1 /* 2 * Copyright (c) 1999, 2025, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "gc/shared/barrierSet.hpp" 26 #include "gc/shared/c2/barrierSetC2.hpp" 27 #include "memory/allocation.inline.hpp" 28 #include "memory/resourceArea.hpp" 29 #include "opto/addnode.hpp" 30 #include "opto/callnode.hpp" 31 #include "opto/castnode.hpp" 32 #include "opto/connode.hpp" 33 #include "opto/divnode.hpp" 34 #include "opto/loopnode.hpp" 35 #include "opto/matcher.hpp" 36 #include "opto/movenode.hpp" 37 #include "opto/mulnode.hpp" 38 #include "opto/opaquenode.hpp" 39 #include "opto/rootnode.hpp" 40 #include "opto/subnode.hpp" 41 #include "opto/subtypenode.hpp" 42 #include "opto/superword.hpp" 43 #include "opto/vectornode.hpp" 44 #include "utilities/checkedCast.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 SplitThruPhiWins wins(region); 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.add_win(i); 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.add_win(i); 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.add_win(i); 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.reset(); 170 break; 171 } 172 } 173 // Too few wins? 174 if (!wins.profitable(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 #ifndef PRODUCT 232 if (TraceLoopOpts) { 233 tty->print("Split %d %s through %d Phi in %d %s", 234 n->_idx, n->Name(), phi->_idx, region->_idx, region->Name()); 235 } 236 #endif // !PRODUCT 237 238 return phi; 239 } 240 241 // Test whether node 'x' can move into an inner loop relative to node 'n'. 242 // Note: The test is not exact. Returns true if 'x' COULD end up in an inner loop, 243 // BUT it can also return true and 'x' is in the outer loop 244 bool PhaseIdealLoop::can_move_to_inner_loop(Node* n, LoopNode* n_loop, Node* x) { 245 IdealLoopTree* n_loop_tree = get_loop(n_loop); 246 IdealLoopTree* x_loop_tree = get_loop(get_early_ctrl(x)); 247 // x_loop_tree should be outer or same loop as n_loop_tree 248 return !x_loop_tree->is_member(n_loop_tree); 249 } 250 251 // Subtype checks that carry profile data don't common so look for a replacement by following edges 252 Node* PhaseIdealLoop::similar_subtype_check(const Node* x, Node* r_in) { 253 if (x->is_SubTypeCheck()) { 254 Node* in1 = x->in(1); 255 for (DUIterator_Fast imax, i = in1->fast_outs(imax); i < imax; i++) { 256 Node* u = in1->fast_out(i); 257 if (u != x && u->is_SubTypeCheck() && u->in(1) == x->in(1) && u->in(2) == x->in(2)) { 258 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) { 259 Node* bol = u->fast_out(j); 260 for (DUIterator_Fast kmax, k = bol->fast_outs(kmax); k < kmax; k++) { 261 Node* iff = bol->fast_out(k); 262 // Only dominating subtype checks are interesting: otherwise we risk replacing a subtype check by another with 263 // unrelated profile 264 if (iff->is_If() && is_dominator(iff, r_in)) { 265 return u; 266 } 267 } 268 } 269 } 270 } 271 } 272 return nullptr; 273 } 274 275 // Return true if 'n' is a Div or Mod node (without zero check If node which was removed earlier) with a loop phi divisor 276 // of a trip-counted (integer or long) loop with a backedge input that could be zero (include zero in its type range). In 277 // this case, we cannot split the division to the backedge as it could freely float above the loop exit check resulting in 278 // a division by zero. This situation is possible because the type of an increment node of an iv phi (trip-counter) could 279 // include zero while the iv phi does not (see PhiNode::Value() for trip-counted loops where we improve types of iv phis). 280 // We also need to check other loop phis as they could have been created in the same split-if pass when applying 281 // PhaseIdealLoop::split_thru_phi() to split nodes through an iv phi. 282 bool PhaseIdealLoop::cannot_split_division(const Node* n, const Node* region) const { 283 const Type* zero; 284 switch (n->Opcode()) { 285 case Op_DivI: 286 case Op_ModI: 287 case Op_UDivI: 288 case Op_UModI: 289 zero = TypeInt::ZERO; 290 break; 291 case Op_DivL: 292 case Op_ModL: 293 case Op_UDivL: 294 case Op_UModL: 295 zero = TypeLong::ZERO; 296 break; 297 default: 298 return false; 299 } 300 301 if (n->in(0) != nullptr) { 302 // Cannot split through phi if Div or Mod node has a control dependency to a zero check. 303 return true; 304 } 305 306 Node* divisor = n->in(2); 307 return is_divisor_loop_phi(divisor, region) && 308 loop_phi_backedge_type_contains_zero(divisor, zero); 309 } 310 311 bool PhaseIdealLoop::is_divisor_loop_phi(const Node* divisor, const Node* loop) { 312 return loop->is_Loop() && divisor->is_Phi() && divisor->in(0) == loop; 313 } 314 315 bool PhaseIdealLoop::loop_phi_backedge_type_contains_zero(const Node* phi_divisor, const Type* zero) const { 316 return _igvn.type(phi_divisor->in(LoopNode::LoopBackControl))->filter_speculative(zero) != Type::TOP; 317 } 318 319 //------------------------------dominated_by------------------------------------ 320 // Replace the dominated test with an obvious true or false. Place it on the 321 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the 322 // live path up to the dominating control. 323 void PhaseIdealLoop::dominated_by(IfProjNode* prevdom, IfNode* iff, bool flip, bool pin_array_access_nodes) { 324 if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); } 325 326 // prevdom is the dominating projection of the dominating test. 327 assert(iff->Opcode() == Op_If || 328 iff->Opcode() == Op_CountedLoopEnd || 329 iff->Opcode() == Op_LongCountedLoopEnd || 330 iff->Opcode() == Op_RangeCheck || 331 iff->Opcode() == Op_ParsePredicate, 332 "Check this code when new subtype is added"); 333 334 int pop = prevdom->Opcode(); 335 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" ); 336 if (flip) { 337 if (pop == Op_IfTrue) 338 pop = Op_IfFalse; 339 else 340 pop = Op_IfTrue; 341 } 342 // 'con' is set to true or false to kill the dominated test. 343 Node* con = makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO); 344 // Hack the dominated test 345 _igvn.replace_input_of(iff, 1, con); 346 347 // If I don't have a reachable TRUE and FALSE path following the IfNode then 348 // I can assume this path reaches an infinite loop. In this case it's not 349 // important to optimize the data Nodes - either the whole compilation will 350 // be tossed or this path (and all data Nodes) will go dead. 351 if (iff->outcnt() != 2) { 352 return; 353 } 354 355 // Make control-dependent data Nodes on the live path (path that will remain 356 // once the dominated IF is removed) become control-dependent on the 357 // dominating projection. 358 Node* dp = iff->proj_out_or_null(pop == Op_IfTrue); 359 360 if (dp == nullptr) { 361 return; 362 } 363 364 rewire_safe_outputs_to_dominator(dp, prevdom, pin_array_access_nodes); 365 } 366 367 void PhaseIdealLoop::rewire_safe_outputs_to_dominator(Node* source, Node* dominator, const bool pin_array_access_nodes) { 368 IdealLoopTree* old_loop = get_loop(source); 369 370 for (DUIterator_Fast imax, i = source->fast_outs(imax); i < imax; i++) { 371 Node* out = source->fast_out(i); // Control-dependent node 372 // Do not rewire Div and Mod nodes which could have a zero divisor to avoid skipping their zero check. 373 if (out->depends_only_on_test() && _igvn.no_dependent_zero_check(out)) { 374 assert(out->in(0) == source, "must be control dependent on source"); 375 _igvn.replace_input_of(out, 0, dominator); 376 if (pin_array_access_nodes) { 377 // Because of Loop Predication, Loads and range check Cast nodes that are control dependent on this range 378 // check (that is about to be removed) now depend on multiple dominating Hoisted Check Predicates. After the 379 // removal of this range check, these control dependent nodes end up at the lowest/nearest dominating predicate 380 // in the graph. To ensure that these Loads/Casts do not float above any of the dominating checks (even when the 381 // lowest dominating check is later replaced by yet another dominating check), we need to pin them at the lowest 382 // dominating check. 383 Node* clone = out->pin_array_access_node(); 384 if (clone != nullptr) { 385 clone = _igvn.register_new_node_with_optimizer(clone, out); 386 _igvn.replace_node(out, clone); 387 out = clone; 388 } 389 } 390 set_early_ctrl(out, false); 391 IdealLoopTree* new_loop = get_loop(get_ctrl(out)); 392 if (old_loop != new_loop) { 393 if (!old_loop->_child) { 394 old_loop->_body.yank(out); 395 } 396 if (!new_loop->_child) { 397 new_loop->_body.push(out); 398 } 399 } 400 --i; 401 --imax; 402 } 403 } 404 } 405 406 //------------------------------has_local_phi_input---------------------------- 407 // Return TRUE if 'n' has Phi inputs from its local block and no other 408 // block-local inputs (all non-local-phi inputs come from earlier blocks) 409 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) { 410 Node *n_ctrl = get_ctrl(n); 411 // See if some inputs come from a Phi in this block, or from before 412 // this block. 413 uint i; 414 for( i = 1; i < n->req(); i++ ) { 415 Node *phi = n->in(i); 416 if( phi->is_Phi() && phi->in(0) == n_ctrl ) 417 break; 418 } 419 if( i >= n->req() ) 420 return nullptr; // No Phi inputs; nowhere to clone thru 421 422 // Check for inputs created between 'n' and the Phi input. These 423 // must split as well; they have already been given the chance 424 // (courtesy of a post-order visit) and since they did not we must 425 // recover the 'cost' of splitting them by being very profitable 426 // when splitting 'n'. Since this is unlikely we simply give up. 427 for( i = 1; i < n->req(); i++ ) { 428 Node *m = n->in(i); 429 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) { 430 // We allow the special case of AddP's with no local inputs. 431 // This allows us to split-up address expressions. 432 if (m->is_AddP() && 433 get_ctrl(m->in(AddPNode::Base)) != n_ctrl && 434 get_ctrl(m->in(AddPNode::Address)) != n_ctrl && 435 get_ctrl(m->in(AddPNode::Offset)) != n_ctrl) { 436 // Move the AddP up to the dominating point. That's fine because control of m's inputs 437 // must dominate get_ctrl(m) == n_ctrl and we just checked that the input controls are != n_ctrl. 438 Node* c = find_non_split_ctrl(idom(n_ctrl)); 439 if (c->is_OuterStripMinedLoop()) { 440 c->as_Loop()->verify_strip_mined(1); 441 c = c->in(LoopNode::EntryControl); 442 } 443 set_ctrl_and_loop(m, c); 444 continue; 445 } 446 return nullptr; 447 } 448 assert(n->is_Phi() || m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control"); 449 } 450 451 return n_ctrl; 452 } 453 454 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2). 455 Node* PhaseIdealLoop::remix_address_expressions_add_left_shift(Node* n, IdealLoopTree* n_loop, Node* n_ctrl, BasicType bt) { 456 assert(bt == T_INT || bt == T_LONG, "only for integers"); 457 int n_op = n->Opcode(); 458 459 if (n_op == Op_LShift(bt)) { 460 // Scale is loop invariant 461 Node* scale = n->in(2); 462 Node* scale_ctrl = get_ctrl(scale); 463 IdealLoopTree* scale_loop = get_loop(scale_ctrl); 464 if (n_loop == scale_loop || !scale_loop->is_member(n_loop)) { 465 return nullptr; 466 } 467 const TypeInt* scale_t = scale->bottom_type()->isa_int(); 468 if (scale_t != nullptr && scale_t->is_con() && scale_t->get_con() >= 16) { 469 return nullptr; // Dont bother with byte/short masking 470 } 471 // Add must vary with loop (else shift would be loop-invariant) 472 Node* add = n->in(1); 473 Node* add_ctrl = get_ctrl(add); 474 IdealLoopTree* add_loop = get_loop(add_ctrl); 475 if (n_loop != add_loop) { 476 return nullptr; // happens w/ evil ZKM loops 477 } 478 479 // Convert I-V into I+ (0-V); same for V-I 480 if (add->Opcode() == Op_Sub(bt) && 481 _igvn.type(add->in(1)) != TypeInteger::zero(bt)) { 482 assert(add->Opcode() == Op_SubI || add->Opcode() == Op_SubL, ""); 483 Node* zero = integercon(0, bt); 484 Node* neg = SubNode::make(zero, add->in(2), bt); 485 register_new_node_with_ctrl_of(neg, add->in(2)); 486 add = AddNode::make(add->in(1), neg, bt); 487 register_new_node(add, add_ctrl); 488 } 489 if (add->Opcode() != Op_Add(bt)) return nullptr; 490 assert(add->Opcode() == Op_AddI || add->Opcode() == Op_AddL, ""); 491 // See if one add input is loop invariant 492 Node* add_var = add->in(1); 493 Node* add_var_ctrl = get_ctrl(add_var); 494 IdealLoopTree* add_var_loop = get_loop(add_var_ctrl); 495 Node* add_invar = add->in(2); 496 Node* add_invar_ctrl = get_ctrl(add_invar); 497 IdealLoopTree* add_invar_loop = get_loop(add_invar_ctrl); 498 if (add_invar_loop == n_loop) { 499 // Swap to find the invariant part 500 add_invar = add_var; 501 add_invar_ctrl = add_var_ctrl; 502 add_invar_loop = add_var_loop; 503 add_var = add->in(2); 504 } else if (add_var_loop != n_loop) { // Else neither input is loop invariant 505 return nullptr; 506 } 507 if (n_loop == add_invar_loop || !add_invar_loop->is_member(n_loop)) { 508 return nullptr; // No invariant part of the add? 509 } 510 511 // Yes! Reshape address expression! 512 Node* inv_scale = LShiftNode::make(add_invar, scale, bt); 513 Node* inv_scale_ctrl = 514 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ? 515 add_invar_ctrl : scale_ctrl; 516 register_new_node(inv_scale, inv_scale_ctrl); 517 Node* var_scale = LShiftNode::make(add_var, scale, bt); 518 register_new_node(var_scale, n_ctrl); 519 Node* var_add = AddNode::make(var_scale, inv_scale, bt); 520 register_new_node(var_add, n_ctrl); 521 _igvn.replace_node(n, var_add); 522 return var_add; 523 } 524 return nullptr; 525 } 526 527 //------------------------------remix_address_expressions---------------------- 528 // Rework addressing expressions to get the most loop-invariant stuff 529 // moved out. We'd like to do all associative operators, but it's especially 530 // important (common) to do address expressions. 531 Node* PhaseIdealLoop::remix_address_expressions(Node* n) { 532 if (!has_ctrl(n)) return nullptr; 533 Node* n_ctrl = get_ctrl(n); 534 IdealLoopTree* n_loop = get_loop(n_ctrl); 535 536 // See if 'n' mixes loop-varying and loop-invariant inputs and 537 // itself is loop-varying. 538 539 // Only interested in binary ops (and AddP) 540 if (n->req() < 3 || n->req() > 4) return nullptr; 541 542 Node* n1_ctrl = get_ctrl(n->in( 1)); 543 Node* n2_ctrl = get_ctrl(n->in( 2)); 544 Node* n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3)); 545 IdealLoopTree* n1_loop = get_loop(n1_ctrl); 546 IdealLoopTree* n2_loop = get_loop(n2_ctrl); 547 IdealLoopTree* n3_loop = get_loop(n3_ctrl); 548 549 // Does one of my inputs spin in a tighter loop than self? 550 if ((n_loop->is_member(n1_loop) && n_loop != n1_loop) || 551 (n_loop->is_member(n2_loop) && n_loop != n2_loop) || 552 (n_loop->is_member(n3_loop) && n_loop != n3_loop)) { 553 return nullptr; // Leave well enough alone 554 } 555 556 // Is at least one of my inputs loop-invariant? 557 if (n1_loop == n_loop && 558 n2_loop == n_loop && 559 n3_loop == n_loop) { 560 return nullptr; // No loop-invariant inputs 561 } 562 563 Node* res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_INT); 564 if (res != nullptr) { 565 return res; 566 } 567 res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_LONG); 568 if (res != nullptr) { 569 return res; 570 } 571 572 int n_op = n->Opcode(); 573 // Replace (I+V) with (V+I) 574 if (n_op == Op_AddI || 575 n_op == Op_AddL || 576 n_op == Op_AddF || 577 n_op == Op_AddD || 578 n_op == Op_MulI || 579 n_op == Op_MulL || 580 n_op == Op_MulF || 581 n_op == Op_MulD) { 582 if (n2_loop == n_loop) { 583 assert(n1_loop != n_loop, ""); 584 n->swap_edges(1, 2); 585 } 586 } 587 588 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V), 589 // but not if I2 is a constant. Skip for irreducible loops. 590 if (n_op == Op_AddP && n_loop->_head->is_Loop()) { 591 if (n2_loop == n_loop && n3_loop != n_loop) { 592 if (n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con()) { 593 Node* n22_ctrl = get_ctrl(n->in(2)->in(2)); 594 Node* n23_ctrl = get_ctrl(n->in(2)->in(3)); 595 IdealLoopTree* n22loop = get_loop(n22_ctrl); 596 IdealLoopTree* n23_loop = get_loop(n23_ctrl); 597 if (n22loop != n_loop && n22loop->is_member(n_loop) && 598 n23_loop == n_loop) { 599 Node* add1 = new AddPNode(n->in(1), n->in(2)->in(2), n->in(3)); 600 // Stuff new AddP in the loop preheader 601 register_new_node(add1, n_loop->_head->as_Loop()->skip_strip_mined(1)->in(LoopNode::EntryControl)); 602 Node* add2 = new AddPNode(n->in(1), add1, n->in(2)->in(3)); 603 register_new_node(add2, n_ctrl); 604 _igvn.replace_node(n, add2); 605 return add2; 606 } 607 } 608 } 609 610 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V) 611 if (n2_loop != n_loop && n3_loop == n_loop) { 612 if (n->in(3)->Opcode() == Op_AddX) { 613 Node* V = n->in(3)->in(1); 614 Node* I = n->in(3)->in(2); 615 if (is_member(n_loop,get_ctrl(V))) { 616 } else { 617 Node *tmp = V; V = I; I = tmp; 618 } 619 if (!is_member(n_loop,get_ctrl(I))) { 620 Node* add1 = new AddPNode(n->in(1), n->in(2), I); 621 // Stuff new AddP in the loop preheader 622 register_new_node(add1, n_loop->_head->as_Loop()->skip_strip_mined(1)->in(LoopNode::EntryControl)); 623 Node* add2 = new AddPNode(n->in(1), add1, V); 624 register_new_node(add2, n_ctrl); 625 _igvn.replace_node(n, add2); 626 return add2; 627 } 628 } 629 } 630 } 631 632 return nullptr; 633 } 634 635 // Optimize ((in1[2*i] * in2[2*i]) + (in1[2*i+1] * in2[2*i+1])) 636 Node *PhaseIdealLoop::convert_add_to_muladd(Node* n) { 637 assert(n->Opcode() == Op_AddI, "sanity"); 638 Node * nn = nullptr; 639 Node * in1 = n->in(1); 640 Node * in2 = n->in(2); 641 if (in1->Opcode() == Op_MulI && in2->Opcode() == Op_MulI) { 642 IdealLoopTree* loop_n = get_loop(get_ctrl(n)); 643 if (loop_n->is_counted() && 644 loop_n->_head->as_Loop()->is_valid_counted_loop(T_INT) && 645 Matcher::match_rule_supported(Op_MulAddVS2VI) && 646 Matcher::match_rule_supported(Op_MulAddS2I)) { 647 Node* mul_in1 = in1->in(1); 648 Node* mul_in2 = in1->in(2); 649 Node* mul_in3 = in2->in(1); 650 Node* mul_in4 = in2->in(2); 651 if (mul_in1->Opcode() == Op_LoadS && 652 mul_in2->Opcode() == Op_LoadS && 653 mul_in3->Opcode() == Op_LoadS && 654 mul_in4->Opcode() == Op_LoadS) { 655 IdealLoopTree* loop1 = get_loop(get_ctrl(mul_in1)); 656 IdealLoopTree* loop2 = get_loop(get_ctrl(mul_in2)); 657 IdealLoopTree* loop3 = get_loop(get_ctrl(mul_in3)); 658 IdealLoopTree* loop4 = get_loop(get_ctrl(mul_in4)); 659 IdealLoopTree* loop5 = get_loop(get_ctrl(in1)); 660 IdealLoopTree* loop6 = get_loop(get_ctrl(in2)); 661 // All nodes should be in the same counted loop. 662 if (loop_n == loop1 && loop_n == loop2 && loop_n == loop3 && 663 loop_n == loop4 && loop_n == loop5 && loop_n == loop6) { 664 Node* adr1 = mul_in1->in(MemNode::Address); 665 Node* adr2 = mul_in2->in(MemNode::Address); 666 Node* adr3 = mul_in3->in(MemNode::Address); 667 Node* adr4 = mul_in4->in(MemNode::Address); 668 if (adr1->is_AddP() && adr2->is_AddP() && adr3->is_AddP() && adr4->is_AddP()) { 669 if ((adr1->in(AddPNode::Base) == adr3->in(AddPNode::Base)) && 670 (adr2->in(AddPNode::Base) == adr4->in(AddPNode::Base))) { 671 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in3, mul_in4); 672 register_new_node_with_ctrl_of(nn, n); 673 _igvn.replace_node(n, nn); 674 return nn; 675 } else if ((adr1->in(AddPNode::Base) == adr4->in(AddPNode::Base)) && 676 (adr2->in(AddPNode::Base) == adr3->in(AddPNode::Base))) { 677 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in4, mul_in3); 678 register_new_node_with_ctrl_of(nn, n); 679 _igvn.replace_node(n, nn); 680 return nn; 681 } 682 } 683 } 684 } 685 } 686 } 687 return nn; 688 } 689 690 //------------------------------conditional_move------------------------------- 691 // Attempt to replace a Phi with a conditional move. We have some pretty 692 // strict profitability requirements. All Phis at the merge point must 693 // be converted, so we can remove the control flow. We need to limit the 694 // number of c-moves to a small handful. All code that was in the side-arms 695 // of the CFG diamond is now speculatively executed. This code has to be 696 // "cheap enough". We are pretty much limited to CFG diamonds that merge 697 // 1 or 2 items with a total of 1 or 2 ops executed speculatively. 698 Node *PhaseIdealLoop::conditional_move( Node *region ) { 699 700 assert(region->is_Region(), "sanity check"); 701 if (region->req() != 3) return nullptr; 702 703 // Check for CFG diamond 704 Node *lp = region->in(1); 705 Node *rp = region->in(2); 706 if (!lp || !rp) return nullptr; 707 Node *lp_c = lp->in(0); 708 if (lp_c == nullptr || lp_c != rp->in(0) || !lp_c->is_If()) return nullptr; 709 IfNode *iff = lp_c->as_If(); 710 711 // Check for ops pinned in an arm of the diamond. 712 // Can't remove the control flow in this case 713 if (lp->outcnt() > 1) return nullptr; 714 if (rp->outcnt() > 1) return nullptr; 715 716 IdealLoopTree* r_loop = get_loop(region); 717 assert(r_loop == get_loop(iff), "sanity"); 718 // Always convert to CMOVE if all results are used only outside this loop. 719 bool used_inside_loop = (r_loop == _ltree_root); 720 721 // Check profitability 722 int cost = 0; 723 int phis = 0; 724 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 725 Node *out = region->fast_out(i); 726 if (!out->is_Phi()) continue; // Ignore other control edges, etc 727 phis++; 728 PhiNode* phi = out->as_Phi(); 729 BasicType bt = phi->type()->basic_type(); 730 switch (bt) { 731 case T_DOUBLE: 732 case T_FLOAT: 733 if (C->use_cmove()) { 734 continue; //TODO: maybe we want to add some cost 735 } 736 cost += Matcher::float_cmove_cost(); // Could be very expensive 737 break; 738 case T_LONG: { 739 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's 740 } 741 case T_INT: // These all CMOV fine 742 case T_ADDRESS: { // (RawPtr) 743 cost++; 744 break; 745 } 746 case T_NARROWOOP: // Fall through 747 case T_OBJECT: { // Base oops are OK, but not derived oops 748 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr(); 749 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a 750 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus 751 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we 752 // have a Phi for the base here that we convert to a CMOVE all is well 753 // and good. But if the base is dead, we'll not make a CMOVE. Later 754 // the allocator will have to produce a base by creating a CMOVE of the 755 // relevant bases. This puts the allocator in the business of 756 // manufacturing expensive instructions, generally a bad plan. 757 // Just Say No to Conditionally-Moved Derived Pointers. 758 if (tp && tp->offset() != 0) 759 return nullptr; 760 cost++; 761 break; 762 } 763 default: 764 return nullptr; // In particular, can't do memory or I/O 765 } 766 // Add in cost any speculative ops 767 for (uint j = 1; j < region->req(); j++) { 768 Node *proj = region->in(j); 769 Node *inp = phi->in(j); 770 if (get_ctrl(inp) == proj) { // Found local op 771 cost++; 772 // Check for a chain of dependent ops; these will all become 773 // speculative in a CMOV. 774 for (uint k = 1; k < inp->req(); k++) 775 if (get_ctrl(inp->in(k)) == proj) 776 cost += ConditionalMoveLimit; // Too much speculative goo 777 } 778 } 779 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode. 780 // This will likely Split-If, a higher-payoff operation. 781 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) { 782 Node* use = phi->fast_out(k); 783 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr()) 784 cost += ConditionalMoveLimit; 785 // Is there a use inside the loop? 786 // Note: check only basic types since CMoveP is pinned. 787 if (!used_inside_loop && is_java_primitive(bt)) { 788 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use); 789 if (r_loop == u_loop || r_loop->is_member(u_loop)) { 790 used_inside_loop = true; 791 } 792 } 793 } 794 }//for 795 Node* bol = iff->in(1); 796 assert(!bol->is_OpaqueInitializedAssertionPredicate(), "Initialized Assertion Predicates cannot form a diamond with Halt"); 797 if (bol->is_OpaqueTemplateAssertionPredicate()) { 798 // Ignore Template Assertion Predicates with OpaqueTemplateAssertionPredicate nodes. 799 return nullptr; 800 } 801 if (bol->is_OpaqueMultiversioning()) { 802 assert(bol->as_OpaqueMultiversioning()->is_useless(), "Must be useless, i.e. fast main loop has already disappeared."); 803 // Ignore multiversion_if that just lost its loops. The OpaqueMultiversioning is marked useless, 804 // and will make the multiversion_if constant fold in the next IGVN round. 805 return nullptr; 806 } 807 if (!bol->is_Bool()) { 808 assert(false, "Expected Bool, but got %s", NodeClassNames[bol->Opcode()]); 809 return nullptr; 810 } 811 int cmp_op = bol->in(1)->Opcode(); 812 if (cmp_op == Op_SubTypeCheck) { // SubTypeCheck expansion expects an IfNode 813 return nullptr; 814 } 815 // It is expensive to generate flags from a float compare. 816 // Avoid duplicated float compare. 817 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return nullptr; 818 819 float infrequent_prob = PROB_UNLIKELY_MAG(3); 820 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop. 821 if (used_inside_loop) { 822 if (cost >= ConditionalMoveLimit) return nullptr; // Too much goo 823 824 // BlockLayoutByFrequency optimization moves infrequent branch 825 // from hot path. No point in CMOV'ing in such case (110 is used 826 // instead of 100 to take into account not exactness of float value). 827 if (BlockLayoutByFrequency) { 828 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f); 829 } 830 } 831 // Check for highly predictable branch. No point in CMOV'ing if 832 // we are going to predict accurately all the time. 833 if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) { 834 //keep going 835 } else if (iff->_prob < infrequent_prob || 836 iff->_prob > (1.0f - infrequent_prob)) 837 return nullptr; 838 839 // -------------- 840 // Now replace all Phis with CMOV's 841 Node *cmov_ctrl = iff->in(0); 842 uint flip = (lp->Opcode() == Op_IfTrue); 843 Node_List wq; 844 while (1) { 845 PhiNode* phi = nullptr; 846 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 847 Node *out = region->fast_out(i); 848 if (out->is_Phi()) { 849 phi = out->as_Phi(); 850 break; 851 } 852 } 853 if (phi == nullptr || _igvn.type(phi) == Type::TOP || !CMoveNode::supported(_igvn.type(phi))) { 854 break; 855 } 856 // Move speculative ops 857 wq.push(phi); 858 while (wq.size() > 0) { 859 Node *n = wq.pop(); 860 for (uint j = 1; j < n->req(); j++) { 861 Node* m = n->in(j); 862 if (m != nullptr && !is_dominator(get_ctrl(m), cmov_ctrl)) { 863 set_ctrl(m, cmov_ctrl); 864 wq.push(m); 865 } 866 } 867 } 868 Node* cmov = CMoveNode::make(iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi)); 869 register_new_node(cmov, cmov_ctrl); 870 _igvn.replace_node(phi, cmov); 871 #ifndef PRODUCT 872 if (TraceLoopOpts) { 873 tty->print("CMOV "); 874 r_loop->dump_head(); 875 if (Verbose) { 876 bol->in(1)->dump(1); 877 cmov->dump(1); 878 } 879 } 880 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } ); 881 #endif 882 } 883 884 // The useless CFG diamond will fold up later; see the optimization in 885 // RegionNode::Ideal. 886 _igvn._worklist.push(region); 887 888 return iff->in(1); 889 } 890 891 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) { 892 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 893 Node* u = m->fast_out(i); 894 if (u->is_CFG()) { 895 if (u->is_NeverBranch()) { 896 u = u->as_NeverBranch()->proj_out(0); 897 enqueue_cfg_uses(u, wq); 898 } else { 899 wq.push(u); 900 } 901 } 902 } 903 } 904 905 // Try moving a store out of a loop, right before the loop 906 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) { 907 // Store has to be first in the loop body 908 IdealLoopTree *n_loop = get_loop(n_ctrl); 909 if (n->is_Store() && n_loop != _ltree_root && 910 n_loop->is_loop() && n_loop->_head->is_Loop() && 911 n->in(0) != nullptr) { 912 Node* address = n->in(MemNode::Address); 913 Node* value = n->in(MemNode::ValueIn); 914 Node* mem = n->in(MemNode::Memory); 915 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 916 IdealLoopTree* value_loop = get_loop(get_ctrl(value)); 917 918 // - address and value must be loop invariant 919 // - memory must be a memory Phi for the loop 920 // - Store must be the only store on this memory slice in the 921 // loop: if there's another store following this one then value 922 // written at iteration i by the second store could be overwritten 923 // at iteration i+n by the first store: it's not safe to move the 924 // first store out of the loop 925 // - nothing must observe the memory Phi: it guarantees no read 926 // before the store, we are also guaranteed the store post 927 // dominates the loop head (ignoring a possible early 928 // exit). Otherwise there would be extra Phi involved between the 929 // loop's Phi and the store. 930 // - there must be no early exit from the loop before the Store 931 // (such an exit most of the time would be an extra use of the 932 // memory Phi but sometimes is a bottom memory Phi that takes the 933 // store as input). 934 935 if (!n_loop->is_member(address_loop) && 936 !n_loop->is_member(value_loop) && 937 mem->is_Phi() && mem->in(0) == n_loop->_head && 938 mem->outcnt() == 1 && 939 mem->in(LoopNode::LoopBackControl) == n) { 940 941 assert(n_loop->_tail != nullptr, "need a tail"); 942 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop"); 943 944 // Verify that there's no early exit of the loop before the store. 945 bool ctrl_ok = false; 946 { 947 // Follow control from loop head until n, we exit the loop or 948 // we reach the tail 949 ResourceMark rm; 950 Unique_Node_List wq; 951 wq.push(n_loop->_head); 952 953 for (uint next = 0; next < wq.size(); ++next) { 954 Node *m = wq.at(next); 955 if (m == n->in(0)) { 956 ctrl_ok = true; 957 continue; 958 } 959 assert(!has_ctrl(m), "should be CFG"); 960 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) { 961 ctrl_ok = false; 962 break; 963 } 964 enqueue_cfg_uses(m, wq); 965 if (wq.size() > 10) { 966 ctrl_ok = false; 967 break; 968 } 969 } 970 } 971 if (ctrl_ok) { 972 // move the Store 973 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem); 974 _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl)); 975 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl)); 976 // Disconnect the phi now. An empty phi can confuse other 977 // optimizations in this pass of loop opts. 978 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl)); 979 n_loop->_body.yank(mem); 980 981 set_ctrl_and_loop(n, n->in(0)); 982 983 return n; 984 } 985 } 986 } 987 return nullptr; 988 } 989 990 // Try moving a store out of a loop, right after the loop 991 void PhaseIdealLoop::try_move_store_after_loop(Node* n) { 992 if (n->is_Store() && n->in(0) != nullptr) { 993 Node *n_ctrl = get_ctrl(n); 994 IdealLoopTree *n_loop = get_loop(n_ctrl); 995 // Store must be in a loop 996 if (n_loop != _ltree_root && !n_loop->_irreducible) { 997 Node* address = n->in(MemNode::Address); 998 Node* value = n->in(MemNode::ValueIn); 999 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 1000 // address must be loop invariant 1001 if (!n_loop->is_member(address_loop)) { 1002 // Store must be last on this memory slice in the loop and 1003 // nothing in the loop must observe it 1004 Node* phi = nullptr; 1005 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1006 Node* u = n->fast_out(i); 1007 if (has_ctrl(u)) { // control use? 1008 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 1009 if (!n_loop->is_member(u_loop)) { 1010 continue; 1011 } 1012 if (u->is_Phi() && u->in(0) == n_loop->_head) { 1013 assert(_igvn.type(u) == Type::MEMORY, "bad phi"); 1014 // multiple phis on the same slice are possible 1015 if (phi != nullptr) { 1016 return; 1017 } 1018 phi = u; 1019 continue; 1020 } 1021 } 1022 return; 1023 } 1024 if (phi != nullptr) { 1025 // Nothing in the loop before the store (next iteration) 1026 // must observe the stored value 1027 bool mem_ok = true; 1028 { 1029 ResourceMark rm; 1030 Unique_Node_List wq; 1031 wq.push(phi); 1032 for (uint next = 0; next < wq.size() && mem_ok; ++next) { 1033 Node *m = wq.at(next); 1034 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) { 1035 Node* u = m->fast_out(i); 1036 if (u->is_Store() || u->is_Phi()) { 1037 if (u != n) { 1038 wq.push(u); 1039 mem_ok = (wq.size() <= 10); 1040 } 1041 } else { 1042 mem_ok = false; 1043 break; 1044 } 1045 } 1046 } 1047 } 1048 if (mem_ok) { 1049 // Move the store out of the loop if the LCA of all 1050 // users (except for the phi) is outside the loop. 1051 Node* hook = new Node(1); 1052 hook->init_req(0, n_ctrl); // Add an input to prevent hook from being dead 1053 _igvn.rehash_node_delayed(phi); 1054 int count = phi->replace_edge(n, hook, &_igvn); 1055 assert(count > 0, "inconsistent phi"); 1056 1057 // Compute latest point this store can go 1058 Node* lca = get_late_ctrl(n, get_ctrl(n)); 1059 if (lca->is_OuterStripMinedLoop()) { 1060 lca = lca->in(LoopNode::EntryControl); 1061 } 1062 if (n_loop->is_member(get_loop(lca))) { 1063 // LCA is in the loop - bail out 1064 _igvn.replace_node(hook, n); 1065 return; 1066 } 1067 #ifdef ASSERT 1068 if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) { 1069 assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined"); 1070 n_loop->_head->as_Loop()->verify_strip_mined(1); 1071 Node* outer = n_loop->_head->as_CountedLoop()->outer_loop(); 1072 IdealLoopTree* outer_loop = get_loop(outer); 1073 assert(n_loop->_parent == outer_loop, "broken loop tree"); 1074 assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state"); 1075 } 1076 #endif 1077 lca = place_outside_loop(lca, n_loop); 1078 assert(!n_loop->is_member(get_loop(lca)), "control must not be back in the loop"); 1079 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"); 1080 1081 // Move store out of the loop 1082 _igvn.replace_node(hook, n->in(MemNode::Memory)); 1083 _igvn.replace_input_of(n, 0, lca); 1084 set_ctrl_and_loop(n, lca); 1085 1086 // Disconnect the phi now. An empty phi can confuse other 1087 // optimizations in this pass of loop opts.. 1088 if (phi->in(LoopNode::LoopBackControl) == phi) { 1089 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl)); 1090 n_loop->_body.yank(phi); 1091 } 1092 } 1093 } 1094 } 1095 } 1096 } 1097 } 1098 1099 //------------------------------split_if_with_blocks_pre----------------------- 1100 // Do the real work in a non-recursive function. Data nodes want to be 1101 // cloned in the pre-order so they can feed each other nicely. 1102 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) { 1103 // Cloning these guys is unlikely to win 1104 int n_op = n->Opcode(); 1105 if (n_op == Op_MergeMem) { 1106 return n; 1107 } 1108 if (n->is_Proj()) { 1109 return n; 1110 } 1111 // Do not clone-up CmpFXXX variations, as these are always 1112 // followed by a CmpI 1113 if (n->is_Cmp()) { 1114 return n; 1115 } 1116 // Attempt to use a conditional move instead of a phi/branch 1117 if (ConditionalMoveLimit > 0 && n_op == Op_Region) { 1118 Node *cmov = conditional_move( n ); 1119 if (cmov) { 1120 return cmov; 1121 } 1122 } 1123 if (n->is_CFG() || n->is_LoadStore()) { 1124 return n; 1125 } 1126 if (n->is_Opaque1()) { // Opaque nodes cannot be mod'd 1127 if (!C->major_progress()) { // If chance of no more loop opts... 1128 _igvn._worklist.push(n); // maybe we'll remove them 1129 } 1130 return n; 1131 } 1132 1133 if (n->is_Con()) { 1134 return n; // No cloning for Con nodes 1135 } 1136 1137 Node *n_ctrl = get_ctrl(n); 1138 if (!n_ctrl) { 1139 return n; // Dead node 1140 } 1141 1142 Node* res = try_move_store_before_loop(n, n_ctrl); 1143 if (res != nullptr) { 1144 return n; 1145 } 1146 1147 // Attempt to remix address expressions for loop invariants 1148 Node *m = remix_address_expressions( n ); 1149 if( m ) return m; 1150 1151 if (n_op == Op_AddI) { 1152 Node *nn = convert_add_to_muladd( n ); 1153 if ( nn ) return nn; 1154 } 1155 1156 if (n->is_ConstraintCast()) { 1157 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this); 1158 // ConstraintCastNode::dominating_cast() uses node control input to determine domination. 1159 // Node control inputs don't necessarily agree with loop control info (due to 1160 // transformations happened in between), thus additional dominance check is needed 1161 // to keep loop info valid. 1162 if (dom_cast != nullptr && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) { 1163 _igvn.replace_node(n, dom_cast); 1164 return dom_cast; 1165 } 1166 } 1167 1168 // Determine if the Node has inputs from some local Phi. 1169 // Returns the block to clone thru. 1170 Node *n_blk = has_local_phi_input( n ); 1171 if( !n_blk ) return n; 1172 1173 // Do not clone the trip counter through on a CountedLoop 1174 // (messes up the canonical shape). 1175 if (((n_blk->is_CountedLoop() || (n_blk->is_Loop() && n_blk->as_Loop()->is_loop_nest_inner_loop())) && n->Opcode() == Op_AddI) || 1176 (n_blk->is_LongCountedLoop() && n->Opcode() == Op_AddL)) { 1177 return n; 1178 } 1179 // Pushing a shift through the iv Phi can get in the way of addressing optimizations or range check elimination 1180 if (n_blk->is_BaseCountedLoop() && n->Opcode() == Op_LShift(n_blk->as_BaseCountedLoop()->bt()) && 1181 n->in(1) == n_blk->as_BaseCountedLoop()->phi()) { 1182 return n; 1183 } 1184 1185 // Check for having no control input; not pinned. Allow 1186 // dominating control. 1187 if (n->in(0)) { 1188 Node *dom = idom(n_blk); 1189 if (dom_lca(n->in(0), dom) != n->in(0)) { 1190 return n; 1191 } 1192 } 1193 // Policy: when is it profitable. You must get more wins than 1194 // policy before it is considered profitable. Policy is usually 0, 1195 // so 1 win is considered profitable. Big merges will require big 1196 // cloning, so get a larger policy. 1197 int policy = n_blk->req() >> 2; 1198 1199 // If the loop is a candidate for range check elimination, 1200 // delay splitting through it's phi until a later loop optimization 1201 if (n_blk->is_BaseCountedLoop()) { 1202 IdealLoopTree *lp = get_loop(n_blk); 1203 if (lp && lp->_rce_candidate) { 1204 return n; 1205 } 1206 } 1207 1208 if (must_throttle_split_if()) return n; 1209 1210 // Split 'n' through the merge point if it is profitable 1211 Node *phi = split_thru_phi( n, n_blk, policy ); 1212 if (!phi) return n; 1213 1214 // Found a Phi to split thru! 1215 // Replace 'n' with the new phi 1216 _igvn.replace_node( n, phi ); 1217 // Moved a load around the loop, 'en-registering' something. 1218 if (n_blk->is_Loop() && n->is_Load() && 1219 !phi->in(LoopNode::LoopBackControl)->is_Load()) 1220 C->set_major_progress(); 1221 1222 return phi; 1223 } 1224 1225 static bool merge_point_too_heavy(Compile* C, Node* region) { 1226 // Bail out if the region and its phis have too many users. 1227 int weight = 0; 1228 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1229 weight += region->fast_out(i)->outcnt(); 1230 } 1231 int nodes_left = C->max_node_limit() - C->live_nodes(); 1232 if (weight * 8 > nodes_left) { 1233 if (PrintOpto) { 1234 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight); 1235 } 1236 return true; 1237 } else { 1238 return false; 1239 } 1240 } 1241 1242 static bool merge_point_safe(Node* region) { 1243 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode 1244 // having a PhiNode input. This sidesteps the dangerous case where the split 1245 // ConvI2LNode may become TOP if the input Value() does not 1246 // overlap the ConvI2L range, leaving a node which may not dominate its 1247 // uses. 1248 // A better fix for this problem can be found in the BugTraq entry, but 1249 // expediency for Mantis demands this hack. 1250 #ifdef _LP64 1251 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1252 Node* n = region->fast_out(i); 1253 if (n->is_Phi()) { 1254 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1255 Node* m = n->fast_out(j); 1256 if (m->Opcode() == Op_ConvI2L) 1257 return false; 1258 if (m->is_CastII()) { 1259 return false; 1260 } 1261 } 1262 } 1263 } 1264 #endif 1265 return true; 1266 } 1267 1268 1269 //------------------------------place_outside_loop--------------------------------- 1270 // Place some computation outside of this loop on the path to the use passed as argument 1271 Node* PhaseIdealLoop::place_outside_loop(Node* useblock, IdealLoopTree* loop) const { 1272 Node* head = loop->_head; 1273 assert(!loop->is_member(get_loop(useblock)), "must be outside loop"); 1274 if (head->is_Loop() && head->as_Loop()->is_strip_mined()) { 1275 loop = loop->_parent; 1276 assert(loop->_head->is_OuterStripMinedLoop(), "malformed strip mined loop"); 1277 } 1278 1279 // Pick control right outside the loop 1280 for (;;) { 1281 Node* dom = idom(useblock); 1282 if (loop->is_member(get_loop(dom))) { 1283 break; 1284 } 1285 useblock = dom; 1286 } 1287 assert(find_non_split_ctrl(useblock) == useblock, "should be non split control"); 1288 return useblock; 1289 } 1290 1291 1292 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) { 1293 if (!n->is_If() || n->is_BaseCountedLoopEnd()) { 1294 return false; 1295 } 1296 if (!n->in(0)->is_Region()) { 1297 return false; 1298 } 1299 1300 Node* region = n->in(0); 1301 Node* dom = idom(region); 1302 if (!dom->is_If() || !n->as_If()->same_condition(dom, &_igvn)) { 1303 return false; 1304 } 1305 IfNode* dom_if = dom->as_If(); 1306 Node* proj_true = dom_if->proj_out(1); 1307 Node* proj_false = dom_if->proj_out(0); 1308 1309 for (uint i = 1; i < region->req(); i++) { 1310 if (is_dominator(proj_true, region->in(i))) { 1311 continue; 1312 } 1313 if (is_dominator(proj_false, region->in(i))) { 1314 continue; 1315 } 1316 return false; 1317 } 1318 1319 return true; 1320 } 1321 1322 1323 bool PhaseIdealLoop::can_split_if(Node* n_ctrl) { 1324 if (must_throttle_split_if()) { 1325 return false; 1326 } 1327 1328 // Do not do 'split-if' if irreducible loops are present. 1329 if (_has_irreducible_loops) { 1330 return false; 1331 } 1332 1333 if (merge_point_too_heavy(C, n_ctrl)) { 1334 return false; 1335 } 1336 1337 // Do not do 'split-if' if some paths are dead. First do dead code 1338 // elimination and then see if its still profitable. 1339 for (uint i = 1; i < n_ctrl->req(); i++) { 1340 if (n_ctrl->in(i) == C->top()) { 1341 return false; 1342 } 1343 } 1344 1345 // If trying to do a 'Split-If' at the loop head, it is only 1346 // profitable if the cmp folds up on BOTH paths. Otherwise we 1347 // risk peeling a loop forever. 1348 1349 // CNC - Disabled for now. Requires careful handling of loop 1350 // body selection for the cloned code. Also, make sure we check 1351 // for any input path not being in the same loop as n_ctrl. For 1352 // irreducible loops we cannot check for 'n_ctrl->is_Loop()' 1353 // because the alternative loop entry points won't be converted 1354 // into LoopNodes. 1355 IdealLoopTree *n_loop = get_loop(n_ctrl); 1356 for (uint j = 1; j < n_ctrl->req(); j++) { 1357 if (get_loop(n_ctrl->in(j)) != n_loop) { 1358 return false; 1359 } 1360 } 1361 1362 // Check for safety of the merge point. 1363 if (!merge_point_safe(n_ctrl)) { 1364 return false; 1365 } 1366 1367 return true; 1368 } 1369 1370 // Detect if the node is the inner strip-mined loop 1371 // Return: null if it's not the case, or the exit of outer strip-mined loop 1372 static Node* is_inner_of_stripmined_loop(const Node* out) { 1373 Node* out_le = nullptr; 1374 1375 if (out->is_CountedLoopEnd()) { 1376 const CountedLoopNode* loop = out->as_CountedLoopEnd()->loopnode(); 1377 1378 if (loop != nullptr && loop->is_strip_mined()) { 1379 out_le = loop->in(LoopNode::EntryControl)->as_OuterStripMinedLoop()->outer_loop_exit(); 1380 } 1381 } 1382 1383 return out_le; 1384 } 1385 1386 //------------------------------split_if_with_blocks_post---------------------- 1387 // Do the real work in a non-recursive function. CFG hackery wants to be 1388 // in the post-order, so it can dirty the I-DOM info and not use the dirtied 1389 // info. 1390 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) { 1391 1392 // Cloning Cmp through Phi's involves the split-if transform. 1393 // FastLock is not used by an If 1394 if (n->is_Cmp() && !n->is_FastLock()) { 1395 Node *n_ctrl = get_ctrl(n); 1396 // Determine if the Node has inputs from some local Phi. 1397 // Returns the block to clone thru. 1398 Node *n_blk = has_local_phi_input(n); 1399 if (n_blk != n_ctrl) { 1400 return; 1401 } 1402 1403 if (!can_split_if(n_ctrl)) { 1404 return; 1405 } 1406 1407 if (n->outcnt() != 1) { 1408 return; // Multiple bool's from 1 compare? 1409 } 1410 Node *bol = n->unique_out(); 1411 assert(bol->is_Bool(), "expect a bool here"); 1412 if (bol->outcnt() != 1) { 1413 return;// Multiple branches from 1 compare? 1414 } 1415 Node *iff = bol->unique_out(); 1416 1417 // Check some safety conditions 1418 if (iff->is_If()) { // Classic split-if? 1419 if (iff->in(0) != n_ctrl) { 1420 return; // Compare must be in same blk as if 1421 } 1422 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE 1423 // Can't split CMove with different control. 1424 if (get_ctrl(iff) != n_ctrl) { 1425 return; 1426 } 1427 if (get_ctrl(iff->in(2)) == n_ctrl || 1428 get_ctrl(iff->in(3)) == n_ctrl) { 1429 return; // Inputs not yet split-up 1430 } 1431 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) { 1432 return; // Loop-invar test gates loop-varying CMOVE 1433 } 1434 } else { 1435 return; // some other kind of node, such as an Allocate 1436 } 1437 1438 // When is split-if profitable? Every 'win' on means some control flow 1439 // goes dead, so it's almost always a win. 1440 int policy = 0; 1441 // Split compare 'n' through the merge point if it is profitable 1442 Node *phi = split_thru_phi( n, n_ctrl, policy); 1443 if (!phi) { 1444 return; 1445 } 1446 1447 // Found a Phi to split thru! 1448 // Replace 'n' with the new phi 1449 _igvn.replace_node(n, phi); 1450 1451 // Now split the bool up thru the phi 1452 Node *bolphi = split_thru_phi(bol, n_ctrl, -1); 1453 guarantee(bolphi != nullptr, "null boolean phi node"); 1454 1455 _igvn.replace_node(bol, bolphi); 1456 assert(iff->in(1) == bolphi, ""); 1457 1458 if (bolphi->Value(&_igvn)->singleton()) { 1459 return; 1460 } 1461 1462 // Conditional-move? Must split up now 1463 if (!iff->is_If()) { 1464 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1); 1465 _igvn.replace_node(iff, cmovphi); 1466 return; 1467 } 1468 1469 // Now split the IF 1470 C->print_method(PHASE_BEFORE_SPLIT_IF, 4, iff); 1471 if (TraceLoopOpts) { 1472 tty->print_cr("Split-If"); 1473 } 1474 do_split_if(iff); 1475 C->print_method(PHASE_AFTER_SPLIT_IF, 4, iff); 1476 return; 1477 } 1478 1479 // Two identical ifs back to back can be merged 1480 if (try_merge_identical_ifs(n)) { 1481 return; 1482 } 1483 1484 // Check for an IF ready to split; one that has its 1485 // condition codes input coming from a Phi at the block start. 1486 int n_op = n->Opcode(); 1487 1488 // Check for an IF being dominated by another IF same test 1489 if (n_op == Op_If || 1490 n_op == Op_RangeCheck) { 1491 Node *bol = n->in(1); 1492 uint max = bol->outcnt(); 1493 // Check for same test used more than once? 1494 if (bol->is_Bool() && (max > 1 || bol->in(1)->is_SubTypeCheck())) { 1495 // Search up IDOMs to see if this IF is dominated. 1496 Node* cmp = bol->in(1); 1497 Node *cutoff = cmp->is_SubTypeCheck() ? dom_lca(get_ctrl(cmp->in(1)), get_ctrl(cmp->in(2))) : get_ctrl(bol); 1498 1499 // Now search up IDOMs till cutoff, looking for a dominating test 1500 Node *prevdom = n; 1501 Node *dom = idom(prevdom); 1502 while (dom != cutoff) { 1503 if (dom->req() > 1 && n->as_If()->same_condition(dom, &_igvn) && prevdom->in(0) == dom && 1504 safe_for_if_replacement(dom)) { 1505 // It's invalid to move control dependent data nodes in the inner 1506 // strip-mined loop, because: 1507 // 1) break validation of LoopNode::verify_strip_mined() 1508 // 2) move code with side-effect in strip-mined loop 1509 // Move to the exit of outer strip-mined loop in that case. 1510 Node* out_le = is_inner_of_stripmined_loop(dom); 1511 if (out_le != nullptr) { 1512 prevdom = out_le; 1513 } 1514 // Replace the dominated test with an obvious true or false. 1515 // Place it on the IGVN worklist for later cleanup. 1516 C->set_major_progress(); 1517 // Split if: pin array accesses that are control dependent on a range check and moved to a regular if, 1518 // to prevent an array load from floating above its range check. There are three cases: 1519 // 1. Move from RangeCheck "a" to RangeCheck "b": don't need to pin. If we ever remove b, then we pin 1520 // all its array accesses at that point. 1521 // 2. We move from RangeCheck "a" to regular if "b": need to pin. If we ever remove b, then its array 1522 // accesses would start to float, since we don't pin at that point. 1523 // 3. If we move from regular if: don't pin. All array accesses are already assumed to be pinned. 1524 bool pin_array_access_nodes = n->Opcode() == Op_RangeCheck && 1525 prevdom->in(0)->Opcode() != Op_RangeCheck; 1526 dominated_by(prevdom->as_IfProj(), n->as_If(), false, pin_array_access_nodes); 1527 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } ); 1528 return; 1529 } 1530 prevdom = dom; 1531 dom = idom(prevdom); 1532 } 1533 } 1534 } 1535 1536 try_sink_out_of_loop(n); 1537 if (C->failing()) { 1538 return; 1539 } 1540 1541 try_move_store_after_loop(n); 1542 } 1543 1544 // Transform: 1545 // 1546 // if (some_condition) { 1547 // // body 1 1548 // } else { 1549 // // body 2 1550 // } 1551 // if (some_condition) { 1552 // // body 3 1553 // } else { 1554 // // body 4 1555 // } 1556 // 1557 // into: 1558 // 1559 // 1560 // if (some_condition) { 1561 // // body 1 1562 // // body 3 1563 // } else { 1564 // // body 2 1565 // // body 4 1566 // } 1567 bool PhaseIdealLoop::try_merge_identical_ifs(Node* n) { 1568 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) { 1569 Node *n_ctrl = n->in(0); 1570 IfNode* dom_if = idom(n_ctrl)->as_If(); 1571 if (n->in(1) != dom_if->in(1)) { 1572 assert(n->in(1)->in(1)->is_SubTypeCheck() && 1573 (n->in(1)->in(1)->as_SubTypeCheck()->method() != nullptr || 1574 dom_if->in(1)->in(1)->as_SubTypeCheck()->method() != nullptr), "only for subtype checks with profile data attached"); 1575 _igvn.replace_input_of(n, 1, dom_if->in(1)); 1576 } 1577 ProjNode* dom_proj_true = dom_if->proj_out(1); 1578 ProjNode* dom_proj_false = dom_if->proj_out(0); 1579 1580 // Now split the IF 1581 RegionNode* new_false_region; 1582 RegionNode* new_true_region; 1583 do_split_if(n, &new_false_region, &new_true_region); 1584 assert(new_false_region->req() == new_true_region->req(), ""); 1585 #ifdef ASSERT 1586 for (uint i = 1; i < new_false_region->req(); ++i) { 1587 assert(new_false_region->in(i)->in(0) == new_true_region->in(i)->in(0), "unexpected shape following split if"); 1588 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"); 1589 } 1590 #endif 1591 assert(new_false_region->in(1)->in(0)->in(1) == dom_if->in(1), "dominating if and dominated if after split must share test"); 1592 1593 // We now have: 1594 // if (some_condition) { 1595 // // body 1 1596 // if (some_condition) { 1597 // body3: // new_true_region 1598 // // body3 1599 // } else { 1600 // goto body4; 1601 // } 1602 // } else { 1603 // // body 2 1604 // if (some_condition) { 1605 // goto body3; 1606 // } else { 1607 // body4: // new_false_region 1608 // // body4; 1609 // } 1610 // } 1611 // 1612 1613 // clone pinned nodes thru the resulting regions 1614 push_pinned_nodes_thru_region(dom_if, new_true_region); 1615 push_pinned_nodes_thru_region(dom_if, new_false_region); 1616 1617 // Optimize out the cloned ifs. Because pinned nodes were cloned, this also allows a CastPP that would be dependent 1618 // 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 1619 // unrelated control dependency. 1620 for (uint i = 1; i < new_false_region->req(); i++) { 1621 if (is_dominator(dom_proj_true, new_false_region->in(i))) { 1622 dominated_by(dom_proj_true->as_IfProj(), new_false_region->in(i)->in(0)->as_If()); 1623 } else { 1624 assert(is_dominator(dom_proj_false, new_false_region->in(i)), "bad if"); 1625 dominated_by(dom_proj_false->as_IfProj(), new_false_region->in(i)->in(0)->as_If()); 1626 } 1627 } 1628 return true; 1629 } 1630 return false; 1631 } 1632 1633 void PhaseIdealLoop::push_pinned_nodes_thru_region(IfNode* dom_if, Node* region) { 1634 for (DUIterator i = region->outs(); region->has_out(i); i++) { 1635 Node* u = region->out(i); 1636 if (!has_ctrl(u) || u->is_Phi() || !u->depends_only_on_test() || !_igvn.no_dependent_zero_check(u)) { 1637 continue; 1638 } 1639 assert(u->in(0) == region, "not a control dependent node?"); 1640 uint j = 1; 1641 for (; j < u->req(); ++j) { 1642 Node* in = u->in(j); 1643 if (!is_dominator(ctrl_or_self(in), dom_if)) { 1644 break; 1645 } 1646 } 1647 if (j == u->req()) { 1648 Node *phi = PhiNode::make_blank(region, u); 1649 for (uint k = 1; k < region->req(); ++k) { 1650 Node* clone = u->clone(); 1651 clone->set_req(0, region->in(k)); 1652 register_new_node(clone, region->in(k)); 1653 phi->init_req(k, clone); 1654 } 1655 register_new_node(phi, region); 1656 _igvn.replace_node(u, phi); 1657 --i; 1658 } 1659 } 1660 } 1661 1662 bool PhaseIdealLoop::safe_for_if_replacement(const Node* dom) const { 1663 if (!dom->is_CountedLoopEnd()) { 1664 return true; 1665 } 1666 CountedLoopEndNode* le = dom->as_CountedLoopEnd(); 1667 CountedLoopNode* cl = le->loopnode(); 1668 if (cl == nullptr) { 1669 return true; 1670 } 1671 if (!cl->is_main_loop()) { 1672 return true; 1673 } 1674 if (cl->is_canonical_loop_entry() == nullptr) { 1675 return true; 1676 } 1677 // Further unrolling is possible so loop exit condition might change 1678 return false; 1679 } 1680 1681 // See if a shared loop-varying computation has no loop-varying uses. 1682 // Happens if something is only used for JVM state in uncommon trap exits, 1683 // like various versions of induction variable+offset. Clone the 1684 // computation per usage to allow it to sink out of the loop. 1685 void PhaseIdealLoop::try_sink_out_of_loop(Node* n) { 1686 bool is_raw_to_oop_cast = n->is_ConstraintCast() && 1687 n->in(1)->bottom_type()->isa_rawptr() && 1688 !n->bottom_type()->isa_rawptr(); 1689 1690 if (has_ctrl(n) && 1691 !n->is_Phi() && 1692 !n->is_Bool() && 1693 !n->is_Proj() && 1694 !n->is_MergeMem() && 1695 !n->is_CMove() && 1696 !n->is_OpaqueNotNull() && 1697 !n->is_OpaqueInitializedAssertionPredicate() && 1698 !n->is_OpaqueTemplateAssertionPredicate() && 1699 !is_raw_to_oop_cast && // don't extend live ranges of raw oops 1700 (KillPathsReachableByDeadTypeNode || !n->is_Type()) 1701 ) { 1702 Node *n_ctrl = get_ctrl(n); 1703 IdealLoopTree *n_loop = get_loop(n_ctrl); 1704 1705 if (n->in(0) != nullptr) { 1706 IdealLoopTree* loop_ctrl = get_loop(n->in(0)); 1707 if (n_loop != loop_ctrl && n_loop->is_member(loop_ctrl)) { 1708 // n has a control input inside a loop but get_ctrl() is member of an outer loop. This could happen, for example, 1709 // for Div nodes inside a loop (control input inside loop) without a use except for an UCT (outside the loop). 1710 // Rewire control of n to right outside of the loop, regardless if its input(s) are later sunk or not. 1711 Node* maybe_pinned_n = n; 1712 Node* outside_ctrl = place_outside_loop(n_ctrl, loop_ctrl); 1713 if (n->depends_only_on_test()) { 1714 Node* pinned_clone = n->pin_array_access_node(); 1715 if (pinned_clone != nullptr) { 1716 // Pin array access nodes: if this is an array load, it's going to be dependent on a condition that's not a 1717 // range check for that access. If that condition is replaced by an identical dominating one, then an 1718 // unpinned load would risk floating above its range check. 1719 register_new_node(pinned_clone, n_ctrl); 1720 maybe_pinned_n = pinned_clone; 1721 _igvn.replace_node(n, pinned_clone); 1722 } 1723 } 1724 _igvn.replace_input_of(maybe_pinned_n, 0, outside_ctrl); 1725 } 1726 } 1727 if (n_loop != _ltree_root && n->outcnt() > 1) { 1728 // Compute early control: needed for anti-dependence analysis. It's also possible that as a result of 1729 // previous transformations in this loop opts round, the node can be hoisted now: early control will tell us. 1730 Node* early_ctrl = compute_early_ctrl(n, n_ctrl); 1731 if (n_loop->is_member(get_loop(early_ctrl)) && // check that this one can't be hoisted now 1732 ctrl_of_all_uses_out_of_loop(n, early_ctrl, n_loop)) { // All uses in outer loops! 1733 if (n->is_Store() || n->is_LoadStore()) { 1734 assert(false, "no node with a side effect"); 1735 C->record_failure("no node with a side effect"); 1736 return; 1737 } 1738 Node* outer_loop_clone = nullptr; 1739 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin;) { 1740 Node* u = n->last_out(j); // Clone private computation per use 1741 _igvn.rehash_node_delayed(u); 1742 Node* x = nullptr; 1743 if (n->depends_only_on_test()) { 1744 // Pin array access nodes: if this is an array load, it's going to be dependent on a condition that's not a 1745 // range check for that access. If that condition is replaced by an identical dominating one, then an 1746 // unpinned load would risk floating above its range check. 1747 x = n->pin_array_access_node(); 1748 } 1749 if (x == nullptr) { 1750 x = n->clone(); 1751 } 1752 Node* x_ctrl = nullptr; 1753 if (u->is_Phi()) { 1754 // Replace all uses of normal nodes. Replace Phi uses 1755 // individually, so the separate Nodes can sink down 1756 // different paths. 1757 uint k = 1; 1758 while (u->in(k) != n) k++; 1759 u->set_req(k, x); 1760 // x goes next to Phi input path 1761 x_ctrl = u->in(0)->in(k); 1762 // Find control for 'x' next to use but not inside inner loops. 1763 x_ctrl = place_outside_loop(x_ctrl, n_loop); 1764 --j; 1765 } else { // Normal use 1766 if (has_ctrl(u)) { 1767 x_ctrl = get_ctrl(u); 1768 } else { 1769 x_ctrl = u->in(0); 1770 } 1771 // Find control for 'x' next to use but not inside inner loops. 1772 x_ctrl = place_outside_loop(x_ctrl, n_loop); 1773 // Replace all uses 1774 if (u->is_ConstraintCast() && _igvn.type(n)->higher_equal(u->bottom_type()) && u->in(0) == x_ctrl) { 1775 // If we're sinking a chain of data nodes, we might have inserted a cast to pin the use which is not necessary 1776 // anymore now that we're going to pin n as well 1777 _igvn.replace_node(u, x); 1778 --j; 1779 } else { 1780 int nb = u->replace_edge(n, x, &_igvn); 1781 j -= nb; 1782 } 1783 } 1784 1785 if (n->is_Load()) { 1786 // For loads, add a control edge to a CFG node outside of the loop 1787 // to force them to not combine and return back inside the loop 1788 // during GVN optimization (4641526). 1789 assert(x_ctrl == get_late_ctrl_with_anti_dep(x->as_Load(), early_ctrl, x_ctrl), "anti-dependences were already checked"); 1790 1791 IdealLoopTree* x_loop = get_loop(x_ctrl); 1792 Node* x_head = x_loop->_head; 1793 if (x_head->is_Loop() && x_head->is_OuterStripMinedLoop()) { 1794 // Do not add duplicate LoadNodes to the outer strip mined loop 1795 if (outer_loop_clone != nullptr) { 1796 _igvn.replace_node(x, outer_loop_clone); 1797 continue; 1798 } 1799 outer_loop_clone = x; 1800 } 1801 x->set_req(0, x_ctrl); 1802 } else if (n->in(0) != nullptr){ 1803 x->set_req(0, x_ctrl); 1804 } 1805 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone"); 1806 assert(!n_loop->is_member(get_loop(x_ctrl)), "should have moved out of loop"); 1807 register_new_node(x, x_ctrl); 1808 1809 // Chain of AddP nodes: (AddP base (AddP base (AddP base ))) 1810 // All AddP nodes must keep the same base after sinking so: 1811 // 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, 1812 // their bases remain the same. 1813 // (see 2- below) 1814 assert(!x->is_AddP() || !x->in(AddPNode::Address)->is_AddP() || 1815 x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base) || 1816 !x->in(AddPNode::Address)->in(AddPNode::Base)->eqv_uncast(x->in(AddPNode::Base)), "unexpected AddP shape"); 1817 if (x->in(0) == nullptr && !x->is_DecodeNarrowPtr() && 1818 !(x->is_AddP() && x->in(AddPNode::Address)->is_AddP() && x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base))) { 1819 assert(!x->is_Load(), "load should be pinned"); 1820 // Use a cast node to pin clone out of loop 1821 Node* cast = nullptr; 1822 for (uint k = 0; k < x->req(); k++) { 1823 Node* in = x->in(k); 1824 if (in != nullptr && n_loop->is_member(get_loop(get_ctrl(in)))) { 1825 const Type* in_t = _igvn.type(in); 1826 cast = ConstraintCastNode::make_cast_for_type(x_ctrl, in, in_t, 1827 ConstraintCastNode::UnconditionalDependency, nullptr); 1828 } 1829 if (cast != nullptr) { 1830 Node* prev = _igvn.hash_find_insert(cast); 1831 if (prev != nullptr && get_ctrl(prev) == x_ctrl) { 1832 cast->destruct(&_igvn); 1833 cast = prev; 1834 } else { 1835 register_new_node(cast, x_ctrl); 1836 } 1837 x->replace_edge(in, cast); 1838 // Chain of AddP nodes: 1839 // 2- A CastPP of the base is only added now that all AddP nodes are sunk 1840 if (x->is_AddP() && k == AddPNode::Base) { 1841 update_addp_chain_base(x, n->in(AddPNode::Base), cast); 1842 } 1843 break; 1844 } 1845 } 1846 assert(cast != nullptr, "must have added a cast to pin the node"); 1847 } 1848 } 1849 _igvn.remove_dead_node(n); 1850 } 1851 _dom_lca_tags_round = 0; 1852 } 1853 } 1854 } 1855 1856 void PhaseIdealLoop::update_addp_chain_base(Node* x, Node* old_base, Node* new_base) { 1857 ResourceMark rm; 1858 Node_List wq; 1859 wq.push(x); 1860 while (wq.size() != 0) { 1861 Node* n = wq.pop(); 1862 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1863 Node* u = n->fast_out(i); 1864 if (u->is_AddP() && u->in(AddPNode::Base) == old_base) { 1865 _igvn.replace_input_of(u, AddPNode::Base, new_base); 1866 wq.push(u); 1867 } 1868 } 1869 } 1870 } 1871 1872 // Compute the early control of a node by following its inputs until we reach 1873 // nodes that are pinned. Then compute the LCA of the control of all pinned nodes. 1874 Node* PhaseIdealLoop::compute_early_ctrl(Node* n, Node* n_ctrl) { 1875 Node* early_ctrl = nullptr; 1876 ResourceMark rm; 1877 Unique_Node_List wq; 1878 wq.push(n); 1879 for (uint i = 0; i < wq.size(); i++) { 1880 Node* m = wq.at(i); 1881 Node* c = nullptr; 1882 if (m->is_CFG()) { 1883 c = m; 1884 } else if (m->pinned()) { 1885 c = m->in(0); 1886 } else { 1887 for (uint j = 0; j < m->req(); j++) { 1888 Node* in = m->in(j); 1889 if (in != nullptr) { 1890 wq.push(in); 1891 } 1892 } 1893 } 1894 if (c != nullptr) { 1895 assert(is_dominator(c, n_ctrl), "control input must dominate current control"); 1896 if (early_ctrl == nullptr || is_dominator(early_ctrl, c)) { 1897 early_ctrl = c; 1898 } 1899 } 1900 } 1901 assert(is_dominator(early_ctrl, n_ctrl), "early control must dominate current control"); 1902 return early_ctrl; 1903 } 1904 1905 bool PhaseIdealLoop::ctrl_of_all_uses_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop) { 1906 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1907 Node* u = n->fast_out(i); 1908 if (u->is_Opaque1()) { 1909 return false; // Found loop limit, bugfix for 4677003 1910 } 1911 // We can't reuse tags in PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal() so make sure calls to 1912 // get_late_ctrl_with_anti_dep() use their own tag 1913 _dom_lca_tags_round++; 1914 assert(_dom_lca_tags_round != 0, "shouldn't wrap around"); 1915 1916 if (u->is_Phi()) { 1917 for (uint j = 1; j < u->req(); ++j) { 1918 if (u->in(j) == n && !ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, u->in(0)->in(j))) { 1919 return false; 1920 } 1921 } 1922 } else { 1923 Node* ctrl = has_ctrl(u) ? get_ctrl(u) : u->in(0); 1924 if (!ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, ctrl)) { 1925 return false; 1926 } 1927 } 1928 } 1929 return true; 1930 } 1931 1932 bool PhaseIdealLoop::ctrl_of_use_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop, Node* ctrl) { 1933 if (n->is_Load()) { 1934 ctrl = get_late_ctrl_with_anti_dep(n->as_Load(), n_ctrl, ctrl); 1935 } 1936 IdealLoopTree *u_loop = get_loop(ctrl); 1937 if (u_loop == n_loop) { 1938 return false; // Found loop-varying use 1939 } 1940 if (n_loop->is_member(u_loop)) { 1941 return false; // Found use in inner loop 1942 } 1943 // 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 1944 // to a check that's eliminated by range check elimination, it becomes input to an expression that feeds into the exit 1945 // test of the pre loop above the point in the graph where it's pinned. 1946 if (n_loop->_head->is_CountedLoop() && n_loop->_head->as_CountedLoop()->is_pre_loop()) { 1947 CountedLoopNode* pre_loop = n_loop->_head->as_CountedLoop(); 1948 if (is_dominator(pre_loop->loopexit(), ctrl)) { 1949 return false; 1950 } 1951 } 1952 return true; 1953 } 1954 1955 //------------------------------split_if_with_blocks--------------------------- 1956 // Check for aggressive application of 'split-if' optimization, 1957 // using basic block level info. 1958 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack) { 1959 Node* root = C->root(); 1960 visited.set(root->_idx); // first, mark root as visited 1961 // Do pre-visit work for root 1962 Node* n = split_if_with_blocks_pre(root); 1963 uint cnt = n->outcnt(); 1964 uint i = 0; 1965 1966 while (true) { 1967 // Visit all children 1968 if (i < cnt) { 1969 Node* use = n->raw_out(i); 1970 ++i; 1971 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) { 1972 // Now do pre-visit work for this use 1973 use = split_if_with_blocks_pre(use); 1974 nstack.push(n, i); // Save parent and next use's index. 1975 n = use; // Process all children of current use. 1976 cnt = use->outcnt(); 1977 i = 0; 1978 } 1979 } 1980 else { 1981 // All of n's children have been processed, complete post-processing. 1982 if (cnt != 0 && !n->is_Con()) { 1983 assert(has_node(n), "no dead nodes"); 1984 split_if_with_blocks_post(n); 1985 if (C->failing()) { 1986 return; 1987 } 1988 } 1989 if (must_throttle_split_if()) { 1990 nstack.clear(); 1991 } 1992 if (nstack.is_empty()) { 1993 // Finished all nodes on stack. 1994 break; 1995 } 1996 // Get saved parent node and next use's index. Visit the rest of uses. 1997 n = nstack.node(); 1998 cnt = n->outcnt(); 1999 i = nstack.index(); 2000 nstack.pop(); 2001 } 2002 } 2003 } 2004 2005 2006 //============================================================================= 2007 // 2008 // C L O N E A L O O P B O D Y 2009 // 2010 2011 //------------------------------clone_iff-------------------------------------- 2012 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 2013 // "Nearly" because all Nodes have been cloned from the original in the loop, 2014 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 2015 // through the Phi recursively, and return a Bool. 2016 Node* PhaseIdealLoop::clone_iff(PhiNode* phi) { 2017 2018 // Convert this Phi into a Phi merging Bools 2019 uint i; 2020 for (i = 1; i < phi->req(); i++) { 2021 Node* b = phi->in(i); 2022 if (b->is_Phi()) { 2023 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi())); 2024 } else { 2025 assert(b->is_Bool() || b->is_OpaqueNotNull() || b->is_OpaqueInitializedAssertionPredicate(), 2026 "bool, non-null check with OpaqueNotNull or Initialized Assertion Predicate with its Opaque node"); 2027 } 2028 } 2029 Node* n = phi->in(1); 2030 Node* sample_opaque = nullptr; 2031 Node *sample_bool = nullptr; 2032 if (n->is_OpaqueNotNull() || n->is_OpaqueInitializedAssertionPredicate()) { 2033 sample_opaque = n; 2034 sample_bool = n->in(1); 2035 assert(sample_bool->is_Bool(), "wrong type"); 2036 } else { 2037 sample_bool = n; 2038 } 2039 Node *sample_cmp = sample_bool->in(1); 2040 2041 // Make Phis to merge the Cmp's inputs. 2042 PhiNode *phi1 = new PhiNode(phi->in(0), Type::TOP); 2043 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP); 2044 for (i = 1; i < phi->req(); i++) { 2045 Node *n1 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1); 2046 Node *n2 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2); 2047 phi1->set_req(i, n1); 2048 phi2->set_req(i, n2); 2049 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 2050 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 2051 } 2052 // See if these Phis have been made before. 2053 // Register with optimizer 2054 Node *hit1 = _igvn.hash_find_insert(phi1); 2055 if (hit1) { // Hit, toss just made Phi 2056 _igvn.remove_dead_node(phi1); // Remove new phi 2057 assert(hit1->is_Phi(), "" ); 2058 phi1 = (PhiNode*)hit1; // Use existing phi 2059 } else { // Miss 2060 _igvn.register_new_node_with_optimizer(phi1); 2061 } 2062 Node *hit2 = _igvn.hash_find_insert(phi2); 2063 if (hit2) { // Hit, toss just made Phi 2064 _igvn.remove_dead_node(phi2); // Remove new phi 2065 assert(hit2->is_Phi(), "" ); 2066 phi2 = (PhiNode*)hit2; // Use existing phi 2067 } else { // Miss 2068 _igvn.register_new_node_with_optimizer(phi2); 2069 } 2070 // Register Phis with loop/block info 2071 set_ctrl(phi1, phi->in(0)); 2072 set_ctrl(phi2, phi->in(0)); 2073 // Make a new Cmp 2074 Node *cmp = sample_cmp->clone(); 2075 cmp->set_req(1, phi1); 2076 cmp->set_req(2, phi2); 2077 _igvn.register_new_node_with_optimizer(cmp); 2078 set_ctrl(cmp, phi->in(0)); 2079 2080 // Make a new Bool 2081 Node *b = sample_bool->clone(); 2082 b->set_req(1,cmp); 2083 _igvn.register_new_node_with_optimizer(b); 2084 set_ctrl(b, phi->in(0)); 2085 2086 if (sample_opaque != nullptr) { 2087 Node* opaque = sample_opaque->clone(); 2088 opaque->set_req(1, b); 2089 _igvn.register_new_node_with_optimizer(opaque); 2090 set_ctrl(opaque, phi->in(0)); 2091 return opaque; 2092 } 2093 2094 assert(b->is_Bool(), ""); 2095 return b; 2096 } 2097 2098 //------------------------------clone_bool------------------------------------- 2099 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 2100 // "Nearly" because all Nodes have been cloned from the original in the loop, 2101 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 2102 // through the Phi recursively, and return a Bool. 2103 CmpNode*PhaseIdealLoop::clone_bool(PhiNode* phi) { 2104 uint i; 2105 // Convert this Phi into a Phi merging Bools 2106 for( i = 1; i < phi->req(); i++ ) { 2107 Node *b = phi->in(i); 2108 if( b->is_Phi() ) { 2109 _igvn.replace_input_of(phi, i, clone_bool(b->as_Phi())); 2110 } else { 2111 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" ); 2112 } 2113 } 2114 2115 Node *sample_cmp = phi->in(1); 2116 2117 // Make Phis to merge the Cmp's inputs. 2118 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP ); 2119 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP ); 2120 for( uint j = 1; j < phi->req(); j++ ) { 2121 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP 2122 Node *n1, *n2; 2123 if( cmp_top->is_Cmp() ) { 2124 n1 = cmp_top->in(1); 2125 n2 = cmp_top->in(2); 2126 } else { 2127 n1 = n2 = cmp_top; 2128 } 2129 phi1->set_req( j, n1 ); 2130 phi2->set_req( j, n2 ); 2131 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 2132 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 2133 } 2134 2135 // See if these Phis have been made before. 2136 // Register with optimizer 2137 Node *hit1 = _igvn.hash_find_insert(phi1); 2138 if( hit1 ) { // Hit, toss just made Phi 2139 _igvn.remove_dead_node(phi1); // Remove new phi 2140 assert( hit1->is_Phi(), "" ); 2141 phi1 = (PhiNode*)hit1; // Use existing phi 2142 } else { // Miss 2143 _igvn.register_new_node_with_optimizer(phi1); 2144 } 2145 Node *hit2 = _igvn.hash_find_insert(phi2); 2146 if( hit2 ) { // Hit, toss just made Phi 2147 _igvn.remove_dead_node(phi2); // Remove new phi 2148 assert( hit2->is_Phi(), "" ); 2149 phi2 = (PhiNode*)hit2; // Use existing phi 2150 } else { // Miss 2151 _igvn.register_new_node_with_optimizer(phi2); 2152 } 2153 // Register Phis with loop/block info 2154 set_ctrl(phi1, phi->in(0)); 2155 set_ctrl(phi2, phi->in(0)); 2156 // Make a new Cmp 2157 Node *cmp = sample_cmp->clone(); 2158 cmp->set_req( 1, phi1 ); 2159 cmp->set_req( 2, phi2 ); 2160 _igvn.register_new_node_with_optimizer(cmp); 2161 set_ctrl(cmp, phi->in(0)); 2162 2163 assert( cmp->is_Cmp(), "" ); 2164 return (CmpNode*)cmp; 2165 } 2166 2167 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new, 2168 IdealLoopTree* loop, IdealLoopTree* outer_loop, 2169 Node_List*& split_if_set, Node_List*& split_bool_set, 2170 Node_List*& split_cex_set, Node_List& worklist, 2171 uint new_counter, CloneLoopMode mode) { 2172 Node* nnn = old_new[old->_idx]; 2173 // Copy uses to a worklist, so I can munge the def-use info 2174 // with impunity. 2175 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 2176 worklist.push(old->fast_out(j)); 2177 2178 while( worklist.size() ) { 2179 Node *use = worklist.pop(); 2180 if (!has_node(use)) continue; // Ignore dead nodes 2181 if (use->in(0) == C->top()) continue; 2182 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 2183 // Check for data-use outside of loop - at least one of OLD or USE 2184 // must not be a CFG node. 2185 #ifdef ASSERT 2186 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == nullptr) { 2187 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint(); 2188 assert(mode != IgnoreStripMined, "incorrect cloning mode"); 2189 assert((mode == ControlAroundStripMined && use == sfpt) || !use->is_reachable_from_root(), "missed a node"); 2190 } 2191 #endif 2192 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) { 2193 2194 // If the Data use is an IF, that means we have an IF outside the 2195 // loop that is switching on a condition that is set inside the 2196 // loop. Happens if people set a loop-exit flag; then test the flag 2197 // in the loop to break the loop, then test is again outside the 2198 // loop to determine which way the loop exited. 2199 // 2200 // For several uses we need to make sure that there is no phi between, 2201 // the use and the Bool/Cmp. We therefore clone the Bool/Cmp down here 2202 // to avoid such a phi in between. 2203 // For example, it is unexpected that there is a Phi between an 2204 // AllocateArray node and its ValidLengthTest input that could cause 2205 // split if to break. 2206 assert(!use->is_OpaqueTemplateAssertionPredicate(), 2207 "should not clone a Template Assertion Predicate which should be removed once it's useless"); 2208 if (use->is_If() || use->is_CMove() || use->is_OpaqueNotNull() || use->is_OpaqueInitializedAssertionPredicate() || 2209 (use->Opcode() == Op_AllocateArray && use->in(AllocateNode::ValidLengthTest) == old)) { 2210 // Since this code is highly unlikely, we lazily build the worklist 2211 // of such Nodes to go split. 2212 if (!split_if_set) { 2213 split_if_set = new Node_List(); 2214 } 2215 split_if_set->push(use); 2216 } 2217 if (use->is_Bool()) { 2218 if (!split_bool_set) { 2219 split_bool_set = new Node_List(); 2220 } 2221 split_bool_set->push(use); 2222 } 2223 if (use->Opcode() == Op_CreateEx) { 2224 if (!split_cex_set) { 2225 split_cex_set = new Node_List(); 2226 } 2227 split_cex_set->push(use); 2228 } 2229 2230 2231 // Get "block" use is in 2232 uint idx = 0; 2233 while( use->in(idx) != old ) idx++; 2234 Node *prev = use->is_CFG() ? use : get_ctrl(use); 2235 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" ); 2236 Node* cfg = (prev->_idx >= new_counter && prev->is_Region()) 2237 ? prev->in(2) 2238 : idom(prev); 2239 if( use->is_Phi() ) // Phi use is in prior block 2240 cfg = prev->in(idx); // NOT in block of Phi itself 2241 if (cfg->is_top()) { // Use is dead? 2242 _igvn.replace_input_of(use, idx, C->top()); 2243 continue; 2244 } 2245 2246 // If use is referenced through control edge... (idx == 0) 2247 if (mode == IgnoreStripMined && idx == 0) { 2248 LoopNode *head = loop->_head->as_Loop(); 2249 if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) { 2250 // That node is outside the inner loop, leave it outside the 2251 // outer loop as well to not confuse verification code. 2252 assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop"); 2253 _igvn.replace_input_of(use, 0, head->outer_loop_exit()); 2254 continue; 2255 } 2256 } 2257 2258 while(!outer_loop->is_member(get_loop(cfg))) { 2259 prev = cfg; 2260 cfg = (cfg->_idx >= new_counter && cfg->is_Region()) ? cfg->in(2) : idom(cfg); 2261 } 2262 // If the use occurs after merging several exits from the loop, then 2263 // old value must have dominated all those exits. Since the same old 2264 // value was used on all those exits we did not need a Phi at this 2265 // merge point. NOW we do need a Phi here. Each loop exit value 2266 // is now merged with the peeled body exit; each exit gets its own 2267 // private Phi and those Phis need to be merged here. 2268 Node *phi; 2269 if( prev->is_Region() ) { 2270 if( idx == 0 ) { // Updating control edge? 2271 phi = prev; // Just use existing control 2272 } else { // Else need a new Phi 2273 phi = PhiNode::make( prev, old ); 2274 // Now recursively fix up the new uses of old! 2275 for( uint i = 1; i < prev->req(); i++ ) { 2276 worklist.push(phi); // Onto worklist once for each 'old' input 2277 } 2278 } 2279 } else { 2280 // Get new RegionNode merging old and new loop exits 2281 prev = old_new[prev->_idx]; 2282 assert( prev, "just made this in step 7" ); 2283 if( idx == 0) { // Updating control edge? 2284 phi = prev; // Just use existing control 2285 } else { // Else need a new Phi 2286 // Make a new Phi merging data values properly 2287 phi = PhiNode::make( prev, old ); 2288 phi->set_req( 1, nnn ); 2289 } 2290 } 2291 // If inserting a new Phi, check for prior hits 2292 if( idx != 0 ) { 2293 Node *hit = _igvn.hash_find_insert(phi); 2294 if( hit == nullptr ) { 2295 _igvn.register_new_node_with_optimizer(phi); // Register new phi 2296 } else { // or 2297 // Remove the new phi from the graph and use the hit 2298 _igvn.remove_dead_node(phi); 2299 phi = hit; // Use existing phi 2300 } 2301 set_ctrl(phi, prev); 2302 } 2303 // Make 'use' use the Phi instead of the old loop body exit value 2304 assert(use->in(idx) == old, "old is still input of use"); 2305 // We notify all uses of old, including use, and the indirect uses, 2306 // that may now be optimized because we have replaced old with phi. 2307 _igvn.add_users_to_worklist(old); 2308 if (idx == 0 && 2309 use->depends_only_on_test()) { 2310 Node* pinned_clone = use->pin_array_access_node(); 2311 if (pinned_clone != nullptr) { 2312 // Pin array access nodes: control is updated here to a region. If, after some transformations, only one path 2313 // into the region is left, an array load could become dependent on a condition that's not a range check for 2314 // that access. If that condition is replaced by an identical dominating one, then an unpinned load would risk 2315 // floating above its range check. 2316 pinned_clone->set_req(0, phi); 2317 register_new_node_with_ctrl_of(pinned_clone, use); 2318 _igvn.replace_node(use, pinned_clone); 2319 continue; 2320 } 2321 } 2322 _igvn.replace_input_of(use, idx, phi); 2323 if( use->_idx >= new_counter ) { // If updating new phis 2324 // Not needed for correctness, but prevents a weak assert 2325 // in AddPNode from tripping (when we end up with different 2326 // base & derived Phis that will become the same after 2327 // IGVN does CSE). 2328 Node *hit = _igvn.hash_find_insert(use); 2329 if( hit ) // Go ahead and re-hash for hits. 2330 _igvn.replace_node( use, hit ); 2331 } 2332 } 2333 } 2334 } 2335 2336 static void collect_nodes_in_outer_loop_not_reachable_from_sfpt(Node* n, const IdealLoopTree *loop, const IdealLoopTree* outer_loop, 2337 const Node_List &old_new, Unique_Node_List& wq, PhaseIdealLoop* phase, 2338 bool check_old_new) { 2339 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2340 Node* u = n->fast_out(j); 2341 assert(check_old_new || old_new[u->_idx] == nullptr, "shouldn't have been cloned"); 2342 if (!u->is_CFG() && (!check_old_new || old_new[u->_idx] == nullptr)) { 2343 Node* c = phase->get_ctrl(u); 2344 IdealLoopTree* u_loop = phase->get_loop(c); 2345 assert(!loop->is_member(u_loop) || !loop->_body.contains(u), "can be in outer loop or out of both loops only"); 2346 if (!loop->is_member(u_loop)) { 2347 if (outer_loop->is_member(u_loop)) { 2348 wq.push(u); 2349 } else { 2350 // nodes pinned with control in the outer loop but not referenced from the safepoint must be moved out of 2351 // the outer loop too 2352 Node* u_c = u->in(0); 2353 if (u_c != nullptr) { 2354 IdealLoopTree* u_c_loop = phase->get_loop(u_c); 2355 if (outer_loop->is_member(u_c_loop) && !loop->is_member(u_c_loop)) { 2356 wq.push(u); 2357 } 2358 } 2359 } 2360 } 2361 } 2362 } 2363 } 2364 2365 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop, 2366 IdealLoopTree* outer_loop, int dd, Node_List &old_new, 2367 Node_List& extra_data_nodes) { 2368 if (head->is_strip_mined() && mode != IgnoreStripMined) { 2369 CountedLoopNode* cl = head->as_CountedLoop(); 2370 Node* l = cl->outer_loop(); 2371 Node* tail = cl->outer_loop_tail(); 2372 IfNode* le = cl->outer_loop_end(); 2373 Node* sfpt = cl->outer_safepoint(); 2374 CountedLoopEndNode* cle = cl->loopexit(); 2375 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop(); 2376 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null(); 2377 Node* cle_out = cle->proj_out(false); 2378 2379 Node* new_sfpt = nullptr; 2380 Node* new_cle_out = cle_out->clone(); 2381 old_new.map(cle_out->_idx, new_cle_out); 2382 if (mode == CloneIncludesStripMined) { 2383 // clone outer loop body 2384 Node* new_l = l->clone(); 2385 Node* new_tail = tail->clone(); 2386 IfNode* new_le = le->clone()->as_If(); 2387 new_sfpt = sfpt->clone(); 2388 2389 set_loop(new_l, outer_loop->_parent); 2390 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd); 2391 set_loop(new_cle_out, outer_loop->_parent); 2392 set_idom(new_cle_out, new_cle, dd); 2393 set_loop(new_sfpt, outer_loop->_parent); 2394 set_idom(new_sfpt, new_cle_out, dd); 2395 set_loop(new_le, outer_loop->_parent); 2396 set_idom(new_le, new_sfpt, dd); 2397 set_loop(new_tail, outer_loop->_parent); 2398 set_idom(new_tail, new_le, dd); 2399 set_idom(new_cl, new_l, dd); 2400 2401 old_new.map(l->_idx, new_l); 2402 old_new.map(tail->_idx, new_tail); 2403 old_new.map(le->_idx, new_le); 2404 old_new.map(sfpt->_idx, new_sfpt); 2405 2406 new_l->set_req(LoopNode::LoopBackControl, new_tail); 2407 new_l->set_req(0, new_l); 2408 new_tail->set_req(0, new_le); 2409 new_le->set_req(0, new_sfpt); 2410 new_sfpt->set_req(0, new_cle_out); 2411 new_cle_out->set_req(0, new_cle); 2412 new_cl->set_req(LoopNode::EntryControl, new_l); 2413 2414 _igvn.register_new_node_with_optimizer(new_l); 2415 _igvn.register_new_node_with_optimizer(new_tail); 2416 _igvn.register_new_node_with_optimizer(new_le); 2417 } else { 2418 Node *newhead = old_new[loop->_head->_idx]; 2419 newhead->as_Loop()->clear_strip_mined(); 2420 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl)); 2421 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 2422 } 2423 // Look at data node that were assigned a control in the outer 2424 // loop: they are kept in the outer loop by the safepoint so start 2425 // from the safepoint node's inputs. 2426 IdealLoopTree* outer_loop = get_loop(l); 2427 Node_Stack stack(2); 2428 stack.push(sfpt, 1); 2429 uint new_counter = C->unique(); 2430 while (stack.size() > 0) { 2431 Node* n = stack.node(); 2432 uint i = stack.index(); 2433 while (i < n->req() && 2434 (n->in(i) == nullptr || 2435 !has_ctrl(n->in(i)) || 2436 get_loop(get_ctrl(n->in(i))) != outer_loop || 2437 (old_new[n->in(i)->_idx] != nullptr && old_new[n->in(i)->_idx]->_idx >= new_counter))) { 2438 i++; 2439 } 2440 if (i < n->req()) { 2441 stack.set_index(i+1); 2442 stack.push(n->in(i), 0); 2443 } else { 2444 assert(old_new[n->_idx] == nullptr || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet"); 2445 Node* m = n == sfpt ? new_sfpt : n->clone(); 2446 if (m != nullptr) { 2447 for (uint i = 0; i < n->req(); i++) { 2448 if (m->in(i) != nullptr && old_new[m->in(i)->_idx] != nullptr) { 2449 m->set_req(i, old_new[m->in(i)->_idx]); 2450 } 2451 } 2452 } else { 2453 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?"); 2454 } 2455 if (n != sfpt) { 2456 extra_data_nodes.push(n); 2457 _igvn.register_new_node_with_optimizer(m); 2458 assert(get_ctrl(n) == cle_out, "what other control?"); 2459 set_ctrl(m, new_cle_out); 2460 old_new.map(n->_idx, m); 2461 } 2462 stack.pop(); 2463 } 2464 } 2465 if (mode == CloneIncludesStripMined) { 2466 _igvn.register_new_node_with_optimizer(new_sfpt); 2467 _igvn.register_new_node_with_optimizer(new_cle_out); 2468 } 2469 // Some other transformation may have pessimistically assigned some 2470 // data nodes to the outer loop. Set their control so they are out 2471 // of the outer loop. 2472 ResourceMark rm; 2473 Unique_Node_List wq; 2474 for (uint i = 0; i < extra_data_nodes.size(); i++) { 2475 Node* old = extra_data_nodes.at(i); 2476 collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true); 2477 } 2478 2479 for (uint i = 0; i < loop->_body.size(); i++) { 2480 Node* old = loop->_body.at(i); 2481 collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true); 2482 } 2483 2484 Node* inner_out = sfpt->in(0); 2485 if (inner_out->outcnt() > 1) { 2486 collect_nodes_in_outer_loop_not_reachable_from_sfpt(inner_out, loop, outer_loop, old_new, wq, this, true); 2487 } 2488 2489 Node* new_ctrl = cl->outer_loop_exit(); 2490 assert(get_loop(new_ctrl) != outer_loop, "must be out of the loop nest"); 2491 for (uint i = 0; i < wq.size(); i++) { 2492 Node* n = wq.at(i); 2493 set_ctrl(n, new_ctrl); 2494 if (n->in(0) != nullptr) { 2495 _igvn.replace_input_of(n, 0, new_ctrl); 2496 } 2497 collect_nodes_in_outer_loop_not_reachable_from_sfpt(n, loop, outer_loop, old_new, wq, this, false); 2498 } 2499 } else { 2500 Node *newhead = old_new[loop->_head->_idx]; 2501 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 2502 } 2503 } 2504 2505 //------------------------------clone_loop------------------------------------- 2506 // 2507 // C L O N E A L O O P B O D Y 2508 // 2509 // This is the basic building block of the loop optimizations. It clones an 2510 // entire loop body. It makes an old_new loop body mapping; with this mapping 2511 // you can find the new-loop equivalent to an old-loop node. All new-loop 2512 // nodes are exactly equal to their old-loop counterparts, all edges are the 2513 // same. All exits from the old-loop now have a RegionNode that merges the 2514 // equivalent new-loop path. This is true even for the normal "loop-exit" 2515 // condition. All uses of loop-invariant old-loop values now come from (one 2516 // or more) Phis that merge their new-loop equivalents. 2517 // 2518 // This operation leaves the graph in an illegal state: there are two valid 2519 // control edges coming from the loop pre-header to both loop bodies. I'll 2520 // definitely have to hack the graph after running this transform. 2521 // 2522 // From this building block I will further edit edges to perform loop peeling 2523 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc. 2524 // 2525 // Parameter side_by_size_idom: 2526 // When side_by_size_idom is null, the dominator tree is constructed for 2527 // the clone loop to dominate the original. Used in construction of 2528 // pre-main-post loop sequence. 2529 // When nonnull, the clone and original are side-by-side, both are 2530 // dominated by the side_by_side_idom node. Used in construction of 2531 // unswitched loops. 2532 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd, 2533 CloneLoopMode mode, Node* side_by_side_idom) { 2534 2535 LoopNode* head = loop->_head->as_Loop(); 2536 head->verify_strip_mined(1); 2537 2538 if (C->do_vector_loop() && PrintOpto) { 2539 const char* mname = C->method()->name()->as_quoted_ascii(); 2540 if (mname != nullptr) { 2541 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname); 2542 } 2543 } 2544 2545 CloneMap& cm = C->clone_map(); 2546 if (C->do_vector_loop()) { 2547 cm.set_clone_idx(cm.max_gen()+1); 2548 #ifndef PRODUCT 2549 if (PrintOpto) { 2550 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx()); 2551 loop->dump_head(); 2552 } 2553 #endif 2554 } 2555 2556 // Step 1: Clone the loop body. Make the old->new mapping. 2557 clone_loop_body(loop->_body, old_new, &cm); 2558 2559 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop; 2560 2561 // Step 2: Fix the edges in the new body. If the old input is outside the 2562 // loop use it. If the old input is INside the loop, use the corresponding 2563 // new node instead. 2564 fix_body_edges(loop->_body, loop, old_new, dd, outer_loop->_parent, false); 2565 2566 Node_List extra_data_nodes; // data nodes in the outer strip mined loop 2567 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes); 2568 2569 // Step 3: Now fix control uses. Loop varying control uses have already 2570 // been fixed up (as part of all input edges in Step 2). Loop invariant 2571 // control uses must be either an IfFalse or an IfTrue. Make a merge 2572 // point to merge the old and new IfFalse/IfTrue nodes; make the use 2573 // refer to this. 2574 Node_List worklist; 2575 uint new_counter = C->unique(); 2576 fix_ctrl_uses(loop->_body, loop, old_new, mode, side_by_side_idom, &cm, worklist); 2577 2578 // Step 4: If loop-invariant use is not control, it must be dominated by a 2579 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region 2580 // there if needed. Make a Phi there merging old and new used values. 2581 Node_List *split_if_set = nullptr; 2582 Node_List *split_bool_set = nullptr; 2583 Node_List *split_cex_set = nullptr; 2584 fix_data_uses(loop->_body, loop, mode, outer_loop, new_counter, old_new, worklist, split_if_set, split_bool_set, split_cex_set); 2585 2586 for (uint i = 0; i < extra_data_nodes.size(); i++) { 2587 Node* old = extra_data_nodes.at(i); 2588 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set, 2589 split_bool_set, split_cex_set, worklist, new_counter, 2590 mode); 2591 } 2592 2593 // Check for IFs that need splitting/cloning. Happens if an IF outside of 2594 // the loop uses a condition set in the loop. The original IF probably 2595 // takes control from one or more OLD Regions (which in turn get from NEW 2596 // Regions). In any case, there will be a set of Phis for each merge point 2597 // from the IF up to where the original BOOL def exists the loop. 2598 finish_clone_loop(split_if_set, split_bool_set, split_cex_set); 2599 2600 } 2601 2602 void PhaseIdealLoop::finish_clone_loop(Node_List* split_if_set, Node_List* split_bool_set, Node_List* split_cex_set) { 2603 if (split_if_set) { 2604 while (split_if_set->size()) { 2605 Node *iff = split_if_set->pop(); 2606 uint input = iff->Opcode() == Op_AllocateArray ? AllocateNode::ValidLengthTest : 1; 2607 if (iff->in(input)->is_Phi()) { 2608 Node *b = clone_iff(iff->in(input)->as_Phi()); 2609 _igvn.replace_input_of(iff, input, b); 2610 } 2611 } 2612 } 2613 if (split_bool_set) { 2614 while (split_bool_set->size()) { 2615 Node *b = split_bool_set->pop(); 2616 Node *phi = b->in(1); 2617 assert(phi->is_Phi(), ""); 2618 CmpNode *cmp = clone_bool((PhiNode*) phi); 2619 _igvn.replace_input_of(b, 1, cmp); 2620 } 2621 } 2622 if (split_cex_set) { 2623 while (split_cex_set->size()) { 2624 Node *b = split_cex_set->pop(); 2625 assert(b->in(0)->is_Region(), ""); 2626 assert(b->in(1)->is_Phi(), ""); 2627 assert(b->in(0)->in(0) == b->in(1)->in(0), ""); 2628 split_up(b, b->in(0), nullptr); 2629 } 2630 } 2631 } 2632 2633 void PhaseIdealLoop::fix_data_uses(Node_List& body, IdealLoopTree* loop, CloneLoopMode mode, IdealLoopTree* outer_loop, 2634 uint new_counter, Node_List &old_new, Node_List &worklist, Node_List*& split_if_set, 2635 Node_List*& split_bool_set, Node_List*& split_cex_set) { 2636 for(uint i = 0; i < body.size(); i++ ) { 2637 Node* old = body.at(i); 2638 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set, 2639 split_bool_set, split_cex_set, worklist, new_counter, 2640 mode); 2641 } 2642 } 2643 2644 void PhaseIdealLoop::fix_ctrl_uses(const Node_List& body, const IdealLoopTree* loop, Node_List &old_new, CloneLoopMode mode, 2645 Node* side_by_side_idom, CloneMap* cm, Node_List &worklist) { 2646 LoopNode* head = loop->_head->as_Loop(); 2647 for(uint i = 0; i < body.size(); i++ ) { 2648 Node* old = body.at(i); 2649 if( !old->is_CFG() ) continue; 2650 2651 // Copy uses to a worklist, so I can munge the def-use info 2652 // with impunity. 2653 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) { 2654 worklist.push(old->fast_out(j)); 2655 } 2656 2657 while (worklist.size()) { // Visit all uses 2658 Node *use = worklist.pop(); 2659 if (!has_node(use)) continue; // Ignore dead nodes 2660 IdealLoopTree *use_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use ); 2661 if (!loop->is_member(use_loop) && use->is_CFG()) { 2662 // Both OLD and USE are CFG nodes here. 2663 assert(use->is_Proj(), "" ); 2664 Node* nnn = old_new[old->_idx]; 2665 2666 Node* newuse = nullptr; 2667 if (head->is_strip_mined() && mode != IgnoreStripMined) { 2668 CountedLoopNode* cl = head->as_CountedLoop(); 2669 CountedLoopEndNode* cle = cl->loopexit(); 2670 Node* cle_out = cle->proj_out_or_null(false); 2671 if (use == cle_out) { 2672 IfNode* le = cl->outer_loop_end(); 2673 use = le->proj_out(false); 2674 use_loop = get_loop(use); 2675 if (mode == CloneIncludesStripMined) { 2676 nnn = old_new[le->_idx]; 2677 } else { 2678 newuse = old_new[cle_out->_idx]; 2679 } 2680 } 2681 } 2682 if (newuse == nullptr) { 2683 newuse = use->clone(); 2684 } 2685 2686 // Clone the loop exit control projection 2687 if (C->do_vector_loop() && cm != nullptr) { 2688 cm->verify_insert_and_clone(use, newuse, cm->clone_idx()); 2689 } 2690 newuse->set_req(0,nnn); 2691 _igvn.register_new_node_with_optimizer(newuse); 2692 set_loop(newuse, use_loop); 2693 set_idom(newuse, nnn, dom_depth(nnn) + 1 ); 2694 2695 // We need a Region to merge the exit from the peeled body and the 2696 // exit from the old loop body. 2697 RegionNode *r = new RegionNode(3); 2698 uint dd_r = MIN2(dom_depth(newuse), dom_depth(use)); 2699 assert(dd_r >= dom_depth(dom_lca(newuse, use)), "" ); 2700 2701 // The original user of 'use' uses 'r' instead. 2702 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) { 2703 Node* useuse = use->last_out(l); 2704 _igvn.rehash_node_delayed(useuse); 2705 uint uses_found = 0; 2706 if (useuse->in(0) == use) { 2707 useuse->set_req(0, r); 2708 uses_found++; 2709 if (useuse->is_CFG()) { 2710 // This is not a dom_depth > dd_r because when new 2711 // control flow is constructed by a loop opt, a node and 2712 // its dominator can end up at the same dom_depth 2713 assert(dom_depth(useuse) >= dd_r, ""); 2714 set_idom(useuse, r, dom_depth(useuse)); 2715 } 2716 } 2717 for (uint k = 1; k < useuse->req(); k++) { 2718 if( useuse->in(k) == use ) { 2719 useuse->set_req(k, r); 2720 uses_found++; 2721 if (useuse->is_Loop() && k == LoopNode::EntryControl) { 2722 // This is not a dom_depth > dd_r because when new 2723 // control flow is constructed by a loop opt, a node 2724 // and its dominator can end up at the same dom_depth 2725 assert(dom_depth(useuse) >= dd_r , ""); 2726 set_idom(useuse, r, dom_depth(useuse)); 2727 } 2728 } 2729 } 2730 l -= uses_found; // we deleted 1 or more copies of this edge 2731 } 2732 2733 assert(use->is_Proj(), "loop exit should be projection"); 2734 // lazy_replace() below moves all nodes that are: 2735 // - control dependent on the loop exit or 2736 // - have control set to the loop exit 2737 // below the post-loop merge point. lazy_replace() takes a dead control as first input. To make it 2738 // possible to use it, the loop exit projection is cloned and becomes the new exit projection. The initial one 2739 // becomes dead and is "replaced" by the region. 2740 Node* use_clone = use->clone(); 2741 register_control(use_clone, use_loop, idom(use), dom_depth(use)); 2742 // Now finish up 'r' 2743 r->set_req(1, newuse); 2744 r->set_req(2, use_clone); 2745 _igvn.register_new_node_with_optimizer(r); 2746 set_loop(r, use_loop); 2747 set_idom(r, (side_by_side_idom == nullptr) ? newuse->in(0) : side_by_side_idom, dd_r); 2748 lazy_replace(use, r); 2749 // Map the (cloned) old use to the new merge point 2750 old_new.map(use_clone->_idx, r); 2751 } // End of if a loop-exit test 2752 } 2753 } 2754 } 2755 2756 void PhaseIdealLoop::fix_body_edges(const Node_List &body, IdealLoopTree* loop, const Node_List &old_new, int dd, 2757 IdealLoopTree* parent, bool partial) { 2758 for(uint i = 0; i < body.size(); i++ ) { 2759 Node *old = body.at(i); 2760 Node *nnn = old_new[old->_idx]; 2761 // Fix CFG/Loop controlling the new node 2762 if (has_ctrl(old)) { 2763 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]); 2764 } else { 2765 set_loop(nnn, parent); 2766 if (old->outcnt() > 0) { 2767 Node* dom = idom(old); 2768 if (old_new[dom->_idx] != nullptr) { 2769 dom = old_new[dom->_idx]; 2770 set_idom(nnn, dom, dd ); 2771 } 2772 } 2773 } 2774 // Correct edges to the new node 2775 for (uint j = 0; j < nnn->req(); j++) { 2776 Node *n = nnn->in(j); 2777 if (n != nullptr) { 2778 IdealLoopTree *old_in_loop = get_loop(has_ctrl(n) ? get_ctrl(n) : n); 2779 if (loop->is_member(old_in_loop)) { 2780 if (old_new[n->_idx] != nullptr) { 2781 nnn->set_req(j, old_new[n->_idx]); 2782 } else { 2783 assert(!body.contains(n), ""); 2784 assert(partial, "node not cloned"); 2785 } 2786 } 2787 } 2788 } 2789 _igvn.hash_find_insert(nnn); 2790 } 2791 } 2792 2793 void PhaseIdealLoop::clone_loop_body(const Node_List& body, Node_List &old_new, CloneMap* cm) { 2794 for (uint i = 0; i < body.size(); i++) { 2795 Node* old = body.at(i); 2796 Node* nnn = old->clone(); 2797 old_new.map(old->_idx, nnn); 2798 if (C->do_vector_loop() && cm != nullptr) { 2799 cm->verify_insert_and_clone(old, nnn, cm->clone_idx()); 2800 } 2801 _igvn.register_new_node_with_optimizer(nnn); 2802 } 2803 } 2804 2805 2806 //---------------------- stride_of_possible_iv ------------------------------------- 2807 // Looks for an iff/bool/comp with one operand of the compare 2808 // being a cycle involving an add and a phi, 2809 // with an optional truncation (left-shift followed by a right-shift) 2810 // of the add. Returns zero if not an iv. 2811 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) { 2812 Node* trunc1 = nullptr; 2813 Node* trunc2 = nullptr; 2814 const TypeInteger* ttype = nullptr; 2815 if (!iff->is_If() || iff->in(1) == nullptr || !iff->in(1)->is_Bool()) { 2816 return 0; 2817 } 2818 BoolNode* bl = iff->in(1)->as_Bool(); 2819 Node* cmp = bl->in(1); 2820 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) { 2821 return 0; 2822 } 2823 // Must have an invariant operand 2824 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) { 2825 return 0; 2826 } 2827 Node* add2 = nullptr; 2828 Node* cmp1 = cmp->in(1); 2829 if (cmp1->is_Phi()) { 2830 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) ))) 2831 Node* phi = cmp1; 2832 for (uint i = 1; i < phi->req(); i++) { 2833 Node* in = phi->in(i); 2834 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in, 2835 &trunc1, &trunc2, &ttype, T_INT); 2836 if (add && add->in(1) == phi) { 2837 add2 = add->in(2); 2838 break; 2839 } 2840 } 2841 } else { 2842 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) ))) 2843 Node* addtrunc = cmp1; 2844 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc, 2845 &trunc1, &trunc2, &ttype, T_INT); 2846 if (add && add->in(1)->is_Phi()) { 2847 Node* phi = add->in(1); 2848 for (uint i = 1; i < phi->req(); i++) { 2849 if (phi->in(i) == addtrunc) { 2850 add2 = add->in(2); 2851 break; 2852 } 2853 } 2854 } 2855 } 2856 if (add2 != nullptr) { 2857 const TypeInt* add2t = _igvn.type(add2)->is_int(); 2858 if (add2t->is_con()) { 2859 return add2t->get_con(); 2860 } 2861 } 2862 return 0; 2863 } 2864 2865 2866 //---------------------- stay_in_loop ------------------------------------- 2867 // Return the (unique) control output node that's in the loop (if it exists.) 2868 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) { 2869 Node* unique = nullptr; 2870 if (!n) return nullptr; 2871 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2872 Node* use = n->fast_out(i); 2873 if (!has_ctrl(use) && loop->is_member(get_loop(use))) { 2874 if (unique != nullptr) { 2875 return nullptr; 2876 } 2877 unique = use; 2878 } 2879 } 2880 return unique; 2881 } 2882 2883 //------------------------------ register_node ------------------------------------- 2884 // Utility to register node "n" with PhaseIdealLoop 2885 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree* loop, Node* pred, uint ddepth) { 2886 _igvn.register_new_node_with_optimizer(n); 2887 loop->_body.push(n); 2888 if (n->is_CFG()) { 2889 set_loop(n, loop); 2890 set_idom(n, pred, ddepth); 2891 } else { 2892 set_ctrl(n, pred); 2893 } 2894 } 2895 2896 //------------------------------ proj_clone ------------------------------------- 2897 // Utility to create an if-projection 2898 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) { 2899 ProjNode* c = p->clone()->as_Proj(); 2900 c->set_req(0, iff); 2901 return c; 2902 } 2903 2904 //------------------------------ short_circuit_if ------------------------------------- 2905 // Force the iff control output to be the live_proj 2906 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) { 2907 guarantee(live_proj != nullptr, "null projection"); 2908 int proj_con = live_proj->_con; 2909 assert(proj_con == 0 || proj_con == 1, "false or true projection"); 2910 Node* con = intcon(proj_con); 2911 if (iff) { 2912 iff->set_req(1, con); 2913 } 2914 return con; 2915 } 2916 2917 //------------------------------ insert_if_before_proj ------------------------------------- 2918 // Insert a new if before an if projection (* - new node) 2919 // 2920 // before 2921 // if(test) 2922 // / \ 2923 // v v 2924 // other-proj proj (arg) 2925 // 2926 // after 2927 // if(test) 2928 // / \ 2929 // / v 2930 // | * proj-clone 2931 // v | 2932 // other-proj v 2933 // * new_if(relop(cmp[IU](left,right))) 2934 // / \ 2935 // v v 2936 // * new-proj proj 2937 // (returned) 2938 // 2939 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) { 2940 IfNode* iff = proj->in(0)->as_If(); 2941 IdealLoopTree *loop = get_loop(proj); 2942 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2943 uint ddepth = dom_depth(proj); 2944 2945 _igvn.rehash_node_delayed(iff); 2946 _igvn.rehash_node_delayed(proj); 2947 2948 proj->set_req(0, nullptr); // temporary disconnect 2949 ProjNode* proj2 = proj_clone(proj, iff); 2950 register_node(proj2, loop, iff, ddepth); 2951 2952 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right); 2953 register_node(cmp, loop, proj2, ddepth); 2954 2955 BoolNode* bol = new BoolNode(cmp, relop); 2956 register_node(bol, loop, proj2, ddepth); 2957 2958 int opcode = iff->Opcode(); 2959 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode"); 2960 IfNode* new_if = IfNode::make_with_same_profile(iff, proj2, bol); 2961 register_node(new_if, loop, proj2, ddepth); 2962 2963 proj->set_req(0, new_if); // reattach 2964 set_idom(proj, new_if, ddepth); 2965 2966 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj(); 2967 guarantee(new_exit != nullptr, "null exit node"); 2968 register_node(new_exit, get_loop(other_proj), new_if, ddepth); 2969 2970 return new_exit; 2971 } 2972 2973 //------------------------------ insert_region_before_proj ------------------------------------- 2974 // Insert a region before an if projection (* - new node) 2975 // 2976 // before 2977 // if(test) 2978 // / | 2979 // v | 2980 // proj v 2981 // other-proj 2982 // 2983 // after 2984 // if(test) 2985 // / | 2986 // v | 2987 // * proj-clone v 2988 // | other-proj 2989 // v 2990 // * new-region 2991 // | 2992 // v 2993 // * dum_if 2994 // / \ 2995 // v \ 2996 // * dum-proj v 2997 // proj 2998 // 2999 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) { 3000 IfNode* iff = proj->in(0)->as_If(); 3001 IdealLoopTree *loop = get_loop(proj); 3002 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 3003 uint ddepth = dom_depth(proj); 3004 3005 _igvn.rehash_node_delayed(iff); 3006 _igvn.rehash_node_delayed(proj); 3007 3008 proj->set_req(0, nullptr); // temporary disconnect 3009 ProjNode* proj2 = proj_clone(proj, iff); 3010 register_node(proj2, loop, iff, ddepth); 3011 3012 RegionNode* reg = new RegionNode(2); 3013 reg->set_req(1, proj2); 3014 register_node(reg, loop, iff, ddepth); 3015 3016 IfNode* dum_if = new IfNode(reg, short_circuit_if(nullptr, proj), iff->_prob, iff->_fcnt); 3017 register_node(dum_if, loop, reg, ddepth); 3018 3019 proj->set_req(0, dum_if); // reattach 3020 set_idom(proj, dum_if, ddepth); 3021 3022 ProjNode* dum_proj = proj_clone(other_proj, dum_if); 3023 register_node(dum_proj, loop, dum_if, ddepth); 3024 3025 return reg; 3026 } 3027 3028 // Idea 3029 // ---- 3030 // Partial Peeling tries to rotate the loop in such a way that it can later be turned into a counted loop. Counted loops 3031 // require a signed loop exit test. When calling this method, we've only found a suitable unsigned test to partial peel 3032 // with. Therefore, we try to split off a signed loop exit test from the unsigned test such that it can be used as new 3033 // loop exit while keeping the unsigned test unchanged and preserving the same behavior as if we've used the unsigned 3034 // test alone instead: 3035 // 3036 // Before Partial Peeling: 3037 // Loop: 3038 // <peeled section> 3039 // Split off signed loop exit test 3040 // <-- CUT HERE --> 3041 // Unchanged unsigned loop exit test 3042 // <rest of unpeeled section> 3043 // goto Loop 3044 // 3045 // After Partial Peeling: 3046 // <cloned peeled section> 3047 // Cloned split off signed loop exit test 3048 // Loop: 3049 // Unchanged unsigned loop exit test 3050 // <rest of unpeeled section> 3051 // <peeled section> 3052 // Split off signed loop exit test 3053 // goto Loop 3054 // 3055 // Details 3056 // ------- 3057 // Before: 3058 // if (i <u limit) Unsigned loop exit condition 3059 // / | 3060 // v v 3061 // exit-proj stay-in-loop-proj 3062 // 3063 // Split off a signed loop exit test (i.e. with CmpI) from an unsigned loop exit test (i.e. with CmpU) and insert it 3064 // before the CmpU on the stay-in-loop path and keep both tests: 3065 // 3066 // if (i <u limit) Signed loop exit test 3067 // / | 3068 // / if (i <u limit) Unsigned loop exit test 3069 // / / | 3070 // v v v 3071 // exit-region stay-in-loop-proj 3072 // 3073 // Implementation 3074 // -------------- 3075 // We need to make sure that the new signed loop exit test is properly inserted into the graph such that the unsigned 3076 // loop exit test still dominates the same set of control nodes, the ctrl() relation from data nodes to both loop 3077 // exit tests is preserved, and their loop nesting is correct. 3078 // 3079 // To achieve that, we clone the unsigned loop exit test completely (leave it unchanged), insert the signed loop exit 3080 // test above it and kill the original unsigned loop exit test by setting it's condition to a constant 3081 // (i.e. stay-in-loop-const in graph below) such that IGVN can fold it later: 3082 // 3083 // if (stay-in-loop-const) Killed original unsigned loop exit test 3084 // / | 3085 // / v 3086 // / if (i < limit) Split off signed loop exit test 3087 // / / | 3088 // / / v 3089 // / / if (i <u limit) Cloned unsigned loop exit test 3090 // / / / | 3091 // v v v | 3092 // exit-region | 3093 // | | 3094 // dummy-if | 3095 // / | | 3096 // dead | | 3097 // v v 3098 // exit-proj stay-in-loop-proj 3099 // 3100 // Note: The dummy-if is inserted to create a region to merge the loop exits between the original to be killed unsigned 3101 // loop exit test and its exit projection while keeping the exit projection (also see insert_region_before_proj()). 3102 // 3103 // Requirements 3104 // ------------ 3105 // Note that we can only split off a signed loop exit test from the unsigned loop exit test when the behavior is exactly 3106 // the same as before with only a single unsigned test. This is only possible if certain requirements are met. 3107 // Otherwise, we need to bail out (see comments in the code below). 3108 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree* loop) { 3109 const bool Signed = true; 3110 const bool Unsigned = false; 3111 3112 BoolNode* bol = if_cmpu->in(1)->as_Bool(); 3113 if (bol->_test._test != BoolTest::lt) { 3114 return nullptr; 3115 } 3116 CmpNode* cmpu = bol->in(1)->as_Cmp(); 3117 assert(cmpu->Opcode() == Op_CmpU, "must be unsigned comparison"); 3118 3119 int stride = stride_of_possible_iv(if_cmpu); 3120 if (stride == 0) { 3121 return nullptr; 3122 } 3123 3124 Node* lp_proj = stay_in_loop(if_cmpu, loop); 3125 guarantee(lp_proj != nullptr, "null loop node"); 3126 3127 ProjNode* lp_continue = lp_proj->as_Proj(); 3128 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj(); 3129 if (!lp_exit->is_IfFalse()) { 3130 // The loop exit condition is (i <u limit) ==> (i >= 0 && i < limit). 3131 // We therefore can't add a single exit condition. 3132 return nullptr; 3133 } 3134 // The unsigned loop exit condition is 3135 // !(i <u limit) 3136 // = i >=u limit 3137 // 3138 // First, we note that for any x for which 3139 // 0 <= x <= INT_MAX 3140 // we can convert x to an unsigned int and still get the same guarantee: 3141 // 0 <= (uint) x <= INT_MAX = (uint) INT_MAX 3142 // 0 <=u (uint) x <=u INT_MAX = (uint) INT_MAX (LEMMA) 3143 // 3144 // With that in mind, if 3145 // limit >= 0 (COND) 3146 // then the unsigned loop exit condition 3147 // i >=u limit (ULE) 3148 // is equivalent to 3149 // i < 0 || i >= limit (SLE-full) 3150 // because either i is negative and therefore always greater than MAX_INT when converting to unsigned 3151 // (uint) i >=u MAX_INT >= limit >= 0 3152 // or otherwise 3153 // i >= limit >= 0 3154 // holds due to (LEMMA). 3155 // 3156 // For completeness, a counterexample with limit < 0: 3157 // Assume i = -3 and limit = -2: 3158 // i < 0 3159 // -2 < 0 3160 // is true and thus also "i < 0 || i >= limit". But 3161 // i >=u limit 3162 // -3 >=u -2 3163 // is false. 3164 Node* limit = cmpu->in(2); 3165 const TypeInt* type_limit = _igvn.type(limit)->is_int(); 3166 if (type_limit->_lo < 0) { 3167 return nullptr; 3168 } 3169 3170 // We prove below that we can extract a single signed loop exit condition from (SLE-full), depending on the stride: 3171 // stride < 0: 3172 // i < 0 (SLE = SLE-negative) 3173 // stride > 0: 3174 // i >= limit (SLE = SLE-positive) 3175 // such that we have the following graph before Partial Peeling with stride > 0 (similar for stride < 0): 3176 // 3177 // Loop: 3178 // <peeled section> 3179 // i >= limit (SLE-positive) 3180 // <-- CUT HERE --> 3181 // i >=u limit (ULE) 3182 // <rest of unpeeled section> 3183 // goto Loop 3184 // 3185 // We exit the loop if: 3186 // (SLE) is true OR (ULE) is true 3187 // However, if (SLE) is true then (ULE) also needs to be true to ensure the exact same behavior. Otherwise, we wrongly 3188 // exit a loop that should not have been exited if we did not apply Partial Peeling. More formally, we need to ensure: 3189 // (SLE) IMPLIES (ULE) 3190 // This indeed holds when (COND) is given: 3191 // - stride > 0: 3192 // i >= limit // (SLE = SLE-positive) 3193 // i >= limit >= 0 // (COND) 3194 // i >=u limit >= 0 // (LEMMA) 3195 // which is the unsigned loop exit condition (ULE). 3196 // - stride < 0: 3197 // i < 0 // (SLE = SLE-negative) 3198 // (uint) i >u MAX_INT // (NEG) all negative values are greater than MAX_INT when converted to unsigned 3199 // MAX_INT >= limit >= 0 // (COND) 3200 // MAX_INT >=u limit >= 0 // (LEMMA) 3201 // and thus from (NEG) and (LEMMA): 3202 // i >=u limit 3203 // which is the unsigned loop exit condition (ULE). 3204 // 3205 // 3206 // After Partial Peeling, we have the following structure for stride > 0 (similar for stride < 0): 3207 // <cloned peeled section> 3208 // i >= limit (SLE-positive) 3209 // Loop: 3210 // i >=u limit (ULE) 3211 // <rest of unpeeled section> 3212 // <peeled section> 3213 // i >= limit (SLE-positive) 3214 // goto Loop 3215 Node* rhs_cmpi; 3216 if (stride > 0) { 3217 rhs_cmpi = limit; // For i >= limit 3218 } else { 3219 rhs_cmpi = makecon(TypeInt::ZERO); // For i < 0 3220 } 3221 // Create a new region on the exit path 3222 RegionNode* reg = insert_region_before_proj(lp_exit); 3223 guarantee(reg != nullptr, "null region node"); 3224 3225 // Clone the if-cmpu-true-false using a signed compare 3226 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge; 3227 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, rhs_cmpi, lp_continue); 3228 reg->add_req(cmpi_exit); 3229 3230 // Clone the if-cmpu-true-false 3231 BoolTest::mask rel_u = bol->_test._test; 3232 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue); 3233 reg->add_req(cmpu_exit); 3234 3235 // Force original if to stay in loop. 3236 short_circuit_if(if_cmpu, lp_continue); 3237 3238 return cmpi_exit->in(0)->as_If(); 3239 } 3240 3241 //------------------------------ remove_cmpi_loop_exit ------------------------------------- 3242 // Remove a previously inserted signed compare loop exit. 3243 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) { 3244 Node* lp_proj = stay_in_loop(if_cmp, loop); 3245 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI && 3246 stay_in_loop(lp_proj, loop)->is_If() && 3247 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu"); 3248 Node* con = makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO); 3249 if_cmp->set_req(1, con); 3250 } 3251 3252 //------------------------------ scheduled_nodelist ------------------------------------- 3253 // Create a post order schedule of nodes that are in the 3254 // "member" set. The list is returned in "sched". 3255 // The first node in "sched" is the loop head, followed by 3256 // nodes which have no inputs in the "member" set, and then 3257 // followed by the nodes that have an immediate input dependence 3258 // on a node in "sched". 3259 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) { 3260 3261 assert(member.test(loop->_head->_idx), "loop head must be in member set"); 3262 VectorSet visited; 3263 Node_Stack nstack(loop->_body.size()); 3264 3265 Node* n = loop->_head; // top of stack is cached in "n" 3266 uint idx = 0; 3267 visited.set(n->_idx); 3268 3269 // Initially push all with no inputs from within member set 3270 for(uint i = 0; i < loop->_body.size(); i++ ) { 3271 Node *elt = loop->_body.at(i); 3272 if (member.test(elt->_idx)) { 3273 bool found = false; 3274 for (uint j = 0; j < elt->req(); j++) { 3275 Node* def = elt->in(j); 3276 if (def && member.test(def->_idx) && def != elt) { 3277 found = true; 3278 break; 3279 } 3280 } 3281 if (!found && elt != loop->_head) { 3282 nstack.push(n, idx); 3283 n = elt; 3284 assert(!visited.test(n->_idx), "not seen yet"); 3285 visited.set(n->_idx); 3286 } 3287 } 3288 } 3289 3290 // traverse out's that are in the member set 3291 while (true) { 3292 if (idx < n->outcnt()) { 3293 Node* use = n->raw_out(idx); 3294 idx++; 3295 if (!visited.test_set(use->_idx)) { 3296 if (member.test(use->_idx)) { 3297 nstack.push(n, idx); 3298 n = use; 3299 idx = 0; 3300 } 3301 } 3302 } else { 3303 // All outputs processed 3304 sched.push(n); 3305 if (nstack.is_empty()) break; 3306 n = nstack.node(); 3307 idx = nstack.index(); 3308 nstack.pop(); 3309 } 3310 } 3311 } 3312 3313 3314 //------------------------------ has_use_in_set ------------------------------------- 3315 // Has a use in the vector set 3316 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) { 3317 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3318 Node* use = n->fast_out(j); 3319 if (vset.test(use->_idx)) { 3320 return true; 3321 } 3322 } 3323 return false; 3324 } 3325 3326 3327 //------------------------------ has_use_internal_to_set ------------------------------------- 3328 // Has use internal to the vector set (ie. not in a phi at the loop head) 3329 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) { 3330 Node* head = loop->_head; 3331 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3332 Node* use = n->fast_out(j); 3333 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) { 3334 return true; 3335 } 3336 } 3337 return false; 3338 } 3339 3340 3341 //------------------------------ clone_for_use_outside_loop ------------------------------------- 3342 // clone "n" for uses that are outside of loop 3343 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) { 3344 int cloned = 0; 3345 assert(worklist.size() == 0, "should be empty"); 3346 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3347 Node* use = n->fast_out(j); 3348 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) { 3349 worklist.push(use); 3350 } 3351 } 3352 3353 if (C->check_node_count(worklist.size() + NodeLimitFudgeFactor, 3354 "Too many clones required in clone_for_use_outside_loop in partial peeling")) { 3355 return -1; 3356 } 3357 3358 while( worklist.size() ) { 3359 Node *use = worklist.pop(); 3360 if (!has_node(use) || use->in(0) == C->top()) continue; 3361 uint j; 3362 for (j = 0; j < use->req(); j++) { 3363 if (use->in(j) == n) break; 3364 } 3365 assert(j < use->req(), "must be there"); 3366 3367 // clone "n" and insert it between the inputs of "n" and the use outside the loop 3368 Node* n_clone = n->clone(); 3369 _igvn.replace_input_of(use, j, n_clone); 3370 cloned++; 3371 Node* use_c; 3372 if (!use->is_Phi()) { 3373 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0); 3374 } else { 3375 // Use in a phi is considered a use in the associated predecessor block 3376 use_c = use->in(0)->in(j); 3377 } 3378 set_ctrl(n_clone, use_c); 3379 assert(!loop->is_member(get_loop(use_c)), "should be outside loop"); 3380 get_loop(use_c)->_body.push(n_clone); 3381 _igvn.register_new_node_with_optimizer(n_clone); 3382 #ifndef PRODUCT 3383 if (TracePartialPeeling) { 3384 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx); 3385 } 3386 #endif 3387 } 3388 return cloned; 3389 } 3390 3391 3392 //------------------------------ clone_for_special_use_inside_loop ------------------------------------- 3393 // clone "n" for special uses that are in the not_peeled region. 3394 // If these def-uses occur in separate blocks, the code generator 3395 // marks the method as not compilable. For example, if a "BoolNode" 3396 // is in a different basic block than the "IfNode" that uses it, then 3397 // the compilation is aborted in the code generator. 3398 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 3399 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) { 3400 if (n->is_Phi() || n->is_Load()) { 3401 return; 3402 } 3403 assert(worklist.size() == 0, "should be empty"); 3404 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3405 Node* use = n->fast_out(j); 3406 if ( not_peel.test(use->_idx) && 3407 (use->is_If() || use->is_CMove() || use->is_Bool() || use->is_OpaqueInitializedAssertionPredicate()) && 3408 use->in(1) == n) { 3409 worklist.push(use); 3410 } 3411 } 3412 if (worklist.size() > 0) { 3413 // clone "n" and insert it between inputs of "n" and the use 3414 Node* n_clone = n->clone(); 3415 loop->_body.push(n_clone); 3416 _igvn.register_new_node_with_optimizer(n_clone); 3417 set_ctrl(n_clone, get_ctrl(n)); 3418 sink_list.push(n_clone); 3419 not_peel.set(n_clone->_idx); 3420 #ifndef PRODUCT 3421 if (TracePartialPeeling) { 3422 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx); 3423 } 3424 #endif 3425 while( worklist.size() ) { 3426 Node *use = worklist.pop(); 3427 _igvn.rehash_node_delayed(use); 3428 for (uint j = 1; j < use->req(); j++) { 3429 if (use->in(j) == n) { 3430 use->set_req(j, n_clone); 3431 } 3432 } 3433 } 3434 } 3435 } 3436 3437 3438 //------------------------------ insert_phi_for_loop ------------------------------------- 3439 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 3440 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) { 3441 Node *phi = PhiNode::make(lp, back_edge_val); 3442 phi->set_req(LoopNode::EntryControl, lp_entry_val); 3443 // Use existing phi if it already exists 3444 Node *hit = _igvn.hash_find_insert(phi); 3445 if( hit == nullptr ) { 3446 _igvn.register_new_node_with_optimizer(phi); 3447 set_ctrl(phi, lp); 3448 } else { 3449 // Remove the new phi from the graph and use the hit 3450 _igvn.remove_dead_node(phi); 3451 phi = hit; 3452 } 3453 _igvn.replace_input_of(use, idx, phi); 3454 } 3455 3456 #ifdef ASSERT 3457 //------------------------------ is_valid_loop_partition ------------------------------------- 3458 // Validate the loop partition sets: peel and not_peel 3459 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, 3460 VectorSet& not_peel ) { 3461 uint i; 3462 // Check that peel_list entries are in the peel set 3463 for (i = 0; i < peel_list.size(); i++) { 3464 if (!peel.test(peel_list.at(i)->_idx)) { 3465 return false; 3466 } 3467 } 3468 // Check at loop members are in one of peel set or not_peel set 3469 for (i = 0; i < loop->_body.size(); i++ ) { 3470 Node *def = loop->_body.at(i); 3471 uint di = def->_idx; 3472 // Check that peel set elements are in peel_list 3473 if (peel.test(di)) { 3474 if (not_peel.test(di)) { 3475 return false; 3476 } 3477 // Must be in peel_list also 3478 bool found = false; 3479 for (uint j = 0; j < peel_list.size(); j++) { 3480 if (peel_list.at(j)->_idx == di) { 3481 found = true; 3482 break; 3483 } 3484 } 3485 if (!found) { 3486 return false; 3487 } 3488 } else if (not_peel.test(di)) { 3489 if (peel.test(di)) { 3490 return false; 3491 } 3492 } else { 3493 return false; 3494 } 3495 } 3496 return true; 3497 } 3498 3499 //------------------------------ is_valid_clone_loop_exit_use ------------------------------------- 3500 // Ensure a use outside of loop is of the right form 3501 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) { 3502 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 3503 return (use->is_Phi() && 3504 use_c->is_Region() && use_c->req() == 3 && 3505 (use_c->in(exit_idx)->Opcode() == Op_IfTrue || 3506 use_c->in(exit_idx)->Opcode() == Op_IfFalse || 3507 use_c->in(exit_idx)->Opcode() == Op_JumpProj) && 3508 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) ); 3509 } 3510 3511 //------------------------------ is_valid_clone_loop_form ------------------------------------- 3512 // Ensure that all uses outside of loop are of the right form 3513 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 3514 uint orig_exit_idx, uint clone_exit_idx) { 3515 uint len = peel_list.size(); 3516 for (uint i = 0; i < len; i++) { 3517 Node *def = peel_list.at(i); 3518 3519 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 3520 Node *use = def->fast_out(j); 3521 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 3522 if (!loop->is_member(get_loop(use_c))) { 3523 // use is not in the loop, check for correct structure 3524 if (use->in(0) == def) { 3525 // Okay 3526 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) { 3527 return false; 3528 } 3529 } 3530 } 3531 } 3532 return true; 3533 } 3534 #endif 3535 3536 //------------------------------ partial_peel ------------------------------------- 3537 // Partially peel (aka loop rotation) the top portion of a loop (called 3538 // the peel section below) by cloning it and placing one copy just before 3539 // the new loop head and the other copy at the bottom of the new loop. 3540 // 3541 // before after where it came from 3542 // 3543 // stmt1 stmt1 3544 // loop: stmt2 clone 3545 // stmt2 if condA goto exitA clone 3546 // if condA goto exitA new_loop: new 3547 // stmt3 stmt3 clone 3548 // if !condB goto loop if condB goto exitB clone 3549 // exitB: stmt2 orig 3550 // stmt4 if !condA goto new_loop orig 3551 // exitA: goto exitA 3552 // exitB: 3553 // stmt4 3554 // exitA: 3555 // 3556 // Step 1: find the cut point: an exit test on probable 3557 // induction variable. 3558 // Step 2: schedule (with cloning) operations in the peel 3559 // section that can be executed after the cut into 3560 // the section that is not peeled. This may need 3561 // to clone operations into exit blocks. For 3562 // instance, a reference to A[i] in the not-peel 3563 // section and a reference to B[i] in an exit block 3564 // may cause a left-shift of i by 2 to be placed 3565 // in the peel block. This step will clone the left 3566 // shift into the exit block and sink the left shift 3567 // from the peel to the not-peel section. 3568 // Step 3: clone the loop, retarget the control, and insert 3569 // phis for values that are live across the new loop 3570 // head. This is very dependent on the graph structure 3571 // from clone_loop. It creates region nodes for 3572 // exit control and associated phi nodes for values 3573 // flow out of the loop through that exit. The region 3574 // node is dominated by the clone's control projection. 3575 // So the clone's peel section is placed before the 3576 // new loop head, and the clone's not-peel section is 3577 // forms the top part of the new loop. The original 3578 // peel section forms the tail of the new loop. 3579 // Step 4: update the dominator tree and recompute the 3580 // dominator depth. 3581 // 3582 // orig 3583 // 3584 // stmt1 3585 // | 3586 // v 3587 // predicates 3588 // | 3589 // v 3590 // loop<----+ 3591 // | | 3592 // stmt2 | 3593 // | | 3594 // v | 3595 // ifA | 3596 // / | | 3597 // v v | 3598 // false true ^ <-- last_peel 3599 // / | | 3600 // / ===|==cut | 3601 // / stmt3 | <-- first_not_peel 3602 // / | | 3603 // | v | 3604 // v ifB | 3605 // exitA: / \ | 3606 // / \ | 3607 // v v | 3608 // false true | 3609 // / \ | 3610 // / ----+ 3611 // | 3612 // v 3613 // exitB: 3614 // stmt4 3615 // 3616 // 3617 // after clone loop 3618 // 3619 // stmt1 3620 // | 3621 // v 3622 // predicates 3623 // / \ 3624 // clone / \ orig 3625 // / \ 3626 // / \ 3627 // v v 3628 // +---->loop loop<----+ 3629 // | | | | 3630 // | stmt2 stmt2 | 3631 // | | | | 3632 // | v v | 3633 // | ifA ifA | 3634 // | | \ / | | 3635 // | v v v v | 3636 // ^ true false false true ^ <-- last_peel 3637 // | | ^ \ / | | 3638 // | cut==|== \ \ / ===|==cut | 3639 // | stmt3 \ \ / stmt3 | <-- first_not_peel 3640 // | | dom | | | | 3641 // | v \ 1v v2 v | 3642 // | ifB regionA ifB | 3643 // | / \ | / \ | 3644 // | / \ v / \ | 3645 // | v v exitA: v v | 3646 // | true false false true | 3647 // | / ^ \ / \ | 3648 // +---- \ \ / ----+ 3649 // dom \ / 3650 // \ 1v v2 3651 // regionB 3652 // | 3653 // v 3654 // exitB: 3655 // stmt4 3656 // 3657 // 3658 // after partial peel 3659 // 3660 // stmt1 3661 // | 3662 // v 3663 // predicates 3664 // / 3665 // clone / orig 3666 // / TOP 3667 // / \ 3668 // v v 3669 // TOP->loop loop----+ 3670 // | | | 3671 // stmt2 stmt2 | 3672 // | | | 3673 // v v | 3674 // ifA ifA | 3675 // | \ / | | 3676 // v v v v | 3677 // true false false true | <-- last_peel 3678 // | ^ \ / +------|---+ 3679 // +->newloop \ \ / === ==cut | | 3680 // | stmt3 \ \ / TOP | | 3681 // | | dom | | stmt3 | | <-- first_not_peel 3682 // | v \ 1v v2 v | | 3683 // | ifB regionA ifB ^ v 3684 // | / \ | / \ | | 3685 // | / \ v / \ | | 3686 // | v v exitA: v v | | 3687 // | true false false true | | 3688 // | / ^ \ / \ | | 3689 // | | \ \ / v | | 3690 // | | dom \ / TOP | | 3691 // | | \ 1v v2 | | 3692 // ^ v regionB | | 3693 // | | | | | 3694 // | | v ^ v 3695 // | | exitB: | | 3696 // | | stmt4 | | 3697 // | +------------>-----------------+ | 3698 // | | 3699 // +-----------------<---------------------+ 3700 // 3701 // 3702 // final graph 3703 // 3704 // stmt1 3705 // | 3706 // v 3707 // predicates 3708 // | 3709 // v 3710 // stmt2 clone 3711 // | 3712 // v 3713 // ........> ifA clone 3714 // : / | 3715 // dom / | 3716 // : v v 3717 // : false true 3718 // : | | 3719 // : | v 3720 // : | newloop<-----+ 3721 // : | | | 3722 // : | stmt3 clone | 3723 // : | | | 3724 // : | v | 3725 // : | ifB | 3726 // : | / \ | 3727 // : | v v | 3728 // : | false true | 3729 // : | | | | 3730 // : | v stmt2 | 3731 // : | exitB: | | 3732 // : | stmt4 v | 3733 // : | ifA orig | 3734 // : | / \ | 3735 // : | / \ | 3736 // : | v v | 3737 // : | false true | 3738 // : | / \ | 3739 // : v v -----+ 3740 // RegionA 3741 // | 3742 // v 3743 // exitA 3744 // 3745 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) { 3746 3747 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only"); 3748 if (!loop->_head->is_Loop()) { 3749 return false; 3750 } 3751 LoopNode *head = loop->_head->as_Loop(); 3752 3753 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) { 3754 return false; 3755 } 3756 3757 // Check for complex exit control 3758 for (uint ii = 0; ii < loop->_body.size(); ii++) { 3759 Node *n = loop->_body.at(ii); 3760 int opc = n->Opcode(); 3761 if (n->is_Call() || 3762 opc == Op_Catch || 3763 opc == Op_CatchProj || 3764 opc == Op_Jump || 3765 opc == Op_JumpProj) { 3766 #ifndef PRODUCT 3767 if (TracePartialPeeling) { 3768 tty->print_cr("\nExit control too complex: lp: %d", head->_idx); 3769 } 3770 #endif 3771 return false; 3772 } 3773 } 3774 3775 int dd = dom_depth(head); 3776 3777 // Step 1: find cut point 3778 3779 // Walk up dominators to loop head looking for first loop exit 3780 // which is executed on every path thru loop. 3781 IfNode *peel_if = nullptr; 3782 IfNode *peel_if_cmpu = nullptr; 3783 3784 Node *iff = loop->tail(); 3785 while (iff != head) { 3786 if (iff->is_If()) { 3787 Node *ctrl = get_ctrl(iff->in(1)); 3788 if (ctrl->is_top()) return false; // Dead test on live IF. 3789 // If loop-varying exit-test, check for induction variable 3790 if (loop->is_member(get_loop(ctrl)) && 3791 loop->is_loop_exit(iff) && 3792 is_possible_iv_test(iff)) { 3793 Node* cmp = iff->in(1)->in(1); 3794 if (cmp->Opcode() == Op_CmpI) { 3795 peel_if = iff->as_If(); 3796 } else { 3797 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU"); 3798 peel_if_cmpu = iff->as_If(); 3799 } 3800 } 3801 } 3802 iff = idom(iff); 3803 } 3804 3805 // Prefer signed compare over unsigned compare. 3806 IfNode* new_peel_if = nullptr; 3807 if (peel_if == nullptr) { 3808 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == nullptr) { 3809 return false; // No peel point found 3810 } 3811 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop); 3812 if (new_peel_if == nullptr) { 3813 return false; // No peel point found 3814 } 3815 peel_if = new_peel_if; 3816 } 3817 Node* last_peel = stay_in_loop(peel_if, loop); 3818 Node* first_not_peeled = stay_in_loop(last_peel, loop); 3819 if (first_not_peeled == nullptr || first_not_peeled == head) { 3820 return false; 3821 } 3822 3823 #ifndef PRODUCT 3824 if (TraceLoopOpts) { 3825 tty->print("PartialPeel "); 3826 loop->dump_head(); 3827 } 3828 3829 if (TracePartialPeeling) { 3830 tty->print_cr("before partial peel one iteration"); 3831 Node_List wl; 3832 Node* t = head->in(2); 3833 while (true) { 3834 wl.push(t); 3835 if (t == head) break; 3836 t = idom(t); 3837 } 3838 while (wl.size() > 0) { 3839 Node* tt = wl.pop(); 3840 tt->dump(); 3841 if (tt == last_peel) tty->print_cr("-- cut --"); 3842 } 3843 } 3844 #endif 3845 3846 C->print_method(PHASE_BEFORE_PARTIAL_PEELING, 4, head); 3847 3848 VectorSet peel; 3849 VectorSet not_peel; 3850 Node_List peel_list; 3851 Node_List worklist; 3852 Node_List sink_list; 3853 3854 uint estimate = loop->est_loop_clone_sz(1); 3855 if (exceeding_node_budget(estimate)) { 3856 return false; 3857 } 3858 3859 // Set of cfg nodes to peel are those that are executable from 3860 // the head through last_peel. 3861 assert(worklist.size() == 0, "should be empty"); 3862 worklist.push(head); 3863 peel.set(head->_idx); 3864 while (worklist.size() > 0) { 3865 Node *n = worklist.pop(); 3866 if (n != last_peel) { 3867 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3868 Node* use = n->fast_out(j); 3869 if (use->is_CFG() && 3870 loop->is_member(get_loop(use)) && 3871 !peel.test_set(use->_idx)) { 3872 worklist.push(use); 3873 } 3874 } 3875 } 3876 } 3877 3878 // Set of non-cfg nodes to peel are those that are control 3879 // dependent on the cfg nodes. 3880 for (uint i = 0; i < loop->_body.size(); i++) { 3881 Node *n = loop->_body.at(i); 3882 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n; 3883 if (peel.test(n_c->_idx)) { 3884 peel.set(n->_idx); 3885 } else { 3886 not_peel.set(n->_idx); 3887 } 3888 } 3889 3890 // Step 2: move operations from the peeled section down into the 3891 // not-peeled section 3892 3893 // Get a post order schedule of nodes in the peel region 3894 // Result in right-most operand. 3895 scheduled_nodelist(loop, peel, peel_list); 3896 3897 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 3898 3899 // For future check for too many new phis 3900 uint old_phi_cnt = 0; 3901 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 3902 Node* use = head->fast_out(j); 3903 if (use->is_Phi()) old_phi_cnt++; 3904 } 3905 3906 #ifndef PRODUCT 3907 if (TracePartialPeeling) { 3908 tty->print_cr("\npeeled list"); 3909 } 3910 #endif 3911 3912 // Evacuate nodes in peel region into the not_peeled region if possible 3913 bool too_many_clones = false; 3914 uint new_phi_cnt = 0; 3915 uint cloned_for_outside_use = 0; 3916 for (uint i = 0; i < peel_list.size();) { 3917 Node* n = peel_list.at(i); 3918 #ifndef PRODUCT 3919 if (TracePartialPeeling) n->dump(); 3920 #endif 3921 bool incr = true; 3922 if (!n->is_CFG()) { 3923 if (has_use_in_set(n, not_peel)) { 3924 // If not used internal to the peeled region, 3925 // move "n" from peeled to not_peeled region. 3926 if (!has_use_internal_to_set(n, peel, loop)) { 3927 // if not pinned and not a load (which maybe anti-dependent on a store) 3928 // and not a CMove (Matcher expects only bool->cmove). 3929 if (n->in(0) == nullptr && !n->is_Load() && !n->is_CMove()) { 3930 int new_clones = clone_for_use_outside_loop(loop, n, worklist); 3931 if (C->failing()) return false; 3932 if (new_clones == -1) { 3933 too_many_clones = true; 3934 break; 3935 } 3936 cloned_for_outside_use += new_clones; 3937 sink_list.push(n); 3938 peel.remove(n->_idx); 3939 not_peel.set(n->_idx); 3940 peel_list.remove(i); 3941 incr = false; 3942 #ifndef PRODUCT 3943 if (TracePartialPeeling) { 3944 tty->print_cr("sink to not_peeled region: %d newbb: %d", 3945 n->_idx, get_ctrl(n)->_idx); 3946 } 3947 #endif 3948 } 3949 } else { 3950 // Otherwise check for special def-use cases that span 3951 // the peel/not_peel boundary such as bool->if 3952 clone_for_special_use_inside_loop(loop, n, not_peel, sink_list, worklist); 3953 new_phi_cnt++; 3954 } 3955 } 3956 } 3957 if (incr) i++; 3958 } 3959 3960 estimate += cloned_for_outside_use + new_phi_cnt; 3961 bool exceed_node_budget = !may_require_nodes(estimate); 3962 bool exceed_phi_limit = new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta; 3963 3964 if (too_many_clones || exceed_node_budget || exceed_phi_limit) { 3965 #ifndef PRODUCT 3966 if (TracePartialPeeling && exceed_phi_limit) { 3967 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c", 3968 new_phi_cnt, old_phi_cnt, new_peel_if != nullptr?'T':'F'); 3969 } 3970 #endif 3971 if (new_peel_if != nullptr) { 3972 remove_cmpi_loop_exit(new_peel_if, loop); 3973 } 3974 // Inhibit more partial peeling on this loop 3975 assert(!head->is_partial_peel_loop(), "not partial peeled"); 3976 head->mark_partial_peel_failed(); 3977 if (cloned_for_outside_use > 0) { 3978 // Terminate this round of loop opts because 3979 // the graph outside this loop was changed. 3980 C->set_major_progress(); 3981 return true; 3982 } 3983 return false; 3984 } 3985 3986 // Step 3: clone loop, retarget control, and insert new phis 3987 3988 // Create new loop head for new phis and to hang 3989 // the nodes being moved (sinked) from the peel region. 3990 LoopNode* new_head = new LoopNode(last_peel, last_peel); 3991 new_head->set_unswitch_count(head->unswitch_count()); // Preserve 3992 _igvn.register_new_node_with_optimizer(new_head); 3993 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled"); 3994 _igvn.replace_input_of(first_not_peeled, 0, new_head); 3995 set_loop(new_head, loop); 3996 loop->_body.push(new_head); 3997 not_peel.set(new_head->_idx); 3998 set_idom(new_head, last_peel, dom_depth(first_not_peeled)); 3999 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled)); 4000 4001 while (sink_list.size() > 0) { 4002 Node* n = sink_list.pop(); 4003 set_ctrl(n, new_head); 4004 } 4005 4006 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 4007 4008 clone_loop(loop, old_new, dd, IgnoreStripMined); 4009 4010 const uint clone_exit_idx = 1; 4011 const uint orig_exit_idx = 2; 4012 assert(is_valid_clone_loop_form(loop, peel_list, orig_exit_idx, clone_exit_idx), "bad clone loop"); 4013 4014 Node* head_clone = old_new[head->_idx]; 4015 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop(); 4016 Node* orig_tail_clone = head_clone->in(2); 4017 4018 // Add phi if "def" node is in peel set and "use" is not 4019 4020 for (uint i = 0; i < peel_list.size(); i++) { 4021 Node *def = peel_list.at(i); 4022 if (!def->is_CFG()) { 4023 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 4024 Node *use = def->fast_out(j); 4025 if (has_node(use) && use->in(0) != C->top() && 4026 (!peel.test(use->_idx) || 4027 (use->is_Phi() && use->in(0) == head)) ) { 4028 worklist.push(use); 4029 } 4030 } 4031 while( worklist.size() ) { 4032 Node *use = worklist.pop(); 4033 for (uint j = 1; j < use->req(); j++) { 4034 Node* n = use->in(j); 4035 if (n == def) { 4036 4037 // "def" is in peel set, "use" is not in peel set 4038 // or "use" is in the entry boundary (a phi) of the peel set 4039 4040 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use; 4041 4042 if ( loop->is_member(get_loop( use_c )) ) { 4043 // use is in loop 4044 if (old_new[use->_idx] != nullptr) { // null for dead code 4045 Node* use_clone = old_new[use->_idx]; 4046 _igvn.replace_input_of(use, j, C->top()); 4047 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone ); 4048 } 4049 } else { 4050 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format"); 4051 // use is not in the loop, check if the live range includes the cut 4052 Node* lp_if = use_c->in(orig_exit_idx)->in(0); 4053 if (not_peel.test(lp_if->_idx)) { 4054 assert(j == orig_exit_idx, "use from original loop"); 4055 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone ); 4056 } 4057 } 4058 } 4059 } 4060 } 4061 } 4062 } 4063 4064 // Step 3b: retarget control 4065 4066 // Redirect control to the new loop head if a cloned node in 4067 // the not_peeled region has control that points into the peeled region. 4068 // This necessary because the cloned peeled region will be outside 4069 // the loop. 4070 // from to 4071 // cloned-peeled <---+ 4072 // new_head_clone: | <--+ 4073 // cloned-not_peeled in(0) in(0) 4074 // orig-peeled 4075 4076 for (uint i = 0; i < loop->_body.size(); i++) { 4077 Node *n = loop->_body.at(i); 4078 if (!n->is_CFG() && n->in(0) != nullptr && 4079 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) { 4080 Node* n_clone = old_new[n->_idx]; 4081 if (n_clone->depends_only_on_test()) { 4082 // Pin array access nodes: control is updated here to the loop head. If, after some transformations, the 4083 // backedge is removed, an array load could become dependent on a condition that's not a range check for that 4084 // access. If that condition is replaced by an identical dominating one, then an unpinned load would risk 4085 // floating above its range check. 4086 Node* pinned_clone = n_clone->pin_array_access_node(); 4087 if (pinned_clone != nullptr) { 4088 register_new_node_with_ctrl_of(pinned_clone, n_clone); 4089 old_new.map(n->_idx, pinned_clone); 4090 _igvn.replace_node(n_clone, pinned_clone); 4091 n_clone = pinned_clone; 4092 } 4093 } 4094 _igvn.replace_input_of(n_clone, 0, new_head_clone); 4095 } 4096 } 4097 4098 // Backedge of the surviving new_head (the clone) is original last_peel 4099 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel); 4100 4101 // Cut first node in original not_peel set 4102 _igvn.rehash_node_delayed(new_head); // Multiple edge updates: 4103 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of 4104 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls 4105 4106 // Copy head_clone back-branch info to original head 4107 // and remove original head's loop entry and 4108 // clone head's back-branch 4109 _igvn.rehash_node_delayed(head); // Multiple edge updates 4110 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl)); 4111 head->set_req(LoopNode::LoopBackControl, C->top()); 4112 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top()); 4113 4114 // Similarly modify the phis 4115 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) { 4116 Node* use = head->fast_out(k); 4117 if (use->is_Phi() && use->outcnt() > 0) { 4118 Node* use_clone = old_new[use->_idx]; 4119 _igvn.rehash_node_delayed(use); // Multiple edge updates 4120 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl)); 4121 use->set_req(LoopNode::LoopBackControl, C->top()); 4122 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top()); 4123 } 4124 } 4125 4126 // Step 4: update dominator tree and dominator depth 4127 4128 set_idom(head, orig_tail_clone, dd); 4129 recompute_dom_depth(); 4130 4131 // Inhibit more partial peeling on this loop 4132 new_head_clone->set_partial_peel_loop(); 4133 C->set_major_progress(); 4134 loop->record_for_igvn(); 4135 4136 #ifndef PRODUCT 4137 if (TracePartialPeeling) { 4138 tty->print_cr("\nafter partial peel one iteration"); 4139 Node_List wl; 4140 Node* t = last_peel; 4141 while (true) { 4142 wl.push(t); 4143 if (t == head_clone) break; 4144 t = idom(t); 4145 } 4146 while (wl.size() > 0) { 4147 Node* tt = wl.pop(); 4148 if (tt == head) tty->print_cr("orig head"); 4149 else if (tt == new_head_clone) tty->print_cr("new head"); 4150 else if (tt == head_clone) tty->print_cr("clone head"); 4151 tt->dump(); 4152 } 4153 } 4154 #endif 4155 4156 C->print_method(PHASE_AFTER_PARTIAL_PEELING, 4, new_head_clone); 4157 4158 return true; 4159 } 4160 4161 // Transform: 4162 // 4163 // loop<-----------------+ 4164 // | | 4165 // stmt1 stmt2 .. stmtn | 4166 // | | | | 4167 // \ | / | 4168 // v v v | 4169 // region | 4170 // | | 4171 // shared_stmt | 4172 // | | 4173 // v | 4174 // if | 4175 // / \ | 4176 // | -----------+ 4177 // v 4178 // 4179 // into: 4180 // 4181 // loop<-------------------+ 4182 // | | 4183 // v | 4184 // +->loop | 4185 // | | | 4186 // | stmt1 stmt2 .. stmtn | 4187 // | | | | | 4188 // | | \ / | 4189 // | | v v | 4190 // | | region1 | 4191 // | | | | 4192 // | shared_stmt shared_stmt | 4193 // | | | | 4194 // | v v | 4195 // | if if | 4196 // | /\ / \ | 4197 // +-- | | -------+ 4198 // \ / 4199 // v v 4200 // region2 4201 // 4202 // (region2 is shown to merge mirrored projections of the loop exit 4203 // ifs to make the diagram clearer but they really merge the same 4204 // projection) 4205 // 4206 // Conditions for this transformation to trigger: 4207 // - the path through stmt1 is frequent enough 4208 // - the inner loop will be turned into a counted loop after transformation 4209 bool PhaseIdealLoop::duplicate_loop_backedge(IdealLoopTree *loop, Node_List &old_new) { 4210 if (!DuplicateBackedge) { 4211 return false; 4212 } 4213 assert(!loop->_head->is_CountedLoop() || StressDuplicateBackedge, "Non-counted loop only"); 4214 if (!loop->_head->is_Loop()) { 4215 return false; 4216 } 4217 4218 uint estimate = loop->est_loop_clone_sz(1); 4219 if (exceeding_node_budget(estimate)) { 4220 return false; 4221 } 4222 4223 LoopNode *head = loop->_head->as_Loop(); 4224 4225 Node* region = nullptr; 4226 IfNode* exit_test = nullptr; 4227 uint inner; 4228 float f; 4229 if (StressDuplicateBackedge) { 4230 if (head->is_strip_mined()) { 4231 return false; 4232 } 4233 Node* c = head->in(LoopNode::LoopBackControl); 4234 4235 while (c != head) { 4236 if (c->is_Region()) { 4237 region = c; 4238 } 4239 c = idom(c); 4240 } 4241 4242 if (region == nullptr) { 4243 return false; 4244 } 4245 4246 inner = 1; 4247 } else { 4248 // Is the shape of the loop that of a counted loop... 4249 Node* back_control = loop_exit_control(head, loop); 4250 if (back_control == nullptr) { 4251 return false; 4252 } 4253 4254 BoolTest::mask bt = BoolTest::illegal; 4255 float cl_prob = 0; 4256 Node* incr = nullptr; 4257 Node* limit = nullptr; 4258 Node* cmp = loop_exit_test(back_control, loop, incr, limit, bt, cl_prob); 4259 if (cmp == nullptr || cmp->Opcode() != Op_CmpI) { 4260 return false; 4261 } 4262 4263 // With an extra phi for the candidate iv? 4264 // Or the region node is the loop head 4265 if (!incr->is_Phi() || incr->in(0) == head) { 4266 return false; 4267 } 4268 4269 PathFrequency pf(head, this); 4270 region = incr->in(0); 4271 4272 // Go over all paths for the extra phi's region and see if that 4273 // path is frequent enough and would match the expected iv shape 4274 // if the extra phi is removed 4275 inner = 0; 4276 for (uint i = 1; i < incr->req(); ++i) { 4277 Node* in = incr->in(i); 4278 Node* trunc1 = nullptr; 4279 Node* trunc2 = nullptr; 4280 const TypeInteger* iv_trunc_t = nullptr; 4281 Node* orig_in = in; 4282 if (!(in = CountedLoopNode::match_incr_with_optional_truncation(in, &trunc1, &trunc2, &iv_trunc_t, T_INT))) { 4283 continue; 4284 } 4285 assert(in->Opcode() == Op_AddI, "wrong increment code"); 4286 Node* xphi = nullptr; 4287 Node* stride = loop_iv_stride(in, xphi); 4288 4289 if (stride == nullptr) { 4290 continue; 4291 } 4292 4293 PhiNode* phi = loop_iv_phi(xphi, nullptr, head); 4294 if (phi == nullptr || 4295 (trunc1 == nullptr && phi->in(LoopNode::LoopBackControl) != incr) || 4296 (trunc1 != nullptr && phi->in(LoopNode::LoopBackControl) != trunc1)) { 4297 return false; 4298 } 4299 4300 f = pf.to(region->in(i)); 4301 if (f > 0.5) { 4302 inner = i; 4303 break; 4304 } 4305 } 4306 4307 if (inner == 0) { 4308 return false; 4309 } 4310 4311 exit_test = back_control->in(0)->as_If(); 4312 } 4313 4314 if (idom(region)->is_Catch()) { 4315 return false; 4316 } 4317 4318 // Collect all control nodes that need to be cloned (shared_stmt in the diagram) 4319 Unique_Node_List wq; 4320 wq.push(head->in(LoopNode::LoopBackControl)); 4321 for (uint i = 0; i < wq.size(); i++) { 4322 Node* c = wq.at(i); 4323 assert(get_loop(c) == loop, "not in the right loop?"); 4324 if (c->is_Region()) { 4325 if (c != region) { 4326 for (uint j = 1; j < c->req(); ++j) { 4327 wq.push(c->in(j)); 4328 } 4329 } 4330 } else { 4331 wq.push(c->in(0)); 4332 } 4333 assert(!is_strict_dominator(c, region), "shouldn't go above region"); 4334 } 4335 4336 Node* region_dom = idom(region); 4337 4338 // Can't do the transformation if this would cause a membar pair to 4339 // be split 4340 for (uint i = 0; i < wq.size(); i++) { 4341 Node* c = wq.at(i); 4342 if (c->is_MemBar() && (c->as_MemBar()->trailing_store() || c->as_MemBar()->trailing_load_store())) { 4343 assert(c->as_MemBar()->leading_membar()->trailing_membar() == c, "bad membar pair"); 4344 if (!wq.member(c->as_MemBar()->leading_membar())) { 4345 return false; 4346 } 4347 } 4348 } 4349 C->print_method(PHASE_BEFORE_DUPLICATE_LOOP_BACKEDGE, 4, head); 4350 4351 // Collect data nodes that need to be clones as well 4352 int dd = dom_depth(head); 4353 4354 for (uint i = 0; i < loop->_body.size(); ++i) { 4355 Node* n = loop->_body.at(i); 4356 if (has_ctrl(n)) { 4357 Node* c = get_ctrl(n); 4358 if (wq.member(c)) { 4359 wq.push(n); 4360 } 4361 } else { 4362 set_idom(n, idom(n), dd); 4363 } 4364 } 4365 4366 // clone shared_stmt 4367 clone_loop_body(wq, old_new, nullptr); 4368 4369 Node* region_clone = old_new[region->_idx]; 4370 region_clone->set_req(inner, C->top()); 4371 set_idom(region, region->in(inner), dd); 4372 4373 // Prepare the outer loop 4374 Node* outer_head = new LoopNode(head->in(LoopNode::EntryControl), old_new[head->in(LoopNode::LoopBackControl)->_idx]); 4375 register_control(outer_head, loop->_parent, outer_head->in(LoopNode::EntryControl)); 4376 _igvn.replace_input_of(head, LoopNode::EntryControl, outer_head); 4377 set_idom(head, outer_head, dd); 4378 4379 fix_body_edges(wq, loop, old_new, dd, loop->_parent, true); 4380 4381 // Make one of the shared_stmt copies only reachable from stmt1, the 4382 // other only from stmt2..stmtn. 4383 Node* dom = nullptr; 4384 for (uint i = 1; i < region->req(); ++i) { 4385 if (i != inner) { 4386 _igvn.replace_input_of(region, i, C->top()); 4387 } 4388 Node* in = region_clone->in(i); 4389 if (in->is_top()) { 4390 continue; 4391 } 4392 if (dom == nullptr) { 4393 dom = in; 4394 } else { 4395 dom = dom_lca(dom, in); 4396 } 4397 } 4398 4399 set_idom(region_clone, dom, dd); 4400 4401 // Set up the outer loop 4402 for (uint i = 0; i < head->outcnt(); i++) { 4403 Node* u = head->raw_out(i); 4404 if (u->is_Phi()) { 4405 Node* outer_phi = u->clone(); 4406 outer_phi->set_req(0, outer_head); 4407 Node* backedge = old_new[u->in(LoopNode::LoopBackControl)->_idx]; 4408 if (backedge == nullptr) { 4409 backedge = u->in(LoopNode::LoopBackControl); 4410 } 4411 outer_phi->set_req(LoopNode::LoopBackControl, backedge); 4412 register_new_node(outer_phi, outer_head); 4413 _igvn.replace_input_of(u, LoopNode::EntryControl, outer_phi); 4414 } 4415 } 4416 4417 // create control and data nodes for out of loop uses (including region2) 4418 Node_List worklist; 4419 uint new_counter = C->unique(); 4420 fix_ctrl_uses(wq, loop, old_new, ControlAroundStripMined, outer_head, nullptr, worklist); 4421 4422 Node_List *split_if_set = nullptr; 4423 Node_List *split_bool_set = nullptr; 4424 Node_List *split_cex_set = nullptr; 4425 fix_data_uses(wq, loop, ControlAroundStripMined, loop->skip_strip_mined(), new_counter, old_new, worklist, 4426 split_if_set, split_bool_set, split_cex_set); 4427 4428 finish_clone_loop(split_if_set, split_bool_set, split_cex_set); 4429 4430 if (exit_test != nullptr) { 4431 float cnt = exit_test->_fcnt; 4432 if (cnt != COUNT_UNKNOWN) { 4433 exit_test->_fcnt = cnt * f; 4434 old_new[exit_test->_idx]->as_If()->_fcnt = cnt * (1 - f); 4435 } 4436 } 4437 4438 C->set_major_progress(); 4439 4440 C->print_method(PHASE_AFTER_DUPLICATE_LOOP_BACKEDGE, 4, outer_head); 4441 4442 return true; 4443 } 4444 4445 // AutoVectorize the loop: replace scalar ops with vector ops. 4446 PhaseIdealLoop::AutoVectorizeStatus 4447 PhaseIdealLoop::auto_vectorize(IdealLoopTree* lpt, VSharedData &vshared) { 4448 // Counted loop only 4449 if (!lpt->is_counted()) { 4450 return AutoVectorizeStatus::Impossible; 4451 } 4452 4453 // Main-loop only 4454 CountedLoopNode* cl = lpt->_head->as_CountedLoop(); 4455 if (!cl->is_main_loop()) { 4456 return AutoVectorizeStatus::Impossible; 4457 } 4458 4459 VLoop vloop(lpt, false); 4460 if (!vloop.check_preconditions()) { 4461 return AutoVectorizeStatus::TriedAndFailed; 4462 } 4463 4464 // Ensure the shared data is cleared before each use 4465 vshared.clear(); 4466 4467 const VLoopAnalyzer vloop_analyzer(vloop, vshared); 4468 if (!vloop_analyzer.success()) { 4469 return AutoVectorizeStatus::TriedAndFailed; 4470 } 4471 4472 SuperWord sw(vloop_analyzer); 4473 if (!sw.transform_loop()) { 4474 return AutoVectorizeStatus::TriedAndFailed; 4475 } 4476 4477 return AutoVectorizeStatus::Success; 4478 } 4479 4480 // Just before insert_pre_post_loops, we can multiversion the loop: 4481 // 4482 // multiversion_if 4483 // | | 4484 // fast_loop slow_loop 4485 // 4486 // In the fast_loop we can make speculative assumptions, and put the 4487 // conditions into the multiversion_if. If the conditions hold at runtime, 4488 // we enter the fast_loop, if the conditions fail, we take the slow_loop 4489 // instead which does not make any of the speculative assumptions. 4490 // 4491 // Note: we only multiversion the loop if the loop does not have any 4492 // auto vectorization check Predicate. If we have that predicate, 4493 // then we can simply add the speculative assumption checks to 4494 // that Predicate. This means we do not need to duplicate the 4495 // loop - we have a smaller graph and save compile time. Should 4496 // the conditions ever fail, then we deopt / trap at the Predicate 4497 // and recompile without that Predicate. At that point we will 4498 // multiversion the loop, so that we can still have speculative 4499 // runtime checks. 4500 // 4501 // We perform the multiversioning when the loop is still in its single 4502 // iteration form, even before we insert pre and post loops. This makes 4503 // the cloning much simpler. However, this means that both the fast 4504 // and the slow loop have to be optimized independently (adding pre 4505 // and post loops, unrolling the main loop, auto-vectorize etc.). And 4506 // we may end up not needing any speculative assumptions in the fast_loop 4507 // and then rejecting the slow_loop by constant folding the multiversion_if. 4508 // 4509 // Therefore, we "delay" the optimization of the slow_loop until we add 4510 // at least one speculative assumption for the fast_loop. If we never 4511 // add such a speculative runtime check, the OpaqueMultiversioningNode 4512 // of the multiversion_if constant folds to true after loop opts, and the 4513 // multiversion_if folds away the "delayed" slow_loop. If we add any 4514 // speculative assumption, then we notify the OpaqueMultiversioningNode 4515 // with "notify_slow_loop_that_it_can_resume_optimizations". 4516 // 4517 // Note: new runtime checks can be added to the multiversion_if with 4518 // PhaseIdealLoop::create_new_if_for_multiversion 4519 void PhaseIdealLoop::maybe_multiversion_for_auto_vectorization_runtime_checks(IdealLoopTree* lpt, Node_List& old_new) { 4520 CountedLoopNode* cl = lpt->_head->as_CountedLoop(); 4521 LoopNode* outer_loop = cl->skip_strip_mined(); 4522 Node* entry = outer_loop->in(LoopNode::EntryControl); 4523 4524 // Check we have multiversioning enabled, and are not already multiversioned. 4525 if (!LoopMultiversioning || cl->is_multiversion()) { return; } 4526 4527 // Check that we do not have a parse-predicate where we can add the runtime checks 4528 // during auto-vectorization. 4529 const Predicates predicates(entry); 4530 const PredicateBlock* predicate_block = predicates.auto_vectorization_check_block(); 4531 if (predicate_block->has_parse_predicate()) { return; } 4532 4533 // Check node budget. 4534 uint estimate = lpt->est_loop_clone_sz(2); 4535 if (!may_require_nodes(estimate)) { return; } 4536 4537 do_multiversioning(lpt, old_new); 4538 } 4539 4540 // Returns true if the Reduction node is unordered. 4541 static bool is_unordered_reduction(Node* n) { 4542 return n->is_Reduction() && !n->as_Reduction()->requires_strict_order(); 4543 } 4544 4545 // Having ReductionNodes in the loop is expensive. They need to recursively 4546 // fold together the vector values, for every vectorized loop iteration. If 4547 // we encounter the following pattern, we can vector accumulate the values 4548 // inside the loop, and only have a single UnorderedReduction after the loop. 4549 // 4550 // Note: UnorderedReduction represents a ReductionNode which does not require 4551 // calculating in strict order. 4552 // 4553 // CountedLoop init 4554 // | | 4555 // +------+ | +-----------------------+ 4556 // | | | | 4557 // PhiNode (s) | 4558 // | | 4559 // | Vector | 4560 // | | | 4561 // UnorderedReduction (first_ur) | 4562 // | | 4563 // ... Vector | 4564 // | | | 4565 // UnorderedReduction (last_ur) | 4566 // | | 4567 // +---------------------+ 4568 // 4569 // We patch the graph to look like this: 4570 // 4571 // CountedLoop identity_vector 4572 // | | 4573 // +-------+ | +---------------+ 4574 // | | | | 4575 // PhiNode (v) | 4576 // | | 4577 // | Vector | 4578 // | | | 4579 // VectorAccumulator | 4580 // | | 4581 // ... Vector | 4582 // | | | 4583 // init VectorAccumulator | 4584 // | | | | 4585 // UnorderedReduction +-----------+ 4586 // 4587 // We turned the scalar (s) Phi into a vectorized one (v). In the loop, we 4588 // use vector_accumulators, which do the same reductions, but only element 4589 // wise. This is a single operation per vector_accumulator, rather than many 4590 // for a UnorderedReduction. We can then reduce the last vector_accumulator 4591 // after the loop, and also reduce the init value into it. 4592 // 4593 // We can not do this with all reductions. Some reductions do not allow the 4594 // reordering of operations (for example float addition/multiplication require 4595 // strict order). 4596 void PhaseIdealLoop::move_unordered_reduction_out_of_loop(IdealLoopTree* loop) { 4597 assert(!C->major_progress() && loop->is_counted() && loop->is_innermost(), "sanity"); 4598 4599 // Find all Phi nodes with an unordered Reduction on backedge. 4600 CountedLoopNode* cl = loop->_head->as_CountedLoop(); 4601 for (DUIterator_Fast jmax, j = cl->fast_outs(jmax); j < jmax; j++) { 4602 Node* phi = cl->fast_out(j); 4603 // We have a phi with a single use, and an unordered Reduction on the backedge. 4604 if (!phi->is_Phi() || phi->outcnt() != 1 || !is_unordered_reduction(phi->in(2))) { 4605 continue; 4606 } 4607 4608 ReductionNode* last_ur = phi->in(2)->as_Reduction(); 4609 assert(!last_ur->requires_strict_order(), "must be"); 4610 4611 // Determine types 4612 const TypeVect* vec_t = last_ur->vect_type(); 4613 uint vector_length = vec_t->length(); 4614 BasicType bt = vec_t->element_basic_type(); 4615 4616 // Convert opcode from vector-reduction -> scalar -> normal-vector-op 4617 const int sopc = VectorNode::scalar_opcode(last_ur->Opcode(), bt); 4618 const int vopc = VectorNode::opcode(sopc, bt); 4619 if (!Matcher::match_rule_supported_vector(vopc, vector_length, bt)) { 4620 DEBUG_ONLY( last_ur->dump(); ) 4621 assert(false, "do not have normal vector op for this reduction"); 4622 continue; // not implemented -> fails 4623 } 4624 4625 // Traverse up the chain of unordered Reductions, checking that it loops back to 4626 // the phi. Check that all unordered Reductions only have a single use, except for 4627 // the last (last_ur), which only has phi as a use in the loop, and all other uses 4628 // are outside the loop. 4629 ReductionNode* current = last_ur; 4630 ReductionNode* first_ur = nullptr; 4631 while (true) { 4632 assert(!current->requires_strict_order(), "sanity"); 4633 4634 // Expect no ctrl and a vector_input from within the loop. 4635 Node* ctrl = current->in(0); 4636 Node* vector_input = current->in(2); 4637 if (ctrl != nullptr || get_ctrl(vector_input) != cl) { 4638 DEBUG_ONLY( current->dump(1); ) 4639 assert(false, "reduction has ctrl or bad vector_input"); 4640 break; // Chain traversal fails. 4641 } 4642 4643 assert(current->vect_type() != nullptr, "must have vector type"); 4644 if (current->vect_type() != last_ur->vect_type()) { 4645 // Reductions do not have the same vector type (length and element type). 4646 break; // Chain traversal fails. 4647 } 4648 4649 // Expect single use of an unordered Reduction, except for last_ur. 4650 if (current == last_ur) { 4651 // Expect all uses to be outside the loop, except phi. 4652 for (DUIterator_Fast kmax, k = current->fast_outs(kmax); k < kmax; k++) { 4653 Node* use = current->fast_out(k); 4654 if (use != phi && ctrl_or_self(use) == cl) { 4655 DEBUG_ONLY( current->dump(-1); ) 4656 assert(false, "reduction has use inside loop"); 4657 // Should not be allowed by SuperWord::mark_reductions 4658 return; // bail out of optimization 4659 } 4660 } 4661 } else { 4662 if (current->outcnt() != 1) { 4663 break; // Chain traversal fails. 4664 } 4665 } 4666 4667 // Expect another unordered Reduction or phi as the scalar input. 4668 Node* scalar_input = current->in(1); 4669 if (is_unordered_reduction(scalar_input) && 4670 scalar_input->Opcode() == current->Opcode()) { 4671 // Move up the unordered Reduction chain. 4672 current = scalar_input->as_Reduction(); 4673 assert(!current->requires_strict_order(), "must be"); 4674 } else if (scalar_input == phi) { 4675 // Chain terminates at phi. 4676 first_ur = current; 4677 current = nullptr; 4678 break; // Success. 4679 } else { 4680 // scalar_input is neither phi nor a matching reduction 4681 // Can for example be scalar reduction when we have 4682 // partial vectorization. 4683 break; // Chain traversal fails. 4684 } 4685 } 4686 if (current != nullptr) { 4687 // Chain traversal was not successful. 4688 continue; 4689 } 4690 assert(first_ur != nullptr, "must have successfully terminated chain traversal"); 4691 4692 Node* identity_scalar = ReductionNode::make_identity_con_scalar(_igvn, sopc, bt); 4693 set_root_as_ctrl(identity_scalar); 4694 VectorNode* identity_vector = VectorNode::scalar2vector(identity_scalar, vector_length, bt); 4695 register_new_node(identity_vector, C->root()); 4696 assert(vec_t == identity_vector->vect_type(), "matching vector type"); 4697 VectorNode::trace_new_vector(identity_vector, "Unordered Reduction"); 4698 4699 // Turn the scalar phi into a vector phi. 4700 _igvn.rehash_node_delayed(phi); 4701 Node* init = phi->in(1); // Remember init before replacing it. 4702 phi->set_req_X(1, identity_vector, &_igvn); 4703 phi->as_Type()->set_type(vec_t); 4704 _igvn.set_type(phi, vec_t); 4705 4706 // Traverse down the chain of unordered Reductions, and replace them with vector_accumulators. 4707 current = first_ur; 4708 while (true) { 4709 // Create vector_accumulator to replace current. 4710 Node* last_vector_accumulator = current->in(1); 4711 Node* vector_input = current->in(2); 4712 VectorNode* vector_accumulator = VectorNode::make(vopc, last_vector_accumulator, vector_input, vec_t); 4713 register_new_node(vector_accumulator, cl); 4714 _igvn.replace_node(current, vector_accumulator); 4715 VectorNode::trace_new_vector(vector_accumulator, "Unordered Reduction"); 4716 if (current == last_ur) { 4717 break; 4718 } 4719 current = vector_accumulator->unique_out()->as_Reduction(); 4720 assert(!current->requires_strict_order(), "must be"); 4721 } 4722 4723 // Create post-loop reduction. 4724 Node* last_accumulator = phi->in(2); 4725 Node* post_loop_reduction = ReductionNode::make(sopc, nullptr, init, last_accumulator, bt); 4726 4727 // Take over uses of last_accumulator that are not in the loop. 4728 for (DUIterator i = last_accumulator->outs(); last_accumulator->has_out(i); i++) { 4729 Node* use = last_accumulator->out(i); 4730 if (use != phi && use != post_loop_reduction) { 4731 assert(ctrl_or_self(use) != cl, "use must be outside loop"); 4732 use->replace_edge(last_accumulator, post_loop_reduction, &_igvn); 4733 --i; 4734 } 4735 } 4736 register_new_node(post_loop_reduction, get_late_ctrl(post_loop_reduction, cl)); 4737 VectorNode::trace_new_vector(post_loop_reduction, "Unordered Reduction"); 4738 4739 assert(last_accumulator->outcnt() == 2, "last_accumulator has 2 uses: phi and post_loop_reduction"); 4740 assert(post_loop_reduction->outcnt() > 0, "should have taken over all non loop uses of last_accumulator"); 4741 assert(phi->outcnt() == 1, "accumulator is the only use of phi"); 4742 } 4743 } 4744 4745 void DataNodeGraph::clone_data_nodes(Node* new_ctrl) { 4746 for (uint i = 0; i < _data_nodes.size(); i++) { 4747 clone(_data_nodes[i], new_ctrl); 4748 } 4749 } 4750 4751 // Clone the given node and set it up properly. Set 'new_ctrl' as ctrl. 4752 void DataNodeGraph::clone(Node* node, Node* new_ctrl) { 4753 Node* clone = node->clone(); 4754 _phase->igvn().register_new_node_with_optimizer(clone); 4755 _orig_to_new.put(node, clone); 4756 _phase->set_ctrl(clone, new_ctrl); 4757 if (node->is_CastII()) { 4758 clone->set_req(0, new_ctrl); 4759 } 4760 } 4761 4762 // Rewire the data inputs of all (unprocessed) cloned nodes, whose inputs are still pointing to the same inputs as their 4763 // corresponding orig nodes, to the newly cloned inputs to create a separate cloned graph. 4764 void DataNodeGraph::rewire_clones_to_cloned_inputs() { 4765 _orig_to_new.iterate_all([&](Node* node, Node* clone) { 4766 for (uint i = 1; i < node->req(); i++) { 4767 Node** cloned_input = _orig_to_new.get(node->in(i)); 4768 if (cloned_input != nullptr) { 4769 // Input was also cloned -> rewire clone to the cloned input. 4770 _phase->igvn().replace_input_of(clone, i, *cloned_input); 4771 } 4772 } 4773 }); 4774 } 4775 4776 // Clone all non-OpaqueLoop* nodes and apply the provided transformation strategy for OpaqueLoop* nodes. 4777 // Set 'new_ctrl' as ctrl for all cloned non-OpaqueLoop* nodes. 4778 void DataNodeGraph::clone_data_nodes_and_transform_opaque_loop_nodes( 4779 const TransformStrategyForOpaqueLoopNodes& transform_strategy, 4780 Node* new_ctrl) { 4781 for (uint i = 0; i < _data_nodes.size(); i++) { 4782 Node* data_node = _data_nodes[i]; 4783 if (data_node->is_Opaque1()) { 4784 transform_opaque_node(transform_strategy, data_node); 4785 } else { 4786 clone(data_node, new_ctrl); 4787 } 4788 } 4789 } 4790 4791 void DataNodeGraph::transform_opaque_node(const TransformStrategyForOpaqueLoopNodes& transform_strategy, Node* node) { 4792 Node* transformed_node; 4793 if (node->is_OpaqueLoopInit()) { 4794 transformed_node = transform_strategy.transform_opaque_init(node->as_OpaqueLoopInit()); 4795 } else { 4796 assert(node->is_OpaqueLoopStride(), "must be OpaqueLoopStrideNode"); 4797 transformed_node = transform_strategy.transform_opaque_stride(node->as_OpaqueLoopStride()); 4798 } 4799 // Add an orig->new mapping to correctly update the inputs of the copied graph in rewire_clones_to_cloned_inputs(). 4800 _orig_to_new.put(node, transformed_node); 4801 }