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 (is_member(n_loop,get_ctrl(V))) { 648 } else { 649 Node *tmp = V; V = I; I = tmp; 650 } 651 if (!is_member(n_loop,get_ctrl(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 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 952 IdealLoopTree* value_loop = get_loop(get_ctrl(value)); 953 954 // - address and value must be loop invariant 955 // - memory must be a memory Phi for the loop 956 // - Store must be the only store on this memory slice in the 957 // loop: if there's another store following this one then value 958 // written at iteration i by the second store could be overwritten 959 // at iteration i+n by the first store: it's not safe to move the 960 // first store out of the loop 961 // - nothing must observe the memory Phi: it guarantees no read 962 // before the store, we are also guaranteed the store post 963 // dominates the loop head (ignoring a possible early 964 // exit). Otherwise there would be extra Phi involved between the 965 // loop's Phi and the store. 966 // - there must be no early exit from the loop before the Store 967 // (such an exit most of the time would be an extra use of the 968 // memory Phi but sometimes is a bottom memory Phi that takes the 969 // store as input). 970 971 if (!n_loop->is_member(address_loop) && 972 !n_loop->is_member(value_loop) && 973 mem->is_Phi() && mem->in(0) == n_loop->_head && 974 mem->outcnt() == 1 && 975 mem->in(LoopNode::LoopBackControl) == n) { 976 977 assert(n_loop->_tail != nullptr, "need a tail"); 978 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop"); 979 980 // Verify that there's no early exit of the loop before the store. 981 bool ctrl_ok = false; 982 { 983 // Follow control from loop head until n, we exit the loop or 984 // we reach the tail 985 ResourceMark rm; 986 Unique_Node_List wq; 987 wq.push(n_loop->_head); 988 989 for (uint next = 0; next < wq.size(); ++next) { 990 Node *m = wq.at(next); 991 if (m == n->in(0)) { 992 ctrl_ok = true; 993 continue; 994 } 995 assert(!has_ctrl(m), "should be CFG"); 996 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) { 997 ctrl_ok = false; 998 break; 999 } 1000 enqueue_cfg_uses(m, wq); 1001 if (wq.size() > 10) { 1002 ctrl_ok = false; 1003 break; 1004 } 1005 } 1006 } 1007 if (ctrl_ok) { 1008 // move the Store 1009 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem); 1010 _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl)); 1011 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl)); 1012 // Disconnect the phi now. An empty phi can confuse other 1013 // optimizations in this pass of loop opts. 1014 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl)); 1015 n_loop->_body.yank(mem); 1016 1017 set_ctrl_and_loop(n, n->in(0)); 1018 1019 return n; 1020 } 1021 } 1022 } 1023 return nullptr; 1024 } 1025 1026 // Try moving a store out of a loop, right after the loop 1027 void PhaseIdealLoop::try_move_store_after_loop(Node* n) { 1028 if (n->is_Store() && n->in(0) != nullptr) { 1029 Node *n_ctrl = get_ctrl(n); 1030 IdealLoopTree *n_loop = get_loop(n_ctrl); 1031 // Store must be in a loop 1032 if (n_loop != _ltree_root && !n_loop->_irreducible) { 1033 Node* address = n->in(MemNode::Address); 1034 Node* value = n->in(MemNode::ValueIn); 1035 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 1036 // address must be loop invariant 1037 if (!n_loop->is_member(address_loop)) { 1038 // Store must be last on this memory slice in the loop and 1039 // nothing in the loop must observe it 1040 Node* phi = nullptr; 1041 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1042 Node* u = n->fast_out(i); 1043 if (has_ctrl(u)) { // control use? 1044 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 1045 if (!n_loop->is_member(u_loop)) { 1046 continue; 1047 } 1048 if (u->is_Phi() && u->in(0) == n_loop->_head) { 1049 assert(_igvn.type(u) == Type::MEMORY, "bad phi"); 1050 // multiple phis on the same slice are possible 1051 if (phi != nullptr) { 1052 return; 1053 } 1054 phi = u; 1055 continue; 1056 } 1057 } 1058 return; 1059 } 1060 if (phi != nullptr) { 1061 // Nothing in the loop before the store (next iteration) 1062 // must observe the stored value 1063 bool mem_ok = true; 1064 { 1065 ResourceMark rm; 1066 Unique_Node_List wq; 1067 wq.push(phi); 1068 for (uint next = 0; next < wq.size() && mem_ok; ++next) { 1069 Node *m = wq.at(next); 1070 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) { 1071 Node* u = m->fast_out(i); 1072 if (u->is_Store() || u->is_Phi()) { 1073 if (u != n) { 1074 wq.push(u); 1075 mem_ok = (wq.size() <= 10); 1076 } 1077 } else { 1078 mem_ok = false; 1079 break; 1080 } 1081 } 1082 } 1083 } 1084 if (mem_ok) { 1085 // Move the store out of the loop if the LCA of all 1086 // users (except for the phi) is outside the loop. 1087 Node* hook = new Node(1); 1088 hook->init_req(0, n_ctrl); // Add an input to prevent hook from being dead 1089 _igvn.rehash_node_delayed(phi); 1090 int count = phi->replace_edge(n, hook, &_igvn); 1091 assert(count > 0, "inconsistent phi"); 1092 1093 // Compute latest point this store can go 1094 Node* lca = get_late_ctrl(n, get_ctrl(n)); 1095 if (lca->is_OuterStripMinedLoop()) { 1096 lca = lca->in(LoopNode::EntryControl); 1097 } 1098 if (n_loop->is_member(get_loop(lca))) { 1099 // LCA is in the loop - bail out 1100 _igvn.replace_node(hook, n); 1101 return; 1102 } 1103 #ifdef ASSERT 1104 if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) { 1105 assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined"); 1106 n_loop->_head->as_Loop()->verify_strip_mined(1); 1107 Node* outer = n_loop->_head->as_CountedLoop()->outer_loop(); 1108 IdealLoopTree* outer_loop = get_loop(outer); 1109 assert(n_loop->_parent == outer_loop, "broken loop tree"); 1110 assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state"); 1111 } 1112 #endif 1113 lca = place_outside_loop(lca, n_loop); 1114 assert(!n_loop->is_member(get_loop(lca)), "control must not be back in the loop"); 1115 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"); 1116 1117 // Move store out of the loop 1118 _igvn.replace_node(hook, n->in(MemNode::Memory)); 1119 _igvn.replace_input_of(n, 0, lca); 1120 set_ctrl_and_loop(n, lca); 1121 1122 // Disconnect the phi now. An empty phi can confuse other 1123 // optimizations in this pass of loop opts.. 1124 if (phi->in(LoopNode::LoopBackControl) == phi) { 1125 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl)); 1126 n_loop->_body.yank(phi); 1127 } 1128 } 1129 } 1130 } 1131 } 1132 } 1133 } 1134 1135 // We can't use immutable memory for the flat array check because we are loading the mark word which is 1136 // mutable. Although the bits we are interested in are immutable (we check for markWord::unlocked_value), 1137 // we need to use raw memory to not break anti dependency analysis. Below code will attempt to still move 1138 // flat array checks out of loops, mainly to enable loop unswitching. 1139 void PhaseIdealLoop::move_flat_array_check_out_of_loop(Node* n) { 1140 // Skip checks for more than one array 1141 if (n->req() > 3) { 1142 return; 1143 } 1144 Node* mem = n->in(FlatArrayCheckNode::Memory); 1145 Node* array = n->in(FlatArrayCheckNode::ArrayOrKlass)->uncast(); 1146 IdealLoopTree* check_loop = get_loop(get_ctrl(n)); 1147 IdealLoopTree* ary_loop = get_loop(get_ctrl(array)); 1148 1149 // Check if array is loop invariant 1150 if (!check_loop->is_member(ary_loop)) { 1151 // Walk up memory graph from the check until we leave the loop 1152 VectorSet wq; 1153 wq.set(mem->_idx); 1154 while (check_loop->is_member(get_loop(ctrl_or_self(mem)))) { 1155 if (mem->is_Phi()) { 1156 mem = mem->in(1); 1157 } else if (mem->is_MergeMem()) { 1158 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw); 1159 } else if (mem->is_Proj()) { 1160 mem = mem->in(0); 1161 } else if (mem->is_MemBar() || mem->is_SafePoint()) { 1162 mem = mem->in(TypeFunc::Memory); 1163 } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) { 1164 mem = mem->in(MemNode::Memory); 1165 } else { 1166 #ifdef ASSERT 1167 mem->dump(); 1168 #endif 1169 ShouldNotReachHere(); 1170 } 1171 if (wq.test_set(mem->_idx)) { 1172 return; 1173 } 1174 } 1175 // Replace memory input and re-compute ctrl to move the check out of the loop 1176 _igvn.replace_input_of(n, 1, mem); 1177 set_ctrl_and_loop(n, get_early_ctrl(n)); 1178 Node* bol = n->unique_out(); 1179 set_ctrl_and_loop(bol, get_early_ctrl(bol)); 1180 } 1181 } 1182 1183 //------------------------------split_if_with_blocks_pre----------------------- 1184 // Do the real work in a non-recursive function. Data nodes want to be 1185 // cloned in the pre-order so they can feed each other nicely. 1186 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) { 1187 // Cloning these guys is unlikely to win 1188 int n_op = n->Opcode(); 1189 if (n_op == Op_MergeMem) { 1190 return n; 1191 } 1192 if (n->is_Proj()) { 1193 return n; 1194 } 1195 1196 if (n->isa_FlatArrayCheck()) { 1197 move_flat_array_check_out_of_loop(n); 1198 return n; 1199 } 1200 1201 // Do not clone-up CmpFXXX variations, as these are always 1202 // followed by a CmpI 1203 if (n->is_Cmp()) { 1204 return n; 1205 } 1206 // Attempt to use a conditional move instead of a phi/branch 1207 if (ConditionalMoveLimit > 0 && n_op == Op_Region) { 1208 Node *cmov = conditional_move( n ); 1209 if (cmov) { 1210 return cmov; 1211 } 1212 } 1213 if (n->is_CFG() || n->is_LoadStore()) { 1214 return n; 1215 } 1216 if (n->is_Opaque1()) { // Opaque nodes cannot be mod'd 1217 if (!C->major_progress()) { // If chance of no more loop opts... 1218 _igvn._worklist.push(n); // maybe we'll remove them 1219 } 1220 return n; 1221 } 1222 1223 if (n->is_Con()) { 1224 return n; // No cloning for Con nodes 1225 } 1226 1227 Node *n_ctrl = get_ctrl(n); 1228 if (!n_ctrl) { 1229 return n; // Dead node 1230 } 1231 1232 Node* res = try_move_store_before_loop(n, n_ctrl); 1233 if (res != nullptr) { 1234 return n; 1235 } 1236 1237 // Attempt to remix address expressions for loop invariants 1238 Node *m = remix_address_expressions( n ); 1239 if( m ) return m; 1240 1241 if (n_op == Op_AddI) { 1242 Node *nn = convert_add_to_muladd( n ); 1243 if ( nn ) return nn; 1244 } 1245 1246 if (n->is_ConstraintCast()) { 1247 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this); 1248 // ConstraintCastNode::dominating_cast() uses node control input to determine domination. 1249 // Node control inputs don't necessarily agree with loop control info (due to 1250 // transformations happened in between), thus additional dominance check is needed 1251 // to keep loop info valid. 1252 if (dom_cast != nullptr && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) { 1253 _igvn.replace_node(n, dom_cast); 1254 return dom_cast; 1255 } 1256 } 1257 1258 // Determine if the Node has inputs from some local Phi. 1259 // Returns the block to clone thru. 1260 Node *n_blk = has_local_phi_input( n ); 1261 if( !n_blk ) return n; 1262 1263 // Do not clone the trip counter through on a CountedLoop 1264 // (messes up the canonical shape). 1265 if (((n_blk->is_CountedLoop() || (n_blk->is_Loop() && n_blk->as_Loop()->is_loop_nest_inner_loop())) && n->Opcode() == Op_AddI) || 1266 (n_blk->is_LongCountedLoop() && n->Opcode() == Op_AddL)) { 1267 return n; 1268 } 1269 // Pushing a shift through the iv Phi can get in the way of addressing optimizations or range check elimination 1270 if (n_blk->is_BaseCountedLoop() && n->Opcode() == Op_LShift(n_blk->as_BaseCountedLoop()->bt()) && 1271 n->in(1) == n_blk->as_BaseCountedLoop()->phi()) { 1272 return n; 1273 } 1274 1275 // Check for having no control input; not pinned. Allow 1276 // dominating control. 1277 if (n->in(0)) { 1278 Node *dom = idom(n_blk); 1279 if (dom_lca(n->in(0), dom) != n->in(0)) { 1280 return n; 1281 } 1282 } 1283 // Policy: when is it profitable. You must get more wins than 1284 // policy before it is considered profitable. Policy is usually 0, 1285 // so 1 win is considered profitable. Big merges will require big 1286 // cloning, so get a larger policy. 1287 int policy = n_blk->req() >> 2; 1288 1289 // If the loop is a candidate for range check elimination, 1290 // delay splitting through it's phi until a later loop optimization 1291 if (n_blk->is_BaseCountedLoop()) { 1292 IdealLoopTree *lp = get_loop(n_blk); 1293 if (lp && lp->_rce_candidate) { 1294 return n; 1295 } 1296 } 1297 1298 if (must_throttle_split_if()) return n; 1299 1300 // Split 'n' through the merge point if it is profitable, replacing it with a new phi. 1301 Node* phi = split_thru_phi(n, n_blk, policy); 1302 if (phi == nullptr) { return n; } 1303 1304 // Moved a load around the loop, 'en-registering' something. 1305 if (n_blk->is_Loop() && n->is_Load() && 1306 !phi->in(LoopNode::LoopBackControl)->is_Load()) 1307 C->set_major_progress(); 1308 1309 return phi; 1310 } 1311 1312 static bool merge_point_too_heavy(Compile* C, Node* region) { 1313 // Bail out if the region and its phis have too many users. 1314 int weight = 0; 1315 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1316 weight += region->fast_out(i)->outcnt(); 1317 } 1318 int nodes_left = C->max_node_limit() - C->live_nodes(); 1319 if (weight * 8 > nodes_left) { 1320 if (PrintOpto) { 1321 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight); 1322 } 1323 return true; 1324 } else { 1325 return false; 1326 } 1327 } 1328 1329 static bool merge_point_safe(Node* region) { 1330 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode 1331 // having a PhiNode input. This sidesteps the dangerous case where the split 1332 // ConvI2LNode may become TOP if the input Value() does not 1333 // overlap the ConvI2L range, leaving a node which may not dominate its 1334 // uses. 1335 // A better fix for this problem can be found in the BugTraq entry, but 1336 // expediency for Mantis demands this hack. 1337 #ifdef _LP64 1338 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1339 Node* n = region->fast_out(i); 1340 if (n->is_Phi()) { 1341 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1342 Node* m = n->fast_out(j); 1343 if (m->Opcode() == Op_ConvI2L) 1344 return false; 1345 if (m->is_CastII()) { 1346 return false; 1347 } 1348 } 1349 } 1350 } 1351 #endif 1352 return true; 1353 } 1354 1355 1356 //------------------------------place_outside_loop--------------------------------- 1357 // Place some computation outside of this loop on the path to the use passed as argument 1358 Node* PhaseIdealLoop::place_outside_loop(Node* useblock, IdealLoopTree* loop) const { 1359 Node* head = loop->_head; 1360 assert(!loop->is_member(get_loop(useblock)), "must be outside loop"); 1361 if (head->is_Loop() && head->as_Loop()->is_strip_mined()) { 1362 loop = loop->_parent; 1363 assert(loop->_head->is_OuterStripMinedLoop(), "malformed strip mined loop"); 1364 } 1365 1366 // Pick control right outside the loop 1367 for (;;) { 1368 Node* dom = idom(useblock); 1369 if (loop->is_member(get_loop(dom))) { 1370 break; 1371 } 1372 useblock = dom; 1373 } 1374 assert(find_non_split_ctrl(useblock) == useblock, "should be non split control"); 1375 return useblock; 1376 } 1377 1378 1379 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) { 1380 if (!n->is_If() || n->is_BaseCountedLoopEnd()) { 1381 return false; 1382 } 1383 if (!n->in(0)->is_Region()) { 1384 return false; 1385 } 1386 1387 Node* region = n->in(0); 1388 Node* dom = idom(region); 1389 if (!dom->is_If() || !n->as_If()->same_condition(dom, &_igvn)) { 1390 return false; 1391 } 1392 IfNode* dom_if = dom->as_If(); 1393 Node* proj_true = dom_if->proj_out(1); 1394 Node* proj_false = dom_if->proj_out(0); 1395 1396 for (uint i = 1; i < region->req(); i++) { 1397 if (is_dominator(proj_true, region->in(i))) { 1398 continue; 1399 } 1400 if (is_dominator(proj_false, region->in(i))) { 1401 continue; 1402 } 1403 return false; 1404 } 1405 1406 return true; 1407 } 1408 1409 1410 bool PhaseIdealLoop::can_split_if(Node* n_ctrl) { 1411 if (must_throttle_split_if()) { 1412 return false; 1413 } 1414 1415 // Do not do 'split-if' if irreducible loops are present. 1416 if (_has_irreducible_loops) { 1417 return false; 1418 } 1419 1420 if (merge_point_too_heavy(C, n_ctrl)) { 1421 return false; 1422 } 1423 1424 // Do not do 'split-if' if some paths are dead. First do dead code 1425 // elimination and then see if its still profitable. 1426 for (uint i = 1; i < n_ctrl->req(); i++) { 1427 if (n_ctrl->in(i) == C->top()) { 1428 return false; 1429 } 1430 } 1431 1432 // If trying to do a 'Split-If' at the loop head, it is only 1433 // profitable if the cmp folds up on BOTH paths. Otherwise we 1434 // risk peeling a loop forever. 1435 1436 // CNC - Disabled for now. Requires careful handling of loop 1437 // body selection for the cloned code. Also, make sure we check 1438 // for any input path not being in the same loop as n_ctrl. For 1439 // irreducible loops we cannot check for 'n_ctrl->is_Loop()' 1440 // because the alternative loop entry points won't be converted 1441 // into LoopNodes. 1442 IdealLoopTree *n_loop = get_loop(n_ctrl); 1443 for (uint j = 1; j < n_ctrl->req(); j++) { 1444 if (get_loop(n_ctrl->in(j)) != n_loop) { 1445 return false; 1446 } 1447 } 1448 1449 // Check for safety of the merge point. 1450 if (!merge_point_safe(n_ctrl)) { 1451 return false; 1452 } 1453 1454 return true; 1455 } 1456 1457 // Detect if the node is the inner strip-mined loop 1458 // Return: null if it's not the case, or the exit of outer strip-mined loop 1459 static Node* is_inner_of_stripmined_loop(const Node* out) { 1460 Node* out_le = nullptr; 1461 1462 if (out->is_CountedLoopEnd()) { 1463 const CountedLoopNode* loop = out->as_CountedLoopEnd()->loopnode(); 1464 1465 if (loop != nullptr && loop->is_strip_mined()) { 1466 out_le = loop->in(LoopNode::EntryControl)->as_OuterStripMinedLoop()->outer_loop_exit(); 1467 } 1468 } 1469 1470 return out_le; 1471 } 1472 1473 bool PhaseIdealLoop::flat_array_element_type_check(Node *n) { 1474 // If the CmpP is a subtype check for a value that has just been 1475 // loaded from an array, the subtype check guarantees the value 1476 // can't be stored in a flat array and the load of the value 1477 // happens with a flat array check then: push the type check 1478 // through the phi of the flat array check. This needs special 1479 // logic because the subtype check's input is not a phi but a 1480 // LoadKlass that must first be cloned through the phi. 1481 if (n->Opcode() != Op_CmpP) { 1482 return false; 1483 } 1484 1485 Node* klassptr = n->in(1); 1486 Node* klasscon = n->in(2); 1487 1488 if (klassptr->is_DecodeNarrowPtr()) { 1489 klassptr = klassptr->in(1); 1490 } 1491 1492 if (klassptr->Opcode() != Op_LoadKlass && klassptr->Opcode() != Op_LoadNKlass) { 1493 return false; 1494 } 1495 1496 if (!klasscon->is_Con()) { 1497 return false; 1498 } 1499 1500 Node* addr = klassptr->in(MemNode::Address); 1501 1502 if (!addr->is_AddP()) { 1503 return false; 1504 } 1505 1506 intptr_t offset; 1507 Node* obj = AddPNode::Ideal_base_and_offset(addr, &_igvn, offset); 1508 1509 if (obj == nullptr) { 1510 return false; 1511 } 1512 1513 assert(obj != nullptr && addr->in(AddPNode::Base) == addr->in(AddPNode::Address), "malformed AddP?"); 1514 if (obj->Opcode() == Op_CastPP) { 1515 obj = obj->in(1); 1516 } 1517 1518 if (!obj->is_Phi()) { 1519 return false; 1520 } 1521 1522 Node* region = obj->in(0); 1523 1524 Node* phi = PhiNode::make_blank(region, n->in(1)); 1525 for (uint i = 1; i < region->req(); i++) { 1526 Node* in = obj->in(i); 1527 Node* ctrl = region->in(i); 1528 if (addr->in(AddPNode::Base) != obj) { 1529 Node* cast = addr->in(AddPNode::Base); 1530 assert(cast->Opcode() == Op_CastPP && cast->in(0) != nullptr, "inconsistent subgraph"); 1531 Node* cast_clone = cast->clone(); 1532 cast_clone->set_req(0, ctrl); 1533 cast_clone->set_req(1, in); 1534 register_new_node(cast_clone, ctrl); 1535 const Type* tcast = cast_clone->Value(&_igvn); 1536 _igvn.set_type(cast_clone, tcast); 1537 cast_clone->as_Type()->set_type(tcast); 1538 in = cast_clone; 1539 } 1540 Node* addr_clone = addr->clone(); 1541 addr_clone->set_req(AddPNode::Base, in); 1542 addr_clone->set_req(AddPNode::Address, in); 1543 register_new_node(addr_clone, ctrl); 1544 _igvn.set_type(addr_clone, addr_clone->Value(&_igvn)); 1545 Node* klassptr_clone = klassptr->clone(); 1546 klassptr_clone->set_req(2, addr_clone); 1547 register_new_node(klassptr_clone, ctrl); 1548 _igvn.set_type(klassptr_clone, klassptr_clone->Value(&_igvn)); 1549 if (klassptr != n->in(1)) { 1550 Node* decode = n->in(1); 1551 assert(decode->is_DecodeNarrowPtr(), "inconsistent subgraph"); 1552 Node* decode_clone = decode->clone(); 1553 decode_clone->set_req(1, klassptr_clone); 1554 register_new_node(decode_clone, ctrl); 1555 _igvn.set_type(decode_clone, decode_clone->Value(&_igvn)); 1556 klassptr_clone = decode_clone; 1557 } 1558 phi->set_req(i, klassptr_clone); 1559 } 1560 register_new_node(phi, region); 1561 Node* orig = n->in(1); 1562 _igvn.replace_input_of(n, 1, phi); 1563 split_if_with_blocks_post(n); 1564 if (n->outcnt() != 0) { 1565 _igvn.replace_input_of(n, 1, orig); 1566 _igvn.remove_dead_node(phi); 1567 } 1568 return true; 1569 } 1570 1571 //------------------------------split_if_with_blocks_post---------------------- 1572 // Do the real work in a non-recursive function. CFG hackery wants to be 1573 // in the post-order, so it can dirty the I-DOM info and not use the dirtied 1574 // info. 1575 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) { 1576 1577 if (flat_array_element_type_check(n)) { 1578 return; 1579 } 1580 1581 // Cloning Cmp through Phi's involves the split-if transform. 1582 // FastLock is not used by an If 1583 if (n->is_Cmp() && !n->is_FastLock()) { 1584 Node *n_ctrl = get_ctrl(n); 1585 // Determine if the Node has inputs from some local Phi. 1586 // Returns the block to clone thru. 1587 Node *n_blk = has_local_phi_input(n); 1588 if (n_blk != n_ctrl) { 1589 return; 1590 } 1591 1592 if (!can_split_if(n_ctrl)) { 1593 return; 1594 } 1595 1596 if (n->outcnt() != 1) { 1597 return; // Multiple bool's from 1 compare? 1598 } 1599 Node *bol = n->unique_out(); 1600 assert(bol->is_Bool(), "expect a bool here"); 1601 if (bol->outcnt() != 1) { 1602 return;// Multiple branches from 1 compare? 1603 } 1604 Node *iff = bol->unique_out(); 1605 1606 // Check some safety conditions 1607 if (iff->is_If()) { // Classic split-if? 1608 if (iff->in(0) != n_ctrl) { 1609 return; // Compare must be in same blk as if 1610 } 1611 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE 1612 // Can't split CMove with different control. 1613 if (get_ctrl(iff) != n_ctrl) { 1614 return; 1615 } 1616 if (get_ctrl(iff->in(2)) == n_ctrl || 1617 get_ctrl(iff->in(3)) == n_ctrl) { 1618 return; // Inputs not yet split-up 1619 } 1620 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) { 1621 return; // Loop-invar test gates loop-varying CMOVE 1622 } 1623 } else { 1624 return; // some other kind of node, such as an Allocate 1625 } 1626 1627 // When is split-if profitable? Every 'win' on means some control flow 1628 // goes dead, so it's almost always a win. 1629 int policy = 0; 1630 // Split compare 'n' through the merge point if it is profitable 1631 Node *phi = split_thru_phi( n, n_ctrl, policy); 1632 if (!phi) { 1633 return; 1634 } 1635 1636 // Now split the bool up thru the phi 1637 Node* bolphi = split_thru_phi(bol, n_ctrl, -1); 1638 guarantee(bolphi != nullptr, "null boolean phi node"); 1639 assert(iff->in(1) == bolphi, ""); 1640 1641 if (bolphi->Value(&_igvn)->singleton()) { 1642 return; 1643 } 1644 1645 // Conditional-move? Must split up now 1646 if (!iff->is_If()) { 1647 Node* cmovphi = split_thru_phi(iff, n_ctrl, -1); 1648 return; 1649 } 1650 1651 // Now split the IF 1652 C->print_method(PHASE_BEFORE_SPLIT_IF, 4, iff); 1653 if (TraceLoopOpts) { 1654 tty->print_cr("Split-If"); 1655 } 1656 do_split_if(iff); 1657 C->print_method(PHASE_AFTER_SPLIT_IF, 4, iff); 1658 return; 1659 } 1660 1661 // Two identical ifs back to back can be merged 1662 if (try_merge_identical_ifs(n)) { 1663 return; 1664 } 1665 1666 // Check for an IF ready to split; one that has its 1667 // condition codes input coming from a Phi at the block start. 1668 int n_op = n->Opcode(); 1669 1670 // Check for an IF being dominated by another IF same test 1671 if (n_op == Op_If || 1672 n_op == Op_RangeCheck) { 1673 Node *bol = n->in(1); 1674 uint max = bol->outcnt(); 1675 // Check for same test used more than once? 1676 if (bol->is_Bool() && (max > 1 || bol->in(1)->is_SubTypeCheck())) { 1677 // Search up IDOMs to see if this IF is dominated. 1678 Node* cmp = bol->in(1); 1679 Node *cutoff = cmp->is_SubTypeCheck() ? dom_lca(get_ctrl(cmp->in(1)), get_ctrl(cmp->in(2))) : get_ctrl(bol); 1680 1681 // Now search up IDOMs till cutoff, looking for a dominating test 1682 Node *prevdom = n; 1683 Node *dom = idom(prevdom); 1684 while (dom != cutoff) { 1685 if (dom->req() > 1 && n->as_If()->same_condition(dom, &_igvn) && prevdom->in(0) == dom && 1686 safe_for_if_replacement(dom)) { 1687 // It's invalid to move control dependent data nodes in the inner 1688 // strip-mined loop, because: 1689 // 1) break validation of LoopNode::verify_strip_mined() 1690 // 2) move code with side-effect in strip-mined loop 1691 // Move to the exit of outer strip-mined loop in that case. 1692 Node* out_le = is_inner_of_stripmined_loop(dom); 1693 if (out_le != nullptr) { 1694 prevdom = out_le; 1695 } 1696 // Replace the dominated test with an obvious true or false. 1697 // Place it on the IGVN worklist for later cleanup. 1698 C->set_major_progress(); 1699 // Split if: pin array accesses that are control dependent on a range check and moved to a regular if, 1700 // to prevent an array load from floating above its range check. There are three cases: 1701 // 1. Move from RangeCheck "a" to RangeCheck "b": don't need to pin. If we ever remove b, then we pin 1702 // all its array accesses at that point. 1703 // 2. We move from RangeCheck "a" to regular if "b": need to pin. If we ever remove b, then its array 1704 // accesses would start to float, since we don't pin at that point. 1705 // 3. If we move from regular if: don't pin. All array accesses are already assumed to be pinned. 1706 bool pin_array_access_nodes = n->Opcode() == Op_RangeCheck && 1707 prevdom->in(0)->Opcode() != Op_RangeCheck; 1708 dominated_by(prevdom->as_IfProj(), n->as_If(), false, pin_array_access_nodes); 1709 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } ); 1710 return; 1711 } 1712 prevdom = dom; 1713 dom = idom(prevdom); 1714 } 1715 } 1716 } 1717 1718 try_sink_out_of_loop(n); 1719 if (C->failing()) { 1720 return; 1721 } 1722 1723 try_move_store_after_loop(n); 1724 1725 // Remove multiple allocations of the same inline type 1726 if (n->is_InlineType()) { 1727 n->as_InlineType()->remove_redundant_allocations(this); 1728 } 1729 } 1730 1731 // Transform: 1732 // 1733 // if (some_condition) { 1734 // // body 1 1735 // } else { 1736 // // body 2 1737 // } 1738 // if (some_condition) { 1739 // // body 3 1740 // } else { 1741 // // body 4 1742 // } 1743 // 1744 // into: 1745 // 1746 // 1747 // if (some_condition) { 1748 // // body 1 1749 // // body 3 1750 // } else { 1751 // // body 2 1752 // // body 4 1753 // } 1754 bool PhaseIdealLoop::try_merge_identical_ifs(Node* n) { 1755 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) { 1756 Node *n_ctrl = n->in(0); 1757 IfNode* dom_if = idom(n_ctrl)->as_If(); 1758 if (n->in(1) != dom_if->in(1)) { 1759 assert(n->in(1)->in(1)->is_SubTypeCheck() && 1760 (n->in(1)->in(1)->as_SubTypeCheck()->method() != nullptr || 1761 dom_if->in(1)->in(1)->as_SubTypeCheck()->method() != nullptr), "only for subtype checks with profile data attached"); 1762 _igvn.replace_input_of(n, 1, dom_if->in(1)); 1763 } 1764 ProjNode* dom_proj_true = dom_if->proj_out(1); 1765 ProjNode* dom_proj_false = dom_if->proj_out(0); 1766 1767 // Now split the IF 1768 RegionNode* new_false_region; 1769 RegionNode* new_true_region; 1770 do_split_if(n, &new_false_region, &new_true_region); 1771 assert(new_false_region->req() == new_true_region->req(), ""); 1772 #ifdef ASSERT 1773 for (uint i = 1; i < new_false_region->req(); ++i) { 1774 assert(new_false_region->in(i)->in(0) == new_true_region->in(i)->in(0), "unexpected shape following split if"); 1775 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"); 1776 } 1777 #endif 1778 assert(new_false_region->in(1)->in(0)->in(1) == dom_if->in(1), "dominating if and dominated if after split must share test"); 1779 1780 // We now have: 1781 // if (some_condition) { 1782 // // body 1 1783 // if (some_condition) { 1784 // body3: // new_true_region 1785 // // body3 1786 // } else { 1787 // goto body4; 1788 // } 1789 // } else { 1790 // // body 2 1791 // if (some_condition) { 1792 // goto body3; 1793 // } else { 1794 // body4: // new_false_region 1795 // // body4; 1796 // } 1797 // } 1798 // 1799 1800 // clone pinned nodes thru the resulting regions 1801 push_pinned_nodes_thru_region(dom_if, new_true_region); 1802 push_pinned_nodes_thru_region(dom_if, new_false_region); 1803 1804 // Optimize out the cloned ifs. Because pinned nodes were cloned, this also allows a CastPP that would be dependent 1805 // 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 1806 // unrelated control dependency. 1807 for (uint i = 1; i < new_false_region->req(); i++) { 1808 if (is_dominator(dom_proj_true, new_false_region->in(i))) { 1809 dominated_by(dom_proj_true->as_IfProj(), new_false_region->in(i)->in(0)->as_If()); 1810 } else { 1811 assert(is_dominator(dom_proj_false, new_false_region->in(i)), "bad if"); 1812 dominated_by(dom_proj_false->as_IfProj(), new_false_region->in(i)->in(0)->as_If()); 1813 } 1814 } 1815 return true; 1816 } 1817 return false; 1818 } 1819 1820 void PhaseIdealLoop::push_pinned_nodes_thru_region(IfNode* dom_if, Node* region) { 1821 for (DUIterator i = region->outs(); region->has_out(i); i++) { 1822 Node* u = region->out(i); 1823 if (!has_ctrl(u) || u->is_Phi() || !u->depends_only_on_test() || !_igvn.no_dependent_zero_check(u)) { 1824 continue; 1825 } 1826 assert(u->in(0) == region, "not a control dependent node?"); 1827 uint j = 1; 1828 for (; j < u->req(); ++j) { 1829 Node* in = u->in(j); 1830 if (!is_dominator(ctrl_or_self(in), dom_if)) { 1831 break; 1832 } 1833 } 1834 if (j == u->req()) { 1835 Node *phi = PhiNode::make_blank(region, u); 1836 for (uint k = 1; k < region->req(); ++k) { 1837 Node* clone = u->clone(); 1838 clone->set_req(0, region->in(k)); 1839 register_new_node(clone, region->in(k)); 1840 phi->init_req(k, clone); 1841 } 1842 register_new_node(phi, region); 1843 _igvn.replace_node(u, phi); 1844 --i; 1845 } 1846 } 1847 } 1848 1849 bool PhaseIdealLoop::safe_for_if_replacement(const Node* dom) const { 1850 if (!dom->is_CountedLoopEnd()) { 1851 return true; 1852 } 1853 CountedLoopEndNode* le = dom->as_CountedLoopEnd(); 1854 CountedLoopNode* cl = le->loopnode(); 1855 if (cl == nullptr) { 1856 return true; 1857 } 1858 if (!cl->is_main_loop()) { 1859 return true; 1860 } 1861 if (cl->is_canonical_loop_entry() == nullptr) { 1862 return true; 1863 } 1864 // Further unrolling is possible so loop exit condition might change 1865 return false; 1866 } 1867 1868 // See if a shared loop-varying computation has no loop-varying uses. 1869 // Happens if something is only used for JVM state in uncommon trap exits, 1870 // like various versions of induction variable+offset. Clone the 1871 // computation per usage to allow it to sink out of the loop. 1872 void PhaseIdealLoop::try_sink_out_of_loop(Node* n) { 1873 bool is_raw_to_oop_cast = n->is_ConstraintCast() && 1874 n->in(1)->bottom_type()->isa_rawptr() && 1875 !n->bottom_type()->isa_rawptr(); 1876 1877 if (has_ctrl(n) && 1878 !n->is_Phi() && 1879 !n->is_Bool() && 1880 !n->is_Proj() && 1881 !n->is_MergeMem() && 1882 !n->is_CMove() && 1883 !n->is_OpaqueNotNull() && 1884 !n->is_OpaqueInitializedAssertionPredicate() && 1885 !n->is_OpaqueTemplateAssertionPredicate() && 1886 !is_raw_to_oop_cast && // don't extend live ranges of raw oops 1887 (KillPathsReachableByDeadTypeNode || !n->is_Type()) 1888 ) { 1889 Node *n_ctrl = get_ctrl(n); 1890 IdealLoopTree *n_loop = get_loop(n_ctrl); 1891 1892 if (n->in(0) != nullptr) { 1893 IdealLoopTree* loop_ctrl = get_loop(n->in(0)); 1894 if (n_loop != loop_ctrl && n_loop->is_member(loop_ctrl)) { 1895 // n has a control input inside a loop but get_ctrl() is member of an outer loop. This could happen, for example, 1896 // for Div nodes inside a loop (control input inside loop) without a use except for an UCT (outside the loop). 1897 // Rewire control of n to right outside of the loop, regardless if its input(s) are later sunk or not. 1898 Node* maybe_pinned_n = n; 1899 Node* outside_ctrl = place_outside_loop(n_ctrl, loop_ctrl); 1900 if (!would_sink_below_pre_loop_exit(loop_ctrl, outside_ctrl)) { 1901 if (n->depends_only_on_test()) { 1902 Node* pinned_clone = n->pin_array_access_node(); 1903 if (pinned_clone != nullptr) { 1904 // Pin array access nodes: if this is an array load, it's going to be dependent on a condition that's not a 1905 // range check for that access. If that condition is replaced by an identical dominating one, then an 1906 // unpinned load would risk floating above its range check. 1907 register_new_node(pinned_clone, n_ctrl); 1908 maybe_pinned_n = pinned_clone; 1909 _igvn.replace_node(n, pinned_clone); 1910 } 1911 } 1912 _igvn.replace_input_of(maybe_pinned_n, 0, outside_ctrl); 1913 } 1914 } 1915 } 1916 if (n_loop != _ltree_root && n->outcnt() > 1) { 1917 // Compute early control: needed for anti-dependence analysis. It's also possible that as a result of 1918 // previous transformations in this loop opts round, the node can be hoisted now: early control will tell us. 1919 Node* early_ctrl = compute_early_ctrl(n, n_ctrl); 1920 if (n_loop->is_member(get_loop(early_ctrl)) && // check that this one can't be hoisted now 1921 ctrl_of_all_uses_out_of_loop(n, early_ctrl, n_loop)) { // All uses in outer loops! 1922 if (n->is_Store() || n->is_LoadStore()) { 1923 assert(false, "no node with a side effect"); 1924 C->record_failure("no node with a side effect"); 1925 return; 1926 } 1927 Node* outer_loop_clone = nullptr; 1928 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin;) { 1929 Node* u = n->last_out(j); // Clone private computation per use 1930 _igvn.rehash_node_delayed(u); 1931 Node* x = nullptr; 1932 if (n->depends_only_on_test()) { 1933 // Pin array access nodes: if this is an array load, it's going to be dependent on a condition that's not a 1934 // range check for that access. If that condition is replaced by an identical dominating one, then an 1935 // unpinned load would risk floating above its range check. 1936 x = n->pin_array_access_node(); 1937 } 1938 if (x == nullptr) { 1939 x = n->clone(); 1940 } 1941 Node* x_ctrl = nullptr; 1942 if (u->is_Phi()) { 1943 // Replace all uses of normal nodes. Replace Phi uses 1944 // individually, so the separate Nodes can sink down 1945 // different paths. 1946 uint k = 1; 1947 while (u->in(k) != n) k++; 1948 u->set_req(k, x); 1949 // x goes next to Phi input path 1950 x_ctrl = u->in(0)->in(k); 1951 // Find control for 'x' next to use but not inside inner loops. 1952 x_ctrl = place_outside_loop(x_ctrl, n_loop); 1953 --j; 1954 } else { // Normal use 1955 if (has_ctrl(u)) { 1956 x_ctrl = get_ctrl(u); 1957 } else { 1958 x_ctrl = u->in(0); 1959 } 1960 // Find control for 'x' next to use but not inside inner loops. 1961 x_ctrl = place_outside_loop(x_ctrl, n_loop); 1962 // Replace all uses 1963 if (u->is_ConstraintCast() && _igvn.type(n)->higher_equal(u->bottom_type()) && u->in(0) == x_ctrl) { 1964 // If we're sinking a chain of data nodes, we might have inserted a cast to pin the use which is not necessary 1965 // anymore now that we're going to pin n as well 1966 _igvn.replace_node(u, x); 1967 --j; 1968 } else { 1969 int nb = u->replace_edge(n, x, &_igvn); 1970 j -= nb; 1971 } 1972 } 1973 1974 if (n->is_Load()) { 1975 // For loads, add a control edge to a CFG node outside of the loop 1976 // to force them to not combine and return back inside the loop 1977 // during GVN optimization (4641526). 1978 assert(x_ctrl == get_late_ctrl_with_anti_dep(x->as_Load(), early_ctrl, x_ctrl), "anti-dependences were already checked"); 1979 1980 IdealLoopTree* x_loop = get_loop(x_ctrl); 1981 Node* x_head = x_loop->_head; 1982 if (x_head->is_Loop() && x_head->is_OuterStripMinedLoop()) { 1983 // Do not add duplicate LoadNodes to the outer strip mined loop 1984 if (outer_loop_clone != nullptr) { 1985 _igvn.replace_node(x, outer_loop_clone); 1986 continue; 1987 } 1988 outer_loop_clone = x; 1989 } 1990 x->set_req(0, x_ctrl); 1991 } else if (n->in(0) != nullptr){ 1992 x->set_req(0, x_ctrl); 1993 } 1994 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone"); 1995 assert(!n_loop->is_member(get_loop(x_ctrl)), "should have moved out of loop"); 1996 register_new_node(x, x_ctrl); 1997 1998 // Chain of AddP nodes: (AddP base (AddP base (AddP base ))) 1999 // All AddP nodes must keep the same base after sinking so: 2000 // 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, 2001 // their bases remain the same. 2002 // (see 2- below) 2003 assert(!x->is_AddP() || !x->in(AddPNode::Address)->is_AddP() || 2004 x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base) || 2005 !x->in(AddPNode::Address)->in(AddPNode::Base)->eqv_uncast(x->in(AddPNode::Base)), "unexpected AddP shape"); 2006 if (x->in(0) == nullptr && !x->is_DecodeNarrowPtr() && 2007 !(x->is_AddP() && x->in(AddPNode::Address)->is_AddP() && x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base))) { 2008 assert(!x->is_Load(), "load should be pinned"); 2009 // Use a cast node to pin clone out of loop 2010 Node* cast = nullptr; 2011 for (uint k = 0; k < x->req(); k++) { 2012 Node* in = x->in(k); 2013 if (in != nullptr && n_loop->is_member(get_loop(get_ctrl(in)))) { 2014 const Type* in_t = _igvn.type(in); 2015 cast = ConstraintCastNode::make_cast_for_type(x_ctrl, in, in_t, 2016 ConstraintCastNode::UnconditionalDependency, nullptr); 2017 } 2018 if (cast != nullptr) { 2019 Node* prev = _igvn.hash_find_insert(cast); 2020 if (prev != nullptr && get_ctrl(prev) == x_ctrl) { 2021 cast->destruct(&_igvn); 2022 cast = prev; 2023 } else { 2024 register_new_node(cast, x_ctrl); 2025 } 2026 x->replace_edge(in, cast); 2027 // Chain of AddP nodes: 2028 // 2- A CastPP of the base is only added now that all AddP nodes are sunk 2029 if (x->is_AddP() && k == AddPNode::Base) { 2030 update_addp_chain_base(x, n->in(AddPNode::Base), cast); 2031 } 2032 break; 2033 } 2034 } 2035 assert(cast != nullptr, "must have added a cast to pin the node"); 2036 } 2037 } 2038 _igvn.remove_dead_node(n); 2039 } 2040 _dom_lca_tags_round = 0; 2041 } 2042 } 2043 } 2044 2045 void PhaseIdealLoop::update_addp_chain_base(Node* x, Node* old_base, Node* new_base) { 2046 ResourceMark rm; 2047 Node_List wq; 2048 wq.push(x); 2049 while (wq.size() != 0) { 2050 Node* n = wq.pop(); 2051 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2052 Node* u = n->fast_out(i); 2053 if (u->is_AddP() && u->in(AddPNode::Base) == old_base) { 2054 _igvn.replace_input_of(u, AddPNode::Base, new_base); 2055 wq.push(u); 2056 } 2057 } 2058 } 2059 } 2060 2061 // Compute the early control of a node by following its inputs until we reach 2062 // nodes that are pinned. Then compute the LCA of the control of all pinned nodes. 2063 Node* PhaseIdealLoop::compute_early_ctrl(Node* n, Node* n_ctrl) { 2064 Node* early_ctrl = nullptr; 2065 ResourceMark rm; 2066 Unique_Node_List wq; 2067 wq.push(n); 2068 for (uint i = 0; i < wq.size(); i++) { 2069 Node* m = wq.at(i); 2070 Node* c = nullptr; 2071 if (m->is_CFG()) { 2072 c = m; 2073 } else if (m->pinned()) { 2074 c = m->in(0); 2075 } else { 2076 for (uint j = 0; j < m->req(); j++) { 2077 Node* in = m->in(j); 2078 if (in != nullptr) { 2079 wq.push(in); 2080 } 2081 } 2082 } 2083 if (c != nullptr) { 2084 assert(is_dominator(c, n_ctrl), "control input must dominate current control"); 2085 if (early_ctrl == nullptr || is_dominator(early_ctrl, c)) { 2086 early_ctrl = c; 2087 } 2088 } 2089 } 2090 assert(is_dominator(early_ctrl, n_ctrl), "early control must dominate current control"); 2091 return early_ctrl; 2092 } 2093 2094 bool PhaseIdealLoop::ctrl_of_all_uses_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop) { 2095 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2096 Node* u = n->fast_out(i); 2097 if (u->is_Opaque1()) { 2098 return false; // Found loop limit, bugfix for 4677003 2099 } 2100 // We can't reuse tags in PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal() so make sure calls to 2101 // get_late_ctrl_with_anti_dep() use their own tag 2102 _dom_lca_tags_round++; 2103 assert(_dom_lca_tags_round != 0, "shouldn't wrap around"); 2104 2105 if (u->is_Phi()) { 2106 for (uint j = 1; j < u->req(); ++j) { 2107 if (u->in(j) == n && !ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, u->in(0)->in(j))) { 2108 return false; 2109 } 2110 } 2111 } else { 2112 Node* ctrl = has_ctrl(u) ? get_ctrl(u) : u->in(0); 2113 if (!ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, ctrl)) { 2114 return false; 2115 } 2116 } 2117 } 2118 return true; 2119 } 2120 2121 // 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 2122 // to a check that's eliminated by range check elimination, it becomes input to an expression that feeds into the exit 2123 // 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 2124 // it would be to not eliminate the check in the main loop. Instead, we prevent sinking of the node here so better code 2125 // is generated for the main loop. 2126 bool PhaseIdealLoop::would_sink_below_pre_loop_exit(IdealLoopTree* n_loop, Node* ctrl) { 2127 if (n_loop->_head->is_CountedLoop() && n_loop->_head->as_CountedLoop()->is_pre_loop()) { 2128 CountedLoopNode* pre_loop = n_loop->_head->as_CountedLoop(); 2129 if (is_dominator(pre_loop->loopexit(), ctrl)) { 2130 return true; 2131 } 2132 } 2133 return false; 2134 } 2135 2136 bool PhaseIdealLoop::ctrl_of_use_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop, Node* ctrl) { 2137 if (n->is_Load()) { 2138 ctrl = get_late_ctrl_with_anti_dep(n->as_Load(), n_ctrl, ctrl); 2139 } 2140 IdealLoopTree *u_loop = get_loop(ctrl); 2141 if (u_loop == n_loop) { 2142 return false; // Found loop-varying use 2143 } 2144 if (n_loop->is_member(u_loop)) { 2145 return false; // Found use in inner loop 2146 } 2147 if (would_sink_below_pre_loop_exit(n_loop, ctrl)) { 2148 return false; 2149 } 2150 return true; 2151 } 2152 2153 //------------------------------split_if_with_blocks--------------------------- 2154 // Check for aggressive application of 'split-if' optimization, 2155 // using basic block level info. 2156 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack) { 2157 Node* root = C->root(); 2158 visited.set(root->_idx); // first, mark root as visited 2159 // Do pre-visit work for root 2160 Node* n = split_if_with_blocks_pre(root); 2161 uint cnt = n->outcnt(); 2162 uint i = 0; 2163 2164 while (true) { 2165 // Visit all children 2166 if (i < cnt) { 2167 Node* use = n->raw_out(i); 2168 ++i; 2169 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) { 2170 // Now do pre-visit work for this use 2171 use = split_if_with_blocks_pre(use); 2172 nstack.push(n, i); // Save parent and next use's index. 2173 n = use; // Process all children of current use. 2174 cnt = use->outcnt(); 2175 i = 0; 2176 } 2177 } 2178 else { 2179 // All of n's children have been processed, complete post-processing. 2180 if (cnt != 0 && !n->is_Con()) { 2181 assert(has_node(n), "no dead nodes"); 2182 split_if_with_blocks_post(n); 2183 if (C->failing()) { 2184 return; 2185 } 2186 } 2187 if (must_throttle_split_if()) { 2188 nstack.clear(); 2189 } 2190 if (nstack.is_empty()) { 2191 // Finished all nodes on stack. 2192 break; 2193 } 2194 // Get saved parent node and next use's index. Visit the rest of uses. 2195 n = nstack.node(); 2196 cnt = n->outcnt(); 2197 i = nstack.index(); 2198 nstack.pop(); 2199 } 2200 } 2201 } 2202 2203 2204 //============================================================================= 2205 // 2206 // C L O N E A L O O P B O D Y 2207 // 2208 2209 //------------------------------clone_iff-------------------------------------- 2210 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 2211 // "Nearly" because all Nodes have been cloned from the original in the loop, 2212 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 2213 // through the Phi recursively, and return a Bool. 2214 Node* PhaseIdealLoop::clone_iff(PhiNode* phi) { 2215 2216 // Convert this Phi into a Phi merging Bools 2217 uint i; 2218 for (i = 1; i < phi->req(); i++) { 2219 Node* b = phi->in(i); 2220 if (b->is_Phi()) { 2221 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi())); 2222 } else { 2223 assert(b->is_Bool() || b->is_OpaqueNotNull() || b->is_OpaqueInitializedAssertionPredicate(), 2224 "bool, non-null check with OpaqueNotNull or Initialized Assertion Predicate with its Opaque node"); 2225 } 2226 } 2227 Node* n = phi->in(1); 2228 Node* sample_opaque = nullptr; 2229 Node *sample_bool = nullptr; 2230 if (n->is_OpaqueNotNull() || n->is_OpaqueInitializedAssertionPredicate()) { 2231 sample_opaque = n; 2232 sample_bool = n->in(1); 2233 assert(sample_bool->is_Bool(), "wrong type"); 2234 } else { 2235 sample_bool = n; 2236 } 2237 Node* sample_cmp = sample_bool->in(1); 2238 const Type* t = Type::TOP; 2239 const TypePtr* at = nullptr; 2240 if (sample_cmp->is_FlatArrayCheck()) { 2241 // Left input of a FlatArrayCheckNode is memory, set the (adr) type of the phi accordingly 2242 assert(sample_cmp->in(1)->bottom_type() == Type::MEMORY, "unexpected input type"); 2243 t = Type::MEMORY; 2244 at = TypeRawPtr::BOTTOM; 2245 } 2246 2247 // Make Phis to merge the Cmp's inputs. 2248 PhiNode *phi1 = new PhiNode(phi->in(0), t, at); 2249 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP); 2250 for (i = 1; i < phi->req(); i++) { 2251 Node *n1 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1); 2252 Node *n2 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2); 2253 phi1->set_req(i, n1); 2254 phi2->set_req(i, n2); 2255 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 2256 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 2257 } 2258 // See if these Phis have been made before. 2259 // Register with optimizer 2260 Node *hit1 = _igvn.hash_find_insert(phi1); 2261 if (hit1) { // Hit, toss just made Phi 2262 _igvn.remove_dead_node(phi1); // Remove new phi 2263 assert(hit1->is_Phi(), "" ); 2264 phi1 = (PhiNode*)hit1; // Use existing phi 2265 } else { // Miss 2266 _igvn.register_new_node_with_optimizer(phi1); 2267 } 2268 Node *hit2 = _igvn.hash_find_insert(phi2); 2269 if (hit2) { // Hit, toss just made Phi 2270 _igvn.remove_dead_node(phi2); // Remove new phi 2271 assert(hit2->is_Phi(), "" ); 2272 phi2 = (PhiNode*)hit2; // Use existing phi 2273 } else { // Miss 2274 _igvn.register_new_node_with_optimizer(phi2); 2275 } 2276 // Register Phis with loop/block info 2277 set_ctrl(phi1, phi->in(0)); 2278 set_ctrl(phi2, phi->in(0)); 2279 // Make a new Cmp 2280 Node *cmp = sample_cmp->clone(); 2281 cmp->set_req(1, phi1); 2282 cmp->set_req(2, phi2); 2283 _igvn.register_new_node_with_optimizer(cmp); 2284 set_ctrl(cmp, phi->in(0)); 2285 2286 // Make a new Bool 2287 Node *b = sample_bool->clone(); 2288 b->set_req(1,cmp); 2289 _igvn.register_new_node_with_optimizer(b); 2290 set_ctrl(b, phi->in(0)); 2291 2292 if (sample_opaque != nullptr) { 2293 Node* opaque = sample_opaque->clone(); 2294 opaque->set_req(1, b); 2295 _igvn.register_new_node_with_optimizer(opaque); 2296 set_ctrl(opaque, phi->in(0)); 2297 return opaque; 2298 } 2299 2300 assert(b->is_Bool(), ""); 2301 return b; 2302 } 2303 2304 //------------------------------clone_bool------------------------------------- 2305 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 2306 // "Nearly" because all Nodes have been cloned from the original in the loop, 2307 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 2308 // through the Phi recursively, and return a Bool. 2309 CmpNode*PhaseIdealLoop::clone_bool(PhiNode* phi) { 2310 uint i; 2311 // Convert this Phi into a Phi merging Bools 2312 for( i = 1; i < phi->req(); i++ ) { 2313 Node *b = phi->in(i); 2314 if( b->is_Phi() ) { 2315 _igvn.replace_input_of(phi, i, clone_bool(b->as_Phi())); 2316 } else { 2317 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" ); 2318 } 2319 } 2320 2321 Node *sample_cmp = phi->in(1); 2322 2323 // Make Phis to merge the Cmp's inputs. 2324 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP ); 2325 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP ); 2326 for( uint j = 1; j < phi->req(); j++ ) { 2327 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP 2328 Node *n1, *n2; 2329 if( cmp_top->is_Cmp() ) { 2330 n1 = cmp_top->in(1); 2331 n2 = cmp_top->in(2); 2332 } else { 2333 n1 = n2 = cmp_top; 2334 } 2335 phi1->set_req( j, n1 ); 2336 phi2->set_req( j, n2 ); 2337 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 2338 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 2339 } 2340 2341 // See if these Phis have been made before. 2342 // Register with optimizer 2343 Node *hit1 = _igvn.hash_find_insert(phi1); 2344 if( hit1 ) { // Hit, toss just made Phi 2345 _igvn.remove_dead_node(phi1); // Remove new phi 2346 assert( hit1->is_Phi(), "" ); 2347 phi1 = (PhiNode*)hit1; // Use existing phi 2348 } else { // Miss 2349 _igvn.register_new_node_with_optimizer(phi1); 2350 } 2351 Node *hit2 = _igvn.hash_find_insert(phi2); 2352 if( hit2 ) { // Hit, toss just made Phi 2353 _igvn.remove_dead_node(phi2); // Remove new phi 2354 assert( hit2->is_Phi(), "" ); 2355 phi2 = (PhiNode*)hit2; // Use existing phi 2356 } else { // Miss 2357 _igvn.register_new_node_with_optimizer(phi2); 2358 } 2359 // Register Phis with loop/block info 2360 set_ctrl(phi1, phi->in(0)); 2361 set_ctrl(phi2, phi->in(0)); 2362 // Make a new Cmp 2363 Node *cmp = sample_cmp->clone(); 2364 cmp->set_req( 1, phi1 ); 2365 cmp->set_req( 2, phi2 ); 2366 _igvn.register_new_node_with_optimizer(cmp); 2367 set_ctrl(cmp, phi->in(0)); 2368 2369 assert( cmp->is_Cmp(), "" ); 2370 return (CmpNode*)cmp; 2371 } 2372 2373 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new, 2374 IdealLoopTree* loop, IdealLoopTree* outer_loop, 2375 Node_List*& split_if_set, Node_List*& split_bool_set, 2376 Node_List*& split_cex_set, Node_List& worklist, 2377 uint new_counter, CloneLoopMode mode) { 2378 Node* nnn = old_new[old->_idx]; 2379 // Copy uses to a worklist, so I can munge the def-use info 2380 // with impunity. 2381 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 2382 worklist.push(old->fast_out(j)); 2383 2384 while( worklist.size() ) { 2385 Node *use = worklist.pop(); 2386 if (!has_node(use)) continue; // Ignore dead nodes 2387 if (use->in(0) == C->top()) continue; 2388 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 2389 // Check for data-use outside of loop - at least one of OLD or USE 2390 // must not be a CFG node. 2391 #ifdef ASSERT 2392 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == nullptr) { 2393 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint(); 2394 assert(mode != IgnoreStripMined, "incorrect cloning mode"); 2395 assert((mode == ControlAroundStripMined && use == sfpt) || !use->is_reachable_from_root(), "missed a node"); 2396 } 2397 #endif 2398 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) { 2399 2400 // If the Data use is an IF, that means we have an IF outside the 2401 // loop that is switching on a condition that is set inside the 2402 // loop. Happens if people set a loop-exit flag; then test the flag 2403 // in the loop to break the loop, then test is again outside the 2404 // loop to determine which way the loop exited. 2405 // 2406 // For several uses we need to make sure that there is no phi between, 2407 // the use and the Bool/Cmp. We therefore clone the Bool/Cmp down here 2408 // to avoid such a phi in between. 2409 // For example, it is unexpected that there is a Phi between an 2410 // AllocateArray node and its ValidLengthTest input that could cause 2411 // split if to break. 2412 assert(!use->is_OpaqueTemplateAssertionPredicate(), 2413 "should not clone a Template Assertion Predicate which should be removed once it's useless"); 2414 if (use->is_If() || use->is_CMove() || use->is_OpaqueNotNull() || use->is_OpaqueInitializedAssertionPredicate() || 2415 (use->Opcode() == Op_AllocateArray && use->in(AllocateNode::ValidLengthTest) == old)) { 2416 // Since this code is highly unlikely, we lazily build the worklist 2417 // of such Nodes to go split. 2418 if (!split_if_set) { 2419 split_if_set = new Node_List(); 2420 } 2421 split_if_set->push(use); 2422 } 2423 if (use->is_Bool()) { 2424 if (!split_bool_set) { 2425 split_bool_set = new Node_List(); 2426 } 2427 split_bool_set->push(use); 2428 } 2429 if (use->Opcode() == Op_CreateEx) { 2430 if (!split_cex_set) { 2431 split_cex_set = new Node_List(); 2432 } 2433 split_cex_set->push(use); 2434 } 2435 2436 2437 // Get "block" use is in 2438 uint idx = 0; 2439 while( use->in(idx) != old ) idx++; 2440 Node *prev = use->is_CFG() ? use : get_ctrl(use); 2441 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" ); 2442 Node* cfg = (prev->_idx >= new_counter && prev->is_Region()) 2443 ? prev->in(2) 2444 : idom(prev); 2445 if( use->is_Phi() ) // Phi use is in prior block 2446 cfg = prev->in(idx); // NOT in block of Phi itself 2447 if (cfg->is_top()) { // Use is dead? 2448 _igvn.replace_input_of(use, idx, C->top()); 2449 continue; 2450 } 2451 2452 // If use is referenced through control edge... (idx == 0) 2453 if (mode == IgnoreStripMined && idx == 0) { 2454 LoopNode *head = loop->_head->as_Loop(); 2455 if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) { 2456 // That node is outside the inner loop, leave it outside the 2457 // outer loop as well to not confuse verification code. 2458 assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop"); 2459 _igvn.replace_input_of(use, 0, head->outer_loop_exit()); 2460 continue; 2461 } 2462 } 2463 2464 while(!outer_loop->is_member(get_loop(cfg))) { 2465 prev = cfg; 2466 cfg = (cfg->_idx >= new_counter && cfg->is_Region()) ? cfg->in(2) : idom(cfg); 2467 } 2468 // If the use occurs after merging several exits from the loop, then 2469 // old value must have dominated all those exits. Since the same old 2470 // value was used on all those exits we did not need a Phi at this 2471 // merge point. NOW we do need a Phi here. Each loop exit value 2472 // is now merged with the peeled body exit; each exit gets its own 2473 // private Phi and those Phis need to be merged here. 2474 Node *phi; 2475 if( prev->is_Region() ) { 2476 if( idx == 0 ) { // Updating control edge? 2477 phi = prev; // Just use existing control 2478 } else { // Else need a new Phi 2479 phi = PhiNode::make( prev, old ); 2480 // Now recursively fix up the new uses of old! 2481 for( uint i = 1; i < prev->req(); i++ ) { 2482 worklist.push(phi); // Onto worklist once for each 'old' input 2483 } 2484 } 2485 } else { 2486 // Get new RegionNode merging old and new loop exits 2487 prev = old_new[prev->_idx]; 2488 assert( prev, "just made this in step 7" ); 2489 if( idx == 0) { // Updating control edge? 2490 phi = prev; // Just use existing control 2491 } else { // Else need a new Phi 2492 // Make a new Phi merging data values properly 2493 phi = PhiNode::make( prev, old ); 2494 phi->set_req( 1, nnn ); 2495 } 2496 } 2497 // If inserting a new Phi, check for prior hits 2498 if( idx != 0 ) { 2499 Node *hit = _igvn.hash_find_insert(phi); 2500 if( hit == nullptr ) { 2501 _igvn.register_new_node_with_optimizer(phi); // Register new phi 2502 } else { // or 2503 // Remove the new phi from the graph and use the hit 2504 _igvn.remove_dead_node(phi); 2505 phi = hit; // Use existing phi 2506 } 2507 set_ctrl(phi, prev); 2508 } 2509 // Make 'use' use the Phi instead of the old loop body exit value 2510 assert(use->in(idx) == old, "old is still input of use"); 2511 // We notify all uses of old, including use, and the indirect uses, 2512 // that may now be optimized because we have replaced old with phi. 2513 _igvn.add_users_to_worklist(old); 2514 if (idx == 0 && 2515 use->depends_only_on_test()) { 2516 Node* pinned_clone = use->pin_array_access_node(); 2517 if (pinned_clone != nullptr) { 2518 // Pin array access nodes: control is updated here to a region. If, after some transformations, only one path 2519 // into the region is left, an array load could become dependent on a condition that's not a range check for 2520 // that access. If that condition is replaced by an identical dominating one, then an unpinned load would risk 2521 // floating above its range check. 2522 pinned_clone->set_req(0, phi); 2523 register_new_node_with_ctrl_of(pinned_clone, use); 2524 _igvn.replace_node(use, pinned_clone); 2525 continue; 2526 } 2527 } 2528 _igvn.replace_input_of(use, idx, phi); 2529 if( use->_idx >= new_counter ) { // If updating new phis 2530 // Not needed for correctness, but prevents a weak assert 2531 // in AddPNode from tripping (when we end up with different 2532 // base & derived Phis that will become the same after 2533 // IGVN does CSE). 2534 Node *hit = _igvn.hash_find_insert(use); 2535 if( hit ) // Go ahead and re-hash for hits. 2536 _igvn.replace_node( use, hit ); 2537 } 2538 } 2539 } 2540 } 2541 2542 static void collect_nodes_in_outer_loop_not_reachable_from_sfpt(Node* n, const IdealLoopTree *loop, const IdealLoopTree* outer_loop, 2543 const Node_List &old_new, Unique_Node_List& wq, PhaseIdealLoop* phase, 2544 bool check_old_new) { 2545 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2546 Node* u = n->fast_out(j); 2547 assert(check_old_new || old_new[u->_idx] == nullptr, "shouldn't have been cloned"); 2548 if (!u->is_CFG() && (!check_old_new || old_new[u->_idx] == nullptr)) { 2549 Node* c = phase->get_ctrl(u); 2550 IdealLoopTree* u_loop = phase->get_loop(c); 2551 assert(!loop->is_member(u_loop) || !loop->_body.contains(u), "can be in outer loop or out of both loops only"); 2552 if (!loop->is_member(u_loop)) { 2553 if (outer_loop->is_member(u_loop)) { 2554 wq.push(u); 2555 } else { 2556 // nodes pinned with control in the outer loop but not referenced from the safepoint must be moved out of 2557 // the outer loop too 2558 Node* u_c = u->in(0); 2559 if (u_c != nullptr) { 2560 IdealLoopTree* u_c_loop = phase->get_loop(u_c); 2561 if (outer_loop->is_member(u_c_loop) && !loop->is_member(u_c_loop)) { 2562 wq.push(u); 2563 } 2564 } 2565 } 2566 } 2567 } 2568 } 2569 } 2570 2571 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop, 2572 IdealLoopTree* outer_loop, int dd, Node_List &old_new, 2573 Node_List& extra_data_nodes) { 2574 if (head->is_strip_mined() && mode != IgnoreStripMined) { 2575 CountedLoopNode* cl = head->as_CountedLoop(); 2576 Node* l = cl->outer_loop(); 2577 Node* tail = cl->outer_loop_tail(); 2578 IfNode* le = cl->outer_loop_end(); 2579 Node* sfpt = cl->outer_safepoint(); 2580 CountedLoopEndNode* cle = cl->loopexit(); 2581 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop(); 2582 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null(); 2583 Node* cle_out = cle->proj_out(false); 2584 2585 Node* new_sfpt = nullptr; 2586 Node* new_cle_out = cle_out->clone(); 2587 old_new.map(cle_out->_idx, new_cle_out); 2588 if (mode == CloneIncludesStripMined) { 2589 // clone outer loop body 2590 Node* new_l = l->clone(); 2591 Node* new_tail = tail->clone(); 2592 IfNode* new_le = le->clone()->as_If(); 2593 new_sfpt = sfpt->clone(); 2594 2595 set_loop(new_l, outer_loop->_parent); 2596 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd); 2597 set_loop(new_cle_out, outer_loop->_parent); 2598 set_idom(new_cle_out, new_cle, dd); 2599 set_loop(new_sfpt, outer_loop->_parent); 2600 set_idom(new_sfpt, new_cle_out, dd); 2601 set_loop(new_le, outer_loop->_parent); 2602 set_idom(new_le, new_sfpt, dd); 2603 set_loop(new_tail, outer_loop->_parent); 2604 set_idom(new_tail, new_le, dd); 2605 set_idom(new_cl, new_l, dd); 2606 2607 old_new.map(l->_idx, new_l); 2608 old_new.map(tail->_idx, new_tail); 2609 old_new.map(le->_idx, new_le); 2610 old_new.map(sfpt->_idx, new_sfpt); 2611 2612 new_l->set_req(LoopNode::LoopBackControl, new_tail); 2613 new_l->set_req(0, new_l); 2614 new_tail->set_req(0, new_le); 2615 new_le->set_req(0, new_sfpt); 2616 new_sfpt->set_req(0, new_cle_out); 2617 new_cle_out->set_req(0, new_cle); 2618 new_cl->set_req(LoopNode::EntryControl, new_l); 2619 2620 _igvn.register_new_node_with_optimizer(new_l); 2621 _igvn.register_new_node_with_optimizer(new_tail); 2622 _igvn.register_new_node_with_optimizer(new_le); 2623 } else { 2624 Node *newhead = old_new[loop->_head->_idx]; 2625 newhead->as_Loop()->clear_strip_mined(); 2626 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl)); 2627 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 2628 } 2629 // Look at data node that were assigned a control in the outer 2630 // loop: they are kept in the outer loop by the safepoint so start 2631 // from the safepoint node's inputs. 2632 IdealLoopTree* outer_loop = get_loop(l); 2633 Node_Stack stack(2); 2634 stack.push(sfpt, 1); 2635 uint new_counter = C->unique(); 2636 while (stack.size() > 0) { 2637 Node* n = stack.node(); 2638 uint i = stack.index(); 2639 while (i < n->req() && 2640 (n->in(i) == nullptr || 2641 !has_ctrl(n->in(i)) || 2642 get_loop(get_ctrl(n->in(i))) != outer_loop || 2643 (old_new[n->in(i)->_idx] != nullptr && old_new[n->in(i)->_idx]->_idx >= new_counter))) { 2644 i++; 2645 } 2646 if (i < n->req()) { 2647 stack.set_index(i+1); 2648 stack.push(n->in(i), 0); 2649 } else { 2650 assert(old_new[n->_idx] == nullptr || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet"); 2651 Node* m = n == sfpt ? new_sfpt : n->clone(); 2652 if (m != nullptr) { 2653 for (uint i = 0; i < n->req(); i++) { 2654 if (m->in(i) != nullptr && old_new[m->in(i)->_idx] != nullptr) { 2655 m->set_req(i, old_new[m->in(i)->_idx]); 2656 } 2657 } 2658 } else { 2659 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?"); 2660 } 2661 if (n != sfpt) { 2662 extra_data_nodes.push(n); 2663 _igvn.register_new_node_with_optimizer(m); 2664 assert(get_ctrl(n) == cle_out, "what other control?"); 2665 set_ctrl(m, new_cle_out); 2666 old_new.map(n->_idx, m); 2667 } 2668 stack.pop(); 2669 } 2670 } 2671 if (mode == CloneIncludesStripMined) { 2672 _igvn.register_new_node_with_optimizer(new_sfpt); 2673 _igvn.register_new_node_with_optimizer(new_cle_out); 2674 } 2675 // Some other transformation may have pessimistically assigned some 2676 // data nodes to the outer loop. Set their control so they are out 2677 // of the outer loop. 2678 ResourceMark rm; 2679 Unique_Node_List wq; 2680 for (uint i = 0; i < extra_data_nodes.size(); i++) { 2681 Node* old = extra_data_nodes.at(i); 2682 collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true); 2683 } 2684 2685 for (uint i = 0; i < loop->_body.size(); i++) { 2686 Node* old = loop->_body.at(i); 2687 collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true); 2688 } 2689 2690 Node* inner_out = sfpt->in(0); 2691 if (inner_out->outcnt() > 1) { 2692 collect_nodes_in_outer_loop_not_reachable_from_sfpt(inner_out, loop, outer_loop, old_new, wq, this, true); 2693 } 2694 2695 Node* new_ctrl = cl->outer_loop_exit(); 2696 assert(get_loop(new_ctrl) != outer_loop, "must be out of the loop nest"); 2697 for (uint i = 0; i < wq.size(); i++) { 2698 Node* n = wq.at(i); 2699 set_ctrl(n, new_ctrl); 2700 if (n->in(0) != nullptr) { 2701 _igvn.replace_input_of(n, 0, new_ctrl); 2702 } 2703 collect_nodes_in_outer_loop_not_reachable_from_sfpt(n, loop, outer_loop, old_new, wq, this, false); 2704 } 2705 } else { 2706 Node *newhead = old_new[loop->_head->_idx]; 2707 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 2708 } 2709 } 2710 2711 //------------------------------clone_loop------------------------------------- 2712 // 2713 // C L O N E A L O O P B O D Y 2714 // 2715 // This is the basic building block of the loop optimizations. It clones an 2716 // entire loop body. It makes an old_new loop body mapping; with this mapping 2717 // you can find the new-loop equivalent to an old-loop node. All new-loop 2718 // nodes are exactly equal to their old-loop counterparts, all edges are the 2719 // same. All exits from the old-loop now have a RegionNode that merges the 2720 // equivalent new-loop path. This is true even for the normal "loop-exit" 2721 // condition. All uses of loop-invariant old-loop values now come from (one 2722 // or more) Phis that merge their new-loop equivalents. 2723 // 2724 // This operation leaves the graph in an illegal state: there are two valid 2725 // control edges coming from the loop pre-header to both loop bodies. I'll 2726 // definitely have to hack the graph after running this transform. 2727 // 2728 // From this building block I will further edit edges to perform loop peeling 2729 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc. 2730 // 2731 // Parameter side_by_size_idom: 2732 // When side_by_size_idom is null, the dominator tree is constructed for 2733 // the clone loop to dominate the original. Used in construction of 2734 // pre-main-post loop sequence. 2735 // When nonnull, the clone and original are side-by-side, both are 2736 // dominated by the side_by_side_idom node. Used in construction of 2737 // unswitched loops. 2738 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd, 2739 CloneLoopMode mode, Node* side_by_side_idom) { 2740 2741 LoopNode* head = loop->_head->as_Loop(); 2742 head->verify_strip_mined(1); 2743 2744 if (C->do_vector_loop() && PrintOpto) { 2745 const char* mname = C->method()->name()->as_quoted_ascii(); 2746 if (mname != nullptr) { 2747 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname); 2748 } 2749 } 2750 2751 CloneMap& cm = C->clone_map(); 2752 if (C->do_vector_loop()) { 2753 cm.set_clone_idx(cm.max_gen()+1); 2754 #ifndef PRODUCT 2755 if (PrintOpto) { 2756 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx()); 2757 loop->dump_head(); 2758 } 2759 #endif 2760 } 2761 2762 // Step 1: Clone the loop body. Make the old->new mapping. 2763 clone_loop_body(loop->_body, old_new, &cm); 2764 2765 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop; 2766 2767 // Step 2: Fix the edges in the new body. If the old input is outside the 2768 // loop use it. If the old input is INside the loop, use the corresponding 2769 // new node instead. 2770 fix_body_edges(loop->_body, loop, old_new, dd, outer_loop->_parent, false); 2771 2772 Node_List extra_data_nodes; // data nodes in the outer strip mined loop 2773 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes); 2774 2775 // Step 3: Now fix control uses. Loop varying control uses have already 2776 // been fixed up (as part of all input edges in Step 2). Loop invariant 2777 // control uses must be either an IfFalse or an IfTrue. Make a merge 2778 // point to merge the old and new IfFalse/IfTrue nodes; make the use 2779 // refer to this. 2780 Node_List worklist; 2781 uint new_counter = C->unique(); 2782 fix_ctrl_uses(loop->_body, loop, old_new, mode, side_by_side_idom, &cm, worklist); 2783 2784 // Step 4: If loop-invariant use is not control, it must be dominated by a 2785 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region 2786 // there if needed. Make a Phi there merging old and new used values. 2787 Node_List *split_if_set = nullptr; 2788 Node_List *split_bool_set = nullptr; 2789 Node_List *split_cex_set = nullptr; 2790 fix_data_uses(loop->_body, loop, mode, outer_loop, new_counter, old_new, worklist, split_if_set, split_bool_set, split_cex_set); 2791 2792 for (uint i = 0; i < extra_data_nodes.size(); i++) { 2793 Node* old = extra_data_nodes.at(i); 2794 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set, 2795 split_bool_set, split_cex_set, worklist, new_counter, 2796 mode); 2797 } 2798 2799 // Check for IFs that need splitting/cloning. Happens if an IF outside of 2800 // the loop uses a condition set in the loop. The original IF probably 2801 // takes control from one or more OLD Regions (which in turn get from NEW 2802 // Regions). In any case, there will be a set of Phis for each merge point 2803 // from the IF up to where the original BOOL def exists the loop. 2804 finish_clone_loop(split_if_set, split_bool_set, split_cex_set); 2805 2806 } 2807 2808 void PhaseIdealLoop::finish_clone_loop(Node_List* split_if_set, Node_List* split_bool_set, Node_List* split_cex_set) { 2809 if (split_if_set) { 2810 while (split_if_set->size()) { 2811 Node *iff = split_if_set->pop(); 2812 uint input = iff->Opcode() == Op_AllocateArray ? AllocateNode::ValidLengthTest : 1; 2813 if (iff->in(input)->is_Phi()) { 2814 Node *b = clone_iff(iff->in(input)->as_Phi()); 2815 _igvn.replace_input_of(iff, input, b); 2816 } 2817 } 2818 } 2819 if (split_bool_set) { 2820 while (split_bool_set->size()) { 2821 Node *b = split_bool_set->pop(); 2822 Node *phi = b->in(1); 2823 assert(phi->is_Phi(), ""); 2824 CmpNode *cmp = clone_bool((PhiNode*) phi); 2825 _igvn.replace_input_of(b, 1, cmp); 2826 } 2827 } 2828 if (split_cex_set) { 2829 while (split_cex_set->size()) { 2830 Node *b = split_cex_set->pop(); 2831 assert(b->in(0)->is_Region(), ""); 2832 assert(b->in(1)->is_Phi(), ""); 2833 assert(b->in(0)->in(0) == b->in(1)->in(0), ""); 2834 split_up(b, b->in(0), nullptr); 2835 } 2836 } 2837 } 2838 2839 void PhaseIdealLoop::fix_data_uses(Node_List& body, IdealLoopTree* loop, CloneLoopMode mode, IdealLoopTree* outer_loop, 2840 uint new_counter, Node_List &old_new, Node_List &worklist, Node_List*& split_if_set, 2841 Node_List*& split_bool_set, Node_List*& split_cex_set) { 2842 for(uint i = 0; i < body.size(); i++ ) { 2843 Node* old = body.at(i); 2844 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set, 2845 split_bool_set, split_cex_set, worklist, new_counter, 2846 mode); 2847 } 2848 } 2849 2850 void PhaseIdealLoop::fix_ctrl_uses(const Node_List& body, const IdealLoopTree* loop, Node_List &old_new, CloneLoopMode mode, 2851 Node* side_by_side_idom, CloneMap* cm, Node_List &worklist) { 2852 LoopNode* head = loop->_head->as_Loop(); 2853 for(uint i = 0; i < body.size(); i++ ) { 2854 Node* old = body.at(i); 2855 if( !old->is_CFG() ) continue; 2856 2857 // Copy uses to a worklist, so I can munge the def-use info 2858 // with impunity. 2859 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) { 2860 worklist.push(old->fast_out(j)); 2861 } 2862 2863 while (worklist.size()) { // Visit all uses 2864 Node *use = worklist.pop(); 2865 if (!has_node(use)) continue; // Ignore dead nodes 2866 IdealLoopTree *use_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use ); 2867 if (!loop->is_member(use_loop) && use->is_CFG()) { 2868 // Both OLD and USE are CFG nodes here. 2869 assert(use->is_Proj(), "" ); 2870 Node* nnn = old_new[old->_idx]; 2871 2872 Node* newuse = nullptr; 2873 if (head->is_strip_mined() && mode != IgnoreStripMined) { 2874 CountedLoopNode* cl = head->as_CountedLoop(); 2875 CountedLoopEndNode* cle = cl->loopexit(); 2876 // is use the projection that exits the loop from the CountedLoopEndNode? 2877 if (use->in(0) == cle) { 2878 IfFalseNode* cle_out = use->as_IfFalse(); 2879 IfNode* le = cl->outer_loop_end(); 2880 use = le->proj_out(false); 2881 use_loop = get_loop(use); 2882 if (mode == CloneIncludesStripMined) { 2883 nnn = old_new[le->_idx]; 2884 } else { 2885 newuse = old_new[cle_out->_idx]; 2886 } 2887 } 2888 } 2889 if (newuse == nullptr) { 2890 newuse = use->clone(); 2891 } 2892 2893 // Clone the loop exit control projection 2894 if (C->do_vector_loop() && cm != nullptr) { 2895 cm->verify_insert_and_clone(use, newuse, cm->clone_idx()); 2896 } 2897 newuse->set_req(0,nnn); 2898 _igvn.register_new_node_with_optimizer(newuse); 2899 set_loop(newuse, use_loop); 2900 set_idom(newuse, nnn, dom_depth(nnn) + 1 ); 2901 2902 // We need a Region to merge the exit from the peeled body and the 2903 // exit from the old loop body. 2904 RegionNode *r = new RegionNode(3); 2905 uint dd_r = MIN2(dom_depth(newuse), dom_depth(use)); 2906 assert(dd_r >= dom_depth(dom_lca(newuse, use)), "" ); 2907 2908 // The original user of 'use' uses 'r' instead. 2909 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) { 2910 Node* useuse = use->last_out(l); 2911 _igvn.rehash_node_delayed(useuse); 2912 uint uses_found = 0; 2913 if (useuse->in(0) == use) { 2914 useuse->set_req(0, r); 2915 uses_found++; 2916 if (useuse->is_CFG()) { 2917 // This is not a dom_depth > dd_r because when new 2918 // control flow is constructed by a loop opt, a node and 2919 // its dominator can end up at the same dom_depth 2920 assert(dom_depth(useuse) >= dd_r, ""); 2921 set_idom(useuse, r, dom_depth(useuse)); 2922 } 2923 } 2924 for (uint k = 1; k < useuse->req(); k++) { 2925 if( useuse->in(k) == use ) { 2926 useuse->set_req(k, r); 2927 uses_found++; 2928 if (useuse->is_Loop() && k == LoopNode::EntryControl) { 2929 // This is not a dom_depth > dd_r because when new 2930 // control flow is constructed by a loop opt, a node 2931 // and its dominator can end up at the same dom_depth 2932 assert(dom_depth(useuse) >= dd_r , ""); 2933 set_idom(useuse, r, dom_depth(useuse)); 2934 } 2935 } 2936 } 2937 l -= uses_found; // we deleted 1 or more copies of this edge 2938 } 2939 2940 assert(use->is_Proj(), "loop exit should be projection"); 2941 // replace_node_and_forward_ctrl() below moves all nodes that are: 2942 // - control dependent on the loop exit or 2943 // - have control set to the loop exit 2944 // below the post-loop merge point. 2945 // replace_node_and_forward_ctrl() takes a dead control as first input. 2946 // To make it possible to use it, the loop exit projection is cloned and becomes the 2947 // new exit projection. The initial one becomes dead and is "replaced" by the region. 2948 Node* use_clone = use->clone(); 2949 register_control(use_clone, use_loop, idom(use), dom_depth(use)); 2950 // Now finish up 'r' 2951 r->set_req(1, newuse); 2952 r->set_req(2, use_clone); 2953 _igvn.register_new_node_with_optimizer(r); 2954 set_loop(r, use_loop); 2955 set_idom(r, (side_by_side_idom == nullptr) ? newuse->in(0) : side_by_side_idom, dd_r); 2956 replace_node_and_forward_ctrl(use, r); 2957 // Map the (cloned) old use to the new merge point 2958 old_new.map(use_clone->_idx, r); 2959 } // End of if a loop-exit test 2960 } 2961 } 2962 } 2963 2964 void PhaseIdealLoop::fix_body_edges(const Node_List &body, IdealLoopTree* loop, const Node_List &old_new, int dd, 2965 IdealLoopTree* parent, bool partial) { 2966 for(uint i = 0; i < body.size(); i++ ) { 2967 Node *old = body.at(i); 2968 Node *nnn = old_new[old->_idx]; 2969 // Fix CFG/Loop controlling the new node 2970 if (has_ctrl(old)) { 2971 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]); 2972 } else { 2973 set_loop(nnn, parent); 2974 if (old->outcnt() > 0) { 2975 Node* dom = idom(old); 2976 if (old_new[dom->_idx] != nullptr) { 2977 dom = old_new[dom->_idx]; 2978 set_idom(nnn, dom, dd ); 2979 } 2980 } 2981 } 2982 // Correct edges to the new node 2983 for (uint j = 0; j < nnn->req(); j++) { 2984 Node *n = nnn->in(j); 2985 if (n != nullptr) { 2986 IdealLoopTree *old_in_loop = get_loop(has_ctrl(n) ? get_ctrl(n) : n); 2987 if (loop->is_member(old_in_loop)) { 2988 if (old_new[n->_idx] != nullptr) { 2989 nnn->set_req(j, old_new[n->_idx]); 2990 } else { 2991 assert(!body.contains(n), ""); 2992 assert(partial, "node not cloned"); 2993 } 2994 } 2995 } 2996 } 2997 _igvn.hash_find_insert(nnn); 2998 } 2999 } 3000 3001 void PhaseIdealLoop::clone_loop_body(const Node_List& body, Node_List &old_new, CloneMap* cm) { 3002 for (uint i = 0; i < body.size(); i++) { 3003 Node* old = body.at(i); 3004 Node* nnn = old->clone(); 3005 old_new.map(old->_idx, nnn); 3006 if (C->do_vector_loop() && cm != nullptr) { 3007 cm->verify_insert_and_clone(old, nnn, cm->clone_idx()); 3008 } 3009 _igvn.register_new_node_with_optimizer(nnn); 3010 } 3011 } 3012 3013 3014 //---------------------- stride_of_possible_iv ------------------------------------- 3015 // Looks for an iff/bool/comp with one operand of the compare 3016 // being a cycle involving an add and a phi, 3017 // with an optional truncation (left-shift followed by a right-shift) 3018 // of the add. Returns zero if not an iv. 3019 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) { 3020 Node* trunc1 = nullptr; 3021 Node* trunc2 = nullptr; 3022 const TypeInteger* ttype = nullptr; 3023 if (!iff->is_If() || iff->in(1) == nullptr || !iff->in(1)->is_Bool()) { 3024 return 0; 3025 } 3026 BoolNode* bl = iff->in(1)->as_Bool(); 3027 Node* cmp = bl->in(1); 3028 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) { 3029 return 0; 3030 } 3031 // Must have an invariant operand 3032 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) { 3033 return 0; 3034 } 3035 Node* add2 = nullptr; 3036 Node* cmp1 = cmp->in(1); 3037 if (cmp1->is_Phi()) { 3038 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) ))) 3039 Node* phi = cmp1; 3040 for (uint i = 1; i < phi->req(); i++) { 3041 Node* in = phi->in(i); 3042 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in, 3043 &trunc1, &trunc2, &ttype, T_INT); 3044 if (add && add->in(1) == phi) { 3045 add2 = add->in(2); 3046 break; 3047 } 3048 } 3049 } else { 3050 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) ))) 3051 Node* addtrunc = cmp1; 3052 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc, 3053 &trunc1, &trunc2, &ttype, T_INT); 3054 if (add && add->in(1)->is_Phi()) { 3055 Node* phi = add->in(1); 3056 for (uint i = 1; i < phi->req(); i++) { 3057 if (phi->in(i) == addtrunc) { 3058 add2 = add->in(2); 3059 break; 3060 } 3061 } 3062 } 3063 } 3064 if (add2 != nullptr) { 3065 const TypeInt* add2t = _igvn.type(add2)->is_int(); 3066 if (add2t->is_con()) { 3067 return add2t->get_con(); 3068 } 3069 } 3070 return 0; 3071 } 3072 3073 3074 //---------------------- stay_in_loop ------------------------------------- 3075 // Return the (unique) control output node that's in the loop (if it exists.) 3076 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) { 3077 Node* unique = nullptr; 3078 if (!n) return nullptr; 3079 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 3080 Node* use = n->fast_out(i); 3081 if (!has_ctrl(use) && loop->is_member(get_loop(use))) { 3082 if (unique != nullptr) { 3083 return nullptr; 3084 } 3085 unique = use; 3086 } 3087 } 3088 return unique; 3089 } 3090 3091 //------------------------------ register_node ------------------------------------- 3092 // Utility to register node "n" with PhaseIdealLoop 3093 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree* loop, Node* pred, uint ddepth) { 3094 _igvn.register_new_node_with_optimizer(n); 3095 loop->_body.push(n); 3096 if (n->is_CFG()) { 3097 set_loop(n, loop); 3098 set_idom(n, pred, ddepth); 3099 } else { 3100 set_ctrl(n, pred); 3101 } 3102 } 3103 3104 //------------------------------ proj_clone ------------------------------------- 3105 // Utility to create an if-projection 3106 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) { 3107 ProjNode* c = p->clone()->as_Proj(); 3108 c->set_req(0, iff); 3109 return c; 3110 } 3111 3112 //------------------------------ short_circuit_if ------------------------------------- 3113 // Force the iff control output to be the live_proj 3114 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) { 3115 guarantee(live_proj != nullptr, "null projection"); 3116 int proj_con = live_proj->_con; 3117 assert(proj_con == 0 || proj_con == 1, "false or true projection"); 3118 Node* con = intcon(proj_con); 3119 if (iff) { 3120 iff->set_req(1, con); 3121 } 3122 return con; 3123 } 3124 3125 //------------------------------ insert_if_before_proj ------------------------------------- 3126 // Insert a new if before an if projection (* - new node) 3127 // 3128 // before 3129 // if(test) 3130 // / \ 3131 // v v 3132 // other-proj proj (arg) 3133 // 3134 // after 3135 // if(test) 3136 // / \ 3137 // / v 3138 // | * proj-clone 3139 // v | 3140 // other-proj v 3141 // * new_if(relop(cmp[IU](left,right))) 3142 // / \ 3143 // v v 3144 // * new-proj proj 3145 // (returned) 3146 // 3147 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) { 3148 IfNode* iff = proj->in(0)->as_If(); 3149 IdealLoopTree *loop = get_loop(proj); 3150 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 3151 uint ddepth = dom_depth(proj); 3152 3153 _igvn.rehash_node_delayed(iff); 3154 _igvn.rehash_node_delayed(proj); 3155 3156 proj->set_req(0, nullptr); // temporary disconnect 3157 ProjNode* proj2 = proj_clone(proj, iff); 3158 register_node(proj2, loop, iff, ddepth); 3159 3160 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right); 3161 register_node(cmp, loop, proj2, ddepth); 3162 3163 BoolNode* bol = new BoolNode(cmp, relop); 3164 register_node(bol, loop, proj2, ddepth); 3165 3166 int opcode = iff->Opcode(); 3167 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode"); 3168 IfNode* new_if = IfNode::make_with_same_profile(iff, proj2, bol); 3169 register_node(new_if, loop, proj2, ddepth); 3170 3171 proj->set_req(0, new_if); // reattach 3172 set_idom(proj, new_if, ddepth); 3173 3174 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj(); 3175 guarantee(new_exit != nullptr, "null exit node"); 3176 register_node(new_exit, get_loop(other_proj), new_if, ddepth); 3177 3178 return new_exit; 3179 } 3180 3181 //------------------------------ insert_region_before_proj ------------------------------------- 3182 // Insert a region before an if projection (* - new node) 3183 // 3184 // before 3185 // if(test) 3186 // / | 3187 // v | 3188 // proj v 3189 // other-proj 3190 // 3191 // after 3192 // if(test) 3193 // / | 3194 // v | 3195 // * proj-clone v 3196 // | other-proj 3197 // v 3198 // * new-region 3199 // | 3200 // v 3201 // * dum_if 3202 // / \ 3203 // v \ 3204 // * dum-proj v 3205 // proj 3206 // 3207 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) { 3208 IfNode* iff = proj->in(0)->as_If(); 3209 IdealLoopTree *loop = get_loop(proj); 3210 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 3211 uint ddepth = dom_depth(proj); 3212 3213 _igvn.rehash_node_delayed(iff); 3214 _igvn.rehash_node_delayed(proj); 3215 3216 proj->set_req(0, nullptr); // temporary disconnect 3217 ProjNode* proj2 = proj_clone(proj, iff); 3218 register_node(proj2, loop, iff, ddepth); 3219 3220 RegionNode* reg = new RegionNode(2); 3221 reg->set_req(1, proj2); 3222 register_node(reg, loop, iff, ddepth); 3223 3224 IfNode* dum_if = new IfNode(reg, short_circuit_if(nullptr, proj), iff->_prob, iff->_fcnt); 3225 register_node(dum_if, loop, reg, ddepth); 3226 3227 proj->set_req(0, dum_if); // reattach 3228 set_idom(proj, dum_if, ddepth); 3229 3230 ProjNode* dum_proj = proj_clone(other_proj, dum_if); 3231 register_node(dum_proj, loop, dum_if, ddepth); 3232 3233 return reg; 3234 } 3235 3236 // Idea 3237 // ---- 3238 // Partial Peeling tries to rotate the loop in such a way that it can later be turned into a counted loop. Counted loops 3239 // require a signed loop exit test. When calling this method, we've only found a suitable unsigned test to partial peel 3240 // with. Therefore, we try to split off a signed loop exit test from the unsigned test such that it can be used as new 3241 // loop exit while keeping the unsigned test unchanged and preserving the same behavior as if we've used the unsigned 3242 // test alone instead: 3243 // 3244 // Before Partial Peeling: 3245 // Loop: 3246 // <peeled section> 3247 // Split off signed loop exit test 3248 // <-- CUT HERE --> 3249 // Unchanged unsigned loop exit test 3250 // <rest of unpeeled section> 3251 // goto Loop 3252 // 3253 // After Partial Peeling: 3254 // <cloned peeled section> 3255 // Cloned split off signed loop exit test 3256 // Loop: 3257 // Unchanged unsigned loop exit test 3258 // <rest of unpeeled section> 3259 // <peeled section> 3260 // Split off signed loop exit test 3261 // goto Loop 3262 // 3263 // Details 3264 // ------- 3265 // Before: 3266 // if (i <u limit) Unsigned loop exit condition 3267 // / | 3268 // v v 3269 // exit-proj stay-in-loop-proj 3270 // 3271 // Split off a signed loop exit test (i.e. with CmpI) from an unsigned loop exit test (i.e. with CmpU) and insert it 3272 // before the CmpU on the stay-in-loop path and keep both tests: 3273 // 3274 // if (i <u limit) Signed loop exit test 3275 // / | 3276 // / if (i <u limit) Unsigned loop exit test 3277 // / / | 3278 // v v v 3279 // exit-region stay-in-loop-proj 3280 // 3281 // Implementation 3282 // -------------- 3283 // We need to make sure that the new signed loop exit test is properly inserted into the graph such that the unsigned 3284 // loop exit test still dominates the same set of control nodes, the ctrl() relation from data nodes to both loop 3285 // exit tests is preserved, and their loop nesting is correct. 3286 // 3287 // To achieve that, we clone the unsigned loop exit test completely (leave it unchanged), insert the signed loop exit 3288 // test above it and kill the original unsigned loop exit test by setting it's condition to a constant 3289 // (i.e. stay-in-loop-const in graph below) such that IGVN can fold it later: 3290 // 3291 // if (stay-in-loop-const) Killed original unsigned loop exit test 3292 // / | 3293 // / v 3294 // / if (i < limit) Split off signed loop exit test 3295 // / / | 3296 // / / v 3297 // / / if (i <u limit) Cloned unsigned loop exit test 3298 // / / / | 3299 // v v v | 3300 // exit-region | 3301 // | | 3302 // dummy-if | 3303 // / | | 3304 // dead | | 3305 // v v 3306 // exit-proj stay-in-loop-proj 3307 // 3308 // Note: The dummy-if is inserted to create a region to merge the loop exits between the original to be killed unsigned 3309 // loop exit test and its exit projection while keeping the exit projection (also see insert_region_before_proj()). 3310 // 3311 // Requirements 3312 // ------------ 3313 // Note that we can only split off a signed loop exit test from the unsigned loop exit test when the behavior is exactly 3314 // the same as before with only a single unsigned test. This is only possible if certain requirements are met. 3315 // Otherwise, we need to bail out (see comments in the code below). 3316 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree* loop) { 3317 const bool Signed = true; 3318 const bool Unsigned = false; 3319 3320 BoolNode* bol = if_cmpu->in(1)->as_Bool(); 3321 if (bol->_test._test != BoolTest::lt) { 3322 return nullptr; 3323 } 3324 CmpNode* cmpu = bol->in(1)->as_Cmp(); 3325 assert(cmpu->Opcode() == Op_CmpU, "must be unsigned comparison"); 3326 3327 int stride = stride_of_possible_iv(if_cmpu); 3328 if (stride == 0) { 3329 return nullptr; 3330 } 3331 3332 Node* lp_proj = stay_in_loop(if_cmpu, loop); 3333 guarantee(lp_proj != nullptr, "null loop node"); 3334 3335 ProjNode* lp_continue = lp_proj->as_Proj(); 3336 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj(); 3337 if (!lp_exit->is_IfFalse()) { 3338 // The loop exit condition is (i <u limit) ==> (i >= 0 && i < limit). 3339 // We therefore can't add a single exit condition. 3340 return nullptr; 3341 } 3342 // The unsigned loop exit condition is 3343 // !(i <u limit) 3344 // = i >=u limit 3345 // 3346 // First, we note that for any x for which 3347 // 0 <= x <= INT_MAX 3348 // we can convert x to an unsigned int and still get the same guarantee: 3349 // 0 <= (uint) x <= INT_MAX = (uint) INT_MAX 3350 // 0 <=u (uint) x <=u INT_MAX = (uint) INT_MAX (LEMMA) 3351 // 3352 // With that in mind, if 3353 // limit >= 0 (COND) 3354 // then the unsigned loop exit condition 3355 // i >=u limit (ULE) 3356 // is equivalent to 3357 // i < 0 || i >= limit (SLE-full) 3358 // because either i is negative and therefore always greater than MAX_INT when converting to unsigned 3359 // (uint) i >=u MAX_INT >= limit >= 0 3360 // or otherwise 3361 // i >= limit >= 0 3362 // holds due to (LEMMA). 3363 // 3364 // For completeness, a counterexample with limit < 0: 3365 // Assume i = -3 and limit = -2: 3366 // i < 0 3367 // -2 < 0 3368 // is true and thus also "i < 0 || i >= limit". But 3369 // i >=u limit 3370 // -3 >=u -2 3371 // is false. 3372 Node* limit = cmpu->in(2); 3373 const TypeInt* type_limit = _igvn.type(limit)->is_int(); 3374 if (type_limit->_lo < 0) { 3375 return nullptr; 3376 } 3377 3378 // We prove below that we can extract a single signed loop exit condition from (SLE-full), depending on the stride: 3379 // stride < 0: 3380 // i < 0 (SLE = SLE-negative) 3381 // stride > 0: 3382 // i >= limit (SLE = SLE-positive) 3383 // such that we have the following graph before Partial Peeling with stride > 0 (similar for stride < 0): 3384 // 3385 // Loop: 3386 // <peeled section> 3387 // i >= limit (SLE-positive) 3388 // <-- CUT HERE --> 3389 // i >=u limit (ULE) 3390 // <rest of unpeeled section> 3391 // goto Loop 3392 // 3393 // We exit the loop if: 3394 // (SLE) is true OR (ULE) is true 3395 // However, if (SLE) is true then (ULE) also needs to be true to ensure the exact same behavior. Otherwise, we wrongly 3396 // exit a loop that should not have been exited if we did not apply Partial Peeling. More formally, we need to ensure: 3397 // (SLE) IMPLIES (ULE) 3398 // This indeed holds when (COND) is given: 3399 // - stride > 0: 3400 // i >= limit // (SLE = SLE-positive) 3401 // i >= limit >= 0 // (COND) 3402 // i >=u limit >= 0 // (LEMMA) 3403 // which is the unsigned loop exit condition (ULE). 3404 // - stride < 0: 3405 // i < 0 // (SLE = SLE-negative) 3406 // (uint) i >u MAX_INT // (NEG) all negative values are greater than MAX_INT when converted to unsigned 3407 // MAX_INT >= limit >= 0 // (COND) 3408 // MAX_INT >=u limit >= 0 // (LEMMA) 3409 // and thus from (NEG) and (LEMMA): 3410 // i >=u limit 3411 // which is the unsigned loop exit condition (ULE). 3412 // 3413 // 3414 // After Partial Peeling, we have the following structure for stride > 0 (similar for stride < 0): 3415 // <cloned peeled section> 3416 // i >= limit (SLE-positive) 3417 // Loop: 3418 // i >=u limit (ULE) 3419 // <rest of unpeeled section> 3420 // <peeled section> 3421 // i >= limit (SLE-positive) 3422 // goto Loop 3423 Node* rhs_cmpi; 3424 if (stride > 0) { 3425 rhs_cmpi = limit; // For i >= limit 3426 } else { 3427 rhs_cmpi = makecon(TypeInt::ZERO); // For i < 0 3428 } 3429 // Create a new region on the exit path 3430 RegionNode* reg = insert_region_before_proj(lp_exit); 3431 guarantee(reg != nullptr, "null region node"); 3432 3433 // Clone the if-cmpu-true-false using a signed compare 3434 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge; 3435 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, rhs_cmpi, lp_continue); 3436 reg->add_req(cmpi_exit); 3437 3438 // Clone the if-cmpu-true-false 3439 BoolTest::mask rel_u = bol->_test._test; 3440 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue); 3441 reg->add_req(cmpu_exit); 3442 3443 // Force original if to stay in loop. 3444 short_circuit_if(if_cmpu, lp_continue); 3445 3446 return cmpi_exit->in(0)->as_If(); 3447 } 3448 3449 //------------------------------ remove_cmpi_loop_exit ------------------------------------- 3450 // Remove a previously inserted signed compare loop exit. 3451 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) { 3452 Node* lp_proj = stay_in_loop(if_cmp, loop); 3453 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI && 3454 stay_in_loop(lp_proj, loop)->is_If() && 3455 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu"); 3456 Node* con = makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO); 3457 if_cmp->set_req(1, con); 3458 } 3459 3460 //------------------------------ scheduled_nodelist ------------------------------------- 3461 // Create a post order schedule of nodes that are in the 3462 // "member" set. The list is returned in "sched". 3463 // The first node in "sched" is the loop head, followed by 3464 // nodes which have no inputs in the "member" set, and then 3465 // followed by the nodes that have an immediate input dependence 3466 // on a node in "sched". 3467 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) { 3468 3469 assert(member.test(loop->_head->_idx), "loop head must be in member set"); 3470 VectorSet visited; 3471 Node_Stack nstack(loop->_body.size()); 3472 3473 Node* n = loop->_head; // top of stack is cached in "n" 3474 uint idx = 0; 3475 visited.set(n->_idx); 3476 3477 // Initially push all with no inputs from within member set 3478 for(uint i = 0; i < loop->_body.size(); i++ ) { 3479 Node *elt = loop->_body.at(i); 3480 if (member.test(elt->_idx)) { 3481 bool found = false; 3482 for (uint j = 0; j < elt->req(); j++) { 3483 Node* def = elt->in(j); 3484 if (def && member.test(def->_idx) && def != elt) { 3485 found = true; 3486 break; 3487 } 3488 } 3489 if (!found && elt != loop->_head) { 3490 nstack.push(n, idx); 3491 n = elt; 3492 assert(!visited.test(n->_idx), "not seen yet"); 3493 visited.set(n->_idx); 3494 } 3495 } 3496 } 3497 3498 // traverse out's that are in the member set 3499 while (true) { 3500 if (idx < n->outcnt()) { 3501 Node* use = n->raw_out(idx); 3502 idx++; 3503 if (!visited.test_set(use->_idx)) { 3504 if (member.test(use->_idx)) { 3505 nstack.push(n, idx); 3506 n = use; 3507 idx = 0; 3508 } 3509 } 3510 } else { 3511 // All outputs processed 3512 sched.push(n); 3513 if (nstack.is_empty()) break; 3514 n = nstack.node(); 3515 idx = nstack.index(); 3516 nstack.pop(); 3517 } 3518 } 3519 } 3520 3521 3522 //------------------------------ has_use_in_set ------------------------------------- 3523 // Has a use in the vector set 3524 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) { 3525 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3526 Node* use = n->fast_out(j); 3527 if (vset.test(use->_idx)) { 3528 return true; 3529 } 3530 } 3531 return false; 3532 } 3533 3534 3535 //------------------------------ has_use_internal_to_set ------------------------------------- 3536 // Has use internal to the vector set (ie. not in a phi at the loop head) 3537 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) { 3538 Node* head = loop->_head; 3539 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3540 Node* use = n->fast_out(j); 3541 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) { 3542 return true; 3543 } 3544 } 3545 return false; 3546 } 3547 3548 3549 //------------------------------ clone_for_use_outside_loop ------------------------------------- 3550 // clone "n" for uses that are outside of loop 3551 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) { 3552 int cloned = 0; 3553 assert(worklist.size() == 0, "should be empty"); 3554 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3555 Node* use = n->fast_out(j); 3556 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) { 3557 worklist.push(use); 3558 } 3559 } 3560 3561 if (C->check_node_count(worklist.size() + NodeLimitFudgeFactor, 3562 "Too many clones required in clone_for_use_outside_loop in partial peeling")) { 3563 return -1; 3564 } 3565 3566 while( worklist.size() ) { 3567 Node *use = worklist.pop(); 3568 if (!has_node(use) || use->in(0) == C->top()) continue; 3569 uint j; 3570 for (j = 0; j < use->req(); j++) { 3571 if (use->in(j) == n) break; 3572 } 3573 assert(j < use->req(), "must be there"); 3574 3575 // clone "n" and insert it between the inputs of "n" and the use outside the loop 3576 Node* n_clone = n->clone(); 3577 _igvn.replace_input_of(use, j, n_clone); 3578 cloned++; 3579 Node* use_c; 3580 if (!use->is_Phi()) { 3581 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0); 3582 } else { 3583 // Use in a phi is considered a use in the associated predecessor block 3584 use_c = use->in(0)->in(j); 3585 } 3586 set_ctrl(n_clone, use_c); 3587 assert(!loop->is_member(get_loop(use_c)), "should be outside loop"); 3588 get_loop(use_c)->_body.push(n_clone); 3589 _igvn.register_new_node_with_optimizer(n_clone); 3590 #ifndef PRODUCT 3591 if (TracePartialPeeling) { 3592 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx); 3593 } 3594 #endif 3595 } 3596 return cloned; 3597 } 3598 3599 3600 //------------------------------ clone_for_special_use_inside_loop ------------------------------------- 3601 // clone "n" for special uses that are in the not_peeled region. 3602 // If these def-uses occur in separate blocks, the code generator 3603 // marks the method as not compilable. For example, if a "BoolNode" 3604 // is in a different basic block than the "IfNode" that uses it, then 3605 // the compilation is aborted in the code generator. 3606 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 3607 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) { 3608 if (n->is_Phi() || n->is_Load()) { 3609 return; 3610 } 3611 assert(worklist.size() == 0, "should be empty"); 3612 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3613 Node* use = n->fast_out(j); 3614 if ( not_peel.test(use->_idx) && 3615 (use->is_If() || use->is_CMove() || use->is_Bool() || use->is_OpaqueInitializedAssertionPredicate()) && 3616 use->in(1) == n) { 3617 worklist.push(use); 3618 } 3619 } 3620 if (worklist.size() > 0) { 3621 // clone "n" and insert it between inputs of "n" and the use 3622 Node* n_clone = n->clone(); 3623 loop->_body.push(n_clone); 3624 _igvn.register_new_node_with_optimizer(n_clone); 3625 set_ctrl(n_clone, get_ctrl(n)); 3626 sink_list.push(n_clone); 3627 not_peel.set(n_clone->_idx); 3628 #ifndef PRODUCT 3629 if (TracePartialPeeling) { 3630 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx); 3631 } 3632 #endif 3633 while( worklist.size() ) { 3634 Node *use = worklist.pop(); 3635 _igvn.rehash_node_delayed(use); 3636 for (uint j = 1; j < use->req(); j++) { 3637 if (use->in(j) == n) { 3638 use->set_req(j, n_clone); 3639 } 3640 } 3641 } 3642 } 3643 } 3644 3645 3646 //------------------------------ insert_phi_for_loop ------------------------------------- 3647 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 3648 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) { 3649 Node *phi = PhiNode::make(lp, back_edge_val); 3650 phi->set_req(LoopNode::EntryControl, lp_entry_val); 3651 // Use existing phi if it already exists 3652 Node *hit = _igvn.hash_find_insert(phi); 3653 if( hit == nullptr ) { 3654 _igvn.register_new_node_with_optimizer(phi); 3655 set_ctrl(phi, lp); 3656 } else { 3657 // Remove the new phi from the graph and use the hit 3658 _igvn.remove_dead_node(phi); 3659 phi = hit; 3660 } 3661 _igvn.replace_input_of(use, idx, phi); 3662 } 3663 3664 #ifdef ASSERT 3665 //------------------------------ is_valid_loop_partition ------------------------------------- 3666 // Validate the loop partition sets: peel and not_peel 3667 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, 3668 VectorSet& not_peel ) { 3669 uint i; 3670 // Check that peel_list entries are in the peel set 3671 for (i = 0; i < peel_list.size(); i++) { 3672 if (!peel.test(peel_list.at(i)->_idx)) { 3673 return false; 3674 } 3675 } 3676 // Check at loop members are in one of peel set or not_peel set 3677 for (i = 0; i < loop->_body.size(); i++ ) { 3678 Node *def = loop->_body.at(i); 3679 uint di = def->_idx; 3680 // Check that peel set elements are in peel_list 3681 if (peel.test(di)) { 3682 if (not_peel.test(di)) { 3683 return false; 3684 } 3685 // Must be in peel_list also 3686 bool found = false; 3687 for (uint j = 0; j < peel_list.size(); j++) { 3688 if (peel_list.at(j)->_idx == di) { 3689 found = true; 3690 break; 3691 } 3692 } 3693 if (!found) { 3694 return false; 3695 } 3696 } else if (not_peel.test(di)) { 3697 if (peel.test(di)) { 3698 return false; 3699 } 3700 } else { 3701 return false; 3702 } 3703 } 3704 return true; 3705 } 3706 3707 //------------------------------ is_valid_clone_loop_exit_use ------------------------------------- 3708 // Ensure a use outside of loop is of the right form 3709 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) { 3710 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 3711 return (use->is_Phi() && 3712 use_c->is_Region() && use_c->req() == 3 && 3713 (use_c->in(exit_idx)->Opcode() == Op_IfTrue || 3714 use_c->in(exit_idx)->Opcode() == Op_IfFalse || 3715 use_c->in(exit_idx)->Opcode() == Op_JumpProj) && 3716 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) ); 3717 } 3718 3719 //------------------------------ is_valid_clone_loop_form ------------------------------------- 3720 // Ensure that all uses outside of loop are of the right form 3721 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 3722 uint orig_exit_idx, uint clone_exit_idx) { 3723 uint len = peel_list.size(); 3724 for (uint i = 0; i < len; i++) { 3725 Node *def = peel_list.at(i); 3726 3727 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 3728 Node *use = def->fast_out(j); 3729 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 3730 if (!loop->is_member(get_loop(use_c))) { 3731 // use is not in the loop, check for correct structure 3732 if (use->in(0) == def) { 3733 // Okay 3734 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) { 3735 return false; 3736 } 3737 } 3738 } 3739 } 3740 return true; 3741 } 3742 #endif 3743 3744 //------------------------------ partial_peel ------------------------------------- 3745 // Partially peel (aka loop rotation) the top portion of a loop (called 3746 // the peel section below) by cloning it and placing one copy just before 3747 // the new loop head and the other copy at the bottom of the new loop. 3748 // 3749 // before after where it came from 3750 // 3751 // stmt1 stmt1 3752 // loop: stmt2 clone 3753 // stmt2 if condA goto exitA clone 3754 // if condA goto exitA new_loop: new 3755 // stmt3 stmt3 clone 3756 // if !condB goto loop if condB goto exitB clone 3757 // exitB: stmt2 orig 3758 // stmt4 if !condA goto new_loop orig 3759 // exitA: goto exitA 3760 // exitB: 3761 // stmt4 3762 // exitA: 3763 // 3764 // Step 1: find the cut point: an exit test on probable 3765 // induction variable. 3766 // Step 2: schedule (with cloning) operations in the peel 3767 // section that can be executed after the cut into 3768 // the section that is not peeled. This may need 3769 // to clone operations into exit blocks. For 3770 // instance, a reference to A[i] in the not-peel 3771 // section and a reference to B[i] in an exit block 3772 // may cause a left-shift of i by 2 to be placed 3773 // in the peel block. This step will clone the left 3774 // shift into the exit block and sink the left shift 3775 // from the peel to the not-peel section. 3776 // Step 3: clone the loop, retarget the control, and insert 3777 // phis for values that are live across the new loop 3778 // head. This is very dependent on the graph structure 3779 // from clone_loop. It creates region nodes for 3780 // exit control and associated phi nodes for values 3781 // flow out of the loop through that exit. The region 3782 // node is dominated by the clone's control projection. 3783 // So the clone's peel section is placed before the 3784 // new loop head, and the clone's not-peel section is 3785 // forms the top part of the new loop. The original 3786 // peel section forms the tail of the new loop. 3787 // Step 4: update the dominator tree and recompute the 3788 // dominator depth. 3789 // 3790 // orig 3791 // 3792 // stmt1 3793 // | 3794 // v 3795 // predicates 3796 // | 3797 // v 3798 // loop<----+ 3799 // | | 3800 // stmt2 | 3801 // | | 3802 // v | 3803 // ifA | 3804 // / | | 3805 // v v | 3806 // false true ^ <-- last_peel 3807 // / | | 3808 // / ===|==cut | 3809 // / stmt3 | <-- first_not_peel 3810 // / | | 3811 // | v | 3812 // v ifB | 3813 // exitA: / \ | 3814 // / \ | 3815 // v v | 3816 // false true | 3817 // / \ | 3818 // / ----+ 3819 // | 3820 // v 3821 // exitB: 3822 // stmt4 3823 // 3824 // 3825 // after clone loop 3826 // 3827 // stmt1 3828 // | 3829 // v 3830 // predicates 3831 // / \ 3832 // clone / \ orig 3833 // / \ 3834 // / \ 3835 // v v 3836 // +---->loop loop<----+ 3837 // | | | | 3838 // | stmt2 stmt2 | 3839 // | | | | 3840 // | v v | 3841 // | ifA ifA | 3842 // | | \ / | | 3843 // | v v v v | 3844 // ^ true false false true ^ <-- last_peel 3845 // | | ^ \ / | | 3846 // | cut==|== \ \ / ===|==cut | 3847 // | stmt3 \ \ / stmt3 | <-- first_not_peel 3848 // | | dom | | | | 3849 // | v \ 1v v2 v | 3850 // | ifB regionA ifB | 3851 // | / \ | / \ | 3852 // | / \ v / \ | 3853 // | v v exitA: v v | 3854 // | true false false true | 3855 // | / ^ \ / \ | 3856 // +---- \ \ / ----+ 3857 // dom \ / 3858 // \ 1v v2 3859 // regionB 3860 // | 3861 // v 3862 // exitB: 3863 // stmt4 3864 // 3865 // 3866 // after partial peel 3867 // 3868 // stmt1 3869 // | 3870 // v 3871 // predicates 3872 // / 3873 // clone / orig 3874 // / TOP 3875 // / \ 3876 // v v 3877 // TOP->loop loop----+ 3878 // | | | 3879 // stmt2 stmt2 | 3880 // | | | 3881 // v v | 3882 // ifA ifA | 3883 // | \ / | | 3884 // v v v v | 3885 // true false false true | <-- last_peel 3886 // | ^ \ / +------|---+ 3887 // +->newloop \ \ / === ==cut | | 3888 // | stmt3 \ \ / TOP | | 3889 // | | dom | | stmt3 | | <-- first_not_peel 3890 // | v \ 1v v2 v | | 3891 // | ifB regionA ifB ^ v 3892 // | / \ | / \ | | 3893 // | / \ v / \ | | 3894 // | v v exitA: v v | | 3895 // | true false false true | | 3896 // | / ^ \ / \ | | 3897 // | | \ \ / v | | 3898 // | | dom \ / TOP | | 3899 // | | \ 1v v2 | | 3900 // ^ v regionB | | 3901 // | | | | | 3902 // | | v ^ v 3903 // | | exitB: | | 3904 // | | stmt4 | | 3905 // | +------------>-----------------+ | 3906 // | | 3907 // +-----------------<---------------------+ 3908 // 3909 // 3910 // final graph 3911 // 3912 // stmt1 3913 // | 3914 // v 3915 // predicates 3916 // | 3917 // v 3918 // stmt2 clone 3919 // | 3920 // v 3921 // ........> ifA clone 3922 // : / | 3923 // dom / | 3924 // : v v 3925 // : false true 3926 // : | | 3927 // : | v 3928 // : | newloop<-----+ 3929 // : | | | 3930 // : | stmt3 clone | 3931 // : | | | 3932 // : | v | 3933 // : | ifB | 3934 // : | / \ | 3935 // : | v v | 3936 // : | false true | 3937 // : | | | | 3938 // : | v stmt2 | 3939 // : | exitB: | | 3940 // : | stmt4 v | 3941 // : | ifA orig | 3942 // : | / \ | 3943 // : | / \ | 3944 // : | v v | 3945 // : | false true | 3946 // : | / \ | 3947 // : v v -----+ 3948 // RegionA 3949 // | 3950 // v 3951 // exitA 3952 // 3953 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) { 3954 3955 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only"); 3956 if (!loop->_head->is_Loop()) { 3957 return false; 3958 } 3959 LoopNode *head = loop->_head->as_Loop(); 3960 3961 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) { 3962 return false; 3963 } 3964 3965 // Check for complex exit control 3966 for (uint ii = 0; ii < loop->_body.size(); ii++) { 3967 Node *n = loop->_body.at(ii); 3968 int opc = n->Opcode(); 3969 if (n->is_Call() || 3970 opc == Op_Catch || 3971 opc == Op_CatchProj || 3972 opc == Op_Jump || 3973 opc == Op_JumpProj) { 3974 #ifndef PRODUCT 3975 if (TracePartialPeeling) { 3976 tty->print_cr("\nExit control too complex: lp: %d", head->_idx); 3977 } 3978 #endif 3979 return false; 3980 } 3981 } 3982 3983 int dd = dom_depth(head); 3984 3985 // Step 1: find cut point 3986 3987 // Walk up dominators to loop head looking for first loop exit 3988 // which is executed on every path thru loop. 3989 IfNode *peel_if = nullptr; 3990 IfNode *peel_if_cmpu = nullptr; 3991 3992 Node *iff = loop->tail(); 3993 while (iff != head) { 3994 if (iff->is_If()) { 3995 Node *ctrl = get_ctrl(iff->in(1)); 3996 if (ctrl->is_top()) return false; // Dead test on live IF. 3997 // If loop-varying exit-test, check for induction variable 3998 if (loop->is_member(get_loop(ctrl)) && 3999 loop->is_loop_exit(iff) && 4000 is_possible_iv_test(iff)) { 4001 Node* cmp = iff->in(1)->in(1); 4002 if (cmp->Opcode() == Op_CmpI) { 4003 peel_if = iff->as_If(); 4004 } else { 4005 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU"); 4006 peel_if_cmpu = iff->as_If(); 4007 } 4008 } 4009 } 4010 iff = idom(iff); 4011 } 4012 4013 // Prefer signed compare over unsigned compare. 4014 IfNode* new_peel_if = nullptr; 4015 if (peel_if == nullptr) { 4016 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == nullptr) { 4017 return false; // No peel point found 4018 } 4019 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop); 4020 if (new_peel_if == nullptr) { 4021 return false; // No peel point found 4022 } 4023 peel_if = new_peel_if; 4024 } 4025 Node* last_peel = stay_in_loop(peel_if, loop); 4026 Node* first_not_peeled = stay_in_loop(last_peel, loop); 4027 if (first_not_peeled == nullptr || first_not_peeled == head) { 4028 return false; 4029 } 4030 4031 #ifndef PRODUCT 4032 if (TraceLoopOpts) { 4033 tty->print("PartialPeel "); 4034 loop->dump_head(); 4035 } 4036 4037 if (TracePartialPeeling) { 4038 tty->print_cr("before partial peel one iteration"); 4039 Node_List wl; 4040 Node* t = head->in(2); 4041 while (true) { 4042 wl.push(t); 4043 if (t == head) break; 4044 t = idom(t); 4045 } 4046 while (wl.size() > 0) { 4047 Node* tt = wl.pop(); 4048 tt->dump(); 4049 if (tt == last_peel) tty->print_cr("-- cut --"); 4050 } 4051 } 4052 #endif 4053 4054 C->print_method(PHASE_BEFORE_PARTIAL_PEELING, 4, head); 4055 4056 VectorSet peel; 4057 VectorSet not_peel; 4058 Node_List peel_list; 4059 Node_List worklist; 4060 Node_List sink_list; 4061 4062 uint estimate = loop->est_loop_clone_sz(1); 4063 if (exceeding_node_budget(estimate)) { 4064 return false; 4065 } 4066 4067 // Set of cfg nodes to peel are those that are executable from 4068 // the head through last_peel. 4069 assert(worklist.size() == 0, "should be empty"); 4070 worklist.push(head); 4071 peel.set(head->_idx); 4072 while (worklist.size() > 0) { 4073 Node *n = worklist.pop(); 4074 if (n != last_peel) { 4075 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 4076 Node* use = n->fast_out(j); 4077 if (use->is_CFG() && 4078 loop->is_member(get_loop(use)) && 4079 !peel.test_set(use->_idx)) { 4080 worklist.push(use); 4081 } 4082 } 4083 } 4084 } 4085 4086 // Set of non-cfg nodes to peel are those that are control 4087 // dependent on the cfg nodes. 4088 for (uint i = 0; i < loop->_body.size(); i++) { 4089 Node *n = loop->_body.at(i); 4090 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n; 4091 if (peel.test(n_c->_idx)) { 4092 peel.set(n->_idx); 4093 } else { 4094 not_peel.set(n->_idx); 4095 } 4096 } 4097 4098 // Step 2: move operations from the peeled section down into the 4099 // not-peeled section 4100 4101 // Get a post order schedule of nodes in the peel region 4102 // Result in right-most operand. 4103 scheduled_nodelist(loop, peel, peel_list); 4104 4105 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 4106 4107 // For future check for too many new phis 4108 uint old_phi_cnt = 0; 4109 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 4110 Node* use = head->fast_out(j); 4111 if (use->is_Phi()) old_phi_cnt++; 4112 } 4113 4114 #ifndef PRODUCT 4115 if (TracePartialPeeling) { 4116 tty->print_cr("\npeeled list"); 4117 } 4118 #endif 4119 4120 // Evacuate nodes in peel region into the not_peeled region if possible 4121 bool too_many_clones = false; 4122 uint new_phi_cnt = 0; 4123 uint cloned_for_outside_use = 0; 4124 for (uint i = 0; i < peel_list.size();) { 4125 Node* n = peel_list.at(i); 4126 #ifndef PRODUCT 4127 if (TracePartialPeeling) n->dump(); 4128 #endif 4129 bool incr = true; 4130 if (!n->is_CFG()) { 4131 if (has_use_in_set(n, not_peel)) { 4132 // If not used internal to the peeled region, 4133 // move "n" from peeled to not_peeled region. 4134 if (!has_use_internal_to_set(n, peel, loop)) { 4135 // if not pinned and not a load (which maybe anti-dependent on a store) 4136 // and not a CMove (Matcher expects only bool->cmove). 4137 if (n->in(0) == nullptr && !n->is_Load() && !n->is_CMove()) { 4138 int new_clones = clone_for_use_outside_loop(loop, n, worklist); 4139 if (C->failing()) return false; 4140 if (new_clones == -1) { 4141 too_many_clones = true; 4142 break; 4143 } 4144 cloned_for_outside_use += new_clones; 4145 sink_list.push(n); 4146 peel.remove(n->_idx); 4147 not_peel.set(n->_idx); 4148 peel_list.remove(i); 4149 incr = false; 4150 #ifndef PRODUCT 4151 if (TracePartialPeeling) { 4152 tty->print_cr("sink to not_peeled region: %d newbb: %d", 4153 n->_idx, get_ctrl(n)->_idx); 4154 } 4155 #endif 4156 } 4157 } else { 4158 // Otherwise check for special def-use cases that span 4159 // the peel/not_peel boundary such as bool->if 4160 clone_for_special_use_inside_loop(loop, n, not_peel, sink_list, worklist); 4161 new_phi_cnt++; 4162 } 4163 } 4164 } 4165 if (incr) i++; 4166 } 4167 4168 estimate += cloned_for_outside_use + new_phi_cnt; 4169 bool exceed_node_budget = !may_require_nodes(estimate); 4170 bool exceed_phi_limit = new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta; 4171 4172 if (too_many_clones || exceed_node_budget || exceed_phi_limit) { 4173 #ifndef PRODUCT 4174 if (TracePartialPeeling && exceed_phi_limit) { 4175 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c", 4176 new_phi_cnt, old_phi_cnt, new_peel_if != nullptr?'T':'F'); 4177 } 4178 #endif 4179 if (new_peel_if != nullptr) { 4180 remove_cmpi_loop_exit(new_peel_if, loop); 4181 } 4182 // Inhibit more partial peeling on this loop 4183 assert(!head->is_partial_peel_loop(), "not partial peeled"); 4184 head->mark_partial_peel_failed(); 4185 if (cloned_for_outside_use > 0) { 4186 // Terminate this round of loop opts because 4187 // the graph outside this loop was changed. 4188 C->set_major_progress(); 4189 return true; 4190 } 4191 return false; 4192 } 4193 4194 // Step 3: clone loop, retarget control, and insert new phis 4195 4196 // Create new loop head for new phis and to hang 4197 // the nodes being moved (sinked) from the peel region. 4198 LoopNode* new_head = new LoopNode(last_peel, last_peel); 4199 new_head->set_unswitch_count(head->unswitch_count()); // Preserve 4200 _igvn.register_new_node_with_optimizer(new_head); 4201 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled"); 4202 _igvn.replace_input_of(first_not_peeled, 0, new_head); 4203 set_loop(new_head, loop); 4204 loop->_body.push(new_head); 4205 not_peel.set(new_head->_idx); 4206 set_idom(new_head, last_peel, dom_depth(first_not_peeled)); 4207 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled)); 4208 4209 while (sink_list.size() > 0) { 4210 Node* n = sink_list.pop(); 4211 set_ctrl(n, new_head); 4212 } 4213 4214 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 4215 4216 clone_loop(loop, old_new, dd, IgnoreStripMined); 4217 4218 const uint clone_exit_idx = 1; 4219 const uint orig_exit_idx = 2; 4220 assert(is_valid_clone_loop_form(loop, peel_list, orig_exit_idx, clone_exit_idx), "bad clone loop"); 4221 4222 Node* head_clone = old_new[head->_idx]; 4223 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop(); 4224 Node* orig_tail_clone = head_clone->in(2); 4225 4226 // Add phi if "def" node is in peel set and "use" is not 4227 4228 for (uint i = 0; i < peel_list.size(); i++) { 4229 Node *def = peel_list.at(i); 4230 if (!def->is_CFG()) { 4231 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 4232 Node *use = def->fast_out(j); 4233 if (has_node(use) && use->in(0) != C->top() && 4234 (!peel.test(use->_idx) || 4235 (use->is_Phi() && use->in(0) == head)) ) { 4236 worklist.push(use); 4237 } 4238 } 4239 while( worklist.size() ) { 4240 Node *use = worklist.pop(); 4241 for (uint j = 1; j < use->req(); j++) { 4242 Node* n = use->in(j); 4243 if (n == def) { 4244 4245 // "def" is in peel set, "use" is not in peel set 4246 // or "use" is in the entry boundary (a phi) of the peel set 4247 4248 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use; 4249 4250 if ( loop->is_member(get_loop( use_c )) ) { 4251 // use is in loop 4252 if (old_new[use->_idx] != nullptr) { // null for dead code 4253 Node* use_clone = old_new[use->_idx]; 4254 _igvn.replace_input_of(use, j, C->top()); 4255 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone ); 4256 } 4257 } else { 4258 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format"); 4259 // use is not in the loop, check if the live range includes the cut 4260 Node* lp_if = use_c->in(orig_exit_idx)->in(0); 4261 if (not_peel.test(lp_if->_idx)) { 4262 assert(j == orig_exit_idx, "use from original loop"); 4263 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone ); 4264 } 4265 } 4266 } 4267 } 4268 } 4269 } 4270 } 4271 4272 // Step 3b: retarget control 4273 4274 // Redirect control to the new loop head if a cloned node in 4275 // the not_peeled region has control that points into the peeled region. 4276 // This necessary because the cloned peeled region will be outside 4277 // the loop. 4278 // from to 4279 // cloned-peeled <---+ 4280 // new_head_clone: | <--+ 4281 // cloned-not_peeled in(0) in(0) 4282 // orig-peeled 4283 4284 for (uint i = 0; i < loop->_body.size(); i++) { 4285 Node *n = loop->_body.at(i); 4286 if (!n->is_CFG() && n->in(0) != nullptr && 4287 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) { 4288 Node* n_clone = old_new[n->_idx]; 4289 if (n_clone->depends_only_on_test()) { 4290 // Pin array access nodes: control is updated here to the loop head. If, after some transformations, the 4291 // backedge is removed, an array load could become dependent on a condition that's not a range check for that 4292 // access. If that condition is replaced by an identical dominating one, then an unpinned load would risk 4293 // floating above its range check. 4294 Node* pinned_clone = n_clone->pin_array_access_node(); 4295 if (pinned_clone != nullptr) { 4296 register_new_node_with_ctrl_of(pinned_clone, n_clone); 4297 old_new.map(n->_idx, pinned_clone); 4298 _igvn.replace_node(n_clone, pinned_clone); 4299 n_clone = pinned_clone; 4300 } 4301 } 4302 _igvn.replace_input_of(n_clone, 0, new_head_clone); 4303 } 4304 } 4305 4306 // Backedge of the surviving new_head (the clone) is original last_peel 4307 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel); 4308 4309 // Cut first node in original not_peel set 4310 _igvn.rehash_node_delayed(new_head); // Multiple edge updates: 4311 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of 4312 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls 4313 4314 // Copy head_clone back-branch info to original head 4315 // and remove original head's loop entry and 4316 // clone head's back-branch 4317 _igvn.rehash_node_delayed(head); // Multiple edge updates 4318 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl)); 4319 head->set_req(LoopNode::LoopBackControl, C->top()); 4320 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top()); 4321 4322 // Similarly modify the phis 4323 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) { 4324 Node* use = head->fast_out(k); 4325 if (use->is_Phi() && use->outcnt() > 0) { 4326 Node* use_clone = old_new[use->_idx]; 4327 _igvn.rehash_node_delayed(use); // Multiple edge updates 4328 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl)); 4329 use->set_req(LoopNode::LoopBackControl, C->top()); 4330 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top()); 4331 } 4332 } 4333 4334 // Step 4: update dominator tree and dominator depth 4335 4336 set_idom(head, orig_tail_clone, dd); 4337 recompute_dom_depth(); 4338 4339 // Inhibit more partial peeling on this loop 4340 new_head_clone->set_partial_peel_loop(); 4341 C->set_major_progress(); 4342 loop->record_for_igvn(); 4343 4344 #ifndef PRODUCT 4345 if (TracePartialPeeling) { 4346 tty->print_cr("\nafter partial peel one iteration"); 4347 Node_List wl; 4348 Node* t = last_peel; 4349 while (true) { 4350 wl.push(t); 4351 if (t == head_clone) break; 4352 t = idom(t); 4353 } 4354 while (wl.size() > 0) { 4355 Node* tt = wl.pop(); 4356 if (tt == head) tty->print_cr("orig head"); 4357 else if (tt == new_head_clone) tty->print_cr("new head"); 4358 else if (tt == head_clone) tty->print_cr("clone head"); 4359 tt->dump(); 4360 } 4361 } 4362 #endif 4363 4364 C->print_method(PHASE_AFTER_PARTIAL_PEELING, 4, new_head_clone); 4365 4366 return true; 4367 } 4368 4369 // Transform: 4370 // 4371 // loop<-----------------+ 4372 // | | 4373 // stmt1 stmt2 .. stmtn | 4374 // | | | | 4375 // \ | / | 4376 // v v v | 4377 // region | 4378 // | | 4379 // shared_stmt | 4380 // | | 4381 // v | 4382 // if | 4383 // / \ | 4384 // | -----------+ 4385 // v 4386 // 4387 // into: 4388 // 4389 // loop<-------------------+ 4390 // | | 4391 // v | 4392 // +->loop | 4393 // | | | 4394 // | stmt1 stmt2 .. stmtn | 4395 // | | | | | 4396 // | | \ / | 4397 // | | v v | 4398 // | | region1 | 4399 // | | | | 4400 // | shared_stmt shared_stmt | 4401 // | | | | 4402 // | v v | 4403 // | if if | 4404 // | /\ / \ | 4405 // +-- | | -------+ 4406 // \ / 4407 // v v 4408 // region2 4409 // 4410 // (region2 is shown to merge mirrored projections of the loop exit 4411 // ifs to make the diagram clearer but they really merge the same 4412 // projection) 4413 // 4414 // Conditions for this transformation to trigger: 4415 // - the path through stmt1 is frequent enough 4416 // - the inner loop will be turned into a counted loop after transformation 4417 bool PhaseIdealLoop::duplicate_loop_backedge(IdealLoopTree *loop, Node_List &old_new) { 4418 if (!DuplicateBackedge) { 4419 return false; 4420 } 4421 assert(!loop->_head->is_CountedLoop() || StressDuplicateBackedge, "Non-counted loop only"); 4422 if (!loop->_head->is_Loop()) { 4423 return false; 4424 } 4425 4426 uint estimate = loop->est_loop_clone_sz(1); 4427 if (exceeding_node_budget(estimate)) { 4428 return false; 4429 } 4430 4431 LoopNode *head = loop->_head->as_Loop(); 4432 4433 Node* region = nullptr; 4434 IfNode* exit_test = nullptr; 4435 uint inner; 4436 float f; 4437 if (StressDuplicateBackedge) { 4438 if (head->is_strip_mined()) { 4439 return false; 4440 } 4441 Node* c = head->in(LoopNode::LoopBackControl); 4442 4443 while (c != head) { 4444 if (c->is_Region()) { 4445 region = c; 4446 } 4447 c = idom(c); 4448 } 4449 4450 if (region == nullptr) { 4451 return false; 4452 } 4453 4454 inner = 1; 4455 } else { 4456 // Is the shape of the loop that of a counted loop... 4457 Node* back_control = loop_exit_control(head, loop); 4458 if (back_control == nullptr) { 4459 return false; 4460 } 4461 4462 BoolTest::mask bt = BoolTest::illegal; 4463 float cl_prob = 0; 4464 Node* incr = nullptr; 4465 Node* limit = nullptr; 4466 Node* cmp = loop_exit_test(back_control, loop, incr, limit, bt, cl_prob); 4467 if (cmp == nullptr || cmp->Opcode() != Op_CmpI) { 4468 return false; 4469 } 4470 4471 // With an extra phi for the candidate iv? 4472 // Or the region node is the loop head 4473 if (!incr->is_Phi() || incr->in(0) == head) { 4474 return false; 4475 } 4476 4477 PathFrequency pf(head, this); 4478 region = incr->in(0); 4479 4480 // Go over all paths for the extra phi's region and see if that 4481 // path is frequent enough and would match the expected iv shape 4482 // if the extra phi is removed 4483 inner = 0; 4484 for (uint i = 1; i < incr->req(); ++i) { 4485 Node* in = incr->in(i); 4486 Node* trunc1 = nullptr; 4487 Node* trunc2 = nullptr; 4488 const TypeInteger* iv_trunc_t = nullptr; 4489 Node* orig_in = in; 4490 if (!(in = CountedLoopNode::match_incr_with_optional_truncation(in, &trunc1, &trunc2, &iv_trunc_t, T_INT))) { 4491 continue; 4492 } 4493 assert(in->Opcode() == Op_AddI, "wrong increment code"); 4494 Node* xphi = nullptr; 4495 Node* stride = loop_iv_stride(in, xphi); 4496 4497 if (stride == nullptr) { 4498 continue; 4499 } 4500 4501 PhiNode* phi = loop_iv_phi(xphi, nullptr, head); 4502 if (phi == nullptr || 4503 (trunc1 == nullptr && phi->in(LoopNode::LoopBackControl) != incr) || 4504 (trunc1 != nullptr && phi->in(LoopNode::LoopBackControl) != trunc1)) { 4505 return false; 4506 } 4507 4508 f = pf.to(region->in(i)); 4509 if (f > 0.5) { 4510 inner = i; 4511 break; 4512 } 4513 } 4514 4515 if (inner == 0) { 4516 return false; 4517 } 4518 4519 exit_test = back_control->in(0)->as_If(); 4520 } 4521 4522 if (idom(region)->is_Catch()) { 4523 return false; 4524 } 4525 4526 // Collect all control nodes that need to be cloned (shared_stmt in the diagram) 4527 Unique_Node_List wq; 4528 wq.push(head->in(LoopNode::LoopBackControl)); 4529 for (uint i = 0; i < wq.size(); i++) { 4530 Node* c = wq.at(i); 4531 assert(get_loop(c) == loop, "not in the right loop?"); 4532 if (c->is_Region()) { 4533 if (c != region) { 4534 for (uint j = 1; j < c->req(); ++j) { 4535 wq.push(c->in(j)); 4536 } 4537 } 4538 } else { 4539 wq.push(c->in(0)); 4540 } 4541 assert(!is_strict_dominator(c, region), "shouldn't go above region"); 4542 } 4543 4544 Node* region_dom = idom(region); 4545 4546 // Can't do the transformation if this would cause a membar pair to 4547 // be split 4548 for (uint i = 0; i < wq.size(); i++) { 4549 Node* c = wq.at(i); 4550 if (c->is_MemBar() && (c->as_MemBar()->trailing_store() || c->as_MemBar()->trailing_load_store())) { 4551 assert(c->as_MemBar()->leading_membar()->trailing_membar() == c, "bad membar pair"); 4552 if (!wq.member(c->as_MemBar()->leading_membar())) { 4553 return false; 4554 } 4555 } 4556 } 4557 C->print_method(PHASE_BEFORE_DUPLICATE_LOOP_BACKEDGE, 4, head); 4558 4559 // Collect data nodes that need to be clones as well 4560 int dd = dom_depth(head); 4561 4562 for (uint i = 0; i < loop->_body.size(); ++i) { 4563 Node* n = loop->_body.at(i); 4564 if (has_ctrl(n)) { 4565 Node* c = get_ctrl(n); 4566 if (wq.member(c)) { 4567 wq.push(n); 4568 } 4569 } else { 4570 set_idom(n, idom(n), dd); 4571 } 4572 } 4573 4574 // clone shared_stmt 4575 clone_loop_body(wq, old_new, nullptr); 4576 4577 Node* region_clone = old_new[region->_idx]; 4578 region_clone->set_req(inner, C->top()); 4579 set_idom(region, region->in(inner), dd); 4580 4581 // Prepare the outer loop 4582 Node* outer_head = new LoopNode(head->in(LoopNode::EntryControl), old_new[head->in(LoopNode::LoopBackControl)->_idx]); 4583 register_control(outer_head, loop->_parent, outer_head->in(LoopNode::EntryControl)); 4584 _igvn.replace_input_of(head, LoopNode::EntryControl, outer_head); 4585 set_idom(head, outer_head, dd); 4586 4587 fix_body_edges(wq, loop, old_new, dd, loop->_parent, true); 4588 4589 // Make one of the shared_stmt copies only reachable from stmt1, the 4590 // other only from stmt2..stmtn. 4591 Node* dom = nullptr; 4592 for (uint i = 1; i < region->req(); ++i) { 4593 if (i != inner) { 4594 _igvn.replace_input_of(region, i, C->top()); 4595 } 4596 Node* in = region_clone->in(i); 4597 if (in->is_top()) { 4598 continue; 4599 } 4600 if (dom == nullptr) { 4601 dom = in; 4602 } else { 4603 dom = dom_lca(dom, in); 4604 } 4605 } 4606 4607 set_idom(region_clone, dom, dd); 4608 4609 // Set up the outer loop 4610 for (uint i = 0; i < head->outcnt(); i++) { 4611 Node* u = head->raw_out(i); 4612 if (u->is_Phi()) { 4613 Node* outer_phi = u->clone(); 4614 outer_phi->set_req(0, outer_head); 4615 Node* backedge = old_new[u->in(LoopNode::LoopBackControl)->_idx]; 4616 if (backedge == nullptr) { 4617 backedge = u->in(LoopNode::LoopBackControl); 4618 } 4619 outer_phi->set_req(LoopNode::LoopBackControl, backedge); 4620 register_new_node(outer_phi, outer_head); 4621 _igvn.replace_input_of(u, LoopNode::EntryControl, outer_phi); 4622 } 4623 } 4624 4625 // create control and data nodes for out of loop uses (including region2) 4626 Node_List worklist; 4627 uint new_counter = C->unique(); 4628 fix_ctrl_uses(wq, loop, old_new, ControlAroundStripMined, outer_head, nullptr, worklist); 4629 4630 Node_List *split_if_set = nullptr; 4631 Node_List *split_bool_set = nullptr; 4632 Node_List *split_cex_set = nullptr; 4633 fix_data_uses(wq, loop, ControlAroundStripMined, loop->skip_strip_mined(), new_counter, old_new, worklist, 4634 split_if_set, split_bool_set, split_cex_set); 4635 4636 finish_clone_loop(split_if_set, split_bool_set, split_cex_set); 4637 4638 if (exit_test != nullptr) { 4639 float cnt = exit_test->_fcnt; 4640 if (cnt != COUNT_UNKNOWN) { 4641 exit_test->_fcnt = cnt * f; 4642 old_new[exit_test->_idx]->as_If()->_fcnt = cnt * (1 - f); 4643 } 4644 } 4645 4646 C->set_major_progress(); 4647 4648 C->print_method(PHASE_AFTER_DUPLICATE_LOOP_BACKEDGE, 4, outer_head); 4649 4650 return true; 4651 } 4652 4653 // AutoVectorize the loop: replace scalar ops with vector ops. 4654 PhaseIdealLoop::AutoVectorizeStatus 4655 PhaseIdealLoop::auto_vectorize(IdealLoopTree* lpt, VSharedData &vshared) { 4656 // Counted loop only 4657 if (!lpt->is_counted()) { 4658 return AutoVectorizeStatus::Impossible; 4659 } 4660 4661 // Main-loop only 4662 CountedLoopNode* cl = lpt->_head->as_CountedLoop(); 4663 if (!cl->is_main_loop()) { 4664 return AutoVectorizeStatus::Impossible; 4665 } 4666 4667 VLoop vloop(lpt, false); 4668 if (!vloop.check_preconditions()) { 4669 return AutoVectorizeStatus::TriedAndFailed; 4670 } 4671 4672 // Ensure the shared data is cleared before each use 4673 vshared.clear(); 4674 4675 const VLoopAnalyzer vloop_analyzer(vloop, vshared); 4676 if (!vloop_analyzer.success()) { 4677 return AutoVectorizeStatus::TriedAndFailed; 4678 } 4679 4680 SuperWord sw(vloop_analyzer); 4681 if (!sw.transform_loop()) { 4682 return AutoVectorizeStatus::TriedAndFailed; 4683 } 4684 4685 return AutoVectorizeStatus::Success; 4686 } 4687 4688 // Just before insert_pre_post_loops, we can multiversion the loop: 4689 // 4690 // multiversion_if 4691 // | | 4692 // fast_loop slow_loop 4693 // 4694 // In the fast_loop we can make speculative assumptions, and put the 4695 // conditions into the multiversion_if. If the conditions hold at runtime, 4696 // we enter the fast_loop, if the conditions fail, we take the slow_loop 4697 // instead which does not make any of the speculative assumptions. 4698 // 4699 // Note: we only multiversion the loop if the loop does not have any 4700 // auto vectorization check Predicate. If we have that predicate, 4701 // then we can simply add the speculative assumption checks to 4702 // that Predicate. This means we do not need to duplicate the 4703 // loop - we have a smaller graph and save compile time. Should 4704 // the conditions ever fail, then we deopt / trap at the Predicate 4705 // and recompile without that Predicate. At that point we will 4706 // multiversion the loop, so that we can still have speculative 4707 // runtime checks. 4708 // 4709 // We perform the multiversioning when the loop is still in its single 4710 // iteration form, even before we insert pre and post loops. This makes 4711 // the cloning much simpler. However, this means that both the fast 4712 // and the slow loop have to be optimized independently (adding pre 4713 // and post loops, unrolling the main loop, auto-vectorize etc.). And 4714 // we may end up not needing any speculative assumptions in the fast_loop 4715 // and then rejecting the slow_loop by constant folding the multiversion_if. 4716 // 4717 // Therefore, we "delay" the optimization of the slow_loop until we add 4718 // at least one speculative assumption for the fast_loop. If we never 4719 // add such a speculative runtime check, the OpaqueMultiversioningNode 4720 // of the multiversion_if constant folds to true after loop opts, and the 4721 // multiversion_if folds away the "delayed" slow_loop. If we add any 4722 // speculative assumption, then we notify the OpaqueMultiversioningNode 4723 // with "notify_slow_loop_that_it_can_resume_optimizations". 4724 // 4725 // Note: new runtime checks can be added to the multiversion_if with 4726 // PhaseIdealLoop::create_new_if_for_multiversion 4727 void PhaseIdealLoop::maybe_multiversion_for_auto_vectorization_runtime_checks(IdealLoopTree* lpt, Node_List& old_new) { 4728 CountedLoopNode* cl = lpt->_head->as_CountedLoop(); 4729 LoopNode* outer_loop = cl->skip_strip_mined(); 4730 Node* entry = outer_loop->in(LoopNode::EntryControl); 4731 4732 // Check we have multiversioning enabled, and are not already multiversioned. 4733 if (!LoopMultiversioning || cl->is_multiversion()) { return; } 4734 4735 // Check that we do not have a parse-predicate where we can add the runtime checks 4736 // during auto-vectorization. 4737 const Predicates predicates(entry); 4738 const PredicateBlock* predicate_block = predicates.auto_vectorization_check_block(); 4739 if (predicate_block->has_parse_predicate()) { return; } 4740 4741 // Check node budget. 4742 uint estimate = lpt->est_loop_clone_sz(2); 4743 if (!may_require_nodes(estimate)) { return; } 4744 4745 do_multiversioning(lpt, old_new); 4746 } 4747 4748 void DataNodeGraph::clone_data_nodes(Node* new_ctrl) { 4749 for (uint i = 0; i < _data_nodes.size(); i++) { 4750 clone(_data_nodes[i], new_ctrl); 4751 } 4752 } 4753 4754 // Clone the given node and set it up properly. Set 'new_ctrl' as ctrl. 4755 void DataNodeGraph::clone(Node* node, Node* new_ctrl) { 4756 Node* clone = node->clone(); 4757 _phase->igvn().register_new_node_with_optimizer(clone); 4758 _orig_to_new.put(node, clone); 4759 _phase->set_ctrl(clone, new_ctrl); 4760 if (node->is_CastII()) { 4761 clone->set_req(0, new_ctrl); 4762 } 4763 } 4764 4765 // Rewire the data inputs of all (unprocessed) cloned nodes, whose inputs are still pointing to the same inputs as their 4766 // corresponding orig nodes, to the newly cloned inputs to create a separate cloned graph. 4767 void DataNodeGraph::rewire_clones_to_cloned_inputs() { 4768 _orig_to_new.iterate_all([&](Node* node, Node* clone) { 4769 for (uint i = 1; i < node->req(); i++) { 4770 Node** cloned_input = _orig_to_new.get(node->in(i)); 4771 if (cloned_input != nullptr) { 4772 // Input was also cloned -> rewire clone to the cloned input. 4773 _phase->igvn().replace_input_of(clone, i, *cloned_input); 4774 } 4775 } 4776 }); 4777 } 4778 4779 // Clone all non-OpaqueLoop* nodes and apply the provided transformation strategy for OpaqueLoop* nodes. 4780 // Set 'new_ctrl' as ctrl for all cloned non-OpaqueLoop* nodes. 4781 void DataNodeGraph::clone_data_nodes_and_transform_opaque_loop_nodes( 4782 const TransformStrategyForOpaqueLoopNodes& transform_strategy, 4783 Node* new_ctrl) { 4784 for (uint i = 0; i < _data_nodes.size(); i++) { 4785 Node* data_node = _data_nodes[i]; 4786 if (data_node->is_Opaque1()) { 4787 transform_opaque_node(transform_strategy, data_node); 4788 } else { 4789 clone(data_node, new_ctrl); 4790 } 4791 } 4792 } 4793 4794 void DataNodeGraph::transform_opaque_node(const TransformStrategyForOpaqueLoopNodes& transform_strategy, Node* node) { 4795 Node* transformed_node; 4796 if (node->is_OpaqueLoopInit()) { 4797 transformed_node = transform_strategy.transform_opaque_init(node->as_OpaqueLoopInit()); 4798 } else { 4799 assert(node->is_OpaqueLoopStride(), "must be OpaqueLoopStrideNode"); 4800 transformed_node = transform_strategy.transform_opaque_stride(node->as_OpaqueLoopStride()); 4801 } 4802 // Add an orig->new mapping to correctly update the inputs of the copied graph in rewire_clones_to_cloned_inputs(). 4803 _orig_to_new.put(node, transformed_node); 4804 } --- EOF ---