1 /* 2 * Copyright (c) 2000, 2023, 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/predicates.hpp" 36 #include "opto/runtime.hpp" 37 #include "opto/rootnode.hpp" 38 #include "opto/subnode.hpp" 39 #include "opto/subtypenode.hpp" 40 41 // Portions of code courtesy of Clifford Click 42 43 // Optimization - Graph Style 44 45 46 #ifndef PRODUCT 47 extern uint explicit_null_checks_elided; 48 #endif 49 50 //============================================================================= 51 //------------------------------Value------------------------------------------ 52 // Return a tuple for whichever arm of the IF is reachable 53 const Type* IfNode::Value(PhaseGVN* phase) const { 54 if( !in(0) ) return Type::TOP; 55 if( phase->type(in(0)) == Type::TOP ) 56 return Type::TOP; 57 const Type *t = phase->type(in(1)); 58 if( t == Type::TOP ) // data is undefined 59 return TypeTuple::IFNEITHER; // unreachable altogether 60 if( t == TypeInt::ZERO ) // zero, or false 61 return TypeTuple::IFFALSE; // only false branch is reachable 62 if( t == TypeInt::ONE ) // 1, or true 63 return TypeTuple::IFTRUE; // only true branch is reachable 64 assert( t == TypeInt::BOOL, "expected boolean type" ); 65 66 return TypeTuple::IFBOTH; // No progress 67 } 68 69 const RegMask &IfNode::out_RegMask() const { 70 return RegMask::Empty; 71 } 72 73 //------------------------------split_if--------------------------------------- 74 // Look for places where we merge constants, then test on the merged value. 75 // If the IF test will be constant folded on the path with the constant, we 76 // win by splitting the IF to before the merge point. 77 static Node* split_if(IfNode *iff, PhaseIterGVN *igvn) { 78 // I could be a lot more general here, but I'm trying to squeeze this 79 // in before the Christmas '98 break so I'm gonna be kinda restrictive 80 // on the patterns I accept. CNC 81 82 // Look for a compare of a constant and a merged value 83 Node *i1 = iff->in(1); 84 if( !i1->is_Bool() ) return nullptr; 85 BoolNode *b = i1->as_Bool(); 86 Node *cmp = b->in(1); 87 if( !cmp->is_Cmp() ) return nullptr; 88 i1 = cmp->in(1); 89 if( i1 == nullptr || !i1->is_Phi() ) return nullptr; 90 PhiNode *phi = i1->as_Phi(); 91 Node *con2 = cmp->in(2); 92 if( !con2->is_Con() ) return nullptr; 93 // See that the merge point contains some constants 94 Node *con1=nullptr; 95 uint i4; 96 RegionNode* phi_region = phi->region(); 97 for (i4 = 1; i4 < phi->req(); i4++ ) { 98 con1 = phi->in(i4); 99 // Do not optimize partially collapsed merges 100 if (con1 == nullptr || phi_region->in(i4) == nullptr || igvn->type(phi_region->in(i4)) == Type::TOP) { 101 igvn->_worklist.push(iff); 102 return nullptr; 103 } 104 if( con1->is_Con() ) break; // Found a constant 105 // Also allow null-vs-not-null checks 106 const TypePtr *tp = igvn->type(con1)->isa_ptr(); 107 if( tp && tp->_ptr == TypePtr::NotNull ) 108 break; 109 } 110 if( i4 >= phi->req() ) return nullptr; // Found no constants 111 112 igvn->C->set_has_split_ifs(true); // Has chance for split-if 113 114 // Make sure that the compare can be constant folded away 115 Node *cmp2 = cmp->clone(); 116 cmp2->set_req(1,con1); 117 cmp2->set_req(2,con2); 118 const Type *t = cmp2->Value(igvn); 119 // This compare is dead, so whack it! 120 igvn->remove_dead_node(cmp2); 121 if( !t->singleton() ) return nullptr; 122 123 // No intervening control, like a simple Call 124 Node* r = iff->in(0); 125 if (!r->is_Region() || r->is_Loop() || phi_region != r || r->as_Region()->is_copy()) { 126 return nullptr; 127 } 128 129 // No other users of the cmp/bool 130 if (b->outcnt() != 1 || cmp->outcnt() != 1) { 131 //tty->print_cr("many users of cmp/bool"); 132 return nullptr; 133 } 134 135 // Make sure we can determine where all the uses of merged values go 136 for (DUIterator_Fast jmax, j = r->fast_outs(jmax); j < jmax; j++) { 137 Node* u = r->fast_out(j); 138 if( u == r ) continue; 139 if( u == iff ) continue; 140 if( u->outcnt() == 0 ) continue; // use is dead & ignorable 141 if( !u->is_Phi() ) { 142 /* 143 if( u->is_Start() ) { 144 tty->print_cr("Region has inlined start use"); 145 } else { 146 tty->print_cr("Region has odd use"); 147 u->dump(2); 148 }*/ 149 return nullptr; 150 } 151 if( u != phi ) { 152 // CNC - do not allow any other merged value 153 //tty->print_cr("Merging another value"); 154 //u->dump(2); 155 return nullptr; 156 } 157 // Make sure we can account for all Phi uses 158 for (DUIterator_Fast kmax, k = u->fast_outs(kmax); k < kmax; k++) { 159 Node* v = u->fast_out(k); // User of the phi 160 // CNC - Allow only really simple patterns. 161 // In particular I disallow AddP of the Phi, a fairly common pattern 162 if (v == cmp) continue; // The compare is OK 163 if (v->is_ConstraintCast()) { 164 // If the cast is derived from data flow edges, it may not have a control edge. 165 // If so, it should be safe to split. But follow-up code can not deal with 166 // this (l. 359). So skip. 167 if (v->in(0) == nullptr) { 168 return nullptr; 169 } 170 if (v->in(0)->in(0) == iff) { 171 continue; // CastPP/II of the IfNode is OK 172 } 173 } 174 // Disabled following code because I cannot tell if exactly one 175 // path dominates without a real dominator check. CNC 9/9/1999 176 //uint vop = v->Opcode(); 177 //if( vop == Op_Phi ) { // Phi from another merge point might be OK 178 // Node *r = v->in(0); // Get controlling point 179 // if( !r ) return nullptr; // Degraded to a copy 180 // // Find exactly one path in (either True or False doms, but not IFF) 181 // int cnt = 0; 182 // for( uint i = 1; i < r->req(); i++ ) 183 // if( r->in(i) && r->in(i)->in(0) == iff ) 184 // cnt++; 185 // if( cnt == 1 ) continue; // Exactly one of True or False guards Phi 186 //} 187 if( !v->is_Call() ) { 188 /* 189 if( v->Opcode() == Op_AddP ) { 190 tty->print_cr("Phi has AddP use"); 191 } else if( v->Opcode() == Op_CastPP ) { 192 tty->print_cr("Phi has CastPP use"); 193 } else if( v->Opcode() == Op_CastII ) { 194 tty->print_cr("Phi has CastII use"); 195 } else { 196 tty->print_cr("Phi has use I can't be bothered with"); 197 } 198 */ 199 } 200 return nullptr; 201 202 /* CNC - Cut out all the fancy acceptance tests 203 // Can we clone this use when doing the transformation? 204 // If all uses are from Phis at this merge or constants, then YES. 205 if( !v->in(0) && v != cmp ) { 206 tty->print_cr("Phi has free-floating use"); 207 v->dump(2); 208 return nullptr; 209 } 210 for( uint l = 1; l < v->req(); l++ ) { 211 if( (!v->in(l)->is_Phi() || v->in(l)->in(0) != r) && 212 !v->in(l)->is_Con() ) { 213 tty->print_cr("Phi has use"); 214 v->dump(2); 215 return nullptr; 216 } // End of if Phi-use input is neither Phi nor Constant 217 } // End of for all inputs to Phi-use 218 */ 219 } // End of for all uses of Phi 220 } // End of for all uses of Region 221 222 // Only do this if the IF node is in a sane state 223 if (iff->outcnt() != 2) 224 return nullptr; 225 226 // Got a hit! Do the Mondo Hack! 227 // 228 //ABC a1c def ghi B 1 e h A C a c d f g i 229 // R - Phi - Phi - Phi Rc - Phi - Phi - Phi Rx - Phi - Phi - Phi 230 // cmp - 2 cmp - 2 cmp - 2 231 // bool bool_c bool_x 232 // if if_c if_x 233 // T F T F T F 234 // ..s.. ..t .. ..s.. ..t.. ..s.. ..t.. 235 // 236 // Split the paths coming into the merge point into 2 separate groups of 237 // merges. On the left will be all the paths feeding constants into the 238 // Cmp's Phi. On the right will be the remaining paths. The Cmp's Phi 239 // will fold up into a constant; this will let the Cmp fold up as well as 240 // all the control flow. Below the original IF we have 2 control 241 // dependent regions, 's' and 't'. Now we will merge the two paths 242 // just prior to 's' and 't' from the two IFs. At least 1 path (and quite 243 // likely 2 or more) will promptly constant fold away. 244 PhaseGVN *phase = igvn; 245 246 // Make a region merging constants and a region merging the rest 247 uint req_c = 0; 248 for (uint ii = 1; ii < r->req(); ii++) { 249 if (phi->in(ii) == con1) { 250 req_c++; 251 } 252 if (Node::may_be_loop_entry(r->in(ii))) { 253 // Bail out if splitting through a region with a Parse Predicate input (could 254 // also be a loop header before loop opts creates a LoopNode for it). 255 return nullptr; 256 } 257 } 258 259 // If all the defs of the phi are the same constant, we already have the desired end state. 260 // Skip the split that would create empty phi and region nodes. 261 if ((r->req() - req_c) == 1) { 262 return nullptr; 263 } 264 265 // At this point we know that we can apply the split if optimization. If the region is still on the worklist, 266 // we should wait until it is processed. The region might be removed which makes this optimization redundant. 267 // This also avoids the creation of dead data loops when rewiring data nodes below when a region is dying. 268 if (igvn->_worklist.member(r)) { 269 igvn->_worklist.push(iff); // retry split if later again 270 return nullptr; 271 } 272 273 Node *region_c = new RegionNode(req_c + 1); 274 Node *phi_c = con1; 275 uint len = r->req(); 276 Node *region_x = new RegionNode(len - req_c); 277 Node *phi_x = PhiNode::make_blank(region_x, phi); 278 for (uint i = 1, i_c = 1, i_x = 1; i < len; i++) { 279 if (phi->in(i) == con1) { 280 region_c->init_req( i_c++, r ->in(i) ); 281 } else { 282 region_x->init_req( i_x, r ->in(i) ); 283 phi_x ->init_req( i_x++, phi->in(i) ); 284 } 285 } 286 287 // Register the new RegionNodes but do not transform them. Cannot 288 // transform until the entire Region/Phi conglomerate has been hacked 289 // as a single huge transform. 290 igvn->register_new_node_with_optimizer( region_c ); 291 igvn->register_new_node_with_optimizer( region_x ); 292 // Prevent the untimely death of phi_x. Currently he has no uses. He is 293 // about to get one. If this only use goes away, then phi_x will look dead. 294 // However, he will be picking up some more uses down below. 295 Node *hook = new Node(4); 296 hook->init_req(0, phi_x); 297 hook->init_req(1, phi_c); 298 phi_x = phase->transform( phi_x ); 299 300 // Make the compare 301 Node *cmp_c = phase->makecon(t); 302 Node *cmp_x = cmp->clone(); 303 cmp_x->set_req(1,phi_x); 304 cmp_x->set_req(2,con2); 305 cmp_x = phase->transform(cmp_x); 306 // Make the bool 307 Node *b_c = phase->transform(new BoolNode(cmp_c,b->_test._test)); 308 Node *b_x = phase->transform(new BoolNode(cmp_x,b->_test._test)); 309 // Make the IfNode 310 IfNode* iff_c = iff->clone()->as_If(); 311 iff_c->set_req(0, region_c); 312 iff_c->set_req(1, b_c); 313 igvn->set_type_bottom(iff_c); 314 igvn->_worklist.push(iff_c); 315 hook->init_req(2, iff_c); 316 317 IfNode* iff_x = iff->clone()->as_If(); 318 iff_x->set_req(0, region_x); 319 iff_x->set_req(1, b_x); 320 igvn->set_type_bottom(iff_x); 321 igvn->_worklist.push(iff_x); 322 hook->init_req(3, iff_x); 323 324 // Make the true/false arms 325 Node *iff_c_t = phase->transform(new IfTrueNode (iff_c)); 326 Node *iff_c_f = phase->transform(new IfFalseNode(iff_c)); 327 Node *iff_x_t = phase->transform(new IfTrueNode (iff_x)); 328 Node *iff_x_f = phase->transform(new IfFalseNode(iff_x)); 329 330 // Merge the TRUE paths 331 Node *region_s = new RegionNode(3); 332 igvn->_worklist.push(region_s); 333 region_s->init_req(1, iff_c_t); 334 region_s->init_req(2, iff_x_t); 335 igvn->register_new_node_with_optimizer( region_s ); 336 337 // Merge the FALSE paths 338 Node *region_f = new RegionNode(3); 339 igvn->_worklist.push(region_f); 340 region_f->init_req(1, iff_c_f); 341 region_f->init_req(2, iff_x_f); 342 igvn->register_new_node_with_optimizer( region_f ); 343 344 igvn->hash_delete(cmp);// Remove soon-to-be-dead node from hash table. 345 cmp->set_req(1,nullptr); // Whack the inputs to cmp because it will be dead 346 cmp->set_req(2,nullptr); 347 // Check for all uses of the Phi and give them a new home. 348 // The 'cmp' got cloned, but CastPP/IIs need to be moved. 349 Node *phi_s = nullptr; // do not construct unless needed 350 Node *phi_f = nullptr; // do not construct unless needed 351 for (DUIterator_Last i2min, i2 = phi->last_outs(i2min); i2 >= i2min; --i2) { 352 Node* v = phi->last_out(i2);// User of the phi 353 igvn->rehash_node_delayed(v); // Have to fixup other Phi users 354 uint vop = v->Opcode(); 355 Node *proj = nullptr; 356 if( vop == Op_Phi ) { // Remote merge point 357 Node *r = v->in(0); 358 for (uint i3 = 1; i3 < r->req(); i3++) 359 if (r->in(i3) && r->in(i3)->in(0) == iff) { 360 proj = r->in(i3); 361 break; 362 } 363 } else if( v->is_ConstraintCast() ) { 364 proj = v->in(0); // Controlling projection 365 } else { 366 assert( 0, "do not know how to handle this guy" ); 367 } 368 guarantee(proj != nullptr, "sanity"); 369 370 Node *proj_path_data, *proj_path_ctrl; 371 if( proj->Opcode() == Op_IfTrue ) { 372 if( phi_s == nullptr ) { 373 // Only construct phi_s if needed, otherwise provides 374 // interfering use. 375 phi_s = PhiNode::make_blank(region_s,phi); 376 phi_s->init_req( 1, phi_c ); 377 phi_s->init_req( 2, phi_x ); 378 hook->add_req(phi_s); 379 phi_s = phase->transform(phi_s); 380 } 381 proj_path_data = phi_s; 382 proj_path_ctrl = region_s; 383 } else { 384 if( phi_f == nullptr ) { 385 // Only construct phi_f if needed, otherwise provides 386 // interfering use. 387 phi_f = PhiNode::make_blank(region_f,phi); 388 phi_f->init_req( 1, phi_c ); 389 phi_f->init_req( 2, phi_x ); 390 hook->add_req(phi_f); 391 phi_f = phase->transform(phi_f); 392 } 393 proj_path_data = phi_f; 394 proj_path_ctrl = region_f; 395 } 396 397 // Fixup 'v' for for the split 398 if( vop == Op_Phi ) { // Remote merge point 399 uint i; 400 for( i = 1; i < v->req(); i++ ) 401 if( v->in(i) == phi ) 402 break; 403 v->set_req(i, proj_path_data ); 404 } else if( v->is_ConstraintCast() ) { 405 v->set_req(0, proj_path_ctrl ); 406 v->set_req(1, proj_path_data ); 407 } else 408 ShouldNotReachHere(); 409 } 410 411 // Now replace the original iff's True/False with region_s/region_t. 412 // This makes the original iff go dead. 413 for (DUIterator_Last i3min, i3 = iff->last_outs(i3min); i3 >= i3min; --i3) { 414 Node* p = iff->last_out(i3); 415 assert( p->Opcode() == Op_IfTrue || p->Opcode() == Op_IfFalse, "" ); 416 Node *u = (p->Opcode() == Op_IfTrue) ? region_s : region_f; 417 // Replace p with u 418 igvn->add_users_to_worklist(p); 419 for (DUIterator_Last lmin, l = p->last_outs(lmin); l >= lmin;) { 420 Node* x = p->last_out(l); 421 igvn->hash_delete(x); 422 uint uses_found = 0; 423 for( uint j = 0; j < x->req(); j++ ) { 424 if( x->in(j) == p ) { 425 x->set_req(j, u); 426 uses_found++; 427 } 428 } 429 l -= uses_found; // we deleted 1 or more copies of this edge 430 } 431 igvn->remove_dead_node(p); 432 } 433 434 // Force the original merge dead 435 igvn->hash_delete(r); 436 // First, remove region's dead users. 437 for (DUIterator_Last lmin, l = r->last_outs(lmin); l >= lmin;) { 438 Node* u = r->last_out(l); 439 if( u == r ) { 440 r->set_req(0, nullptr); 441 } else { 442 assert(u->outcnt() == 0, "only dead users"); 443 igvn->remove_dead_node(u); 444 } 445 l -= 1; 446 } 447 igvn->remove_dead_node(r); 448 449 // Now remove the bogus extra edges used to keep things alive 450 igvn->remove_dead_node( hook ); 451 452 // Must return either the original node (now dead) or a new node 453 // (Do not return a top here, since that would break the uniqueness of top.) 454 return new ConINode(TypeInt::ZERO); 455 } 456 457 // if this IfNode follows a range check pattern return the projection 458 // for the failed path 459 ProjNode* IfNode::range_check_trap_proj(int& flip_test, Node*& l, Node*& r) { 460 if (outcnt() != 2) { 461 return nullptr; 462 } 463 Node* b = in(1); 464 if (b == nullptr || !b->is_Bool()) return nullptr; 465 BoolNode* bn = b->as_Bool(); 466 Node* cmp = bn->in(1); 467 if (cmp == nullptr) return nullptr; 468 if (cmp->Opcode() != Op_CmpU) return nullptr; 469 470 l = cmp->in(1); 471 r = cmp->in(2); 472 flip_test = 1; 473 if (bn->_test._test == BoolTest::le) { 474 l = cmp->in(2); 475 r = cmp->in(1); 476 flip_test = 2; 477 } else if (bn->_test._test != BoolTest::lt) { 478 return nullptr; 479 } 480 if (l->is_top()) return nullptr; // Top input means dead test 481 if (r->Opcode() != Op_LoadRange && !is_RangeCheck()) return nullptr; 482 483 // We have recognized one of these forms: 484 // Flip 1: If (Bool[<] CmpU(l, LoadRange)) ... 485 // Flip 2: If (Bool[<=] CmpU(LoadRange, l)) ... 486 487 ProjNode* iftrap = proj_out_or_null(flip_test == 2 ? true : false); 488 return iftrap; 489 } 490 491 492 //------------------------------is_range_check--------------------------------- 493 // Return 0 if not a range check. Return 1 if a range check and set index and 494 // offset. Return 2 if we had to negate the test. Index is null if the check 495 // is versus a constant. 496 int RangeCheckNode::is_range_check(Node* &range, Node* &index, jint &offset) { 497 int flip_test = 0; 498 Node* l = nullptr; 499 Node* r = nullptr; 500 ProjNode* iftrap = range_check_trap_proj(flip_test, l, r); 501 502 if (iftrap == nullptr) { 503 return 0; 504 } 505 506 // Make sure it's a real range check by requiring an uncommon trap 507 // along the OOB path. Otherwise, it's possible that the user wrote 508 // something which optimized to look like a range check but behaves 509 // in some other way. 510 if (iftrap->is_uncommon_trap_proj(Deoptimization::Reason_range_check) == nullptr) { 511 return 0; 512 } 513 514 // Look for index+offset form 515 Node* ind = l; 516 jint off = 0; 517 if (l->is_top()) { 518 return 0; 519 } else if (l->Opcode() == Op_AddI) { 520 if ((off = l->in(1)->find_int_con(0)) != 0) { 521 ind = l->in(2)->uncast(); 522 } else if ((off = l->in(2)->find_int_con(0)) != 0) { 523 ind = l->in(1)->uncast(); 524 } 525 } else if ((off = l->find_int_con(-1)) >= 0) { 526 // constant offset with no variable index 527 ind = nullptr; 528 } else { 529 // variable index with no constant offset (or dead negative index) 530 off = 0; 531 } 532 533 // Return all the values: 534 index = ind; 535 offset = off; 536 range = r; 537 return flip_test; 538 } 539 540 //------------------------------adjust_check----------------------------------- 541 // Adjust (widen) a prior range check 542 static void adjust_check(IfProjNode* proj, Node* range, Node* index, 543 int flip, jint off_lo, PhaseIterGVN* igvn) { 544 PhaseGVN *gvn = igvn; 545 // Break apart the old check 546 Node *iff = proj->in(0); 547 Node *bol = iff->in(1); 548 if( bol->is_top() ) return; // In case a partially dead range check appears 549 // bail (or bomb[ASSERT/DEBUG]) if NOT projection-->IfNode-->BoolNode 550 DEBUG_ONLY( if (!bol->is_Bool()) { proj->dump(3); fatal("Expect projection-->IfNode-->BoolNode"); } ) 551 if (!bol->is_Bool()) return; 552 553 Node *cmp = bol->in(1); 554 // Compute a new check 555 Node *new_add = gvn->intcon(off_lo); 556 if (index) { 557 new_add = off_lo ? gvn->transform(new AddINode(index, new_add)) : index; 558 } 559 Node *new_cmp = (flip == 1) 560 ? new CmpUNode(new_add, range) 561 : new CmpUNode(range, new_add); 562 new_cmp = gvn->transform(new_cmp); 563 // See if no need to adjust the existing check 564 if (new_cmp == cmp) return; 565 // Else, adjust existing check 566 Node* new_bol = gvn->transform(new BoolNode(new_cmp, bol->as_Bool()->_test._test)); 567 igvn->rehash_node_delayed(iff); 568 iff->set_req_X(1, new_bol, igvn); 569 // As part of range check smearing, this range check is widened. Loads and range check Cast nodes that are control 570 // dependent on this range check now depend on multiple dominating range checks. These control dependent nodes end up 571 // at the lowest/nearest dominating check in the graph. To ensure that these Loads/Casts do not float above any of the 572 // dominating checks (even when the lowest dominating check is later replaced by yet another dominating check), we 573 // need to pin them at the lowest dominating check. 574 proj->pin_array_access_nodes(igvn); 575 } 576 577 //------------------------------up_one_dom------------------------------------- 578 // Walk up the dominator tree one step. Return null at root or true 579 // complex merges. Skips through small diamonds. 580 Node* IfNode::up_one_dom(Node *curr, bool linear_only) { 581 Node *dom = curr->in(0); 582 if( !dom ) // Found a Region degraded to a copy? 583 return curr->nonnull_req(); // Skip thru it 584 585 if( curr != dom ) // Normal walk up one step? 586 return dom; 587 588 // Use linear_only if we are still parsing, since we cannot 589 // trust the regions to be fully filled in. 590 if (linear_only) 591 return nullptr; 592 593 if( dom->is_Root() ) 594 return nullptr; 595 596 // Else hit a Region. Check for a loop header 597 if( dom->is_Loop() ) 598 return dom->in(1); // Skip up thru loops 599 600 // Check for small diamonds 601 Node *din1, *din2, *din3, *din4; 602 if( dom->req() == 3 && // 2-path merge point 603 (din1 = dom ->in(1)) && // Left path exists 604 (din2 = dom ->in(2)) && // Right path exists 605 (din3 = din1->in(0)) && // Left path up one 606 (din4 = din2->in(0)) ) { // Right path up one 607 if( din3->is_Call() && // Handle a slow-path call on either arm 608 (din3 = din3->in(0)) ) 609 din3 = din3->in(0); 610 if( din4->is_Call() && // Handle a slow-path call on either arm 611 (din4 = din4->in(0)) ) 612 din4 = din4->in(0); 613 if (din3 != nullptr && din3 == din4 && din3->is_If()) // Regions not degraded to a copy 614 return din3; // Skip around diamonds 615 } 616 617 // Give up the search at true merges 618 return nullptr; // Dead loop? Or hit root? 619 } 620 621 622 //------------------------------filtered_int_type-------------------------------- 623 // Return a possibly more restrictive type for val based on condition control flow for an if 624 const TypeInt* IfNode::filtered_int_type(PhaseGVN* gvn, Node* val, Node* if_proj) { 625 assert(if_proj && 626 (if_proj->Opcode() == Op_IfTrue || if_proj->Opcode() == Op_IfFalse), "expecting an if projection"); 627 if (if_proj->in(0) && if_proj->in(0)->is_If()) { 628 IfNode* iff = if_proj->in(0)->as_If(); 629 if (iff->in(1) && iff->in(1)->is_Bool()) { 630 BoolNode* bol = iff->in(1)->as_Bool(); 631 if (bol->in(1) && bol->in(1)->is_Cmp()) { 632 const CmpNode* cmp = bol->in(1)->as_Cmp(); 633 if (cmp->in(1) == val) { 634 const TypeInt* cmp2_t = gvn->type(cmp->in(2))->isa_int(); 635 if (cmp2_t != nullptr) { 636 jint lo = cmp2_t->_lo; 637 jint hi = cmp2_t->_hi; 638 BoolTest::mask msk = if_proj->Opcode() == Op_IfTrue ? bol->_test._test : bol->_test.negate(); 639 switch (msk) { 640 case BoolTest::ne: { 641 // If val is compared to its lower or upper bound, we can narrow the type 642 const TypeInt* val_t = gvn->type(val)->isa_int(); 643 if (val_t != nullptr && !val_t->singleton() && cmp2_t->is_con()) { 644 if (val_t->_lo == lo) { 645 return TypeInt::make(val_t->_lo + 1, val_t->_hi, val_t->_widen); 646 } else if (val_t->_hi == hi) { 647 return TypeInt::make(val_t->_lo, val_t->_hi - 1, val_t->_widen); 648 } 649 } 650 // Can't refine type 651 return nullptr; 652 } 653 case BoolTest::eq: 654 return cmp2_t; 655 case BoolTest::lt: 656 lo = TypeInt::INT->_lo; 657 if (hi != min_jint) { 658 hi = hi - 1; 659 } 660 break; 661 case BoolTest::le: 662 lo = TypeInt::INT->_lo; 663 break; 664 case BoolTest::gt: 665 if (lo != max_jint) { 666 lo = lo + 1; 667 } 668 hi = TypeInt::INT->_hi; 669 break; 670 case BoolTest::ge: 671 // lo unchanged 672 hi = TypeInt::INT->_hi; 673 break; 674 default: 675 break; 676 } 677 const TypeInt* rtn_t = TypeInt::make(lo, hi, cmp2_t->_widen); 678 return rtn_t; 679 } 680 } 681 } 682 } 683 } 684 return nullptr; 685 } 686 687 //------------------------------fold_compares---------------------------- 688 // See if a pair of CmpIs can be converted into a CmpU. In some cases 689 // the direction of this if is determined by the preceding if so it 690 // can be eliminate entirely. 691 // 692 // Given an if testing (CmpI n v) check for an immediately control 693 // dependent if that is testing (CmpI n v2) and has one projection 694 // leading to this if and the other projection leading to a region 695 // that merges one of this ifs control projections. 696 // 697 // If 698 // / | 699 // / | 700 // / | 701 // If | 702 // /\ | 703 // / \ | 704 // / \ | 705 // / Region 706 // 707 // Or given an if testing (CmpI n v) check for a dominating if that is 708 // testing (CmpI n v2), both having one projection leading to an 709 // uncommon trap. Allow Another independent guard in between to cover 710 // an explicit range check: 711 // if (index < 0 || index >= array.length) { 712 // which may need a null check to guard the LoadRange 713 // 714 // If 715 // / \ 716 // / \ 717 // / \ 718 // If unc 719 // /\ 720 // / \ 721 // / \ 722 // / unc 723 // 724 725 // Is the comparison for this If suitable for folding? 726 bool IfNode::cmpi_folds(PhaseIterGVN* igvn, bool fold_ne) { 727 return in(1) != nullptr && 728 in(1)->is_Bool() && 729 in(1)->in(1) != nullptr && 730 in(1)->in(1)->Opcode() == Op_CmpI && 731 in(1)->in(1)->in(2) != nullptr && 732 in(1)->in(1)->in(2) != igvn->C->top() && 733 (in(1)->as_Bool()->_test.is_less() || 734 in(1)->as_Bool()->_test.is_greater() || 735 (fold_ne && in(1)->as_Bool()->_test._test == BoolTest::ne)); 736 } 737 738 // Is a dominating control suitable for folding with this if? 739 bool IfNode::is_ctrl_folds(Node* ctrl, PhaseIterGVN* igvn) { 740 return ctrl != nullptr && 741 ctrl->is_Proj() && 742 ctrl->in(0) != nullptr && 743 ctrl->in(0)->Opcode() == Op_If && 744 ctrl->in(0)->outcnt() == 2 && 745 ctrl->in(0)->as_If()->cmpi_folds(igvn, true) && 746 // Must compare same value 747 ctrl->in(0)->in(1)->in(1)->in(1) != nullptr && 748 ctrl->in(0)->in(1)->in(1)->in(1) != igvn->C->top() && 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_or_null(); 756 RegionNode* region = (otherproj_ctrl_use != nullptr && otherproj_ctrl_use->is_Region()) ? otherproj_ctrl_use->as_Region() : nullptr; 757 success = nullptr; 758 fail = nullptr; 759 760 if (otherproj->outcnt() == 1 && region != nullptr && !region->has_phi()) { 761 for (int i = 0; i < 2; i++) { 762 ProjNode* proj = proj_out(i); 763 if (success == nullptr && proj->outcnt() == 1 && proj->unique_out() == region) { 764 success = proj; 765 } else if (fail == nullptr) { 766 fail = proj; 767 } else { 768 success = fail = nullptr; 769 } 770 } 771 } 772 return success != nullptr && fail != nullptr; 773 } 774 775 bool IfNode::is_dominator_unc(CallStaticJavaNode* dom_unc, CallStaticJavaNode* unc) { 776 // Different methods and methods containing jsrs are not supported. 777 ciMethod* method = unc->jvms()->method(); 778 ciMethod* dom_method = dom_unc->jvms()->method(); 779 if (method != dom_method || method->has_jsrs()) { 780 return false; 781 } 782 // Check that both traps are in the same activation of the method (instead 783 // of two activations being inlined through different call sites) by verifying 784 // that the call stacks are equal for both JVMStates. 785 JVMState* dom_caller = dom_unc->jvms()->caller(); 786 JVMState* caller = unc->jvms()->caller(); 787 if ((dom_caller == nullptr) != (caller == nullptr)) { 788 // The current method must either be inlined into both dom_caller and 789 // caller or must not be inlined at all (top method). Bail out otherwise. 790 return false; 791 } else if (dom_caller != nullptr && !dom_caller->same_calls_as(caller)) { 792 return false; 793 } 794 // Check that the bci of the dominating uncommon trap dominates the bci 795 // of the dominated uncommon trap. Otherwise we may not re-execute 796 // the dominated check after deoptimization from the merged uncommon trap. 797 ciTypeFlow* flow = dom_method->get_flow_analysis(); 798 int bci = unc->jvms()->bci(); 799 int dom_bci = dom_unc->jvms()->bci(); 800 if (!flow->is_dominated_by(bci, dom_bci)) { 801 return false; 802 } 803 804 return true; 805 } 806 807 // Return projection that leads to an uncommon trap if any 808 ProjNode* IfNode::uncommon_trap_proj(CallStaticJavaNode*& call) const { 809 for (int i = 0; i < 2; i++) { 810 call = proj_out(i)->is_uncommon_trap_proj(); 811 if (call != nullptr) { 812 return proj_out(i); 813 } 814 } 815 return nullptr; 816 } 817 818 // Do this If and the dominating If both branch out to an uncommon trap 819 bool IfNode::has_only_uncommon_traps(ProjNode* proj, ProjNode*& success, ProjNode*& fail, PhaseIterGVN* igvn) { 820 ProjNode* otherproj = proj->other_if_proj(); 821 CallStaticJavaNode* dom_unc = otherproj->is_uncommon_trap_proj(); 822 823 if (otherproj->outcnt() == 1 && dom_unc != nullptr) { 824 // We need to re-execute the folded Ifs after deoptimization from the merged traps 825 if (!dom_unc->jvms()->should_reexecute()) { 826 return false; 827 } 828 829 CallStaticJavaNode* unc = nullptr; 830 ProjNode* unc_proj = uncommon_trap_proj(unc); 831 if (unc_proj != nullptr && unc_proj->outcnt() == 1) { 832 if (dom_unc == unc) { 833 // Allow the uncommon trap to be shared through a region 834 RegionNode* r = unc->in(0)->as_Region(); 835 if (r->outcnt() != 2 || r->req() != 3 || r->find_edge(otherproj) == -1 || r->find_edge(unc_proj) == -1) { 836 return false; 837 } 838 assert(r->has_phi() == nullptr, "simple region shouldn't have a phi"); 839 } else if (dom_unc->in(0) != otherproj || unc->in(0) != unc_proj) { 840 return false; 841 } 842 843 if (!is_dominator_unc(dom_unc, unc)) { 844 return false; 845 } 846 847 // See merge_uncommon_traps: the reason of the uncommon trap 848 // will be changed and the state of the dominating If will be 849 // used. Checked that we didn't apply this transformation in a 850 // previous compilation and it didn't cause too many traps 851 ciMethod* dom_method = dom_unc->jvms()->method(); 852 int dom_bci = dom_unc->jvms()->bci(); 853 if (!igvn->C->too_many_traps(dom_method, dom_bci, Deoptimization::Reason_unstable_fused_if) && 854 !igvn->C->too_many_traps(dom_method, dom_bci, Deoptimization::Reason_range_check) && 855 // Return true if c2 manages to reconcile with UnstableIf optimization. See the comments for it. 856 igvn->C->remove_unstable_if_trap(dom_unc, true/*yield*/)) { 857 success = unc_proj; 858 fail = unc_proj->other_if_proj(); 859 return true; 860 } 861 } 862 } 863 return false; 864 } 865 866 // Check that the 2 CmpI can be folded into as single CmpU and proceed with the folding 867 bool IfNode::fold_compares_helper(ProjNode* proj, ProjNode* success, ProjNode* fail, PhaseIterGVN* igvn) { 868 Node* this_cmp = in(1)->in(1); 869 BoolNode* this_bool = in(1)->as_Bool(); 870 IfNode* dom_iff = proj->in(0)->as_If(); 871 BoolNode* dom_bool = dom_iff->in(1)->as_Bool(); 872 Node* lo = dom_iff->in(1)->in(1)->in(2); 873 Node* hi = this_cmp->in(2); 874 Node* n = this_cmp->in(1); 875 ProjNode* otherproj = proj->other_if_proj(); 876 877 const TypeInt* lo_type = IfNode::filtered_int_type(igvn, n, otherproj); 878 const TypeInt* hi_type = IfNode::filtered_int_type(igvn, n, success); 879 880 BoolTest::mask lo_test = dom_bool->_test._test; 881 BoolTest::mask hi_test = this_bool->_test._test; 882 BoolTest::mask cond = hi_test; 883 884 // convert: 885 // 886 // dom_bool = x {<,<=,>,>=} a 887 // / \ 888 // proj = {True,False} / \ otherproj = {False,True} 889 // / 890 // this_bool = x {<,<=} b 891 // / \ 892 // fail = {True,False} / \ success = {False,True} 893 // / 894 // 895 // (Second test guaranteed canonicalized, first one may not have 896 // been canonicalized yet) 897 // 898 // into: 899 // 900 // cond = (x - lo) {<u,<=u,>u,>=u} adjusted_lim 901 // / \ 902 // fail / \ success 903 // / 904 // 905 906 // Figure out which of the two tests sets the upper bound and which 907 // sets the lower bound if any. 908 Node* adjusted_lim = nullptr; 909 if (lo_type != nullptr && hi_type != nullptr && hi_type->_lo > lo_type->_hi && 910 hi_type->_hi == max_jint && lo_type->_lo == min_jint && lo_test != BoolTest::ne) { 911 assert((dom_bool->_test.is_less() && !proj->_con) || 912 (dom_bool->_test.is_greater() && proj->_con), "incorrect test"); 913 914 // this_bool = < 915 // dom_bool = >= (proj = True) or dom_bool = < (proj = False) 916 // x in [a, b[ on the fail (= True) projection, b > a-1 (because of hi_type->_lo > lo_type->_hi test above): 917 // lo = a, hi = b, adjusted_lim = b-a, cond = <u 918 // dom_bool = > (proj = True) or dom_bool = <= (proj = False) 919 // x in ]a, b[ on the fail (= True) projection, b > a: 920 // lo = a+1, hi = b, adjusted_lim = b-a-1, cond = <u 921 // this_bool = <= 922 // dom_bool = >= (proj = True) or dom_bool = < (proj = False) 923 // x in [a, b] on the fail (= True) projection, b+1 > a-1: 924 // lo = a, hi = b, adjusted_lim = b-a+1, cond = <u 925 // lo = a, hi = b, adjusted_lim = b-a, cond = <=u doesn't work because b = a - 1 is possible, then b-a = -1 926 // dom_bool = > (proj = True) or dom_bool = <= (proj = False) 927 // x in ]a, b] on the fail (= True) projection b+1 > a: 928 // lo = a+1, hi = b, adjusted_lim = b-a, cond = <u 929 // lo = a+1, hi = b, adjusted_lim = b-a-1, cond = <=u doesn't work because a = b is possible, then b-a-1 = -1 930 931 if (hi_test == BoolTest::lt) { 932 if (lo_test == BoolTest::gt || lo_test == BoolTest::le) { 933 lo = igvn->transform(new AddINode(lo, igvn->intcon(1))); 934 } 935 } else if (hi_test == BoolTest::le) { 936 if (lo_test == BoolTest::ge || lo_test == BoolTest::lt) { 937 adjusted_lim = igvn->transform(new SubINode(hi, lo)); 938 adjusted_lim = igvn->transform(new AddINode(adjusted_lim, igvn->intcon(1))); 939 cond = BoolTest::lt; 940 } else if (lo_test == BoolTest::gt || lo_test == BoolTest::le) { 941 adjusted_lim = igvn->transform(new SubINode(hi, lo)); 942 lo = igvn->transform(new AddINode(lo, igvn->intcon(1))); 943 cond = BoolTest::lt; 944 } else { 945 assert(false, "unhandled lo_test: %d", lo_test); 946 return false; 947 } 948 } else { 949 assert(igvn->_worklist.member(in(1)) && in(1)->Value(igvn) != igvn->type(in(1)), "unhandled hi_test: %d", hi_test); 950 return false; 951 } 952 // this test was canonicalized 953 assert(this_bool->_test.is_less() && fail->_con, "incorrect test"); 954 } else if (lo_type != nullptr && hi_type != nullptr && lo_type->_lo > hi_type->_hi && 955 lo_type->_hi == max_jint && hi_type->_lo == min_jint && lo_test != BoolTest::ne) { 956 957 // this_bool = < 958 // dom_bool = < (proj = True) or dom_bool = >= (proj = False) 959 // x in [b, a[ on the fail (= False) projection, a > b-1 (because of lo_type->_lo > hi_type->_hi above): 960 // lo = b, hi = a, adjusted_lim = a-b, cond = >=u 961 // dom_bool = <= (proj = True) or dom_bool = > (proj = False) 962 // x in [b, a] on the fail (= False) projection, a+1 > b-1: 963 // lo = b, hi = a, adjusted_lim = a-b+1, cond = >=u 964 // lo = b, hi = a, adjusted_lim = a-b, cond = >u doesn't work because a = b - 1 is possible, then b-a = -1 965 // this_bool = <= 966 // dom_bool = < (proj = True) or dom_bool = >= (proj = False) 967 // x in ]b, a[ on the fail (= False) projection, a > b: 968 // lo = b+1, hi = a, adjusted_lim = a-b-1, cond = >=u 969 // dom_bool = <= (proj = True) or dom_bool = > (proj = False) 970 // x in ]b, a] on the fail (= False) projection, a+1 > b: 971 // lo = b+1, hi = a, adjusted_lim = a-b, cond = >=u 972 // lo = b+1, hi = a, adjusted_lim = a-b-1, cond = >u doesn't work because a = b is possible, then b-a-1 = -1 973 974 swap(lo, hi); 975 swap(lo_type, hi_type); 976 swap(lo_test, hi_test); 977 978 assert((dom_bool->_test.is_less() && proj->_con) || 979 (dom_bool->_test.is_greater() && !proj->_con), "incorrect test"); 980 981 cond = (hi_test == BoolTest::le || hi_test == BoolTest::gt) ? BoolTest::gt : BoolTest::ge; 982 983 if (lo_test == BoolTest::lt) { 984 if (hi_test == BoolTest::lt || hi_test == BoolTest::ge) { 985 cond = BoolTest::ge; 986 } else if (hi_test == BoolTest::le || hi_test == BoolTest::gt) { 987 adjusted_lim = igvn->transform(new SubINode(hi, lo)); 988 adjusted_lim = igvn->transform(new AddINode(adjusted_lim, igvn->intcon(1))); 989 cond = BoolTest::ge; 990 } else { 991 assert(false, "unhandled hi_test: %d", hi_test); 992 return false; 993 } 994 } else if (lo_test == BoolTest::le) { 995 if (hi_test == BoolTest::lt || hi_test == BoolTest::ge) { 996 lo = igvn->transform(new AddINode(lo, igvn->intcon(1))); 997 cond = BoolTest::ge; 998 } else if (hi_test == BoolTest::le || hi_test == BoolTest::gt) { 999 adjusted_lim = igvn->transform(new SubINode(hi, lo)); 1000 lo = igvn->transform(new AddINode(lo, igvn->intcon(1))); 1001 cond = BoolTest::ge; 1002 } else { 1003 assert(false, "unhandled hi_test: %d", hi_test); 1004 return false; 1005 } 1006 } else { 1007 assert(igvn->_worklist.member(in(1)) && in(1)->Value(igvn) != igvn->type(in(1)), "unhandled lo_test: %d", lo_test); 1008 return false; 1009 } 1010 // this test was canonicalized 1011 assert(this_bool->_test.is_less() && !fail->_con, "incorrect test"); 1012 } else { 1013 const TypeInt* failtype = filtered_int_type(igvn, n, proj); 1014 if (failtype != nullptr) { 1015 const TypeInt* type2 = filtered_int_type(igvn, n, fail); 1016 if (type2 != nullptr) { 1017 failtype = failtype->join(type2)->is_int(); 1018 if (failtype->_lo > failtype->_hi) { 1019 // previous if determines the result of this if so 1020 // replace Bool with constant 1021 igvn->replace_input_of(this, 1, igvn->intcon(success->_con)); 1022 return true; 1023 } 1024 } 1025 } 1026 lo = nullptr; 1027 hi = nullptr; 1028 } 1029 1030 if (lo && hi) { 1031 Node* hook = new Node(1); 1032 hook->init_req(0, lo); // Add a use to lo to prevent him from dying 1033 // Merge the two compares into a single unsigned compare by building (CmpU (n - lo) (hi - lo)) 1034 Node* adjusted_val = igvn->transform(new SubINode(n, lo)); 1035 if (adjusted_lim == nullptr) { 1036 adjusted_lim = igvn->transform(new SubINode(hi, lo)); 1037 } 1038 hook->destruct(igvn); 1039 1040 int lo = igvn->type(adjusted_lim)->is_int()->_lo; 1041 if (lo < 0) { 1042 // If range check elimination applies to this comparison, it includes code to protect from overflows that may 1043 // cause the main loop to be skipped entirely. Delay this transformation. 1044 // Example: 1045 // for (int i = 0; i < limit; i++) { 1046 // if (i < max_jint && i > min_jint) {... 1047 // } 1048 // Comparisons folded as: 1049 // i - min_jint - 1 <u -2 1050 // when RC applies, main loop limit becomes: 1051 // min(limit, max(-2 + min_jint + 1, min_jint)) 1052 // = min(limit, min_jint) 1053 // = min_jint 1054 if (!igvn->C->post_loop_opts_phase()) { 1055 if (adjusted_val->outcnt() == 0) { 1056 igvn->remove_dead_node(adjusted_val); 1057 } 1058 if (adjusted_lim->outcnt() == 0) { 1059 igvn->remove_dead_node(adjusted_lim); 1060 } 1061 igvn->C->record_for_post_loop_opts_igvn(this); 1062 return false; 1063 } 1064 } 1065 1066 Node* newcmp = igvn->transform(new CmpUNode(adjusted_val, adjusted_lim)); 1067 Node* newbool = igvn->transform(new BoolNode(newcmp, cond)); 1068 1069 igvn->replace_input_of(dom_iff, 1, igvn->intcon(proj->_con)); 1070 igvn->replace_input_of(this, 1, newbool); 1071 1072 return true; 1073 } 1074 return false; 1075 } 1076 1077 // Merge the branches that trap for this If and the dominating If into 1078 // a single region that branches to the uncommon trap for the 1079 // dominating If 1080 Node* IfNode::merge_uncommon_traps(ProjNode* proj, ProjNode* success, ProjNode* fail, PhaseIterGVN* igvn) { 1081 Node* res = this; 1082 assert(success->in(0) == this, "bad projection"); 1083 1084 ProjNode* otherproj = proj->other_if_proj(); 1085 1086 CallStaticJavaNode* unc = success->is_uncommon_trap_proj(); 1087 CallStaticJavaNode* dom_unc = otherproj->is_uncommon_trap_proj(); 1088 1089 if (unc != dom_unc) { 1090 Node* r = new RegionNode(3); 1091 1092 r->set_req(1, otherproj); 1093 r->set_req(2, success); 1094 r = igvn->transform(r); 1095 assert(r->is_Region(), "can't go away"); 1096 1097 // Make both If trap at the state of the first If: once the CmpI 1098 // nodes are merged, if we trap we don't know which of the CmpI 1099 // nodes would have caused the trap so we have to restart 1100 // execution at the first one 1101 igvn->replace_input_of(dom_unc, 0, r); 1102 igvn->replace_input_of(unc, 0, igvn->C->top()); 1103 } 1104 int trap_request = dom_unc->uncommon_trap_request(); 1105 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 1106 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 1107 1108 int flip_test = 0; 1109 Node* l = nullptr; 1110 Node* r = nullptr; 1111 1112 if (success->in(0)->as_If()->range_check_trap_proj(flip_test, l, r) != nullptr) { 1113 // If this looks like a range check, change the trap to 1114 // Reason_range_check so the compiler recognizes it as a range 1115 // check and applies the corresponding optimizations 1116 trap_request = Deoptimization::make_trap_request(Deoptimization::Reason_range_check, action); 1117 1118 improve_address_types(l, r, fail, igvn); 1119 1120 res = igvn->transform(new RangeCheckNode(in(0), in(1), _prob, _fcnt)); 1121 } else if (unc != dom_unc) { 1122 // If we trap we won't know what CmpI would have caused the trap 1123 // so use a special trap reason to mark this pair of CmpI nodes as 1124 // bad candidate for folding. On recompilation we won't fold them 1125 // and we may trap again but this time we'll know what branch 1126 // traps 1127 trap_request = Deoptimization::make_trap_request(Deoptimization::Reason_unstable_fused_if, action); 1128 } 1129 igvn->replace_input_of(dom_unc, TypeFunc::Parms, igvn->intcon(trap_request)); 1130 return res; 1131 } 1132 1133 // If we are turning 2 CmpI nodes into a CmpU that follows the pattern 1134 // of a rangecheck on index i, on 64 bit the compares may be followed 1135 // by memory accesses using i as index. In that case, the CmpU tells 1136 // us something about the values taken by i that can help the compiler 1137 // (see Compile::conv_I2X_index()) 1138 void IfNode::improve_address_types(Node* l, Node* r, ProjNode* fail, PhaseIterGVN* igvn) { 1139 #ifdef _LP64 1140 ResourceMark rm; 1141 Node_Stack stack(2); 1142 1143 assert(r->Opcode() == Op_LoadRange, "unexpected range check"); 1144 const TypeInt* array_size = igvn->type(r)->is_int(); 1145 1146 stack.push(l, 0); 1147 1148 while(stack.size() > 0) { 1149 Node* n = stack.node(); 1150 uint start = stack.index(); 1151 1152 uint i = start; 1153 for (; i < n->outcnt(); i++) { 1154 Node* use = n->raw_out(i); 1155 if (stack.size() == 1) { 1156 if (use->Opcode() == Op_ConvI2L) { 1157 const TypeLong* bounds = use->as_Type()->type()->is_long(); 1158 if (bounds->_lo <= array_size->_lo && bounds->_hi >= array_size->_hi && 1159 (bounds->_lo != array_size->_lo || bounds->_hi != array_size->_hi)) { 1160 stack.set_index(i+1); 1161 stack.push(use, 0); 1162 break; 1163 } 1164 } 1165 } else if (use->is_Mem()) { 1166 Node* ctrl = use->in(0); 1167 for (int i = 0; i < 10 && ctrl != nullptr && ctrl != fail; i++) { 1168 ctrl = up_one_dom(ctrl); 1169 } 1170 if (ctrl == fail) { 1171 Node* init_n = stack.node_at(1); 1172 assert(init_n->Opcode() == Op_ConvI2L, "unexpected first node"); 1173 // Create a new narrow ConvI2L node that is dependent on the range check 1174 Node* new_n = igvn->C->conv_I2X_index(igvn, l, array_size, fail); 1175 1176 // The type of the ConvI2L may be widen and so the new 1177 // ConvI2L may not be better than an existing ConvI2L 1178 if (new_n != init_n) { 1179 for (uint j = 2; j < stack.size(); j++) { 1180 Node* n = stack.node_at(j); 1181 Node* clone = n->clone(); 1182 int rep = clone->replace_edge(init_n, new_n, igvn); 1183 assert(rep > 0, "can't find expected node?"); 1184 clone = igvn->transform(clone); 1185 init_n = n; 1186 new_n = clone; 1187 } 1188 igvn->hash_delete(use); 1189 int rep = use->replace_edge(init_n, new_n, igvn); 1190 assert(rep > 0, "can't find expected node?"); 1191 igvn->transform(use); 1192 if (init_n->outcnt() == 0) { 1193 igvn->_worklist.push(init_n); 1194 } 1195 } 1196 } 1197 } else if (use->in(0) == nullptr && (igvn->type(use)->isa_long() || 1198 igvn->type(use)->isa_ptr())) { 1199 stack.set_index(i+1); 1200 stack.push(use, 0); 1201 break; 1202 } 1203 } 1204 if (i == n->outcnt()) { 1205 stack.pop(); 1206 } 1207 } 1208 #endif 1209 } 1210 1211 bool IfNode::is_cmp_with_loadrange(ProjNode* proj) { 1212 if (in(1) != nullptr && 1213 in(1)->in(1) != nullptr && 1214 in(1)->in(1)->in(2) != nullptr) { 1215 Node* other = in(1)->in(1)->in(2); 1216 if (other->Opcode() == Op_LoadRange && 1217 ((other->in(0) != nullptr && other->in(0) == proj) || 1218 (other->in(0) == nullptr && 1219 other->in(2) != nullptr && 1220 other->in(2)->is_AddP() && 1221 other->in(2)->in(1) != nullptr && 1222 other->in(2)->in(1)->Opcode() == Op_CastPP && 1223 other->in(2)->in(1)->in(0) == proj))) { 1224 return true; 1225 } 1226 } 1227 return false; 1228 } 1229 1230 bool IfNode::is_null_check(ProjNode* proj, PhaseIterGVN* igvn) { 1231 Node* other = in(1)->in(1)->in(2); 1232 if (other->in(MemNode::Address) != nullptr && 1233 proj->in(0)->in(1) != nullptr && 1234 proj->in(0)->in(1)->is_Bool() && 1235 proj->in(0)->in(1)->in(1) != nullptr && 1236 proj->in(0)->in(1)->in(1)->Opcode() == Op_CmpP && 1237 proj->in(0)->in(1)->in(1)->in(2) != nullptr && 1238 proj->in(0)->in(1)->in(1)->in(1) == other->in(MemNode::Address)->in(AddPNode::Address)->uncast() && 1239 igvn->type(proj->in(0)->in(1)->in(1)->in(2)) == TypePtr::NULL_PTR) { 1240 return true; 1241 } 1242 return false; 1243 } 1244 1245 // Check that the If that is in between the 2 integer comparisons has 1246 // no side effect 1247 bool IfNode::is_side_effect_free_test(ProjNode* proj, PhaseIterGVN* igvn) { 1248 if (proj == nullptr) { 1249 return false; 1250 } 1251 CallStaticJavaNode* unc = proj->is_uncommon_trap_if_pattern(); 1252 if (unc != nullptr && proj->outcnt() <= 2) { 1253 if (proj->outcnt() == 1 || 1254 // Allow simple null check from LoadRange 1255 (is_cmp_with_loadrange(proj) && is_null_check(proj, igvn))) { 1256 CallStaticJavaNode* unc = proj->is_uncommon_trap_if_pattern(); 1257 CallStaticJavaNode* dom_unc = proj->in(0)->in(0)->as_Proj()->is_uncommon_trap_if_pattern(); 1258 assert(dom_unc != nullptr, "is_uncommon_trap_if_pattern returned null"); 1259 1260 // reroute_side_effect_free_unc changes the state of this 1261 // uncommon trap to restart execution at the previous 1262 // CmpI. Check that this change in a previous compilation didn't 1263 // cause too many traps. 1264 int trap_request = unc->uncommon_trap_request(); 1265 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 1266 1267 if (igvn->C->too_many_traps(dom_unc->jvms()->method(), dom_unc->jvms()->bci(), reason)) { 1268 return false; 1269 } 1270 1271 if (!is_dominator_unc(dom_unc, unc)) { 1272 return false; 1273 } 1274 1275 return true; 1276 } 1277 } 1278 return false; 1279 } 1280 1281 // Make the If between the 2 integer comparisons trap at the state of 1282 // the first If: the last CmpI is the one replaced by a CmpU and the 1283 // first CmpI is eliminated, so the test between the 2 CmpI nodes 1284 // won't be guarded by the first CmpI anymore. It can trap in cases 1285 // where the first CmpI would have prevented it from executing: on a 1286 // trap, we need to restart execution at the state of the first CmpI 1287 void IfNode::reroute_side_effect_free_unc(ProjNode* proj, ProjNode* dom_proj, PhaseIterGVN* igvn) { 1288 CallStaticJavaNode* dom_unc = dom_proj->is_uncommon_trap_if_pattern(); 1289 ProjNode* otherproj = proj->other_if_proj(); 1290 CallStaticJavaNode* unc = proj->is_uncommon_trap_if_pattern(); 1291 Node* call_proj = dom_unc->unique_ctrl_out(); 1292 Node* halt = call_proj->unique_ctrl_out(); 1293 1294 Node* new_unc = dom_unc->clone(); 1295 call_proj = call_proj->clone(); 1296 halt = halt->clone(); 1297 Node* c = otherproj->clone(); 1298 1299 c = igvn->transform(c); 1300 new_unc->set_req(TypeFunc::Parms, unc->in(TypeFunc::Parms)); 1301 new_unc->set_req(0, c); 1302 new_unc = igvn->transform(new_unc); 1303 call_proj->set_req(0, new_unc); 1304 call_proj = igvn->transform(call_proj); 1305 halt->set_req(0, call_proj); 1306 halt = igvn->transform(halt); 1307 1308 igvn->replace_node(otherproj, igvn->C->top()); 1309 igvn->C->root()->add_req(halt); 1310 } 1311 1312 Node* IfNode::fold_compares(PhaseIterGVN* igvn) { 1313 if (Opcode() != Op_If) return nullptr; 1314 1315 if (cmpi_folds(igvn)) { 1316 Node* ctrl = in(0); 1317 if (is_ctrl_folds(ctrl, igvn) && ctrl->outcnt() == 1) { 1318 // A integer comparison immediately dominated by another integer 1319 // comparison 1320 ProjNode* success = nullptr; 1321 ProjNode* fail = nullptr; 1322 ProjNode* dom_cmp = ctrl->as_Proj(); 1323 if (has_shared_region(dom_cmp, success, fail) && 1324 // Next call modifies graph so must be last 1325 fold_compares_helper(dom_cmp, success, fail, igvn)) { 1326 return this; 1327 } 1328 if (has_only_uncommon_traps(dom_cmp, success, fail, igvn) && 1329 // Next call modifies graph so must be last 1330 fold_compares_helper(dom_cmp, success, fail, igvn)) { 1331 return merge_uncommon_traps(dom_cmp, success, fail, igvn); 1332 } 1333 return nullptr; 1334 } else if (ctrl->in(0) != nullptr && 1335 ctrl->in(0)->in(0) != nullptr) { 1336 ProjNode* success = nullptr; 1337 ProjNode* fail = nullptr; 1338 Node* dom = ctrl->in(0)->in(0); 1339 ProjNode* dom_cmp = dom->isa_Proj(); 1340 ProjNode* other_cmp = ctrl->isa_Proj(); 1341 1342 // Check if it's an integer comparison dominated by another 1343 // integer comparison with another test in between 1344 if (is_ctrl_folds(dom, igvn) && 1345 has_only_uncommon_traps(dom_cmp, success, fail, igvn) && 1346 is_side_effect_free_test(other_cmp, igvn) && 1347 // Next call modifies graph so must be last 1348 fold_compares_helper(dom_cmp, success, fail, igvn)) { 1349 reroute_side_effect_free_unc(other_cmp, dom_cmp, igvn); 1350 return merge_uncommon_traps(dom_cmp, success, fail, igvn); 1351 } 1352 } 1353 } 1354 return nullptr; 1355 } 1356 1357 //------------------------------remove_useless_bool---------------------------- 1358 // Check for people making a useless boolean: things like 1359 // if( (x < y ? true : false) ) { ... } 1360 // Replace with if( x < y ) { ... } 1361 static Node *remove_useless_bool(IfNode *iff, PhaseGVN *phase) { 1362 Node *i1 = iff->in(1); 1363 if( !i1->is_Bool() ) return nullptr; 1364 BoolNode *bol = i1->as_Bool(); 1365 1366 Node *cmp = bol->in(1); 1367 if( cmp->Opcode() != Op_CmpI ) return nullptr; 1368 1369 // Must be comparing against a bool 1370 const Type *cmp2_t = phase->type( cmp->in(2) ); 1371 if( cmp2_t != TypeInt::ZERO && 1372 cmp2_t != TypeInt::ONE ) 1373 return nullptr; 1374 1375 // Find a prior merge point merging the boolean 1376 i1 = cmp->in(1); 1377 if( !i1->is_Phi() ) return nullptr; 1378 PhiNode *phi = i1->as_Phi(); 1379 if( phase->type( phi ) != TypeInt::BOOL ) 1380 return nullptr; 1381 1382 // Check for diamond pattern 1383 int true_path = phi->is_diamond_phi(); 1384 if( true_path == 0 ) return nullptr; 1385 1386 // Make sure that iff and the control of the phi are different. This 1387 // should really only happen for dead control flow since it requires 1388 // an illegal cycle. 1389 if (phi->in(0)->in(1)->in(0) == iff) return nullptr; 1390 1391 // phi->region->if_proj->ifnode->bool->cmp 1392 BoolNode *bol2 = phi->in(0)->in(1)->in(0)->in(1)->as_Bool(); 1393 1394 // Now get the 'sense' of the test correct so we can plug in 1395 // either iff2->in(1) or its complement. 1396 int flip = 0; 1397 if( bol->_test._test == BoolTest::ne ) flip = 1-flip; 1398 else if( bol->_test._test != BoolTest::eq ) return nullptr; 1399 if( cmp2_t == TypeInt::ZERO ) flip = 1-flip; 1400 1401 const Type *phi1_t = phase->type( phi->in(1) ); 1402 const Type *phi2_t = phase->type( phi->in(2) ); 1403 // Check for Phi(0,1) and flip 1404 if( phi1_t == TypeInt::ZERO ) { 1405 if( phi2_t != TypeInt::ONE ) return nullptr; 1406 flip = 1-flip; 1407 } else { 1408 // Check for Phi(1,0) 1409 if( phi1_t != TypeInt::ONE ) return nullptr; 1410 if( phi2_t != TypeInt::ZERO ) return nullptr; 1411 } 1412 if( true_path == 2 ) { 1413 flip = 1-flip; 1414 } 1415 1416 Node* new_bol = (flip ? phase->transform( bol2->negate(phase) ) : bol2); 1417 assert(new_bol != iff->in(1), "must make progress"); 1418 iff->set_req_X(1, new_bol, phase); 1419 // Intervening diamond probably goes dead 1420 phase->C->set_major_progress(); 1421 return iff; 1422 } 1423 1424 static IfNode* idealize_test(PhaseGVN* phase, IfNode* iff); 1425 1426 struct RangeCheck { 1427 IfProjNode* ctl; 1428 jint off; 1429 }; 1430 1431 Node* IfNode::Ideal_common(PhaseGVN *phase, bool can_reshape) { 1432 if (remove_dead_region(phase, can_reshape)) return this; 1433 // No Def-Use info? 1434 if (!can_reshape) return nullptr; 1435 1436 // Don't bother trying to transform a dead if 1437 if (in(0)->is_top()) return nullptr; 1438 // Don't bother trying to transform an if with a dead test 1439 if (in(1)->is_top()) return nullptr; 1440 // Another variation of a dead test 1441 if (in(1)->is_Con()) return nullptr; 1442 // Another variation of a dead if 1443 if (outcnt() < 2) return nullptr; 1444 1445 // Canonicalize the test. 1446 Node* idt_if = idealize_test(phase, this); 1447 if (idt_if != nullptr) return idt_if; 1448 1449 // Try to split the IF 1450 PhaseIterGVN *igvn = phase->is_IterGVN(); 1451 Node *s = split_if(this, igvn); 1452 if (s != nullptr) return s; 1453 1454 return NodeSentinel; 1455 } 1456 1457 //------------------------------Ideal------------------------------------------ 1458 // Return a node which is more "ideal" than the current node. Strip out 1459 // control copies 1460 Node* IfNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1461 Node* res = Ideal_common(phase, can_reshape); 1462 if (res != NodeSentinel) { 1463 return res; 1464 } 1465 1466 // Check for people making a useless boolean: things like 1467 // if( (x < y ? true : false) ) { ... } 1468 // Replace with if( x < y ) { ... } 1469 Node* bol2 = remove_useless_bool(this, phase); 1470 if (bol2) return bol2; 1471 1472 if (in(0) == nullptr) return nullptr; // Dead loop? 1473 1474 PhaseIterGVN* igvn = phase->is_IterGVN(); 1475 Node* result = fold_compares(igvn); 1476 if (result != nullptr) { 1477 return result; 1478 } 1479 1480 // Scan for an equivalent test 1481 int dist = 4; // Cutoff limit for search 1482 if (is_If() && in(1)->is_Bool()) { 1483 Node* cmp = in(1)->in(1); 1484 if (cmp->Opcode() == Op_CmpP && 1485 cmp->in(2) != nullptr && // make sure cmp is not already dead 1486 cmp->in(2)->bottom_type() == TypePtr::NULL_PTR) { 1487 dist = 64; // Limit for null-pointer scans 1488 } 1489 } 1490 1491 Node* prev_dom = search_identical(dist, igvn); 1492 1493 if (prev_dom != nullptr) { 1494 // Replace dominated IfNode 1495 return dominated_by(prev_dom, igvn, false); 1496 } 1497 1498 return simple_subsuming(igvn); 1499 } 1500 1501 //------------------------------dominated_by----------------------------------- 1502 Node* IfNode::dominated_by(Node* prev_dom, PhaseIterGVN* igvn, bool pin_array_access_nodes) { 1503 #ifndef PRODUCT 1504 if (TraceIterativeGVN) { 1505 tty->print(" Removing IfNode: "); this->dump(); 1506 } 1507 #endif 1508 1509 igvn->hash_delete(this); // Remove self to prevent spurious V-N 1510 Node *idom = in(0); 1511 // Need opcode to decide which way 'this' test goes 1512 int prev_op = prev_dom->Opcode(); 1513 Node *top = igvn->C->top(); // Shortcut to top 1514 1515 // Now walk the current IfNode's projections. 1516 // Loop ends when 'this' has no more uses. 1517 for (DUIterator_Last imin, i = last_outs(imin); i >= imin; --i) { 1518 Node *ifp = last_out(i); // Get IfTrue/IfFalse 1519 igvn->add_users_to_worklist(ifp); 1520 // Check which projection it is and set target. 1521 // Data-target is either the dominating projection of the same type 1522 // or TOP if the dominating projection is of opposite type. 1523 // Data-target will be used as the new control edge for the non-CFG 1524 // nodes like Casts and Loads. 1525 Node *data_target = (ifp->Opcode() == prev_op) ? prev_dom : top; 1526 // Control-target is just the If's immediate dominator or TOP. 1527 Node *ctrl_target = (ifp->Opcode() == prev_op) ? idom : top; 1528 1529 // For each child of an IfTrue/IfFalse projection, reroute. 1530 // Loop ends when projection has no more uses. 1531 for (DUIterator_Last jmin, j = ifp->last_outs(jmin); j >= jmin; --j) { 1532 Node* s = ifp->last_out(j); // Get child of IfTrue/IfFalse 1533 if (s->depends_only_on_test() && igvn->no_dependent_zero_check(s)) { 1534 // For control producers. 1535 // Do not rewire Div and Mod nodes which could have a zero divisor to avoid skipping their zero check. 1536 igvn->replace_input_of(s, 0, data_target); // Move child to data-target 1537 if (pin_array_access_nodes && data_target != top) { 1538 // As a result of range check smearing, Loads and range check Cast nodes that are control dependent on this 1539 // range check (that is about to be removed) now depend on multiple dominating range checks. After the removal 1540 // of this range check, these control dependent nodes end up at the lowest/nearest dominating check in the 1541 // graph. To ensure that these Loads/Casts do not float above any of the dominating checks (even when the 1542 // lowest dominating check is later replaced by yet another dominating check), we need to pin them at the 1543 // lowest dominating check. 1544 Node* clone = s->pin_array_access_node(); 1545 if (clone != nullptr) { 1546 clone = igvn->transform(clone); 1547 igvn->replace_node(s, clone); 1548 } 1549 } 1550 } else { 1551 // Find the control input matching this def-use edge. 1552 // For Regions it may not be in slot 0. 1553 uint l; 1554 for (l = 0; s->in(l) != ifp; l++) { } 1555 igvn->replace_input_of(s, l, ctrl_target); 1556 } 1557 } // End for each child of a projection 1558 1559 igvn->remove_dead_node(ifp); 1560 } // End for each IfTrue/IfFalse child of If 1561 1562 // Kill the IfNode 1563 igvn->remove_dead_node(this); 1564 1565 // Must return either the original node (now dead) or a new node 1566 // (Do not return a top here, since that would break the uniqueness of top.) 1567 return new ConINode(TypeInt::ZERO); 1568 } 1569 1570 Node* IfNode::search_identical(int dist, PhaseIterGVN* igvn) { 1571 // Setup to scan up the CFG looking for a dominating test 1572 Node* dom = in(0); 1573 Node* prev_dom = this; 1574 int op = Opcode(); 1575 // Search up the dominator tree for an If with an identical test 1576 while (dom->Opcode() != op || // Not same opcode? 1577 !same_condition(dom, igvn) || // Not same input 1? 1578 prev_dom->in(0) != dom) { // One path of test does not dominate? 1579 if (dist < 0) return nullptr; 1580 1581 dist--; 1582 prev_dom = dom; 1583 dom = up_one_dom(dom); 1584 if (!dom) return nullptr; 1585 } 1586 1587 // Check that we did not follow a loop back to ourselves 1588 if (this == dom) { 1589 return nullptr; 1590 } 1591 1592 #ifndef PRODUCT 1593 if (dist > 2) { // Add to count of null checks elided 1594 explicit_null_checks_elided++; 1595 } 1596 #endif 1597 1598 return prev_dom; 1599 } 1600 1601 bool IfNode::same_condition(const Node* dom, PhaseIterGVN* igvn) const { 1602 Node* dom_bool = dom->in(1); 1603 Node* this_bool = in(1); 1604 if (dom_bool == this_bool) { 1605 return true; 1606 } 1607 1608 if (dom_bool == nullptr || !dom_bool->is_Bool() || 1609 this_bool == nullptr || !this_bool->is_Bool()) { 1610 return false; 1611 } 1612 Node* dom_cmp = dom_bool->in(1); 1613 Node* this_cmp = this_bool->in(1); 1614 1615 // If the comparison is a subtype check, then SubTypeCheck nodes may have profile data attached to them and may be 1616 // different nodes even-though they perform the same subtype check 1617 if (dom_cmp == nullptr || !dom_cmp->is_SubTypeCheck() || 1618 this_cmp == nullptr || !this_cmp->is_SubTypeCheck()) { 1619 return false; 1620 } 1621 1622 if (dom_cmp->in(1) != this_cmp->in(1) || 1623 dom_cmp->in(2) != this_cmp->in(2) || 1624 dom_bool->as_Bool()->_test._test != this_bool->as_Bool()->_test._test) { 1625 return false; 1626 } 1627 1628 return true; 1629 } 1630 1631 1632 static int subsuming_bool_test_encode(Node*); 1633 1634 // Check if dominating test is subsuming 'this' one. 1635 // 1636 // cmp 1637 // / \ 1638 // (r1) bool \ 1639 // / bool (r2) 1640 // (dom) if \ 1641 // \ ) 1642 // (pre) if[TF] / 1643 // \ / 1644 // if (this) 1645 // \r1 1646 // r2\ eqT eqF neT neF ltT ltF leT leF gtT gtF geT geF 1647 // eq t f f t f - - f f - - f 1648 // ne f t t f t - - t t - - t 1649 // lt f - - f t f - f f - f t 1650 // le t - - t t - t f f t - t 1651 // gt f - - f f - f t t f - f 1652 // ge t - - t f t - t t - t f 1653 // 1654 Node* IfNode::simple_subsuming(PhaseIterGVN* igvn) { 1655 // Table encoding: N/A (na), True-branch (tb), False-branch (fb). 1656 static enum { na, tb, fb } s_short_circuit_map[6][12] = { 1657 /*rel: eq+T eq+F ne+T ne+F lt+T lt+F le+T le+F gt+T gt+F ge+T ge+F*/ 1658 /*eq*/{ tb, fb, fb, tb, fb, na, na, fb, fb, na, na, fb }, 1659 /*ne*/{ fb, tb, tb, fb, tb, na, na, tb, tb, na, na, tb }, 1660 /*lt*/{ fb, na, na, fb, tb, fb, na, fb, fb, na, fb, tb }, 1661 /*le*/{ tb, na, na, tb, tb, na, tb, fb, fb, tb, na, tb }, 1662 /*gt*/{ fb, na, na, fb, fb, na, fb, tb, tb, fb, na, fb }, 1663 /*ge*/{ tb, na, na, tb, fb, tb, na, tb, tb, na, tb, fb }}; 1664 1665 Node* pre = in(0); 1666 if (!pre->is_IfTrue() && !pre->is_IfFalse()) { 1667 return nullptr; 1668 } 1669 Node* dom = pre->in(0); 1670 if (!dom->is_If()) { 1671 return nullptr; 1672 } 1673 Node* bol = in(1); 1674 if (!bol->is_Bool()) { 1675 return nullptr; 1676 } 1677 Node* cmp = in(1)->in(1); 1678 if (!cmp->is_Cmp()) { 1679 return nullptr; 1680 } 1681 1682 if (!dom->in(1)->is_Bool()) { 1683 return nullptr; 1684 } 1685 if (dom->in(1)->in(1) != cmp) { // Not same cond? 1686 return nullptr; 1687 } 1688 1689 int drel = subsuming_bool_test_encode(dom->in(1)); 1690 int trel = subsuming_bool_test_encode(bol); 1691 int bout = pre->is_IfFalse() ? 1 : 0; 1692 1693 if (drel < 0 || trel < 0) { 1694 return nullptr; 1695 } 1696 int br = s_short_circuit_map[trel][2*drel+bout]; 1697 if (br == na) { 1698 return nullptr; 1699 } 1700 #ifndef PRODUCT 1701 if (TraceIterativeGVN) { 1702 tty->print(" Subsumed IfNode: "); dump(); 1703 } 1704 #endif 1705 // Replace condition with constant True(1)/False(0). 1706 bool is_always_true = br == tb; 1707 set_req(1, igvn->intcon(is_always_true ? 1 : 0)); 1708 1709 // Update any data dependencies to the directly dominating test. This subsumed test is not immediately removed by igvn 1710 // and therefore subsequent optimizations might miss these data dependencies otherwise. There might be a dead loop 1711 // ('always_taken_proj' == 'pre') that is cleaned up later. Skip this case to make the iterator work properly. 1712 Node* always_taken_proj = proj_out(is_always_true); 1713 if (always_taken_proj != pre) { 1714 for (DUIterator_Fast imax, i = always_taken_proj->fast_outs(imax); i < imax; i++) { 1715 Node* u = always_taken_proj->fast_out(i); 1716 if (!u->is_CFG()) { 1717 igvn->replace_input_of(u, 0, pre); 1718 --i; 1719 --imax; 1720 } 1721 } 1722 } 1723 1724 if (bol->outcnt() == 0) { 1725 igvn->remove_dead_node(bol); // Kill the BoolNode. 1726 } 1727 return this; 1728 } 1729 1730 // Map BoolTest to local table encoding. The BoolTest (e)numerals 1731 // { eq = 0, ne = 4, le = 5, ge = 7, lt = 3, gt = 1 } 1732 // are mapped to table indices, while the remaining (e)numerals in BoolTest 1733 // { overflow = 2, no_overflow = 6, never = 8, illegal = 9 } 1734 // are ignored (these are not modeled in the table). 1735 // 1736 static int subsuming_bool_test_encode(Node* node) { 1737 precond(node->is_Bool()); 1738 BoolTest::mask x = node->as_Bool()->_test._test; 1739 switch (x) { 1740 case BoolTest::eq: return 0; 1741 case BoolTest::ne: return 1; 1742 case BoolTest::lt: return 2; 1743 case BoolTest::le: return 3; 1744 case BoolTest::gt: return 4; 1745 case BoolTest::ge: return 5; 1746 case BoolTest::overflow: 1747 case BoolTest::no_overflow: 1748 case BoolTest::never: 1749 case BoolTest::illegal: 1750 default: 1751 return -1; 1752 } 1753 } 1754 1755 //------------------------------Identity--------------------------------------- 1756 // If the test is constant & we match, then we are the input Control 1757 Node* IfProjNode::Identity(PhaseGVN* phase) { 1758 // Can only optimize if cannot go the other way 1759 const TypeTuple *t = phase->type(in(0))->is_tuple(); 1760 if (t == TypeTuple::IFNEITHER || (always_taken(t) && 1761 // During parsing (GVN) we don't remove dead code aggressively. 1762 // Cut off dead branch and let PhaseRemoveUseless take care of it. 1763 (!phase->is_IterGVN() || 1764 // During IGVN, first wait for the dead branch to be killed. 1765 // Otherwise, the IfNode's control will have two control uses (the IfNode 1766 // that doesn't go away because it still has uses and this branch of the 1767 // If) which breaks other optimizations. Node::has_special_unique_user() 1768 // will cause this node to be reprocessed once the dead branch is killed. 1769 in(0)->outcnt() == 1))) { 1770 // IfNode control 1771 if (in(0)->is_BaseCountedLoopEnd()) { 1772 // CountedLoopEndNode may be eliminated by if subsuming, replace CountedLoopNode with LoopNode to 1773 // avoid mismatching between CountedLoopNode and CountedLoopEndNode in the following optimization. 1774 Node* head = unique_ctrl_out_or_null(); 1775 if (head != nullptr && head->is_BaseCountedLoop() && head->in(LoopNode::LoopBackControl) == this) { 1776 Node* new_head = new LoopNode(head->in(LoopNode::EntryControl), this); 1777 phase->is_IterGVN()->register_new_node_with_optimizer(new_head); 1778 phase->is_IterGVN()->replace_node(head, new_head); 1779 } 1780 } 1781 return in(0)->in(0); 1782 } 1783 // no progress 1784 return this; 1785 } 1786 1787 bool IfNode::is_zero_trip_guard() const { 1788 if (in(1)->is_Bool() && in(1)->in(1)->is_Cmp()) { 1789 return in(1)->in(1)->in(1)->Opcode() == Op_OpaqueZeroTripGuard; 1790 } 1791 return false; 1792 } 1793 1794 void IfProjNode::pin_array_access_nodes(PhaseIterGVN* igvn) { 1795 for (DUIterator i = outs(); has_out(i); i++) { 1796 Node* u = out(i); 1797 if (!u->depends_only_on_test()) { 1798 continue; 1799 } 1800 Node* clone = u->pin_array_access_node(); 1801 if (clone != nullptr) { 1802 clone = igvn->transform(clone); 1803 assert(clone != u, "shouldn't common"); 1804 igvn->replace_node(u, clone); 1805 --i; 1806 } 1807 } 1808 } 1809 1810 #ifndef PRODUCT 1811 //------------------------------dump_spec-------------------------------------- 1812 void IfNode::dump_spec(outputStream *st) const { 1813 st->print("P=%f, C=%f",_prob,_fcnt); 1814 } 1815 #endif 1816 1817 //------------------------------idealize_test---------------------------------- 1818 // Try to canonicalize tests better. Peek at the Cmp/Bool/If sequence and 1819 // come up with a canonical sequence. Bools getting 'eq', 'gt' and 'ge' forms 1820 // converted to 'ne', 'le' and 'lt' forms. IfTrue/IfFalse get swapped as 1821 // needed. 1822 static IfNode* idealize_test(PhaseGVN* phase, IfNode* iff) { 1823 assert(iff->in(0) != nullptr, "If must be live"); 1824 1825 if (iff->outcnt() != 2) return nullptr; // Malformed projections. 1826 Node* old_if_f = iff->proj_out(false); 1827 Node* old_if_t = iff->proj_out(true); 1828 1829 // CountedLoopEnds want the back-control test to be TRUE, regardless of 1830 // whether they are testing a 'gt' or 'lt' condition. The 'gt' condition 1831 // happens in count-down loops 1832 if (iff->is_BaseCountedLoopEnd()) return nullptr; 1833 if (!iff->in(1)->is_Bool()) return nullptr; // Happens for partially optimized IF tests 1834 BoolNode *b = iff->in(1)->as_Bool(); 1835 BoolTest bt = b->_test; 1836 // Test already in good order? 1837 if( bt.is_canonical() ) 1838 return nullptr; 1839 1840 // Flip test to be canonical. Requires flipping the IfFalse/IfTrue and 1841 // cloning the IfNode. 1842 Node* new_b = phase->transform( new BoolNode(b->in(1), bt.negate()) ); 1843 if( !new_b->is_Bool() ) return nullptr; 1844 b = new_b->as_Bool(); 1845 1846 PhaseIterGVN *igvn = phase->is_IterGVN(); 1847 assert( igvn, "Test is not canonical in parser?" ); 1848 1849 // The IF node never really changes, but it needs to be cloned 1850 iff = iff->clone()->as_If(); 1851 iff->set_req(1, b); 1852 iff->_prob = 1.0-iff->_prob; 1853 1854 Node *prior = igvn->hash_find_insert(iff); 1855 if( prior ) { 1856 igvn->remove_dead_node(iff); 1857 iff = (IfNode*)prior; 1858 } else { 1859 // Cannot call transform on it just yet 1860 igvn->set_type_bottom(iff); 1861 } 1862 igvn->_worklist.push(iff); 1863 1864 // Now handle projections. Cloning not required. 1865 Node* new_if_f = (Node*)(new IfFalseNode( iff )); 1866 Node* new_if_t = (Node*)(new IfTrueNode ( iff )); 1867 1868 igvn->register_new_node_with_optimizer(new_if_f); 1869 igvn->register_new_node_with_optimizer(new_if_t); 1870 // Flip test, so flip trailing control 1871 igvn->replace_node(old_if_f, new_if_t); 1872 igvn->replace_node(old_if_t, new_if_f); 1873 1874 // Progress 1875 return iff; 1876 } 1877 1878 Node* RangeCheckNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1879 Node* res = Ideal_common(phase, can_reshape); 1880 if (res != NodeSentinel) { 1881 return res; 1882 } 1883 1884 PhaseIterGVN *igvn = phase->is_IterGVN(); 1885 // Setup to scan up the CFG looking for a dominating test 1886 Node* prev_dom = this; 1887 1888 // Check for range-check vs other kinds of tests 1889 Node* index1; 1890 Node* range1; 1891 jint offset1; 1892 int flip1 = is_range_check(range1, index1, offset1); 1893 if (flip1) { 1894 Node* dom = in(0); 1895 // Try to remove extra range checks. All 'up_one_dom' gives up at merges 1896 // so all checks we inspect post-dominate the top-most check we find. 1897 // If we are going to fail the current check and we reach the top check 1898 // then we are guaranteed to fail, so just start interpreting there. 1899 // We 'expand' the top 3 range checks to include all post-dominating 1900 // checks. 1901 // 1902 // Example: 1903 // a[i+x] // (1) 1 < x < 6 1904 // a[i+3] // (2) 1905 // a[i+4] // (3) 1906 // a[i+6] // max = max of all constants 1907 // a[i+2] 1908 // a[i+1] // min = min of all constants 1909 // 1910 // If x < 3: 1911 // (1) a[i+x]: Leave unchanged 1912 // (2) a[i+3]: Replace with a[i+max] = a[i+6]: i+x < i+3 <= i+6 -> (2) is covered 1913 // (3) a[i+4]: Replace with a[i+min] = a[i+1]: i+1 < i+4 <= i+6 -> (3) and all following checks are covered 1914 // Remove all other a[i+c] checks 1915 // 1916 // If x >= 3: 1917 // (1) a[i+x]: Leave unchanged 1918 // (2) a[i+3]: Replace with a[i+min] = a[i+1]: i+1 < i+3 <= i+x -> (2) is covered 1919 // (3) a[i+4]: Replace with a[i+max] = a[i+6]: i+1 < i+4 <= i+6 -> (3) and all following checks are covered 1920 // Remove all other a[i+c] checks 1921 // 1922 // We only need the top 2 range checks if x is the min or max of all constants. 1923 // 1924 // This, however, only works if the interval [i+min,i+max] is not larger than max_int (i.e. abs(max - min) < max_int): 1925 // The theoretical max size of an array is max_int with: 1926 // - Valid index space: [0,max_int-1] 1927 // - Invalid index space: [max_int,-1] // max_int, min_int, min_int - 1 ..., -1 1928 // 1929 // The size of the consecutive valid index space is smaller than the size of the consecutive invalid index space. 1930 // If we choose min and max in such a way that: 1931 // - abs(max - min) < max_int 1932 // - i+max and i+min are inside the valid index space 1933 // then all indices [i+min,i+max] must be in the valid index space. Otherwise, the invalid index space must be 1934 // smaller than the valid index space which is never the case for any array size. 1935 // 1936 // Choosing a smaller array size only makes the valid index space smaller and the invalid index space larger and 1937 // the argument above still holds. 1938 // 1939 // Note that the same optimization with the same maximal accepted interval size can also be found in C1. 1940 const jlong maximum_number_of_min_max_interval_indices = (jlong)max_jint; 1941 1942 // The top 3 range checks seen 1943 const int NRC = 3; 1944 RangeCheck prev_checks[NRC]; 1945 int nb_checks = 0; 1946 1947 // Low and high offsets seen so far 1948 jint off_lo = offset1; 1949 jint off_hi = offset1; 1950 1951 bool found_immediate_dominator = false; 1952 1953 // Scan for the top checks and collect range of offsets 1954 for (int dist = 0; dist < 999; dist++) { // Range-Check scan limit 1955 if (dom->Opcode() == Op_RangeCheck && // Not same opcode? 1956 prev_dom->in(0) == dom) { // One path of test does dominate? 1957 if (dom == this) return nullptr; // dead loop 1958 // See if this is a range check 1959 Node* index2; 1960 Node* range2; 1961 jint offset2; 1962 int flip2 = dom->as_RangeCheck()->is_range_check(range2, index2, offset2); 1963 // See if this is a _matching_ range check, checking against 1964 // the same array bounds. 1965 if (flip2 == flip1 && range2 == range1 && index2 == index1 && 1966 dom->outcnt() == 2) { 1967 if (nb_checks == 0 && dom->in(1) == in(1)) { 1968 // Found an immediately dominating test at the same offset. 1969 // This kind of back-to-back test can be eliminated locally, 1970 // and there is no need to search further for dominating tests. 1971 assert(offset2 == offset1, "Same test but different offsets"); 1972 found_immediate_dominator = true; 1973 break; 1974 } 1975 1976 // "x - y" -> must add one to the difference for number of elements in [x,y] 1977 const jlong diff = (jlong)MIN2(offset2, off_lo) - (jlong)MAX2(offset2, off_hi); 1978 if (ABS(diff) < maximum_number_of_min_max_interval_indices) { 1979 // Gather expanded bounds 1980 off_lo = MIN2(off_lo, offset2); 1981 off_hi = MAX2(off_hi, offset2); 1982 // Record top NRC range checks 1983 prev_checks[nb_checks % NRC].ctl = prev_dom->as_IfProj(); 1984 prev_checks[nb_checks % NRC].off = offset2; 1985 nb_checks++; 1986 } 1987 } 1988 } 1989 prev_dom = dom; 1990 dom = up_one_dom(dom); 1991 if (!dom) break; 1992 } 1993 1994 if (!found_immediate_dominator) { 1995 // Attempt to widen the dominating range check to cover some later 1996 // ones. Since range checks "fail" by uncommon-trapping to the 1997 // interpreter, widening a check can make us speculatively enter 1998 // the interpreter. If we see range-check deopt's, do not widen! 1999 if (!phase->C->allow_range_check_smearing()) return nullptr; 2000 2001 if (can_reshape && !phase->C->post_loop_opts_phase()) { 2002 // We are about to perform range check smearing (i.e. remove this RangeCheck if it is dominated by 2003 // a series of RangeChecks which have a range that covers this RangeCheck). This can cause array access nodes to 2004 // be pinned. We want to avoid that and first allow range check elimination a chance to remove the RangeChecks 2005 // from loops. Hence, we delay range check smearing until after loop opts. 2006 phase->C->record_for_post_loop_opts_igvn(this); 2007 return nullptr; 2008 } 2009 2010 // Didn't find prior covering check, so cannot remove anything. 2011 if (nb_checks == 0) { 2012 return nullptr; 2013 } 2014 // Constant indices only need to check the upper bound. 2015 // Non-constant indices must check both low and high. 2016 int chk0 = (nb_checks - 1) % NRC; 2017 if (index1) { 2018 if (nb_checks == 1) { 2019 return nullptr; 2020 } else { 2021 // If the top range check's constant is the min or max of 2022 // all constants we widen the next one to cover the whole 2023 // range of constants. 2024 RangeCheck rc0 = prev_checks[chk0]; 2025 int chk1 = (nb_checks - 2) % NRC; 2026 RangeCheck rc1 = prev_checks[chk1]; 2027 if (rc0.off == off_lo) { 2028 adjust_check(rc1.ctl, range1, index1, flip1, off_hi, igvn); 2029 prev_dom = rc1.ctl; 2030 } else if (rc0.off == off_hi) { 2031 adjust_check(rc1.ctl, range1, index1, flip1, off_lo, igvn); 2032 prev_dom = rc1.ctl; 2033 } else { 2034 // If the top test's constant is not the min or max of all 2035 // constants, we need 3 range checks. We must leave the 2036 // top test unchanged because widening it would allow the 2037 // accesses it protects to successfully read/write out of 2038 // bounds. 2039 if (nb_checks == 2) { 2040 return nullptr; 2041 } 2042 int chk2 = (nb_checks - 3) % NRC; 2043 RangeCheck rc2 = prev_checks[chk2]; 2044 // The top range check a+i covers interval: -a <= i < length-a 2045 // The second range check b+i covers interval: -b <= i < length-b 2046 if (rc1.off <= rc0.off) { 2047 // if b <= a, we change the second range check to: 2048 // -min_of_all_constants <= i < length-min_of_all_constants 2049 // Together top and second range checks now cover: 2050 // -min_of_all_constants <= i < length-a 2051 // which is more restrictive than -b <= i < length-b: 2052 // -b <= -min_of_all_constants <= i < length-a <= length-b 2053 // The third check is then changed to: 2054 // -max_of_all_constants <= i < length-max_of_all_constants 2055 // so 2nd and 3rd checks restrict allowed values of i to: 2056 // -min_of_all_constants <= i < length-max_of_all_constants 2057 adjust_check(rc1.ctl, range1, index1, flip1, off_lo, igvn); 2058 adjust_check(rc2.ctl, range1, index1, flip1, off_hi, igvn); 2059 } else { 2060 // if b > a, we change the second range check to: 2061 // -max_of_all_constants <= i < length-max_of_all_constants 2062 // Together top and second range checks now cover: 2063 // -a <= i < length-max_of_all_constants 2064 // which is more restrictive than -b <= i < length-b: 2065 // -b < -a <= i < length-max_of_all_constants <= length-b 2066 // The third check is then changed to: 2067 // -max_of_all_constants <= i < length-max_of_all_constants 2068 // so 2nd and 3rd checks restrict allowed values of i to: 2069 // -min_of_all_constants <= i < length-max_of_all_constants 2070 adjust_check(rc1.ctl, range1, index1, flip1, off_hi, igvn); 2071 adjust_check(rc2.ctl, range1, index1, flip1, off_lo, igvn); 2072 } 2073 prev_dom = rc2.ctl; 2074 } 2075 } 2076 } else { 2077 RangeCheck rc0 = prev_checks[chk0]; 2078 // 'Widen' the offset of the 1st and only covering check 2079 adjust_check(rc0.ctl, range1, index1, flip1, off_hi, igvn); 2080 // Test is now covered by prior checks, dominate it out 2081 prev_dom = rc0.ctl; 2082 } 2083 // The last RangeCheck is found to be redundant with a sequence of n (n >= 2) preceding RangeChecks. 2084 // If an array load is control dependent on the eliminated range check, the array load nodes (CastII and Load) 2085 // become control dependent on the last range check of the sequence, but they are really dependent on the entire 2086 // sequence of RangeChecks. If RangeCheck#n is later replaced by a dominating identical check, the array load 2087 // nodes must not float above the n-1 other RangeCheck in the sequence. We pin the array load nodes here to 2088 // guarantee it doesn't happen. 2089 // 2090 // RangeCheck#1 RangeCheck#1 2091 // | \ | \ 2092 // | uncommon trap | uncommon trap 2093 // .. .. 2094 // RangeCheck#n -> RangeCheck#n 2095 // | \ | \ 2096 // | uncommon trap CastII uncommon trap 2097 // RangeCheck Load 2098 // | \ 2099 // CastII uncommon trap 2100 // Load 2101 2102 return dominated_by(prev_dom, igvn, true); 2103 } 2104 } else { 2105 prev_dom = search_identical(4, igvn); 2106 2107 if (prev_dom == nullptr) { 2108 return nullptr; 2109 } 2110 } 2111 2112 // Replace dominated IfNode 2113 return dominated_by(prev_dom, igvn, false); 2114 } 2115 2116 ParsePredicateNode::ParsePredicateNode(Node* control, Deoptimization::DeoptReason deopt_reason, PhaseGVN* gvn) 2117 : IfNode(control, gvn->intcon(1), PROB_MAX, COUNT_UNKNOWN), 2118 _deopt_reason(deopt_reason), 2119 _useless(false) { 2120 init_class_id(Class_ParsePredicate); 2121 gvn->C->add_parse_predicate(this); 2122 gvn->C->record_for_post_loop_opts_igvn(this); 2123 #ifdef ASSERT 2124 switch (deopt_reason) { 2125 case Deoptimization::Reason_predicate: 2126 case Deoptimization::Reason_profile_predicate: 2127 case Deoptimization::Reason_loop_limit_check: 2128 break; 2129 default: 2130 assert(false, "unsupported deoptimization reason for Parse Predicate"); 2131 } 2132 #endif // ASSERT 2133 } 2134 2135 Node* ParsePredicateNode::uncommon_trap() const { 2136 ParsePredicateUncommonProj* uncommon_proj = proj_out(0)->as_IfFalse(); 2137 Node* uct_region_or_call = uncommon_proj->unique_ctrl_out(); 2138 assert(uct_region_or_call->is_Region() || uct_region_or_call->is_Call(), "must be a region or call uct"); 2139 return uct_region_or_call; 2140 } 2141 2142 // Fold this node away once it becomes useless or at latest in post loop opts IGVN. 2143 const Type* ParsePredicateNode::Value(PhaseGVN* phase) const { 2144 if (phase->type(in(0)) == Type::TOP) { 2145 return Type::TOP; 2146 } 2147 if (_useless || phase->C->post_loop_opts_phase()) { 2148 return TypeTuple::IFTRUE; 2149 } else { 2150 return bottom_type(); 2151 } 2152 } 2153 2154 #ifndef PRODUCT 2155 void ParsePredicateNode::dump_spec(outputStream* st) const { 2156 st->print(" #"); 2157 switch (_deopt_reason) { 2158 case Deoptimization::DeoptReason::Reason_predicate: 2159 st->print("Loop "); 2160 break; 2161 case Deoptimization::DeoptReason::Reason_profile_predicate: 2162 st->print("Profiled_Loop "); 2163 break; 2164 case Deoptimization::DeoptReason::Reason_loop_limit_check: 2165 st->print("Loop_Limit_Check "); 2166 break; 2167 default: 2168 fatal("unknown kind"); 2169 } 2170 } 2171 2172 #endif // NOT PRODUCT