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/runtime.hpp" 36 #include "opto/rootnode.hpp" 37 #include "opto/subnode.hpp" 38 39 // Portions of code courtesy of Clifford Click 40 41 // Optimization - Graph Style 42 43 44 #ifndef PRODUCT 45 extern uint explicit_null_checks_elided; 46 #endif 47 48 //============================================================================= 49 //------------------------------Value------------------------------------------ 50 // Return a tuple for whichever arm of the IF is reachable 51 const Type* IfNode::Value(PhaseGVN* phase) const { 52 if( !in(0) ) return Type::TOP; 53 if( phase->type(in(0)) == Type::TOP ) 54 return Type::TOP; 55 const Type *t = phase->type(in(1)); 56 if( t == Type::TOP ) // data is undefined 57 return TypeTuple::IFNEITHER; // unreachable altogether 58 if( t == TypeInt::ZERO ) // zero, or false 59 return TypeTuple::IFFALSE; // only false branch is reachable 60 if( t == TypeInt::ONE ) // 1, or true 61 return TypeTuple::IFTRUE; // only true branch is reachable 62 assert( t == TypeInt::BOOL, "expected boolean type" ); 63 64 return TypeTuple::IFBOTH; // No progress 65 } 66 67 const RegMask &IfNode::out_RegMask() const { 68 return RegMask::Empty; 69 } 70 71 //------------------------------split_if--------------------------------------- 72 // Look for places where we merge constants, then test on the merged value. 73 // If the IF test will be constant folded on the path with the constant, we 74 // win by splitting the IF to before the merge point. 75 static Node* split_if(IfNode *iff, PhaseIterGVN *igvn) { 76 // I could be a lot more general here, but I'm trying to squeeze this 77 // in before the Christmas '98 break so I'm gonna be kinda restrictive 78 // on the patterns I accept. CNC 79 80 // Look for a compare of a constant and a merged value 81 Node *i1 = iff->in(1); 82 if( !i1->is_Bool() ) return nullptr; 83 BoolNode *b = i1->as_Bool(); 84 Node *cmp = b->in(1); 85 if( !cmp->is_Cmp() ) return nullptr; 86 i1 = cmp->in(1); 87 if( i1 == nullptr || !i1->is_Phi() ) return nullptr; 88 PhiNode *phi = i1->as_Phi(); 89 Node *con2 = cmp->in(2); 90 if( !con2->is_Con() ) return nullptr; 91 // See that the merge point contains some constants 92 Node *con1=nullptr; 93 uint i4; 94 RegionNode* phi_region = phi->region(); 95 for (i4 = 1; i4 < phi->req(); i4++ ) { 96 con1 = phi->in(i4); 97 // Do not optimize partially collapsed merges 98 if (con1 == nullptr || phi_region->in(i4) == nullptr || igvn->type(phi_region->in(i4)) == Type::TOP) { 99 igvn->_worklist.push(iff); 100 return nullptr; 101 } 102 if( con1->is_Con() ) break; // Found a constant 103 // Also allow null-vs-not-null checks 104 const TypePtr *tp = igvn->type(con1)->isa_ptr(); 105 if( tp && tp->_ptr == TypePtr::NotNull ) 106 break; 107 } 108 if( i4 >= phi->req() ) return nullptr; // Found no constants 109 110 igvn->C->set_has_split_ifs(true); // Has chance for split-if 111 112 // Make sure that the compare can be constant folded away 113 Node *cmp2 = cmp->clone(); 114 cmp2->set_req(1,con1); 115 cmp2->set_req(2,con2); 116 const Type *t = cmp2->Value(igvn); 117 // This compare is dead, so whack it! 118 igvn->remove_dead_node(cmp2); 119 if( !t->singleton() ) return nullptr; 120 121 // No intervening control, like a simple Call 122 Node* r = iff->in(0); 123 if (!r->is_Region() || r->is_Loop() || phi_region != r || r->as_Region()->is_copy()) { 124 return nullptr; 125 } 126 127 // No other users of the cmp/bool 128 if (b->outcnt() != 1 || cmp->outcnt() != 1) { 129 //tty->print_cr("many users of cmp/bool"); 130 return nullptr; 131 } 132 133 // Make sure we can determine where all the uses of merged values go 134 for (DUIterator_Fast jmax, j = r->fast_outs(jmax); j < jmax; j++) { 135 Node* u = r->fast_out(j); 136 if( u == r ) continue; 137 if( u == iff ) continue; 138 if( u->outcnt() == 0 ) continue; // use is dead & ignorable 139 if( !u->is_Phi() ) { 140 /* 141 if( u->is_Start() ) { 142 tty->print_cr("Region has inlined start use"); 143 } else { 144 tty->print_cr("Region has odd use"); 145 u->dump(2); 146 }*/ 147 return nullptr; 148 } 149 if( u != phi ) { 150 // CNC - do not allow any other merged value 151 //tty->print_cr("Merging another value"); 152 //u->dump(2); 153 return nullptr; 154 } 155 // Make sure we can account for all Phi uses 156 for (DUIterator_Fast kmax, k = u->fast_outs(kmax); k < kmax; k++) { 157 Node* v = u->fast_out(k); // User of the phi 158 // CNC - Allow only really simple patterns. 159 // In particular I disallow AddP of the Phi, a fairly common pattern 160 if (v == cmp) continue; // The compare is OK 161 if (v->is_ConstraintCast()) { 162 // If the cast is derived from data flow edges, it may not have a control edge. 163 // If so, it should be safe to split. But follow-up code can not deal with 164 // this (l. 359). So skip. 165 if (v->in(0) == nullptr) { 166 return nullptr; 167 } 168 if (v->in(0)->in(0) == iff) { 169 continue; // CastPP/II of the IfNode is OK 170 } 171 } 172 // Disabled following code because I cannot tell if exactly one 173 // path dominates without a real dominator check. CNC 9/9/1999 174 //uint vop = v->Opcode(); 175 //if( vop == Op_Phi ) { // Phi from another merge point might be OK 176 // Node *r = v->in(0); // Get controlling point 177 // if( !r ) return nullptr; // Degraded to a copy 178 // // Find exactly one path in (either True or False doms, but not IFF) 179 // int cnt = 0; 180 // for( uint i = 1; i < r->req(); i++ ) 181 // if( r->in(i) && r->in(i)->in(0) == iff ) 182 // cnt++; 183 // if( cnt == 1 ) continue; // Exactly one of True or False guards Phi 184 //} 185 if( !v->is_Call() ) { 186 /* 187 if( v->Opcode() == Op_AddP ) { 188 tty->print_cr("Phi has AddP use"); 189 } else if( v->Opcode() == Op_CastPP ) { 190 tty->print_cr("Phi has CastPP use"); 191 } else if( v->Opcode() == Op_CastII ) { 192 tty->print_cr("Phi has CastII use"); 193 } else { 194 tty->print_cr("Phi has use I can't be bothered with"); 195 } 196 */ 197 } 198 return nullptr; 199 200 /* CNC - Cut out all the fancy acceptance tests 201 // Can we clone this use when doing the transformation? 202 // If all uses are from Phis at this merge or constants, then YES. 203 if( !v->in(0) && v != cmp ) { 204 tty->print_cr("Phi has free-floating use"); 205 v->dump(2); 206 return nullptr; 207 } 208 for( uint l = 1; l < v->req(); l++ ) { 209 if( (!v->in(l)->is_Phi() || v->in(l)->in(0) != r) && 210 !v->in(l)->is_Con() ) { 211 tty->print_cr("Phi has use"); 212 v->dump(2); 213 return nullptr; 214 } // End of if Phi-use input is neither Phi nor Constant 215 } // End of for all inputs to Phi-use 216 */ 217 } // End of for all uses of Phi 218 } // End of for all uses of Region 219 220 // Only do this if the IF node is in a sane state 221 if (iff->outcnt() != 2) 222 return nullptr; 223 224 // Got a hit! Do the Mondo Hack! 225 // 226 //ABC a1c def ghi B 1 e h A C a c d f g i 227 // R - Phi - Phi - Phi Rc - Phi - Phi - Phi Rx - Phi - Phi - Phi 228 // cmp - 2 cmp - 2 cmp - 2 229 // bool bool_c bool_x 230 // if if_c if_x 231 // T F T F T F 232 // ..s.. ..t .. ..s.. ..t.. ..s.. ..t.. 233 // 234 // Split the paths coming into the merge point into 2 separate groups of 235 // merges. On the left will be all the paths feeding constants into the 236 // Cmp's Phi. On the right will be the remaining paths. The Cmp's Phi 237 // will fold up into a constant; this will let the Cmp fold up as well as 238 // all the control flow. Below the original IF we have 2 control 239 // dependent regions, 's' and 't'. Now we will merge the two paths 240 // just prior to 's' and 't' from the two IFs. At least 1 path (and quite 241 // likely 2 or more) will promptly constant fold away. 242 PhaseGVN *phase = igvn; 243 244 // Make a region merging constants and a region merging the rest 245 uint req_c = 0; 246 for (uint ii = 1; ii < r->req(); ii++) { 247 if (phi->in(ii) == con1) { 248 req_c++; 249 } 250 if (Node::may_be_loop_entry(r->in(ii))) { 251 // Bail out if splitting through a region with a Parse Predicate input (could 252 // also be a loop header before loop opts creates a LoopNode for it). 253 return nullptr; 254 } 255 } 256 257 // If all the defs of the phi are the same constant, we already have the desired end state. 258 // Skip the split that would create empty phi and region nodes. 259 if ((r->req() - req_c) == 1) { 260 return nullptr; 261 } 262 263 // At this point we know that we can apply the split if optimization. If the region is still on the worklist, 264 // we should wait until it is processed. The region might be removed which makes this optimization redundant. 265 // This also avoids the creation of dead data loops when rewiring data nodes below when a region is dying. 266 if (igvn->_worklist.member(r)) { 267 igvn->_worklist.push(iff); // retry split if later again 268 return nullptr; 269 } 270 271 Node *region_c = new RegionNode(req_c + 1); 272 Node *phi_c = con1; 273 uint len = r->req(); 274 Node *region_x = new RegionNode(len - req_c); 275 Node *phi_x = PhiNode::make_blank(region_x, phi); 276 for (uint i = 1, i_c = 1, i_x = 1; i < len; i++) { 277 if (phi->in(i) == con1) { 278 region_c->init_req( i_c++, r ->in(i) ); 279 } else { 280 region_x->init_req( i_x, r ->in(i) ); 281 phi_x ->init_req( i_x++, phi->in(i) ); 282 } 283 } 284 285 // Register the new RegionNodes but do not transform them. Cannot 286 // transform until the entire Region/Phi conglomerate has been hacked 287 // as a single huge transform. 288 igvn->register_new_node_with_optimizer( region_c ); 289 igvn->register_new_node_with_optimizer( region_x ); 290 // Prevent the untimely death of phi_x. Currently he has no uses. He is 291 // about to get one. If this only use goes away, then phi_x will look dead. 292 // However, he will be picking up some more uses down below. 293 Node *hook = new Node(4); 294 hook->init_req(0, phi_x); 295 hook->init_req(1, phi_c); 296 phi_x = phase->transform( phi_x ); 297 298 // Make the compare 299 Node *cmp_c = phase->makecon(t); 300 Node *cmp_x = cmp->clone(); 301 cmp_x->set_req(1,phi_x); 302 cmp_x->set_req(2,con2); 303 cmp_x = phase->transform(cmp_x); 304 // Make the bool 305 Node *b_c = phase->transform(new BoolNode(cmp_c,b->_test._test)); 306 Node *b_x = phase->transform(new BoolNode(cmp_x,b->_test._test)); 307 // Make the IfNode 308 IfNode* iff_c = iff->clone()->as_If(); 309 iff_c->set_req(0, region_c); 310 iff_c->set_req(1, b_c); 311 igvn->set_type_bottom(iff_c); 312 igvn->_worklist.push(iff_c); 313 hook->init_req(2, iff_c); 314 315 IfNode* iff_x = iff->clone()->as_If(); 316 iff_x->set_req(0, region_x); 317 iff_x->set_req(1, b_x); 318 igvn->set_type_bottom(iff_x); 319 igvn->_worklist.push(iff_x); 320 hook->init_req(3, iff_x); 321 322 // Make the true/false arms 323 Node *iff_c_t = phase->transform(new IfTrueNode (iff_c)); 324 Node *iff_c_f = phase->transform(new IfFalseNode(iff_c)); 325 Node *iff_x_t = phase->transform(new IfTrueNode (iff_x)); 326 Node *iff_x_f = phase->transform(new IfFalseNode(iff_x)); 327 328 // Merge the TRUE paths 329 Node *region_s = new RegionNode(3); 330 igvn->_worklist.push(region_s); 331 region_s->init_req(1, iff_c_t); 332 region_s->init_req(2, iff_x_t); 333 igvn->register_new_node_with_optimizer( region_s ); 334 335 // Merge the FALSE paths 336 Node *region_f = new RegionNode(3); 337 igvn->_worklist.push(region_f); 338 region_f->init_req(1, iff_c_f); 339 region_f->init_req(2, iff_x_f); 340 igvn->register_new_node_with_optimizer( region_f ); 341 342 igvn->hash_delete(cmp);// Remove soon-to-be-dead node from hash table. 343 cmp->set_req(1,nullptr); // Whack the inputs to cmp because it will be dead 344 cmp->set_req(2,nullptr); 345 // Check for all uses of the Phi and give them a new home. 346 // The 'cmp' got cloned, but CastPP/IIs need to be moved. 347 Node *phi_s = nullptr; // do not construct unless needed 348 Node *phi_f = nullptr; // do not construct unless needed 349 for (DUIterator_Last i2min, i2 = phi->last_outs(i2min); i2 >= i2min; --i2) { 350 Node* v = phi->last_out(i2);// User of the phi 351 igvn->rehash_node_delayed(v); // Have to fixup other Phi users 352 uint vop = v->Opcode(); 353 Node *proj = nullptr; 354 if( vop == Op_Phi ) { // Remote merge point 355 Node *r = v->in(0); 356 for (uint i3 = 1; i3 < r->req(); i3++) 357 if (r->in(i3) && r->in(i3)->in(0) == iff) { 358 proj = r->in(i3); 359 break; 360 } 361 } else if( v->is_ConstraintCast() ) { 362 proj = v->in(0); // Controlling projection 363 } else { 364 assert( 0, "do not know how to handle this guy" ); 365 } 366 guarantee(proj != nullptr, "sanity"); 367 368 Node *proj_path_data, *proj_path_ctrl; 369 if( proj->Opcode() == Op_IfTrue ) { 370 if( phi_s == nullptr ) { 371 // Only construct phi_s if needed, otherwise provides 372 // interfering use. 373 phi_s = PhiNode::make_blank(region_s,phi); 374 phi_s->init_req( 1, phi_c ); 375 phi_s->init_req( 2, phi_x ); 376 hook->add_req(phi_s); 377 phi_s = phase->transform(phi_s); 378 } 379 proj_path_data = phi_s; 380 proj_path_ctrl = region_s; 381 } else { 382 if( phi_f == nullptr ) { 383 // Only construct phi_f if needed, otherwise provides 384 // interfering use. 385 phi_f = PhiNode::make_blank(region_f,phi); 386 phi_f->init_req( 1, phi_c ); 387 phi_f->init_req( 2, phi_x ); 388 hook->add_req(phi_f); 389 phi_f = phase->transform(phi_f); 390 } 391 proj_path_data = phi_f; 392 proj_path_ctrl = region_f; 393 } 394 395 // Fixup 'v' for for the split 396 if( vop == Op_Phi ) { // Remote merge point 397 uint i; 398 for( i = 1; i < v->req(); i++ ) 399 if( v->in(i) == phi ) 400 break; 401 v->set_req(i, proj_path_data ); 402 } else if( v->is_ConstraintCast() ) { 403 v->set_req(0, proj_path_ctrl ); 404 v->set_req(1, proj_path_data ); 405 } else 406 ShouldNotReachHere(); 407 } 408 409 // Now replace the original iff's True/False with region_s/region_t. 410 // This makes the original iff go dead. 411 for (DUIterator_Last i3min, i3 = iff->last_outs(i3min); i3 >= i3min; --i3) { 412 Node* p = iff->last_out(i3); 413 assert( p->Opcode() == Op_IfTrue || p->Opcode() == Op_IfFalse, "" ); 414 Node *u = (p->Opcode() == Op_IfTrue) ? region_s : region_f; 415 // Replace p with u 416 igvn->add_users_to_worklist(p); 417 for (DUIterator_Last lmin, l = p->last_outs(lmin); l >= lmin;) { 418 Node* x = p->last_out(l); 419 igvn->hash_delete(x); 420 uint uses_found = 0; 421 for( uint j = 0; j < x->req(); j++ ) { 422 if( x->in(j) == p ) { 423 x->set_req(j, u); 424 uses_found++; 425 } 426 } 427 l -= uses_found; // we deleted 1 or more copies of this edge 428 } 429 igvn->remove_dead_node(p); 430 } 431 432 // Force the original merge dead 433 igvn->hash_delete(r); 434 // First, remove region's dead users. 435 for (DUIterator_Last lmin, l = r->last_outs(lmin); l >= lmin;) { 436 Node* u = r->last_out(l); 437 if( u == r ) { 438 r->set_req(0, nullptr); 439 } else { 440 assert(u->outcnt() == 0, "only dead users"); 441 igvn->remove_dead_node(u); 442 } 443 l -= 1; 444 } 445 igvn->remove_dead_node(r); 446 447 // Now remove the bogus extra edges used to keep things alive 448 igvn->remove_dead_node( hook ); 449 450 // Must return either the original node (now dead) or a new node 451 // (Do not return a top here, since that would break the uniqueness of top.) 452 return new ConINode(TypeInt::ZERO); 453 } 454 455 // if this IfNode follows a range check pattern return the projection 456 // for the failed path 457 ProjNode* IfNode::range_check_trap_proj(int& flip_test, Node*& l, Node*& r) { 458 if (outcnt() != 2) { 459 return nullptr; 460 } 461 Node* b = in(1); 462 if (b == nullptr || !b->is_Bool()) return nullptr; 463 BoolNode* bn = b->as_Bool(); 464 Node* cmp = bn->in(1); 465 if (cmp == nullptr) return nullptr; 466 if (cmp->Opcode() != Op_CmpU) return nullptr; 467 468 l = cmp->in(1); 469 r = cmp->in(2); 470 flip_test = 1; 471 if (bn->_test._test == BoolTest::le) { 472 l = cmp->in(2); 473 r = cmp->in(1); 474 flip_test = 2; 475 } else if (bn->_test._test != BoolTest::lt) { 476 return nullptr; 477 } 478 if (l->is_top()) return nullptr; // Top input means dead test 479 if (r->Opcode() != Op_LoadRange && !is_RangeCheck()) return nullptr; 480 481 // We have recognized one of these forms: 482 // Flip 1: If (Bool[<] CmpU(l, LoadRange)) ... 483 // Flip 2: If (Bool[<=] CmpU(LoadRange, l)) ... 484 485 ProjNode* iftrap = proj_out_or_null(flip_test == 2 ? true : false); 486 return iftrap; 487 } 488 489 490 //------------------------------is_range_check--------------------------------- 491 // Return 0 if not a range check. Return 1 if a range check and set index and 492 // offset. Return 2 if we had to negate the test. Index is null if the check 493 // is versus a constant. 494 int RangeCheckNode::is_range_check(Node* &range, Node* &index, jint &offset) { 495 int flip_test = 0; 496 Node* l = nullptr; 497 Node* r = nullptr; 498 ProjNode* iftrap = range_check_trap_proj(flip_test, l, r); 499 500 if (iftrap == nullptr) { 501 return 0; 502 } 503 504 // Make sure it's a real range check by requiring an uncommon trap 505 // along the OOB path. Otherwise, it's possible that the user wrote 506 // something which optimized to look like a range check but behaves 507 // in some other way. 508 if (iftrap->is_uncommon_trap_proj(Deoptimization::Reason_range_check) == nullptr) { 509 return 0; 510 } 511 512 // Look for index+offset form 513 Node* ind = l; 514 jint off = 0; 515 if (l->is_top()) { 516 return 0; 517 } else if (l->Opcode() == Op_AddI) { 518 if ((off = l->in(1)->find_int_con(0)) != 0) { 519 ind = l->in(2)->uncast(); 520 } else if ((off = l->in(2)->find_int_con(0)) != 0) { 521 ind = l->in(1)->uncast(); 522 } 523 } else if ((off = l->find_int_con(-1)) >= 0) { 524 // constant offset with no variable index 525 ind = nullptr; 526 } else { 527 // variable index with no constant offset (or dead negative index) 528 off = 0; 529 } 530 531 // Return all the values: 532 index = ind; 533 offset = off; 534 range = r; 535 return flip_test; 536 } 537 538 //------------------------------adjust_check----------------------------------- 539 // Adjust (widen) a prior range check 540 static void adjust_check(Node* proj, Node* range, Node* index, 541 int flip, jint off_lo, PhaseIterGVN* igvn) { 542 PhaseGVN *gvn = igvn; 543 // Break apart the old check 544 Node *iff = proj->in(0); 545 Node *bol = iff->in(1); 546 if( bol->is_top() ) return; // In case a partially dead range check appears 547 // bail (or bomb[ASSERT/DEBUG]) if NOT projection-->IfNode-->BoolNode 548 DEBUG_ONLY( if( !bol->is_Bool() ) { proj->dump(3); fatal("Expect projection-->IfNode-->BoolNode"); } ) 549 if( !bol->is_Bool() ) return; 550 551 Node *cmp = bol->in(1); 552 // Compute a new check 553 Node *new_add = gvn->intcon(off_lo); 554 if( index ) { 555 new_add = off_lo ? gvn->transform(new AddINode( index, new_add )) : index; 556 } 557 Node *new_cmp = (flip == 1) 558 ? new CmpUNode( new_add, range ) 559 : new CmpUNode( range, new_add ); 560 new_cmp = gvn->transform(new_cmp); 561 // See if no need to adjust the existing check 562 if( new_cmp == cmp ) return; 563 // Else, adjust existing check 564 Node *new_bol = gvn->transform( new BoolNode( new_cmp, bol->as_Bool()->_test._test ) ); 565 igvn->rehash_node_delayed( iff ); 566 iff->set_req_X( 1, new_bol, igvn ); 567 } 568 569 //------------------------------up_one_dom------------------------------------- 570 // Walk up the dominator tree one step. Return null at root or true 571 // complex merges. Skips through small diamonds. 572 Node* IfNode::up_one_dom(Node *curr, bool linear_only) { 573 Node *dom = curr->in(0); 574 if( !dom ) // Found a Region degraded to a copy? 575 return curr->nonnull_req(); // Skip thru it 576 577 if( curr != dom ) // Normal walk up one step? 578 return dom; 579 580 // Use linear_only if we are still parsing, since we cannot 581 // trust the regions to be fully filled in. 582 if (linear_only) 583 return nullptr; 584 585 if( dom->is_Root() ) 586 return nullptr; 587 588 // Else hit a Region. Check for a loop header 589 if( dom->is_Loop() ) 590 return dom->in(1); // Skip up thru loops 591 592 // Check for small diamonds 593 Node *din1, *din2, *din3, *din4; 594 if( dom->req() == 3 && // 2-path merge point 595 (din1 = dom ->in(1)) && // Left path exists 596 (din2 = dom ->in(2)) && // Right path exists 597 (din3 = din1->in(0)) && // Left path up one 598 (din4 = din2->in(0)) ) { // Right path up one 599 if( din3->is_Call() && // Handle a slow-path call on either arm 600 (din3 = din3->in(0)) ) 601 din3 = din3->in(0); 602 if( din4->is_Call() && // Handle a slow-path call on either arm 603 (din4 = din4->in(0)) ) 604 din4 = din4->in(0); 605 if (din3 != nullptr && din3 == din4 && din3->is_If()) // Regions not degraded to a copy 606 return din3; // Skip around diamonds 607 } 608 609 // Give up the search at true merges 610 return nullptr; // Dead loop? Or hit root? 611 } 612 613 614 //------------------------------filtered_int_type-------------------------------- 615 // Return a possibly more restrictive type for val based on condition control flow for an if 616 const TypeInt* IfNode::filtered_int_type(PhaseGVN* gvn, Node* val, Node* if_proj) { 617 assert(if_proj && 618 (if_proj->Opcode() == Op_IfTrue || if_proj->Opcode() == Op_IfFalse), "expecting an if projection"); 619 if (if_proj->in(0) && if_proj->in(0)->is_If()) { 620 IfNode* iff = if_proj->in(0)->as_If(); 621 if (iff->in(1) && iff->in(1)->is_Bool()) { 622 BoolNode* bol = iff->in(1)->as_Bool(); 623 if (bol->in(1) && bol->in(1)->is_Cmp()) { 624 const CmpNode* cmp = bol->in(1)->as_Cmp(); 625 if (cmp->in(1) == val) { 626 const TypeInt* cmp2_t = gvn->type(cmp->in(2))->isa_int(); 627 if (cmp2_t != nullptr) { 628 jint lo = cmp2_t->_lo; 629 jint hi = cmp2_t->_hi; 630 BoolTest::mask msk = if_proj->Opcode() == Op_IfTrue ? bol->_test._test : bol->_test.negate(); 631 switch (msk) { 632 case BoolTest::ne: { 633 // If val is compared to its lower or upper bound, we can narrow the type 634 const TypeInt* val_t = gvn->type(val)->isa_int(); 635 if (val_t != nullptr && !val_t->singleton() && cmp2_t->is_con()) { 636 if (val_t->_lo == lo) { 637 return TypeInt::make(val_t->_lo + 1, val_t->_hi, val_t->_widen); 638 } else if (val_t->_hi == hi) { 639 return TypeInt::make(val_t->_lo, val_t->_hi - 1, val_t->_widen); 640 } 641 } 642 // Can't refine type 643 return nullptr; 644 } 645 case BoolTest::eq: 646 return cmp2_t; 647 case BoolTest::lt: 648 lo = TypeInt::INT->_lo; 649 if (hi != min_jint) { 650 hi = hi - 1; 651 } 652 break; 653 case BoolTest::le: 654 lo = TypeInt::INT->_lo; 655 break; 656 case BoolTest::gt: 657 if (lo != max_jint) { 658 lo = lo + 1; 659 } 660 hi = TypeInt::INT->_hi; 661 break; 662 case BoolTest::ge: 663 // lo unchanged 664 hi = TypeInt::INT->_hi; 665 break; 666 default: 667 break; 668 } 669 const TypeInt* rtn_t = TypeInt::make(lo, hi, cmp2_t->_widen); 670 return rtn_t; 671 } 672 } 673 } 674 } 675 } 676 return nullptr; 677 } 678 679 //------------------------------fold_compares---------------------------- 680 // See if a pair of CmpIs can be converted into a CmpU. In some cases 681 // the direction of this if is determined by the preceding if so it 682 // can be eliminate entirely. 683 // 684 // Given an if testing (CmpI n v) check for an immediately control 685 // dependent if that is testing (CmpI n v2) and has one projection 686 // leading to this if and the other projection leading to a region 687 // that merges one of this ifs control projections. 688 // 689 // If 690 // / | 691 // / | 692 // / | 693 // If | 694 // /\ | 695 // / \ | 696 // / \ | 697 // / Region 698 // 699 // Or given an if testing (CmpI n v) check for a dominating if that is 700 // testing (CmpI n v2), both having one projection leading to an 701 // uncommon trap. Allow Another independent guard in between to cover 702 // an explicit range check: 703 // if (index < 0 || index >= array.length) { 704 // which may need a null check to guard the LoadRange 705 // 706 // If 707 // / \ 708 // / \ 709 // / \ 710 // If unc 711 // /\ 712 // / \ 713 // / \ 714 // / unc 715 // 716 717 // Is the comparison for this If suitable for folding? 718 bool IfNode::cmpi_folds(PhaseIterGVN* igvn, bool fold_ne) { 719 return in(1) != nullptr && 720 in(1)->is_Bool() && 721 in(1)->in(1) != nullptr && 722 in(1)->in(1)->Opcode() == Op_CmpI && 723 in(1)->in(1)->in(2) != nullptr && 724 in(1)->in(1)->in(2) != igvn->C->top() && 725 (in(1)->as_Bool()->_test.is_less() || 726 in(1)->as_Bool()->_test.is_greater() || 727 (fold_ne && in(1)->as_Bool()->_test._test == BoolTest::ne)); 728 } 729 730 // Is a dominating control suitable for folding with this if? 731 bool IfNode::is_ctrl_folds(Node* ctrl, PhaseIterGVN* igvn) { 732 return ctrl != nullptr && 733 ctrl->is_Proj() && 734 ctrl->in(0) != nullptr && 735 ctrl->in(0)->Opcode() == Op_If && 736 ctrl->in(0)->outcnt() == 2 && 737 ctrl->in(0)->as_If()->cmpi_folds(igvn, true) && 738 // Must compare same value 739 ctrl->in(0)->in(1)->in(1)->in(1) != nullptr && 740 ctrl->in(0)->in(1)->in(1)->in(1) != igvn->C->top() && 741 ctrl->in(0)->in(1)->in(1)->in(1) == in(1)->in(1)->in(1); 742 } 743 744 // Do this If and the dominating If share a region? 745 bool IfNode::has_shared_region(ProjNode* proj, ProjNode*& success, ProjNode*& fail) { 746 ProjNode* otherproj = proj->other_if_proj(); 747 Node* otherproj_ctrl_use = otherproj->unique_ctrl_out_or_null(); 748 RegionNode* region = (otherproj_ctrl_use != nullptr && otherproj_ctrl_use->is_Region()) ? otherproj_ctrl_use->as_Region() : nullptr; 749 success = nullptr; 750 fail = nullptr; 751 752 if (otherproj->outcnt() == 1 && region != nullptr && !region->has_phi()) { 753 for (int i = 0; i < 2; i++) { 754 ProjNode* proj = proj_out(i); 755 if (success == nullptr && proj->outcnt() == 1 && proj->unique_out() == region) { 756 success = proj; 757 } else if (fail == nullptr) { 758 fail = proj; 759 } else { 760 success = fail = nullptr; 761 } 762 } 763 } 764 return success != nullptr && fail != nullptr; 765 } 766 767 bool IfNode::is_dominator_unc(CallStaticJavaNode* dom_unc, CallStaticJavaNode* unc) { 768 // Different methods and methods containing jsrs are not supported. 769 ciMethod* method = unc->jvms()->method(); 770 ciMethod* dom_method = dom_unc->jvms()->method(); 771 if (method != dom_method || method->has_jsrs()) { 772 return false; 773 } 774 // Check that both traps are in the same activation of the method (instead 775 // of two activations being inlined through different call sites) by verifying 776 // that the call stacks are equal for both JVMStates. 777 JVMState* dom_caller = dom_unc->jvms()->caller(); 778 JVMState* caller = unc->jvms()->caller(); 779 if ((dom_caller == nullptr) != (caller == nullptr)) { 780 // The current method must either be inlined into both dom_caller and 781 // caller or must not be inlined at all (top method). Bail out otherwise. 782 return false; 783 } else if (dom_caller != nullptr && !dom_caller->same_calls_as(caller)) { 784 return false; 785 } 786 // Check that the bci of the dominating uncommon trap dominates the bci 787 // of the dominated uncommon trap. Otherwise we may not re-execute 788 // the dominated check after deoptimization from the merged uncommon trap. 789 ciTypeFlow* flow = dom_method->get_flow_analysis(); 790 int bci = unc->jvms()->bci(); 791 int dom_bci = dom_unc->jvms()->bci(); 792 if (!flow->is_dominated_by(bci, dom_bci)) { 793 return false; 794 } 795 796 return true; 797 } 798 799 // Return projection that leads to an uncommon trap if any 800 ProjNode* IfNode::uncommon_trap_proj(CallStaticJavaNode*& call) const { 801 for (int i = 0; i < 2; i++) { 802 call = proj_out(i)->is_uncommon_trap_proj(Deoptimization::Reason_none); 803 if (call != nullptr) { 804 return proj_out(i); 805 } 806 } 807 return nullptr; 808 } 809 810 // Do this If and the dominating If both branch out to an uncommon trap 811 bool IfNode::has_only_uncommon_traps(ProjNode* proj, ProjNode*& success, ProjNode*& fail, PhaseIterGVN* igvn) { 812 ProjNode* otherproj = proj->other_if_proj(); 813 CallStaticJavaNode* dom_unc = otherproj->is_uncommon_trap_proj(Deoptimization::Reason_none); 814 815 if (otherproj->outcnt() == 1 && dom_unc != nullptr) { 816 // We need to re-execute the folded Ifs after deoptimization from the merged traps 817 if (!dom_unc->jvms()->should_reexecute()) { 818 return false; 819 } 820 821 CallStaticJavaNode* unc = nullptr; 822 ProjNode* unc_proj = uncommon_trap_proj(unc); 823 if (unc_proj != nullptr && unc_proj->outcnt() == 1) { 824 if (dom_unc == unc) { 825 // Allow the uncommon trap to be shared through a region 826 RegionNode* r = unc->in(0)->as_Region(); 827 if (r->outcnt() != 2 || r->req() != 3 || r->find_edge(otherproj) == -1 || r->find_edge(unc_proj) == -1) { 828 return false; 829 } 830 assert(r->has_phi() == nullptr, "simple region shouldn't have a phi"); 831 } else if (dom_unc->in(0) != otherproj || unc->in(0) != unc_proj) { 832 return false; 833 } 834 835 if (!is_dominator_unc(dom_unc, unc)) { 836 return false; 837 } 838 839 // See merge_uncommon_traps: the reason of the uncommon trap 840 // will be changed and the state of the dominating If will be 841 // used. Checked that we didn't apply this transformation in a 842 // previous compilation and it didn't cause too many traps 843 ciMethod* dom_method = dom_unc->jvms()->method(); 844 int dom_bci = dom_unc->jvms()->bci(); 845 if (!igvn->C->too_many_traps(dom_method, dom_bci, Deoptimization::Reason_unstable_fused_if) && 846 !igvn->C->too_many_traps(dom_method, dom_bci, Deoptimization::Reason_range_check) && 847 // Return true if c2 manages to reconcile with UnstableIf optimization. See the comments for it. 848 igvn->C->remove_unstable_if_trap(dom_unc, true/*yield*/)) { 849 success = unc_proj; 850 fail = unc_proj->other_if_proj(); 851 return true; 852 } 853 } 854 } 855 return false; 856 } 857 858 // Check that the 2 CmpI can be folded into as single CmpU and proceed with the folding 859 bool IfNode::fold_compares_helper(ProjNode* proj, ProjNode* success, ProjNode* fail, PhaseIterGVN* igvn) { 860 Node* this_cmp = in(1)->in(1); 861 BoolNode* this_bool = in(1)->as_Bool(); 862 IfNode* dom_iff = proj->in(0)->as_If(); 863 BoolNode* dom_bool = dom_iff->in(1)->as_Bool(); 864 Node* lo = dom_iff->in(1)->in(1)->in(2); 865 Node* hi = this_cmp->in(2); 866 Node* n = this_cmp->in(1); 867 ProjNode* otherproj = proj->other_if_proj(); 868 869 const TypeInt* lo_type = IfNode::filtered_int_type(igvn, n, otherproj); 870 const TypeInt* hi_type = IfNode::filtered_int_type(igvn, n, success); 871 872 BoolTest::mask lo_test = dom_bool->_test._test; 873 BoolTest::mask hi_test = this_bool->_test._test; 874 BoolTest::mask cond = hi_test; 875 876 // convert: 877 // 878 // dom_bool = x {<,<=,>,>=} a 879 // / \ 880 // proj = {True,False} / \ otherproj = {False,True} 881 // / 882 // this_bool = x {<,<=} b 883 // / \ 884 // fail = {True,False} / \ success = {False,True} 885 // / 886 // 887 // (Second test guaranteed canonicalized, first one may not have 888 // been canonicalized yet) 889 // 890 // into: 891 // 892 // cond = (x - lo) {<u,<=u,>u,>=u} adjusted_lim 893 // / \ 894 // fail / \ success 895 // / 896 // 897 898 // Figure out which of the two tests sets the upper bound and which 899 // sets the lower bound if any. 900 Node* adjusted_lim = nullptr; 901 if (lo_type != nullptr && hi_type != nullptr && hi_type->_lo > lo_type->_hi && 902 hi_type->_hi == max_jint && lo_type->_lo == min_jint && lo_test != BoolTest::ne) { 903 assert((dom_bool->_test.is_less() && !proj->_con) || 904 (dom_bool->_test.is_greater() && proj->_con), "incorrect test"); 905 906 // this_bool = < 907 // dom_bool = >= (proj = True) or dom_bool = < (proj = False) 908 // x in [a, b[ on the fail (= True) projection, b > a-1 (because of hi_type->_lo > lo_type->_hi test above): 909 // lo = a, hi = b, adjusted_lim = b-a, cond = <u 910 // dom_bool = > (proj = True) or dom_bool = <= (proj = False) 911 // x in ]a, b[ on the fail (= True) projection, b > a: 912 // lo = a+1, hi = b, adjusted_lim = b-a-1, cond = <u 913 // this_bool = <= 914 // dom_bool = >= (proj = True) or dom_bool = < (proj = False) 915 // x in [a, b] on the fail (= True) projection, b+1 > a-1: 916 // lo = a, hi = b, adjusted_lim = b-a+1, cond = <u 917 // lo = a, hi = b, adjusted_lim = b-a, cond = <=u doesn't work because b = a - 1 is possible, then b-a = -1 918 // dom_bool = > (proj = True) or dom_bool = <= (proj = False) 919 // x in ]a, b] on the fail (= True) projection b+1 > a: 920 // lo = a+1, hi = b, adjusted_lim = b-a, cond = <u 921 // lo = a+1, hi = b, adjusted_lim = b-a-1, cond = <=u doesn't work because a = b is possible, then b-a-1 = -1 922 923 if (hi_test == BoolTest::lt) { 924 if (lo_test == BoolTest::gt || lo_test == BoolTest::le) { 925 lo = igvn->transform(new AddINode(lo, igvn->intcon(1))); 926 } 927 } else if (hi_test == BoolTest::le) { 928 if (lo_test == BoolTest::ge || lo_test == BoolTest::lt) { 929 adjusted_lim = igvn->transform(new SubINode(hi, lo)); 930 adjusted_lim = igvn->transform(new AddINode(adjusted_lim, igvn->intcon(1))); 931 cond = BoolTest::lt; 932 } else if (lo_test == BoolTest::gt || lo_test == BoolTest::le) { 933 adjusted_lim = igvn->transform(new SubINode(hi, lo)); 934 lo = igvn->transform(new AddINode(lo, igvn->intcon(1))); 935 cond = BoolTest::lt; 936 } else { 937 assert(false, "unhandled lo_test: %d", lo_test); 938 return false; 939 } 940 } else { 941 assert(igvn->_worklist.member(in(1)) && in(1)->Value(igvn) != igvn->type(in(1)), "unhandled hi_test: %d", hi_test); 942 return false; 943 } 944 // this test was canonicalized 945 assert(this_bool->_test.is_less() && fail->_con, "incorrect test"); 946 } else if (lo_type != nullptr && hi_type != nullptr && lo_type->_lo > hi_type->_hi && 947 lo_type->_hi == max_jint && hi_type->_lo == min_jint && lo_test != BoolTest::ne) { 948 949 // this_bool = < 950 // dom_bool = < (proj = True) or dom_bool = >= (proj = False) 951 // x in [b, a[ on the fail (= False) projection, a > b-1 (because of lo_type->_lo > hi_type->_hi above): 952 // lo = b, hi = a, adjusted_lim = a-b, cond = >=u 953 // dom_bool = <= (proj = True) or dom_bool = > (proj = False) 954 // x in [b, a] on the fail (= False) projection, a+1 > b-1: 955 // lo = b, hi = a, adjusted_lim = a-b+1, cond = >=u 956 // lo = b, hi = a, adjusted_lim = a-b, cond = >u doesn't work because a = b - 1 is possible, then b-a = -1 957 // this_bool = <= 958 // dom_bool = < (proj = True) or dom_bool = >= (proj = False) 959 // x in ]b, a[ on the fail (= False) projection, a > b: 960 // lo = b+1, hi = a, adjusted_lim = a-b-1, 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: 963 // lo = b+1, hi = a, adjusted_lim = a-b, cond = >=u 964 // lo = b+1, hi = a, adjusted_lim = a-b-1, cond = >u doesn't work because a = b is possible, then b-a-1 = -1 965 966 swap(lo, hi); 967 swap(lo_type, hi_type); 968 swap(lo_test, hi_test); 969 970 assert((dom_bool->_test.is_less() && proj->_con) || 971 (dom_bool->_test.is_greater() && !proj->_con), "incorrect test"); 972 973 cond = (hi_test == BoolTest::le || hi_test == BoolTest::gt) ? BoolTest::gt : BoolTest::ge; 974 975 if (lo_test == BoolTest::lt) { 976 if (hi_test == BoolTest::lt || hi_test == BoolTest::ge) { 977 cond = BoolTest::ge; 978 } else if (hi_test == BoolTest::le || hi_test == BoolTest::gt) { 979 adjusted_lim = igvn->transform(new SubINode(hi, lo)); 980 adjusted_lim = igvn->transform(new AddINode(adjusted_lim, igvn->intcon(1))); 981 cond = BoolTest::ge; 982 } else { 983 assert(false, "unhandled hi_test: %d", hi_test); 984 return false; 985 } 986 } else if (lo_test == BoolTest::le) { 987 if (hi_test == BoolTest::lt || hi_test == BoolTest::ge) { 988 lo = igvn->transform(new AddINode(lo, igvn->intcon(1))); 989 cond = BoolTest::ge; 990 } else if (hi_test == BoolTest::le || hi_test == BoolTest::gt) { 991 adjusted_lim = igvn->transform(new SubINode(hi, lo)); 992 lo = igvn->transform(new AddINode(lo, igvn->intcon(1))); 993 cond = BoolTest::ge; 994 } else { 995 assert(false, "unhandled hi_test: %d", hi_test); 996 return false; 997 } 998 } else { 999 assert(igvn->_worklist.member(in(1)) && in(1)->Value(igvn) != igvn->type(in(1)), "unhandled lo_test: %d", lo_test); 1000 return false; 1001 } 1002 // this test was canonicalized 1003 assert(this_bool->_test.is_less() && !fail->_con, "incorrect test"); 1004 } else { 1005 const TypeInt* failtype = filtered_int_type(igvn, n, proj); 1006 if (failtype != nullptr) { 1007 const TypeInt* type2 = filtered_int_type(igvn, n, fail); 1008 if (type2 != nullptr) { 1009 failtype = failtype->join(type2)->is_int(); 1010 if (failtype->_lo > failtype->_hi) { 1011 // previous if determines the result of this if so 1012 // replace Bool with constant 1013 igvn->replace_input_of(this, 1, igvn->intcon(success->_con)); 1014 return true; 1015 } 1016 } 1017 } 1018 lo = nullptr; 1019 hi = nullptr; 1020 } 1021 1022 if (lo && hi) { 1023 Node* hook = new Node(1); 1024 hook->init_req(0, lo); // Add a use to lo to prevent him from dying 1025 // Merge the two compares into a single unsigned compare by building (CmpU (n - lo) (hi - lo)) 1026 Node* adjusted_val = igvn->transform(new SubINode(n, lo)); 1027 if (adjusted_lim == nullptr) { 1028 adjusted_lim = igvn->transform(new SubINode(hi, lo)); 1029 } 1030 hook->destruct(igvn); 1031 1032 int lo = igvn->type(adjusted_lim)->is_int()->_lo; 1033 if (lo < 0) { 1034 // If range check elimination applies to this comparison, it includes code to protect from overflows that may 1035 // cause the main loop to be skipped entirely. Delay this transformation. 1036 // Example: 1037 // for (int i = 0; i < limit; i++) { 1038 // if (i < max_jint && i > min_jint) {... 1039 // } 1040 // Comparisons folded as: 1041 // i - min_jint - 1 <u -2 1042 // when RC applies, main loop limit becomes: 1043 // min(limit, max(-2 + min_jint + 1, min_jint)) 1044 // = min(limit, min_jint) 1045 // = min_jint 1046 if (!igvn->C->post_loop_opts_phase()) { 1047 if (adjusted_val->outcnt() == 0) { 1048 igvn->remove_dead_node(adjusted_val); 1049 } 1050 if (adjusted_lim->outcnt() == 0) { 1051 igvn->remove_dead_node(adjusted_lim); 1052 } 1053 igvn->C->record_for_post_loop_opts_igvn(this); 1054 return false; 1055 } 1056 } 1057 1058 Node* newcmp = igvn->transform(new CmpUNode(adjusted_val, adjusted_lim)); 1059 Node* newbool = igvn->transform(new BoolNode(newcmp, cond)); 1060 1061 igvn->replace_input_of(dom_iff, 1, igvn->intcon(proj->_con)); 1062 igvn->replace_input_of(this, 1, newbool); 1063 1064 return true; 1065 } 1066 return false; 1067 } 1068 1069 // Merge the branches that trap for this If and the dominating If into 1070 // a single region that branches to the uncommon trap for the 1071 // dominating If 1072 Node* IfNode::merge_uncommon_traps(ProjNode* proj, ProjNode* success, ProjNode* fail, PhaseIterGVN* igvn) { 1073 Node* res = this; 1074 assert(success->in(0) == this, "bad projection"); 1075 1076 ProjNode* otherproj = proj->other_if_proj(); 1077 1078 CallStaticJavaNode* unc = success->is_uncommon_trap_proj(Deoptimization::Reason_none); 1079 CallStaticJavaNode* dom_unc = otherproj->is_uncommon_trap_proj(Deoptimization::Reason_none); 1080 1081 if (unc != dom_unc) { 1082 Node* r = new RegionNode(3); 1083 1084 r->set_req(1, otherproj); 1085 r->set_req(2, success); 1086 r = igvn->transform(r); 1087 assert(r->is_Region(), "can't go away"); 1088 1089 // Make both If trap at the state of the first If: once the CmpI 1090 // nodes are merged, if we trap we don't know which of the CmpI 1091 // nodes would have caused the trap so we have to restart 1092 // execution at the first one 1093 igvn->replace_input_of(dom_unc, 0, r); 1094 igvn->replace_input_of(unc, 0, igvn->C->top()); 1095 } 1096 int trap_request = dom_unc->uncommon_trap_request(); 1097 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 1098 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 1099 1100 int flip_test = 0; 1101 Node* l = nullptr; 1102 Node* r = nullptr; 1103 1104 if (success->in(0)->as_If()->range_check_trap_proj(flip_test, l, r) != nullptr) { 1105 // If this looks like a range check, change the trap to 1106 // Reason_range_check so the compiler recognizes it as a range 1107 // check and applies the corresponding optimizations 1108 trap_request = Deoptimization::make_trap_request(Deoptimization::Reason_range_check, action); 1109 1110 improve_address_types(l, r, fail, igvn); 1111 1112 res = igvn->transform(new RangeCheckNode(in(0), in(1), _prob, _fcnt)); 1113 } else if (unc != dom_unc) { 1114 // If we trap we won't know what CmpI would have caused the trap 1115 // so use a special trap reason to mark this pair of CmpI nodes as 1116 // bad candidate for folding. On recompilation we won't fold them 1117 // and we may trap again but this time we'll know what branch 1118 // traps 1119 trap_request = Deoptimization::make_trap_request(Deoptimization::Reason_unstable_fused_if, action); 1120 } 1121 igvn->replace_input_of(dom_unc, TypeFunc::Parms, igvn->intcon(trap_request)); 1122 return res; 1123 } 1124 1125 // If we are turning 2 CmpI nodes into a CmpU that follows the pattern 1126 // of a rangecheck on index i, on 64 bit the compares may be followed 1127 // by memory accesses using i as index. In that case, the CmpU tells 1128 // us something about the values taken by i that can help the compiler 1129 // (see Compile::conv_I2X_index()) 1130 void IfNode::improve_address_types(Node* l, Node* r, ProjNode* fail, PhaseIterGVN* igvn) { 1131 #ifdef _LP64 1132 ResourceMark rm; 1133 Node_Stack stack(2); 1134 1135 assert(r->Opcode() == Op_LoadRange, "unexpected range check"); 1136 const TypeInt* array_size = igvn->type(r)->is_int(); 1137 1138 stack.push(l, 0); 1139 1140 while(stack.size() > 0) { 1141 Node* n = stack.node(); 1142 uint start = stack.index(); 1143 1144 uint i = start; 1145 for (; i < n->outcnt(); i++) { 1146 Node* use = n->raw_out(i); 1147 if (stack.size() == 1) { 1148 if (use->Opcode() == Op_ConvI2L) { 1149 const TypeLong* bounds = use->as_Type()->type()->is_long(); 1150 if (bounds->_lo <= array_size->_lo && bounds->_hi >= array_size->_hi && 1151 (bounds->_lo != array_size->_lo || bounds->_hi != array_size->_hi)) { 1152 stack.set_index(i+1); 1153 stack.push(use, 0); 1154 break; 1155 } 1156 } 1157 } else if (use->is_Mem()) { 1158 Node* ctrl = use->in(0); 1159 for (int i = 0; i < 10 && ctrl != nullptr && ctrl != fail; i++) { 1160 ctrl = up_one_dom(ctrl); 1161 } 1162 if (ctrl == fail) { 1163 Node* init_n = stack.node_at(1); 1164 assert(init_n->Opcode() == Op_ConvI2L, "unexpected first node"); 1165 // Create a new narrow ConvI2L node that is dependent on the range check 1166 Node* new_n = igvn->C->conv_I2X_index(igvn, l, array_size, fail); 1167 1168 // The type of the ConvI2L may be widen and so the new 1169 // ConvI2L may not be better than an existing ConvI2L 1170 if (new_n != init_n) { 1171 for (uint j = 2; j < stack.size(); j++) { 1172 Node* n = stack.node_at(j); 1173 Node* clone = n->clone(); 1174 int rep = clone->replace_edge(init_n, new_n, igvn); 1175 assert(rep > 0, "can't find expected node?"); 1176 clone = igvn->transform(clone); 1177 init_n = n; 1178 new_n = clone; 1179 } 1180 igvn->hash_delete(use); 1181 int rep = use->replace_edge(init_n, new_n, igvn); 1182 assert(rep > 0, "can't find expected node?"); 1183 igvn->transform(use); 1184 if (init_n->outcnt() == 0) { 1185 igvn->_worklist.push(init_n); 1186 } 1187 } 1188 } 1189 } else if (use->in(0) == nullptr && (igvn->type(use)->isa_long() || 1190 igvn->type(use)->isa_ptr())) { 1191 stack.set_index(i+1); 1192 stack.push(use, 0); 1193 break; 1194 } 1195 } 1196 if (i == n->outcnt()) { 1197 stack.pop(); 1198 } 1199 } 1200 #endif 1201 } 1202 1203 bool IfNode::is_cmp_with_loadrange(ProjNode* proj) { 1204 if (in(1) != nullptr && 1205 in(1)->in(1) != nullptr && 1206 in(1)->in(1)->in(2) != nullptr) { 1207 Node* other = in(1)->in(1)->in(2); 1208 if (other->Opcode() == Op_LoadRange && 1209 ((other->in(0) != nullptr && other->in(0) == proj) || 1210 (other->in(0) == nullptr && 1211 other->in(2) != nullptr && 1212 other->in(2)->is_AddP() && 1213 other->in(2)->in(1) != nullptr && 1214 other->in(2)->in(1)->Opcode() == Op_CastPP && 1215 other->in(2)->in(1)->in(0) == proj))) { 1216 return true; 1217 } 1218 } 1219 return false; 1220 } 1221 1222 bool IfNode::is_null_check(ProjNode* proj, PhaseIterGVN* igvn) { 1223 Node* other = in(1)->in(1)->in(2); 1224 if (other->in(MemNode::Address) != nullptr && 1225 proj->in(0)->in(1) != nullptr && 1226 proj->in(0)->in(1)->is_Bool() && 1227 proj->in(0)->in(1)->in(1) != nullptr && 1228 proj->in(0)->in(1)->in(1)->Opcode() == Op_CmpP && 1229 proj->in(0)->in(1)->in(1)->in(2) != nullptr && 1230 proj->in(0)->in(1)->in(1)->in(1) == other->in(MemNode::Address)->in(AddPNode::Address)->uncast() && 1231 igvn->type(proj->in(0)->in(1)->in(1)->in(2)) == TypePtr::NULL_PTR) { 1232 return true; 1233 } 1234 return false; 1235 } 1236 1237 // Returns true if this IfNode belongs to a flat array check 1238 // and returns the corresponding array in the 'array' parameter. 1239 bool IfNode::is_flat_array_check(PhaseTransform* phase, Node** array) { 1240 Node* bol = in(1); 1241 if (!bol->is_Bool()) { 1242 return false; 1243 } 1244 Node* cmp = bol->in(1); 1245 if (cmp->isa_FlatArrayCheck()) { 1246 if (array != nullptr) { 1247 *array = cmp->in(FlatArrayCheckNode::ArrayOrKlass); 1248 } 1249 return true; 1250 } 1251 return false; 1252 } 1253 1254 // Check that the If that is in between the 2 integer comparisons has 1255 // no side effect 1256 bool IfNode::is_side_effect_free_test(ProjNode* proj, PhaseIterGVN* igvn) { 1257 if (proj == nullptr) { 1258 return false; 1259 } 1260 CallStaticJavaNode* unc = proj->is_uncommon_trap_if_pattern(Deoptimization::Reason_none); 1261 if (unc != nullptr && proj->outcnt() <= 2) { 1262 if (proj->outcnt() == 1 || 1263 // Allow simple null check from LoadRange 1264 (is_cmp_with_loadrange(proj) && is_null_check(proj, igvn))) { 1265 CallStaticJavaNode* unc = proj->is_uncommon_trap_if_pattern(Deoptimization::Reason_none); 1266 CallStaticJavaNode* dom_unc = proj->in(0)->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none); 1267 assert(dom_unc != nullptr, "is_uncommon_trap_if_pattern returned null"); 1268 1269 // reroute_side_effect_free_unc changes the state of this 1270 // uncommon trap to restart execution at the previous 1271 // CmpI. Check that this change in a previous compilation didn't 1272 // cause too many traps. 1273 int trap_request = unc->uncommon_trap_request(); 1274 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 1275 1276 if (igvn->C->too_many_traps(dom_unc->jvms()->method(), dom_unc->jvms()->bci(), reason)) { 1277 return false; 1278 } 1279 1280 if (!is_dominator_unc(dom_unc, unc)) { 1281 return false; 1282 } 1283 1284 return true; 1285 } 1286 } 1287 return false; 1288 } 1289 1290 // Make the If between the 2 integer comparisons trap at the state of 1291 // the first If: the last CmpI is the one replaced by a CmpU and the 1292 // first CmpI is eliminated, so the test between the 2 CmpI nodes 1293 // won't be guarded by the first CmpI anymore. It can trap in cases 1294 // where the first CmpI would have prevented it from executing: on a 1295 // trap, we need to restart execution at the state of the first CmpI 1296 void IfNode::reroute_side_effect_free_unc(ProjNode* proj, ProjNode* dom_proj, PhaseIterGVN* igvn) { 1297 CallStaticJavaNode* dom_unc = dom_proj->is_uncommon_trap_if_pattern(Deoptimization::Reason_none); 1298 ProjNode* otherproj = proj->other_if_proj(); 1299 CallStaticJavaNode* unc = proj->is_uncommon_trap_if_pattern(Deoptimization::Reason_none); 1300 Node* call_proj = dom_unc->unique_ctrl_out(); 1301 Node* halt = call_proj->unique_ctrl_out(); 1302 1303 Node* new_unc = dom_unc->clone(); 1304 call_proj = call_proj->clone(); 1305 halt = halt->clone(); 1306 Node* c = otherproj->clone(); 1307 1308 c = igvn->transform(c); 1309 new_unc->set_req(TypeFunc::Parms, unc->in(TypeFunc::Parms)); 1310 new_unc->set_req(0, c); 1311 new_unc = igvn->transform(new_unc); 1312 call_proj->set_req(0, new_unc); 1313 call_proj = igvn->transform(call_proj); 1314 halt->set_req(0, call_proj); 1315 halt = igvn->transform(halt); 1316 1317 igvn->replace_node(otherproj, igvn->C->top()); 1318 igvn->C->root()->add_req(halt); 1319 } 1320 1321 Node* IfNode::fold_compares(PhaseIterGVN* igvn) { 1322 if (Opcode() != Op_If) return nullptr; 1323 1324 if (cmpi_folds(igvn)) { 1325 Node* ctrl = in(0); 1326 if (is_ctrl_folds(ctrl, igvn) && ctrl->outcnt() == 1) { 1327 // A integer comparison immediately dominated by another integer 1328 // comparison 1329 ProjNode* success = nullptr; 1330 ProjNode* fail = nullptr; 1331 ProjNode* dom_cmp = ctrl->as_Proj(); 1332 if (has_shared_region(dom_cmp, success, fail) && 1333 // Next call modifies graph so must be last 1334 fold_compares_helper(dom_cmp, success, fail, igvn)) { 1335 return this; 1336 } 1337 if (has_only_uncommon_traps(dom_cmp, success, fail, igvn) && 1338 // Next call modifies graph so must be last 1339 fold_compares_helper(dom_cmp, success, fail, igvn)) { 1340 return merge_uncommon_traps(dom_cmp, success, fail, igvn); 1341 } 1342 return nullptr; 1343 } else if (ctrl->in(0) != nullptr && 1344 ctrl->in(0)->in(0) != nullptr) { 1345 ProjNode* success = nullptr; 1346 ProjNode* fail = nullptr; 1347 Node* dom = ctrl->in(0)->in(0); 1348 ProjNode* dom_cmp = dom->isa_Proj(); 1349 ProjNode* other_cmp = ctrl->isa_Proj(); 1350 1351 // Check if it's an integer comparison dominated by another 1352 // integer comparison with another test in between 1353 if (is_ctrl_folds(dom, igvn) && 1354 has_only_uncommon_traps(dom_cmp, success, fail, igvn) && 1355 is_side_effect_free_test(other_cmp, igvn) && 1356 // Next call modifies graph so must be last 1357 fold_compares_helper(dom_cmp, success, fail, igvn)) { 1358 reroute_side_effect_free_unc(other_cmp, dom_cmp, igvn); 1359 return merge_uncommon_traps(dom_cmp, success, fail, igvn); 1360 } 1361 } 1362 } 1363 return nullptr; 1364 } 1365 1366 //------------------------------remove_useless_bool---------------------------- 1367 // Check for people making a useless boolean: things like 1368 // if( (x < y ? true : false) ) { ... } 1369 // Replace with if( x < y ) { ... } 1370 static Node *remove_useless_bool(IfNode *iff, PhaseGVN *phase) { 1371 Node *i1 = iff->in(1); 1372 if( !i1->is_Bool() ) return nullptr; 1373 BoolNode *bol = i1->as_Bool(); 1374 1375 Node *cmp = bol->in(1); 1376 if( cmp->Opcode() != Op_CmpI ) return nullptr; 1377 1378 // Must be comparing against a bool 1379 const Type *cmp2_t = phase->type( cmp->in(2) ); 1380 if( cmp2_t != TypeInt::ZERO && 1381 cmp2_t != TypeInt::ONE ) 1382 return nullptr; 1383 1384 // Find a prior merge point merging the boolean 1385 i1 = cmp->in(1); 1386 if( !i1->is_Phi() ) return nullptr; 1387 PhiNode *phi = i1->as_Phi(); 1388 if( phase->type( phi ) != TypeInt::BOOL ) 1389 return nullptr; 1390 1391 // Check for diamond pattern 1392 int true_path = phi->is_diamond_phi(); 1393 if( true_path == 0 ) return nullptr; 1394 1395 // Make sure that iff and the control of the phi are different. This 1396 // should really only happen for dead control flow since it requires 1397 // an illegal cycle. 1398 if (phi->in(0)->in(1)->in(0) == iff) return nullptr; 1399 1400 // phi->region->if_proj->ifnode->bool->cmp 1401 BoolNode *bol2 = phi->in(0)->in(1)->in(0)->in(1)->as_Bool(); 1402 1403 // Now get the 'sense' of the test correct so we can plug in 1404 // either iff2->in(1) or its complement. 1405 int flip = 0; 1406 if( bol->_test._test == BoolTest::ne ) flip = 1-flip; 1407 else if( bol->_test._test != BoolTest::eq ) return nullptr; 1408 if( cmp2_t == TypeInt::ZERO ) flip = 1-flip; 1409 1410 const Type *phi1_t = phase->type( phi->in(1) ); 1411 const Type *phi2_t = phase->type( phi->in(2) ); 1412 // Check for Phi(0,1) and flip 1413 if( phi1_t == TypeInt::ZERO ) { 1414 if( phi2_t != TypeInt::ONE ) return nullptr; 1415 flip = 1-flip; 1416 } else { 1417 // Check for Phi(1,0) 1418 if( phi1_t != TypeInt::ONE ) return nullptr; 1419 if( phi2_t != TypeInt::ZERO ) return nullptr; 1420 } 1421 if( true_path == 2 ) { 1422 flip = 1-flip; 1423 } 1424 1425 Node* new_bol = (flip ? phase->transform( bol2->negate(phase) ) : bol2); 1426 assert(new_bol != iff->in(1), "must make progress"); 1427 iff->set_req_X(1, new_bol, phase); 1428 // Intervening diamond probably goes dead 1429 phase->C->set_major_progress(); 1430 return iff; 1431 } 1432 1433 static IfNode* idealize_test(PhaseGVN* phase, IfNode* iff); 1434 1435 struct RangeCheck { 1436 Node* ctl; 1437 jint off; 1438 }; 1439 1440 Node* IfNode::Ideal_common(PhaseGVN *phase, bool can_reshape) { 1441 if (remove_dead_region(phase, can_reshape)) return this; 1442 // No Def-Use info? 1443 if (!can_reshape) return nullptr; 1444 1445 // Don't bother trying to transform a dead if 1446 if (in(0)->is_top()) return nullptr; 1447 // Don't bother trying to transform an if with a dead test 1448 if (in(1)->is_top()) return nullptr; 1449 // Another variation of a dead test 1450 if (in(1)->is_Con()) return nullptr; 1451 // Another variation of a dead if 1452 if (outcnt() < 2) return nullptr; 1453 1454 // Canonicalize the test. 1455 Node* idt_if = idealize_test(phase, this); 1456 if (idt_if != nullptr) return idt_if; 1457 1458 // Try to split the IF 1459 PhaseIterGVN *igvn = phase->is_IterGVN(); 1460 Node *s = split_if(this, igvn); 1461 if (s != nullptr) return s; 1462 1463 return NodeSentinel; 1464 } 1465 1466 //------------------------------Ideal------------------------------------------ 1467 // Return a node which is more "ideal" than the current node. Strip out 1468 // control copies 1469 Node* IfNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1470 Node* res = Ideal_common(phase, can_reshape); 1471 if (res != NodeSentinel) { 1472 return res; 1473 } 1474 1475 // Check for people making a useless boolean: things like 1476 // if( (x < y ? true : false) ) { ... } 1477 // Replace with if( x < y ) { ... } 1478 Node* bol2 = remove_useless_bool(this, phase); 1479 if (bol2) return bol2; 1480 1481 if (in(0) == nullptr) return nullptr; // Dead loop? 1482 1483 PhaseIterGVN* igvn = phase->is_IterGVN(); 1484 Node* result = fold_compares(igvn); 1485 if (result != nullptr) { 1486 return result; 1487 } 1488 1489 // Scan for an equivalent test 1490 int dist = 4; // Cutoff limit for search 1491 if (is_If() && in(1)->is_Bool()) { 1492 Node* cmp = in(1)->in(1); 1493 if (cmp->Opcode() == Op_CmpP && 1494 cmp->in(2) != nullptr && // make sure cmp is not already dead 1495 cmp->in(2)->bottom_type() == TypePtr::NULL_PTR) { 1496 dist = 64; // Limit for null-pointer scans 1497 } 1498 } 1499 1500 Node* prev_dom = search_identical(dist); 1501 1502 if (prev_dom != nullptr) { 1503 // Replace dominated IfNode 1504 return dominated_by(prev_dom, igvn); 1505 } 1506 1507 return simple_subsuming(igvn); 1508 } 1509 1510 //------------------------------dominated_by----------------------------------- 1511 Node* IfNode::dominated_by(Node* prev_dom, PhaseIterGVN *igvn) { 1512 #ifndef PRODUCT 1513 if (TraceIterativeGVN) { 1514 tty->print(" Removing IfNode: "); this->dump(); 1515 } 1516 #endif 1517 1518 igvn->hash_delete(this); // Remove self to prevent spurious V-N 1519 Node *idom = in(0); 1520 // Need opcode to decide which way 'this' test goes 1521 int prev_op = prev_dom->Opcode(); 1522 Node *top = igvn->C->top(); // Shortcut to top 1523 1524 // Loop predicates may have depending checks which should not 1525 // be skipped. For example, range check predicate has two checks 1526 // for lower and upper bounds. 1527 ProjNode* unc_proj = proj_out(1 - prev_dom->as_Proj()->_con)->as_Proj(); 1528 if (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) != nullptr || 1529 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_profile_predicate) != nullptr) { 1530 prev_dom = idom; 1531 } 1532 1533 // Now walk the current IfNode's projections. 1534 // Loop ends when 'this' has no more uses. 1535 for (DUIterator_Last imin, i = last_outs(imin); i >= imin; --i) { 1536 Node *ifp = last_out(i); // Get IfTrue/IfFalse 1537 igvn->add_users_to_worklist(ifp); 1538 // Check which projection it is and set target. 1539 // Data-target is either the dominating projection of the same type 1540 // or TOP if the dominating projection is of opposite type. 1541 // Data-target will be used as the new control edge for the non-CFG 1542 // nodes like Casts and Loads. 1543 Node *data_target = (ifp->Opcode() == prev_op) ? prev_dom : top; 1544 // Control-target is just the If's immediate dominator or TOP. 1545 Node *ctrl_target = (ifp->Opcode() == prev_op) ? idom : top; 1546 1547 // For each child of an IfTrue/IfFalse projection, reroute. 1548 // Loop ends when projection has no more uses. 1549 for (DUIterator_Last jmin, j = ifp->last_outs(jmin); j >= jmin; --j) { 1550 Node* s = ifp->last_out(j); // Get child of IfTrue/IfFalse 1551 if (s->depends_only_on_test() && igvn->no_dependent_zero_check(s)) { 1552 // For control producers. 1553 // Do not rewire Div and Mod nodes which could have a zero divisor to avoid skipping their zero check. 1554 igvn->replace_input_of(s, 0, data_target); // Move child to data-target 1555 } else { 1556 // Find the control input matching this def-use edge. 1557 // For Regions it may not be in slot 0. 1558 uint l; 1559 for (l = 0; s->in(l) != ifp; l++) { } 1560 igvn->replace_input_of(s, l, ctrl_target); 1561 } 1562 } // End for each child of a projection 1563 1564 igvn->remove_dead_node(ifp); 1565 } // End for each IfTrue/IfFalse child of If 1566 1567 // Kill the IfNode 1568 igvn->remove_dead_node(this); 1569 1570 // Must return either the original node (now dead) or a new node 1571 // (Do not return a top here, since that would break the uniqueness of top.) 1572 return new ConINode(TypeInt::ZERO); 1573 } 1574 1575 Node* IfNode::search_identical(int dist) { 1576 // Setup to scan up the CFG looking for a dominating test 1577 Node* dom = in(0); 1578 Node* prev_dom = this; 1579 int op = Opcode(); 1580 // Search up the dominator tree for an If with an identical test 1581 while (dom->Opcode() != op || // Not same opcode? 1582 dom->in(1) != in(1) || // Not same input 1? 1583 prev_dom->in(0) != dom) { // One path of test does not dominate? 1584 if (dist < 0) return nullptr; 1585 1586 dist--; 1587 prev_dom = dom; 1588 dom = up_one_dom(dom); 1589 if (!dom) return nullptr; 1590 } 1591 1592 // Check that we did not follow a loop back to ourselves 1593 if (this == dom) { 1594 return nullptr; 1595 } 1596 1597 #ifndef PRODUCT 1598 if (dist > 2) { // Add to count of null checks elided 1599 explicit_null_checks_elided++; 1600 } 1601 #endif 1602 1603 return prev_dom; 1604 } 1605 1606 1607 static int subsuming_bool_test_encode(Node*); 1608 1609 // Check if dominating test is subsuming 'this' one. 1610 // 1611 // cmp 1612 // / \ 1613 // (r1) bool \ 1614 // / bool (r2) 1615 // (dom) if \ 1616 // \ ) 1617 // (pre) if[TF] / 1618 // \ / 1619 // if (this) 1620 // \r1 1621 // r2\ eqT eqF neT neF ltT ltF leT leF gtT gtF geT geF 1622 // eq t f f t f - - f f - - f 1623 // ne f t t f t - - t t - - t 1624 // lt f - - f t f - f f - f t 1625 // le t - - t t - t f f t - t 1626 // gt f - - f f - f t t f - f 1627 // ge t - - t f t - t t - t f 1628 // 1629 Node* IfNode::simple_subsuming(PhaseIterGVN* igvn) { 1630 // Table encoding: N/A (na), True-branch (tb), False-branch (fb). 1631 static enum { na, tb, fb } s_short_circuit_map[6][12] = { 1632 /*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*/ 1633 /*eq*/{ tb, fb, fb, tb, fb, na, na, fb, fb, na, na, fb }, 1634 /*ne*/{ fb, tb, tb, fb, tb, na, na, tb, tb, na, na, tb }, 1635 /*lt*/{ fb, na, na, fb, tb, fb, na, fb, fb, na, fb, tb }, 1636 /*le*/{ tb, na, na, tb, tb, na, tb, fb, fb, tb, na, tb }, 1637 /*gt*/{ fb, na, na, fb, fb, na, fb, tb, tb, fb, na, fb }, 1638 /*ge*/{ tb, na, na, tb, fb, tb, na, tb, tb, na, tb, fb }}; 1639 1640 Node* pre = in(0); 1641 if (!pre->is_IfTrue() && !pre->is_IfFalse()) { 1642 return nullptr; 1643 } 1644 Node* dom = pre->in(0); 1645 if (!dom->is_If()) { 1646 return nullptr; 1647 } 1648 Node* bol = in(1); 1649 if (!bol->is_Bool()) { 1650 return nullptr; 1651 } 1652 Node* cmp = in(1)->in(1); 1653 if (!cmp->is_Cmp()) { 1654 return nullptr; 1655 } 1656 1657 if (!dom->in(1)->is_Bool()) { 1658 return nullptr; 1659 } 1660 if (dom->in(1)->in(1) != cmp) { // Not same cond? 1661 return nullptr; 1662 } 1663 1664 int drel = subsuming_bool_test_encode(dom->in(1)); 1665 int trel = subsuming_bool_test_encode(bol); 1666 int bout = pre->is_IfFalse() ? 1 : 0; 1667 1668 if (drel < 0 || trel < 0) { 1669 return nullptr; 1670 } 1671 int br = s_short_circuit_map[trel][2*drel+bout]; 1672 if (br == na) { 1673 return nullptr; 1674 } 1675 #ifndef PRODUCT 1676 if (TraceIterativeGVN) { 1677 tty->print(" Subsumed IfNode: "); dump(); 1678 } 1679 #endif 1680 // Replace condition with constant True(1)/False(0). 1681 bool is_always_true = br == tb; 1682 set_req(1, igvn->intcon(is_always_true ? 1 : 0)); 1683 1684 // Update any data dependencies to the directly dominating test. This subsumed test is not immediately removed by igvn 1685 // and therefore subsequent optimizations might miss these data dependencies otherwise. There might be a dead loop 1686 // ('always_taken_proj' == 'pre') that is cleaned up later. Skip this case to make the iterator work properly. 1687 Node* always_taken_proj = proj_out(is_always_true); 1688 if (always_taken_proj != pre) { 1689 for (DUIterator_Fast imax, i = always_taken_proj->fast_outs(imax); i < imax; i++) { 1690 Node* u = always_taken_proj->fast_out(i); 1691 if (!u->is_CFG()) { 1692 igvn->replace_input_of(u, 0, pre); 1693 --i; 1694 --imax; 1695 } 1696 } 1697 } 1698 1699 if (bol->outcnt() == 0) { 1700 igvn->remove_dead_node(bol); // Kill the BoolNode. 1701 } 1702 return this; 1703 } 1704 1705 // Map BoolTest to local table encoding. The BoolTest (e)numerals 1706 // { eq = 0, ne = 4, le = 5, ge = 7, lt = 3, gt = 1 } 1707 // are mapped to table indices, while the remaining (e)numerals in BoolTest 1708 // { overflow = 2, no_overflow = 6, never = 8, illegal = 9 } 1709 // are ignored (these are not modeled in the table). 1710 // 1711 static int subsuming_bool_test_encode(Node* node) { 1712 precond(node->is_Bool()); 1713 BoolTest::mask x = node->as_Bool()->_test._test; 1714 switch (x) { 1715 case BoolTest::eq: return 0; 1716 case BoolTest::ne: return 1; 1717 case BoolTest::lt: return 2; 1718 case BoolTest::le: return 3; 1719 case BoolTest::gt: return 4; 1720 case BoolTest::ge: return 5; 1721 case BoolTest::overflow: 1722 case BoolTest::no_overflow: 1723 case BoolTest::never: 1724 case BoolTest::illegal: 1725 default: 1726 return -1; 1727 } 1728 } 1729 1730 //------------------------------Identity--------------------------------------- 1731 // If the test is constant & we match, then we are the input Control 1732 Node* IfProjNode::Identity(PhaseGVN* phase) { 1733 // Can only optimize if cannot go the other way 1734 const TypeTuple *t = phase->type(in(0))->is_tuple(); 1735 if (t == TypeTuple::IFNEITHER || (always_taken(t) && 1736 // During parsing (GVN) we don't remove dead code aggressively. 1737 // Cut off dead branch and let PhaseRemoveUseless take care of it. 1738 (!phase->is_IterGVN() || 1739 // During IGVN, first wait for the dead branch to be killed. 1740 // Otherwise, the IfNode's control will have two control uses (the IfNode 1741 // that doesn't go away because it still has uses and this branch of the 1742 // If) which breaks other optimizations. Node::has_special_unique_user() 1743 // will cause this node to be reprocessed once the dead branch is killed. 1744 in(0)->outcnt() == 1))) { 1745 // IfNode control 1746 if (in(0)->is_BaseCountedLoopEnd()) { 1747 // CountedLoopEndNode may be eliminated by if subsuming, replace CountedLoopNode with LoopNode to 1748 // avoid mismatching between CountedLoopNode and CountedLoopEndNode in the following optimization. 1749 Node* head = unique_ctrl_out_or_null(); 1750 if (head != nullptr && head->is_BaseCountedLoop() && head->in(LoopNode::LoopBackControl) == this) { 1751 Node* new_head = new LoopNode(head->in(LoopNode::EntryControl), this); 1752 phase->is_IterGVN()->register_new_node_with_optimizer(new_head); 1753 phase->is_IterGVN()->replace_node(head, new_head); 1754 } 1755 } 1756 return in(0)->in(0); 1757 } 1758 // no progress 1759 return this; 1760 } 1761 1762 #ifndef PRODUCT 1763 //------------------------------dump_spec-------------------------------------- 1764 void IfNode::dump_spec(outputStream *st) const { 1765 st->print("P=%f, C=%f",_prob,_fcnt); 1766 } 1767 #endif 1768 1769 //------------------------------idealize_test---------------------------------- 1770 // Try to canonicalize tests better. Peek at the Cmp/Bool/If sequence and 1771 // come up with a canonical sequence. Bools getting 'eq', 'gt' and 'ge' forms 1772 // converted to 'ne', 'le' and 'lt' forms. IfTrue/IfFalse get swapped as 1773 // needed. 1774 static IfNode* idealize_test(PhaseGVN* phase, IfNode* iff) { 1775 assert(iff->in(0) != nullptr, "If must be live"); 1776 1777 if (iff->outcnt() != 2) return nullptr; // Malformed projections. 1778 Node* old_if_f = iff->proj_out(false); 1779 Node* old_if_t = iff->proj_out(true); 1780 1781 // CountedLoopEnds want the back-control test to be TRUE, regardless of 1782 // whether they are testing a 'gt' or 'lt' condition. The 'gt' condition 1783 // happens in count-down loops 1784 if (iff->is_BaseCountedLoopEnd()) return nullptr; 1785 if (!iff->in(1)->is_Bool()) return nullptr; // Happens for partially optimized IF tests 1786 BoolNode *b = iff->in(1)->as_Bool(); 1787 BoolTest bt = b->_test; 1788 // Test already in good order? 1789 if( bt.is_canonical() ) 1790 return nullptr; 1791 1792 // Flip test to be canonical. Requires flipping the IfFalse/IfTrue and 1793 // cloning the IfNode. 1794 Node* new_b = phase->transform( new BoolNode(b->in(1), bt.negate()) ); 1795 if( !new_b->is_Bool() ) return nullptr; 1796 b = new_b->as_Bool(); 1797 1798 PhaseIterGVN *igvn = phase->is_IterGVN(); 1799 assert( igvn, "Test is not canonical in parser?" ); 1800 1801 // The IF node never really changes, but it needs to be cloned 1802 iff = iff->clone()->as_If(); 1803 iff->set_req(1, b); 1804 iff->_prob = 1.0-iff->_prob; 1805 1806 Node *prior = igvn->hash_find_insert(iff); 1807 if( prior ) { 1808 igvn->remove_dead_node(iff); 1809 iff = (IfNode*)prior; 1810 } else { 1811 // Cannot call transform on it just yet 1812 igvn->set_type_bottom(iff); 1813 } 1814 igvn->_worklist.push(iff); 1815 1816 // Now handle projections. Cloning not required. 1817 Node* new_if_f = (Node*)(new IfFalseNode( iff )); 1818 Node* new_if_t = (Node*)(new IfTrueNode ( iff )); 1819 1820 igvn->register_new_node_with_optimizer(new_if_f); 1821 igvn->register_new_node_with_optimizer(new_if_t); 1822 // Flip test, so flip trailing control 1823 igvn->replace_node(old_if_f, new_if_t); 1824 igvn->replace_node(old_if_t, new_if_f); 1825 1826 // Progress 1827 return iff; 1828 } 1829 1830 Node* RangeCheckNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1831 Node* res = Ideal_common(phase, can_reshape); 1832 if (res != NodeSentinel) { 1833 return res; 1834 } 1835 1836 PhaseIterGVN *igvn = phase->is_IterGVN(); 1837 // Setup to scan up the CFG looking for a dominating test 1838 Node* prev_dom = this; 1839 1840 // Check for range-check vs other kinds of tests 1841 Node* index1; 1842 Node* range1; 1843 jint offset1; 1844 int flip1 = is_range_check(range1, index1, offset1); 1845 if (flip1) { 1846 Node* dom = in(0); 1847 // Try to remove extra range checks. All 'up_one_dom' gives up at merges 1848 // so all checks we inspect post-dominate the top-most check we find. 1849 // If we are going to fail the current check and we reach the top check 1850 // then we are guaranteed to fail, so just start interpreting there. 1851 // We 'expand' the top 3 range checks to include all post-dominating 1852 // checks. 1853 1854 // The top 3 range checks seen 1855 const int NRC = 3; 1856 RangeCheck prev_checks[NRC]; 1857 int nb_checks = 0; 1858 1859 // Low and high offsets seen so far 1860 jint off_lo = offset1; 1861 jint off_hi = offset1; 1862 1863 bool found_immediate_dominator = false; 1864 1865 // Scan for the top checks and collect range of offsets 1866 for (int dist = 0; dist < 999; dist++) { // Range-Check scan limit 1867 if (dom->Opcode() == Op_RangeCheck && // Not same opcode? 1868 prev_dom->in(0) == dom) { // One path of test does dominate? 1869 if (dom == this) return nullptr; // dead loop 1870 // See if this is a range check 1871 Node* index2; 1872 Node* range2; 1873 jint offset2; 1874 int flip2 = dom->as_RangeCheck()->is_range_check(range2, index2, offset2); 1875 // See if this is a _matching_ range check, checking against 1876 // the same array bounds. 1877 if (flip2 == flip1 && range2 == range1 && index2 == index1 && 1878 dom->outcnt() == 2) { 1879 if (nb_checks == 0 && dom->in(1) == in(1)) { 1880 // Found an immediately dominating test at the same offset. 1881 // This kind of back-to-back test can be eliminated locally, 1882 // and there is no need to search further for dominating tests. 1883 assert(offset2 == offset1, "Same test but different offsets"); 1884 found_immediate_dominator = true; 1885 break; 1886 } 1887 // Gather expanded bounds 1888 off_lo = MIN2(off_lo,offset2); 1889 off_hi = MAX2(off_hi,offset2); 1890 // Record top NRC range checks 1891 prev_checks[nb_checks%NRC].ctl = prev_dom; 1892 prev_checks[nb_checks%NRC].off = offset2; 1893 nb_checks++; 1894 } 1895 } 1896 prev_dom = dom; 1897 dom = up_one_dom(dom); 1898 if (!dom) break; 1899 } 1900 1901 if (!found_immediate_dominator) { 1902 // Attempt to widen the dominating range check to cover some later 1903 // ones. Since range checks "fail" by uncommon-trapping to the 1904 // interpreter, widening a check can make us speculatively enter 1905 // the interpreter. If we see range-check deopt's, do not widen! 1906 if (!phase->C->allow_range_check_smearing()) return nullptr; 1907 1908 // Didn't find prior covering check, so cannot remove anything. 1909 if (nb_checks == 0) { 1910 return nullptr; 1911 } 1912 // Constant indices only need to check the upper bound. 1913 // Non-constant indices must check both low and high. 1914 int chk0 = (nb_checks - 1) % NRC; 1915 if (index1) { 1916 if (nb_checks == 1) { 1917 return nullptr; 1918 } else { 1919 // If the top range check's constant is the min or max of 1920 // all constants we widen the next one to cover the whole 1921 // range of constants. 1922 RangeCheck rc0 = prev_checks[chk0]; 1923 int chk1 = (nb_checks - 2) % NRC; 1924 RangeCheck rc1 = prev_checks[chk1]; 1925 if (rc0.off == off_lo) { 1926 adjust_check(rc1.ctl, range1, index1, flip1, off_hi, igvn); 1927 prev_dom = rc1.ctl; 1928 } else if (rc0.off == off_hi) { 1929 adjust_check(rc1.ctl, range1, index1, flip1, off_lo, igvn); 1930 prev_dom = rc1.ctl; 1931 } else { 1932 // If the top test's constant is not the min or max of all 1933 // constants, we need 3 range checks. We must leave the 1934 // top test unchanged because widening it would allow the 1935 // accesses it protects to successfully read/write out of 1936 // bounds. 1937 if (nb_checks == 2) { 1938 return nullptr; 1939 } 1940 int chk2 = (nb_checks - 3) % NRC; 1941 RangeCheck rc2 = prev_checks[chk2]; 1942 // The top range check a+i covers interval: -a <= i < length-a 1943 // The second range check b+i covers interval: -b <= i < length-b 1944 if (rc1.off <= rc0.off) { 1945 // if b <= a, we change the second range check to: 1946 // -min_of_all_constants <= i < length-min_of_all_constants 1947 // Together top and second range checks now cover: 1948 // -min_of_all_constants <= i < length-a 1949 // which is more restrictive than -b <= i < length-b: 1950 // -b <= -min_of_all_constants <= i < length-a <= length-b 1951 // The third check is then changed to: 1952 // -max_of_all_constants <= i < length-max_of_all_constants 1953 // so 2nd and 3rd checks restrict allowed values of i to: 1954 // -min_of_all_constants <= i < length-max_of_all_constants 1955 adjust_check(rc1.ctl, range1, index1, flip1, off_lo, igvn); 1956 adjust_check(rc2.ctl, range1, index1, flip1, off_hi, igvn); 1957 } else { 1958 // if b > a, we change the second range check to: 1959 // -max_of_all_constants <= i < length-max_of_all_constants 1960 // Together top and second range checks now cover: 1961 // -a <= i < length-max_of_all_constants 1962 // which is more restrictive than -b <= i < length-b: 1963 // -b < -a <= i < length-max_of_all_constants <= length-b 1964 // The third check is then changed to: 1965 // -max_of_all_constants <= i < length-max_of_all_constants 1966 // so 2nd and 3rd checks restrict allowed values of i to: 1967 // -min_of_all_constants <= i < length-max_of_all_constants 1968 adjust_check(rc1.ctl, range1, index1, flip1, off_hi, igvn); 1969 adjust_check(rc2.ctl, range1, index1, flip1, off_lo, igvn); 1970 } 1971 prev_dom = rc2.ctl; 1972 } 1973 } 1974 } else { 1975 RangeCheck rc0 = prev_checks[chk0]; 1976 // 'Widen' the offset of the 1st and only covering check 1977 adjust_check(rc0.ctl, range1, index1, flip1, off_hi, igvn); 1978 // Test is now covered by prior checks, dominate it out 1979 prev_dom = rc0.ctl; 1980 } 1981 } 1982 } else { 1983 prev_dom = search_identical(4); 1984 1985 if (prev_dom == nullptr) { 1986 return nullptr; 1987 } 1988 } 1989 1990 // Replace dominated IfNode 1991 return dominated_by(prev_dom, igvn); 1992 } 1993 1994 ParsePredicateNode::ParsePredicateNode(Node* control, Node* bol, Deoptimization::DeoptReason deopt_reason) 1995 : IfNode(control, bol, PROB_MAX, COUNT_UNKNOWN), 1996 _deopt_reason(deopt_reason) { 1997 init_class_id(Class_ParsePredicate); 1998 assert(bol->Opcode() == Op_Conv2B && bol->in(1) != nullptr && bol->in(1)->is_Opaque1(), "wrong boolean input"); 1999 #ifdef ASSERT 2000 switch (deopt_reason) { 2001 case Deoptimization::Reason_predicate: 2002 case Deoptimization::Reason_profile_predicate: 2003 case Deoptimization::Reason_loop_limit_check: 2004 break; 2005 default: 2006 assert(false, "unsupported deoptimization reason for Parse Predicate"); 2007 } 2008 #endif // ASSERT 2009 } 2010 2011 #ifndef PRODUCT 2012 void ParsePredicateNode::dump_spec(outputStream* st) const { 2013 st->print(" #"); 2014 switch (_deopt_reason) { 2015 case Deoptimization::DeoptReason::Reason_predicate: 2016 st->print("Loop "); 2017 break; 2018 case Deoptimization::DeoptReason::Reason_profile_predicate: 2019 st->print("Profiled_Loop "); 2020 break; 2021 case Deoptimization::DeoptReason::Reason_loop_limit_check: 2022 st->print("Loop_Limit_Check "); 2023 break; 2024 default: 2025 fatal("unknown kind"); 2026 } 2027 } 2028 #endif // NOT PRODUCT