1 /* 2 * Copyright (c) 1999, 2025, 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); 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); 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 pin_array_access_nodes) { 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, pin_array_access_nodes); 397 } 398 399 void PhaseIdealLoop::rewire_safe_outputs_to_dominator(Node* source, Node* dominator, const bool pin_array_access_nodes) { 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 // Do not rewire Div and Mod nodes which could have a zero divisor to avoid skipping their zero check. 405 if (out->depends_only_on_test() && _igvn.no_dependent_zero_check(out)) { 406 assert(out->in(0) == source, "must be control dependent on source"); 407 _igvn.replace_input_of(out, 0, dominator); 408 if (pin_array_access_nodes) { 409 // Because of Loop Predication, Loads and range check Cast nodes that are control dependent on this range 410 // check (that is about to be removed) now depend on multiple dominating Hoisted Check Predicates. After the 411 // removal of this range check, these control dependent nodes end up at the lowest/nearest dominating predicate 412 // in the graph. To ensure that these Loads/Casts do not float above any of the dominating checks (even when the 413 // lowest dominating check is later replaced by yet another dominating check), we need to pin them at the lowest 414 // dominating check. 415 Node* clone = out->pin_array_access_node(); 416 if (clone != nullptr) { 417 clone = _igvn.register_new_node_with_optimizer(clone, out); 418 _igvn.replace_node(out, clone); 419 out = clone; 420 } 421 } 422 set_early_ctrl(out, false); 423 IdealLoopTree* new_loop = get_loop(get_ctrl(out)); 424 if (old_loop != new_loop) { 425 if (!old_loop->_child) { 426 old_loop->_body.yank(out); 427 } 428 if (!new_loop->_child) { 429 new_loop->_body.push(out); 430 } 431 } 432 --i; 433 --imax; 434 } 435 } 436 } 437 438 //------------------------------has_local_phi_input---------------------------- 439 // Return TRUE if 'n' has Phi inputs from its local block and no other 440 // block-local inputs (all non-local-phi inputs come from earlier blocks) 441 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) { 442 Node *n_ctrl = get_ctrl(n); 443 // See if some inputs come from a Phi in this block, or from before 444 // this block. 445 uint i; 446 for( i = 1; i < n->req(); i++ ) { 447 Node *phi = n->in(i); 448 if( phi->is_Phi() && phi->in(0) == n_ctrl ) 449 break; 450 } 451 if( i >= n->req() ) 452 return nullptr; // No Phi inputs; nowhere to clone thru 453 454 // Check for inputs created between 'n' and the Phi input. These 455 // must split as well; they have already been given the chance 456 // (courtesy of a post-order visit) and since they did not we must 457 // recover the 'cost' of splitting them by being very profitable 458 // when splitting 'n'. Since this is unlikely we simply give up. 459 for( i = 1; i < n->req(); i++ ) { 460 Node *m = n->in(i); 461 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) { 462 // We allow the special case of AddP's with no local inputs. 463 // This allows us to split-up address expressions. 464 if (m->is_AddP() && 465 get_ctrl(m->in(AddPNode::Base)) != n_ctrl && 466 get_ctrl(m->in(AddPNode::Address)) != n_ctrl && 467 get_ctrl(m->in(AddPNode::Offset)) != n_ctrl) { 468 // Move the AddP up to the dominating point. That's fine because control of m's inputs 469 // must dominate get_ctrl(m) == n_ctrl and we just checked that the input controls are != n_ctrl. 470 Node* c = find_non_split_ctrl(idom(n_ctrl)); 471 if (c->is_OuterStripMinedLoop()) { 472 c->as_Loop()->verify_strip_mined(1); 473 c = c->in(LoopNode::EntryControl); 474 } 475 set_ctrl_and_loop(m, c); 476 continue; 477 } 478 return nullptr; 479 } 480 assert(n->is_Phi() || m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control"); 481 } 482 483 return n_ctrl; 484 } 485 486 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2). 487 Node* PhaseIdealLoop::remix_address_expressions_add_left_shift(Node* n, IdealLoopTree* n_loop, Node* n_ctrl, BasicType bt) { 488 assert(bt == T_INT || bt == T_LONG, "only for integers"); 489 int n_op = n->Opcode(); 490 491 if (n_op == Op_LShift(bt)) { 492 // Scale is loop invariant 493 Node* scale = n->in(2); 494 Node* scale_ctrl = get_ctrl(scale); 495 IdealLoopTree* scale_loop = get_loop(scale_ctrl); 496 if (n_loop == scale_loop || !scale_loop->is_member(n_loop)) { 497 return nullptr; 498 } 499 const TypeInt* scale_t = scale->bottom_type()->isa_int(); 500 if (scale_t != nullptr && scale_t->is_con() && scale_t->get_con() >= 16) { 501 return nullptr; // Dont bother with byte/short masking 502 } 503 // Add must vary with loop (else shift would be loop-invariant) 504 Node* add = n->in(1); 505 Node* add_ctrl = get_ctrl(add); 506 IdealLoopTree* add_loop = get_loop(add_ctrl); 507 if (n_loop != add_loop) { 508 return nullptr; // happens w/ evil ZKM loops 509 } 510 511 // Convert I-V into I+ (0-V); same for V-I 512 if (add->Opcode() == Op_Sub(bt) && 513 _igvn.type(add->in(1)) != TypeInteger::zero(bt)) { 514 assert(add->Opcode() == Op_SubI || add->Opcode() == Op_SubL, ""); 515 Node* zero = integercon(0, bt); 516 Node* neg = SubNode::make(zero, add->in(2), bt); 517 register_new_node_with_ctrl_of(neg, add->in(2)); 518 add = AddNode::make(add->in(1), neg, bt); 519 register_new_node(add, add_ctrl); 520 } 521 if (add->Opcode() != Op_Add(bt)) return nullptr; 522 assert(add->Opcode() == Op_AddI || add->Opcode() == Op_AddL, ""); 523 // See if one add input is loop invariant 524 Node* add_var = add->in(1); 525 Node* add_var_ctrl = get_ctrl(add_var); 526 IdealLoopTree* add_var_loop = get_loop(add_var_ctrl); 527 Node* add_invar = add->in(2); 528 Node* add_invar_ctrl = get_ctrl(add_invar); 529 IdealLoopTree* add_invar_loop = get_loop(add_invar_ctrl); 530 if (add_invar_loop == n_loop) { 531 // Swap to find the invariant part 532 add_invar = add_var; 533 add_invar_ctrl = add_var_ctrl; 534 add_invar_loop = add_var_loop; 535 add_var = add->in(2); 536 } else if (add_var_loop != n_loop) { // Else neither input is loop invariant 537 return nullptr; 538 } 539 if (n_loop == add_invar_loop || !add_invar_loop->is_member(n_loop)) { 540 return nullptr; // No invariant part of the add? 541 } 542 543 // Yes! Reshape address expression! 544 Node* inv_scale = LShiftNode::make(add_invar, scale, bt); 545 Node* inv_scale_ctrl = 546 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ? 547 add_invar_ctrl : scale_ctrl; 548 register_new_node(inv_scale, inv_scale_ctrl); 549 Node* var_scale = LShiftNode::make(add_var, scale, bt); 550 register_new_node(var_scale, n_ctrl); 551 Node* var_add = AddNode::make(var_scale, inv_scale, bt); 552 register_new_node(var_add, n_ctrl); 553 _igvn.replace_node(n, var_add); 554 return var_add; 555 } 556 return nullptr; 557 } 558 559 //------------------------------remix_address_expressions---------------------- 560 // Rework addressing expressions to get the most loop-invariant stuff 561 // moved out. We'd like to do all associative operators, but it's especially 562 // important (common) to do address expressions. 563 Node* PhaseIdealLoop::remix_address_expressions(Node* n) { 564 if (!has_ctrl(n)) return nullptr; 565 Node* n_ctrl = get_ctrl(n); 566 IdealLoopTree* n_loop = get_loop(n_ctrl); 567 568 // See if 'n' mixes loop-varying and loop-invariant inputs and 569 // itself is loop-varying. 570 571 // Only interested in binary ops (and AddP) 572 if (n->req() < 3 || n->req() > 4) return nullptr; 573 574 Node* n1_ctrl = get_ctrl(n->in( 1)); 575 Node* n2_ctrl = get_ctrl(n->in( 2)); 576 Node* n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3)); 577 IdealLoopTree* n1_loop = get_loop(n1_ctrl); 578 IdealLoopTree* n2_loop = get_loop(n2_ctrl); 579 IdealLoopTree* n3_loop = get_loop(n3_ctrl); 580 581 // Does one of my inputs spin in a tighter loop than self? 582 if ((n_loop->is_member(n1_loop) && n_loop != n1_loop) || 583 (n_loop->is_member(n2_loop) && n_loop != n2_loop) || 584 (n_loop->is_member(n3_loop) && n_loop != n3_loop)) { 585 return nullptr; // Leave well enough alone 586 } 587 588 // Is at least one of my inputs loop-invariant? 589 if (n1_loop == n_loop && 590 n2_loop == n_loop && 591 n3_loop == n_loop) { 592 return nullptr; // No loop-invariant inputs 593 } 594 595 Node* res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_INT); 596 if (res != nullptr) { 597 return res; 598 } 599 res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_LONG); 600 if (res != nullptr) { 601 return res; 602 } 603 604 int n_op = n->Opcode(); 605 // Replace (I+V) with (V+I) 606 if (n_op == Op_AddI || 607 n_op == Op_AddL || 608 n_op == Op_AddF || 609 n_op == Op_AddD || 610 n_op == Op_MulI || 611 n_op == Op_MulL || 612 n_op == Op_MulF || 613 n_op == Op_MulD) { 614 if (n2_loop == n_loop) { 615 assert(n1_loop != n_loop, ""); 616 n->swap_edges(1, 2); 617 } 618 } 619 620 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V), 621 // but not if I2 is a constant. Skip for irreducible loops. 622 if (n_op == Op_AddP && n_loop->_head->is_Loop()) { 623 if (n2_loop == n_loop && n3_loop != n_loop) { 624 if (n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con()) { 625 Node* n22_ctrl = get_ctrl(n->in(2)->in(2)); 626 Node* n23_ctrl = get_ctrl(n->in(2)->in(3)); 627 IdealLoopTree* n22loop = get_loop(n22_ctrl); 628 IdealLoopTree* n23_loop = get_loop(n23_ctrl); 629 if (n22loop != n_loop && n22loop->is_member(n_loop) && 630 n23_loop == n_loop) { 631 Node* add1 = new AddPNode(n->in(1), n->in(2)->in(2), n->in(3)); 632 // Stuff new AddP in the loop preheader 633 register_new_node(add1, n_loop->_head->as_Loop()->skip_strip_mined(1)->in(LoopNode::EntryControl)); 634 Node* add2 = new AddPNode(n->in(1), add1, n->in(2)->in(3)); 635 register_new_node(add2, n_ctrl); 636 _igvn.replace_node(n, add2); 637 return add2; 638 } 639 } 640 } 641 642 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V) 643 if (n2_loop != n_loop && n3_loop == n_loop) { 644 if (n->in(3)->Opcode() == Op_AddX) { 645 Node* V = n->in(3)->in(1); 646 Node* I = n->in(3)->in(2); 647 if (ctrl_is_member(n_loop, V)) { 648 } else { 649 Node *tmp = V; V = I; I = tmp; 650 } 651 if (!ctrl_is_member(n_loop, I)) { 652 Node* add1 = new AddPNode(n->in(1), n->in(2), I); 653 // Stuff new AddP in the loop preheader 654 register_new_node(add1, n_loop->_head->as_Loop()->skip_strip_mined(1)->in(LoopNode::EntryControl)); 655 Node* add2 = new AddPNode(n->in(1), add1, V); 656 register_new_node(add2, n_ctrl); 657 _igvn.replace_node(n, add2); 658 return add2; 659 } 660 } 661 } 662 } 663 664 return nullptr; 665 } 666 667 // Optimize ((in1[2*i] * in2[2*i]) + (in1[2*i+1] * in2[2*i+1])) 668 Node *PhaseIdealLoop::convert_add_to_muladd(Node* n) { 669 assert(n->Opcode() == Op_AddI, "sanity"); 670 Node * nn = nullptr; 671 Node * in1 = n->in(1); 672 Node * in2 = n->in(2); 673 if (in1->Opcode() == Op_MulI && in2->Opcode() == Op_MulI) { 674 IdealLoopTree* loop_n = get_loop(get_ctrl(n)); 675 if (loop_n->is_counted() && 676 loop_n->_head->as_Loop()->is_valid_counted_loop(T_INT) && 677 Matcher::match_rule_supported(Op_MulAddVS2VI) && 678 Matcher::match_rule_supported(Op_MulAddS2I)) { 679 Node* mul_in1 = in1->in(1); 680 Node* mul_in2 = in1->in(2); 681 Node* mul_in3 = in2->in(1); 682 Node* mul_in4 = in2->in(2); 683 if (mul_in1->Opcode() == Op_LoadS && 684 mul_in2->Opcode() == Op_LoadS && 685 mul_in3->Opcode() == Op_LoadS && 686 mul_in4->Opcode() == Op_LoadS) { 687 IdealLoopTree* loop1 = get_loop(get_ctrl(mul_in1)); 688 IdealLoopTree* loop2 = get_loop(get_ctrl(mul_in2)); 689 IdealLoopTree* loop3 = get_loop(get_ctrl(mul_in3)); 690 IdealLoopTree* loop4 = get_loop(get_ctrl(mul_in4)); 691 IdealLoopTree* loop5 = get_loop(get_ctrl(in1)); 692 IdealLoopTree* loop6 = get_loop(get_ctrl(in2)); 693 // All nodes should be in the same counted loop. 694 if (loop_n == loop1 && loop_n == loop2 && loop_n == loop3 && 695 loop_n == loop4 && loop_n == loop5 && loop_n == loop6) { 696 Node* adr1 = mul_in1->in(MemNode::Address); 697 Node* adr2 = mul_in2->in(MemNode::Address); 698 Node* adr3 = mul_in3->in(MemNode::Address); 699 Node* adr4 = mul_in4->in(MemNode::Address); 700 if (adr1->is_AddP() && adr2->is_AddP() && adr3->is_AddP() && adr4->is_AddP()) { 701 if ((adr1->in(AddPNode::Base) == adr3->in(AddPNode::Base)) && 702 (adr2->in(AddPNode::Base) == adr4->in(AddPNode::Base))) { 703 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in3, mul_in4); 704 register_new_node_with_ctrl_of(nn, n); 705 _igvn.replace_node(n, nn); 706 return nn; 707 } else if ((adr1->in(AddPNode::Base) == adr4->in(AddPNode::Base)) && 708 (adr2->in(AddPNode::Base) == adr3->in(AddPNode::Base))) { 709 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in4, mul_in3); 710 register_new_node_with_ctrl_of(nn, n); 711 _igvn.replace_node(n, nn); 712 return nn; 713 } 714 } 715 } 716 } 717 } 718 } 719 return nn; 720 } 721 722 //------------------------------conditional_move------------------------------- 723 // Attempt to replace a Phi with a conditional move. We have some pretty 724 // strict profitability requirements. All Phis at the merge point must 725 // be converted, so we can remove the control flow. We need to limit the 726 // number of c-moves to a small handful. All code that was in the side-arms 727 // of the CFG diamond is now speculatively executed. This code has to be 728 // "cheap enough". We are pretty much limited to CFG diamonds that merge 729 // 1 or 2 items with a total of 1 or 2 ops executed speculatively. 730 Node *PhaseIdealLoop::conditional_move( Node *region ) { 731 732 assert(region->is_Region(), "sanity check"); 733 if (region->req() != 3) return nullptr; 734 735 // Check for CFG diamond 736 Node *lp = region->in(1); 737 Node *rp = region->in(2); 738 if (!lp || !rp) return nullptr; 739 Node *lp_c = lp->in(0); 740 if (lp_c == nullptr || lp_c != rp->in(0) || !lp_c->is_If()) return nullptr; 741 IfNode *iff = lp_c->as_If(); 742 743 // Check for ops pinned in an arm of the diamond. 744 // Can't remove the control flow in this case 745 if (lp->outcnt() > 1) return nullptr; 746 if (rp->outcnt() > 1) return nullptr; 747 748 IdealLoopTree* r_loop = get_loop(region); 749 assert(r_loop == get_loop(iff), "sanity"); 750 // Always convert to CMOVE if all results are used only outside this loop. 751 bool used_inside_loop = (r_loop == _ltree_root); 752 753 // Check profitability 754 int cost = 0; 755 int phis = 0; 756 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 757 Node *out = region->fast_out(i); 758 if (!out->is_Phi()) continue; // Ignore other control edges, etc 759 phis++; 760 PhiNode* phi = out->as_Phi(); 761 BasicType bt = phi->type()->basic_type(); 762 switch (bt) { 763 case T_DOUBLE: 764 case T_FLOAT: 765 if (C->use_cmove()) { 766 continue; //TODO: maybe we want to add some cost 767 } 768 cost += Matcher::float_cmove_cost(); // Could be very expensive 769 break; 770 case T_LONG: { 771 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's 772 } 773 case T_INT: // These all CMOV fine 774 case T_ADDRESS: { // (RawPtr) 775 cost++; 776 break; 777 } 778 case T_NARROWOOP: // Fall through 779 case T_OBJECT: { // Base oops are OK, but not derived oops 780 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr(); 781 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a 782 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus 783 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we 784 // have a Phi for the base here that we convert to a CMOVE all is well 785 // and good. But if the base is dead, we'll not make a CMOVE. Later 786 // the allocator will have to produce a base by creating a CMOVE of the 787 // relevant bases. This puts the allocator in the business of 788 // manufacturing expensive instructions, generally a bad plan. 789 // Just Say No to Conditionally-Moved Derived Pointers. 790 if (tp && tp->offset() != 0) 791 return nullptr; 792 cost++; 793 break; 794 } 795 default: 796 return nullptr; // In particular, can't do memory or I/O 797 } 798 // Add in cost any speculative ops 799 for (uint j = 1; j < region->req(); j++) { 800 Node *proj = region->in(j); 801 Node *inp = phi->in(j); 802 if (inp->isa_InlineType()) { 803 // TODO 8302217 This prevents PhiNode::push_inline_types_through 804 return nullptr; 805 } 806 if (get_ctrl(inp) == proj) { // Found local op 807 cost++; 808 // Check for a chain of dependent ops; these will all become 809 // speculative in a CMOV. 810 for (uint k = 1; k < inp->req(); k++) 811 if (get_ctrl(inp->in(k)) == proj) 812 cost += ConditionalMoveLimit; // Too much speculative goo 813 } 814 } 815 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode. 816 // This will likely Split-If, a higher-payoff operation. 817 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) { 818 Node* use = phi->fast_out(k); 819 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr()) 820 cost += ConditionalMoveLimit; 821 // Is there a use inside the loop? 822 // Note: check only basic types since CMoveP is pinned. 823 if (!used_inside_loop && is_java_primitive(bt)) { 824 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use); 825 if (r_loop == u_loop || r_loop->is_member(u_loop)) { 826 used_inside_loop = true; 827 } 828 } 829 } 830 }//for 831 Node* bol = iff->in(1); 832 assert(!bol->is_OpaqueInitializedAssertionPredicate(), "Initialized Assertion Predicates cannot form a diamond with Halt"); 833 if (bol->is_OpaqueTemplateAssertionPredicate()) { 834 // Ignore Template Assertion Predicates with OpaqueTemplateAssertionPredicate nodes. 835 return nullptr; 836 } 837 if (bol->is_OpaqueMultiversioning()) { 838 assert(bol->as_OpaqueMultiversioning()->is_useless(), "Must be useless, i.e. fast main loop has already disappeared."); 839 // Ignore multiversion_if that just lost its loops. The OpaqueMultiversioning is marked useless, 840 // and will make the multiversion_if constant fold in the next IGVN round. 841 return nullptr; 842 } 843 if (!bol->is_Bool()) { 844 assert(false, "Expected Bool, but got %s", NodeClassNames[bol->Opcode()]); 845 return nullptr; 846 } 847 int cmp_op = bol->in(1)->Opcode(); 848 if (cmp_op == Op_SubTypeCheck) { // SubTypeCheck expansion expects an IfNode 849 return nullptr; 850 } 851 // It is expensive to generate flags from a float compare. 852 // Avoid duplicated float compare. 853 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return nullptr; 854 855 float infrequent_prob = PROB_UNLIKELY_MAG(3); 856 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop. 857 if (used_inside_loop) { 858 if (cost >= ConditionalMoveLimit) return nullptr; // Too much goo 859 860 // BlockLayoutByFrequency optimization moves infrequent branch 861 // from hot path. No point in CMOV'ing in such case (110 is used 862 // instead of 100 to take into account not exactness of float value). 863 if (BlockLayoutByFrequency) { 864 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f); 865 } 866 } 867 // Check for highly predictable branch. No point in CMOV'ing if 868 // we are going to predict accurately all the time. 869 if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) { 870 //keep going 871 } else if (iff->_prob < infrequent_prob || 872 iff->_prob > (1.0f - infrequent_prob)) 873 return nullptr; 874 875 // -------------- 876 // Now replace all Phis with CMOV's 877 Node *cmov_ctrl = iff->in(0); 878 uint flip = (lp->Opcode() == Op_IfTrue); 879 Node_List wq; 880 while (1) { 881 PhiNode* phi = nullptr; 882 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 883 Node *out = region->fast_out(i); 884 if (out->is_Phi()) { 885 phi = out->as_Phi(); 886 break; 887 } 888 } 889 if (phi == nullptr || _igvn.type(phi) == Type::TOP || !CMoveNode::supported(_igvn.type(phi))) { 890 break; 891 } 892 // Move speculative ops 893 wq.push(phi); 894 while (wq.size() > 0) { 895 Node *n = wq.pop(); 896 for (uint j = 1; j < n->req(); j++) { 897 Node* m = n->in(j); 898 if (m != nullptr && !is_dominator(get_ctrl(m), cmov_ctrl)) { 899 set_ctrl(m, cmov_ctrl); 900 wq.push(m); 901 } 902 } 903 } 904 Node* cmov = CMoveNode::make(iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi)); 905 register_new_node(cmov, cmov_ctrl); 906 _igvn.replace_node(phi, cmov); 907 #ifndef PRODUCT 908 if (TraceLoopOpts) { 909 tty->print("CMOV "); 910 r_loop->dump_head(); 911 if (Verbose) { 912 bol->in(1)->dump(1); 913 cmov->dump(1); 914 } 915 } 916 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } ); 917 #endif 918 } 919 920 // The useless CFG diamond will fold up later; see the optimization in 921 // RegionNode::Ideal. 922 _igvn._worklist.push(region); 923 924 return iff->in(1); 925 } 926 927 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) { 928 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 929 Node* u = m->fast_out(i); 930 if (u->is_CFG()) { 931 if (u->is_NeverBranch()) { 932 u = u->as_NeverBranch()->proj_out(0); 933 enqueue_cfg_uses(u, wq); 934 } else { 935 wq.push(u); 936 } 937 } 938 } 939 } 940 941 // Try moving a store out of a loop, right before the loop 942 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) { 943 // Store has to be first in the loop body 944 IdealLoopTree *n_loop = get_loop(n_ctrl); 945 if (n->is_Store() && n_loop != _ltree_root && 946 n_loop->is_loop() && n_loop->_head->is_Loop() && 947 n->in(0) != nullptr) { 948 Node* address = n->in(MemNode::Address); 949 Node* value = n->in(MemNode::ValueIn); 950 Node* mem = n->in(MemNode::Memory); 951 952 // - address and value must be loop invariant 953 // - memory must be a memory Phi for the loop 954 // - Store must be the only store on this memory slice in the 955 // loop: if there's another store following this one then value 956 // written at iteration i by the second store could be overwritten 957 // at iteration i+n by the first store: it's not safe to move the 958 // first store out of the loop 959 // - nothing must observe the memory Phi: it guarantees no read 960 // before the store, we are also guaranteed the store post 961 // dominates the loop head (ignoring a possible early 962 // exit). Otherwise there would be extra Phi involved between the 963 // loop's Phi and the store. 964 // - there must be no early exit from the loop before the Store 965 // (such an exit most of the time would be an extra use of the 966 // memory Phi but sometimes is a bottom memory Phi that takes the 967 // store as input). 968 969 if (!ctrl_is_member(n_loop, address) && 970 !ctrl_is_member(n_loop, value) && 971 mem->is_Phi() && mem->in(0) == n_loop->_head && 972 mem->outcnt() == 1 && 973 mem->in(LoopNode::LoopBackControl) == n) { 974 975 assert(n_loop->_tail != nullptr, "need a tail"); 976 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop"); 977 978 // Verify that there's no early exit of the loop before the store. 979 bool ctrl_ok = false; 980 { 981 // Follow control from loop head until n, we exit the loop or 982 // we reach the tail 983 ResourceMark rm; 984 Unique_Node_List wq; 985 wq.push(n_loop->_head); 986 987 for (uint next = 0; next < wq.size(); ++next) { 988 Node *m = wq.at(next); 989 if (m == n->in(0)) { 990 ctrl_ok = true; 991 continue; 992 } 993 assert(!has_ctrl(m), "should be CFG"); 994 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) { 995 ctrl_ok = false; 996 break; 997 } 998 enqueue_cfg_uses(m, wq); 999 if (wq.size() > 10) { 1000 ctrl_ok = false; 1001 break; 1002 } 1003 } 1004 } 1005 if (ctrl_ok) { 1006 // move the Store 1007 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem); 1008 _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl)); 1009 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl)); 1010 // Disconnect the phi now. An empty phi can confuse other 1011 // optimizations in this pass of loop opts. 1012 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl)); 1013 n_loop->_body.yank(mem); 1014 1015 set_ctrl_and_loop(n, n->in(0)); 1016 1017 return n; 1018 } 1019 } 1020 } 1021 return nullptr; 1022 } 1023 1024 // Try moving a store out of a loop, right after the loop 1025 void PhaseIdealLoop::try_move_store_after_loop(Node* n) { 1026 if (n->is_Store() && n->in(0) != nullptr) { 1027 Node *n_ctrl = get_ctrl(n); 1028 IdealLoopTree *n_loop = get_loop(n_ctrl); 1029 // Store must be in a loop 1030 if (n_loop != _ltree_root && !n_loop->_irreducible) { 1031 Node* address = n->in(MemNode::Address); 1032 Node* value = n->in(MemNode::ValueIn); 1033 // address must be loop invariant 1034 if (!ctrl_is_member(n_loop, address)) { 1035 // Store must be last on this memory slice in the loop and 1036 // nothing in the loop must observe it 1037 Node* phi = nullptr; 1038 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1039 Node* u = n->fast_out(i); 1040 if (has_ctrl(u)) { // control use? 1041 if (!ctrl_is_member(n_loop, u)) { 1042 continue; 1043 } 1044 if (u->is_Phi() && u->in(0) == n_loop->_head) { 1045 assert(_igvn.type(u) == Type::MEMORY, "bad phi"); 1046 // multiple phis on the same slice are possible 1047 if (phi != nullptr) { 1048 return; 1049 } 1050 phi = u; 1051 continue; 1052 } 1053 } 1054 return; 1055 } 1056 if (phi != nullptr) { 1057 // Nothing in the loop before the store (next iteration) 1058 // must observe the stored value 1059 bool mem_ok = true; 1060 { 1061 ResourceMark rm; 1062 Unique_Node_List wq; 1063 wq.push(phi); 1064 for (uint next = 0; next < wq.size() && mem_ok; ++next) { 1065 Node *m = wq.at(next); 1066 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) { 1067 Node* u = m->fast_out(i); 1068 if (u->is_Store() || u->is_Phi()) { 1069 if (u != n) { 1070 wq.push(u); 1071 mem_ok = (wq.size() <= 10); 1072 } 1073 } else { 1074 mem_ok = false; 1075 break; 1076 } 1077 } 1078 } 1079 } 1080 if (mem_ok) { 1081 // Move the store out of the loop if the LCA of all 1082 // users (except for the phi) is outside the loop. 1083 Node* hook = new Node(1); 1084 hook->init_req(0, n_ctrl); // Add an input to prevent hook from being dead 1085 _igvn.rehash_node_delayed(phi); 1086 int count = phi->replace_edge(n, hook, &_igvn); 1087 assert(count > 0, "inconsistent phi"); 1088 1089 // Compute latest point this store can go 1090 Node* lca = get_late_ctrl(n, get_ctrl(n)); 1091 if (lca->is_OuterStripMinedLoop()) { 1092 lca = lca->in(LoopNode::EntryControl); 1093 } 1094 if (n_loop->is_member(get_loop(lca))) { 1095 // LCA is in the loop - bail out 1096 _igvn.replace_node(hook, n); 1097 return; 1098 } 1099 #ifdef ASSERT 1100 if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) { 1101 assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined"); 1102 n_loop->_head->as_Loop()->verify_strip_mined(1); 1103 Node* outer = n_loop->_head->as_CountedLoop()->outer_loop(); 1104 IdealLoopTree* outer_loop = get_loop(outer); 1105 assert(n_loop->_parent == outer_loop, "broken loop tree"); 1106 assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state"); 1107 } 1108 #endif 1109 lca = place_outside_loop(lca, n_loop); 1110 assert(!n_loop->is_member(get_loop(lca)), "control must not be back in the loop"); 1111 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"); 1112 1113 // Move store out of the loop 1114 _igvn.replace_node(hook, n->in(MemNode::Memory)); 1115 _igvn.replace_input_of(n, 0, lca); 1116 set_ctrl_and_loop(n, lca); 1117 1118 // Disconnect the phi now. An empty phi can confuse other 1119 // optimizations in this pass of loop opts.. 1120 if (phi->in(LoopNode::LoopBackControl) == phi) { 1121 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl)); 1122 n_loop->_body.yank(phi); 1123 } 1124 } 1125 } 1126 } 1127 } 1128 } 1129 } 1130 1131 // We can't use immutable memory for the flat array check because we are loading the mark word which is 1132 // mutable. Although the bits we are interested in are immutable (we check for markWord::unlocked_value), 1133 // we need to use raw memory to not break anti dependency analysis. Below code will attempt to still move 1134 // flat array checks out of loops, mainly to enable loop unswitching. 1135 void PhaseIdealLoop::move_flat_array_check_out_of_loop(Node* n) { 1136 // Skip checks for more than one array 1137 if (n->req() > 3) { 1138 return; 1139 } 1140 Node* mem = n->in(FlatArrayCheckNode::Memory); 1141 Node* array = n->in(FlatArrayCheckNode::ArrayOrKlass)->uncast(); 1142 IdealLoopTree* check_loop = get_loop(get_ctrl(n)); 1143 IdealLoopTree* ary_loop = get_loop(get_ctrl(array)); 1144 1145 // Check if array is loop invariant 1146 if (!check_loop->is_member(ary_loop)) { 1147 // Walk up memory graph from the check until we leave the loop 1148 VectorSet wq; 1149 wq.set(mem->_idx); 1150 while (check_loop->is_member(get_loop(ctrl_or_self(mem)))) { 1151 if (mem->is_Phi()) { 1152 mem = mem->in(1); 1153 } else if (mem->is_MergeMem()) { 1154 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw); 1155 } else if (mem->is_Proj()) { 1156 mem = mem->in(0); 1157 } else if (mem->is_MemBar() || mem->is_SafePoint()) { 1158 mem = mem->in(TypeFunc::Memory); 1159 } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) { 1160 mem = mem->in(MemNode::Memory); 1161 } else { 1162 #ifdef ASSERT 1163 mem->dump(); 1164 #endif 1165 ShouldNotReachHere(); 1166 } 1167 if (wq.test_set(mem->_idx)) { 1168 return; 1169 } 1170 } 1171 // Replace memory input and re-compute ctrl to move the check out of the loop 1172 _igvn.replace_input_of(n, 1, mem); 1173 set_ctrl_and_loop(n, get_early_ctrl(n)); 1174 Node* bol = n->unique_out(); 1175 set_ctrl_and_loop(bol, get_early_ctrl(bol)); 1176 } 1177 } 1178 1179 //------------------------------split_if_with_blocks_pre----------------------- 1180 // Do the real work in a non-recursive function. Data nodes want to be 1181 // cloned in the pre-order so they can feed each other nicely. 1182 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) { 1183 // Cloning these guys is unlikely to win 1184 int n_op = n->Opcode(); 1185 if (n_op == Op_MergeMem) { 1186 return n; 1187 } 1188 if (n->is_Proj()) { 1189 return n; 1190 } 1191 1192 if (n->isa_FlatArrayCheck()) { 1193 move_flat_array_check_out_of_loop(n); 1194 return n; 1195 } 1196 1197 // Do not clone-up CmpFXXX variations, as these are always 1198 // followed by a CmpI 1199 if (n->is_Cmp()) { 1200 return n; 1201 } 1202 // Attempt to use a conditional move instead of a phi/branch 1203 if (ConditionalMoveLimit > 0 && n_op == Op_Region) { 1204 Node *cmov = conditional_move( n ); 1205 if (cmov) { 1206 return cmov; 1207 } 1208 } 1209 if (n->is_CFG() || n->is_LoadStore()) { 1210 return n; 1211 } 1212 if (n->is_Opaque1()) { // Opaque nodes cannot be mod'd 1213 if (!C->major_progress()) { // If chance of no more loop opts... 1214 _igvn._worklist.push(n); // maybe we'll remove them 1215 } 1216 return n; 1217 } 1218 1219 if (n->is_Con()) { 1220 return n; // No cloning for Con nodes 1221 } 1222 1223 Node *n_ctrl = get_ctrl(n); 1224 if (!n_ctrl) { 1225 return n; // Dead node 1226 } 1227 1228 Node* res = try_move_store_before_loop(n, n_ctrl); 1229 if (res != nullptr) { 1230 return n; 1231 } 1232 1233 // Attempt to remix address expressions for loop invariants 1234 Node *m = remix_address_expressions( n ); 1235 if( m ) return m; 1236 1237 if (n_op == Op_AddI) { 1238 Node *nn = convert_add_to_muladd( n ); 1239 if ( nn ) return nn; 1240 } 1241 1242 if (n->is_ConstraintCast() && n->as_ConstraintCast()->dependency().narrows_type()) { 1243 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this); 1244 // ConstraintCastNode::dominating_cast() uses node control input to determine domination. 1245 // Node control inputs don't necessarily agree with loop control info (due to 1246 // transformations happened in between), thus additional dominance check is needed 1247 // to keep loop info valid. 1248 if (dom_cast != nullptr && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) { 1249 _igvn.replace_node(n, dom_cast); 1250 return dom_cast; 1251 } 1252 } 1253 1254 // Determine if the Node has inputs from some local Phi. 1255 // Returns the block to clone thru. 1256 Node *n_blk = has_local_phi_input( n ); 1257 if( !n_blk ) return n; 1258 1259 // Do not clone the trip counter through on a CountedLoop 1260 // (messes up the canonical shape). 1261 if (((n_blk->is_CountedLoop() || (n_blk->is_Loop() && n_blk->as_Loop()->is_loop_nest_inner_loop())) && n->Opcode() == Op_AddI) || 1262 (n_blk->is_LongCountedLoop() && n->Opcode() == Op_AddL)) { 1263 return n; 1264 } 1265 // Pushing a shift through the iv Phi can get in the way of addressing optimizations or range check elimination 1266 if (n_blk->is_BaseCountedLoop() && n->Opcode() == Op_LShift(n_blk->as_BaseCountedLoop()->bt()) && 1267 n->in(1) == n_blk->as_BaseCountedLoop()->phi()) { 1268 return n; 1269 } 1270 1271 // Check for having no control input; not pinned. Allow 1272 // dominating control. 1273 if (n->in(0)) { 1274 Node *dom = idom(n_blk); 1275 if (dom_lca(n->in(0), dom) != n->in(0)) { 1276 return n; 1277 } 1278 } 1279 // Policy: when is it profitable. You must get more wins than 1280 // policy before it is considered profitable. Policy is usually 0, 1281 // so 1 win is considered profitable. Big merges will require big 1282 // cloning, so get a larger policy. 1283 int policy = n_blk->req() >> 2; 1284 1285 // If the loop is a candidate for range check elimination, 1286 // delay splitting through it's phi until a later loop optimization 1287 if (n_blk->is_BaseCountedLoop()) { 1288 IdealLoopTree *lp = get_loop(n_blk); 1289 if (lp && lp->_rce_candidate) { 1290 return n; 1291 } 1292 } 1293 1294 if (must_throttle_split_if()) return n; 1295 1296 // Split 'n' through the merge point if it is profitable, replacing it with a new phi. 1297 Node* phi = split_thru_phi(n, n_blk, policy); 1298 if (phi == nullptr) { return n; } 1299 1300 // Moved a load around the loop, 'en-registering' something. 1301 if (n_blk->is_Loop() && n->is_Load() && 1302 !phi->in(LoopNode::LoopBackControl)->is_Load()) 1303 C->set_major_progress(); 1304 1305 return phi; 1306 } 1307 1308 static bool merge_point_too_heavy(Compile* C, Node* region) { 1309 // Bail out if the region and its phis have too many users. 1310 int weight = 0; 1311 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1312 weight += region->fast_out(i)->outcnt(); 1313 } 1314 int nodes_left = C->max_node_limit() - C->live_nodes(); 1315 if (weight * 8 > nodes_left) { 1316 if (PrintOpto) { 1317 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight); 1318 } 1319 return true; 1320 } else { 1321 return false; 1322 } 1323 } 1324 1325 static bool merge_point_safe(Node* region) { 1326 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode 1327 // having a PhiNode input. This sidesteps the dangerous case where the split 1328 // ConvI2LNode may become TOP if the input Value() does not 1329 // overlap the ConvI2L range, leaving a node which may not dominate its 1330 // uses. 1331 // A better fix for this problem can be found in the BugTraq entry, but 1332 // expediency for Mantis demands this hack. 1333 #ifdef _LP64 1334 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1335 Node* n = region->fast_out(i); 1336 if (n->is_Phi()) { 1337 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1338 Node* m = n->fast_out(j); 1339 if (m->Opcode() == Op_ConvI2L) 1340 return false; 1341 if (m->is_CastII()) { 1342 return false; 1343 } 1344 } 1345 } 1346 } 1347 #endif 1348 return true; 1349 } 1350 1351 1352 //------------------------------place_outside_loop--------------------------------- 1353 // Place some computation outside of this loop on the path to the use passed as argument 1354 Node* PhaseIdealLoop::place_outside_loop(Node* useblock, IdealLoopTree* loop) const { 1355 Node* head = loop->_head; 1356 assert(!loop->is_member(get_loop(useblock)), "must be outside loop"); 1357 if (head->is_Loop() && head->as_Loop()->is_strip_mined()) { 1358 loop = loop->_parent; 1359 assert(loop->_head->is_OuterStripMinedLoop(), "malformed strip mined loop"); 1360 } 1361 1362 // Pick control right outside the loop 1363 for (;;) { 1364 Node* dom = idom(useblock); 1365 if (loop->is_member(get_loop(dom))) { 1366 break; 1367 } 1368 useblock = dom; 1369 } 1370 assert(find_non_split_ctrl(useblock) == useblock, "should be non split control"); 1371 return useblock; 1372 } 1373 1374 1375 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) { 1376 if (!n->is_If() || n->is_BaseCountedLoopEnd()) { 1377 return false; 1378 } 1379 if (!n->in(0)->is_Region()) { 1380 return false; 1381 } 1382 1383 Node* region = n->in(0); 1384 Node* dom = idom(region); 1385 if (!dom->is_If() || !n->as_If()->same_condition(dom, &_igvn)) { 1386 return false; 1387 } 1388 IfNode* dom_if = dom->as_If(); 1389 IfTrueNode* proj_true = dom_if->true_proj(); 1390 IfFalseNode* proj_false = dom_if->false_proj(); 1391 1392 for (uint i = 1; i < region->req(); i++) { 1393 if (is_dominator(proj_true, region->in(i))) { 1394 continue; 1395 } 1396 if (is_dominator(proj_false, region->in(i))) { 1397 continue; 1398 } 1399 return false; 1400 } 1401 1402 return true; 1403 } 1404 1405 1406 bool PhaseIdealLoop::can_split_if(Node* n_ctrl) { 1407 if (must_throttle_split_if()) { 1408 return false; 1409 } 1410 1411 // Do not do 'split-if' if irreducible loops are present. 1412 if (_has_irreducible_loops) { 1413 return false; 1414 } 1415 1416 if (merge_point_too_heavy(C, n_ctrl)) { 1417 return false; 1418 } 1419 1420 // Do not do 'split-if' if some paths are dead. First do dead code 1421 // elimination and then see if its still profitable. 1422 for (uint i = 1; i < n_ctrl->req(); i++) { 1423 if (n_ctrl->in(i) == C->top()) { 1424 return false; 1425 } 1426 } 1427 1428 // If trying to do a 'Split-If' at the loop head, it is only 1429 // profitable if the cmp folds up on BOTH paths. Otherwise we 1430 // risk peeling a loop forever. 1431 1432 // CNC - Disabled for now. Requires careful handling of loop 1433 // body selection for the cloned code. Also, make sure we check 1434 // for any input path not being in the same loop as n_ctrl. For 1435 // irreducible loops we cannot check for 'n_ctrl->is_Loop()' 1436 // because the alternative loop entry points won't be converted 1437 // into LoopNodes. 1438 IdealLoopTree *n_loop = get_loop(n_ctrl); 1439 for (uint j = 1; j < n_ctrl->req(); j++) { 1440 if (get_loop(n_ctrl->in(j)) != n_loop) { 1441 return false; 1442 } 1443 } 1444 1445 // Check for safety of the merge point. 1446 if (!merge_point_safe(n_ctrl)) { 1447 return false; 1448 } 1449 1450 return true; 1451 } 1452 1453 // Detect if the node is the inner strip-mined loop 1454 // Return: null if it's not the case, or the exit of outer strip-mined loop 1455 static Node* is_inner_of_stripmined_loop(const Node* out) { 1456 Node* out_le = nullptr; 1457 1458 if (out->is_CountedLoopEnd()) { 1459 const CountedLoopNode* loop = out->as_CountedLoopEnd()->loopnode(); 1460 1461 if (loop != nullptr && loop->is_strip_mined()) { 1462 out_le = loop->in(LoopNode::EntryControl)->as_OuterStripMinedLoop()->outer_loop_exit(); 1463 } 1464 } 1465 1466 return out_le; 1467 } 1468 1469 bool PhaseIdealLoop::flat_array_element_type_check(Node *n) { 1470 // If the CmpP is a subtype check for a value that has just been 1471 // loaded from an array, the subtype check guarantees the value 1472 // can't be stored in a flat array and the load of the value 1473 // happens with a flat array check then: push the type check 1474 // through the phi of the flat array check. This needs special 1475 // logic because the subtype check's input is not a phi but a 1476 // LoadKlass that must first be cloned through the phi. 1477 if (n->Opcode() != Op_CmpP) { 1478 return false; 1479 } 1480 1481 Node* klassptr = n->in(1); 1482 Node* klasscon = n->in(2); 1483 1484 if (klassptr->is_DecodeNarrowPtr()) { 1485 klassptr = klassptr->in(1); 1486 } 1487 1488 if (klassptr->Opcode() != Op_LoadKlass && klassptr->Opcode() != Op_LoadNKlass) { 1489 return false; 1490 } 1491 1492 if (!klasscon->is_Con()) { 1493 return false; 1494 } 1495 1496 Node* addr = klassptr->in(MemNode::Address); 1497 1498 if (!addr->is_AddP()) { 1499 return false; 1500 } 1501 1502 intptr_t offset; 1503 Node* obj = AddPNode::Ideal_base_and_offset(addr, &_igvn, offset); 1504 1505 if (obj == nullptr) { 1506 return false; 1507 } 1508 1509 // TODO 8378077: The code below does not work anymore with off-heap accesses which set their bases to top with 1510 // JDK-8373343. Also: flat_array_element_type_check() was introduced with JDK-8228622 for a specific check to enable 1511 // split-if but JDK-8245729 changed how that check looks like. Is it still relevant? This should be revisited. 1512 if (addr->in(AddPNode::Base)->is_top()) { 1513 return false; 1514 } 1515 1516 if (obj->Opcode() == Op_CastPP) { 1517 obj = obj->in(1); 1518 } 1519 1520 if (!obj->is_Phi()) { 1521 return false; 1522 } 1523 1524 Node* region = obj->in(0); 1525 1526 Node* phi = PhiNode::make_blank(region, n->in(1)); 1527 for (uint i = 1; i < region->req(); i++) { 1528 Node* in = obj->in(i); 1529 Node* ctrl = region->in(i); 1530 if (addr->in(AddPNode::Base) != obj) { 1531 Node* cast = addr->in(AddPNode::Base); 1532 assert(cast->Opcode() == Op_CastPP && cast->in(0) != nullptr, "inconsistent subgraph"); 1533 Node* cast_clone = cast->clone(); 1534 cast_clone->set_req(0, ctrl); 1535 cast_clone->set_req(1, in); 1536 register_new_node(cast_clone, ctrl); 1537 const Type* tcast = cast_clone->Value(&_igvn); 1538 _igvn.set_type(cast_clone, tcast); 1539 cast_clone->as_Type()->set_type(tcast); 1540 in = cast_clone; 1541 } 1542 Node* addr_clone = addr->clone(); 1543 addr_clone->set_req(AddPNode::Base, in); 1544 addr_clone->set_req(AddPNode::Address, in); 1545 register_new_node(addr_clone, ctrl); 1546 _igvn.set_type(addr_clone, addr_clone->Value(&_igvn)); 1547 Node* klassptr_clone = klassptr->clone(); 1548 klassptr_clone->set_req(2, addr_clone); 1549 register_new_node(klassptr_clone, ctrl); 1550 _igvn.set_type(klassptr_clone, klassptr_clone->Value(&_igvn)); 1551 if (klassptr != n->in(1)) { 1552 Node* decode = n->in(1); 1553 assert(decode->is_DecodeNarrowPtr(), "inconsistent subgraph"); 1554 Node* decode_clone = decode->clone(); 1555 decode_clone->set_req(1, klassptr_clone); 1556 register_new_node(decode_clone, ctrl); 1557 _igvn.set_type(decode_clone, decode_clone->Value(&_igvn)); 1558 klassptr_clone = decode_clone; 1559 } 1560 phi->set_req(i, klassptr_clone); 1561 } 1562 register_new_node(phi, region); 1563 Node* orig = n->in(1); 1564 _igvn.replace_input_of(n, 1, phi); 1565 split_if_with_blocks_post(n); 1566 if (n->outcnt() != 0) { 1567 _igvn.replace_input_of(n, 1, orig); 1568 _igvn.remove_dead_node(phi); 1569 } 1570 return true; 1571 } 1572 1573 //------------------------------split_if_with_blocks_post---------------------- 1574 // Do the real work in a non-recursive function. CFG hackery wants to be 1575 // in the post-order, so it can dirty the I-DOM info and not use the dirtied 1576 // info. 1577 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) { 1578 1579 if (flat_array_element_type_check(n)) { 1580 return; 1581 } 1582 1583 // Cloning Cmp through Phi's involves the split-if transform. 1584 // FastLock is not used by an If 1585 if (n->is_Cmp() && !n->is_FastLock()) { 1586 Node *n_ctrl = get_ctrl(n); 1587 // Determine if the Node has inputs from some local Phi. 1588 // Returns the block to clone thru. 1589 Node *n_blk = has_local_phi_input(n); 1590 if (n_blk != n_ctrl) { 1591 return; 1592 } 1593 1594 if (!can_split_if(n_ctrl)) { 1595 return; 1596 } 1597 1598 if (n->outcnt() != 1) { 1599 return; // Multiple bool's from 1 compare? 1600 } 1601 Node *bol = n->unique_out(); 1602 assert(bol->is_Bool(), "expect a bool here"); 1603 if (bol->outcnt() != 1) { 1604 return;// Multiple branches from 1 compare? 1605 } 1606 Node *iff = bol->unique_out(); 1607 1608 // Check some safety conditions 1609 if (iff->is_If()) { // Classic split-if? 1610 if (iff->in(0) != n_ctrl) { 1611 return; // Compare must be in same blk as if 1612 } 1613 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE 1614 // Can't split CMove with different control. 1615 if (get_ctrl(iff) != n_ctrl) { 1616 return; 1617 } 1618 if (get_ctrl(iff->in(2)) == n_ctrl || 1619 get_ctrl(iff->in(3)) == n_ctrl) { 1620 return; // Inputs not yet split-up 1621 } 1622 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) { 1623 return; // Loop-invar test gates loop-varying CMOVE 1624 } 1625 } else { 1626 return; // some other kind of node, such as an Allocate 1627 } 1628 1629 // When is split-if profitable? Every 'win' on means some control flow 1630 // goes dead, so it's almost always a win. 1631 int policy = 0; 1632 // Split compare 'n' through the merge point if it is profitable 1633 Node *phi = split_thru_phi( n, n_ctrl, policy); 1634 if (!phi) { 1635 return; 1636 } 1637 1638 // Now split the bool up thru the phi 1639 Node* bolphi = split_thru_phi(bol, n_ctrl, -1); 1640 guarantee(bolphi != nullptr, "null boolean phi node"); 1641 assert(iff->in(1) == bolphi, ""); 1642 1643 if (bolphi->Value(&_igvn)->singleton()) { 1644 return; 1645 } 1646 1647 // Conditional-move? Must split up now 1648 if (!iff->is_If()) { 1649 Node* cmovphi = split_thru_phi(iff, n_ctrl, -1); 1650 return; 1651 } 1652 1653 // Now split the IF 1654 C->print_method(PHASE_BEFORE_SPLIT_IF, 4, iff); 1655 if (TraceLoopOpts) { 1656 tty->print_cr("Split-If"); 1657 } 1658 do_split_if(iff); 1659 C->print_method(PHASE_AFTER_SPLIT_IF, 4, iff); 1660 return; 1661 } 1662 1663 // Two identical ifs back to back can be merged 1664 if (try_merge_identical_ifs(n)) { 1665 return; 1666 } 1667 1668 // Check for an IF ready to split; one that has its 1669 // condition codes input coming from a Phi at the block start. 1670 int n_op = n->Opcode(); 1671 1672 // Check for an IF being dominated by another IF same test 1673 if (n_op == Op_If || 1674 n_op == Op_RangeCheck) { 1675 Node *bol = n->in(1); 1676 uint max = bol->outcnt(); 1677 // Check for same test used more than once? 1678 if (bol->is_Bool() && (max > 1 || bol->in(1)->is_SubTypeCheck())) { 1679 // Search up IDOMs to see if this IF is dominated. 1680 Node* cmp = bol->in(1); 1681 Node *cutoff = cmp->is_SubTypeCheck() ? dom_lca(get_ctrl(cmp->in(1)), get_ctrl(cmp->in(2))) : get_ctrl(bol); 1682 1683 // Now search up IDOMs till cutoff, looking for a dominating test 1684 Node *prevdom = n; 1685 Node *dom = idom(prevdom); 1686 while (dom != cutoff) { 1687 if (dom->req() > 1 && n->as_If()->same_condition(dom, &_igvn) && prevdom->in(0) == dom && 1688 safe_for_if_replacement(dom)) { 1689 // It's invalid to move control dependent data nodes in the inner 1690 // strip-mined loop, because: 1691 // 1) break validation of LoopNode::verify_strip_mined() 1692 // 2) move code with side-effect in strip-mined loop 1693 // Move to the exit of outer strip-mined loop in that case. 1694 Node* out_le = is_inner_of_stripmined_loop(dom); 1695 if (out_le != nullptr) { 1696 prevdom = out_le; 1697 } 1698 // Replace the dominated test with an obvious true or false. 1699 // Place it on the IGVN worklist for later cleanup. 1700 C->set_major_progress(); 1701 // Split if: pin array accesses that are control dependent on a range check and moved to a regular if, 1702 // to prevent an array load from floating above its range check. There are three cases: 1703 // 1. Move from RangeCheck "a" to RangeCheck "b": don't need to pin. If we ever remove b, then we pin 1704 // all its array accesses at that point. 1705 // 2. We move from RangeCheck "a" to regular if "b": need to pin. If we ever remove b, then its array 1706 // accesses would start to float, since we don't pin at that point. 1707 // 3. If we move from regular if: don't pin. All array accesses are already assumed to be pinned. 1708 bool pin_array_access_nodes = n->Opcode() == Op_RangeCheck && 1709 prevdom->in(0)->Opcode() != Op_RangeCheck; 1710 dominated_by(prevdom->as_IfProj(), n->as_If(), false, pin_array_access_nodes); 1711 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } ); 1712 return; 1713 } 1714 prevdom = dom; 1715 dom = idom(prevdom); 1716 } 1717 } 1718 } 1719 1720 try_sink_out_of_loop(n); 1721 if (C->failing()) { 1722 return; 1723 } 1724 1725 try_move_store_after_loop(n); 1726 1727 // Remove multiple allocations of the same inline type 1728 if (n->is_InlineType()) { 1729 n->as_InlineType()->remove_redundant_allocations(this); 1730 } 1731 } 1732 1733 // Transform: 1734 // 1735 // if (some_condition) { 1736 // // body 1 1737 // } else { 1738 // // body 2 1739 // } 1740 // if (some_condition) { 1741 // // body 3 1742 // } else { 1743 // // body 4 1744 // } 1745 // 1746 // into: 1747 // 1748 // 1749 // if (some_condition) { 1750 // // body 1 1751 // // body 3 1752 // } else { 1753 // // body 2 1754 // // body 4 1755 // } 1756 bool PhaseIdealLoop::try_merge_identical_ifs(Node* n) { 1757 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) { 1758 Node *n_ctrl = n->in(0); 1759 IfNode* dom_if = idom(n_ctrl)->as_If(); 1760 if (n->in(1) != dom_if->in(1)) { 1761 assert(n->in(1)->in(1)->is_SubTypeCheck() && 1762 (n->in(1)->in(1)->as_SubTypeCheck()->method() != nullptr || 1763 dom_if->in(1)->in(1)->as_SubTypeCheck()->method() != nullptr), "only for subtype checks with profile data attached"); 1764 _igvn.replace_input_of(n, 1, dom_if->in(1)); 1765 } 1766 IfTrueNode* dom_proj_true = dom_if->true_proj(); 1767 IfFalseNode* dom_proj_false = dom_if->false_proj(); 1768 1769 // Now split the IF 1770 RegionNode* new_false_region; 1771 RegionNode* new_true_region; 1772 do_split_if(n, &new_false_region, &new_true_region); 1773 assert(new_false_region->req() == new_true_region->req(), ""); 1774 #ifdef ASSERT 1775 for (uint i = 1; i < new_false_region->req(); ++i) { 1776 assert(new_false_region->in(i)->in(0) == new_true_region->in(i)->in(0), "unexpected shape following split if"); 1777 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"); 1778 } 1779 #endif 1780 assert(new_false_region->in(1)->in(0)->in(1) == dom_if->in(1), "dominating if and dominated if after split must share test"); 1781 1782 // We now have: 1783 // if (some_condition) { 1784 // // body 1 1785 // if (some_condition) { 1786 // body3: // new_true_region 1787 // // body3 1788 // } else { 1789 // goto body4; 1790 // } 1791 // } else { 1792 // // body 2 1793 // if (some_condition) { 1794 // goto body3; 1795 // } else { 1796 // body4: // new_false_region 1797 // // body4; 1798 // } 1799 // } 1800 // 1801 1802 // clone pinned nodes thru the resulting regions 1803 push_pinned_nodes_thru_region(dom_if, new_true_region); 1804 push_pinned_nodes_thru_region(dom_if, new_false_region); 1805 1806 // Optimize out the cloned ifs. Because pinned nodes were cloned, this also allows a CastPP that would be dependent 1807 // 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 1808 // unrelated control dependency. 1809 for (uint i = 1; i < new_false_region->req(); i++) { 1810 if (is_dominator(dom_proj_true, new_false_region->in(i))) { 1811 dominated_by(dom_proj_true, new_false_region->in(i)->in(0)->as_If()); 1812 } else { 1813 assert(is_dominator(dom_proj_false, new_false_region->in(i)), "bad if"); 1814 dominated_by(dom_proj_false, new_false_region->in(i)->in(0)->as_If()); 1815 } 1816 } 1817 return true; 1818 } 1819 return false; 1820 } 1821 1822 void PhaseIdealLoop::push_pinned_nodes_thru_region(IfNode* dom_if, Node* region) { 1823 for (DUIterator i = region->outs(); region->has_out(i); i++) { 1824 Node* u = region->out(i); 1825 if (!has_ctrl(u) || u->is_Phi() || !u->depends_only_on_test() || !_igvn.no_dependent_zero_check(u)) { 1826 continue; 1827 } 1828 assert(u->in(0) == region, "not a control dependent node?"); 1829 uint j = 1; 1830 for (; j < u->req(); ++j) { 1831 Node* in = u->in(j); 1832 if (!is_dominator(ctrl_or_self(in), dom_if)) { 1833 break; 1834 } 1835 } 1836 if (j == u->req()) { 1837 Node *phi = PhiNode::make_blank(region, u); 1838 for (uint k = 1; k < region->req(); ++k) { 1839 Node* clone = u->clone(); 1840 clone->set_req(0, region->in(k)); 1841 register_new_node(clone, region->in(k)); 1842 phi->init_req(k, clone); 1843 } 1844 register_new_node(phi, region); 1845 _igvn.replace_node(u, phi); 1846 --i; 1847 } 1848 } 1849 } 1850 1851 bool PhaseIdealLoop::safe_for_if_replacement(const Node* dom) const { 1852 if (!dom->is_CountedLoopEnd()) { 1853 return true; 1854 } 1855 CountedLoopEndNode* le = dom->as_CountedLoopEnd(); 1856 CountedLoopNode* cl = le->loopnode(); 1857 if (cl == nullptr) { 1858 return true; 1859 } 1860 if (!cl->is_main_loop()) { 1861 return true; 1862 } 1863 if (cl->is_canonical_loop_entry() == nullptr) { 1864 return true; 1865 } 1866 // Further unrolling is possible so loop exit condition might change 1867 return false; 1868 } 1869 1870 // See if a shared loop-varying computation has no loop-varying uses. 1871 // Happens if something is only used for JVM state in uncommon trap exits, 1872 // like various versions of induction variable+offset. Clone the 1873 // computation per usage to allow it to sink out of the loop. 1874 void PhaseIdealLoop::try_sink_out_of_loop(Node* n) { 1875 bool is_raw_to_oop_cast = n->is_ConstraintCast() && 1876 n->in(1)->bottom_type()->isa_rawptr() && 1877 !n->bottom_type()->isa_rawptr(); 1878 1879 if (has_ctrl(n) && 1880 !n->is_Phi() && 1881 !n->is_Bool() && 1882 !n->is_Proj() && 1883 !n->is_MergeMem() && 1884 !n->is_CMove() && 1885 !n->is_OpaqueNotNull() && 1886 !n->is_OpaqueInitializedAssertionPredicate() && 1887 !n->is_OpaqueTemplateAssertionPredicate() && 1888 !is_raw_to_oop_cast && // don't extend live ranges of raw oops 1889 n->Opcode() != Op_CreateEx && 1890 (KillPathsReachableByDeadTypeNode || !n->is_Type()) 1891 ) { 1892 Node *n_ctrl = get_ctrl(n); 1893 IdealLoopTree *n_loop = get_loop(n_ctrl); 1894 1895 if (n->in(0) != nullptr) { 1896 IdealLoopTree* loop_ctrl = get_loop(n->in(0)); 1897 if (n_loop != loop_ctrl && n_loop->is_member(loop_ctrl)) { 1898 // n has a control input inside a loop but get_ctrl() is member of an outer loop. This could happen, for example, 1899 // for Div nodes inside a loop (control input inside loop) without a use except for an UCT (outside the loop). 1900 // Rewire control of n to right outside of the loop, regardless if its input(s) are later sunk or not. 1901 Node* maybe_pinned_n = n; 1902 Node* outside_ctrl = place_outside_loop(n_ctrl, loop_ctrl); 1903 if (!would_sink_below_pre_loop_exit(loop_ctrl, outside_ctrl)) { 1904 if (n->depends_only_on_test()) { 1905 Node* pinned_clone = n->pin_array_access_node(); 1906 if (pinned_clone != nullptr) { 1907 // Pin array access nodes: if this is an array load, it's going to be dependent on a condition that's not a 1908 // range check for that access. If that condition is replaced by an identical dominating one, then an 1909 // unpinned load would risk floating above its range check. 1910 register_new_node(pinned_clone, n_ctrl); 1911 maybe_pinned_n = pinned_clone; 1912 _igvn.replace_node(n, pinned_clone); 1913 } 1914 } 1915 _igvn.replace_input_of(maybe_pinned_n, 0, outside_ctrl); 1916 } 1917 } 1918 } 1919 if (n_loop != _ltree_root && n->outcnt() > 1) { 1920 // Compute early control: needed for anti-dependence analysis. It's also possible that as a result of 1921 // previous transformations in this loop opts round, the node can be hoisted now: early control will tell us. 1922 Node* early_ctrl = compute_early_ctrl(n, n_ctrl); 1923 if (n_loop->is_member(get_loop(early_ctrl)) && // check that this one can't be hoisted now 1924 ctrl_of_all_uses_out_of_loop(n, early_ctrl, n_loop)) { // All uses in outer loops! 1925 if (n->is_Store() || n->is_LoadStore()) { 1926 assert(false, "no node with a side effect"); 1927 C->record_failure("no node with a side effect"); 1928 return; 1929 } 1930 Node* outer_loop_clone = nullptr; 1931 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin;) { 1932 Node* u = n->last_out(j); // Clone private computation per use 1933 _igvn.rehash_node_delayed(u); 1934 Node* x = nullptr; 1935 if (n->depends_only_on_test()) { 1936 // Pin array access nodes: if this is an array load, it's going to be dependent on a condition that's not a 1937 // range check for that access. If that condition is replaced by an identical dominating one, then an 1938 // unpinned load would risk floating above its range check. 1939 x = n->pin_array_access_node(); 1940 } 1941 if (x == nullptr) { 1942 x = n->clone(); 1943 } 1944 Node* x_ctrl = nullptr; 1945 if (u->is_Phi()) { 1946 // Replace all uses of normal nodes. Replace Phi uses 1947 // individually, so the separate Nodes can sink down 1948 // different paths. 1949 uint k = 1; 1950 while (u->in(k) != n) k++; 1951 u->set_req(k, x); 1952 // x goes next to Phi input path 1953 x_ctrl = u->in(0)->in(k); 1954 // Find control for 'x' next to use but not inside inner loops. 1955 x_ctrl = place_outside_loop(x_ctrl, n_loop); 1956 --j; 1957 } else { // Normal use 1958 if (has_ctrl(u)) { 1959 x_ctrl = get_ctrl(u); 1960 } else { 1961 x_ctrl = u->in(0); 1962 } 1963 // Find control for 'x' next to use but not inside inner loops. 1964 x_ctrl = place_outside_loop(x_ctrl, n_loop); 1965 // Replace all uses 1966 if (u->is_ConstraintCast() && _igvn.type(n)->higher_equal(u->bottom_type()) && u->in(0) == x_ctrl) { 1967 // If we're sinking a chain of data nodes, we might have inserted a cast to pin the use which is not necessary 1968 // anymore now that we're going to pin n as well 1969 _igvn.replace_node(u, x); 1970 --j; 1971 } else { 1972 int nb = u->replace_edge(n, x, &_igvn); 1973 j -= nb; 1974 } 1975 } 1976 1977 if (n->is_Load()) { 1978 // For loads, add a control edge to a CFG node outside of the loop 1979 // to force them to not combine and return back inside the loop 1980 // during GVN optimization (4641526). 1981 assert(x_ctrl == get_late_ctrl_with_anti_dep(x->as_Load(), early_ctrl, x_ctrl), "anti-dependences were already checked"); 1982 1983 IdealLoopTree* x_loop = get_loop(x_ctrl); 1984 Node* x_head = x_loop->_head; 1985 if (x_head->is_Loop() && x_head->is_OuterStripMinedLoop()) { 1986 // Do not add duplicate LoadNodes to the outer strip mined loop 1987 if (outer_loop_clone != nullptr) { 1988 _igvn.replace_node(x, outer_loop_clone); 1989 continue; 1990 } 1991 outer_loop_clone = x; 1992 } 1993 x->set_req(0, x_ctrl); 1994 } else if (n->in(0) != nullptr){ 1995 x->set_req(0, x_ctrl); 1996 } 1997 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone"); 1998 assert(!n_loop->is_member(get_loop(x_ctrl)), "should have moved out of loop"); 1999 register_new_node(x, x_ctrl); 2000 2001 // Chain of AddP nodes: (AddP base (AddP base (AddP base ))) 2002 // All AddP nodes must keep the same base after sinking so: 2003 // 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, 2004 // their bases remain the same. 2005 // (see 2- below) 2006 assert(!x->is_AddP() || !x->in(AddPNode::Address)->is_AddP() || 2007 x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base) || 2008 !x->in(AddPNode::Address)->in(AddPNode::Base)->eqv_uncast(x->in(AddPNode::Base)), "unexpected AddP shape"); 2009 if (x->in(0) == nullptr && !x->is_DecodeNarrowPtr() && 2010 !(x->is_AddP() && x->in(AddPNode::Address)->is_AddP() && x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base))) { 2011 assert(!x->is_Load(), "load should be pinned"); 2012 // Use a cast node to pin clone out of loop 2013 Node* cast = nullptr; 2014 for (uint k = 0; k < x->req(); k++) { 2015 Node* in = x->in(k); 2016 if (in != nullptr && ctrl_is_member(n_loop, in)) { 2017 const Type* in_t = _igvn.type(in); 2018 cast = ConstraintCastNode::make_cast_for_type(x_ctrl, in, in_t, 2019 ConstraintCastNode::DependencyType::NonFloatingNonNarrowing, nullptr); 2020 } 2021 if (cast != nullptr) { 2022 Node* prev = _igvn.hash_find_insert(cast); 2023 if (prev != nullptr && get_ctrl(prev) == x_ctrl) { 2024 cast->destruct(&_igvn); 2025 cast = prev; 2026 } else { 2027 register_new_node(cast, x_ctrl); 2028 } 2029 x->replace_edge(in, cast); 2030 // Chain of AddP nodes: 2031 // 2- A CastPP of the base is only added now that all AddP nodes are sunk 2032 if (x->is_AddP() && k == AddPNode::Base) { 2033 update_addp_chain_base(x, n->in(AddPNode::Base), cast); 2034 } 2035 break; 2036 } 2037 } 2038 assert(cast != nullptr, "must have added a cast to pin the node"); 2039 } 2040 } 2041 _igvn.remove_dead_node(n); 2042 } 2043 _dom_lca_tags_round = 0; 2044 } 2045 } 2046 } 2047 2048 void PhaseIdealLoop::update_addp_chain_base(Node* x, Node* old_base, Node* new_base) { 2049 ResourceMark rm; 2050 Node_List wq; 2051 wq.push(x); 2052 while (wq.size() != 0) { 2053 Node* n = wq.pop(); 2054 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2055 Node* u = n->fast_out(i); 2056 if (u->is_AddP() && u->in(AddPNode::Base) == old_base) { 2057 _igvn.replace_input_of(u, AddPNode::Base, new_base); 2058 wq.push(u); 2059 } 2060 } 2061 } 2062 } 2063 2064 // Compute the early control of a node by following its inputs until we reach 2065 // nodes that are pinned. Then compute the LCA of the control of all pinned nodes. 2066 Node* PhaseIdealLoop::compute_early_ctrl(Node* n, Node* n_ctrl) { 2067 Node* early_ctrl = nullptr; 2068 ResourceMark rm; 2069 Unique_Node_List wq; 2070 wq.push(n); 2071 for (uint i = 0; i < wq.size(); i++) { 2072 Node* m = wq.at(i); 2073 Node* c = nullptr; 2074 if (m->is_CFG()) { 2075 c = m; 2076 } else if (m->pinned()) { 2077 c = m->in(0); 2078 } else { 2079 for (uint j = 0; j < m->req(); j++) { 2080 Node* in = m->in(j); 2081 if (in != nullptr) { 2082 wq.push(in); 2083 } 2084 } 2085 } 2086 if (c != nullptr) { 2087 assert(is_dominator(c, n_ctrl), "control input must dominate current control"); 2088 if (early_ctrl == nullptr || is_dominator(early_ctrl, c)) { 2089 early_ctrl = c; 2090 } 2091 } 2092 } 2093 assert(is_dominator(early_ctrl, n_ctrl), "early control must dominate current control"); 2094 return early_ctrl; 2095 } 2096 2097 bool PhaseIdealLoop::ctrl_of_all_uses_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop) { 2098 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2099 Node* u = n->fast_out(i); 2100 if (u->is_Opaque1()) { 2101 return false; // Found loop limit, bugfix for 4677003 2102 } 2103 if (u->is_Phi()) { 2104 for (uint j = 1; j < u->req(); ++j) { 2105 if (u->in(j) == n && !ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, u->in(0)->in(j))) { 2106 return false; 2107 } 2108 } 2109 } else { 2110 Node* ctrl = has_ctrl(u) ? get_ctrl(u) : u->in(0); 2111 if (!ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, ctrl)) { 2112 return false; 2113 } 2114 } 2115 } 2116 return true; 2117 } 2118 2119 // 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 2120 // to a check that's eliminated by range check elimination, it becomes input to an expression that feeds into the exit 2121 // 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 2122 // it would be to not eliminate the check in the main loop. Instead, we prevent sinking of the node here so better code 2123 // is generated for the main loop. 2124 bool PhaseIdealLoop::would_sink_below_pre_loop_exit(IdealLoopTree* n_loop, Node* ctrl) { 2125 if (n_loop->_head->is_CountedLoop() && n_loop->_head->as_CountedLoop()->is_pre_loop()) { 2126 CountedLoopNode* pre_loop = n_loop->_head->as_CountedLoop(); 2127 if (is_dominator(pre_loop->loopexit(), ctrl)) { 2128 return true; 2129 } 2130 } 2131 return false; 2132 } 2133 2134 bool PhaseIdealLoop::ctrl_of_use_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop, Node* ctrl) { 2135 if (n->is_Load()) { 2136 // We can't reuse tags in PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal() so make sure each call to 2137 // get_late_ctrl_with_anti_dep() uses its own tag 2138 _dom_lca_tags_round++; 2139 assert(_dom_lca_tags_round != 0, "shouldn't wrap around"); 2140 2141 ctrl = get_late_ctrl_with_anti_dep(n->as_Load(), n_ctrl, ctrl); 2142 } 2143 IdealLoopTree *u_loop = get_loop(ctrl); 2144 if (u_loop == n_loop) { 2145 return false; // Found loop-varying use 2146 } 2147 if (n_loop->is_member(u_loop)) { 2148 return false; // Found use in inner loop 2149 } 2150 if (would_sink_below_pre_loop_exit(n_loop, ctrl)) { 2151 return false; 2152 } 2153 return true; 2154 } 2155 2156 //------------------------------split_if_with_blocks--------------------------- 2157 // Check for aggressive application of 'split-if' optimization, 2158 // using basic block level info. 2159 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack) { 2160 Node* root = C->root(); 2161 visited.set(root->_idx); // first, mark root as visited 2162 // Do pre-visit work for root 2163 Node* n = split_if_with_blocks_pre(root); 2164 uint cnt = n->outcnt(); 2165 uint i = 0; 2166 2167 while (true) { 2168 // Visit all children 2169 if (i < cnt) { 2170 Node* use = n->raw_out(i); 2171 ++i; 2172 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) { 2173 // Now do pre-visit work for this use 2174 use = split_if_with_blocks_pre(use); 2175 nstack.push(n, i); // Save parent and next use's index. 2176 n = use; // Process all children of current use. 2177 cnt = use->outcnt(); 2178 i = 0; 2179 } 2180 } 2181 else { 2182 // All of n's children have been processed, complete post-processing. 2183 if (cnt != 0 && !n->is_Con()) { 2184 assert(has_node(n), "no dead nodes"); 2185 split_if_with_blocks_post(n); 2186 if (C->failing()) { 2187 return; 2188 } 2189 } 2190 if (must_throttle_split_if()) { 2191 nstack.clear(); 2192 } 2193 if (nstack.is_empty()) { 2194 // Finished all nodes on stack. 2195 break; 2196 } 2197 // Get saved parent node and next use's index. Visit the rest of uses. 2198 n = nstack.node(); 2199 cnt = n->outcnt(); 2200 i = nstack.index(); 2201 nstack.pop(); 2202 } 2203 } 2204 } 2205 2206 2207 //============================================================================= 2208 // 2209 // C L O N E A L O O P B O D Y 2210 // 2211 2212 //------------------------------clone_iff-------------------------------------- 2213 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 2214 // "Nearly" because all Nodes have been cloned from the original in the loop, 2215 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 2216 // through the Phi recursively, and return a Bool. 2217 Node* PhaseIdealLoop::clone_iff(PhiNode* phi) { 2218 2219 // Convert this Phi into a Phi merging Bools 2220 uint i; 2221 for (i = 1; i < phi->req(); i++) { 2222 Node* b = phi->in(i); 2223 if (b->is_Phi()) { 2224 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi())); 2225 } else { 2226 assert(b->is_Bool() || b->is_OpaqueNotNull() || b->is_OpaqueInitializedAssertionPredicate(), 2227 "bool, non-null check with OpaqueNotNull or Initialized Assertion Predicate with its Opaque node"); 2228 } 2229 } 2230 Node* n = phi->in(1); 2231 Node* sample_opaque = nullptr; 2232 Node *sample_bool = nullptr; 2233 if (n->is_OpaqueNotNull() || n->is_OpaqueInitializedAssertionPredicate()) { 2234 sample_opaque = n; 2235 sample_bool = n->in(1); 2236 assert(sample_bool->is_Bool(), "wrong type"); 2237 } else { 2238 sample_bool = n; 2239 } 2240 Node* sample_cmp = sample_bool->in(1); 2241 const Type* t = Type::TOP; 2242 const TypePtr* at = nullptr; 2243 if (sample_cmp->is_FlatArrayCheck()) { 2244 // Left input of a FlatArrayCheckNode is memory, set the (adr) type of the phi accordingly 2245 assert(sample_cmp->in(1)->bottom_type() == Type::MEMORY, "unexpected input type"); 2246 t = Type::MEMORY; 2247 at = TypeRawPtr::BOTTOM; 2248 } 2249 2250 // Make Phis to merge the Cmp's inputs. 2251 PhiNode *phi1 = new PhiNode(phi->in(0), t, at); 2252 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP); 2253 for (i = 1; i < phi->req(); i++) { 2254 Node *n1 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1); 2255 Node *n2 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2); 2256 phi1->set_req(i, n1); 2257 phi2->set_req(i, n2); 2258 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 2259 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 2260 } 2261 // See if these Phis have been made before. 2262 // Register with optimizer 2263 Node *hit1 = _igvn.hash_find_insert(phi1); 2264 if (hit1) { // Hit, toss just made Phi 2265 _igvn.remove_dead_node(phi1); // Remove new phi 2266 assert(hit1->is_Phi(), "" ); 2267 phi1 = (PhiNode*)hit1; // Use existing phi 2268 } else { // Miss 2269 _igvn.register_new_node_with_optimizer(phi1); 2270 } 2271 Node *hit2 = _igvn.hash_find_insert(phi2); 2272 if (hit2) { // Hit, toss just made Phi 2273 _igvn.remove_dead_node(phi2); // Remove new phi 2274 assert(hit2->is_Phi(), "" ); 2275 phi2 = (PhiNode*)hit2; // Use existing phi 2276 } else { // Miss 2277 _igvn.register_new_node_with_optimizer(phi2); 2278 } 2279 // Register Phis with loop/block info 2280 set_ctrl(phi1, phi->in(0)); 2281 set_ctrl(phi2, phi->in(0)); 2282 // Make a new Cmp 2283 Node *cmp = sample_cmp->clone(); 2284 cmp->set_req(1, phi1); 2285 cmp->set_req(2, phi2); 2286 _igvn.register_new_node_with_optimizer(cmp); 2287 set_ctrl(cmp, phi->in(0)); 2288 2289 // Make a new Bool 2290 Node *b = sample_bool->clone(); 2291 b->set_req(1,cmp); 2292 _igvn.register_new_node_with_optimizer(b); 2293 set_ctrl(b, phi->in(0)); 2294 2295 if (sample_opaque != nullptr) { 2296 Node* opaque = sample_opaque->clone(); 2297 opaque->set_req(1, b); 2298 _igvn.register_new_node_with_optimizer(opaque); 2299 set_ctrl(opaque, phi->in(0)); 2300 return opaque; 2301 } 2302 2303 assert(b->is_Bool(), ""); 2304 return b; 2305 } 2306 2307 //------------------------------clone_bool------------------------------------- 2308 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 2309 // "Nearly" because all Nodes have been cloned from the original in the loop, 2310 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 2311 // through the Phi recursively, and return a Bool. 2312 CmpNode*PhaseIdealLoop::clone_bool(PhiNode* phi) { 2313 uint i; 2314 // Convert this Phi into a Phi merging Bools 2315 for( i = 1; i < phi->req(); i++ ) { 2316 Node *b = phi->in(i); 2317 if( b->is_Phi() ) { 2318 _igvn.replace_input_of(phi, i, clone_bool(b->as_Phi())); 2319 } else { 2320 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" ); 2321 } 2322 } 2323 2324 Node *sample_cmp = phi->in(1); 2325 2326 // Make Phis to merge the Cmp's inputs. 2327 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP ); 2328 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP ); 2329 for( uint j = 1; j < phi->req(); j++ ) { 2330 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP 2331 Node *n1, *n2; 2332 if( cmp_top->is_Cmp() ) { 2333 n1 = cmp_top->in(1); 2334 n2 = cmp_top->in(2); 2335 } else { 2336 n1 = n2 = cmp_top; 2337 } 2338 phi1->set_req( j, n1 ); 2339 phi2->set_req( j, n2 ); 2340 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 2341 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 2342 } 2343 2344 // See if these Phis have been made before. 2345 // Register with optimizer 2346 Node *hit1 = _igvn.hash_find_insert(phi1); 2347 if( hit1 ) { // Hit, toss just made Phi 2348 _igvn.remove_dead_node(phi1); // Remove new phi 2349 assert( hit1->is_Phi(), "" ); 2350 phi1 = (PhiNode*)hit1; // Use existing phi 2351 } else { // Miss 2352 _igvn.register_new_node_with_optimizer(phi1); 2353 } 2354 Node *hit2 = _igvn.hash_find_insert(phi2); 2355 if( hit2 ) { // Hit, toss just made Phi 2356 _igvn.remove_dead_node(phi2); // Remove new phi 2357 assert( hit2->is_Phi(), "" ); 2358 phi2 = (PhiNode*)hit2; // Use existing phi 2359 } else { // Miss 2360 _igvn.register_new_node_with_optimizer(phi2); 2361 } 2362 // Register Phis with loop/block info 2363 set_ctrl(phi1, phi->in(0)); 2364 set_ctrl(phi2, phi->in(0)); 2365 // Make a new Cmp 2366 Node *cmp = sample_cmp->clone(); 2367 cmp->set_req( 1, phi1 ); 2368 cmp->set_req( 2, phi2 ); 2369 _igvn.register_new_node_with_optimizer(cmp); 2370 set_ctrl(cmp, phi->in(0)); 2371 2372 assert( cmp->is_Cmp(), "" ); 2373 return (CmpNode*)cmp; 2374 } 2375 2376 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new, 2377 IdealLoopTree* loop, IdealLoopTree* outer_loop, 2378 Node_List*& split_if_set, Node_List*& split_bool_set, 2379 Node_List*& split_cex_set, Node_List& worklist, 2380 uint new_counter, CloneLoopMode mode) { 2381 Node* nnn = old_new[old->_idx]; 2382 // Copy uses to a worklist, so I can munge the def-use info 2383 // with impunity. 2384 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 2385 worklist.push(old->fast_out(j)); 2386 2387 while( worklist.size() ) { 2388 Node *use = worklist.pop(); 2389 if (!has_node(use)) continue; // Ignore dead nodes 2390 if (use->in(0) == C->top()) continue; 2391 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 2392 // Check for data-use outside of loop - at least one of OLD or USE 2393 // must not be a CFG node. 2394 #ifdef ASSERT 2395 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == nullptr) { 2396 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint(); 2397 assert(mode != IgnoreStripMined, "incorrect cloning mode"); 2398 assert((mode == ControlAroundStripMined && use == sfpt) || !use->is_reachable_from_root(), "missed a node"); 2399 } 2400 #endif 2401 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) { 2402 2403 // If the Data use is an IF, that means we have an IF outside the 2404 // loop that is switching on a condition that is set inside the 2405 // loop. Happens if people set a loop-exit flag; then test the flag 2406 // in the loop to break the loop, then test is again outside the 2407 // loop to determine which way the loop exited. 2408 // 2409 // For several uses we need to make sure that there is no phi between, 2410 // the use and the Bool/Cmp. We therefore clone the Bool/Cmp down here 2411 // to avoid such a phi in between. 2412 // For example, it is unexpected that there is a Phi between an 2413 // AllocateArray node and its ValidLengthTest input that could cause 2414 // split if to break. 2415 assert(!use->is_OpaqueTemplateAssertionPredicate(), 2416 "should not clone a Template Assertion Predicate which should be removed once it's useless"); 2417 if (use->is_If() || use->is_CMove() || use->is_OpaqueNotNull() || use->is_OpaqueInitializedAssertionPredicate() || 2418 (use->Opcode() == Op_AllocateArray && use->in(AllocateNode::ValidLengthTest) == old)) { 2419 // Since this code is highly unlikely, we lazily build the worklist 2420 // of such Nodes to go split. 2421 if (!split_if_set) { 2422 split_if_set = new Node_List(); 2423 } 2424 split_if_set->push(use); 2425 } 2426 if (use->is_Bool()) { 2427 if (!split_bool_set) { 2428 split_bool_set = new Node_List(); 2429 } 2430 split_bool_set->push(use); 2431 } 2432 if (use->Opcode() == Op_CreateEx) { 2433 if (!split_cex_set) { 2434 split_cex_set = new Node_List(); 2435 } 2436 split_cex_set->push(use); 2437 } 2438 2439 2440 // Get "block" use is in 2441 uint idx = 0; 2442 while( use->in(idx) != old ) idx++; 2443 Node *prev = use->is_CFG() ? use : get_ctrl(use); 2444 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" ); 2445 Node* cfg = (prev->_idx >= new_counter && prev->is_Region()) 2446 ? prev->in(2) 2447 : idom(prev); 2448 if( use->is_Phi() ) // Phi use is in prior block 2449 cfg = prev->in(idx); // NOT in block of Phi itself 2450 if (cfg->is_top()) { // Use is dead? 2451 _igvn.replace_input_of(use, idx, C->top()); 2452 continue; 2453 } 2454 2455 // If use is referenced through control edge... (idx == 0) 2456 if (mode == IgnoreStripMined && idx == 0) { 2457 LoopNode *head = loop->_head->as_Loop(); 2458 if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) { 2459 // That node is outside the inner loop, leave it outside the 2460 // outer loop as well to not confuse verification code. 2461 assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop"); 2462 _igvn.replace_input_of(use, 0, head->outer_loop_exit()); 2463 continue; 2464 } 2465 } 2466 2467 while(!outer_loop->is_member(get_loop(cfg))) { 2468 prev = cfg; 2469 cfg = (cfg->_idx >= new_counter && cfg->is_Region()) ? cfg->in(2) : idom(cfg); 2470 } 2471 // If the use occurs after merging several exits from the loop, then 2472 // old value must have dominated all those exits. Since the same old 2473 // value was used on all those exits we did not need a Phi at this 2474 // merge point. NOW we do need a Phi here. Each loop exit value 2475 // is now merged with the peeled body exit; each exit gets its own 2476 // private Phi and those Phis need to be merged here. 2477 Node *phi; 2478 if( prev->is_Region() ) { 2479 if( idx == 0 ) { // Updating control edge? 2480 phi = prev; // Just use existing control 2481 } else { // Else need a new Phi 2482 phi = PhiNode::make( prev, old ); 2483 // Now recursively fix up the new uses of old! 2484 for( uint i = 1; i < prev->req(); i++ ) { 2485 worklist.push(phi); // Onto worklist once for each 'old' input 2486 } 2487 } 2488 } else { 2489 // Get new RegionNode merging old and new loop exits 2490 prev = old_new[prev->_idx]; 2491 assert( prev, "just made this in step 7" ); 2492 if( idx == 0) { // Updating control edge? 2493 phi = prev; // Just use existing control 2494 } else { // Else need a new Phi 2495 // Make a new Phi merging data values properly 2496 phi = PhiNode::make( prev, old ); 2497 phi->set_req( 1, nnn ); 2498 } 2499 } 2500 // If inserting a new Phi, check for prior hits 2501 if( idx != 0 ) { 2502 Node *hit = _igvn.hash_find_insert(phi); 2503 if( hit == nullptr ) { 2504 _igvn.register_new_node_with_optimizer(phi); // Register new phi 2505 } else { // or 2506 // Remove the new phi from the graph and use the hit 2507 _igvn.remove_dead_node(phi); 2508 phi = hit; // Use existing phi 2509 } 2510 set_ctrl(phi, prev); 2511 } 2512 // Make 'use' use the Phi instead of the old loop body exit value 2513 assert(use->in(idx) == old, "old is still input of use"); 2514 // We notify all uses of old, including use, and the indirect uses, 2515 // that may now be optimized because we have replaced old with phi. 2516 _igvn.add_users_to_worklist(old); 2517 if (idx == 0 && 2518 use->depends_only_on_test()) { 2519 Node* pinned_clone = use->pin_array_access_node(); 2520 if (pinned_clone != nullptr) { 2521 // Pin array access nodes: control is updated here to a region. If, after some transformations, only one path 2522 // into the region is left, an array load could become dependent on a condition that's not a range check for 2523 // that access. If that condition is replaced by an identical dominating one, then an unpinned load would risk 2524 // floating above its range check. 2525 pinned_clone->set_req(0, phi); 2526 register_new_node_with_ctrl_of(pinned_clone, use); 2527 _igvn.replace_node(use, pinned_clone); 2528 continue; 2529 } 2530 } 2531 _igvn.replace_input_of(use, idx, phi); 2532 if( use->_idx >= new_counter ) { // If updating new phis 2533 // Not needed for correctness, but prevents a weak assert 2534 // in AddPNode from tripping (when we end up with different 2535 // base & derived Phis that will become the same after 2536 // IGVN does CSE). 2537 Node *hit = _igvn.hash_find_insert(use); 2538 if( hit ) // Go ahead and re-hash for hits. 2539 _igvn.replace_node( use, hit ); 2540 } 2541 } 2542 } 2543 } 2544 2545 static void collect_nodes_in_outer_loop_not_reachable_from_sfpt(Node* n, const IdealLoopTree *loop, const IdealLoopTree* outer_loop, 2546 const Node_List &old_new, Unique_Node_List& wq, PhaseIdealLoop* phase, 2547 bool check_old_new) { 2548 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2549 Node* u = n->fast_out(j); 2550 assert(check_old_new || old_new[u->_idx] == nullptr, "shouldn't have been cloned"); 2551 if (!u->is_CFG() && (!check_old_new || old_new[u->_idx] == nullptr)) { 2552 assert(!phase->ctrl_is_member(loop, u) || !loop->_body.contains(u), "can be in outer loop or out of both loops only"); 2553 if (!phase->ctrl_is_member(loop, u)) { 2554 if (phase->ctrl_is_member(outer_loop, u)) { 2555 wq.push(u); 2556 } else { 2557 // nodes pinned with control in the outer loop but not referenced from the safepoint must be moved out of 2558 // the outer loop too 2559 Node* u_c = u->in(0); 2560 if (u_c != nullptr) { 2561 IdealLoopTree* u_c_loop = phase->get_loop(u_c); 2562 if (outer_loop->is_member(u_c_loop) && !loop->is_member(u_c_loop)) { 2563 wq.push(u); 2564 } 2565 } 2566 } 2567 } 2568 } 2569 } 2570 } 2571 2572 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop, 2573 IdealLoopTree* outer_loop, int dd, Node_List &old_new, 2574 Node_List& extra_data_nodes) { 2575 if (head->is_strip_mined() && mode != IgnoreStripMined) { 2576 CountedLoopNode* cl = head->as_CountedLoop(); 2577 Node* l = cl->outer_loop(); 2578 Node* tail = cl->outer_loop_tail(); 2579 IfNode* le = cl->outer_loop_end(); 2580 Node* sfpt = cl->outer_safepoint(); 2581 CountedLoopEndNode* cle = cl->loopexit(); 2582 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop(); 2583 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null(); 2584 IfFalseNode* cle_out = cle->false_proj(); 2585 2586 Node* new_sfpt = nullptr; 2587 Node* new_cle_out = cle_out->clone(); 2588 old_new.map(cle_out->_idx, new_cle_out); 2589 if (mode == CloneIncludesStripMined) { 2590 // clone outer loop body 2591 Node* new_l = l->clone(); 2592 Node* new_tail = tail->clone(); 2593 IfNode* new_le = le->clone()->as_If(); 2594 new_sfpt = sfpt->clone(); 2595 2596 set_loop(new_l, outer_loop->_parent); 2597 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd); 2598 set_loop(new_cle_out, outer_loop->_parent); 2599 set_idom(new_cle_out, new_cle, dd); 2600 set_loop(new_sfpt, outer_loop->_parent); 2601 set_idom(new_sfpt, new_cle_out, dd); 2602 set_loop(new_le, outer_loop->_parent); 2603 set_idom(new_le, new_sfpt, dd); 2604 set_loop(new_tail, outer_loop->_parent); 2605 set_idom(new_tail, new_le, dd); 2606 set_idom(new_cl, new_l, dd); 2607 2608 old_new.map(l->_idx, new_l); 2609 old_new.map(tail->_idx, new_tail); 2610 old_new.map(le->_idx, new_le); 2611 old_new.map(sfpt->_idx, new_sfpt); 2612 2613 new_l->set_req(LoopNode::LoopBackControl, new_tail); 2614 new_l->set_req(0, new_l); 2615 new_tail->set_req(0, new_le); 2616 new_le->set_req(0, new_sfpt); 2617 new_sfpt->set_req(0, new_cle_out); 2618 new_cle_out->set_req(0, new_cle); 2619 new_cl->set_req(LoopNode::EntryControl, new_l); 2620 2621 _igvn.register_new_node_with_optimizer(new_l); 2622 _igvn.register_new_node_with_optimizer(new_tail); 2623 _igvn.register_new_node_with_optimizer(new_le); 2624 } else { 2625 Node *newhead = old_new[loop->_head->_idx]; 2626 newhead->as_Loop()->clear_strip_mined(); 2627 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl)); 2628 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 2629 } 2630 // Look at data node that were assigned a control in the outer 2631 // loop: they are kept in the outer loop by the safepoint so start 2632 // from the safepoint node's inputs. 2633 IdealLoopTree* outer_loop = get_loop(l); 2634 Node_Stack stack(2); 2635 stack.push(sfpt, 1); 2636 uint new_counter = C->unique(); 2637 while (stack.size() > 0) { 2638 Node* n = stack.node(); 2639 uint i = stack.index(); 2640 while (i < n->req() && 2641 (n->in(i) == nullptr || 2642 !has_ctrl(n->in(i)) || 2643 get_loop(get_ctrl(n->in(i))) != outer_loop || 2644 (old_new[n->in(i)->_idx] != nullptr && old_new[n->in(i)->_idx]->_idx >= new_counter))) { 2645 i++; 2646 } 2647 if (i < n->req()) { 2648 stack.set_index(i+1); 2649 stack.push(n->in(i), 0); 2650 } else { 2651 assert(old_new[n->_idx] == nullptr || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet"); 2652 Node* m = n == sfpt ? new_sfpt : n->clone(); 2653 if (m != nullptr) { 2654 for (uint i = 0; i < n->req(); i++) { 2655 if (m->in(i) != nullptr && old_new[m->in(i)->_idx] != nullptr) { 2656 m->set_req(i, old_new[m->in(i)->_idx]); 2657 } 2658 } 2659 } else { 2660 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?"); 2661 } 2662 if (n != sfpt) { 2663 extra_data_nodes.push(n); 2664 _igvn.register_new_node_with_optimizer(m); 2665 assert(get_ctrl(n) == cle_out, "what other control?"); 2666 set_ctrl(m, new_cle_out); 2667 old_new.map(n->_idx, m); 2668 } 2669 stack.pop(); 2670 } 2671 } 2672 if (mode == CloneIncludesStripMined) { 2673 _igvn.register_new_node_with_optimizer(new_sfpt); 2674 _igvn.register_new_node_with_optimizer(new_cle_out); 2675 } 2676 // Some other transformation may have pessimistically assigned some 2677 // data nodes to the outer loop. Set their control so they are out 2678 // of the outer loop. 2679 ResourceMark rm; 2680 Unique_Node_List wq; 2681 for (uint i = 0; i < extra_data_nodes.size(); i++) { 2682 Node* old = extra_data_nodes.at(i); 2683 collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true); 2684 } 2685 2686 for (uint i = 0; i < loop->_body.size(); i++) { 2687 Node* old = loop->_body.at(i); 2688 collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true); 2689 } 2690 2691 Node* inner_out = sfpt->in(0); 2692 if (inner_out->outcnt() > 1) { 2693 collect_nodes_in_outer_loop_not_reachable_from_sfpt(inner_out, loop, outer_loop, old_new, wq, this, true); 2694 } 2695 2696 Node* new_ctrl = cl->outer_loop_exit(); 2697 assert(get_loop(new_ctrl) != outer_loop, "must be out of the loop nest"); 2698 for (uint i = 0; i < wq.size(); i++) { 2699 Node* n = wq.at(i); 2700 set_ctrl(n, new_ctrl); 2701 if (n->in(0) != nullptr) { 2702 _igvn.replace_input_of(n, 0, new_ctrl); 2703 } 2704 collect_nodes_in_outer_loop_not_reachable_from_sfpt(n, loop, outer_loop, old_new, wq, this, false); 2705 } 2706 } else { 2707 Node *newhead = old_new[loop->_head->_idx]; 2708 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 2709 } 2710 } 2711 2712 //------------------------------clone_loop------------------------------------- 2713 // 2714 // C L O N E A L O O P B O D Y 2715 // 2716 // This is the basic building block of the loop optimizations. It clones an 2717 // entire loop body. It makes an old_new loop body mapping; with this mapping 2718 // you can find the new-loop equivalent to an old-loop node. All new-loop 2719 // nodes are exactly equal to their old-loop counterparts, all edges are the 2720 // same. All exits from the old-loop now have a RegionNode that merges the 2721 // equivalent new-loop path. This is true even for the normal "loop-exit" 2722 // condition. All uses of loop-invariant old-loop values now come from (one 2723 // or more) Phis that merge their new-loop equivalents. 2724 // 2725 // This operation leaves the graph in an illegal state: there are two valid 2726 // control edges coming from the loop pre-header to both loop bodies. I'll 2727 // definitely have to hack the graph after running this transform. 2728 // 2729 // From this building block I will further edit edges to perform loop peeling 2730 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc. 2731 // 2732 // Parameter side_by_size_idom: 2733 // When side_by_size_idom is null, the dominator tree is constructed for 2734 // the clone loop to dominate the original. Used in construction of 2735 // pre-main-post loop sequence. 2736 // When nonnull, the clone and original are side-by-side, both are 2737 // dominated by the side_by_side_idom node. Used in construction of 2738 // unswitched loops. 2739 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd, 2740 CloneLoopMode mode, Node* side_by_side_idom) { 2741 2742 LoopNode* head = loop->_head->as_Loop(); 2743 head->verify_strip_mined(1); 2744 2745 if (C->do_vector_loop() && PrintOpto) { 2746 const char* mname = C->method()->name()->as_quoted_ascii(); 2747 if (mname != nullptr) { 2748 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname); 2749 } 2750 } 2751 2752 CloneMap& cm = C->clone_map(); 2753 if (C->do_vector_loop()) { 2754 cm.set_clone_idx(cm.max_gen()+1); 2755 #ifndef PRODUCT 2756 if (PrintOpto) { 2757 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx()); 2758 loop->dump_head(); 2759 } 2760 #endif 2761 } 2762 2763 // Step 1: Clone the loop body. Make the old->new mapping. 2764 clone_loop_body(loop->_body, old_new, &cm); 2765 2766 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop; 2767 2768 // Step 2: Fix the edges in the new body. If the old input is outside the 2769 // loop use it. If the old input is INside the loop, use the corresponding 2770 // new node instead. 2771 fix_body_edges(loop->_body, loop, old_new, dd, outer_loop->_parent, false); 2772 2773 Node_List extra_data_nodes; // data nodes in the outer strip mined loop 2774 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes); 2775 2776 // Step 3: Now fix control uses. Loop varying control uses have already 2777 // been fixed up (as part of all input edges in Step 2). Loop invariant 2778 // control uses must be either an IfFalse or an IfTrue. Make a merge 2779 // point to merge the old and new IfFalse/IfTrue nodes; make the use 2780 // refer to this. 2781 Node_List worklist; 2782 uint new_counter = C->unique(); 2783 fix_ctrl_uses(loop->_body, loop, old_new, mode, side_by_side_idom, &cm, worklist); 2784 2785 // Step 4: If loop-invariant use is not control, it must be dominated by a 2786 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region 2787 // there if needed. Make a Phi there merging old and new used values. 2788 Node_List *split_if_set = nullptr; 2789 Node_List *split_bool_set = nullptr; 2790 Node_List *split_cex_set = nullptr; 2791 fix_data_uses(loop->_body, loop, mode, outer_loop, new_counter, old_new, worklist, split_if_set, split_bool_set, split_cex_set); 2792 2793 for (uint i = 0; i < extra_data_nodes.size(); i++) { 2794 Node* old = extra_data_nodes.at(i); 2795 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set, 2796 split_bool_set, split_cex_set, worklist, new_counter, 2797 mode); 2798 } 2799 2800 // Check for IFs that need splitting/cloning. Happens if an IF outside of 2801 // the loop uses a condition set in the loop. The original IF probably 2802 // takes control from one or more OLD Regions (which in turn get from NEW 2803 // Regions). In any case, there will be a set of Phis for each merge point 2804 // from the IF up to where the original BOOL def exists the loop. 2805 finish_clone_loop(split_if_set, split_bool_set, split_cex_set); 2806 2807 } 2808 2809 void PhaseIdealLoop::finish_clone_loop(Node_List* split_if_set, Node_List* split_bool_set, Node_List* split_cex_set) { 2810 if (split_if_set) { 2811 while (split_if_set->size()) { 2812 Node *iff = split_if_set->pop(); 2813 uint input = iff->Opcode() == Op_AllocateArray ? AllocateNode::ValidLengthTest : 1; 2814 if (iff->in(input)->is_Phi()) { 2815 Node *b = clone_iff(iff->in(input)->as_Phi()); 2816 _igvn.replace_input_of(iff, input, b); 2817 } 2818 } 2819 } 2820 if (split_bool_set) { 2821 while (split_bool_set->size()) { 2822 Node *b = split_bool_set->pop(); 2823 Node *phi = b->in(1); 2824 assert(phi->is_Phi(), ""); 2825 CmpNode *cmp = clone_bool((PhiNode*) phi); 2826 _igvn.replace_input_of(b, 1, cmp); 2827 } 2828 } 2829 if (split_cex_set) { 2830 while (split_cex_set->size()) { 2831 Node *b = split_cex_set->pop(); 2832 assert(b->in(0)->is_Region(), ""); 2833 assert(b->in(1)->is_Phi(), ""); 2834 assert(b->in(0)->in(0) == b->in(1)->in(0), ""); 2835 split_up(b, b->in(0), nullptr); 2836 } 2837 } 2838 } 2839 2840 void PhaseIdealLoop::fix_data_uses(Node_List& body, IdealLoopTree* loop, CloneLoopMode mode, IdealLoopTree* outer_loop, 2841 uint new_counter, Node_List &old_new, Node_List &worklist, Node_List*& split_if_set, 2842 Node_List*& split_bool_set, Node_List*& split_cex_set) { 2843 for(uint i = 0; i < body.size(); i++ ) { 2844 Node* old = body.at(i); 2845 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set, 2846 split_bool_set, split_cex_set, worklist, new_counter, 2847 mode); 2848 } 2849 } 2850 2851 void PhaseIdealLoop::fix_ctrl_uses(const Node_List& body, const IdealLoopTree* loop, Node_List &old_new, CloneLoopMode mode, 2852 Node* side_by_side_idom, CloneMap* cm, Node_List &worklist) { 2853 LoopNode* head = loop->_head->as_Loop(); 2854 for(uint i = 0; i < body.size(); i++ ) { 2855 Node* old = body.at(i); 2856 if( !old->is_CFG() ) continue; 2857 2858 // Copy uses to a worklist, so I can munge the def-use info 2859 // with impunity. 2860 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) { 2861 worklist.push(old->fast_out(j)); 2862 } 2863 2864 while (worklist.size()) { // Visit all uses 2865 Node *use = worklist.pop(); 2866 if (!has_node(use)) continue; // Ignore dead nodes 2867 IdealLoopTree *use_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use ); 2868 if (!loop->is_member(use_loop) && use->is_CFG()) { 2869 // Both OLD and USE are CFG nodes here. 2870 assert(use->is_Proj(), "" ); 2871 Node* nnn = old_new[old->_idx]; 2872 2873 Node* newuse = nullptr; 2874 if (head->is_strip_mined() && mode != IgnoreStripMined) { 2875 CountedLoopNode* cl = head->as_CountedLoop(); 2876 CountedLoopEndNode* cle = cl->loopexit(); 2877 // is use the projection that exits the loop from the CountedLoopEndNode? 2878 if (use->in(0) == cle) { 2879 IfFalseNode* cle_out = use->as_IfFalse(); 2880 IfNode* le = cl->outer_loop_end(); 2881 use = le->false_proj(); 2882 use_loop = get_loop(use); 2883 if (mode == CloneIncludesStripMined) { 2884 nnn = old_new[le->_idx]; 2885 } else { 2886 newuse = old_new[cle_out->_idx]; 2887 } 2888 } 2889 } 2890 if (newuse == nullptr) { 2891 newuse = use->clone(); 2892 } 2893 2894 // Clone the loop exit control projection 2895 if (C->do_vector_loop() && cm != nullptr) { 2896 cm->verify_insert_and_clone(use, newuse, cm->clone_idx()); 2897 } 2898 newuse->set_req(0,nnn); 2899 _igvn.register_new_node_with_optimizer(newuse); 2900 set_loop(newuse, use_loop); 2901 set_idom(newuse, nnn, dom_depth(nnn) + 1 ); 2902 2903 // We need a Region to merge the exit from the peeled body and the 2904 // exit from the old loop body. 2905 RegionNode *r = new RegionNode(3); 2906 uint dd_r = MIN2(dom_depth(newuse), dom_depth(use)); 2907 assert(dd_r >= dom_depth(dom_lca(newuse, use)), "" ); 2908 2909 // The original user of 'use' uses 'r' instead. 2910 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) { 2911 Node* useuse = use->last_out(l); 2912 _igvn.rehash_node_delayed(useuse); 2913 uint uses_found = 0; 2914 if (useuse->in(0) == use) { 2915 useuse->set_req(0, r); 2916 uses_found++; 2917 if (useuse->is_CFG()) { 2918 // This is not a dom_depth > dd_r because when new 2919 // control flow is constructed by a loop opt, a node and 2920 // its dominator can end up at the same dom_depth 2921 assert(dom_depth(useuse) >= dd_r, ""); 2922 set_idom(useuse, r, dom_depth(useuse)); 2923 } 2924 } 2925 for (uint k = 1; k < useuse->req(); k++) { 2926 if( useuse->in(k) == use ) { 2927 useuse->set_req(k, r); 2928 uses_found++; 2929 if (useuse->is_Loop() && k == LoopNode::EntryControl) { 2930 // This is not a dom_depth > dd_r because when new 2931 // control flow is constructed by a loop opt, a node 2932 // and its dominator can end up at the same dom_depth 2933 assert(dom_depth(useuse) >= dd_r , ""); 2934 set_idom(useuse, r, dom_depth(useuse)); 2935 } 2936 } 2937 } 2938 l -= uses_found; // we deleted 1 or more copies of this edge 2939 } 2940 2941 assert(use->is_Proj(), "loop exit should be projection"); 2942 // replace_node_and_forward_ctrl() below moves all nodes that are: 2943 // - control dependent on the loop exit or 2944 // - have control set to the loop exit 2945 // below the post-loop merge point. 2946 // replace_node_and_forward_ctrl() takes a dead control as first input. 2947 // To make it possible to use it, the loop exit projection is cloned and becomes the 2948 // new exit projection. The initial one becomes dead and is "replaced" by the region. 2949 Node* use_clone = use->clone(); 2950 register_control(use_clone, use_loop, idom(use), dom_depth(use)); 2951 // Now finish up 'r' 2952 r->set_req(1, newuse); 2953 r->set_req(2, use_clone); 2954 _igvn.register_new_node_with_optimizer(r); 2955 set_loop(r, use_loop); 2956 set_idom(r, (side_by_side_idom == nullptr) ? newuse->in(0) : side_by_side_idom, dd_r); 2957 replace_node_and_forward_ctrl(use, r); 2958 // Map the (cloned) old use to the new merge point 2959 old_new.map(use_clone->_idx, r); 2960 } // End of if a loop-exit test 2961 } 2962 } 2963 } 2964 2965 void PhaseIdealLoop::fix_body_edges(const Node_List &body, IdealLoopTree* loop, const Node_List &old_new, int dd, 2966 IdealLoopTree* parent, bool partial) { 2967 for(uint i = 0; i < body.size(); i++ ) { 2968 Node *old = body.at(i); 2969 Node *nnn = old_new[old->_idx]; 2970 // Fix CFG/Loop controlling the new node 2971 if (has_ctrl(old)) { 2972 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]); 2973 } else { 2974 set_loop(nnn, parent); 2975 if (old->outcnt() > 0) { 2976 Node* dom = idom(old); 2977 if (old_new[dom->_idx] != nullptr) { 2978 dom = old_new[dom->_idx]; 2979 set_idom(nnn, dom, dd ); 2980 } 2981 } 2982 } 2983 // Correct edges to the new node 2984 for (uint j = 0; j < nnn->req(); j++) { 2985 Node *n = nnn->in(j); 2986 if (n != nullptr) { 2987 IdealLoopTree *old_in_loop = get_loop(has_ctrl(n) ? get_ctrl(n) : n); 2988 if (loop->is_member(old_in_loop)) { 2989 if (old_new[n->_idx] != nullptr) { 2990 nnn->set_req(j, old_new[n->_idx]); 2991 } else { 2992 assert(!body.contains(n), ""); 2993 assert(partial, "node not cloned"); 2994 } 2995 } 2996 } 2997 } 2998 _igvn.hash_find_insert(nnn); 2999 } 3000 } 3001 3002 void PhaseIdealLoop::clone_loop_body(const Node_List& body, Node_List &old_new, CloneMap* cm) { 3003 for (uint i = 0; i < body.size(); i++) { 3004 Node* old = body.at(i); 3005 Node* nnn = old->clone(); 3006 old_new.map(old->_idx, nnn); 3007 if (C->do_vector_loop() && cm != nullptr) { 3008 cm->verify_insert_and_clone(old, nnn, cm->clone_idx()); 3009 } 3010 _igvn.register_new_node_with_optimizer(nnn); 3011 } 3012 } 3013 3014 3015 //---------------------- stride_of_possible_iv ------------------------------------- 3016 // Looks for an iff/bool/comp with one operand of the compare 3017 // being a cycle involving an add and a phi, 3018 // with an optional truncation (left-shift followed by a right-shift) 3019 // of the add. Returns zero if not an iv. 3020 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) { 3021 Node* trunc1 = nullptr; 3022 Node* trunc2 = nullptr; 3023 const TypeInteger* ttype = nullptr; 3024 if (!iff->is_If() || iff->in(1) == nullptr || !iff->in(1)->is_Bool()) { 3025 return 0; 3026 } 3027 BoolNode* bl = iff->in(1)->as_Bool(); 3028 Node* cmp = bl->in(1); 3029 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) { 3030 return 0; 3031 } 3032 // Must have an invariant operand 3033 if (ctrl_is_member(get_loop(iff), cmp->in(2))) { 3034 return 0; 3035 } 3036 Node* add2 = nullptr; 3037 Node* cmp1 = cmp->in(1); 3038 if (cmp1->is_Phi()) { 3039 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) ))) 3040 Node* phi = cmp1; 3041 for (uint i = 1; i < phi->req(); i++) { 3042 Node* in = phi->in(i); 3043 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in, 3044 &trunc1, &trunc2, &ttype, T_INT); 3045 if (add && add->in(1) == phi) { 3046 add2 = add->in(2); 3047 break; 3048 } 3049 } 3050 } else { 3051 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) ))) 3052 Node* addtrunc = cmp1; 3053 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc, 3054 &trunc1, &trunc2, &ttype, T_INT); 3055 if (add && add->in(1)->is_Phi()) { 3056 Node* phi = add->in(1); 3057 for (uint i = 1; i < phi->req(); i++) { 3058 if (phi->in(i) == addtrunc) { 3059 add2 = add->in(2); 3060 break; 3061 } 3062 } 3063 } 3064 } 3065 if (add2 != nullptr) { 3066 const TypeInt* add2t = _igvn.type(add2)->is_int(); 3067 if (add2t->is_con()) { 3068 return add2t->get_con(); 3069 } 3070 } 3071 return 0; 3072 } 3073 3074 3075 //---------------------- stay_in_loop ------------------------------------- 3076 // Return the (unique) control output node that's in the loop (if it exists.) 3077 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) { 3078 Node* unique = nullptr; 3079 if (!n) return nullptr; 3080 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 3081 Node* use = n->fast_out(i); 3082 if (!has_ctrl(use) && loop->is_member(get_loop(use))) { 3083 if (unique != nullptr) { 3084 return nullptr; 3085 } 3086 unique = use; 3087 } 3088 } 3089 return unique; 3090 } 3091 3092 //------------------------------ register_node ------------------------------------- 3093 // Utility to register node "n" with PhaseIdealLoop 3094 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree* loop, Node* pred, uint ddepth) { 3095 _igvn.register_new_node_with_optimizer(n); 3096 loop->_body.push(n); 3097 if (n->is_CFG()) { 3098 set_loop(n, loop); 3099 set_idom(n, pred, ddepth); 3100 } else { 3101 set_ctrl(n, pred); 3102 } 3103 } 3104 3105 //------------------------------ proj_clone ------------------------------------- 3106 // Utility to create an if-projection 3107 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) { 3108 ProjNode* c = p->clone()->as_Proj(); 3109 c->set_req(0, iff); 3110 return c; 3111 } 3112 3113 //------------------------------ short_circuit_if ------------------------------------- 3114 // Force the iff control output to be the live_proj 3115 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) { 3116 guarantee(live_proj != nullptr, "null projection"); 3117 int proj_con = live_proj->_con; 3118 assert(proj_con == 0 || proj_con == 1, "false or true projection"); 3119 Node* con = intcon(proj_con); 3120 if (iff) { 3121 iff->set_req(1, con); 3122 } 3123 return con; 3124 } 3125 3126 //------------------------------ insert_if_before_proj ------------------------------------- 3127 // Insert a new if before an if projection (* - new node) 3128 // 3129 // before 3130 // if(test) 3131 // / \ 3132 // v v 3133 // other-proj proj (arg) 3134 // 3135 // after 3136 // if(test) 3137 // / \ 3138 // / v 3139 // | * proj-clone 3140 // v | 3141 // other-proj v 3142 // * new_if(relop(cmp[IU](left,right))) 3143 // / \ 3144 // v v 3145 // * new-proj proj 3146 // (returned) 3147 // 3148 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) { 3149 IfNode* iff = proj->in(0)->as_If(); 3150 IdealLoopTree *loop = get_loop(proj); 3151 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 3152 uint ddepth = dom_depth(proj); 3153 3154 _igvn.rehash_node_delayed(iff); 3155 _igvn.rehash_node_delayed(proj); 3156 3157 proj->set_req(0, nullptr); // temporary disconnect 3158 ProjNode* proj2 = proj_clone(proj, iff); 3159 register_node(proj2, loop, iff, ddepth); 3160 3161 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right); 3162 register_node(cmp, loop, proj2, ddepth); 3163 3164 BoolNode* bol = new BoolNode(cmp, relop); 3165 register_node(bol, loop, proj2, ddepth); 3166 3167 int opcode = iff->Opcode(); 3168 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode"); 3169 IfNode* new_if = IfNode::make_with_same_profile(iff, proj2, bol); 3170 register_node(new_if, loop, proj2, ddepth); 3171 3172 proj->set_req(0, new_if); // reattach 3173 set_idom(proj, new_if, ddepth); 3174 3175 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj(); 3176 guarantee(new_exit != nullptr, "null exit node"); 3177 register_node(new_exit, get_loop(other_proj), new_if, ddepth); 3178 3179 return new_exit; 3180 } 3181 3182 //------------------------------ insert_region_before_proj ------------------------------------- 3183 // Insert a region before an if projection (* - new node) 3184 // 3185 // before 3186 // if(test) 3187 // / | 3188 // v | 3189 // proj v 3190 // other-proj 3191 // 3192 // after 3193 // if(test) 3194 // / | 3195 // v | 3196 // * proj-clone v 3197 // | other-proj 3198 // v 3199 // * new-region 3200 // | 3201 // v 3202 // * dum_if 3203 // / \ 3204 // v \ 3205 // * dum-proj v 3206 // proj 3207 // 3208 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) { 3209 IfNode* iff = proj->in(0)->as_If(); 3210 IdealLoopTree *loop = get_loop(proj); 3211 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 3212 uint ddepth = dom_depth(proj); 3213 3214 _igvn.rehash_node_delayed(iff); 3215 _igvn.rehash_node_delayed(proj); 3216 3217 proj->set_req(0, nullptr); // temporary disconnect 3218 ProjNode* proj2 = proj_clone(proj, iff); 3219 register_node(proj2, loop, iff, ddepth); 3220 3221 RegionNode* reg = new RegionNode(2); 3222 reg->set_req(1, proj2); 3223 register_node(reg, loop, iff, ddepth); 3224 3225 IfNode* dum_if = new IfNode(reg, short_circuit_if(nullptr, proj), iff->_prob, iff->_fcnt); 3226 register_node(dum_if, loop, reg, ddepth); 3227 3228 proj->set_req(0, dum_if); // reattach 3229 set_idom(proj, dum_if, ddepth); 3230 3231 ProjNode* dum_proj = proj_clone(other_proj, dum_if); 3232 register_node(dum_proj, loop, dum_if, ddepth); 3233 3234 return reg; 3235 } 3236 3237 // Idea 3238 // ---- 3239 // Partial Peeling tries to rotate the loop in such a way that it can later be turned into a counted loop. Counted loops 3240 // require a signed loop exit test. When calling this method, we've only found a suitable unsigned test to partial peel 3241 // with. Therefore, we try to split off a signed loop exit test from the unsigned test such that it can be used as new 3242 // loop exit while keeping the unsigned test unchanged and preserving the same behavior as if we've used the unsigned 3243 // test alone instead: 3244 // 3245 // Before Partial Peeling: 3246 // Loop: 3247 // <peeled section> 3248 // Split off signed loop exit test 3249 // <-- CUT HERE --> 3250 // Unchanged unsigned loop exit test 3251 // <rest of unpeeled section> 3252 // goto Loop 3253 // 3254 // After Partial Peeling: 3255 // <cloned peeled section> 3256 // Cloned split off signed loop exit test 3257 // Loop: 3258 // Unchanged unsigned loop exit test 3259 // <rest of unpeeled section> 3260 // <peeled section> 3261 // Split off signed loop exit test 3262 // goto Loop 3263 // 3264 // Details 3265 // ------- 3266 // Before: 3267 // if (i <u limit) Unsigned loop exit condition 3268 // / | 3269 // v v 3270 // exit-proj stay-in-loop-proj 3271 // 3272 // Split off a signed loop exit test (i.e. with CmpI) from an unsigned loop exit test (i.e. with CmpU) and insert it 3273 // before the CmpU on the stay-in-loop path and keep both tests: 3274 // 3275 // if (i <u limit) Signed loop exit test 3276 // / | 3277 // / if (i <u limit) Unsigned loop exit test 3278 // / / | 3279 // v v v 3280 // exit-region stay-in-loop-proj 3281 // 3282 // Implementation 3283 // -------------- 3284 // We need to make sure that the new signed loop exit test is properly inserted into the graph such that the unsigned 3285 // loop exit test still dominates the same set of control nodes, the ctrl() relation from data nodes to both loop 3286 // exit tests is preserved, and their loop nesting is correct. 3287 // 3288 // To achieve that, we clone the unsigned loop exit test completely (leave it unchanged), insert the signed loop exit 3289 // test above it and kill the original unsigned loop exit test by setting it's condition to a constant 3290 // (i.e. stay-in-loop-const in graph below) such that IGVN can fold it later: 3291 // 3292 // if (stay-in-loop-const) Killed original unsigned loop exit test 3293 // / | 3294 // / v 3295 // / if (i < limit) Split off signed loop exit test 3296 // / / | 3297 // / / v 3298 // / / if (i <u limit) Cloned unsigned loop exit test 3299 // / / / | 3300 // v v v | 3301 // exit-region | 3302 // | | 3303 // dummy-if | 3304 // / | | 3305 // dead | | 3306 // v v 3307 // exit-proj stay-in-loop-proj 3308 // 3309 // Note: The dummy-if is inserted to create a region to merge the loop exits between the original to be killed unsigned 3310 // loop exit test and its exit projection while keeping the exit projection (also see insert_region_before_proj()). 3311 // 3312 // Requirements 3313 // ------------ 3314 // Note that we can only split off a signed loop exit test from the unsigned loop exit test when the behavior is exactly 3315 // the same as before with only a single unsigned test. This is only possible if certain requirements are met. 3316 // Otherwise, we need to bail out (see comments in the code below). 3317 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree* loop) { 3318 const bool Signed = true; 3319 const bool Unsigned = false; 3320 3321 BoolNode* bol = if_cmpu->in(1)->as_Bool(); 3322 if (bol->_test._test != BoolTest::lt) { 3323 return nullptr; 3324 } 3325 CmpNode* cmpu = bol->in(1)->as_Cmp(); 3326 assert(cmpu->Opcode() == Op_CmpU, "must be unsigned comparison"); 3327 3328 int stride = stride_of_possible_iv(if_cmpu); 3329 if (stride == 0) { 3330 return nullptr; 3331 } 3332 3333 Node* lp_proj = stay_in_loop(if_cmpu, loop); 3334 guarantee(lp_proj != nullptr, "null loop node"); 3335 3336 ProjNode* lp_continue = lp_proj->as_Proj(); 3337 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj(); 3338 if (!lp_exit->is_IfFalse()) { 3339 // The loop exit condition is (i <u limit) ==> (i >= 0 && i < limit). 3340 // We therefore can't add a single exit condition. 3341 return nullptr; 3342 } 3343 // The unsigned loop exit condition is 3344 // !(i <u limit) 3345 // = i >=u limit 3346 // 3347 // First, we note that for any x for which 3348 // 0 <= x <= INT_MAX 3349 // we can convert x to an unsigned int and still get the same guarantee: 3350 // 0 <= (uint) x <= INT_MAX = (uint) INT_MAX 3351 // 0 <=u (uint) x <=u INT_MAX = (uint) INT_MAX (LEMMA) 3352 // 3353 // With that in mind, if 3354 // limit >= 0 (COND) 3355 // then the unsigned loop exit condition 3356 // i >=u limit (ULE) 3357 // is equivalent to 3358 // i < 0 || i >= limit (SLE-full) 3359 // because either i is negative and therefore always greater than MAX_INT when converting to unsigned 3360 // (uint) i >=u MAX_INT >= limit >= 0 3361 // or otherwise 3362 // i >= limit >= 0 3363 // holds due to (LEMMA). 3364 // 3365 // For completeness, a counterexample with limit < 0: 3366 // Assume i = -3 and limit = -2: 3367 // i < 0 3368 // -2 < 0 3369 // is true and thus also "i < 0 || i >= limit". But 3370 // i >=u limit 3371 // -3 >=u -2 3372 // is false. 3373 Node* limit = cmpu->in(2); 3374 const TypeInt* type_limit = _igvn.type(limit)->is_int(); 3375 if (type_limit->_lo < 0) { 3376 return nullptr; 3377 } 3378 3379 // We prove below that we can extract a single signed loop exit condition from (SLE-full), depending on the stride: 3380 // stride < 0: 3381 // i < 0 (SLE = SLE-negative) 3382 // stride > 0: 3383 // i >= limit (SLE = SLE-positive) 3384 // such that we have the following graph before Partial Peeling with stride > 0 (similar for stride < 0): 3385 // 3386 // Loop: 3387 // <peeled section> 3388 // i >= limit (SLE-positive) 3389 // <-- CUT HERE --> 3390 // i >=u limit (ULE) 3391 // <rest of unpeeled section> 3392 // goto Loop 3393 // 3394 // We exit the loop if: 3395 // (SLE) is true OR (ULE) is true 3396 // However, if (SLE) is true then (ULE) also needs to be true to ensure the exact same behavior. Otherwise, we wrongly 3397 // exit a loop that should not have been exited if we did not apply Partial Peeling. More formally, we need to ensure: 3398 // (SLE) IMPLIES (ULE) 3399 // This indeed holds when (COND) is given: 3400 // - stride > 0: 3401 // i >= limit // (SLE = SLE-positive) 3402 // i >= limit >= 0 // (COND) 3403 // i >=u limit >= 0 // (LEMMA) 3404 // which is the unsigned loop exit condition (ULE). 3405 // - stride < 0: 3406 // i < 0 // (SLE = SLE-negative) 3407 // (uint) i >u MAX_INT // (NEG) all negative values are greater than MAX_INT when converted to unsigned 3408 // MAX_INT >= limit >= 0 // (COND) 3409 // MAX_INT >=u limit >= 0 // (LEMMA) 3410 // and thus from (NEG) and (LEMMA): 3411 // i >=u limit 3412 // which is the unsigned loop exit condition (ULE). 3413 // 3414 // 3415 // After Partial Peeling, we have the following structure for stride > 0 (similar for stride < 0): 3416 // <cloned peeled section> 3417 // i >= limit (SLE-positive) 3418 // Loop: 3419 // i >=u limit (ULE) 3420 // <rest of unpeeled section> 3421 // <peeled section> 3422 // i >= limit (SLE-positive) 3423 // goto Loop 3424 Node* rhs_cmpi; 3425 if (stride > 0) { 3426 rhs_cmpi = limit; // For i >= limit 3427 } else { 3428 rhs_cmpi = makecon(TypeInt::ZERO); // For i < 0 3429 } 3430 // Create a new region on the exit path 3431 RegionNode* reg = insert_region_before_proj(lp_exit); 3432 guarantee(reg != nullptr, "null region node"); 3433 3434 // Clone the if-cmpu-true-false using a signed compare 3435 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge; 3436 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, rhs_cmpi, lp_continue); 3437 reg->add_req(cmpi_exit); 3438 3439 // Clone the if-cmpu-true-false 3440 BoolTest::mask rel_u = bol->_test._test; 3441 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue); 3442 reg->add_req(cmpu_exit); 3443 3444 // Force original if to stay in loop. 3445 short_circuit_if(if_cmpu, lp_continue); 3446 3447 return cmpi_exit->in(0)->as_If(); 3448 } 3449 3450 //------------------------------ remove_cmpi_loop_exit ------------------------------------- 3451 // Remove a previously inserted signed compare loop exit. 3452 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) { 3453 Node* lp_proj = stay_in_loop(if_cmp, loop); 3454 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI && 3455 stay_in_loop(lp_proj, loop)->is_If() && 3456 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu"); 3457 Node* con = makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO); 3458 if_cmp->set_req(1, con); 3459 } 3460 3461 //------------------------------ scheduled_nodelist ------------------------------------- 3462 // Create a post order schedule of nodes that are in the 3463 // "member" set. The list is returned in "sched". 3464 // The first node in "sched" is the loop head, followed by 3465 // nodes which have no inputs in the "member" set, and then 3466 // followed by the nodes that have an immediate input dependence 3467 // on a node in "sched". 3468 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) { 3469 3470 assert(member.test(loop->_head->_idx), "loop head must be in member set"); 3471 VectorSet visited; 3472 Node_Stack nstack(loop->_body.size()); 3473 3474 Node* n = loop->_head; // top of stack is cached in "n" 3475 uint idx = 0; 3476 visited.set(n->_idx); 3477 3478 // Initially push all with no inputs from within member set 3479 for(uint i = 0; i < loop->_body.size(); i++ ) { 3480 Node *elt = loop->_body.at(i); 3481 if (member.test(elt->_idx)) { 3482 bool found = false; 3483 for (uint j = 0; j < elt->req(); j++) { 3484 Node* def = elt->in(j); 3485 if (def && member.test(def->_idx) && def != elt) { 3486 found = true; 3487 break; 3488 } 3489 } 3490 if (!found && elt != loop->_head) { 3491 nstack.push(n, idx); 3492 n = elt; 3493 assert(!visited.test(n->_idx), "not seen yet"); 3494 visited.set(n->_idx); 3495 } 3496 } 3497 } 3498 3499 // traverse out's that are in the member set 3500 while (true) { 3501 if (idx < n->outcnt()) { 3502 Node* use = n->raw_out(idx); 3503 idx++; 3504 if (!visited.test_set(use->_idx)) { 3505 if (member.test(use->_idx)) { 3506 nstack.push(n, idx); 3507 n = use; 3508 idx = 0; 3509 } 3510 } 3511 } else { 3512 // All outputs processed 3513 sched.push(n); 3514 if (nstack.is_empty()) break; 3515 n = nstack.node(); 3516 idx = nstack.index(); 3517 nstack.pop(); 3518 } 3519 } 3520 } 3521 3522 3523 //------------------------------ has_use_in_set ------------------------------------- 3524 // Has a use in the vector set 3525 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) { 3526 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3527 Node* use = n->fast_out(j); 3528 if (vset.test(use->_idx)) { 3529 return true; 3530 } 3531 } 3532 return false; 3533 } 3534 3535 3536 //------------------------------ has_use_internal_to_set ------------------------------------- 3537 // Has use internal to the vector set (ie. not in a phi at the loop head) 3538 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) { 3539 Node* head = loop->_head; 3540 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3541 Node* use = n->fast_out(j); 3542 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) { 3543 return true; 3544 } 3545 } 3546 return false; 3547 } 3548 3549 3550 //------------------------------ clone_for_use_outside_loop ------------------------------------- 3551 // clone "n" for uses that are outside of loop 3552 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) { 3553 int cloned = 0; 3554 assert(worklist.size() == 0, "should be empty"); 3555 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3556 Node* use = n->fast_out(j); 3557 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) { 3558 worklist.push(use); 3559 } 3560 } 3561 3562 if (C->check_node_count(worklist.size() + NodeLimitFudgeFactor, 3563 "Too many clones required in clone_for_use_outside_loop in partial peeling")) { 3564 return -1; 3565 } 3566 3567 while( worklist.size() ) { 3568 Node *use = worklist.pop(); 3569 if (!has_node(use) || use->in(0) == C->top()) continue; 3570 uint j; 3571 for (j = 0; j < use->req(); j++) { 3572 if (use->in(j) == n) break; 3573 } 3574 assert(j < use->req(), "must be there"); 3575 3576 // clone "n" and insert it between the inputs of "n" and the use outside the loop 3577 Node* n_clone = n->clone(); 3578 _igvn.replace_input_of(use, j, n_clone); 3579 cloned++; 3580 Node* use_c; 3581 if (!use->is_Phi()) { 3582 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0); 3583 } else { 3584 // Use in a phi is considered a use in the associated predecessor block 3585 use_c = use->in(0)->in(j); 3586 } 3587 set_ctrl(n_clone, use_c); 3588 assert(!loop->is_member(get_loop(use_c)), "should be outside loop"); 3589 get_loop(use_c)->_body.push(n_clone); 3590 _igvn.register_new_node_with_optimizer(n_clone); 3591 #ifndef PRODUCT 3592 if (TracePartialPeeling) { 3593 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx); 3594 } 3595 #endif 3596 } 3597 return cloned; 3598 } 3599 3600 3601 //------------------------------ clone_for_special_use_inside_loop ------------------------------------- 3602 // clone "n" for special uses that are in the not_peeled region. 3603 // If these def-uses occur in separate blocks, the code generator 3604 // marks the method as not compilable. For example, if a "BoolNode" 3605 // is in a different basic block than the "IfNode" that uses it, then 3606 // the compilation is aborted in the code generator. 3607 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 3608 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) { 3609 if (n->is_Phi() || n->is_Load()) { 3610 return; 3611 } 3612 assert(worklist.size() == 0, "should be empty"); 3613 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3614 Node* use = n->fast_out(j); 3615 if ( not_peel.test(use->_idx) && 3616 (use->is_If() || use->is_CMove() || use->is_Bool() || use->is_OpaqueInitializedAssertionPredicate()) && 3617 use->in(1) == n) { 3618 worklist.push(use); 3619 } 3620 } 3621 if (worklist.size() > 0) { 3622 // clone "n" and insert it between inputs of "n" and the use 3623 Node* n_clone = n->clone(); 3624 loop->_body.push(n_clone); 3625 _igvn.register_new_node_with_optimizer(n_clone); 3626 set_ctrl(n_clone, get_ctrl(n)); 3627 sink_list.push(n_clone); 3628 not_peel.set(n_clone->_idx); 3629 #ifndef PRODUCT 3630 if (TracePartialPeeling) { 3631 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx); 3632 } 3633 #endif 3634 while( worklist.size() ) { 3635 Node *use = worklist.pop(); 3636 _igvn.rehash_node_delayed(use); 3637 for (uint j = 1; j < use->req(); j++) { 3638 if (use->in(j) == n) { 3639 use->set_req(j, n_clone); 3640 } 3641 } 3642 } 3643 } 3644 } 3645 3646 3647 //------------------------------ insert_phi_for_loop ------------------------------------- 3648 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 3649 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) { 3650 Node *phi = PhiNode::make(lp, back_edge_val); 3651 phi->set_req(LoopNode::EntryControl, lp_entry_val); 3652 // Use existing phi if it already exists 3653 Node *hit = _igvn.hash_find_insert(phi); 3654 if( hit == nullptr ) { 3655 _igvn.register_new_node_with_optimizer(phi); 3656 set_ctrl(phi, lp); 3657 } else { 3658 // Remove the new phi from the graph and use the hit 3659 _igvn.remove_dead_node(phi); 3660 phi = hit; 3661 } 3662 _igvn.replace_input_of(use, idx, phi); 3663 } 3664 3665 #ifdef ASSERT 3666 //------------------------------ is_valid_loop_partition ------------------------------------- 3667 // Validate the loop partition sets: peel and not_peel 3668 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, 3669 VectorSet& not_peel ) { 3670 uint i; 3671 // Check that peel_list entries are in the peel set 3672 for (i = 0; i < peel_list.size(); i++) { 3673 if (!peel.test(peel_list.at(i)->_idx)) { 3674 return false; 3675 } 3676 } 3677 // Check at loop members are in one of peel set or not_peel set 3678 for (i = 0; i < loop->_body.size(); i++ ) { 3679 Node *def = loop->_body.at(i); 3680 uint di = def->_idx; 3681 // Check that peel set elements are in peel_list 3682 if (peel.test(di)) { 3683 if (not_peel.test(di)) { 3684 return false; 3685 } 3686 // Must be in peel_list also 3687 bool found = false; 3688 for (uint j = 0; j < peel_list.size(); j++) { 3689 if (peel_list.at(j)->_idx == di) { 3690 found = true; 3691 break; 3692 } 3693 } 3694 if (!found) { 3695 return false; 3696 } 3697 } else if (not_peel.test(di)) { 3698 if (peel.test(di)) { 3699 return false; 3700 } 3701 } else { 3702 return false; 3703 } 3704 } 3705 return true; 3706 } 3707 3708 //------------------------------ is_valid_clone_loop_exit_use ------------------------------------- 3709 // Ensure a use outside of loop is of the right form 3710 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) { 3711 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 3712 return (use->is_Phi() && 3713 use_c->is_Region() && use_c->req() == 3 && 3714 (use_c->in(exit_idx)->Opcode() == Op_IfTrue || 3715 use_c->in(exit_idx)->Opcode() == Op_IfFalse || 3716 use_c->in(exit_idx)->Opcode() == Op_JumpProj) && 3717 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) ); 3718 } 3719 3720 //------------------------------ is_valid_clone_loop_form ------------------------------------- 3721 // Ensure that all uses outside of loop are of the right form 3722 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 3723 uint orig_exit_idx, uint clone_exit_idx) { 3724 uint len = peel_list.size(); 3725 for (uint i = 0; i < len; i++) { 3726 Node *def = peel_list.at(i); 3727 3728 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 3729 Node *use = def->fast_out(j); 3730 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 3731 if (!loop->is_member(get_loop(use_c))) { 3732 // use is not in the loop, check for correct structure 3733 if (use->in(0) == def) { 3734 // Okay 3735 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) { 3736 return false; 3737 } 3738 } 3739 } 3740 } 3741 return true; 3742 } 3743 #endif 3744 3745 //------------------------------ partial_peel ------------------------------------- 3746 // Partially peel (aka loop rotation) the top portion of a loop (called 3747 // the peel section below) by cloning it and placing one copy just before 3748 // the new loop head and the other copy at the bottom of the new loop. 3749 // 3750 // before after where it came from 3751 // 3752 // stmt1 stmt1 3753 // loop: stmt2 clone 3754 // stmt2 if condA goto exitA clone 3755 // if condA goto exitA new_loop: new 3756 // stmt3 stmt3 clone 3757 // if !condB goto loop if condB goto exitB clone 3758 // exitB: stmt2 orig 3759 // stmt4 if !condA goto new_loop orig 3760 // exitA: goto exitA 3761 // exitB: 3762 // stmt4 3763 // exitA: 3764 // 3765 // Step 1: find the cut point: an exit test on probable 3766 // induction variable. 3767 // Step 2: schedule (with cloning) operations in the peel 3768 // section that can be executed after the cut into 3769 // the section that is not peeled. This may need 3770 // to clone operations into exit blocks. For 3771 // instance, a reference to A[i] in the not-peel 3772 // section and a reference to B[i] in an exit block 3773 // may cause a left-shift of i by 2 to be placed 3774 // in the peel block. This step will clone the left 3775 // shift into the exit block and sink the left shift 3776 // from the peel to the not-peel section. 3777 // Step 3: clone the loop, retarget the control, and insert 3778 // phis for values that are live across the new loop 3779 // head. This is very dependent on the graph structure 3780 // from clone_loop. It creates region nodes for 3781 // exit control and associated phi nodes for values 3782 // flow out of the loop through that exit. The region 3783 // node is dominated by the clone's control projection. 3784 // So the clone's peel section is placed before the 3785 // new loop head, and the clone's not-peel section is 3786 // forms the top part of the new loop. The original 3787 // peel section forms the tail of the new loop. 3788 // Step 4: update the dominator tree and recompute the 3789 // dominator depth. 3790 // 3791 // orig 3792 // 3793 // stmt1 3794 // | 3795 // v 3796 // predicates 3797 // | 3798 // v 3799 // loop<----+ 3800 // | | 3801 // stmt2 | 3802 // | | 3803 // v | 3804 // ifA | 3805 // / | | 3806 // v v | 3807 // false true ^ <-- last_peel 3808 // / | | 3809 // / ===|==cut | 3810 // / stmt3 | <-- first_not_peel 3811 // / | | 3812 // | v | 3813 // v ifB | 3814 // exitA: / \ | 3815 // / \ | 3816 // v v | 3817 // false true | 3818 // / \ | 3819 // / ----+ 3820 // | 3821 // v 3822 // exitB: 3823 // stmt4 3824 // 3825 // 3826 // after clone loop 3827 // 3828 // stmt1 3829 // | 3830 // v 3831 // predicates 3832 // / \ 3833 // clone / \ orig 3834 // / \ 3835 // / \ 3836 // v v 3837 // +---->loop loop<----+ 3838 // | | | | 3839 // | stmt2 stmt2 | 3840 // | | | | 3841 // | v v | 3842 // | ifA ifA | 3843 // | | \ / | | 3844 // | v v v v | 3845 // ^ true false false true ^ <-- last_peel 3846 // | | ^ \ / | | 3847 // | cut==|== \ \ / ===|==cut | 3848 // | stmt3 \ \ / stmt3 | <-- first_not_peel 3849 // | | dom | | | | 3850 // | v \ 1v v2 v | 3851 // | ifB regionA ifB | 3852 // | / \ | / \ | 3853 // | / \ v / \ | 3854 // | v v exitA: v v | 3855 // | true false false true | 3856 // | / ^ \ / \ | 3857 // +---- \ \ / ----+ 3858 // dom \ / 3859 // \ 1v v2 3860 // regionB 3861 // | 3862 // v 3863 // exitB: 3864 // stmt4 3865 // 3866 // 3867 // after partial peel 3868 // 3869 // stmt1 3870 // | 3871 // v 3872 // predicates 3873 // / 3874 // clone / orig 3875 // / TOP 3876 // / \ 3877 // v v 3878 // TOP->loop loop----+ 3879 // | | | 3880 // stmt2 stmt2 | 3881 // | | | 3882 // v v | 3883 // ifA ifA | 3884 // | \ / | | 3885 // v v v v | 3886 // true false false true | <-- last_peel 3887 // | ^ \ / +------|---+ 3888 // +->newloop \ \ / === ==cut | | 3889 // | stmt3 \ \ / TOP | | 3890 // | | dom | | stmt3 | | <-- first_not_peel 3891 // | v \ 1v v2 v | | 3892 // | ifB regionA ifB ^ v 3893 // | / \ | / \ | | 3894 // | / \ v / \ | | 3895 // | v v exitA: v v | | 3896 // | true false false true | | 3897 // | / ^ \ / \ | | 3898 // | | \ \ / v | | 3899 // | | dom \ / TOP | | 3900 // | | \ 1v v2 | | 3901 // ^ v regionB | | 3902 // | | | | | 3903 // | | v ^ v 3904 // | | exitB: | | 3905 // | | stmt4 | | 3906 // | +------------>-----------------+ | 3907 // | | 3908 // +-----------------<---------------------+ 3909 // 3910 // 3911 // final graph 3912 // 3913 // stmt1 3914 // | 3915 // v 3916 // predicates 3917 // | 3918 // v 3919 // stmt2 clone 3920 // | 3921 // v 3922 // ........> ifA clone 3923 // : / | 3924 // dom / | 3925 // : v v 3926 // : false true 3927 // : | | 3928 // : | v 3929 // : | newloop<-----+ 3930 // : | | | 3931 // : | stmt3 clone | 3932 // : | | | 3933 // : | v | 3934 // : | ifB | 3935 // : | / \ | 3936 // : | v v | 3937 // : | false true | 3938 // : | | | | 3939 // : | v stmt2 | 3940 // : | exitB: | | 3941 // : | stmt4 v | 3942 // : | ifA orig | 3943 // : | / \ | 3944 // : | / \ | 3945 // : | v v | 3946 // : | false true | 3947 // : | / \ | 3948 // : v v -----+ 3949 // RegionA 3950 // | 3951 // v 3952 // exitA 3953 // 3954 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) { 3955 3956 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only"); 3957 if (!loop->_head->is_Loop()) { 3958 return false; 3959 } 3960 LoopNode *head = loop->_head->as_Loop(); 3961 3962 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) { 3963 return false; 3964 } 3965 3966 // Check for complex exit control 3967 for (uint ii = 0; ii < loop->_body.size(); ii++) { 3968 Node *n = loop->_body.at(ii); 3969 int opc = n->Opcode(); 3970 if (n->is_Call() || 3971 opc == Op_Catch || 3972 opc == Op_CatchProj || 3973 opc == Op_Jump || 3974 opc == Op_JumpProj) { 3975 #ifndef PRODUCT 3976 if (TracePartialPeeling) { 3977 tty->print_cr("\nExit control too complex: lp: %d", head->_idx); 3978 } 3979 #endif 3980 return false; 3981 } 3982 } 3983 3984 int dd = dom_depth(head); 3985 3986 // Step 1: find cut point 3987 3988 // Walk up dominators to loop head looking for first loop exit 3989 // which is executed on every path thru loop. 3990 IfNode *peel_if = nullptr; 3991 IfNode *peel_if_cmpu = nullptr; 3992 3993 Node *iff = loop->tail(); 3994 while (iff != head) { 3995 if (iff->is_If()) { 3996 Node *ctrl = get_ctrl(iff->in(1)); 3997 if (ctrl->is_top()) return false; // Dead test on live IF. 3998 // If loop-varying exit-test, check for induction variable 3999 if (loop->is_member(get_loop(ctrl)) && 4000 loop->is_loop_exit(iff) && 4001 is_possible_iv_test(iff)) { 4002 Node* cmp = iff->in(1)->in(1); 4003 if (cmp->Opcode() == Op_CmpI) { 4004 peel_if = iff->as_If(); 4005 } else { 4006 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU"); 4007 peel_if_cmpu = iff->as_If(); 4008 } 4009 } 4010 } 4011 iff = idom(iff); 4012 } 4013 4014 // Prefer signed compare over unsigned compare. 4015 IfNode* new_peel_if = nullptr; 4016 if (peel_if == nullptr) { 4017 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == nullptr) { 4018 return false; // No peel point found 4019 } 4020 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop); 4021 if (new_peel_if == nullptr) { 4022 return false; // No peel point found 4023 } 4024 peel_if = new_peel_if; 4025 } 4026 Node* last_peel = stay_in_loop(peel_if, loop); 4027 Node* first_not_peeled = stay_in_loop(last_peel, loop); 4028 if (first_not_peeled == nullptr || first_not_peeled == head) { 4029 return false; 4030 } 4031 4032 #ifndef PRODUCT 4033 if (TraceLoopOpts) { 4034 tty->print("PartialPeel "); 4035 loop->dump_head(); 4036 } 4037 4038 if (TracePartialPeeling) { 4039 tty->print_cr("before partial peel one iteration"); 4040 Node_List wl; 4041 Node* t = head->in(2); 4042 while (true) { 4043 wl.push(t); 4044 if (t == head) break; 4045 t = idom(t); 4046 } 4047 while (wl.size() > 0) { 4048 Node* tt = wl.pop(); 4049 tt->dump(); 4050 if (tt == last_peel) tty->print_cr("-- cut --"); 4051 } 4052 } 4053 #endif 4054 4055 C->print_method(PHASE_BEFORE_PARTIAL_PEELING, 4, head); 4056 4057 VectorSet peel; 4058 VectorSet not_peel; 4059 Node_List peel_list; 4060 Node_List worklist; 4061 Node_List sink_list; 4062 4063 uint estimate = loop->est_loop_clone_sz(1); 4064 if (exceeding_node_budget(estimate)) { 4065 return false; 4066 } 4067 4068 // Set of cfg nodes to peel are those that are executable from 4069 // the head through last_peel. 4070 assert(worklist.size() == 0, "should be empty"); 4071 worklist.push(head); 4072 peel.set(head->_idx); 4073 while (worklist.size() > 0) { 4074 Node *n = worklist.pop(); 4075 if (n != last_peel) { 4076 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 4077 Node* use = n->fast_out(j); 4078 if (use->is_CFG() && 4079 loop->is_member(get_loop(use)) && 4080 !peel.test_set(use->_idx)) { 4081 worklist.push(use); 4082 } 4083 } 4084 } 4085 } 4086 4087 // Set of non-cfg nodes to peel are those that are control 4088 // dependent on the cfg nodes. 4089 for (uint i = 0; i < loop->_body.size(); i++) { 4090 Node *n = loop->_body.at(i); 4091 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n; 4092 if (peel.test(n_c->_idx)) { 4093 peel.set(n->_idx); 4094 } else { 4095 not_peel.set(n->_idx); 4096 } 4097 } 4098 4099 // Step 2: move operations from the peeled section down into the 4100 // not-peeled section 4101 4102 // Get a post order schedule of nodes in the peel region 4103 // Result in right-most operand. 4104 scheduled_nodelist(loop, peel, peel_list); 4105 4106 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 4107 4108 // For future check for too many new phis 4109 uint old_phi_cnt = 0; 4110 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 4111 Node* use = head->fast_out(j); 4112 if (use->is_Phi()) old_phi_cnt++; 4113 } 4114 4115 #ifndef PRODUCT 4116 if (TracePartialPeeling) { 4117 tty->print_cr("\npeeled list"); 4118 } 4119 #endif 4120 4121 // Evacuate nodes in peel region into the not_peeled region if possible 4122 bool too_many_clones = false; 4123 uint new_phi_cnt = 0; 4124 uint cloned_for_outside_use = 0; 4125 for (uint i = 0; i < peel_list.size();) { 4126 Node* n = peel_list.at(i); 4127 #ifndef PRODUCT 4128 if (TracePartialPeeling) n->dump(); 4129 #endif 4130 bool incr = true; 4131 if (!n->is_CFG()) { 4132 if (has_use_in_set(n, not_peel)) { 4133 // If not used internal to the peeled region, 4134 // move "n" from peeled to not_peeled region. 4135 if (!has_use_internal_to_set(n, peel, loop)) { 4136 // if not pinned and not a load (which maybe anti-dependent on a store) 4137 // and not a CMove (Matcher expects only bool->cmove). 4138 if (n->in(0) == nullptr && !n->is_Load() && !n->is_CMove()) { 4139 int new_clones = clone_for_use_outside_loop(loop, n, worklist); 4140 if (C->failing()) return false; 4141 if (new_clones == -1) { 4142 too_many_clones = true; 4143 break; 4144 } 4145 cloned_for_outside_use += new_clones; 4146 sink_list.push(n); 4147 peel.remove(n->_idx); 4148 not_peel.set(n->_idx); 4149 peel_list.remove(i); 4150 incr = false; 4151 #ifndef PRODUCT 4152 if (TracePartialPeeling) { 4153 tty->print_cr("sink to not_peeled region: %d newbb: %d", 4154 n->_idx, get_ctrl(n)->_idx); 4155 } 4156 #endif 4157 } 4158 } else { 4159 // Otherwise check for special def-use cases that span 4160 // the peel/not_peel boundary such as bool->if 4161 clone_for_special_use_inside_loop(loop, n, not_peel, sink_list, worklist); 4162 new_phi_cnt++; 4163 } 4164 } 4165 } 4166 if (incr) i++; 4167 } 4168 4169 estimate += cloned_for_outside_use + new_phi_cnt; 4170 bool exceed_node_budget = !may_require_nodes(estimate); 4171 bool exceed_phi_limit = new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta; 4172 4173 if (too_many_clones || exceed_node_budget || exceed_phi_limit) { 4174 #ifndef PRODUCT 4175 if (TracePartialPeeling && exceed_phi_limit) { 4176 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c", 4177 new_phi_cnt, old_phi_cnt, new_peel_if != nullptr?'T':'F'); 4178 } 4179 #endif 4180 if (new_peel_if != nullptr) { 4181 remove_cmpi_loop_exit(new_peel_if, loop); 4182 } 4183 // Inhibit more partial peeling on this loop 4184 assert(!head->is_partial_peel_loop(), "not partial peeled"); 4185 head->mark_partial_peel_failed(); 4186 if (cloned_for_outside_use > 0) { 4187 // Terminate this round of loop opts because 4188 // the graph outside this loop was changed. 4189 C->set_major_progress(); 4190 return true; 4191 } 4192 return false; 4193 } 4194 4195 // Step 3: clone loop, retarget control, and insert new phis 4196 4197 // Create new loop head for new phis and to hang 4198 // the nodes being moved (sinked) from the peel region. 4199 LoopNode* new_head = new LoopNode(last_peel, last_peel); 4200 new_head->set_unswitch_count(head->unswitch_count()); // Preserve 4201 _igvn.register_new_node_with_optimizer(new_head); 4202 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled"); 4203 _igvn.replace_input_of(first_not_peeled, 0, new_head); 4204 set_loop(new_head, loop); 4205 loop->_body.push(new_head); 4206 not_peel.set(new_head->_idx); 4207 set_idom(new_head, last_peel, dom_depth(first_not_peeled)); 4208 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled)); 4209 4210 while (sink_list.size() > 0) { 4211 Node* n = sink_list.pop(); 4212 set_ctrl(n, new_head); 4213 } 4214 4215 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 4216 4217 clone_loop(loop, old_new, dd, IgnoreStripMined); 4218 4219 const uint clone_exit_idx = 1; 4220 const uint orig_exit_idx = 2; 4221 assert(is_valid_clone_loop_form(loop, peel_list, orig_exit_idx, clone_exit_idx), "bad clone loop"); 4222 4223 Node* head_clone = old_new[head->_idx]; 4224 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop(); 4225 Node* orig_tail_clone = head_clone->in(2); 4226 4227 // Add phi if "def" node is in peel set and "use" is not 4228 4229 for (uint i = 0; i < peel_list.size(); i++) { 4230 Node *def = peel_list.at(i); 4231 if (!def->is_CFG()) { 4232 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 4233 Node *use = def->fast_out(j); 4234 if (has_node(use) && use->in(0) != C->top() && 4235 (!peel.test(use->_idx) || 4236 (use->is_Phi() && use->in(0) == head)) ) { 4237 worklist.push(use); 4238 } 4239 } 4240 while( worklist.size() ) { 4241 Node *use = worklist.pop(); 4242 for (uint j = 1; j < use->req(); j++) { 4243 Node* n = use->in(j); 4244 if (n == def) { 4245 4246 // "def" is in peel set, "use" is not in peel set 4247 // or "use" is in the entry boundary (a phi) of the peel set 4248 4249 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use; 4250 4251 if ( loop->is_member(get_loop( use_c )) ) { 4252 // use is in loop 4253 if (old_new[use->_idx] != nullptr) { // null for dead code 4254 Node* use_clone = old_new[use->_idx]; 4255 _igvn.replace_input_of(use, j, C->top()); 4256 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone ); 4257 } 4258 } else { 4259 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format"); 4260 // use is not in the loop, check if the live range includes the cut 4261 Node* lp_if = use_c->in(orig_exit_idx)->in(0); 4262 if (not_peel.test(lp_if->_idx)) { 4263 assert(j == orig_exit_idx, "use from original loop"); 4264 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone ); 4265 } 4266 } 4267 } 4268 } 4269 } 4270 } 4271 } 4272 4273 // Step 3b: retarget control 4274 4275 // Redirect control to the new loop head if a cloned node in 4276 // the not_peeled region has control that points into the peeled region. 4277 // This necessary because the cloned peeled region will be outside 4278 // the loop. 4279 // from to 4280 // cloned-peeled <---+ 4281 // new_head_clone: | <--+ 4282 // cloned-not_peeled in(0) in(0) 4283 // orig-peeled 4284 4285 for (uint i = 0; i < loop->_body.size(); i++) { 4286 Node *n = loop->_body.at(i); 4287 if (!n->is_CFG() && n->in(0) != nullptr && 4288 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) { 4289 Node* n_clone = old_new[n->_idx]; 4290 if (n_clone->depends_only_on_test()) { 4291 // Pin array access nodes: control is updated here to the loop head. If, after some transformations, the 4292 // backedge is removed, an array load could become dependent on a condition that's not a range check for that 4293 // access. If that condition is replaced by an identical dominating one, then an unpinned load would risk 4294 // floating above its range check. 4295 Node* pinned_clone = n_clone->pin_array_access_node(); 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(head, loop); 4489 if (back_control == nullptr) { 4490 return false; 4491 } 4492 4493 BoolTest::mask bt = BoolTest::illegal; 4494 float cl_prob = 0; 4495 Node* incr = nullptr; 4496 Node* limit = nullptr; 4497 Node* cmp = loop_exit_test(back_control, loop, incr, limit, bt, cl_prob); 4498 if (cmp == nullptr || cmp->Opcode() != Op_CmpI) { 4499 return false; 4500 } 4501 4502 // With an extra phi for the candidate iv? 4503 // Or the region node is the loop head 4504 if (!incr->is_Phi() || incr->in(0) == head) { 4505 return false; 4506 } 4507 4508 PathFrequency pf(head, this); 4509 region = incr->in(0); 4510 4511 // Go over all paths for the extra phi's region and see if that 4512 // path is frequent enough and would match the expected iv shape 4513 // if the extra phi is removed 4514 inner = 0; 4515 for (uint i = 1; i < incr->req(); ++i) { 4516 Node* in = incr->in(i); 4517 Node* trunc1 = nullptr; 4518 Node* trunc2 = nullptr; 4519 const TypeInteger* iv_trunc_t = nullptr; 4520 Node* orig_in = in; 4521 if (!(in = CountedLoopNode::match_incr_with_optional_truncation(in, &trunc1, &trunc2, &iv_trunc_t, T_INT))) { 4522 continue; 4523 } 4524 assert(in->Opcode() == Op_AddI, "wrong increment code"); 4525 Node* xphi = nullptr; 4526 Node* stride = loop_iv_stride(in, xphi); 4527 4528 if (stride == nullptr) { 4529 continue; 4530 } 4531 4532 PhiNode* phi = loop_iv_phi(xphi, nullptr, head); 4533 if (phi == nullptr || 4534 (trunc1 == nullptr && phi->in(LoopNode::LoopBackControl) != incr) || 4535 (trunc1 != nullptr && phi->in(LoopNode::LoopBackControl) != trunc1)) { 4536 return false; 4537 } 4538 4539 f = pf.to(region->in(i)); 4540 if (f > 0.5) { 4541 inner = i; 4542 break; 4543 } 4544 } 4545 4546 if (inner == 0) { 4547 return false; 4548 } 4549 4550 exit_test = back_control->in(0)->as_If(); 4551 } 4552 4553 if (idom(region)->is_Catch()) { 4554 return false; 4555 } 4556 4557 // Collect all control nodes that need to be cloned (shared_stmt in the diagram) 4558 Unique_Node_List wq; 4559 wq.push(head->in(LoopNode::LoopBackControl)); 4560 for (uint i = 0; i < wq.size(); i++) { 4561 Node* c = wq.at(i); 4562 assert(get_loop(c) == loop, "not in the right loop?"); 4563 if (c->is_Region()) { 4564 if (c != region) { 4565 for (uint j = 1; j < c->req(); ++j) { 4566 wq.push(c->in(j)); 4567 } 4568 } 4569 } else { 4570 wq.push(c->in(0)); 4571 } 4572 assert(!is_strict_dominator(c, region), "shouldn't go above region"); 4573 } 4574 4575 Node* region_dom = idom(region); 4576 4577 // Can't do the transformation if this would cause a membar pair to 4578 // be split 4579 for (uint i = 0; i < wq.size(); i++) { 4580 Node* c = wq.at(i); 4581 if (c->is_MemBar() && (c->as_MemBar()->trailing_store() || c->as_MemBar()->trailing_load_store())) { 4582 assert(c->as_MemBar()->leading_membar()->trailing_membar() == c, "bad membar pair"); 4583 if (!wq.member(c->as_MemBar()->leading_membar())) { 4584 return false; 4585 } 4586 } 4587 } 4588 C->print_method(PHASE_BEFORE_DUPLICATE_LOOP_BACKEDGE, 4, head); 4589 4590 // Collect data nodes that need to be clones as well 4591 int dd = dom_depth(head); 4592 4593 for (uint i = 0; i < loop->_body.size(); ++i) { 4594 Node* n = loop->_body.at(i); 4595 if (has_ctrl(n)) { 4596 Node* c = get_ctrl(n); 4597 if (wq.member(c)) { 4598 wq.push(n); 4599 } 4600 } else { 4601 set_idom(n, idom(n), dd); 4602 } 4603 } 4604 4605 // clone shared_stmt 4606 clone_loop_body(wq, old_new, nullptr); 4607 4608 Node* region_clone = old_new[region->_idx]; 4609 region_clone->set_req(inner, C->top()); 4610 set_idom(region, region->in(inner), dd); 4611 4612 // Prepare the outer loop 4613 Node* outer_head = new LoopNode(head->in(LoopNode::EntryControl), old_new[head->in(LoopNode::LoopBackControl)->_idx]); 4614 register_control(outer_head, loop->_parent, outer_head->in(LoopNode::EntryControl)); 4615 _igvn.replace_input_of(head, LoopNode::EntryControl, outer_head); 4616 set_idom(head, outer_head, dd); 4617 4618 fix_body_edges(wq, loop, old_new, dd, loop->_parent, true); 4619 4620 // Make one of the shared_stmt copies only reachable from stmt1, the 4621 // other only from stmt2..stmtn. 4622 Node* dom = nullptr; 4623 for (uint i = 1; i < region->req(); ++i) { 4624 if (i != inner) { 4625 _igvn.replace_input_of(region, i, C->top()); 4626 } 4627 Node* in = region_clone->in(i); 4628 if (in->is_top()) { 4629 continue; 4630 } 4631 if (dom == nullptr) { 4632 dom = in; 4633 } else { 4634 dom = dom_lca(dom, in); 4635 } 4636 } 4637 4638 set_idom(region_clone, dom, dd); 4639 4640 // Set up the outer loop 4641 for (uint i = 0; i < head->outcnt(); i++) { 4642 Node* u = head->raw_out(i); 4643 if (u->is_Phi()) { 4644 Node* outer_phi = u->clone(); 4645 outer_phi->set_req(0, outer_head); 4646 Node* backedge = old_new[u->in(LoopNode::LoopBackControl)->_idx]; 4647 if (backedge == nullptr) { 4648 backedge = u->in(LoopNode::LoopBackControl); 4649 } 4650 outer_phi->set_req(LoopNode::LoopBackControl, backedge); 4651 register_new_node(outer_phi, outer_head); 4652 _igvn.replace_input_of(u, LoopNode::EntryControl, outer_phi); 4653 } 4654 } 4655 4656 // create control and data nodes for out of loop uses (including region2) 4657 Node_List worklist; 4658 uint new_counter = C->unique(); 4659 fix_ctrl_uses(wq, loop, old_new, ControlAroundStripMined, outer_head, nullptr, worklist); 4660 4661 Node_List *split_if_set = nullptr; 4662 Node_List *split_bool_set = nullptr; 4663 Node_List *split_cex_set = nullptr; 4664 fix_data_uses(wq, loop, ControlAroundStripMined, loop->skip_strip_mined(), new_counter, old_new, worklist, 4665 split_if_set, split_bool_set, split_cex_set); 4666 4667 finish_clone_loop(split_if_set, split_bool_set, split_cex_set); 4668 4669 if (exit_test != nullptr) { 4670 float cnt = exit_test->_fcnt; 4671 if (cnt != COUNT_UNKNOWN) { 4672 exit_test->_fcnt = cnt * f; 4673 old_new[exit_test->_idx]->as_If()->_fcnt = cnt * (1 - f); 4674 } 4675 } 4676 4677 #ifdef ASSERT 4678 if (StressDuplicateBackedge && head->is_CountedLoop()) { 4679 // The Template Assertion Predicates from the old counted loop are now at the new outer loop - clone them to 4680 // the inner counted loop and kill the old ones. We only need to do this with debug builds because 4681 // StressDuplicateBackedge is a devlop flag and false by default. Without StressDuplicateBackedge 'head' will be a 4682 // non-counted loop, and thus we have no Template Assertion Predicates above the old loop to move down. 4683 PredicateIterator predicate_iterator(outer_head->in(LoopNode::EntryControl)); 4684 NodeInSingleLoopBody node_in_body(this, loop); 4685 MoveAssertionPredicatesVisitor move_assertion_predicates_visitor(head, node_in_body, this); 4686 predicate_iterator.for_each(move_assertion_predicates_visitor); 4687 } 4688 #endif // ASSERT 4689 4690 C->set_major_progress(); 4691 4692 C->print_method(PHASE_AFTER_DUPLICATE_LOOP_BACKEDGE, 4, outer_head); 4693 4694 return true; 4695 } 4696 4697 // AutoVectorize the loop: replace scalar ops with vector ops. 4698 PhaseIdealLoop::AutoVectorizeStatus 4699 PhaseIdealLoop::auto_vectorize(IdealLoopTree* lpt, VSharedData &vshared) { 4700 // Counted loop only 4701 if (!lpt->is_counted()) { 4702 return AutoVectorizeStatus::Impossible; 4703 } 4704 4705 // Main-loop only 4706 CountedLoopNode* cl = lpt->_head->as_CountedLoop(); 4707 if (!cl->is_main_loop()) { 4708 return AutoVectorizeStatus::Impossible; 4709 } 4710 4711 VLoop vloop(lpt, false); 4712 if (!vloop.check_preconditions()) { 4713 return AutoVectorizeStatus::TriedAndFailed; 4714 } 4715 4716 // Ensure the shared data is cleared before each use 4717 vshared.clear(); 4718 4719 const VLoopAnalyzer vloop_analyzer(vloop, vshared); 4720 if (!vloop_analyzer.success()) { 4721 return AutoVectorizeStatus::TriedAndFailed; 4722 } 4723 4724 SuperWord sw(vloop_analyzer); 4725 if (!sw.transform_loop()) { 4726 return AutoVectorizeStatus::TriedAndFailed; 4727 } 4728 4729 return AutoVectorizeStatus::Success; 4730 } 4731 4732 // Just before insert_pre_post_loops, we can multiversion the loop: 4733 // 4734 // multiversion_if 4735 // | | 4736 // fast_loop slow_loop 4737 // 4738 // In the fast_loop we can make speculative assumptions, and put the 4739 // conditions into the multiversion_if. If the conditions hold at runtime, 4740 // we enter the fast_loop, if the conditions fail, we take the slow_loop 4741 // instead which does not make any of the speculative assumptions. 4742 // 4743 // Note: we only multiversion the loop if the loop does not have any 4744 // auto vectorization check Predicate. If we have that predicate, 4745 // then we can simply add the speculative assumption checks to 4746 // that Predicate. This means we do not need to duplicate the 4747 // loop - we have a smaller graph and save compile time. Should 4748 // the conditions ever fail, then we deopt / trap at the Predicate 4749 // and recompile without that Predicate. At that point we will 4750 // multiversion the loop, so that we can still have speculative 4751 // runtime checks. 4752 // 4753 // We perform the multiversioning when the loop is still in its single 4754 // iteration form, even before we insert pre and post loops. This makes 4755 // the cloning much simpler. However, this means that both the fast 4756 // and the slow loop have to be optimized independently (adding pre 4757 // and post loops, unrolling the main loop, auto-vectorize etc.). And 4758 // we may end up not needing any speculative assumptions in the fast_loop 4759 // and then rejecting the slow_loop by constant folding the multiversion_if. 4760 // 4761 // Therefore, we "delay" the optimization of the slow_loop until we add 4762 // at least one speculative assumption for the fast_loop. If we never 4763 // add such a speculative runtime check, the OpaqueMultiversioningNode 4764 // of the multiversion_if constant folds to true after loop opts, and the 4765 // multiversion_if folds away the "delayed" slow_loop. If we add any 4766 // speculative assumption, then we notify the OpaqueMultiversioningNode 4767 // with "notify_slow_loop_that_it_can_resume_optimizations". 4768 // 4769 // Note: new runtime checks can be added to the multiversion_if with 4770 // PhaseIdealLoop::create_new_if_for_multiversion 4771 void PhaseIdealLoop::maybe_multiversion_for_auto_vectorization_runtime_checks(IdealLoopTree* lpt, Node_List& old_new) { 4772 CountedLoopNode* cl = lpt->_head->as_CountedLoop(); 4773 LoopNode* outer_loop = cl->skip_strip_mined(); 4774 Node* entry = outer_loop->in(LoopNode::EntryControl); 4775 4776 // Check we have multiversioning enabled, and are not already multiversioned. 4777 if (!LoopMultiversioning || cl->is_multiversion()) { return; } 4778 4779 // Check that we do not have a parse-predicate where we can add the runtime checks 4780 // during auto-vectorization. 4781 const Predicates predicates(entry); 4782 const PredicateBlock* predicate_block = predicates.auto_vectorization_check_block(); 4783 if (predicate_block->has_parse_predicate()) { return; } 4784 4785 // Check node budget. 4786 uint estimate = lpt->est_loop_clone_sz(2); 4787 if (!may_require_nodes(estimate)) { return; } 4788 4789 do_multiversioning(lpt, old_new); 4790 } 4791 4792 void DataNodeGraph::clone_data_nodes(Node* new_ctrl) { 4793 for (uint i = 0; i < _data_nodes.size(); i++) { 4794 clone(_data_nodes[i], new_ctrl); 4795 } 4796 } 4797 4798 // Clone the given node and set it up properly. Set 'new_ctrl' as ctrl. 4799 void DataNodeGraph::clone(Node* node, Node* new_ctrl) { 4800 Node* clone = node->clone(); 4801 _phase->igvn().register_new_node_with_optimizer(clone); 4802 _orig_to_new.put(node, clone); 4803 _phase->set_ctrl(clone, new_ctrl); 4804 if (node->is_CastII()) { 4805 clone->set_req(0, new_ctrl); 4806 } 4807 } 4808 4809 // Rewire the data inputs of all (unprocessed) cloned nodes, whose inputs are still pointing to the same inputs as their 4810 // corresponding orig nodes, to the newly cloned inputs to create a separate cloned graph. 4811 void DataNodeGraph::rewire_clones_to_cloned_inputs() { 4812 _orig_to_new.iterate_all([&](Node* node, Node* clone) { 4813 for (uint i = 1; i < node->req(); i++) { 4814 Node** cloned_input = _orig_to_new.get(node->in(i)); 4815 if (cloned_input != nullptr) { 4816 // Input was also cloned -> rewire clone to the cloned input. 4817 _phase->igvn().replace_input_of(clone, i, *cloned_input); 4818 } 4819 } 4820 }); 4821 } 4822 4823 // Clone all non-OpaqueLoop* nodes and apply the provided transformation strategy for OpaqueLoop* nodes. 4824 // Set 'new_ctrl' as ctrl for all cloned non-OpaqueLoop* nodes. 4825 void DataNodeGraph::clone_data_nodes_and_transform_opaque_loop_nodes( 4826 const TransformStrategyForOpaqueLoopNodes& transform_strategy, 4827 Node* new_ctrl) { 4828 for (uint i = 0; i < _data_nodes.size(); i++) { 4829 Node* data_node = _data_nodes[i]; 4830 if (data_node->is_Opaque1()) { 4831 transform_opaque_node(transform_strategy, data_node); 4832 } else { 4833 clone(data_node, new_ctrl); 4834 } 4835 } 4836 } 4837 4838 void DataNodeGraph::transform_opaque_node(const TransformStrategyForOpaqueLoopNodes& transform_strategy, Node* node) { 4839 Node* transformed_node; 4840 if (node->is_OpaqueLoopInit()) { 4841 transformed_node = transform_strategy.transform_opaque_init(node->as_OpaqueLoopInit()); 4842 } else { 4843 assert(node->is_OpaqueLoopStride(), "must be OpaqueLoopStrideNode"); 4844 transformed_node = transform_strategy.transform_opaque_stride(node->as_OpaqueLoopStride()); 4845 } 4846 // Add an orig->new mapping to correctly update the inputs of the copied graph in rewire_clones_to_cloned_inputs(). 4847 _orig_to_new.put(node, transformed_node); 4848 } --- EOF ---