1 /* 2 * Copyright (c) 2000, 2017, 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/ciTypeFlow.hpp" 27 #include "memory/allocation.inline.hpp" 28 #include "memory/resourceArea.hpp" 29 #include "opto/addnode.hpp" 30 #include "opto/castnode.hpp" 31 #include "opto/cfgnode.hpp" 32 #include "opto/connode.hpp" 33 #include "opto/loopnode.hpp" 34 #include "opto/phaseX.hpp" 35 #include "opto/runtime.hpp" 36 #include "opto/rootnode.hpp" 37 #include "opto/subnode.hpp" 38 39 // Portions of code courtesy of Clifford Click 40 41 // Optimization - Graph Style 42 43 44 #ifndef PRODUCT 45 extern int explicit_null_checks_elided; 46 #endif 47 48 //============================================================================= 49 //------------------------------Value------------------------------------------ 50 // Return a tuple for whichever arm of the IF is reachable 51 const Type* IfNode::Value(PhaseGVN* phase) const { 52 if( !in(0) ) return Type::TOP; 53 if( phase->type(in(0)) == Type::TOP ) 54 return Type::TOP; 55 const Type *t = phase->type(in(1)); 56 if( t == Type::TOP ) // data is undefined 57 return TypeTuple::IFNEITHER; // unreachable altogether 58 if( t == TypeInt::ZERO ) // zero, or false 59 return TypeTuple::IFFALSE; // only false branch is reachable 60 if( t == TypeInt::ONE ) // 1, or true 61 return TypeTuple::IFTRUE; // only true branch is reachable 62 assert( t == TypeInt::BOOL, "expected boolean type" ); 63 64 return TypeTuple::IFBOTH; // No progress 65 } 66 67 const RegMask &IfNode::out_RegMask() const { 68 return RegMask::Empty; 69 } 70 71 //------------------------------split_if--------------------------------------- 72 // Look for places where we merge constants, then test on the merged value. 73 // If the IF test will be constant folded on the path with the constant, we 74 // win by splitting the IF to before the merge point. 75 static Node* split_if(IfNode *iff, PhaseIterGVN *igvn) { 76 // I could be a lot more general here, but I'm trying to squeeze this 77 // in before the Christmas '98 break so I'm gonna be kinda restrictive 78 // on the patterns I accept. CNC 79 80 // Look for a compare of a constant and a merged value 81 Node *i1 = iff->in(1); 82 if( !i1->is_Bool() ) return NULL; 83 BoolNode *b = i1->as_Bool(); 84 Node *cmp = b->in(1); 85 if( !cmp->is_Cmp() ) return NULL; 86 i1 = cmp->in(1); 87 if( i1 == NULL || !i1->is_Phi() ) return NULL; 88 PhiNode *phi = i1->as_Phi(); 89 if( phi->is_copy() ) return NULL; 90 Node *con2 = cmp->in(2); 91 if( !con2->is_Con() ) return NULL; 92 // See that the merge point contains some constants 93 Node *con1=NULL; 94 uint i4; 95 for( i4 = 1; i4 < phi->req(); i4++ ) { 96 con1 = phi->in(i4); 97 if( !con1 ) return NULL; // Do not optimize partially collapsed merges 98 if( con1->is_Con() ) break; // Found a constant 99 // Also allow null-vs-not-null checks 100 const TypePtr *tp = igvn->type(con1)->isa_ptr(); 101 if( tp && tp->_ptr == TypePtr::NotNull ) 102 break; 103 } 104 if( i4 >= phi->req() ) return NULL; // Found no constants 105 106 igvn->C->set_has_split_ifs(true); // Has chance for split-if 107 108 // Make sure that the compare can be constant folded away 109 Node *cmp2 = cmp->clone(); 110 cmp2->set_req(1,con1); 111 cmp2->set_req(2,con2); 112 const Type *t = cmp2->Value(igvn); 113 // This compare is dead, so whack it! 114 igvn->remove_dead_node(cmp2); 115 if( !t->singleton() ) return NULL; 116 117 // No intervening control, like a simple Call 118 Node *r = iff->in(0); 119 if( !r->is_Region() ) return NULL; 120 if (r->is_Loop() && r->in(LoopNode::LoopBackControl)->is_top()) return NULL; // going away anyway 121 if( phi->region() != r ) return NULL; 122 // No other users of the cmp/bool 123 if (b->outcnt() != 1 || cmp->outcnt() != 1) { 124 //tty->print_cr("many users of cmp/bool"); 125 return NULL; 126 } 127 128 // Make sure we can determine where all the uses of merged values go 129 for (DUIterator_Fast jmax, j = r->fast_outs(jmax); j < jmax; j++) { 130 Node* u = r->fast_out(j); 131 if( u == r ) continue; 132 if( u == iff ) continue; 133 if( u->outcnt() == 0 ) continue; // use is dead & ignorable 134 if( !u->is_Phi() ) { 135 /* 136 if( u->is_Start() ) { 137 tty->print_cr("Region has inlined start use"); 138 } else { 139 tty->print_cr("Region has odd use"); 140 u->dump(2); 141 }*/ 142 return NULL; 143 } 144 if( u != phi ) { 145 // CNC - do not allow any other merged value 146 //tty->print_cr("Merging another value"); 147 //u->dump(2); 148 return NULL; 149 } 150 // Make sure we can account for all Phi uses 151 for (DUIterator_Fast kmax, k = u->fast_outs(kmax); k < kmax; k++) { 152 Node* v = u->fast_out(k); // User of the phi 153 // CNC - Allow only really simple patterns. 154 // In particular I disallow AddP of the Phi, a fairly common pattern 155 if (v == cmp) continue; // The compare is OK 156 if (v->is_ConstraintCast()) { 157 // If the cast is derived from data flow edges, it may not have a control edge. 158 // If so, it should be safe to split. But follow-up code can not deal with 159 // this (l. 359). So skip. 160 if (v->in(0) == NULL) { 161 return NULL; 162 } 163 if (v->in(0)->in(0) == iff) { 164 continue; // CastPP/II of the IfNode is OK 165 } 166 } 167 // Disabled following code because I cannot tell if exactly one 168 // path dominates without a real dominator check. CNC 9/9/1999 169 //uint vop = v->Opcode(); 170 //if( vop == Op_Phi ) { // Phi from another merge point might be OK 171 // Node *r = v->in(0); // Get controlling point 172 // if( !r ) return NULL; // Degraded to a copy 173 // // Find exactly one path in (either True or False doms, but not IFF) 174 // int cnt = 0; 175 // for( uint i = 1; i < r->req(); i++ ) 176 // if( r->in(i) && r->in(i)->in(0) == iff ) 177 // cnt++; 178 // if( cnt == 1 ) continue; // Exactly one of True or False guards Phi 179 //} 180 if( !v->is_Call() ) { 181 /* 182 if( v->Opcode() == Op_AddP ) { 183 tty->print_cr("Phi has AddP use"); 184 } else if( v->Opcode() == Op_CastPP ) { 185 tty->print_cr("Phi has CastPP use"); 186 } else if( v->Opcode() == Op_CastII ) { 187 tty->print_cr("Phi has CastII use"); 188 } else { 189 tty->print_cr("Phi has use I cant be bothered with"); 190 } 191 */ 192 } 193 return NULL; 194 195 /* CNC - Cut out all the fancy acceptance tests 196 // Can we clone this use when doing the transformation? 197 // If all uses are from Phis at this merge or constants, then YES. 198 if( !v->in(0) && v != cmp ) { 199 tty->print_cr("Phi has free-floating use"); 200 v->dump(2); 201 return NULL; 202 } 203 for( uint l = 1; l < v->req(); l++ ) { 204 if( (!v->in(l)->is_Phi() || v->in(l)->in(0) != r) && 205 !v->in(l)->is_Con() ) { 206 tty->print_cr("Phi has use"); 207 v->dump(2); 208 return NULL; 209 } // End of if Phi-use input is neither Phi nor Constant 210 } // End of for all inputs to Phi-use 211 */ 212 } // End of for all uses of Phi 213 } // End of for all uses of Region 214 215 // Only do this if the IF node is in a sane state 216 if (iff->outcnt() != 2) 217 return NULL; 218 219 // Got a hit! Do the Mondo Hack! 220 // 221 //ABC a1c def ghi B 1 e h A C a c d f g i 222 // R - Phi - Phi - Phi Rc - Phi - Phi - Phi Rx - Phi - Phi - Phi 223 // cmp - 2 cmp - 2 cmp - 2 224 // bool bool_c bool_x 225 // if if_c if_x 226 // T F T F T F 227 // ..s.. ..t .. ..s.. ..t.. ..s.. ..t.. 228 // 229 // Split the paths coming into the merge point into 2 separate groups of 230 // merges. On the left will be all the paths feeding constants into the 231 // Cmp's Phi. On the right will be the remaining paths. The Cmp's Phi 232 // will fold up into a constant; this will let the Cmp fold up as well as 233 // all the control flow. Below the original IF we have 2 control 234 // dependent regions, 's' and 't'. Now we will merge the two paths 235 // just prior to 's' and 't' from the two IFs. At least 1 path (and quite 236 // likely 2 or more) will promptly constant fold away. 237 PhaseGVN *phase = igvn; 238 239 // Make a region merging constants and a region merging the rest 240 uint req_c = 0; 241 Node* predicate_proj = NULL; 242 int nb_predicate_proj = 0; 243 for (uint ii = 1; ii < r->req(); ii++) { 244 if (phi->in(ii) == con1) { 245 req_c++; 246 } 247 Node* proj = PhaseIdealLoop::find_predicate(r->in(ii)); 248 if (proj != NULL) { 249 nb_predicate_proj++; 250 predicate_proj = proj; 251 } 252 } 253 254 // If all the defs of the phi are the same constant, we already have the desired end state. 255 // Skip the split that would create empty phi and region nodes. 256 if((r->req() - req_c) == 1) { 257 return NULL; 258 } 259 260 if (nb_predicate_proj > 1) { 261 // Can happen in case of loop unswitching and when the loop is 262 // optimized out: it's not a loop anymore so we don't care about 263 // predicates. 264 assert(!r->is_Loop(), "this must not be a loop anymore"); 265 predicate_proj = NULL; 266 } 267 Node* predicate_c = NULL; 268 Node* predicate_x = NULL; 269 bool counted_loop = r->is_CountedLoop(); 270 271 Node *region_c = new RegionNode(req_c + 1); 272 Node *phi_c = con1; 273 uint len = r->req(); 274 Node *region_x = new RegionNode(len - req_c); 275 Node *phi_x = PhiNode::make_blank(region_x, phi); 276 for (uint i = 1, i_c = 1, i_x = 1; i < len; i++) { 277 if (phi->in(i) == con1) { 278 region_c->init_req( i_c++, r ->in(i) ); 279 if (r->in(i) == predicate_proj) 280 predicate_c = predicate_proj; 281 } else { 282 region_x->init_req( i_x, r ->in(i) ); 283 phi_x ->init_req( i_x++, phi->in(i) ); 284 if (r->in(i) == predicate_proj) 285 predicate_x = predicate_proj; 286 } 287 } 288 if (predicate_c != NULL && (req_c > 1)) { 289 assert(predicate_x == NULL, "only one predicate entry expected"); 290 predicate_c = NULL; // Do not clone predicate below merge point 291 } 292 if (predicate_x != NULL && ((len - req_c) > 2)) { 293 assert(predicate_c == NULL, "only one predicate entry expected"); 294 predicate_x = NULL; // Do not clone predicate below merge point 295 } 296 297 // Register the new RegionNodes but do not transform them. Cannot 298 // transform until the entire Region/Phi conglomerate has been hacked 299 // as a single huge transform. 300 igvn->register_new_node_with_optimizer( region_c ); 301 igvn->register_new_node_with_optimizer( region_x ); 302 // Prevent the untimely death of phi_x. Currently he has no uses. He is 303 // about to get one. If this only use goes away, then phi_x will look dead. 304 // However, he will be picking up some more uses down below. 305 Node *hook = new Node(4); 306 hook->init_req(0, phi_x); 307 hook->init_req(1, phi_c); 308 phi_x = phase->transform( phi_x ); 309 310 // Make the compare 311 Node *cmp_c = phase->makecon(t); 312 Node *cmp_x = cmp->clone(); 313 cmp_x->set_req(1,phi_x); 314 cmp_x->set_req(2,con2); 315 cmp_x = phase->transform(cmp_x); 316 // Make the bool 317 Node *b_c = phase->transform(new BoolNode(cmp_c,b->_test._test)); 318 Node *b_x = phase->transform(new BoolNode(cmp_x,b->_test._test)); 319 // Make the IfNode 320 IfNode* iff_c = iff->clone()->as_If(); 321 iff_c->set_req(0, region_c); 322 iff_c->set_req(1, b_c); 323 igvn->set_type_bottom(iff_c); 324 igvn->_worklist.push(iff_c); 325 hook->init_req(2, iff_c); 326 327 IfNode* iff_x = iff->clone()->as_If(); 328 iff_x->set_req(0, region_x); 329 iff_x->set_req(1, b_x); 330 igvn->set_type_bottom(iff_x); 331 igvn->_worklist.push(iff_x); 332 hook->init_req(3, iff_x); 333 334 // Make the true/false arms 335 Node *iff_c_t = phase->transform(new IfTrueNode (iff_c)); 336 Node *iff_c_f = phase->transform(new IfFalseNode(iff_c)); 337 if (predicate_c != NULL) { 338 assert(predicate_x == NULL, "only one predicate entry expected"); 339 // Clone loop predicates to each path 340 iff_c_t = igvn->clone_loop_predicates(predicate_c, iff_c_t, !counted_loop); 341 iff_c_f = igvn->clone_loop_predicates(predicate_c, iff_c_f, !counted_loop); 342 } 343 Node *iff_x_t = phase->transform(new IfTrueNode (iff_x)); 344 Node *iff_x_f = phase->transform(new IfFalseNode(iff_x)); 345 if (predicate_x != NULL) { 346 assert(predicate_c == NULL, "only one predicate entry expected"); 347 // Clone loop predicates to each path 348 iff_x_t = igvn->clone_loop_predicates(predicate_x, iff_x_t, !counted_loop); 349 iff_x_f = igvn->clone_loop_predicates(predicate_x, iff_x_f, !counted_loop); 350 } 351 352 // Merge the TRUE paths 353 Node *region_s = new RegionNode(3); 354 igvn->_worklist.push(region_s); 355 region_s->init_req(1, iff_c_t); 356 region_s->init_req(2, iff_x_t); 357 igvn->register_new_node_with_optimizer( region_s ); 358 359 // Merge the FALSE paths 360 Node *region_f = new RegionNode(3); 361 igvn->_worklist.push(region_f); 362 region_f->init_req(1, iff_c_f); 363 region_f->init_req(2, iff_x_f); 364 igvn->register_new_node_with_optimizer( region_f ); 365 366 igvn->hash_delete(cmp);// Remove soon-to-be-dead node from hash table. 367 cmp->set_req(1,NULL); // Whack the inputs to cmp because it will be dead 368 cmp->set_req(2,NULL); 369 // Check for all uses of the Phi and give them a new home. 370 // The 'cmp' got cloned, but CastPP/IIs need to be moved. 371 Node *phi_s = NULL; // do not construct unless needed 372 Node *phi_f = NULL; // do not construct unless needed 373 for (DUIterator_Last i2min, i2 = phi->last_outs(i2min); i2 >= i2min; --i2) { 374 Node* v = phi->last_out(i2);// User of the phi 375 igvn->rehash_node_delayed(v); // Have to fixup other Phi users 376 uint vop = v->Opcode(); 377 Node *proj = NULL; 378 if( vop == Op_Phi ) { // Remote merge point 379 Node *r = v->in(0); 380 for (uint i3 = 1; i3 < r->req(); i3++) 381 if (r->in(i3) && r->in(i3)->in(0) == iff) { 382 proj = r->in(i3); 383 break; 384 } 385 } else if( v->is_ConstraintCast() ) { 386 proj = v->in(0); // Controlling projection 387 } else { 388 assert( 0, "do not know how to handle this guy" ); 389 } 390 391 Node *proj_path_data, *proj_path_ctrl; 392 if( proj->Opcode() == Op_IfTrue ) { 393 if( phi_s == NULL ) { 394 // Only construct phi_s if needed, otherwise provides 395 // interfering use. 396 phi_s = PhiNode::make_blank(region_s,phi); 397 phi_s->init_req( 1, phi_c ); 398 phi_s->init_req( 2, phi_x ); 399 hook->add_req(phi_s); 400 phi_s = phase->transform(phi_s); 401 } 402 proj_path_data = phi_s; 403 proj_path_ctrl = region_s; 404 } else { 405 if( phi_f == NULL ) { 406 // Only construct phi_f if needed, otherwise provides 407 // interfering use. 408 phi_f = PhiNode::make_blank(region_f,phi); 409 phi_f->init_req( 1, phi_c ); 410 phi_f->init_req( 2, phi_x ); 411 hook->add_req(phi_f); 412 phi_f = phase->transform(phi_f); 413 } 414 proj_path_data = phi_f; 415 proj_path_ctrl = region_f; 416 } 417 418 // Fixup 'v' for for the split 419 if( vop == Op_Phi ) { // Remote merge point 420 uint i; 421 for( i = 1; i < v->req(); i++ ) 422 if( v->in(i) == phi ) 423 break; 424 v->set_req(i, proj_path_data ); 425 } else if( v->is_ConstraintCast() ) { 426 v->set_req(0, proj_path_ctrl ); 427 v->set_req(1, proj_path_data ); 428 } else 429 ShouldNotReachHere(); 430 } 431 432 // Now replace the original iff's True/False with region_s/region_t. 433 // This makes the original iff go dead. 434 for (DUIterator_Last i3min, i3 = iff->last_outs(i3min); i3 >= i3min; --i3) { 435 Node* p = iff->last_out(i3); 436 assert( p->Opcode() == Op_IfTrue || p->Opcode() == Op_IfFalse, "" ); 437 Node *u = (p->Opcode() == Op_IfTrue) ? region_s : region_f; 438 // Replace p with u 439 igvn->add_users_to_worklist(p); 440 for (DUIterator_Last lmin, l = p->last_outs(lmin); l >= lmin;) { 441 Node* x = p->last_out(l); 442 igvn->hash_delete(x); 443 uint uses_found = 0; 444 for( uint j = 0; j < x->req(); j++ ) { 445 if( x->in(j) == p ) { 446 x->set_req(j, u); 447 uses_found++; 448 } 449 } 450 l -= uses_found; // we deleted 1 or more copies of this edge 451 } 452 igvn->remove_dead_node(p); 453 } 454 455 // Force the original merge dead 456 igvn->hash_delete(r); 457 // First, remove region's dead users. 458 for (DUIterator_Last lmin, l = r->last_outs(lmin); l >= lmin;) { 459 Node* u = r->last_out(l); 460 if( u == r ) { 461 r->set_req(0, NULL); 462 } else { 463 assert(u->outcnt() == 0, "only dead users"); 464 igvn->remove_dead_node(u); 465 } 466 l -= 1; 467 } 468 igvn->remove_dead_node(r); 469 470 // Now remove the bogus extra edges used to keep things alive 471 igvn->remove_dead_node( hook ); 472 473 // Must return either the original node (now dead) or a new node 474 // (Do not return a top here, since that would break the uniqueness of top.) 475 return new ConINode(TypeInt::ZERO); 476 } 477 478 // if this IfNode follows a range check pattern return the projection 479 // for the failed path 480 ProjNode* IfNode::range_check_trap_proj(int& flip_test, Node*& l, Node*& r) { 481 Node* b = in(1); 482 if (b == NULL || !b->is_Bool()) return NULL; 483 BoolNode* bn = b->as_Bool(); 484 Node* cmp = bn->in(1); 485 if (cmp == NULL) return NULL; 486 if (cmp->Opcode() != Op_CmpU) return NULL; 487 488 l = cmp->in(1); 489 r = cmp->in(2); 490 flip_test = 1; 491 if (bn->_test._test == BoolTest::le) { 492 l = cmp->in(2); 493 r = cmp->in(1); 494 flip_test = 2; 495 } else if (bn->_test._test != BoolTest::lt) { 496 return NULL; 497 } 498 if (l->is_top()) return NULL; // Top input means dead test 499 if (r->Opcode() != Op_LoadRange && !is_RangeCheck()) return NULL; 500 501 // We have recognized one of these forms: 502 // Flip 1: If (Bool[<] CmpU(l, LoadRange)) ... 503 // Flip 2: If (Bool[<=] CmpU(LoadRange, l)) ... 504 505 ProjNode* iftrap = proj_out(flip_test == 2 ? true : false); 506 return iftrap; 507 } 508 509 510 //------------------------------is_range_check--------------------------------- 511 // Return 0 if not a range check. Return 1 if a range check and set index and 512 // offset. Return 2 if we had to negate the test. Index is NULL if the check 513 // is versus a constant. 514 int RangeCheckNode::is_range_check(Node* &range, Node* &index, jint &offset) { 515 int flip_test = 0; 516 Node* l = NULL; 517 Node* r = NULL; 518 ProjNode* iftrap = range_check_trap_proj(flip_test, l, r); 519 520 if (iftrap == NULL) { 521 return 0; 522 } 523 524 // Make sure it's a real range check by requiring an uncommon trap 525 // along the OOB path. Otherwise, it's possible that the user wrote 526 // something which optimized to look like a range check but behaves 527 // in some other way. 528 if (iftrap->is_uncommon_trap_proj(Deoptimization::Reason_range_check) == NULL) { 529 return 0; 530 } 531 532 // Look for index+offset form 533 Node* ind = l; 534 jint off = 0; 535 if (l->is_top()) { 536 return 0; 537 } else if (l->Opcode() == Op_AddI) { 538 if ((off = l->in(1)->find_int_con(0)) != 0) { 539 ind = l->in(2)->uncast(); 540 } else if ((off = l->in(2)->find_int_con(0)) != 0) { 541 ind = l->in(1)->uncast(); 542 } 543 } else if ((off = l->find_int_con(-1)) >= 0) { 544 // constant offset with no variable index 545 ind = NULL; 546 } else { 547 // variable index with no constant offset (or dead negative index) 548 off = 0; 549 } 550 551 // Return all the values: 552 index = ind; 553 offset = off; 554 range = r; 555 return flip_test; 556 } 557 558 //------------------------------adjust_check----------------------------------- 559 // Adjust (widen) a prior range check 560 static void adjust_check(Node* proj, Node* range, Node* index, 561 int flip, jint off_lo, PhaseIterGVN* igvn) { 562 PhaseGVN *gvn = igvn; 563 // Break apart the old check 564 Node *iff = proj->in(0); 565 Node *bol = iff->in(1); 566 if( bol->is_top() ) return; // In case a partially dead range check appears 567 // bail (or bomb[ASSERT/DEBUG]) if NOT projection-->IfNode-->BoolNode 568 DEBUG_ONLY( if( !bol->is_Bool() ) { proj->dump(3); fatal("Expect projection-->IfNode-->BoolNode"); } ) 569 if( !bol->is_Bool() ) return; 570 571 Node *cmp = bol->in(1); 572 // Compute a new check 573 Node *new_add = gvn->intcon(off_lo); 574 if( index ) { 575 new_add = off_lo ? gvn->transform(new AddINode( index, new_add )) : index; 576 } 577 Node *new_cmp = (flip == 1) 578 ? new CmpUNode( new_add, range ) 579 : new CmpUNode( range, new_add ); 580 new_cmp = gvn->transform(new_cmp); 581 // See if no need to adjust the existing check 582 if( new_cmp == cmp ) return; 583 // Else, adjust existing check 584 Node *new_bol = gvn->transform( new BoolNode( new_cmp, bol->as_Bool()->_test._test ) ); 585 igvn->rehash_node_delayed( iff ); 586 iff->set_req_X( 1, new_bol, igvn ); 587 } 588 589 //------------------------------up_one_dom------------------------------------- 590 // Walk up the dominator tree one step. Return NULL at root or true 591 // complex merges. Skips through small diamonds. 592 Node* IfNode::up_one_dom(Node *curr, bool linear_only) { 593 Node *dom = curr->in(0); 594 if( !dom ) // Found a Region degraded to a copy? 595 return curr->nonnull_req(); // Skip thru it 596 597 if( curr != dom ) // Normal walk up one step? 598 return dom; 599 600 // Use linear_only if we are still parsing, since we cannot 601 // trust the regions to be fully filled in. 602 if (linear_only) 603 return NULL; 604 605 if( dom->is_Root() ) 606 return NULL; 607 608 // Else hit a Region. Check for a loop header 609 if( dom->is_Loop() ) 610 return dom->in(1); // Skip up thru loops 611 612 // Check for small diamonds 613 Node *din1, *din2, *din3, *din4; 614 if( dom->req() == 3 && // 2-path merge point 615 (din1 = dom ->in(1)) && // Left path exists 616 (din2 = dom ->in(2)) && // Right path exists 617 (din3 = din1->in(0)) && // Left path up one 618 (din4 = din2->in(0)) ) { // Right path up one 619 if( din3->is_Call() && // Handle a slow-path call on either arm 620 (din3 = din3->in(0)) ) 621 din3 = din3->in(0); 622 if( din4->is_Call() && // Handle a slow-path call on either arm 623 (din4 = din4->in(0)) ) 624 din4 = din4->in(0); 625 if( din3 == din4 && din3->is_If() ) 626 return din3; // Skip around diamonds 627 } 628 629 // Give up the search at true merges 630 return NULL; // Dead loop? Or hit root? 631 } 632 633 634 //------------------------------filtered_int_type-------------------------------- 635 // Return a possibly more restrictive type for val based on condition control flow for an if 636 const TypeInt* IfNode::filtered_int_type(PhaseGVN* gvn, Node *val, Node* if_proj) { 637 assert(if_proj && 638 (if_proj->Opcode() == Op_IfTrue || if_proj->Opcode() == Op_IfFalse), "expecting an if projection"); 639 if (if_proj->in(0) && if_proj->in(0)->is_If()) { 640 IfNode* iff = if_proj->in(0)->as_If(); 641 if (iff->in(1) && iff->in(1)->is_Bool()) { 642 BoolNode* bol = iff->in(1)->as_Bool(); 643 if (bol->in(1) && bol->in(1)->is_Cmp()) { 644 const CmpNode* cmp = bol->in(1)->as_Cmp(); 645 if (cmp->in(1) == val) { 646 const TypeInt* cmp2_t = gvn->type(cmp->in(2))->isa_int(); 647 if (cmp2_t != NULL) { 648 jint lo = cmp2_t->_lo; 649 jint hi = cmp2_t->_hi; 650 BoolTest::mask msk = if_proj->Opcode() == Op_IfTrue ? bol->_test._test : bol->_test.negate(); 651 switch (msk) { 652 case BoolTest::ne: 653 // Can't refine type 654 return NULL; 655 case BoolTest::eq: 656 return cmp2_t; 657 case BoolTest::lt: 658 lo = TypeInt::INT->_lo; 659 if (hi - 1 < hi) { 660 hi = hi - 1; 661 } 662 break; 663 case BoolTest::le: 664 lo = TypeInt::INT->_lo; 665 break; 666 case BoolTest::gt: 667 if (lo + 1 > lo) { 668 lo = lo + 1; 669 } 670 hi = TypeInt::INT->_hi; 671 break; 672 case BoolTest::ge: 673 // lo unchanged 674 hi = TypeInt::INT->_hi; 675 break; 676 default: 677 break; 678 } 679 const TypeInt* rtn_t = TypeInt::make(lo, hi, cmp2_t->_widen); 680 return rtn_t; 681 } 682 } 683 } 684 } 685 } 686 return NULL; 687 } 688 689 //------------------------------fold_compares---------------------------- 690 // See if a pair of CmpIs can be converted into a CmpU. In some cases 691 // the direction of this if is determined by the preceding if so it 692 // can be eliminate entirely. 693 // 694 // Given an if testing (CmpI n v) check for an immediately control 695 // dependent if that is testing (CmpI n v2) and has one projection 696 // leading to this if and the other projection leading to a region 697 // that merges one of this ifs control projections. 698 // 699 // If 700 // / | 701 // / | 702 // / | 703 // If | 704 // /\ | 705 // / \ | 706 // / \ | 707 // / Region 708 // 709 // Or given an if testing (CmpI n v) check for a dominating if that is 710 // testing (CmpI n v2), both having one projection leading to an 711 // uncommon trap. Allow Another independent guard in between to cover 712 // an explicit range check: 713 // if (index < 0 || index >= array.length) { 714 // which may need a null check to guard the LoadRange 715 // 716 // If 717 // / \ 718 // / \ 719 // / \ 720 // If unc 721 // /\ 722 // / \ 723 // / \ 724 // / unc 725 // 726 727 // Is the comparison for this If suitable for folding? 728 bool IfNode::cmpi_folds(PhaseIterGVN* igvn) { 729 return in(1) != NULL && 730 in(1)->is_Bool() && 731 in(1)->in(1) != NULL && 732 in(1)->in(1)->Opcode() == Op_CmpI && 733 in(1)->in(1)->in(2) != NULL && 734 in(1)->in(1)->in(2) != igvn->C->top() && 735 (in(1)->as_Bool()->_test.is_less() || 736 in(1)->as_Bool()->_test.is_greater()); 737 } 738 739 // Is a dominating control suitable for folding with this if? 740 bool IfNode::is_ctrl_folds(Node* ctrl, PhaseIterGVN* igvn) { 741 return ctrl != NULL && 742 ctrl->is_Proj() && 743 ctrl->in(0) != NULL && 744 ctrl->in(0)->Opcode() == Op_If && 745 ctrl->in(0)->outcnt() == 2 && 746 ctrl->in(0)->as_If()->cmpi_folds(igvn) && 747 // Must compare same value 748 ctrl->in(0)->in(1)->in(1)->in(1) != NULL && 749 ctrl->in(0)->in(1)->in(1)->in(1) == in(1)->in(1)->in(1); 750 } 751 752 // Do this If and the dominating If share a region? 753 bool IfNode::has_shared_region(ProjNode* proj, ProjNode*& success, ProjNode*& fail) { 754 ProjNode* otherproj = proj->other_if_proj(); 755 Node* otherproj_ctrl_use = otherproj->unique_ctrl_out(); 756 RegionNode* region = (otherproj_ctrl_use != NULL && otherproj_ctrl_use->is_Region()) ? otherproj_ctrl_use->as_Region() : NULL; 757 success = NULL; 758 fail = NULL; 759 760 if (otherproj->outcnt() == 1 && region != NULL && !region->has_phi()) { 761 for (int i = 0; i < 2; i++) { 762 ProjNode* proj = proj_out(i); 763 if (success == NULL && proj->outcnt() == 1 && proj->unique_out() == region) { 764 success = proj; 765 } else if (fail == NULL) { 766 fail = proj; 767 } else { 768 success = fail = NULL; 769 } 770 } 771 } 772 return success != NULL && fail != NULL; 773 } 774 775 // Return projection that leads to an uncommon trap if any 776 ProjNode* IfNode::uncommon_trap_proj(CallStaticJavaNode*& call) const { 777 for (int i = 0; i < 2; i++) { 778 call = proj_out(i)->is_uncommon_trap_proj(Deoptimization::Reason_none); 779 if (call != NULL) { 780 return proj_out(i); 781 } 782 } 783 return NULL; 784 } 785 786 // Do this If and the dominating If both branch out to an uncommon trap 787 bool IfNode::has_only_uncommon_traps(ProjNode* proj, ProjNode*& success, ProjNode*& fail, PhaseIterGVN* igvn) { 788 ProjNode* otherproj = proj->other_if_proj(); 789 CallStaticJavaNode* dom_unc = otherproj->is_uncommon_trap_proj(Deoptimization::Reason_none); 790 791 if (otherproj->outcnt() == 1 && dom_unc != NULL) { 792 // We need to re-execute the folded Ifs after deoptimization from the merged traps 793 if (!dom_unc->jvms()->should_reexecute()) { 794 return false; 795 } 796 797 CallStaticJavaNode* unc = NULL; 798 ProjNode* unc_proj = uncommon_trap_proj(unc); 799 if (unc_proj != NULL && unc_proj->outcnt() == 1) { 800 if (dom_unc == unc) { 801 // Allow the uncommon trap to be shared through a region 802 RegionNode* r = unc->in(0)->as_Region(); 803 if (r->outcnt() != 2 || r->req() != 3 || r->find_edge(otherproj) == -1 || r->find_edge(unc_proj) == -1) { 804 return false; 805 } 806 assert(r->has_phi() == NULL, "simple region shouldn't have a phi"); 807 } else if (dom_unc->in(0) != otherproj || unc->in(0) != unc_proj) { 808 return false; 809 } 810 811 // Different methods and methods containing jsrs are not supported. 812 ciMethod* method = unc->jvms()->method(); 813 ciMethod* dom_method = dom_unc->jvms()->method(); 814 if (method != dom_method || method->has_jsrs()) { 815 return false; 816 } 817 // Check that both traps are in the same activation of the method (instead 818 // of two activations being inlined through different call sites) by verifying 819 // that the call stacks are equal for both JVMStates. 820 JVMState* dom_caller = dom_unc->jvms()->caller(); 821 JVMState* caller = unc->jvms()->caller(); 822 if ((dom_caller == NULL) != (caller == NULL)) { 823 // The current method must either be inlined into both dom_caller and 824 // caller or must not be inlined at all (top method). Bail out otherwise. 825 return false; 826 } else if (dom_caller != NULL && !dom_caller->same_calls_as(caller)) { 827 return false; 828 } 829 // Check that the bci of the dominating uncommon trap dominates the bci 830 // of the dominated uncommon trap. Otherwise we may not re-execute 831 // the dominated check after deoptimization from the merged uncommon trap. 832 ciTypeFlow* flow = dom_method->get_flow_analysis(); 833 int bci = unc->jvms()->bci(); 834 int dom_bci = dom_unc->jvms()->bci(); 835 if (!flow->is_dominated_by(bci, dom_bci)) { 836 return false; 837 } 838 839 // See merge_uncommon_traps: the reason of the uncommon trap 840 // will be changed and the state of the dominating If will be 841 // used. Checked that we didn't apply this transformation in a 842 // previous compilation and it didn't cause too many traps 843 if (!igvn->C->too_many_traps(dom_method, dom_bci, Deoptimization::Reason_unstable_fused_if) && 844 !igvn->C->too_many_traps(dom_method, dom_bci, Deoptimization::Reason_range_check)) { 845 success = unc_proj; 846 fail = unc_proj->other_if_proj(); 847 return true; 848 } 849 } 850 } 851 return false; 852 } 853 854 // Check that the 2 CmpI can be folded into as single CmpU and proceed with the folding 855 bool IfNode::fold_compares_helper(ProjNode* proj, ProjNode* success, ProjNode* fail, PhaseIterGVN* igvn) { 856 Node* this_cmp = in(1)->in(1); 857 BoolNode* this_bool = in(1)->as_Bool(); 858 IfNode* dom_iff = proj->in(0)->as_If(); 859 BoolNode* dom_bool = dom_iff->in(1)->as_Bool(); 860 Node* lo = dom_iff->in(1)->in(1)->in(2); 861 Node* hi = this_cmp->in(2); 862 Node* n = this_cmp->in(1); 863 ProjNode* otherproj = proj->other_if_proj(); 864 865 const TypeInt* lo_type = IfNode::filtered_int_type(igvn, n, otherproj); 866 const TypeInt* hi_type = IfNode::filtered_int_type(igvn, n, success); 867 868 BoolTest::mask lo_test = dom_bool->_test._test; 869 BoolTest::mask hi_test = this_bool->_test._test; 870 BoolTest::mask cond = hi_test; 871 872 // convert: 873 // 874 // dom_bool = x {<,<=,>,>=} a 875 // / \ 876 // proj = {True,False} / \ otherproj = {False,True} 877 // / 878 // this_bool = x {<,<=} b 879 // / \ 880 // fail = {True,False} / \ success = {False,True} 881 // / 882 // 883 // (Second test guaranteed canonicalized, first one may not have 884 // been canonicalized yet) 885 // 886 // into: 887 // 888 // cond = (x - lo) {<u,<=u,>u,>=u} adjusted_lim 889 // / \ 890 // fail / \ success 891 // / 892 // 893 894 // Figure out which of the two tests sets the upper bound and which 895 // sets the lower bound if any. 896 Node* adjusted_lim = NULL; 897 if (hi_type->_lo > lo_type->_hi && hi_type->_hi == max_jint && lo_type->_lo == min_jint) { 898 assert((dom_bool->_test.is_less() && !proj->_con) || 899 (dom_bool->_test.is_greater() && proj->_con), "incorrect test"); 900 // this test was canonicalized 901 assert(this_bool->_test.is_less() && fail->_con, "incorrect test"); 902 903 // this_bool = < 904 // dom_bool = >= (proj = True) or dom_bool = < (proj = False) 905 // x in [a, b[ on the fail (= True) projection, b > a-1 (because of hi_type->_lo > lo_type->_hi test above): 906 // lo = a, hi = b, adjusted_lim = b-a, cond = <u 907 // dom_bool = > (proj = True) or dom_bool = <= (proj = False) 908 // x in ]a, b[ on the fail (= True) projection, b > a: 909 // lo = a+1, hi = b, adjusted_lim = b-a-1, cond = <u 910 // this_bool = <= 911 // dom_bool = >= (proj = True) or dom_bool = < (proj = False) 912 // x in [a, b] on the fail (= True) projection, b+1 > a-1: 913 // lo = a, hi = b, adjusted_lim = b-a+1, cond = <u 914 // lo = a, hi = b, adjusted_lim = b-a, cond = <=u doesn't work because b = a - 1 is possible, then b-a = -1 915 // dom_bool = > (proj = True) or dom_bool = <= (proj = False) 916 // x in ]a, b] on the fail (= True) projection b+1 > a: 917 // lo = a+1, hi = b, adjusted_lim = b-a, cond = <u 918 // lo = a+1, hi = b, adjusted_lim = b-a-1, cond = <=u doesn't work because a = b is possible, then b-a-1 = -1 919 920 if (hi_test == BoolTest::lt) { 921 if (lo_test == BoolTest::gt || lo_test == BoolTest::le) { 922 lo = igvn->transform(new AddINode(lo, igvn->intcon(1))); 923 } 924 } else { 925 assert(hi_test == BoolTest::le, "bad test"); 926 if (lo_test == BoolTest::ge || lo_test == BoolTest::lt) { 927 adjusted_lim = igvn->transform(new SubINode(hi, lo)); 928 adjusted_lim = igvn->transform(new AddINode(adjusted_lim, igvn->intcon(1))); 929 cond = BoolTest::lt; 930 } else { 931 assert(lo_test == BoolTest::gt || lo_test == BoolTest::le, "bad test"); 932 adjusted_lim = igvn->transform(new SubINode(hi, lo)); 933 lo = igvn->transform(new AddINode(lo, igvn->intcon(1))); 934 cond = BoolTest::lt; 935 } 936 } 937 } else if (lo_type->_lo > hi_type->_hi && lo_type->_hi == max_jint && hi_type->_lo == min_jint) { 938 939 // this_bool = < 940 // dom_bool = < (proj = True) or dom_bool = >= (proj = False) 941 // x in [b, a[ on the fail (= False) projection, a > b-1 (because of lo_type->_lo > hi_type->_hi above): 942 // lo = b, hi = a, adjusted_lim = a-b, cond = >=u 943 // dom_bool = <= (proj = True) or dom_bool = > (proj = False) 944 // x in [b, a] on the fail (= False) projection, a+1 > b-1: 945 // lo = b, hi = a, adjusted_lim = a-b+1, cond = >=u 946 // lo = b, hi = a, adjusted_lim = a-b, cond = >u doesn't work because a = b - 1 is possible, then b-a = -1 947 // this_bool = <= 948 // dom_bool = < (proj = True) or dom_bool = >= (proj = False) 949 // x in ]b, a[ on the fail (= False) projection, a > b: 950 // lo = b+1, hi = a, adjusted_lim = a-b-1, cond = >=u 951 // dom_bool = <= (proj = True) or dom_bool = > (proj = False) 952 // x in ]b, a] on the fail (= False) projection, a+1 > b: 953 // lo = b+1, hi = a, adjusted_lim = a-b, cond = >=u 954 // lo = b+1, hi = a, adjusted_lim = a-b-1, cond = >u doesn't work because a = b is possible, then b-a-1 = -1 955 956 swap(lo, hi); 957 swap(lo_type, hi_type); 958 swap(lo_test, hi_test); 959 960 assert((dom_bool->_test.is_less() && proj->_con) || 961 (dom_bool->_test.is_greater() && !proj->_con), "incorrect test"); 962 // this test was canonicalized 963 assert(this_bool->_test.is_less() && !fail->_con, "incorrect test"); 964 965 cond = (hi_test == BoolTest::le || hi_test == BoolTest::gt) ? BoolTest::gt : BoolTest::ge; 966 967 if (lo_test == BoolTest::lt) { 968 if (hi_test == BoolTest::lt || hi_test == BoolTest::ge) { 969 cond = BoolTest::ge; 970 } else { 971 assert(hi_test == BoolTest::le || hi_test == BoolTest::gt, "bad test"); 972 adjusted_lim = igvn->transform(new SubINode(hi, lo)); 973 adjusted_lim = igvn->transform(new AddINode(adjusted_lim, igvn->intcon(1))); 974 cond = BoolTest::ge; 975 } 976 } else if (lo_test == BoolTest::le) { 977 if (hi_test == BoolTest::lt || hi_test == BoolTest::ge) { 978 lo = igvn->transform(new AddINode(lo, igvn->intcon(1))); 979 cond = BoolTest::ge; 980 } else { 981 assert(hi_test == BoolTest::le || hi_test == BoolTest::gt, "bad test"); 982 adjusted_lim = igvn->transform(new SubINode(hi, lo)); 983 lo = igvn->transform(new AddINode(lo, igvn->intcon(1))); 984 cond = BoolTest::ge; 985 } 986 } 987 } else { 988 const TypeInt* failtype = filtered_int_type(igvn, n, proj); 989 if (failtype != NULL) { 990 const TypeInt* type2 = filtered_int_type(igvn, n, fail); 991 if (type2 != NULL) { 992 failtype = failtype->join(type2)->is_int(); 993 if (failtype->_lo > failtype->_hi) { 994 // previous if determines the result of this if so 995 // replace Bool with constant 996 igvn->_worklist.push(in(1)); 997 igvn->replace_input_of(this, 1, igvn->intcon(success->_con)); 998 return true; 999 } 1000 } 1001 } 1002 lo = NULL; 1003 hi = NULL; 1004 } 1005 1006 if (lo && hi) { 1007 // Merge the two compares into a single unsigned compare by building (CmpU (n - lo) (hi - lo)) 1008 Node* adjusted_val = igvn->transform(new SubINode(n, lo)); 1009 if (adjusted_lim == NULL) { 1010 adjusted_lim = igvn->transform(new SubINode(hi, lo)); 1011 } 1012 Node* newcmp = igvn->transform(new CmpUNode(adjusted_val, adjusted_lim)); 1013 Node* newbool = igvn->transform(new BoolNode(newcmp, cond)); 1014 1015 igvn->replace_input_of(dom_iff, 1, igvn->intcon(proj->_con)); 1016 igvn->_worklist.push(in(1)); 1017 igvn->replace_input_of(this, 1, newbool); 1018 1019 return true; 1020 } 1021 return false; 1022 } 1023 1024 // Merge the branches that trap for this If and the dominating If into 1025 // a single region that branches to the uncommon trap for the 1026 // dominating If 1027 Node* IfNode::merge_uncommon_traps(ProjNode* proj, ProjNode* success, ProjNode* fail, PhaseIterGVN* igvn) { 1028 Node* res = this; 1029 assert(success->in(0) == this, "bad projection"); 1030 1031 ProjNode* otherproj = proj->other_if_proj(); 1032 1033 CallStaticJavaNode* unc = success->is_uncommon_trap_proj(Deoptimization::Reason_none); 1034 CallStaticJavaNode* dom_unc = otherproj->is_uncommon_trap_proj(Deoptimization::Reason_none); 1035 1036 if (unc != dom_unc) { 1037 Node* r = new RegionNode(3); 1038 1039 r->set_req(1, otherproj); 1040 r->set_req(2, success); 1041 r = igvn->transform(r); 1042 assert(r->is_Region(), "can't go away"); 1043 1044 // Make both If trap at the state of the first If: once the CmpI 1045 // nodes are merged, if we trap we don't know which of the CmpI 1046 // nodes would have caused the trap so we have to restart 1047 // execution at the first one 1048 igvn->replace_input_of(dom_unc, 0, r); 1049 igvn->replace_input_of(unc, 0, igvn->C->top()); 1050 } 1051 int trap_request = dom_unc->uncommon_trap_request(); 1052 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 1053 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 1054 1055 int flip_test = 0; 1056 Node* l = NULL; 1057 Node* r = NULL; 1058 1059 if (success->in(0)->as_If()->range_check_trap_proj(flip_test, l, r) != NULL) { 1060 // If this looks like a range check, change the trap to 1061 // Reason_range_check so the compiler recognizes it as a range 1062 // check and applies the corresponding optimizations 1063 trap_request = Deoptimization::make_trap_request(Deoptimization::Reason_range_check, action); 1064 1065 improve_address_types(l, r, fail, igvn); 1066 1067 res = igvn->transform(new RangeCheckNode(in(0), in(1), _prob, _fcnt)); 1068 } else if (unc != dom_unc) { 1069 // If we trap we won't know what CmpI would have caused the trap 1070 // so use a special trap reason to mark this pair of CmpI nodes as 1071 // bad candidate for folding. On recompilation we won't fold them 1072 // and we may trap again but this time we'll know what branch 1073 // traps 1074 trap_request = Deoptimization::make_trap_request(Deoptimization::Reason_unstable_fused_if, action); 1075 } 1076 igvn->replace_input_of(dom_unc, TypeFunc::Parms, igvn->intcon(trap_request)); 1077 return res; 1078 } 1079 1080 // If we are turning 2 CmpI nodes into a CmpU that follows the pattern 1081 // of a rangecheck on index i, on 64 bit the compares may be followed 1082 // by memory accesses using i as index. In that case, the CmpU tells 1083 // us something about the values taken by i that can help the compiler 1084 // (see Compile::conv_I2X_index()) 1085 void IfNode::improve_address_types(Node* l, Node* r, ProjNode* fail, PhaseIterGVN* igvn) { 1086 #ifdef _LP64 1087 ResourceMark rm; 1088 Node_Stack stack(2); 1089 1090 assert(r->Opcode() == Op_LoadRange, "unexpected range check"); 1091 const TypeInt* array_size = igvn->type(r)->is_int(); 1092 1093 stack.push(l, 0); 1094 1095 while(stack.size() > 0) { 1096 Node* n = stack.node(); 1097 uint start = stack.index(); 1098 1099 uint i = start; 1100 for (; i < n->outcnt(); i++) { 1101 Node* use = n->raw_out(i); 1102 if (stack.size() == 1) { 1103 if (use->Opcode() == Op_ConvI2L) { 1104 const TypeLong* bounds = use->as_Type()->type()->is_long(); 1105 if (bounds->_lo <= array_size->_lo && bounds->_hi >= array_size->_hi && 1106 (bounds->_lo != array_size->_lo || bounds->_hi != array_size->_hi)) { 1107 stack.set_index(i+1); 1108 stack.push(use, 0); 1109 break; 1110 } 1111 } 1112 } else if (use->is_Mem()) { 1113 Node* ctrl = use->in(0); 1114 for (int i = 0; i < 10 && ctrl != NULL && ctrl != fail; i++) { 1115 ctrl = up_one_dom(ctrl); 1116 } 1117 if (ctrl == fail) { 1118 Node* init_n = stack.node_at(1); 1119 assert(init_n->Opcode() == Op_ConvI2L, "unexpected first node"); 1120 // Create a new narrow ConvI2L node that is dependent on the range check 1121 Node* new_n = igvn->C->conv_I2X_index(igvn, l, array_size, fail); 1122 1123 // The type of the ConvI2L may be widen and so the new 1124 // ConvI2L may not be better than an existing ConvI2L 1125 if (new_n != init_n) { 1126 for (uint j = 2; j < stack.size(); j++) { 1127 Node* n = stack.node_at(j); 1128 Node* clone = n->clone(); 1129 int rep = clone->replace_edge(init_n, new_n); 1130 assert(rep > 0, "can't find expected node?"); 1131 clone = igvn->transform(clone); 1132 init_n = n; 1133 new_n = clone; 1134 } 1135 igvn->hash_delete(use); 1136 int rep = use->replace_edge(init_n, new_n); 1137 assert(rep > 0, "can't find expected node?"); 1138 igvn->transform(use); 1139 if (init_n->outcnt() == 0) { 1140 igvn->_worklist.push(init_n); 1141 } 1142 } 1143 } 1144 } else if (use->in(0) == NULL && (igvn->type(use)->isa_long() || 1145 igvn->type(use)->isa_ptr())) { 1146 stack.set_index(i+1); 1147 stack.push(use, 0); 1148 break; 1149 } 1150 } 1151 if (i == n->outcnt()) { 1152 stack.pop(); 1153 } 1154 } 1155 #endif 1156 } 1157 1158 bool IfNode::is_cmp_with_loadrange(ProjNode* proj) { 1159 if (in(1) != NULL && 1160 in(1)->in(1) != NULL && 1161 in(1)->in(1)->in(2) != NULL) { 1162 Node* other = in(1)->in(1)->in(2); 1163 if (other->Opcode() == Op_LoadRange && 1164 ((other->in(0) != NULL && other->in(0) == proj) || 1165 (other->in(0) == NULL && 1166 other->in(2) != NULL && 1167 other->in(2)->is_AddP() && 1168 other->in(2)->in(1) != NULL && 1169 other->in(2)->in(1)->Opcode() == Op_CastPP && 1170 other->in(2)->in(1)->in(0) == proj))) { 1171 return true; 1172 } 1173 } 1174 return false; 1175 } 1176 1177 bool IfNode::is_null_check(ProjNode* proj, PhaseIterGVN* igvn) { 1178 Node* other = in(1)->in(1)->in(2); 1179 if (other->in(MemNode::Address) != NULL && 1180 proj->in(0)->in(1) != NULL && 1181 proj->in(0)->in(1)->is_Bool() && 1182 proj->in(0)->in(1)->in(1) != NULL && 1183 proj->in(0)->in(1)->in(1)->Opcode() == Op_CmpP && 1184 proj->in(0)->in(1)->in(1)->in(2) != NULL && 1185 proj->in(0)->in(1)->in(1)->in(1) == other->in(MemNode::Address)->in(AddPNode::Address)->uncast() && 1186 igvn->type(proj->in(0)->in(1)->in(1)->in(2)) == TypePtr::NULL_PTR) { 1187 return true; 1188 } 1189 return false; 1190 } 1191 1192 // Check that the If that is in between the 2 integer comparisons has 1193 // no side effect 1194 bool IfNode::is_side_effect_free_test(ProjNode* proj, PhaseIterGVN* igvn) { 1195 if (proj != NULL && 1196 proj->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) && 1197 proj->outcnt() <= 2) { 1198 if (proj->outcnt() == 1 || 1199 // Allow simple null check from LoadRange 1200 (is_cmp_with_loadrange(proj) && is_null_check(proj, igvn))) { 1201 CallStaticJavaNode* unc = proj->is_uncommon_trap_if_pattern(Deoptimization::Reason_none); 1202 CallStaticJavaNode* dom_unc = proj->in(0)->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none); 1203 1204 // reroute_side_effect_free_unc changes the state of this 1205 // uncommon trap to restart execution at the previous 1206 // CmpI. Check that this change in a previous compilation didn't 1207 // cause too many traps. 1208 int trap_request = unc->uncommon_trap_request(); 1209 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 1210 1211 if (igvn->C->too_many_traps(dom_unc->jvms()->method(), dom_unc->jvms()->bci(), reason)) { 1212 return false; 1213 } 1214 1215 return true; 1216 } 1217 } 1218 return false; 1219 } 1220 1221 // Make the If between the 2 integer comparisons trap at the state of 1222 // the first If: the last CmpI is the one replaced by a CmpU and the 1223 // first CmpI is eliminated, so the test between the 2 CmpI nodes 1224 // won't be guarded by the first CmpI anymore. It can trap in cases 1225 // where the first CmpI would have prevented it from executing: on a 1226 // trap, we need to restart execution at the state of the first CmpI 1227 void IfNode::reroute_side_effect_free_unc(ProjNode* proj, ProjNode* dom_proj, PhaseIterGVN* igvn) { 1228 CallStaticJavaNode* dom_unc = dom_proj->is_uncommon_trap_if_pattern(Deoptimization::Reason_none); 1229 ProjNode* otherproj = proj->other_if_proj(); 1230 CallStaticJavaNode* unc = proj->is_uncommon_trap_if_pattern(Deoptimization::Reason_none); 1231 Node* call_proj = dom_unc->unique_ctrl_out(); 1232 Node* halt = call_proj->unique_ctrl_out(); 1233 1234 Node* new_unc = dom_unc->clone(); 1235 call_proj = call_proj->clone(); 1236 halt = halt->clone(); 1237 Node* c = otherproj->clone(); 1238 1239 c = igvn->transform(c); 1240 new_unc->set_req(TypeFunc::Parms, unc->in(TypeFunc::Parms)); 1241 new_unc->set_req(0, c); 1242 new_unc = igvn->transform(new_unc); 1243 call_proj->set_req(0, new_unc); 1244 call_proj = igvn->transform(call_proj); 1245 halt->set_req(0, call_proj); 1246 halt = igvn->transform(halt); 1247 1248 igvn->replace_node(otherproj, igvn->C->top()); 1249 igvn->C->root()->add_req(halt); 1250 } 1251 1252 Node* IfNode::fold_compares(PhaseIterGVN* igvn) { 1253 if (Opcode() != Op_If) return NULL; 1254 1255 if (cmpi_folds(igvn)) { 1256 Node* ctrl = in(0); 1257 if (is_ctrl_folds(ctrl, igvn) && 1258 ctrl->outcnt() == 1) { 1259 // A integer comparison immediately dominated by another integer 1260 // comparison 1261 ProjNode* success = NULL; 1262 ProjNode* fail = NULL; 1263 ProjNode* dom_cmp = ctrl->as_Proj(); 1264 if (has_shared_region(dom_cmp, success, fail) && 1265 // Next call modifies graph so must be last 1266 fold_compares_helper(dom_cmp, success, fail, igvn)) { 1267 return this; 1268 } 1269 if (has_only_uncommon_traps(dom_cmp, success, fail, igvn) && 1270 // Next call modifies graph so must be last 1271 fold_compares_helper(dom_cmp, success, fail, igvn)) { 1272 return merge_uncommon_traps(dom_cmp, success, fail, igvn); 1273 } 1274 return NULL; 1275 } else if (ctrl->in(0) != NULL && 1276 ctrl->in(0)->in(0) != NULL) { 1277 ProjNode* success = NULL; 1278 ProjNode* fail = NULL; 1279 Node* dom = ctrl->in(0)->in(0); 1280 ProjNode* dom_cmp = dom->isa_Proj(); 1281 ProjNode* other_cmp = ctrl->isa_Proj(); 1282 1283 // Check if it's an integer comparison dominated by another 1284 // integer comparison with another test in between 1285 if (is_ctrl_folds(dom, igvn) && 1286 has_only_uncommon_traps(dom_cmp, success, fail, igvn) && 1287 is_side_effect_free_test(other_cmp, igvn) && 1288 // Next call modifies graph so must be last 1289 fold_compares_helper(dom_cmp, success, fail, igvn)) { 1290 reroute_side_effect_free_unc(other_cmp, dom_cmp, igvn); 1291 return merge_uncommon_traps(dom_cmp, success, fail, igvn); 1292 } 1293 } 1294 } 1295 return NULL; 1296 } 1297 1298 //------------------------------remove_useless_bool---------------------------- 1299 // Check for people making a useless boolean: things like 1300 // if( (x < y ? true : false) ) { ... } 1301 // Replace with if( x < y ) { ... } 1302 static Node *remove_useless_bool(IfNode *iff, PhaseGVN *phase) { 1303 Node *i1 = iff->in(1); 1304 if( !i1->is_Bool() ) return NULL; 1305 BoolNode *bol = i1->as_Bool(); 1306 1307 Node *cmp = bol->in(1); 1308 if( cmp->Opcode() != Op_CmpI ) return NULL; 1309 1310 // Must be comparing against a bool 1311 const Type *cmp2_t = phase->type( cmp->in(2) ); 1312 if( cmp2_t != TypeInt::ZERO && 1313 cmp2_t != TypeInt::ONE ) 1314 return NULL; 1315 1316 // Find a prior merge point merging the boolean 1317 i1 = cmp->in(1); 1318 if( !i1->is_Phi() ) return NULL; 1319 PhiNode *phi = i1->as_Phi(); 1320 if( phase->type( phi ) != TypeInt::BOOL ) 1321 return NULL; 1322 1323 // Check for diamond pattern 1324 int true_path = phi->is_diamond_phi(); 1325 if( true_path == 0 ) return NULL; 1326 1327 // Make sure that iff and the control of the phi are different. This 1328 // should really only happen for dead control flow since it requires 1329 // an illegal cycle. 1330 if (phi->in(0)->in(1)->in(0) == iff) return NULL; 1331 1332 // phi->region->if_proj->ifnode->bool->cmp 1333 BoolNode *bol2 = phi->in(0)->in(1)->in(0)->in(1)->as_Bool(); 1334 1335 // Now get the 'sense' of the test correct so we can plug in 1336 // either iff2->in(1) or its complement. 1337 int flip = 0; 1338 if( bol->_test._test == BoolTest::ne ) flip = 1-flip; 1339 else if( bol->_test._test != BoolTest::eq ) return NULL; 1340 if( cmp2_t == TypeInt::ZERO ) flip = 1-flip; 1341 1342 const Type *phi1_t = phase->type( phi->in(1) ); 1343 const Type *phi2_t = phase->type( phi->in(2) ); 1344 // Check for Phi(0,1) and flip 1345 if( phi1_t == TypeInt::ZERO ) { 1346 if( phi2_t != TypeInt::ONE ) return NULL; 1347 flip = 1-flip; 1348 } else { 1349 // Check for Phi(1,0) 1350 if( phi1_t != TypeInt::ONE ) return NULL; 1351 if( phi2_t != TypeInt::ZERO ) return NULL; 1352 } 1353 if( true_path == 2 ) { 1354 flip = 1-flip; 1355 } 1356 1357 Node* new_bol = (flip ? phase->transform( bol2->negate(phase) ) : bol2); 1358 assert(new_bol != iff->in(1), "must make progress"); 1359 iff->set_req(1, new_bol); 1360 // Intervening diamond probably goes dead 1361 phase->C->set_major_progress(); 1362 return iff; 1363 } 1364 1365 static IfNode* idealize_test(PhaseGVN* phase, IfNode* iff); 1366 1367 struct RangeCheck { 1368 Node* ctl; 1369 jint off; 1370 }; 1371 1372 Node* IfNode::Ideal_common(PhaseGVN *phase, bool can_reshape) { 1373 if (remove_dead_region(phase, can_reshape)) return this; 1374 // No Def-Use info? 1375 if (!can_reshape) return NULL; 1376 1377 // Don't bother trying to transform a dead if 1378 if (in(0)->is_top()) return NULL; 1379 // Don't bother trying to transform an if with a dead test 1380 if (in(1)->is_top()) return NULL; 1381 // Another variation of a dead test 1382 if (in(1)->is_Con()) return NULL; 1383 // Another variation of a dead if 1384 if (outcnt() < 2) return NULL; 1385 1386 // Canonicalize the test. 1387 Node* idt_if = idealize_test(phase, this); 1388 if (idt_if != NULL) return idt_if; 1389 1390 // Try to split the IF 1391 PhaseIterGVN *igvn = phase->is_IterGVN(); 1392 Node *s = split_if(this, igvn); 1393 if (s != NULL) return s; 1394 1395 return NodeSentinel; 1396 } 1397 1398 //------------------------------Ideal------------------------------------------ 1399 // Return a node which is more "ideal" than the current node. Strip out 1400 // control copies 1401 Node* IfNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1402 Node* res = Ideal_common(phase, can_reshape); 1403 if (res != NodeSentinel) { 1404 return res; 1405 } 1406 1407 // Check for people making a useless boolean: things like 1408 // if( (x < y ? true : false) ) { ... } 1409 // Replace with if( x < y ) { ... } 1410 Node *bol2 = remove_useless_bool(this, phase); 1411 if( bol2 ) return bol2; 1412 1413 if (in(0) == NULL) return NULL; // Dead loop? 1414 1415 PhaseIterGVN *igvn = phase->is_IterGVN(); 1416 Node* result = fold_compares(igvn); 1417 if (result != NULL) { 1418 return result; 1419 } 1420 1421 // Scan for an equivalent test 1422 Node *cmp; 1423 int dist = 0; // Cutoff limit for search 1424 int op = Opcode(); 1425 if( op == Op_If && 1426 (cmp=in(1)->in(1))->Opcode() == Op_CmpP ) { 1427 if( cmp->in(2) != NULL && // make sure cmp is not already dead 1428 cmp->in(2)->bottom_type() == TypePtr::NULL_PTR ) { 1429 dist = 64; // Limit for null-pointer scans 1430 } else { 1431 dist = 4; // Do not bother for random pointer tests 1432 } 1433 } else { 1434 dist = 4; // Limit for random junky scans 1435 } 1436 1437 Node* prev_dom = search_identical(dist); 1438 1439 if (prev_dom == NULL) { 1440 return NULL; 1441 } 1442 1443 // Replace dominated IfNode 1444 return dominated_by(prev_dom, igvn); 1445 } 1446 1447 //------------------------------dominated_by----------------------------------- 1448 Node* IfNode::dominated_by(Node* prev_dom, PhaseIterGVN *igvn) { 1449 #ifndef PRODUCT 1450 if (TraceIterativeGVN) { 1451 tty->print(" Removing IfNode: "); this->dump(); 1452 } 1453 if (VerifyOpto && !igvn->allow_progress()) { 1454 // Found an equivalent dominating test, 1455 // we can not guarantee reaching a fix-point for these during iterativeGVN 1456 // since intervening nodes may not change. 1457 return NULL; 1458 } 1459 #endif 1460 1461 igvn->hash_delete(this); // Remove self to prevent spurious V-N 1462 Node *idom = in(0); 1463 // Need opcode to decide which way 'this' test goes 1464 int prev_op = prev_dom->Opcode(); 1465 Node *top = igvn->C->top(); // Shortcut to top 1466 1467 // Loop predicates may have depending checks which should not 1468 // be skipped. For example, range check predicate has two checks 1469 // for lower and upper bounds. 1470 ProjNode* unc_proj = proj_out(1 - prev_dom->as_Proj()->_con)->as_Proj(); 1471 if ((unc_proj != NULL) && (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) != NULL)) { 1472 prev_dom = idom; 1473 } 1474 1475 // Now walk the current IfNode's projections. 1476 // Loop ends when 'this' has no more uses. 1477 for (DUIterator_Last imin, i = last_outs(imin); i >= imin; --i) { 1478 Node *ifp = last_out(i); // Get IfTrue/IfFalse 1479 igvn->add_users_to_worklist(ifp); 1480 // Check which projection it is and set target. 1481 // Data-target is either the dominating projection of the same type 1482 // or TOP if the dominating projection is of opposite type. 1483 // Data-target will be used as the new control edge for the non-CFG 1484 // nodes like Casts and Loads. 1485 Node *data_target = (ifp->Opcode() == prev_op) ? prev_dom : top; 1486 // Control-target is just the If's immediate dominator or TOP. 1487 Node *ctrl_target = (ifp->Opcode() == prev_op) ? idom : top; 1488 1489 // For each child of an IfTrue/IfFalse projection, reroute. 1490 // Loop ends when projection has no more uses. 1491 for (DUIterator_Last jmin, j = ifp->last_outs(jmin); j >= jmin; --j) { 1492 Node* s = ifp->last_out(j); // Get child of IfTrue/IfFalse 1493 if( !s->depends_only_on_test() ) { 1494 // Find the control input matching this def-use edge. 1495 // For Regions it may not be in slot 0. 1496 uint l; 1497 for( l = 0; s->in(l) != ifp; l++ ) { } 1498 igvn->replace_input_of(s, l, ctrl_target); 1499 } else { // Else, for control producers, 1500 igvn->replace_input_of(s, 0, data_target); // Move child to data-target 1501 } 1502 } // End for each child of a projection 1503 1504 igvn->remove_dead_node(ifp); 1505 } // End for each IfTrue/IfFalse child of If 1506 1507 // Kill the IfNode 1508 igvn->remove_dead_node(this); 1509 1510 // Must return either the original node (now dead) or a new node 1511 // (Do not return a top here, since that would break the uniqueness of top.) 1512 return new ConINode(TypeInt::ZERO); 1513 } 1514 1515 Node* IfNode::search_identical(int dist) { 1516 // Setup to scan up the CFG looking for a dominating test 1517 Node* dom = in(0); 1518 Node* prev_dom = this; 1519 int op = Opcode(); 1520 // Search up the dominator tree for an If with an identical test 1521 while (dom->Opcode() != op || // Not same opcode? 1522 dom->in(1) != in(1) || // Not same input 1? 1523 (req() == 3 && dom->in(2) != in(2)) || // Not same input 2? 1524 prev_dom->in(0) != dom) { // One path of test does not dominate? 1525 if (dist < 0) return NULL; 1526 1527 dist--; 1528 prev_dom = dom; 1529 dom = up_one_dom(dom); 1530 if (!dom) return NULL; 1531 } 1532 1533 // Check that we did not follow a loop back to ourselves 1534 if (this == dom) { 1535 return NULL; 1536 } 1537 1538 #ifndef PRODUCT 1539 if (dist > 2) { // Add to count of NULL checks elided 1540 explicit_null_checks_elided++; 1541 } 1542 #endif 1543 1544 return prev_dom; 1545 } 1546 1547 //------------------------------Identity--------------------------------------- 1548 // If the test is constant & we match, then we are the input Control 1549 Node* IfProjNode::Identity(PhaseGVN* phase) { 1550 // Can only optimize if cannot go the other way 1551 const TypeTuple *t = phase->type(in(0))->is_tuple(); 1552 if (t == TypeTuple::IFNEITHER || (always_taken(t) && 1553 // During parsing (GVN) we don't remove dead code aggressively. 1554 // Cut off dead branch and let PhaseRemoveUseless take care of it. 1555 (!phase->is_IterGVN() || 1556 // During IGVN, first wait for the dead branch to be killed. 1557 // Otherwise, the IfNode's control will have two control uses (the IfNode 1558 // that doesn't go away because it still has uses and this branch of the 1559 // If) which breaks other optimizations. Node::has_special_unique_user() 1560 // will cause this node to be reprocessed once the dead branch is killed. 1561 in(0)->outcnt() == 1))) { 1562 // IfNode control 1563 return in(0)->in(0); 1564 } 1565 // no progress 1566 return this; 1567 } 1568 1569 #ifndef PRODUCT 1570 //-------------------------------related--------------------------------------- 1571 // An IfProjNode's related node set consists of its input (an IfNode) including 1572 // the IfNode's condition, plus all of its outputs at level 1. In compact mode, 1573 // the restrictions for IfNode apply (see IfNode::rel). 1574 void IfProjNode::related(GrowableArray<Node*> *in_rel, GrowableArray<Node*> *out_rel, bool compact) const { 1575 Node* ifNode = this->in(0); 1576 in_rel->append(ifNode); 1577 if (compact) { 1578 ifNode->collect_nodes(in_rel, 3, false, true); 1579 } else { 1580 ifNode->collect_nodes_in_all_data(in_rel, false); 1581 } 1582 this->collect_nodes(out_rel, -1, false, false); 1583 } 1584 1585 //------------------------------dump_spec-------------------------------------- 1586 void IfNode::dump_spec(outputStream *st) const { 1587 st->print("P=%f, C=%f",_prob,_fcnt); 1588 } 1589 1590 //-------------------------------related--------------------------------------- 1591 // For an IfNode, the set of related output nodes is just the output nodes till 1592 // depth 2, i.e, the IfTrue/IfFalse projection nodes plus the nodes they refer. 1593 // The related input nodes contain no control nodes, but all data nodes 1594 // pertaining to the condition. In compact mode, the input nodes are collected 1595 // up to a depth of 3. 1596 void IfNode::related(GrowableArray <Node *> *in_rel, GrowableArray <Node *> *out_rel, bool compact) const { 1597 if (compact) { 1598 this->collect_nodes(in_rel, 3, false, true); 1599 } else { 1600 this->collect_nodes_in_all_data(in_rel, false); 1601 } 1602 this->collect_nodes(out_rel, -2, false, false); 1603 } 1604 #endif 1605 1606 //------------------------------idealize_test---------------------------------- 1607 // Try to canonicalize tests better. Peek at the Cmp/Bool/If sequence and 1608 // come up with a canonical sequence. Bools getting 'eq', 'gt' and 'ge' forms 1609 // converted to 'ne', 'le' and 'lt' forms. IfTrue/IfFalse get swapped as 1610 // needed. 1611 static IfNode* idealize_test(PhaseGVN* phase, IfNode* iff) { 1612 assert(iff->in(0) != NULL, "If must be live"); 1613 1614 if (iff->outcnt() != 2) return NULL; // Malformed projections. 1615 Node* old_if_f = iff->proj_out(false); 1616 Node* old_if_t = iff->proj_out(true); 1617 1618 // CountedLoopEnds want the back-control test to be TRUE, irregardless of 1619 // whether they are testing a 'gt' or 'lt' condition. The 'gt' condition 1620 // happens in count-down loops 1621 if (iff->is_CountedLoopEnd()) return NULL; 1622 Node* proj_true = iff->proj_out(true); 1623 if (proj_true->outcnt() == 1) { 1624 Node* c = proj_true->unique_out(); 1625 // Leave test of outer strip mined loop alone 1626 if (c != NULL && c->is_Loop() && 1627 c->in(LoopNode::LoopBackControl) == proj_true && 1628 c->as_Loop()->is_strip_mined()) { 1629 return NULL; 1630 } 1631 } 1632 if (!iff->in(1)->is_Bool()) return NULL; // Happens for partially optimized IF tests 1633 BoolNode *b = iff->in(1)->as_Bool(); 1634 BoolTest bt = b->_test; 1635 // Test already in good order? 1636 if( bt.is_canonical() ) 1637 return NULL; 1638 1639 // Flip test to be canonical. Requires flipping the IfFalse/IfTrue and 1640 // cloning the IfNode. 1641 Node* new_b = phase->transform( new BoolNode(b->in(1), bt.negate()) ); 1642 if( !new_b->is_Bool() ) return NULL; 1643 b = new_b->as_Bool(); 1644 1645 PhaseIterGVN *igvn = phase->is_IterGVN(); 1646 assert( igvn, "Test is not canonical in parser?" ); 1647 1648 // The IF node never really changes, but it needs to be cloned 1649 iff = iff->clone()->as_If(); 1650 iff->set_req(1, b); 1651 iff->_prob = 1.0-iff->_prob; 1652 1653 Node *prior = igvn->hash_find_insert(iff); 1654 if( prior ) { 1655 igvn->remove_dead_node(iff); 1656 iff = (IfNode*)prior; 1657 } else { 1658 // Cannot call transform on it just yet 1659 igvn->set_type_bottom(iff); 1660 } 1661 igvn->_worklist.push(iff); 1662 1663 // Now handle projections. Cloning not required. 1664 Node* new_if_f = (Node*)(new IfFalseNode( iff )); 1665 Node* new_if_t = (Node*)(new IfTrueNode ( iff )); 1666 1667 igvn->register_new_node_with_optimizer(new_if_f); 1668 igvn->register_new_node_with_optimizer(new_if_t); 1669 // Flip test, so flip trailing control 1670 igvn->replace_node(old_if_f, new_if_t); 1671 igvn->replace_node(old_if_t, new_if_f); 1672 1673 // Progress 1674 return iff; 1675 } 1676 1677 Node* RangeCheckNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1678 Node* res = Ideal_common(phase, can_reshape); 1679 if (res != NodeSentinel) { 1680 return res; 1681 } 1682 1683 PhaseIterGVN *igvn = phase->is_IterGVN(); 1684 // Setup to scan up the CFG looking for a dominating test 1685 Node* prev_dom = this; 1686 1687 // Check for range-check vs other kinds of tests 1688 Node* index1; 1689 Node* range1; 1690 jint offset1; 1691 int flip1 = is_range_check(range1, index1, offset1); 1692 if (flip1) { 1693 Node* dom = in(0); 1694 // Try to remove extra range checks. All 'up_one_dom' gives up at merges 1695 // so all checks we inspect post-dominate the top-most check we find. 1696 // If we are going to fail the current check and we reach the top check 1697 // then we are guaranteed to fail, so just start interpreting there. 1698 // We 'expand' the top 3 range checks to include all post-dominating 1699 // checks. 1700 1701 // The top 3 range checks seen 1702 const int NRC =3; 1703 RangeCheck prev_checks[NRC]; 1704 int nb_checks = 0; 1705 1706 // Low and high offsets seen so far 1707 jint off_lo = offset1; 1708 jint off_hi = offset1; 1709 1710 bool found_immediate_dominator = false; 1711 1712 // Scan for the top checks and collect range of offsets 1713 for (int dist = 0; dist < 999; dist++) { // Range-Check scan limit 1714 if (dom->Opcode() == Op_RangeCheck && // Not same opcode? 1715 prev_dom->in(0) == dom) { // One path of test does dominate? 1716 if (dom == this) return NULL; // dead loop 1717 // See if this is a range check 1718 Node* index2; 1719 Node* range2; 1720 jint offset2; 1721 int flip2 = dom->as_RangeCheck()->is_range_check(range2, index2, offset2); 1722 // See if this is a _matching_ range check, checking against 1723 // the same array bounds. 1724 if (flip2 == flip1 && range2 == range1 && index2 == index1 && 1725 dom->outcnt() == 2) { 1726 if (nb_checks == 0 && dom->in(1) == in(1)) { 1727 // Found an immediately dominating test at the same offset. 1728 // This kind of back-to-back test can be eliminated locally, 1729 // and there is no need to search further for dominating tests. 1730 assert(offset2 == offset1, "Same test but different offsets"); 1731 found_immediate_dominator = true; 1732 break; 1733 } 1734 // Gather expanded bounds 1735 off_lo = MIN2(off_lo,offset2); 1736 off_hi = MAX2(off_hi,offset2); 1737 // Record top NRC range checks 1738 prev_checks[nb_checks%NRC].ctl = prev_dom; 1739 prev_checks[nb_checks%NRC].off = offset2; 1740 nb_checks++; 1741 } 1742 } 1743 prev_dom = dom; 1744 dom = up_one_dom(dom); 1745 if (!dom) break; 1746 } 1747 1748 if (!found_immediate_dominator) { 1749 // Attempt to widen the dominating range check to cover some later 1750 // ones. Since range checks "fail" by uncommon-trapping to the 1751 // interpreter, widening a check can make us speculatively enter 1752 // the interpreter. If we see range-check deopt's, do not widen! 1753 if (!phase->C->allow_range_check_smearing()) return NULL; 1754 1755 // Didn't find prior covering check, so cannot remove anything. 1756 if (nb_checks == 0) { 1757 return NULL; 1758 } 1759 // Constant indices only need to check the upper bound. 1760 // Non-constant indices must check both low and high. 1761 int chk0 = (nb_checks - 1) % NRC; 1762 if (index1) { 1763 if (nb_checks == 1) { 1764 return NULL; 1765 } else { 1766 // If the top range check's constant is the min or max of 1767 // all constants we widen the next one to cover the whole 1768 // range of constants. 1769 RangeCheck rc0 = prev_checks[chk0]; 1770 int chk1 = (nb_checks - 2) % NRC; 1771 RangeCheck rc1 = prev_checks[chk1]; 1772 if (rc0.off == off_lo) { 1773 adjust_check(rc1.ctl, range1, index1, flip1, off_hi, igvn); 1774 prev_dom = rc1.ctl; 1775 } else if (rc0.off == off_hi) { 1776 adjust_check(rc1.ctl, range1, index1, flip1, off_lo, igvn); 1777 prev_dom = rc1.ctl; 1778 } else { 1779 // If the top test's constant is not the min or max of all 1780 // constants, we need 3 range checks. We must leave the 1781 // top test unchanged because widening it would allow the 1782 // accesses it protects to successfully read/write out of 1783 // bounds. 1784 if (nb_checks == 2) { 1785 return NULL; 1786 } 1787 int chk2 = (nb_checks - 3) % NRC; 1788 RangeCheck rc2 = prev_checks[chk2]; 1789 // The top range check a+i covers interval: -a <= i < length-a 1790 // The second range check b+i covers interval: -b <= i < length-b 1791 if (rc1.off <= rc0.off) { 1792 // if b <= a, we change the second range check to: 1793 // -min_of_all_constants <= i < length-min_of_all_constants 1794 // Together top and second range checks now cover: 1795 // -min_of_all_constants <= i < length-a 1796 // which is more restrictive than -b <= i < length-b: 1797 // -b <= -min_of_all_constants <= i < length-a <= length-b 1798 // The third check is then changed to: 1799 // -max_of_all_constants <= i < length-max_of_all_constants 1800 // so 2nd and 3rd checks restrict allowed values of i to: 1801 // -min_of_all_constants <= i < length-max_of_all_constants 1802 adjust_check(rc1.ctl, range1, index1, flip1, off_lo, igvn); 1803 adjust_check(rc2.ctl, range1, index1, flip1, off_hi, igvn); 1804 } else { 1805 // if b > a, we change the second range check to: 1806 // -max_of_all_constants <= i < length-max_of_all_constants 1807 // Together top and second range checks now cover: 1808 // -a <= i < length-max_of_all_constants 1809 // which is more restrictive than -b <= i < length-b: 1810 // -b < -a <= i < length-max_of_all_constants <= length-b 1811 // The third check is then changed to: 1812 // -max_of_all_constants <= i < length-max_of_all_constants 1813 // so 2nd and 3rd checks restrict allowed values of i to: 1814 // -min_of_all_constants <= i < length-max_of_all_constants 1815 adjust_check(rc1.ctl, range1, index1, flip1, off_hi, igvn); 1816 adjust_check(rc2.ctl, range1, index1, flip1, off_lo, igvn); 1817 } 1818 prev_dom = rc2.ctl; 1819 } 1820 } 1821 } else { 1822 RangeCheck rc0 = prev_checks[chk0]; 1823 // 'Widen' the offset of the 1st and only covering check 1824 adjust_check(rc0.ctl, range1, index1, flip1, off_hi, igvn); 1825 // Test is now covered by prior checks, dominate it out 1826 prev_dom = rc0.ctl; 1827 } 1828 } 1829 } else { 1830 prev_dom = search_identical(4); 1831 1832 if (prev_dom == NULL) { 1833 return NULL; 1834 } 1835 } 1836 1837 // Replace dominated IfNode 1838 return dominated_by(prev_dom, igvn); 1839 }