1 /* 2 * Copyright (c) 1998, 2022, 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/vmSymbols.hpp" 28 #include "compiler/compileLog.hpp" 29 #include "interpreter/linkResolver.hpp" 30 #include "jvm_io.h" 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(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->instance_klass()->unique_concrete_subklass()) { 128 // If we load from "AbstractClass[]" we must see "ConcreteSubClass". 129 const Type* subklass = Type::get_const_type(toop->instance_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 if (!arytype->is_loaded()) { 147 // Only fails for some -Xcomp runs 148 // The class is unloaded. We have to run this bytecode in the interpreter. 149 ciKlass* klass = arytype->unloaded_klass(); 150 151 uncommon_trap(Deoptimization::Reason_unloaded, 152 Deoptimization::Action_reinterpret, 153 klass, "!loaded array"); 154 return top(); 155 } 156 157 // Do the range check 158 if (GenerateRangeChecks && need_range_check) { 159 Node* tst; 160 if (sizetype->_hi <= 0) { 161 // The greatest array bound is negative, so we can conclude that we're 162 // compiling unreachable code, but the unsigned compare trick used below 163 // only works with non-negative lengths. Instead, hack "tst" to be zero so 164 // the uncommon_trap path will always be taken. 165 tst = _gvn.intcon(0); 166 } else { 167 // Range is constant in array-oop, so we can use the original state of mem 168 Node* len = load_array_length(ary); 169 170 // Test length vs index (standard trick using unsigned compare) 171 Node* chk = _gvn.transform( new CmpUNode(idx, len) ); 172 BoolTest::mask btest = BoolTest::lt; 173 tst = _gvn.transform( new BoolNode(chk, btest) ); 174 } 175 RangeCheckNode* rc = new RangeCheckNode(control(), tst, PROB_MAX, COUNT_UNKNOWN); 176 _gvn.set_type(rc, rc->Value(&_gvn)); 177 if (!tst->is_Con()) { 178 record_for_igvn(rc); 179 } 180 set_control(_gvn.transform(new IfTrueNode(rc))); 181 // Branch to failure if out of bounds 182 { 183 PreserveJVMState pjvms(this); 184 set_control(_gvn.transform(new IfFalseNode(rc))); 185 if (C->allow_range_check_smearing()) { 186 // Do not use builtin_throw, since range checks are sometimes 187 // made more stringent by an optimistic transformation. 188 // This creates "tentative" range checks at this point, 189 // which are not guaranteed to throw exceptions. 190 // See IfNode::Ideal, is_range_check, adjust_check. 191 uncommon_trap(Deoptimization::Reason_range_check, 192 Deoptimization::Action_make_not_entrant, 193 NULL, "range_check"); 194 } else { 195 // If we have already recompiled with the range-check-widening 196 // heroic optimization turned off, then we must really be throwing 197 // range check exceptions. 198 builtin_throw(Deoptimization::Reason_range_check); 199 } 200 } 201 } 202 // Check for always knowing you are throwing a range-check exception 203 if (stopped()) return top(); 204 205 // Make array address computation control dependent to prevent it 206 // from floating above the range check during loop optimizations. 207 Node* ptr = array_element_address(ary, idx, type, sizetype, control()); 208 assert(ptr != top(), "top should go hand-in-hand with stopped"); 209 210 return ptr; 211 } 212 213 214 // returns IfNode 215 IfNode* Parse::jump_if_fork_int(Node* a, Node* b, BoolTest::mask mask, float prob, float cnt) { 216 Node *cmp = _gvn.transform(new CmpINode(a, b)); // two cases: shiftcount > 32 and shiftcount <= 32 217 Node *tst = _gvn.transform(new BoolNode(cmp, mask)); 218 IfNode *iff = create_and_map_if(control(), tst, prob, cnt); 219 return iff; 220 } 221 222 223 // sentinel value for the target bci to mark never taken branches 224 // (according to profiling) 225 static const int never_reached = INT_MAX; 226 227 //------------------------------helper for tableswitch------------------------- 228 void Parse::jump_if_true_fork(IfNode *iff, int dest_bci_if_true, bool unc) { 229 // True branch, use existing map info 230 { PreserveJVMState pjvms(this); 231 Node *iftrue = _gvn.transform( new IfTrueNode (iff) ); 232 set_control( iftrue ); 233 if (unc) { 234 repush_if_args(); 235 uncommon_trap(Deoptimization::Reason_unstable_if, 236 Deoptimization::Action_reinterpret, 237 NULL, 238 "taken always"); 239 } else { 240 assert(dest_bci_if_true != never_reached, "inconsistent dest"); 241 merge_new_path(dest_bci_if_true); 242 } 243 } 244 245 // False branch 246 Node *iffalse = _gvn.transform( new IfFalseNode(iff) ); 247 set_control( iffalse ); 248 } 249 250 void Parse::jump_if_false_fork(IfNode *iff, int dest_bci_if_true, bool unc) { 251 // True branch, use existing map info 252 { PreserveJVMState pjvms(this); 253 Node *iffalse = _gvn.transform( new IfFalseNode (iff) ); 254 set_control( iffalse ); 255 if (unc) { 256 repush_if_args(); 257 uncommon_trap(Deoptimization::Reason_unstable_if, 258 Deoptimization::Action_reinterpret, 259 NULL, 260 "taken never"); 261 } else { 262 assert(dest_bci_if_true != never_reached, "inconsistent dest"); 263 merge_new_path(dest_bci_if_true); 264 } 265 } 266 267 // False branch 268 Node *iftrue = _gvn.transform( new IfTrueNode(iff) ); 269 set_control( iftrue ); 270 } 271 272 void Parse::jump_if_always_fork(int dest_bci, bool unc) { 273 // False branch, use existing map and control() 274 if (unc) { 275 repush_if_args(); 276 uncommon_trap(Deoptimization::Reason_unstable_if, 277 Deoptimization::Action_reinterpret, 278 NULL, 279 "taken never"); 280 } else { 281 assert(dest_bci != never_reached, "inconsistent dest"); 282 merge_new_path(dest_bci); 283 } 284 } 285 286 287 extern "C" { 288 static int jint_cmp(const void *i, const void *j) { 289 int a = *(jint *)i; 290 int b = *(jint *)j; 291 return a > b ? 1 : a < b ? -1 : 0; 292 } 293 } 294 295 296 class SwitchRange : public StackObj { 297 // a range of integers coupled with a bci destination 298 jint _lo; // inclusive lower limit 299 jint _hi; // inclusive upper limit 300 int _dest; 301 float _cnt; // how many times this range was hit according to profiling 302 303 public: 304 jint lo() const { return _lo; } 305 jint hi() const { return _hi; } 306 int dest() const { return _dest; } 307 bool is_singleton() const { return _lo == _hi; } 308 float cnt() const { return _cnt; } 309 310 void setRange(jint lo, jint hi, int dest, float cnt) { 311 assert(lo <= hi, "must be a non-empty range"); 312 _lo = lo, _hi = hi; _dest = dest; _cnt = cnt; 313 assert(_cnt >= 0, ""); 314 } 315 bool adjoinRange(jint lo, jint hi, int dest, float cnt, bool trim_ranges) { 316 assert(lo <= hi, "must be a non-empty range"); 317 if (lo == _hi+1) { 318 // see merge_ranges() comment below 319 if (trim_ranges) { 320 if (cnt == 0) { 321 if (_cnt != 0) { 322 return false; 323 } 324 if (dest != _dest) { 325 _dest = never_reached; 326 } 327 } else { 328 if (_cnt == 0) { 329 return false; 330 } 331 if (dest != _dest) { 332 return false; 333 } 334 } 335 } else { 336 if (dest != _dest) { 337 return false; 338 } 339 } 340 _hi = hi; 341 _cnt += cnt; 342 return true; 343 } 344 return false; 345 } 346 347 void set (jint value, int dest, float cnt) { 348 setRange(value, value, dest, cnt); 349 } 350 bool adjoin(jint value, int dest, float cnt, bool trim_ranges) { 351 return adjoinRange(value, value, dest, cnt, trim_ranges); 352 } 353 bool adjoin(SwitchRange& other) { 354 return adjoinRange(other._lo, other._hi, other._dest, other._cnt, false); 355 } 356 357 void print() { 358 if (is_singleton()) 359 tty->print(" {%d}=>%d (cnt=%f)", lo(), dest(), cnt()); 360 else if (lo() == min_jint) 361 tty->print(" {..%d}=>%d (cnt=%f)", hi(), dest(), cnt()); 362 else if (hi() == max_jint) 363 tty->print(" {%d..}=>%d (cnt=%f)", lo(), dest(), cnt()); 364 else 365 tty->print(" {%d..%d}=>%d (cnt=%f)", lo(), hi(), dest(), cnt()); 366 } 367 }; 368 369 // We try to minimize the number of ranges and the size of the taken 370 // ones using profiling data. When ranges are created, 371 // SwitchRange::adjoinRange() only allows 2 adjoining ranges to merge 372 // if both were never hit or both were hit to build longer unreached 373 // ranges. Here, we now merge adjoining ranges with the same 374 // destination and finally set destination of unreached ranges to the 375 // special value never_reached because it can help minimize the number 376 // of tests that are necessary. 377 // 378 // For instance: 379 // [0, 1] to target1 sometimes taken 380 // [1, 2] to target1 never taken 381 // [2, 3] to target2 never taken 382 // would lead to: 383 // [0, 1] to target1 sometimes taken 384 // [1, 3] never taken 385 // 386 // (first 2 ranges to target1 are not merged) 387 static void merge_ranges(SwitchRange* ranges, int& rp) { 388 if (rp == 0) { 389 return; 390 } 391 int shift = 0; 392 for (int j = 0; j < rp; j++) { 393 SwitchRange& r1 = ranges[j-shift]; 394 SwitchRange& r2 = ranges[j+1]; 395 if (r1.adjoin(r2)) { 396 shift++; 397 } else if (shift > 0) { 398 ranges[j+1-shift] = r2; 399 } 400 } 401 rp -= shift; 402 for (int j = 0; j <= rp; j++) { 403 SwitchRange& r = ranges[j]; 404 if (r.cnt() == 0 && r.dest() != never_reached) { 405 r.setRange(r.lo(), r.hi(), never_reached, r.cnt()); 406 } 407 } 408 } 409 410 //-------------------------------do_tableswitch-------------------------------- 411 void Parse::do_tableswitch() { 412 // Get information about tableswitch 413 int default_dest = iter().get_dest_table(0); 414 jint lo_index = iter().get_int_table(1); 415 jint hi_index = iter().get_int_table(2); 416 int len = hi_index - lo_index + 1; 417 418 if (len < 1) { 419 // If this is a backward branch, add safepoint 420 maybe_add_safepoint(default_dest); 421 pop(); // the effect of the instruction execution on the operand stack 422 merge(default_dest); 423 return; 424 } 425 426 ciMethodData* methodData = method()->method_data(); 427 ciMultiBranchData* profile = NULL; 428 if (methodData->is_mature() && UseSwitchProfiling) { 429 ciProfileData* data = methodData->bci_to_data(bci()); 430 if (data != NULL && data->is_MultiBranchData()) { 431 profile = (ciMultiBranchData*)data; 432 } 433 } 434 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if); 435 436 // generate decision tree, using trichotomy when possible 437 int rnum = len+2; 438 bool makes_backward_branch = false; 439 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum); 440 int rp = -1; 441 if (lo_index != min_jint) { 442 float cnt = 1.0F; 443 if (profile != NULL) { 444 cnt = (float)profile->default_count() / (hi_index != max_jint ? 2.0F : 1.0F); 445 } 446 ranges[++rp].setRange(min_jint, lo_index-1, default_dest, cnt); 447 } 448 for (int j = 0; j < len; j++) { 449 jint match_int = lo_index+j; 450 int dest = iter().get_dest_table(j+3); 451 makes_backward_branch |= (dest <= bci()); 452 float cnt = 1.0F; 453 if (profile != NULL) { 454 cnt = (float)profile->count_at(j); 455 } 456 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, cnt, trim_ranges)) { 457 ranges[++rp].set(match_int, dest, cnt); 458 } 459 } 460 jint highest = lo_index+(len-1); 461 assert(ranges[rp].hi() == highest, ""); 462 if (highest != max_jint) { 463 float cnt = 1.0F; 464 if (profile != NULL) { 465 cnt = (float)profile->default_count() / (lo_index != min_jint ? 2.0F : 1.0F); 466 } 467 if (!ranges[rp].adjoinRange(highest+1, max_jint, default_dest, cnt, trim_ranges)) { 468 ranges[++rp].setRange(highest+1, max_jint, default_dest, cnt); 469 } 470 } 471 assert(rp < len+2, "not too many ranges"); 472 473 if (trim_ranges) { 474 merge_ranges(ranges, rp); 475 } 476 477 // Safepoint in case if backward branch observed 478 if (makes_backward_branch) { 479 add_safepoint(); 480 } 481 482 Node* lookup = pop(); // lookup value 483 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]); 484 } 485 486 487 //------------------------------do_lookupswitch-------------------------------- 488 void Parse::do_lookupswitch() { 489 // Get information about lookupswitch 490 int default_dest = iter().get_dest_table(0); 491 jint len = iter().get_int_table(1); 492 493 if (len < 1) { // If this is a backward branch, add safepoint 494 maybe_add_safepoint(default_dest); 495 pop(); // the effect of the instruction execution on the operand stack 496 merge(default_dest); 497 return; 498 } 499 500 ciMethodData* methodData = method()->method_data(); 501 ciMultiBranchData* profile = NULL; 502 if (methodData->is_mature() && UseSwitchProfiling) { 503 ciProfileData* data = methodData->bci_to_data(bci()); 504 if (data != NULL && data->is_MultiBranchData()) { 505 profile = (ciMultiBranchData*)data; 506 } 507 } 508 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if); 509 510 // generate decision tree, using trichotomy when possible 511 jint* table = NEW_RESOURCE_ARRAY(jint, len*3); 512 { 513 for (int j = 0; j < len; j++) { 514 table[3*j+0] = iter().get_int_table(2+2*j); 515 table[3*j+1] = iter().get_dest_table(2+2*j+1); 516 // Handle overflow when converting from uint to jint 517 table[3*j+2] = (profile == NULL) ? 1 : (jint)MIN2<uint>((uint)max_jint, profile->count_at(j)); 518 } 519 qsort(table, len, 3*sizeof(table[0]), jint_cmp); 520 } 521 522 float default_cnt = 1.0F; 523 if (profile != NULL) { 524 juint defaults = max_juint - len; 525 default_cnt = (float)profile->default_count()/(float)defaults; 526 } 527 528 int rnum = len*2+1; 529 bool makes_backward_branch = false; 530 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum); 531 int rp = -1; 532 for (int j = 0; j < len; j++) { 533 jint match_int = table[3*j+0]; 534 jint dest = table[3*j+1]; 535 jint cnt = table[3*j+2]; 536 jint next_lo = rp < 0 ? min_jint : ranges[rp].hi()+1; 537 makes_backward_branch |= (dest <= bci()); 538 float c = default_cnt * ((float)match_int - (float)next_lo); 539 if (match_int != next_lo && (rp < 0 || !ranges[rp].adjoinRange(next_lo, match_int-1, default_dest, c, trim_ranges))) { 540 assert(default_dest != never_reached, "sentinel value for dead destinations"); 541 ranges[++rp].setRange(next_lo, match_int-1, default_dest, c); 542 } 543 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, (float)cnt, trim_ranges)) { 544 assert(dest != never_reached, "sentinel value for dead destinations"); 545 ranges[++rp].set(match_int, dest, (float)cnt); 546 } 547 } 548 jint highest = table[3*(len-1)]; 549 assert(ranges[rp].hi() == highest, ""); 550 if (highest != max_jint && 551 !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, default_cnt * ((float)max_jint - (float)highest), trim_ranges)) { 552 ranges[++rp].setRange(highest+1, max_jint, default_dest, default_cnt * ((float)max_jint - (float)highest)); 553 } 554 assert(rp < rnum, "not too many ranges"); 555 556 if (trim_ranges) { 557 merge_ranges(ranges, rp); 558 } 559 560 // Safepoint in case backward branch observed 561 if (makes_backward_branch) { 562 add_safepoint(); 563 } 564 565 Node *lookup = pop(); // lookup value 566 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]); 567 } 568 569 static float if_prob(float taken_cnt, float total_cnt) { 570 assert(taken_cnt <= total_cnt, ""); 571 if (total_cnt == 0) { 572 return PROB_FAIR; 573 } 574 float p = taken_cnt / total_cnt; 575 return clamp(p, PROB_MIN, PROB_MAX); 576 } 577 578 static float if_cnt(float cnt) { 579 if (cnt == 0) { 580 return COUNT_UNKNOWN; 581 } 582 return cnt; 583 } 584 585 static float sum_of_cnts(SwitchRange *lo, SwitchRange *hi) { 586 float total_cnt = 0; 587 for (SwitchRange* sr = lo; sr <= hi; sr++) { 588 total_cnt += sr->cnt(); 589 } 590 return total_cnt; 591 } 592 593 class SwitchRanges : public ResourceObj { 594 public: 595 SwitchRange* _lo; 596 SwitchRange* _hi; 597 SwitchRange* _mid; 598 float _cost; 599 600 enum { 601 Start, 602 LeftDone, 603 RightDone, 604 Done 605 } _state; 606 607 SwitchRanges(SwitchRange *lo, SwitchRange *hi) 608 : _lo(lo), _hi(hi), _mid(NULL), 609 _cost(0), _state(Start) { 610 } 611 612 SwitchRanges() 613 : _lo(NULL), _hi(NULL), _mid(NULL), 614 _cost(0), _state(Start) {} 615 }; 616 617 // Estimate cost of performing a binary search on lo..hi 618 static float compute_tree_cost(SwitchRange *lo, SwitchRange *hi, float total_cnt) { 619 GrowableArray<SwitchRanges> tree; 620 SwitchRanges root(lo, hi); 621 tree.push(root); 622 623 float cost = 0; 624 do { 625 SwitchRanges& r = *tree.adr_at(tree.length()-1); 626 if (r._hi != r._lo) { 627 if (r._mid == NULL) { 628 float r_cnt = sum_of_cnts(r._lo, r._hi); 629 630 if (r_cnt == 0) { 631 tree.pop(); 632 cost = 0; 633 continue; 634 } 635 636 SwitchRange* mid = NULL; 637 mid = r._lo; 638 for (float cnt = 0; ; ) { 639 assert(mid <= r._hi, "out of bounds"); 640 cnt += mid->cnt(); 641 if (cnt > r_cnt / 2) { 642 break; 643 } 644 mid++; 645 } 646 assert(mid <= r._hi, "out of bounds"); 647 r._mid = mid; 648 r._cost = r_cnt / total_cnt; 649 } 650 r._cost += cost; 651 if (r._state < SwitchRanges::LeftDone && r._mid > r._lo) { 652 cost = 0; 653 r._state = SwitchRanges::LeftDone; 654 tree.push(SwitchRanges(r._lo, r._mid-1)); 655 } else if (r._state < SwitchRanges::RightDone) { 656 cost = 0; 657 r._state = SwitchRanges::RightDone; 658 tree.push(SwitchRanges(r._mid == r._lo ? r._mid+1 : r._mid, r._hi)); 659 } else { 660 tree.pop(); 661 cost = r._cost; 662 } 663 } else { 664 tree.pop(); 665 cost = r._cost; 666 } 667 } while (tree.length() > 0); 668 669 670 return cost; 671 } 672 673 // It sometimes pays off to test most common ranges before the binary search 674 void Parse::linear_search_switch_ranges(Node* key_val, SwitchRange*& lo, SwitchRange*& hi) { 675 uint nr = hi - lo + 1; 676 float total_cnt = sum_of_cnts(lo, hi); 677 678 float min = compute_tree_cost(lo, hi, total_cnt); 679 float extra = 1; 680 float sub = 0; 681 682 SwitchRange* array1 = lo; 683 SwitchRange* array2 = NEW_RESOURCE_ARRAY(SwitchRange, nr); 684 685 SwitchRange* ranges = NULL; 686 687 while (nr >= 2) { 688 assert(lo == array1 || lo == array2, "one the 2 already allocated arrays"); 689 ranges = (lo == array1) ? array2 : array1; 690 691 // Find highest frequency range 692 SwitchRange* candidate = lo; 693 for (SwitchRange* sr = lo+1; sr <= hi; sr++) { 694 if (sr->cnt() > candidate->cnt()) { 695 candidate = sr; 696 } 697 } 698 SwitchRange most_freq = *candidate; 699 if (most_freq.cnt() == 0) { 700 break; 701 } 702 703 // Copy remaining ranges into another array 704 int shift = 0; 705 for (uint i = 0; i < nr; i++) { 706 SwitchRange* sr = &lo[i]; 707 if (sr != candidate) { 708 ranges[i-shift] = *sr; 709 } else { 710 shift++; 711 if (i > 0 && i < nr-1) { 712 SwitchRange prev = lo[i-1]; 713 prev.setRange(prev.lo(), sr->hi(), prev.dest(), prev.cnt()); 714 if (prev.adjoin(lo[i+1])) { 715 shift++; 716 i++; 717 } 718 ranges[i-shift] = prev; 719 } 720 } 721 } 722 nr -= shift; 723 724 // Evaluate cost of testing the most common range and performing a 725 // binary search on the other ranges 726 float cost = extra + compute_tree_cost(&ranges[0], &ranges[nr-1], total_cnt); 727 if (cost >= min) { 728 break; 729 } 730 // swap arrays 731 lo = &ranges[0]; 732 hi = &ranges[nr-1]; 733 734 // It pays off: emit the test for the most common range 735 assert(most_freq.cnt() > 0, "must be taken"); 736 Node* val = _gvn.transform(new SubINode(key_val, _gvn.intcon(most_freq.lo()))); 737 Node* cmp = _gvn.transform(new CmpUNode(val, _gvn.intcon(most_freq.hi() - most_freq.lo()))); 738 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::le)); 739 IfNode* iff = create_and_map_if(control(), tst, if_prob(most_freq.cnt(), total_cnt), if_cnt(most_freq.cnt())); 740 jump_if_true_fork(iff, most_freq.dest(), false); 741 742 sub += most_freq.cnt() / total_cnt; 743 extra += 1 - sub; 744 min = cost; 745 } 746 } 747 748 //----------------------------create_jump_tables------------------------------- 749 bool Parse::create_jump_tables(Node* key_val, SwitchRange* lo, SwitchRange* hi) { 750 // Are jumptables enabled 751 if (!UseJumpTables) return false; 752 753 // Are jumptables supported 754 if (!Matcher::has_match_rule(Op_Jump)) return false; 755 756 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if); 757 758 // Decide if a guard is needed to lop off big ranges at either (or 759 // both) end(s) of the input set. We'll call this the default target 760 // even though we can't be sure that it is the true "default". 761 762 bool needs_guard = false; 763 int default_dest; 764 int64_t total_outlier_size = 0; 765 int64_t hi_size = ((int64_t)hi->hi()) - ((int64_t)hi->lo()) + 1; 766 int64_t lo_size = ((int64_t)lo->hi()) - ((int64_t)lo->lo()) + 1; 767 768 if (lo->dest() == hi->dest()) { 769 total_outlier_size = hi_size + lo_size; 770 default_dest = lo->dest(); 771 } else if (lo_size > hi_size) { 772 total_outlier_size = lo_size; 773 default_dest = lo->dest(); 774 } else { 775 total_outlier_size = hi_size; 776 default_dest = hi->dest(); 777 } 778 779 float total = sum_of_cnts(lo, hi); 780 float cost = compute_tree_cost(lo, hi, total); 781 782 // If a guard test will eliminate very sparse end ranges, then 783 // it is worth the cost of an extra jump. 784 float trimmed_cnt = 0; 785 if (total_outlier_size > (MaxJumpTableSparseness * 4)) { 786 needs_guard = true; 787 if (default_dest == lo->dest()) { 788 trimmed_cnt += lo->cnt(); 789 lo++; 790 } 791 if (default_dest == hi->dest()) { 792 trimmed_cnt += hi->cnt(); 793 hi--; 794 } 795 } 796 797 // Find the total number of cases and ranges 798 int64_t num_cases = ((int64_t)hi->hi()) - ((int64_t)lo->lo()) + 1; 799 int num_range = hi - lo + 1; 800 801 // Don't create table if: too large, too small, or too sparse. 802 if (num_cases > MaxJumpTableSize) 803 return false; 804 if (UseSwitchProfiling) { 805 // MinJumpTableSize is set so with a well balanced binary tree, 806 // when the number of ranges is MinJumpTableSize, it's cheaper to 807 // go through a JumpNode that a tree of IfNodes. Average cost of a 808 // tree of IfNodes with MinJumpTableSize is 809 // log2f(MinJumpTableSize) comparisons. So if the cost computed 810 // from profile data is less than log2f(MinJumpTableSize) then 811 // going with the binary search is cheaper. 812 if (cost < log2f(MinJumpTableSize)) { 813 return false; 814 } 815 } else { 816 if (num_cases < MinJumpTableSize) 817 return false; 818 } 819 if (num_cases > (MaxJumpTableSparseness * num_range)) 820 return false; 821 822 // Normalize table lookups to zero 823 int lowval = lo->lo(); 824 key_val = _gvn.transform( new SubINode(key_val, _gvn.intcon(lowval)) ); 825 826 // Generate a guard to protect against input keyvals that aren't 827 // in the switch domain. 828 if (needs_guard) { 829 Node* size = _gvn.intcon(num_cases); 830 Node* cmp = _gvn.transform(new CmpUNode(key_val, size)); 831 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::ge)); 832 IfNode* iff = create_and_map_if(control(), tst, if_prob(trimmed_cnt, total), if_cnt(trimmed_cnt)); 833 jump_if_true_fork(iff, default_dest, trim_ranges && trimmed_cnt == 0); 834 835 total -= trimmed_cnt; 836 } 837 838 // Create an ideal node JumpTable that has projections 839 // of all possible ranges for a switch statement 840 // The key_val input must be converted to a pointer offset and scaled. 841 // Compare Parse::array_addressing above. 842 843 // Clean the 32-bit int into a real 64-bit offset. 844 // Otherwise, the jint value 0 might turn into an offset of 0x0800000000. 845 // Make I2L conversion control dependent to prevent it from 846 // floating above the range check during loop optimizations. 847 // Do not use a narrow int type here to prevent the data path from dying 848 // while the control path is not removed. This can happen if the type of key_val 849 // is later known to be out of bounds of [0, num_cases] and therefore a narrow cast 850 // would be replaced by TOP while C2 is not able to fold the corresponding range checks. 851 // Set _carry_dependency for the cast to avoid being removed by IGVN. 852 #ifdef _LP64 853 key_val = C->constrained_convI2L(&_gvn, key_val, TypeInt::INT, control(), true /* carry_dependency */); 854 #endif 855 856 // Shift the value by wordsize so we have an index into the table, rather 857 // than a switch value 858 Node *shiftWord = _gvn.MakeConX(wordSize); 859 key_val = _gvn.transform( new MulXNode( key_val, shiftWord)); 860 861 // Create the JumpNode 862 Arena* arena = C->comp_arena(); 863 float* probs = (float*)arena->Amalloc(sizeof(float)*num_cases); 864 int i = 0; 865 if (total == 0) { 866 for (SwitchRange* r = lo; r <= hi; r++) { 867 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) { 868 probs[i] = 1.0F / num_cases; 869 } 870 } 871 } else { 872 for (SwitchRange* r = lo; r <= hi; r++) { 873 float prob = r->cnt()/total; 874 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) { 875 probs[i] = prob / (r->hi() - r->lo() + 1); 876 } 877 } 878 } 879 880 ciMethodData* methodData = method()->method_data(); 881 ciMultiBranchData* profile = NULL; 882 if (methodData->is_mature()) { 883 ciProfileData* data = methodData->bci_to_data(bci()); 884 if (data != NULL && data->is_MultiBranchData()) { 885 profile = (ciMultiBranchData*)data; 886 } 887 } 888 889 Node* jtn = _gvn.transform(new JumpNode(control(), key_val, num_cases, probs, profile == NULL ? COUNT_UNKNOWN : total)); 890 891 // These are the switch destinations hanging off the jumpnode 892 i = 0; 893 for (SwitchRange* r = lo; r <= hi; r++) { 894 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) { 895 Node* input = _gvn.transform(new JumpProjNode(jtn, i, r->dest(), (int)(j - lowval))); 896 { 897 PreserveJVMState pjvms(this); 898 set_control(input); 899 jump_if_always_fork(r->dest(), trim_ranges && r->cnt() == 0); 900 } 901 } 902 } 903 assert(i == num_cases, "miscount of cases"); 904 stop_and_kill_map(); // no more uses for this JVMS 905 return true; 906 } 907 908 //----------------------------jump_switch_ranges------------------------------- 909 void Parse::jump_switch_ranges(Node* key_val, SwitchRange *lo, SwitchRange *hi, int switch_depth) { 910 Block* switch_block = block(); 911 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if); 912 913 if (switch_depth == 0) { 914 // Do special processing for the top-level call. 915 assert(lo->lo() == min_jint, "initial range must exhaust Type::INT"); 916 assert(hi->hi() == max_jint, "initial range must exhaust Type::INT"); 917 918 // Decrement pred-numbers for the unique set of nodes. 919 #ifdef ASSERT 920 if (!trim_ranges) { 921 // Ensure that the block's successors are a (duplicate-free) set. 922 int successors_counted = 0; // block occurrences in [hi..lo] 923 int unique_successors = switch_block->num_successors(); 924 for (int i = 0; i < unique_successors; i++) { 925 Block* target = switch_block->successor_at(i); 926 927 // Check that the set of successors is the same in both places. 928 int successors_found = 0; 929 for (SwitchRange* p = lo; p <= hi; p++) { 930 if (p->dest() == target->start()) successors_found++; 931 } 932 assert(successors_found > 0, "successor must be known"); 933 successors_counted += successors_found; 934 } 935 assert(successors_counted == (hi-lo)+1, "no unexpected successors"); 936 } 937 #endif 938 939 // Maybe prune the inputs, based on the type of key_val. 940 jint min_val = min_jint; 941 jint max_val = max_jint; 942 const TypeInt* ti = key_val->bottom_type()->isa_int(); 943 if (ti != NULL) { 944 min_val = ti->_lo; 945 max_val = ti->_hi; 946 assert(min_val <= max_val, "invalid int type"); 947 } 948 while (lo->hi() < min_val) { 949 lo++; 950 } 951 if (lo->lo() < min_val) { 952 lo->setRange(min_val, lo->hi(), lo->dest(), lo->cnt()); 953 } 954 while (hi->lo() > max_val) { 955 hi--; 956 } 957 if (hi->hi() > max_val) { 958 hi->setRange(hi->lo(), max_val, hi->dest(), hi->cnt()); 959 } 960 961 linear_search_switch_ranges(key_val, lo, hi); 962 } 963 964 #ifndef PRODUCT 965 if (switch_depth == 0) { 966 _max_switch_depth = 0; 967 _est_switch_depth = log2i_graceful((hi - lo + 1) - 1) + 1; 968 } 969 #endif 970 971 assert(lo <= hi, "must be a non-empty set of ranges"); 972 if (lo == hi) { 973 jump_if_always_fork(lo->dest(), trim_ranges && lo->cnt() == 0); 974 } else { 975 assert(lo->hi() == (lo+1)->lo()-1, "contiguous ranges"); 976 assert(hi->lo() == (hi-1)->hi()+1, "contiguous ranges"); 977 978 if (create_jump_tables(key_val, lo, hi)) return; 979 980 SwitchRange* mid = NULL; 981 float total_cnt = sum_of_cnts(lo, hi); 982 983 int nr = hi - lo + 1; 984 if (UseSwitchProfiling) { 985 // Don't keep the binary search tree balanced: pick up mid point 986 // that split frequencies in half. 987 float cnt = 0; 988 for (SwitchRange* sr = lo; sr <= hi; sr++) { 989 cnt += sr->cnt(); 990 if (cnt >= total_cnt / 2) { 991 mid = sr; 992 break; 993 } 994 } 995 } else { 996 mid = lo + nr/2; 997 998 // if there is an easy choice, pivot at a singleton: 999 if (nr > 3 && !mid->is_singleton() && (mid-1)->is_singleton()) mid--; 1000 1001 assert(lo < mid && mid <= hi, "good pivot choice"); 1002 assert(nr != 2 || mid == hi, "should pick higher of 2"); 1003 assert(nr != 3 || mid == hi-1, "should pick middle of 3"); 1004 } 1005 1006 1007 Node *test_val = _gvn.intcon(mid == lo ? mid->hi() : mid->lo()); 1008 1009 if (mid->is_singleton()) { 1010 IfNode *iff_ne = jump_if_fork_int(key_val, test_val, BoolTest::ne, 1-if_prob(mid->cnt(), total_cnt), if_cnt(mid->cnt())); 1011 jump_if_false_fork(iff_ne, mid->dest(), trim_ranges && mid->cnt() == 0); 1012 1013 // Special Case: If there are exactly three ranges, and the high 1014 // and low range each go to the same place, omit the "gt" test, 1015 // since it will not discriminate anything. 1016 bool eq_test_only = (hi == lo+2 && hi->dest() == lo->dest() && mid == hi-1) || mid == lo; 1017 1018 // if there is a higher range, test for it and process it: 1019 if (mid < hi && !eq_test_only) { 1020 // two comparisons of same values--should enable 1 test for 2 branches 1021 // Use BoolTest::lt instead of BoolTest::gt 1022 float cnt = sum_of_cnts(lo, mid-1); 1023 IfNode *iff_lt = jump_if_fork_int(key_val, test_val, BoolTest::lt, if_prob(cnt, total_cnt), if_cnt(cnt)); 1024 Node *iftrue = _gvn.transform( new IfTrueNode(iff_lt) ); 1025 Node *iffalse = _gvn.transform( new IfFalseNode(iff_lt) ); 1026 { PreserveJVMState pjvms(this); 1027 set_control(iffalse); 1028 jump_switch_ranges(key_val, mid+1, hi, switch_depth+1); 1029 } 1030 set_control(iftrue); 1031 } 1032 1033 } else { 1034 // mid is a range, not a singleton, so treat mid..hi as a unit 1035 float cnt = sum_of_cnts(mid == lo ? mid+1 : mid, hi); 1036 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)); 1037 1038 // if there is a higher range, test for it and process it: 1039 if (mid == hi) { 1040 jump_if_true_fork(iff_ge, mid->dest(), trim_ranges && cnt == 0); 1041 } else { 1042 Node *iftrue = _gvn.transform( new IfTrueNode(iff_ge) ); 1043 Node *iffalse = _gvn.transform( new IfFalseNode(iff_ge) ); 1044 { PreserveJVMState pjvms(this); 1045 set_control(iftrue); 1046 jump_switch_ranges(key_val, mid == lo ? mid+1 : mid, hi, switch_depth+1); 1047 } 1048 set_control(iffalse); 1049 } 1050 } 1051 1052 // in any case, process the lower range 1053 if (mid == lo) { 1054 if (mid->is_singleton()) { 1055 jump_switch_ranges(key_val, lo+1, hi, switch_depth+1); 1056 } else { 1057 jump_if_always_fork(lo->dest(), trim_ranges && lo->cnt() == 0); 1058 } 1059 } else { 1060 jump_switch_ranges(key_val, lo, mid-1, switch_depth+1); 1061 } 1062 } 1063 1064 // Decrease pred_count for each successor after all is done. 1065 if (switch_depth == 0) { 1066 int unique_successors = switch_block->num_successors(); 1067 for (int i = 0; i < unique_successors; i++) { 1068 Block* target = switch_block->successor_at(i); 1069 // Throw away the pre-allocated path for each unique successor. 1070 target->next_path_num(); 1071 } 1072 } 1073 1074 #ifndef PRODUCT 1075 _max_switch_depth = MAX2(switch_depth, _max_switch_depth); 1076 if (TraceOptoParse && Verbose && WizardMode && switch_depth == 0) { 1077 SwitchRange* r; 1078 int nsing = 0; 1079 for( r = lo; r <= hi; r++ ) { 1080 if( r->is_singleton() ) nsing++; 1081 } 1082 tty->print(">>> "); 1083 _method->print_short_name(); 1084 tty->print_cr(" switch decision tree"); 1085 tty->print_cr(" %d ranges (%d singletons), max_depth=%d, est_depth=%d", 1086 (int) (hi-lo+1), nsing, _max_switch_depth, _est_switch_depth); 1087 if (_max_switch_depth > _est_switch_depth) { 1088 tty->print_cr("******** BAD SWITCH DEPTH ********"); 1089 } 1090 tty->print(" "); 1091 for( r = lo; r <= hi; r++ ) { 1092 r->print(); 1093 } 1094 tty->cr(); 1095 } 1096 #endif 1097 } 1098 1099 void Parse::modf() { 1100 Node *f2 = pop(); 1101 Node *f1 = pop(); 1102 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::modf_Type(), 1103 CAST_FROM_FN_PTR(address, SharedRuntime::frem), 1104 "frem", NULL, //no memory effects 1105 f1, f2); 1106 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0)); 1107 1108 push(res); 1109 } 1110 1111 void Parse::modd() { 1112 Node *d2 = pop_pair(); 1113 Node *d1 = pop_pair(); 1114 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::Math_DD_D_Type(), 1115 CAST_FROM_FN_PTR(address, SharedRuntime::drem), 1116 "drem", NULL, //no memory effects 1117 d1, top(), d2, top()); 1118 Node* res_d = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0)); 1119 1120 #ifdef ASSERT 1121 Node* res_top = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 1)); 1122 assert(res_top == top(), "second value must be top"); 1123 #endif 1124 1125 push_pair(res_d); 1126 } 1127 1128 void Parse::l2f() { 1129 Node* f2 = pop(); 1130 Node* f1 = pop(); 1131 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::l2f_Type(), 1132 CAST_FROM_FN_PTR(address, SharedRuntime::l2f), 1133 "l2f", NULL, //no memory effects 1134 f1, f2); 1135 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0)); 1136 1137 push(res); 1138 } 1139 1140 // Handle jsr and jsr_w bytecode 1141 void Parse::do_jsr() { 1142 assert(bc() == Bytecodes::_jsr || bc() == Bytecodes::_jsr_w, "wrong bytecode"); 1143 1144 // Store information about current state, tagged with new _jsr_bci 1145 int return_bci = iter().next_bci(); 1146 int jsr_bci = (bc() == Bytecodes::_jsr) ? iter().get_dest() : iter().get_far_dest(); 1147 1148 // The way we do things now, there is only one successor block 1149 // for the jsr, because the target code is cloned by ciTypeFlow. 1150 Block* target = successor_for_bci(jsr_bci); 1151 1152 // What got pushed? 1153 const Type* ret_addr = target->peek(); 1154 assert(ret_addr->singleton(), "must be a constant (cloned jsr body)"); 1155 1156 // Effect on jsr on stack 1157 push(_gvn.makecon(ret_addr)); 1158 1159 // Flow to the jsr. 1160 merge(jsr_bci); 1161 } 1162 1163 // Handle ret bytecode 1164 void Parse::do_ret() { 1165 // Find to whom we return. 1166 assert(block()->num_successors() == 1, "a ret can only go one place now"); 1167 Block* target = block()->successor_at(0); 1168 assert(!target->is_ready(), "our arrival must be expected"); 1169 int pnum = target->next_path_num(); 1170 merge_common(target, pnum); 1171 } 1172 1173 static bool has_injected_profile(BoolTest::mask btest, Node* test, int& taken, int& not_taken) { 1174 if (btest != BoolTest::eq && btest != BoolTest::ne) { 1175 // Only ::eq and ::ne are supported for profile injection. 1176 return false; 1177 } 1178 if (test->is_Cmp() && 1179 test->in(1)->Opcode() == Op_ProfileBoolean) { 1180 ProfileBooleanNode* profile = (ProfileBooleanNode*)test->in(1); 1181 int false_cnt = profile->false_count(); 1182 int true_cnt = profile->true_count(); 1183 1184 // Counts matching depends on the actual test operation (::eq or ::ne). 1185 // No need to scale the counts because profile injection was designed 1186 // to feed exact counts into VM. 1187 taken = (btest == BoolTest::eq) ? false_cnt : true_cnt; 1188 not_taken = (btest == BoolTest::eq) ? true_cnt : false_cnt; 1189 1190 profile->consume(); 1191 return true; 1192 } 1193 return false; 1194 } 1195 //--------------------------dynamic_branch_prediction-------------------------- 1196 // Try to gather dynamic branch prediction behavior. Return a probability 1197 // of the branch being taken and set the "cnt" field. Returns a -1.0 1198 // if we need to use static prediction for some reason. 1199 float Parse::dynamic_branch_prediction(float &cnt, BoolTest::mask btest, Node* test) { 1200 ResourceMark rm; 1201 1202 cnt = COUNT_UNKNOWN; 1203 1204 int taken = 0; 1205 int not_taken = 0; 1206 1207 bool use_mdo = !has_injected_profile(btest, test, taken, not_taken); 1208 1209 if (use_mdo) { 1210 // Use MethodData information if it is available 1211 // FIXME: free the ProfileData structure 1212 ciMethodData* methodData = method()->method_data(); 1213 if (!methodData->is_mature()) return PROB_UNKNOWN; 1214 ciProfileData* data = methodData->bci_to_data(bci()); 1215 if (data == NULL) { 1216 return PROB_UNKNOWN; 1217 } 1218 if (!data->is_JumpData()) return PROB_UNKNOWN; 1219 1220 // get taken and not taken values 1221 taken = data->as_JumpData()->taken(); 1222 not_taken = 0; 1223 if (data->is_BranchData()) { 1224 not_taken = data->as_BranchData()->not_taken(); 1225 } 1226 1227 // scale the counts to be commensurate with invocation counts: 1228 taken = method()->scale_count(taken); 1229 not_taken = method()->scale_count(not_taken); 1230 } 1231 1232 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful. 1233 // We also check that individual counters are positive first, otherwise the sum can become positive. 1234 if (taken < 0 || not_taken < 0 || taken + not_taken < 40) { 1235 if (C->log() != NULL) { 1236 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d'", iter().get_dest(), taken, not_taken); 1237 } 1238 return PROB_UNKNOWN; 1239 } 1240 1241 // Compute frequency that we arrive here 1242 float sum = taken + not_taken; 1243 // Adjust, if this block is a cloned private block but the 1244 // Jump counts are shared. Taken the private counts for 1245 // just this path instead of the shared counts. 1246 if( block()->count() > 0 ) 1247 sum = block()->count(); 1248 cnt = sum / FreqCountInvocations; 1249 1250 // Pin probability to sane limits 1251 float prob; 1252 if( !taken ) 1253 prob = (0+PROB_MIN) / 2; 1254 else if( !not_taken ) 1255 prob = (1+PROB_MAX) / 2; 1256 else { // Compute probability of true path 1257 prob = (float)taken / (float)(taken + not_taken); 1258 if (prob > PROB_MAX) prob = PROB_MAX; 1259 if (prob < PROB_MIN) prob = PROB_MIN; 1260 } 1261 1262 assert((cnt > 0.0f) && (prob > 0.0f), 1263 "Bad frequency assignment in if"); 1264 1265 if (C->log() != NULL) { 1266 const char* prob_str = NULL; 1267 if (prob >= PROB_MAX) prob_str = (prob == PROB_MAX) ? "max" : "always"; 1268 if (prob <= PROB_MIN) prob_str = (prob == PROB_MIN) ? "min" : "never"; 1269 char prob_str_buf[30]; 1270 if (prob_str == NULL) { 1271 jio_snprintf(prob_str_buf, sizeof(prob_str_buf), "%20.2f", prob); 1272 prob_str = prob_str_buf; 1273 } 1274 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d' cnt='%f' prob='%s'", 1275 iter().get_dest(), taken, not_taken, cnt, prob_str); 1276 } 1277 return prob; 1278 } 1279 1280 //-----------------------------branch_prediction------------------------------- 1281 float Parse::branch_prediction(float& cnt, 1282 BoolTest::mask btest, 1283 int target_bci, 1284 Node* test) { 1285 float prob = dynamic_branch_prediction(cnt, btest, test); 1286 // If prob is unknown, switch to static prediction 1287 if (prob != PROB_UNKNOWN) return prob; 1288 1289 prob = PROB_FAIR; // Set default value 1290 if (btest == BoolTest::eq) // Exactly equal test? 1291 prob = PROB_STATIC_INFREQUENT; // Assume its relatively infrequent 1292 else if (btest == BoolTest::ne) 1293 prob = PROB_STATIC_FREQUENT; // Assume its relatively frequent 1294 1295 // If this is a conditional test guarding a backwards branch, 1296 // assume its a loop-back edge. Make it a likely taken branch. 1297 if (target_bci < bci()) { 1298 if (is_osr_parse()) { // Could be a hot OSR'd loop; force deopt 1299 // Since it's an OSR, we probably have profile data, but since 1300 // branch_prediction returned PROB_UNKNOWN, the counts are too small. 1301 // Let's make a special check here for completely zero counts. 1302 ciMethodData* methodData = method()->method_data(); 1303 if (!methodData->is_empty()) { 1304 ciProfileData* data = methodData->bci_to_data(bci()); 1305 // Only stop for truly zero counts, which mean an unknown part 1306 // of the OSR-ed method, and we want to deopt to gather more stats. 1307 // If you have ANY counts, then this loop is simply 'cold' relative 1308 // to the OSR loop. 1309 if (data == NULL || 1310 (data->as_BranchData()->taken() + data->as_BranchData()->not_taken() == 0)) { 1311 // This is the only way to return PROB_UNKNOWN: 1312 return PROB_UNKNOWN; 1313 } 1314 } 1315 } 1316 prob = PROB_STATIC_FREQUENT; // Likely to take backwards branch 1317 } 1318 1319 assert(prob != PROB_UNKNOWN, "must have some guess at this point"); 1320 return prob; 1321 } 1322 1323 // The magic constants are chosen so as to match the output of 1324 // branch_prediction() when the profile reports a zero taken count. 1325 // It is important to distinguish zero counts unambiguously, because 1326 // some branches (e.g., _213_javac.Assembler.eliminate) validly produce 1327 // very small but nonzero probabilities, which if confused with zero 1328 // counts would keep the program recompiling indefinitely. 1329 bool Parse::seems_never_taken(float prob) const { 1330 return prob < PROB_MIN; 1331 } 1332 1333 // True if the comparison seems to be the kind that will not change its 1334 // statistics from true to false. See comments in adjust_map_after_if. 1335 // This question is only asked along paths which are already 1336 // classified as untaken (by seems_never_taken), so really, 1337 // if a path is never taken, its controlling comparison is 1338 // already acting in a stable fashion. If the comparison 1339 // seems stable, we will put an expensive uncommon trap 1340 // on the untaken path. 1341 bool Parse::seems_stable_comparison() const { 1342 if (C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if)) { 1343 return false; 1344 } 1345 return true; 1346 } 1347 1348 //-------------------------------repush_if_args-------------------------------- 1349 // Push arguments of an "if" bytecode back onto the stack by adjusting _sp. 1350 inline int Parse::repush_if_args() { 1351 if (PrintOpto && WizardMode) { 1352 tty->print("defending against excessive implicit null exceptions on %s @%d in ", 1353 Bytecodes::name(iter().cur_bc()), iter().cur_bci()); 1354 method()->print_name(); tty->cr(); 1355 } 1356 int bc_depth = - Bytecodes::depth(iter().cur_bc()); 1357 assert(bc_depth == 1 || bc_depth == 2, "only two kinds of branches"); 1358 DEBUG_ONLY(sync_jvms()); // argument(n) requires a synced jvms 1359 assert(argument(0) != NULL, "must exist"); 1360 assert(bc_depth == 1 || argument(1) != NULL, "two must exist"); 1361 inc_sp(bc_depth); 1362 return bc_depth; 1363 } 1364 1365 //----------------------------------do_ifnull---------------------------------- 1366 void Parse::do_ifnull(BoolTest::mask btest, Node *c) { 1367 int target_bci = iter().get_dest(); 1368 1369 Block* branch_block = successor_for_bci(target_bci); 1370 Block* next_block = successor_for_bci(iter().next_bci()); 1371 1372 float cnt; 1373 float prob = branch_prediction(cnt, btest, target_bci, c); 1374 if (prob == PROB_UNKNOWN) { 1375 // (An earlier version of do_ifnull omitted this trap for OSR methods.) 1376 if (PrintOpto && Verbose) { 1377 tty->print_cr("Never-taken edge stops compilation at bci %d", bci()); 1378 } 1379 repush_if_args(); // to gather stats on loop 1380 uncommon_trap(Deoptimization::Reason_unreached, 1381 Deoptimization::Action_reinterpret, 1382 NULL, "cold"); 1383 if (C->eliminate_boxing()) { 1384 // Mark the successor blocks as parsed 1385 branch_block->next_path_num(); 1386 next_block->next_path_num(); 1387 } 1388 return; 1389 } 1390 1391 NOT_PRODUCT(explicit_null_checks_inserted++); 1392 1393 // Generate real control flow 1394 Node *tst = _gvn.transform( new BoolNode( c, btest ) ); 1395 1396 // Sanity check the probability value 1397 assert(prob > 0.0f,"Bad probability in Parser"); 1398 // Need xform to put node in hash table 1399 IfNode *iff = create_and_xform_if( control(), tst, prob, cnt ); 1400 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser"); 1401 // True branch 1402 { PreserveJVMState pjvms(this); 1403 Node* iftrue = _gvn.transform( new IfTrueNode (iff) ); 1404 set_control(iftrue); 1405 1406 if (stopped()) { // Path is dead? 1407 NOT_PRODUCT(explicit_null_checks_elided++); 1408 if (C->eliminate_boxing()) { 1409 // Mark the successor block as parsed 1410 branch_block->next_path_num(); 1411 } 1412 } else { // Path is live. 1413 adjust_map_after_if(btest, c, prob, branch_block); 1414 if (!stopped()) { 1415 merge(target_bci); 1416 } 1417 } 1418 } 1419 1420 // False branch 1421 Node* iffalse = _gvn.transform( new IfFalseNode(iff) ); 1422 set_control(iffalse); 1423 1424 if (stopped()) { // Path is dead? 1425 NOT_PRODUCT(explicit_null_checks_elided++); 1426 if (C->eliminate_boxing()) { 1427 // Mark the successor block as parsed 1428 next_block->next_path_num(); 1429 } 1430 } else { // Path is live. 1431 adjust_map_after_if(BoolTest(btest).negate(), c, 1.0-prob, next_block); 1432 } 1433 } 1434 1435 //------------------------------------do_if------------------------------------ 1436 void Parse::do_if(BoolTest::mask btest, Node* c) { 1437 int target_bci = iter().get_dest(); 1438 1439 Block* branch_block = successor_for_bci(target_bci); 1440 Block* next_block = successor_for_bci(iter().next_bci()); 1441 1442 float cnt; 1443 float prob = branch_prediction(cnt, btest, target_bci, c); 1444 float untaken_prob = 1.0 - prob; 1445 1446 if (prob == PROB_UNKNOWN) { 1447 if (PrintOpto && Verbose) { 1448 tty->print_cr("Never-taken edge stops compilation at bci %d", bci()); 1449 } 1450 repush_if_args(); // to gather stats on loop 1451 uncommon_trap(Deoptimization::Reason_unreached, 1452 Deoptimization::Action_reinterpret, 1453 NULL, "cold"); 1454 if (C->eliminate_boxing()) { 1455 // Mark the successor blocks as parsed 1456 branch_block->next_path_num(); 1457 next_block->next_path_num(); 1458 } 1459 return; 1460 } 1461 1462 // Sanity check the probability value 1463 assert(0.0f < prob && prob < 1.0f,"Bad probability in Parser"); 1464 1465 bool taken_if_true = true; 1466 // Convert BoolTest to canonical form: 1467 if (!BoolTest(btest).is_canonical()) { 1468 btest = BoolTest(btest).negate(); 1469 taken_if_true = false; 1470 // prob is NOT updated here; it remains the probability of the taken 1471 // path (as opposed to the prob of the path guarded by an 'IfTrueNode'). 1472 } 1473 assert(btest != BoolTest::eq, "!= is the only canonical exact test"); 1474 1475 Node* tst0 = new BoolNode(c, btest); 1476 Node* tst = _gvn.transform(tst0); 1477 BoolTest::mask taken_btest = BoolTest::illegal; 1478 BoolTest::mask untaken_btest = BoolTest::illegal; 1479 1480 if (tst->is_Bool()) { 1481 // Refresh c from the transformed bool node, since it may be 1482 // simpler than the original c. Also re-canonicalize btest. 1483 // This wins when (Bool ne (Conv2B p) 0) => (Bool ne (CmpP p NULL)). 1484 // That can arise from statements like: if (x instanceof C) ... 1485 if (tst != tst0) { 1486 // Canonicalize one more time since transform can change it. 1487 btest = tst->as_Bool()->_test._test; 1488 if (!BoolTest(btest).is_canonical()) { 1489 // Reverse edges one more time... 1490 tst = _gvn.transform( tst->as_Bool()->negate(&_gvn) ); 1491 btest = tst->as_Bool()->_test._test; 1492 assert(BoolTest(btest).is_canonical(), "sanity"); 1493 taken_if_true = !taken_if_true; 1494 } 1495 c = tst->in(1); 1496 } 1497 BoolTest::mask neg_btest = BoolTest(btest).negate(); 1498 taken_btest = taken_if_true ? btest : neg_btest; 1499 untaken_btest = taken_if_true ? neg_btest : btest; 1500 } 1501 1502 // Generate real control flow 1503 float true_prob = (taken_if_true ? prob : untaken_prob); 1504 IfNode* iff = create_and_map_if(control(), tst, true_prob, cnt); 1505 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser"); 1506 Node* taken_branch = new IfTrueNode(iff); 1507 Node* untaken_branch = new IfFalseNode(iff); 1508 if (!taken_if_true) { // Finish conversion to canonical form 1509 Node* tmp = taken_branch; 1510 taken_branch = untaken_branch; 1511 untaken_branch = tmp; 1512 } 1513 1514 // Branch is taken: 1515 { PreserveJVMState pjvms(this); 1516 taken_branch = _gvn.transform(taken_branch); 1517 set_control(taken_branch); 1518 1519 if (stopped()) { 1520 if (C->eliminate_boxing()) { 1521 // Mark the successor block as parsed 1522 branch_block->next_path_num(); 1523 } 1524 } else { 1525 adjust_map_after_if(taken_btest, c, prob, branch_block); 1526 if (!stopped()) { 1527 merge(target_bci); 1528 } 1529 } 1530 } 1531 1532 untaken_branch = _gvn.transform(untaken_branch); 1533 set_control(untaken_branch); 1534 1535 // Branch not taken. 1536 if (stopped()) { 1537 if (C->eliminate_boxing()) { 1538 // Mark the successor block as parsed 1539 next_block->next_path_num(); 1540 } 1541 } else { 1542 adjust_map_after_if(untaken_btest, c, untaken_prob, next_block); 1543 } 1544 } 1545 1546 bool Parse::path_is_suitable_for_uncommon_trap(float prob) const { 1547 // Don't want to speculate on uncommon traps when running with -Xcomp 1548 if (!UseInterpreter) { 1549 return false; 1550 } 1551 return (seems_never_taken(prob) && seems_stable_comparison()); 1552 } 1553 1554 void Parse::maybe_add_predicate_after_if(Block* path) { 1555 if (path->is_SEL_head() && path->preds_parsed() == 0) { 1556 // Add predicates at bci of if dominating the loop so traps can be 1557 // recorded on the if's profile data 1558 int bc_depth = repush_if_args(); 1559 add_empty_predicates(); 1560 dec_sp(bc_depth); 1561 path->set_has_predicates(); 1562 } 1563 } 1564 1565 1566 //----------------------------adjust_map_after_if------------------------------ 1567 // Adjust the JVM state to reflect the result of taking this path. 1568 // Basically, it means inspecting the CmpNode controlling this 1569 // branch, seeing how it constrains a tested value, and then 1570 // deciding if it's worth our while to encode this constraint 1571 // as graph nodes in the current abstract interpretation map. 1572 void Parse::adjust_map_after_if(BoolTest::mask btest, Node* c, float prob, Block* path) { 1573 if (!c->is_Cmp()) { 1574 maybe_add_predicate_after_if(path); 1575 return; 1576 } 1577 1578 if (stopped() || btest == BoolTest::illegal) { 1579 return; // nothing to do 1580 } 1581 1582 bool is_fallthrough = (path == successor_for_bci(iter().next_bci())); 1583 1584 if (path_is_suitable_for_uncommon_trap(prob)) { 1585 repush_if_args(); 1586 Node* call = uncommon_trap(Deoptimization::Reason_unstable_if, 1587 Deoptimization::Action_reinterpret, 1588 NULL, 1589 (is_fallthrough ? "taken always" : "taken never")); 1590 1591 if (call != nullptr) { 1592 C->record_unstable_if_trap(new UnstableIfTrap(call->as_CallStaticJava(), path)); 1593 } 1594 return; 1595 } 1596 1597 Node* val = c->in(1); 1598 Node* con = c->in(2); 1599 const Type* tcon = _gvn.type(con); 1600 const Type* tval = _gvn.type(val); 1601 bool have_con = tcon->singleton(); 1602 if (tval->singleton()) { 1603 if (!have_con) { 1604 // Swap, so constant is in con. 1605 con = val; 1606 tcon = tval; 1607 val = c->in(2); 1608 tval = _gvn.type(val); 1609 btest = BoolTest(btest).commute(); 1610 have_con = true; 1611 } else { 1612 // Do we have two constants? Then leave well enough alone. 1613 have_con = false; 1614 } 1615 } 1616 if (!have_con) { // remaining adjustments need a con 1617 maybe_add_predicate_after_if(path); 1618 return; 1619 } 1620 1621 sharpen_type_after_if(btest, con, tcon, val, tval); 1622 maybe_add_predicate_after_if(path); 1623 } 1624 1625 1626 static Node* extract_obj_from_klass_load(PhaseGVN* gvn, Node* n) { 1627 Node* ldk; 1628 if (n->is_DecodeNKlass()) { 1629 if (n->in(1)->Opcode() != Op_LoadNKlass) { 1630 return NULL; 1631 } else { 1632 ldk = n->in(1); 1633 } 1634 } else if (n->Opcode() != Op_LoadKlass) { 1635 return NULL; 1636 } else { 1637 ldk = n; 1638 } 1639 assert(ldk != NULL && ldk->is_Load(), "should have found a LoadKlass or LoadNKlass node"); 1640 1641 Node* adr = ldk->in(MemNode::Address); 1642 intptr_t off = 0; 1643 Node* obj = AddPNode::Ideal_base_and_offset(adr, gvn, off); 1644 if (obj == NULL || off != oopDesc::klass_offset_in_bytes()) // loading oopDesc::_klass? 1645 return NULL; 1646 const TypePtr* tp = gvn->type(obj)->is_ptr(); 1647 if (tp == NULL || !(tp->isa_instptr() || tp->isa_aryptr())) // is obj a Java object ptr? 1648 return NULL; 1649 1650 return obj; 1651 } 1652 1653 void Parse::sharpen_type_after_if(BoolTest::mask btest, 1654 Node* con, const Type* tcon, 1655 Node* val, const Type* tval) { 1656 // Look for opportunities to sharpen the type of a node 1657 // whose klass is compared with a constant klass. 1658 if (btest == BoolTest::eq && tcon->isa_klassptr()) { 1659 Node* obj = extract_obj_from_klass_load(&_gvn, val); 1660 const TypeOopPtr* con_type = tcon->isa_klassptr()->as_instance_type(); 1661 if (obj != NULL && (con_type->isa_instptr() || con_type->isa_aryptr())) { 1662 // Found: 1663 // Bool(CmpP(LoadKlass(obj._klass), ConP(Foo.klass)), [eq]) 1664 // or the narrowOop equivalent. 1665 const Type* obj_type = _gvn.type(obj); 1666 const TypeOopPtr* tboth = obj_type->join_speculative(con_type)->isa_oopptr(); 1667 if (tboth != NULL && tboth->klass_is_exact() && tboth != obj_type && 1668 tboth->higher_equal(obj_type)) { 1669 // obj has to be of the exact type Foo if the CmpP succeeds. 1670 int obj_in_map = map()->find_edge(obj); 1671 JVMState* jvms = this->jvms(); 1672 if (obj_in_map >= 0 && 1673 (jvms->is_loc(obj_in_map) || jvms->is_stk(obj_in_map))) { 1674 TypeNode* ccast = new CheckCastPPNode(control(), obj, tboth); 1675 const Type* tcc = ccast->as_Type()->type(); 1676 assert(tcc != obj_type && tcc->higher_equal(obj_type), "must improve"); 1677 // Delay transform() call to allow recovery of pre-cast value 1678 // at the control merge. 1679 _gvn.set_type_bottom(ccast); 1680 record_for_igvn(ccast); 1681 // Here's the payoff. 1682 replace_in_map(obj, ccast); 1683 } 1684 } 1685 } 1686 } 1687 1688 int val_in_map = map()->find_edge(val); 1689 if (val_in_map < 0) return; // replace_in_map would be useless 1690 { 1691 JVMState* jvms = this->jvms(); 1692 if (!(jvms->is_loc(val_in_map) || 1693 jvms->is_stk(val_in_map))) 1694 return; // again, it would be useless 1695 } 1696 1697 // Check for a comparison to a constant, and "know" that the compared 1698 // value is constrained on this path. 1699 assert(tcon->singleton(), ""); 1700 ConstraintCastNode* ccast = NULL; 1701 Node* cast = NULL; 1702 1703 switch (btest) { 1704 case BoolTest::eq: // Constant test? 1705 { 1706 const Type* tboth = tcon->join_speculative(tval); 1707 if (tboth == tval) break; // Nothing to gain. 1708 if (tcon->isa_int()) { 1709 ccast = new CastIINode(val, tboth); 1710 } else if (tcon == TypePtr::NULL_PTR) { 1711 // Cast to null, but keep the pointer identity temporarily live. 1712 ccast = new CastPPNode(val, tboth); 1713 } else { 1714 const TypeF* tf = tcon->isa_float_constant(); 1715 const TypeD* td = tcon->isa_double_constant(); 1716 // Exclude tests vs float/double 0 as these could be 1717 // either +0 or -0. Just because you are equal to +0 1718 // doesn't mean you ARE +0! 1719 // Note, following code also replaces Long and Oop values. 1720 if ((!tf || tf->_f != 0.0) && 1721 (!td || td->_d != 0.0)) 1722 cast = con; // Replace non-constant val by con. 1723 } 1724 } 1725 break; 1726 1727 case BoolTest::ne: 1728 if (tcon == TypePtr::NULL_PTR) { 1729 cast = cast_not_null(val, false); 1730 } 1731 break; 1732 1733 default: 1734 // (At this point we could record int range types with CastII.) 1735 break; 1736 } 1737 1738 if (ccast != NULL) { 1739 const Type* tcc = ccast->as_Type()->type(); 1740 assert(tcc != tval && tcc->higher_equal(tval), "must improve"); 1741 // Delay transform() call to allow recovery of pre-cast value 1742 // at the control merge. 1743 ccast->set_req(0, control()); 1744 _gvn.set_type_bottom(ccast); 1745 record_for_igvn(ccast); 1746 cast = ccast; 1747 } 1748 1749 if (cast != NULL) { // Here's the payoff. 1750 replace_in_map(val, cast); 1751 } 1752 } 1753 1754 /** 1755 * Use speculative type to optimize CmpP node: if comparison is 1756 * against the low level class, cast the object to the speculative 1757 * type if any. CmpP should then go away. 1758 * 1759 * @param c expected CmpP node 1760 * @return result of CmpP on object casted to speculative type 1761 * 1762 */ 1763 Node* Parse::optimize_cmp_with_klass(Node* c) { 1764 // If this is transformed by the _gvn to a comparison with the low 1765 // level klass then we may be able to use speculation 1766 if (c->Opcode() == Op_CmpP && 1767 (c->in(1)->Opcode() == Op_LoadKlass || c->in(1)->Opcode() == Op_DecodeNKlass) && 1768 c->in(2)->is_Con()) { 1769 Node* load_klass = NULL; 1770 Node* decode = NULL; 1771 if (c->in(1)->Opcode() == Op_DecodeNKlass) { 1772 decode = c->in(1); 1773 load_klass = c->in(1)->in(1); 1774 } else { 1775 load_klass = c->in(1); 1776 } 1777 if (load_klass->in(2)->is_AddP()) { 1778 Node* addp = load_klass->in(2); 1779 Node* obj = addp->in(AddPNode::Address); 1780 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 1781 if (obj_type->speculative_type_not_null() != NULL) { 1782 ciKlass* k = obj_type->speculative_type(); 1783 inc_sp(2); 1784 obj = maybe_cast_profiled_obj(obj, k); 1785 dec_sp(2); 1786 // Make the CmpP use the casted obj 1787 addp = basic_plus_adr(obj, addp->in(AddPNode::Offset)); 1788 load_klass = load_klass->clone(); 1789 load_klass->set_req(2, addp); 1790 load_klass = _gvn.transform(load_klass); 1791 if (decode != NULL) { 1792 decode = decode->clone(); 1793 decode->set_req(1, load_klass); 1794 load_klass = _gvn.transform(decode); 1795 } 1796 c = c->clone(); 1797 c->set_req(1, load_klass); 1798 c = _gvn.transform(c); 1799 } 1800 } 1801 } 1802 return c; 1803 } 1804 1805 //------------------------------do_one_bytecode-------------------------------- 1806 // Parse this bytecode, and alter the Parsers JVM->Node mapping 1807 void Parse::do_one_bytecode() { 1808 Node *a, *b, *c, *d; // Handy temps 1809 BoolTest::mask btest; 1810 int i; 1811 1812 assert(!has_exceptions(), "bytecode entry state must be clear of throws"); 1813 1814 if (C->check_node_count(NodeLimitFudgeFactor * 5, 1815 "out of nodes parsing method")) { 1816 return; 1817 } 1818 1819 #ifdef ASSERT 1820 // for setting breakpoints 1821 if (TraceOptoParse) { 1822 tty->print(" @"); 1823 dump_bci(bci()); 1824 tty->print(" %s", Bytecodes::name(bc())); 1825 tty->cr(); 1826 } 1827 #endif 1828 1829 switch (bc()) { 1830 case Bytecodes::_nop: 1831 // do nothing 1832 break; 1833 case Bytecodes::_lconst_0: 1834 push_pair(longcon(0)); 1835 break; 1836 1837 case Bytecodes::_lconst_1: 1838 push_pair(longcon(1)); 1839 break; 1840 1841 case Bytecodes::_fconst_0: 1842 push(zerocon(T_FLOAT)); 1843 break; 1844 1845 case Bytecodes::_fconst_1: 1846 push(makecon(TypeF::ONE)); 1847 break; 1848 1849 case Bytecodes::_fconst_2: 1850 push(makecon(TypeF::make(2.0f))); 1851 break; 1852 1853 case Bytecodes::_dconst_0: 1854 push_pair(zerocon(T_DOUBLE)); 1855 break; 1856 1857 case Bytecodes::_dconst_1: 1858 push_pair(makecon(TypeD::ONE)); 1859 break; 1860 1861 case Bytecodes::_iconst_m1:push(intcon(-1)); break; 1862 case Bytecodes::_iconst_0: push(intcon( 0)); break; 1863 case Bytecodes::_iconst_1: push(intcon( 1)); break; 1864 case Bytecodes::_iconst_2: push(intcon( 2)); break; 1865 case Bytecodes::_iconst_3: push(intcon( 3)); break; 1866 case Bytecodes::_iconst_4: push(intcon( 4)); break; 1867 case Bytecodes::_iconst_5: push(intcon( 5)); break; 1868 case Bytecodes::_bipush: push(intcon(iter().get_constant_u1())); break; 1869 case Bytecodes::_sipush: push(intcon(iter().get_constant_u2())); break; 1870 case Bytecodes::_aconst_null: push(null()); break; 1871 1872 case Bytecodes::_ldc: 1873 case Bytecodes::_ldc_w: 1874 case Bytecodes::_ldc2_w: { 1875 ciConstant constant = iter().get_constant(); 1876 if (constant.is_loaded()) { 1877 const Type* con_type = Type::make_from_constant(constant); 1878 if (con_type != NULL) { 1879 push_node(con_type->basic_type(), makecon(con_type)); 1880 } 1881 } else { 1882 // If the constant is unresolved or in error state, run this BC in the interpreter. 1883 if (iter().is_in_error()) { 1884 uncommon_trap(Deoptimization::make_trap_request(Deoptimization::Reason_unhandled, 1885 Deoptimization::Action_none), 1886 NULL, "constant in error state", true /* must_throw */); 1887 1888 } else { 1889 int index = iter().get_constant_pool_index(); 1890 uncommon_trap(Deoptimization::make_trap_request(Deoptimization::Reason_unloaded, 1891 Deoptimization::Action_reinterpret, 1892 index), 1893 NULL, "unresolved constant", false /* must_throw */); 1894 } 1895 } 1896 break; 1897 } 1898 1899 case Bytecodes::_aload_0: 1900 push( local(0) ); 1901 break; 1902 case Bytecodes::_aload_1: 1903 push( local(1) ); 1904 break; 1905 case Bytecodes::_aload_2: 1906 push( local(2) ); 1907 break; 1908 case Bytecodes::_aload_3: 1909 push( local(3) ); 1910 break; 1911 case Bytecodes::_aload: 1912 push( local(iter().get_index()) ); 1913 break; 1914 1915 case Bytecodes::_fload_0: 1916 case Bytecodes::_iload_0: 1917 push( local(0) ); 1918 break; 1919 case Bytecodes::_fload_1: 1920 case Bytecodes::_iload_1: 1921 push( local(1) ); 1922 break; 1923 case Bytecodes::_fload_2: 1924 case Bytecodes::_iload_2: 1925 push( local(2) ); 1926 break; 1927 case Bytecodes::_fload_3: 1928 case Bytecodes::_iload_3: 1929 push( local(3) ); 1930 break; 1931 case Bytecodes::_fload: 1932 case Bytecodes::_iload: 1933 push( local(iter().get_index()) ); 1934 break; 1935 case Bytecodes::_lload_0: 1936 push_pair_local( 0 ); 1937 break; 1938 case Bytecodes::_lload_1: 1939 push_pair_local( 1 ); 1940 break; 1941 case Bytecodes::_lload_2: 1942 push_pair_local( 2 ); 1943 break; 1944 case Bytecodes::_lload_3: 1945 push_pair_local( 3 ); 1946 break; 1947 case Bytecodes::_lload: 1948 push_pair_local( iter().get_index() ); 1949 break; 1950 1951 case Bytecodes::_dload_0: 1952 push_pair_local(0); 1953 break; 1954 case Bytecodes::_dload_1: 1955 push_pair_local(1); 1956 break; 1957 case Bytecodes::_dload_2: 1958 push_pair_local(2); 1959 break; 1960 case Bytecodes::_dload_3: 1961 push_pair_local(3); 1962 break; 1963 case Bytecodes::_dload: 1964 push_pair_local(iter().get_index()); 1965 break; 1966 case Bytecodes::_fstore_0: 1967 case Bytecodes::_istore_0: 1968 case Bytecodes::_astore_0: 1969 set_local( 0, pop() ); 1970 break; 1971 case Bytecodes::_fstore_1: 1972 case Bytecodes::_istore_1: 1973 case Bytecodes::_astore_1: 1974 set_local( 1, pop() ); 1975 break; 1976 case Bytecodes::_fstore_2: 1977 case Bytecodes::_istore_2: 1978 case Bytecodes::_astore_2: 1979 set_local( 2, pop() ); 1980 break; 1981 case Bytecodes::_fstore_3: 1982 case Bytecodes::_istore_3: 1983 case Bytecodes::_astore_3: 1984 set_local( 3, pop() ); 1985 break; 1986 case Bytecodes::_fstore: 1987 case Bytecodes::_istore: 1988 case Bytecodes::_astore: 1989 set_local( iter().get_index(), pop() ); 1990 break; 1991 // long stores 1992 case Bytecodes::_lstore_0: 1993 set_pair_local( 0, pop_pair() ); 1994 break; 1995 case Bytecodes::_lstore_1: 1996 set_pair_local( 1, pop_pair() ); 1997 break; 1998 case Bytecodes::_lstore_2: 1999 set_pair_local( 2, pop_pair() ); 2000 break; 2001 case Bytecodes::_lstore_3: 2002 set_pair_local( 3, pop_pair() ); 2003 break; 2004 case Bytecodes::_lstore: 2005 set_pair_local( iter().get_index(), pop_pair() ); 2006 break; 2007 2008 // double stores 2009 case Bytecodes::_dstore_0: 2010 set_pair_local( 0, dprecision_rounding(pop_pair()) ); 2011 break; 2012 case Bytecodes::_dstore_1: 2013 set_pair_local( 1, dprecision_rounding(pop_pair()) ); 2014 break; 2015 case Bytecodes::_dstore_2: 2016 set_pair_local( 2, dprecision_rounding(pop_pair()) ); 2017 break; 2018 case Bytecodes::_dstore_3: 2019 set_pair_local( 3, dprecision_rounding(pop_pair()) ); 2020 break; 2021 case Bytecodes::_dstore: 2022 set_pair_local( iter().get_index(), dprecision_rounding(pop_pair()) ); 2023 break; 2024 2025 case Bytecodes::_pop: dec_sp(1); break; 2026 case Bytecodes::_pop2: dec_sp(2); break; 2027 case Bytecodes::_swap: 2028 a = pop(); 2029 b = pop(); 2030 push(a); 2031 push(b); 2032 break; 2033 case Bytecodes::_dup: 2034 a = pop(); 2035 push(a); 2036 push(a); 2037 break; 2038 case Bytecodes::_dup_x1: 2039 a = pop(); 2040 b = pop(); 2041 push( a ); 2042 push( b ); 2043 push( a ); 2044 break; 2045 case Bytecodes::_dup_x2: 2046 a = pop(); 2047 b = pop(); 2048 c = pop(); 2049 push( a ); 2050 push( c ); 2051 push( b ); 2052 push( a ); 2053 break; 2054 case Bytecodes::_dup2: 2055 a = pop(); 2056 b = pop(); 2057 push( b ); 2058 push( a ); 2059 push( b ); 2060 push( a ); 2061 break; 2062 2063 case Bytecodes::_dup2_x1: 2064 // before: .. c, b, a 2065 // after: .. b, a, c, b, a 2066 // not tested 2067 a = pop(); 2068 b = pop(); 2069 c = pop(); 2070 push( b ); 2071 push( a ); 2072 push( c ); 2073 push( b ); 2074 push( a ); 2075 break; 2076 case Bytecodes::_dup2_x2: 2077 // before: .. d, c, b, a 2078 // after: .. b, a, d, c, b, a 2079 // not tested 2080 a = pop(); 2081 b = pop(); 2082 c = pop(); 2083 d = pop(); 2084 push( b ); 2085 push( a ); 2086 push( d ); 2087 push( c ); 2088 push( b ); 2089 push( a ); 2090 break; 2091 2092 case Bytecodes::_arraylength: { 2093 // Must do null-check with value on expression stack 2094 Node *ary = null_check(peek(), T_ARRAY); 2095 // Compile-time detect of null-exception? 2096 if (stopped()) return; 2097 a = pop(); 2098 push(load_array_length(a)); 2099 break; 2100 } 2101 2102 case Bytecodes::_baload: array_load(T_BYTE); break; 2103 case Bytecodes::_caload: array_load(T_CHAR); break; 2104 case Bytecodes::_iaload: array_load(T_INT); break; 2105 case Bytecodes::_saload: array_load(T_SHORT); break; 2106 case Bytecodes::_faload: array_load(T_FLOAT); break; 2107 case Bytecodes::_aaload: array_load(T_OBJECT); break; 2108 case Bytecodes::_laload: array_load(T_LONG); break; 2109 case Bytecodes::_daload: array_load(T_DOUBLE); break; 2110 case Bytecodes::_bastore: array_store(T_BYTE); break; 2111 case Bytecodes::_castore: array_store(T_CHAR); break; 2112 case Bytecodes::_iastore: array_store(T_INT); break; 2113 case Bytecodes::_sastore: array_store(T_SHORT); break; 2114 case Bytecodes::_fastore: array_store(T_FLOAT); break; 2115 case Bytecodes::_aastore: array_store(T_OBJECT); break; 2116 case Bytecodes::_lastore: array_store(T_LONG); break; 2117 case Bytecodes::_dastore: array_store(T_DOUBLE); break; 2118 2119 case Bytecodes::_getfield: 2120 do_getfield(); 2121 break; 2122 2123 case Bytecodes::_getstatic: 2124 do_getstatic(); 2125 break; 2126 2127 case Bytecodes::_putfield: 2128 do_putfield(); 2129 break; 2130 2131 case Bytecodes::_putstatic: 2132 do_putstatic(); 2133 break; 2134 2135 case Bytecodes::_irem: 2136 // Must keep both values on the expression-stack during null-check 2137 zero_check_int(peek()); 2138 // Compile-time detect of null-exception? 2139 if (stopped()) return; 2140 b = pop(); 2141 a = pop(); 2142 push(_gvn.transform(new ModINode(control(), a, b))); 2143 break; 2144 case Bytecodes::_idiv: 2145 // Must keep both values on the expression-stack during null-check 2146 zero_check_int(peek()); 2147 // Compile-time detect of null-exception? 2148 if (stopped()) return; 2149 b = pop(); 2150 a = pop(); 2151 push( _gvn.transform( new DivINode(control(),a,b) ) ); 2152 break; 2153 case Bytecodes::_imul: 2154 b = pop(); a = pop(); 2155 push( _gvn.transform( new MulINode(a,b) ) ); 2156 break; 2157 case Bytecodes::_iadd: 2158 b = pop(); a = pop(); 2159 push( _gvn.transform( new AddINode(a,b) ) ); 2160 break; 2161 case Bytecodes::_ineg: 2162 a = pop(); 2163 push( _gvn.transform( new SubINode(_gvn.intcon(0),a)) ); 2164 break; 2165 case Bytecodes::_isub: 2166 b = pop(); a = pop(); 2167 push( _gvn.transform( new SubINode(a,b) ) ); 2168 break; 2169 case Bytecodes::_iand: 2170 b = pop(); a = pop(); 2171 push( _gvn.transform( new AndINode(a,b) ) ); 2172 break; 2173 case Bytecodes::_ior: 2174 b = pop(); a = pop(); 2175 push( _gvn.transform( new OrINode(a,b) ) ); 2176 break; 2177 case Bytecodes::_ixor: 2178 b = pop(); a = pop(); 2179 push( _gvn.transform( new XorINode(a,b) ) ); 2180 break; 2181 case Bytecodes::_ishl: 2182 b = pop(); a = pop(); 2183 push( _gvn.transform( new LShiftINode(a,b) ) ); 2184 break; 2185 case Bytecodes::_ishr: 2186 b = pop(); a = pop(); 2187 push( _gvn.transform( new RShiftINode(a,b) ) ); 2188 break; 2189 case Bytecodes::_iushr: 2190 b = pop(); a = pop(); 2191 push( _gvn.transform( new URShiftINode(a,b) ) ); 2192 break; 2193 2194 case Bytecodes::_fneg: 2195 a = pop(); 2196 b = _gvn.transform(new NegFNode (a)); 2197 push(b); 2198 break; 2199 2200 case Bytecodes::_fsub: 2201 b = pop(); 2202 a = pop(); 2203 c = _gvn.transform( new SubFNode(a,b) ); 2204 d = precision_rounding(c); 2205 push( d ); 2206 break; 2207 2208 case Bytecodes::_fadd: 2209 b = pop(); 2210 a = pop(); 2211 c = _gvn.transform( new AddFNode(a,b) ); 2212 d = precision_rounding(c); 2213 push( d ); 2214 break; 2215 2216 case Bytecodes::_fmul: 2217 b = pop(); 2218 a = pop(); 2219 c = _gvn.transform( new MulFNode(a,b) ); 2220 d = precision_rounding(c); 2221 push( d ); 2222 break; 2223 2224 case Bytecodes::_fdiv: 2225 b = pop(); 2226 a = pop(); 2227 c = _gvn.transform( new DivFNode(0,a,b) ); 2228 d = precision_rounding(c); 2229 push( d ); 2230 break; 2231 2232 case Bytecodes::_frem: 2233 if (Matcher::has_match_rule(Op_ModF)) { 2234 // Generate a ModF node. 2235 b = pop(); 2236 a = pop(); 2237 c = _gvn.transform( new ModFNode(0,a,b) ); 2238 d = precision_rounding(c); 2239 push( d ); 2240 } 2241 else { 2242 // Generate a call. 2243 modf(); 2244 } 2245 break; 2246 2247 case Bytecodes::_fcmpl: 2248 b = pop(); 2249 a = pop(); 2250 c = _gvn.transform( new CmpF3Node( a, b)); 2251 push(c); 2252 break; 2253 case Bytecodes::_fcmpg: 2254 b = pop(); 2255 a = pop(); 2256 2257 // Same as fcmpl but need to flip the unordered case. Swap the inputs, 2258 // which negates the result sign except for unordered. Flip the unordered 2259 // as well by using CmpF3 which implements unordered-lesser instead of 2260 // unordered-greater semantics. Finally, commute the result bits. Result 2261 // is same as using a CmpF3Greater except we did it with CmpF3 alone. 2262 c = _gvn.transform( new CmpF3Node( b, a)); 2263 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) ); 2264 push(c); 2265 break; 2266 2267 case Bytecodes::_f2i: 2268 a = pop(); 2269 push(_gvn.transform(new ConvF2INode(a))); 2270 break; 2271 2272 case Bytecodes::_d2i: 2273 a = pop_pair(); 2274 b = _gvn.transform(new ConvD2INode(a)); 2275 push( b ); 2276 break; 2277 2278 case Bytecodes::_f2d: 2279 a = pop(); 2280 b = _gvn.transform( new ConvF2DNode(a)); 2281 push_pair( b ); 2282 break; 2283 2284 case Bytecodes::_d2f: 2285 a = pop_pair(); 2286 b = _gvn.transform( new ConvD2FNode(a)); 2287 // This breaks _227_mtrt (speed & correctness) and _222_mpegaudio (speed) 2288 //b = _gvn.transform(new RoundFloatNode(0, b) ); 2289 push( b ); 2290 break; 2291 2292 case Bytecodes::_l2f: 2293 if (Matcher::convL2FSupported()) { 2294 a = pop_pair(); 2295 b = _gvn.transform( new ConvL2FNode(a)); 2296 // For x86_32.ad, FILD doesn't restrict precision to 24 or 53 bits. 2297 // Rather than storing the result into an FP register then pushing 2298 // out to memory to round, the machine instruction that implements 2299 // ConvL2D is responsible for rounding. 2300 // c = precision_rounding(b); 2301 push(b); 2302 } else { 2303 l2f(); 2304 } 2305 break; 2306 2307 case Bytecodes::_l2d: 2308 a = pop_pair(); 2309 b = _gvn.transform( new ConvL2DNode(a)); 2310 // For x86_32.ad, rounding is always necessary (see _l2f above). 2311 // c = dprecision_rounding(b); 2312 push_pair(b); 2313 break; 2314 2315 case Bytecodes::_f2l: 2316 a = pop(); 2317 b = _gvn.transform( new ConvF2LNode(a)); 2318 push_pair(b); 2319 break; 2320 2321 case Bytecodes::_d2l: 2322 a = pop_pair(); 2323 b = _gvn.transform( new ConvD2LNode(a)); 2324 push_pair(b); 2325 break; 2326 2327 case Bytecodes::_dsub: 2328 b = pop_pair(); 2329 a = pop_pair(); 2330 c = _gvn.transform( new SubDNode(a,b) ); 2331 d = dprecision_rounding(c); 2332 push_pair( d ); 2333 break; 2334 2335 case Bytecodes::_dadd: 2336 b = pop_pair(); 2337 a = pop_pair(); 2338 c = _gvn.transform( new AddDNode(a,b) ); 2339 d = dprecision_rounding(c); 2340 push_pair( d ); 2341 break; 2342 2343 case Bytecodes::_dmul: 2344 b = pop_pair(); 2345 a = pop_pair(); 2346 c = _gvn.transform( new MulDNode(a,b) ); 2347 d = dprecision_rounding(c); 2348 push_pair( d ); 2349 break; 2350 2351 case Bytecodes::_ddiv: 2352 b = pop_pair(); 2353 a = pop_pair(); 2354 c = _gvn.transform( new DivDNode(0,a,b) ); 2355 d = dprecision_rounding(c); 2356 push_pair( d ); 2357 break; 2358 2359 case Bytecodes::_dneg: 2360 a = pop_pair(); 2361 b = _gvn.transform(new NegDNode (a)); 2362 push_pair(b); 2363 break; 2364 2365 case Bytecodes::_drem: 2366 if (Matcher::has_match_rule(Op_ModD)) { 2367 // Generate a ModD node. 2368 b = pop_pair(); 2369 a = pop_pair(); 2370 // a % b 2371 2372 c = _gvn.transform( new ModDNode(0,a,b) ); 2373 d = dprecision_rounding(c); 2374 push_pair( d ); 2375 } 2376 else { 2377 // Generate a call. 2378 modd(); 2379 } 2380 break; 2381 2382 case Bytecodes::_dcmpl: 2383 b = pop_pair(); 2384 a = pop_pair(); 2385 c = _gvn.transform( new CmpD3Node( a, b)); 2386 push(c); 2387 break; 2388 2389 case Bytecodes::_dcmpg: 2390 b = pop_pair(); 2391 a = pop_pair(); 2392 // Same as dcmpl but need to flip the unordered case. 2393 // Commute the inputs, which negates the result sign except for unordered. 2394 // Flip the unordered as well by using CmpD3 which implements 2395 // unordered-lesser instead of unordered-greater semantics. 2396 // Finally, negate the result bits. Result is same as using a 2397 // CmpD3Greater except we did it with CmpD3 alone. 2398 c = _gvn.transform( new CmpD3Node( b, a)); 2399 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) ); 2400 push(c); 2401 break; 2402 2403 2404 // Note for longs -> lo word is on TOS, hi word is on TOS - 1 2405 case Bytecodes::_land: 2406 b = pop_pair(); 2407 a = pop_pair(); 2408 c = _gvn.transform( new AndLNode(a,b) ); 2409 push_pair(c); 2410 break; 2411 case Bytecodes::_lor: 2412 b = pop_pair(); 2413 a = pop_pair(); 2414 c = _gvn.transform( new OrLNode(a,b) ); 2415 push_pair(c); 2416 break; 2417 case Bytecodes::_lxor: 2418 b = pop_pair(); 2419 a = pop_pair(); 2420 c = _gvn.transform( new XorLNode(a,b) ); 2421 push_pair(c); 2422 break; 2423 2424 case Bytecodes::_lshl: 2425 b = pop(); // the shift count 2426 a = pop_pair(); // value to be shifted 2427 c = _gvn.transform( new LShiftLNode(a,b) ); 2428 push_pair(c); 2429 break; 2430 case Bytecodes::_lshr: 2431 b = pop(); // the shift count 2432 a = pop_pair(); // value to be shifted 2433 c = _gvn.transform( new RShiftLNode(a,b) ); 2434 push_pair(c); 2435 break; 2436 case Bytecodes::_lushr: 2437 b = pop(); // the shift count 2438 a = pop_pair(); // value to be shifted 2439 c = _gvn.transform( new URShiftLNode(a,b) ); 2440 push_pair(c); 2441 break; 2442 case Bytecodes::_lmul: 2443 b = pop_pair(); 2444 a = pop_pair(); 2445 c = _gvn.transform( new MulLNode(a,b) ); 2446 push_pair(c); 2447 break; 2448 2449 case Bytecodes::_lrem: 2450 // Must keep both values on the expression-stack during null-check 2451 assert(peek(0) == top(), "long word order"); 2452 zero_check_long(peek(1)); 2453 // Compile-time detect of null-exception? 2454 if (stopped()) return; 2455 b = pop_pair(); 2456 a = pop_pair(); 2457 c = _gvn.transform( new ModLNode(control(),a,b) ); 2458 push_pair(c); 2459 break; 2460 2461 case Bytecodes::_ldiv: 2462 // Must keep both values on the expression-stack during null-check 2463 assert(peek(0) == top(), "long word order"); 2464 zero_check_long(peek(1)); 2465 // Compile-time detect of null-exception? 2466 if (stopped()) return; 2467 b = pop_pair(); 2468 a = pop_pair(); 2469 c = _gvn.transform( new DivLNode(control(),a,b) ); 2470 push_pair(c); 2471 break; 2472 2473 case Bytecodes::_ladd: 2474 b = pop_pair(); 2475 a = pop_pair(); 2476 c = _gvn.transform( new AddLNode(a,b) ); 2477 push_pair(c); 2478 break; 2479 case Bytecodes::_lsub: 2480 b = pop_pair(); 2481 a = pop_pair(); 2482 c = _gvn.transform( new SubLNode(a,b) ); 2483 push_pair(c); 2484 break; 2485 case Bytecodes::_lcmp: 2486 // Safepoints are now inserted _before_ branches. The long-compare 2487 // bytecode painfully produces a 3-way value (-1,0,+1) which requires a 2488 // slew of control flow. These are usually followed by a CmpI vs zero and 2489 // a branch; this pattern then optimizes to the obvious long-compare and 2490 // branch. However, if the branch is backwards there's a Safepoint 2491 // inserted. The inserted Safepoint captures the JVM state at the 2492 // pre-branch point, i.e. it captures the 3-way value. Thus if a 2493 // long-compare is used to control a loop the debug info will force 2494 // computation of the 3-way value, even though the generated code uses a 2495 // long-compare and branch. We try to rectify the situation by inserting 2496 // a SafePoint here and have it dominate and kill the safepoint added at a 2497 // following backwards branch. At this point the JVM state merely holds 2 2498 // longs but not the 3-way value. 2499 switch (iter().next_bc()) { 2500 case Bytecodes::_ifgt: 2501 case Bytecodes::_iflt: 2502 case Bytecodes::_ifge: 2503 case Bytecodes::_ifle: 2504 case Bytecodes::_ifne: 2505 case Bytecodes::_ifeq: 2506 // If this is a backwards branch in the bytecodes, add Safepoint 2507 maybe_add_safepoint(iter().next_get_dest()); 2508 default: 2509 break; 2510 } 2511 b = pop_pair(); 2512 a = pop_pair(); 2513 c = _gvn.transform( new CmpL3Node( a, b )); 2514 push(c); 2515 break; 2516 2517 case Bytecodes::_lneg: 2518 a = pop_pair(); 2519 b = _gvn.transform( new SubLNode(longcon(0),a)); 2520 push_pair(b); 2521 break; 2522 case Bytecodes::_l2i: 2523 a = pop_pair(); 2524 push( _gvn.transform( new ConvL2INode(a))); 2525 break; 2526 case Bytecodes::_i2l: 2527 a = pop(); 2528 b = _gvn.transform( new ConvI2LNode(a)); 2529 push_pair(b); 2530 break; 2531 case Bytecodes::_i2b: 2532 // Sign extend 2533 a = pop(); 2534 a = Compile::narrow_value(T_BYTE, a, NULL, &_gvn, true); 2535 push(a); 2536 break; 2537 case Bytecodes::_i2s: 2538 a = pop(); 2539 a = Compile::narrow_value(T_SHORT, a, NULL, &_gvn, true); 2540 push(a); 2541 break; 2542 case Bytecodes::_i2c: 2543 a = pop(); 2544 a = Compile::narrow_value(T_CHAR, a, NULL, &_gvn, true); 2545 push(a); 2546 break; 2547 2548 case Bytecodes::_i2f: 2549 a = pop(); 2550 b = _gvn.transform( new ConvI2FNode(a) ) ; 2551 c = precision_rounding(b); 2552 push (b); 2553 break; 2554 2555 case Bytecodes::_i2d: 2556 a = pop(); 2557 b = _gvn.transform( new ConvI2DNode(a)); 2558 push_pair(b); 2559 break; 2560 2561 case Bytecodes::_iinc: // Increment local 2562 i = iter().get_index(); // Get local index 2563 set_local( i, _gvn.transform( new AddINode( _gvn.intcon(iter().get_iinc_con()), local(i) ) ) ); 2564 break; 2565 2566 // Exit points of synchronized methods must have an unlock node 2567 case Bytecodes::_return: 2568 return_current(NULL); 2569 break; 2570 2571 case Bytecodes::_ireturn: 2572 case Bytecodes::_areturn: 2573 case Bytecodes::_freturn: 2574 return_current(pop()); 2575 break; 2576 case Bytecodes::_lreturn: 2577 return_current(pop_pair()); 2578 break; 2579 case Bytecodes::_dreturn: 2580 return_current(pop_pair()); 2581 break; 2582 2583 case Bytecodes::_athrow: 2584 // null exception oop throws NULL pointer exception 2585 null_check(peek()); 2586 if (stopped()) return; 2587 // Hook the thrown exception directly to subsequent handlers. 2588 if (BailoutToInterpreterForThrows) { 2589 // Keep method interpreted from now on. 2590 uncommon_trap(Deoptimization::Reason_unhandled, 2591 Deoptimization::Action_make_not_compilable); 2592 return; 2593 } 2594 if (env()->jvmti_can_post_on_exceptions()) { 2595 // check if we must post exception events, take uncommon trap if so (with must_throw = false) 2596 uncommon_trap_if_should_post_on_exceptions(Deoptimization::Reason_unhandled, false); 2597 } 2598 // Here if either can_post_on_exceptions or should_post_on_exceptions is false 2599 add_exception_state(make_exception_state(peek())); 2600 break; 2601 2602 case Bytecodes::_goto: // fall through 2603 case Bytecodes::_goto_w: { 2604 int target_bci = (bc() == Bytecodes::_goto) ? iter().get_dest() : iter().get_far_dest(); 2605 2606 // If this is a backwards branch in the bytecodes, add Safepoint 2607 maybe_add_safepoint(target_bci); 2608 2609 // Merge the current control into the target basic block 2610 merge(target_bci); 2611 2612 // See if we can get some profile data and hand it off to the next block 2613 Block *target_block = block()->successor_for_bci(target_bci); 2614 if (target_block->pred_count() != 1) break; 2615 ciMethodData* methodData = method()->method_data(); 2616 if (!methodData->is_mature()) break; 2617 ciProfileData* data = methodData->bci_to_data(bci()); 2618 assert(data != NULL && data->is_JumpData(), "need JumpData for taken branch"); 2619 int taken = ((ciJumpData*)data)->taken(); 2620 taken = method()->scale_count(taken); 2621 target_block->set_count(taken); 2622 break; 2623 } 2624 2625 case Bytecodes::_ifnull: btest = BoolTest::eq; goto handle_if_null; 2626 case Bytecodes::_ifnonnull: btest = BoolTest::ne; goto handle_if_null; 2627 handle_if_null: 2628 // If this is a backwards branch in the bytecodes, add Safepoint 2629 maybe_add_safepoint(iter().get_dest()); 2630 a = null(); 2631 b = pop(); 2632 if (!_gvn.type(b)->speculative_maybe_null() && 2633 !too_many_traps(Deoptimization::Reason_speculate_null_check)) { 2634 inc_sp(1); 2635 Node* null_ctl = top(); 2636 b = null_check_oop(b, &null_ctl, true, true, true); 2637 assert(null_ctl->is_top(), "no null control here"); 2638 dec_sp(1); 2639 } else if (_gvn.type(b)->speculative_always_null() && 2640 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) { 2641 inc_sp(1); 2642 b = null_assert(b); 2643 dec_sp(1); 2644 } 2645 c = _gvn.transform( new CmpPNode(b, a) ); 2646 do_ifnull(btest, c); 2647 break; 2648 2649 case Bytecodes::_if_acmpeq: btest = BoolTest::eq; goto handle_if_acmp; 2650 case Bytecodes::_if_acmpne: btest = BoolTest::ne; goto handle_if_acmp; 2651 handle_if_acmp: 2652 // If this is a backwards branch in the bytecodes, add Safepoint 2653 maybe_add_safepoint(iter().get_dest()); 2654 a = pop(); 2655 b = pop(); 2656 c = _gvn.transform( new CmpPNode(b, a) ); 2657 c = optimize_cmp_with_klass(c); 2658 do_if(btest, c); 2659 break; 2660 2661 case Bytecodes::_ifeq: btest = BoolTest::eq; goto handle_ifxx; 2662 case Bytecodes::_ifne: btest = BoolTest::ne; goto handle_ifxx; 2663 case Bytecodes::_iflt: btest = BoolTest::lt; goto handle_ifxx; 2664 case Bytecodes::_ifle: btest = BoolTest::le; goto handle_ifxx; 2665 case Bytecodes::_ifgt: btest = BoolTest::gt; goto handle_ifxx; 2666 case Bytecodes::_ifge: btest = BoolTest::ge; goto handle_ifxx; 2667 handle_ifxx: 2668 // If this is a backwards branch in the bytecodes, add Safepoint 2669 maybe_add_safepoint(iter().get_dest()); 2670 a = _gvn.intcon(0); 2671 b = pop(); 2672 c = _gvn.transform( new CmpINode(b, a) ); 2673 do_if(btest, c); 2674 break; 2675 2676 case Bytecodes::_if_icmpeq: btest = BoolTest::eq; goto handle_if_icmp; 2677 case Bytecodes::_if_icmpne: btest = BoolTest::ne; goto handle_if_icmp; 2678 case Bytecodes::_if_icmplt: btest = BoolTest::lt; goto handle_if_icmp; 2679 case Bytecodes::_if_icmple: btest = BoolTest::le; goto handle_if_icmp; 2680 case Bytecodes::_if_icmpgt: btest = BoolTest::gt; goto handle_if_icmp; 2681 case Bytecodes::_if_icmpge: btest = BoolTest::ge; goto handle_if_icmp; 2682 handle_if_icmp: 2683 // If this is a backwards branch in the bytecodes, add Safepoint 2684 maybe_add_safepoint(iter().get_dest()); 2685 a = pop(); 2686 b = pop(); 2687 c = _gvn.transform( new CmpINode( b, a ) ); 2688 do_if(btest, c); 2689 break; 2690 2691 case Bytecodes::_tableswitch: 2692 do_tableswitch(); 2693 break; 2694 2695 case Bytecodes::_lookupswitch: 2696 do_lookupswitch(); 2697 break; 2698 2699 case Bytecodes::_invokestatic: 2700 case Bytecodes::_invokedynamic: 2701 case Bytecodes::_invokespecial: 2702 case Bytecodes::_invokevirtual: 2703 case Bytecodes::_invokeinterface: 2704 do_call(); 2705 break; 2706 case Bytecodes::_checkcast: 2707 do_checkcast(); 2708 break; 2709 case Bytecodes::_instanceof: 2710 do_instanceof(); 2711 break; 2712 case Bytecodes::_anewarray: 2713 do_anewarray(); 2714 break; 2715 case Bytecodes::_newarray: 2716 do_newarray((BasicType)iter().get_index()); 2717 break; 2718 case Bytecodes::_multianewarray: 2719 do_multianewarray(); 2720 break; 2721 case Bytecodes::_new: 2722 do_new(); 2723 break; 2724 2725 case Bytecodes::_jsr: 2726 case Bytecodes::_jsr_w: 2727 do_jsr(); 2728 break; 2729 2730 case Bytecodes::_ret: 2731 do_ret(); 2732 break; 2733 2734 2735 case Bytecodes::_monitorenter: 2736 do_monitor_enter(); 2737 break; 2738 2739 case Bytecodes::_monitorexit: 2740 do_monitor_exit(); 2741 break; 2742 2743 case Bytecodes::_breakpoint: 2744 // Breakpoint set concurrently to compile 2745 // %%% use an uncommon trap? 2746 C->record_failure("breakpoint in method"); 2747 return; 2748 2749 default: 2750 #ifndef PRODUCT 2751 map()->dump(99); 2752 #endif 2753 tty->print("\nUnhandled bytecode %s\n", Bytecodes::name(bc()) ); 2754 ShouldNotReachHere(); 2755 } 2756 2757 #ifndef PRODUCT 2758 if (C->should_print_igv(1)) { 2759 IdealGraphPrinter* printer = C->igv_printer(); 2760 char buffer[256]; 2761 jio_snprintf(buffer, sizeof(buffer), "Bytecode %d: %s", bci(), Bytecodes::name(bc())); 2762 bool old = printer->traverse_outs(); 2763 printer->set_traverse_outs(true); 2764 printer->print_method(buffer, 4); 2765 printer->set_traverse_outs(old); 2766 } 2767 #endif 2768 }