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(Node* 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 } 570 571 //------------------------------up_one_dom------------------------------------- 572 // Walk up the dominator tree one step. Return null at root or true 573 // complex merges. Skips through small diamonds. 574 Node* IfNode::up_one_dom(Node *curr, bool linear_only) { 575 Node *dom = curr->in(0); 576 if( !dom ) // Found a Region degraded to a copy? 577 return curr->nonnull_req(); // Skip thru it 578 579 if( curr != dom ) // Normal walk up one step? 580 return dom; 581 582 // Use linear_only if we are still parsing, since we cannot 583 // trust the regions to be fully filled in. 584 if (linear_only) 585 return nullptr; 586 587 if( dom->is_Root() ) 588 return nullptr; 589 590 // Else hit a Region. Check for a loop header 591 if( dom->is_Loop() ) 592 return dom->in(1); // Skip up thru loops 593 594 // Check for small diamonds 595 Node *din1, *din2, *din3, *din4; 596 if( dom->req() == 3 && // 2-path merge point 597 (din1 = dom ->in(1)) && // Left path exists 598 (din2 = dom ->in(2)) && // Right path exists 599 (din3 = din1->in(0)) && // Left path up one 600 (din4 = din2->in(0)) ) { // Right path up one 601 if( din3->is_Call() && // Handle a slow-path call on either arm 602 (din3 = din3->in(0)) ) 603 din3 = din3->in(0); 604 if( din4->is_Call() && // Handle a slow-path call on either arm 605 (din4 = din4->in(0)) ) 606 din4 = din4->in(0); 607 if (din3 != nullptr && din3 == din4 && din3->is_If()) // Regions not degraded to a copy 608 return din3; // Skip around diamonds 609 } 610 611 // Give up the search at true merges 612 return nullptr; // Dead loop? Or hit root? 613 } 614 615 616 //------------------------------filtered_int_type-------------------------------- 617 // Return a possibly more restrictive type for val based on condition control flow for an if 618 const TypeInt* IfNode::filtered_int_type(PhaseGVN* gvn, Node* val, Node* if_proj) { 619 assert(if_proj && 620 (if_proj->Opcode() == Op_IfTrue || if_proj->Opcode() == Op_IfFalse), "expecting an if projection"); 621 if (if_proj->in(0) && if_proj->in(0)->is_If()) { 622 IfNode* iff = if_proj->in(0)->as_If(); 623 if (iff->in(1) && iff->in(1)->is_Bool()) { 624 BoolNode* bol = iff->in(1)->as_Bool(); 625 if (bol->in(1) && bol->in(1)->is_Cmp()) { 626 const CmpNode* cmp = bol->in(1)->as_Cmp(); 627 if (cmp->in(1) == val) { 628 const TypeInt* cmp2_t = gvn->type(cmp->in(2))->isa_int(); 629 if (cmp2_t != nullptr) { 630 jint lo = cmp2_t->_lo; 631 jint hi = cmp2_t->_hi; 632 BoolTest::mask msk = if_proj->Opcode() == Op_IfTrue ? bol->_test._test : bol->_test.negate(); 633 switch (msk) { 634 case BoolTest::ne: { 635 // If val is compared to its lower or upper bound, we can narrow the type 636 const TypeInt* val_t = gvn->type(val)->isa_int(); 637 if (val_t != nullptr && !val_t->singleton() && cmp2_t->is_con()) { 638 if (val_t->_lo == lo) { 639 return TypeInt::make(val_t->_lo + 1, val_t->_hi, val_t->_widen); 640 } else if (val_t->_hi == hi) { 641 return TypeInt::make(val_t->_lo, val_t->_hi - 1, val_t->_widen); 642 } 643 } 644 // Can't refine type 645 return nullptr; 646 } 647 case BoolTest::eq: 648 return cmp2_t; 649 case BoolTest::lt: 650 lo = TypeInt::INT->_lo; 651 if (hi != min_jint) { 652 hi = hi - 1; 653 } 654 break; 655 case BoolTest::le: 656 lo = TypeInt::INT->_lo; 657 break; 658 case BoolTest::gt: 659 if (lo != max_jint) { 660 lo = lo + 1; 661 } 662 hi = TypeInt::INT->_hi; 663 break; 664 case BoolTest::ge: 665 // lo unchanged 666 hi = TypeInt::INT->_hi; 667 break; 668 default: 669 break; 670 } 671 const TypeInt* rtn_t = TypeInt::make(lo, hi, cmp2_t->_widen); 672 return rtn_t; 673 } 674 } 675 } 676 } 677 } 678 return nullptr; 679 } 680 681 //------------------------------fold_compares---------------------------- 682 // See if a pair of CmpIs can be converted into a CmpU. In some cases 683 // the direction of this if is determined by the preceding if so it 684 // can be eliminate entirely. 685 // 686 // Given an if testing (CmpI n v) check for an immediately control 687 // dependent if that is testing (CmpI n v2) and has one projection 688 // leading to this if and the other projection leading to a region 689 // that merges one of this ifs control projections. 690 // 691 // If 692 // / | 693 // / | 694 // / | 695 // If | 696 // /\ | 697 // / \ | 698 // / \ | 699 // / Region 700 // 701 // Or given an if testing (CmpI n v) check for a dominating if that is 702 // testing (CmpI n v2), both having one projection leading to an 703 // uncommon trap. Allow Another independent guard in between to cover 704 // an explicit range check: 705 // if (index < 0 || index >= array.length) { 706 // which may need a null check to guard the LoadRange 707 // 708 // If 709 // / \ 710 // / \ 711 // / \ 712 // If unc 713 // /\ 714 // / \ 715 // / \ 716 // / unc 717 // 718 719 // Is the comparison for this If suitable for folding? 720 bool IfNode::cmpi_folds(PhaseIterGVN* igvn, bool fold_ne) { 721 return in(1) != nullptr && 722 in(1)->is_Bool() && 723 in(1)->in(1) != nullptr && 724 in(1)->in(1)->Opcode() == Op_CmpI && 725 in(1)->in(1)->in(2) != nullptr && 726 in(1)->in(1)->in(2) != igvn->C->top() && 727 (in(1)->as_Bool()->_test.is_less() || 728 in(1)->as_Bool()->_test.is_greater() || 729 (fold_ne && in(1)->as_Bool()->_test._test == BoolTest::ne)); 730 } 731 732 // Is a dominating control suitable for folding with this if? 733 bool IfNode::is_ctrl_folds(Node* ctrl, PhaseIterGVN* igvn) { 734 return ctrl != nullptr && 735 ctrl->is_Proj() && 736 ctrl->in(0) != nullptr && 737 ctrl->in(0)->Opcode() == Op_If && 738 ctrl->in(0)->outcnt() == 2 && 739 ctrl->in(0)->as_If()->cmpi_folds(igvn, true) && 740 // Must compare same value 741 ctrl->in(0)->in(1)->in(1)->in(1) != nullptr && 742 ctrl->in(0)->in(1)->in(1)->in(1) != igvn->C->top() && 743 ctrl->in(0)->in(1)->in(1)->in(1) == in(1)->in(1)->in(1); 744 } 745 746 // Do this If and the dominating If share a region? 747 bool IfNode::has_shared_region(ProjNode* proj, ProjNode*& success, ProjNode*& fail) { 748 ProjNode* otherproj = proj->other_if_proj(); 749 Node* otherproj_ctrl_use = otherproj->unique_ctrl_out_or_null(); 750 RegionNode* region = (otherproj_ctrl_use != nullptr && otherproj_ctrl_use->is_Region()) ? otherproj_ctrl_use->as_Region() : nullptr; 751 success = nullptr; 752 fail = nullptr; 753 754 if (otherproj->outcnt() == 1 && region != nullptr && !region->has_phi()) { 755 for (int i = 0; i < 2; i++) { 756 ProjNode* proj = proj_out(i); 757 if (success == nullptr && proj->outcnt() == 1 && proj->unique_out() == region) { 758 success = proj; 759 } else if (fail == nullptr) { 760 fail = proj; 761 } else { 762 success = fail = nullptr; 763 } 764 } 765 } 766 return success != nullptr && fail != nullptr; 767 } 768 769 bool IfNode::is_dominator_unc(CallStaticJavaNode* dom_unc, CallStaticJavaNode* unc) { 770 // Different methods and methods containing jsrs are not supported. 771 ciMethod* method = unc->jvms()->method(); 772 ciMethod* dom_method = dom_unc->jvms()->method(); 773 if (method != dom_method || method->has_jsrs()) { 774 return false; 775 } 776 // Check that both traps are in the same activation of the method (instead 777 // of two activations being inlined through different call sites) by verifying 778 // that the call stacks are equal for both JVMStates. 779 JVMState* dom_caller = dom_unc->jvms()->caller(); 780 JVMState* caller = unc->jvms()->caller(); 781 if ((dom_caller == nullptr) != (caller == nullptr)) { 782 // The current method must either be inlined into both dom_caller and 783 // caller or must not be inlined at all (top method). Bail out otherwise. 784 return false; 785 } else if (dom_caller != nullptr && !dom_caller->same_calls_as(caller)) { 786 return false; 787 } 788 // Check that the bci of the dominating uncommon trap dominates the bci 789 // of the dominated uncommon trap. Otherwise we may not re-execute 790 // the dominated check after deoptimization from the merged uncommon trap. 791 ciTypeFlow* flow = dom_method->get_flow_analysis(); 792 int bci = unc->jvms()->bci(); 793 int dom_bci = dom_unc->jvms()->bci(); 794 if (!flow->is_dominated_by(bci, dom_bci)) { 795 return false; 796 } 797 798 return true; 799 } 800 801 // Return projection that leads to an uncommon trap if any 802 ProjNode* IfNode::uncommon_trap_proj(CallStaticJavaNode*& call) const { 803 for (int i = 0; i < 2; i++) { 804 call = proj_out(i)->is_uncommon_trap_proj(); 805 if (call != nullptr) { 806 return proj_out(i); 807 } 808 } 809 return nullptr; 810 } 811 812 // Do this If and the dominating If both branch out to an uncommon trap 813 bool IfNode::has_only_uncommon_traps(ProjNode* proj, ProjNode*& success, ProjNode*& fail, PhaseIterGVN* igvn) { 814 ProjNode* otherproj = proj->other_if_proj(); 815 CallStaticJavaNode* dom_unc = otherproj->is_uncommon_trap_proj(); 816 817 if (otherproj->outcnt() == 1 && dom_unc != nullptr) { 818 // We need to re-execute the folded Ifs after deoptimization from the merged traps 819 if (!dom_unc->jvms()->should_reexecute()) { 820 return false; 821 } 822 823 CallStaticJavaNode* unc = nullptr; 824 ProjNode* unc_proj = uncommon_trap_proj(unc); 825 if (unc_proj != nullptr && unc_proj->outcnt() == 1) { 826 if (dom_unc == unc) { 827 // Allow the uncommon trap to be shared through a region 828 RegionNode* r = unc->in(0)->as_Region(); 829 if (r->outcnt() != 2 || r->req() != 3 || r->find_edge(otherproj) == -1 || r->find_edge(unc_proj) == -1) { 830 return false; 831 } 832 assert(r->has_phi() == nullptr, "simple region shouldn't have a phi"); 833 } else if (dom_unc->in(0) != otherproj || unc->in(0) != unc_proj) { 834 return false; 835 } 836 837 if (!is_dominator_unc(dom_unc, unc)) { 838 return false; 839 } 840 841 // See merge_uncommon_traps: the reason of the uncommon trap 842 // will be changed and the state of the dominating If will be 843 // used. Checked that we didn't apply this transformation in a 844 // previous compilation and it didn't cause too many traps 845 ciMethod* dom_method = dom_unc->jvms()->method(); 846 int dom_bci = dom_unc->jvms()->bci(); 847 if (!igvn->C->too_many_traps(dom_method, dom_bci, Deoptimization::Reason_unstable_fused_if) && 848 !igvn->C->too_many_traps(dom_method, dom_bci, Deoptimization::Reason_range_check) && 849 // Return true if c2 manages to reconcile with UnstableIf optimization. See the comments for it. 850 igvn->C->remove_unstable_if_trap(dom_unc, true/*yield*/)) { 851 success = unc_proj; 852 fail = unc_proj->other_if_proj(); 853 return true; 854 } 855 } 856 } 857 return false; 858 } 859 860 // Check that the 2 CmpI can be folded into as single CmpU and proceed with the folding 861 bool IfNode::fold_compares_helper(ProjNode* proj, ProjNode* success, ProjNode* fail, PhaseIterGVN* igvn) { 862 Node* this_cmp = in(1)->in(1); 863 BoolNode* this_bool = in(1)->as_Bool(); 864 IfNode* dom_iff = proj->in(0)->as_If(); 865 BoolNode* dom_bool = dom_iff->in(1)->as_Bool(); 866 Node* lo = dom_iff->in(1)->in(1)->in(2); 867 Node* hi = this_cmp->in(2); 868 Node* n = this_cmp->in(1); 869 ProjNode* otherproj = proj->other_if_proj(); 870 871 const TypeInt* lo_type = IfNode::filtered_int_type(igvn, n, otherproj); 872 const TypeInt* hi_type = IfNode::filtered_int_type(igvn, n, success); 873 874 BoolTest::mask lo_test = dom_bool->_test._test; 875 BoolTest::mask hi_test = this_bool->_test._test; 876 BoolTest::mask cond = hi_test; 877 878 // convert: 879 // 880 // dom_bool = x {<,<=,>,>=} a 881 // / \ 882 // proj = {True,False} / \ otherproj = {False,True} 883 // / 884 // this_bool = x {<,<=} b 885 // / \ 886 // fail = {True,False} / \ success = {False,True} 887 // / 888 // 889 // (Second test guaranteed canonicalized, first one may not have 890 // been canonicalized yet) 891 // 892 // into: 893 // 894 // cond = (x - lo) {<u,<=u,>u,>=u} adjusted_lim 895 // / \ 896 // fail / \ success 897 // / 898 // 899 900 // Figure out which of the two tests sets the upper bound and which 901 // sets the lower bound if any. 902 Node* adjusted_lim = nullptr; 903 if (lo_type != nullptr && hi_type != nullptr && hi_type->_lo > lo_type->_hi && 904 hi_type->_hi == max_jint && lo_type->_lo == min_jint && lo_test != BoolTest::ne) { 905 assert((dom_bool->_test.is_less() && !proj->_con) || 906 (dom_bool->_test.is_greater() && proj->_con), "incorrect test"); 907 908 // this_bool = < 909 // dom_bool = >= (proj = True) or dom_bool = < (proj = False) 910 // x in [a, b[ on the fail (= True) projection, b > a-1 (because of hi_type->_lo > lo_type->_hi test above): 911 // lo = a, hi = b, adjusted_lim = b-a, cond = <u 912 // dom_bool = > (proj = True) or dom_bool = <= (proj = False) 913 // x in ]a, b[ on the fail (= True) projection, b > a: 914 // lo = a+1, hi = b, adjusted_lim = b-a-1, cond = <u 915 // this_bool = <= 916 // dom_bool = >= (proj = True) or dom_bool = < (proj = False) 917 // x in [a, b] on the fail (= True) projection, b+1 > a-1: 918 // lo = a, hi = b, adjusted_lim = b-a+1, cond = <u 919 // lo = a, hi = b, adjusted_lim = b-a, cond = <=u doesn't work because b = a - 1 is possible, then b-a = -1 920 // dom_bool = > (proj = True) or dom_bool = <= (proj = False) 921 // x in ]a, b] on the fail (= True) projection b+1 > a: 922 // lo = a+1, hi = b, adjusted_lim = b-a, cond = <u 923 // lo = a+1, hi = b, adjusted_lim = b-a-1, cond = <=u doesn't work because a = b is possible, then b-a-1 = -1 924 925 if (hi_test == BoolTest::lt) { 926 if (lo_test == BoolTest::gt || lo_test == BoolTest::le) { 927 lo = igvn->transform(new AddINode(lo, igvn->intcon(1))); 928 } 929 } else if (hi_test == BoolTest::le) { 930 if (lo_test == BoolTest::ge || lo_test == BoolTest::lt) { 931 adjusted_lim = igvn->transform(new SubINode(hi, lo)); 932 adjusted_lim = igvn->transform(new AddINode(adjusted_lim, igvn->intcon(1))); 933 cond = BoolTest::lt; 934 } else if (lo_test == BoolTest::gt || lo_test == BoolTest::le) { 935 adjusted_lim = igvn->transform(new SubINode(hi, lo)); 936 lo = igvn->transform(new AddINode(lo, igvn->intcon(1))); 937 cond = BoolTest::lt; 938 } else { 939 assert(false, "unhandled lo_test: %d", lo_test); 940 return false; 941 } 942 } else { 943 assert(igvn->_worklist.member(in(1)) && in(1)->Value(igvn) != igvn->type(in(1)), "unhandled hi_test: %d", hi_test); 944 return false; 945 } 946 // this test was canonicalized 947 assert(this_bool->_test.is_less() && fail->_con, "incorrect test"); 948 } else if (lo_type != nullptr && hi_type != nullptr && lo_type->_lo > hi_type->_hi && 949 lo_type->_hi == max_jint && hi_type->_lo == min_jint && lo_test != BoolTest::ne) { 950 951 // this_bool = < 952 // dom_bool = < (proj = True) or dom_bool = >= (proj = False) 953 // x in [b, a[ on the fail (= False) projection, a > b-1 (because of lo_type->_lo > hi_type->_hi above): 954 // lo = b, hi = a, adjusted_lim = a-b, cond = >=u 955 // dom_bool = <= (proj = True) or dom_bool = > (proj = False) 956 // x in [b, a] on the fail (= False) projection, a+1 > b-1: 957 // lo = b, hi = a, adjusted_lim = a-b+1, cond = >=u 958 // lo = b, hi = a, adjusted_lim = a-b, cond = >u doesn't work because a = b - 1 is possible, then b-a = -1 959 // this_bool = <= 960 // dom_bool = < (proj = True) or dom_bool = >= (proj = False) 961 // x in ]b, a[ on the fail (= False) projection, a > b: 962 // lo = b+1, hi = a, adjusted_lim = a-b-1, cond = >=u 963 // dom_bool = <= (proj = True) or dom_bool = > (proj = False) 964 // x in ]b, a] on the fail (= False) projection, a+1 > b: 965 // lo = b+1, hi = a, adjusted_lim = a-b, cond = >=u 966 // lo = b+1, hi = a, adjusted_lim = a-b-1, cond = >u doesn't work because a = b is possible, then b-a-1 = -1 967 968 swap(lo, hi); 969 swap(lo_type, hi_type); 970 swap(lo_test, hi_test); 971 972 assert((dom_bool->_test.is_less() && proj->_con) || 973 (dom_bool->_test.is_greater() && !proj->_con), "incorrect test"); 974 975 cond = (hi_test == BoolTest::le || hi_test == BoolTest::gt) ? BoolTest::gt : BoolTest::ge; 976 977 if (lo_test == BoolTest::lt) { 978 if (hi_test == BoolTest::lt || hi_test == BoolTest::ge) { 979 cond = BoolTest::ge; 980 } else if (hi_test == BoolTest::le || hi_test == BoolTest::gt) { 981 adjusted_lim = igvn->transform(new SubINode(hi, lo)); 982 adjusted_lim = igvn->transform(new AddINode(adjusted_lim, igvn->intcon(1))); 983 cond = BoolTest::ge; 984 } else { 985 assert(false, "unhandled hi_test: %d", hi_test); 986 return false; 987 } 988 } else if (lo_test == BoolTest::le) { 989 if (hi_test == BoolTest::lt || hi_test == BoolTest::ge) { 990 lo = igvn->transform(new AddINode(lo, igvn->intcon(1))); 991 cond = BoolTest::ge; 992 } else if (hi_test == BoolTest::le || hi_test == BoolTest::gt) { 993 adjusted_lim = igvn->transform(new SubINode(hi, lo)); 994 lo = igvn->transform(new AddINode(lo, igvn->intcon(1))); 995 cond = BoolTest::ge; 996 } else { 997 assert(false, "unhandled hi_test: %d", hi_test); 998 return false; 999 } 1000 } else { 1001 assert(igvn->_worklist.member(in(1)) && in(1)->Value(igvn) != igvn->type(in(1)), "unhandled lo_test: %d", lo_test); 1002 return false; 1003 } 1004 // this test was canonicalized 1005 assert(this_bool->_test.is_less() && !fail->_con, "incorrect test"); 1006 } else { 1007 const TypeInt* failtype = filtered_int_type(igvn, n, proj); 1008 if (failtype != nullptr) { 1009 const TypeInt* type2 = filtered_int_type(igvn, n, fail); 1010 if (type2 != nullptr) { 1011 failtype = failtype->join(type2)->is_int(); 1012 if (failtype->_lo > failtype->_hi) { 1013 // previous if determines the result of this if so 1014 // replace Bool with constant 1015 igvn->replace_input_of(this, 1, igvn->intcon(success->_con)); 1016 return true; 1017 } 1018 } 1019 } 1020 lo = nullptr; 1021 hi = nullptr; 1022 } 1023 1024 if (lo && hi) { 1025 Node* hook = new Node(1); 1026 hook->init_req(0, lo); // Add a use to lo to prevent him from dying 1027 // Merge the two compares into a single unsigned compare by building (CmpU (n - lo) (hi - lo)) 1028 Node* adjusted_val = igvn->transform(new SubINode(n, lo)); 1029 if (adjusted_lim == nullptr) { 1030 adjusted_lim = igvn->transform(new SubINode(hi, lo)); 1031 } 1032 hook->destruct(igvn); 1033 1034 int lo = igvn->type(adjusted_lim)->is_int()->_lo; 1035 if (lo < 0) { 1036 // If range check elimination applies to this comparison, it includes code to protect from overflows that may 1037 // cause the main loop to be skipped entirely. Delay this transformation. 1038 // Example: 1039 // for (int i = 0; i < limit; i++) { 1040 // if (i < max_jint && i > min_jint) {... 1041 // } 1042 // Comparisons folded as: 1043 // i - min_jint - 1 <u -2 1044 // when RC applies, main loop limit becomes: 1045 // min(limit, max(-2 + min_jint + 1, min_jint)) 1046 // = min(limit, min_jint) 1047 // = min_jint 1048 if (!igvn->C->post_loop_opts_phase()) { 1049 if (adjusted_val->outcnt() == 0) { 1050 igvn->remove_dead_node(adjusted_val); 1051 } 1052 if (adjusted_lim->outcnt() == 0) { 1053 igvn->remove_dead_node(adjusted_lim); 1054 } 1055 igvn->C->record_for_post_loop_opts_igvn(this); 1056 return false; 1057 } 1058 } 1059 1060 Node* newcmp = igvn->transform(new CmpUNode(adjusted_val, adjusted_lim)); 1061 Node* newbool = igvn->transform(new BoolNode(newcmp, cond)); 1062 1063 igvn->replace_input_of(dom_iff, 1, igvn->intcon(proj->_con)); 1064 igvn->replace_input_of(this, 1, newbool); 1065 1066 return true; 1067 } 1068 return false; 1069 } 1070 1071 // Merge the branches that trap for this If and the dominating If into 1072 // a single region that branches to the uncommon trap for the 1073 // dominating If 1074 Node* IfNode::merge_uncommon_traps(ProjNode* proj, ProjNode* success, ProjNode* fail, PhaseIterGVN* igvn) { 1075 Node* res = this; 1076 assert(success->in(0) == this, "bad projection"); 1077 1078 ProjNode* otherproj = proj->other_if_proj(); 1079 1080 CallStaticJavaNode* unc = success->is_uncommon_trap_proj(); 1081 CallStaticJavaNode* dom_unc = otherproj->is_uncommon_trap_proj(); 1082 1083 if (unc != dom_unc) { 1084 Node* r = new RegionNode(3); 1085 1086 r->set_req(1, otherproj); 1087 r->set_req(2, success); 1088 r = igvn->transform(r); 1089 assert(r->is_Region(), "can't go away"); 1090 1091 // Make both If trap at the state of the first If: once the CmpI 1092 // nodes are merged, if we trap we don't know which of the CmpI 1093 // nodes would have caused the trap so we have to restart 1094 // execution at the first one 1095 igvn->replace_input_of(dom_unc, 0, r); 1096 igvn->replace_input_of(unc, 0, igvn->C->top()); 1097 } 1098 int trap_request = dom_unc->uncommon_trap_request(); 1099 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 1100 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 1101 1102 int flip_test = 0; 1103 Node* l = nullptr; 1104 Node* r = nullptr; 1105 1106 if (success->in(0)->as_If()->range_check_trap_proj(flip_test, l, r) != nullptr) { 1107 // If this looks like a range check, change the trap to 1108 // Reason_range_check so the compiler recognizes it as a range 1109 // check and applies the corresponding optimizations 1110 trap_request = Deoptimization::make_trap_request(Deoptimization::Reason_range_check, action); 1111 1112 improve_address_types(l, r, fail, igvn); 1113 1114 res = igvn->transform(new RangeCheckNode(in(0), in(1), _prob, _fcnt)); 1115 } else if (unc != dom_unc) { 1116 // If we trap we won't know what CmpI would have caused the trap 1117 // so use a special trap reason to mark this pair of CmpI nodes as 1118 // bad candidate for folding. On recompilation we won't fold them 1119 // and we may trap again but this time we'll know what branch 1120 // traps 1121 trap_request = Deoptimization::make_trap_request(Deoptimization::Reason_unstable_fused_if, action); 1122 } 1123 igvn->replace_input_of(dom_unc, TypeFunc::Parms, igvn->intcon(trap_request)); 1124 return res; 1125 } 1126 1127 // If we are turning 2 CmpI nodes into a CmpU that follows the pattern 1128 // of a rangecheck on index i, on 64 bit the compares may be followed 1129 // by memory accesses using i as index. In that case, the CmpU tells 1130 // us something about the values taken by i that can help the compiler 1131 // (see Compile::conv_I2X_index()) 1132 void IfNode::improve_address_types(Node* l, Node* r, ProjNode* fail, PhaseIterGVN* igvn) { 1133 #ifdef _LP64 1134 ResourceMark rm; 1135 Node_Stack stack(2); 1136 1137 assert(r->Opcode() == Op_LoadRange, "unexpected range check"); 1138 const TypeInt* array_size = igvn->type(r)->is_int(); 1139 1140 stack.push(l, 0); 1141 1142 while(stack.size() > 0) { 1143 Node* n = stack.node(); 1144 uint start = stack.index(); 1145 1146 uint i = start; 1147 for (; i < n->outcnt(); i++) { 1148 Node* use = n->raw_out(i); 1149 if (stack.size() == 1) { 1150 if (use->Opcode() == Op_ConvI2L) { 1151 const TypeLong* bounds = use->as_Type()->type()->is_long(); 1152 if (bounds->_lo <= array_size->_lo && bounds->_hi >= array_size->_hi && 1153 (bounds->_lo != array_size->_lo || bounds->_hi != array_size->_hi)) { 1154 stack.set_index(i+1); 1155 stack.push(use, 0); 1156 break; 1157 } 1158 } 1159 } else if (use->is_Mem()) { 1160 Node* ctrl = use->in(0); 1161 for (int i = 0; i < 10 && ctrl != nullptr && ctrl != fail; i++) { 1162 ctrl = up_one_dom(ctrl); 1163 } 1164 if (ctrl == fail) { 1165 Node* init_n = stack.node_at(1); 1166 assert(init_n->Opcode() == Op_ConvI2L, "unexpected first node"); 1167 // Create a new narrow ConvI2L node that is dependent on the range check 1168 Node* new_n = igvn->C->conv_I2X_index(igvn, l, array_size, fail); 1169 1170 // The type of the ConvI2L may be widen and so the new 1171 // ConvI2L may not be better than an existing ConvI2L 1172 if (new_n != init_n) { 1173 for (uint j = 2; j < stack.size(); j++) { 1174 Node* n = stack.node_at(j); 1175 Node* clone = n->clone(); 1176 int rep = clone->replace_edge(init_n, new_n, igvn); 1177 assert(rep > 0, "can't find expected node?"); 1178 clone = igvn->transform(clone); 1179 init_n = n; 1180 new_n = clone; 1181 } 1182 igvn->hash_delete(use); 1183 int rep = use->replace_edge(init_n, new_n, igvn); 1184 assert(rep > 0, "can't find expected node?"); 1185 igvn->transform(use); 1186 if (init_n->outcnt() == 0) { 1187 igvn->_worklist.push(init_n); 1188 } 1189 } 1190 } 1191 } else if (use->in(0) == nullptr && (igvn->type(use)->isa_long() || 1192 igvn->type(use)->isa_ptr())) { 1193 stack.set_index(i+1); 1194 stack.push(use, 0); 1195 break; 1196 } 1197 } 1198 if (i == n->outcnt()) { 1199 stack.pop(); 1200 } 1201 } 1202 #endif 1203 } 1204 1205 bool IfNode::is_cmp_with_loadrange(ProjNode* proj) { 1206 if (in(1) != nullptr && 1207 in(1)->in(1) != nullptr && 1208 in(1)->in(1)->in(2) != nullptr) { 1209 Node* other = in(1)->in(1)->in(2); 1210 if (other->Opcode() == Op_LoadRange && 1211 ((other->in(0) != nullptr && other->in(0) == proj) || 1212 (other->in(0) == nullptr && 1213 other->in(2) != nullptr && 1214 other->in(2)->is_AddP() && 1215 other->in(2)->in(1) != nullptr && 1216 other->in(2)->in(1)->Opcode() == Op_CastPP && 1217 other->in(2)->in(1)->in(0) == proj))) { 1218 return true; 1219 } 1220 } 1221 return false; 1222 } 1223 1224 bool IfNode::is_null_check(ProjNode* proj, PhaseIterGVN* igvn) { 1225 Node* other = in(1)->in(1)->in(2); 1226 if (other->in(MemNode::Address) != nullptr && 1227 proj->in(0)->in(1) != nullptr && 1228 proj->in(0)->in(1)->is_Bool() && 1229 proj->in(0)->in(1)->in(1) != nullptr && 1230 proj->in(0)->in(1)->in(1)->Opcode() == Op_CmpP && 1231 proj->in(0)->in(1)->in(1)->in(2) != nullptr && 1232 proj->in(0)->in(1)->in(1)->in(1) == other->in(MemNode::Address)->in(AddPNode::Address)->uncast() && 1233 igvn->type(proj->in(0)->in(1)->in(1)->in(2)) == TypePtr::NULL_PTR) { 1234 return true; 1235 } 1236 return false; 1237 } 1238 1239 // Check that the If that is in between the 2 integer comparisons has 1240 // no side effect 1241 bool IfNode::is_side_effect_free_test(ProjNode* proj, PhaseIterGVN* igvn) { 1242 if (proj == nullptr) { 1243 return false; 1244 } 1245 CallStaticJavaNode* unc = proj->is_uncommon_trap_if_pattern(); 1246 if (unc != nullptr && proj->outcnt() <= 2) { 1247 if (proj->outcnt() == 1 || 1248 // Allow simple null check from LoadRange 1249 (is_cmp_with_loadrange(proj) && is_null_check(proj, igvn))) { 1250 CallStaticJavaNode* unc = proj->is_uncommon_trap_if_pattern(); 1251 CallStaticJavaNode* dom_unc = proj->in(0)->in(0)->as_Proj()->is_uncommon_trap_if_pattern(); 1252 assert(dom_unc != nullptr, "is_uncommon_trap_if_pattern returned null"); 1253 1254 // reroute_side_effect_free_unc changes the state of this 1255 // uncommon trap to restart execution at the previous 1256 // CmpI. Check that this change in a previous compilation didn't 1257 // cause too many traps. 1258 int trap_request = unc->uncommon_trap_request(); 1259 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 1260 1261 if (igvn->C->too_many_traps(dom_unc->jvms()->method(), dom_unc->jvms()->bci(), reason)) { 1262 return false; 1263 } 1264 1265 if (!is_dominator_unc(dom_unc, unc)) { 1266 return false; 1267 } 1268 1269 return true; 1270 } 1271 } 1272 return false; 1273 } 1274 1275 // Make the If between the 2 integer comparisons trap at the state of 1276 // the first If: the last CmpI is the one replaced by a CmpU and the 1277 // first CmpI is eliminated, so the test between the 2 CmpI nodes 1278 // won't be guarded by the first CmpI anymore. It can trap in cases 1279 // where the first CmpI would have prevented it from executing: on a 1280 // trap, we need to restart execution at the state of the first CmpI 1281 void IfNode::reroute_side_effect_free_unc(ProjNode* proj, ProjNode* dom_proj, PhaseIterGVN* igvn) { 1282 CallStaticJavaNode* dom_unc = dom_proj->is_uncommon_trap_if_pattern(); 1283 ProjNode* otherproj = proj->other_if_proj(); 1284 CallStaticJavaNode* unc = proj->is_uncommon_trap_if_pattern(); 1285 Node* call_proj = dom_unc->unique_ctrl_out(); 1286 Node* halt = call_proj->unique_ctrl_out(); 1287 1288 Node* new_unc = dom_unc->clone(); 1289 call_proj = call_proj->clone(); 1290 halt = halt->clone(); 1291 Node* c = otherproj->clone(); 1292 1293 c = igvn->transform(c); 1294 new_unc->set_req(TypeFunc::Parms, unc->in(TypeFunc::Parms)); 1295 new_unc->set_req(0, c); 1296 new_unc = igvn->transform(new_unc); 1297 call_proj->set_req(0, new_unc); 1298 call_proj = igvn->transform(call_proj); 1299 halt->set_req(0, call_proj); 1300 halt = igvn->transform(halt); 1301 1302 igvn->replace_node(otherproj, igvn->C->top()); 1303 igvn->C->root()->add_req(halt); 1304 } 1305 1306 Node* IfNode::fold_compares(PhaseIterGVN* igvn) { 1307 if (Opcode() != Op_If) return nullptr; 1308 1309 if (cmpi_folds(igvn)) { 1310 Node* ctrl = in(0); 1311 if (is_ctrl_folds(ctrl, igvn) && ctrl->outcnt() == 1) { 1312 // A integer comparison immediately dominated by another integer 1313 // comparison 1314 ProjNode* success = nullptr; 1315 ProjNode* fail = nullptr; 1316 ProjNode* dom_cmp = ctrl->as_Proj(); 1317 if (has_shared_region(dom_cmp, success, fail) && 1318 // Next call modifies graph so must be last 1319 fold_compares_helper(dom_cmp, success, fail, igvn)) { 1320 return this; 1321 } 1322 if (has_only_uncommon_traps(dom_cmp, success, fail, igvn) && 1323 // Next call modifies graph so must be last 1324 fold_compares_helper(dom_cmp, success, fail, igvn)) { 1325 return merge_uncommon_traps(dom_cmp, success, fail, igvn); 1326 } 1327 return nullptr; 1328 } else if (ctrl->in(0) != nullptr && 1329 ctrl->in(0)->in(0) != nullptr) { 1330 ProjNode* success = nullptr; 1331 ProjNode* fail = nullptr; 1332 Node* dom = ctrl->in(0)->in(0); 1333 ProjNode* dom_cmp = dom->isa_Proj(); 1334 ProjNode* other_cmp = ctrl->isa_Proj(); 1335 1336 // Check if it's an integer comparison dominated by another 1337 // integer comparison with another test in between 1338 if (is_ctrl_folds(dom, igvn) && 1339 has_only_uncommon_traps(dom_cmp, success, fail, igvn) && 1340 is_side_effect_free_test(other_cmp, igvn) && 1341 // Next call modifies graph so must be last 1342 fold_compares_helper(dom_cmp, success, fail, igvn)) { 1343 reroute_side_effect_free_unc(other_cmp, dom_cmp, igvn); 1344 return merge_uncommon_traps(dom_cmp, success, fail, igvn); 1345 } 1346 } 1347 } 1348 return nullptr; 1349 } 1350 1351 //------------------------------remove_useless_bool---------------------------- 1352 // Check for people making a useless boolean: things like 1353 // if( (x < y ? true : false) ) { ... } 1354 // Replace with if( x < y ) { ... } 1355 static Node *remove_useless_bool(IfNode *iff, PhaseGVN *phase) { 1356 Node *i1 = iff->in(1); 1357 if( !i1->is_Bool() ) return nullptr; 1358 BoolNode *bol = i1->as_Bool(); 1359 1360 Node *cmp = bol->in(1); 1361 if( cmp->Opcode() != Op_CmpI ) return nullptr; 1362 1363 // Must be comparing against a bool 1364 const Type *cmp2_t = phase->type( cmp->in(2) ); 1365 if( cmp2_t != TypeInt::ZERO && 1366 cmp2_t != TypeInt::ONE ) 1367 return nullptr; 1368 1369 // Find a prior merge point merging the boolean 1370 i1 = cmp->in(1); 1371 if( !i1->is_Phi() ) return nullptr; 1372 PhiNode *phi = i1->as_Phi(); 1373 if( phase->type( phi ) != TypeInt::BOOL ) 1374 return nullptr; 1375 1376 // Check for diamond pattern 1377 int true_path = phi->is_diamond_phi(); 1378 if( true_path == 0 ) return nullptr; 1379 1380 // Make sure that iff and the control of the phi are different. This 1381 // should really only happen for dead control flow since it requires 1382 // an illegal cycle. 1383 if (phi->in(0)->in(1)->in(0) == iff) return nullptr; 1384 1385 // phi->region->if_proj->ifnode->bool->cmp 1386 BoolNode *bol2 = phi->in(0)->in(1)->in(0)->in(1)->as_Bool(); 1387 1388 // Now get the 'sense' of the test correct so we can plug in 1389 // either iff2->in(1) or its complement. 1390 int flip = 0; 1391 if( bol->_test._test == BoolTest::ne ) flip = 1-flip; 1392 else if( bol->_test._test != BoolTest::eq ) return nullptr; 1393 if( cmp2_t == TypeInt::ZERO ) flip = 1-flip; 1394 1395 const Type *phi1_t = phase->type( phi->in(1) ); 1396 const Type *phi2_t = phase->type( phi->in(2) ); 1397 // Check for Phi(0,1) and flip 1398 if( phi1_t == TypeInt::ZERO ) { 1399 if( phi2_t != TypeInt::ONE ) return nullptr; 1400 flip = 1-flip; 1401 } else { 1402 // Check for Phi(1,0) 1403 if( phi1_t != TypeInt::ONE ) return nullptr; 1404 if( phi2_t != TypeInt::ZERO ) return nullptr; 1405 } 1406 if( true_path == 2 ) { 1407 flip = 1-flip; 1408 } 1409 1410 Node* new_bol = (flip ? phase->transform( bol2->negate(phase) ) : bol2); 1411 assert(new_bol != iff->in(1), "must make progress"); 1412 iff->set_req_X(1, new_bol, phase); 1413 // Intervening diamond probably goes dead 1414 phase->C->set_major_progress(); 1415 return iff; 1416 } 1417 1418 static IfNode* idealize_test(PhaseGVN* phase, IfNode* iff); 1419 1420 struct RangeCheck { 1421 Node* ctl; 1422 jint off; 1423 }; 1424 1425 Node* IfNode::Ideal_common(PhaseGVN *phase, bool can_reshape) { 1426 if (remove_dead_region(phase, can_reshape)) return this; 1427 // No Def-Use info? 1428 if (!can_reshape) return nullptr; 1429 1430 // Don't bother trying to transform a dead if 1431 if (in(0)->is_top()) return nullptr; 1432 // Don't bother trying to transform an if with a dead test 1433 if (in(1)->is_top()) return nullptr; 1434 // Another variation of a dead test 1435 if (in(1)->is_Con()) return nullptr; 1436 // Another variation of a dead if 1437 if (outcnt() < 2) return nullptr; 1438 1439 // Canonicalize the test. 1440 Node* idt_if = idealize_test(phase, this); 1441 if (idt_if != nullptr) return idt_if; 1442 1443 // Try to split the IF 1444 PhaseIterGVN *igvn = phase->is_IterGVN(); 1445 Node *s = split_if(this, igvn); 1446 if (s != nullptr) return s; 1447 1448 return NodeSentinel; 1449 } 1450 1451 //------------------------------Ideal------------------------------------------ 1452 // Return a node which is more "ideal" than the current node. Strip out 1453 // control copies 1454 Node* IfNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1455 Node* res = Ideal_common(phase, can_reshape); 1456 if (res != NodeSentinel) { 1457 return res; 1458 } 1459 1460 // Check for people making a useless boolean: things like 1461 // if( (x < y ? true : false) ) { ... } 1462 // Replace with if( x < y ) { ... } 1463 Node* bol2 = remove_useless_bool(this, phase); 1464 if (bol2) return bol2; 1465 1466 if (in(0) == nullptr) return nullptr; // Dead loop? 1467 1468 PhaseIterGVN* igvn = phase->is_IterGVN(); 1469 Node* result = fold_compares(igvn); 1470 if (result != nullptr) { 1471 return result; 1472 } 1473 1474 // Scan for an equivalent test 1475 int dist = 4; // Cutoff limit for search 1476 if (is_If() && in(1)->is_Bool()) { 1477 Node* cmp = in(1)->in(1); 1478 if (cmp->Opcode() == Op_CmpP && 1479 cmp->in(2) != nullptr && // make sure cmp is not already dead 1480 cmp->in(2)->bottom_type() == TypePtr::NULL_PTR) { 1481 dist = 64; // Limit for null-pointer scans 1482 } 1483 } 1484 1485 Node* prev_dom = search_identical(dist, igvn); 1486 1487 if (prev_dom != nullptr) { 1488 // Replace dominated IfNode 1489 return dominated_by(prev_dom, igvn); 1490 } 1491 1492 return simple_subsuming(igvn); 1493 } 1494 1495 //------------------------------dominated_by----------------------------------- 1496 Node* IfNode::dominated_by(Node* prev_dom, PhaseIterGVN *igvn) { 1497 #ifndef PRODUCT 1498 if (TraceIterativeGVN) { 1499 tty->print(" Removing IfNode: "); this->dump(); 1500 } 1501 #endif 1502 1503 igvn->hash_delete(this); // Remove self to prevent spurious V-N 1504 Node *idom = in(0); 1505 // Need opcode to decide which way 'this' test goes 1506 int prev_op = prev_dom->Opcode(); 1507 Node *top = igvn->C->top(); // Shortcut to top 1508 1509 // Loop predicates may have depending checks which should not 1510 // be skipped. For example, range check predicate has two checks 1511 // for lower and upper bounds. 1512 ProjNode* unc_proj = proj_out(1 - prev_dom->as_Proj()->_con)->as_Proj(); 1513 if (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) != nullptr || 1514 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_profile_predicate) != nullptr) { 1515 prev_dom = idom; 1516 } 1517 1518 // Now walk the current IfNode's projections. 1519 // Loop ends when 'this' has no more uses. 1520 for (DUIterator_Last imin, i = last_outs(imin); i >= imin; --i) { 1521 Node *ifp = last_out(i); // Get IfTrue/IfFalse 1522 igvn->add_users_to_worklist(ifp); 1523 // Check which projection it is and set target. 1524 // Data-target is either the dominating projection of the same type 1525 // or TOP if the dominating projection is of opposite type. 1526 // Data-target will be used as the new control edge for the non-CFG 1527 // nodes like Casts and Loads. 1528 Node *data_target = (ifp->Opcode() == prev_op) ? prev_dom : top; 1529 // Control-target is just the If's immediate dominator or TOP. 1530 Node *ctrl_target = (ifp->Opcode() == prev_op) ? idom : top; 1531 1532 // For each child of an IfTrue/IfFalse projection, reroute. 1533 // Loop ends when projection has no more uses. 1534 for (DUIterator_Last jmin, j = ifp->last_outs(jmin); j >= jmin; --j) { 1535 Node* s = ifp->last_out(j); // Get child of IfTrue/IfFalse 1536 if (s->depends_only_on_test() && igvn->no_dependent_zero_check(s)) { 1537 // For control producers. 1538 // Do not rewire Div and Mod nodes which could have a zero divisor to avoid skipping their zero check. 1539 igvn->replace_input_of(s, 0, data_target); // Move child to data-target 1540 } else { 1541 // Find the control input matching this def-use edge. 1542 // For Regions it may not be in slot 0. 1543 uint l; 1544 for (l = 0; s->in(l) != ifp; l++) { } 1545 igvn->replace_input_of(s, l, ctrl_target); 1546 } 1547 } // End for each child of a projection 1548 1549 igvn->remove_dead_node(ifp); 1550 } // End for each IfTrue/IfFalse child of If 1551 1552 // Kill the IfNode 1553 igvn->remove_dead_node(this); 1554 1555 // Must return either the original node (now dead) or a new node 1556 // (Do not return a top here, since that would break the uniqueness of top.) 1557 return new ConINode(TypeInt::ZERO); 1558 } 1559 1560 Node* IfNode::search_identical(int dist, PhaseIterGVN* igvn) { 1561 // Setup to scan up the CFG looking for a dominating test 1562 Node* dom = in(0); 1563 Node* prev_dom = this; 1564 int op = Opcode(); 1565 // Search up the dominator tree for an If with an identical test 1566 while (dom->Opcode() != op || // Not same opcode? 1567 !same_condition(dom, igvn) || // Not same input 1? 1568 prev_dom->in(0) != dom) { // One path of test does not dominate? 1569 if (dist < 0) return nullptr; 1570 1571 dist--; 1572 prev_dom = dom; 1573 dom = up_one_dom(dom); 1574 if (!dom) return nullptr; 1575 } 1576 1577 // Check that we did not follow a loop back to ourselves 1578 if (this == dom) { 1579 return nullptr; 1580 } 1581 1582 #ifndef PRODUCT 1583 if (dist > 2) { // Add to count of null checks elided 1584 explicit_null_checks_elided++; 1585 } 1586 #endif 1587 1588 return prev_dom; 1589 } 1590 1591 bool IfNode::same_condition(const Node* dom, PhaseIterGVN* igvn) const { 1592 Node* dom_bool = dom->in(1); 1593 Node* this_bool = in(1); 1594 if (dom_bool == this_bool) { 1595 return true; 1596 } 1597 1598 if (dom_bool == nullptr || !dom_bool->is_Bool() || 1599 this_bool == nullptr || !this_bool->is_Bool()) { 1600 return false; 1601 } 1602 Node* dom_cmp = dom_bool->in(1); 1603 Node* this_cmp = this_bool->in(1); 1604 1605 // If the comparison is a subtype check, then SubTypeCheck nodes may have profile data attached to them and may be 1606 // different nodes even-though they perform the same subtype check 1607 if (dom_cmp == nullptr || !dom_cmp->is_SubTypeCheck() || 1608 this_cmp == nullptr || !this_cmp->is_SubTypeCheck()) { 1609 return false; 1610 } 1611 1612 if (dom_cmp->in(1) != this_cmp->in(1) || 1613 dom_cmp->in(2) != this_cmp->in(2) || 1614 dom_bool->as_Bool()->_test._test != this_bool->as_Bool()->_test._test) { 1615 return false; 1616 } 1617 1618 return true; 1619 } 1620 1621 1622 static int subsuming_bool_test_encode(Node*); 1623 1624 // Check if dominating test is subsuming 'this' one. 1625 // 1626 // cmp 1627 // / \ 1628 // (r1) bool \ 1629 // / bool (r2) 1630 // (dom) if \ 1631 // \ ) 1632 // (pre) if[TF] / 1633 // \ / 1634 // if (this) 1635 // \r1 1636 // r2\ eqT eqF neT neF ltT ltF leT leF gtT gtF geT geF 1637 // eq t f f t f - - f f - - f 1638 // ne f t t f t - - t t - - t 1639 // lt f - - f t f - f f - f t 1640 // le t - - t t - t f f t - t 1641 // gt f - - f f - f t t f - f 1642 // ge t - - t f t - t t - t f 1643 // 1644 Node* IfNode::simple_subsuming(PhaseIterGVN* igvn) { 1645 // Table encoding: N/A (na), True-branch (tb), False-branch (fb). 1646 static enum { na, tb, fb } s_short_circuit_map[6][12] = { 1647 /*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*/ 1648 /*eq*/{ tb, fb, fb, tb, fb, na, na, fb, fb, na, na, fb }, 1649 /*ne*/{ fb, tb, tb, fb, tb, na, na, tb, tb, na, na, tb }, 1650 /*lt*/{ fb, na, na, fb, tb, fb, na, fb, fb, na, fb, tb }, 1651 /*le*/{ tb, na, na, tb, tb, na, tb, fb, fb, tb, na, tb }, 1652 /*gt*/{ fb, na, na, fb, fb, na, fb, tb, tb, fb, na, fb }, 1653 /*ge*/{ tb, na, na, tb, fb, tb, na, tb, tb, na, tb, fb }}; 1654 1655 Node* pre = in(0); 1656 if (!pre->is_IfTrue() && !pre->is_IfFalse()) { 1657 return nullptr; 1658 } 1659 Node* dom = pre->in(0); 1660 if (!dom->is_If()) { 1661 return nullptr; 1662 } 1663 Node* bol = in(1); 1664 if (!bol->is_Bool()) { 1665 return nullptr; 1666 } 1667 Node* cmp = in(1)->in(1); 1668 if (!cmp->is_Cmp()) { 1669 return nullptr; 1670 } 1671 1672 if (!dom->in(1)->is_Bool()) { 1673 return nullptr; 1674 } 1675 if (dom->in(1)->in(1) != cmp) { // Not same cond? 1676 return nullptr; 1677 } 1678 1679 int drel = subsuming_bool_test_encode(dom->in(1)); 1680 int trel = subsuming_bool_test_encode(bol); 1681 int bout = pre->is_IfFalse() ? 1 : 0; 1682 1683 if (drel < 0 || trel < 0) { 1684 return nullptr; 1685 } 1686 int br = s_short_circuit_map[trel][2*drel+bout]; 1687 if (br == na) { 1688 return nullptr; 1689 } 1690 #ifndef PRODUCT 1691 if (TraceIterativeGVN) { 1692 tty->print(" Subsumed IfNode: "); dump(); 1693 } 1694 #endif 1695 // Replace condition with constant True(1)/False(0). 1696 bool is_always_true = br == tb; 1697 set_req(1, igvn->intcon(is_always_true ? 1 : 0)); 1698 1699 // Update any data dependencies to the directly dominating test. This subsumed test is not immediately removed by igvn 1700 // and therefore subsequent optimizations might miss these data dependencies otherwise. There might be a dead loop 1701 // ('always_taken_proj' == 'pre') that is cleaned up later. Skip this case to make the iterator work properly. 1702 Node* always_taken_proj = proj_out(is_always_true); 1703 if (always_taken_proj != pre) { 1704 for (DUIterator_Fast imax, i = always_taken_proj->fast_outs(imax); i < imax; i++) { 1705 Node* u = always_taken_proj->fast_out(i); 1706 if (!u->is_CFG()) { 1707 igvn->replace_input_of(u, 0, pre); 1708 --i; 1709 --imax; 1710 } 1711 } 1712 } 1713 1714 if (bol->outcnt() == 0) { 1715 igvn->remove_dead_node(bol); // Kill the BoolNode. 1716 } 1717 return this; 1718 } 1719 1720 // Map BoolTest to local table encoding. The BoolTest (e)numerals 1721 // { eq = 0, ne = 4, le = 5, ge = 7, lt = 3, gt = 1 } 1722 // are mapped to table indices, while the remaining (e)numerals in BoolTest 1723 // { overflow = 2, no_overflow = 6, never = 8, illegal = 9 } 1724 // are ignored (these are not modeled in the table). 1725 // 1726 static int subsuming_bool_test_encode(Node* node) { 1727 precond(node->is_Bool()); 1728 BoolTest::mask x = node->as_Bool()->_test._test; 1729 switch (x) { 1730 case BoolTest::eq: return 0; 1731 case BoolTest::ne: return 1; 1732 case BoolTest::lt: return 2; 1733 case BoolTest::le: return 3; 1734 case BoolTest::gt: return 4; 1735 case BoolTest::ge: return 5; 1736 case BoolTest::overflow: 1737 case BoolTest::no_overflow: 1738 case BoolTest::never: 1739 case BoolTest::illegal: 1740 default: 1741 return -1; 1742 } 1743 } 1744 1745 //------------------------------Identity--------------------------------------- 1746 // If the test is constant & we match, then we are the input Control 1747 Node* IfProjNode::Identity(PhaseGVN* phase) { 1748 // Can only optimize if cannot go the other way 1749 const TypeTuple *t = phase->type(in(0))->is_tuple(); 1750 if (t == TypeTuple::IFNEITHER || (always_taken(t) && 1751 // During parsing (GVN) we don't remove dead code aggressively. 1752 // Cut off dead branch and let PhaseRemoveUseless take care of it. 1753 (!phase->is_IterGVN() || 1754 // During IGVN, first wait for the dead branch to be killed. 1755 // Otherwise, the IfNode's control will have two control uses (the IfNode 1756 // that doesn't go away because it still has uses and this branch of the 1757 // If) which breaks other optimizations. Node::has_special_unique_user() 1758 // will cause this node to be reprocessed once the dead branch is killed. 1759 in(0)->outcnt() == 1))) { 1760 // IfNode control 1761 if (in(0)->is_BaseCountedLoopEnd()) { 1762 // CountedLoopEndNode may be eliminated by if subsuming, replace CountedLoopNode with LoopNode to 1763 // avoid mismatching between CountedLoopNode and CountedLoopEndNode in the following optimization. 1764 Node* head = unique_ctrl_out_or_null(); 1765 if (head != nullptr && head->is_BaseCountedLoop() && head->in(LoopNode::LoopBackControl) == this) { 1766 Node* new_head = new LoopNode(head->in(LoopNode::EntryControl), this); 1767 phase->is_IterGVN()->register_new_node_with_optimizer(new_head); 1768 phase->is_IterGVN()->replace_node(head, new_head); 1769 } 1770 } 1771 return in(0)->in(0); 1772 } 1773 // no progress 1774 return this; 1775 } 1776 1777 #ifndef PRODUCT 1778 //------------------------------dump_spec-------------------------------------- 1779 void IfNode::dump_spec(outputStream *st) const { 1780 st->print("P=%f, C=%f",_prob,_fcnt); 1781 } 1782 #endif 1783 1784 //------------------------------idealize_test---------------------------------- 1785 // Try to canonicalize tests better. Peek at the Cmp/Bool/If sequence and 1786 // come up with a canonical sequence. Bools getting 'eq', 'gt' and 'ge' forms 1787 // converted to 'ne', 'le' and 'lt' forms. IfTrue/IfFalse get swapped as 1788 // needed. 1789 static IfNode* idealize_test(PhaseGVN* phase, IfNode* iff) { 1790 assert(iff->in(0) != nullptr, "If must be live"); 1791 1792 if (iff->outcnt() != 2) return nullptr; // Malformed projections. 1793 Node* old_if_f = iff->proj_out(false); 1794 Node* old_if_t = iff->proj_out(true); 1795 1796 // CountedLoopEnds want the back-control test to be TRUE, regardless of 1797 // whether they are testing a 'gt' or 'lt' condition. The 'gt' condition 1798 // happens in count-down loops 1799 if (iff->is_BaseCountedLoopEnd()) return nullptr; 1800 if (!iff->in(1)->is_Bool()) return nullptr; // Happens for partially optimized IF tests 1801 BoolNode *b = iff->in(1)->as_Bool(); 1802 BoolTest bt = b->_test; 1803 // Test already in good order? 1804 if( bt.is_canonical() ) 1805 return nullptr; 1806 1807 // Flip test to be canonical. Requires flipping the IfFalse/IfTrue and 1808 // cloning the IfNode. 1809 Node* new_b = phase->transform( new BoolNode(b->in(1), bt.negate()) ); 1810 if( !new_b->is_Bool() ) return nullptr; 1811 b = new_b->as_Bool(); 1812 1813 PhaseIterGVN *igvn = phase->is_IterGVN(); 1814 assert( igvn, "Test is not canonical in parser?" ); 1815 1816 // The IF node never really changes, but it needs to be cloned 1817 iff = iff->clone()->as_If(); 1818 iff->set_req(1, b); 1819 iff->_prob = 1.0-iff->_prob; 1820 1821 Node *prior = igvn->hash_find_insert(iff); 1822 if( prior ) { 1823 igvn->remove_dead_node(iff); 1824 iff = (IfNode*)prior; 1825 } else { 1826 // Cannot call transform on it just yet 1827 igvn->set_type_bottom(iff); 1828 } 1829 igvn->_worklist.push(iff); 1830 1831 // Now handle projections. Cloning not required. 1832 Node* new_if_f = (Node*)(new IfFalseNode( iff )); 1833 Node* new_if_t = (Node*)(new IfTrueNode ( iff )); 1834 1835 igvn->register_new_node_with_optimizer(new_if_f); 1836 igvn->register_new_node_with_optimizer(new_if_t); 1837 // Flip test, so flip trailing control 1838 igvn->replace_node(old_if_f, new_if_t); 1839 igvn->replace_node(old_if_t, new_if_f); 1840 1841 // Progress 1842 return iff; 1843 } 1844 1845 Node* RangeCheckNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1846 Node* res = Ideal_common(phase, can_reshape); 1847 if (res != NodeSentinel) { 1848 return res; 1849 } 1850 1851 PhaseIterGVN *igvn = phase->is_IterGVN(); 1852 // Setup to scan up the CFG looking for a dominating test 1853 Node* prev_dom = this; 1854 1855 // Check for range-check vs other kinds of tests 1856 Node* index1; 1857 Node* range1; 1858 jint offset1; 1859 int flip1 = is_range_check(range1, index1, offset1); 1860 if (flip1) { 1861 Node* dom = in(0); 1862 // Try to remove extra range checks. All 'up_one_dom' gives up at merges 1863 // so all checks we inspect post-dominate the top-most check we find. 1864 // If we are going to fail the current check and we reach the top check 1865 // then we are guaranteed to fail, so just start interpreting there. 1866 // We 'expand' the top 3 range checks to include all post-dominating 1867 // checks. 1868 1869 // The top 3 range checks seen 1870 const int NRC = 3; 1871 RangeCheck prev_checks[NRC]; 1872 int nb_checks = 0; 1873 1874 // Low and high offsets seen so far 1875 jint off_lo = offset1; 1876 jint off_hi = offset1; 1877 1878 bool found_immediate_dominator = false; 1879 1880 // Scan for the top checks and collect range of offsets 1881 for (int dist = 0; dist < 999; dist++) { // Range-Check scan limit 1882 if (dom->Opcode() == Op_RangeCheck && // Not same opcode? 1883 prev_dom->in(0) == dom) { // One path of test does dominate? 1884 if (dom == this) return nullptr; // dead loop 1885 // See if this is a range check 1886 Node* index2; 1887 Node* range2; 1888 jint offset2; 1889 int flip2 = dom->as_RangeCheck()->is_range_check(range2, index2, offset2); 1890 // See if this is a _matching_ range check, checking against 1891 // the same array bounds. 1892 if (flip2 == flip1 && range2 == range1 && index2 == index1 && 1893 dom->outcnt() == 2) { 1894 if (nb_checks == 0 && dom->in(1) == in(1)) { 1895 // Found an immediately dominating test at the same offset. 1896 // This kind of back-to-back test can be eliminated locally, 1897 // and there is no need to search further for dominating tests. 1898 assert(offset2 == offset1, "Same test but different offsets"); 1899 found_immediate_dominator = true; 1900 break; 1901 } 1902 // Gather expanded bounds 1903 off_lo = MIN2(off_lo,offset2); 1904 off_hi = MAX2(off_hi,offset2); 1905 // Record top NRC range checks 1906 prev_checks[nb_checks%NRC].ctl = prev_dom; 1907 prev_checks[nb_checks%NRC].off = offset2; 1908 nb_checks++; 1909 } 1910 } 1911 prev_dom = dom; 1912 dom = up_one_dom(dom); 1913 if (!dom) break; 1914 } 1915 1916 if (!found_immediate_dominator) { 1917 // Attempt to widen the dominating range check to cover some later 1918 // ones. Since range checks "fail" by uncommon-trapping to the 1919 // interpreter, widening a check can make us speculatively enter 1920 // the interpreter. If we see range-check deopt's, do not widen! 1921 if (!phase->C->allow_range_check_smearing()) return nullptr; 1922 1923 // Didn't find prior covering check, so cannot remove anything. 1924 if (nb_checks == 0) { 1925 return nullptr; 1926 } 1927 // Constant indices only need to check the upper bound. 1928 // Non-constant indices must check both low and high. 1929 int chk0 = (nb_checks - 1) % NRC; 1930 if (index1) { 1931 if (nb_checks == 1) { 1932 return nullptr; 1933 } else { 1934 // If the top range check's constant is the min or max of 1935 // all constants we widen the next one to cover the whole 1936 // range of constants. 1937 RangeCheck rc0 = prev_checks[chk0]; 1938 int chk1 = (nb_checks - 2) % NRC; 1939 RangeCheck rc1 = prev_checks[chk1]; 1940 if (rc0.off == off_lo) { 1941 adjust_check(rc1.ctl, range1, index1, flip1, off_hi, igvn); 1942 prev_dom = rc1.ctl; 1943 } else if (rc0.off == off_hi) { 1944 adjust_check(rc1.ctl, range1, index1, flip1, off_lo, igvn); 1945 prev_dom = rc1.ctl; 1946 } else { 1947 // If the top test's constant is not the min or max of all 1948 // constants, we need 3 range checks. We must leave the 1949 // top test unchanged because widening it would allow the 1950 // accesses it protects to successfully read/write out of 1951 // bounds. 1952 if (nb_checks == 2) { 1953 return nullptr; 1954 } 1955 int chk2 = (nb_checks - 3) % NRC; 1956 RangeCheck rc2 = prev_checks[chk2]; 1957 // The top range check a+i covers interval: -a <= i < length-a 1958 // The second range check b+i covers interval: -b <= i < length-b 1959 if (rc1.off <= rc0.off) { 1960 // if b <= a, we change the second range check to: 1961 // -min_of_all_constants <= i < length-min_of_all_constants 1962 // Together top and second range checks now cover: 1963 // -min_of_all_constants <= i < length-a 1964 // which is more restrictive than -b <= i < length-b: 1965 // -b <= -min_of_all_constants <= i < length-a <= length-b 1966 // The third check is then changed to: 1967 // -max_of_all_constants <= i < length-max_of_all_constants 1968 // so 2nd and 3rd checks restrict allowed values of i to: 1969 // -min_of_all_constants <= i < length-max_of_all_constants 1970 adjust_check(rc1.ctl, range1, index1, flip1, off_lo, igvn); 1971 adjust_check(rc2.ctl, range1, index1, flip1, off_hi, igvn); 1972 } else { 1973 // if b > a, we change the second range check to: 1974 // -max_of_all_constants <= i < length-max_of_all_constants 1975 // Together top and second range checks now cover: 1976 // -a <= i < length-max_of_all_constants 1977 // which is more restrictive than -b <= i < length-b: 1978 // -b < -a <= i < length-max_of_all_constants <= length-b 1979 // The third check is then changed to: 1980 // -max_of_all_constants <= i < length-max_of_all_constants 1981 // so 2nd and 3rd checks restrict allowed values of i to: 1982 // -min_of_all_constants <= i < length-max_of_all_constants 1983 adjust_check(rc1.ctl, range1, index1, flip1, off_hi, igvn); 1984 adjust_check(rc2.ctl, range1, index1, flip1, off_lo, igvn); 1985 } 1986 prev_dom = rc2.ctl; 1987 } 1988 } 1989 } else { 1990 RangeCheck rc0 = prev_checks[chk0]; 1991 // 'Widen' the offset of the 1st and only covering check 1992 adjust_check(rc0.ctl, range1, index1, flip1, off_hi, igvn); 1993 // Test is now covered by prior checks, dominate it out 1994 prev_dom = rc0.ctl; 1995 } 1996 } 1997 } else { 1998 prev_dom = search_identical(4, igvn); 1999 2000 if (prev_dom == nullptr) { 2001 return nullptr; 2002 } 2003 } 2004 2005 // Replace dominated IfNode 2006 return dominated_by(prev_dom, igvn); 2007 } 2008 2009 ParsePredicateNode::ParsePredicateNode(Node* control, Deoptimization::DeoptReason deopt_reason, PhaseGVN* gvn) 2010 : IfNode(control, gvn->intcon(1), PROB_MAX, COUNT_UNKNOWN), 2011 _deopt_reason(deopt_reason), 2012 _useless(false) { 2013 init_class_id(Class_ParsePredicate); 2014 gvn->C->add_parse_predicate(this); 2015 gvn->C->record_for_post_loop_opts_igvn(this); 2016 #ifdef ASSERT 2017 switch (deopt_reason) { 2018 case Deoptimization::Reason_predicate: 2019 case Deoptimization::Reason_profile_predicate: 2020 case Deoptimization::Reason_loop_limit_check: 2021 break; 2022 default: 2023 assert(false, "unsupported deoptimization reason for Parse Predicate"); 2024 } 2025 #endif // ASSERT 2026 } 2027 2028 Node* ParsePredicateNode::uncommon_trap() const { 2029 ParsePredicateUncommonProj* uncommon_proj = proj_out(0)->as_IfFalse(); 2030 Node* uct_region_or_call = uncommon_proj->unique_ctrl_out(); 2031 assert(uct_region_or_call->is_Region() || uct_region_or_call->is_Call(), "must be a region or call uct"); 2032 return uct_region_or_call; 2033 } 2034 2035 // Fold this node away once it becomes useless or at latest in post loop opts IGVN. 2036 const Type* ParsePredicateNode::Value(PhaseGVN* phase) const { 2037 if (phase->type(in(0)) == Type::TOP) { 2038 return Type::TOP; 2039 } 2040 if (_useless || phase->C->post_loop_opts_phase()) { 2041 return TypeTuple::IFTRUE; 2042 } else { 2043 return bottom_type(); 2044 } 2045 } 2046 2047 #ifndef PRODUCT 2048 void ParsePredicateNode::dump_spec(outputStream* st) const { 2049 st->print(" #"); 2050 switch (_deopt_reason) { 2051 case Deoptimization::DeoptReason::Reason_predicate: 2052 st->print("Loop "); 2053 break; 2054 case Deoptimization::DeoptReason::Reason_profile_predicate: 2055 st->print("Profiled_Loop "); 2056 break; 2057 case Deoptimization::DeoptReason::Reason_loop_limit_check: 2058 st->print("Loop_Limit_Check "); 2059 break; 2060 default: 2061 fatal("unknown kind"); 2062 } 2063 } 2064 2065 #endif // NOT PRODUCT