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 bt) { 55 const Type* elemtype = Type::TOP; 56 bool big_val = bt == T_DOUBLE || bt == T_LONG; 57 Node* adr = array_addressing(bt, 0, elemtype); 58 if (stopped()) return; // guaranteed null or range check 59 60 pop(); // index (already used) 61 Node* array = pop(); // the array itself 62 63 if (elemtype == TypeInt::BOOL) { 64 bt = T_BOOLEAN; 65 } 66 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(bt); 67 68 Node* ld = access_load_at(array, adr, adr_type, elemtype, bt, 69 IN_HEAP | IS_ARRAY | C2_CONTROL_DEPENDENT_LOAD); 70 if (big_val) { 71 push_pair(ld); 72 } else { 73 push(ld); 74 } 75 } 76 77 78 //--------------------------------array_store---------------------------------- 79 void Parse::array_store(BasicType bt) { 80 const Type* elemtype = Type::TOP; 81 bool big_val = bt == T_DOUBLE || bt == T_LONG; 82 Node* adr = array_addressing(bt, big_val ? 2 : 1, elemtype); 83 if (stopped()) return; // guaranteed null or range check 84 if (bt == T_OBJECT) { 85 array_store_check(); 86 if (stopped()) { 87 return; 88 } 89 } 90 Node* val; // Oop to store 91 if (big_val) { 92 val = pop_pair(); 93 } else { 94 val = pop(); 95 } 96 pop(); // index (already used) 97 Node* array = pop(); // the array itself 98 99 if (elemtype == TypeInt::BOOL) { 100 bt = T_BOOLEAN; 101 } 102 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(bt); 103 104 access_store_at(control(), array, adr, adr_type, val, elemtype, bt, MO_UNORDERED | IN_HEAP | IS_ARRAY); 105 } 106 107 108 //------------------------------array_addressing------------------------------- 109 // Pull array and index from the stack. Compute pointer-to-element. 110 Node* Parse::array_addressing(BasicType type, int vals, const Type*& elemtype) { 111 Node *idx = peek(0+vals); // Get from stack without popping 112 Node *ary = peek(1+vals); // in case of exception 113 114 // Null check the array base, with correct stack contents 115 ary = null_check(ary, T_ARRAY); 116 // Compile-time detect of null-exception? 117 if (stopped()) return top(); 118 119 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr(); 120 const TypeInt* sizetype = arytype->size(); 121 elemtype = arytype->elem(); 122 123 if (UseUniqueSubclasses) { 124 const Type* el = elemtype->make_ptr(); 125 if (el && el->isa_instptr()) { 126 const TypeInstPtr* toop = el->is_instptr(); 127 if (toop->klass()->as_instance_klass()->unique_concrete_subklass()) { 128 // If we load from "AbstractClass[]" we must see "ConcreteSubClass". 129 const Type* subklass = Type::get_const_type(toop->klass()); 130 elemtype = subklass->join_speculative(el); 131 } 132 } 133 } 134 135 // Check for big class initializers with all constant offsets 136 // feeding into a known-size array. 137 const TypeInt* idxtype = _gvn.type(idx)->is_int(); 138 // See if the highest idx value is less than the lowest array bound, 139 // and if the idx value cannot be negative: 140 bool need_range_check = true; 141 if (idxtype->_hi < sizetype->_lo && idxtype->_lo >= 0) { 142 need_range_check = false; 143 if (C->log() != NULL) C->log()->elem("observe that='!need_range_check'"); 144 } 145 146 ciKlass * arytype_klass = arytype->klass(); 147 if ((arytype_klass != NULL) && (!arytype_klass->is_loaded())) { 148 // Only fails for some -Xcomp runs 149 // The class is unloaded. We have to run this bytecode in the interpreter. 150 uncommon_trap(Deoptimization::Reason_unloaded, 151 Deoptimization::Action_reinterpret, 152 arytype->klass(), "!loaded array"); 153 return top(); 154 } 155 156 // Do the range check 157 if (GenerateRangeChecks && need_range_check) { 158 Node* tst; 159 if (sizetype->_hi <= 0) { 160 // The greatest array bound is negative, so we can conclude that we're 161 // compiling unreachable code, but the unsigned compare trick used below 162 // only works with non-negative lengths. Instead, hack "tst" to be zero so 163 // the uncommon_trap path will always be taken. 164 tst = _gvn.intcon(0); 165 } else { 166 // Range is constant in array-oop, so we can use the original state of mem 167 Node* len = load_array_length(ary); 168 169 // Test length vs index (standard trick using unsigned compare) 170 Node* chk = _gvn.transform( new CmpUNode(idx, len) ); 171 BoolTest::mask btest = BoolTest::lt; 172 tst = _gvn.transform( new BoolNode(chk, btest) ); 173 } 174 RangeCheckNode* rc = new RangeCheckNode(control(), tst, PROB_MAX, COUNT_UNKNOWN); 175 _gvn.set_type(rc, rc->Value(&_gvn)); 176 if (!tst->is_Con()) { 177 record_for_igvn(rc); 178 } 179 set_control(_gvn.transform(new IfTrueNode(rc))); 180 // Branch to failure if out of bounds 181 { 182 PreserveJVMState pjvms(this); 183 set_control(_gvn.transform(new IfFalseNode(rc))); 184 if (C->allow_range_check_smearing()) { 185 // Do not use builtin_throw, since range checks are sometimes 186 // made more stringent by an optimistic transformation. 187 // This creates "tentative" range checks at this point, 188 // which are not guaranteed to throw exceptions. 189 // See IfNode::Ideal, is_range_check, adjust_check. 190 uncommon_trap(Deoptimization::Reason_range_check, 191 Deoptimization::Action_make_not_entrant, 192 NULL, "range_check"); 193 } else { 194 // If we have already recompiled with the range-check-widening 195 // heroic optimization turned off, then we must really be throwing 196 // range check exceptions. 197 builtin_throw(Deoptimization::Reason_range_check, idx); 198 } 199 } 200 } 201 // Check for always knowing you are throwing a range-check exception 202 if (stopped()) return top(); 203 204 // Make array address computation control dependent to prevent it 205 // from floating above the range check during loop optimizations. 206 Node* ptr = array_element_address(ary, idx, type, sizetype, control()); 207 assert(ptr != top(), "top should go hand-in-hand with stopped"); 208 209 return ptr; 210 } 211 212 213 // returns IfNode 214 IfNode* Parse::jump_if_fork_int(Node* a, Node* b, BoolTest::mask mask, float prob, float cnt) { 215 Node *cmp = _gvn.transform(new CmpINode(a, b)); // two cases: shiftcount > 32 and shiftcount <= 32 216 Node *tst = _gvn.transform(new BoolNode(cmp, mask)); 217 IfNode *iff = create_and_map_if(control(), tst, prob, cnt); 218 return iff; 219 } 220 221 // return Region node 222 Node* Parse::jump_if_join(Node* iffalse, Node* iftrue) { 223 Node *region = new RegionNode(3); // 2 results 224 record_for_igvn(region); 225 region->init_req(1, iffalse); 226 region->init_req(2, iftrue ); 227 _gvn.set_type(region, Type::CONTROL); 228 region = _gvn.transform(region); 229 set_control (region); 230 return region; 231 } 232 233 // sentinel value for the target bci to mark never taken branches 234 // (according to profiling) 235 static const int never_reached = INT_MAX; 236 237 //------------------------------helper for tableswitch------------------------- 238 void Parse::jump_if_true_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index, bool unc) { 239 // True branch, use existing map info 240 { PreserveJVMState pjvms(this); 241 Node *iftrue = _gvn.transform( new IfTrueNode (iff) ); 242 set_control( iftrue ); 243 if (unc) { 244 repush_if_args(); 245 uncommon_trap(Deoptimization::Reason_unstable_if, 246 Deoptimization::Action_reinterpret, 247 NULL, 248 "taken always"); 249 } else { 250 assert(dest_bci_if_true != never_reached, "inconsistent dest"); 251 profile_switch_case(prof_table_index); 252 merge_new_path(dest_bci_if_true); 253 } 254 } 255 256 // False branch 257 Node *iffalse = _gvn.transform( new IfFalseNode(iff) ); 258 set_control( iffalse ); 259 } 260 261 void Parse::jump_if_false_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index, bool unc) { 262 // True branch, use existing map info 263 { PreserveJVMState pjvms(this); 264 Node *iffalse = _gvn.transform( new IfFalseNode (iff) ); 265 set_control( iffalse ); 266 if (unc) { 267 repush_if_args(); 268 uncommon_trap(Deoptimization::Reason_unstable_if, 269 Deoptimization::Action_reinterpret, 270 NULL, 271 "taken never"); 272 } else { 273 assert(dest_bci_if_true != never_reached, "inconsistent dest"); 274 profile_switch_case(prof_table_index); 275 merge_new_path(dest_bci_if_true); 276 } 277 } 278 279 // False branch 280 Node *iftrue = _gvn.transform( new IfTrueNode(iff) ); 281 set_control( iftrue ); 282 } 283 284 void Parse::jump_if_always_fork(int dest_bci, int prof_table_index, bool unc) { 285 // False branch, use existing map and control() 286 if (unc) { 287 repush_if_args(); 288 uncommon_trap(Deoptimization::Reason_unstable_if, 289 Deoptimization::Action_reinterpret, 290 NULL, 291 "taken never"); 292 } else { 293 assert(dest_bci != never_reached, "inconsistent dest"); 294 profile_switch_case(prof_table_index); 295 merge_new_path(dest_bci); 296 } 297 } 298 299 300 extern "C" { 301 static int jint_cmp(const void *i, const void *j) { 302 int a = *(jint *)i; 303 int b = *(jint *)j; 304 return a > b ? 1 : a < b ? -1 : 0; 305 } 306 } 307 308 309 // Default value for methodData switch indexing. Must be a negative value to avoid 310 // conflict with any legal switch index. 311 #define NullTableIndex -1 312 313 class SwitchRange : public StackObj { 314 // a range of integers coupled with a bci destination 315 jint _lo; // inclusive lower limit 316 jint _hi; // inclusive upper limit 317 int _dest; 318 int _table_index; // index into method data table 319 float _cnt; // how many times this range was hit according to profiling 320 321 public: 322 jint lo() const { return _lo; } 323 jint hi() const { return _hi; } 324 int dest() const { return _dest; } 325 int table_index() const { return _table_index; } 326 bool is_singleton() const { return _lo == _hi; } 327 float cnt() const { return _cnt; } 328 329 void setRange(jint lo, jint hi, int dest, int table_index, float cnt) { 330 assert(lo <= hi, "must be a non-empty range"); 331 _lo = lo, _hi = hi; _dest = dest; _table_index = table_index; _cnt = cnt; 332 assert(_cnt >= 0, ""); 333 } 334 bool adjoinRange(jint lo, jint hi, int dest, int table_index, float cnt, bool trim_ranges) { 335 assert(lo <= hi, "must be a non-empty range"); 336 if (lo == _hi+1 && table_index == _table_index) { 337 // see merge_ranges() comment below 338 if (trim_ranges) { 339 if (cnt == 0) { 340 if (_cnt != 0) { 341 return false; 342 } 343 if (dest != _dest) { 344 _dest = never_reached; 345 } 346 } else { 347 if (_cnt == 0) { 348 return false; 349 } 350 if (dest != _dest) { 351 return false; 352 } 353 } 354 } else { 355 if (dest != _dest) { 356 return false; 357 } 358 } 359 _hi = hi; 360 _cnt += cnt; 361 return true; 362 } 363 return false; 364 } 365 366 void set (jint value, int dest, int table_index, float cnt) { 367 setRange(value, value, dest, table_index, cnt); 368 } 369 bool adjoin(jint value, int dest, int table_index, float cnt, bool trim_ranges) { 370 return adjoinRange(value, value, dest, table_index, cnt, trim_ranges); 371 } 372 bool adjoin(SwitchRange& other) { 373 return adjoinRange(other._lo, other._hi, other._dest, other._table_index, other._cnt, false); 374 } 375 376 void print() { 377 if (is_singleton()) 378 tty->print(" {%d}=>%d (cnt=%f)", lo(), dest(), cnt()); 379 else if (lo() == min_jint) 380 tty->print(" {..%d}=>%d (cnt=%f)", hi(), dest(), cnt()); 381 else if (hi() == max_jint) 382 tty->print(" {%d..}=>%d (cnt=%f)", lo(), dest(), cnt()); 383 else 384 tty->print(" {%d..%d}=>%d (cnt=%f)", lo(), hi(), dest(), cnt()); 385 } 386 }; 387 388 // We try to minimize the number of ranges and the size of the taken 389 // ones using profiling data. When ranges are created, 390 // SwitchRange::adjoinRange() only allows 2 adjoining ranges to merge 391 // if both were never hit or both were hit to build longer unreached 392 // ranges. Here, we now merge adjoining ranges with the same 393 // destination and finally set destination of unreached ranges to the 394 // special value never_reached because it can help minimize the number 395 // of tests that are necessary. 396 // 397 // For instance: 398 // [0, 1] to target1 sometimes taken 399 // [1, 2] to target1 never taken 400 // [2, 3] to target2 never taken 401 // would lead to: 402 // [0, 1] to target1 sometimes taken 403 // [1, 3] never taken 404 // 405 // (first 2 ranges to target1 are not merged) 406 static void merge_ranges(SwitchRange* ranges, int& rp) { 407 if (rp == 0) { 408 return; 409 } 410 int shift = 0; 411 for (int j = 0; j < rp; j++) { 412 SwitchRange& r1 = ranges[j-shift]; 413 SwitchRange& r2 = ranges[j+1]; 414 if (r1.adjoin(r2)) { 415 shift++; 416 } else if (shift > 0) { 417 ranges[j+1-shift] = r2; 418 } 419 } 420 rp -= shift; 421 for (int j = 0; j <= rp; j++) { 422 SwitchRange& r = ranges[j]; 423 if (r.cnt() == 0 && r.dest() != never_reached) { 424 r.setRange(r.lo(), r.hi(), never_reached, r.table_index(), r.cnt()); 425 } 426 } 427 } 428 429 //-------------------------------do_tableswitch-------------------------------- 430 void Parse::do_tableswitch() { 431 // Get information about tableswitch 432 int default_dest = iter().get_dest_table(0); 433 int lo_index = iter().get_int_table(1); 434 int hi_index = iter().get_int_table(2); 435 int len = hi_index - lo_index + 1; 436 437 if (len < 1) { 438 // If this is a backward branch, add safepoint 439 maybe_add_safepoint(default_dest); 440 pop(); // the effect of the instruction execution on the operand stack 441 merge(default_dest); 442 return; 443 } 444 445 ciMethodData* methodData = method()->method_data(); 446 ciMultiBranchData* profile = NULL; 447 if (methodData->is_mature() && UseSwitchProfiling) { 448 ciProfileData* data = methodData->bci_to_data(bci()); 449 if (data != NULL && data->is_MultiBranchData()) { 450 profile = (ciMultiBranchData*)data; 451 } 452 } 453 bool trim_ranges = !method_data_update() && !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if); 454 455 // generate decision tree, using trichotomy when possible 456 int rnum = len+2; 457 bool makes_backward_branch = false; 458 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum); 459 int rp = -1; 460 if (lo_index != min_jint) { 461 uint cnt = 1; 462 if (profile != NULL) { 463 cnt = profile->default_count() / (hi_index != max_jint ? 2 : 1); 464 } 465 ranges[++rp].setRange(min_jint, lo_index-1, default_dest, NullTableIndex, cnt); 466 } 467 for (int j = 0; j < len; j++) { 468 jint match_int = lo_index+j; 469 int dest = iter().get_dest_table(j+3); 470 makes_backward_branch |= (dest <= bci()); 471 int table_index = method_data_update() ? j : NullTableIndex; 472 uint cnt = 1; 473 if (profile != NULL) { 474 cnt = profile->count_at(j); 475 } 476 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index, cnt, trim_ranges)) { 477 ranges[++rp].set(match_int, dest, table_index, cnt); 478 } 479 } 480 jint highest = lo_index+(len-1); 481 assert(ranges[rp].hi() == highest, ""); 482 if (highest != max_jint) { 483 uint cnt = 1; 484 if (profile != NULL) { 485 cnt = profile->default_count() / (lo_index != min_jint ? 2 : 1); 486 } 487 if (!ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex, cnt, trim_ranges)) { 488 ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex, cnt); 489 } 490 } 491 assert(rp < len+2, "not too many ranges"); 492 493 if (trim_ranges) { 494 merge_ranges(ranges, rp); 495 } 496 497 // Safepoint in case if backward branch observed 498 if (makes_backward_branch && UseLoopSafepoints) { 499 add_safepoint(); 500 } 501 502 Node* lookup = pop(); // lookup value 503 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]); 504 } 505 506 507 //------------------------------do_lookupswitch-------------------------------- 508 void Parse::do_lookupswitch() { 509 // Get information about lookupswitch 510 int default_dest = iter().get_dest_table(0); 511 int len = iter().get_int_table(1); 512 513 if (len < 1) { // If this is a backward branch, add safepoint 514 maybe_add_safepoint(default_dest); 515 pop(); // the effect of the instruction execution on the operand stack 516 merge(default_dest); 517 return; 518 } 519 520 ciMethodData* methodData = method()->method_data(); 521 ciMultiBranchData* profile = NULL; 522 if (methodData->is_mature() && UseSwitchProfiling) { 523 ciProfileData* data = methodData->bci_to_data(bci()); 524 if (data != NULL && data->is_MultiBranchData()) { 525 profile = (ciMultiBranchData*)data; 526 } 527 } 528 bool trim_ranges = !method_data_update() && !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if); 529 530 // generate decision tree, using trichotomy when possible 531 jint* table = NEW_RESOURCE_ARRAY(jint, len*3); 532 { 533 for (int j = 0; j < len; j++) { 534 table[3*j+0] = iter().get_int_table(2+2*j); 535 table[3*j+1] = iter().get_dest_table(2+2*j+1); 536 // Handle overflow when converting from uint to jint 537 table[3*j+2] = (profile == NULL) ? 1 : MIN2<uint>(max_jint, profile->count_at(j)); 538 } 539 qsort(table, len, 3*sizeof(table[0]), jint_cmp); 540 } 541 542 float defaults = 0; 543 jint prev = min_jint; 544 for (int j = 0; j < len; j++) { 545 jint match_int = table[3*j+0]; 546 if (match_int != prev) { 547 defaults += (float)match_int - prev; 548 } 549 prev = match_int+1; 550 } 551 if (prev-1 != max_jint) { 552 defaults += (float)max_jint - prev + 1; 553 } 554 float default_cnt = 1; 555 if (profile != NULL) { 556 default_cnt = profile->default_count()/defaults; 557 } 558 559 int rnum = len*2+1; 560 bool makes_backward_branch = false; 561 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum); 562 int rp = -1; 563 for (int j = 0; j < len; j++) { 564 jint match_int = table[3*j+0]; 565 int dest = table[3*j+1]; 566 int cnt = table[3*j+2]; 567 int next_lo = rp < 0 ? min_jint : ranges[rp].hi()+1; 568 int table_index = method_data_update() ? j : NullTableIndex; 569 makes_backward_branch |= (dest <= bci()); 570 float c = default_cnt * ((float)match_int - next_lo); 571 if (match_int != next_lo && (rp < 0 || !ranges[rp].adjoinRange(next_lo, match_int-1, default_dest, NullTableIndex, c, trim_ranges))) { 572 assert(default_dest != never_reached, "sentinel value for dead destinations"); 573 ranges[++rp].setRange(next_lo, match_int-1, default_dest, NullTableIndex, c); 574 } 575 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index, cnt, trim_ranges)) { 576 assert(dest != never_reached, "sentinel value for dead destinations"); 577 ranges[++rp].set(match_int, dest, table_index, cnt); 578 } 579 } 580 jint highest = table[3*(len-1)]; 581 assert(ranges[rp].hi() == highest, ""); 582 if (highest != max_jint && 583 !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex, default_cnt * ((float)max_jint - highest), trim_ranges)) { 584 ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex, default_cnt * ((float)max_jint - highest)); 585 } 586 assert(rp < rnum, "not too many ranges"); 587 588 if (trim_ranges) { 589 merge_ranges(ranges, rp); 590 } 591 592 // Safepoint in case backward branch observed 593 if (makes_backward_branch && UseLoopSafepoints) { 594 add_safepoint(); 595 } 596 597 Node *lookup = pop(); // lookup value 598 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]); 599 } 600 601 static float if_prob(float taken_cnt, float total_cnt) { 602 assert(taken_cnt <= total_cnt, ""); 603 if (total_cnt == 0) { 604 return PROB_FAIR; 605 } 606 float p = taken_cnt / total_cnt; 607 return MIN2(MAX2(p, PROB_MIN), PROB_MAX); 608 } 609 610 static float if_cnt(float cnt) { 611 if (cnt == 0) { 612 return COUNT_UNKNOWN; 613 } 614 return cnt; 615 } 616 617 static float sum_of_cnts(SwitchRange *lo, SwitchRange *hi) { 618 float total_cnt = 0; 619 for (SwitchRange* sr = lo; sr <= hi; sr++) { 620 total_cnt += sr->cnt(); 621 } 622 return total_cnt; 623 } 624 625 class SwitchRanges : public ResourceObj { 626 public: 627 SwitchRange* _lo; 628 SwitchRange* _hi; 629 SwitchRange* _mid; 630 float _cost; 631 632 enum { 633 Start, 634 LeftDone, 635 RightDone, 636 Done 637 } _state; 638 639 SwitchRanges(SwitchRange *lo, SwitchRange *hi) 640 : _lo(lo), _hi(hi), _mid(NULL), 641 _cost(0), _state(Start) { 642 } 643 644 SwitchRanges() 645 : _lo(NULL), _hi(NULL), _mid(NULL), 646 _cost(0), _state(Start) {} 647 }; 648 649 // Estimate cost of performing a binary search on lo..hi 650 static float compute_tree_cost(SwitchRange *lo, SwitchRange *hi, float total_cnt) { 651 GrowableArray<SwitchRanges> tree; 652 SwitchRanges root(lo, hi); 653 tree.push(root); 654 655 float cost = 0; 656 do { 657 SwitchRanges& r = *tree.adr_at(tree.length()-1); 658 if (r._hi != r._lo) { 659 if (r._mid == NULL) { 660 float r_cnt = sum_of_cnts(r._lo, r._hi); 661 662 if (r_cnt == 0) { 663 tree.pop(); 664 cost = 0; 665 continue; 666 } 667 668 SwitchRange* mid = NULL; 669 mid = r._lo; 670 for (float cnt = 0; ; ) { 671 assert(mid <= r._hi, "out of bounds"); 672 cnt += mid->cnt(); 673 if (cnt > r_cnt / 2) { 674 break; 675 } 676 mid++; 677 } 678 assert(mid <= r._hi, "out of bounds"); 679 r._mid = mid; 680 r._cost = r_cnt / total_cnt; 681 } 682 r._cost += cost; 683 if (r._state < SwitchRanges::LeftDone && r._mid > r._lo) { 684 cost = 0; 685 r._state = SwitchRanges::LeftDone; 686 tree.push(SwitchRanges(r._lo, r._mid-1)); 687 } else if (r._state < SwitchRanges::RightDone) { 688 cost = 0; 689 r._state = SwitchRanges::RightDone; 690 tree.push(SwitchRanges(r._mid == r._lo ? r._mid+1 : r._mid, r._hi)); 691 } else { 692 tree.pop(); 693 cost = r._cost; 694 } 695 } else { 696 tree.pop(); 697 cost = r._cost; 698 } 699 } while (tree.length() > 0); 700 701 702 return cost; 703 } 704 705 // It sometimes pays off to test most common ranges before the binary search 706 void Parse::linear_search_switch_ranges(Node* key_val, SwitchRange*& lo, SwitchRange*& hi) { 707 uint nr = hi - lo + 1; 708 float total_cnt = sum_of_cnts(lo, hi); 709 710 float min = compute_tree_cost(lo, hi, total_cnt); 711 float extra = 1; 712 float sub = 0; 713 714 SwitchRange* array1 = lo; 715 SwitchRange* array2 = NEW_RESOURCE_ARRAY(SwitchRange, nr); 716 717 SwitchRange* ranges = NULL; 718 719 while (nr >= 2) { 720 assert(lo == array1 || lo == array2, "one the 2 already allocated arrays"); 721 ranges = (lo == array1) ? array2 : array1; 722 723 // Find highest frequency range 724 SwitchRange* candidate = lo; 725 for (SwitchRange* sr = lo+1; sr <= hi; sr++) { 726 if (sr->cnt() > candidate->cnt()) { 727 candidate = sr; 728 } 729 } 730 SwitchRange most_freq = *candidate; 731 if (most_freq.cnt() == 0) { 732 break; 733 } 734 735 // Copy remaining ranges into another array 736 int shift = 0; 737 for (uint i = 0; i < nr; i++) { 738 SwitchRange* sr = &lo[i]; 739 if (sr != candidate) { 740 ranges[i-shift] = *sr; 741 } else { 742 shift++; 743 if (i > 0 && i < nr-1) { 744 SwitchRange prev = lo[i-1]; 745 prev.setRange(prev.lo(), sr->hi(), prev.dest(), prev.table_index(), prev.cnt()); 746 if (prev.adjoin(lo[i+1])) { 747 shift++; 748 i++; 749 } 750 ranges[i-shift] = prev; 751 } 752 } 753 } 754 nr -= shift; 755 756 // Evaluate cost of testing the most common range and performing a 757 // binary search on the other ranges 758 float cost = extra + compute_tree_cost(&ranges[0], &ranges[nr-1], total_cnt); 759 if (cost >= min) { 760 break; 761 } 762 // swap arrays 763 lo = &ranges[0]; 764 hi = &ranges[nr-1]; 765 766 // It pays off: emit the test for the most common range 767 assert(most_freq.cnt() > 0, "must be taken"); 768 Node* val = _gvn.transform(new SubINode(key_val, _gvn.intcon(most_freq.lo()))); 769 Node* cmp = _gvn.transform(new CmpUNode(val, _gvn.intcon(most_freq.hi() - most_freq.lo()))); 770 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::le)); 771 IfNode* iff = create_and_map_if(control(), tst, if_prob(most_freq.cnt(), total_cnt), if_cnt(most_freq.cnt())); 772 jump_if_true_fork(iff, most_freq.dest(), most_freq.table_index(), false); 773 774 sub += most_freq.cnt() / total_cnt; 775 extra += 1 - sub; 776 min = cost; 777 } 778 } 779 780 //----------------------------create_jump_tables------------------------------- 781 bool Parse::create_jump_tables(Node* key_val, SwitchRange* lo, SwitchRange* hi) { 782 // Are jumptables enabled 783 if (!UseJumpTables) return false; 784 785 // Are jumptables supported 786 if (!Matcher::has_match_rule(Op_Jump)) return false; 787 788 // Don't make jump table if profiling 789 if (method_data_update()) return false; 790 791 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if); 792 793 // Decide if a guard is needed to lop off big ranges at either (or 794 // both) end(s) of the input set. We'll call this the default target 795 // even though we can't be sure that it is the true "default". 796 797 bool needs_guard = false; 798 int default_dest; 799 int64_t total_outlier_size = 0; 800 int64_t hi_size = ((int64_t)hi->hi()) - ((int64_t)hi->lo()) + 1; 801 int64_t lo_size = ((int64_t)lo->hi()) - ((int64_t)lo->lo()) + 1; 802 803 if (lo->dest() == hi->dest()) { 804 total_outlier_size = hi_size + lo_size; 805 default_dest = lo->dest(); 806 } else if (lo_size > hi_size) { 807 total_outlier_size = lo_size; 808 default_dest = lo->dest(); 809 } else { 810 total_outlier_size = hi_size; 811 default_dest = hi->dest(); 812 } 813 814 float total = sum_of_cnts(lo, hi); 815 float cost = compute_tree_cost(lo, hi, total); 816 817 // If a guard test will eliminate very sparse end ranges, then 818 // it is worth the cost of an extra jump. 819 float trimmed_cnt = 0; 820 if (total_outlier_size > (MaxJumpTableSparseness * 4)) { 821 needs_guard = true; 822 if (default_dest == lo->dest()) { 823 trimmed_cnt += lo->cnt(); 824 lo++; 825 } 826 if (default_dest == hi->dest()) { 827 trimmed_cnt += hi->cnt(); 828 hi--; 829 } 830 } 831 832 // Find the total number of cases and ranges 833 int64_t num_cases = ((int64_t)hi->hi()) - ((int64_t)lo->lo()) + 1; 834 int num_range = hi - lo + 1; 835 836 // Don't create table if: too large, too small, or too sparse. 837 if (num_cases > MaxJumpTableSize) 838 return false; 839 if (UseSwitchProfiling) { 840 // MinJumpTableSize is set so with a well balanced binary tree, 841 // when the number of ranges is MinJumpTableSize, it's cheaper to 842 // go through a JumpNode that a tree of IfNodes. Average cost of a 843 // tree of IfNodes with MinJumpTableSize is 844 // log2f(MinJumpTableSize) comparisons. So if the cost computed 845 // from profile data is less than log2f(MinJumpTableSize) then 846 // going with the binary search is cheaper. 847 if (cost < log2f(MinJumpTableSize)) { 848 return false; 849 } 850 } else { 851 if (num_cases < MinJumpTableSize) 852 return false; 853 } 854 if (num_cases > (MaxJumpTableSparseness * num_range)) 855 return false; 856 857 // Normalize table lookups to zero 858 int lowval = lo->lo(); 859 key_val = _gvn.transform( new SubINode(key_val, _gvn.intcon(lowval)) ); 860 861 // Generate a guard to protect against input keyvals that aren't 862 // in the switch domain. 863 if (needs_guard) { 864 Node* size = _gvn.intcon(num_cases); 865 Node* cmp = _gvn.transform(new CmpUNode(key_val, size)); 866 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::ge)); 867 IfNode* iff = create_and_map_if(control(), tst, if_prob(trimmed_cnt, total), if_cnt(trimmed_cnt)); 868 jump_if_true_fork(iff, default_dest, NullTableIndex, trim_ranges && trimmed_cnt == 0); 869 870 total -= trimmed_cnt; 871 } 872 873 // Create an ideal node JumpTable that has projections 874 // of all possible ranges for a switch statement 875 // The key_val input must be converted to a pointer offset and scaled. 876 // Compare Parse::array_addressing above. 877 878 // Clean the 32-bit int into a real 64-bit offset. 879 // Otherwise, the jint value 0 might turn into an offset of 0x0800000000. 880 // Make I2L conversion control dependent to prevent it from 881 // floating above the range check during loop optimizations. 882 // Do not use a narrow int type here to prevent the data path from dying 883 // while the control path is not removed. This can happen if the type of key_val 884 // is later known to be out of bounds of [0, num_cases] and therefore a narrow cast 885 // would be replaced by TOP while C2 is not able to fold the corresponding range checks. 886 // Set _carry_dependency for the cast to avoid being removed by IGVN. 887 #ifdef _LP64 888 key_val = C->constrained_convI2L(&_gvn, key_val, TypeInt::INT, control(), true /* carry_dependency */); 889 #endif 890 891 // Shift the value by wordsize so we have an index into the table, rather 892 // than a switch value 893 Node *shiftWord = _gvn.MakeConX(wordSize); 894 key_val = _gvn.transform( new MulXNode( key_val, shiftWord)); 895 896 // Create the JumpNode 897 Arena* arena = C->comp_arena(); 898 float* probs = (float*)arena->Amalloc(sizeof(float)*num_cases); 899 int i = 0; 900 if (total == 0) { 901 for (SwitchRange* r = lo; r <= hi; r++) { 902 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) { 903 probs[i] = 1.0F / num_cases; 904 } 905 } 906 } else { 907 for (SwitchRange* r = lo; r <= hi; r++) { 908 float prob = r->cnt()/total; 909 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) { 910 probs[i] = prob / (r->hi() - r->lo() + 1); 911 } 912 } 913 } 914 915 ciMethodData* methodData = method()->method_data(); 916 ciMultiBranchData* profile = NULL; 917 if (methodData->is_mature()) { 918 ciProfileData* data = methodData->bci_to_data(bci()); 919 if (data != NULL && data->is_MultiBranchData()) { 920 profile = (ciMultiBranchData*)data; 921 } 922 } 923 924 Node* jtn = _gvn.transform(new JumpNode(control(), key_val, num_cases, probs, profile == NULL ? COUNT_UNKNOWN : total)); 925 926 // These are the switch destinations hanging off the jumpnode 927 i = 0; 928 for (SwitchRange* r = lo; r <= hi; r++) { 929 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) { 930 Node* input = _gvn.transform(new JumpProjNode(jtn, i, r->dest(), (int)(j - lowval))); 931 { 932 PreserveJVMState pjvms(this); 933 set_control(input); 934 jump_if_always_fork(r->dest(), r->table_index(), trim_ranges && r->cnt() == 0); 935 } 936 } 937 } 938 assert(i == num_cases, "miscount of cases"); 939 stop_and_kill_map(); // no more uses for this JVMS 940 return true; 941 } 942 943 //----------------------------jump_switch_ranges------------------------------- 944 void Parse::jump_switch_ranges(Node* key_val, SwitchRange *lo, SwitchRange *hi, int switch_depth) { 945 Block* switch_block = block(); 946 bool trim_ranges = !method_data_update() && !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if); 947 948 if (switch_depth == 0) { 949 // Do special processing for the top-level call. 950 assert(lo->lo() == min_jint, "initial range must exhaust Type::INT"); 951 assert(hi->hi() == max_jint, "initial range must exhaust Type::INT"); 952 953 // Decrement pred-numbers for the unique set of nodes. 954 #ifdef ASSERT 955 if (!trim_ranges) { 956 // Ensure that the block's successors are a (duplicate-free) set. 957 int successors_counted = 0; // block occurrences in [hi..lo] 958 int unique_successors = switch_block->num_successors(); 959 for (int i = 0; i < unique_successors; i++) { 960 Block* target = switch_block->successor_at(i); 961 962 // Check that the set of successors is the same in both places. 963 int successors_found = 0; 964 for (SwitchRange* p = lo; p <= hi; p++) { 965 if (p->dest() == target->start()) successors_found++; 966 } 967 assert(successors_found > 0, "successor must be known"); 968 successors_counted += successors_found; 969 } 970 assert(successors_counted == (hi-lo)+1, "no unexpected successors"); 971 } 972 #endif 973 974 // Maybe prune the inputs, based on the type of key_val. 975 jint min_val = min_jint; 976 jint max_val = max_jint; 977 const TypeInt* ti = key_val->bottom_type()->isa_int(); 978 if (ti != NULL) { 979 min_val = ti->_lo; 980 max_val = ti->_hi; 981 assert(min_val <= max_val, "invalid int type"); 982 } 983 while (lo->hi() < min_val) { 984 lo++; 985 } 986 if (lo->lo() < min_val) { 987 lo->setRange(min_val, lo->hi(), lo->dest(), lo->table_index(), lo->cnt()); 988 } 989 while (hi->lo() > max_val) { 990 hi--; 991 } 992 if (hi->hi() > max_val) { 993 hi->setRange(hi->lo(), max_val, hi->dest(), hi->table_index(), hi->cnt()); 994 } 995 996 linear_search_switch_ranges(key_val, lo, hi); 997 } 998 999 #ifndef PRODUCT 1000 if (switch_depth == 0) { 1001 _max_switch_depth = 0; 1002 _est_switch_depth = log2_intptr((hi-lo+1)-1)+1; 1003 } 1004 #endif 1005 1006 assert(lo <= hi, "must be a non-empty set of ranges"); 1007 if (lo == hi) { 1008 jump_if_always_fork(lo->dest(), lo->table_index(), trim_ranges && lo->cnt() == 0); 1009 } else { 1010 assert(lo->hi() == (lo+1)->lo()-1, "contiguous ranges"); 1011 assert(hi->lo() == (hi-1)->hi()+1, "contiguous ranges"); 1012 1013 if (create_jump_tables(key_val, lo, hi)) return; 1014 1015 SwitchRange* mid = NULL; 1016 float total_cnt = sum_of_cnts(lo, hi); 1017 1018 int nr = hi - lo + 1; 1019 if (UseSwitchProfiling) { 1020 // Don't keep the binary search tree balanced: pick up mid point 1021 // that split frequencies in half. 1022 float cnt = 0; 1023 for (SwitchRange* sr = lo; sr <= hi; sr++) { 1024 cnt += sr->cnt(); 1025 if (cnt >= total_cnt / 2) { 1026 mid = sr; 1027 break; 1028 } 1029 } 1030 } else { 1031 mid = lo + nr/2; 1032 1033 // if there is an easy choice, pivot at a singleton: 1034 if (nr > 3 && !mid->is_singleton() && (mid-1)->is_singleton()) mid--; 1035 1036 assert(lo < mid && mid <= hi, "good pivot choice"); 1037 assert(nr != 2 || mid == hi, "should pick higher of 2"); 1038 assert(nr != 3 || mid == hi-1, "should pick middle of 3"); 1039 } 1040 1041 1042 Node *test_val = _gvn.intcon(mid == lo ? mid->hi() : mid->lo()); 1043 1044 if (mid->is_singleton()) { 1045 IfNode *iff_ne = jump_if_fork_int(key_val, test_val, BoolTest::ne, 1-if_prob(mid->cnt(), total_cnt), if_cnt(mid->cnt())); 1046 jump_if_false_fork(iff_ne, mid->dest(), mid->table_index(), trim_ranges && mid->cnt() == 0); 1047 1048 // Special Case: If there are exactly three ranges, and the high 1049 // and low range each go to the same place, omit the "gt" test, 1050 // since it will not discriminate anything. 1051 bool eq_test_only = (hi == lo+2 && hi->dest() == lo->dest() && mid == hi-1) || mid == lo; 1052 1053 // if there is a higher range, test for it and process it: 1054 if (mid < hi && !eq_test_only) { 1055 // two comparisons of same values--should enable 1 test for 2 branches 1056 // Use BoolTest::lt instead of BoolTest::gt 1057 float cnt = sum_of_cnts(lo, mid-1); 1058 IfNode *iff_lt = jump_if_fork_int(key_val, test_val, BoolTest::lt, if_prob(cnt, total_cnt), if_cnt(cnt)); 1059 Node *iftrue = _gvn.transform( new IfTrueNode(iff_lt) ); 1060 Node *iffalse = _gvn.transform( new IfFalseNode(iff_lt) ); 1061 { PreserveJVMState pjvms(this); 1062 set_control(iffalse); 1063 jump_switch_ranges(key_val, mid+1, hi, switch_depth+1); 1064 } 1065 set_control(iftrue); 1066 } 1067 1068 } else { 1069 // mid is a range, not a singleton, so treat mid..hi as a unit 1070 float cnt = sum_of_cnts(mid == lo ? mid+1 : mid, hi); 1071 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)); 1072 1073 // if there is a higher range, test for it and process it: 1074 if (mid == hi) { 1075 jump_if_true_fork(iff_ge, mid->dest(), mid->table_index(), trim_ranges && cnt == 0); 1076 } else { 1077 Node *iftrue = _gvn.transform( new IfTrueNode(iff_ge) ); 1078 Node *iffalse = _gvn.transform( new IfFalseNode(iff_ge) ); 1079 { PreserveJVMState pjvms(this); 1080 set_control(iftrue); 1081 jump_switch_ranges(key_val, mid == lo ? mid+1 : mid, hi, switch_depth+1); 1082 } 1083 set_control(iffalse); 1084 } 1085 } 1086 1087 // in any case, process the lower range 1088 if (mid == lo) { 1089 if (mid->is_singleton()) { 1090 jump_switch_ranges(key_val, lo+1, hi, switch_depth+1); 1091 } else { 1092 jump_if_always_fork(lo->dest(), lo->table_index(), trim_ranges && lo->cnt() == 0); 1093 } 1094 } else { 1095 jump_switch_ranges(key_val, lo, mid-1, switch_depth+1); 1096 } 1097 } 1098 1099 // Decrease pred_count for each successor after all is done. 1100 if (switch_depth == 0) { 1101 int unique_successors = switch_block->num_successors(); 1102 for (int i = 0; i < unique_successors; i++) { 1103 Block* target = switch_block->successor_at(i); 1104 // Throw away the pre-allocated path for each unique successor. 1105 target->next_path_num(); 1106 } 1107 } 1108 1109 #ifndef PRODUCT 1110 _max_switch_depth = MAX2(switch_depth, _max_switch_depth); 1111 if (TraceOptoParse && Verbose && WizardMode && switch_depth == 0) { 1112 SwitchRange* r; 1113 int nsing = 0; 1114 for( r = lo; r <= hi; r++ ) { 1115 if( r->is_singleton() ) nsing++; 1116 } 1117 tty->print(">>> "); 1118 _method->print_short_name(); 1119 tty->print_cr(" switch decision tree"); 1120 tty->print_cr(" %d ranges (%d singletons), max_depth=%d, est_depth=%d", 1121 (int) (hi-lo+1), nsing, _max_switch_depth, _est_switch_depth); 1122 if (_max_switch_depth > _est_switch_depth) { 1123 tty->print_cr("******** BAD SWITCH DEPTH ********"); 1124 } 1125 tty->print(" "); 1126 for( r = lo; r <= hi; r++ ) { 1127 r->print(); 1128 } 1129 tty->cr(); 1130 } 1131 #endif 1132 } 1133 1134 void Parse::modf() { 1135 Node *f2 = pop(); 1136 Node *f1 = pop(); 1137 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::modf_Type(), 1138 CAST_FROM_FN_PTR(address, SharedRuntime::frem), 1139 "frem", NULL, //no memory effects 1140 f1, f2); 1141 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0)); 1142 1143 push(res); 1144 } 1145 1146 void Parse::modd() { 1147 Node *d2 = pop_pair(); 1148 Node *d1 = pop_pair(); 1149 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::Math_DD_D_Type(), 1150 CAST_FROM_FN_PTR(address, SharedRuntime::drem), 1151 "drem", NULL, //no memory effects 1152 d1, top(), d2, top()); 1153 Node* res_d = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0)); 1154 1155 #ifdef ASSERT 1156 Node* res_top = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 1)); 1157 assert(res_top == top(), "second value must be top"); 1158 #endif 1159 1160 push_pair(res_d); 1161 } 1162 1163 void Parse::l2f() { 1164 Node* f2 = pop(); 1165 Node* f1 = pop(); 1166 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::l2f_Type(), 1167 CAST_FROM_FN_PTR(address, SharedRuntime::l2f), 1168 "l2f", NULL, //no memory effects 1169 f1, f2); 1170 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0)); 1171 1172 push(res); 1173 } 1174 1175 void Parse::do_irem() { 1176 // Must keep both values on the expression-stack during null-check 1177 zero_check_int(peek()); 1178 // Compile-time detect of null-exception? 1179 if (stopped()) return; 1180 1181 Node* b = pop(); 1182 Node* a = pop(); 1183 1184 const Type *t = _gvn.type(b); 1185 if (t != Type::TOP) { 1186 const TypeInt *ti = t->is_int(); 1187 if (ti->is_con()) { 1188 int divisor = ti->get_con(); 1189 // check for positive power of 2 1190 if (divisor > 0 && 1191 (divisor & ~(divisor-1)) == divisor) { 1192 // yes ! 1193 Node *mask = _gvn.intcon((divisor - 1)); 1194 // Sigh, must handle negative dividends 1195 Node *zero = _gvn.intcon(0); 1196 IfNode *ifff = jump_if_fork_int(a, zero, BoolTest::lt, PROB_FAIR, COUNT_UNKNOWN); 1197 Node *iff = _gvn.transform( new IfFalseNode(ifff) ); 1198 Node *ift = _gvn.transform( new IfTrueNode (ifff) ); 1199 Node *reg = jump_if_join(ift, iff); 1200 Node *phi = PhiNode::make(reg, NULL, TypeInt::INT); 1201 // Negative path; negate/and/negate 1202 Node *neg = _gvn.transform( new SubINode(zero, a) ); 1203 Node *andn= _gvn.transform( new AndINode(neg, mask) ); 1204 Node *negn= _gvn.transform( new SubINode(zero, andn) ); 1205 phi->init_req(1, negn); 1206 // Fast positive case 1207 Node *andx = _gvn.transform( new AndINode(a, mask) ); 1208 phi->init_req(2, andx); 1209 // Push the merge 1210 push( _gvn.transform(phi) ); 1211 return; 1212 } 1213 } 1214 } 1215 // Default case 1216 push( _gvn.transform( new ModINode(control(),a,b) ) ); 1217 } 1218 1219 // Handle jsr and jsr_w bytecode 1220 void Parse::do_jsr() { 1221 assert(bc() == Bytecodes::_jsr || bc() == Bytecodes::_jsr_w, "wrong bytecode"); 1222 1223 // Store information about current state, tagged with new _jsr_bci 1224 int return_bci = iter().next_bci(); 1225 int jsr_bci = (bc() == Bytecodes::_jsr) ? iter().get_dest() : iter().get_far_dest(); 1226 1227 // Update method data 1228 profile_taken_branch(jsr_bci); 1229 1230 // The way we do things now, there is only one successor block 1231 // for the jsr, because the target code is cloned by ciTypeFlow. 1232 Block* target = successor_for_bci(jsr_bci); 1233 1234 // What got pushed? 1235 const Type* ret_addr = target->peek(); 1236 assert(ret_addr->singleton(), "must be a constant (cloned jsr body)"); 1237 1238 // Effect on jsr on stack 1239 push(_gvn.makecon(ret_addr)); 1240 1241 // Flow to the jsr. 1242 merge(jsr_bci); 1243 } 1244 1245 // Handle ret bytecode 1246 void Parse::do_ret() { 1247 // Find to whom we return. 1248 assert(block()->num_successors() == 1, "a ret can only go one place now"); 1249 Block* target = block()->successor_at(0); 1250 assert(!target->is_ready(), "our arrival must be expected"); 1251 profile_ret(target->flow()->start()); 1252 int pnum = target->next_path_num(); 1253 merge_common(target, pnum); 1254 } 1255 1256 static bool has_injected_profile(BoolTest::mask btest, Node* test, int& taken, int& not_taken) { 1257 if (btest != BoolTest::eq && btest != BoolTest::ne) { 1258 // Only ::eq and ::ne are supported for profile injection. 1259 return false; 1260 } 1261 if (test->is_Cmp() && 1262 test->in(1)->Opcode() == Op_ProfileBoolean) { 1263 ProfileBooleanNode* profile = (ProfileBooleanNode*)test->in(1); 1264 int false_cnt = profile->false_count(); 1265 int true_cnt = profile->true_count(); 1266 1267 // Counts matching depends on the actual test operation (::eq or ::ne). 1268 // No need to scale the counts because profile injection was designed 1269 // to feed exact counts into VM. 1270 taken = (btest == BoolTest::eq) ? false_cnt : true_cnt; 1271 not_taken = (btest == BoolTest::eq) ? true_cnt : false_cnt; 1272 1273 profile->consume(); 1274 return true; 1275 } 1276 return false; 1277 } 1278 //--------------------------dynamic_branch_prediction-------------------------- 1279 // Try to gather dynamic branch prediction behavior. Return a probability 1280 // of the branch being taken and set the "cnt" field. Returns a -1.0 1281 // if we need to use static prediction for some reason. 1282 float Parse::dynamic_branch_prediction(float &cnt, BoolTest::mask btest, Node* test) { 1283 ResourceMark rm; 1284 1285 cnt = COUNT_UNKNOWN; 1286 1287 int taken = 0; 1288 int not_taken = 0; 1289 1290 bool use_mdo = !has_injected_profile(btest, test, taken, not_taken); 1291 1292 if (use_mdo) { 1293 // Use MethodData information if it is available 1294 // FIXME: free the ProfileData structure 1295 ciMethodData* methodData = method()->method_data(); 1296 if (!methodData->is_mature()) return PROB_UNKNOWN; 1297 ciProfileData* data = methodData->bci_to_data(bci()); 1298 if (data == NULL) { 1299 return PROB_UNKNOWN; 1300 } 1301 if (!data->is_JumpData()) return PROB_UNKNOWN; 1302 1303 // get taken and not taken values 1304 taken = data->as_JumpData()->taken(); 1305 not_taken = 0; 1306 if (data->is_BranchData()) { 1307 not_taken = data->as_BranchData()->not_taken(); 1308 } 1309 1310 // scale the counts to be commensurate with invocation counts: 1311 taken = method()->scale_count(taken); 1312 not_taken = method()->scale_count(not_taken); 1313 } 1314 1315 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful. 1316 // We also check that individual counters are positive first, otherwise the sum can become positive. 1317 if (taken < 0 || not_taken < 0 || taken + not_taken < 40) { 1318 if (C->log() != NULL) { 1319 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d'", iter().get_dest(), taken, not_taken); 1320 } 1321 return PROB_UNKNOWN; 1322 } 1323 1324 // Compute frequency that we arrive here 1325 float sum = taken + not_taken; 1326 // Adjust, if this block is a cloned private block but the 1327 // Jump counts are shared. Taken the private counts for 1328 // just this path instead of the shared counts. 1329 if( block()->count() > 0 ) 1330 sum = block()->count(); 1331 cnt = sum / FreqCountInvocations; 1332 1333 // Pin probability to sane limits 1334 float prob; 1335 if( !taken ) 1336 prob = (0+PROB_MIN) / 2; 1337 else if( !not_taken ) 1338 prob = (1+PROB_MAX) / 2; 1339 else { // Compute probability of true path 1340 prob = (float)taken / (float)(taken + not_taken); 1341 if (prob > PROB_MAX) prob = PROB_MAX; 1342 if (prob < PROB_MIN) prob = PROB_MIN; 1343 } 1344 1345 assert((cnt > 0.0f) && (prob > 0.0f), 1346 "Bad frequency assignment in if"); 1347 1348 if (C->log() != NULL) { 1349 const char* prob_str = NULL; 1350 if (prob >= PROB_MAX) prob_str = (prob == PROB_MAX) ? "max" : "always"; 1351 if (prob <= PROB_MIN) prob_str = (prob == PROB_MIN) ? "min" : "never"; 1352 char prob_str_buf[30]; 1353 if (prob_str == NULL) { 1354 sprintf(prob_str_buf, "%g", prob); 1355 prob_str = prob_str_buf; 1356 } 1357 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d' cnt='%f' prob='%s'", 1358 iter().get_dest(), taken, not_taken, cnt, prob_str); 1359 } 1360 return prob; 1361 } 1362 1363 //-----------------------------branch_prediction------------------------------- 1364 float Parse::branch_prediction(float& cnt, 1365 BoolTest::mask btest, 1366 int target_bci, 1367 Node* test) { 1368 float prob = dynamic_branch_prediction(cnt, btest, test); 1369 // If prob is unknown, switch to static prediction 1370 if (prob != PROB_UNKNOWN) return prob; 1371 1372 prob = PROB_FAIR; // Set default value 1373 if (btest == BoolTest::eq) // Exactly equal test? 1374 prob = PROB_STATIC_INFREQUENT; // Assume its relatively infrequent 1375 else if (btest == BoolTest::ne) 1376 prob = PROB_STATIC_FREQUENT; // Assume its relatively frequent 1377 1378 // If this is a conditional test guarding a backwards branch, 1379 // assume its a loop-back edge. Make it a likely taken branch. 1380 if (target_bci < bci()) { 1381 if (is_osr_parse()) { // Could be a hot OSR'd loop; force deopt 1382 // Since it's an OSR, we probably have profile data, but since 1383 // branch_prediction returned PROB_UNKNOWN, the counts are too small. 1384 // Let's make a special check here for completely zero counts. 1385 ciMethodData* methodData = method()->method_data(); 1386 if (!methodData->is_empty()) { 1387 ciProfileData* data = methodData->bci_to_data(bci()); 1388 // Only stop for truly zero counts, which mean an unknown part 1389 // of the OSR-ed method, and we want to deopt to gather more stats. 1390 // If you have ANY counts, then this loop is simply 'cold' relative 1391 // to the OSR loop. 1392 if (data == NULL || 1393 (data->as_BranchData()->taken() + data->as_BranchData()->not_taken() == 0)) { 1394 // This is the only way to return PROB_UNKNOWN: 1395 return PROB_UNKNOWN; 1396 } 1397 } 1398 } 1399 prob = PROB_STATIC_FREQUENT; // Likely to take backwards branch 1400 } 1401 1402 assert(prob != PROB_UNKNOWN, "must have some guess at this point"); 1403 return prob; 1404 } 1405 1406 // The magic constants are chosen so as to match the output of 1407 // branch_prediction() when the profile reports a zero taken count. 1408 // It is important to distinguish zero counts unambiguously, because 1409 // some branches (e.g., _213_javac.Assembler.eliminate) validly produce 1410 // very small but nonzero probabilities, which if confused with zero 1411 // counts would keep the program recompiling indefinitely. 1412 bool Parse::seems_never_taken(float prob) const { 1413 return prob < PROB_MIN; 1414 } 1415 1416 // True if the comparison seems to be the kind that will not change its 1417 // statistics from true to false. See comments in adjust_map_after_if. 1418 // This question is only asked along paths which are already 1419 // classifed as untaken (by seems_never_taken), so really, 1420 // if a path is never taken, its controlling comparison is 1421 // already acting in a stable fashion. If the comparison 1422 // seems stable, we will put an expensive uncommon trap 1423 // on the untaken path. 1424 bool Parse::seems_stable_comparison() const { 1425 if (C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if)) { 1426 return false; 1427 } 1428 return true; 1429 } 1430 1431 //-------------------------------repush_if_args-------------------------------- 1432 // Push arguments of an "if" bytecode back onto the stack by adjusting _sp. 1433 inline int Parse::repush_if_args() { 1434 if (PrintOpto && WizardMode) { 1435 tty->print("defending against excessive implicit null exceptions on %s @%d in ", 1436 Bytecodes::name(iter().cur_bc()), iter().cur_bci()); 1437 method()->print_name(); tty->cr(); 1438 } 1439 int bc_depth = - Bytecodes::depth(iter().cur_bc()); 1440 assert(bc_depth == 1 || bc_depth == 2, "only two kinds of branches"); 1441 DEBUG_ONLY(sync_jvms()); // argument(n) requires a synced jvms 1442 assert(argument(0) != NULL, "must exist"); 1443 assert(bc_depth == 1 || argument(1) != NULL, "two must exist"); 1444 inc_sp(bc_depth); 1445 return bc_depth; 1446 } 1447 1448 //----------------------------------do_ifnull---------------------------------- 1449 void Parse::do_ifnull(BoolTest::mask btest, Node *c) { 1450 int target_bci = iter().get_dest(); 1451 1452 Block* branch_block = successor_for_bci(target_bci); 1453 Block* next_block = successor_for_bci(iter().next_bci()); 1454 1455 float cnt; 1456 float prob = branch_prediction(cnt, btest, target_bci, c); 1457 if (prob == PROB_UNKNOWN) { 1458 // (An earlier version of do_ifnull omitted this trap for OSR methods.) 1459 if (PrintOpto && Verbose) { 1460 tty->print_cr("Never-taken edge stops compilation at bci %d", bci()); 1461 } 1462 repush_if_args(); // to gather stats on loop 1463 // We need to mark this branch as taken so that if we recompile we will 1464 // see that it is possible. In the tiered system the interpreter doesn't 1465 // do profiling and by the time we get to the lower tier from the interpreter 1466 // the path may be cold again. Make sure it doesn't look untaken 1467 profile_taken_branch(target_bci, !ProfileInterpreter); 1468 uncommon_trap(Deoptimization::Reason_unreached, 1469 Deoptimization::Action_reinterpret, 1470 NULL, "cold"); 1471 if (C->eliminate_boxing()) { 1472 // Mark the successor blocks as parsed 1473 branch_block->next_path_num(); 1474 next_block->next_path_num(); 1475 } 1476 return; 1477 } 1478 1479 NOT_PRODUCT(explicit_null_checks_inserted++); 1480 1481 // Generate real control flow 1482 Node *tst = _gvn.transform( new BoolNode( c, btest ) ); 1483 1484 // Sanity check the probability value 1485 assert(prob > 0.0f,"Bad probability in Parser"); 1486 // Need xform to put node in hash table 1487 IfNode *iff = create_and_xform_if( control(), tst, prob, cnt ); 1488 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser"); 1489 // True branch 1490 { PreserveJVMState pjvms(this); 1491 Node* iftrue = _gvn.transform( new IfTrueNode (iff) ); 1492 set_control(iftrue); 1493 1494 if (stopped()) { // Path is dead? 1495 NOT_PRODUCT(explicit_null_checks_elided++); 1496 if (C->eliminate_boxing()) { 1497 // Mark the successor block as parsed 1498 branch_block->next_path_num(); 1499 } 1500 } else { // Path is live. 1501 // Update method data 1502 profile_taken_branch(target_bci); 1503 adjust_map_after_if(btest, c, prob, branch_block, next_block); 1504 if (!stopped()) { 1505 merge(target_bci); 1506 } 1507 } 1508 } 1509 1510 // False branch 1511 Node* iffalse = _gvn.transform( new IfFalseNode(iff) ); 1512 set_control(iffalse); 1513 1514 if (stopped()) { // Path is dead? 1515 NOT_PRODUCT(explicit_null_checks_elided++); 1516 if (C->eliminate_boxing()) { 1517 // Mark the successor block as parsed 1518 next_block->next_path_num(); 1519 } 1520 } else { // Path is live. 1521 // Update method data 1522 profile_not_taken_branch(); 1523 adjust_map_after_if(BoolTest(btest).negate(), c, 1.0-prob, 1524 next_block, branch_block); 1525 } 1526 } 1527 1528 //------------------------------------do_if------------------------------------ 1529 void Parse::do_if(BoolTest::mask btest, Node* c) { 1530 int target_bci = iter().get_dest(); 1531 1532 Block* branch_block = successor_for_bci(target_bci); 1533 Block* next_block = successor_for_bci(iter().next_bci()); 1534 1535 float cnt; 1536 float prob = branch_prediction(cnt, btest, target_bci, c); 1537 float untaken_prob = 1.0 - prob; 1538 1539 if (prob == PROB_UNKNOWN) { 1540 if (PrintOpto && Verbose) { 1541 tty->print_cr("Never-taken edge stops compilation at bci %d", bci()); 1542 } 1543 repush_if_args(); // to gather stats on loop 1544 // We need to mark this branch as taken so that if we recompile we will 1545 // see that it is possible. In the tiered system the interpreter doesn't 1546 // do profiling and by the time we get to the lower tier from the interpreter 1547 // the path may be cold again. Make sure it doesn't look untaken 1548 profile_taken_branch(target_bci, !ProfileInterpreter); 1549 uncommon_trap(Deoptimization::Reason_unreached, 1550 Deoptimization::Action_reinterpret, 1551 NULL, "cold"); 1552 if (C->eliminate_boxing()) { 1553 // Mark the successor blocks as parsed 1554 branch_block->next_path_num(); 1555 next_block->next_path_num(); 1556 } 1557 return; 1558 } 1559 1560 // Sanity check the probability value 1561 assert(0.0f < prob && prob < 1.0f,"Bad probability in Parser"); 1562 1563 bool taken_if_true = true; 1564 // Convert BoolTest to canonical form: 1565 if (!BoolTest(btest).is_canonical()) { 1566 btest = BoolTest(btest).negate(); 1567 taken_if_true = false; 1568 // prob is NOT updated here; it remains the probability of the taken 1569 // path (as opposed to the prob of the path guarded by an 'IfTrueNode'). 1570 } 1571 assert(btest != BoolTest::eq, "!= is the only canonical exact test"); 1572 1573 Node* tst0 = new BoolNode(c, btest); 1574 Node* tst = _gvn.transform(tst0); 1575 BoolTest::mask taken_btest = BoolTest::illegal; 1576 BoolTest::mask untaken_btest = BoolTest::illegal; 1577 1578 if (tst->is_Bool()) { 1579 // Refresh c from the transformed bool node, since it may be 1580 // simpler than the original c. Also re-canonicalize btest. 1581 // This wins when (Bool ne (Conv2B p) 0) => (Bool ne (CmpP p NULL)). 1582 // That can arise from statements like: if (x instanceof C) ... 1583 if (tst != tst0) { 1584 // Canonicalize one more time since transform can change it. 1585 btest = tst->as_Bool()->_test._test; 1586 if (!BoolTest(btest).is_canonical()) { 1587 // Reverse edges one more time... 1588 tst = _gvn.transform( tst->as_Bool()->negate(&_gvn) ); 1589 btest = tst->as_Bool()->_test._test; 1590 assert(BoolTest(btest).is_canonical(), "sanity"); 1591 taken_if_true = !taken_if_true; 1592 } 1593 c = tst->in(1); 1594 } 1595 BoolTest::mask neg_btest = BoolTest(btest).negate(); 1596 taken_btest = taken_if_true ? btest : neg_btest; 1597 untaken_btest = taken_if_true ? neg_btest : btest; 1598 } 1599 1600 // Generate real control flow 1601 float true_prob = (taken_if_true ? prob : untaken_prob); 1602 IfNode* iff = create_and_map_if(control(), tst, true_prob, cnt); 1603 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser"); 1604 Node* taken_branch = new IfTrueNode(iff); 1605 Node* untaken_branch = new IfFalseNode(iff); 1606 if (!taken_if_true) { // Finish conversion to canonical form 1607 Node* tmp = taken_branch; 1608 taken_branch = untaken_branch; 1609 untaken_branch = tmp; 1610 } 1611 1612 // Branch is taken: 1613 { PreserveJVMState pjvms(this); 1614 taken_branch = _gvn.transform(taken_branch); 1615 set_control(taken_branch); 1616 1617 if (stopped()) { 1618 if (C->eliminate_boxing()) { 1619 // Mark the successor block as parsed 1620 branch_block->next_path_num(); 1621 } 1622 } else { 1623 // Update method data 1624 profile_taken_branch(target_bci); 1625 adjust_map_after_if(taken_btest, c, prob, branch_block, next_block); 1626 if (!stopped()) { 1627 merge(target_bci); 1628 } 1629 } 1630 } 1631 1632 untaken_branch = _gvn.transform(untaken_branch); 1633 set_control(untaken_branch); 1634 1635 // Branch not taken. 1636 if (stopped()) { 1637 if (C->eliminate_boxing()) { 1638 // Mark the successor block as parsed 1639 next_block->next_path_num(); 1640 } 1641 } else { 1642 // Update method data 1643 profile_not_taken_branch(); 1644 adjust_map_after_if(untaken_btest, c, untaken_prob, 1645 next_block, branch_block); 1646 } 1647 } 1648 1649 bool Parse::path_is_suitable_for_uncommon_trap(float prob) const { 1650 // Don't want to speculate on uncommon traps when running with -Xcomp 1651 if (!UseInterpreter) { 1652 return false; 1653 } 1654 return (seems_never_taken(prob) && seems_stable_comparison()); 1655 } 1656 1657 void Parse::maybe_add_predicate_after_if(Block* path) { 1658 if (path->is_SEL_head() && path->preds_parsed() == 0) { 1659 // Add predicates at bci of if dominating the loop so traps can be 1660 // recorded on the if's profile data 1661 int bc_depth = repush_if_args(); 1662 add_empty_predicates(); 1663 dec_sp(bc_depth); 1664 path->set_has_predicates(); 1665 } 1666 } 1667 1668 1669 //----------------------------adjust_map_after_if------------------------------ 1670 // Adjust the JVM state to reflect the result of taking this path. 1671 // Basically, it means inspecting the CmpNode controlling this 1672 // branch, seeing how it constrains a tested value, and then 1673 // deciding if it's worth our while to encode this constraint 1674 // as graph nodes in the current abstract interpretation map. 1675 void Parse::adjust_map_after_if(BoolTest::mask btest, Node* c, float prob, 1676 Block* path, Block* other_path) { 1677 if (!c->is_Cmp()) { 1678 maybe_add_predicate_after_if(path); 1679 return; 1680 } 1681 1682 if (stopped() || btest == BoolTest::illegal) { 1683 return; // nothing to do 1684 } 1685 1686 bool is_fallthrough = (path == successor_for_bci(iter().next_bci())); 1687 1688 if (path_is_suitable_for_uncommon_trap(prob)) { 1689 repush_if_args(); 1690 uncommon_trap(Deoptimization::Reason_unstable_if, 1691 Deoptimization::Action_reinterpret, 1692 NULL, 1693 (is_fallthrough ? "taken always" : "taken never")); 1694 return; 1695 } 1696 1697 Node* val = c->in(1); 1698 Node* con = c->in(2); 1699 const Type* tcon = _gvn.type(con); 1700 const Type* tval = _gvn.type(val); 1701 bool have_con = tcon->singleton(); 1702 if (tval->singleton()) { 1703 if (!have_con) { 1704 // Swap, so constant is in con. 1705 con = val; 1706 tcon = tval; 1707 val = c->in(2); 1708 tval = _gvn.type(val); 1709 btest = BoolTest(btest).commute(); 1710 have_con = true; 1711 } else { 1712 // Do we have two constants? Then leave well enough alone. 1713 have_con = false; 1714 } 1715 } 1716 if (!have_con) { // remaining adjustments need a con 1717 maybe_add_predicate_after_if(path); 1718 return; 1719 } 1720 1721 sharpen_type_after_if(btest, con, tcon, val, tval); 1722 maybe_add_predicate_after_if(path); 1723 } 1724 1725 1726 static Node* extract_obj_from_klass_load(PhaseGVN* gvn, Node* n) { 1727 Node* ldk; 1728 if (n->is_DecodeNKlass()) { 1729 if (n->in(1)->Opcode() != Op_LoadNKlass) { 1730 return NULL; 1731 } else { 1732 ldk = n->in(1); 1733 } 1734 } else if (n->Opcode() != Op_LoadKlass) { 1735 return NULL; 1736 } else { 1737 ldk = n; 1738 } 1739 assert(ldk != NULL && ldk->is_Load(), "should have found a LoadKlass or LoadNKlass node"); 1740 1741 Node* adr = ldk->in(MemNode::Address); 1742 intptr_t off = 0; 1743 Node* obj = AddPNode::Ideal_base_and_offset(adr, gvn, off); 1744 if (obj == NULL || off != oopDesc::klass_offset_in_bytes()) // loading oopDesc::_klass? 1745 return NULL; 1746 const TypePtr* tp = gvn->type(obj)->is_ptr(); 1747 if (tp == NULL || !(tp->isa_instptr() || tp->isa_aryptr())) // is obj a Java object ptr? 1748 return NULL; 1749 1750 return obj; 1751 } 1752 1753 void Parse::sharpen_type_after_if(BoolTest::mask btest, 1754 Node* con, const Type* tcon, 1755 Node* val, const Type* tval) { 1756 // Look for opportunities to sharpen the type of a node 1757 // whose klass is compared with a constant klass. 1758 if (btest == BoolTest::eq && tcon->isa_klassptr()) { 1759 Node* obj = extract_obj_from_klass_load(&_gvn, val); 1760 const TypeOopPtr* con_type = tcon->isa_klassptr()->as_instance_type(); 1761 if (obj != NULL && (con_type->isa_instptr() || con_type->isa_aryptr())) { 1762 // Found: 1763 // Bool(CmpP(LoadKlass(obj._klass), ConP(Foo.klass)), [eq]) 1764 // or the narrowOop equivalent. 1765 const Type* obj_type = _gvn.type(obj); 1766 const TypeOopPtr* tboth = obj_type->join_speculative(con_type)->isa_oopptr(); 1767 if (tboth != NULL && tboth->klass_is_exact() && tboth != obj_type && 1768 tboth->higher_equal(obj_type)) { 1769 // obj has to be of the exact type Foo if the CmpP succeeds. 1770 int obj_in_map = map()->find_edge(obj); 1771 JVMState* jvms = this->jvms(); 1772 if (obj_in_map >= 0 && 1773 (jvms->is_loc(obj_in_map) || jvms->is_stk(obj_in_map))) { 1774 TypeNode* ccast = new CheckCastPPNode(control(), obj, tboth); 1775 const Type* tcc = ccast->as_Type()->type(); 1776 assert(tcc != obj_type && tcc->higher_equal(obj_type), "must improve"); 1777 // Delay transform() call to allow recovery of pre-cast value 1778 // at the control merge. 1779 _gvn.set_type_bottom(ccast); 1780 record_for_igvn(ccast); 1781 // Here's the payoff. 1782 replace_in_map(obj, ccast); 1783 } 1784 } 1785 } 1786 } 1787 1788 int val_in_map = map()->find_edge(val); 1789 if (val_in_map < 0) return; // replace_in_map would be useless 1790 { 1791 JVMState* jvms = this->jvms(); 1792 if (!(jvms->is_loc(val_in_map) || 1793 jvms->is_stk(val_in_map))) 1794 return; // again, it would be useless 1795 } 1796 1797 // Check for a comparison to a constant, and "know" that the compared 1798 // value is constrained on this path. 1799 assert(tcon->singleton(), ""); 1800 ConstraintCastNode* ccast = NULL; 1801 Node* cast = NULL; 1802 1803 switch (btest) { 1804 case BoolTest::eq: // Constant test? 1805 { 1806 const Type* tboth = tcon->join_speculative(tval); 1807 if (tboth == tval) break; // Nothing to gain. 1808 if (tcon->isa_int()) { 1809 ccast = new CastIINode(val, tboth); 1810 } else if (tcon == TypePtr::NULL_PTR) { 1811 // Cast to null, but keep the pointer identity temporarily live. 1812 ccast = new CastPPNode(val, tboth); 1813 } else { 1814 const TypeF* tf = tcon->isa_float_constant(); 1815 const TypeD* td = tcon->isa_double_constant(); 1816 // Exclude tests vs float/double 0 as these could be 1817 // either +0 or -0. Just because you are equal to +0 1818 // doesn't mean you ARE +0! 1819 // Note, following code also replaces Long and Oop values. 1820 if ((!tf || tf->_f != 0.0) && 1821 (!td || td->_d != 0.0)) 1822 cast = con; // Replace non-constant val by con. 1823 } 1824 } 1825 break; 1826 1827 case BoolTest::ne: 1828 if (tcon == TypePtr::NULL_PTR) { 1829 cast = cast_not_null(val, false); 1830 } 1831 break; 1832 1833 default: 1834 // (At this point we could record int range types with CastII.) 1835 break; 1836 } 1837 1838 if (ccast != NULL) { 1839 const Type* tcc = ccast->as_Type()->type(); 1840 assert(tcc != tval && tcc->higher_equal(tval), "must improve"); 1841 // Delay transform() call to allow recovery of pre-cast value 1842 // at the control merge. 1843 ccast->set_req(0, control()); 1844 _gvn.set_type_bottom(ccast); 1845 record_for_igvn(ccast); 1846 cast = ccast; 1847 } 1848 1849 if (cast != NULL) { // Here's the payoff. 1850 replace_in_map(val, cast); 1851 } 1852 } 1853 1854 /** 1855 * Use speculative type to optimize CmpP node: if comparison is 1856 * against the low level class, cast the object to the speculative 1857 * type if any. CmpP should then go away. 1858 * 1859 * @param c expected CmpP node 1860 * @return result of CmpP on object casted to speculative type 1861 * 1862 */ 1863 Node* Parse::optimize_cmp_with_klass(Node* c) { 1864 // If this is transformed by the _gvn to a comparison with the low 1865 // level klass then we may be able to use speculation 1866 if (c->Opcode() == Op_CmpP && 1867 (c->in(1)->Opcode() == Op_LoadKlass || c->in(1)->Opcode() == Op_DecodeNKlass) && 1868 c->in(2)->is_Con()) { 1869 Node* load_klass = NULL; 1870 Node* decode = NULL; 1871 if (c->in(1)->Opcode() == Op_DecodeNKlass) { 1872 decode = c->in(1); 1873 load_klass = c->in(1)->in(1); 1874 } else { 1875 load_klass = c->in(1); 1876 } 1877 if (load_klass->in(2)->is_AddP()) { 1878 Node* addp = load_klass->in(2); 1879 Node* obj = addp->in(AddPNode::Address); 1880 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 1881 if (obj_type->speculative_type_not_null() != NULL) { 1882 ciKlass* k = obj_type->speculative_type(); 1883 inc_sp(2); 1884 obj = maybe_cast_profiled_obj(obj, k); 1885 dec_sp(2); 1886 // Make the CmpP use the casted obj 1887 addp = basic_plus_adr(obj, addp->in(AddPNode::Offset)); 1888 load_klass = load_klass->clone(); 1889 load_klass->set_req(2, addp); 1890 load_klass = _gvn.transform(load_klass); 1891 if (decode != NULL) { 1892 decode = decode->clone(); 1893 decode->set_req(1, load_klass); 1894 load_klass = _gvn.transform(decode); 1895 } 1896 c = c->clone(); 1897 c->set_req(1, load_klass); 1898 c = _gvn.transform(c); 1899 } 1900 } 1901 } 1902 return c; 1903 } 1904 1905 //------------------------------do_one_bytecode-------------------------------- 1906 // Parse this bytecode, and alter the Parsers JVM->Node mapping 1907 void Parse::do_one_bytecode() { 1908 Node *a, *b, *c, *d; // Handy temps 1909 BoolTest::mask btest; 1910 int i; 1911 1912 assert(!has_exceptions(), "bytecode entry state must be clear of throws"); 1913 1914 if (C->check_node_count(NodeLimitFudgeFactor * 5, 1915 "out of nodes parsing method")) { 1916 return; 1917 } 1918 1919 #ifdef ASSERT 1920 // for setting breakpoints 1921 if (TraceOptoParse) { 1922 tty->print(" @"); 1923 dump_bci(bci()); 1924 tty->cr(); 1925 } 1926 #endif 1927 1928 switch (bc()) { 1929 case Bytecodes::_nop: 1930 // do nothing 1931 break; 1932 case Bytecodes::_lconst_0: 1933 push_pair(longcon(0)); 1934 break; 1935 1936 case Bytecodes::_lconst_1: 1937 push_pair(longcon(1)); 1938 break; 1939 1940 case Bytecodes::_fconst_0: 1941 push(zerocon(T_FLOAT)); 1942 break; 1943 1944 case Bytecodes::_fconst_1: 1945 push(makecon(TypeF::ONE)); 1946 break; 1947 1948 case Bytecodes::_fconst_2: 1949 push(makecon(TypeF::make(2.0f))); 1950 break; 1951 1952 case Bytecodes::_dconst_0: 1953 push_pair(zerocon(T_DOUBLE)); 1954 break; 1955 1956 case Bytecodes::_dconst_1: 1957 push_pair(makecon(TypeD::ONE)); 1958 break; 1959 1960 case Bytecodes::_iconst_m1:push(intcon(-1)); break; 1961 case Bytecodes::_iconst_0: push(intcon( 0)); break; 1962 case Bytecodes::_iconst_1: push(intcon( 1)); break; 1963 case Bytecodes::_iconst_2: push(intcon( 2)); break; 1964 case Bytecodes::_iconst_3: push(intcon( 3)); break; 1965 case Bytecodes::_iconst_4: push(intcon( 4)); break; 1966 case Bytecodes::_iconst_5: push(intcon( 5)); break; 1967 case Bytecodes::_bipush: push(intcon(iter().get_constant_u1())); break; 1968 case Bytecodes::_sipush: push(intcon(iter().get_constant_u2())); break; 1969 case Bytecodes::_aconst_null: push(null()); break; 1970 case Bytecodes::_ldc: 1971 case Bytecodes::_ldc_w: 1972 case Bytecodes::_ldc2_w: 1973 // If the constant is unresolved, run this BC once in the interpreter. 1974 { 1975 ciConstant constant = iter().get_constant(); 1976 if (!constant.is_valid() || 1977 (constant.basic_type() == T_OBJECT && 1978 !constant.as_object()->is_loaded())) { 1979 int index = iter().get_constant_pool_index(); 1980 constantTag tag = iter().get_constant_pool_tag(index); 1981 uncommon_trap(Deoptimization::make_trap_request 1982 (Deoptimization::Reason_unloaded, 1983 Deoptimization::Action_reinterpret, 1984 index), 1985 NULL, tag.internal_name()); 1986 break; 1987 } 1988 assert(constant.basic_type() != T_OBJECT || constant.as_object()->is_instance(), 1989 "must be java_mirror of klass"); 1990 const Type* con_type = Type::make_from_constant(constant); 1991 if (con_type != NULL) { 1992 push_node(con_type->basic_type(), makecon(con_type)); 1993 } 1994 } 1995 1996 break; 1997 1998 case Bytecodes::_aload_0: 1999 push( local(0) ); 2000 break; 2001 case Bytecodes::_aload_1: 2002 push( local(1) ); 2003 break; 2004 case Bytecodes::_aload_2: 2005 push( local(2) ); 2006 break; 2007 case Bytecodes::_aload_3: 2008 push( local(3) ); 2009 break; 2010 case Bytecodes::_aload: 2011 push( local(iter().get_index()) ); 2012 break; 2013 2014 case Bytecodes::_fload_0: 2015 case Bytecodes::_iload_0: 2016 push( local(0) ); 2017 break; 2018 case Bytecodes::_fload_1: 2019 case Bytecodes::_iload_1: 2020 push( local(1) ); 2021 break; 2022 case Bytecodes::_fload_2: 2023 case Bytecodes::_iload_2: 2024 push( local(2) ); 2025 break; 2026 case Bytecodes::_fload_3: 2027 case Bytecodes::_iload_3: 2028 push( local(3) ); 2029 break; 2030 case Bytecodes::_fload: 2031 case Bytecodes::_iload: 2032 push( local(iter().get_index()) ); 2033 break; 2034 case Bytecodes::_lload_0: 2035 push_pair_local( 0 ); 2036 break; 2037 case Bytecodes::_lload_1: 2038 push_pair_local( 1 ); 2039 break; 2040 case Bytecodes::_lload_2: 2041 push_pair_local( 2 ); 2042 break; 2043 case Bytecodes::_lload_3: 2044 push_pair_local( 3 ); 2045 break; 2046 case Bytecodes::_lload: 2047 push_pair_local( iter().get_index() ); 2048 break; 2049 2050 case Bytecodes::_dload_0: 2051 push_pair_local(0); 2052 break; 2053 case Bytecodes::_dload_1: 2054 push_pair_local(1); 2055 break; 2056 case Bytecodes::_dload_2: 2057 push_pair_local(2); 2058 break; 2059 case Bytecodes::_dload_3: 2060 push_pair_local(3); 2061 break; 2062 case Bytecodes::_dload: 2063 push_pair_local(iter().get_index()); 2064 break; 2065 case Bytecodes::_fstore_0: 2066 case Bytecodes::_istore_0: 2067 case Bytecodes::_astore_0: 2068 set_local( 0, pop() ); 2069 break; 2070 case Bytecodes::_fstore_1: 2071 case Bytecodes::_istore_1: 2072 case Bytecodes::_astore_1: 2073 set_local( 1, pop() ); 2074 break; 2075 case Bytecodes::_fstore_2: 2076 case Bytecodes::_istore_2: 2077 case Bytecodes::_astore_2: 2078 set_local( 2, pop() ); 2079 break; 2080 case Bytecodes::_fstore_3: 2081 case Bytecodes::_istore_3: 2082 case Bytecodes::_astore_3: 2083 set_local( 3, pop() ); 2084 break; 2085 case Bytecodes::_fstore: 2086 case Bytecodes::_istore: 2087 case Bytecodes::_astore: 2088 set_local( iter().get_index(), pop() ); 2089 break; 2090 // long stores 2091 case Bytecodes::_lstore_0: 2092 set_pair_local( 0, pop_pair() ); 2093 break; 2094 case Bytecodes::_lstore_1: 2095 set_pair_local( 1, pop_pair() ); 2096 break; 2097 case Bytecodes::_lstore_2: 2098 set_pair_local( 2, pop_pair() ); 2099 break; 2100 case Bytecodes::_lstore_3: 2101 set_pair_local( 3, pop_pair() ); 2102 break; 2103 case Bytecodes::_lstore: 2104 set_pair_local( iter().get_index(), pop_pair() ); 2105 break; 2106 2107 // double stores 2108 case Bytecodes::_dstore_0: 2109 set_pair_local( 0, dstore_rounding(pop_pair()) ); 2110 break; 2111 case Bytecodes::_dstore_1: 2112 set_pair_local( 1, dstore_rounding(pop_pair()) ); 2113 break; 2114 case Bytecodes::_dstore_2: 2115 set_pair_local( 2, dstore_rounding(pop_pair()) ); 2116 break; 2117 case Bytecodes::_dstore_3: 2118 set_pair_local( 3, dstore_rounding(pop_pair()) ); 2119 break; 2120 case Bytecodes::_dstore: 2121 set_pair_local( iter().get_index(), dstore_rounding(pop_pair()) ); 2122 break; 2123 2124 case Bytecodes::_pop: dec_sp(1); break; 2125 case Bytecodes::_pop2: dec_sp(2); break; 2126 case Bytecodes::_swap: 2127 a = pop(); 2128 b = pop(); 2129 push(a); 2130 push(b); 2131 break; 2132 case Bytecodes::_dup: 2133 a = pop(); 2134 push(a); 2135 push(a); 2136 break; 2137 case Bytecodes::_dup_x1: 2138 a = pop(); 2139 b = pop(); 2140 push( a ); 2141 push( b ); 2142 push( a ); 2143 break; 2144 case Bytecodes::_dup_x2: 2145 a = pop(); 2146 b = pop(); 2147 c = pop(); 2148 push( a ); 2149 push( c ); 2150 push( b ); 2151 push( a ); 2152 break; 2153 case Bytecodes::_dup2: 2154 a = pop(); 2155 b = pop(); 2156 push( b ); 2157 push( a ); 2158 push( b ); 2159 push( a ); 2160 break; 2161 2162 case Bytecodes::_dup2_x1: 2163 // before: .. c, b, a 2164 // after: .. b, a, c, b, a 2165 // not tested 2166 a = pop(); 2167 b = pop(); 2168 c = pop(); 2169 push( b ); 2170 push( a ); 2171 push( c ); 2172 push( b ); 2173 push( a ); 2174 break; 2175 case Bytecodes::_dup2_x2: 2176 // before: .. d, c, b, a 2177 // after: .. b, a, d, c, b, a 2178 // not tested 2179 a = pop(); 2180 b = pop(); 2181 c = pop(); 2182 d = pop(); 2183 push( b ); 2184 push( a ); 2185 push( d ); 2186 push( c ); 2187 push( b ); 2188 push( a ); 2189 break; 2190 2191 case Bytecodes::_arraylength: { 2192 // Must do null-check with value on expression stack 2193 Node *ary = null_check(peek(), T_ARRAY); 2194 // Compile-time detect of null-exception? 2195 if (stopped()) return; 2196 a = pop(); 2197 push(load_array_length(a)); 2198 break; 2199 } 2200 2201 case Bytecodes::_baload: array_load(T_BYTE); break; 2202 case Bytecodes::_caload: array_load(T_CHAR); break; 2203 case Bytecodes::_iaload: array_load(T_INT); break; 2204 case Bytecodes::_saload: array_load(T_SHORT); break; 2205 case Bytecodes::_faload: array_load(T_FLOAT); break; 2206 case Bytecodes::_aaload: array_load(T_OBJECT); break; 2207 case Bytecodes::_laload: array_load(T_LONG); break; 2208 case Bytecodes::_daload: array_load(T_DOUBLE); break; 2209 case Bytecodes::_bastore: array_store(T_BYTE); break; 2210 case Bytecodes::_castore: array_store(T_CHAR); break; 2211 case Bytecodes::_iastore: array_store(T_INT); break; 2212 case Bytecodes::_sastore: array_store(T_SHORT); break; 2213 case Bytecodes::_fastore: array_store(T_FLOAT); break; 2214 case Bytecodes::_aastore: array_store(T_OBJECT); break; 2215 case Bytecodes::_lastore: array_store(T_LONG); break; 2216 case Bytecodes::_dastore: array_store(T_DOUBLE); break; 2217 2218 case Bytecodes::_getfield: 2219 do_getfield(); 2220 break; 2221 2222 case Bytecodes::_getstatic: 2223 do_getstatic(); 2224 break; 2225 2226 case Bytecodes::_putfield: 2227 do_putfield(); 2228 break; 2229 2230 case Bytecodes::_putstatic: 2231 do_putstatic(); 2232 break; 2233 2234 case Bytecodes::_irem: 2235 do_irem(); 2236 break; 2237 case Bytecodes::_idiv: 2238 // Must keep both values on the expression-stack during null-check 2239 zero_check_int(peek()); 2240 // Compile-time detect of null-exception? 2241 if (stopped()) return; 2242 b = pop(); 2243 a = pop(); 2244 push( _gvn.transform( new DivINode(control(),a,b) ) ); 2245 break; 2246 case Bytecodes::_imul: 2247 b = pop(); a = pop(); 2248 push( _gvn.transform( new MulINode(a,b) ) ); 2249 break; 2250 case Bytecodes::_iadd: 2251 b = pop(); a = pop(); 2252 push( _gvn.transform( new AddINode(a,b) ) ); 2253 break; 2254 case Bytecodes::_ineg: 2255 a = pop(); 2256 push( _gvn.transform( new SubINode(_gvn.intcon(0),a)) ); 2257 break; 2258 case Bytecodes::_isub: 2259 b = pop(); a = pop(); 2260 push( _gvn.transform( new SubINode(a,b) ) ); 2261 break; 2262 case Bytecodes::_iand: 2263 b = pop(); a = pop(); 2264 push( _gvn.transform( new AndINode(a,b) ) ); 2265 break; 2266 case Bytecodes::_ior: 2267 b = pop(); a = pop(); 2268 push( _gvn.transform( new OrINode(a,b) ) ); 2269 break; 2270 case Bytecodes::_ixor: 2271 b = pop(); a = pop(); 2272 push( _gvn.transform( new XorINode(a,b) ) ); 2273 break; 2274 case Bytecodes::_ishl: 2275 b = pop(); a = pop(); 2276 push( _gvn.transform( new LShiftINode(a,b) ) ); 2277 break; 2278 case Bytecodes::_ishr: 2279 b = pop(); a = pop(); 2280 push( _gvn.transform( new RShiftINode(a,b) ) ); 2281 break; 2282 case Bytecodes::_iushr: 2283 b = pop(); a = pop(); 2284 push( _gvn.transform( new URShiftINode(a,b) ) ); 2285 break; 2286 2287 case Bytecodes::_fneg: 2288 a = pop(); 2289 b = _gvn.transform(new NegFNode (a)); 2290 push(b); 2291 break; 2292 2293 case Bytecodes::_fsub: 2294 b = pop(); 2295 a = pop(); 2296 c = _gvn.transform( new SubFNode(a,b) ); 2297 d = precision_rounding(c); 2298 push( d ); 2299 break; 2300 2301 case Bytecodes::_fadd: 2302 b = pop(); 2303 a = pop(); 2304 c = _gvn.transform( new AddFNode(a,b) ); 2305 d = precision_rounding(c); 2306 push( d ); 2307 break; 2308 2309 case Bytecodes::_fmul: 2310 b = pop(); 2311 a = pop(); 2312 c = _gvn.transform( new MulFNode(a,b) ); 2313 d = precision_rounding(c); 2314 push( d ); 2315 break; 2316 2317 case Bytecodes::_fdiv: 2318 b = pop(); 2319 a = pop(); 2320 c = _gvn.transform( new DivFNode(0,a,b) ); 2321 d = precision_rounding(c); 2322 push( d ); 2323 break; 2324 2325 case Bytecodes::_frem: 2326 if (Matcher::has_match_rule(Op_ModF)) { 2327 // Generate a ModF node. 2328 b = pop(); 2329 a = pop(); 2330 c = _gvn.transform( new ModFNode(0,a,b) ); 2331 d = precision_rounding(c); 2332 push( d ); 2333 } 2334 else { 2335 // Generate a call. 2336 modf(); 2337 } 2338 break; 2339 2340 case Bytecodes::_fcmpl: 2341 b = pop(); 2342 a = pop(); 2343 c = _gvn.transform( new CmpF3Node( a, b)); 2344 push(c); 2345 break; 2346 case Bytecodes::_fcmpg: 2347 b = pop(); 2348 a = pop(); 2349 2350 // Same as fcmpl but need to flip the unordered case. Swap the inputs, 2351 // which negates the result sign except for unordered. Flip the unordered 2352 // as well by using CmpF3 which implements unordered-lesser instead of 2353 // unordered-greater semantics. Finally, commute the result bits. Result 2354 // is same as using a CmpF3Greater except we did it with CmpF3 alone. 2355 c = _gvn.transform( new CmpF3Node( b, a)); 2356 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) ); 2357 push(c); 2358 break; 2359 2360 case Bytecodes::_f2i: 2361 a = pop(); 2362 push(_gvn.transform(new ConvF2INode(a))); 2363 break; 2364 2365 case Bytecodes::_d2i: 2366 a = pop_pair(); 2367 b = _gvn.transform(new ConvD2INode(a)); 2368 push( b ); 2369 break; 2370 2371 case Bytecodes::_f2d: 2372 a = pop(); 2373 b = _gvn.transform( new ConvF2DNode(a)); 2374 push_pair( b ); 2375 break; 2376 2377 case Bytecodes::_d2f: 2378 a = pop_pair(); 2379 b = _gvn.transform( new ConvD2FNode(a)); 2380 // This breaks _227_mtrt (speed & correctness) and _222_mpegaudio (speed) 2381 //b = _gvn.transform(new RoundFloatNode(0, b) ); 2382 push( b ); 2383 break; 2384 2385 case Bytecodes::_l2f: 2386 if (Matcher::convL2FSupported()) { 2387 a = pop_pair(); 2388 b = _gvn.transform( new ConvL2FNode(a)); 2389 // For i486.ad, FILD doesn't restrict precision to 24 or 53 bits. 2390 // Rather than storing the result into an FP register then pushing 2391 // out to memory to round, the machine instruction that implements 2392 // ConvL2D is responsible for rounding. 2393 // c = precision_rounding(b); 2394 c = _gvn.transform(b); 2395 push(c); 2396 } else { 2397 l2f(); 2398 } 2399 break; 2400 2401 case Bytecodes::_l2d: 2402 a = pop_pair(); 2403 b = _gvn.transform( new ConvL2DNode(a)); 2404 // For i486.ad, rounding is always necessary (see _l2f above). 2405 // c = dprecision_rounding(b); 2406 c = _gvn.transform(b); 2407 push_pair(c); 2408 break; 2409 2410 case Bytecodes::_f2l: 2411 a = pop(); 2412 b = _gvn.transform( new ConvF2LNode(a)); 2413 push_pair(b); 2414 break; 2415 2416 case Bytecodes::_d2l: 2417 a = pop_pair(); 2418 b = _gvn.transform( new ConvD2LNode(a)); 2419 push_pair(b); 2420 break; 2421 2422 case Bytecodes::_dsub: 2423 b = pop_pair(); 2424 a = pop_pair(); 2425 c = _gvn.transform( new SubDNode(a,b) ); 2426 d = dprecision_rounding(c); 2427 push_pair( d ); 2428 break; 2429 2430 case Bytecodes::_dadd: 2431 b = pop_pair(); 2432 a = pop_pair(); 2433 c = _gvn.transform( new AddDNode(a,b) ); 2434 d = dprecision_rounding(c); 2435 push_pair( d ); 2436 break; 2437 2438 case Bytecodes::_dmul: 2439 b = pop_pair(); 2440 a = pop_pair(); 2441 c = _gvn.transform( new MulDNode(a,b) ); 2442 d = dprecision_rounding(c); 2443 push_pair( d ); 2444 break; 2445 2446 case Bytecodes::_ddiv: 2447 b = pop_pair(); 2448 a = pop_pair(); 2449 c = _gvn.transform( new DivDNode(0,a,b) ); 2450 d = dprecision_rounding(c); 2451 push_pair( d ); 2452 break; 2453 2454 case Bytecodes::_dneg: 2455 a = pop_pair(); 2456 b = _gvn.transform(new NegDNode (a)); 2457 push_pair(b); 2458 break; 2459 2460 case Bytecodes::_drem: 2461 if (Matcher::has_match_rule(Op_ModD)) { 2462 // Generate a ModD node. 2463 b = pop_pair(); 2464 a = pop_pair(); 2465 // a % b 2466 2467 c = _gvn.transform( new ModDNode(0,a,b) ); 2468 d = dprecision_rounding(c); 2469 push_pair( d ); 2470 } 2471 else { 2472 // Generate a call. 2473 modd(); 2474 } 2475 break; 2476 2477 case Bytecodes::_dcmpl: 2478 b = pop_pair(); 2479 a = pop_pair(); 2480 c = _gvn.transform( new CmpD3Node( a, b)); 2481 push(c); 2482 break; 2483 2484 case Bytecodes::_dcmpg: 2485 b = pop_pair(); 2486 a = pop_pair(); 2487 // Same as dcmpl but need to flip the unordered case. 2488 // Commute the inputs, which negates the result sign except for unordered. 2489 // Flip the unordered as well by using CmpD3 which implements 2490 // unordered-lesser instead of unordered-greater semantics. 2491 // Finally, negate the result bits. Result is same as using a 2492 // CmpD3Greater except we did it with CmpD3 alone. 2493 c = _gvn.transform( new CmpD3Node( b, a)); 2494 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) ); 2495 push(c); 2496 break; 2497 2498 2499 // Note for longs -> lo word is on TOS, hi word is on TOS - 1 2500 case Bytecodes::_land: 2501 b = pop_pair(); 2502 a = pop_pair(); 2503 c = _gvn.transform( new AndLNode(a,b) ); 2504 push_pair(c); 2505 break; 2506 case Bytecodes::_lor: 2507 b = pop_pair(); 2508 a = pop_pair(); 2509 c = _gvn.transform( new OrLNode(a,b) ); 2510 push_pair(c); 2511 break; 2512 case Bytecodes::_lxor: 2513 b = pop_pair(); 2514 a = pop_pair(); 2515 c = _gvn.transform( new XorLNode(a,b) ); 2516 push_pair(c); 2517 break; 2518 2519 case Bytecodes::_lshl: 2520 b = pop(); // the shift count 2521 a = pop_pair(); // value to be shifted 2522 c = _gvn.transform( new LShiftLNode(a,b) ); 2523 push_pair(c); 2524 break; 2525 case Bytecodes::_lshr: 2526 b = pop(); // the shift count 2527 a = pop_pair(); // value to be shifted 2528 c = _gvn.transform( new RShiftLNode(a,b) ); 2529 push_pair(c); 2530 break; 2531 case Bytecodes::_lushr: 2532 b = pop(); // the shift count 2533 a = pop_pair(); // value to be shifted 2534 c = _gvn.transform( new URShiftLNode(a,b) ); 2535 push_pair(c); 2536 break; 2537 case Bytecodes::_lmul: 2538 b = pop_pair(); 2539 a = pop_pair(); 2540 c = _gvn.transform( new MulLNode(a,b) ); 2541 push_pair(c); 2542 break; 2543 2544 case Bytecodes::_lrem: 2545 // Must keep both values on the expression-stack during null-check 2546 assert(peek(0) == top(), "long word order"); 2547 zero_check_long(peek(1)); 2548 // Compile-time detect of null-exception? 2549 if (stopped()) return; 2550 b = pop_pair(); 2551 a = pop_pair(); 2552 c = _gvn.transform( new ModLNode(control(),a,b) ); 2553 push_pair(c); 2554 break; 2555 2556 case Bytecodes::_ldiv: 2557 // Must keep both values on the expression-stack during null-check 2558 assert(peek(0) == top(), "long word order"); 2559 zero_check_long(peek(1)); 2560 // Compile-time detect of null-exception? 2561 if (stopped()) return; 2562 b = pop_pair(); 2563 a = pop_pair(); 2564 c = _gvn.transform( new DivLNode(control(),a,b) ); 2565 push_pair(c); 2566 break; 2567 2568 case Bytecodes::_ladd: 2569 b = pop_pair(); 2570 a = pop_pair(); 2571 c = _gvn.transform( new AddLNode(a,b) ); 2572 push_pair(c); 2573 break; 2574 case Bytecodes::_lsub: 2575 b = pop_pair(); 2576 a = pop_pair(); 2577 c = _gvn.transform( new SubLNode(a,b) ); 2578 push_pair(c); 2579 break; 2580 case Bytecodes::_lcmp: 2581 // Safepoints are now inserted _before_ branches. The long-compare 2582 // bytecode painfully produces a 3-way value (-1,0,+1) which requires a 2583 // slew of control flow. These are usually followed by a CmpI vs zero and 2584 // a branch; this pattern then optimizes to the obvious long-compare and 2585 // branch. However, if the branch is backwards there's a Safepoint 2586 // inserted. The inserted Safepoint captures the JVM state at the 2587 // pre-branch point, i.e. it captures the 3-way value. Thus if a 2588 // long-compare is used to control a loop the debug info will force 2589 // computation of the 3-way value, even though the generated code uses a 2590 // long-compare and branch. We try to rectify the situation by inserting 2591 // a SafePoint here and have it dominate and kill the safepoint added at a 2592 // following backwards branch. At this point the JVM state merely holds 2 2593 // longs but not the 3-way value. 2594 if( UseLoopSafepoints ) { 2595 switch( iter().next_bc() ) { 2596 case Bytecodes::_ifgt: 2597 case Bytecodes::_iflt: 2598 case Bytecodes::_ifge: 2599 case Bytecodes::_ifle: 2600 case Bytecodes::_ifne: 2601 case Bytecodes::_ifeq: 2602 // If this is a backwards branch in the bytecodes, add Safepoint 2603 maybe_add_safepoint(iter().next_get_dest()); 2604 default: 2605 break; 2606 } 2607 } 2608 b = pop_pair(); 2609 a = pop_pair(); 2610 c = _gvn.transform( new CmpL3Node( a, b )); 2611 push(c); 2612 break; 2613 2614 case Bytecodes::_lneg: 2615 a = pop_pair(); 2616 b = _gvn.transform( new SubLNode(longcon(0),a)); 2617 push_pair(b); 2618 break; 2619 case Bytecodes::_l2i: 2620 a = pop_pair(); 2621 push( _gvn.transform( new ConvL2INode(a))); 2622 break; 2623 case Bytecodes::_i2l: 2624 a = pop(); 2625 b = _gvn.transform( new ConvI2LNode(a)); 2626 push_pair(b); 2627 break; 2628 case Bytecodes::_i2b: 2629 // Sign extend 2630 a = pop(); 2631 a = _gvn.transform( new LShiftINode(a,_gvn.intcon(24)) ); 2632 a = _gvn.transform( new RShiftINode(a,_gvn.intcon(24)) ); 2633 push( a ); 2634 break; 2635 case Bytecodes::_i2s: 2636 a = pop(); 2637 a = _gvn.transform( new LShiftINode(a,_gvn.intcon(16)) ); 2638 a = _gvn.transform( new RShiftINode(a,_gvn.intcon(16)) ); 2639 push( a ); 2640 break; 2641 case Bytecodes::_i2c: 2642 a = pop(); 2643 push( _gvn.transform( new AndINode(a,_gvn.intcon(0xFFFF)) ) ); 2644 break; 2645 2646 case Bytecodes::_i2f: 2647 a = pop(); 2648 b = _gvn.transform( new ConvI2FNode(a) ) ; 2649 c = precision_rounding(b); 2650 push (b); 2651 break; 2652 2653 case Bytecodes::_i2d: 2654 a = pop(); 2655 b = _gvn.transform( new ConvI2DNode(a)); 2656 push_pair(b); 2657 break; 2658 2659 case Bytecodes::_iinc: // Increment local 2660 i = iter().get_index(); // Get local index 2661 set_local( i, _gvn.transform( new AddINode( _gvn.intcon(iter().get_iinc_con()), local(i) ) ) ); 2662 break; 2663 2664 // Exit points of synchronized methods must have an unlock node 2665 case Bytecodes::_return: 2666 return_current(NULL); 2667 break; 2668 2669 case Bytecodes::_ireturn: 2670 case Bytecodes::_areturn: 2671 case Bytecodes::_freturn: 2672 return_current(pop()); 2673 break; 2674 case Bytecodes::_lreturn: 2675 return_current(pop_pair()); 2676 break; 2677 case Bytecodes::_dreturn: 2678 return_current(pop_pair()); 2679 break; 2680 2681 case Bytecodes::_athrow: 2682 // null exception oop throws NULL pointer exception 2683 null_check(peek()); 2684 if (stopped()) return; 2685 // Hook the thrown exception directly to subsequent handlers. 2686 if (BailoutToInterpreterForThrows) { 2687 // Keep method interpreted from now on. 2688 uncommon_trap(Deoptimization::Reason_unhandled, 2689 Deoptimization::Action_make_not_compilable); 2690 return; 2691 } 2692 if (env()->jvmti_can_post_on_exceptions()) { 2693 // check if we must post exception events, take uncommon trap if so (with must_throw = false) 2694 uncommon_trap_if_should_post_on_exceptions(Deoptimization::Reason_unhandled, false); 2695 } 2696 // Here if either can_post_on_exceptions or should_post_on_exceptions is false 2697 add_exception_state(make_exception_state(peek())); 2698 break; 2699 2700 case Bytecodes::_goto: // fall through 2701 case Bytecodes::_goto_w: { 2702 int target_bci = (bc() == Bytecodes::_goto) ? iter().get_dest() : iter().get_far_dest(); 2703 2704 // If this is a backwards branch in the bytecodes, add Safepoint 2705 maybe_add_safepoint(target_bci); 2706 2707 // Update method data 2708 profile_taken_branch(target_bci); 2709 2710 // Merge the current control into the target basic block 2711 merge(target_bci); 2712 2713 // See if we can get some profile data and hand it off to the next block 2714 Block *target_block = block()->successor_for_bci(target_bci); 2715 if (target_block->pred_count() != 1) break; 2716 ciMethodData* methodData = method()->method_data(); 2717 if (!methodData->is_mature()) break; 2718 ciProfileData* data = methodData->bci_to_data(bci()); 2719 assert(data != NULL && data->is_JumpData(), "need JumpData for taken branch"); 2720 int taken = ((ciJumpData*)data)->taken(); 2721 taken = method()->scale_count(taken); 2722 target_block->set_count(taken); 2723 break; 2724 } 2725 2726 case Bytecodes::_ifnull: btest = BoolTest::eq; goto handle_if_null; 2727 case Bytecodes::_ifnonnull: btest = BoolTest::ne; goto handle_if_null; 2728 handle_if_null: 2729 // If this is a backwards branch in the bytecodes, add Safepoint 2730 maybe_add_safepoint(iter().get_dest()); 2731 a = null(); 2732 b = pop(); 2733 if (!_gvn.type(b)->speculative_maybe_null() && 2734 !too_many_traps(Deoptimization::Reason_speculate_null_check)) { 2735 inc_sp(1); 2736 Node* null_ctl = top(); 2737 b = null_check_oop(b, &null_ctl, true, true, true); 2738 assert(null_ctl->is_top(), "no null control here"); 2739 dec_sp(1); 2740 } else if (_gvn.type(b)->speculative_always_null() && 2741 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) { 2742 inc_sp(1); 2743 b = null_assert(b); 2744 dec_sp(1); 2745 } 2746 c = _gvn.transform( new CmpPNode(b, a) ); 2747 do_ifnull(btest, c); 2748 break; 2749 2750 case Bytecodes::_if_acmpeq: btest = BoolTest::eq; goto handle_if_acmp; 2751 case Bytecodes::_if_acmpne: btest = BoolTest::ne; goto handle_if_acmp; 2752 handle_if_acmp: 2753 // If this is a backwards branch in the bytecodes, add Safepoint 2754 maybe_add_safepoint(iter().get_dest()); 2755 a = pop(); 2756 b = pop(); 2757 c = _gvn.transform( new CmpPNode(b, a) ); 2758 c = optimize_cmp_with_klass(c); 2759 do_if(btest, c); 2760 break; 2761 2762 case Bytecodes::_ifeq: btest = BoolTest::eq; goto handle_ifxx; 2763 case Bytecodes::_ifne: btest = BoolTest::ne; goto handle_ifxx; 2764 case Bytecodes::_iflt: btest = BoolTest::lt; goto handle_ifxx; 2765 case Bytecodes::_ifle: btest = BoolTest::le; goto handle_ifxx; 2766 case Bytecodes::_ifgt: btest = BoolTest::gt; goto handle_ifxx; 2767 case Bytecodes::_ifge: btest = BoolTest::ge; goto handle_ifxx; 2768 handle_ifxx: 2769 // If this is a backwards branch in the bytecodes, add Safepoint 2770 maybe_add_safepoint(iter().get_dest()); 2771 a = _gvn.intcon(0); 2772 b = pop(); 2773 c = _gvn.transform( new CmpINode(b, a) ); 2774 do_if(btest, c); 2775 break; 2776 2777 case Bytecodes::_if_icmpeq: btest = BoolTest::eq; goto handle_if_icmp; 2778 case Bytecodes::_if_icmpne: btest = BoolTest::ne; goto handle_if_icmp; 2779 case Bytecodes::_if_icmplt: btest = BoolTest::lt; goto handle_if_icmp; 2780 case Bytecodes::_if_icmple: btest = BoolTest::le; goto handle_if_icmp; 2781 case Bytecodes::_if_icmpgt: btest = BoolTest::gt; goto handle_if_icmp; 2782 case Bytecodes::_if_icmpge: btest = BoolTest::ge; goto handle_if_icmp; 2783 handle_if_icmp: 2784 // If this is a backwards branch in the bytecodes, add Safepoint 2785 maybe_add_safepoint(iter().get_dest()); 2786 a = pop(); 2787 b = pop(); 2788 c = _gvn.transform( new CmpINode( b, a ) ); 2789 do_if(btest, c); 2790 break; 2791 2792 case Bytecodes::_tableswitch: 2793 do_tableswitch(); 2794 break; 2795 2796 case Bytecodes::_lookupswitch: 2797 do_lookupswitch(); 2798 break; 2799 2800 case Bytecodes::_invokestatic: 2801 case Bytecodes::_invokedynamic: 2802 case Bytecodes::_invokespecial: 2803 case Bytecodes::_invokevirtual: 2804 case Bytecodes::_invokeinterface: 2805 do_call(); 2806 break; 2807 case Bytecodes::_checkcast: 2808 do_checkcast(); 2809 break; 2810 case Bytecodes::_instanceof: 2811 do_instanceof(); 2812 break; 2813 case Bytecodes::_anewarray: 2814 do_anewarray(); 2815 break; 2816 case Bytecodes::_newarray: 2817 do_newarray((BasicType)iter().get_index()); 2818 break; 2819 case Bytecodes::_multianewarray: 2820 do_multianewarray(); 2821 break; 2822 case Bytecodes::_new: 2823 do_new(); 2824 break; 2825 2826 case Bytecodes::_jsr: 2827 case Bytecodes::_jsr_w: 2828 do_jsr(); 2829 break; 2830 2831 case Bytecodes::_ret: 2832 do_ret(); 2833 break; 2834 2835 2836 case Bytecodes::_monitorenter: 2837 do_monitor_enter(); 2838 break; 2839 2840 case Bytecodes::_monitorexit: 2841 do_monitor_exit(); 2842 break; 2843 2844 case Bytecodes::_breakpoint: 2845 // Breakpoint set concurrently to compile 2846 // %%% use an uncommon trap? 2847 C->record_failure("breakpoint in method"); 2848 return; 2849 2850 default: 2851 #ifndef PRODUCT 2852 map()->dump(99); 2853 #endif 2854 tty->print("\nUnhandled bytecode %s\n", Bytecodes::name(bc()) ); 2855 ShouldNotReachHere(); 2856 } 2857 2858 #ifndef PRODUCT 2859 IdealGraphPrinter *printer = C->printer(); 2860 if (printer && printer->should_print(1)) { 2861 char buffer[256]; 2862 sprintf(buffer, "Bytecode %d: %s", bci(), Bytecodes::name(bc())); 2863 bool old = printer->traverse_outs(); 2864 printer->set_traverse_outs(true); 2865 printer->print_method(buffer, 4); 2866 printer->set_traverse_outs(old); 2867 } 2868 #endif 2869 }