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/castnode.hpp" 34 #include "opto/divnode.hpp" 35 #include "opto/loopnode.hpp" 36 #include "opto/matcher.hpp" 37 #include "opto/mulnode.hpp" 38 #include "opto/movenode.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/macros.hpp" 46 47 //============================================================================= 48 //------------------------------split_thru_phi--------------------------------- 49 // Split Node 'n' through merge point if there is enough win. 50 Node* PhaseIdealLoop::split_thru_phi(Node* n, Node* region, int policy) { 51 if ((n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) || 52 (n->Opcode() == Op_ConvL2I && n->bottom_type() != TypeInt::INT)) { 53 // ConvI2L/ConvL2I may have type information on it which is unsafe to push up 54 // so disable this for now 55 return nullptr; 56 } 57 58 // Splitting range check CastIIs through a loop induction Phi can 59 // cause new Phis to be created that are left unrelated to the loop 60 // induction Phi and prevent optimizations (vectorization) 61 if (n->Opcode() == Op_CastII && region->is_CountedLoop() && 62 n->in(1) == region->as_CountedLoop()->phi()) { 63 return nullptr; 64 } 65 66 if (cannot_split_division(n, region)) { 67 return nullptr; 68 } 69 70 int wins = 0; 71 assert(!n->is_CFG(), ""); 72 assert(region->is_Region(), ""); 73 74 const Type* type = n->bottom_type(); 75 const TypeOopPtr* t_oop = _igvn.type(n)->isa_oopptr(); 76 Node* phi; 77 if (t_oop != nullptr && t_oop->is_known_instance_field()) { 78 int iid = t_oop->instance_id(); 79 int index = C->get_alias_index(t_oop); 80 int offset = t_oop->offset(); 81 phi = new PhiNode(region, type, nullptr, iid, index, offset); 82 } else { 83 phi = PhiNode::make_blank(region, n); 84 } 85 uint old_unique = C->unique(); 86 for (uint i = 1; i < region->req(); i++) { 87 Node* x; 88 Node* the_clone = nullptr; 89 if (region->in(i) == C->top()) { 90 x = C->top(); // Dead path? Use a dead data op 91 } else { 92 x = n->clone(); // Else clone up the data op 93 the_clone = x; // Remember for possible deletion. 94 // Alter data node to use pre-phi inputs 95 if (n->in(0) == region) 96 x->set_req( 0, region->in(i) ); 97 for (uint j = 1; j < n->req(); j++) { 98 Node* in = n->in(j); 99 if (in->is_Phi() && in->in(0) == region) 100 x->set_req(j, in->in(i)); // Use pre-Phi input for the clone 101 } 102 } 103 // Check for a 'win' on some paths 104 const Type* t = x->Value(&_igvn); 105 106 bool singleton = t->singleton(); 107 108 // A TOP singleton indicates that there are no possible values incoming 109 // along a particular edge. In most cases, this is OK, and the Phi will 110 // be eliminated later in an Ideal call. However, we can't allow this to 111 // happen if the singleton occurs on loop entry, as the elimination of 112 // the PhiNode may cause the resulting node to migrate back to a previous 113 // loop iteration. 114 if (singleton && t == Type::TOP) { 115 // Is_Loop() == false does not confirm the absence of a loop (e.g., an 116 // irreducible loop may not be indicated by an affirmative is_Loop()); 117 // therefore, the only top we can split thru a phi is on a backedge of 118 // a loop. 119 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 120 } 121 122 if (singleton) { 123 wins++; 124 x = makecon(t); 125 } else { 126 // We now call Identity to try to simplify the cloned node. 127 // Note that some Identity methods call phase->type(this). 128 // Make sure that the type array is big enough for 129 // our new node, even though we may throw the node away. 130 // (Note: This tweaking with igvn only works because x is a new node.) 131 _igvn.set_type(x, t); 132 // If x is a TypeNode, capture any more-precise type permanently into Node 133 // otherwise it will be not updated during igvn->transform since 134 // igvn->type(x) is set to x->Value() already. 135 x->raise_bottom_type(t); 136 Node* y = x->Identity(&_igvn); 137 if (y != x) { 138 wins++; 139 x = y; 140 } else { 141 y = _igvn.hash_find(x); 142 if (y == nullptr) { 143 y = similar_subtype_check(x, region->in(i)); 144 } 145 if (y) { 146 wins++; 147 x = y; 148 } else { 149 // Else x is a new node we are keeping 150 // We do not need register_new_node_with_optimizer 151 // because set_type has already been called. 152 _igvn._worklist.push(x); 153 } 154 } 155 } 156 157 phi->set_req( i, x ); 158 159 if (the_clone == nullptr) { 160 continue; 161 } 162 163 if (the_clone != x) { 164 _igvn.remove_dead_node(the_clone); 165 } else if (region->is_Loop() && i == LoopNode::LoopBackControl && 166 n->is_Load() && can_move_to_inner_loop(n, region->as_Loop(), x)) { 167 // it is not a win if 'x' moved from an outer to an inner loop 168 // this edge case can only happen for Load nodes 169 wins = 0; 170 break; 171 } 172 } 173 // Too few wins? 174 if (wins <= policy) { 175 _igvn.remove_dead_node(phi); 176 return nullptr; 177 } 178 179 // Record Phi 180 register_new_node( phi, region ); 181 182 for (uint i2 = 1; i2 < phi->req(); i2++) { 183 Node *x = phi->in(i2); 184 // If we commoned up the cloned 'x' with another existing Node, 185 // the existing Node picks up a new use. We need to make the 186 // existing Node occur higher up so it dominates its uses. 187 Node *old_ctrl; 188 IdealLoopTree *old_loop; 189 190 if (x->is_Con()) { 191 assert(get_ctrl(x) == C->root(), "constant control is not root"); 192 continue; 193 } 194 // The occasional new node 195 if (x->_idx >= old_unique) { // Found a new, unplaced node? 196 old_ctrl = nullptr; 197 old_loop = nullptr; // Not in any prior loop 198 } else { 199 old_ctrl = get_ctrl(x); 200 old_loop = get_loop(old_ctrl); // Get prior loop 201 } 202 // New late point must dominate new use 203 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2)); 204 if (new_ctrl == old_ctrl) // Nothing is changed 205 continue; 206 207 IdealLoopTree *new_loop = get_loop(new_ctrl); 208 209 // Don't move x into a loop if its uses are 210 // outside of loop. Otherwise x will be cloned 211 // for each use outside of this loop. 212 IdealLoopTree *use_loop = get_loop(region); 213 if (!new_loop->is_member(use_loop) && 214 (old_loop == nullptr || !new_loop->is_member(old_loop))) { 215 // Take early control, later control will be recalculated 216 // during next iteration of loop optimizations. 217 new_ctrl = get_early_ctrl(x); 218 new_loop = get_loop(new_ctrl); 219 } 220 // Set new location 221 set_ctrl(x, new_ctrl); 222 // If changing loop bodies, see if we need to collect into new body 223 if (old_loop != new_loop) { 224 if (old_loop && !old_loop->_child) 225 old_loop->_body.yank(x); 226 if (!new_loop->_child) 227 new_loop->_body.push(x); // Collect body info 228 } 229 } 230 231 return phi; 232 } 233 234 // Test whether node 'x' can move into an inner loop relative to node 'n'. 235 // Note: The test is not exact. Returns true if 'x' COULD end up in an inner loop, 236 // BUT it can also return true and 'x' is in the outer loop 237 bool PhaseIdealLoop::can_move_to_inner_loop(Node* n, LoopNode* n_loop, Node* x) { 238 IdealLoopTree* n_loop_tree = get_loop(n_loop); 239 IdealLoopTree* x_loop_tree = get_loop(get_early_ctrl(x)); 240 // x_loop_tree should be outer or same loop as n_loop_tree 241 return !x_loop_tree->is_member(n_loop_tree); 242 } 243 244 // Subtype checks that carry profile data don't common so look for a replacement by following edges 245 Node* PhaseIdealLoop::similar_subtype_check(const Node* x, Node* r_in) { 246 if (x->is_SubTypeCheck()) { 247 Node* in1 = x->in(1); 248 for (DUIterator_Fast imax, i = in1->fast_outs(imax); i < imax; i++) { 249 Node* u = in1->fast_out(i); 250 if (u != x && u->is_SubTypeCheck() && u->in(1) == x->in(1) && u->in(2) == x->in(2)) { 251 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) { 252 Node* bol = u->fast_out(j); 253 for (DUIterator_Fast kmax, k = bol->fast_outs(kmax); k < kmax; k++) { 254 Node* iff = bol->fast_out(k); 255 // Only dominating subtype checks are interesting: otherwise we risk replacing a subtype check by another with 256 // unrelated profile 257 if (iff->is_If() && is_dominator(iff, r_in)) { 258 return u; 259 } 260 } 261 } 262 } 263 } 264 } 265 return nullptr; 266 } 267 268 // Return true if 'n' is a Div or Mod node (without zero check If node which was removed earlier) with a loop phi divisor 269 // of a trip-counted (integer or long) loop with a backedge input that could be zero (include zero in its type range). In 270 // this case, we cannot split the division to the backedge as it could freely float above the loop exit check resulting in 271 // a division by zero. This situation is possible because the type of an increment node of an iv phi (trip-counter) could 272 // include zero while the iv phi does not (see PhiNode::Value() for trip-counted loops where we improve types of iv phis). 273 // We also need to check other loop phis as they could have been created in the same split-if pass when applying 274 // PhaseIdealLoop::split_thru_phi() to split nodes through an iv phi. 275 bool PhaseIdealLoop::cannot_split_division(const Node* n, const Node* region) const { 276 const Type* zero; 277 switch (n->Opcode()) { 278 case Op_DivI: 279 case Op_ModI: 280 case Op_UDivI: 281 case Op_UModI: 282 zero = TypeInt::ZERO; 283 break; 284 case Op_DivL: 285 case Op_ModL: 286 case Op_UDivL: 287 case Op_UModL: 288 zero = TypeLong::ZERO; 289 break; 290 default: 291 return false; 292 } 293 294 if (n->in(0) != nullptr) { 295 // Cannot split through phi if Div or Mod node has a control dependency to a zero check. 296 return true; 297 } 298 299 Node* divisor = n->in(2); 300 return is_divisor_loop_phi(divisor, region) && 301 loop_phi_backedge_type_contains_zero(divisor, zero); 302 } 303 304 bool PhaseIdealLoop::is_divisor_loop_phi(const Node* divisor, const Node* loop) { 305 return loop->is_Loop() && divisor->is_Phi() && divisor->in(0) == loop; 306 } 307 308 bool PhaseIdealLoop::loop_phi_backedge_type_contains_zero(const Node* phi_divisor, const Type* zero) const { 309 return _igvn.type(phi_divisor->in(LoopNode::LoopBackControl))->filter_speculative(zero) != Type::TOP; 310 } 311 312 //------------------------------dominated_by------------------------------------ 313 // Replace the dominated test with an obvious true or false. Place it on the 314 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the 315 // live path up to the dominating control. 316 void PhaseIdealLoop::dominated_by(IfProjNode* prevdom, IfNode* iff, bool flip, bool pin_array_access_nodes) { 317 if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); } 318 319 // prevdom is the dominating projection of the dominating test. 320 assert(iff->Opcode() == Op_If || 321 iff->Opcode() == Op_CountedLoopEnd || 322 iff->Opcode() == Op_LongCountedLoopEnd || 323 iff->Opcode() == Op_RangeCheck || 324 iff->Opcode() == Op_ParsePredicate, 325 "Check this code when new subtype is added"); 326 327 int pop = prevdom->Opcode(); 328 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" ); 329 if (flip) { 330 if (pop == Op_IfTrue) 331 pop = Op_IfFalse; 332 else 333 pop = Op_IfTrue; 334 } 335 // 'con' is set to true or false to kill the dominated test. 336 Node* con = makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO); 337 // Hack the dominated test 338 _igvn.replace_input_of(iff, 1, con); 339 340 // If I don't have a reachable TRUE and FALSE path following the IfNode then 341 // I can assume this path reaches an infinite loop. In this case it's not 342 // important to optimize the data Nodes - either the whole compilation will 343 // be tossed or this path (and all data Nodes) will go dead. 344 if (iff->outcnt() != 2) { 345 return; 346 } 347 348 // Make control-dependent data Nodes on the live path (path that will remain 349 // once the dominated IF is removed) become control-dependent on the 350 // dominating projection. 351 Node* dp = iff->proj_out_or_null(pop == Op_IfTrue); 352 353 if (dp == nullptr) { 354 return; 355 } 356 357 rewire_safe_outputs_to_dominator(dp, prevdom, pin_array_access_nodes); 358 } 359 360 void PhaseIdealLoop::rewire_safe_outputs_to_dominator(Node* source, Node* dominator, const bool pin_array_access_nodes) { 361 IdealLoopTree* old_loop = get_loop(source); 362 363 for (DUIterator_Fast imax, i = source->fast_outs(imax); i < imax; i++) { 364 Node* out = source->fast_out(i); // Control-dependent node 365 // Do not rewire Div and Mod nodes which could have a zero divisor to avoid skipping their zero check. 366 if (out->depends_only_on_test() && _igvn.no_dependent_zero_check(out)) { 367 assert(out->in(0) == source, "must be control dependent on source"); 368 _igvn.replace_input_of(out, 0, dominator); 369 if (pin_array_access_nodes) { 370 // Because of Loop Predication, Loads and range check Cast nodes that are control dependent on this range 371 // check (that is about to be removed) now depend on multiple dominating Hoisted Check Predicates. After the 372 // removal of this range check, these control dependent nodes end up at the lowest/nearest dominating predicate 373 // in the graph. To ensure that these Loads/Casts do not float above any of the dominating checks (even when the 374 // lowest dominating check is later replaced by yet another dominating check), we need to pin them at the lowest 375 // dominating check. 376 Node* clone = out->pin_array_access_node(); 377 if (clone != nullptr) { 378 clone = _igvn.register_new_node_with_optimizer(clone, out); 379 _igvn.replace_node(out, clone); 380 out = clone; 381 } 382 } 383 set_early_ctrl(out, false); 384 IdealLoopTree* new_loop = get_loop(get_ctrl(out)); 385 if (old_loop != new_loop) { 386 if (!old_loop->_child) { 387 old_loop->_body.yank(out); 388 } 389 if (!new_loop->_child) { 390 new_loop->_body.push(out); 391 } 392 } 393 --i; 394 --imax; 395 } 396 } 397 } 398 399 //------------------------------has_local_phi_input---------------------------- 400 // Return TRUE if 'n' has Phi inputs from its local block and no other 401 // block-local inputs (all non-local-phi inputs come from earlier blocks) 402 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) { 403 Node *n_ctrl = get_ctrl(n); 404 // See if some inputs come from a Phi in this block, or from before 405 // this block. 406 uint i; 407 for( i = 1; i < n->req(); i++ ) { 408 Node *phi = n->in(i); 409 if( phi->is_Phi() && phi->in(0) == n_ctrl ) 410 break; 411 } 412 if( i >= n->req() ) 413 return nullptr; // No Phi inputs; nowhere to clone thru 414 415 // Check for inputs created between 'n' and the Phi input. These 416 // must split as well; they have already been given the chance 417 // (courtesy of a post-order visit) and since they did not we must 418 // recover the 'cost' of splitting them by being very profitable 419 // when splitting 'n'. Since this is unlikely we simply give up. 420 for( i = 1; i < n->req(); i++ ) { 421 Node *m = n->in(i); 422 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) { 423 // We allow the special case of AddP's with no local inputs. 424 // This allows us to split-up address expressions. 425 if (m->is_AddP() && 426 get_ctrl(m->in(AddPNode::Base)) != n_ctrl && 427 get_ctrl(m->in(AddPNode::Address)) != n_ctrl && 428 get_ctrl(m->in(AddPNode::Offset)) != n_ctrl) { 429 // Move the AddP up to the dominating point. That's fine because control of m's inputs 430 // must dominate get_ctrl(m) == n_ctrl and we just checked that the input controls are != n_ctrl. 431 Node* c = find_non_split_ctrl(idom(n_ctrl)); 432 if (c->is_OuterStripMinedLoop()) { 433 c->as_Loop()->verify_strip_mined(1); 434 c = c->in(LoopNode::EntryControl); 435 } 436 set_ctrl_and_loop(m, c); 437 continue; 438 } 439 return nullptr; 440 } 441 assert(n->is_Phi() || m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control"); 442 } 443 444 return n_ctrl; 445 } 446 447 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2). 448 Node* PhaseIdealLoop::remix_address_expressions_add_left_shift(Node* n, IdealLoopTree* n_loop, Node* n_ctrl, BasicType bt) { 449 assert(bt == T_INT || bt == T_LONG, "only for integers"); 450 int n_op = n->Opcode(); 451 452 if (n_op == Op_LShift(bt)) { 453 // Scale is loop invariant 454 Node* scale = n->in(2); 455 Node* scale_ctrl = get_ctrl(scale); 456 IdealLoopTree* scale_loop = get_loop(scale_ctrl); 457 if (n_loop == scale_loop || !scale_loop->is_member(n_loop)) { 458 return nullptr; 459 } 460 const TypeInt* scale_t = scale->bottom_type()->isa_int(); 461 if (scale_t != nullptr && scale_t->is_con() && scale_t->get_con() >= 16) { 462 return nullptr; // Dont bother with byte/short masking 463 } 464 // Add must vary with loop (else shift would be loop-invariant) 465 Node* add = n->in(1); 466 Node* add_ctrl = get_ctrl(add); 467 IdealLoopTree* add_loop = get_loop(add_ctrl); 468 if (n_loop != add_loop) { 469 return nullptr; // happens w/ evil ZKM loops 470 } 471 472 // Convert I-V into I+ (0-V); same for V-I 473 if (add->Opcode() == Op_Sub(bt) && 474 _igvn.type(add->in(1)) != TypeInteger::zero(bt)) { 475 assert(add->Opcode() == Op_SubI || add->Opcode() == Op_SubL, ""); 476 Node* zero = integercon(0, bt); 477 Node* neg = SubNode::make(zero, add->in(2), bt); 478 register_new_node_with_ctrl_of(neg, add->in(2)); 479 add = AddNode::make(add->in(1), neg, bt); 480 register_new_node(add, add_ctrl); 481 } 482 if (add->Opcode() != Op_Add(bt)) return nullptr; 483 assert(add->Opcode() == Op_AddI || add->Opcode() == Op_AddL, ""); 484 // See if one add input is loop invariant 485 Node* add_var = add->in(1); 486 Node* add_var_ctrl = get_ctrl(add_var); 487 IdealLoopTree* add_var_loop = get_loop(add_var_ctrl); 488 Node* add_invar = add->in(2); 489 Node* add_invar_ctrl = get_ctrl(add_invar); 490 IdealLoopTree* add_invar_loop = get_loop(add_invar_ctrl); 491 if (add_invar_loop == n_loop) { 492 // Swap to find the invariant part 493 add_invar = add_var; 494 add_invar_ctrl = add_var_ctrl; 495 add_invar_loop = add_var_loop; 496 add_var = add->in(2); 497 } else if (add_var_loop != n_loop) { // Else neither input is loop invariant 498 return nullptr; 499 } 500 if (n_loop == add_invar_loop || !add_invar_loop->is_member(n_loop)) { 501 return nullptr; // No invariant part of the add? 502 } 503 504 // Yes! Reshape address expression! 505 Node* inv_scale = LShiftNode::make(add_invar, scale, bt); 506 Node* inv_scale_ctrl = 507 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ? 508 add_invar_ctrl : scale_ctrl; 509 register_new_node(inv_scale, inv_scale_ctrl); 510 Node* var_scale = LShiftNode::make(add_var, scale, bt); 511 register_new_node(var_scale, n_ctrl); 512 Node* var_add = AddNode::make(var_scale, inv_scale, bt); 513 register_new_node(var_add, n_ctrl); 514 _igvn.replace_node(n, var_add); 515 return var_add; 516 } 517 return nullptr; 518 } 519 520 //------------------------------remix_address_expressions---------------------- 521 // Rework addressing expressions to get the most loop-invariant stuff 522 // moved out. We'd like to do all associative operators, but it's especially 523 // important (common) to do address expressions. 524 Node* PhaseIdealLoop::remix_address_expressions(Node* n) { 525 if (!has_ctrl(n)) return nullptr; 526 Node* n_ctrl = get_ctrl(n); 527 IdealLoopTree* n_loop = get_loop(n_ctrl); 528 529 // See if 'n' mixes loop-varying and loop-invariant inputs and 530 // itself is loop-varying. 531 532 // Only interested in binary ops (and AddP) 533 if (n->req() < 3 || n->req() > 4) return nullptr; 534 535 Node* n1_ctrl = get_ctrl(n->in( 1)); 536 Node* n2_ctrl = get_ctrl(n->in( 2)); 537 Node* n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3)); 538 IdealLoopTree* n1_loop = get_loop(n1_ctrl); 539 IdealLoopTree* n2_loop = get_loop(n2_ctrl); 540 IdealLoopTree* n3_loop = get_loop(n3_ctrl); 541 542 // Does one of my inputs spin in a tighter loop than self? 543 if ((n_loop->is_member(n1_loop) && n_loop != n1_loop) || 544 (n_loop->is_member(n2_loop) && n_loop != n2_loop) || 545 (n_loop->is_member(n3_loop) && n_loop != n3_loop)) { 546 return nullptr; // Leave well enough alone 547 } 548 549 // Is at least one of my inputs loop-invariant? 550 if (n1_loop == n_loop && 551 n2_loop == n_loop && 552 n3_loop == n_loop) { 553 return nullptr; // No loop-invariant inputs 554 } 555 556 Node* res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_INT); 557 if (res != nullptr) { 558 return res; 559 } 560 res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_LONG); 561 if (res != nullptr) { 562 return res; 563 } 564 565 int n_op = n->Opcode(); 566 // Replace (I+V) with (V+I) 567 if (n_op == Op_AddI || 568 n_op == Op_AddL || 569 n_op == Op_AddF || 570 n_op == Op_AddD || 571 n_op == Op_MulI || 572 n_op == Op_MulL || 573 n_op == Op_MulF || 574 n_op == Op_MulD) { 575 if (n2_loop == n_loop) { 576 assert(n1_loop != n_loop, ""); 577 n->swap_edges(1, 2); 578 } 579 } 580 581 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V), 582 // but not if I2 is a constant. Skip for irreducible loops. 583 if (n_op == Op_AddP && n_loop->_head->is_Loop()) { 584 if (n2_loop == n_loop && n3_loop != n_loop) { 585 if (n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con()) { 586 Node* n22_ctrl = get_ctrl(n->in(2)->in(2)); 587 Node* n23_ctrl = get_ctrl(n->in(2)->in(3)); 588 IdealLoopTree* n22loop = get_loop(n22_ctrl); 589 IdealLoopTree* n23_loop = get_loop(n23_ctrl); 590 if (n22loop != n_loop && n22loop->is_member(n_loop) && 591 n23_loop == n_loop) { 592 Node* add1 = new AddPNode(n->in(1), n->in(2)->in(2), n->in(3)); 593 // Stuff new AddP in the loop preheader 594 register_new_node(add1, n_loop->_head->as_Loop()->skip_strip_mined(1)->in(LoopNode::EntryControl)); 595 Node* add2 = new AddPNode(n->in(1), add1, n->in(2)->in(3)); 596 register_new_node(add2, n_ctrl); 597 _igvn.replace_node(n, add2); 598 return add2; 599 } 600 } 601 } 602 603 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V) 604 if (n2_loop != n_loop && n3_loop == n_loop) { 605 if (n->in(3)->Opcode() == Op_AddX) { 606 Node* V = n->in(3)->in(1); 607 Node* I = n->in(3)->in(2); 608 if (is_member(n_loop,get_ctrl(V))) { 609 } else { 610 Node *tmp = V; V = I; I = tmp; 611 } 612 if (!is_member(n_loop,get_ctrl(I))) { 613 Node* add1 = new AddPNode(n->in(1), n->in(2), I); 614 // Stuff new AddP in the loop preheader 615 register_new_node(add1, n_loop->_head->as_Loop()->skip_strip_mined(1)->in(LoopNode::EntryControl)); 616 Node* add2 = new AddPNode(n->in(1), add1, V); 617 register_new_node(add2, n_ctrl); 618 _igvn.replace_node(n, add2); 619 return add2; 620 } 621 } 622 } 623 } 624 625 return nullptr; 626 } 627 628 // Optimize ((in1[2*i] * in2[2*i]) + (in1[2*i+1] * in2[2*i+1])) 629 Node *PhaseIdealLoop::convert_add_to_muladd(Node* n) { 630 assert(n->Opcode() == Op_AddI, "sanity"); 631 Node * nn = nullptr; 632 Node * in1 = n->in(1); 633 Node * in2 = n->in(2); 634 if (in1->Opcode() == Op_MulI && in2->Opcode() == Op_MulI) { 635 IdealLoopTree* loop_n = get_loop(get_ctrl(n)); 636 if (loop_n->is_counted() && 637 loop_n->_head->as_Loop()->is_valid_counted_loop(T_INT) && 638 Matcher::match_rule_supported(Op_MulAddVS2VI) && 639 Matcher::match_rule_supported(Op_MulAddS2I)) { 640 Node* mul_in1 = in1->in(1); 641 Node* mul_in2 = in1->in(2); 642 Node* mul_in3 = in2->in(1); 643 Node* mul_in4 = in2->in(2); 644 if (mul_in1->Opcode() == Op_LoadS && 645 mul_in2->Opcode() == Op_LoadS && 646 mul_in3->Opcode() == Op_LoadS && 647 mul_in4->Opcode() == Op_LoadS) { 648 IdealLoopTree* loop1 = get_loop(get_ctrl(mul_in1)); 649 IdealLoopTree* loop2 = get_loop(get_ctrl(mul_in2)); 650 IdealLoopTree* loop3 = get_loop(get_ctrl(mul_in3)); 651 IdealLoopTree* loop4 = get_loop(get_ctrl(mul_in4)); 652 IdealLoopTree* loop5 = get_loop(get_ctrl(in1)); 653 IdealLoopTree* loop6 = get_loop(get_ctrl(in2)); 654 // All nodes should be in the same counted loop. 655 if (loop_n == loop1 && loop_n == loop2 && loop_n == loop3 && 656 loop_n == loop4 && loop_n == loop5 && loop_n == loop6) { 657 Node* adr1 = mul_in1->in(MemNode::Address); 658 Node* adr2 = mul_in2->in(MemNode::Address); 659 Node* adr3 = mul_in3->in(MemNode::Address); 660 Node* adr4 = mul_in4->in(MemNode::Address); 661 if (adr1->is_AddP() && adr2->is_AddP() && adr3->is_AddP() && adr4->is_AddP()) { 662 if ((adr1->in(AddPNode::Base) == adr3->in(AddPNode::Base)) && 663 (adr2->in(AddPNode::Base) == adr4->in(AddPNode::Base))) { 664 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in3, mul_in4); 665 register_new_node_with_ctrl_of(nn, n); 666 _igvn.replace_node(n, nn); 667 return nn; 668 } else if ((adr1->in(AddPNode::Base) == adr4->in(AddPNode::Base)) && 669 (adr2->in(AddPNode::Base) == adr3->in(AddPNode::Base))) { 670 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in4, mul_in3); 671 register_new_node_with_ctrl_of(nn, n); 672 _igvn.replace_node(n, nn); 673 return nn; 674 } 675 } 676 } 677 } 678 } 679 } 680 return nn; 681 } 682 683 //------------------------------conditional_move------------------------------- 684 // Attempt to replace a Phi with a conditional move. We have some pretty 685 // strict profitability requirements. All Phis at the merge point must 686 // be converted, so we can remove the control flow. We need to limit the 687 // number of c-moves to a small handful. All code that was in the side-arms 688 // of the CFG diamond is now speculatively executed. This code has to be 689 // "cheap enough". We are pretty much limited to CFG diamonds that merge 690 // 1 or 2 items with a total of 1 or 2 ops executed speculatively. 691 Node *PhaseIdealLoop::conditional_move( Node *region ) { 692 693 assert(region->is_Region(), "sanity check"); 694 if (region->req() != 3) return nullptr; 695 696 // Check for CFG diamond 697 Node *lp = region->in(1); 698 Node *rp = region->in(2); 699 if (!lp || !rp) return nullptr; 700 Node *lp_c = lp->in(0); 701 if (lp_c == nullptr || lp_c != rp->in(0) || !lp_c->is_If()) return nullptr; 702 IfNode *iff = lp_c->as_If(); 703 704 // Check for ops pinned in an arm of the diamond. 705 // Can't remove the control flow in this case 706 if (lp->outcnt() > 1) return nullptr; 707 if (rp->outcnt() > 1) return nullptr; 708 709 IdealLoopTree* r_loop = get_loop(region); 710 assert(r_loop == get_loop(iff), "sanity"); 711 // Always convert to CMOVE if all results are used only outside this loop. 712 bool used_inside_loop = (r_loop == _ltree_root); 713 714 // Check profitability 715 int cost = 0; 716 int phis = 0; 717 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 718 Node *out = region->fast_out(i); 719 if (!out->is_Phi()) continue; // Ignore other control edges, etc 720 phis++; 721 PhiNode* phi = out->as_Phi(); 722 BasicType bt = phi->type()->basic_type(); 723 switch (bt) { 724 case T_DOUBLE: 725 case T_FLOAT: 726 if (C->use_cmove()) { 727 continue; //TODO: maybe we want to add some cost 728 } 729 cost += Matcher::float_cmove_cost(); // Could be very expensive 730 break; 731 case T_LONG: { 732 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's 733 } 734 case T_INT: // These all CMOV fine 735 case T_ADDRESS: { // (RawPtr) 736 cost++; 737 break; 738 } 739 case T_NARROWOOP: // Fall through 740 case T_OBJECT: { // Base oops are OK, but not derived oops 741 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr(); 742 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a 743 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus 744 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we 745 // have a Phi for the base here that we convert to a CMOVE all is well 746 // and good. But if the base is dead, we'll not make a CMOVE. Later 747 // the allocator will have to produce a base by creating a CMOVE of the 748 // relevant bases. This puts the allocator in the business of 749 // manufacturing expensive instructions, generally a bad plan. 750 // Just Say No to Conditionally-Moved Derived Pointers. 751 if (tp && tp->offset() != 0) 752 return nullptr; 753 cost++; 754 break; 755 } 756 default: 757 return nullptr; // In particular, can't do memory or I/O 758 } 759 // Add in cost any speculative ops 760 for (uint j = 1; j < region->req(); j++) { 761 Node *proj = region->in(j); 762 Node *inp = phi->in(j); 763 if (get_ctrl(inp) == proj) { // Found local op 764 cost++; 765 // Check for a chain of dependent ops; these will all become 766 // speculative in a CMOV. 767 for (uint k = 1; k < inp->req(); k++) 768 if (get_ctrl(inp->in(k)) == proj) 769 cost += ConditionalMoveLimit; // Too much speculative goo 770 } 771 } 772 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode. 773 // This will likely Split-If, a higher-payoff operation. 774 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) { 775 Node* use = phi->fast_out(k); 776 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr()) 777 cost += ConditionalMoveLimit; 778 // Is there a use inside the loop? 779 // Note: check only basic types since CMoveP is pinned. 780 if (!used_inside_loop && is_java_primitive(bt)) { 781 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use); 782 if (r_loop == u_loop || r_loop->is_member(u_loop)) { 783 used_inside_loop = true; 784 } 785 } 786 } 787 }//for 788 Node* bol = iff->in(1); 789 assert(!bol->is_OpaqueInitializedAssertionPredicate(), "Initialized Assertion Predicates cannot form a diamond with Halt"); 790 if (bol->is_OpaqueTemplateAssertionPredicate()) { 791 // Ignore Template Assertion Predicates with OpaqueTemplateAssertionPredicate nodes. 792 return nullptr; 793 } 794 if (bol->is_OpaqueMultiversioning()) { 795 assert(bol->as_OpaqueMultiversioning()->is_useless(), "Must be useless, i.e. fast main loop has already disappeared."); 796 // Ignore multiversion_if that just lost its loops. The OpaqueMultiversioning is marked useless, 797 // and will make the multiversion_if constant fold in the next IGVN round. 798 return nullptr; 799 } 800 if (!bol->is_Bool()) { 801 assert(false, "Expected Bool, but got %s", NodeClassNames[bol->Opcode()]); 802 return nullptr; 803 } 804 int cmp_op = bol->in(1)->Opcode(); 805 if (cmp_op == Op_SubTypeCheck) { // SubTypeCheck expansion expects an IfNode 806 return nullptr; 807 } 808 // It is expensive to generate flags from a float compare. 809 // Avoid duplicated float compare. 810 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return nullptr; 811 812 float infrequent_prob = PROB_UNLIKELY_MAG(3); 813 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop. 814 if (used_inside_loop) { 815 if (cost >= ConditionalMoveLimit) return nullptr; // Too much goo 816 817 // BlockLayoutByFrequency optimization moves infrequent branch 818 // from hot path. No point in CMOV'ing in such case (110 is used 819 // instead of 100 to take into account not exactness of float value). 820 if (BlockLayoutByFrequency) { 821 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f); 822 } 823 } 824 // Check for highly predictable branch. No point in CMOV'ing if 825 // we are going to predict accurately all the time. 826 if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) { 827 //keep going 828 } else if (iff->_prob < infrequent_prob || 829 iff->_prob > (1.0f - infrequent_prob)) 830 return nullptr; 831 832 // -------------- 833 // Now replace all Phis with CMOV's 834 Node *cmov_ctrl = iff->in(0); 835 uint flip = (lp->Opcode() == Op_IfTrue); 836 Node_List wq; 837 while (1) { 838 PhiNode* phi = nullptr; 839 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 840 Node *out = region->fast_out(i); 841 if (out->is_Phi()) { 842 phi = out->as_Phi(); 843 break; 844 } 845 } 846 if (phi == nullptr || _igvn.type(phi) == Type::TOP || !CMoveNode::supported(_igvn.type(phi))) { 847 break; 848 } 849 // Move speculative ops 850 wq.push(phi); 851 while (wq.size() > 0) { 852 Node *n = wq.pop(); 853 for (uint j = 1; j < n->req(); j++) { 854 Node* m = n->in(j); 855 if (m != nullptr && !is_dominator(get_ctrl(m), cmov_ctrl)) { 856 set_ctrl(m, cmov_ctrl); 857 wq.push(m); 858 } 859 } 860 } 861 Node* cmov = CMoveNode::make(iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi)); 862 register_new_node(cmov, cmov_ctrl); 863 _igvn.replace_node(phi, cmov); 864 #ifndef PRODUCT 865 if (TraceLoopOpts) { 866 tty->print("CMOV "); 867 r_loop->dump_head(); 868 if (Verbose) { 869 bol->in(1)->dump(1); 870 cmov->dump(1); 871 } 872 } 873 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } ); 874 #endif 875 } 876 877 // The useless CFG diamond will fold up later; see the optimization in 878 // RegionNode::Ideal. 879 _igvn._worklist.push(region); 880 881 return iff->in(1); 882 } 883 884 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) { 885 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 886 Node* u = m->fast_out(i); 887 if (u->is_CFG()) { 888 if (u->is_NeverBranch()) { 889 u = u->as_NeverBranch()->proj_out(0); 890 enqueue_cfg_uses(u, wq); 891 } else { 892 wq.push(u); 893 } 894 } 895 } 896 } 897 898 // Try moving a store out of a loop, right before the loop 899 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) { 900 // Store has to be first in the loop body 901 IdealLoopTree *n_loop = get_loop(n_ctrl); 902 if (n->is_Store() && n_loop != _ltree_root && 903 n_loop->is_loop() && n_loop->_head->is_Loop() && 904 n->in(0) != nullptr) { 905 Node* address = n->in(MemNode::Address); 906 Node* value = n->in(MemNode::ValueIn); 907 Node* mem = n->in(MemNode::Memory); 908 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 909 IdealLoopTree* value_loop = get_loop(get_ctrl(value)); 910 911 // - address and value must be loop invariant 912 // - memory must be a memory Phi for the loop 913 // - Store must be the only store on this memory slice in the 914 // loop: if there's another store following this one then value 915 // written at iteration i by the second store could be overwritten 916 // at iteration i+n by the first store: it's not safe to move the 917 // first store out of the loop 918 // - nothing must observe the memory Phi: it guarantees no read 919 // before the store, we are also guaranteed the store post 920 // dominates the loop head (ignoring a possible early 921 // exit). Otherwise there would be extra Phi involved between the 922 // loop's Phi and the store. 923 // - there must be no early exit from the loop before the Store 924 // (such an exit most of the time would be an extra use of the 925 // memory Phi but sometimes is a bottom memory Phi that takes the 926 // store as input). 927 928 if (!n_loop->is_member(address_loop) && 929 !n_loop->is_member(value_loop) && 930 mem->is_Phi() && mem->in(0) == n_loop->_head && 931 mem->outcnt() == 1 && 932 mem->in(LoopNode::LoopBackControl) == n) { 933 934 assert(n_loop->_tail != nullptr, "need a tail"); 935 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop"); 936 937 // Verify that there's no early exit of the loop before the store. 938 bool ctrl_ok = false; 939 { 940 // Follow control from loop head until n, we exit the loop or 941 // we reach the tail 942 ResourceMark rm; 943 Unique_Node_List wq; 944 wq.push(n_loop->_head); 945 946 for (uint next = 0; next < wq.size(); ++next) { 947 Node *m = wq.at(next); 948 if (m == n->in(0)) { 949 ctrl_ok = true; 950 continue; 951 } 952 assert(!has_ctrl(m), "should be CFG"); 953 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) { 954 ctrl_ok = false; 955 break; 956 } 957 enqueue_cfg_uses(m, wq); 958 if (wq.size() > 10) { 959 ctrl_ok = false; 960 break; 961 } 962 } 963 } 964 if (ctrl_ok) { 965 // move the Store 966 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem); 967 _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl)); 968 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl)); 969 // Disconnect the phi now. An empty phi can confuse other 970 // optimizations in this pass of loop opts. 971 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl)); 972 n_loop->_body.yank(mem); 973 974 set_ctrl_and_loop(n, n->in(0)); 975 976 return n; 977 } 978 } 979 } 980 return nullptr; 981 } 982 983 // Try moving a store out of a loop, right after the loop 984 void PhaseIdealLoop::try_move_store_after_loop(Node* n) { 985 if (n->is_Store() && n->in(0) != nullptr) { 986 Node *n_ctrl = get_ctrl(n); 987 IdealLoopTree *n_loop = get_loop(n_ctrl); 988 // Store must be in a loop 989 if (n_loop != _ltree_root && !n_loop->_irreducible) { 990 Node* address = n->in(MemNode::Address); 991 Node* value = n->in(MemNode::ValueIn); 992 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 993 // address must be loop invariant 994 if (!n_loop->is_member(address_loop)) { 995 // Store must be last on this memory slice in the loop and 996 // nothing in the loop must observe it 997 Node* phi = nullptr; 998 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 999 Node* u = n->fast_out(i); 1000 if (has_ctrl(u)) { // control use? 1001 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 1002 if (!n_loop->is_member(u_loop)) { 1003 continue; 1004 } 1005 if (u->is_Phi() && u->in(0) == n_loop->_head) { 1006 assert(_igvn.type(u) == Type::MEMORY, "bad phi"); 1007 // multiple phis on the same slice are possible 1008 if (phi != nullptr) { 1009 return; 1010 } 1011 phi = u; 1012 continue; 1013 } 1014 } 1015 return; 1016 } 1017 if (phi != nullptr) { 1018 // Nothing in the loop before the store (next iteration) 1019 // must observe the stored value 1020 bool mem_ok = true; 1021 { 1022 ResourceMark rm; 1023 Unique_Node_List wq; 1024 wq.push(phi); 1025 for (uint next = 0; next < wq.size() && mem_ok; ++next) { 1026 Node *m = wq.at(next); 1027 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) { 1028 Node* u = m->fast_out(i); 1029 if (u->is_Store() || u->is_Phi()) { 1030 if (u != n) { 1031 wq.push(u); 1032 mem_ok = (wq.size() <= 10); 1033 } 1034 } else { 1035 mem_ok = false; 1036 break; 1037 } 1038 } 1039 } 1040 } 1041 if (mem_ok) { 1042 // Move the store out of the loop if the LCA of all 1043 // users (except for the phi) is outside the loop. 1044 Node* hook = new Node(1); 1045 hook->init_req(0, n_ctrl); // Add an input to prevent hook from being dead 1046 _igvn.rehash_node_delayed(phi); 1047 int count = phi->replace_edge(n, hook, &_igvn); 1048 assert(count > 0, "inconsistent phi"); 1049 1050 // Compute latest point this store can go 1051 Node* lca = get_late_ctrl(n, get_ctrl(n)); 1052 if (lca->is_OuterStripMinedLoop()) { 1053 lca = lca->in(LoopNode::EntryControl); 1054 } 1055 if (n_loop->is_member(get_loop(lca))) { 1056 // LCA is in the loop - bail out 1057 _igvn.replace_node(hook, n); 1058 return; 1059 } 1060 #ifdef ASSERT 1061 if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) { 1062 assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined"); 1063 n_loop->_head->as_Loop()->verify_strip_mined(1); 1064 Node* outer = n_loop->_head->as_CountedLoop()->outer_loop(); 1065 IdealLoopTree* outer_loop = get_loop(outer); 1066 assert(n_loop->_parent == outer_loop, "broken loop tree"); 1067 assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state"); 1068 } 1069 #endif 1070 lca = place_outside_loop(lca, n_loop); 1071 assert(!n_loop->is_member(get_loop(lca)), "control must not be back in the loop"); 1072 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"); 1073 1074 // Move store out of the loop 1075 _igvn.replace_node(hook, n->in(MemNode::Memory)); 1076 _igvn.replace_input_of(n, 0, lca); 1077 set_ctrl_and_loop(n, lca); 1078 1079 // Disconnect the phi now. An empty phi can confuse other 1080 // optimizations in this pass of loop opts.. 1081 if (phi->in(LoopNode::LoopBackControl) == phi) { 1082 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl)); 1083 n_loop->_body.yank(phi); 1084 } 1085 } 1086 } 1087 } 1088 } 1089 } 1090 } 1091 1092 //------------------------------split_if_with_blocks_pre----------------------- 1093 // Do the real work in a non-recursive function. Data nodes want to be 1094 // cloned in the pre-order so they can feed each other nicely. 1095 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) { 1096 // Cloning these guys is unlikely to win 1097 int n_op = n->Opcode(); 1098 if (n_op == Op_MergeMem) { 1099 return n; 1100 } 1101 if (n->is_Proj()) { 1102 return n; 1103 } 1104 // Do not clone-up CmpFXXX variations, as these are always 1105 // followed by a CmpI 1106 if (n->is_Cmp()) { 1107 return n; 1108 } 1109 // Attempt to use a conditional move instead of a phi/branch 1110 if (ConditionalMoveLimit > 0 && n_op == Op_Region) { 1111 Node *cmov = conditional_move( n ); 1112 if (cmov) { 1113 return cmov; 1114 } 1115 } 1116 if (n->is_CFG() || n->is_LoadStore()) { 1117 return n; 1118 } 1119 if (n->is_Opaque1()) { // Opaque nodes cannot be mod'd 1120 if (!C->major_progress()) { // If chance of no more loop opts... 1121 _igvn._worklist.push(n); // maybe we'll remove them 1122 } 1123 return n; 1124 } 1125 1126 if (n->is_Con()) { 1127 return n; // No cloning for Con nodes 1128 } 1129 1130 Node *n_ctrl = get_ctrl(n); 1131 if (!n_ctrl) { 1132 return n; // Dead node 1133 } 1134 1135 Node* res = try_move_store_before_loop(n, n_ctrl); 1136 if (res != nullptr) { 1137 return n; 1138 } 1139 1140 // Attempt to remix address expressions for loop invariants 1141 Node *m = remix_address_expressions( n ); 1142 if( m ) return m; 1143 1144 if (n_op == Op_AddI) { 1145 Node *nn = convert_add_to_muladd( n ); 1146 if ( nn ) return nn; 1147 } 1148 1149 if (n->is_ConstraintCast()) { 1150 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this); 1151 // ConstraintCastNode::dominating_cast() uses node control input to determine domination. 1152 // Node control inputs don't necessarily agree with loop control info (due to 1153 // transformations happened in between), thus additional dominance check is needed 1154 // to keep loop info valid. 1155 if (dom_cast != nullptr && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) { 1156 _igvn.replace_node(n, dom_cast); 1157 return dom_cast; 1158 } 1159 } 1160 1161 // Determine if the Node has inputs from some local Phi. 1162 // Returns the block to clone thru. 1163 Node *n_blk = has_local_phi_input( n ); 1164 if( !n_blk ) return n; 1165 1166 // Do not clone the trip counter through on a CountedLoop 1167 // (messes up the canonical shape). 1168 if (((n_blk->is_CountedLoop() || (n_blk->is_Loop() && n_blk->as_Loop()->is_loop_nest_inner_loop())) && n->Opcode() == Op_AddI) || 1169 (n_blk->is_LongCountedLoop() && n->Opcode() == Op_AddL)) { 1170 return n; 1171 } 1172 // Pushing a shift through the iv Phi can get in the way of addressing optimizations or range check elimination 1173 if (n_blk->is_BaseCountedLoop() && n->Opcode() == Op_LShift(n_blk->as_BaseCountedLoop()->bt()) && 1174 n->in(1) == n_blk->as_BaseCountedLoop()->phi()) { 1175 return n; 1176 } 1177 1178 // Check for having no control input; not pinned. Allow 1179 // dominating control. 1180 if (n->in(0)) { 1181 Node *dom = idom(n_blk); 1182 if (dom_lca(n->in(0), dom) != n->in(0)) { 1183 return n; 1184 } 1185 } 1186 // Policy: when is it profitable. You must get more wins than 1187 // policy before it is considered profitable. Policy is usually 0, 1188 // so 1 win is considered profitable. Big merges will require big 1189 // cloning, so get a larger policy. 1190 int policy = n_blk->req() >> 2; 1191 1192 // If the loop is a candidate for range check elimination, 1193 // delay splitting through it's phi until a later loop optimization 1194 if (n_blk->is_BaseCountedLoop()) { 1195 IdealLoopTree *lp = get_loop(n_blk); 1196 if (lp && lp->_rce_candidate) { 1197 return n; 1198 } 1199 } 1200 1201 if (must_throttle_split_if()) return n; 1202 1203 // Split 'n' through the merge point if it is profitable 1204 Node *phi = split_thru_phi( n, n_blk, policy ); 1205 if (!phi) return n; 1206 1207 // Found a Phi to split thru! 1208 // Replace 'n' with the new phi 1209 _igvn.replace_node( n, phi ); 1210 // Moved a load around the loop, 'en-registering' something. 1211 if (n_blk->is_Loop() && n->is_Load() && 1212 !phi->in(LoopNode::LoopBackControl)->is_Load()) 1213 C->set_major_progress(); 1214 1215 return phi; 1216 } 1217 1218 static bool merge_point_too_heavy(Compile* C, Node* region) { 1219 // Bail out if the region and its phis have too many users. 1220 int weight = 0; 1221 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1222 weight += region->fast_out(i)->outcnt(); 1223 } 1224 int nodes_left = C->max_node_limit() - C->live_nodes(); 1225 if (weight * 8 > nodes_left) { 1226 if (PrintOpto) { 1227 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight); 1228 } 1229 return true; 1230 } else { 1231 return false; 1232 } 1233 } 1234 1235 static bool merge_point_safe(Node* region) { 1236 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode 1237 // having a PhiNode input. This sidesteps the dangerous case where the split 1238 // ConvI2LNode may become TOP if the input Value() does not 1239 // overlap the ConvI2L range, leaving a node which may not dominate its 1240 // uses. 1241 // A better fix for this problem can be found in the BugTraq entry, but 1242 // expediency for Mantis demands this hack. 1243 #ifdef _LP64 1244 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1245 Node* n = region->fast_out(i); 1246 if (n->is_Phi()) { 1247 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1248 Node* m = n->fast_out(j); 1249 if (m->Opcode() == Op_ConvI2L) 1250 return false; 1251 if (m->is_CastII()) { 1252 return false; 1253 } 1254 } 1255 } 1256 } 1257 #endif 1258 return true; 1259 } 1260 1261 1262 //------------------------------place_outside_loop--------------------------------- 1263 // Place some computation outside of this loop on the path to the use passed as argument 1264 Node* PhaseIdealLoop::place_outside_loop(Node* useblock, IdealLoopTree* loop) const { 1265 Node* head = loop->_head; 1266 assert(!loop->is_member(get_loop(useblock)), "must be outside loop"); 1267 if (head->is_Loop() && head->as_Loop()->is_strip_mined()) { 1268 loop = loop->_parent; 1269 assert(loop->_head->is_OuterStripMinedLoop(), "malformed strip mined loop"); 1270 } 1271 1272 // Pick control right outside the loop 1273 for (;;) { 1274 Node* dom = idom(useblock); 1275 if (loop->is_member(get_loop(dom))) { 1276 break; 1277 } 1278 useblock = dom; 1279 } 1280 assert(find_non_split_ctrl(useblock) == useblock, "should be non split control"); 1281 return useblock; 1282 } 1283 1284 1285 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) { 1286 if (!n->is_If() || n->is_BaseCountedLoopEnd()) { 1287 return false; 1288 } 1289 if (!n->in(0)->is_Region()) { 1290 return false; 1291 } 1292 1293 Node* region = n->in(0); 1294 Node* dom = idom(region); 1295 if (!dom->is_If() || !n->as_If()->same_condition(dom, &_igvn)) { 1296 return false; 1297 } 1298 IfNode* dom_if = dom->as_If(); 1299 Node* proj_true = dom_if->proj_out(1); 1300 Node* proj_false = dom_if->proj_out(0); 1301 1302 for (uint i = 1; i < region->req(); i++) { 1303 if (is_dominator(proj_true, region->in(i))) { 1304 continue; 1305 } 1306 if (is_dominator(proj_false, region->in(i))) { 1307 continue; 1308 } 1309 return false; 1310 } 1311 1312 return true; 1313 } 1314 1315 1316 bool PhaseIdealLoop::can_split_if(Node* n_ctrl) { 1317 if (must_throttle_split_if()) { 1318 return false; 1319 } 1320 1321 // Do not do 'split-if' if irreducible loops are present. 1322 if (_has_irreducible_loops) { 1323 return false; 1324 } 1325 1326 if (merge_point_too_heavy(C, n_ctrl)) { 1327 return false; 1328 } 1329 1330 // Do not do 'split-if' if some paths are dead. First do dead code 1331 // elimination and then see if its still profitable. 1332 for (uint i = 1; i < n_ctrl->req(); i++) { 1333 if (n_ctrl->in(i) == C->top()) { 1334 return false; 1335 } 1336 } 1337 1338 // If trying to do a 'Split-If' at the loop head, it is only 1339 // profitable if the cmp folds up on BOTH paths. Otherwise we 1340 // risk peeling a loop forever. 1341 1342 // CNC - Disabled for now. Requires careful handling of loop 1343 // body selection for the cloned code. Also, make sure we check 1344 // for any input path not being in the same loop as n_ctrl. For 1345 // irreducible loops we cannot check for 'n_ctrl->is_Loop()' 1346 // because the alternative loop entry points won't be converted 1347 // into LoopNodes. 1348 IdealLoopTree *n_loop = get_loop(n_ctrl); 1349 for (uint j = 1; j < n_ctrl->req(); j++) { 1350 if (get_loop(n_ctrl->in(j)) != n_loop) { 1351 return false; 1352 } 1353 } 1354 1355 // Check for safety of the merge point. 1356 if (!merge_point_safe(n_ctrl)) { 1357 return false; 1358 } 1359 1360 return true; 1361 } 1362 1363 // Detect if the node is the inner strip-mined loop 1364 // Return: null if it's not the case, or the exit of outer strip-mined loop 1365 static Node* is_inner_of_stripmined_loop(const Node* out) { 1366 Node* out_le = nullptr; 1367 1368 if (out->is_CountedLoopEnd()) { 1369 const CountedLoopNode* loop = out->as_CountedLoopEnd()->loopnode(); 1370 1371 if (loop != nullptr && loop->is_strip_mined()) { 1372 out_le = loop->in(LoopNode::EntryControl)->as_OuterStripMinedLoop()->outer_loop_exit(); 1373 } 1374 } 1375 1376 return out_le; 1377 } 1378 1379 //------------------------------split_if_with_blocks_post---------------------- 1380 // Do the real work in a non-recursive function. CFG hackery wants to be 1381 // in the post-order, so it can dirty the I-DOM info and not use the dirtied 1382 // info. 1383 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) { 1384 1385 // Cloning Cmp through Phi's involves the split-if transform. 1386 // FastLock is not used by an If 1387 if (n->is_Cmp() && !n->is_FastLock()) { 1388 Node *n_ctrl = get_ctrl(n); 1389 // Determine if the Node has inputs from some local Phi. 1390 // Returns the block to clone thru. 1391 Node *n_blk = has_local_phi_input(n); 1392 if (n_blk != n_ctrl) { 1393 return; 1394 } 1395 1396 if (!can_split_if(n_ctrl)) { 1397 return; 1398 } 1399 1400 if (n->outcnt() != 1) { 1401 return; // Multiple bool's from 1 compare? 1402 } 1403 Node *bol = n->unique_out(); 1404 assert(bol->is_Bool(), "expect a bool here"); 1405 if (bol->outcnt() != 1) { 1406 return;// Multiple branches from 1 compare? 1407 } 1408 Node *iff = bol->unique_out(); 1409 1410 // Check some safety conditions 1411 if (iff->is_If()) { // Classic split-if? 1412 if (iff->in(0) != n_ctrl) { 1413 return; // Compare must be in same blk as if 1414 } 1415 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE 1416 // Can't split CMove with different control. 1417 if (get_ctrl(iff) != n_ctrl) { 1418 return; 1419 } 1420 if (get_ctrl(iff->in(2)) == n_ctrl || 1421 get_ctrl(iff->in(3)) == n_ctrl) { 1422 return; // Inputs not yet split-up 1423 } 1424 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) { 1425 return; // Loop-invar test gates loop-varying CMOVE 1426 } 1427 } else { 1428 return; // some other kind of node, such as an Allocate 1429 } 1430 1431 // When is split-if profitable? Every 'win' on means some control flow 1432 // goes dead, so it's almost always a win. 1433 int policy = 0; 1434 // Split compare 'n' through the merge point if it is profitable 1435 Node *phi = split_thru_phi( n, n_ctrl, policy); 1436 if (!phi) { 1437 return; 1438 } 1439 1440 // Found a Phi to split thru! 1441 // Replace 'n' with the new phi 1442 _igvn.replace_node(n, phi); 1443 1444 // Now split the bool up thru the phi 1445 Node *bolphi = split_thru_phi(bol, n_ctrl, -1); 1446 guarantee(bolphi != nullptr, "null boolean phi node"); 1447 1448 _igvn.replace_node(bol, bolphi); 1449 assert(iff->in(1) == bolphi, ""); 1450 1451 if (bolphi->Value(&_igvn)->singleton()) { 1452 return; 1453 } 1454 1455 // Conditional-move? Must split up now 1456 if (!iff->is_If()) { 1457 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1); 1458 _igvn.replace_node(iff, cmovphi); 1459 return; 1460 } 1461 1462 // Now split the IF 1463 C->print_method(PHASE_BEFORE_SPLIT_IF, 4, iff); 1464 if (TraceLoopOpts) { 1465 tty->print_cr("Split-If"); 1466 } 1467 do_split_if(iff); 1468 C->print_method(PHASE_AFTER_SPLIT_IF, 4, iff); 1469 return; 1470 } 1471 1472 // Two identical ifs back to back can be merged 1473 if (try_merge_identical_ifs(n)) { 1474 return; 1475 } 1476 1477 // Check for an IF ready to split; one that has its 1478 // condition codes input coming from a Phi at the block start. 1479 int n_op = n->Opcode(); 1480 1481 // Check for an IF being dominated by another IF same test 1482 if (n_op == Op_If || 1483 n_op == Op_RangeCheck) { 1484 Node *bol = n->in(1); 1485 uint max = bol->outcnt(); 1486 // Check for same test used more than once? 1487 if (bol->is_Bool() && (max > 1 || bol->in(1)->is_SubTypeCheck())) { 1488 // Search up IDOMs to see if this IF is dominated. 1489 Node* cmp = bol->in(1); 1490 Node *cutoff = cmp->is_SubTypeCheck() ? dom_lca(get_ctrl(cmp->in(1)), get_ctrl(cmp->in(2))) : get_ctrl(bol); 1491 1492 // Now search up IDOMs till cutoff, looking for a dominating test 1493 Node *prevdom = n; 1494 Node *dom = idom(prevdom); 1495 while (dom != cutoff) { 1496 if (dom->req() > 1 && n->as_If()->same_condition(dom, &_igvn) && prevdom->in(0) == dom && 1497 safe_for_if_replacement(dom)) { 1498 // It's invalid to move control dependent data nodes in the inner 1499 // strip-mined loop, because: 1500 // 1) break validation of LoopNode::verify_strip_mined() 1501 // 2) move code with side-effect in strip-mined loop 1502 // Move to the exit of outer strip-mined loop in that case. 1503 Node* out_le = is_inner_of_stripmined_loop(dom); 1504 if (out_le != nullptr) { 1505 prevdom = out_le; 1506 } 1507 // Replace the dominated test with an obvious true or false. 1508 // Place it on the IGVN worklist for later cleanup. 1509 C->set_major_progress(); 1510 // Split if: pin array accesses that are control dependent on a range check and moved to a regular if, 1511 // to prevent an array load from floating above its range check. There are three cases: 1512 // 1. Move from RangeCheck "a" to RangeCheck "b": don't need to pin. If we ever remove b, then we pin 1513 // all its array accesses at that point. 1514 // 2. We move from RangeCheck "a" to regular if "b": need to pin. If we ever remove b, then its array 1515 // accesses would start to float, since we don't pin at that point. 1516 // 3. If we move from regular if: don't pin. All array accesses are already assumed to be pinned. 1517 bool pin_array_access_nodes = n->Opcode() == Op_RangeCheck && 1518 prevdom->in(0)->Opcode() != Op_RangeCheck; 1519 dominated_by(prevdom->as_IfProj(), n->as_If(), false, pin_array_access_nodes); 1520 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } ); 1521 return; 1522 } 1523 prevdom = dom; 1524 dom = idom(prevdom); 1525 } 1526 } 1527 } 1528 1529 try_sink_out_of_loop(n); 1530 if (C->failing()) { 1531 return; 1532 } 1533 1534 try_move_store_after_loop(n); 1535 } 1536 1537 // Transform: 1538 // 1539 // if (some_condition) { 1540 // // body 1 1541 // } else { 1542 // // body 2 1543 // } 1544 // if (some_condition) { 1545 // // body 3 1546 // } else { 1547 // // body 4 1548 // } 1549 // 1550 // into: 1551 // 1552 // 1553 // if (some_condition) { 1554 // // body 1 1555 // // body 3 1556 // } else { 1557 // // body 2 1558 // // body 4 1559 // } 1560 bool PhaseIdealLoop::try_merge_identical_ifs(Node* n) { 1561 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) { 1562 Node *n_ctrl = n->in(0); 1563 IfNode* dom_if = idom(n_ctrl)->as_If(); 1564 if (n->in(1) != dom_if->in(1)) { 1565 assert(n->in(1)->in(1)->is_SubTypeCheck() && 1566 (n->in(1)->in(1)->as_SubTypeCheck()->method() != nullptr || 1567 dom_if->in(1)->in(1)->as_SubTypeCheck()->method() != nullptr), "only for subtype checks with profile data attached"); 1568 _igvn.replace_input_of(n, 1, dom_if->in(1)); 1569 } 1570 ProjNode* dom_proj_true = dom_if->proj_out(1); 1571 ProjNode* dom_proj_false = dom_if->proj_out(0); 1572 1573 // Now split the IF 1574 RegionNode* new_false_region; 1575 RegionNode* new_true_region; 1576 do_split_if(n, &new_false_region, &new_true_region); 1577 assert(new_false_region->req() == new_true_region->req(), ""); 1578 #ifdef ASSERT 1579 for (uint i = 1; i < new_false_region->req(); ++i) { 1580 assert(new_false_region->in(i)->in(0) == new_true_region->in(i)->in(0), "unexpected shape following split if"); 1581 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"); 1582 } 1583 #endif 1584 assert(new_false_region->in(1)->in(0)->in(1) == dom_if->in(1), "dominating if and dominated if after split must share test"); 1585 1586 // We now have: 1587 // if (some_condition) { 1588 // // body 1 1589 // if (some_condition) { 1590 // body3: // new_true_region 1591 // // body3 1592 // } else { 1593 // goto body4; 1594 // } 1595 // } else { 1596 // // body 2 1597 // if (some_condition) { 1598 // goto body3; 1599 // } else { 1600 // body4: // new_false_region 1601 // // body4; 1602 // } 1603 // } 1604 // 1605 1606 // clone pinned nodes thru the resulting regions 1607 push_pinned_nodes_thru_region(dom_if, new_true_region); 1608 push_pinned_nodes_thru_region(dom_if, new_false_region); 1609 1610 // Optimize out the cloned ifs. Because pinned nodes were cloned, this also allows a CastPP that would be dependent 1611 // 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 1612 // unrelated control dependency. 1613 for (uint i = 1; i < new_false_region->req(); i++) { 1614 if (is_dominator(dom_proj_true, new_false_region->in(i))) { 1615 dominated_by(dom_proj_true->as_IfProj(), new_false_region->in(i)->in(0)->as_If()); 1616 } else { 1617 assert(is_dominator(dom_proj_false, new_false_region->in(i)), "bad if"); 1618 dominated_by(dom_proj_false->as_IfProj(), new_false_region->in(i)->in(0)->as_If()); 1619 } 1620 } 1621 return true; 1622 } 1623 return false; 1624 } 1625 1626 void PhaseIdealLoop::push_pinned_nodes_thru_region(IfNode* dom_if, Node* region) { 1627 for (DUIterator i = region->outs(); region->has_out(i); i++) { 1628 Node* u = region->out(i); 1629 if (!has_ctrl(u) || u->is_Phi() || !u->depends_only_on_test() || !_igvn.no_dependent_zero_check(u)) { 1630 continue; 1631 } 1632 assert(u->in(0) == region, "not a control dependent node?"); 1633 uint j = 1; 1634 for (; j < u->req(); ++j) { 1635 Node* in = u->in(j); 1636 if (!is_dominator(ctrl_or_self(in), dom_if)) { 1637 break; 1638 } 1639 } 1640 if (j == u->req()) { 1641 Node *phi = PhiNode::make_blank(region, u); 1642 for (uint k = 1; k < region->req(); ++k) { 1643 Node* clone = u->clone(); 1644 clone->set_req(0, region->in(k)); 1645 register_new_node(clone, region->in(k)); 1646 phi->init_req(k, clone); 1647 } 1648 register_new_node(phi, region); 1649 _igvn.replace_node(u, phi); 1650 --i; 1651 } 1652 } 1653 } 1654 1655 bool PhaseIdealLoop::safe_for_if_replacement(const Node* dom) const { 1656 if (!dom->is_CountedLoopEnd()) { 1657 return true; 1658 } 1659 CountedLoopEndNode* le = dom->as_CountedLoopEnd(); 1660 CountedLoopNode* cl = le->loopnode(); 1661 if (cl == nullptr) { 1662 return true; 1663 } 1664 if (!cl->is_main_loop()) { 1665 return true; 1666 } 1667 if (cl->is_canonical_loop_entry() == nullptr) { 1668 return true; 1669 } 1670 // Further unrolling is possible so loop exit condition might change 1671 return false; 1672 } 1673 1674 // See if a shared loop-varying computation has no loop-varying uses. 1675 // Happens if something is only used for JVM state in uncommon trap exits, 1676 // like various versions of induction variable+offset. Clone the 1677 // computation per usage to allow it to sink out of the loop. 1678 void PhaseIdealLoop::try_sink_out_of_loop(Node* n) { 1679 if (has_ctrl(n) && 1680 !n->is_Phi() && 1681 !n->is_Bool() && 1682 !n->is_Proj() && 1683 !n->is_MergeMem() && 1684 !n->is_CMove() && 1685 !n->is_OpaqueNotNull() && 1686 !n->is_OpaqueInitializedAssertionPredicate() && 1687 !n->is_OpaqueTemplateAssertionPredicate() && 1688 !n->is_Type()) { 1689 Node *n_ctrl = get_ctrl(n); 1690 IdealLoopTree *n_loop = get_loop(n_ctrl); 1691 1692 if (n->in(0) != nullptr) { 1693 IdealLoopTree* loop_ctrl = get_loop(n->in(0)); 1694 if (n_loop != loop_ctrl && n_loop->is_member(loop_ctrl)) { 1695 // n has a control input inside a loop but get_ctrl() is member of an outer loop. This could happen, for example, 1696 // for Div nodes inside a loop (control input inside loop) without a use except for an UCT (outside the loop). 1697 // Rewire control of n to right outside of the loop, regardless if its input(s) are later sunk or not. 1698 Node* maybe_pinned_n = n; 1699 Node* outside_ctrl = place_outside_loop(n_ctrl, loop_ctrl); 1700 if (n->depends_only_on_test()) { 1701 Node* pinned_clone = n->pin_array_access_node(); 1702 if (pinned_clone != nullptr) { 1703 // Pin array access nodes: if this is an array load, it's going to be dependent on a condition that's not a 1704 // range check for that access. If that condition is replaced by an identical dominating one, then an 1705 // unpinned load would risk floating above its range check. 1706 register_new_node(pinned_clone, n_ctrl); 1707 maybe_pinned_n = pinned_clone; 1708 _igvn.replace_node(n, pinned_clone); 1709 } 1710 } 1711 _igvn.replace_input_of(maybe_pinned_n, 0, outside_ctrl); 1712 } 1713 } 1714 if (n_loop != _ltree_root && n->outcnt() > 1) { 1715 // Compute early control: needed for anti-dependence analysis. It's also possible that as a result of 1716 // previous transformations in this loop opts round, the node can be hoisted now: early control will tell us. 1717 Node* early_ctrl = compute_early_ctrl(n, n_ctrl); 1718 if (n_loop->is_member(get_loop(early_ctrl)) && // check that this one can't be hoisted now 1719 ctrl_of_all_uses_out_of_loop(n, early_ctrl, n_loop)) { // All uses in outer loops! 1720 if (n->is_Store() || n->is_LoadStore()) { 1721 assert(false, "no node with a side effect"); 1722 C->record_failure("no node with a side effect"); 1723 return; 1724 } 1725 Node* outer_loop_clone = nullptr; 1726 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin;) { 1727 Node* u = n->last_out(j); // Clone private computation per use 1728 _igvn.rehash_node_delayed(u); 1729 Node* x = nullptr; 1730 if (n->depends_only_on_test()) { 1731 // Pin array access nodes: if this is an array load, it's going to be dependent on a condition that's not a 1732 // range check for that access. If that condition is replaced by an identical dominating one, then an 1733 // unpinned load would risk floating above its range check. 1734 x = n->pin_array_access_node(); 1735 } 1736 if (x == nullptr) { 1737 x = n->clone(); 1738 } 1739 Node* x_ctrl = nullptr; 1740 if (u->is_Phi()) { 1741 // Replace all uses of normal nodes. Replace Phi uses 1742 // individually, so the separate Nodes can sink down 1743 // different paths. 1744 uint k = 1; 1745 while (u->in(k) != n) k++; 1746 u->set_req(k, x); 1747 // x goes next to Phi input path 1748 x_ctrl = u->in(0)->in(k); 1749 // Find control for 'x' next to use but not inside inner loops. 1750 x_ctrl = place_outside_loop(x_ctrl, n_loop); 1751 --j; 1752 } else { // Normal use 1753 if (has_ctrl(u)) { 1754 x_ctrl = get_ctrl(u); 1755 } else { 1756 x_ctrl = u->in(0); 1757 } 1758 // Find control for 'x' next to use but not inside inner loops. 1759 x_ctrl = place_outside_loop(x_ctrl, n_loop); 1760 // Replace all uses 1761 if (u->is_ConstraintCast() && _igvn.type(n)->higher_equal(u->bottom_type()) && u->in(0) == x_ctrl) { 1762 // If we're sinking a chain of data nodes, we might have inserted a cast to pin the use which is not necessary 1763 // anymore now that we're going to pin n as well 1764 _igvn.replace_node(u, x); 1765 --j; 1766 } else { 1767 int nb = u->replace_edge(n, x, &_igvn); 1768 j -= nb; 1769 } 1770 } 1771 1772 if (n->is_Load()) { 1773 // For loads, add a control edge to a CFG node outside of the loop 1774 // to force them to not combine and return back inside the loop 1775 // during GVN optimization (4641526). 1776 assert(x_ctrl == get_late_ctrl_with_anti_dep(x->as_Load(), early_ctrl, x_ctrl), "anti-dependences were already checked"); 1777 1778 IdealLoopTree* x_loop = get_loop(x_ctrl); 1779 Node* x_head = x_loop->_head; 1780 if (x_head->is_Loop() && x_head->is_OuterStripMinedLoop()) { 1781 // Do not add duplicate LoadNodes to the outer strip mined loop 1782 if (outer_loop_clone != nullptr) { 1783 _igvn.replace_node(x, outer_loop_clone); 1784 continue; 1785 } 1786 outer_loop_clone = x; 1787 } 1788 x->set_req(0, x_ctrl); 1789 } else if (n->in(0) != nullptr){ 1790 x->set_req(0, x_ctrl); 1791 } 1792 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone"); 1793 assert(!n_loop->is_member(get_loop(x_ctrl)), "should have moved out of loop"); 1794 register_new_node(x, x_ctrl); 1795 1796 // Chain of AddP nodes: (AddP base (AddP base (AddP base ))) 1797 // All AddP nodes must keep the same base after sinking so: 1798 // 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, 1799 // their bases remain the same. 1800 // (see 2- below) 1801 assert(!x->is_AddP() || !x->in(AddPNode::Address)->is_AddP() || 1802 x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base) || 1803 !x->in(AddPNode::Address)->in(AddPNode::Base)->eqv_uncast(x->in(AddPNode::Base)), "unexpected AddP shape"); 1804 if (x->in(0) == nullptr && !x->is_DecodeNarrowPtr() && 1805 !(x->is_AddP() && x->in(AddPNode::Address)->is_AddP() && x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base))) { 1806 assert(!x->is_Load(), "load should be pinned"); 1807 // Use a cast node to pin clone out of loop 1808 Node* cast = nullptr; 1809 for (uint k = 0; k < x->req(); k++) { 1810 Node* in = x->in(k); 1811 if (in != nullptr && n_loop->is_member(get_loop(get_ctrl(in)))) { 1812 const Type* in_t = _igvn.type(in); 1813 cast = ConstraintCastNode::make_cast_for_type(x_ctrl, in, in_t, 1814 ConstraintCastNode::UnconditionalDependency, nullptr); 1815 } 1816 if (cast != nullptr) { 1817 Node* prev = _igvn.hash_find_insert(cast); 1818 if (prev != nullptr && get_ctrl(prev) == x_ctrl) { 1819 cast->destruct(&_igvn); 1820 cast = prev; 1821 } else { 1822 register_new_node(cast, x_ctrl); 1823 } 1824 x->replace_edge(in, cast); 1825 // Chain of AddP nodes: 1826 // 2- A CastPP of the base is only added now that all AddP nodes are sunk 1827 if (x->is_AddP() && k == AddPNode::Base) { 1828 update_addp_chain_base(x, n->in(AddPNode::Base), cast); 1829 } 1830 break; 1831 } 1832 } 1833 assert(cast != nullptr, "must have added a cast to pin the node"); 1834 } 1835 } 1836 _igvn.remove_dead_node(n); 1837 } 1838 _dom_lca_tags_round = 0; 1839 } 1840 } 1841 } 1842 1843 void PhaseIdealLoop::update_addp_chain_base(Node* x, Node* old_base, Node* new_base) { 1844 ResourceMark rm; 1845 Node_List wq; 1846 wq.push(x); 1847 while (wq.size() != 0) { 1848 Node* n = wq.pop(); 1849 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1850 Node* u = n->fast_out(i); 1851 if (u->is_AddP() && u->in(AddPNode::Base) == old_base) { 1852 _igvn.replace_input_of(u, AddPNode::Base, new_base); 1853 wq.push(u); 1854 } 1855 } 1856 } 1857 } 1858 1859 // Compute the early control of a node by following its inputs until we reach 1860 // nodes that are pinned. Then compute the LCA of the control of all pinned nodes. 1861 Node* PhaseIdealLoop::compute_early_ctrl(Node* n, Node* n_ctrl) { 1862 Node* early_ctrl = nullptr; 1863 ResourceMark rm; 1864 Unique_Node_List wq; 1865 wq.push(n); 1866 for (uint i = 0; i < wq.size(); i++) { 1867 Node* m = wq.at(i); 1868 Node* c = nullptr; 1869 if (m->is_CFG()) { 1870 c = m; 1871 } else if (m->pinned()) { 1872 c = m->in(0); 1873 } else { 1874 for (uint j = 0; j < m->req(); j++) { 1875 Node* in = m->in(j); 1876 if (in != nullptr) { 1877 wq.push(in); 1878 } 1879 } 1880 } 1881 if (c != nullptr) { 1882 assert(is_dominator(c, n_ctrl), "control input must dominate current control"); 1883 if (early_ctrl == nullptr || is_dominator(early_ctrl, c)) { 1884 early_ctrl = c; 1885 } 1886 } 1887 } 1888 assert(is_dominator(early_ctrl, n_ctrl), "early control must dominate current control"); 1889 return early_ctrl; 1890 } 1891 1892 bool PhaseIdealLoop::ctrl_of_all_uses_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop) { 1893 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1894 Node* u = n->fast_out(i); 1895 if (u->is_Opaque1()) { 1896 return false; // Found loop limit, bugfix for 4677003 1897 } 1898 // We can't reuse tags in PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal() so make sure calls to 1899 // get_late_ctrl_with_anti_dep() use their own tag 1900 _dom_lca_tags_round++; 1901 assert(_dom_lca_tags_round != 0, "shouldn't wrap around"); 1902 1903 if (u->is_Phi()) { 1904 for (uint j = 1; j < u->req(); ++j) { 1905 if (u->in(j) == n && !ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, u->in(0)->in(j))) { 1906 return false; 1907 } 1908 } 1909 } else { 1910 Node* ctrl = has_ctrl(u) ? get_ctrl(u) : u->in(0); 1911 if (!ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, ctrl)) { 1912 return false; 1913 } 1914 } 1915 } 1916 return true; 1917 } 1918 1919 bool PhaseIdealLoop::ctrl_of_use_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop, Node* ctrl) { 1920 if (n->is_Load()) { 1921 ctrl = get_late_ctrl_with_anti_dep(n->as_Load(), n_ctrl, ctrl); 1922 } 1923 IdealLoopTree *u_loop = get_loop(ctrl); 1924 if (u_loop == n_loop) { 1925 return false; // Found loop-varying use 1926 } 1927 if (n_loop->is_member(u_loop)) { 1928 return false; // Found use in inner loop 1929 } 1930 // 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 1931 // to a check that's eliminated by range check elimination, it becomes input to an expression that feeds into the exit 1932 // test of the pre loop above the point in the graph where it's pinned. 1933 if (n_loop->_head->is_CountedLoop() && n_loop->_head->as_CountedLoop()->is_pre_loop()) { 1934 CountedLoopNode* pre_loop = n_loop->_head->as_CountedLoop(); 1935 if (is_dominator(pre_loop->loopexit(), ctrl)) { 1936 return false; 1937 } 1938 } 1939 return true; 1940 } 1941 1942 //------------------------------split_if_with_blocks--------------------------- 1943 // Check for aggressive application of 'split-if' optimization, 1944 // using basic block level info. 1945 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack) { 1946 Node* root = C->root(); 1947 visited.set(root->_idx); // first, mark root as visited 1948 // Do pre-visit work for root 1949 Node* n = split_if_with_blocks_pre(root); 1950 uint cnt = n->outcnt(); 1951 uint i = 0; 1952 1953 while (true) { 1954 // Visit all children 1955 if (i < cnt) { 1956 Node* use = n->raw_out(i); 1957 ++i; 1958 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) { 1959 // Now do pre-visit work for this use 1960 use = split_if_with_blocks_pre(use); 1961 nstack.push(n, i); // Save parent and next use's index. 1962 n = use; // Process all children of current use. 1963 cnt = use->outcnt(); 1964 i = 0; 1965 } 1966 } 1967 else { 1968 // All of n's children have been processed, complete post-processing. 1969 if (cnt != 0 && !n->is_Con()) { 1970 assert(has_node(n), "no dead nodes"); 1971 split_if_with_blocks_post(n); 1972 if (C->failing()) { 1973 return; 1974 } 1975 } 1976 if (must_throttle_split_if()) { 1977 nstack.clear(); 1978 } 1979 if (nstack.is_empty()) { 1980 // Finished all nodes on stack. 1981 break; 1982 } 1983 // Get saved parent node and next use's index. Visit the rest of uses. 1984 n = nstack.node(); 1985 cnt = n->outcnt(); 1986 i = nstack.index(); 1987 nstack.pop(); 1988 } 1989 } 1990 } 1991 1992 1993 //============================================================================= 1994 // 1995 // C L O N E A L O O P B O D Y 1996 // 1997 1998 //------------------------------clone_iff-------------------------------------- 1999 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 2000 // "Nearly" because all Nodes have been cloned from the original in the loop, 2001 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 2002 // through the Phi recursively, and return a Bool. 2003 Node* PhaseIdealLoop::clone_iff(PhiNode* phi) { 2004 2005 // Convert this Phi into a Phi merging Bools 2006 uint i; 2007 for (i = 1; i < phi->req(); i++) { 2008 Node* b = phi->in(i); 2009 if (b->is_Phi()) { 2010 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi())); 2011 } else { 2012 assert(b->is_Bool() || b->is_OpaqueNotNull() || b->is_OpaqueInitializedAssertionPredicate(), 2013 "bool, non-null check with OpaqueNotNull or Initialized Assertion Predicate with its Opaque node"); 2014 } 2015 } 2016 Node* n = phi->in(1); 2017 Node* sample_opaque = nullptr; 2018 Node *sample_bool = nullptr; 2019 if (n->is_OpaqueNotNull() || n->is_OpaqueInitializedAssertionPredicate()) { 2020 sample_opaque = n; 2021 sample_bool = n->in(1); 2022 assert(sample_bool->is_Bool(), "wrong type"); 2023 } else { 2024 sample_bool = n; 2025 } 2026 Node *sample_cmp = sample_bool->in(1); 2027 2028 // Make Phis to merge the Cmp's inputs. 2029 PhiNode *phi1 = new PhiNode(phi->in(0), Type::TOP); 2030 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP); 2031 for (i = 1; i < phi->req(); i++) { 2032 Node *n1 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1); 2033 Node *n2 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2); 2034 phi1->set_req(i, n1); 2035 phi2->set_req(i, n2); 2036 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 2037 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 2038 } 2039 // See if these Phis have been made before. 2040 // Register with optimizer 2041 Node *hit1 = _igvn.hash_find_insert(phi1); 2042 if (hit1) { // Hit, toss just made Phi 2043 _igvn.remove_dead_node(phi1); // Remove new phi 2044 assert(hit1->is_Phi(), "" ); 2045 phi1 = (PhiNode*)hit1; // Use existing phi 2046 } else { // Miss 2047 _igvn.register_new_node_with_optimizer(phi1); 2048 } 2049 Node *hit2 = _igvn.hash_find_insert(phi2); 2050 if (hit2) { // Hit, toss just made Phi 2051 _igvn.remove_dead_node(phi2); // Remove new phi 2052 assert(hit2->is_Phi(), "" ); 2053 phi2 = (PhiNode*)hit2; // Use existing phi 2054 } else { // Miss 2055 _igvn.register_new_node_with_optimizer(phi2); 2056 } 2057 // Register Phis with loop/block info 2058 set_ctrl(phi1, phi->in(0)); 2059 set_ctrl(phi2, phi->in(0)); 2060 // Make a new Cmp 2061 Node *cmp = sample_cmp->clone(); 2062 cmp->set_req(1, phi1); 2063 cmp->set_req(2, phi2); 2064 _igvn.register_new_node_with_optimizer(cmp); 2065 set_ctrl(cmp, phi->in(0)); 2066 2067 // Make a new Bool 2068 Node *b = sample_bool->clone(); 2069 b->set_req(1,cmp); 2070 _igvn.register_new_node_with_optimizer(b); 2071 set_ctrl(b, phi->in(0)); 2072 2073 if (sample_opaque != nullptr) { 2074 Node* opaque = sample_opaque->clone(); 2075 opaque->set_req(1, b); 2076 _igvn.register_new_node_with_optimizer(opaque); 2077 set_ctrl(opaque, phi->in(0)); 2078 return opaque; 2079 } 2080 2081 assert(b->is_Bool(), ""); 2082 return b; 2083 } 2084 2085 //------------------------------clone_bool------------------------------------- 2086 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 2087 // "Nearly" because all Nodes have been cloned from the original in the loop, 2088 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 2089 // through the Phi recursively, and return a Bool. 2090 CmpNode*PhaseIdealLoop::clone_bool(PhiNode* phi) { 2091 uint i; 2092 // Convert this Phi into a Phi merging Bools 2093 for( i = 1; i < phi->req(); i++ ) { 2094 Node *b = phi->in(i); 2095 if( b->is_Phi() ) { 2096 _igvn.replace_input_of(phi, i, clone_bool(b->as_Phi())); 2097 } else { 2098 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" ); 2099 } 2100 } 2101 2102 Node *sample_cmp = phi->in(1); 2103 2104 // Make Phis to merge the Cmp's inputs. 2105 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP ); 2106 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP ); 2107 for( uint j = 1; j < phi->req(); j++ ) { 2108 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP 2109 Node *n1, *n2; 2110 if( cmp_top->is_Cmp() ) { 2111 n1 = cmp_top->in(1); 2112 n2 = cmp_top->in(2); 2113 } else { 2114 n1 = n2 = cmp_top; 2115 } 2116 phi1->set_req( j, n1 ); 2117 phi2->set_req( j, n2 ); 2118 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 2119 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 2120 } 2121 2122 // See if these Phis have been made before. 2123 // Register with optimizer 2124 Node *hit1 = _igvn.hash_find_insert(phi1); 2125 if( hit1 ) { // Hit, toss just made Phi 2126 _igvn.remove_dead_node(phi1); // Remove new phi 2127 assert( hit1->is_Phi(), "" ); 2128 phi1 = (PhiNode*)hit1; // Use existing phi 2129 } else { // Miss 2130 _igvn.register_new_node_with_optimizer(phi1); 2131 } 2132 Node *hit2 = _igvn.hash_find_insert(phi2); 2133 if( hit2 ) { // Hit, toss just made Phi 2134 _igvn.remove_dead_node(phi2); // Remove new phi 2135 assert( hit2->is_Phi(), "" ); 2136 phi2 = (PhiNode*)hit2; // Use existing phi 2137 } else { // Miss 2138 _igvn.register_new_node_with_optimizer(phi2); 2139 } 2140 // Register Phis with loop/block info 2141 set_ctrl(phi1, phi->in(0)); 2142 set_ctrl(phi2, phi->in(0)); 2143 // Make a new Cmp 2144 Node *cmp = sample_cmp->clone(); 2145 cmp->set_req( 1, phi1 ); 2146 cmp->set_req( 2, phi2 ); 2147 _igvn.register_new_node_with_optimizer(cmp); 2148 set_ctrl(cmp, phi->in(0)); 2149 2150 assert( cmp->is_Cmp(), "" ); 2151 return (CmpNode*)cmp; 2152 } 2153 2154 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new, 2155 IdealLoopTree* loop, IdealLoopTree* outer_loop, 2156 Node_List*& split_if_set, Node_List*& split_bool_set, 2157 Node_List*& split_cex_set, Node_List& worklist, 2158 uint new_counter, CloneLoopMode mode) { 2159 Node* nnn = old_new[old->_idx]; 2160 // Copy uses to a worklist, so I can munge the def-use info 2161 // with impunity. 2162 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 2163 worklist.push(old->fast_out(j)); 2164 2165 while( worklist.size() ) { 2166 Node *use = worklist.pop(); 2167 if (!has_node(use)) continue; // Ignore dead nodes 2168 if (use->in(0) == C->top()) continue; 2169 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 2170 // Check for data-use outside of loop - at least one of OLD or USE 2171 // must not be a CFG node. 2172 #ifdef ASSERT 2173 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == nullptr) { 2174 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint(); 2175 assert(mode != IgnoreStripMined, "incorrect cloning mode"); 2176 assert((mode == ControlAroundStripMined && use == sfpt) || !use->is_reachable_from_root(), "missed a node"); 2177 } 2178 #endif 2179 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) { 2180 2181 // If the Data use is an IF, that means we have an IF outside the 2182 // loop that is switching on a condition that is set inside the 2183 // loop. Happens if people set a loop-exit flag; then test the flag 2184 // in the loop to break the loop, then test is again outside the 2185 // loop to determine which way the loop exited. 2186 // 2187 // For several uses we need to make sure that there is no phi between, 2188 // the use and the Bool/Cmp. We therefore clone the Bool/Cmp down here 2189 // to avoid such a phi in between. 2190 // For example, it is unexpected that there is a Phi between an 2191 // AllocateArray node and its ValidLengthTest input that could cause 2192 // split if to break. 2193 assert(!use->is_OpaqueTemplateAssertionPredicate(), 2194 "should not clone a Template Assertion Predicate which should be removed once it's useless"); 2195 if (use->is_If() || use->is_CMove() || use->is_OpaqueNotNull() || use->is_OpaqueInitializedAssertionPredicate() || 2196 (use->Opcode() == Op_AllocateArray && use->in(AllocateNode::ValidLengthTest) == old)) { 2197 // Since this code is highly unlikely, we lazily build the worklist 2198 // of such Nodes to go split. 2199 if (!split_if_set) { 2200 split_if_set = new Node_List(); 2201 } 2202 split_if_set->push(use); 2203 } 2204 if (use->is_Bool()) { 2205 if (!split_bool_set) { 2206 split_bool_set = new Node_List(); 2207 } 2208 split_bool_set->push(use); 2209 } 2210 if (use->Opcode() == Op_CreateEx) { 2211 if (!split_cex_set) { 2212 split_cex_set = new Node_List(); 2213 } 2214 split_cex_set->push(use); 2215 } 2216 2217 2218 // Get "block" use is in 2219 uint idx = 0; 2220 while( use->in(idx) != old ) idx++; 2221 Node *prev = use->is_CFG() ? use : get_ctrl(use); 2222 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" ); 2223 Node* cfg = (prev->_idx >= new_counter && prev->is_Region()) 2224 ? prev->in(2) 2225 : idom(prev); 2226 if( use->is_Phi() ) // Phi use is in prior block 2227 cfg = prev->in(idx); // NOT in block of Phi itself 2228 if (cfg->is_top()) { // Use is dead? 2229 _igvn.replace_input_of(use, idx, C->top()); 2230 continue; 2231 } 2232 2233 // If use is referenced through control edge... (idx == 0) 2234 if (mode == IgnoreStripMined && idx == 0) { 2235 LoopNode *head = loop->_head->as_Loop(); 2236 if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) { 2237 // That node is outside the inner loop, leave it outside the 2238 // outer loop as well to not confuse verification code. 2239 assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop"); 2240 _igvn.replace_input_of(use, 0, head->outer_loop_exit()); 2241 continue; 2242 } 2243 } 2244 2245 while(!outer_loop->is_member(get_loop(cfg))) { 2246 prev = cfg; 2247 cfg = (cfg->_idx >= new_counter && cfg->is_Region()) ? cfg->in(2) : idom(cfg); 2248 } 2249 // If the use occurs after merging several exits from the loop, then 2250 // old value must have dominated all those exits. Since the same old 2251 // value was used on all those exits we did not need a Phi at this 2252 // merge point. NOW we do need a Phi here. Each loop exit value 2253 // is now merged with the peeled body exit; each exit gets its own 2254 // private Phi and those Phis need to be merged here. 2255 Node *phi; 2256 if( prev->is_Region() ) { 2257 if( idx == 0 ) { // Updating control edge? 2258 phi = prev; // Just use existing control 2259 } else { // Else need a new Phi 2260 phi = PhiNode::make( prev, old ); 2261 // Now recursively fix up the new uses of old! 2262 for( uint i = 1; i < prev->req(); i++ ) { 2263 worklist.push(phi); // Onto worklist once for each 'old' input 2264 } 2265 } 2266 } else { 2267 // Get new RegionNode merging old and new loop exits 2268 prev = old_new[prev->_idx]; 2269 assert( prev, "just made this in step 7" ); 2270 if( idx == 0) { // Updating control edge? 2271 phi = prev; // Just use existing control 2272 } else { // Else need a new Phi 2273 // Make a new Phi merging data values properly 2274 phi = PhiNode::make( prev, old ); 2275 phi->set_req( 1, nnn ); 2276 } 2277 } 2278 // If inserting a new Phi, check for prior hits 2279 if( idx != 0 ) { 2280 Node *hit = _igvn.hash_find_insert(phi); 2281 if( hit == nullptr ) { 2282 _igvn.register_new_node_with_optimizer(phi); // Register new phi 2283 } else { // or 2284 // Remove the new phi from the graph and use the hit 2285 _igvn.remove_dead_node(phi); 2286 phi = hit; // Use existing phi 2287 } 2288 set_ctrl(phi, prev); 2289 } 2290 // Make 'use' use the Phi instead of the old loop body exit value 2291 assert(use->in(idx) == old, "old is still input of use"); 2292 // We notify all uses of old, including use, and the indirect uses, 2293 // that may now be optimized because we have replaced old with phi. 2294 _igvn.add_users_to_worklist(old); 2295 if (idx == 0 && 2296 use->depends_only_on_test()) { 2297 Node* pinned_clone = use->pin_array_access_node(); 2298 if (pinned_clone != nullptr) { 2299 // Pin array access nodes: control is updated here to a region. If, after some transformations, only one path 2300 // into the region is left, an array load could become dependent on a condition that's not a range check for 2301 // that access. If that condition is replaced by an identical dominating one, then an unpinned load would risk 2302 // floating above its range check. 2303 pinned_clone->set_req(0, phi); 2304 register_new_node_with_ctrl_of(pinned_clone, use); 2305 _igvn.replace_node(use, pinned_clone); 2306 continue; 2307 } 2308 } 2309 _igvn.replace_input_of(use, idx, phi); 2310 if( use->_idx >= new_counter ) { // If updating new phis 2311 // Not needed for correctness, but prevents a weak assert 2312 // in AddPNode from tripping (when we end up with different 2313 // base & derived Phis that will become the same after 2314 // IGVN does CSE). 2315 Node *hit = _igvn.hash_find_insert(use); 2316 if( hit ) // Go ahead and re-hash for hits. 2317 _igvn.replace_node( use, hit ); 2318 } 2319 } 2320 } 2321 } 2322 2323 static void collect_nodes_in_outer_loop_not_reachable_from_sfpt(Node* n, const IdealLoopTree *loop, const IdealLoopTree* outer_loop, 2324 const Node_List &old_new, Unique_Node_List& wq, PhaseIdealLoop* phase, 2325 bool check_old_new) { 2326 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2327 Node* u = n->fast_out(j); 2328 assert(check_old_new || old_new[u->_idx] == nullptr, "shouldn't have been cloned"); 2329 if (!u->is_CFG() && (!check_old_new || old_new[u->_idx] == nullptr)) { 2330 Node* c = phase->get_ctrl(u); 2331 IdealLoopTree* u_loop = phase->get_loop(c); 2332 assert(!loop->is_member(u_loop) || !loop->_body.contains(u), "can be in outer loop or out of both loops only"); 2333 if (!loop->is_member(u_loop)) { 2334 if (outer_loop->is_member(u_loop)) { 2335 wq.push(u); 2336 } else { 2337 // nodes pinned with control in the outer loop but not referenced from the safepoint must be moved out of 2338 // the outer loop too 2339 Node* u_c = u->in(0); 2340 if (u_c != nullptr) { 2341 IdealLoopTree* u_c_loop = phase->get_loop(u_c); 2342 if (outer_loop->is_member(u_c_loop) && !loop->is_member(u_c_loop)) { 2343 wq.push(u); 2344 } 2345 } 2346 } 2347 } 2348 } 2349 } 2350 } 2351 2352 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop, 2353 IdealLoopTree* outer_loop, int dd, Node_List &old_new, 2354 Node_List& extra_data_nodes) { 2355 if (head->is_strip_mined() && mode != IgnoreStripMined) { 2356 CountedLoopNode* cl = head->as_CountedLoop(); 2357 Node* l = cl->outer_loop(); 2358 Node* tail = cl->outer_loop_tail(); 2359 IfNode* le = cl->outer_loop_end(); 2360 Node* sfpt = cl->outer_safepoint(); 2361 CountedLoopEndNode* cle = cl->loopexit(); 2362 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop(); 2363 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null(); 2364 Node* cle_out = cle->proj_out(false); 2365 2366 Node* new_sfpt = nullptr; 2367 Node* new_cle_out = cle_out->clone(); 2368 old_new.map(cle_out->_idx, new_cle_out); 2369 if (mode == CloneIncludesStripMined) { 2370 // clone outer loop body 2371 Node* new_l = l->clone(); 2372 Node* new_tail = tail->clone(); 2373 IfNode* new_le = le->clone()->as_If(); 2374 new_sfpt = sfpt->clone(); 2375 2376 set_loop(new_l, outer_loop->_parent); 2377 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd); 2378 set_loop(new_cle_out, outer_loop->_parent); 2379 set_idom(new_cle_out, new_cle, dd); 2380 set_loop(new_sfpt, outer_loop->_parent); 2381 set_idom(new_sfpt, new_cle_out, dd); 2382 set_loop(new_le, outer_loop->_parent); 2383 set_idom(new_le, new_sfpt, dd); 2384 set_loop(new_tail, outer_loop->_parent); 2385 set_idom(new_tail, new_le, dd); 2386 set_idom(new_cl, new_l, dd); 2387 2388 old_new.map(l->_idx, new_l); 2389 old_new.map(tail->_idx, new_tail); 2390 old_new.map(le->_idx, new_le); 2391 old_new.map(sfpt->_idx, new_sfpt); 2392 2393 new_l->set_req(LoopNode::LoopBackControl, new_tail); 2394 new_l->set_req(0, new_l); 2395 new_tail->set_req(0, new_le); 2396 new_le->set_req(0, new_sfpt); 2397 new_sfpt->set_req(0, new_cle_out); 2398 new_cle_out->set_req(0, new_cle); 2399 new_cl->set_req(LoopNode::EntryControl, new_l); 2400 2401 _igvn.register_new_node_with_optimizer(new_l); 2402 _igvn.register_new_node_with_optimizer(new_tail); 2403 _igvn.register_new_node_with_optimizer(new_le); 2404 } else { 2405 Node *newhead = old_new[loop->_head->_idx]; 2406 newhead->as_Loop()->clear_strip_mined(); 2407 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl)); 2408 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 2409 } 2410 // Look at data node that were assigned a control in the outer 2411 // loop: they are kept in the outer loop by the safepoint so start 2412 // from the safepoint node's inputs. 2413 IdealLoopTree* outer_loop = get_loop(l); 2414 Node_Stack stack(2); 2415 stack.push(sfpt, 1); 2416 uint new_counter = C->unique(); 2417 while (stack.size() > 0) { 2418 Node* n = stack.node(); 2419 uint i = stack.index(); 2420 while (i < n->req() && 2421 (n->in(i) == nullptr || 2422 !has_ctrl(n->in(i)) || 2423 get_loop(get_ctrl(n->in(i))) != outer_loop || 2424 (old_new[n->in(i)->_idx] != nullptr && old_new[n->in(i)->_idx]->_idx >= new_counter))) { 2425 i++; 2426 } 2427 if (i < n->req()) { 2428 stack.set_index(i+1); 2429 stack.push(n->in(i), 0); 2430 } else { 2431 assert(old_new[n->_idx] == nullptr || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet"); 2432 Node* m = n == sfpt ? new_sfpt : n->clone(); 2433 if (m != nullptr) { 2434 for (uint i = 0; i < n->req(); i++) { 2435 if (m->in(i) != nullptr && old_new[m->in(i)->_idx] != nullptr) { 2436 m->set_req(i, old_new[m->in(i)->_idx]); 2437 } 2438 } 2439 } else { 2440 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?"); 2441 } 2442 if (n != sfpt) { 2443 extra_data_nodes.push(n); 2444 _igvn.register_new_node_with_optimizer(m); 2445 assert(get_ctrl(n) == cle_out, "what other control?"); 2446 set_ctrl(m, new_cle_out); 2447 old_new.map(n->_idx, m); 2448 } 2449 stack.pop(); 2450 } 2451 } 2452 if (mode == CloneIncludesStripMined) { 2453 _igvn.register_new_node_with_optimizer(new_sfpt); 2454 _igvn.register_new_node_with_optimizer(new_cle_out); 2455 } 2456 // Some other transformation may have pessimistically assigned some 2457 // data nodes to the outer loop. Set their control so they are out 2458 // of the outer loop. 2459 ResourceMark rm; 2460 Unique_Node_List wq; 2461 for (uint i = 0; i < extra_data_nodes.size(); i++) { 2462 Node* old = extra_data_nodes.at(i); 2463 collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true); 2464 } 2465 2466 for (uint i = 0; i < loop->_body.size(); i++) { 2467 Node* old = loop->_body.at(i); 2468 collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true); 2469 } 2470 2471 Node* inner_out = sfpt->in(0); 2472 if (inner_out->outcnt() > 1) { 2473 collect_nodes_in_outer_loop_not_reachable_from_sfpt(inner_out, loop, outer_loop, old_new, wq, this, true); 2474 } 2475 2476 Node* new_ctrl = cl->outer_loop_exit(); 2477 assert(get_loop(new_ctrl) != outer_loop, "must be out of the loop nest"); 2478 for (uint i = 0; i < wq.size(); i++) { 2479 Node* n = wq.at(i); 2480 set_ctrl(n, new_ctrl); 2481 if (n->in(0) != nullptr) { 2482 _igvn.replace_input_of(n, 0, new_ctrl); 2483 } 2484 collect_nodes_in_outer_loop_not_reachable_from_sfpt(n, loop, outer_loop, old_new, wq, this, false); 2485 } 2486 } else { 2487 Node *newhead = old_new[loop->_head->_idx]; 2488 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 2489 } 2490 } 2491 2492 //------------------------------clone_loop------------------------------------- 2493 // 2494 // C L O N E A L O O P B O D Y 2495 // 2496 // This is the basic building block of the loop optimizations. It clones an 2497 // entire loop body. It makes an old_new loop body mapping; with this mapping 2498 // you can find the new-loop equivalent to an old-loop node. All new-loop 2499 // nodes are exactly equal to their old-loop counterparts, all edges are the 2500 // same. All exits from the old-loop now have a RegionNode that merges the 2501 // equivalent new-loop path. This is true even for the normal "loop-exit" 2502 // condition. All uses of loop-invariant old-loop values now come from (one 2503 // or more) Phis that merge their new-loop equivalents. 2504 // 2505 // This operation leaves the graph in an illegal state: there are two valid 2506 // control edges coming from the loop pre-header to both loop bodies. I'll 2507 // definitely have to hack the graph after running this transform. 2508 // 2509 // From this building block I will further edit edges to perform loop peeling 2510 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc. 2511 // 2512 // Parameter side_by_size_idom: 2513 // When side_by_size_idom is null, the dominator tree is constructed for 2514 // the clone loop to dominate the original. Used in construction of 2515 // pre-main-post loop sequence. 2516 // When nonnull, the clone and original are side-by-side, both are 2517 // dominated by the side_by_side_idom node. Used in construction of 2518 // unswitched loops. 2519 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd, 2520 CloneLoopMode mode, Node* side_by_side_idom) { 2521 2522 LoopNode* head = loop->_head->as_Loop(); 2523 head->verify_strip_mined(1); 2524 2525 if (C->do_vector_loop() && PrintOpto) { 2526 const char* mname = C->method()->name()->as_quoted_ascii(); 2527 if (mname != nullptr) { 2528 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname); 2529 } 2530 } 2531 2532 CloneMap& cm = C->clone_map(); 2533 if (C->do_vector_loop()) { 2534 cm.set_clone_idx(cm.max_gen()+1); 2535 #ifndef PRODUCT 2536 if (PrintOpto) { 2537 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx()); 2538 loop->dump_head(); 2539 } 2540 #endif 2541 } 2542 2543 // Step 1: Clone the loop body. Make the old->new mapping. 2544 clone_loop_body(loop->_body, old_new, &cm); 2545 2546 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop; 2547 2548 // Step 2: Fix the edges in the new body. If the old input is outside the 2549 // loop use it. If the old input is INside the loop, use the corresponding 2550 // new node instead. 2551 fix_body_edges(loop->_body, loop, old_new, dd, outer_loop->_parent, false); 2552 2553 Node_List extra_data_nodes; // data nodes in the outer strip mined loop 2554 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes); 2555 2556 // Step 3: Now fix control uses. Loop varying control uses have already 2557 // been fixed up (as part of all input edges in Step 2). Loop invariant 2558 // control uses must be either an IfFalse or an IfTrue. Make a merge 2559 // point to merge the old and new IfFalse/IfTrue nodes; make the use 2560 // refer to this. 2561 Node_List worklist; 2562 uint new_counter = C->unique(); 2563 fix_ctrl_uses(loop->_body, loop, old_new, mode, side_by_side_idom, &cm, worklist); 2564 2565 // Step 4: If loop-invariant use is not control, it must be dominated by a 2566 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region 2567 // there if needed. Make a Phi there merging old and new used values. 2568 Node_List *split_if_set = nullptr; 2569 Node_List *split_bool_set = nullptr; 2570 Node_List *split_cex_set = nullptr; 2571 fix_data_uses(loop->_body, loop, mode, outer_loop, new_counter, old_new, worklist, split_if_set, split_bool_set, split_cex_set); 2572 2573 for (uint i = 0; i < extra_data_nodes.size(); i++) { 2574 Node* old = extra_data_nodes.at(i); 2575 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set, 2576 split_bool_set, split_cex_set, worklist, new_counter, 2577 mode); 2578 } 2579 2580 // Check for IFs that need splitting/cloning. Happens if an IF outside of 2581 // the loop uses a condition set in the loop. The original IF probably 2582 // takes control from one or more OLD Regions (which in turn get from NEW 2583 // Regions). In any case, there will be a set of Phis for each merge point 2584 // from the IF up to where the original BOOL def exists the loop. 2585 finish_clone_loop(split_if_set, split_bool_set, split_cex_set); 2586 2587 } 2588 2589 void PhaseIdealLoop::finish_clone_loop(Node_List* split_if_set, Node_List* split_bool_set, Node_List* split_cex_set) { 2590 if (split_if_set) { 2591 while (split_if_set->size()) { 2592 Node *iff = split_if_set->pop(); 2593 uint input = iff->Opcode() == Op_AllocateArray ? AllocateNode::ValidLengthTest : 1; 2594 if (iff->in(input)->is_Phi()) { 2595 Node *b = clone_iff(iff->in(input)->as_Phi()); 2596 _igvn.replace_input_of(iff, input, b); 2597 } 2598 } 2599 } 2600 if (split_bool_set) { 2601 while (split_bool_set->size()) { 2602 Node *b = split_bool_set->pop(); 2603 Node *phi = b->in(1); 2604 assert(phi->is_Phi(), ""); 2605 CmpNode *cmp = clone_bool((PhiNode*) phi); 2606 _igvn.replace_input_of(b, 1, cmp); 2607 } 2608 } 2609 if (split_cex_set) { 2610 while (split_cex_set->size()) { 2611 Node *b = split_cex_set->pop(); 2612 assert(b->in(0)->is_Region(), ""); 2613 assert(b->in(1)->is_Phi(), ""); 2614 assert(b->in(0)->in(0) == b->in(1)->in(0), ""); 2615 split_up(b, b->in(0), nullptr); 2616 } 2617 } 2618 } 2619 2620 void PhaseIdealLoop::fix_data_uses(Node_List& body, IdealLoopTree* loop, CloneLoopMode mode, IdealLoopTree* outer_loop, 2621 uint new_counter, Node_List &old_new, Node_List &worklist, Node_List*& split_if_set, 2622 Node_List*& split_bool_set, Node_List*& split_cex_set) { 2623 for(uint i = 0; i < body.size(); i++ ) { 2624 Node* old = body.at(i); 2625 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set, 2626 split_bool_set, split_cex_set, worklist, new_counter, 2627 mode); 2628 } 2629 } 2630 2631 void PhaseIdealLoop::fix_ctrl_uses(const Node_List& body, const IdealLoopTree* loop, Node_List &old_new, CloneLoopMode mode, 2632 Node* side_by_side_idom, CloneMap* cm, Node_List &worklist) { 2633 LoopNode* head = loop->_head->as_Loop(); 2634 for(uint i = 0; i < body.size(); i++ ) { 2635 Node* old = body.at(i); 2636 if( !old->is_CFG() ) continue; 2637 2638 // Copy uses to a worklist, so I can munge the def-use info 2639 // with impunity. 2640 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) { 2641 worklist.push(old->fast_out(j)); 2642 } 2643 2644 while (worklist.size()) { // Visit all uses 2645 Node *use = worklist.pop(); 2646 if (!has_node(use)) continue; // Ignore dead nodes 2647 IdealLoopTree *use_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use ); 2648 if (!loop->is_member(use_loop) && use->is_CFG()) { 2649 // Both OLD and USE are CFG nodes here. 2650 assert(use->is_Proj(), "" ); 2651 Node* nnn = old_new[old->_idx]; 2652 2653 Node* newuse = nullptr; 2654 if (head->is_strip_mined() && mode != IgnoreStripMined) { 2655 CountedLoopNode* cl = head->as_CountedLoop(); 2656 CountedLoopEndNode* cle = cl->loopexit(); 2657 Node* cle_out = cle->proj_out_or_null(false); 2658 if (use == cle_out) { 2659 IfNode* le = cl->outer_loop_end(); 2660 use = le->proj_out(false); 2661 use_loop = get_loop(use); 2662 if (mode == CloneIncludesStripMined) { 2663 nnn = old_new[le->_idx]; 2664 } else { 2665 newuse = old_new[cle_out->_idx]; 2666 } 2667 } 2668 } 2669 if (newuse == nullptr) { 2670 newuse = use->clone(); 2671 } 2672 2673 // Clone the loop exit control projection 2674 if (C->do_vector_loop() && cm != nullptr) { 2675 cm->verify_insert_and_clone(use, newuse, cm->clone_idx()); 2676 } 2677 newuse->set_req(0,nnn); 2678 _igvn.register_new_node_with_optimizer(newuse); 2679 set_loop(newuse, use_loop); 2680 set_idom(newuse, nnn, dom_depth(nnn) + 1 ); 2681 2682 // We need a Region to merge the exit from the peeled body and the 2683 // exit from the old loop body. 2684 RegionNode *r = new RegionNode(3); 2685 uint dd_r = MIN2(dom_depth(newuse), dom_depth(use)); 2686 assert(dd_r >= dom_depth(dom_lca(newuse, use)), "" ); 2687 2688 // The original user of 'use' uses 'r' instead. 2689 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) { 2690 Node* useuse = use->last_out(l); 2691 _igvn.rehash_node_delayed(useuse); 2692 uint uses_found = 0; 2693 if (useuse->in(0) == use) { 2694 useuse->set_req(0, r); 2695 uses_found++; 2696 if (useuse->is_CFG()) { 2697 // This is not a dom_depth > dd_r because when new 2698 // control flow is constructed by a loop opt, a node and 2699 // its dominator can end up at the same dom_depth 2700 assert(dom_depth(useuse) >= dd_r, ""); 2701 set_idom(useuse, r, dom_depth(useuse)); 2702 } 2703 } 2704 for (uint k = 1; k < useuse->req(); k++) { 2705 if( useuse->in(k) == use ) { 2706 useuse->set_req(k, r); 2707 uses_found++; 2708 if (useuse->is_Loop() && k == LoopNode::EntryControl) { 2709 // This is not a dom_depth > dd_r because when new 2710 // control flow is constructed by a loop opt, a node 2711 // and its dominator can end up at the same dom_depth 2712 assert(dom_depth(useuse) >= dd_r , ""); 2713 set_idom(useuse, r, dom_depth(useuse)); 2714 } 2715 } 2716 } 2717 l -= uses_found; // we deleted 1 or more copies of this edge 2718 } 2719 2720 assert(use->is_Proj(), "loop exit should be projection"); 2721 // lazy_replace() below moves all nodes that are: 2722 // - control dependent on the loop exit or 2723 // - have control set to the loop exit 2724 // below the post-loop merge point. lazy_replace() takes a dead control as first input. To make it 2725 // possible to use it, the loop exit projection is cloned and becomes the new exit projection. The initial one 2726 // becomes dead and is "replaced" by the region. 2727 Node* use_clone = use->clone(); 2728 register_control(use_clone, use_loop, idom(use), dom_depth(use)); 2729 // Now finish up 'r' 2730 r->set_req(1, newuse); 2731 r->set_req(2, use_clone); 2732 _igvn.register_new_node_with_optimizer(r); 2733 set_loop(r, use_loop); 2734 set_idom(r, (side_by_side_idom == nullptr) ? newuse->in(0) : side_by_side_idom, dd_r); 2735 lazy_replace(use, r); 2736 // Map the (cloned) old use to the new merge point 2737 old_new.map(use_clone->_idx, r); 2738 } // End of if a loop-exit test 2739 } 2740 } 2741 } 2742 2743 void PhaseIdealLoop::fix_body_edges(const Node_List &body, IdealLoopTree* loop, const Node_List &old_new, int dd, 2744 IdealLoopTree* parent, bool partial) { 2745 for(uint i = 0; i < body.size(); i++ ) { 2746 Node *old = body.at(i); 2747 Node *nnn = old_new[old->_idx]; 2748 // Fix CFG/Loop controlling the new node 2749 if (has_ctrl(old)) { 2750 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]); 2751 } else { 2752 set_loop(nnn, parent); 2753 if (old->outcnt() > 0) { 2754 Node* dom = idom(old); 2755 if (old_new[dom->_idx] != nullptr) { 2756 dom = old_new[dom->_idx]; 2757 set_idom(nnn, dom, dd ); 2758 } 2759 } 2760 } 2761 // Correct edges to the new node 2762 for (uint j = 0; j < nnn->req(); j++) { 2763 Node *n = nnn->in(j); 2764 if (n != nullptr) { 2765 IdealLoopTree *old_in_loop = get_loop(has_ctrl(n) ? get_ctrl(n) : n); 2766 if (loop->is_member(old_in_loop)) { 2767 if (old_new[n->_idx] != nullptr) { 2768 nnn->set_req(j, old_new[n->_idx]); 2769 } else { 2770 assert(!body.contains(n), ""); 2771 assert(partial, "node not cloned"); 2772 } 2773 } 2774 } 2775 } 2776 _igvn.hash_find_insert(nnn); 2777 } 2778 } 2779 2780 void PhaseIdealLoop::clone_loop_body(const Node_List& body, Node_List &old_new, CloneMap* cm) { 2781 for (uint i = 0; i < body.size(); i++) { 2782 Node* old = body.at(i); 2783 Node* nnn = old->clone(); 2784 old_new.map(old->_idx, nnn); 2785 if (C->do_vector_loop() && cm != nullptr) { 2786 cm->verify_insert_and_clone(old, nnn, cm->clone_idx()); 2787 } 2788 _igvn.register_new_node_with_optimizer(nnn); 2789 } 2790 } 2791 2792 2793 //---------------------- stride_of_possible_iv ------------------------------------- 2794 // Looks for an iff/bool/comp with one operand of the compare 2795 // being a cycle involving an add and a phi, 2796 // with an optional truncation (left-shift followed by a right-shift) 2797 // of the add. Returns zero if not an iv. 2798 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) { 2799 Node* trunc1 = nullptr; 2800 Node* trunc2 = nullptr; 2801 const TypeInteger* ttype = nullptr; 2802 if (!iff->is_If() || iff->in(1) == nullptr || !iff->in(1)->is_Bool()) { 2803 return 0; 2804 } 2805 BoolNode* bl = iff->in(1)->as_Bool(); 2806 Node* cmp = bl->in(1); 2807 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) { 2808 return 0; 2809 } 2810 // Must have an invariant operand 2811 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) { 2812 return 0; 2813 } 2814 Node* add2 = nullptr; 2815 Node* cmp1 = cmp->in(1); 2816 if (cmp1->is_Phi()) { 2817 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) ))) 2818 Node* phi = cmp1; 2819 for (uint i = 1; i < phi->req(); i++) { 2820 Node* in = phi->in(i); 2821 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in, 2822 &trunc1, &trunc2, &ttype, T_INT); 2823 if (add && add->in(1) == phi) { 2824 add2 = add->in(2); 2825 break; 2826 } 2827 } 2828 } else { 2829 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) ))) 2830 Node* addtrunc = cmp1; 2831 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc, 2832 &trunc1, &trunc2, &ttype, T_INT); 2833 if (add && add->in(1)->is_Phi()) { 2834 Node* phi = add->in(1); 2835 for (uint i = 1; i < phi->req(); i++) { 2836 if (phi->in(i) == addtrunc) { 2837 add2 = add->in(2); 2838 break; 2839 } 2840 } 2841 } 2842 } 2843 if (add2 != nullptr) { 2844 const TypeInt* add2t = _igvn.type(add2)->is_int(); 2845 if (add2t->is_con()) { 2846 return add2t->get_con(); 2847 } 2848 } 2849 return 0; 2850 } 2851 2852 2853 //---------------------- stay_in_loop ------------------------------------- 2854 // Return the (unique) control output node that's in the loop (if it exists.) 2855 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) { 2856 Node* unique = nullptr; 2857 if (!n) return nullptr; 2858 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2859 Node* use = n->fast_out(i); 2860 if (!has_ctrl(use) && loop->is_member(get_loop(use))) { 2861 if (unique != nullptr) { 2862 return nullptr; 2863 } 2864 unique = use; 2865 } 2866 } 2867 return unique; 2868 } 2869 2870 //------------------------------ register_node ------------------------------------- 2871 // Utility to register node "n" with PhaseIdealLoop 2872 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree* loop, Node* pred, uint ddepth) { 2873 _igvn.register_new_node_with_optimizer(n); 2874 loop->_body.push(n); 2875 if (n->is_CFG()) { 2876 set_loop(n, loop); 2877 set_idom(n, pred, ddepth); 2878 } else { 2879 set_ctrl(n, pred); 2880 } 2881 } 2882 2883 //------------------------------ proj_clone ------------------------------------- 2884 // Utility to create an if-projection 2885 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) { 2886 ProjNode* c = p->clone()->as_Proj(); 2887 c->set_req(0, iff); 2888 return c; 2889 } 2890 2891 //------------------------------ short_circuit_if ------------------------------------- 2892 // Force the iff control output to be the live_proj 2893 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) { 2894 guarantee(live_proj != nullptr, "null projection"); 2895 int proj_con = live_proj->_con; 2896 assert(proj_con == 0 || proj_con == 1, "false or true projection"); 2897 Node* con = intcon(proj_con); 2898 if (iff) { 2899 iff->set_req(1, con); 2900 } 2901 return con; 2902 } 2903 2904 //------------------------------ insert_if_before_proj ------------------------------------- 2905 // Insert a new if before an if projection (* - new node) 2906 // 2907 // before 2908 // if(test) 2909 // / \ 2910 // v v 2911 // other-proj proj (arg) 2912 // 2913 // after 2914 // if(test) 2915 // / \ 2916 // / v 2917 // | * proj-clone 2918 // v | 2919 // other-proj v 2920 // * new_if(relop(cmp[IU](left,right))) 2921 // / \ 2922 // v v 2923 // * new-proj proj 2924 // (returned) 2925 // 2926 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) { 2927 IfNode* iff = proj->in(0)->as_If(); 2928 IdealLoopTree *loop = get_loop(proj); 2929 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2930 uint ddepth = dom_depth(proj); 2931 2932 _igvn.rehash_node_delayed(iff); 2933 _igvn.rehash_node_delayed(proj); 2934 2935 proj->set_req(0, nullptr); // temporary disconnect 2936 ProjNode* proj2 = proj_clone(proj, iff); 2937 register_node(proj2, loop, iff, ddepth); 2938 2939 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right); 2940 register_node(cmp, loop, proj2, ddepth); 2941 2942 BoolNode* bol = new BoolNode(cmp, relop); 2943 register_node(bol, loop, proj2, ddepth); 2944 2945 int opcode = iff->Opcode(); 2946 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode"); 2947 IfNode* new_if = IfNode::make_with_same_profile(iff, proj2, bol); 2948 register_node(new_if, loop, proj2, ddepth); 2949 2950 proj->set_req(0, new_if); // reattach 2951 set_idom(proj, new_if, ddepth); 2952 2953 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj(); 2954 guarantee(new_exit != nullptr, "null exit node"); 2955 register_node(new_exit, get_loop(other_proj), new_if, ddepth); 2956 2957 return new_exit; 2958 } 2959 2960 //------------------------------ insert_region_before_proj ------------------------------------- 2961 // Insert a region before an if projection (* - new node) 2962 // 2963 // before 2964 // if(test) 2965 // / | 2966 // v | 2967 // proj v 2968 // other-proj 2969 // 2970 // after 2971 // if(test) 2972 // / | 2973 // v | 2974 // * proj-clone v 2975 // | other-proj 2976 // v 2977 // * new-region 2978 // | 2979 // v 2980 // * dum_if 2981 // / \ 2982 // v \ 2983 // * dum-proj v 2984 // proj 2985 // 2986 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) { 2987 IfNode* iff = proj->in(0)->as_If(); 2988 IdealLoopTree *loop = get_loop(proj); 2989 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2990 uint ddepth = dom_depth(proj); 2991 2992 _igvn.rehash_node_delayed(iff); 2993 _igvn.rehash_node_delayed(proj); 2994 2995 proj->set_req(0, nullptr); // temporary disconnect 2996 ProjNode* proj2 = proj_clone(proj, iff); 2997 register_node(proj2, loop, iff, ddepth); 2998 2999 RegionNode* reg = new RegionNode(2); 3000 reg->set_req(1, proj2); 3001 register_node(reg, loop, iff, ddepth); 3002 3003 IfNode* dum_if = new IfNode(reg, short_circuit_if(nullptr, proj), iff->_prob, iff->_fcnt); 3004 register_node(dum_if, loop, reg, ddepth); 3005 3006 proj->set_req(0, dum_if); // reattach 3007 set_idom(proj, dum_if, ddepth); 3008 3009 ProjNode* dum_proj = proj_clone(other_proj, dum_if); 3010 register_node(dum_proj, loop, dum_if, ddepth); 3011 3012 return reg; 3013 } 3014 3015 // Idea 3016 // ---- 3017 // Partial Peeling tries to rotate the loop in such a way that it can later be turned into a counted loop. Counted loops 3018 // require a signed loop exit test. When calling this method, we've only found a suitable unsigned test to partial peel 3019 // with. Therefore, we try to split off a signed loop exit test from the unsigned test such that it can be used as new 3020 // loop exit while keeping the unsigned test unchanged and preserving the same behavior as if we've used the unsigned 3021 // test alone instead: 3022 // 3023 // Before Partial Peeling: 3024 // Loop: 3025 // <peeled section> 3026 // Split off signed loop exit test 3027 // <-- CUT HERE --> 3028 // Unchanged unsigned loop exit test 3029 // <rest of unpeeled section> 3030 // goto Loop 3031 // 3032 // After Partial Peeling: 3033 // <cloned peeled section> 3034 // Cloned split off signed loop exit test 3035 // Loop: 3036 // Unchanged unsigned loop exit test 3037 // <rest of unpeeled section> 3038 // <peeled section> 3039 // Split off signed loop exit test 3040 // goto Loop 3041 // 3042 // Details 3043 // ------- 3044 // Before: 3045 // if (i <u limit) Unsigned loop exit condition 3046 // / | 3047 // v v 3048 // exit-proj stay-in-loop-proj 3049 // 3050 // Split off a signed loop exit test (i.e. with CmpI) from an unsigned loop exit test (i.e. with CmpU) and insert it 3051 // before the CmpU on the stay-in-loop path and keep both tests: 3052 // 3053 // if (i <u limit) Signed loop exit test 3054 // / | 3055 // / if (i <u limit) Unsigned loop exit test 3056 // / / | 3057 // v v v 3058 // exit-region stay-in-loop-proj 3059 // 3060 // Implementation 3061 // -------------- 3062 // We need to make sure that the new signed loop exit test is properly inserted into the graph such that the unsigned 3063 // loop exit test still dominates the same set of control nodes, the ctrl() relation from data nodes to both loop 3064 // exit tests is preserved, and their loop nesting is correct. 3065 // 3066 // To achieve that, we clone the unsigned loop exit test completely (leave it unchanged), insert the signed loop exit 3067 // test above it and kill the original unsigned loop exit test by setting it's condition to a constant 3068 // (i.e. stay-in-loop-const in graph below) such that IGVN can fold it later: 3069 // 3070 // if (stay-in-loop-const) Killed original unsigned loop exit test 3071 // / | 3072 // / v 3073 // / if (i < limit) Split off signed loop exit test 3074 // / / | 3075 // / / v 3076 // / / if (i <u limit) Cloned unsigned loop exit test 3077 // / / / | 3078 // v v v | 3079 // exit-region | 3080 // | | 3081 // dummy-if | 3082 // / | | 3083 // dead | | 3084 // v v 3085 // exit-proj stay-in-loop-proj 3086 // 3087 // Note: The dummy-if is inserted to create a region to merge the loop exits between the original to be killed unsigned 3088 // loop exit test and its exit projection while keeping the exit projection (also see insert_region_before_proj()). 3089 // 3090 // Requirements 3091 // ------------ 3092 // Note that we can only split off a signed loop exit test from the unsigned loop exit test when the behavior is exactly 3093 // the same as before with only a single unsigned test. This is only possible if certain requirements are met. 3094 // Otherwise, we need to bail out (see comments in the code below). 3095 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree* loop) { 3096 const bool Signed = true; 3097 const bool Unsigned = false; 3098 3099 BoolNode* bol = if_cmpu->in(1)->as_Bool(); 3100 if (bol->_test._test != BoolTest::lt) { 3101 return nullptr; 3102 } 3103 CmpNode* cmpu = bol->in(1)->as_Cmp(); 3104 assert(cmpu->Opcode() == Op_CmpU, "must be unsigned comparison"); 3105 3106 int stride = stride_of_possible_iv(if_cmpu); 3107 if (stride == 0) { 3108 return nullptr; 3109 } 3110 3111 Node* lp_proj = stay_in_loop(if_cmpu, loop); 3112 guarantee(lp_proj != nullptr, "null loop node"); 3113 3114 ProjNode* lp_continue = lp_proj->as_Proj(); 3115 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj(); 3116 if (!lp_exit->is_IfFalse()) { 3117 // The loop exit condition is (i <u limit) ==> (i >= 0 && i < limit). 3118 // We therefore can't add a single exit condition. 3119 return nullptr; 3120 } 3121 // The unsigned loop exit condition is 3122 // !(i <u limit) 3123 // = i >=u limit 3124 // 3125 // First, we note that for any x for which 3126 // 0 <= x <= INT_MAX 3127 // we can convert x to an unsigned int and still get the same guarantee: 3128 // 0 <= (uint) x <= INT_MAX = (uint) INT_MAX 3129 // 0 <=u (uint) x <=u INT_MAX = (uint) INT_MAX (LEMMA) 3130 // 3131 // With that in mind, if 3132 // limit >= 0 (COND) 3133 // then the unsigned loop exit condition 3134 // i >=u limit (ULE) 3135 // is equivalent to 3136 // i < 0 || i >= limit (SLE-full) 3137 // because either i is negative and therefore always greater than MAX_INT when converting to unsigned 3138 // (uint) i >=u MAX_INT >= limit >= 0 3139 // or otherwise 3140 // i >= limit >= 0 3141 // holds due to (LEMMA). 3142 // 3143 // For completeness, a counterexample with limit < 0: 3144 // Assume i = -3 and limit = -2: 3145 // i < 0 3146 // -2 < 0 3147 // is true and thus also "i < 0 || i >= limit". But 3148 // i >=u limit 3149 // -3 >=u -2 3150 // is false. 3151 Node* limit = cmpu->in(2); 3152 const TypeInt* type_limit = _igvn.type(limit)->is_int(); 3153 if (type_limit->_lo < 0) { 3154 return nullptr; 3155 } 3156 3157 // We prove below that we can extract a single signed loop exit condition from (SLE-full), depending on the stride: 3158 // stride < 0: 3159 // i < 0 (SLE = SLE-negative) 3160 // stride > 0: 3161 // i >= limit (SLE = SLE-positive) 3162 // such that we have the following graph before Partial Peeling with stride > 0 (similar for stride < 0): 3163 // 3164 // Loop: 3165 // <peeled section> 3166 // i >= limit (SLE-positive) 3167 // <-- CUT HERE --> 3168 // i >=u limit (ULE) 3169 // <rest of unpeeled section> 3170 // goto Loop 3171 // 3172 // We exit the loop if: 3173 // (SLE) is true OR (ULE) is true 3174 // However, if (SLE) is true then (ULE) also needs to be true to ensure the exact same behavior. Otherwise, we wrongly 3175 // exit a loop that should not have been exited if we did not apply Partial Peeling. More formally, we need to ensure: 3176 // (SLE) IMPLIES (ULE) 3177 // This indeed holds when (COND) is given: 3178 // - stride > 0: 3179 // i >= limit // (SLE = SLE-positive) 3180 // i >= limit >= 0 // (COND) 3181 // i >=u limit >= 0 // (LEMMA) 3182 // which is the unsigned loop exit condition (ULE). 3183 // - stride < 0: 3184 // i < 0 // (SLE = SLE-negative) 3185 // (uint) i >u MAX_INT // (NEG) all negative values are greater than MAX_INT when converted to unsigned 3186 // MAX_INT >= limit >= 0 // (COND) 3187 // MAX_INT >=u limit >= 0 // (LEMMA) 3188 // and thus from (NEG) and (LEMMA): 3189 // i >=u limit 3190 // which is the unsigned loop exit condition (ULE). 3191 // 3192 // 3193 // After Partial Peeling, we have the following structure for stride > 0 (similar for stride < 0): 3194 // <cloned peeled section> 3195 // i >= limit (SLE-positive) 3196 // Loop: 3197 // i >=u limit (ULE) 3198 // <rest of unpeeled section> 3199 // <peeled section> 3200 // i >= limit (SLE-positive) 3201 // goto Loop 3202 Node* rhs_cmpi; 3203 if (stride > 0) { 3204 rhs_cmpi = limit; // For i >= limit 3205 } else { 3206 rhs_cmpi = makecon(TypeInt::ZERO); // For i < 0 3207 } 3208 // Create a new region on the exit path 3209 RegionNode* reg = insert_region_before_proj(lp_exit); 3210 guarantee(reg != nullptr, "null region node"); 3211 3212 // Clone the if-cmpu-true-false using a signed compare 3213 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge; 3214 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, rhs_cmpi, lp_continue); 3215 reg->add_req(cmpi_exit); 3216 3217 // Clone the if-cmpu-true-false 3218 BoolTest::mask rel_u = bol->_test._test; 3219 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue); 3220 reg->add_req(cmpu_exit); 3221 3222 // Force original if to stay in loop. 3223 short_circuit_if(if_cmpu, lp_continue); 3224 3225 return cmpi_exit->in(0)->as_If(); 3226 } 3227 3228 //------------------------------ remove_cmpi_loop_exit ------------------------------------- 3229 // Remove a previously inserted signed compare loop exit. 3230 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) { 3231 Node* lp_proj = stay_in_loop(if_cmp, loop); 3232 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI && 3233 stay_in_loop(lp_proj, loop)->is_If() && 3234 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu"); 3235 Node* con = makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO); 3236 if_cmp->set_req(1, con); 3237 } 3238 3239 //------------------------------ scheduled_nodelist ------------------------------------- 3240 // Create a post order schedule of nodes that are in the 3241 // "member" set. The list is returned in "sched". 3242 // The first node in "sched" is the loop head, followed by 3243 // nodes which have no inputs in the "member" set, and then 3244 // followed by the nodes that have an immediate input dependence 3245 // on a node in "sched". 3246 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) { 3247 3248 assert(member.test(loop->_head->_idx), "loop head must be in member set"); 3249 VectorSet visited; 3250 Node_Stack nstack(loop->_body.size()); 3251 3252 Node* n = loop->_head; // top of stack is cached in "n" 3253 uint idx = 0; 3254 visited.set(n->_idx); 3255 3256 // Initially push all with no inputs from within member set 3257 for(uint i = 0; i < loop->_body.size(); i++ ) { 3258 Node *elt = loop->_body.at(i); 3259 if (member.test(elt->_idx)) { 3260 bool found = false; 3261 for (uint j = 0; j < elt->req(); j++) { 3262 Node* def = elt->in(j); 3263 if (def && member.test(def->_idx) && def != elt) { 3264 found = true; 3265 break; 3266 } 3267 } 3268 if (!found && elt != loop->_head) { 3269 nstack.push(n, idx); 3270 n = elt; 3271 assert(!visited.test(n->_idx), "not seen yet"); 3272 visited.set(n->_idx); 3273 } 3274 } 3275 } 3276 3277 // traverse out's that are in the member set 3278 while (true) { 3279 if (idx < n->outcnt()) { 3280 Node* use = n->raw_out(idx); 3281 idx++; 3282 if (!visited.test_set(use->_idx)) { 3283 if (member.test(use->_idx)) { 3284 nstack.push(n, idx); 3285 n = use; 3286 idx = 0; 3287 } 3288 } 3289 } else { 3290 // All outputs processed 3291 sched.push(n); 3292 if (nstack.is_empty()) break; 3293 n = nstack.node(); 3294 idx = nstack.index(); 3295 nstack.pop(); 3296 } 3297 } 3298 } 3299 3300 3301 //------------------------------ has_use_in_set ------------------------------------- 3302 // Has a use in the vector set 3303 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) { 3304 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3305 Node* use = n->fast_out(j); 3306 if (vset.test(use->_idx)) { 3307 return true; 3308 } 3309 } 3310 return false; 3311 } 3312 3313 3314 //------------------------------ has_use_internal_to_set ------------------------------------- 3315 // Has use internal to the vector set (ie. not in a phi at the loop head) 3316 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) { 3317 Node* head = loop->_head; 3318 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3319 Node* use = n->fast_out(j); 3320 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) { 3321 return true; 3322 } 3323 } 3324 return false; 3325 } 3326 3327 3328 //------------------------------ clone_for_use_outside_loop ------------------------------------- 3329 // clone "n" for uses that are outside of loop 3330 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) { 3331 int cloned = 0; 3332 assert(worklist.size() == 0, "should be empty"); 3333 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3334 Node* use = n->fast_out(j); 3335 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) { 3336 worklist.push(use); 3337 } 3338 } 3339 3340 if (C->check_node_count(worklist.size() + NodeLimitFudgeFactor, 3341 "Too many clones required in clone_for_use_outside_loop in partial peeling")) { 3342 return -1; 3343 } 3344 3345 while( worklist.size() ) { 3346 Node *use = worklist.pop(); 3347 if (!has_node(use) || use->in(0) == C->top()) continue; 3348 uint j; 3349 for (j = 0; j < use->req(); j++) { 3350 if (use->in(j) == n) break; 3351 } 3352 assert(j < use->req(), "must be there"); 3353 3354 // clone "n" and insert it between the inputs of "n" and the use outside the loop 3355 Node* n_clone = n->clone(); 3356 _igvn.replace_input_of(use, j, n_clone); 3357 cloned++; 3358 Node* use_c; 3359 if (!use->is_Phi()) { 3360 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0); 3361 } else { 3362 // Use in a phi is considered a use in the associated predecessor block 3363 use_c = use->in(0)->in(j); 3364 } 3365 set_ctrl(n_clone, use_c); 3366 assert(!loop->is_member(get_loop(use_c)), "should be outside loop"); 3367 get_loop(use_c)->_body.push(n_clone); 3368 _igvn.register_new_node_with_optimizer(n_clone); 3369 #ifndef PRODUCT 3370 if (TracePartialPeeling) { 3371 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx); 3372 } 3373 #endif 3374 } 3375 return cloned; 3376 } 3377 3378 3379 //------------------------------ clone_for_special_use_inside_loop ------------------------------------- 3380 // clone "n" for special uses that are in the not_peeled region. 3381 // If these def-uses occur in separate blocks, the code generator 3382 // marks the method as not compilable. For example, if a "BoolNode" 3383 // is in a different basic block than the "IfNode" that uses it, then 3384 // the compilation is aborted in the code generator. 3385 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 3386 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) { 3387 if (n->is_Phi() || n->is_Load()) { 3388 return; 3389 } 3390 assert(worklist.size() == 0, "should be empty"); 3391 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3392 Node* use = n->fast_out(j); 3393 if ( not_peel.test(use->_idx) && 3394 (use->is_If() || use->is_CMove() || use->is_Bool()) && 3395 use->in(1) == n) { 3396 worklist.push(use); 3397 } 3398 } 3399 if (worklist.size() > 0) { 3400 // clone "n" and insert it between inputs of "n" and the use 3401 Node* n_clone = n->clone(); 3402 loop->_body.push(n_clone); 3403 _igvn.register_new_node_with_optimizer(n_clone); 3404 set_ctrl(n_clone, get_ctrl(n)); 3405 sink_list.push(n_clone); 3406 not_peel.set(n_clone->_idx); 3407 #ifndef PRODUCT 3408 if (TracePartialPeeling) { 3409 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx); 3410 } 3411 #endif 3412 while( worklist.size() ) { 3413 Node *use = worklist.pop(); 3414 _igvn.rehash_node_delayed(use); 3415 for (uint j = 1; j < use->req(); j++) { 3416 if (use->in(j) == n) { 3417 use->set_req(j, n_clone); 3418 } 3419 } 3420 } 3421 } 3422 } 3423 3424 3425 //------------------------------ insert_phi_for_loop ------------------------------------- 3426 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 3427 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) { 3428 Node *phi = PhiNode::make(lp, back_edge_val); 3429 phi->set_req(LoopNode::EntryControl, lp_entry_val); 3430 // Use existing phi if it already exists 3431 Node *hit = _igvn.hash_find_insert(phi); 3432 if( hit == nullptr ) { 3433 _igvn.register_new_node_with_optimizer(phi); 3434 set_ctrl(phi, lp); 3435 } else { 3436 // Remove the new phi from the graph and use the hit 3437 _igvn.remove_dead_node(phi); 3438 phi = hit; 3439 } 3440 _igvn.replace_input_of(use, idx, phi); 3441 } 3442 3443 #ifdef ASSERT 3444 //------------------------------ is_valid_loop_partition ------------------------------------- 3445 // Validate the loop partition sets: peel and not_peel 3446 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, 3447 VectorSet& not_peel ) { 3448 uint i; 3449 // Check that peel_list entries are in the peel set 3450 for (i = 0; i < peel_list.size(); i++) { 3451 if (!peel.test(peel_list.at(i)->_idx)) { 3452 return false; 3453 } 3454 } 3455 // Check at loop members are in one of peel set or not_peel set 3456 for (i = 0; i < loop->_body.size(); i++ ) { 3457 Node *def = loop->_body.at(i); 3458 uint di = def->_idx; 3459 // Check that peel set elements are in peel_list 3460 if (peel.test(di)) { 3461 if (not_peel.test(di)) { 3462 return false; 3463 } 3464 // Must be in peel_list also 3465 bool found = false; 3466 for (uint j = 0; j < peel_list.size(); j++) { 3467 if (peel_list.at(j)->_idx == di) { 3468 found = true; 3469 break; 3470 } 3471 } 3472 if (!found) { 3473 return false; 3474 } 3475 } else if (not_peel.test(di)) { 3476 if (peel.test(di)) { 3477 return false; 3478 } 3479 } else { 3480 return false; 3481 } 3482 } 3483 return true; 3484 } 3485 3486 //------------------------------ is_valid_clone_loop_exit_use ------------------------------------- 3487 // Ensure a use outside of loop is of the right form 3488 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) { 3489 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 3490 return (use->is_Phi() && 3491 use_c->is_Region() && use_c->req() == 3 && 3492 (use_c->in(exit_idx)->Opcode() == Op_IfTrue || 3493 use_c->in(exit_idx)->Opcode() == Op_IfFalse || 3494 use_c->in(exit_idx)->Opcode() == Op_JumpProj) && 3495 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) ); 3496 } 3497 3498 //------------------------------ is_valid_clone_loop_form ------------------------------------- 3499 // Ensure that all uses outside of loop are of the right form 3500 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 3501 uint orig_exit_idx, uint clone_exit_idx) { 3502 uint len = peel_list.size(); 3503 for (uint i = 0; i < len; i++) { 3504 Node *def = peel_list.at(i); 3505 3506 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 3507 Node *use = def->fast_out(j); 3508 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 3509 if (!loop->is_member(get_loop(use_c))) { 3510 // use is not in the loop, check for correct structure 3511 if (use->in(0) == def) { 3512 // Okay 3513 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) { 3514 return false; 3515 } 3516 } 3517 } 3518 } 3519 return true; 3520 } 3521 #endif 3522 3523 //------------------------------ partial_peel ------------------------------------- 3524 // Partially peel (aka loop rotation) the top portion of a loop (called 3525 // the peel section below) by cloning it and placing one copy just before 3526 // the new loop head and the other copy at the bottom of the new loop. 3527 // 3528 // before after where it came from 3529 // 3530 // stmt1 stmt1 3531 // loop: stmt2 clone 3532 // stmt2 if condA goto exitA clone 3533 // if condA goto exitA new_loop: new 3534 // stmt3 stmt3 clone 3535 // if !condB goto loop if condB goto exitB clone 3536 // exitB: stmt2 orig 3537 // stmt4 if !condA goto new_loop orig 3538 // exitA: goto exitA 3539 // exitB: 3540 // stmt4 3541 // exitA: 3542 // 3543 // Step 1: find the cut point: an exit test on probable 3544 // induction variable. 3545 // Step 2: schedule (with cloning) operations in the peel 3546 // section that can be executed after the cut into 3547 // the section that is not peeled. This may need 3548 // to clone operations into exit blocks. For 3549 // instance, a reference to A[i] in the not-peel 3550 // section and a reference to B[i] in an exit block 3551 // may cause a left-shift of i by 2 to be placed 3552 // in the peel block. This step will clone the left 3553 // shift into the exit block and sink the left shift 3554 // from the peel to the not-peel section. 3555 // Step 3: clone the loop, retarget the control, and insert 3556 // phis for values that are live across the new loop 3557 // head. This is very dependent on the graph structure 3558 // from clone_loop. It creates region nodes for 3559 // exit control and associated phi nodes for values 3560 // flow out of the loop through that exit. The region 3561 // node is dominated by the clone's control projection. 3562 // So the clone's peel section is placed before the 3563 // new loop head, and the clone's not-peel section is 3564 // forms the top part of the new loop. The original 3565 // peel section forms the tail of the new loop. 3566 // Step 4: update the dominator tree and recompute the 3567 // dominator depth. 3568 // 3569 // orig 3570 // 3571 // stmt1 3572 // | 3573 // v 3574 // predicates 3575 // | 3576 // v 3577 // loop<----+ 3578 // | | 3579 // stmt2 | 3580 // | | 3581 // v | 3582 // ifA | 3583 // / | | 3584 // v v | 3585 // false true ^ <-- last_peel 3586 // / | | 3587 // / ===|==cut | 3588 // / stmt3 | <-- first_not_peel 3589 // / | | 3590 // | v | 3591 // v ifB | 3592 // exitA: / \ | 3593 // / \ | 3594 // v v | 3595 // false true | 3596 // / \ | 3597 // / ----+ 3598 // | 3599 // v 3600 // exitB: 3601 // stmt4 3602 // 3603 // 3604 // after clone loop 3605 // 3606 // stmt1 3607 // | 3608 // v 3609 // predicates 3610 // / \ 3611 // clone / \ orig 3612 // / \ 3613 // / \ 3614 // v v 3615 // +---->loop loop<----+ 3616 // | | | | 3617 // | stmt2 stmt2 | 3618 // | | | | 3619 // | v v | 3620 // | ifA ifA | 3621 // | | \ / | | 3622 // | v v v v | 3623 // ^ true false false true ^ <-- last_peel 3624 // | | ^ \ / | | 3625 // | cut==|== \ \ / ===|==cut | 3626 // | stmt3 \ \ / stmt3 | <-- first_not_peel 3627 // | | dom | | | | 3628 // | v \ 1v v2 v | 3629 // | ifB regionA ifB | 3630 // | / \ | / \ | 3631 // | / \ v / \ | 3632 // | v v exitA: v v | 3633 // | true false false true | 3634 // | / ^ \ / \ | 3635 // +---- \ \ / ----+ 3636 // dom \ / 3637 // \ 1v v2 3638 // regionB 3639 // | 3640 // v 3641 // exitB: 3642 // stmt4 3643 // 3644 // 3645 // after partial peel 3646 // 3647 // stmt1 3648 // | 3649 // v 3650 // predicates 3651 // / 3652 // clone / orig 3653 // / TOP 3654 // / \ 3655 // v v 3656 // TOP->loop loop----+ 3657 // | | | 3658 // stmt2 stmt2 | 3659 // | | | 3660 // v v | 3661 // ifA ifA | 3662 // | \ / | | 3663 // v v v v | 3664 // true false false true | <-- last_peel 3665 // | ^ \ / +------|---+ 3666 // +->newloop \ \ / === ==cut | | 3667 // | stmt3 \ \ / TOP | | 3668 // | | dom | | stmt3 | | <-- first_not_peel 3669 // | v \ 1v v2 v | | 3670 // | ifB regionA ifB ^ v 3671 // | / \ | / \ | | 3672 // | / \ v / \ | | 3673 // | v v exitA: v v | | 3674 // | true false false true | | 3675 // | / ^ \ / \ | | 3676 // | | \ \ / v | | 3677 // | | dom \ / TOP | | 3678 // | | \ 1v v2 | | 3679 // ^ v regionB | | 3680 // | | | | | 3681 // | | v ^ v 3682 // | | exitB: | | 3683 // | | stmt4 | | 3684 // | +------------>-----------------+ | 3685 // | | 3686 // +-----------------<---------------------+ 3687 // 3688 // 3689 // final graph 3690 // 3691 // stmt1 3692 // | 3693 // v 3694 // predicates 3695 // | 3696 // v 3697 // stmt2 clone 3698 // | 3699 // v 3700 // ........> ifA clone 3701 // : / | 3702 // dom / | 3703 // : v v 3704 // : false true 3705 // : | | 3706 // : | v 3707 // : | newloop<-----+ 3708 // : | | | 3709 // : | stmt3 clone | 3710 // : | | | 3711 // : | v | 3712 // : | ifB | 3713 // : | / \ | 3714 // : | v v | 3715 // : | false true | 3716 // : | | | | 3717 // : | v stmt2 | 3718 // : | exitB: | | 3719 // : | stmt4 v | 3720 // : | ifA orig | 3721 // : | / \ | 3722 // : | / \ | 3723 // : | v v | 3724 // : | false true | 3725 // : | / \ | 3726 // : v v -----+ 3727 // RegionA 3728 // | 3729 // v 3730 // exitA 3731 // 3732 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) { 3733 3734 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only"); 3735 if (!loop->_head->is_Loop()) { 3736 return false; 3737 } 3738 LoopNode *head = loop->_head->as_Loop(); 3739 3740 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) { 3741 return false; 3742 } 3743 3744 // Check for complex exit control 3745 for (uint ii = 0; ii < loop->_body.size(); ii++) { 3746 Node *n = loop->_body.at(ii); 3747 int opc = n->Opcode(); 3748 if (n->is_Call() || 3749 opc == Op_Catch || 3750 opc == Op_CatchProj || 3751 opc == Op_Jump || 3752 opc == Op_JumpProj) { 3753 #ifndef PRODUCT 3754 if (TracePartialPeeling) { 3755 tty->print_cr("\nExit control too complex: lp: %d", head->_idx); 3756 } 3757 #endif 3758 return false; 3759 } 3760 } 3761 3762 int dd = dom_depth(head); 3763 3764 // Step 1: find cut point 3765 3766 // Walk up dominators to loop head looking for first loop exit 3767 // which is executed on every path thru loop. 3768 IfNode *peel_if = nullptr; 3769 IfNode *peel_if_cmpu = nullptr; 3770 3771 Node *iff = loop->tail(); 3772 while (iff != head) { 3773 if (iff->is_If()) { 3774 Node *ctrl = get_ctrl(iff->in(1)); 3775 if (ctrl->is_top()) return false; // Dead test on live IF. 3776 // If loop-varying exit-test, check for induction variable 3777 if (loop->is_member(get_loop(ctrl)) && 3778 loop->is_loop_exit(iff) && 3779 is_possible_iv_test(iff)) { 3780 Node* cmp = iff->in(1)->in(1); 3781 if (cmp->Opcode() == Op_CmpI) { 3782 peel_if = iff->as_If(); 3783 } else { 3784 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU"); 3785 peel_if_cmpu = iff->as_If(); 3786 } 3787 } 3788 } 3789 iff = idom(iff); 3790 } 3791 3792 // Prefer signed compare over unsigned compare. 3793 IfNode* new_peel_if = nullptr; 3794 if (peel_if == nullptr) { 3795 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == nullptr) { 3796 return false; // No peel point found 3797 } 3798 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop); 3799 if (new_peel_if == nullptr) { 3800 return false; // No peel point found 3801 } 3802 peel_if = new_peel_if; 3803 } 3804 Node* last_peel = stay_in_loop(peel_if, loop); 3805 Node* first_not_peeled = stay_in_loop(last_peel, loop); 3806 if (first_not_peeled == nullptr || first_not_peeled == head) { 3807 return false; 3808 } 3809 3810 #ifndef PRODUCT 3811 if (TraceLoopOpts) { 3812 tty->print("PartialPeel "); 3813 loop->dump_head(); 3814 } 3815 3816 if (TracePartialPeeling) { 3817 tty->print_cr("before partial peel one iteration"); 3818 Node_List wl; 3819 Node* t = head->in(2); 3820 while (true) { 3821 wl.push(t); 3822 if (t == head) break; 3823 t = idom(t); 3824 } 3825 while (wl.size() > 0) { 3826 Node* tt = wl.pop(); 3827 tt->dump(); 3828 if (tt == last_peel) tty->print_cr("-- cut --"); 3829 } 3830 } 3831 #endif 3832 3833 C->print_method(PHASE_BEFORE_PARTIAL_PEELING, 4, head); 3834 3835 VectorSet peel; 3836 VectorSet not_peel; 3837 Node_List peel_list; 3838 Node_List worklist; 3839 Node_List sink_list; 3840 3841 uint estimate = loop->est_loop_clone_sz(1); 3842 if (exceeding_node_budget(estimate)) { 3843 return false; 3844 } 3845 3846 // Set of cfg nodes to peel are those that are executable from 3847 // the head through last_peel. 3848 assert(worklist.size() == 0, "should be empty"); 3849 worklist.push(head); 3850 peel.set(head->_idx); 3851 while (worklist.size() > 0) { 3852 Node *n = worklist.pop(); 3853 if (n != last_peel) { 3854 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3855 Node* use = n->fast_out(j); 3856 if (use->is_CFG() && 3857 loop->is_member(get_loop(use)) && 3858 !peel.test_set(use->_idx)) { 3859 worklist.push(use); 3860 } 3861 } 3862 } 3863 } 3864 3865 // Set of non-cfg nodes to peel are those that are control 3866 // dependent on the cfg nodes. 3867 for (uint i = 0; i < loop->_body.size(); i++) { 3868 Node *n = loop->_body.at(i); 3869 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n; 3870 if (peel.test(n_c->_idx)) { 3871 peel.set(n->_idx); 3872 } else { 3873 not_peel.set(n->_idx); 3874 } 3875 } 3876 3877 // Step 2: move operations from the peeled section down into the 3878 // not-peeled section 3879 3880 // Get a post order schedule of nodes in the peel region 3881 // Result in right-most operand. 3882 scheduled_nodelist(loop, peel, peel_list); 3883 3884 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 3885 3886 // For future check for too many new phis 3887 uint old_phi_cnt = 0; 3888 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 3889 Node* use = head->fast_out(j); 3890 if (use->is_Phi()) old_phi_cnt++; 3891 } 3892 3893 #ifndef PRODUCT 3894 if (TracePartialPeeling) { 3895 tty->print_cr("\npeeled list"); 3896 } 3897 #endif 3898 3899 // Evacuate nodes in peel region into the not_peeled region if possible 3900 bool too_many_clones = false; 3901 uint new_phi_cnt = 0; 3902 uint cloned_for_outside_use = 0; 3903 for (uint i = 0; i < peel_list.size();) { 3904 Node* n = peel_list.at(i); 3905 #ifndef PRODUCT 3906 if (TracePartialPeeling) n->dump(); 3907 #endif 3908 bool incr = true; 3909 if (!n->is_CFG()) { 3910 if (has_use_in_set(n, not_peel)) { 3911 // If not used internal to the peeled region, 3912 // move "n" from peeled to not_peeled region. 3913 if (!has_use_internal_to_set(n, peel, loop)) { 3914 // if not pinned and not a load (which maybe anti-dependent on a store) 3915 // and not a CMove (Matcher expects only bool->cmove). 3916 if (n->in(0) == nullptr && !n->is_Load() && !n->is_CMove()) { 3917 int new_clones = clone_for_use_outside_loop(loop, n, worklist); 3918 if (C->failing()) return false; 3919 if (new_clones == -1) { 3920 too_many_clones = true; 3921 break; 3922 } 3923 cloned_for_outside_use += new_clones; 3924 sink_list.push(n); 3925 peel.remove(n->_idx); 3926 not_peel.set(n->_idx); 3927 peel_list.remove(i); 3928 incr = false; 3929 #ifndef PRODUCT 3930 if (TracePartialPeeling) { 3931 tty->print_cr("sink to not_peeled region: %d newbb: %d", 3932 n->_idx, get_ctrl(n)->_idx); 3933 } 3934 #endif 3935 } 3936 } else { 3937 // Otherwise check for special def-use cases that span 3938 // the peel/not_peel boundary such as bool->if 3939 clone_for_special_use_inside_loop(loop, n, not_peel, sink_list, worklist); 3940 new_phi_cnt++; 3941 } 3942 } 3943 } 3944 if (incr) i++; 3945 } 3946 3947 estimate += cloned_for_outside_use + new_phi_cnt; 3948 bool exceed_node_budget = !may_require_nodes(estimate); 3949 bool exceed_phi_limit = new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta; 3950 3951 if (too_many_clones || exceed_node_budget || exceed_phi_limit) { 3952 #ifndef PRODUCT 3953 if (TracePartialPeeling && exceed_phi_limit) { 3954 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c", 3955 new_phi_cnt, old_phi_cnt, new_peel_if != nullptr?'T':'F'); 3956 } 3957 #endif 3958 if (new_peel_if != nullptr) { 3959 remove_cmpi_loop_exit(new_peel_if, loop); 3960 } 3961 // Inhibit more partial peeling on this loop 3962 assert(!head->is_partial_peel_loop(), "not partial peeled"); 3963 head->mark_partial_peel_failed(); 3964 if (cloned_for_outside_use > 0) { 3965 // Terminate this round of loop opts because 3966 // the graph outside this loop was changed. 3967 C->set_major_progress(); 3968 return true; 3969 } 3970 return false; 3971 } 3972 3973 // Step 3: clone loop, retarget control, and insert new phis 3974 3975 // Create new loop head for new phis and to hang 3976 // the nodes being moved (sinked) from the peel region. 3977 LoopNode* new_head = new LoopNode(last_peel, last_peel); 3978 new_head->set_unswitch_count(head->unswitch_count()); // Preserve 3979 _igvn.register_new_node_with_optimizer(new_head); 3980 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled"); 3981 _igvn.replace_input_of(first_not_peeled, 0, new_head); 3982 set_loop(new_head, loop); 3983 loop->_body.push(new_head); 3984 not_peel.set(new_head->_idx); 3985 set_idom(new_head, last_peel, dom_depth(first_not_peeled)); 3986 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled)); 3987 3988 while (sink_list.size() > 0) { 3989 Node* n = sink_list.pop(); 3990 set_ctrl(n, new_head); 3991 } 3992 3993 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 3994 3995 clone_loop(loop, old_new, dd, IgnoreStripMined); 3996 3997 const uint clone_exit_idx = 1; 3998 const uint orig_exit_idx = 2; 3999 assert(is_valid_clone_loop_form(loop, peel_list, orig_exit_idx, clone_exit_idx), "bad clone loop"); 4000 4001 Node* head_clone = old_new[head->_idx]; 4002 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop(); 4003 Node* orig_tail_clone = head_clone->in(2); 4004 4005 // Add phi if "def" node is in peel set and "use" is not 4006 4007 for (uint i = 0; i < peel_list.size(); i++) { 4008 Node *def = peel_list.at(i); 4009 if (!def->is_CFG()) { 4010 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 4011 Node *use = def->fast_out(j); 4012 if (has_node(use) && use->in(0) != C->top() && 4013 (!peel.test(use->_idx) || 4014 (use->is_Phi() && use->in(0) == head)) ) { 4015 worklist.push(use); 4016 } 4017 } 4018 while( worklist.size() ) { 4019 Node *use = worklist.pop(); 4020 for (uint j = 1; j < use->req(); j++) { 4021 Node* n = use->in(j); 4022 if (n == def) { 4023 4024 // "def" is in peel set, "use" is not in peel set 4025 // or "use" is in the entry boundary (a phi) of the peel set 4026 4027 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use; 4028 4029 if ( loop->is_member(get_loop( use_c )) ) { 4030 // use is in loop 4031 if (old_new[use->_idx] != nullptr) { // null for dead code 4032 Node* use_clone = old_new[use->_idx]; 4033 _igvn.replace_input_of(use, j, C->top()); 4034 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone ); 4035 } 4036 } else { 4037 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format"); 4038 // use is not in the loop, check if the live range includes the cut 4039 Node* lp_if = use_c->in(orig_exit_idx)->in(0); 4040 if (not_peel.test(lp_if->_idx)) { 4041 assert(j == orig_exit_idx, "use from original loop"); 4042 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone ); 4043 } 4044 } 4045 } 4046 } 4047 } 4048 } 4049 } 4050 4051 // Step 3b: retarget control 4052 4053 // Redirect control to the new loop head if a cloned node in 4054 // the not_peeled region has control that points into the peeled region. 4055 // This necessary because the cloned peeled region will be outside 4056 // the loop. 4057 // from to 4058 // cloned-peeled <---+ 4059 // new_head_clone: | <--+ 4060 // cloned-not_peeled in(0) in(0) 4061 // orig-peeled 4062 4063 for (uint i = 0; i < loop->_body.size(); i++) { 4064 Node *n = loop->_body.at(i); 4065 if (!n->is_CFG() && n->in(0) != nullptr && 4066 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) { 4067 Node* n_clone = old_new[n->_idx]; 4068 if (n_clone->depends_only_on_test()) { 4069 // Pin array access nodes: control is updated here to the loop head. If, after some transformations, the 4070 // backedge is removed, an array load could become dependent on a condition that's not a range check for that 4071 // access. If that condition is replaced by an identical dominating one, then an unpinned load would risk 4072 // floating above its range check. 4073 Node* pinned_clone = n_clone->pin_array_access_node(); 4074 if (pinned_clone != nullptr) { 4075 register_new_node_with_ctrl_of(pinned_clone, n_clone); 4076 old_new.map(n->_idx, pinned_clone); 4077 _igvn.replace_node(n_clone, pinned_clone); 4078 n_clone = pinned_clone; 4079 } 4080 } 4081 _igvn.replace_input_of(n_clone, 0, new_head_clone); 4082 } 4083 } 4084 4085 // Backedge of the surviving new_head (the clone) is original last_peel 4086 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel); 4087 4088 // Cut first node in original not_peel set 4089 _igvn.rehash_node_delayed(new_head); // Multiple edge updates: 4090 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of 4091 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls 4092 4093 // Copy head_clone back-branch info to original head 4094 // and remove original head's loop entry and 4095 // clone head's back-branch 4096 _igvn.rehash_node_delayed(head); // Multiple edge updates 4097 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl)); 4098 head->set_req(LoopNode::LoopBackControl, C->top()); 4099 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top()); 4100 4101 // Similarly modify the phis 4102 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) { 4103 Node* use = head->fast_out(k); 4104 if (use->is_Phi() && use->outcnt() > 0) { 4105 Node* use_clone = old_new[use->_idx]; 4106 _igvn.rehash_node_delayed(use); // Multiple edge updates 4107 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl)); 4108 use->set_req(LoopNode::LoopBackControl, C->top()); 4109 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top()); 4110 } 4111 } 4112 4113 // Step 4: update dominator tree and dominator depth 4114 4115 set_idom(head, orig_tail_clone, dd); 4116 recompute_dom_depth(); 4117 4118 // Inhibit more partial peeling on this loop 4119 new_head_clone->set_partial_peel_loop(); 4120 C->set_major_progress(); 4121 loop->record_for_igvn(); 4122 4123 #ifndef PRODUCT 4124 if (TracePartialPeeling) { 4125 tty->print_cr("\nafter partial peel one iteration"); 4126 Node_List wl; 4127 Node* t = last_peel; 4128 while (true) { 4129 wl.push(t); 4130 if (t == head_clone) break; 4131 t = idom(t); 4132 } 4133 while (wl.size() > 0) { 4134 Node* tt = wl.pop(); 4135 if (tt == head) tty->print_cr("orig head"); 4136 else if (tt == new_head_clone) tty->print_cr("new head"); 4137 else if (tt == head_clone) tty->print_cr("clone head"); 4138 tt->dump(); 4139 } 4140 } 4141 #endif 4142 4143 C->print_method(PHASE_AFTER_PARTIAL_PEELING, 4, new_head_clone); 4144 4145 return true; 4146 } 4147 4148 // Transform: 4149 // 4150 // loop<-----------------+ 4151 // | | 4152 // stmt1 stmt2 .. stmtn | 4153 // | | | | 4154 // \ | / | 4155 // v v v | 4156 // region | 4157 // | | 4158 // shared_stmt | 4159 // | | 4160 // v | 4161 // if | 4162 // / \ | 4163 // | -----------+ 4164 // v 4165 // 4166 // into: 4167 // 4168 // loop<-------------------+ 4169 // | | 4170 // v | 4171 // +->loop | 4172 // | | | 4173 // | stmt1 stmt2 .. stmtn | 4174 // | | | | | 4175 // | | \ / | 4176 // | | v v | 4177 // | | region1 | 4178 // | | | | 4179 // | shared_stmt shared_stmt | 4180 // | | | | 4181 // | v v | 4182 // | if if | 4183 // | /\ / \ | 4184 // +-- | | -------+ 4185 // \ / 4186 // v v 4187 // region2 4188 // 4189 // (region2 is shown to merge mirrored projections of the loop exit 4190 // ifs to make the diagram clearer but they really merge the same 4191 // projection) 4192 // 4193 // Conditions for this transformation to trigger: 4194 // - the path through stmt1 is frequent enough 4195 // - the inner loop will be turned into a counted loop after transformation 4196 bool PhaseIdealLoop::duplicate_loop_backedge(IdealLoopTree *loop, Node_List &old_new) { 4197 if (!DuplicateBackedge) { 4198 return false; 4199 } 4200 assert(!loop->_head->is_CountedLoop() || StressDuplicateBackedge, "Non-counted loop only"); 4201 if (!loop->_head->is_Loop()) { 4202 return false; 4203 } 4204 4205 uint estimate = loop->est_loop_clone_sz(1); 4206 if (exceeding_node_budget(estimate)) { 4207 return false; 4208 } 4209 4210 LoopNode *head = loop->_head->as_Loop(); 4211 4212 Node* region = nullptr; 4213 IfNode* exit_test = nullptr; 4214 uint inner; 4215 float f; 4216 if (StressDuplicateBackedge) { 4217 if (head->is_strip_mined()) { 4218 return false; 4219 } 4220 Node* c = head->in(LoopNode::LoopBackControl); 4221 4222 while (c != head) { 4223 if (c->is_Region()) { 4224 region = c; 4225 } 4226 c = idom(c); 4227 } 4228 4229 if (region == nullptr) { 4230 return false; 4231 } 4232 4233 inner = 1; 4234 } else { 4235 // Is the shape of the loop that of a counted loop... 4236 Node* back_control = loop_exit_control(head, loop); 4237 if (back_control == nullptr) { 4238 return false; 4239 } 4240 4241 BoolTest::mask bt = BoolTest::illegal; 4242 float cl_prob = 0; 4243 Node* incr = nullptr; 4244 Node* limit = nullptr; 4245 Node* cmp = loop_exit_test(back_control, loop, incr, limit, bt, cl_prob); 4246 if (cmp == nullptr || cmp->Opcode() != Op_CmpI) { 4247 return false; 4248 } 4249 4250 // With an extra phi for the candidate iv? 4251 // Or the region node is the loop head 4252 if (!incr->is_Phi() || incr->in(0) == head) { 4253 return false; 4254 } 4255 4256 PathFrequency pf(head, this); 4257 region = incr->in(0); 4258 4259 // Go over all paths for the extra phi's region and see if that 4260 // path is frequent enough and would match the expected iv shape 4261 // if the extra phi is removed 4262 inner = 0; 4263 for (uint i = 1; i < incr->req(); ++i) { 4264 Node* in = incr->in(i); 4265 Node* trunc1 = nullptr; 4266 Node* trunc2 = nullptr; 4267 const TypeInteger* iv_trunc_t = nullptr; 4268 Node* orig_in = in; 4269 if (!(in = CountedLoopNode::match_incr_with_optional_truncation(in, &trunc1, &trunc2, &iv_trunc_t, T_INT))) { 4270 continue; 4271 } 4272 assert(in->Opcode() == Op_AddI, "wrong increment code"); 4273 Node* xphi = nullptr; 4274 Node* stride = loop_iv_stride(in, loop, xphi); 4275 4276 if (stride == nullptr) { 4277 continue; 4278 } 4279 4280 PhiNode* phi = loop_iv_phi(xphi, nullptr, head, loop); 4281 if (phi == nullptr || 4282 (trunc1 == nullptr && phi->in(LoopNode::LoopBackControl) != incr) || 4283 (trunc1 != nullptr && phi->in(LoopNode::LoopBackControl) != trunc1)) { 4284 return false; 4285 } 4286 4287 f = pf.to(region->in(i)); 4288 if (f > 0.5) { 4289 inner = i; 4290 break; 4291 } 4292 } 4293 4294 if (inner == 0) { 4295 return false; 4296 } 4297 4298 exit_test = back_control->in(0)->as_If(); 4299 } 4300 4301 if (idom(region)->is_Catch()) { 4302 return false; 4303 } 4304 4305 // Collect all control nodes that need to be cloned (shared_stmt in the diagram) 4306 Unique_Node_List wq; 4307 wq.push(head->in(LoopNode::LoopBackControl)); 4308 for (uint i = 0; i < wq.size(); i++) { 4309 Node* c = wq.at(i); 4310 assert(get_loop(c) == loop, "not in the right loop?"); 4311 if (c->is_Region()) { 4312 if (c != region) { 4313 for (uint j = 1; j < c->req(); ++j) { 4314 wq.push(c->in(j)); 4315 } 4316 } 4317 } else { 4318 wq.push(c->in(0)); 4319 } 4320 assert(!is_strict_dominator(c, region), "shouldn't go above region"); 4321 } 4322 4323 Node* region_dom = idom(region); 4324 4325 // Can't do the transformation if this would cause a membar pair to 4326 // be split 4327 for (uint i = 0; i < wq.size(); i++) { 4328 Node* c = wq.at(i); 4329 if (c->is_MemBar() && (c->as_MemBar()->trailing_store() || c->as_MemBar()->trailing_load_store())) { 4330 assert(c->as_MemBar()->leading_membar()->trailing_membar() == c, "bad membar pair"); 4331 if (!wq.member(c->as_MemBar()->leading_membar())) { 4332 return false; 4333 } 4334 } 4335 } 4336 4337 // Collect data nodes that need to be clones as well 4338 int dd = dom_depth(head); 4339 4340 for (uint i = 0; i < loop->_body.size(); ++i) { 4341 Node* n = loop->_body.at(i); 4342 if (has_ctrl(n)) { 4343 Node* c = get_ctrl(n); 4344 if (wq.member(c)) { 4345 wq.push(n); 4346 } 4347 } else { 4348 set_idom(n, idom(n), dd); 4349 } 4350 } 4351 4352 // clone shared_stmt 4353 clone_loop_body(wq, old_new, nullptr); 4354 4355 Node* region_clone = old_new[region->_idx]; 4356 region_clone->set_req(inner, C->top()); 4357 set_idom(region, region->in(inner), dd); 4358 4359 // Prepare the outer loop 4360 Node* outer_head = new LoopNode(head->in(LoopNode::EntryControl), old_new[head->in(LoopNode::LoopBackControl)->_idx]); 4361 register_control(outer_head, loop->_parent, outer_head->in(LoopNode::EntryControl)); 4362 _igvn.replace_input_of(head, LoopNode::EntryControl, outer_head); 4363 set_idom(head, outer_head, dd); 4364 4365 fix_body_edges(wq, loop, old_new, dd, loop->_parent, true); 4366 4367 // Make one of the shared_stmt copies only reachable from stmt1, the 4368 // other only from stmt2..stmtn. 4369 Node* dom = nullptr; 4370 for (uint i = 1; i < region->req(); ++i) { 4371 if (i != inner) { 4372 _igvn.replace_input_of(region, i, C->top()); 4373 } 4374 Node* in = region_clone->in(i); 4375 if (in->is_top()) { 4376 continue; 4377 } 4378 if (dom == nullptr) { 4379 dom = in; 4380 } else { 4381 dom = dom_lca(dom, in); 4382 } 4383 } 4384 4385 set_idom(region_clone, dom, dd); 4386 4387 // Set up the outer loop 4388 for (uint i = 0; i < head->outcnt(); i++) { 4389 Node* u = head->raw_out(i); 4390 if (u->is_Phi()) { 4391 Node* outer_phi = u->clone(); 4392 outer_phi->set_req(0, outer_head); 4393 Node* backedge = old_new[u->in(LoopNode::LoopBackControl)->_idx]; 4394 if (backedge == nullptr) { 4395 backedge = u->in(LoopNode::LoopBackControl); 4396 } 4397 outer_phi->set_req(LoopNode::LoopBackControl, backedge); 4398 register_new_node(outer_phi, outer_head); 4399 _igvn.replace_input_of(u, LoopNode::EntryControl, outer_phi); 4400 } 4401 } 4402 4403 // create control and data nodes for out of loop uses (including region2) 4404 Node_List worklist; 4405 uint new_counter = C->unique(); 4406 fix_ctrl_uses(wq, loop, old_new, ControlAroundStripMined, outer_head, nullptr, worklist); 4407 4408 Node_List *split_if_set = nullptr; 4409 Node_List *split_bool_set = nullptr; 4410 Node_List *split_cex_set = nullptr; 4411 fix_data_uses(wq, loop, ControlAroundStripMined, loop->skip_strip_mined(), new_counter, old_new, worklist, 4412 split_if_set, split_bool_set, split_cex_set); 4413 4414 finish_clone_loop(split_if_set, split_bool_set, split_cex_set); 4415 4416 if (exit_test != nullptr) { 4417 float cnt = exit_test->_fcnt; 4418 if (cnt != COUNT_UNKNOWN) { 4419 exit_test->_fcnt = cnt * f; 4420 old_new[exit_test->_idx]->as_If()->_fcnt = cnt * (1 - f); 4421 } 4422 } 4423 4424 C->set_major_progress(); 4425 4426 return true; 4427 } 4428 4429 // AutoVectorize the loop: replace scalar ops with vector ops. 4430 PhaseIdealLoop::AutoVectorizeStatus 4431 PhaseIdealLoop::auto_vectorize(IdealLoopTree* lpt, VSharedData &vshared) { 4432 // Counted loop only 4433 if (!lpt->is_counted()) { 4434 return AutoVectorizeStatus::Impossible; 4435 } 4436 4437 // Main-loop only 4438 CountedLoopNode* cl = lpt->_head->as_CountedLoop(); 4439 if (!cl->is_main_loop()) { 4440 return AutoVectorizeStatus::Impossible; 4441 } 4442 4443 VLoop vloop(lpt, false); 4444 if (!vloop.check_preconditions()) { 4445 return AutoVectorizeStatus::TriedAndFailed; 4446 } 4447 4448 // Ensure the shared data is cleared before each use 4449 vshared.clear(); 4450 4451 const VLoopAnalyzer vloop_analyzer(vloop, vshared); 4452 if (!vloop_analyzer.success()) { 4453 return AutoVectorizeStatus::TriedAndFailed; 4454 } 4455 4456 SuperWord sw(vloop_analyzer); 4457 if (!sw.transform_loop()) { 4458 return AutoVectorizeStatus::TriedAndFailed; 4459 } 4460 4461 return AutoVectorizeStatus::Success; 4462 } 4463 4464 // Just before insert_pre_post_loops, we can multiversion the loop: 4465 // 4466 // multiversion_if 4467 // | | 4468 // fast_loop slow_loop 4469 // 4470 // In the fast_loop we can make speculative assumptions, and put the 4471 // conditions into the multiversion_if. If the conditions hold at runtime, 4472 // we enter the fast_loop, if the conditions fail, we take the slow_loop 4473 // instead which does not make any of the speculative assumptions. 4474 // 4475 // Note: we only multiversion the loop if the loop does not have any 4476 // auto vectorization check Predicate. If we have that predicate, 4477 // then we can simply add the speculative assumption checks to 4478 // that Predicate. This means we do not need to duplicate the 4479 // loop - we have a smaller graph and save compile time. Should 4480 // the conditions ever fail, then we deopt / trap at the Predicate 4481 // and recompile without that Predicate. At that point we will 4482 // multiversion the loop, so that we can still have speculative 4483 // runtime checks. 4484 // 4485 // We perform the multiversioning when the loop is still in its single 4486 // iteration form, even before we insert pre and post loops. This makes 4487 // the cloning much simpler. However, this means that both the fast 4488 // and the slow loop have to be optimized independently (adding pre 4489 // and post loops, unrolling the main loop, auto-vectorize etc.). And 4490 // we may end up not needing any speculative assumptions in the fast_loop 4491 // and then rejecting the slow_loop by constant folding the multiversion_if. 4492 // 4493 // Therefore, we "delay" the optimization of the slow_loop until we add 4494 // at least one speculative assumption for the fast_loop. If we never 4495 // add such a speculative runtime check, the OpaqueMultiversioningNode 4496 // of the multiversion_if constant folds to true after loop opts, and the 4497 // multiversion_if folds away the "delayed" slow_loop. If we add any 4498 // speculative assumption, then we notify the OpaqueMultiversioningNode 4499 // with "notify_slow_loop_that_it_can_resume_optimizations". 4500 // 4501 // Note: new runtime checks can be added to the multiversion_if with 4502 // PhaseIdealLoop::create_new_if_for_multiversion 4503 void PhaseIdealLoop::maybe_multiversion_for_auto_vectorization_runtime_checks(IdealLoopTree* lpt, Node_List& old_new) { 4504 CountedLoopNode* cl = lpt->_head->as_CountedLoop(); 4505 LoopNode* outer_loop = cl->skip_strip_mined(); 4506 Node* entry = outer_loop->in(LoopNode::EntryControl); 4507 4508 // Check we have multiversioning enabled, and are not already multiversioned. 4509 if (!LoopMultiversioning || cl->is_multiversion()) { return; } 4510 4511 // Check that we do not have a parse-predicate where we can add the runtime checks 4512 // during auto-vectorization. 4513 const Predicates predicates(entry); 4514 const PredicateBlock* predicate_block = predicates.auto_vectorization_check_block(); 4515 if (predicate_block->has_parse_predicate()) { return; } 4516 4517 // Check node budget. 4518 uint estimate = lpt->est_loop_clone_sz(2); 4519 if (!may_require_nodes(estimate)) { return; } 4520 4521 do_multiversioning(lpt, old_new); 4522 } 4523 4524 // Returns true if the Reduction node is unordered. 4525 static bool is_unordered_reduction(Node* n) { 4526 return n->is_Reduction() && !n->as_Reduction()->requires_strict_order(); 4527 } 4528 4529 // Having ReductionNodes in the loop is expensive. They need to recursively 4530 // fold together the vector values, for every vectorized loop iteration. If 4531 // we encounter the following pattern, we can vector accumulate the values 4532 // inside the loop, and only have a single UnorderedReduction after the loop. 4533 // 4534 // Note: UnorderedReduction represents a ReductionNode which does not require 4535 // calculating in strict order. 4536 // 4537 // CountedLoop init 4538 // | | 4539 // +------+ | +-----------------------+ 4540 // | | | | 4541 // PhiNode (s) | 4542 // | | 4543 // | Vector | 4544 // | | | 4545 // UnorderedReduction (first_ur) | 4546 // | | 4547 // ... Vector | 4548 // | | | 4549 // UnorderedReduction (last_ur) | 4550 // | | 4551 // +---------------------+ 4552 // 4553 // We patch the graph to look like this: 4554 // 4555 // CountedLoop identity_vector 4556 // | | 4557 // +-------+ | +---------------+ 4558 // | | | | 4559 // PhiNode (v) | 4560 // | | 4561 // | Vector | 4562 // | | | 4563 // VectorAccumulator | 4564 // | | 4565 // ... Vector | 4566 // | | | 4567 // init VectorAccumulator | 4568 // | | | | 4569 // UnorderedReduction +-----------+ 4570 // 4571 // We turned the scalar (s) Phi into a vectorized one (v). In the loop, we 4572 // use vector_accumulators, which do the same reductions, but only element 4573 // wise. This is a single operation per vector_accumulator, rather than many 4574 // for a UnorderedReduction. We can then reduce the last vector_accumulator 4575 // after the loop, and also reduce the init value into it. 4576 // 4577 // We can not do this with all reductions. Some reductions do not allow the 4578 // reordering of operations (for example float addition/multiplication require 4579 // strict order). 4580 void PhaseIdealLoop::move_unordered_reduction_out_of_loop(IdealLoopTree* loop) { 4581 assert(!C->major_progress() && loop->is_counted() && loop->is_innermost(), "sanity"); 4582 4583 // Find all Phi nodes with an unordered Reduction on backedge. 4584 CountedLoopNode* cl = loop->_head->as_CountedLoop(); 4585 for (DUIterator_Fast jmax, j = cl->fast_outs(jmax); j < jmax; j++) { 4586 Node* phi = cl->fast_out(j); 4587 // We have a phi with a single use, and an unordered Reduction on the backedge. 4588 if (!phi->is_Phi() || phi->outcnt() != 1 || !is_unordered_reduction(phi->in(2))) { 4589 continue; 4590 } 4591 4592 ReductionNode* last_ur = phi->in(2)->as_Reduction(); 4593 assert(!last_ur->requires_strict_order(), "must be"); 4594 4595 // Determine types 4596 const TypeVect* vec_t = last_ur->vect_type(); 4597 uint vector_length = vec_t->length(); 4598 BasicType bt = vec_t->element_basic_type(); 4599 4600 // Convert opcode from vector-reduction -> scalar -> normal-vector-op 4601 const int sopc = VectorNode::scalar_opcode(last_ur->Opcode(), bt); 4602 const int vopc = VectorNode::opcode(sopc, bt); 4603 if (!Matcher::match_rule_supported_vector(vopc, vector_length, bt)) { 4604 DEBUG_ONLY( last_ur->dump(); ) 4605 assert(false, "do not have normal vector op for this reduction"); 4606 continue; // not implemented -> fails 4607 } 4608 4609 // Traverse up the chain of unordered Reductions, checking that it loops back to 4610 // the phi. Check that all unordered Reductions only have a single use, except for 4611 // the last (last_ur), which only has phi as a use in the loop, and all other uses 4612 // are outside the loop. 4613 ReductionNode* current = last_ur; 4614 ReductionNode* first_ur = nullptr; 4615 while (true) { 4616 assert(!current->requires_strict_order(), "sanity"); 4617 4618 // Expect no ctrl and a vector_input from within the loop. 4619 Node* ctrl = current->in(0); 4620 Node* vector_input = current->in(2); 4621 if (ctrl != nullptr || get_ctrl(vector_input) != cl) { 4622 DEBUG_ONLY( current->dump(1); ) 4623 assert(false, "reduction has ctrl or bad vector_input"); 4624 break; // Chain traversal fails. 4625 } 4626 4627 assert(current->vect_type() != nullptr, "must have vector type"); 4628 if (current->vect_type() != last_ur->vect_type()) { 4629 // Reductions do not have the same vector type (length and element type). 4630 break; // Chain traversal fails. 4631 } 4632 4633 // Expect single use of an unordered Reduction, except for last_ur. 4634 if (current == last_ur) { 4635 // Expect all uses to be outside the loop, except phi. 4636 for (DUIterator_Fast kmax, k = current->fast_outs(kmax); k < kmax; k++) { 4637 Node* use = current->fast_out(k); 4638 if (use != phi && ctrl_or_self(use) == cl) { 4639 DEBUG_ONLY( current->dump(-1); ) 4640 assert(false, "reduction has use inside loop"); 4641 // Should not be allowed by SuperWord::mark_reductions 4642 return; // bail out of optimization 4643 } 4644 } 4645 } else { 4646 if (current->outcnt() != 1) { 4647 break; // Chain traversal fails. 4648 } 4649 } 4650 4651 // Expect another unordered Reduction or phi as the scalar input. 4652 Node* scalar_input = current->in(1); 4653 if (is_unordered_reduction(scalar_input) && 4654 scalar_input->Opcode() == current->Opcode()) { 4655 // Move up the unordered Reduction chain. 4656 current = scalar_input->as_Reduction(); 4657 assert(!current->requires_strict_order(), "must be"); 4658 } else if (scalar_input == phi) { 4659 // Chain terminates at phi. 4660 first_ur = current; 4661 current = nullptr; 4662 break; // Success. 4663 } else { 4664 // scalar_input is neither phi nor a matching reduction 4665 // Can for example be scalar reduction when we have 4666 // partial vectorization. 4667 break; // Chain traversal fails. 4668 } 4669 } 4670 if (current != nullptr) { 4671 // Chain traversal was not successful. 4672 continue; 4673 } 4674 assert(first_ur != nullptr, "must have successfully terminated chain traversal"); 4675 4676 Node* identity_scalar = ReductionNode::make_identity_con_scalar(_igvn, sopc, bt); 4677 set_root_as_ctrl(identity_scalar); 4678 VectorNode* identity_vector = VectorNode::scalar2vector(identity_scalar, vector_length, bt); 4679 register_new_node(identity_vector, C->root()); 4680 assert(vec_t == identity_vector->vect_type(), "matching vector type"); 4681 VectorNode::trace_new_vector(identity_vector, "Unordered Reduction"); 4682 4683 // Turn the scalar phi into a vector phi. 4684 _igvn.rehash_node_delayed(phi); 4685 Node* init = phi->in(1); // Remember init before replacing it. 4686 phi->set_req_X(1, identity_vector, &_igvn); 4687 phi->as_Type()->set_type(vec_t); 4688 _igvn.set_type(phi, vec_t); 4689 4690 // Traverse down the chain of unordered Reductions, and replace them with vector_accumulators. 4691 current = first_ur; 4692 while (true) { 4693 // Create vector_accumulator to replace current. 4694 Node* last_vector_accumulator = current->in(1); 4695 Node* vector_input = current->in(2); 4696 VectorNode* vector_accumulator = VectorNode::make(vopc, last_vector_accumulator, vector_input, vec_t); 4697 register_new_node(vector_accumulator, cl); 4698 _igvn.replace_node(current, vector_accumulator); 4699 VectorNode::trace_new_vector(vector_accumulator, "Unordered Reduction"); 4700 if (current == last_ur) { 4701 break; 4702 } 4703 current = vector_accumulator->unique_out()->as_Reduction(); 4704 assert(!current->requires_strict_order(), "must be"); 4705 } 4706 4707 // Create post-loop reduction. 4708 Node* last_accumulator = phi->in(2); 4709 Node* post_loop_reduction = ReductionNode::make(sopc, nullptr, init, last_accumulator, bt); 4710 4711 // Take over uses of last_accumulator that are not in the loop. 4712 for (DUIterator i = last_accumulator->outs(); last_accumulator->has_out(i); i++) { 4713 Node* use = last_accumulator->out(i); 4714 if (use != phi && use != post_loop_reduction) { 4715 assert(ctrl_or_self(use) != cl, "use must be outside loop"); 4716 use->replace_edge(last_accumulator, post_loop_reduction, &_igvn); 4717 --i; 4718 } 4719 } 4720 register_new_node(post_loop_reduction, get_late_ctrl(post_loop_reduction, cl)); 4721 VectorNode::trace_new_vector(post_loop_reduction, "Unordered Reduction"); 4722 4723 assert(last_accumulator->outcnt() == 2, "last_accumulator has 2 uses: phi and post_loop_reduction"); 4724 assert(post_loop_reduction->outcnt() > 0, "should have taken over all non loop uses of last_accumulator"); 4725 assert(phi->outcnt() == 1, "accumulator is the only use of phi"); 4726 } 4727 } 4728 4729 void DataNodeGraph::clone_data_nodes(Node* new_ctrl) { 4730 for (uint i = 0; i < _data_nodes.size(); i++) { 4731 clone(_data_nodes[i], new_ctrl); 4732 } 4733 } 4734 4735 // Clone the given node and set it up properly. Set 'new_ctrl' as ctrl. 4736 void DataNodeGraph::clone(Node* node, Node* new_ctrl) { 4737 Node* clone = node->clone(); 4738 _phase->igvn().register_new_node_with_optimizer(clone); 4739 _orig_to_new.put(node, clone); 4740 _phase->set_ctrl(clone, new_ctrl); 4741 if (node->is_CastII()) { 4742 clone->set_req(0, new_ctrl); 4743 } 4744 } 4745 4746 // Rewire the data inputs of all (unprocessed) cloned nodes, whose inputs are still pointing to the same inputs as their 4747 // corresponding orig nodes, to the newly cloned inputs to create a separate cloned graph. 4748 void DataNodeGraph::rewire_clones_to_cloned_inputs() { 4749 _orig_to_new.iterate_all([&](Node* node, Node* clone) { 4750 for (uint i = 1; i < node->req(); i++) { 4751 Node** cloned_input = _orig_to_new.get(node->in(i)); 4752 if (cloned_input != nullptr) { 4753 // Input was also cloned -> rewire clone to the cloned input. 4754 _phase->igvn().replace_input_of(clone, i, *cloned_input); 4755 } 4756 } 4757 }); 4758 } 4759 4760 // Clone all non-OpaqueLoop* nodes and apply the provided transformation strategy for OpaqueLoop* nodes. 4761 // Set 'new_ctrl' as ctrl for all cloned non-OpaqueLoop* nodes. 4762 void DataNodeGraph::clone_data_nodes_and_transform_opaque_loop_nodes( 4763 const TransformStrategyForOpaqueLoopNodes& transform_strategy, 4764 Node* new_ctrl) { 4765 for (uint i = 0; i < _data_nodes.size(); i++) { 4766 Node* data_node = _data_nodes[i]; 4767 if (data_node->is_Opaque1()) { 4768 transform_opaque_node(transform_strategy, data_node); 4769 } else { 4770 clone(data_node, new_ctrl); 4771 } 4772 } 4773 } 4774 4775 void DataNodeGraph::transform_opaque_node(const TransformStrategyForOpaqueLoopNodes& transform_strategy, Node* node) { 4776 Node* transformed_node; 4777 if (node->is_OpaqueLoopInit()) { 4778 transformed_node = transform_strategy.transform_opaque_init(node->as_OpaqueLoopInit()); 4779 } else { 4780 assert(node->is_OpaqueLoopStride(), "must be OpaqueLoopStrideNode"); 4781 transformed_node = transform_strategy.transform_opaque_stride(node->as_OpaqueLoopStride()); 4782 } 4783 // Add an orig->new mapping to correctly update the inputs of the copied graph in rewire_clones_to_cloned_inputs(). 4784 _orig_to_new.put(node, transformed_node); 4785 }