1 /* 2 * Copyright (c) 1999, 2026, 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 "gc/shared/barrierSet.hpp" 26 #include "gc/shared/c2/barrierSetC2.hpp" 27 #include "memory/allocation.inline.hpp" 28 #include "memory/resourceArea.hpp" 29 #include "opto/addnode.hpp" 30 #include "opto/callnode.hpp" 31 #include "opto/castnode.hpp" 32 #include "opto/connode.hpp" 33 #include "opto/divnode.hpp" 34 #include "opto/inlinetypenode.hpp" 35 #include "opto/loopnode.hpp" 36 #include "opto/matcher.hpp" 37 #include "opto/movenode.hpp" 38 #include "opto/mulnode.hpp" 39 #include "opto/opaquenode.hpp" 40 #include "opto/rootnode.hpp" 41 #include "opto/subnode.hpp" 42 #include "opto/subtypenode.hpp" 43 #include "opto/superword.hpp" 44 #include "opto/vectornode.hpp" 45 #include "utilities/checkedCast.hpp" 46 #include "utilities/macros.hpp" 47 48 //============================================================================= 49 //------------------------------split_thru_phi--------------------------------- 50 // Split Node 'n' through merge point if there is enough win. 51 Node* PhaseIdealLoop::split_thru_phi(Node* n, Node* region, int policy) { 52 if ((n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) || 53 (n->Opcode() == Op_ConvL2I && n->bottom_type() != TypeInt::INT)) { 54 // ConvI2L/ConvL2I may have type information on it which is unsafe to push up 55 // so disable this for now 56 return nullptr; 57 } 58 59 // Splitting range check CastIIs through a loop induction Phi can 60 // cause new Phis to be created that are left unrelated to the loop 61 // induction Phi and prevent optimizations (vectorization) 62 if (n->Opcode() == Op_CastII && region->is_CountedLoop() && 63 n->in(1) == region->as_CountedLoop()->phi()) { 64 return nullptr; 65 } 66 67 // Inline types should not be split through Phis because they cannot be merged 68 // through Phi nodes but each value input needs to be merged individually. 69 if (n->is_InlineType()) { 70 return nullptr; 71 } 72 73 if (cannot_split_division(n, region)) { 74 return nullptr; 75 } 76 77 SplitThruPhiWins wins(region); 78 assert(!n->is_CFG(), ""); 79 assert(region->is_Region(), ""); 80 81 const Type* type = n->bottom_type(); 82 const TypeOopPtr* t_oop = _igvn.type(n)->isa_oopptr(); 83 Node* phi; 84 if (t_oop != nullptr && t_oop->is_known_instance_field()) { 85 int iid = t_oop->instance_id(); 86 int index = C->get_alias_index(t_oop); 87 int offset = t_oop->offset(); 88 phi = new PhiNode(region, type, nullptr, iid, index, offset); 89 } else { 90 phi = PhiNode::make_blank(region, n); 91 } 92 uint old_unique = C->unique(); 93 for (uint i = 1; i < region->req(); i++) { 94 Node* x; 95 Node* the_clone = nullptr; 96 if (region->in(i) == C->top()) { 97 x = C->top(); // Dead path? Use a dead data op 98 } else { 99 x = n->clone(); // Else clone up the data op 100 the_clone = x; // Remember for possible deletion. 101 // Alter data node to use pre-phi inputs 102 if (n->in(0) == region) 103 x->set_req( 0, region->in(i) ); 104 for (uint j = 1; j < n->req(); j++) { 105 Node* in = n->in(j); 106 if (in->is_Phi() && in->in(0) == region) 107 x->set_req(j, in->in(i)); // Use pre-Phi input for the clone 108 } 109 } 110 // Check for a 'win' on some paths 111 const Type* t = x->Value(&_igvn); 112 113 bool singleton = t->singleton(); 114 115 // A TOP singleton indicates that there are no possible values incoming 116 // along a particular edge. In most cases, this is OK, and the Phi will 117 // be eliminated later in an Ideal call. However, we can't allow this to 118 // happen if the singleton occurs on loop entry, as the elimination of 119 // the PhiNode may cause the resulting node to migrate back to a previous 120 // loop iteration. 121 if (singleton && t == Type::TOP) { 122 // Is_Loop() == false does not confirm the absence of a loop (e.g., an 123 // irreducible loop may not be indicated by an affirmative is_Loop()); 124 // therefore, the only top we can split thru a phi is on a backedge of 125 // a loop. 126 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 127 } 128 129 if (singleton) { 130 wins.add_win(i); 131 x = makecon(t); 132 } else { 133 // We now call Identity to try to simplify the cloned node. 134 // Note that some Identity methods call phase->type(this). 135 // Make sure that the type array is big enough for 136 // our new node, even though we may throw the node away. 137 // (Note: This tweaking with igvn only works because x is a new node.) 138 _igvn.set_type(x, t); 139 // If x is a TypeNode, capture any more-precise type permanently into Node 140 // otherwise it will be not updated during igvn->transform since 141 // igvn->type(x) is set to x->Value() already. 142 x->raise_bottom_type(t); 143 Node* y = x->Identity(&_igvn); 144 if (y != x) { 145 wins.add_win(i); 146 x = y; 147 } else { 148 y = _igvn.hash_find(x); 149 if (y == nullptr) { 150 y = similar_subtype_check(x, region->in(i)); 151 } 152 if (y) { 153 wins.add_win(i); 154 x = y; 155 } else { 156 // Else x is a new node we are keeping 157 // We do not need register_new_node_with_optimizer 158 // because set_type has already been called. 159 _igvn._worklist.push(x); 160 } 161 } 162 } 163 164 phi->set_req( i, x ); 165 166 if (the_clone == nullptr) { 167 continue; 168 } 169 170 if (the_clone != x) { 171 _igvn.remove_dead_node(the_clone, PhaseIterGVN::NodeOrigin::Speculative); 172 } else if (region->is_Loop() && i == LoopNode::LoopBackControl && 173 n->is_Load() && can_move_to_inner_loop(n, region->as_Loop(), x)) { 174 // it is not a win if 'x' moved from an outer to an inner loop 175 // this edge case can only happen for Load nodes 176 wins.reset(); 177 break; 178 } 179 } 180 // Too few wins? 181 if (!wins.profitable(policy)) { 182 _igvn.remove_dead_node(phi, PhaseIterGVN::NodeOrigin::Speculative); 183 return nullptr; 184 } 185 186 // Record Phi 187 register_new_node( phi, region ); 188 189 for (uint i2 = 1; i2 < phi->req(); i2++) { 190 Node *x = phi->in(i2); 191 // If we commoned up the cloned 'x' with another existing Node, 192 // the existing Node picks up a new use. We need to make the 193 // existing Node occur higher up so it dominates its uses. 194 Node *old_ctrl; 195 IdealLoopTree *old_loop; 196 197 if (x->is_Con()) { 198 assert(get_ctrl(x) == C->root(), "constant control is not root"); 199 continue; 200 } 201 // The occasional new node 202 if (x->_idx >= old_unique) { // Found a new, unplaced node? 203 old_ctrl = nullptr; 204 old_loop = nullptr; // Not in any prior loop 205 } else { 206 old_ctrl = get_ctrl(x); 207 old_loop = get_loop(old_ctrl); // Get prior loop 208 } 209 // New late point must dominate new use 210 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2)); 211 if (new_ctrl == old_ctrl) // Nothing is changed 212 continue; 213 214 IdealLoopTree *new_loop = get_loop(new_ctrl); 215 216 // Don't move x into a loop if its uses are 217 // outside of loop. Otherwise x will be cloned 218 // for each use outside of this loop. 219 IdealLoopTree *use_loop = get_loop(region); 220 if (!new_loop->is_member(use_loop) && 221 (old_loop == nullptr || !new_loop->is_member(old_loop))) { 222 // Take early control, later control will be recalculated 223 // during next iteration of loop optimizations. 224 new_ctrl = get_early_ctrl(x); 225 new_loop = get_loop(new_ctrl); 226 } 227 // Set new location 228 set_ctrl(x, new_ctrl); 229 // If changing loop bodies, see if we need to collect into new body 230 if (old_loop != new_loop) { 231 if (old_loop && !old_loop->_child) 232 old_loop->_body.yank(x); 233 if (!new_loop->_child) 234 new_loop->_body.push(x); // Collect body info 235 } 236 } 237 238 split_thru_phi_yank_old_nodes(n, region); 239 _igvn.replace_node(n, phi); 240 241 #ifndef PRODUCT 242 if (TraceLoopOpts) { 243 tty->print_cr("Split %d %s through %d Phi in %d %s", 244 n->_idx, n->Name(), phi->_idx, region->_idx, region->Name()); 245 } 246 #endif // !PRODUCT 247 248 return phi; 249 } 250 251 // If the region is a Loop, we are removing the old n, 252 // and need to yank it from the _body. If any phi we 253 // just split through now has no use any more, it also 254 // has to be removed. 255 void PhaseIdealLoop::split_thru_phi_yank_old_nodes(Node* n, Node* region) { 256 IdealLoopTree* region_loop = get_loop(region); 257 if (region->is_Loop() && region_loop->is_innermost()) { 258 region_loop->_body.yank(n); 259 for (uint j = 1; j < n->req(); j++) { 260 PhiNode* phi = n->in(j)->isa_Phi(); 261 // Check that phi belongs to the region and only has n as a use. 262 if (phi != nullptr && 263 phi->in(0) == region && 264 phi->unique_multiple_edges_out_or_null() == n) { 265 assert(get_ctrl(phi) == region, "sanity"); 266 assert(get_ctrl(n) == region, "sanity"); 267 region_loop->_body.yank(phi); 268 } 269 } 270 } 271 } 272 273 // Test whether node 'x' can move into an inner loop relative to node 'n'. 274 // Note: The test is not exact. Returns true if 'x' COULD end up in an inner loop, 275 // BUT it can also return true and 'x' is in the outer loop 276 bool PhaseIdealLoop::can_move_to_inner_loop(Node* n, LoopNode* n_loop, Node* x) { 277 IdealLoopTree* n_loop_tree = get_loop(n_loop); 278 IdealLoopTree* x_loop_tree = get_loop(get_early_ctrl(x)); 279 // x_loop_tree should be outer or same loop as n_loop_tree 280 return !x_loop_tree->is_member(n_loop_tree); 281 } 282 283 // Subtype checks that carry profile data don't common so look for a replacement by following edges 284 Node* PhaseIdealLoop::similar_subtype_check(const Node* x, Node* r_in) { 285 if (x->is_SubTypeCheck()) { 286 Node* in1 = x->in(1); 287 for (DUIterator_Fast imax, i = in1->fast_outs(imax); i < imax; i++) { 288 Node* u = in1->fast_out(i); 289 if (u != x && u->is_SubTypeCheck() && u->in(1) == x->in(1) && u->in(2) == x->in(2)) { 290 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) { 291 Node* bol = u->fast_out(j); 292 for (DUIterator_Fast kmax, k = bol->fast_outs(kmax); k < kmax; k++) { 293 Node* iff = bol->fast_out(k); 294 // Only dominating subtype checks are interesting: otherwise we risk replacing a subtype check by another with 295 // unrelated profile 296 if (iff->is_If() && is_dominator(iff, r_in)) { 297 return u; 298 } 299 } 300 } 301 } 302 } 303 } 304 return nullptr; 305 } 306 307 // Return true if 'n' is a Div or Mod node (without zero check If node which was removed earlier) with a loop phi divisor 308 // of a trip-counted (integer or long) loop with a backedge input that could be zero (include zero in its type range). In 309 // this case, we cannot split the division to the backedge as it could freely float above the loop exit check resulting in 310 // a division by zero. This situation is possible because the type of an increment node of an iv phi (trip-counter) could 311 // include zero while the iv phi does not (see PhiNode::Value() for trip-counted loops where we improve types of iv phis). 312 // We also need to check other loop phis as they could have been created in the same split-if pass when applying 313 // PhaseIdealLoop::split_thru_phi() to split nodes through an iv phi. 314 bool PhaseIdealLoop::cannot_split_division(const Node* n, const Node* region) const { 315 const Type* zero; 316 switch (n->Opcode()) { 317 case Op_DivI: 318 case Op_ModI: 319 case Op_UDivI: 320 case Op_UModI: 321 zero = TypeInt::ZERO; 322 break; 323 case Op_DivL: 324 case Op_ModL: 325 case Op_UDivL: 326 case Op_UModL: 327 zero = TypeLong::ZERO; 328 break; 329 default: 330 return false; 331 } 332 333 if (n->in(0) != nullptr) { 334 // Cannot split through phi if Div or Mod node has a control dependency to a zero check. 335 return true; 336 } 337 338 Node* divisor = n->in(2); 339 return is_divisor_loop_phi(divisor, region) && 340 loop_phi_backedge_type_contains_zero(divisor, zero); 341 } 342 343 bool PhaseIdealLoop::is_divisor_loop_phi(const Node* divisor, const Node* loop) { 344 return loop->is_Loop() && divisor->is_Phi() && divisor->in(0) == loop; 345 } 346 347 bool PhaseIdealLoop::loop_phi_backedge_type_contains_zero(const Node* phi_divisor, const Type* zero) const { 348 return _igvn.type(phi_divisor->in(LoopNode::LoopBackControl))->filter_speculative(zero) != Type::TOP; 349 } 350 351 //------------------------------dominated_by------------------------------------ 352 // Replace the dominated test with an obvious true or false. Place it on the 353 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the 354 // live path up to the dominating control. 355 void PhaseIdealLoop::dominated_by(IfProjNode* prevdom, IfNode* iff, bool flip, bool prevdom_not_imply_this) { 356 if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); } 357 358 // prevdom is the dominating projection of the dominating test. 359 assert(iff->Opcode() == Op_If || 360 iff->Opcode() == Op_CountedLoopEnd || 361 iff->Opcode() == Op_LongCountedLoopEnd || 362 iff->Opcode() == Op_RangeCheck || 363 iff->Opcode() == Op_ParsePredicate, 364 "Check this code when new subtype is added"); 365 366 int pop = prevdom->Opcode(); 367 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" ); 368 if (flip) { 369 if (pop == Op_IfTrue) 370 pop = Op_IfFalse; 371 else 372 pop = Op_IfTrue; 373 } 374 // 'con' is set to true or false to kill the dominated test. 375 Node* con = makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO); 376 // Hack the dominated test 377 _igvn.replace_input_of(iff, 1, con); 378 379 // If I don't have a reachable TRUE and FALSE path following the IfNode then 380 // I can assume this path reaches an infinite loop. In this case it's not 381 // important to optimize the data Nodes - either the whole compilation will 382 // be tossed or this path (and all data Nodes) will go dead. 383 if (iff->outcnt() != 2) { 384 return; 385 } 386 387 // Make control-dependent data Nodes on the live path (path that will remain 388 // once the dominated IF is removed) become control-dependent on the 389 // dominating projection. 390 Node* dp = iff->proj_out_or_null(pop == Op_IfTrue); 391 392 if (dp == nullptr) { 393 return; 394 } 395 396 rewire_safe_outputs_to_dominator(dp, prevdom, prevdom_not_imply_this); 397 } 398 399 void PhaseIdealLoop::rewire_safe_outputs_to_dominator(Node* source, Node* dominator, const bool dominator_not_imply_source) { 400 IdealLoopTree* old_loop = get_loop(source); 401 402 for (DUIterator_Fast imax, i = source->fast_outs(imax); i < imax; i++) { 403 Node* out = source->fast_out(i); // Control-dependent node 404 if (out->depends_only_on_test()) { 405 assert(out->in(0) == source, "must be control dependent on source"); 406 _igvn.replace_input_of(out, 0, dominator); 407 if (dominator_not_imply_source) { 408 // Because of Loop Predication, Loads and range check Cast nodes that are control dependent on this range 409 // check (that is about to be removed) now depend on multiple dominating Hoisted Check Predicates. After the 410 // removal of this range check, these control dependent nodes end up at the lowest/nearest dominating predicate 411 // in the graph. To ensure that these Loads/Casts do not float above any of the dominating checks (even when the 412 // lowest dominating check is later replaced by yet another dominating check), we need to pin them at the lowest 413 // dominating check. 414 Node* clone = out->pin_node_under_control(); 415 if (clone != nullptr) { 416 clone = _igvn.register_new_node_with_optimizer(clone, out); 417 _igvn.replace_node(out, clone); 418 out = clone; 419 } 420 } 421 set_early_ctrl(out, false); 422 IdealLoopTree* new_loop = get_loop(get_ctrl(out)); 423 if (old_loop != new_loop) { 424 if (!old_loop->_child) { 425 old_loop->_body.yank(out); 426 } 427 if (!new_loop->_child) { 428 new_loop->_body.push(out); 429 } 430 } 431 --i; 432 --imax; 433 } 434 } 435 } 436 437 //------------------------------has_local_phi_input---------------------------- 438 // Return TRUE if 'n' has Phi inputs from its local block and no other 439 // block-local inputs (all non-local-phi inputs come from earlier blocks) 440 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) { 441 Node *n_ctrl = get_ctrl(n); 442 // See if some inputs come from a Phi in this block, or from before 443 // this block. 444 uint i; 445 for( i = 1; i < n->req(); i++ ) { 446 Node *phi = n->in(i); 447 if( phi->is_Phi() && phi->in(0) == n_ctrl ) 448 break; 449 } 450 if( i >= n->req() ) 451 return nullptr; // No Phi inputs; nowhere to clone thru 452 453 // Check for inputs created between 'n' and the Phi input. These 454 // must split as well; they have already been given the chance 455 // (courtesy of a post-order visit) and since they did not we must 456 // recover the 'cost' of splitting them by being very profitable 457 // when splitting 'n'. Since this is unlikely we simply give up. 458 for( i = 1; i < n->req(); i++ ) { 459 Node *m = n->in(i); 460 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) { 461 // We allow the special case of AddP's with no local inputs. 462 // This allows us to split-up address expressions. 463 if (m->is_AddP() && 464 get_ctrl(m->in(AddPNode::Base)) != n_ctrl && 465 get_ctrl(m->in(AddPNode::Address)) != n_ctrl && 466 get_ctrl(m->in(AddPNode::Offset)) != n_ctrl) { 467 // Move the AddP up to the dominating point. That's fine because control of m's inputs 468 // must dominate get_ctrl(m) == n_ctrl and we just checked that the input controls are != n_ctrl. 469 Node* c = find_non_split_ctrl(idom(n_ctrl)); 470 if (c->is_OuterStripMinedLoop()) { 471 c->as_Loop()->verify_strip_mined(1); 472 c = c->in(LoopNode::EntryControl); 473 } 474 set_ctrl_and_loop(m, c); 475 continue; 476 } 477 return nullptr; 478 } 479 assert(n->is_Phi() || m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control"); 480 } 481 482 return n_ctrl; 483 } 484 485 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2). 486 Node* PhaseIdealLoop::remix_address_expressions_add_left_shift(Node* n, IdealLoopTree* n_loop, Node* n_ctrl, BasicType bt) { 487 assert(bt == T_INT || bt == T_LONG, "only for integers"); 488 int n_op = n->Opcode(); 489 490 if (n_op == Op_LShift(bt)) { 491 // Scale is loop invariant 492 Node* scale = n->in(2); 493 Node* scale_ctrl = get_ctrl(scale); 494 IdealLoopTree* scale_loop = get_loop(scale_ctrl); 495 if (n_loop == scale_loop || !scale_loop->is_member(n_loop)) { 496 return nullptr; 497 } 498 const TypeInt* scale_t = scale->bottom_type()->isa_int(); 499 if (scale_t != nullptr && scale_t->is_con() && scale_t->get_con() >= 16) { 500 return nullptr; // Dont bother with byte/short masking 501 } 502 // Add must vary with loop (else shift would be loop-invariant) 503 Node* add = n->in(1); 504 Node* add_ctrl = get_ctrl(add); 505 IdealLoopTree* add_loop = get_loop(add_ctrl); 506 if (n_loop != add_loop) { 507 return nullptr; // happens w/ evil ZKM loops 508 } 509 510 // Convert I-V into I+ (0-V); same for V-I 511 if (add->Opcode() == Op_Sub(bt) && 512 _igvn.type(add->in(1)) != TypeInteger::zero(bt)) { 513 assert(add->Opcode() == Op_SubI || add->Opcode() == Op_SubL, ""); 514 Node* zero = integercon(0, bt); 515 Node* neg = SubNode::make(zero, add->in(2), bt); 516 register_new_node_with_ctrl_of(neg, add->in(2)); 517 add = AddNode::make(add->in(1), neg, bt); 518 register_new_node(add, add_ctrl); 519 } 520 if (add->Opcode() != Op_Add(bt)) return nullptr; 521 assert(add->Opcode() == Op_AddI || add->Opcode() == Op_AddL, ""); 522 // See if one add input is loop invariant 523 Node* add_var = add->in(1); 524 Node* add_var_ctrl = get_ctrl(add_var); 525 IdealLoopTree* add_var_loop = get_loop(add_var_ctrl); 526 Node* add_invar = add->in(2); 527 Node* add_invar_ctrl = get_ctrl(add_invar); 528 IdealLoopTree* add_invar_loop = get_loop(add_invar_ctrl); 529 if (add_invar_loop == n_loop) { 530 // Swap to find the invariant part 531 add_invar = add_var; 532 add_invar_ctrl = add_var_ctrl; 533 add_invar_loop = add_var_loop; 534 add_var = add->in(2); 535 } else if (add_var_loop != n_loop) { // Else neither input is loop invariant 536 return nullptr; 537 } 538 if (n_loop == add_invar_loop || !add_invar_loop->is_member(n_loop)) { 539 return nullptr; // No invariant part of the add? 540 } 541 542 // Yes! Reshape address expression! 543 Node* inv_scale = LShiftNode::make(add_invar, scale, bt); 544 Node* inv_scale_ctrl = 545 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ? 546 add_invar_ctrl : scale_ctrl; 547 register_new_node(inv_scale, inv_scale_ctrl); 548 Node* var_scale = LShiftNode::make(add_var, scale, bt); 549 register_new_node(var_scale, n_ctrl); 550 Node* var_add = AddNode::make(var_scale, inv_scale, bt); 551 register_new_node(var_add, n_ctrl); 552 _igvn.replace_node(n, var_add); 553 return var_add; 554 } 555 return nullptr; 556 } 557 558 //------------------------------remix_address_expressions---------------------- 559 // Rework addressing expressions to get the most loop-invariant stuff 560 // moved out. We'd like to do all associative operators, but it's especially 561 // important (common) to do address expressions. 562 Node* PhaseIdealLoop::remix_address_expressions(Node* n) { 563 if (!has_ctrl(n)) return nullptr; 564 Node* n_ctrl = get_ctrl(n); 565 IdealLoopTree* n_loop = get_loop(n_ctrl); 566 567 // See if 'n' mixes loop-varying and loop-invariant inputs and 568 // itself is loop-varying. 569 570 // Only interested in binary ops (and AddP) 571 if (n->req() < 3 || n->req() > 4) return nullptr; 572 573 Node* n1_ctrl = get_ctrl(n->in( 1)); 574 Node* n2_ctrl = get_ctrl(n->in( 2)); 575 Node* n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3)); 576 IdealLoopTree* n1_loop = get_loop(n1_ctrl); 577 IdealLoopTree* n2_loop = get_loop(n2_ctrl); 578 IdealLoopTree* n3_loop = get_loop(n3_ctrl); 579 580 // Does one of my inputs spin in a tighter loop than self? 581 if ((n_loop->is_member(n1_loop) && n_loop != n1_loop) || 582 (n_loop->is_member(n2_loop) && n_loop != n2_loop) || 583 (n_loop->is_member(n3_loop) && n_loop != n3_loop)) { 584 return nullptr; // Leave well enough alone 585 } 586 587 // Is at least one of my inputs loop-invariant? 588 if (n1_loop == n_loop && 589 n2_loop == n_loop && 590 n3_loop == n_loop) { 591 return nullptr; // No loop-invariant inputs 592 } 593 594 Node* res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_INT); 595 if (res != nullptr) { 596 return res; 597 } 598 res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_LONG); 599 if (res != nullptr) { 600 return res; 601 } 602 603 int n_op = n->Opcode(); 604 // Replace (I+V) with (V+I) 605 if (n_op == Op_AddI || 606 n_op == Op_AddL || 607 n_op == Op_AddF || 608 n_op == Op_AddD || 609 n_op == Op_MulI || 610 n_op == Op_MulL || 611 n_op == Op_MulF || 612 n_op == Op_MulD) { 613 if (n2_loop == n_loop) { 614 assert(n1_loop != n_loop, ""); 615 n->swap_edges(1, 2); 616 } 617 } 618 619 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V), 620 // but not if I2 is a constant. Skip for irreducible loops. 621 if (n_op == Op_AddP && n_loop->_head->is_Loop()) { 622 if (n2_loop == n_loop && n3_loop != n_loop) { 623 if (n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con()) { 624 Node* n22_ctrl = get_ctrl(n->in(2)->in(2)); 625 Node* n23_ctrl = get_ctrl(n->in(2)->in(3)); 626 IdealLoopTree* n22loop = get_loop(n22_ctrl); 627 IdealLoopTree* n23_loop = get_loop(n23_ctrl); 628 if (n22loop != n_loop && n22loop->is_member(n_loop) && 629 n23_loop == n_loop) { 630 Node* add1 = AddPNode::make_with_base(n->in(1), n->in(2)->in(2), n->in(3)); 631 // Stuff new AddP in the loop preheader 632 register_new_node(add1, n_loop->_head->as_Loop()->skip_strip_mined(1)->in(LoopNode::EntryControl)); 633 Node* add2 = AddPNode::make_with_base(n->in(1), add1, n->in(2)->in(3)); 634 register_new_node(add2, n_ctrl); 635 _igvn.replace_node(n, add2); 636 return add2; 637 } 638 } 639 } 640 641 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V) 642 if (n2_loop != n_loop && n3_loop == n_loop) { 643 if (n->in(3)->Opcode() == Op_AddX) { 644 Node* V = n->in(3)->in(1); 645 Node* I = n->in(3)->in(2); 646 if (ctrl_is_member(n_loop, V)) { 647 } else { 648 Node *tmp = V; V = I; I = tmp; 649 } 650 if (!ctrl_is_member(n_loop, I)) { 651 Node* add1 = AddPNode::make_with_base(n->in(1), n->in(2), I); 652 // Stuff new AddP in the loop preheader 653 register_new_node(add1, n_loop->_head->as_Loop()->skip_strip_mined(1)->in(LoopNode::EntryControl)); 654 Node* add2 = AddPNode::make_with_base(n->in(1), add1, V); 655 register_new_node(add2, n_ctrl); 656 _igvn.replace_node(n, add2); 657 return add2; 658 } 659 } 660 } 661 } 662 663 return nullptr; 664 } 665 666 // Optimize ((in1[2*i] * in2[2*i]) + (in1[2*i+1] * in2[2*i+1])) 667 Node *PhaseIdealLoop::convert_add_to_muladd(Node* n) { 668 assert(n->Opcode() == Op_AddI, "sanity"); 669 Node * nn = nullptr; 670 Node * in1 = n->in(1); 671 Node * in2 = n->in(2); 672 if (in1->Opcode() == Op_MulI && in2->Opcode() == Op_MulI) { 673 IdealLoopTree* loop_n = get_loop(get_ctrl(n)); 674 if (loop_n->is_counted() && 675 loop_n->_head->as_Loop()->is_valid_counted_loop(T_INT) && 676 Matcher::match_rule_supported(Op_MulAddVS2VI) && 677 Matcher::match_rule_supported(Op_MulAddS2I)) { 678 Node* mul_in1 = in1->in(1); 679 Node* mul_in2 = in1->in(2); 680 Node* mul_in3 = in2->in(1); 681 Node* mul_in4 = in2->in(2); 682 if (mul_in1->Opcode() == Op_LoadS && 683 mul_in2->Opcode() == Op_LoadS && 684 mul_in3->Opcode() == Op_LoadS && 685 mul_in4->Opcode() == Op_LoadS) { 686 IdealLoopTree* loop1 = get_loop(get_ctrl(mul_in1)); 687 IdealLoopTree* loop2 = get_loop(get_ctrl(mul_in2)); 688 IdealLoopTree* loop3 = get_loop(get_ctrl(mul_in3)); 689 IdealLoopTree* loop4 = get_loop(get_ctrl(mul_in4)); 690 IdealLoopTree* loop5 = get_loop(get_ctrl(in1)); 691 IdealLoopTree* loop6 = get_loop(get_ctrl(in2)); 692 // All nodes should be in the same counted loop. 693 if (loop_n == loop1 && loop_n == loop2 && loop_n == loop3 && 694 loop_n == loop4 && loop_n == loop5 && loop_n == loop6) { 695 Node* adr1 = mul_in1->in(MemNode::Address); 696 Node* adr2 = mul_in2->in(MemNode::Address); 697 Node* adr3 = mul_in3->in(MemNode::Address); 698 Node* adr4 = mul_in4->in(MemNode::Address); 699 if (adr1->is_AddP() && adr2->is_AddP() && adr3->is_AddP() && adr4->is_AddP()) { 700 if ((adr1->in(AddPNode::Base) == adr3->in(AddPNode::Base)) && 701 (adr2->in(AddPNode::Base) == adr4->in(AddPNode::Base))) { 702 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in3, mul_in4); 703 register_new_node_with_ctrl_of(nn, n); 704 _igvn.replace_node(n, nn); 705 return nn; 706 } else if ((adr1->in(AddPNode::Base) == adr4->in(AddPNode::Base)) && 707 (adr2->in(AddPNode::Base) == adr3->in(AddPNode::Base))) { 708 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in4, mul_in3); 709 register_new_node_with_ctrl_of(nn, n); 710 _igvn.replace_node(n, nn); 711 return nn; 712 } 713 } 714 } 715 } 716 } 717 } 718 return nn; 719 } 720 721 //------------------------------conditional_move------------------------------- 722 // Attempt to replace a Phi with a conditional move. We have some pretty 723 // strict profitability requirements. All Phis at the merge point must 724 // be converted, so we can remove the control flow. We need to limit the 725 // number of c-moves to a small handful. All code that was in the side-arms 726 // of the CFG diamond is now speculatively executed. This code has to be 727 // "cheap enough". We are pretty much limited to CFG diamonds that merge 728 // 1 or 2 items with a total of 1 or 2 ops executed speculatively. 729 Node *PhaseIdealLoop::conditional_move( Node *region ) { 730 731 assert(region->is_Region(), "sanity check"); 732 if (region->req() != 3) return nullptr; 733 734 // Check for CFG diamond 735 Node *lp = region->in(1); 736 Node *rp = region->in(2); 737 if (!lp || !rp) return nullptr; 738 Node *lp_c = lp->in(0); 739 if (lp_c == nullptr || lp_c != rp->in(0) || !lp_c->is_If()) return nullptr; 740 IfNode *iff = lp_c->as_If(); 741 742 // Check for ops pinned in an arm of the diamond. 743 // Can't remove the control flow in this case 744 if (lp->outcnt() > 1) return nullptr; 745 if (rp->outcnt() > 1) return nullptr; 746 747 IdealLoopTree* r_loop = get_loop(region); 748 assert(r_loop == get_loop(iff), "sanity"); 749 // Always convert to CMOVE if all results are used only outside this loop. 750 bool used_inside_loop = (r_loop == _ltree_root); 751 752 // Check profitability 753 int cost = 0; 754 int phis = 0; 755 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 756 Node *out = region->fast_out(i); 757 if (!out->is_Phi()) continue; // Ignore other control edges, etc 758 phis++; 759 PhiNode* phi = out->as_Phi(); 760 BasicType bt = phi->type()->basic_type(); 761 switch (bt) { 762 case T_DOUBLE: 763 case T_FLOAT: 764 if (C->use_cmove()) { 765 continue; //TODO: maybe we want to add some cost 766 } 767 cost += Matcher::float_cmove_cost(); // Could be very expensive 768 break; 769 case T_LONG: { 770 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's 771 } 772 case T_INT: // These all CMOV fine 773 case T_ADDRESS: { // (RawPtr) 774 cost++; 775 break; 776 } 777 case T_NARROWOOP: // Fall through 778 case T_OBJECT: { // Base oops are OK, but not derived oops 779 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr(); 780 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a 781 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus 782 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we 783 // have a Phi for the base here that we convert to a CMOVE all is well 784 // and good. But if the base is dead, we'll not make a CMOVE. Later 785 // the allocator will have to produce a base by creating a CMOVE of the 786 // relevant bases. This puts the allocator in the business of 787 // manufacturing expensive instructions, generally a bad plan. 788 // Just Say No to Conditionally-Moved Derived Pointers. 789 if (tp && tp->offset() != 0) 790 return nullptr; 791 cost++; 792 break; 793 } 794 default: 795 return nullptr; // In particular, can't do memory or I/O 796 } 797 // Add in cost any speculative ops 798 for (uint j = 1; j < region->req(); j++) { 799 Node *proj = region->in(j); 800 Node *inp = phi->in(j); 801 if (inp->isa_InlineType()) { 802 // TODO 8302217 This prevents PhiNode::push_inline_types_through 803 return nullptr; 804 } 805 if (get_ctrl(inp) == proj) { // Found local op 806 cost++; 807 // Check for a chain of dependent ops; these will all become 808 // speculative in a CMOV. 809 for (uint k = 1; k < inp->req(); k++) 810 if (get_ctrl(inp->in(k)) == proj) 811 cost += ConditionalMoveLimit; // Too much speculative goo 812 } 813 } 814 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode. 815 // This will likely Split-If, a higher-payoff operation. 816 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) { 817 Node* use = phi->fast_out(k); 818 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr()) 819 cost += ConditionalMoveLimit; 820 // Is there a use inside the loop? 821 // Note: check only basic types since CMoveP is pinned. 822 if (!used_inside_loop && is_java_primitive(bt)) { 823 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use); 824 if (r_loop == u_loop || r_loop->is_member(u_loop)) { 825 used_inside_loop = true; 826 } 827 } 828 } 829 }//for 830 Node* bol = iff->in(1); 831 assert(!bol->is_OpaqueInitializedAssertionPredicate(), "Initialized Assertion Predicates cannot form a diamond with Halt"); 832 if (bol->is_OpaqueTemplateAssertionPredicate()) { 833 // Ignore Template Assertion Predicates with OpaqueTemplateAssertionPredicate nodes. 834 return nullptr; 835 } 836 if (bol->is_OpaqueMultiversioning()) { 837 assert(bol->as_OpaqueMultiversioning()->is_useless(), "Must be useless, i.e. fast main loop has already disappeared."); 838 // Ignore multiversion_if that just lost its loops. The OpaqueMultiversioning is marked useless, 839 // and will make the multiversion_if constant fold in the next IGVN round. 840 return nullptr; 841 } 842 if (!bol->is_Bool()) { 843 assert(false, "Expected Bool, but got %s", NodeClassNames[bol->Opcode()]); 844 return nullptr; 845 } 846 int cmp_op = bol->in(1)->Opcode(); 847 if (cmp_op == Op_SubTypeCheck) { // SubTypeCheck expansion expects an IfNode 848 return nullptr; 849 } 850 // It is expensive to generate flags from a float compare. 851 // Avoid duplicated float compare. 852 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return nullptr; 853 854 float infrequent_prob = PROB_UNLIKELY_MAG(3); 855 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop. 856 if (used_inside_loop) { 857 if (cost >= ConditionalMoveLimit) return nullptr; // Too much goo 858 859 // BlockLayoutByFrequency optimization moves infrequent branch 860 // from hot path. No point in CMOV'ing in such case (110 is used 861 // instead of 100 to take into account not exactness of float value). 862 if (BlockLayoutByFrequency) { 863 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f); 864 } 865 } 866 // Check for highly predictable branch. No point in CMOV'ing if 867 // we are going to predict accurately all the time. 868 if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) { 869 //keep going 870 } else if (iff->_prob < infrequent_prob || 871 iff->_prob > (1.0f - infrequent_prob)) 872 return nullptr; 873 874 // -------------- 875 // Now replace all Phis with CMOV's 876 Node *cmov_ctrl = iff->in(0); 877 uint flip = (lp->Opcode() == Op_IfTrue); 878 Node_List wq; 879 while (1) { 880 PhiNode* phi = nullptr; 881 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 882 Node *out = region->fast_out(i); 883 if (out->is_Phi()) { 884 phi = out->as_Phi(); 885 break; 886 } 887 } 888 if (phi == nullptr || _igvn.type(phi) == Type::TOP || !CMoveNode::supported(_igvn.type(phi))) { 889 break; 890 } 891 // Move speculative ops 892 wq.push(phi); 893 while (wq.size() > 0) { 894 Node *n = wq.pop(); 895 for (uint j = 1; j < n->req(); j++) { 896 Node* m = n->in(j); 897 if (m != nullptr && !is_dominator(get_ctrl(m), cmov_ctrl)) { 898 set_ctrl(m, cmov_ctrl); 899 wq.push(m); 900 } 901 } 902 } 903 Node* cmov = CMoveNode::make(iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi)); 904 register_new_node(cmov, cmov_ctrl); 905 _igvn.replace_node(phi, cmov); 906 #ifndef PRODUCT 907 if (TraceLoopOpts) { 908 tty->print("CMOV "); 909 r_loop->dump_head(); 910 if (Verbose) { 911 bol->in(1)->dump(1); 912 cmov->dump(1); 913 } 914 } 915 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } ); 916 #endif 917 } 918 919 // The useless CFG diamond will fold up later; see the optimization in 920 // RegionNode::Ideal. 921 _igvn._worklist.push(region); 922 923 return iff->in(1); 924 } 925 926 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) { 927 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 928 Node* u = m->fast_out(i); 929 if (u->is_CFG()) { 930 if (u->is_NeverBranch()) { 931 u = u->as_NeverBranch()->proj_out(0); 932 enqueue_cfg_uses(u, wq); 933 } else { 934 wq.push(u); 935 } 936 } 937 } 938 } 939 940 // Try moving a store out of a loop, right before the loop 941 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) { 942 // Store has to be first in the loop body 943 IdealLoopTree *n_loop = get_loop(n_ctrl); 944 if (n->is_Store() && n_loop != _ltree_root && 945 n_loop->is_loop() && n_loop->_head->is_Loop() && 946 n->in(0) != nullptr) { 947 Node* address = n->in(MemNode::Address); 948 Node* value = n->in(MemNode::ValueIn); 949 Node* mem = n->in(MemNode::Memory); 950 951 // - address and value must be loop invariant 952 // - memory must be a memory Phi for the loop 953 // - Store must be the only store on this memory slice in the 954 // loop: if there's another store following this one then value 955 // written at iteration i by the second store could be overwritten 956 // at iteration i+n by the first store: it's not safe to move the 957 // first store out of the loop 958 // - nothing must observe the memory Phi: it guarantees no read 959 // before the store, we are also guaranteed the store post 960 // dominates the loop head (ignoring a possible early 961 // exit). Otherwise there would be extra Phi involved between the 962 // loop's Phi and the store. 963 // - there must be no early exit from the loop before the Store 964 // (such an exit most of the time would be an extra use of the 965 // memory Phi but sometimes is a bottom memory Phi that takes the 966 // store as input). 967 968 if (!ctrl_is_member(n_loop, address) && 969 !ctrl_is_member(n_loop, value) && 970 mem->is_Phi() && mem->in(0) == n_loop->_head && 971 mem->outcnt() == 1 && 972 mem->in(LoopNode::LoopBackControl) == n) { 973 974 assert(n_loop->_tail != nullptr, "need a tail"); 975 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop"); 976 977 // Verify that there's no early exit of the loop before the store. 978 bool ctrl_ok = false; 979 { 980 // Follow control from loop head until n, we exit the loop or 981 // we reach the tail 982 ResourceMark rm; 983 Unique_Node_List wq; 984 wq.push(n_loop->_head); 985 986 for (uint next = 0; next < wq.size(); ++next) { 987 Node *m = wq.at(next); 988 if (m == n->in(0)) { 989 ctrl_ok = true; 990 continue; 991 } 992 assert(!has_ctrl(m), "should be CFG"); 993 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) { 994 ctrl_ok = false; 995 break; 996 } 997 enqueue_cfg_uses(m, wq); 998 if (wq.size() > 10) { 999 ctrl_ok = false; 1000 break; 1001 } 1002 } 1003 } 1004 if (ctrl_ok) { 1005 // move the Store 1006 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem); 1007 _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl)); 1008 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl)); 1009 // Disconnect the phi now. An empty phi can confuse other 1010 // optimizations in this pass of loop opts. 1011 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl)); 1012 n_loop->_body.yank(mem); 1013 1014 set_ctrl_and_loop(n, n->in(0)); 1015 1016 return n; 1017 } 1018 } 1019 } 1020 return nullptr; 1021 } 1022 1023 // Try moving a store out of a loop, right after the loop 1024 void PhaseIdealLoop::try_move_store_after_loop(Node* n) { 1025 if (n->is_Store() && n->in(0) != nullptr) { 1026 Node *n_ctrl = get_ctrl(n); 1027 IdealLoopTree *n_loop = get_loop(n_ctrl); 1028 // Store must be in a loop 1029 if (n_loop != _ltree_root && !n_loop->_irreducible) { 1030 Node* address = n->in(MemNode::Address); 1031 Node* value = n->in(MemNode::ValueIn); 1032 // address must be loop invariant 1033 if (!ctrl_is_member(n_loop, address)) { 1034 // Store must be last on this memory slice in the loop and 1035 // nothing in the loop must observe it 1036 Node* phi = nullptr; 1037 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1038 Node* u = n->fast_out(i); 1039 if (has_ctrl(u)) { // control use? 1040 if (!ctrl_is_member(n_loop, u)) { 1041 continue; 1042 } 1043 if (u->is_Phi() && u->in(0) == n_loop->_head) { 1044 assert(_igvn.type(u) == Type::MEMORY, "bad phi"); 1045 // multiple phis on the same slice are possible 1046 if (phi != nullptr) { 1047 return; 1048 } 1049 phi = u; 1050 continue; 1051 } 1052 } 1053 return; 1054 } 1055 if (phi != nullptr) { 1056 // Nothing in the loop before the store (next iteration) 1057 // must observe the stored value 1058 bool mem_ok = true; 1059 { 1060 ResourceMark rm; 1061 Unique_Node_List wq; 1062 wq.push(phi); 1063 for (uint next = 0; next < wq.size() && mem_ok; ++next) { 1064 Node *m = wq.at(next); 1065 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) { 1066 Node* u = m->fast_out(i); 1067 if (u->is_Store() || u->is_Phi()) { 1068 if (u != n) { 1069 wq.push(u); 1070 mem_ok = (wq.size() <= 10); 1071 } 1072 } else { 1073 mem_ok = false; 1074 break; 1075 } 1076 } 1077 } 1078 } 1079 if (mem_ok) { 1080 // Move the store out of the loop if the LCA of all 1081 // users (except for the phi) is outside the loop. 1082 Node* hook = new Node(1); 1083 hook->init_req(0, n_ctrl); // Add an input to prevent hook from being dead 1084 _igvn.rehash_node_delayed(phi); 1085 int count = phi->replace_edge(n, hook, &_igvn); 1086 assert(count > 0, "inconsistent phi"); 1087 1088 // Compute latest point this store can go 1089 Node* lca = get_late_ctrl(n, get_ctrl(n)); 1090 if (lca->is_OuterStripMinedLoop()) { 1091 lca = lca->in(LoopNode::EntryControl); 1092 } 1093 if (n_loop->is_member(get_loop(lca))) { 1094 // LCA is in the loop - bail out 1095 _igvn.replace_node(hook, n); 1096 return; 1097 } 1098 #ifdef ASSERT 1099 if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) { 1100 assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined"); 1101 n_loop->_head->as_Loop()->verify_strip_mined(1); 1102 Node* outer = n_loop->_head->as_CountedLoop()->outer_loop(); 1103 IdealLoopTree* outer_loop = get_loop(outer); 1104 assert(n_loop->_parent == outer_loop, "broken loop tree"); 1105 assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state"); 1106 } 1107 #endif 1108 lca = place_outside_loop(lca, n_loop); 1109 assert(!n_loop->is_member(get_loop(lca)), "control must not be back in the loop"); 1110 assert(get_loop(lca)->_nest < n_loop->_nest || get_loop(lca)->_head->as_Loop()->is_in_infinite_subgraph(), "must not be moved into inner loop"); 1111 1112 // Move store out of the loop 1113 _igvn.replace_node(hook, n->in(MemNode::Memory)); 1114 _igvn.replace_input_of(n, 0, lca); 1115 set_ctrl_and_loop(n, lca); 1116 1117 // Disconnect the phi now. An empty phi can confuse other 1118 // optimizations in this pass of loop opts.. 1119 if (phi->in(LoopNode::LoopBackControl) == phi) { 1120 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl)); 1121 n_loop->_body.yank(phi); 1122 } 1123 } 1124 } 1125 } 1126 } 1127 } 1128 } 1129 1130 // We can't use immutable memory for the flat array check because we are loading the mark word which is 1131 // mutable. Although the bits we are interested in are immutable (we check for markWord::unlocked_value), 1132 // we need to use raw memory to not break anti dependency analysis. Below code will attempt to still move 1133 // flat array checks out of loops, mainly to enable loop unswitching. 1134 void PhaseIdealLoop::move_flat_array_check_out_of_loop(Node* n) { 1135 // Skip checks for more than one array 1136 if (n->req() > 3) { 1137 return; 1138 } 1139 Node* mem = n->in(FlatArrayCheckNode::Memory); 1140 Node* array = n->in(FlatArrayCheckNode::ArrayOrKlass)->uncast(); 1141 IdealLoopTree* check_loop = get_loop(get_ctrl(n)); 1142 IdealLoopTree* ary_loop = get_loop(get_ctrl(array)); 1143 1144 // Check if array is loop invariant 1145 if (!check_loop->is_member(ary_loop)) { 1146 // Walk up memory graph from the check until we leave the loop 1147 VectorSet wq; 1148 wq.set(mem->_idx); 1149 while (check_loop->is_member(get_loop(ctrl_or_self(mem)))) { 1150 if (mem->is_Phi()) { 1151 mem = mem->in(1); 1152 } else if (mem->is_MergeMem()) { 1153 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw); 1154 } else if (mem->is_Proj()) { 1155 mem = mem->in(0); 1156 } else if (mem->is_MemBar() || mem->is_SafePoint()) { 1157 mem = mem->in(TypeFunc::Memory); 1158 } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) { 1159 mem = mem->in(MemNode::Memory); 1160 } else { 1161 #ifdef ASSERT 1162 mem->dump(); 1163 #endif 1164 ShouldNotReachHere(); 1165 } 1166 if (wq.test_set(mem->_idx)) { 1167 return; 1168 } 1169 } 1170 // Replace memory input and re-compute ctrl to move the check out of the loop 1171 _igvn.replace_input_of(n, 1, mem); 1172 set_ctrl_and_loop(n, get_early_ctrl(n)); 1173 Node* bol = n->unique_out(); 1174 set_ctrl_and_loop(bol, get_early_ctrl(bol)); 1175 } 1176 } 1177 1178 //------------------------------split_if_with_blocks_pre----------------------- 1179 // Do the real work in a non-recursive function. Data nodes want to be 1180 // cloned in the pre-order so they can feed each other nicely. 1181 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) { 1182 // Cloning these guys is unlikely to win 1183 int n_op = n->Opcode(); 1184 if (n_op == Op_MergeMem) { 1185 return n; 1186 } 1187 if (n->is_Proj()) { 1188 return n; 1189 } 1190 1191 if (n->isa_FlatArrayCheck()) { 1192 move_flat_array_check_out_of_loop(n); 1193 return n; 1194 } 1195 1196 // Do not clone-up CmpFXXX variations, as these are always 1197 // followed by a CmpI 1198 if (n->is_Cmp()) { 1199 return n; 1200 } 1201 // Attempt to use a conditional move instead of a phi/branch 1202 if (ConditionalMoveLimit > 0 && n_op == Op_Region) { 1203 Node *cmov = conditional_move( n ); 1204 if (cmov) { 1205 return cmov; 1206 } 1207 } 1208 if (n->is_CFG() || n->is_LoadStore()) { 1209 return n; 1210 } 1211 if (n->is_Opaque1()) { // Opaque nodes cannot be mod'd 1212 if (!C->major_progress()) { // If chance of no more loop opts... 1213 _igvn._worklist.push(n); // maybe we'll remove them 1214 } 1215 return n; 1216 } 1217 1218 if (n->is_Con()) { 1219 return n; // No cloning for Con nodes 1220 } 1221 1222 Node *n_ctrl = get_ctrl(n); 1223 if (!n_ctrl) { 1224 return n; // Dead node 1225 } 1226 1227 Node* res = try_move_store_before_loop(n, n_ctrl); 1228 if (res != nullptr) { 1229 return n; 1230 } 1231 1232 // Attempt to remix address expressions for loop invariants 1233 Node *m = remix_address_expressions( n ); 1234 if( m ) return m; 1235 1236 if (n_op == Op_AddI) { 1237 Node *nn = convert_add_to_muladd( n ); 1238 if ( nn ) return nn; 1239 } 1240 1241 if (n->is_ConstraintCast() && n->as_ConstraintCast()->dependency().narrows_type()) { 1242 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this); 1243 // ConstraintCastNode::dominating_cast() uses node control input to determine domination. 1244 // Node control inputs don't necessarily agree with loop control info (due to 1245 // transformations happened in between), thus additional dominance check is needed 1246 // to keep loop info valid. 1247 if (dom_cast != nullptr && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) { 1248 _igvn.replace_node(n, dom_cast); 1249 return dom_cast; 1250 } 1251 } 1252 1253 // Determine if the Node has inputs from some local Phi. 1254 // Returns the block to clone thru. 1255 Node *n_blk = has_local_phi_input( n ); 1256 if( !n_blk ) return n; 1257 1258 // Do not clone the trip counter through on a CountedLoop 1259 // (messes up the canonical shape). 1260 if (((n_blk->is_CountedLoop() || (n_blk->is_Loop() && n_blk->as_Loop()->is_loop_nest_inner_loop())) && n->Opcode() == Op_AddI) || 1261 (n_blk->is_LongCountedLoop() && n->Opcode() == Op_AddL)) { 1262 return n; 1263 } 1264 // Pushing a shift through the iv Phi can get in the way of addressing optimizations or range check elimination 1265 if (n_blk->is_BaseCountedLoop() && n->Opcode() == Op_LShift(n_blk->as_BaseCountedLoop()->bt()) && 1266 n->in(1) == n_blk->as_BaseCountedLoop()->phi()) { 1267 return n; 1268 } 1269 1270 // Check for having no control input; not pinned. Allow 1271 // dominating control. 1272 if (n->in(0)) { 1273 Node *dom = idom(n_blk); 1274 if (dom_lca(n->in(0), dom) != n->in(0)) { 1275 return n; 1276 } 1277 } 1278 // Policy: when is it profitable. You must get more wins than 1279 // policy before it is considered profitable. Policy is usually 0, 1280 // so 1 win is considered profitable. Big merges will require big 1281 // cloning, so get a larger policy. 1282 int policy = n_blk->req() >> 2; 1283 1284 // If the loop is a candidate for range check elimination, 1285 // delay splitting through it's phi until a later loop optimization 1286 if (n_blk->is_BaseCountedLoop()) { 1287 IdealLoopTree *lp = get_loop(n_blk); 1288 if (lp && lp->_rce_candidate) { 1289 return n; 1290 } 1291 } 1292 1293 if (must_throttle_split_if()) return n; 1294 1295 // Split 'n' through the merge point if it is profitable, replacing it with a new phi. 1296 Node* phi = split_thru_phi(n, n_blk, policy); 1297 if (phi == nullptr) { return n; } 1298 1299 // Moved a load around the loop, 'en-registering' something. 1300 if (n_blk->is_Loop() && n->is_Load() && 1301 !phi->in(LoopNode::LoopBackControl)->is_Load()) 1302 C->set_major_progress(); 1303 1304 return phi; 1305 } 1306 1307 static bool merge_point_too_heavy(Compile* C, Node* region) { 1308 // Bail out if the region and its phis have too many users. 1309 int weight = 0; 1310 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1311 weight += region->fast_out(i)->outcnt(); 1312 } 1313 int nodes_left = C->max_node_limit() - C->live_nodes(); 1314 if (weight * 8 > nodes_left) { 1315 if (PrintOpto) { 1316 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight); 1317 } 1318 return true; 1319 } else { 1320 return false; 1321 } 1322 } 1323 1324 static bool merge_point_safe(Node* region) { 1325 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode 1326 // having a PhiNode input. This sidesteps the dangerous case where the split 1327 // ConvI2LNode may become TOP if the input Value() does not 1328 // overlap the ConvI2L range, leaving a node which may not dominate its 1329 // uses. 1330 // A better fix for this problem can be found in the BugTraq entry, but 1331 // expediency for Mantis demands this hack. 1332 #ifdef _LP64 1333 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1334 Node* n = region->fast_out(i); 1335 if (n->is_Phi()) { 1336 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1337 Node* m = n->fast_out(j); 1338 if (m->Opcode() == Op_ConvI2L) 1339 return false; 1340 if (m->is_CastII()) { 1341 return false; 1342 } 1343 } 1344 } 1345 } 1346 #endif 1347 return true; 1348 } 1349 1350 1351 //------------------------------place_outside_loop--------------------------------- 1352 // Place some computation outside of this loop on the path to the use passed as argument 1353 Node* PhaseIdealLoop::place_outside_loop(Node* useblock, IdealLoopTree* loop) const { 1354 Node* head = loop->_head; 1355 assert(!loop->is_member(get_loop(useblock)), "must be outside loop"); 1356 if (head->is_Loop() && head->as_Loop()->is_strip_mined()) { 1357 loop = loop->_parent; 1358 assert(loop->_head->is_OuterStripMinedLoop(), "malformed strip mined loop"); 1359 } 1360 1361 // Pick control right outside the loop 1362 for (;;) { 1363 Node* dom = idom(useblock); 1364 if (loop->is_member(get_loop(dom))) { 1365 break; 1366 } 1367 useblock = dom; 1368 } 1369 assert(find_non_split_ctrl(useblock) == useblock, "should be non split control"); 1370 return useblock; 1371 } 1372 1373 1374 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) { 1375 if (!n->is_If() || n->is_BaseCountedLoopEnd()) { 1376 return false; 1377 } 1378 if (!n->in(0)->is_Region()) { 1379 return false; 1380 } 1381 1382 Node* region = n->in(0); 1383 Node* dom = idom(region); 1384 if (!dom->is_If() || !n->as_If()->same_condition(dom, &_igvn)) { 1385 return false; 1386 } 1387 IfNode* dom_if = dom->as_If(); 1388 IfTrueNode* proj_true = dom_if->true_proj(); 1389 IfFalseNode* proj_false = dom_if->false_proj(); 1390 1391 for (uint i = 1; i < region->req(); i++) { 1392 if (is_dominator(proj_true, region->in(i))) { 1393 continue; 1394 } 1395 if (is_dominator(proj_false, region->in(i))) { 1396 continue; 1397 } 1398 return false; 1399 } 1400 1401 return true; 1402 } 1403 1404 1405 bool PhaseIdealLoop::can_split_if(Node* n_ctrl) { 1406 if (must_throttle_split_if()) { 1407 return false; 1408 } 1409 1410 // Do not do 'split-if' if irreducible loops are present. 1411 if (_has_irreducible_loops) { 1412 return false; 1413 } 1414 1415 if (merge_point_too_heavy(C, n_ctrl)) { 1416 return false; 1417 } 1418 1419 // Do not do 'split-if' if some paths are dead. First do dead code 1420 // elimination and then see if its still profitable. 1421 for (uint i = 1; i < n_ctrl->req(); i++) { 1422 if (n_ctrl->in(i) == C->top()) { 1423 return false; 1424 } 1425 } 1426 1427 // If trying to do a 'Split-If' at the loop head, it is only 1428 // profitable if the cmp folds up on BOTH paths. Otherwise we 1429 // risk peeling a loop forever. 1430 1431 // CNC - Disabled for now. Requires careful handling of loop 1432 // body selection for the cloned code. Also, make sure we check 1433 // for any input path not being in the same loop as n_ctrl. For 1434 // irreducible loops we cannot check for 'n_ctrl->is_Loop()' 1435 // because the alternative loop entry points won't be converted 1436 // into LoopNodes. 1437 IdealLoopTree *n_loop = get_loop(n_ctrl); 1438 for (uint j = 1; j < n_ctrl->req(); j++) { 1439 if (get_loop(n_ctrl->in(j)) != n_loop) { 1440 return false; 1441 } 1442 } 1443 1444 // Check for safety of the merge point. 1445 if (!merge_point_safe(n_ctrl)) { 1446 return false; 1447 } 1448 1449 return true; 1450 } 1451 1452 // Detect if the node is the inner strip-mined loop 1453 // Return: null if it's not the case, or the exit of outer strip-mined loop 1454 static Node* is_inner_of_stripmined_loop(const Node* out) { 1455 Node* out_le = nullptr; 1456 1457 if (out->is_CountedLoopEnd()) { 1458 const CountedLoopNode* loop = out->as_CountedLoopEnd()->loopnode(); 1459 1460 if (loop != nullptr && loop->is_strip_mined()) { 1461 out_le = loop->in(LoopNode::EntryControl)->as_OuterStripMinedLoop()->outer_loop_exit(); 1462 } 1463 } 1464 1465 return out_le; 1466 } 1467 1468 bool PhaseIdealLoop::flat_array_element_type_check(Node *n) { 1469 // If the CmpP is a subtype check for a value that has just been 1470 // loaded from an array, the subtype check guarantees the value 1471 // can't be stored in a flat array and the load of the value 1472 // happens with a flat array check then: push the type check 1473 // through the phi of the flat array check. This needs special 1474 // logic because the subtype check's input is not a phi but a 1475 // LoadKlass that must first be cloned through the phi. 1476 if (n->Opcode() != Op_CmpP) { 1477 return false; 1478 } 1479 1480 Node* klassptr = n->in(1); 1481 Node* klasscon = n->in(2); 1482 1483 if (klassptr->is_DecodeNarrowPtr()) { 1484 klassptr = klassptr->in(1); 1485 } 1486 1487 if (klassptr->Opcode() != Op_LoadKlass && klassptr->Opcode() != Op_LoadNKlass) { 1488 return false; 1489 } 1490 1491 if (!klasscon->is_Con()) { 1492 return false; 1493 } 1494 1495 Node* addr = klassptr->in(MemNode::Address); 1496 1497 if (!addr->is_AddP()) { 1498 return false; 1499 } 1500 1501 intptr_t offset; 1502 Node* obj = AddPNode::Ideal_base_and_offset(addr, &_igvn, offset); 1503 1504 if (obj == nullptr) { 1505 return false; 1506 } 1507 1508 // TODO 8378077: The code below does not work anymore with off-heap accesses which set their bases to top with 1509 // JDK-8373343. Also: flat_array_element_type_check() was introduced with JDK-8228622 for a specific check to enable 1510 // split-if but JDK-8245729 changed how that check looks like. Is it still relevant? This should be revisited. 1511 if (addr->in(AddPNode::Base)->is_top()) { 1512 return false; 1513 } 1514 1515 if (obj->Opcode() == Op_CastPP) { 1516 obj = obj->in(1); 1517 } 1518 1519 if (!obj->is_Phi()) { 1520 return false; 1521 } 1522 1523 Node* region = obj->in(0); 1524 1525 Node* phi = PhiNode::make_blank(region, n->in(1)); 1526 for (uint i = 1; i < region->req(); i++) { 1527 Node* in = obj->in(i); 1528 Node* ctrl = region->in(i); 1529 if (addr->in(AddPNode::Base) != obj) { 1530 Node* cast = addr->in(AddPNode::Base); 1531 assert(cast->Opcode() == Op_CastPP && cast->in(0) != nullptr, "inconsistent subgraph"); 1532 Node* cast_clone = cast->clone(); 1533 cast_clone->set_req(0, ctrl); 1534 cast_clone->set_req(1, in); 1535 register_new_node(cast_clone, ctrl); 1536 const Type* tcast = cast_clone->Value(&_igvn); 1537 _igvn.set_type(cast_clone, tcast); 1538 cast_clone->as_Type()->set_type(tcast); 1539 in = cast_clone; 1540 } 1541 Node* addr_clone = addr->clone(); 1542 addr_clone->set_req(AddPNode::Base, in); 1543 addr_clone->set_req(AddPNode::Address, in); 1544 register_new_node(addr_clone, ctrl); 1545 _igvn.set_type(addr_clone, addr_clone->Value(&_igvn)); 1546 Node* klassptr_clone = klassptr->clone(); 1547 klassptr_clone->set_req(2, addr_clone); 1548 register_new_node(klassptr_clone, ctrl); 1549 _igvn.set_type(klassptr_clone, klassptr_clone->Value(&_igvn)); 1550 if (klassptr != n->in(1)) { 1551 Node* decode = n->in(1); 1552 assert(decode->is_DecodeNarrowPtr(), "inconsistent subgraph"); 1553 Node* decode_clone = decode->clone(); 1554 decode_clone->set_req(1, klassptr_clone); 1555 register_new_node(decode_clone, ctrl); 1556 _igvn.set_type(decode_clone, decode_clone->Value(&_igvn)); 1557 klassptr_clone = decode_clone; 1558 } 1559 phi->set_req(i, klassptr_clone); 1560 } 1561 register_new_node(phi, region); 1562 Node* orig = n->in(1); 1563 _igvn.replace_input_of(n, 1, phi); 1564 split_if_with_blocks_post(n); 1565 if (n->outcnt() != 0) { 1566 _igvn.replace_input_of(n, 1, orig); 1567 _igvn.remove_dead_node(phi, PhaseIterGVN::NodeOrigin::Graph); 1568 } 1569 return true; 1570 } 1571 1572 //------------------------------split_if_with_blocks_post---------------------- 1573 // Do the real work in a non-recursive function. CFG hackery wants to be 1574 // in the post-order, so it can dirty the I-DOM info and not use the dirtied 1575 // info. 1576 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) { 1577 1578 if (flat_array_element_type_check(n)) { 1579 return; 1580 } 1581 1582 // Cloning Cmp through Phi's involves the split-if transform. 1583 // FastLock is not used by an If 1584 if (n->is_Cmp() && !n->is_FastLock()) { 1585 Node *n_ctrl = get_ctrl(n); 1586 // Determine if the Node has inputs from some local Phi. 1587 // Returns the block to clone thru. 1588 Node *n_blk = has_local_phi_input(n); 1589 if (n_blk != n_ctrl) { 1590 return; 1591 } 1592 1593 if (!can_split_if(n_ctrl)) { 1594 return; 1595 } 1596 1597 if (n->outcnt() != 1) { 1598 return; // Multiple bool's from 1 compare? 1599 } 1600 Node *bol = n->unique_out(); 1601 assert(bol->is_Bool(), "expect a bool here"); 1602 if (bol->outcnt() != 1) { 1603 return;// Multiple branches from 1 compare? 1604 } 1605 Node *iff = bol->unique_out(); 1606 1607 // Check some safety conditions 1608 if (iff->is_If()) { // Classic split-if? 1609 if (iff->in(0) != n_ctrl) { 1610 return; // Compare must be in same blk as if 1611 } 1612 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE 1613 // Can't split CMove with different control. 1614 if (get_ctrl(iff) != n_ctrl) { 1615 return; 1616 } 1617 if (get_ctrl(iff->in(2)) == n_ctrl || 1618 get_ctrl(iff->in(3)) == n_ctrl) { 1619 return; // Inputs not yet split-up 1620 } 1621 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) { 1622 return; // Loop-invar test gates loop-varying CMOVE 1623 } 1624 } else { 1625 return; // some other kind of node, such as an Allocate 1626 } 1627 1628 // When is split-if profitable? Every 'win' on means some control flow 1629 // goes dead, so it's almost always a win. 1630 int policy = 0; 1631 // Split compare 'n' through the merge point if it is profitable 1632 Node *phi = split_thru_phi( n, n_ctrl, policy); 1633 if (!phi) { 1634 return; 1635 } 1636 1637 // Now split the bool up thru the phi 1638 Node* bolphi = split_thru_phi(bol, n_ctrl, -1); 1639 guarantee(bolphi != nullptr, "null boolean phi node"); 1640 assert(iff->in(1) == bolphi, ""); 1641 1642 if (bolphi->Value(&_igvn)->singleton()) { 1643 return; 1644 } 1645 1646 // Conditional-move? Must split up now 1647 if (!iff->is_If()) { 1648 Node* cmovphi = split_thru_phi(iff, n_ctrl, -1); 1649 return; 1650 } 1651 1652 // Now split the IF 1653 C->print_method(PHASE_BEFORE_SPLIT_IF, 4, iff); 1654 #ifndef PRODUCT 1655 if (TraceLoopOpts || TraceSplitIf) { 1656 tty->print_cr("Split-If: %d %s", iff->_idx, iff->Name()); 1657 } 1658 #endif 1659 do_split_if(iff); 1660 C->print_method(PHASE_AFTER_SPLIT_IF, 4, iff); 1661 return; 1662 } 1663 1664 // Two identical ifs back to back can be merged 1665 if (try_merge_identical_ifs(n)) { 1666 return; 1667 } 1668 1669 // Check for an IF ready to split; one that has its 1670 // condition codes input coming from a Phi at the block start. 1671 int n_op = n->Opcode(); 1672 1673 // Check for an IF being dominated by another IF same test 1674 if (n_op == Op_If || 1675 n_op == Op_RangeCheck) { 1676 Node *bol = n->in(1); 1677 uint max = bol->outcnt(); 1678 // Check for same test used more than once? 1679 if (bol->is_Bool() && (max > 1 || bol->in(1)->is_SubTypeCheck())) { 1680 // Search up IDOMs to see if this IF is dominated. 1681 Node* cmp = bol->in(1); 1682 Node *cutoff = cmp->is_SubTypeCheck() ? dom_lca(get_ctrl(cmp->in(1)), get_ctrl(cmp->in(2))) : get_ctrl(bol); 1683 1684 // Now search up IDOMs till cutoff, looking for a dominating test 1685 Node *prevdom = n; 1686 Node *dom = idom(prevdom); 1687 while (dom != cutoff) { 1688 if (dom->req() > 1 && n->as_If()->same_condition(dom, &_igvn) && prevdom->in(0) == dom && 1689 safe_for_if_replacement(dom)) { 1690 // It's invalid to move control dependent data nodes in the inner 1691 // strip-mined loop, because: 1692 // 1) break validation of LoopNode::verify_strip_mined() 1693 // 2) move code with side-effect in strip-mined loop 1694 // Move to the exit of outer strip-mined loop in that case. 1695 Node* out_le = is_inner_of_stripmined_loop(dom); 1696 if (out_le != nullptr) { 1697 prevdom = out_le; 1698 } 1699 // Replace the dominated test with an obvious true or false. 1700 // Place it on the IGVN worklist for later cleanup. 1701 C->set_major_progress(); 1702 // Split if: pin array accesses that are control dependent on a range check and moved to a regular if, 1703 // to prevent an array load from floating above its range check. There are three cases: 1704 // 1. Move from RangeCheck "a" to RangeCheck "b": don't need to pin. If we ever remove b, then we pin 1705 // all its array accesses at that point. 1706 // 2. We move from RangeCheck "a" to regular if "b": need to pin. If we ever remove b, then its array 1707 // accesses would start to float, since we don't pin at that point. 1708 // 3. If we move from regular if: don't pin. All array accesses are already assumed to be pinned. 1709 bool pin_array_access_nodes = n->Opcode() == Op_RangeCheck && 1710 prevdom->in(0)->Opcode() != Op_RangeCheck; 1711 dominated_by(prevdom->as_IfProj(), n->as_If(), false, pin_array_access_nodes); 1712 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } ); 1713 return; 1714 } 1715 prevdom = dom; 1716 dom = idom(prevdom); 1717 } 1718 } 1719 } 1720 1721 try_sink_out_of_loop(n); 1722 if (C->failing()) { 1723 return; 1724 } 1725 1726 try_move_store_after_loop(n); 1727 1728 // Remove multiple allocations of the same inline type 1729 if (n->is_InlineType()) { 1730 n->as_InlineType()->remove_redundant_allocations(this); 1731 } 1732 } 1733 1734 // Transform: 1735 // 1736 // if (some_condition) { 1737 // // body 1 1738 // } else { 1739 // // body 2 1740 // } 1741 // if (some_condition) { 1742 // // body 3 1743 // } else { 1744 // // body 4 1745 // } 1746 // 1747 // into: 1748 // 1749 // 1750 // if (some_condition) { 1751 // // body 1 1752 // // body 3 1753 // } else { 1754 // // body 2 1755 // // body 4 1756 // } 1757 bool PhaseIdealLoop::try_merge_identical_ifs(Node* n) { 1758 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) { 1759 Node *n_ctrl = n->in(0); 1760 IfNode* dom_if = idom(n_ctrl)->as_If(); 1761 if (n->in(1) != dom_if->in(1)) { 1762 assert(n->in(1)->in(1)->is_SubTypeCheck() && 1763 (n->in(1)->in(1)->as_SubTypeCheck()->method() != nullptr || 1764 dom_if->in(1)->in(1)->as_SubTypeCheck()->method() != nullptr), "only for subtype checks with profile data attached"); 1765 _igvn.replace_input_of(n, 1, dom_if->in(1)); 1766 } 1767 IfTrueNode* dom_proj_true = dom_if->true_proj(); 1768 IfFalseNode* dom_proj_false = dom_if->false_proj(); 1769 1770 // Now split the IF 1771 RegionNode* new_false_region; 1772 RegionNode* new_true_region; 1773 #ifndef PRODUCT 1774 if (TraceLoopOpts || TraceSplitIf) { 1775 tty->print_cr("Split-If Merging Identical Ifs: Dom-If: %d %s, If: %d %s", dom_if->_idx, dom_if->Name(), n->_idx, n->Name()); 1776 } 1777 #endif 1778 do_split_if(n, &new_false_region, &new_true_region); 1779 assert(new_false_region->req() == new_true_region->req(), ""); 1780 #ifdef ASSERT 1781 for (uint i = 1; i < new_false_region->req(); ++i) { 1782 assert(new_false_region->in(i)->in(0) == new_true_region->in(i)->in(0), "unexpected shape following split if"); 1783 assert(i == new_false_region->req() - 1 || new_false_region->in(i)->in(0)->in(1) == new_false_region->in(i + 1)->in(0)->in(1), "unexpected shape following split if"); 1784 } 1785 #endif 1786 assert(new_false_region->in(1)->in(0)->in(1) == dom_if->in(1), "dominating if and dominated if after split must share test"); 1787 1788 // We now have: 1789 // if (some_condition) { 1790 // // body 1 1791 // if (some_condition) { 1792 // body3: // new_true_region 1793 // // body3 1794 // } else { 1795 // goto body4; 1796 // } 1797 // } else { 1798 // // body 2 1799 // if (some_condition) { 1800 // goto body3; 1801 // } else { 1802 // body4: // new_false_region 1803 // // body4; 1804 // } 1805 // } 1806 // 1807 1808 // clone pinned nodes thru the resulting regions 1809 push_pinned_nodes_thru_region(dom_if, new_true_region); 1810 push_pinned_nodes_thru_region(dom_if, new_false_region); 1811 1812 // Optimize out the cloned ifs. Because pinned nodes were cloned, this also allows a CastPP that would be dependent 1813 // on a projection of n to have the dom_if as a control dependency. We don't want the CastPP to end up with an 1814 // unrelated control dependency. 1815 for (uint i = 1; i < new_false_region->req(); i++) { 1816 if (is_dominator(dom_proj_true, new_false_region->in(i))) { 1817 dominated_by(dom_proj_true, new_false_region->in(i)->in(0)->as_If()); 1818 } else { 1819 assert(is_dominator(dom_proj_false, new_false_region->in(i)), "bad if"); 1820 dominated_by(dom_proj_false, new_false_region->in(i)->in(0)->as_If()); 1821 } 1822 } 1823 return true; 1824 } 1825 return false; 1826 } 1827 1828 void PhaseIdealLoop::push_pinned_nodes_thru_region(IfNode* dom_if, Node* region) { 1829 for (DUIterator i = region->outs(); region->has_out(i); i++) { 1830 Node* u = region->out(i); 1831 if (!has_ctrl(u) || u->is_Phi() || !u->depends_only_on_test()) { 1832 continue; 1833 } 1834 assert(u->in(0) == region, "not a control dependent node?"); 1835 uint j = 1; 1836 for (; j < u->req(); ++j) { 1837 Node* in = u->in(j); 1838 if (!is_dominator(ctrl_or_self(in), dom_if)) { 1839 break; 1840 } 1841 } 1842 if (j == u->req()) { 1843 Node *phi = PhiNode::make_blank(region, u); 1844 for (uint k = 1; k < region->req(); ++k) { 1845 Node* clone = u->clone(); 1846 clone->set_req(0, region->in(k)); 1847 register_new_node(clone, region->in(k)); 1848 phi->init_req(k, clone); 1849 } 1850 register_new_node(phi, region); 1851 _igvn.replace_node(u, phi); 1852 --i; 1853 } 1854 } 1855 } 1856 1857 bool PhaseIdealLoop::safe_for_if_replacement(const Node* dom) const { 1858 if (!dom->is_CountedLoopEnd()) { 1859 return true; 1860 } 1861 CountedLoopEndNode* le = dom->as_CountedLoopEnd(); 1862 CountedLoopNode* cl = le->loopnode(); 1863 if (cl == nullptr) { 1864 return true; 1865 } 1866 if (!cl->is_main_loop()) { 1867 return true; 1868 } 1869 if (cl->is_canonical_loop_entry() == nullptr) { 1870 return true; 1871 } 1872 // Further unrolling is possible so loop exit condition might change 1873 return false; 1874 } 1875 1876 // See if a shared loop-varying computation has no loop-varying uses. 1877 // Happens if something is only used for JVM state in uncommon trap exits, 1878 // like various versions of induction variable+offset. Clone the 1879 // computation per usage to allow it to sink out of the loop. 1880 void PhaseIdealLoop::try_sink_out_of_loop(Node* n) { 1881 bool is_raw_to_oop_cast = n->is_ConstraintCast() && 1882 n->in(1)->bottom_type()->isa_rawptr() && 1883 !n->bottom_type()->isa_rawptr(); 1884 1885 if (has_ctrl(n) && 1886 !n->is_Phi() && 1887 !n->is_Bool() && 1888 !n->is_Proj() && 1889 !n->is_MergeMem() && 1890 !n->is_CMove() && 1891 !n->is_OpaqueConstantBool() && 1892 !n->is_OpaqueInitializedAssertionPredicate() && 1893 !n->is_OpaqueTemplateAssertionPredicate() && 1894 !is_raw_to_oop_cast && // don't extend live ranges of raw oops 1895 n->Opcode() != Op_CreateEx && 1896 (KillPathsReachableByDeadTypeNode || !n->is_Type()) 1897 ) { 1898 Node *n_ctrl = get_ctrl(n); 1899 IdealLoopTree *n_loop = get_loop(n_ctrl); 1900 1901 if (n->in(0) != nullptr) { 1902 IdealLoopTree* loop_ctrl = get_loop(n->in(0)); 1903 if (n_loop != loop_ctrl && n_loop->is_member(loop_ctrl)) { 1904 // n has a control input inside a loop but get_ctrl() is member of an outer loop. This could happen, for example, 1905 // for Div nodes inside a loop (control input inside loop) without a use except for an UCT (outside the loop). 1906 // Rewire control of n to right outside of the loop, regardless if its input(s) are later sunk or not. 1907 Node* maybe_pinned_n = n; 1908 Node* outside_ctrl = place_outside_loop(n_ctrl, loop_ctrl); 1909 if (!would_sink_below_pre_loop_exit(loop_ctrl, outside_ctrl)) { 1910 if (n->depends_only_on_test()) { 1911 // If this node depends_only_on_test, it will be rewired to a control input that is not 1912 // the correct test. As a result, it must be pinned otherwise it can be incorrectly 1913 // rewired to a dominating test equivalent to the new control. 1914 Node* pinned_clone = n->pin_node_under_control(); 1915 if (pinned_clone != nullptr) { 1916 register_new_node(pinned_clone, n_ctrl); 1917 maybe_pinned_n = pinned_clone; 1918 _igvn.replace_node(n, pinned_clone); 1919 } 1920 } 1921 _igvn.replace_input_of(maybe_pinned_n, 0, outside_ctrl); 1922 } 1923 } 1924 } 1925 if (n_loop != _ltree_root && n->outcnt() > 1) { 1926 // Compute early control: needed for anti-dependence analysis. It's also possible that as a result of 1927 // previous transformations in this loop opts round, the node can be hoisted now: early control will tell us. 1928 Node* early_ctrl = compute_early_ctrl(n, n_ctrl); 1929 if (n_loop->is_member(get_loop(early_ctrl)) && // check that this one can't be hoisted now 1930 ctrl_of_all_uses_out_of_loop(n, early_ctrl, n_loop)) { // All uses in outer loops! 1931 if (n->is_Store() || n->is_LoadStore()) { 1932 assert(false, "no node with a side effect"); 1933 C->record_failure("no node with a side effect"); 1934 return; 1935 } 1936 Node* outer_loop_clone = nullptr; 1937 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin;) { 1938 Node* u = n->last_out(j); // Clone private computation per use 1939 _igvn.rehash_node_delayed(u); 1940 Node* x = nullptr; 1941 if (n->in(0) != nullptr && n->depends_only_on_test()) { 1942 // If this node depends_only_on_test, it will be rewired to a control input that is not 1943 // the correct test. As a result, it must be pinned otherwise it can be incorrectly 1944 // rewired to a dominating test equivalent to the new control. 1945 x = n->pin_node_under_control(); 1946 } 1947 if (x == nullptr) { 1948 x = n->clone(); 1949 } 1950 Node* x_ctrl = nullptr; 1951 if (u->is_Phi()) { 1952 // Replace all uses of normal nodes. Replace Phi uses 1953 // individually, so the separate Nodes can sink down 1954 // different paths. 1955 uint k = 1; 1956 while (u->in(k) != n) k++; 1957 u->set_req(k, x); 1958 // x goes next to Phi input path 1959 x_ctrl = u->in(0)->in(k); 1960 // Find control for 'x' next to use but not inside inner loops. 1961 x_ctrl = place_outside_loop(x_ctrl, n_loop); 1962 --j; 1963 } else { // Normal use 1964 if (has_ctrl(u)) { 1965 x_ctrl = get_ctrl(u); 1966 } else { 1967 x_ctrl = u->in(0); 1968 } 1969 // Find control for 'x' next to use but not inside inner loops. 1970 x_ctrl = place_outside_loop(x_ctrl, n_loop); 1971 // Replace all uses 1972 if (u->is_ConstraintCast() && _igvn.type(n)->higher_equal(u->bottom_type()) && u->in(0) == x_ctrl) { 1973 // If we're sinking a chain of data nodes, we might have inserted a cast to pin the use which is not necessary 1974 // anymore now that we're going to pin n as well 1975 _igvn.replace_node(u, x); 1976 --j; 1977 } else { 1978 int nb = u->replace_edge(n, x, &_igvn); 1979 j -= nb; 1980 } 1981 } 1982 1983 if (n->is_Load()) { 1984 // For loads, add a control edge to a CFG node outside of the loop 1985 // to force them to not combine and return back inside the loop 1986 // during GVN optimization (4641526). 1987 assert(x_ctrl == get_late_ctrl_with_anti_dep(x->as_Load(), early_ctrl, x_ctrl), "anti-dependences were already checked"); 1988 1989 IdealLoopTree* x_loop = get_loop(x_ctrl); 1990 Node* x_head = x_loop->_head; 1991 if (x_head->is_Loop() && x_head->is_OuterStripMinedLoop()) { 1992 // Do not add duplicate LoadNodes to the outer strip mined loop 1993 if (outer_loop_clone != nullptr) { 1994 _igvn.replace_node(x, outer_loop_clone); 1995 continue; 1996 } 1997 outer_loop_clone = x; 1998 } 1999 x->set_req(0, x_ctrl); 2000 } else if (n->in(0) != nullptr){ 2001 x->set_req(0, x_ctrl); 2002 } 2003 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone"); 2004 assert(!n_loop->is_member(get_loop(x_ctrl)), "should have moved out of loop"); 2005 register_new_node(x, x_ctrl); 2006 2007 // Chain of AddP nodes: (AddP base (AddP base (AddP base ))) 2008 // All AddP nodes must keep the same base after sinking so: 2009 // 1- We don't add a CastPP here until the last one of the chain is sunk: if part of the chain is not sunk, 2010 // their bases remain the same. 2011 // (see 2- below) 2012 assert(!x->is_AddP() || !x->in(AddPNode::Address)->is_AddP() || 2013 x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base) || 2014 !x->in(AddPNode::Address)->in(AddPNode::Base)->eqv_uncast(x->in(AddPNode::Base)), "unexpected AddP shape"); 2015 if (x->in(0) == nullptr && !x->is_DecodeNarrowPtr() && 2016 !(x->is_AddP() && x->in(AddPNode::Address)->is_AddP() && x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base))) { 2017 assert(!x->is_Load(), "load should be pinned"); 2018 // Use a cast node to pin clone out of loop 2019 Node* cast = nullptr; 2020 for (uint k = 0; k < x->req(); k++) { 2021 Node* in = x->in(k); 2022 if (in != nullptr && ctrl_is_member(n_loop, in)) { 2023 const Type* in_t = _igvn.type(in); 2024 cast = ConstraintCastNode::make_cast_for_type(x_ctrl, in, in_t, 2025 ConstraintCastNode::DependencyType::NonFloatingNonNarrowing, nullptr); 2026 } 2027 if (cast != nullptr) { 2028 Node* prev = _igvn.hash_find_insert(cast); 2029 if (prev != nullptr && get_ctrl(prev) == x_ctrl) { 2030 cast->destruct(&_igvn); 2031 cast = prev; 2032 } else { 2033 register_new_node(cast, x_ctrl); 2034 } 2035 x->replace_edge(in, cast); 2036 // Chain of AddP nodes: 2037 // 2- A CastPP of the base is only added now that all AddP nodes are sunk 2038 if (x->is_AddP() && k == AddPNode::Base) { 2039 update_addp_chain_base(x, n->in(AddPNode::Base), cast); 2040 } 2041 break; 2042 } 2043 } 2044 assert(cast != nullptr, "must have added a cast to pin the node"); 2045 } 2046 } 2047 _igvn.remove_dead_node(n, PhaseIterGVN::NodeOrigin::Graph); 2048 } 2049 _dom_lca_tags_round = 0; 2050 } 2051 } 2052 } 2053 2054 void PhaseIdealLoop::update_addp_chain_base(Node* x, Node* old_base, Node* new_base) { 2055 ResourceMark rm; 2056 Node_List wq; 2057 wq.push(x); 2058 while (wq.size() != 0) { 2059 Node* n = wq.pop(); 2060 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2061 Node* u = n->fast_out(i); 2062 if (u->is_AddP() && u->in(AddPNode::Base) == old_base) { 2063 _igvn.replace_input_of(u, AddPNode::Base, new_base); 2064 wq.push(u); 2065 } 2066 } 2067 } 2068 } 2069 2070 // Compute the early control of a node by following its inputs until we reach 2071 // nodes that are pinned. Then compute the LCA of the control of all pinned nodes. 2072 Node* PhaseIdealLoop::compute_early_ctrl(Node* n, Node* n_ctrl) { 2073 Node* early_ctrl = nullptr; 2074 ResourceMark rm; 2075 Unique_Node_List wq; 2076 wq.push(n); 2077 for (uint i = 0; i < wq.size(); i++) { 2078 Node* m = wq.at(i); 2079 Node* c = nullptr; 2080 if (m->is_CFG()) { 2081 c = m; 2082 } else if (m->pinned()) { 2083 c = m->in(0); 2084 } else { 2085 for (uint j = 0; j < m->req(); j++) { 2086 Node* in = m->in(j); 2087 if (in != nullptr) { 2088 wq.push(in); 2089 } 2090 } 2091 } 2092 if (c != nullptr) { 2093 assert(is_dominator(c, n_ctrl), "control input must dominate current control"); 2094 if (early_ctrl == nullptr || is_dominator(early_ctrl, c)) { 2095 early_ctrl = c; 2096 } 2097 } 2098 } 2099 assert(is_dominator(early_ctrl, n_ctrl), "early control must dominate current control"); 2100 return early_ctrl; 2101 } 2102 2103 bool PhaseIdealLoop::ctrl_of_all_uses_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop) { 2104 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2105 Node* u = n->fast_out(i); 2106 if (u->is_Opaque1()) { 2107 return false; // Found loop limit, bugfix for 4677003 2108 } 2109 if (u->is_Phi()) { 2110 for (uint j = 1; j < u->req(); ++j) { 2111 if (u->in(j) == n && !ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, u->in(0)->in(j))) { 2112 return false; 2113 } 2114 } 2115 } else { 2116 Node* ctrl = has_ctrl(u) ? get_ctrl(u) : u->in(0); 2117 if (!ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, ctrl)) { 2118 return false; 2119 } 2120 } 2121 } 2122 return true; 2123 } 2124 2125 // Sinking a node from a pre loop to its main loop pins the node between the pre and main loops. If that node is input 2126 // to a check that's eliminated by range check elimination, it becomes input to an expression that feeds into the exit 2127 // test of the pre loop above the point in the graph where it's pinned. This results in a broken graph. One way to avoid 2128 // it would be to not eliminate the check in the main loop. Instead, we prevent sinking of the node here so better code 2129 // is generated for the main loop. 2130 bool PhaseIdealLoop::would_sink_below_pre_loop_exit(IdealLoopTree* n_loop, Node* ctrl) { 2131 if (n_loop->_head->is_CountedLoop() && n_loop->_head->as_CountedLoop()->is_pre_loop()) { 2132 CountedLoopNode* pre_loop = n_loop->_head->as_CountedLoop(); 2133 if (is_dominator(pre_loop->loopexit(), ctrl)) { 2134 return true; 2135 } 2136 } 2137 return false; 2138 } 2139 2140 bool PhaseIdealLoop::ctrl_of_use_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop, Node* ctrl) { 2141 if (n->is_Load()) { 2142 // We can't reuse tags in PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal() so make sure each call to 2143 // get_late_ctrl_with_anti_dep() uses its own tag 2144 _dom_lca_tags_round++; 2145 assert(_dom_lca_tags_round != 0, "shouldn't wrap around"); 2146 2147 ctrl = get_late_ctrl_with_anti_dep(n->as_Load(), n_ctrl, ctrl); 2148 } 2149 IdealLoopTree *u_loop = get_loop(ctrl); 2150 if (u_loop == n_loop) { 2151 return false; // Found loop-varying use 2152 } 2153 if (n_loop->is_member(u_loop)) { 2154 return false; // Found use in inner loop 2155 } 2156 if (would_sink_below_pre_loop_exit(n_loop, ctrl)) { 2157 return false; 2158 } 2159 return true; 2160 } 2161 2162 //------------------------------split_if_with_blocks--------------------------- 2163 // Check for aggressive application of 'split-if' optimization, 2164 // using basic block level info. 2165 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack) { 2166 Node* root = C->root(); 2167 visited.set(root->_idx); // first, mark root as visited 2168 // Do pre-visit work for root 2169 Node* n = split_if_with_blocks_pre(root); 2170 uint cnt = n->outcnt(); 2171 uint i = 0; 2172 2173 while (true) { 2174 // Visit all children 2175 if (i < cnt) { 2176 Node* use = n->raw_out(i); 2177 ++i; 2178 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) { 2179 // Now do pre-visit work for this use 2180 use = split_if_with_blocks_pre(use); 2181 nstack.push(n, i); // Save parent and next use's index. 2182 n = use; // Process all children of current use. 2183 cnt = use->outcnt(); 2184 i = 0; 2185 } 2186 } 2187 else { 2188 // All of n's children have been processed, complete post-processing. 2189 if (cnt != 0 && !n->is_Con()) { 2190 assert(has_node(n), "no dead nodes"); 2191 split_if_with_blocks_post(n); 2192 if (C->failing()) { 2193 return; 2194 } 2195 } 2196 if (must_throttle_split_if()) { 2197 nstack.clear(); 2198 } 2199 if (nstack.is_empty()) { 2200 // Finished all nodes on stack. 2201 break; 2202 } 2203 // Get saved parent node and next use's index. Visit the rest of uses. 2204 n = nstack.node(); 2205 cnt = n->outcnt(); 2206 i = nstack.index(); 2207 nstack.pop(); 2208 } 2209 } 2210 } 2211 2212 2213 //============================================================================= 2214 // 2215 // C L O N E A L O O P B O D Y 2216 // 2217 2218 //------------------------------clone_iff-------------------------------------- 2219 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 2220 // "Nearly" because all Nodes have been cloned from the original in the loop, 2221 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 2222 // through the Phi recursively, and return a Bool. 2223 Node* PhaseIdealLoop::clone_iff(PhiNode* phi) { 2224 2225 // Convert this Phi into a Phi merging Bools 2226 uint i; 2227 for (i = 1; i < phi->req(); i++) { 2228 Node* b = phi->in(i); 2229 if (b->is_Phi()) { 2230 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi())); 2231 } else { 2232 assert(b->is_Bool() || b->is_OpaqueConstantBool() || b->is_OpaqueInitializedAssertionPredicate(), 2233 "bool, non-null check with OpaqueConstantBool or Initialized Assertion Predicate with its Opaque node"); 2234 } 2235 } 2236 Node* n = phi->in(1); 2237 Node* sample_opaque = nullptr; 2238 Node *sample_bool = nullptr; 2239 if (n->is_OpaqueConstantBool() || n->is_OpaqueInitializedAssertionPredicate()) { 2240 sample_opaque = n; 2241 sample_bool = n->in(1); 2242 assert(sample_bool->is_Bool(), "wrong type"); 2243 } else { 2244 sample_bool = n; 2245 } 2246 Node* sample_cmp = sample_bool->in(1); 2247 const Type* t = Type::TOP; 2248 const TypePtr* at = nullptr; 2249 if (sample_cmp->is_FlatArrayCheck()) { 2250 // Left input of a FlatArrayCheckNode is memory, set the (adr) type of the phi accordingly 2251 assert(sample_cmp->in(1)->bottom_type() == Type::MEMORY, "unexpected input type"); 2252 t = Type::MEMORY; 2253 at = TypeRawPtr::BOTTOM; 2254 } 2255 2256 // Make Phis to merge the Cmp's inputs. 2257 PhiNode *phi1 = new PhiNode(phi->in(0), t, at); 2258 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP); 2259 for (i = 1; i < phi->req(); i++) { 2260 Node *n1 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1); 2261 Node *n2 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2); 2262 phi1->set_req(i, n1); 2263 phi2->set_req(i, n2); 2264 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 2265 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 2266 } 2267 // See if these Phis have been made before. 2268 // Register with optimizer 2269 Node *hit1 = _igvn.hash_find_insert(phi1); 2270 if (hit1) { // Hit, toss just made Phi 2271 _igvn.remove_dead_node(phi1, PhaseIterGVN::NodeOrigin::Speculative); // Remove new phi 2272 assert(hit1->is_Phi(), "" ); 2273 phi1 = (PhiNode*)hit1; // Use existing phi 2274 } else { // Miss 2275 _igvn.register_new_node_with_optimizer(phi1); 2276 } 2277 Node *hit2 = _igvn.hash_find_insert(phi2); 2278 if (hit2) { // Hit, toss just made Phi 2279 _igvn.remove_dead_node(phi2, PhaseIterGVN::NodeOrigin::Speculative); // Remove new phi 2280 assert(hit2->is_Phi(), "" ); 2281 phi2 = (PhiNode*)hit2; // Use existing phi 2282 } else { // Miss 2283 _igvn.register_new_node_with_optimizer(phi2); 2284 } 2285 // Register Phis with loop/block info 2286 set_ctrl(phi1, phi->in(0)); 2287 set_ctrl(phi2, phi->in(0)); 2288 // Make a new Cmp 2289 Node *cmp = sample_cmp->clone(); 2290 cmp->set_req(1, phi1); 2291 cmp->set_req(2, phi2); 2292 _igvn.register_new_node_with_optimizer(cmp); 2293 set_ctrl(cmp, phi->in(0)); 2294 2295 // Make a new Bool 2296 Node *b = sample_bool->clone(); 2297 b->set_req(1,cmp); 2298 _igvn.register_new_node_with_optimizer(b); 2299 set_ctrl(b, phi->in(0)); 2300 2301 if (sample_opaque != nullptr) { 2302 Node* opaque = sample_opaque->clone(); 2303 opaque->set_req(1, b); 2304 _igvn.register_new_node_with_optimizer(opaque); 2305 set_ctrl(opaque, phi->in(0)); 2306 return opaque; 2307 } 2308 2309 assert(b->is_Bool(), ""); 2310 return b; 2311 } 2312 2313 //------------------------------clone_bool------------------------------------- 2314 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 2315 // "Nearly" because all Nodes have been cloned from the original in the loop, 2316 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 2317 // through the Phi recursively, and return a Bool. 2318 CmpNode*PhaseIdealLoop::clone_bool(PhiNode* phi) { 2319 uint i; 2320 // Convert this Phi into a Phi merging Bools 2321 for( i = 1; i < phi->req(); i++ ) { 2322 Node *b = phi->in(i); 2323 if( b->is_Phi() ) { 2324 _igvn.replace_input_of(phi, i, clone_bool(b->as_Phi())); 2325 } else { 2326 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" ); 2327 } 2328 } 2329 2330 Node *sample_cmp = phi->in(1); 2331 2332 // Make Phis to merge the Cmp's inputs. 2333 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP ); 2334 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP ); 2335 for( uint j = 1; j < phi->req(); j++ ) { 2336 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP 2337 Node *n1, *n2; 2338 if( cmp_top->is_Cmp() ) { 2339 n1 = cmp_top->in(1); 2340 n2 = cmp_top->in(2); 2341 } else { 2342 n1 = n2 = cmp_top; 2343 } 2344 phi1->set_req( j, n1 ); 2345 phi2->set_req( j, n2 ); 2346 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 2347 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 2348 } 2349 2350 // See if these Phis have been made before. 2351 // Register with optimizer 2352 Node *hit1 = _igvn.hash_find_insert(phi1); 2353 if( hit1 ) { // Hit, toss just made Phi 2354 _igvn.remove_dead_node(phi1, PhaseIterGVN::NodeOrigin::Speculative); // Remove new phi 2355 assert( hit1->is_Phi(), "" ); 2356 phi1 = (PhiNode*)hit1; // Use existing phi 2357 } else { // Miss 2358 _igvn.register_new_node_with_optimizer(phi1); 2359 } 2360 Node *hit2 = _igvn.hash_find_insert(phi2); 2361 if( hit2 ) { // Hit, toss just made Phi 2362 _igvn.remove_dead_node(phi2, PhaseIterGVN::NodeOrigin::Speculative); // Remove new phi 2363 assert( hit2->is_Phi(), "" ); 2364 phi2 = (PhiNode*)hit2; // Use existing phi 2365 } else { // Miss 2366 _igvn.register_new_node_with_optimizer(phi2); 2367 } 2368 // Register Phis with loop/block info 2369 set_ctrl(phi1, phi->in(0)); 2370 set_ctrl(phi2, phi->in(0)); 2371 // Make a new Cmp 2372 Node *cmp = sample_cmp->clone(); 2373 cmp->set_req( 1, phi1 ); 2374 cmp->set_req( 2, phi2 ); 2375 _igvn.register_new_node_with_optimizer(cmp); 2376 set_ctrl(cmp, phi->in(0)); 2377 2378 assert( cmp->is_Cmp(), "" ); 2379 return (CmpNode*)cmp; 2380 } 2381 2382 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new, 2383 IdealLoopTree* loop, IdealLoopTree* outer_loop, 2384 Node_List*& split_if_set, Node_List*& split_bool_set, 2385 Node_List*& split_cex_set, Node_List& worklist, 2386 uint new_counter, CloneLoopMode mode) { 2387 Node* nnn = old_new[old->_idx]; 2388 // Copy uses to a worklist, so I can munge the def-use info 2389 // with impunity. 2390 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 2391 worklist.push(old->fast_out(j)); 2392 2393 while( worklist.size() ) { 2394 Node *use = worklist.pop(); 2395 if (!has_node(use)) continue; // Ignore dead nodes 2396 if (use->in(0) == C->top()) continue; 2397 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 2398 // Check for data-use outside of loop - at least one of OLD or USE 2399 // must not be a CFG node. 2400 #ifdef ASSERT 2401 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == nullptr) { 2402 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint(); 2403 assert(mode != IgnoreStripMined, "incorrect cloning mode"); 2404 assert((mode == ControlAroundStripMined && use == sfpt) || !use->is_reachable_from_root(), "missed a node"); 2405 } 2406 #endif 2407 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) { 2408 2409 // If the Data use is an IF, that means we have an IF outside the 2410 // loop that is switching on a condition that is set inside the 2411 // loop. Happens if people set a loop-exit flag; then test the flag 2412 // in the loop to break the loop, then test is again outside the 2413 // loop to determine which way the loop exited. 2414 // 2415 // For several uses we need to make sure that there is no phi between, 2416 // the use and the Bool/Cmp. We therefore clone the Bool/Cmp down here 2417 // to avoid such a phi in between. 2418 // For example, it is unexpected that there is a Phi between an 2419 // AllocateArray node and its ValidLengthTest input that could cause 2420 // split if to break. 2421 assert(!use->is_OpaqueTemplateAssertionPredicate(), 2422 "should not clone a Template Assertion Predicate which should be removed once it's useless"); 2423 if (use->is_If() || use->is_CMove() || use->is_OpaqueConstantBool() || use->is_OpaqueInitializedAssertionPredicate() || 2424 (use->Opcode() == Op_AllocateArray && use->in(AllocateNode::ValidLengthTest) == old)) { 2425 // Since this code is highly unlikely, we lazily build the worklist 2426 // of such Nodes to go split. 2427 if (!split_if_set) { 2428 split_if_set = new Node_List(); 2429 } 2430 split_if_set->push(use); 2431 } 2432 if (use->is_Bool()) { 2433 if (!split_bool_set) { 2434 split_bool_set = new Node_List(); 2435 } 2436 split_bool_set->push(use); 2437 } 2438 if (use->Opcode() == Op_CreateEx) { 2439 if (!split_cex_set) { 2440 split_cex_set = new Node_List(); 2441 } 2442 split_cex_set->push(use); 2443 } 2444 2445 2446 // Get "block" use is in 2447 uint idx = 0; 2448 while( use->in(idx) != old ) idx++; 2449 Node *prev = use->is_CFG() ? use : get_ctrl(use); 2450 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" ); 2451 Node* cfg = (prev->_idx >= new_counter && prev->is_Region()) 2452 ? prev->in(2) 2453 : idom(prev); 2454 if( use->is_Phi() ) // Phi use is in prior block 2455 cfg = prev->in(idx); // NOT in block of Phi itself 2456 if (cfg->is_top()) { // Use is dead? 2457 _igvn.replace_input_of(use, idx, C->top()); 2458 continue; 2459 } 2460 2461 // If use is referenced through control edge... (idx == 0) 2462 if (mode == IgnoreStripMined && idx == 0) { 2463 LoopNode *head = loop->_head->as_Loop(); 2464 if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) { 2465 // That node is outside the inner loop, leave it outside the 2466 // outer loop as well to not confuse verification code. 2467 assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop"); 2468 _igvn.replace_input_of(use, 0, head->outer_loop_exit()); 2469 continue; 2470 } 2471 } 2472 2473 while(!outer_loop->is_member(get_loop(cfg))) { 2474 prev = cfg; 2475 cfg = (cfg->_idx >= new_counter && cfg->is_Region()) ? cfg->in(2) : idom(cfg); 2476 } 2477 // If the use occurs after merging several exits from the loop, then 2478 // old value must have dominated all those exits. Since the same old 2479 // value was used on all those exits we did not need a Phi at this 2480 // merge point. NOW we do need a Phi here. Each loop exit value 2481 // is now merged with the peeled body exit; each exit gets its own 2482 // private Phi and those Phis need to be merged here. 2483 Node *phi; 2484 if( prev->is_Region() ) { 2485 if( idx == 0 ) { // Updating control edge? 2486 phi = prev; // Just use existing control 2487 } else { // Else need a new Phi 2488 phi = PhiNode::make( prev, old ); 2489 // Now recursively fix up the new uses of old! 2490 for( uint i = 1; i < prev->req(); i++ ) { 2491 worklist.push(phi); // Onto worklist once for each 'old' input 2492 } 2493 } 2494 } else { 2495 // Get new RegionNode merging old and new loop exits 2496 prev = old_new[prev->_idx]; 2497 assert( prev, "just made this in step 7" ); 2498 if( idx == 0) { // Updating control edge? 2499 phi = prev; // Just use existing control 2500 } else { // Else need a new Phi 2501 // Make a new Phi merging data values properly 2502 phi = PhiNode::make( prev, old ); 2503 phi->set_req( 1, nnn ); 2504 } 2505 } 2506 // If inserting a new Phi, check for prior hits 2507 if( idx != 0 ) { 2508 Node *hit = _igvn.hash_find_insert(phi); 2509 if( hit == nullptr ) { 2510 _igvn.register_new_node_with_optimizer(phi); // Register new phi 2511 } else { // or 2512 // Remove the new phi from the graph and use the hit 2513 _igvn.remove_dead_node(phi, phi == prev ? PhaseIterGVN::NodeOrigin::Graph : PhaseIterGVN::NodeOrigin::Speculative); 2514 phi = hit; // Use existing phi 2515 } 2516 set_ctrl(phi, prev); 2517 } 2518 // Make 'use' use the Phi instead of the old loop body exit value 2519 assert(use->in(idx) == old, "old is still input of use"); 2520 // We notify all uses of old, including use, and the indirect uses, 2521 // that may now be optimized because we have replaced old with phi. 2522 _igvn.add_users_to_worklist(old); 2523 if (idx == 0 && use->depends_only_on_test()) { 2524 // If this node depends_only_on_test, it will be rewired to a control input that is not the 2525 // correct test. As a result, it must be pinned otherwise it can be incorrectly rewired to 2526 // a dominating test equivalent to the new control. 2527 Node* pinned_clone = use->pin_node_under_control(); 2528 if (pinned_clone != nullptr) { 2529 pinned_clone->set_req(0, phi); 2530 register_new_node_with_ctrl_of(pinned_clone, use); 2531 _igvn.replace_node(use, pinned_clone); 2532 continue; 2533 } 2534 } 2535 _igvn.replace_input_of(use, idx, phi); 2536 if( use->_idx >= new_counter ) { // If updating new phis 2537 // Not needed for correctness, but prevents a weak assert 2538 // in AddPNode from tripping (when we end up with different 2539 // base & derived Phis that will become the same after 2540 // IGVN does CSE). 2541 Node *hit = _igvn.hash_find_insert(use); 2542 if( hit ) // Go ahead and re-hash for hits. 2543 _igvn.replace_node( use, hit ); 2544 } 2545 } 2546 } 2547 } 2548 2549 static void collect_nodes_in_outer_loop_not_reachable_from_sfpt(Node* n, const IdealLoopTree *loop, const IdealLoopTree* outer_loop, 2550 const Node_List &old_new, Unique_Node_List& wq, PhaseIdealLoop* phase, 2551 bool check_old_new) { 2552 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2553 Node* u = n->fast_out(j); 2554 assert(check_old_new || old_new[u->_idx] == nullptr, "shouldn't have been cloned"); 2555 if (!u->is_CFG() && (!check_old_new || old_new[u->_idx] == nullptr)) { 2556 assert(!phase->ctrl_is_member(loop, u) || !loop->_body.contains(u), "can be in outer loop or out of both loops only"); 2557 if (!phase->ctrl_is_member(loop, u)) { 2558 if (phase->ctrl_is_member(outer_loop, u)) { 2559 wq.push(u); 2560 } else { 2561 // nodes pinned with control in the outer loop but not referenced from the safepoint must be moved out of 2562 // the outer loop too 2563 Node* u_c = u->in(0); 2564 if (u_c != nullptr) { 2565 IdealLoopTree* u_c_loop = phase->get_loop(u_c); 2566 if (outer_loop->is_member(u_c_loop) && !loop->is_member(u_c_loop)) { 2567 wq.push(u); 2568 } 2569 } 2570 } 2571 } 2572 } 2573 } 2574 } 2575 2576 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop, 2577 IdealLoopTree* outer_loop, int dd, Node_List &old_new, 2578 Node_List& extra_data_nodes) { 2579 if (head->is_strip_mined() && mode != IgnoreStripMined) { 2580 CountedLoopNode* cl = head->as_CountedLoop(); 2581 Node* l = cl->outer_loop(); 2582 Node* tail = cl->outer_loop_tail(); 2583 IfNode* le = cl->outer_loop_end(); 2584 Node* sfpt = cl->outer_safepoint(); 2585 CountedLoopEndNode* cle = cl->loopexit(); 2586 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop(); 2587 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null(); 2588 IfFalseNode* cle_out = cle->false_proj(); 2589 2590 Node* new_sfpt = nullptr; 2591 Node* new_cle_out = cle_out->clone(); 2592 old_new.map(cle_out->_idx, new_cle_out); 2593 if (mode == CloneIncludesStripMined) { 2594 // clone outer loop body 2595 Node* new_l = l->clone(); 2596 Node* new_tail = tail->clone(); 2597 IfNode* new_le = le->clone()->as_If(); 2598 new_sfpt = sfpt->clone(); 2599 2600 set_loop(new_l, outer_loop->_parent); 2601 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd); 2602 set_loop(new_cle_out, outer_loop->_parent); 2603 set_idom(new_cle_out, new_cle, dd); 2604 set_loop(new_sfpt, outer_loop->_parent); 2605 set_idom(new_sfpt, new_cle_out, dd); 2606 set_loop(new_le, outer_loop->_parent); 2607 set_idom(new_le, new_sfpt, dd); 2608 set_loop(new_tail, outer_loop->_parent); 2609 set_idom(new_tail, new_le, dd); 2610 set_idom(new_cl, new_l, dd); 2611 2612 old_new.map(l->_idx, new_l); 2613 old_new.map(tail->_idx, new_tail); 2614 old_new.map(le->_idx, new_le); 2615 old_new.map(sfpt->_idx, new_sfpt); 2616 2617 new_l->set_req(LoopNode::LoopBackControl, new_tail); 2618 new_l->set_req(0, new_l); 2619 new_tail->set_req(0, new_le); 2620 new_le->set_req(0, new_sfpt); 2621 new_sfpt->set_req(0, new_cle_out); 2622 new_cle_out->set_req(0, new_cle); 2623 new_cl->set_req(LoopNode::EntryControl, new_l); 2624 2625 _igvn.register_new_node_with_optimizer(new_l); 2626 _igvn.register_new_node_with_optimizer(new_tail); 2627 _igvn.register_new_node_with_optimizer(new_le); 2628 } else { 2629 Node *newhead = old_new[loop->_head->_idx]; 2630 newhead->as_Loop()->clear_strip_mined(); 2631 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl)); 2632 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 2633 } 2634 // Look at data node that were assigned a control in the outer 2635 // loop: they are kept in the outer loop by the safepoint so start 2636 // from the safepoint node's inputs. 2637 IdealLoopTree* outer_loop = get_loop(l); 2638 Node_Stack stack(2); 2639 stack.push(sfpt, 1); 2640 uint new_counter = C->unique(); 2641 while (stack.size() > 0) { 2642 Node* n = stack.node(); 2643 uint i = stack.index(); 2644 while (i < n->req() && 2645 (n->in(i) == nullptr || 2646 !has_ctrl(n->in(i)) || 2647 get_loop(get_ctrl(n->in(i))) != outer_loop || 2648 (old_new[n->in(i)->_idx] != nullptr && old_new[n->in(i)->_idx]->_idx >= new_counter))) { 2649 i++; 2650 } 2651 if (i < n->req()) { 2652 stack.set_index(i+1); 2653 stack.push(n->in(i), 0); 2654 } else { 2655 assert(old_new[n->_idx] == nullptr || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet"); 2656 Node* m = n == sfpt ? new_sfpt : n->clone(); 2657 if (m != nullptr) { 2658 for (uint i = 0; i < n->req(); i++) { 2659 if (m->in(i) != nullptr && old_new[m->in(i)->_idx] != nullptr) { 2660 m->set_req(i, old_new[m->in(i)->_idx]); 2661 } 2662 } 2663 } else { 2664 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?"); 2665 } 2666 if (n != sfpt) { 2667 extra_data_nodes.push(n); 2668 _igvn.register_new_node_with_optimizer(m); 2669 assert(get_ctrl(n) == cle_out, "what other control?"); 2670 set_ctrl(m, new_cle_out); 2671 old_new.map(n->_idx, m); 2672 } 2673 stack.pop(); 2674 } 2675 } 2676 if (mode == CloneIncludesStripMined) { 2677 _igvn.register_new_node_with_optimizer(new_sfpt); 2678 _igvn.register_new_node_with_optimizer(new_cle_out); 2679 } 2680 // Some other transformation may have pessimistically assigned some 2681 // data nodes to the outer loop. Set their control so they are out 2682 // of the outer loop. 2683 ResourceMark rm; 2684 Unique_Node_List wq; 2685 for (uint i = 0; i < extra_data_nodes.size(); i++) { 2686 Node* old = extra_data_nodes.at(i); 2687 collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true); 2688 } 2689 2690 for (uint i = 0; i < loop->_body.size(); i++) { 2691 Node* old = loop->_body.at(i); 2692 collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true); 2693 } 2694 2695 Node* inner_out = sfpt->in(0); 2696 if (inner_out->outcnt() > 1) { 2697 collect_nodes_in_outer_loop_not_reachable_from_sfpt(inner_out, loop, outer_loop, old_new, wq, this, true); 2698 } 2699 2700 Node* new_ctrl = cl->outer_loop_exit(); 2701 assert(get_loop(new_ctrl) != outer_loop, "must be out of the loop nest"); 2702 for (uint i = 0; i < wq.size(); i++) { 2703 Node* n = wq.at(i); 2704 set_ctrl(n, new_ctrl); 2705 if (n->in(0) != nullptr) { 2706 _igvn.replace_input_of(n, 0, new_ctrl); 2707 } 2708 collect_nodes_in_outer_loop_not_reachable_from_sfpt(n, loop, outer_loop, old_new, wq, this, false); 2709 } 2710 } else { 2711 Node *newhead = old_new[loop->_head->_idx]; 2712 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 2713 } 2714 } 2715 2716 //------------------------------clone_loop------------------------------------- 2717 // 2718 // C L O N E A L O O P B O D Y 2719 // 2720 // This is the basic building block of the loop optimizations. It clones an 2721 // entire loop body. It makes an old_new loop body mapping; with this mapping 2722 // you can find the new-loop equivalent to an old-loop node. All new-loop 2723 // nodes are exactly equal to their old-loop counterparts, all edges are the 2724 // same. All exits from the old-loop now have a RegionNode that merges the 2725 // equivalent new-loop path. This is true even for the normal "loop-exit" 2726 // condition. All uses of loop-invariant old-loop values now come from (one 2727 // or more) Phis that merge their new-loop equivalents. 2728 // 2729 // This operation leaves the graph in an illegal state: there are two valid 2730 // control edges coming from the loop pre-header to both loop bodies. I'll 2731 // definitely have to hack the graph after running this transform. 2732 // 2733 // From this building block I will further edit edges to perform loop peeling 2734 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc. 2735 // 2736 // Parameter side_by_size_idom: 2737 // When side_by_size_idom is null, the dominator tree is constructed for 2738 // the clone loop to dominate the original. Used in construction of 2739 // pre-main-post loop sequence. 2740 // When nonnull, the clone and original are side-by-side, both are 2741 // dominated by the side_by_side_idom node. Used in construction of 2742 // unswitched loops. 2743 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd, 2744 CloneLoopMode mode, Node* side_by_side_idom) { 2745 2746 LoopNode* head = loop->_head->as_Loop(); 2747 head->verify_strip_mined(1); 2748 2749 if (C->do_vector_loop() && PrintOpto) { 2750 const char* mname = C->method()->name()->as_quoted_ascii(); 2751 if (mname != nullptr) { 2752 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname); 2753 } 2754 } 2755 2756 CloneMap& cm = C->clone_map(); 2757 if (C->do_vector_loop()) { 2758 cm.set_clone_idx(cm.max_gen()+1); 2759 #ifndef PRODUCT 2760 if (PrintOpto) { 2761 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx()); 2762 loop->dump_head(); 2763 } 2764 #endif 2765 } 2766 2767 // Step 1: Clone the loop body. Make the old->new mapping. 2768 clone_loop_body(loop->_body, old_new, &cm); 2769 2770 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop; 2771 2772 // Step 2: Fix the edges in the new body. If the old input is outside the 2773 // loop use it. If the old input is INside the loop, use the corresponding 2774 // new node instead. 2775 fix_body_edges(loop->_body, loop, old_new, dd, outer_loop->_parent, false); 2776 2777 Node_List extra_data_nodes; // data nodes in the outer strip mined loop 2778 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes); 2779 2780 // Step 3: Now fix control uses. Loop varying control uses have already 2781 // been fixed up (as part of all input edges in Step 2). Loop invariant 2782 // control uses must be either an IfFalse or an IfTrue. Make a merge 2783 // point to merge the old and new IfFalse/IfTrue nodes; make the use 2784 // refer to this. 2785 Node_List worklist; 2786 uint new_counter = C->unique(); 2787 fix_ctrl_uses(loop->_body, loop, old_new, mode, side_by_side_idom, &cm, worklist); 2788 2789 // Step 4: If loop-invariant use is not control, it must be dominated by a 2790 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region 2791 // there if needed. Make a Phi there merging old and new used values. 2792 Node_List *split_if_set = nullptr; 2793 Node_List *split_bool_set = nullptr; 2794 Node_List *split_cex_set = nullptr; 2795 fix_data_uses(loop->_body, loop, mode, outer_loop, new_counter, old_new, worklist, split_if_set, split_bool_set, split_cex_set); 2796 2797 for (uint i = 0; i < extra_data_nodes.size(); i++) { 2798 Node* old = extra_data_nodes.at(i); 2799 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set, 2800 split_bool_set, split_cex_set, worklist, new_counter, 2801 mode); 2802 } 2803 2804 // Check for IFs that need splitting/cloning. Happens if an IF outside of 2805 // the loop uses a condition set in the loop. The original IF probably 2806 // takes control from one or more OLD Regions (which in turn get from NEW 2807 // Regions). In any case, there will be a set of Phis for each merge point 2808 // from the IF up to where the original BOOL def exists the loop. 2809 finish_clone_loop(split_if_set, split_bool_set, split_cex_set); 2810 2811 } 2812 2813 void PhaseIdealLoop::finish_clone_loop(Node_List* split_if_set, Node_List* split_bool_set, Node_List* split_cex_set) { 2814 if (split_if_set) { 2815 while (split_if_set->size()) { 2816 Node *iff = split_if_set->pop(); 2817 uint input = iff->Opcode() == Op_AllocateArray ? AllocateNode::ValidLengthTest : 1; 2818 if (iff->in(input)->is_Phi()) { 2819 Node *b = clone_iff(iff->in(input)->as_Phi()); 2820 _igvn.replace_input_of(iff, input, b); 2821 } 2822 } 2823 } 2824 if (split_bool_set) { 2825 while (split_bool_set->size()) { 2826 Node *b = split_bool_set->pop(); 2827 Node *phi = b->in(1); 2828 assert(phi->is_Phi(), ""); 2829 CmpNode *cmp = clone_bool((PhiNode*) phi); 2830 _igvn.replace_input_of(b, 1, cmp); 2831 } 2832 } 2833 if (split_cex_set) { 2834 while (split_cex_set->size()) { 2835 Node *b = split_cex_set->pop(); 2836 assert(b->in(0)->is_Region(), ""); 2837 assert(b->in(1)->is_Phi(), ""); 2838 assert(b->in(0)->in(0) == b->in(1)->in(0), ""); 2839 split_up(b, b->in(0), nullptr); 2840 } 2841 } 2842 } 2843 2844 void PhaseIdealLoop::fix_data_uses(Node_List& body, IdealLoopTree* loop, CloneLoopMode mode, IdealLoopTree* outer_loop, 2845 uint new_counter, Node_List &old_new, Node_List &worklist, Node_List*& split_if_set, 2846 Node_List*& split_bool_set, Node_List*& split_cex_set) { 2847 for(uint i = 0; i < body.size(); i++ ) { 2848 Node* old = body.at(i); 2849 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set, 2850 split_bool_set, split_cex_set, worklist, new_counter, 2851 mode); 2852 } 2853 } 2854 2855 void PhaseIdealLoop::fix_ctrl_uses(const Node_List& body, const IdealLoopTree* loop, Node_List &old_new, CloneLoopMode mode, 2856 Node* side_by_side_idom, CloneMap* cm, Node_List &worklist) { 2857 LoopNode* head = loop->_head->as_Loop(); 2858 for(uint i = 0; i < body.size(); i++ ) { 2859 Node* old = body.at(i); 2860 if( !old->is_CFG() ) continue; 2861 2862 // Copy uses to a worklist, so I can munge the def-use info 2863 // with impunity. 2864 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) { 2865 worklist.push(old->fast_out(j)); 2866 } 2867 2868 while (worklist.size()) { // Visit all uses 2869 Node *use = worklist.pop(); 2870 if (!has_node(use)) continue; // Ignore dead nodes 2871 IdealLoopTree *use_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use ); 2872 if (!loop->is_member(use_loop) && use->is_CFG()) { 2873 // Both OLD and USE are CFG nodes here. 2874 assert(use->is_Proj(), "" ); 2875 Node* nnn = old_new[old->_idx]; 2876 2877 Node* newuse = nullptr; 2878 if (head->is_strip_mined() && mode != IgnoreStripMined) { 2879 CountedLoopNode* cl = head->as_CountedLoop(); 2880 CountedLoopEndNode* cle = cl->loopexit(); 2881 // is use the projection that exits the loop from the CountedLoopEndNode? 2882 if (use->in(0) == cle) { 2883 IfFalseNode* cle_out = use->as_IfFalse(); 2884 IfNode* le = cl->outer_loop_end(); 2885 use = le->false_proj(); 2886 use_loop = get_loop(use); 2887 if (mode == CloneIncludesStripMined) { 2888 nnn = old_new[le->_idx]; 2889 } else { 2890 newuse = old_new[cle_out->_idx]; 2891 } 2892 } 2893 } 2894 if (newuse == nullptr) { 2895 newuse = use->clone(); 2896 } 2897 2898 // Clone the loop exit control projection 2899 if (C->do_vector_loop() && cm != nullptr) { 2900 cm->verify_insert_and_clone(use, newuse, cm->clone_idx()); 2901 } 2902 newuse->set_req(0,nnn); 2903 _igvn.register_new_node_with_optimizer(newuse); 2904 set_loop(newuse, use_loop); 2905 set_idom(newuse, nnn, dom_depth(nnn) + 1 ); 2906 2907 // We need a Region to merge the exit from the peeled body and the 2908 // exit from the old loop body. 2909 RegionNode *r = new RegionNode(3); 2910 uint dd_r = MIN2(dom_depth(newuse), dom_depth(use)); 2911 assert(dd_r >= dom_depth(dom_lca(newuse, use)), "" ); 2912 2913 // The original user of 'use' uses 'r' instead. 2914 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) { 2915 Node* useuse = use->last_out(l); 2916 _igvn.rehash_node_delayed(useuse); 2917 uint uses_found = 0; 2918 if (useuse->in(0) == use) { 2919 useuse->set_req(0, r); 2920 uses_found++; 2921 if (useuse->is_CFG()) { 2922 // This is not a dom_depth > dd_r because when new 2923 // control flow is constructed by a loop opt, a node and 2924 // its dominator can end up at the same dom_depth 2925 assert(dom_depth(useuse) >= dd_r, ""); 2926 set_idom(useuse, r, dom_depth(useuse)); 2927 } 2928 } 2929 for (uint k = 1; k < useuse->req(); k++) { 2930 if( useuse->in(k) == use ) { 2931 useuse->set_req(k, r); 2932 uses_found++; 2933 if (useuse->is_Loop() && k == LoopNode::EntryControl) { 2934 // This is not a dom_depth > dd_r because when new 2935 // control flow is constructed by a loop opt, a node 2936 // and its dominator can end up at the same dom_depth 2937 assert(dom_depth(useuse) >= dd_r , ""); 2938 set_idom(useuse, r, dom_depth(useuse)); 2939 } 2940 } 2941 } 2942 l -= uses_found; // we deleted 1 or more copies of this edge 2943 } 2944 2945 assert(use->is_Proj(), "loop exit should be projection"); 2946 // replace_node_and_forward_ctrl() below moves all nodes that are: 2947 // - control dependent on the loop exit or 2948 // - have control set to the loop exit 2949 // below the post-loop merge point. 2950 // replace_node_and_forward_ctrl() takes a dead control as first input. 2951 // To make it possible to use it, the loop exit projection is cloned and becomes the 2952 // new exit projection. The initial one becomes dead and is "replaced" by the region. 2953 Node* use_clone = use->clone(); 2954 register_control(use_clone, use_loop, idom(use), dom_depth(use)); 2955 // Now finish up 'r' 2956 r->set_req(1, newuse); 2957 r->set_req(2, use_clone); 2958 _igvn.register_new_node_with_optimizer(r); 2959 set_loop(r, use_loop); 2960 set_idom(r, (side_by_side_idom == nullptr) ? newuse->in(0) : side_by_side_idom, dd_r); 2961 replace_node_and_forward_ctrl(use, r); 2962 // Map the (cloned) old use to the new merge point 2963 old_new.map(use_clone->_idx, r); 2964 } // End of if a loop-exit test 2965 } 2966 } 2967 } 2968 2969 void PhaseIdealLoop::fix_body_edges(const Node_List &body, IdealLoopTree* loop, const Node_List &old_new, int dd, 2970 IdealLoopTree* parent, bool partial) { 2971 for(uint i = 0; i < body.size(); i++ ) { 2972 Node *old = body.at(i); 2973 Node *nnn = old_new[old->_idx]; 2974 // Fix CFG/Loop controlling the new node 2975 if (has_ctrl(old)) { 2976 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]); 2977 } else { 2978 set_loop(nnn, parent); 2979 if (old->outcnt() > 0) { 2980 Node* dom = idom(old); 2981 if (old_new[dom->_idx] != nullptr) { 2982 dom = old_new[dom->_idx]; 2983 set_idom(nnn, dom, dd ); 2984 } 2985 } 2986 } 2987 // Correct edges to the new node 2988 for (uint j = 0; j < nnn->req(); j++) { 2989 Node *n = nnn->in(j); 2990 if (n != nullptr) { 2991 IdealLoopTree *old_in_loop = get_loop(has_ctrl(n) ? get_ctrl(n) : n); 2992 if (loop->is_member(old_in_loop)) { 2993 if (old_new[n->_idx] != nullptr) { 2994 nnn->set_req(j, old_new[n->_idx]); 2995 } else { 2996 assert(!body.contains(n), ""); 2997 assert(partial, "node not cloned"); 2998 } 2999 } 3000 } 3001 } 3002 _igvn.hash_find_insert(nnn); 3003 } 3004 } 3005 3006 void PhaseIdealLoop::clone_loop_body(const Node_List& body, Node_List &old_new, CloneMap* cm) { 3007 for (uint i = 0; i < body.size(); i++) { 3008 Node* old = body.at(i); 3009 Node* nnn = old->clone(); 3010 old_new.map(old->_idx, nnn); 3011 if (C->do_vector_loop() && cm != nullptr) { 3012 cm->verify_insert_and_clone(old, nnn, cm->clone_idx()); 3013 } 3014 _igvn.register_new_node_with_optimizer(nnn); 3015 } 3016 } 3017 3018 3019 //---------------------- stride_of_possible_iv ------------------------------------- 3020 // Looks for an iff/bool/comp with one operand of the compare 3021 // being a cycle involving an add and a phi, 3022 // with an optional truncation (left-shift followed by a right-shift) 3023 // of the add. Returns zero if not an iv. 3024 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) { 3025 const TypeInteger* ttype = nullptr; 3026 if (!iff->is_If() || iff->in(1) == nullptr || !iff->in(1)->is_Bool()) { 3027 return 0; 3028 } 3029 BoolNode* bl = iff->in(1)->as_Bool(); 3030 Node* cmp = bl->in(1); 3031 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) { 3032 return 0; 3033 } 3034 // Must have an invariant operand 3035 if (ctrl_is_member(get_loop(iff), cmp->in(2))) { 3036 return 0; 3037 } 3038 Node* add2 = nullptr; 3039 Node* cmp1 = cmp->in(1); 3040 if (cmp1->is_Phi()) { 3041 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) ))) 3042 Node* phi = cmp1; 3043 for (uint i = 1; i < phi->req(); i++) { 3044 Node* in = phi->in(i); 3045 CountedLoopConverter::TruncatedIncrement add(T_INT); 3046 add.build(in); 3047 if (add.is_valid() && add.incr()->in(1) == phi) { 3048 add2 = add.incr()->in(2); 3049 break; 3050 } 3051 } 3052 } else { 3053 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) ))) 3054 Node* addtrunc = cmp1; 3055 CountedLoopConverter::TruncatedIncrement add(T_INT); 3056 add.build(addtrunc); 3057 if (add.is_valid() && add.incr()->in(1)->is_Phi()) { 3058 Node* phi = add.incr()->in(1); 3059 for (uint i = 1; i < phi->req(); i++) { 3060 if (phi->in(i) == addtrunc) { 3061 add2 = add.incr()->in(2); 3062 break; 3063 } 3064 } 3065 } 3066 } 3067 if (add2 != nullptr) { 3068 const TypeInt* add2t = _igvn.type(add2)->is_int(); 3069 if (add2t->is_con()) { 3070 return add2t->get_con(); 3071 } 3072 } 3073 return 0; 3074 } 3075 3076 3077 //---------------------- stay_in_loop ------------------------------------- 3078 // Return the (unique) control output node that's in the loop (if it exists.) 3079 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) { 3080 Node* unique = nullptr; 3081 if (!n) return nullptr; 3082 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 3083 Node* use = n->fast_out(i); 3084 if (!has_ctrl(use) && loop->is_member(get_loop(use))) { 3085 if (unique != nullptr) { 3086 return nullptr; 3087 } 3088 unique = use; 3089 } 3090 } 3091 return unique; 3092 } 3093 3094 //------------------------------ register_node ------------------------------------- 3095 // Utility to register node "n" with PhaseIdealLoop 3096 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree* loop, Node* pred, uint ddepth) { 3097 _igvn.register_new_node_with_optimizer(n); 3098 loop->_body.push(n); 3099 if (n->is_CFG()) { 3100 set_loop(n, loop); 3101 set_idom(n, pred, ddepth); 3102 } else { 3103 set_ctrl(n, pred); 3104 } 3105 } 3106 3107 //------------------------------ proj_clone ------------------------------------- 3108 // Utility to create an if-projection 3109 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) { 3110 ProjNode* c = p->clone()->as_Proj(); 3111 c->set_req(0, iff); 3112 return c; 3113 } 3114 3115 //------------------------------ short_circuit_if ------------------------------------- 3116 // Force the iff control output to be the live_proj 3117 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) { 3118 guarantee(live_proj != nullptr, "null projection"); 3119 int proj_con = live_proj->_con; 3120 assert(proj_con == 0 || proj_con == 1, "false or true projection"); 3121 Node* con = intcon(proj_con); 3122 if (iff) { 3123 iff->set_req(1, con); 3124 } 3125 return con; 3126 } 3127 3128 //------------------------------ insert_if_before_proj ------------------------------------- 3129 // Insert a new if before an if projection (* - new node) 3130 // 3131 // before 3132 // if(test) 3133 // / \ 3134 // v v 3135 // other-proj proj (arg) 3136 // 3137 // after 3138 // if(test) 3139 // / \ 3140 // / v 3141 // | * proj-clone 3142 // v | 3143 // other-proj v 3144 // * new_if(relop(cmp[IU](left,right))) 3145 // / \ 3146 // v v 3147 // * new-proj proj 3148 // (returned) 3149 // 3150 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) { 3151 IfNode* iff = proj->in(0)->as_If(); 3152 IdealLoopTree *loop = get_loop(proj); 3153 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 3154 uint ddepth = dom_depth(proj); 3155 3156 _igvn.rehash_node_delayed(iff); 3157 _igvn.rehash_node_delayed(proj); 3158 3159 proj->set_req(0, nullptr); // temporary disconnect 3160 ProjNode* proj2 = proj_clone(proj, iff); 3161 register_node(proj2, loop, iff, ddepth); 3162 3163 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right); 3164 register_node(cmp, loop, proj2, ddepth); 3165 3166 BoolNode* bol = new BoolNode(cmp, relop); 3167 register_node(bol, loop, proj2, ddepth); 3168 3169 int opcode = iff->Opcode(); 3170 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode"); 3171 IfNode* new_if = IfNode::make_with_same_profile(iff, proj2, bol); 3172 register_node(new_if, loop, proj2, ddepth); 3173 3174 proj->set_req(0, new_if); // reattach 3175 set_idom(proj, new_if, ddepth); 3176 3177 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj(); 3178 guarantee(new_exit != nullptr, "null exit node"); 3179 register_node(new_exit, get_loop(other_proj), new_if, ddepth); 3180 3181 return new_exit; 3182 } 3183 3184 //------------------------------ insert_region_before_proj ------------------------------------- 3185 // Insert a region before an if projection (* - new node) 3186 // 3187 // before 3188 // if(test) 3189 // / | 3190 // v | 3191 // proj v 3192 // other-proj 3193 // 3194 // after 3195 // if(test) 3196 // / | 3197 // v | 3198 // * proj-clone v 3199 // | other-proj 3200 // v 3201 // * new-region 3202 // | 3203 // v 3204 // * dum_if 3205 // / \ 3206 // v \ 3207 // * dum-proj v 3208 // proj 3209 // 3210 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) { 3211 IfNode* iff = proj->in(0)->as_If(); 3212 IdealLoopTree *loop = get_loop(proj); 3213 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 3214 uint ddepth = dom_depth(proj); 3215 3216 _igvn.rehash_node_delayed(iff); 3217 _igvn.rehash_node_delayed(proj); 3218 3219 proj->set_req(0, nullptr); // temporary disconnect 3220 ProjNode* proj2 = proj_clone(proj, iff); 3221 register_node(proj2, loop, iff, ddepth); 3222 3223 RegionNode* reg = new RegionNode(2); 3224 reg->set_req(1, proj2); 3225 register_node(reg, loop, iff, ddepth); 3226 3227 IfNode* dum_if = new IfNode(reg, short_circuit_if(nullptr, proj), iff->_prob, iff->_fcnt); 3228 register_node(dum_if, loop, reg, ddepth); 3229 3230 proj->set_req(0, dum_if); // reattach 3231 set_idom(proj, dum_if, ddepth); 3232 3233 ProjNode* dum_proj = proj_clone(other_proj, dum_if); 3234 register_node(dum_proj, loop, dum_if, ddepth); 3235 3236 return reg; 3237 } 3238 3239 // Idea 3240 // ---- 3241 // Partial Peeling tries to rotate the loop in such a way that it can later be turned into a counted loop. Counted loops 3242 // require a signed loop exit test. When calling this method, we've only found a suitable unsigned test to partial peel 3243 // with. Therefore, we try to split off a signed loop exit test from the unsigned test such that it can be used as new 3244 // loop exit while keeping the unsigned test unchanged and preserving the same behavior as if we've used the unsigned 3245 // test alone instead: 3246 // 3247 // Before Partial Peeling: 3248 // Loop: 3249 // <peeled section> 3250 // Split off signed loop exit test 3251 // <-- CUT HERE --> 3252 // Unchanged unsigned loop exit test 3253 // <rest of unpeeled section> 3254 // goto Loop 3255 // 3256 // After Partial Peeling: 3257 // <cloned peeled section> 3258 // Cloned split off signed loop exit test 3259 // Loop: 3260 // Unchanged unsigned loop exit test 3261 // <rest of unpeeled section> 3262 // <peeled section> 3263 // Split off signed loop exit test 3264 // goto Loop 3265 // 3266 // Details 3267 // ------- 3268 // Before: 3269 // if (i <u limit) Unsigned loop exit condition 3270 // / | 3271 // v v 3272 // exit-proj stay-in-loop-proj 3273 // 3274 // Split off a signed loop exit test (i.e. with CmpI) from an unsigned loop exit test (i.e. with CmpU) and insert it 3275 // before the CmpU on the stay-in-loop path and keep both tests: 3276 // 3277 // if (i <u limit) Signed loop exit test 3278 // / | 3279 // / if (i <u limit) Unsigned loop exit test 3280 // / / | 3281 // v v v 3282 // exit-region stay-in-loop-proj 3283 // 3284 // Implementation 3285 // -------------- 3286 // We need to make sure that the new signed loop exit test is properly inserted into the graph such that the unsigned 3287 // loop exit test still dominates the same set of control nodes, the ctrl() relation from data nodes to both loop 3288 // exit tests is preserved, and their loop nesting is correct. 3289 // 3290 // To achieve that, we clone the unsigned loop exit test completely (leave it unchanged), insert the signed loop exit 3291 // test above it and kill the original unsigned loop exit test by setting it's condition to a constant 3292 // (i.e. stay-in-loop-const in graph below) such that IGVN can fold it later: 3293 // 3294 // if (stay-in-loop-const) Killed original unsigned loop exit test 3295 // / | 3296 // / v 3297 // / if (i < limit) Split off signed loop exit test 3298 // / / | 3299 // / / v 3300 // / / if (i <u limit) Cloned unsigned loop exit test 3301 // / / / | 3302 // v v v | 3303 // exit-region | 3304 // | | 3305 // dummy-if | 3306 // / | | 3307 // dead | | 3308 // v v 3309 // exit-proj stay-in-loop-proj 3310 // 3311 // Note: The dummy-if is inserted to create a region to merge the loop exits between the original to be killed unsigned 3312 // loop exit test and its exit projection while keeping the exit projection (also see insert_region_before_proj()). 3313 // 3314 // Requirements 3315 // ------------ 3316 // Note that we can only split off a signed loop exit test from the unsigned loop exit test when the behavior is exactly 3317 // the same as before with only a single unsigned test. This is only possible if certain requirements are met. 3318 // Otherwise, we need to bail out (see comments in the code below). 3319 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree* loop) { 3320 const bool Signed = true; 3321 const bool Unsigned = false; 3322 3323 BoolNode* bol = if_cmpu->in(1)->as_Bool(); 3324 if (bol->_test._test != BoolTest::lt) { 3325 return nullptr; 3326 } 3327 CmpNode* cmpu = bol->in(1)->as_Cmp(); 3328 assert(cmpu->Opcode() == Op_CmpU, "must be unsigned comparison"); 3329 3330 int stride = stride_of_possible_iv(if_cmpu); 3331 if (stride == 0) { 3332 return nullptr; 3333 } 3334 3335 Node* lp_proj = stay_in_loop(if_cmpu, loop); 3336 guarantee(lp_proj != nullptr, "null loop node"); 3337 3338 ProjNode* lp_continue = lp_proj->as_Proj(); 3339 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj(); 3340 if (!lp_exit->is_IfFalse()) { 3341 // The loop exit condition is (i <u limit) ==> (i >= 0 && i < limit). 3342 // We therefore can't add a single exit condition. 3343 return nullptr; 3344 } 3345 // The unsigned loop exit condition is 3346 // !(i <u limit) 3347 // = i >=u limit 3348 // 3349 // First, we note that for any x for which 3350 // 0 <= x <= INT_MAX 3351 // we can convert x to an unsigned int and still get the same guarantee: 3352 // 0 <= (uint) x <= INT_MAX = (uint) INT_MAX 3353 // 0 <=u (uint) x <=u INT_MAX = (uint) INT_MAX (LEMMA) 3354 // 3355 // With that in mind, if 3356 // limit >= 0 (COND) 3357 // then the unsigned loop exit condition 3358 // i >=u limit (ULE) 3359 // is equivalent to 3360 // i < 0 || i >= limit (SLE-full) 3361 // because either i is negative and therefore always greater than MAX_INT when converting to unsigned 3362 // (uint) i >=u MAX_INT >= limit >= 0 3363 // or otherwise 3364 // i >= limit >= 0 3365 // holds due to (LEMMA). 3366 // 3367 // For completeness, a counterexample with limit < 0: 3368 // Assume i = -3 and limit = -2: 3369 // i < 0 3370 // -2 < 0 3371 // is true and thus also "i < 0 || i >= limit". But 3372 // i >=u limit 3373 // -3 >=u -2 3374 // is false. 3375 Node* limit = cmpu->in(2); 3376 const TypeInt* type_limit = _igvn.type(limit)->is_int(); 3377 if (type_limit->_lo < 0) { 3378 return nullptr; 3379 } 3380 3381 // We prove below that we can extract a single signed loop exit condition from (SLE-full), depending on the stride: 3382 // stride < 0: 3383 // i < 0 (SLE = SLE-negative) 3384 // stride > 0: 3385 // i >= limit (SLE = SLE-positive) 3386 // such that we have the following graph before Partial Peeling with stride > 0 (similar for stride < 0): 3387 // 3388 // Loop: 3389 // <peeled section> 3390 // i >= limit (SLE-positive) 3391 // <-- CUT HERE --> 3392 // i >=u limit (ULE) 3393 // <rest of unpeeled section> 3394 // goto Loop 3395 // 3396 // We exit the loop if: 3397 // (SLE) is true OR (ULE) is true 3398 // However, if (SLE) is true then (ULE) also needs to be true to ensure the exact same behavior. Otherwise, we wrongly 3399 // exit a loop that should not have been exited if we did not apply Partial Peeling. More formally, we need to ensure: 3400 // (SLE) IMPLIES (ULE) 3401 // This indeed holds when (COND) is given: 3402 // - stride > 0: 3403 // i >= limit // (SLE = SLE-positive) 3404 // i >= limit >= 0 // (COND) 3405 // i >=u limit >= 0 // (LEMMA) 3406 // which is the unsigned loop exit condition (ULE). 3407 // - stride < 0: 3408 // i < 0 // (SLE = SLE-negative) 3409 // (uint) i >u MAX_INT // (NEG) all negative values are greater than MAX_INT when converted to unsigned 3410 // MAX_INT >= limit >= 0 // (COND) 3411 // MAX_INT >=u limit >= 0 // (LEMMA) 3412 // and thus from (NEG) and (LEMMA): 3413 // i >=u limit 3414 // which is the unsigned loop exit condition (ULE). 3415 // 3416 // 3417 // After Partial Peeling, we have the following structure for stride > 0 (similar for stride < 0): 3418 // <cloned peeled section> 3419 // i >= limit (SLE-positive) 3420 // Loop: 3421 // i >=u limit (ULE) 3422 // <rest of unpeeled section> 3423 // <peeled section> 3424 // i >= limit (SLE-positive) 3425 // goto Loop 3426 Node* rhs_cmpi; 3427 if (stride > 0) { 3428 rhs_cmpi = limit; // For i >= limit 3429 } else { 3430 rhs_cmpi = makecon(TypeInt::ZERO); // For i < 0 3431 } 3432 // Create a new region on the exit path 3433 RegionNode* reg = insert_region_before_proj(lp_exit); 3434 guarantee(reg != nullptr, "null region node"); 3435 3436 // Clone the if-cmpu-true-false using a signed compare 3437 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge; 3438 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, rhs_cmpi, lp_continue); 3439 reg->add_req(cmpi_exit); 3440 3441 // Clone the if-cmpu-true-false 3442 BoolTest::mask rel_u = bol->_test._test; 3443 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue); 3444 reg->add_req(cmpu_exit); 3445 3446 // Force original if to stay in loop. 3447 short_circuit_if(if_cmpu, lp_continue); 3448 3449 return cmpi_exit->in(0)->as_If(); 3450 } 3451 3452 //------------------------------ remove_cmpi_loop_exit ------------------------------------- 3453 // Remove a previously inserted signed compare loop exit. 3454 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) { 3455 Node* lp_proj = stay_in_loop(if_cmp, loop); 3456 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI && 3457 stay_in_loop(lp_proj, loop)->is_If() && 3458 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu"); 3459 Node* con = makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO); 3460 if_cmp->set_req(1, con); 3461 } 3462 3463 //------------------------------ scheduled_nodelist ------------------------------------- 3464 // Create a post order schedule of nodes that are in the 3465 // "member" set. The list is returned in "sched". 3466 // The first node in "sched" is the loop head, followed by 3467 // nodes which have no inputs in the "member" set, and then 3468 // followed by the nodes that have an immediate input dependence 3469 // on a node in "sched". 3470 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) { 3471 3472 assert(member.test(loop->_head->_idx), "loop head must be in member set"); 3473 VectorSet visited; 3474 Node_Stack nstack(loop->_body.size()); 3475 3476 Node* n = loop->_head; // top of stack is cached in "n" 3477 uint idx = 0; 3478 visited.set(n->_idx); 3479 3480 // Initially push all with no inputs from within member set 3481 for(uint i = 0; i < loop->_body.size(); i++ ) { 3482 Node *elt = loop->_body.at(i); 3483 if (member.test(elt->_idx)) { 3484 bool found = false; 3485 for (uint j = 0; j < elt->req(); j++) { 3486 Node* def = elt->in(j); 3487 if (def && member.test(def->_idx) && def != elt) { 3488 found = true; 3489 break; 3490 } 3491 } 3492 if (!found && elt != loop->_head) { 3493 nstack.push(n, idx); 3494 n = elt; 3495 assert(!visited.test(n->_idx), "not seen yet"); 3496 visited.set(n->_idx); 3497 } 3498 } 3499 } 3500 3501 // traverse out's that are in the member set 3502 while (true) { 3503 if (idx < n->outcnt()) { 3504 Node* use = n->raw_out(idx); 3505 idx++; 3506 if (!visited.test_set(use->_idx)) { 3507 if (member.test(use->_idx)) { 3508 nstack.push(n, idx); 3509 n = use; 3510 idx = 0; 3511 } 3512 } 3513 } else { 3514 // All outputs processed 3515 sched.push(n); 3516 if (nstack.is_empty()) break; 3517 n = nstack.node(); 3518 idx = nstack.index(); 3519 nstack.pop(); 3520 } 3521 } 3522 } 3523 3524 3525 //------------------------------ has_use_in_set ------------------------------------- 3526 // Has a use in the vector set 3527 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) { 3528 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3529 Node* use = n->fast_out(j); 3530 if (vset.test(use->_idx)) { 3531 return true; 3532 } 3533 } 3534 return false; 3535 } 3536 3537 3538 //------------------------------ has_use_internal_to_set ------------------------------------- 3539 // Has use internal to the vector set (ie. not in a phi at the loop head) 3540 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) { 3541 Node* head = loop->_head; 3542 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3543 Node* use = n->fast_out(j); 3544 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) { 3545 return true; 3546 } 3547 } 3548 return false; 3549 } 3550 3551 3552 //------------------------------ clone_for_use_outside_loop ------------------------------------- 3553 // clone "n" for uses that are outside of loop 3554 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) { 3555 int cloned = 0; 3556 assert(worklist.size() == 0, "should be empty"); 3557 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3558 Node* use = n->fast_out(j); 3559 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) { 3560 worklist.push(use); 3561 } 3562 } 3563 3564 if (C->check_node_count(worklist.size() + NodeLimitFudgeFactor, 3565 "Too many clones required in clone_for_use_outside_loop in partial peeling")) { 3566 return -1; 3567 } 3568 3569 while( worklist.size() ) { 3570 Node *use = worklist.pop(); 3571 if (!has_node(use) || use->in(0) == C->top()) continue; 3572 uint j; 3573 for (j = 0; j < use->req(); j++) { 3574 if (use->in(j) == n) break; 3575 } 3576 assert(j < use->req(), "must be there"); 3577 3578 // clone "n" and insert it between the inputs of "n" and the use outside the loop 3579 Node* n_clone = n->clone(); 3580 _igvn.replace_input_of(use, j, n_clone); 3581 cloned++; 3582 Node* use_c; 3583 if (!use->is_Phi()) { 3584 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0); 3585 } else { 3586 // Use in a phi is considered a use in the associated predecessor block 3587 use_c = use->in(0)->in(j); 3588 } 3589 set_ctrl(n_clone, use_c); 3590 assert(!loop->is_member(get_loop(use_c)), "should be outside loop"); 3591 get_loop(use_c)->_body.push(n_clone); 3592 _igvn.register_new_node_with_optimizer(n_clone); 3593 #ifndef PRODUCT 3594 if (TracePartialPeeling) { 3595 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx); 3596 } 3597 #endif 3598 } 3599 return cloned; 3600 } 3601 3602 3603 //------------------------------ clone_for_special_use_inside_loop ------------------------------------- 3604 // clone "n" for special uses that are in the not_peeled region. 3605 // If these def-uses occur in separate blocks, the code generator 3606 // marks the method as not compilable. For example, if a "BoolNode" 3607 // is in a different basic block than the "IfNode" that uses it, then 3608 // the compilation is aborted in the code generator. 3609 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 3610 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) { 3611 if (n->is_Phi() || n->is_Load()) { 3612 return; 3613 } 3614 assert(worklist.size() == 0, "should be empty"); 3615 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3616 Node* use = n->fast_out(j); 3617 if ( not_peel.test(use->_idx) && 3618 (use->is_If() || use->is_CMove() || use->is_Bool() || use->is_OpaqueInitializedAssertionPredicate()) && 3619 use->in(1) == n) { 3620 worklist.push(use); 3621 } 3622 } 3623 if (worklist.size() > 0) { 3624 // clone "n" and insert it between inputs of "n" and the use 3625 Node* n_clone = n->clone(); 3626 loop->_body.push(n_clone); 3627 _igvn.register_new_node_with_optimizer(n_clone); 3628 set_ctrl(n_clone, get_ctrl(n)); 3629 sink_list.push(n_clone); 3630 not_peel.set(n_clone->_idx); 3631 #ifndef PRODUCT 3632 if (TracePartialPeeling) { 3633 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx); 3634 } 3635 #endif 3636 while( worklist.size() ) { 3637 Node *use = worklist.pop(); 3638 _igvn.rehash_node_delayed(use); 3639 for (uint j = 1; j < use->req(); j++) { 3640 if (use->in(j) == n) { 3641 use->set_req(j, n_clone); 3642 } 3643 } 3644 } 3645 } 3646 } 3647 3648 3649 //------------------------------ insert_phi_for_loop ------------------------------------- 3650 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 3651 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) { 3652 Node *phi = PhiNode::make(lp, back_edge_val); 3653 phi->set_req(LoopNode::EntryControl, lp_entry_val); 3654 // Use existing phi if it already exists 3655 Node *hit = _igvn.hash_find_insert(phi); 3656 if( hit == nullptr ) { 3657 _igvn.register_new_node_with_optimizer(phi); 3658 set_ctrl(phi, lp); 3659 } else { 3660 // Remove the new phi from the graph and use the hit 3661 _igvn.remove_dead_node(phi, PhaseIterGVN::NodeOrigin::Speculative); 3662 phi = hit; 3663 } 3664 _igvn.replace_input_of(use, idx, phi); 3665 } 3666 3667 #ifdef ASSERT 3668 //------------------------------ is_valid_loop_partition ------------------------------------- 3669 // Validate the loop partition sets: peel and not_peel 3670 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, 3671 VectorSet& not_peel ) { 3672 uint i; 3673 // Check that peel_list entries are in the peel set 3674 for (i = 0; i < peel_list.size(); i++) { 3675 if (!peel.test(peel_list.at(i)->_idx)) { 3676 return false; 3677 } 3678 } 3679 // Check at loop members are in one of peel set or not_peel set 3680 for (i = 0; i < loop->_body.size(); i++ ) { 3681 Node *def = loop->_body.at(i); 3682 uint di = def->_idx; 3683 // Check that peel set elements are in peel_list 3684 if (peel.test(di)) { 3685 if (not_peel.test(di)) { 3686 return false; 3687 } 3688 // Must be in peel_list also 3689 bool found = false; 3690 for (uint j = 0; j < peel_list.size(); j++) { 3691 if (peel_list.at(j)->_idx == di) { 3692 found = true; 3693 break; 3694 } 3695 } 3696 if (!found) { 3697 return false; 3698 } 3699 } else if (not_peel.test(di)) { 3700 if (peel.test(di)) { 3701 return false; 3702 } 3703 } else { 3704 return false; 3705 } 3706 } 3707 return true; 3708 } 3709 3710 //------------------------------ is_valid_clone_loop_exit_use ------------------------------------- 3711 // Ensure a use outside of loop is of the right form 3712 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) { 3713 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 3714 return (use->is_Phi() && 3715 use_c->is_Region() && use_c->req() == 3 && 3716 (use_c->in(exit_idx)->Opcode() == Op_IfTrue || 3717 use_c->in(exit_idx)->Opcode() == Op_IfFalse || 3718 use_c->in(exit_idx)->Opcode() == Op_JumpProj) && 3719 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) ); 3720 } 3721 3722 //------------------------------ is_valid_clone_loop_form ------------------------------------- 3723 // Ensure that all uses outside of loop are of the right form 3724 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 3725 uint orig_exit_idx, uint clone_exit_idx) { 3726 uint len = peel_list.size(); 3727 for (uint i = 0; i < len; i++) { 3728 Node *def = peel_list.at(i); 3729 3730 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 3731 Node *use = def->fast_out(j); 3732 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 3733 if (!loop->is_member(get_loop(use_c))) { 3734 // use is not in the loop, check for correct structure 3735 if (use->in(0) == def) { 3736 // Okay 3737 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) { 3738 return false; 3739 } 3740 } 3741 } 3742 } 3743 return true; 3744 } 3745 #endif 3746 3747 //------------------------------ partial_peel ------------------------------------- 3748 // Partially peel (aka loop rotation) the top portion of a loop (called 3749 // the peel section below) by cloning it and placing one copy just before 3750 // the new loop head and the other copy at the bottom of the new loop. 3751 // 3752 // before after where it came from 3753 // 3754 // stmt1 stmt1 3755 // loop: stmt2 clone 3756 // stmt2 if condA goto exitA clone 3757 // if condA goto exitA new_loop: new 3758 // stmt3 stmt3 clone 3759 // if !condB goto loop if condB goto exitB clone 3760 // exitB: stmt2 orig 3761 // stmt4 if !condA goto new_loop orig 3762 // exitA: goto exitA 3763 // exitB: 3764 // stmt4 3765 // exitA: 3766 // 3767 // Step 1: find the cut point: an exit test on probable 3768 // induction variable. 3769 // Step 2: schedule (with cloning) operations in the peel 3770 // section that can be executed after the cut into 3771 // the section that is not peeled. This may need 3772 // to clone operations into exit blocks. For 3773 // instance, a reference to A[i] in the not-peel 3774 // section and a reference to B[i] in an exit block 3775 // may cause a left-shift of i by 2 to be placed 3776 // in the peel block. This step will clone the left 3777 // shift into the exit block and sink the left shift 3778 // from the peel to the not-peel section. 3779 // Step 3: clone the loop, retarget the control, and insert 3780 // phis for values that are live across the new loop 3781 // head. This is very dependent on the graph structure 3782 // from clone_loop. It creates region nodes for 3783 // exit control and associated phi nodes for values 3784 // flow out of the loop through that exit. The region 3785 // node is dominated by the clone's control projection. 3786 // So the clone's peel section is placed before the 3787 // new loop head, and the clone's not-peel section is 3788 // forms the top part of the new loop. The original 3789 // peel section forms the tail of the new loop. 3790 // Step 4: update the dominator tree and recompute the 3791 // dominator depth. 3792 // 3793 // orig 3794 // 3795 // stmt1 3796 // | 3797 // v 3798 // predicates 3799 // | 3800 // v 3801 // loop<----+ 3802 // | | 3803 // stmt2 | 3804 // | | 3805 // v | 3806 // ifA | 3807 // / | | 3808 // v v | 3809 // false true ^ <-- last_peel 3810 // / | | 3811 // / ===|==cut | 3812 // / stmt3 | <-- first_not_peel 3813 // / | | 3814 // | v | 3815 // v ifB | 3816 // exitA: / \ | 3817 // / \ | 3818 // v v | 3819 // false true | 3820 // / \ | 3821 // / ----+ 3822 // | 3823 // v 3824 // exitB: 3825 // stmt4 3826 // 3827 // 3828 // after clone loop 3829 // 3830 // stmt1 3831 // | 3832 // v 3833 // predicates 3834 // / \ 3835 // clone / \ orig 3836 // / \ 3837 // / \ 3838 // v v 3839 // +---->loop loop<----+ 3840 // | | | | 3841 // | stmt2 stmt2 | 3842 // | | | | 3843 // | v v | 3844 // | ifA ifA | 3845 // | | \ / | | 3846 // | v v v v | 3847 // ^ true false false true ^ <-- last_peel 3848 // | | ^ \ / | | 3849 // | cut==|== \ \ / ===|==cut | 3850 // | stmt3 \ \ / stmt3 | <-- first_not_peel 3851 // | | dom | | | | 3852 // | v \ 1v v2 v | 3853 // | ifB regionA ifB | 3854 // | / \ | / \ | 3855 // | / \ v / \ | 3856 // | v v exitA: v v | 3857 // | true false false true | 3858 // | / ^ \ / \ | 3859 // +---- \ \ / ----+ 3860 // dom \ / 3861 // \ 1v v2 3862 // regionB 3863 // | 3864 // v 3865 // exitB: 3866 // stmt4 3867 // 3868 // 3869 // after partial peel 3870 // 3871 // stmt1 3872 // | 3873 // v 3874 // predicates 3875 // / 3876 // clone / orig 3877 // / TOP 3878 // / \ 3879 // v v 3880 // TOP->loop loop----+ 3881 // | | | 3882 // stmt2 stmt2 | 3883 // | | | 3884 // v v | 3885 // ifA ifA | 3886 // | \ / | | 3887 // v v v v | 3888 // true false false true | <-- last_peel 3889 // | ^ \ / +------|---+ 3890 // +->newloop \ \ / === ==cut | | 3891 // | stmt3 \ \ / TOP | | 3892 // | | dom | | stmt3 | | <-- first_not_peel 3893 // | v \ 1v v2 v | | 3894 // | ifB regionA ifB ^ v 3895 // | / \ | / \ | | 3896 // | / \ v / \ | | 3897 // | v v exitA: v v | | 3898 // | true false false true | | 3899 // | / ^ \ / \ | | 3900 // | | \ \ / v | | 3901 // | | dom \ / TOP | | 3902 // | | \ 1v v2 | | 3903 // ^ v regionB | | 3904 // | | | | | 3905 // | | v ^ v 3906 // | | exitB: | | 3907 // | | stmt4 | | 3908 // | +------------>-----------------+ | 3909 // | | 3910 // +-----------------<---------------------+ 3911 // 3912 // 3913 // final graph 3914 // 3915 // stmt1 3916 // | 3917 // v 3918 // predicates 3919 // | 3920 // v 3921 // stmt2 clone 3922 // | 3923 // v 3924 // ........> ifA clone 3925 // : / | 3926 // dom / | 3927 // : v v 3928 // : false true 3929 // : | | 3930 // : | v 3931 // : | newloop<-----+ 3932 // : | | | 3933 // : | stmt3 clone | 3934 // : | | | 3935 // : | v | 3936 // : | ifB | 3937 // : | / \ | 3938 // : | v v | 3939 // : | false true | 3940 // : | | | | 3941 // : | v stmt2 | 3942 // : | exitB: | | 3943 // : | stmt4 v | 3944 // : | ifA orig | 3945 // : | / \ | 3946 // : | / \ | 3947 // : | v v | 3948 // : | false true | 3949 // : | / \ | 3950 // : v v -----+ 3951 // RegionA 3952 // | 3953 // v 3954 // exitA 3955 // 3956 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) { 3957 3958 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only"); 3959 if (!loop->_head->is_Loop()) { 3960 return false; 3961 } 3962 LoopNode *head = loop->_head->as_Loop(); 3963 3964 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) { 3965 return false; 3966 } 3967 3968 // Check for complex exit control 3969 for (uint ii = 0; ii < loop->_body.size(); ii++) { 3970 Node *n = loop->_body.at(ii); 3971 int opc = n->Opcode(); 3972 if (n->is_Call() || 3973 opc == Op_Catch || 3974 opc == Op_CatchProj || 3975 opc == Op_Jump || 3976 opc == Op_JumpProj) { 3977 #ifndef PRODUCT 3978 if (TracePartialPeeling) { 3979 tty->print_cr("\nExit control too complex: lp: %d", head->_idx); 3980 } 3981 #endif 3982 return false; 3983 } 3984 } 3985 3986 int dd = dom_depth(head); 3987 3988 // Step 1: find cut point 3989 3990 // Walk up dominators to loop head looking for first loop exit 3991 // which is executed on every path thru loop. 3992 IfNode *peel_if = nullptr; 3993 IfNode *peel_if_cmpu = nullptr; 3994 3995 Node *iff = loop->tail(); 3996 while (iff != head) { 3997 if (iff->is_If()) { 3998 Node *ctrl = get_ctrl(iff->in(1)); 3999 if (ctrl->is_top()) return false; // Dead test on live IF. 4000 // If loop-varying exit-test, check for induction variable 4001 if (loop->is_member(get_loop(ctrl)) && 4002 loop->is_loop_exit(iff) && 4003 is_possible_iv_test(iff)) { 4004 Node* cmp = iff->in(1)->in(1); 4005 if (cmp->Opcode() == Op_CmpI) { 4006 peel_if = iff->as_If(); 4007 } else { 4008 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU"); 4009 peel_if_cmpu = iff->as_If(); 4010 } 4011 } 4012 } 4013 iff = idom(iff); 4014 } 4015 4016 // Prefer signed compare over unsigned compare. 4017 IfNode* new_peel_if = nullptr; 4018 if (peel_if == nullptr) { 4019 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == nullptr) { 4020 return false; // No peel point found 4021 } 4022 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop); 4023 if (new_peel_if == nullptr) { 4024 return false; // No peel point found 4025 } 4026 peel_if = new_peel_if; 4027 } 4028 Node* last_peel = stay_in_loop(peel_if, loop); 4029 Node* first_not_peeled = stay_in_loop(last_peel, loop); 4030 if (first_not_peeled == nullptr || first_not_peeled == head) { 4031 return false; 4032 } 4033 4034 #ifndef PRODUCT 4035 if (TraceLoopOpts) { 4036 tty->print("PartialPeel "); 4037 loop->dump_head(); 4038 } 4039 4040 if (TracePartialPeeling) { 4041 tty->print_cr("before partial peel one iteration"); 4042 Node_List wl; 4043 Node* t = head->in(2); 4044 while (true) { 4045 wl.push(t); 4046 if (t == head) break; 4047 t = idom(t); 4048 } 4049 while (wl.size() > 0) { 4050 Node* tt = wl.pop(); 4051 tt->dump(); 4052 if (tt == last_peel) tty->print_cr("-- cut --"); 4053 } 4054 } 4055 #endif 4056 4057 C->print_method(PHASE_BEFORE_PARTIAL_PEELING, 4, head); 4058 4059 VectorSet peel; 4060 VectorSet not_peel; 4061 Node_List peel_list; 4062 Node_List worklist; 4063 Node_List sink_list; 4064 4065 uint estimate = loop->est_loop_clone_sz(1); 4066 if (exceeding_node_budget(estimate)) { 4067 return false; 4068 } 4069 4070 // Set of cfg nodes to peel are those that are executable from 4071 // the head through last_peel. 4072 assert(worklist.size() == 0, "should be empty"); 4073 worklist.push(head); 4074 peel.set(head->_idx); 4075 while (worklist.size() > 0) { 4076 Node *n = worklist.pop(); 4077 if (n != last_peel) { 4078 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 4079 Node* use = n->fast_out(j); 4080 if (use->is_CFG() && 4081 loop->is_member(get_loop(use)) && 4082 !peel.test_set(use->_idx)) { 4083 worklist.push(use); 4084 } 4085 } 4086 } 4087 } 4088 4089 // Set of non-cfg nodes to peel are those that are control 4090 // dependent on the cfg nodes. 4091 for (uint i = 0; i < loop->_body.size(); i++) { 4092 Node *n = loop->_body.at(i); 4093 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n; 4094 if (peel.test(n_c->_idx)) { 4095 peel.set(n->_idx); 4096 } else { 4097 not_peel.set(n->_idx); 4098 } 4099 } 4100 4101 // Step 2: move operations from the peeled section down into the 4102 // not-peeled section 4103 4104 // Get a post order schedule of nodes in the peel region 4105 // Result in right-most operand. 4106 scheduled_nodelist(loop, peel, peel_list); 4107 4108 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 4109 4110 // For future check for too many new phis 4111 uint old_phi_cnt = 0; 4112 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 4113 Node* use = head->fast_out(j); 4114 if (use->is_Phi()) old_phi_cnt++; 4115 } 4116 4117 #ifndef PRODUCT 4118 if (TracePartialPeeling) { 4119 tty->print_cr("\npeeled list"); 4120 } 4121 #endif 4122 4123 // Evacuate nodes in peel region into the not_peeled region if possible 4124 bool too_many_clones = false; 4125 uint new_phi_cnt = 0; 4126 uint cloned_for_outside_use = 0; 4127 for (uint i = 0; i < peel_list.size();) { 4128 Node* n = peel_list.at(i); 4129 #ifndef PRODUCT 4130 if (TracePartialPeeling) n->dump(); 4131 #endif 4132 bool incr = true; 4133 if (!n->is_CFG()) { 4134 if (has_use_in_set(n, not_peel)) { 4135 // If not used internal to the peeled region, 4136 // move "n" from peeled to not_peeled region. 4137 if (!has_use_internal_to_set(n, peel, loop)) { 4138 // if not pinned and not a load (which maybe anti-dependent on a store) 4139 // and not a CMove (Matcher expects only bool->cmove). 4140 if (n->in(0) == nullptr && !n->is_Load() && !n->is_CMove()) { 4141 int new_clones = clone_for_use_outside_loop(loop, n, worklist); 4142 if (C->failing()) return false; 4143 if (new_clones == -1) { 4144 too_many_clones = true; 4145 break; 4146 } 4147 cloned_for_outside_use += new_clones; 4148 sink_list.push(n); 4149 peel.remove(n->_idx); 4150 not_peel.set(n->_idx); 4151 peel_list.remove(i); 4152 incr = false; 4153 #ifndef PRODUCT 4154 if (TracePartialPeeling) { 4155 tty->print_cr("sink to not_peeled region: %d newbb: %d", 4156 n->_idx, get_ctrl(n)->_idx); 4157 } 4158 #endif 4159 } 4160 } else { 4161 // Otherwise check for special def-use cases that span 4162 // the peel/not_peel boundary such as bool->if 4163 clone_for_special_use_inside_loop(loop, n, not_peel, sink_list, worklist); 4164 new_phi_cnt++; 4165 } 4166 } 4167 } 4168 if (incr) i++; 4169 } 4170 4171 estimate += cloned_for_outside_use + new_phi_cnt; 4172 bool exceed_node_budget = !may_require_nodes(estimate); 4173 bool exceed_phi_limit = new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta; 4174 4175 if (too_many_clones || exceed_node_budget || exceed_phi_limit) { 4176 #ifndef PRODUCT 4177 if (TracePartialPeeling && exceed_phi_limit) { 4178 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c", 4179 new_phi_cnt, old_phi_cnt, new_peel_if != nullptr?'T':'F'); 4180 } 4181 #endif 4182 if (new_peel_if != nullptr) { 4183 remove_cmpi_loop_exit(new_peel_if, loop); 4184 } 4185 // Inhibit more partial peeling on this loop 4186 assert(!head->is_partial_peel_loop(), "not partial peeled"); 4187 head->mark_partial_peel_failed(); 4188 if (cloned_for_outside_use > 0) { 4189 // Terminate this round of loop opts because 4190 // the graph outside this loop was changed. 4191 C->set_major_progress(); 4192 return true; 4193 } 4194 return false; 4195 } 4196 4197 // Step 3: clone loop, retarget control, and insert new phis 4198 4199 // Create new loop head for new phis and to hang 4200 // the nodes being moved (sinked) from the peel region. 4201 LoopNode* new_head = new LoopNode(last_peel, last_peel); 4202 new_head->set_unswitch_count(head->unswitch_count()); // Preserve 4203 _igvn.register_new_node_with_optimizer(new_head); 4204 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled"); 4205 _igvn.replace_input_of(first_not_peeled, 0, new_head); 4206 set_loop(new_head, loop); 4207 loop->_body.push(new_head); 4208 not_peel.set(new_head->_idx); 4209 set_idom(new_head, last_peel, dom_depth(first_not_peeled)); 4210 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled)); 4211 4212 while (sink_list.size() > 0) { 4213 Node* n = sink_list.pop(); 4214 set_ctrl(n, new_head); 4215 } 4216 4217 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 4218 4219 clone_loop(loop, old_new, dd, IgnoreStripMined); 4220 4221 const uint clone_exit_idx = 1; 4222 const uint orig_exit_idx = 2; 4223 assert(is_valid_clone_loop_form(loop, peel_list, orig_exit_idx, clone_exit_idx), "bad clone loop"); 4224 4225 Node* head_clone = old_new[head->_idx]; 4226 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop(); 4227 Node* orig_tail_clone = head_clone->in(2); 4228 4229 // Add phi if "def" node is in peel set and "use" is not 4230 4231 for (uint i = 0; i < peel_list.size(); i++) { 4232 Node *def = peel_list.at(i); 4233 if (!def->is_CFG()) { 4234 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 4235 Node *use = def->fast_out(j); 4236 if (has_node(use) && use->in(0) != C->top() && 4237 (!peel.test(use->_idx) || 4238 (use->is_Phi() && use->in(0) == head)) ) { 4239 worklist.push(use); 4240 } 4241 } 4242 while( worklist.size() ) { 4243 Node *use = worklist.pop(); 4244 for (uint j = 1; j < use->req(); j++) { 4245 Node* n = use->in(j); 4246 if (n == def) { 4247 4248 // "def" is in peel set, "use" is not in peel set 4249 // or "use" is in the entry boundary (a phi) of the peel set 4250 4251 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use; 4252 4253 if ( loop->is_member(get_loop( use_c )) ) { 4254 // use is in loop 4255 if (old_new[use->_idx] != nullptr) { // null for dead code 4256 Node* use_clone = old_new[use->_idx]; 4257 _igvn.replace_input_of(use, j, C->top()); 4258 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone ); 4259 } 4260 } else { 4261 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format"); 4262 // use is not in the loop, check if the live range includes the cut 4263 Node* lp_if = use_c->in(orig_exit_idx)->in(0); 4264 if (not_peel.test(lp_if->_idx)) { 4265 assert(j == orig_exit_idx, "use from original loop"); 4266 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone ); 4267 } 4268 } 4269 } 4270 } 4271 } 4272 } 4273 } 4274 4275 // Step 3b: retarget control 4276 4277 // Redirect control to the new loop head if a cloned node in 4278 // the not_peeled region has control that points into the peeled region. 4279 // This necessary because the cloned peeled region will be outside 4280 // the loop. 4281 // from to 4282 // cloned-peeled <---+ 4283 // new_head_clone: | <--+ 4284 // cloned-not_peeled in(0) in(0) 4285 // orig-peeled 4286 4287 for (uint i = 0; i < loop->_body.size(); i++) { 4288 Node *n = loop->_body.at(i); 4289 if (!n->is_CFG() && n->in(0) != nullptr && 4290 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) { 4291 Node* n_clone = old_new[n->_idx]; 4292 if (n_clone->depends_only_on_test()) { 4293 // If this node depends_only_on_test, it will be rewire to the loop head, which is not the 4294 // correct test 4295 Node* pinned_clone = n_clone->pin_node_under_control(); 4296 if (pinned_clone != nullptr) { 4297 register_new_node_with_ctrl_of(pinned_clone, n_clone); 4298 old_new.map(n->_idx, pinned_clone); 4299 _igvn.replace_node(n_clone, pinned_clone); 4300 n_clone = pinned_clone; 4301 } 4302 } 4303 _igvn.replace_input_of(n_clone, 0, new_head_clone); 4304 } 4305 } 4306 4307 // Backedge of the surviving new_head (the clone) is original last_peel 4308 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel); 4309 4310 // Cut first node in original not_peel set 4311 _igvn.rehash_node_delayed(new_head); // Multiple edge updates: 4312 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of 4313 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls 4314 4315 // Copy head_clone back-branch info to original head 4316 // and remove original head's loop entry and 4317 // clone head's back-branch 4318 _igvn.rehash_node_delayed(head); // Multiple edge updates 4319 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl)); 4320 head->set_req(LoopNode::LoopBackControl, C->top()); 4321 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top()); 4322 4323 // Similarly modify the phis 4324 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) { 4325 Node* use = head->fast_out(k); 4326 if (use->is_Phi() && use->outcnt() > 0) { 4327 Node* use_clone = old_new[use->_idx]; 4328 _igvn.rehash_node_delayed(use); // Multiple edge updates 4329 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl)); 4330 use->set_req(LoopNode::LoopBackControl, C->top()); 4331 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top()); 4332 } 4333 } 4334 4335 // Step 4: update dominator tree and dominator depth 4336 4337 set_idom(head, orig_tail_clone, dd); 4338 recompute_dom_depth(); 4339 4340 // Inhibit more partial peeling on this loop 4341 new_head_clone->set_partial_peel_loop(); 4342 C->set_major_progress(); 4343 loop->record_for_igvn(); 4344 4345 #ifndef PRODUCT 4346 if (TracePartialPeeling) { 4347 tty->print_cr("\nafter partial peel one iteration"); 4348 Node_List wl; 4349 Node* t = last_peel; 4350 while (true) { 4351 wl.push(t); 4352 if (t == head_clone) break; 4353 t = idom(t); 4354 } 4355 while (wl.size() > 0) { 4356 Node* tt = wl.pop(); 4357 if (tt == head) tty->print_cr("orig head"); 4358 else if (tt == new_head_clone) tty->print_cr("new head"); 4359 else if (tt == head_clone) tty->print_cr("clone head"); 4360 tt->dump(); 4361 } 4362 } 4363 #endif 4364 4365 C->print_method(PHASE_AFTER_PARTIAL_PEELING, 4, new_head_clone); 4366 4367 return true; 4368 } 4369 4370 #ifdef ASSERT 4371 4372 // Moves Template Assertion Predicates to a target loop by cloning and killing the old ones. The target loop is the 4373 // original, not-cloned loop. This is currently only used with StressLoopBackedge which is a develop flag only and 4374 // false with product builds. We can therefore guard it with an ifdef. More details can be found at the use-site. 4375 class MoveAssertionPredicatesVisitor : public PredicateVisitor { 4376 ClonePredicateToTargetLoop _clone_predicate_to_loop; 4377 PhaseIdealLoop* const _phase; 4378 4379 public: 4380 MoveAssertionPredicatesVisitor(LoopNode* target_loop_head, 4381 const NodeInSingleLoopBody &node_in_loop_body, 4382 PhaseIdealLoop* phase) 4383 : _clone_predicate_to_loop(target_loop_head, node_in_loop_body, phase), 4384 _phase(phase) { 4385 } 4386 NONCOPYABLE(MoveAssertionPredicatesVisitor); 4387 4388 using PredicateVisitor::visit; 4389 4390 void visit(const TemplateAssertionPredicate& template_assertion_predicate) override { 4391 _clone_predicate_to_loop.clone_template_assertion_predicate(template_assertion_predicate); 4392 template_assertion_predicate.kill(_phase->igvn()); 4393 } 4394 }; 4395 #endif // ASSERT 4396 4397 // Transform: 4398 // 4399 // loop<-----------------+ 4400 // | | 4401 // stmt1 stmt2 .. stmtn | 4402 // | | | | 4403 // \ | / | 4404 // v v v | 4405 // region | 4406 // | | 4407 // shared_stmt | 4408 // | | 4409 // v | 4410 // if | 4411 // / \ | 4412 // | -----------+ 4413 // v 4414 // 4415 // into: 4416 // 4417 // loop<-------------------+ 4418 // | | 4419 // v | 4420 // +->loop | 4421 // | | | 4422 // | stmt1 stmt2 .. stmtn | 4423 // | | | | | 4424 // | | \ / | 4425 // | | v v | 4426 // | | region1 | 4427 // | | | | 4428 // | shared_stmt shared_stmt | 4429 // | | | | 4430 // | v v | 4431 // | if if | 4432 // | /\ / \ | 4433 // +-- | | -------+ 4434 // \ / 4435 // v v 4436 // region2 4437 // 4438 // (region2 is shown to merge mirrored projections of the loop exit 4439 // ifs to make the diagram clearer but they really merge the same 4440 // projection) 4441 // 4442 // Conditions for this transformation to trigger: 4443 // - the path through stmt1 is frequent enough 4444 // - the inner loop will be turned into a counted loop after transformation 4445 bool PhaseIdealLoop::duplicate_loop_backedge(IdealLoopTree *loop, Node_List &old_new) { 4446 if (!DuplicateBackedge) { 4447 return false; 4448 } 4449 assert(!loop->_head->is_CountedLoop() || StressDuplicateBackedge, "Non-counted loop only"); 4450 if (!loop->_head->is_Loop()) { 4451 return false; 4452 } 4453 4454 uint estimate = loop->est_loop_clone_sz(1); 4455 if (exceeding_node_budget(estimate)) { 4456 return false; 4457 } 4458 4459 LoopNode *head = loop->_head->as_Loop(); 4460 4461 Node* region = nullptr; 4462 IfNode* exit_test = nullptr; 4463 uint inner; 4464 float f; 4465 #ifdef ASSERT 4466 if (StressDuplicateBackedge) { 4467 if (head->is_strip_mined()) { 4468 return false; 4469 } 4470 Node* c = head->in(LoopNode::LoopBackControl); 4471 4472 while (c != head) { 4473 if (c->is_Region()) { 4474 region = c; 4475 } 4476 c = idom(c); 4477 } 4478 4479 if (region == nullptr) { 4480 return false; 4481 } 4482 4483 inner = 1; 4484 } else 4485 #endif //ASSERT 4486 { 4487 // Is the shape of the loop that of a counted loop... 4488 Node* back_control = loop_exit_control(loop); 4489 if (back_control == nullptr) { 4490 return false; 4491 } 4492 4493 LoopExitTest loop_exit(back_control, loop, this); 4494 loop_exit.build(); 4495 if (!loop_exit.is_valid_with_bt(T_INT)) { 4496 return false; 4497 } 4498 4499 const Node* loop_incr = loop_exit.incr(); 4500 4501 // With an extra phi for the candidate iv? 4502 // Or the region node is the loop head 4503 if (!loop_incr->is_Phi() || loop_incr->in(0) == head) { 4504 return false; 4505 } 4506 4507 PathFrequency pf(head, this); 4508 region = loop_incr->in(0); 4509 4510 // Go over all paths for the extra phi's region and see if that 4511 // path is frequent enough and would match the expected iv shape 4512 // if the extra phi is removed 4513 inner = 0; 4514 for (uint i = 1; i < loop_incr->req(); ++i) { 4515 CountedLoopConverter::TruncatedIncrement increment(T_INT); 4516 increment.build(loop_incr->in(i)); 4517 if (!increment.is_valid()) { 4518 continue; 4519 } 4520 assert(increment.incr()->Opcode() == Op_AddI, "wrong increment code"); 4521 4522 LoopIVStride stride = LoopIVStride(T_INT); 4523 stride.build(increment.incr()); 4524 if (!stride.is_valid()) { 4525 continue; 4526 } 4527 4528 PhiNode* phi = loop_iv_phi(stride.xphi(), nullptr, head); 4529 if (phi == nullptr || 4530 (increment.outer_trunc() == nullptr && phi->in(LoopNode::LoopBackControl) != loop_exit.incr()) || 4531 (increment.outer_trunc() != nullptr && phi->in(LoopNode::LoopBackControl) != increment.outer_trunc())) { 4532 return false; 4533 } 4534 4535 f = pf.to(region->in(i)); 4536 if (f > 0.5) { 4537 inner = i; 4538 break; 4539 } 4540 } 4541 4542 if (inner == 0) { 4543 return false; 4544 } 4545 4546 exit_test = back_control->in(0)->as_If(); 4547 } 4548 4549 if (idom(region)->is_Catch()) { 4550 return false; 4551 } 4552 4553 // Collect all control nodes that need to be cloned (shared_stmt in the diagram) 4554 Unique_Node_List wq; 4555 wq.push(head->in(LoopNode::LoopBackControl)); 4556 for (uint i = 0; i < wq.size(); i++) { 4557 Node* c = wq.at(i); 4558 assert(get_loop(c) == loop, "not in the right loop?"); 4559 if (c->is_Region()) { 4560 if (c != region) { 4561 for (uint j = 1; j < c->req(); ++j) { 4562 wq.push(c->in(j)); 4563 } 4564 } 4565 } else { 4566 wq.push(c->in(0)); 4567 } 4568 assert(!is_strict_dominator(c, region), "shouldn't go above region"); 4569 } 4570 4571 Node* region_dom = idom(region); 4572 4573 // Can't do the transformation if this would cause a membar pair to 4574 // be split 4575 for (uint i = 0; i < wq.size(); i++) { 4576 Node* c = wq.at(i); 4577 if (c->is_MemBar() && (c->as_MemBar()->trailing_store() || c->as_MemBar()->trailing_load_store())) { 4578 assert(c->as_MemBar()->leading_membar()->trailing_membar() == c, "bad membar pair"); 4579 if (!wq.member(c->as_MemBar()->leading_membar())) { 4580 return false; 4581 } 4582 } 4583 } 4584 C->print_method(PHASE_BEFORE_DUPLICATE_LOOP_BACKEDGE, 4, head); 4585 4586 // Collect data nodes that need to be clones as well 4587 int dd = dom_depth(head); 4588 4589 for (uint i = 0; i < loop->_body.size(); ++i) { 4590 Node* n = loop->_body.at(i); 4591 if (has_ctrl(n)) { 4592 Node* c = get_ctrl(n); 4593 if (wq.member(c)) { 4594 wq.push(n); 4595 } 4596 } else { 4597 set_idom(n, idom(n), dd); 4598 } 4599 } 4600 4601 // clone shared_stmt 4602 clone_loop_body(wq, old_new, nullptr); 4603 4604 Node* region_clone = old_new[region->_idx]; 4605 region_clone->set_req(inner, C->top()); 4606 set_idom(region, region->in(inner), dd); 4607 4608 // Prepare the outer loop 4609 Node* outer_head = new LoopNode(head->in(LoopNode::EntryControl), old_new[head->in(LoopNode::LoopBackControl)->_idx]); 4610 register_control(outer_head, loop->_parent, outer_head->in(LoopNode::EntryControl)); 4611 _igvn.replace_input_of(head, LoopNode::EntryControl, outer_head); 4612 set_idom(head, outer_head, dd); 4613 4614 fix_body_edges(wq, loop, old_new, dd, loop->_parent, true); 4615 4616 // Make one of the shared_stmt copies only reachable from stmt1, the 4617 // other only from stmt2..stmtn. 4618 Node* dom = nullptr; 4619 for (uint i = 1; i < region->req(); ++i) { 4620 if (i != inner) { 4621 _igvn.replace_input_of(region, i, C->top()); 4622 } 4623 Node* in = region_clone->in(i); 4624 if (in->is_top()) { 4625 continue; 4626 } 4627 if (dom == nullptr) { 4628 dom = in; 4629 } else { 4630 dom = dom_lca(dom, in); 4631 } 4632 } 4633 4634 set_idom(region_clone, dom, dd); 4635 4636 // Set up the outer loop 4637 for (uint i = 0; i < head->outcnt(); i++) { 4638 Node* u = head->raw_out(i); 4639 if (u->is_Phi()) { 4640 Node* outer_phi = u->clone(); 4641 outer_phi->set_req(0, outer_head); 4642 Node* backedge = old_new[u->in(LoopNode::LoopBackControl)->_idx]; 4643 if (backedge == nullptr) { 4644 backedge = u->in(LoopNode::LoopBackControl); 4645 } 4646 outer_phi->set_req(LoopNode::LoopBackControl, backedge); 4647 register_new_node(outer_phi, outer_head); 4648 _igvn.replace_input_of(u, LoopNode::EntryControl, outer_phi); 4649 } 4650 } 4651 4652 // create control and data nodes for out of loop uses (including region2) 4653 Node_List worklist; 4654 uint new_counter = C->unique(); 4655 fix_ctrl_uses(wq, loop, old_new, ControlAroundStripMined, outer_head, nullptr, worklist); 4656 4657 Node_List *split_if_set = nullptr; 4658 Node_List *split_bool_set = nullptr; 4659 Node_List *split_cex_set = nullptr; 4660 fix_data_uses(wq, loop, ControlAroundStripMined, loop->skip_strip_mined(), new_counter, old_new, worklist, 4661 split_if_set, split_bool_set, split_cex_set); 4662 4663 finish_clone_loop(split_if_set, split_bool_set, split_cex_set); 4664 4665 if (exit_test != nullptr) { 4666 float cnt = exit_test->_fcnt; 4667 if (cnt != COUNT_UNKNOWN) { 4668 exit_test->_fcnt = cnt * f; 4669 old_new[exit_test->_idx]->as_If()->_fcnt = cnt * (1 - f); 4670 } 4671 } 4672 4673 #ifdef ASSERT 4674 if (StressDuplicateBackedge && head->is_CountedLoop()) { 4675 // The Template Assertion Predicates from the old counted loop are now at the new outer loop - clone them to 4676 // the inner counted loop and kill the old ones. We only need to do this with debug builds because 4677 // StressDuplicateBackedge is a devlop flag and false by default. Without StressDuplicateBackedge 'head' will be a 4678 // non-counted loop, and thus we have no Template Assertion Predicates above the old loop to move down. 4679 PredicateIterator predicate_iterator(outer_head->in(LoopNode::EntryControl)); 4680 NodeInSingleLoopBody node_in_body(this, loop); 4681 MoveAssertionPredicatesVisitor move_assertion_predicates_visitor(head, node_in_body, this); 4682 predicate_iterator.for_each(move_assertion_predicates_visitor); 4683 } 4684 #endif // ASSERT 4685 4686 C->set_major_progress(); 4687 4688 C->print_method(PHASE_AFTER_DUPLICATE_LOOP_BACKEDGE, 4, outer_head); 4689 4690 return true; 4691 } 4692 4693 // AutoVectorize the loop: replace scalar ops with vector ops. 4694 PhaseIdealLoop::AutoVectorizeStatus 4695 PhaseIdealLoop::auto_vectorize(IdealLoopTree* lpt, VSharedData &vshared) { 4696 // Counted loop only 4697 if (!lpt->is_counted()) { 4698 return AutoVectorizeStatus::Impossible; 4699 } 4700 4701 // Main-loop only 4702 CountedLoopNode* cl = lpt->_head->as_CountedLoop(); 4703 if (!cl->is_main_loop()) { 4704 return AutoVectorizeStatus::Impossible; 4705 } 4706 4707 VLoop vloop(lpt, false); 4708 if (!vloop.check_preconditions()) { 4709 return AutoVectorizeStatus::TriedAndFailed; 4710 } 4711 4712 // Ensure the shared data is cleared before each use 4713 vshared.clear(); 4714 4715 const VLoopAnalyzer vloop_analyzer(vloop, vshared); 4716 if (!vloop_analyzer.success()) { 4717 return AutoVectorizeStatus::TriedAndFailed; 4718 } 4719 4720 SuperWord sw(vloop_analyzer); 4721 if (!sw.transform_loop()) { 4722 return AutoVectorizeStatus::TriedAndFailed; 4723 } 4724 4725 return AutoVectorizeStatus::Success; 4726 } 4727 4728 // Just before insert_pre_post_loops, we can multiversion the loop: 4729 // 4730 // multiversion_if 4731 // | | 4732 // fast_loop slow_loop 4733 // 4734 // In the fast_loop we can make speculative assumptions, and put the 4735 // conditions into the multiversion_if. If the conditions hold at runtime, 4736 // we enter the fast_loop, if the conditions fail, we take the slow_loop 4737 // instead which does not make any of the speculative assumptions. 4738 // 4739 // Note: we only multiversion the loop if the loop does not have any 4740 // auto vectorization check Predicate. If we have that predicate, 4741 // then we can simply add the speculative assumption checks to 4742 // that Predicate. This means we do not need to duplicate the 4743 // loop - we have a smaller graph and save compile time. Should 4744 // the conditions ever fail, then we deopt / trap at the Predicate 4745 // and recompile without that Predicate. At that point we will 4746 // multiversion the loop, so that we can still have speculative 4747 // runtime checks. 4748 // 4749 // We perform the multiversioning when the loop is still in its single 4750 // iteration form, even before we insert pre and post loops. This makes 4751 // the cloning much simpler. However, this means that both the fast 4752 // and the slow loop have to be optimized independently (adding pre 4753 // and post loops, unrolling the main loop, auto-vectorize etc.). And 4754 // we may end up not needing any speculative assumptions in the fast_loop 4755 // and then rejecting the slow_loop by constant folding the multiversion_if. 4756 // 4757 // Therefore, we "delay" the optimization of the slow_loop until we add 4758 // at least one speculative assumption for the fast_loop. If we never 4759 // add such a speculative runtime check, the OpaqueMultiversioningNode 4760 // of the multiversion_if constant folds to true after loop opts, and the 4761 // multiversion_if folds away the "delayed" slow_loop. If we add any 4762 // speculative assumption, then we notify the OpaqueMultiversioningNode 4763 // with "notify_slow_loop_that_it_can_resume_optimizations". 4764 // 4765 // Note: new runtime checks can be added to the multiversion_if with 4766 // PhaseIdealLoop::create_new_if_for_multiversion 4767 void PhaseIdealLoop::maybe_multiversion_for_auto_vectorization_runtime_checks(IdealLoopTree* lpt, Node_List& old_new) { 4768 CountedLoopNode* cl = lpt->_head->as_CountedLoop(); 4769 LoopNode* outer_loop = cl->skip_strip_mined(); 4770 Node* entry = outer_loop->in(LoopNode::EntryControl); 4771 4772 // Check we have multiversioning enabled, and are not already multiversioned. 4773 if (!LoopMultiversioning || cl->is_multiversion()) { return; } 4774 4775 // Check that we do not have a parse-predicate where we can add the runtime checks 4776 // during auto-vectorization. 4777 const Predicates predicates(entry); 4778 const PredicateBlock* predicate_block = predicates.auto_vectorization_check_block(); 4779 if (predicate_block->has_parse_predicate()) { return; } 4780 4781 // Check node budget. 4782 uint estimate = lpt->est_loop_clone_sz(2); 4783 if (!may_require_nodes(estimate)) { return; } 4784 4785 do_multiversioning(lpt, old_new); 4786 } 4787 4788 void DataNodeGraph::clone_data_nodes(Node* new_ctrl) { 4789 for (uint i = 0; i < _data_nodes.size(); i++) { 4790 clone(_data_nodes[i], new_ctrl); 4791 } 4792 } 4793 4794 // Clone the given node and set it up properly. Set 'new_ctrl' as ctrl. 4795 void DataNodeGraph::clone(Node* node, Node* new_ctrl) { 4796 Node* clone = node->clone(); 4797 _phase->igvn().register_new_node_with_optimizer(clone); 4798 _orig_to_new.put(node, clone); 4799 _phase->set_ctrl(clone, new_ctrl); 4800 if (node->is_CastII()) { 4801 clone->set_req(0, new_ctrl); 4802 } 4803 } 4804 4805 // Rewire the data inputs of all (unprocessed) cloned nodes, whose inputs are still pointing to the same inputs as their 4806 // corresponding orig nodes, to the newly cloned inputs to create a separate cloned graph. 4807 void DataNodeGraph::rewire_clones_to_cloned_inputs() { 4808 _orig_to_new.iterate_all([&](Node* node, Node* clone) { 4809 for (uint i = 1; i < node->req(); i++) { 4810 Node** cloned_input = _orig_to_new.get(node->in(i)); 4811 if (cloned_input != nullptr) { 4812 // Input was also cloned -> rewire clone to the cloned input. 4813 _phase->igvn().replace_input_of(clone, i, *cloned_input); 4814 } 4815 } 4816 }); 4817 } 4818 4819 // Clone all non-OpaqueLoop* nodes and apply the provided transformation strategy for OpaqueLoop* nodes. 4820 // Set 'new_ctrl' as ctrl for all cloned non-OpaqueLoop* nodes. 4821 void DataNodeGraph::clone_data_nodes_and_transform_opaque_loop_nodes( 4822 const TransformStrategyForOpaqueLoopNodes& transform_strategy, 4823 Node* new_ctrl) { 4824 for (uint i = 0; i < _data_nodes.size(); i++) { 4825 Node* data_node = _data_nodes[i]; 4826 if (data_node->is_Opaque1()) { 4827 transform_opaque_node(transform_strategy, data_node); 4828 } else { 4829 clone(data_node, new_ctrl); 4830 } 4831 } 4832 } 4833 4834 void DataNodeGraph::transform_opaque_node(const TransformStrategyForOpaqueLoopNodes& transform_strategy, Node* node) { 4835 Node* transformed_node; 4836 if (node->is_OpaqueLoopInit()) { 4837 transformed_node = transform_strategy.transform_opaque_init(node->as_OpaqueLoopInit()); 4838 } else { 4839 assert(node->is_OpaqueLoopStride(), "must be OpaqueLoopStrideNode"); 4840 transformed_node = transform_strategy.transform_opaque_stride(node->as_OpaqueLoopStride()); 4841 } 4842 // Add an orig->new mapping to correctly update the inputs of the copied graph in rewire_clones_to_cloned_inputs(). 4843 _orig_to_new.put(node, transformed_node); 4844 } --- EOF ---