1488 next_block, branch_block); 1489 } 1490 } 1491 1492 //------------------------------------do_if------------------------------------ 1493 void Parse::do_if(BoolTest::mask btest, Node* c) { 1494 int target_bci = iter().get_dest(); 1495 1496 Block* branch_block = successor_for_bci(target_bci); 1497 Block* next_block = successor_for_bci(iter().next_bci()); 1498 1499 float cnt; 1500 float prob = branch_prediction(cnt, btest, target_bci, c); 1501 float untaken_prob = 1.0 - prob; 1502 1503 if (prob == PROB_UNKNOWN) { 1504 if (PrintOpto && Verbose) { 1505 tty->print_cr("Never-taken edge stops compilation at bci %d", bci()); 1506 } 1507 repush_if_args(); // to gather stats on loop 1508 // We need to mark this branch as taken so that if we recompile we will 1509 // see that it is possible. In the tiered system the interpreter doesn't 1510 // do profiling and by the time we get to the lower tier from the interpreter 1511 // the path may be cold again. Make sure it doesn't look untaken 1512 profile_taken_branch(target_bci, !ProfileInterpreter); 1513 uncommon_trap(Deoptimization::Reason_unreached, 1514 Deoptimization::Action_reinterpret, 1515 NULL, "cold"); 1516 if (C->eliminate_boxing()) { 1517 // Mark the successor blocks as parsed 1518 branch_block->next_path_num(); 1519 next_block->next_path_num(); 1520 } 1521 return; 1522 } 1523 1524 // Sanity check the probability value 1525 assert(0.0f < prob && prob < 1.0f,"Bad probability in Parser"); 1526 1527 bool taken_if_true = true; 1528 // Convert BoolTest to canonical form: 1529 if (!BoolTest(btest).is_canonical()) { 1530 btest = BoolTest(btest).negate(); 1531 taken_if_true = false; 1532 // prob is NOT updated here; it remains the probability of the taken 1533 // path (as opposed to the prob of the path guarded by an 'IfTrueNode'). 1534 } 1535 assert(btest != BoolTest::eq, "!= is the only canonical exact test"); 1536 1537 Node* tst0 = new BoolNode(c, btest); 1538 Node* tst = _gvn.transform(tst0); 1539 BoolTest::mask taken_btest = BoolTest::illegal; 1540 BoolTest::mask untaken_btest = BoolTest::illegal; 1541 1542 if (tst->is_Bool()) { 1543 // Refresh c from the transformed bool node, since it may be 1544 // simpler than the original c. Also re-canonicalize btest. 1545 // This wins when (Bool ne (Conv2B p) 0) => (Bool ne (CmpP p NULL)). 1546 // That can arise from statements like: if (x instanceof C) ... 1547 if (tst != tst0) { 1548 // Canonicalize one more time since transform can change it. 1549 btest = tst->as_Bool()->_test._test; 1550 if (!BoolTest(btest).is_canonical()) { 1551 // Reverse edges one more time... 1552 tst = _gvn.transform( tst->as_Bool()->negate(&_gvn) ); 1553 btest = tst->as_Bool()->_test._test; 1554 assert(BoolTest(btest).is_canonical(), "sanity"); 1555 taken_if_true = !taken_if_true; 1556 } 1557 c = tst->in(1); 1558 } 1559 BoolTest::mask neg_btest = BoolTest(btest).negate(); 1560 taken_btest = taken_if_true ? btest : neg_btest; 1561 untaken_btest = taken_if_true ? neg_btest : btest; 1562 } 1563 1564 // Generate real control flow 1565 float true_prob = (taken_if_true ? prob : untaken_prob); 1566 IfNode* iff = create_and_map_if(control(), tst, true_prob, cnt); 1567 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser"); 1568 Node* taken_branch = new IfTrueNode(iff); 1569 Node* untaken_branch = new IfFalseNode(iff); 1570 if (!taken_if_true) { // Finish conversion to canonical form 1571 Node* tmp = taken_branch; 1572 taken_branch = untaken_branch; 1573 untaken_branch = tmp; 1574 } 1575 1576 // Branch is taken: 1577 { PreserveJVMState pjvms(this); 1578 taken_branch = _gvn.transform(taken_branch); 1579 set_control(taken_branch); 1580 1581 if (stopped()) { 1582 if (C->eliminate_boxing()) { 1583 // Mark the successor block as parsed 1584 branch_block->next_path_num(); 1585 } 1586 } else { 1587 // Update method data 1588 profile_taken_branch(target_bci); 1589 adjust_map_after_if(taken_btest, c, prob, branch_block, next_block); 1590 if (!stopped()) { 1591 merge(target_bci); 1592 } 1593 } 1594 } 1595 1596 untaken_branch = _gvn.transform(untaken_branch); 1597 set_control(untaken_branch); 1598 1599 // Branch not taken. 1600 if (stopped()) { 1601 if (C->eliminate_boxing()) { 1602 // Mark the successor block as parsed 1603 next_block->next_path_num(); 1604 } 1605 } else { 1606 // Update method data 1607 profile_not_taken_branch(); 1608 adjust_map_after_if(untaken_btest, c, untaken_prob, 1609 next_block, branch_block); 1610 } 1611 } 1612 1613 bool Parse::path_is_suitable_for_uncommon_trap(float prob) const { 1614 // Don't want to speculate on uncommon traps when running with -Xcomp 1615 if (!UseInterpreter) { 1616 return false; 1617 } 1618 return (seems_never_taken(prob) && seems_stable_comparison()); 1619 } 1620 1621 //----------------------------adjust_map_after_if------------------------------ 1622 // Adjust the JVM state to reflect the result of taking this path. 1623 // Basically, it means inspecting the CmpNode controlling this 1624 // branch, seeing how it constrains a tested value, and then 1625 // deciding if it's worth our while to encode this constraint 1626 // as graph nodes in the current abstract interpretation map. 1627 void Parse::adjust_map_after_if(BoolTest::mask btest, Node* c, float prob, 1628 Block* path, Block* other_path) { 1629 if (stopped() || !c->is_Cmp() || btest == BoolTest::illegal) 1630 return; // nothing to do 1631 1632 bool is_fallthrough = (path == successor_for_bci(iter().next_bci())); 1633 1634 if (path_is_suitable_for_uncommon_trap(prob)) { 1635 repush_if_args(); 1636 uncommon_trap(Deoptimization::Reason_unstable_if, 1637 Deoptimization::Action_reinterpret, 1638 NULL, 1639 (is_fallthrough ? "taken always" : "taken never")); 1640 return; 1641 } 1642 1643 Node* val = c->in(1); 1644 Node* con = c->in(2); 1645 const Type* tcon = _gvn.type(con); 1646 const Type* tval = _gvn.type(val); 1647 bool have_con = tcon->singleton(); 1648 if (tval->singleton()) { 1649 if (!have_con) { 1650 // Swap, so constant is in con. 1651 con = val; 1652 tcon = tval; 1653 val = c->in(2); 1654 tval = _gvn.type(val); 1655 btest = BoolTest(btest).commute(); 1656 have_con = true; 1657 } else { 1658 // Do we have two constants? Then leave well enough alone. 1659 have_con = false; 1660 } 1661 } 1662 if (!have_con) // remaining adjustments need a con 1663 return; 1664 1665 sharpen_type_after_if(btest, con, tcon, val, tval); 1666 } 1667 1668 1669 static Node* extract_obj_from_klass_load(PhaseGVN* gvn, Node* n) { 1670 Node* ldk; 1671 if (n->is_DecodeNKlass()) { 1672 if (n->in(1)->Opcode() != Op_LoadNKlass) { 1673 return NULL; 1674 } else { 1675 ldk = n->in(1); 1676 } 1677 } else if (n->Opcode() != Op_LoadKlass) { 1678 return NULL; 1679 } else { 1680 ldk = n; 1681 } 1682 assert(ldk != NULL && ldk->is_Load(), "should have found a LoadKlass or LoadNKlass node"); 1683 1684 Node* adr = ldk->in(MemNode::Address); 1685 intptr_t off = 0; 1686 Node* obj = AddPNode::Ideal_base_and_offset(adr, gvn, off); 1687 if (obj == NULL || off != oopDesc::klass_offset_in_bytes()) // loading oopDesc::_klass? 1688 return NULL; 1689 const TypePtr* tp = gvn->type(obj)->is_ptr(); 1690 if (tp == NULL || !(tp->isa_instptr() || tp->isa_aryptr())) // is obj a Java object ptr? 1691 return NULL; 1692 1693 return obj; 1694 } 1695 1696 void Parse::sharpen_type_after_if(BoolTest::mask btest, 1697 Node* con, const Type* tcon, 1698 Node* val, const Type* tval) { 1699 // Look for opportunities to sharpen the type of a node 1700 // whose klass is compared with a constant klass. 1701 if (btest == BoolTest::eq && tcon->isa_klassptr()) { 1702 Node* obj = extract_obj_from_klass_load(&_gvn, val); 1703 const TypeOopPtr* con_type = tcon->isa_klassptr()->as_instance_type(); 1704 if (obj != NULL && (con_type->isa_instptr() || con_type->isa_aryptr())) { 1705 // Found: 1706 // Bool(CmpP(LoadKlass(obj._klass), ConP(Foo.klass)), [eq]) 1707 // or the narrowOop equivalent. 1708 const Type* obj_type = _gvn.type(obj); 1709 const TypeOopPtr* tboth = obj_type->join_speculative(con_type)->isa_oopptr(); 1710 if (tboth != NULL && tboth->klass_is_exact() && tboth != obj_type && 1711 tboth->higher_equal(obj_type)) { 1712 // obj has to be of the exact type Foo if the CmpP succeeds. 1713 int obj_in_map = map()->find_edge(obj); 1714 JVMState* jvms = this->jvms(); 1715 if (obj_in_map >= 0 && 1716 (jvms->is_loc(obj_in_map) || jvms->is_stk(obj_in_map))) { 1717 TypeNode* ccast = new CheckCastPPNode(control(), obj, tboth); 1718 const Type* tcc = ccast->as_Type()->type(); 1719 assert(tcc != obj_type && tcc->higher_equal(obj_type), "must improve"); 1720 // Delay transform() call to allow recovery of pre-cast value 1721 // at the control merge. 1722 _gvn.set_type_bottom(ccast); 1723 record_for_igvn(ccast); 1724 // Here's the payoff. 1725 replace_in_map(obj, ccast); 1726 } 1727 } 1728 } 1729 } 1730 1731 int val_in_map = map()->find_edge(val); 1732 if (val_in_map < 0) return; // replace_in_map would be useless 1733 { 1734 JVMState* jvms = this->jvms(); 1735 if (!(jvms->is_loc(val_in_map) || 1736 jvms->is_stk(val_in_map))) 1737 return; // again, it would be useless 1738 } 1739 1740 // Check for a comparison to a constant, and "know" that the compared 1741 // value is constrained on this path. 1742 assert(tcon->singleton(), ""); 1743 ConstraintCastNode* ccast = NULL; 1744 Node* cast = NULL; 1745 1746 switch (btest) { 1747 case BoolTest::eq: // Constant test? 1748 { 1749 const Type* tboth = tcon->join_speculative(tval); 1750 if (tboth == tval) break; // Nothing to gain. 1751 if (tcon->isa_int()) { 1752 ccast = new CastIINode(val, tboth); 1753 } else if (tcon == TypePtr::NULL_PTR) { 1754 // Cast to null, but keep the pointer identity temporarily live. 1755 ccast = new CastPPNode(val, tboth); 1756 } else { 1757 const TypeF* tf = tcon->isa_float_constant(); 1758 const TypeD* td = tcon->isa_double_constant(); 1759 // Exclude tests vs float/double 0 as these could be 1760 // either +0 or -0. Just because you are equal to +0 1761 // doesn't mean you ARE +0! 1762 // Note, following code also replaces Long and Oop values. 1763 if ((!tf || tf->_f != 0.0) && 1764 (!td || td->_d != 0.0)) 1765 cast = con; // Replace non-constant val by con. 1766 } 1767 } 1768 break; 1769 1770 case BoolTest::ne: 1771 if (tcon == TypePtr::NULL_PTR) { 1772 cast = cast_not_null(val, false); 1773 } 1774 break; 1775 1776 default: 1777 // (At this point we could record int range types with CastII.) 1778 break; 1779 } 1780 1781 if (ccast != NULL) { 1782 const Type* tcc = ccast->as_Type()->type(); 1783 assert(tcc != tval && tcc->higher_equal(tval), "must improve"); 1784 // Delay transform() call to allow recovery of pre-cast value 1785 // at the control merge. 1786 ccast->set_req(0, control()); 1787 _gvn.set_type_bottom(ccast); 1788 record_for_igvn(ccast); 1789 cast = ccast; 1790 } 1791 1792 if (cast != NULL) { // Here's the payoff. 1793 replace_in_map(val, cast); 1794 } 1795 } 1796 1797 /** 1798 * Use speculative type to optimize CmpP node: if comparison is 1799 * against the low level class, cast the object to the speculative 1800 * type if any. CmpP should then go away. 1801 * 1802 * @param c expected CmpP node 1803 * @return result of CmpP on object casted to speculative type 1804 * 1805 */ 1806 Node* Parse::optimize_cmp_with_klass(Node* c) { 1807 // If this is transformed by the _gvn to a comparison with the low 1808 // level klass then we may be able to use speculation 1809 if (c->Opcode() == Op_CmpP && 1810 (c->in(1)->Opcode() == Op_LoadKlass || c->in(1)->Opcode() == Op_DecodeNKlass) && 1811 c->in(2)->is_Con()) { 1812 Node* load_klass = NULL; 1813 Node* decode = NULL; 1814 if (c->in(1)->Opcode() == Op_DecodeNKlass) { 1815 decode = c->in(1); 1816 load_klass = c->in(1)->in(1); 1817 } else { 1818 load_klass = c->in(1); 1819 } 1820 if (load_klass->in(2)->is_AddP()) { 1821 Node* addp = load_klass->in(2); 1822 Node* obj = addp->in(AddPNode::Address); 1823 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 1824 if (obj_type->speculative_type_not_null() != NULL) { 1825 ciKlass* k = obj_type->speculative_type(); 1826 inc_sp(2); 1827 obj = maybe_cast_profiled_obj(obj, k); 1828 dec_sp(2); 1829 // Make the CmpP use the casted obj 1830 addp = basic_plus_adr(obj, addp->in(AddPNode::Offset)); 1831 load_klass = load_klass->clone(); 1832 load_klass->set_req(2, addp); 1833 load_klass = _gvn.transform(load_klass); 1834 if (decode != NULL) { 1835 decode = decode->clone(); 1836 decode->set_req(1, load_klass); 1837 load_klass = _gvn.transform(decode); 1838 } 1839 c = c->clone(); 1840 c->set_req(1, load_klass); 1841 c = _gvn.transform(c); 1842 } 1843 } 1844 } 1845 return c; 1846 } 1847 1848 //------------------------------do_one_bytecode-------------------------------- 1849 // Parse this bytecode, and alter the Parsers JVM->Node mapping 1850 void Parse::do_one_bytecode() { 1851 Node *a, *b, *c, *d; // Handy temps 1852 BoolTest::mask btest; 1853 int i; 1854 1855 assert(!has_exceptions(), "bytecode entry state must be clear of throws"); 1856 1857 if (C->check_node_count(NodeLimitFudgeFactor * 5, 1858 "out of nodes parsing method")) { 1859 return; 1860 } 1861 1862 #ifdef ASSERT 1863 // for setting breakpoints 1864 if (TraceOptoParse) { 1865 tty->print(" @"); 1866 dump_bci(bci()); 1867 tty->cr(); 1868 } 1869 #endif 1870 1871 switch (bc()) { 1872 case Bytecodes::_nop: 1873 // do nothing 1874 break; 1875 case Bytecodes::_lconst_0: 1876 push_pair(longcon(0)); 1877 break; 1878 1879 case Bytecodes::_lconst_1: 1880 push_pair(longcon(1)); 1881 break; 1882 1883 case Bytecodes::_fconst_0: 1884 push(zerocon(T_FLOAT)); 1885 break; 1886 1887 case Bytecodes::_fconst_1: 1888 push(makecon(TypeF::ONE)); 1889 break; 1890 1891 case Bytecodes::_fconst_2: 1892 push(makecon(TypeF::make(2.0f))); 1893 break; 1894 1895 case Bytecodes::_dconst_0: 1896 push_pair(zerocon(T_DOUBLE)); 1897 break; 1898 1899 case Bytecodes::_dconst_1: 1900 push_pair(makecon(TypeD::ONE)); 1901 break; 1902 1903 case Bytecodes::_iconst_m1:push(intcon(-1)); break; 1904 case Bytecodes::_iconst_0: push(intcon( 0)); break; 1905 case Bytecodes::_iconst_1: push(intcon( 1)); break; 1906 case Bytecodes::_iconst_2: push(intcon( 2)); break; 1907 case Bytecodes::_iconst_3: push(intcon( 3)); break; 1908 case Bytecodes::_iconst_4: push(intcon( 4)); break; 1909 case Bytecodes::_iconst_5: push(intcon( 5)); break; 1910 case Bytecodes::_bipush: push(intcon(iter().get_constant_u1())); break; 1911 case Bytecodes::_sipush: push(intcon(iter().get_constant_u2())); break; 1912 case Bytecodes::_aconst_null: push(null()); break; 1913 case Bytecodes::_ldc: 1914 case Bytecodes::_ldc_w: 1915 case Bytecodes::_ldc2_w: 1916 // If the constant is unresolved, run this BC once in the interpreter. 1917 { 1918 ciConstant constant = iter().get_constant(); 1919 if (!constant.is_valid() || 1920 (constant.basic_type() == T_OBJECT && 1921 !constant.as_object()->is_loaded())) { 1922 int index = iter().get_constant_pool_index(); 1923 constantTag tag = iter().get_constant_pool_tag(index); 1924 uncommon_trap(Deoptimization::make_trap_request 1925 (Deoptimization::Reason_unloaded, 1926 Deoptimization::Action_reinterpret, 1927 index), 1928 NULL, tag.internal_name()); 1929 break; 1930 } 1931 assert(constant.basic_type() != T_OBJECT || constant.as_object()->is_instance(), 1932 "must be java_mirror of klass"); 1933 const Type* con_type = Type::make_from_constant(constant); 1934 if (con_type != NULL) { 1935 push_node(con_type->basic_type(), makecon(con_type)); 1936 } 1937 } 1938 1939 break; 1940 1941 case Bytecodes::_aload_0: 1942 push( local(0) ); 1943 break; 1944 case Bytecodes::_aload_1: 1945 push( local(1) ); 1946 break; 1947 case Bytecodes::_aload_2: 1948 push( local(2) ); 1949 break; 1950 case Bytecodes::_aload_3: 1951 push( local(3) ); 1952 break; 1953 case Bytecodes::_aload: 1954 push( local(iter().get_index()) ); 1955 break; 1956 1957 case Bytecodes::_fload_0: 1958 case Bytecodes::_iload_0: 1959 push( local(0) ); 1960 break; 1961 case Bytecodes::_fload_1: 1962 case Bytecodes::_iload_1: 1963 push( local(1) ); 1964 break; 1965 case Bytecodes::_fload_2: 1966 case Bytecodes::_iload_2: 1967 push( local(2) ); 1968 break; 1969 case Bytecodes::_fload_3: 1970 case Bytecodes::_iload_3: 1971 push( local(3) ); 1972 break; 1973 case Bytecodes::_fload: 1974 case Bytecodes::_iload: 1975 push( local(iter().get_index()) ); 1976 break; 1977 case Bytecodes::_lload_0: 1978 push_pair_local( 0 ); 1979 break; 1980 case Bytecodes::_lload_1: 1981 push_pair_local( 1 ); 1982 break; 1983 case Bytecodes::_lload_2: 1984 push_pair_local( 2 ); 1985 break; 1986 case Bytecodes::_lload_3: 1987 push_pair_local( 3 ); 1988 break; 1989 case Bytecodes::_lload: 1990 push_pair_local( iter().get_index() ); 1991 break; 1992 1993 case Bytecodes::_dload_0: 1994 push_pair_local(0); 1995 break; 1996 case Bytecodes::_dload_1: 1997 push_pair_local(1); 1998 break; 1999 case Bytecodes::_dload_2: 2000 push_pair_local(2); 2001 break; 2002 case Bytecodes::_dload_3: 2003 push_pair_local(3); 2004 break; 2005 case Bytecodes::_dload: 2006 push_pair_local(iter().get_index()); 2007 break; 2008 case Bytecodes::_fstore_0: 2009 case Bytecodes::_istore_0: 2010 case Bytecodes::_astore_0: 2011 set_local( 0, pop() ); 2012 break; 2013 case Bytecodes::_fstore_1: 2014 case Bytecodes::_istore_1: 2015 case Bytecodes::_astore_1: 2016 set_local( 1, pop() ); 2017 break; 2018 case Bytecodes::_fstore_2: 2019 case Bytecodes::_istore_2: 2020 case Bytecodes::_astore_2: 2021 set_local( 2, pop() ); 2022 break; 2023 case Bytecodes::_fstore_3: 2024 case Bytecodes::_istore_3: 2025 case Bytecodes::_astore_3: 2026 set_local( 3, pop() ); 2027 break; 2028 case Bytecodes::_fstore: 2029 case Bytecodes::_istore: 2030 case Bytecodes::_astore: 2031 set_local( iter().get_index(), pop() ); 2032 break; 2033 // long stores 2034 case Bytecodes::_lstore_0: 2035 set_pair_local( 0, pop_pair() ); 2036 break; 2037 case Bytecodes::_lstore_1: 2038 set_pair_local( 1, pop_pair() ); 2039 break; 2040 case Bytecodes::_lstore_2: 2041 set_pair_local( 2, pop_pair() ); 2042 break; 2043 case Bytecodes::_lstore_3: 2044 set_pair_local( 3, pop_pair() ); 2045 break; 2046 case Bytecodes::_lstore: 2047 set_pair_local( iter().get_index(), pop_pair() ); 2048 break; 2049 2050 // double stores 2051 case Bytecodes::_dstore_0: 2052 set_pair_local( 0, dstore_rounding(pop_pair()) ); 2053 break; 2054 case Bytecodes::_dstore_1: 2055 set_pair_local( 1, dstore_rounding(pop_pair()) ); 2056 break; 2057 case Bytecodes::_dstore_2: 2058 set_pair_local( 2, dstore_rounding(pop_pair()) ); 2059 break; 2060 case Bytecodes::_dstore_3: 2061 set_pair_local( 3, dstore_rounding(pop_pair()) ); 2062 break; 2063 case Bytecodes::_dstore: 2064 set_pair_local( iter().get_index(), dstore_rounding(pop_pair()) ); 2065 break; 2066 2067 case Bytecodes::_pop: dec_sp(1); break; 2068 case Bytecodes::_pop2: dec_sp(2); break; 2069 case Bytecodes::_swap: 2070 a = pop(); 2071 b = pop(); 2072 push(a); 2073 push(b); 2074 break; 2075 case Bytecodes::_dup: 2076 a = pop(); 2077 push(a); 2078 push(a); 2079 break; 2080 case Bytecodes::_dup_x1: 2081 a = pop(); 2082 b = pop(); 2083 push( a ); 2084 push( b ); 2085 push( a ); 2086 break; 2087 case Bytecodes::_dup_x2: 2088 a = pop(); 2089 b = pop(); 2090 c = pop(); 2091 push( a ); 2092 push( c ); 2093 push( b ); 2094 push( a ); 2095 break; 2096 case Bytecodes::_dup2: 2097 a = pop(); 2098 b = pop(); 2099 push( b ); 2100 push( a ); 2101 push( b ); 2102 push( a ); 2103 break; 2104 2105 case Bytecodes::_dup2_x1: 2106 // before: .. c, b, a 2107 // after: .. b, a, c, b, a 2108 // not tested 2109 a = pop(); 2110 b = pop(); 2111 c = pop(); 2112 push( b ); 2113 push( a ); 2114 push( c ); 2115 push( b ); 2116 push( a ); 2117 break; 2118 case Bytecodes::_dup2_x2: 2119 // before: .. d, c, b, a 2120 // after: .. b, a, d, c, b, a 2121 // not tested 2122 a = pop(); 2123 b = pop(); 2124 c = pop(); 2125 d = pop(); 2126 push( b ); 2127 push( a ); 2128 push( d ); 2129 push( c ); 2130 push( b ); 2131 push( a ); 2132 break; 2133 2134 case Bytecodes::_arraylength: { 2135 // Must do null-check with value on expression stack 2136 Node *ary = null_check(peek(), T_ARRAY); 2137 // Compile-time detect of null-exception? 2138 if (stopped()) return; 2139 a = pop(); 2140 push(load_array_length(a)); 2141 break; 2142 } 2143 2144 case Bytecodes::_baload: array_load(T_BYTE); break; 2145 case Bytecodes::_caload: array_load(T_CHAR); break; 2146 case Bytecodes::_iaload: array_load(T_INT); break; 2147 case Bytecodes::_saload: array_load(T_SHORT); break; 2148 case Bytecodes::_faload: array_load(T_FLOAT); break; 2149 case Bytecodes::_aaload: array_load(T_OBJECT); break; 2150 case Bytecodes::_laload: { 2151 a = array_addressing(T_LONG, 0); 2152 if (stopped()) return; // guaranteed null or range check 2153 dec_sp(2); // Pop array and index 2154 push_pair(make_load(control(), a, TypeLong::LONG, T_LONG, TypeAryPtr::LONGS, MemNode::unordered)); 2155 break; 2156 } 2157 case Bytecodes::_daload: { 2158 a = array_addressing(T_DOUBLE, 0); 2159 if (stopped()) return; // guaranteed null or range check 2160 dec_sp(2); // Pop array and index 2161 push_pair(make_load(control(), a, Type::DOUBLE, T_DOUBLE, TypeAryPtr::DOUBLES, MemNode::unordered)); 2162 break; 2163 } 2164 case Bytecodes::_bastore: array_store(T_BYTE); break; 2165 case Bytecodes::_castore: array_store(T_CHAR); break; 2166 case Bytecodes::_iastore: array_store(T_INT); break; 2167 case Bytecodes::_sastore: array_store(T_SHORT); break; 2168 case Bytecodes::_fastore: array_store(T_FLOAT); break; 2169 case Bytecodes::_aastore: { 2170 d = array_addressing(T_OBJECT, 1); 2171 if (stopped()) return; // guaranteed null or range check 2172 array_store_check(); 2173 c = pop(); // Oop to store 2174 b = pop(); // index (already used) 2175 a = pop(); // the array itself 2176 const TypeOopPtr* elemtype = _gvn.type(a)->is_aryptr()->elem()->make_oopptr(); 2177 const TypeAryPtr* adr_type = TypeAryPtr::OOPS; 2178 Node* store = store_oop_to_array(control(), a, d, adr_type, c, elemtype, T_OBJECT, 2179 StoreNode::release_if_reference(T_OBJECT)); 2180 break; 2181 } 2182 case Bytecodes::_lastore: { 2183 a = array_addressing(T_LONG, 2); 2184 if (stopped()) return; // guaranteed null or range check 2185 c = pop_pair(); 2186 dec_sp(2); // Pop array and index 2187 store_to_memory(control(), a, c, T_LONG, TypeAryPtr::LONGS, MemNode::unordered); 2188 break; 2189 } 2190 case Bytecodes::_dastore: { 2191 a = array_addressing(T_DOUBLE, 2); 2192 if (stopped()) return; // guaranteed null or range check 2193 c = pop_pair(); 2194 dec_sp(2); // Pop array and index 2195 c = dstore_rounding(c); 2196 store_to_memory(control(), a, c, T_DOUBLE, TypeAryPtr::DOUBLES, MemNode::unordered); 2197 break; 2198 } 2199 case Bytecodes::_getfield: 2200 do_getfield(); 2201 break; 2202 2203 case Bytecodes::_getstatic: 2204 do_getstatic(); 2205 break; 2206 2207 case Bytecodes::_putfield: 2208 do_putfield(); 2209 break; 2210 2211 case Bytecodes::_putstatic: 2212 do_putstatic(); 2213 break; 2214 2215 case Bytecodes::_irem: 2216 do_irem(); 2217 break; 2218 case Bytecodes::_idiv: 2219 // Must keep both values on the expression-stack during null-check 2220 zero_check_int(peek()); 2221 // Compile-time detect of null-exception? 2222 if (stopped()) return; 2223 b = pop(); 2224 a = pop(); 2225 push( _gvn.transform( new DivINode(control(),a,b) ) ); 2226 break; 2227 case Bytecodes::_imul: 2228 b = pop(); a = pop(); 2229 push( _gvn.transform( new MulINode(a,b) ) ); 2230 break; 2231 case Bytecodes::_iadd: 2232 b = pop(); a = pop(); 2233 push( _gvn.transform( new AddINode(a,b) ) ); 2234 break; 2235 case Bytecodes::_ineg: 2236 a = pop(); 2237 push( _gvn.transform( new SubINode(_gvn.intcon(0),a)) ); 2238 break; 2239 case Bytecodes::_isub: 2240 b = pop(); a = pop(); 2241 push( _gvn.transform( new SubINode(a,b) ) ); 2242 break; 2243 case Bytecodes::_iand: 2244 b = pop(); a = pop(); 2245 push( _gvn.transform( new AndINode(a,b) ) ); 2246 break; 2247 case Bytecodes::_ior: 2248 b = pop(); a = pop(); 2249 push( _gvn.transform( new OrINode(a,b) ) ); 2250 break; 2251 case Bytecodes::_ixor: 2252 b = pop(); a = pop(); 2253 push( _gvn.transform( new XorINode(a,b) ) ); 2254 break; 2255 case Bytecodes::_ishl: 2256 b = pop(); a = pop(); 2257 push( _gvn.transform( new LShiftINode(a,b) ) ); 2258 break; 2259 case Bytecodes::_ishr: 2260 b = pop(); a = pop(); 2261 push( _gvn.transform( new RShiftINode(a,b) ) ); 2262 break; 2263 case Bytecodes::_iushr: 2264 b = pop(); a = pop(); 2265 push( _gvn.transform( new URShiftINode(a,b) ) ); 2266 break; 2267 2268 case Bytecodes::_fneg: 2269 a = pop(); 2270 b = _gvn.transform(new NegFNode (a)); 2271 push(b); 2272 break; 2273 2274 case Bytecodes::_fsub: 2275 b = pop(); 2276 a = pop(); 2277 c = _gvn.transform( new SubFNode(a,b) ); 2278 d = precision_rounding(c); 2279 push( d ); 2280 break; 2281 2282 case Bytecodes::_fadd: 2283 b = pop(); 2284 a = pop(); 2285 c = _gvn.transform( new AddFNode(a,b) ); 2286 d = precision_rounding(c); 2287 push( d ); 2288 break; 2289 2290 case Bytecodes::_fmul: 2291 b = pop(); 2292 a = pop(); 2293 c = _gvn.transform( new MulFNode(a,b) ); 2294 d = precision_rounding(c); 2295 push( d ); 2296 break; 2297 2298 case Bytecodes::_fdiv: 2299 b = pop(); 2300 a = pop(); 2301 c = _gvn.transform( new DivFNode(0,a,b) ); 2302 d = precision_rounding(c); 2303 push( d ); 2304 break; 2305 2306 case Bytecodes::_frem: 2307 if (Matcher::has_match_rule(Op_ModF)) { 2308 // Generate a ModF node. 2309 b = pop(); 2310 a = pop(); 2311 c = _gvn.transform( new ModFNode(0,a,b) ); 2312 d = precision_rounding(c); 2313 push( d ); 2314 } 2315 else { 2316 // Generate a call. 2317 modf(); 2318 } 2319 break; 2320 2321 case Bytecodes::_fcmpl: 2322 b = pop(); 2323 a = pop(); 2324 c = _gvn.transform( new CmpF3Node( a, b)); 2325 push(c); 2326 break; 2327 case Bytecodes::_fcmpg: 2328 b = pop(); 2329 a = pop(); 2330 2331 // Same as fcmpl but need to flip the unordered case. Swap the inputs, 2332 // which negates the result sign except for unordered. Flip the unordered 2333 // as well by using CmpF3 which implements unordered-lesser instead of 2334 // unordered-greater semantics. Finally, commute the result bits. Result 2335 // is same as using a CmpF3Greater except we did it with CmpF3 alone. 2336 c = _gvn.transform( new CmpF3Node( b, a)); 2337 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) ); 2338 push(c); 2339 break; 2340 2341 case Bytecodes::_f2i: 2342 a = pop(); 2343 push(_gvn.transform(new ConvF2INode(a))); 2344 break; 2345 2346 case Bytecodes::_d2i: 2347 a = pop_pair(); 2348 b = _gvn.transform(new ConvD2INode(a)); 2349 push( b ); 2350 break; 2351 2352 case Bytecodes::_f2d: 2353 a = pop(); 2354 b = _gvn.transform( new ConvF2DNode(a)); 2355 push_pair( b ); 2356 break; 2357 2358 case Bytecodes::_d2f: 2359 a = pop_pair(); 2360 b = _gvn.transform( new ConvD2FNode(a)); 2361 // This breaks _227_mtrt (speed & correctness) and _222_mpegaudio (speed) 2362 //b = _gvn.transform(new RoundFloatNode(0, b) ); 2363 push( b ); 2364 break; 2365 2366 case Bytecodes::_l2f: 2367 if (Matcher::convL2FSupported()) { 2368 a = pop_pair(); 2369 b = _gvn.transform( new ConvL2FNode(a)); 2370 // For i486.ad, FILD doesn't restrict precision to 24 or 53 bits. 2371 // Rather than storing the result into an FP register then pushing 2372 // out to memory to round, the machine instruction that implements 2373 // ConvL2D is responsible for rounding. 2374 // c = precision_rounding(b); 2375 c = _gvn.transform(b); 2376 push(c); 2377 } else { 2378 l2f(); 2379 } 2380 break; 2381 2382 case Bytecodes::_l2d: 2383 a = pop_pair(); 2384 b = _gvn.transform( new ConvL2DNode(a)); 2385 // For i486.ad, rounding is always necessary (see _l2f above). 2386 // c = dprecision_rounding(b); 2387 c = _gvn.transform(b); 2388 push_pair(c); 2389 break; 2390 2391 case Bytecodes::_f2l: 2392 a = pop(); 2393 b = _gvn.transform( new ConvF2LNode(a)); 2394 push_pair(b); 2395 break; 2396 2397 case Bytecodes::_d2l: 2398 a = pop_pair(); 2399 b = _gvn.transform( new ConvD2LNode(a)); 2400 push_pair(b); 2401 break; 2402 2403 case Bytecodes::_dsub: 2404 b = pop_pair(); 2405 a = pop_pair(); 2406 c = _gvn.transform( new SubDNode(a,b) ); 2407 d = dprecision_rounding(c); 2408 push_pair( d ); 2409 break; 2410 2411 case Bytecodes::_dadd: 2412 b = pop_pair(); 2413 a = pop_pair(); 2414 c = _gvn.transform( new AddDNode(a,b) ); 2415 d = dprecision_rounding(c); 2416 push_pair( d ); 2417 break; 2418 2419 case Bytecodes::_dmul: 2420 b = pop_pair(); 2421 a = pop_pair(); 2422 c = _gvn.transform( new MulDNode(a,b) ); 2423 d = dprecision_rounding(c); 2424 push_pair( d ); 2425 break; 2426 2427 case Bytecodes::_ddiv: 2428 b = pop_pair(); 2429 a = pop_pair(); 2430 c = _gvn.transform( new DivDNode(0,a,b) ); 2431 d = dprecision_rounding(c); 2432 push_pair( d ); 2433 break; 2434 2435 case Bytecodes::_dneg: 2436 a = pop_pair(); 2437 b = _gvn.transform(new NegDNode (a)); 2438 push_pair(b); 2439 break; 2440 2441 case Bytecodes::_drem: 2442 if (Matcher::has_match_rule(Op_ModD)) { 2443 // Generate a ModD node. 2444 b = pop_pair(); 2445 a = pop_pair(); 2446 // a % b 2447 2448 c = _gvn.transform( new ModDNode(0,a,b) ); 2449 d = dprecision_rounding(c); 2450 push_pair( d ); 2451 } 2452 else { 2453 // Generate a call. 2454 modd(); 2455 } 2456 break; 2457 2458 case Bytecodes::_dcmpl: 2459 b = pop_pair(); 2460 a = pop_pair(); 2461 c = _gvn.transform( new CmpD3Node( a, b)); 2462 push(c); 2463 break; 2464 2465 case Bytecodes::_dcmpg: 2466 b = pop_pair(); 2467 a = pop_pair(); 2468 // Same as dcmpl but need to flip the unordered case. 2469 // Commute the inputs, which negates the result sign except for unordered. 2470 // Flip the unordered as well by using CmpD3 which implements 2471 // unordered-lesser instead of unordered-greater semantics. 2472 // Finally, negate the result bits. Result is same as using a 2473 // CmpD3Greater except we did it with CmpD3 alone. 2474 c = _gvn.transform( new CmpD3Node( b, a)); 2475 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) ); 2476 push(c); 2477 break; 2478 2479 2480 // Note for longs -> lo word is on TOS, hi word is on TOS - 1 2481 case Bytecodes::_land: 2482 b = pop_pair(); 2483 a = pop_pair(); 2484 c = _gvn.transform( new AndLNode(a,b) ); 2485 push_pair(c); 2486 break; 2487 case Bytecodes::_lor: 2488 b = pop_pair(); 2489 a = pop_pair(); 2490 c = _gvn.transform( new OrLNode(a,b) ); 2491 push_pair(c); 2492 break; 2493 case Bytecodes::_lxor: 2494 b = pop_pair(); 2495 a = pop_pair(); 2496 c = _gvn.transform( new XorLNode(a,b) ); 2497 push_pair(c); 2498 break; 2499 2500 case Bytecodes::_lshl: 2501 b = pop(); // the shift count 2502 a = pop_pair(); // value to be shifted 2503 c = _gvn.transform( new LShiftLNode(a,b) ); 2504 push_pair(c); 2505 break; 2506 case Bytecodes::_lshr: 2507 b = pop(); // the shift count 2508 a = pop_pair(); // value to be shifted 2509 c = _gvn.transform( new RShiftLNode(a,b) ); 2510 push_pair(c); 2511 break; 2512 case Bytecodes::_lushr: 2513 b = pop(); // the shift count 2514 a = pop_pair(); // value to be shifted 2515 c = _gvn.transform( new URShiftLNode(a,b) ); 2516 push_pair(c); 2517 break; 2518 case Bytecodes::_lmul: 2519 b = pop_pair(); 2520 a = pop_pair(); 2521 c = _gvn.transform( new MulLNode(a,b) ); 2522 push_pair(c); 2523 break; 2524 2525 case Bytecodes::_lrem: 2526 // Must keep both values on the expression-stack during null-check 2527 assert(peek(0) == top(), "long word order"); 2528 zero_check_long(peek(1)); 2529 // Compile-time detect of null-exception? 2530 if (stopped()) return; 2531 b = pop_pair(); 2532 a = pop_pair(); 2533 c = _gvn.transform( new ModLNode(control(),a,b) ); 2534 push_pair(c); 2535 break; 2536 2537 case Bytecodes::_ldiv: 2538 // Must keep both values on the expression-stack during null-check 2539 assert(peek(0) == top(), "long word order"); 2540 zero_check_long(peek(1)); 2541 // Compile-time detect of null-exception? 2542 if (stopped()) return; 2543 b = pop_pair(); 2544 a = pop_pair(); 2545 c = _gvn.transform( new DivLNode(control(),a,b) ); 2546 push_pair(c); 2547 break; 2548 2549 case Bytecodes::_ladd: 2550 b = pop_pair(); 2551 a = pop_pair(); 2552 c = _gvn.transform( new AddLNode(a,b) ); 2553 push_pair(c); 2554 break; 2555 case Bytecodes::_lsub: 2556 b = pop_pair(); 2557 a = pop_pair(); 2558 c = _gvn.transform( new SubLNode(a,b) ); 2559 push_pair(c); 2560 break; 2561 case Bytecodes::_lcmp: 2562 // Safepoints are now inserted _before_ branches. The long-compare 2563 // bytecode painfully produces a 3-way value (-1,0,+1) which requires a 2564 // slew of control flow. These are usually followed by a CmpI vs zero and 2565 // a branch; this pattern then optimizes to the obvious long-compare and 2566 // branch. However, if the branch is backwards there's a Safepoint 2567 // inserted. The inserted Safepoint captures the JVM state at the 2568 // pre-branch point, i.e. it captures the 3-way value. Thus if a 2569 // long-compare is used to control a loop the debug info will force 2570 // computation of the 3-way value, even though the generated code uses a 2571 // long-compare and branch. We try to rectify the situation by inserting 2572 // a SafePoint here and have it dominate and kill the safepoint added at a 2573 // following backwards branch. At this point the JVM state merely holds 2 2574 // longs but not the 3-way value. 2575 if( UseLoopSafepoints ) { 2576 switch( iter().next_bc() ) { 2577 case Bytecodes::_ifgt: 2578 case Bytecodes::_iflt: 2579 case Bytecodes::_ifge: 2580 case Bytecodes::_ifle: 2581 case Bytecodes::_ifne: 2582 case Bytecodes::_ifeq: 2583 // If this is a backwards branch in the bytecodes, add Safepoint 2584 maybe_add_safepoint(iter().next_get_dest()); 2585 default: 2586 break; 2587 } 2588 } 2589 b = pop_pair(); 2590 a = pop_pair(); 2591 c = _gvn.transform( new CmpL3Node( a, b )); 2592 push(c); 2593 break; 2594 2595 case Bytecodes::_lneg: 2596 a = pop_pair(); 2597 b = _gvn.transform( new SubLNode(longcon(0),a)); 2598 push_pair(b); 2599 break; 2600 case Bytecodes::_l2i: 2601 a = pop_pair(); 2602 push( _gvn.transform( new ConvL2INode(a))); 2603 break; 2604 case Bytecodes::_i2l: 2605 a = pop(); 2606 b = _gvn.transform( new ConvI2LNode(a)); 2607 push_pair(b); 2608 break; 2609 case Bytecodes::_i2b: 2610 // Sign extend 2611 a = pop(); 2612 a = _gvn.transform( new LShiftINode(a,_gvn.intcon(24)) ); 2613 a = _gvn.transform( new RShiftINode(a,_gvn.intcon(24)) ); 2614 push( a ); 2615 break; 2616 case Bytecodes::_i2s: 2617 a = pop(); 2618 a = _gvn.transform( new LShiftINode(a,_gvn.intcon(16)) ); 2619 a = _gvn.transform( new RShiftINode(a,_gvn.intcon(16)) ); 2620 push( a ); 2621 break; 2622 case Bytecodes::_i2c: 2623 a = pop(); 2624 push( _gvn.transform( new AndINode(a,_gvn.intcon(0xFFFF)) ) ); 2625 break; 2626 2627 case Bytecodes::_i2f: 2628 a = pop(); 2629 b = _gvn.transform( new ConvI2FNode(a) ) ; 2630 c = precision_rounding(b); 2631 push (b); 2632 break; 2633 2634 case Bytecodes::_i2d: 2635 a = pop(); 2636 b = _gvn.transform( new ConvI2DNode(a)); 2637 push_pair(b); 2638 break; 2639 2640 case Bytecodes::_iinc: // Increment local 2641 i = iter().get_index(); // Get local index 2642 set_local( i, _gvn.transform( new AddINode( _gvn.intcon(iter().get_iinc_con()), local(i) ) ) ); 2643 break; 2644 2645 // Exit points of synchronized methods must have an unlock node 2646 case Bytecodes::_return: 2647 return_current(NULL); 2648 break; 2649 2650 case Bytecodes::_ireturn: 2651 case Bytecodes::_areturn: 2652 case Bytecodes::_freturn: 2653 return_current(pop()); 2654 break; 2655 case Bytecodes::_lreturn: 2656 return_current(pop_pair()); 2657 break; 2658 case Bytecodes::_dreturn: 2659 return_current(pop_pair()); 2660 break; 2661 2662 case Bytecodes::_athrow: 2663 // null exception oop throws NULL pointer exception 2664 null_check(peek()); 2665 if (stopped()) return; 2666 // Hook the thrown exception directly to subsequent handlers. 2667 if (BailoutToInterpreterForThrows) { 2668 // Keep method interpreted from now on. 2669 uncommon_trap(Deoptimization::Reason_unhandled, 2670 Deoptimization::Action_make_not_compilable); 2671 return; 2672 } 2673 if (env()->jvmti_can_post_on_exceptions()) { 2674 // check if we must post exception events, take uncommon trap if so (with must_throw = false) 2675 uncommon_trap_if_should_post_on_exceptions(Deoptimization::Reason_unhandled, false); 2676 } 2677 // Here if either can_post_on_exceptions or should_post_on_exceptions is false 2678 add_exception_state(make_exception_state(peek())); 2679 break; 2680 2681 case Bytecodes::_goto: // fall through 2682 case Bytecodes::_goto_w: { 2683 int target_bci = (bc() == Bytecodes::_goto) ? iter().get_dest() : iter().get_far_dest(); 2684 2685 // If this is a backwards branch in the bytecodes, add Safepoint 2686 maybe_add_safepoint(target_bci); 2687 2688 // Update method data 2689 profile_taken_branch(target_bci); 2690 2691 // Merge the current control into the target basic block 2692 merge(target_bci); 2693 2694 // See if we can get some profile data and hand it off to the next block 2695 Block *target_block = block()->successor_for_bci(target_bci); 2696 if (target_block->pred_count() != 1) break; 2697 ciMethodData* methodData = method()->method_data(); 2698 if (!methodData->is_mature()) break; 2699 ciProfileData* data = methodData->bci_to_data(bci()); 2700 assert(data != NULL && data->is_JumpData(), "need JumpData for taken branch"); 2701 int taken = ((ciJumpData*)data)->taken(); 2702 taken = method()->scale_count(taken); 2703 target_block->set_count(taken); 2704 break; 2705 } 2706 2707 case Bytecodes::_ifnull: btest = BoolTest::eq; goto handle_if_null; 2708 case Bytecodes::_ifnonnull: btest = BoolTest::ne; goto handle_if_null; 2709 handle_if_null: 2710 // If this is a backwards branch in the bytecodes, add Safepoint 2711 maybe_add_safepoint(iter().get_dest()); 2712 a = null(); 2713 b = pop(); 2714 if (!_gvn.type(b)->speculative_maybe_null() && 2715 !too_many_traps(Deoptimization::Reason_speculate_null_check)) { 2716 inc_sp(1); 2717 Node* null_ctl = top(); 2718 b = null_check_oop(b, &null_ctl, true, true, true); 2719 assert(null_ctl->is_top(), "no null control here"); 2720 dec_sp(1); 2721 } else if (_gvn.type(b)->speculative_always_null() && 2722 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) { 2723 inc_sp(1); 2724 b = null_assert(b); 2725 dec_sp(1); 2726 } 2727 c = _gvn.transform( new CmpPNode(b, a) ); 2728 do_ifnull(btest, c); 2729 break; 2730 2731 case Bytecodes::_if_acmpeq: btest = BoolTest::eq; goto handle_if_acmp; 2732 case Bytecodes::_if_acmpne: btest = BoolTest::ne; goto handle_if_acmp; 2733 handle_if_acmp: 2734 // If this is a backwards branch in the bytecodes, add Safepoint 2735 maybe_add_safepoint(iter().get_dest()); 2736 a = pop(); 2737 b = pop(); 2738 c = _gvn.transform( new CmpPNode(b, a) ); 2739 c = optimize_cmp_with_klass(c); 2740 do_if(btest, c); 2741 break; 2742 2743 case Bytecodes::_ifeq: btest = BoolTest::eq; goto handle_ifxx; 2744 case Bytecodes::_ifne: btest = BoolTest::ne; goto handle_ifxx; 2745 case Bytecodes::_iflt: btest = BoolTest::lt; goto handle_ifxx; 2746 case Bytecodes::_ifle: btest = BoolTest::le; goto handle_ifxx; 2747 case Bytecodes::_ifgt: btest = BoolTest::gt; goto handle_ifxx; 2748 case Bytecodes::_ifge: btest = BoolTest::ge; goto handle_ifxx; 2749 handle_ifxx: 2750 // If this is a backwards branch in the bytecodes, add Safepoint 2751 maybe_add_safepoint(iter().get_dest()); 2752 a = _gvn.intcon(0); 2753 b = pop(); 2754 c = _gvn.transform( new CmpINode(b, a) ); 2755 do_if(btest, c); 2756 break; 2757 2758 case Bytecodes::_if_icmpeq: btest = BoolTest::eq; goto handle_if_icmp; 2759 case Bytecodes::_if_icmpne: btest = BoolTest::ne; goto handle_if_icmp; 2760 case Bytecodes::_if_icmplt: btest = BoolTest::lt; goto handle_if_icmp; 2761 case Bytecodes::_if_icmple: btest = BoolTest::le; goto handle_if_icmp; 2762 case Bytecodes::_if_icmpgt: btest = BoolTest::gt; goto handle_if_icmp; 2763 case Bytecodes::_if_icmpge: btest = BoolTest::ge; goto handle_if_icmp; 2764 handle_if_icmp: 2765 // If this is a backwards branch in the bytecodes, add Safepoint 2766 maybe_add_safepoint(iter().get_dest()); 2767 a = pop(); 2768 b = pop(); 2769 c = _gvn.transform( new CmpINode( b, a ) ); 2770 do_if(btest, c); 2771 break; 2772 2773 case Bytecodes::_tableswitch: 2774 do_tableswitch(); 2775 break; 2776 2777 case Bytecodes::_lookupswitch: 2778 do_lookupswitch(); 2779 break; 2780 2781 case Bytecodes::_invokestatic: 2782 case Bytecodes::_invokedynamic: 2783 case Bytecodes::_invokespecial: 2784 case Bytecodes::_invokevirtual: 2785 case Bytecodes::_invokeinterface: 2786 do_call(); 2787 break; 2788 case Bytecodes::_checkcast: 2789 do_checkcast(); 2790 break; 2791 case Bytecodes::_instanceof: 2792 do_instanceof(); 2793 break; 2794 case Bytecodes::_anewarray: 2795 do_anewarray(); 2796 break; 2797 case Bytecodes::_newarray: 2798 do_newarray((BasicType)iter().get_index()); 2799 break; 2800 case Bytecodes::_multianewarray: 2801 do_multianewarray(); 2802 break; 2803 case Bytecodes::_new: 2804 do_new(); 2805 break; 2806 2807 case Bytecodes::_jsr: 2808 case Bytecodes::_jsr_w: 2809 do_jsr(); 2810 break; 2811 2812 case Bytecodes::_ret: 2813 do_ret(); 2814 break; 2815 2816 2817 case Bytecodes::_monitorenter: 2818 do_monitor_enter(); 2819 break; 2820 2821 case Bytecodes::_monitorexit: 2822 do_monitor_exit(); 2823 break; 2824 2825 case Bytecodes::_breakpoint: 2826 // Breakpoint set concurrently to compile 2827 // %%% use an uncommon trap? 2828 C->record_failure("breakpoint in method"); 2829 return; 2830 2831 default: 2832 #ifndef PRODUCT 2833 map()->dump(99); 2834 #endif 2835 tty->print("\nUnhandled bytecode %s\n", Bytecodes::name(bc()) ); 2836 ShouldNotReachHere(); 2837 } 2838 2839 #ifndef PRODUCT 2840 IdealGraphPrinter *printer = C->printer(); 2841 if (printer && printer->should_print(1)) { 2842 char buffer[256]; 2843 sprintf(buffer, "Bytecode %d: %s", bci(), Bytecodes::name(bc())); 2844 bool old = printer->traverse_outs(); 2845 printer->set_traverse_outs(true); 2846 printer->print_method(buffer, 4); 2847 printer->set_traverse_outs(old); 2848 } 2849 #endif 2850 }