1 /* 2 * Copyright (c) 1999, 2023, 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 "precompiled.hpp" 26 #include "gc/shared/barrierSet.hpp" 27 #include "gc/shared/c2/barrierSetC2.hpp" 28 #include "memory/allocation.inline.hpp" 29 #include "memory/resourceArea.hpp" 30 #include "opto/addnode.hpp" 31 #include "opto/callnode.hpp" 32 #include "opto/castnode.hpp" 33 #include "opto/connode.hpp" 34 #include "opto/castnode.hpp" 35 #include "opto/divnode.hpp" 36 #include "opto/inlinetypenode.hpp" 37 #include "opto/loopnode.hpp" 38 #include "opto/matcher.hpp" 39 #include "opto/mulnode.hpp" 40 #include "opto/movenode.hpp" 41 #include "opto/opaquenode.hpp" 42 #include "opto/rootnode.hpp" 43 #include "opto/subnode.hpp" 44 #include "opto/subtypenode.hpp" 45 #include "opto/vectornode.hpp" 46 #include "utilities/macros.hpp" 47 48 //============================================================================= 49 //------------------------------split_thru_phi--------------------------------- 50 // Split Node 'n' through merge point if there is enough win. 51 Node* PhaseIdealLoop::split_thru_phi(Node* n, Node* region, int policy) { 52 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) { 53 // ConvI2L 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 // Inline types should not be split through Phis because they cannot be merged 67 // through Phi nodes but each value input needs to be merged individually. 68 if (n->is_InlineType()) { 69 return nullptr; 70 } 71 72 if (cannot_split_division(n, region)) { 73 return nullptr; 74 } 75 76 int wins = 0; 77 assert(!n->is_CFG(), ""); 78 assert(region->is_Region(), ""); 79 80 const Type* type = n->bottom_type(); 81 const TypeOopPtr* t_oop = _igvn.type(n)->isa_oopptr(); 82 Node* phi; 83 if (t_oop != nullptr && t_oop->is_known_instance_field()) { 84 int iid = t_oop->instance_id(); 85 int index = C->get_alias_index(t_oop); 86 int offset = t_oop->offset(); 87 phi = new PhiNode(region, type, nullptr, iid, index, offset); 88 } else { 89 phi = PhiNode::make_blank(region, n); 90 } 91 uint old_unique = C->unique(); 92 for (uint i = 1; i < region->req(); i++) { 93 Node* x; 94 Node* the_clone = nullptr; 95 if (region->in(i) == C->top()) { 96 x = C->top(); // Dead path? Use a dead data op 97 } else { 98 x = n->clone(); // Else clone up the data op 99 the_clone = x; // Remember for possible deletion. 100 // Alter data node to use pre-phi inputs 101 if (n->in(0) == region) 102 x->set_req( 0, region->in(i) ); 103 for (uint j = 1; j < n->req(); j++) { 104 Node* in = n->in(j); 105 if (in->is_Phi() && in->in(0) == region) 106 x->set_req(j, in->in(i)); // Use pre-Phi input for the clone 107 } 108 } 109 // Check for a 'win' on some paths 110 const Type* t = x->Value(&_igvn); 111 112 bool singleton = t->singleton(); 113 114 // A TOP singleton indicates that there are no possible values incoming 115 // along a particular edge. In most cases, this is OK, and the Phi will 116 // be eliminated later in an Ideal call. However, we can't allow this to 117 // happen if the singleton occurs on loop entry, as the elimination of 118 // the PhiNode may cause the resulting node to migrate back to a previous 119 // loop iteration. 120 if (singleton && t == Type::TOP) { 121 // Is_Loop() == false does not confirm the absence of a loop (e.g., an 122 // irreducible loop may not be indicated by an affirmative is_Loop()); 123 // therefore, the only top we can split thru a phi is on a backedge of 124 // a loop. 125 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 126 } 127 128 if (singleton) { 129 wins++; 130 x = ((PhaseGVN&)_igvn).makecon(t); 131 } else { 132 // We now call Identity to try to simplify the cloned node. 133 // Note that some Identity methods call phase->type(this). 134 // Make sure that the type array is big enough for 135 // our new node, even though we may throw the node away. 136 // (Note: This tweaking with igvn only works because x is a new node.) 137 _igvn.set_type(x, t); 138 // If x is a TypeNode, capture any more-precise type permanently into Node 139 // otherwise it will be not updated during igvn->transform since 140 // igvn->type(x) is set to x->Value() already. 141 x->raise_bottom_type(t); 142 Node* y = x->Identity(&_igvn); 143 if (y != x) { 144 wins++; 145 x = y; 146 } else { 147 y = _igvn.hash_find(x); 148 if (y == nullptr) { 149 y = similar_subtype_check(x, region->in(i)); 150 } 151 if (y) { 152 wins++; 153 x = y; 154 } else { 155 // Else x is a new node we are keeping 156 // We do not need register_new_node_with_optimizer 157 // because set_type has already been called. 158 _igvn._worklist.push(x); 159 } 160 } 161 } 162 if (x != the_clone && the_clone != nullptr) 163 _igvn.remove_dead_node(the_clone); 164 phi->set_req( i, x ); 165 } 166 // Too few wins? 167 if (wins <= policy) { 168 _igvn.remove_dead_node(phi); 169 return nullptr; 170 } 171 172 // Record Phi 173 register_new_node( phi, region ); 174 175 for (uint i2 = 1; i2 < phi->req(); i2++) { 176 Node *x = phi->in(i2); 177 // If we commoned up the cloned 'x' with another existing Node, 178 // the existing Node picks up a new use. We need to make the 179 // existing Node occur higher up so it dominates its uses. 180 Node *old_ctrl; 181 IdealLoopTree *old_loop; 182 183 if (x->is_Con()) { 184 // Constant's control is always root. 185 set_ctrl(x, C->root()); 186 continue; 187 } 188 // The occasional new node 189 if (x->_idx >= old_unique) { // Found a new, unplaced node? 190 old_ctrl = nullptr; 191 old_loop = nullptr; // Not in any prior loop 192 } else { 193 old_ctrl = get_ctrl(x); 194 old_loop = get_loop(old_ctrl); // Get prior loop 195 } 196 // New late point must dominate new use 197 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2)); 198 if (new_ctrl == old_ctrl) // Nothing is changed 199 continue; 200 201 IdealLoopTree *new_loop = get_loop(new_ctrl); 202 203 // Don't move x into a loop if its uses are 204 // outside of loop. Otherwise x will be cloned 205 // for each use outside of this loop. 206 IdealLoopTree *use_loop = get_loop(region); 207 if (!new_loop->is_member(use_loop) && 208 (old_loop == nullptr || !new_loop->is_member(old_loop))) { 209 // Take early control, later control will be recalculated 210 // during next iteration of loop optimizations. 211 new_ctrl = get_early_ctrl(x); 212 new_loop = get_loop(new_ctrl); 213 } 214 // Set new location 215 set_ctrl(x, new_ctrl); 216 // If changing loop bodies, see if we need to collect into new body 217 if (old_loop != new_loop) { 218 if (old_loop && !old_loop->_child) 219 old_loop->_body.yank(x); 220 if (!new_loop->_child) 221 new_loop->_body.push(x); // Collect body info 222 } 223 } 224 225 return phi; 226 } 227 228 // Subtype checks that carry profile data don't common so look for a replacement by following edges 229 Node* PhaseIdealLoop::similar_subtype_check(const Node* x, Node* r_in) { 230 if (x->is_SubTypeCheck()) { 231 Node* in1 = x->in(1); 232 for (DUIterator_Fast imax, i = in1->fast_outs(imax); i < imax; i++) { 233 Node* u = in1->fast_out(i); 234 if (u != x && u->is_SubTypeCheck() && u->in(1) == x->in(1) && u->in(2) == x->in(2)) { 235 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) { 236 Node* bol = u->fast_out(j); 237 for (DUIterator_Fast kmax, k = bol->fast_outs(kmax); k < kmax; k++) { 238 Node* iff = bol->fast_out(k); 239 // Only dominating subtype checks are interesting: otherwise we risk replacing a subtype check by another with 240 // unrelated profile 241 if (iff->is_If() && is_dominator(iff, r_in)) { 242 return u; 243 } 244 } 245 } 246 } 247 } 248 } 249 return nullptr; 250 } 251 252 // Return true if 'n' is a Div or Mod node (without zero check If node which was removed earlier) with a loop phi divisor 253 // of a trip-counted (integer or long) loop with a backedge input that could be zero (include zero in its type range). In 254 // this case, we cannot split the division to the backedge as it could freely float above the loop exit check resulting in 255 // a division by zero. This situation is possible because the type of an increment node of an iv phi (trip-counter) could 256 // include zero while the iv phi does not (see PhiNode::Value() for trip-counted loops where we improve types of iv phis). 257 // We also need to check other loop phis as they could have been created in the same split-if pass when applying 258 // PhaseIdealLoop::split_thru_phi() to split nodes through an iv phi. 259 bool PhaseIdealLoop::cannot_split_division(const Node* n, const Node* region) const { 260 const Type* zero; 261 switch (n->Opcode()) { 262 case Op_DivI: 263 case Op_ModI: 264 zero = TypeInt::ZERO; 265 break; 266 case Op_DivL: 267 case Op_ModL: 268 zero = TypeLong::ZERO; 269 break; 270 default: 271 return false; 272 } 273 274 assert(n->in(0) == nullptr, "divisions with zero check should already have bailed out earlier in split-if"); 275 Node* divisor = n->in(2); 276 return is_divisor_counted_loop_phi(divisor, region) && 277 loop_phi_backedge_type_contains_zero(divisor, zero); 278 } 279 280 bool PhaseIdealLoop::is_divisor_counted_loop_phi(const Node* divisor, const Node* loop) { 281 return loop->is_BaseCountedLoop() && divisor->is_Phi() && divisor->in(0) == loop; 282 } 283 284 bool PhaseIdealLoop::loop_phi_backedge_type_contains_zero(const Node* phi_divisor, const Type* zero) const { 285 return _igvn.type(phi_divisor->in(LoopNode::LoopBackControl))->filter_speculative(zero) != Type::TOP; 286 } 287 288 //------------------------------dominated_by------------------------------------ 289 // Replace the dominated test with an obvious true or false. Place it on the 290 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the 291 // live path up to the dominating control. 292 void PhaseIdealLoop::dominated_by(IfProjNode* prevdom, IfNode* iff, bool flip, bool exclude_loop_predicate) { 293 if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); } 294 295 // prevdom is the dominating projection of the dominating test. 296 assert(iff->Opcode() == Op_If || 297 iff->Opcode() == Op_CountedLoopEnd || 298 iff->Opcode() == Op_LongCountedLoopEnd || 299 iff->Opcode() == Op_RangeCheck || 300 iff->Opcode() == Op_ParsePredicate, 301 "Check this code when new subtype is added"); 302 303 int pop = prevdom->Opcode(); 304 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" ); 305 if (flip) { 306 if (pop == Op_IfTrue) 307 pop = Op_IfFalse; 308 else 309 pop = Op_IfTrue; 310 } 311 // 'con' is set to true or false to kill the dominated test. 312 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO); 313 set_ctrl(con, C->root()); // Constant gets a new use 314 // Hack the dominated test 315 _igvn.replace_input_of(iff, 1, con); 316 317 // If I dont have a reachable TRUE and FALSE path following the IfNode then 318 // I can assume this path reaches an infinite loop. In this case it's not 319 // important to optimize the data Nodes - either the whole compilation will 320 // be tossed or this path (and all data Nodes) will go dead. 321 if (iff->outcnt() != 2) return; 322 323 // Make control-dependent data Nodes on the live path (path that will remain 324 // once the dominated IF is removed) become control-dependent on the 325 // dominating projection. 326 Node* dp = iff->proj_out_or_null(pop == Op_IfTrue); 327 328 // Loop predicates may have depending checks which should not 329 // be skipped. For example, range check predicate has two checks 330 // for lower and upper bounds. 331 if (dp == nullptr) 332 return; 333 334 ProjNode* dp_proj = dp->as_Proj(); 335 ProjNode* unc_proj = iff->proj_out(1 - dp_proj->_con)->as_Proj(); 336 if (exclude_loop_predicate && 337 (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) != nullptr || 338 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_profile_predicate) != nullptr || 339 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check) != nullptr)) { 340 // If this is a range check (IfNode::is_range_check), do not 341 // reorder because Compile::allow_range_check_smearing might have 342 // changed the check. 343 return; // Let IGVN transformation change control dependence. 344 } 345 346 IdealLoopTree* old_loop = get_loop(dp); 347 348 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { 349 Node* cd = dp->fast_out(i); // Control-dependent node 350 // Do not rewire Div and Mod nodes which could have a zero divisor to avoid skipping their zero check. 351 if (cd->depends_only_on_test() && _igvn.no_dependent_zero_check(cd)) { 352 assert(cd->in(0) == dp, ""); 353 _igvn.replace_input_of(cd, 0, prevdom); 354 set_early_ctrl(cd, false); 355 IdealLoopTree* new_loop = get_loop(get_ctrl(cd)); 356 if (old_loop != new_loop) { 357 if (!old_loop->_child) { 358 old_loop->_body.yank(cd); 359 } 360 if (!new_loop->_child) { 361 new_loop->_body.push(cd); 362 } 363 } 364 --i; 365 --imax; 366 } 367 } 368 } 369 370 //------------------------------has_local_phi_input---------------------------- 371 // Return TRUE if 'n' has Phi inputs from its local block and no other 372 // block-local inputs (all non-local-phi inputs come from earlier blocks) 373 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) { 374 Node *n_ctrl = get_ctrl(n); 375 // See if some inputs come from a Phi in this block, or from before 376 // this block. 377 uint i; 378 for( i = 1; i < n->req(); i++ ) { 379 Node *phi = n->in(i); 380 if( phi->is_Phi() && phi->in(0) == n_ctrl ) 381 break; 382 } 383 if( i >= n->req() ) 384 return nullptr; // No Phi inputs; nowhere to clone thru 385 386 // Check for inputs created between 'n' and the Phi input. These 387 // must split as well; they have already been given the chance 388 // (courtesy of a post-order visit) and since they did not we must 389 // recover the 'cost' of splitting them by being very profitable 390 // when splitting 'n'. Since this is unlikely we simply give up. 391 for( i = 1; i < n->req(); i++ ) { 392 Node *m = n->in(i); 393 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) { 394 // We allow the special case of AddP's with no local inputs. 395 // This allows us to split-up address expressions. 396 if (m->is_AddP() && 397 get_ctrl(m->in(AddPNode::Base)) != n_ctrl && 398 get_ctrl(m->in(AddPNode::Address)) != n_ctrl && 399 get_ctrl(m->in(AddPNode::Offset)) != n_ctrl) { 400 // Move the AddP up to the dominating point. That's fine because control of m's inputs 401 // must dominate get_ctrl(m) == n_ctrl and we just checked that the input controls are != n_ctrl. 402 Node* c = find_non_split_ctrl(idom(n_ctrl)); 403 if (c->is_OuterStripMinedLoop()) { 404 c->as_Loop()->verify_strip_mined(1); 405 c = c->in(LoopNode::EntryControl); 406 } 407 set_ctrl_and_loop(m, c); 408 continue; 409 } 410 return nullptr; 411 } 412 assert(n->is_Phi() || m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control"); 413 } 414 415 return n_ctrl; 416 } 417 418 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2). 419 Node* PhaseIdealLoop::remix_address_expressions_add_left_shift(Node* n, IdealLoopTree* n_loop, Node* n_ctrl, BasicType bt) { 420 assert(bt == T_INT || bt == T_LONG, "only for integers"); 421 int n_op = n->Opcode(); 422 423 if (n_op == Op_LShift(bt)) { 424 // Scale is loop invariant 425 Node* scale = n->in(2); 426 Node* scale_ctrl = get_ctrl(scale); 427 IdealLoopTree* scale_loop = get_loop(scale_ctrl); 428 if (n_loop == scale_loop || !scale_loop->is_member(n_loop)) { 429 return nullptr; 430 } 431 const TypeInt* scale_t = scale->bottom_type()->isa_int(); 432 if (scale_t != nullptr && scale_t->is_con() && scale_t->get_con() >= 16) { 433 return nullptr; // Dont bother with byte/short masking 434 } 435 // Add must vary with loop (else shift would be loop-invariant) 436 Node* add = n->in(1); 437 Node* add_ctrl = get_ctrl(add); 438 IdealLoopTree* add_loop = get_loop(add_ctrl); 439 if (n_loop != add_loop) { 440 return nullptr; // happens w/ evil ZKM loops 441 } 442 443 // Convert I-V into I+ (0-V); same for V-I 444 if (add->Opcode() == Op_Sub(bt) && 445 _igvn.type(add->in(1)) != TypeInteger::zero(bt)) { 446 assert(add->Opcode() == Op_SubI || add->Opcode() == Op_SubL, ""); 447 Node* zero = _igvn.integercon(0, bt); 448 set_ctrl(zero, C->root()); 449 Node* neg = SubNode::make(zero, add->in(2), bt); 450 register_new_node(neg, get_ctrl(add->in(2))); 451 add = AddNode::make(add->in(1), neg, bt); 452 register_new_node(add, add_ctrl); 453 } 454 if (add->Opcode() != Op_Add(bt)) return nullptr; 455 assert(add->Opcode() == Op_AddI || add->Opcode() == Op_AddL, ""); 456 // See if one add input is loop invariant 457 Node* add_var = add->in(1); 458 Node* add_var_ctrl = get_ctrl(add_var); 459 IdealLoopTree* add_var_loop = get_loop(add_var_ctrl); 460 Node* add_invar = add->in(2); 461 Node* add_invar_ctrl = get_ctrl(add_invar); 462 IdealLoopTree* add_invar_loop = get_loop(add_invar_ctrl); 463 if (add_invar_loop == n_loop) { 464 // Swap to find the invariant part 465 add_invar = add_var; 466 add_invar_ctrl = add_var_ctrl; 467 add_invar_loop = add_var_loop; 468 add_var = add->in(2); 469 } else if (add_var_loop != n_loop) { // Else neither input is loop invariant 470 return nullptr; 471 } 472 if (n_loop == add_invar_loop || !add_invar_loop->is_member(n_loop)) { 473 return nullptr; // No invariant part of the add? 474 } 475 476 // Yes! Reshape address expression! 477 Node* inv_scale = LShiftNode::make(add_invar, scale, bt); 478 Node* inv_scale_ctrl = 479 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ? 480 add_invar_ctrl : scale_ctrl; 481 register_new_node(inv_scale, inv_scale_ctrl); 482 Node* var_scale = LShiftNode::make(add_var, scale, bt); 483 register_new_node(var_scale, n_ctrl); 484 Node* var_add = AddNode::make(var_scale, inv_scale, bt); 485 register_new_node(var_add, n_ctrl); 486 _igvn.replace_node(n, var_add); 487 return var_add; 488 } 489 return nullptr; 490 } 491 492 //------------------------------remix_address_expressions---------------------- 493 // Rework addressing expressions to get the most loop-invariant stuff 494 // moved out. We'd like to do all associative operators, but it's especially 495 // important (common) to do address expressions. 496 Node* PhaseIdealLoop::remix_address_expressions(Node* n) { 497 if (!has_ctrl(n)) return nullptr; 498 Node* n_ctrl = get_ctrl(n); 499 IdealLoopTree* n_loop = get_loop(n_ctrl); 500 501 // See if 'n' mixes loop-varying and loop-invariant inputs and 502 // itself is loop-varying. 503 504 // Only interested in binary ops (and AddP) 505 if (n->req() < 3 || n->req() > 4) return nullptr; 506 507 Node* n1_ctrl = get_ctrl(n->in( 1)); 508 Node* n2_ctrl = get_ctrl(n->in( 2)); 509 Node* n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3)); 510 IdealLoopTree* n1_loop = get_loop(n1_ctrl); 511 IdealLoopTree* n2_loop = get_loop(n2_ctrl); 512 IdealLoopTree* n3_loop = get_loop(n3_ctrl); 513 514 // Does one of my inputs spin in a tighter loop than self? 515 if ((n_loop->is_member(n1_loop) && n_loop != n1_loop) || 516 (n_loop->is_member(n2_loop) && n_loop != n2_loop) || 517 (n_loop->is_member(n3_loop) && n_loop != n3_loop)) { 518 return nullptr; // Leave well enough alone 519 } 520 521 // Is at least one of my inputs loop-invariant? 522 if (n1_loop == n_loop && 523 n2_loop == n_loop && 524 n3_loop == n_loop) { 525 return nullptr; // No loop-invariant inputs 526 } 527 528 Node* res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_INT); 529 if (res != nullptr) { 530 return res; 531 } 532 res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_LONG); 533 if (res != nullptr) { 534 return res; 535 } 536 537 int n_op = n->Opcode(); 538 // Replace (I+V) with (V+I) 539 if (n_op == Op_AddI || 540 n_op == Op_AddL || 541 n_op == Op_AddF || 542 n_op == Op_AddD || 543 n_op == Op_MulI || 544 n_op == Op_MulL || 545 n_op == Op_MulF || 546 n_op == Op_MulD) { 547 if (n2_loop == n_loop) { 548 assert(n1_loop != n_loop, ""); 549 n->swap_edges(1, 2); 550 } 551 } 552 553 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V), 554 // but not if I2 is a constant. 555 if (n_op == Op_AddP) { 556 if (n2_loop == n_loop && n3_loop != n_loop) { 557 if (n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con()) { 558 Node* n22_ctrl = get_ctrl(n->in(2)->in(2)); 559 Node* n23_ctrl = get_ctrl(n->in(2)->in(3)); 560 IdealLoopTree* n22loop = get_loop(n22_ctrl); 561 IdealLoopTree* n23_loop = get_loop(n23_ctrl); 562 if (n22loop != n_loop && n22loop->is_member(n_loop) && 563 n23_loop == n_loop) { 564 Node* add1 = new AddPNode(n->in(1), n->in(2)->in(2), n->in(3)); 565 // Stuff new AddP in the loop preheader 566 register_new_node(add1, n_loop->_head->as_Loop()->skip_strip_mined(1)->in(LoopNode::EntryControl)); 567 Node* add2 = new AddPNode(n->in(1), add1, n->in(2)->in(3)); 568 register_new_node(add2, n_ctrl); 569 _igvn.replace_node(n, add2); 570 return add2; 571 } 572 } 573 } 574 575 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V) 576 if (n2_loop != n_loop && n3_loop == n_loop) { 577 if (n->in(3)->Opcode() == Op_AddX) { 578 Node* V = n->in(3)->in(1); 579 Node* I = n->in(3)->in(2); 580 if (is_member(n_loop,get_ctrl(V))) { 581 } else { 582 Node *tmp = V; V = I; I = tmp; 583 } 584 if (!is_member(n_loop,get_ctrl(I))) { 585 Node* add1 = new AddPNode(n->in(1), n->in(2), I); 586 // Stuff new AddP in the loop preheader 587 register_new_node(add1, n_loop->_head->as_Loop()->skip_strip_mined(1)->in(LoopNode::EntryControl)); 588 Node* add2 = new AddPNode(n->in(1), add1, V); 589 register_new_node(add2, n_ctrl); 590 _igvn.replace_node(n, add2); 591 return add2; 592 } 593 } 594 } 595 } 596 597 return nullptr; 598 } 599 600 // Optimize ((in1[2*i] * in2[2*i]) + (in1[2*i+1] * in2[2*i+1])) 601 Node *PhaseIdealLoop::convert_add_to_muladd(Node* n) { 602 assert(n->Opcode() == Op_AddI, "sanity"); 603 Node * nn = nullptr; 604 Node * in1 = n->in(1); 605 Node * in2 = n->in(2); 606 if (in1->Opcode() == Op_MulI && in2->Opcode() == Op_MulI) { 607 IdealLoopTree* loop_n = get_loop(get_ctrl(n)); 608 if (loop_n->is_counted() && 609 loop_n->_head->as_Loop()->is_valid_counted_loop(T_INT) && 610 Matcher::match_rule_supported(Op_MulAddVS2VI) && 611 Matcher::match_rule_supported(Op_MulAddS2I)) { 612 Node* mul_in1 = in1->in(1); 613 Node* mul_in2 = in1->in(2); 614 Node* mul_in3 = in2->in(1); 615 Node* mul_in4 = in2->in(2); 616 if (mul_in1->Opcode() == Op_LoadS && 617 mul_in2->Opcode() == Op_LoadS && 618 mul_in3->Opcode() == Op_LoadS && 619 mul_in4->Opcode() == Op_LoadS) { 620 IdealLoopTree* loop1 = get_loop(get_ctrl(mul_in1)); 621 IdealLoopTree* loop2 = get_loop(get_ctrl(mul_in2)); 622 IdealLoopTree* loop3 = get_loop(get_ctrl(mul_in3)); 623 IdealLoopTree* loop4 = get_loop(get_ctrl(mul_in4)); 624 IdealLoopTree* loop5 = get_loop(get_ctrl(in1)); 625 IdealLoopTree* loop6 = get_loop(get_ctrl(in2)); 626 // All nodes should be in the same counted loop. 627 if (loop_n == loop1 && loop_n == loop2 && loop_n == loop3 && 628 loop_n == loop4 && loop_n == loop5 && loop_n == loop6) { 629 Node* adr1 = mul_in1->in(MemNode::Address); 630 Node* adr2 = mul_in2->in(MemNode::Address); 631 Node* adr3 = mul_in3->in(MemNode::Address); 632 Node* adr4 = mul_in4->in(MemNode::Address); 633 if (adr1->is_AddP() && adr2->is_AddP() && adr3->is_AddP() && adr4->is_AddP()) { 634 if ((adr1->in(AddPNode::Base) == adr3->in(AddPNode::Base)) && 635 (adr2->in(AddPNode::Base) == adr4->in(AddPNode::Base))) { 636 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in3, mul_in4); 637 register_new_node(nn, get_ctrl(n)); 638 _igvn.replace_node(n, nn); 639 return nn; 640 } else if ((adr1->in(AddPNode::Base) == adr4->in(AddPNode::Base)) && 641 (adr2->in(AddPNode::Base) == adr3->in(AddPNode::Base))) { 642 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in4, mul_in3); 643 register_new_node(nn, get_ctrl(n)); 644 _igvn.replace_node(n, nn); 645 return nn; 646 } 647 } 648 } 649 } 650 } 651 } 652 return nn; 653 } 654 655 //------------------------------conditional_move------------------------------- 656 // Attempt to replace a Phi with a conditional move. We have some pretty 657 // strict profitability requirements. All Phis at the merge point must 658 // be converted, so we can remove the control flow. We need to limit the 659 // number of c-moves to a small handful. All code that was in the side-arms 660 // of the CFG diamond is now speculatively executed. This code has to be 661 // "cheap enough". We are pretty much limited to CFG diamonds that merge 662 // 1 or 2 items with a total of 1 or 2 ops executed speculatively. 663 Node *PhaseIdealLoop::conditional_move( Node *region ) { 664 665 assert(region->is_Region(), "sanity check"); 666 if (region->req() != 3) return nullptr; 667 668 // Check for CFG diamond 669 Node *lp = region->in(1); 670 Node *rp = region->in(2); 671 if (!lp || !rp) return nullptr; 672 Node *lp_c = lp->in(0); 673 if (lp_c == nullptr || lp_c != rp->in(0) || !lp_c->is_If()) return nullptr; 674 IfNode *iff = lp_c->as_If(); 675 676 // Check for ops pinned in an arm of the diamond. 677 // Can't remove the control flow in this case 678 if (lp->outcnt() > 1) return nullptr; 679 if (rp->outcnt() > 1) return nullptr; 680 681 IdealLoopTree* r_loop = get_loop(region); 682 assert(r_loop == get_loop(iff), "sanity"); 683 // Always convert to CMOVE if all results are used only outside this loop. 684 bool used_inside_loop = (r_loop == _ltree_root); 685 686 // Check profitability 687 int cost = 0; 688 int phis = 0; 689 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 690 Node *out = region->fast_out(i); 691 if (!out->is_Phi()) continue; // Ignore other control edges, etc 692 phis++; 693 PhiNode* phi = out->as_Phi(); 694 BasicType bt = phi->type()->basic_type(); 695 switch (bt) { 696 case T_DOUBLE: 697 case T_FLOAT: 698 if (C->use_cmove()) { 699 continue; //TODO: maybe we want to add some cost 700 } 701 cost += Matcher::float_cmove_cost(); // Could be very expensive 702 break; 703 case T_LONG: { 704 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's 705 } 706 case T_INT: // These all CMOV fine 707 case T_ADDRESS: { // (RawPtr) 708 cost++; 709 break; 710 } 711 case T_NARROWOOP: // Fall through 712 case T_OBJECT: { // Base oops are OK, but not derived oops 713 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr(); 714 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a 715 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus 716 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we 717 // have a Phi for the base here that we convert to a CMOVE all is well 718 // and good. But if the base is dead, we'll not make a CMOVE. Later 719 // the allocator will have to produce a base by creating a CMOVE of the 720 // relevant bases. This puts the allocator in the business of 721 // manufacturing expensive instructions, generally a bad plan. 722 // Just Say No to Conditionally-Moved Derived Pointers. 723 if (tp && tp->offset() != 0) 724 return nullptr; 725 cost++; 726 break; 727 } 728 default: 729 return nullptr; // In particular, can't do memory or I/O 730 } 731 // Add in cost any speculative ops 732 for (uint j = 1; j < region->req(); j++) { 733 Node *proj = region->in(j); 734 Node *inp = phi->in(j); 735 if (inp->isa_InlineType()) { 736 // TODO 8302217 This prevents PhiNode::push_inline_types_through 737 return nullptr; 738 } 739 if (get_ctrl(inp) == proj) { // Found local op 740 cost++; 741 // Check for a chain of dependent ops; these will all become 742 // speculative in a CMOV. 743 for (uint k = 1; k < inp->req(); k++) 744 if (get_ctrl(inp->in(k)) == proj) 745 cost += ConditionalMoveLimit; // Too much speculative goo 746 } 747 } 748 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode. 749 // This will likely Split-If, a higher-payoff operation. 750 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) { 751 Node* use = phi->fast_out(k); 752 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr()) 753 cost += ConditionalMoveLimit; 754 // Is there a use inside the loop? 755 // Note: check only basic types since CMoveP is pinned. 756 if (!used_inside_loop && is_java_primitive(bt)) { 757 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use); 758 if (r_loop == u_loop || r_loop->is_member(u_loop)) { 759 used_inside_loop = true; 760 } 761 } 762 } 763 }//for 764 Node* bol = iff->in(1); 765 if (bol->Opcode() == Op_Opaque4) { 766 return nullptr; // Ignore loop predicate checks (the Opaque4 ensures they will go away) 767 } 768 assert(bol->Opcode() == Op_Bool, "Unexpected node"); 769 int cmp_op = bol->in(1)->Opcode(); 770 if (cmp_op == Op_SubTypeCheck) { // SubTypeCheck expansion expects an IfNode 771 return nullptr; 772 } 773 // It is expensive to generate flags from a float compare. 774 // Avoid duplicated float compare. 775 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return nullptr; 776 777 float infrequent_prob = PROB_UNLIKELY_MAG(3); 778 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop. 779 if (used_inside_loop) { 780 if (cost >= ConditionalMoveLimit) return nullptr; // Too much goo 781 782 // BlockLayoutByFrequency optimization moves infrequent branch 783 // from hot path. No point in CMOV'ing in such case (110 is used 784 // instead of 100 to take into account not exactness of float value). 785 if (BlockLayoutByFrequency) { 786 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f); 787 } 788 } 789 // Check for highly predictable branch. No point in CMOV'ing if 790 // we are going to predict accurately all the time. 791 if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) { 792 //keep going 793 } else if (iff->_prob < infrequent_prob || 794 iff->_prob > (1.0f - infrequent_prob)) 795 return nullptr; 796 797 // -------------- 798 // Now replace all Phis with CMOV's 799 Node *cmov_ctrl = iff->in(0); 800 uint flip = (lp->Opcode() == Op_IfTrue); 801 Node_List wq; 802 while (1) { 803 PhiNode* phi = nullptr; 804 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 805 Node *out = region->fast_out(i); 806 if (out->is_Phi()) { 807 phi = out->as_Phi(); 808 break; 809 } 810 } 811 if (phi == nullptr || _igvn.type(phi) == Type::TOP) { 812 break; 813 } 814 if (PrintOpto && VerifyLoopOptimizations) { tty->print_cr("CMOV"); } 815 // Move speculative ops 816 wq.push(phi); 817 while (wq.size() > 0) { 818 Node *n = wq.pop(); 819 for (uint j = 1; j < n->req(); j++) { 820 Node* m = n->in(j); 821 if (m != nullptr && !is_dominator(get_ctrl(m), cmov_ctrl)) { 822 #ifndef PRODUCT 823 if (PrintOpto && VerifyLoopOptimizations) { 824 tty->print(" speculate: "); 825 m->dump(); 826 } 827 #endif 828 set_ctrl(m, cmov_ctrl); 829 wq.push(m); 830 } 831 } 832 } 833 Node *cmov = CMoveNode::make(cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi)); 834 register_new_node( cmov, cmov_ctrl ); 835 _igvn.replace_node( phi, cmov ); 836 #ifndef PRODUCT 837 if (TraceLoopOpts) { 838 tty->print("CMOV "); 839 r_loop->dump_head(); 840 if (Verbose) { 841 bol->in(1)->dump(1); 842 cmov->dump(1); 843 } 844 } 845 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } ); 846 #endif 847 } 848 849 // The useless CFG diamond will fold up later; see the optimization in 850 // RegionNode::Ideal. 851 _igvn._worklist.push(region); 852 853 return iff->in(1); 854 } 855 856 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) { 857 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 858 Node* u = m->fast_out(i); 859 if (u->is_CFG()) { 860 if (u->is_NeverBranch()) { 861 u = u->as_NeverBranch()->proj_out(0); 862 enqueue_cfg_uses(u, wq); 863 } else { 864 wq.push(u); 865 } 866 } 867 } 868 } 869 870 // Try moving a store out of a loop, right before the loop 871 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) { 872 // Store has to be first in the loop body 873 IdealLoopTree *n_loop = get_loop(n_ctrl); 874 if (n->is_Store() && n_loop != _ltree_root && 875 n_loop->is_loop() && n_loop->_head->is_Loop() && 876 n->in(0) != nullptr) { 877 Node* address = n->in(MemNode::Address); 878 Node* value = n->in(MemNode::ValueIn); 879 Node* mem = n->in(MemNode::Memory); 880 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 881 IdealLoopTree* value_loop = get_loop(get_ctrl(value)); 882 883 // - address and value must be loop invariant 884 // - memory must be a memory Phi for the loop 885 // - Store must be the only store on this memory slice in the 886 // loop: if there's another store following this one then value 887 // written at iteration i by the second store could be overwritten 888 // at iteration i+n by the first store: it's not safe to move the 889 // first store out of the loop 890 // - nothing must observe the memory Phi: it guarantees no read 891 // before the store, we are also guaranteed the store post 892 // dominates the loop head (ignoring a possible early 893 // exit). Otherwise there would be extra Phi involved between the 894 // loop's Phi and the store. 895 // - there must be no early exit from the loop before the Store 896 // (such an exit most of the time would be an extra use of the 897 // memory Phi but sometimes is a bottom memory Phi that takes the 898 // store as input). 899 900 if (!n_loop->is_member(address_loop) && 901 !n_loop->is_member(value_loop) && 902 mem->is_Phi() && mem->in(0) == n_loop->_head && 903 mem->outcnt() == 1 && 904 mem->in(LoopNode::LoopBackControl) == n) { 905 906 assert(n_loop->_tail != nullptr, "need a tail"); 907 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop"); 908 909 // Verify that there's no early exit of the loop before the store. 910 bool ctrl_ok = false; 911 { 912 // Follow control from loop head until n, we exit the loop or 913 // we reach the tail 914 ResourceMark rm; 915 Unique_Node_List wq; 916 wq.push(n_loop->_head); 917 918 for (uint next = 0; next < wq.size(); ++next) { 919 Node *m = wq.at(next); 920 if (m == n->in(0)) { 921 ctrl_ok = true; 922 continue; 923 } 924 assert(!has_ctrl(m), "should be CFG"); 925 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) { 926 ctrl_ok = false; 927 break; 928 } 929 enqueue_cfg_uses(m, wq); 930 if (wq.size() > 10) { 931 ctrl_ok = false; 932 break; 933 } 934 } 935 } 936 if (ctrl_ok) { 937 // move the Store 938 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem); 939 _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl)); 940 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl)); 941 // Disconnect the phi now. An empty phi can confuse other 942 // optimizations in this pass of loop opts. 943 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl)); 944 n_loop->_body.yank(mem); 945 946 set_ctrl_and_loop(n, n->in(0)); 947 948 return n; 949 } 950 } 951 } 952 return nullptr; 953 } 954 955 // Try moving a store out of a loop, right after the loop 956 void PhaseIdealLoop::try_move_store_after_loop(Node* n) { 957 if (n->is_Store() && n->in(0) != nullptr) { 958 Node *n_ctrl = get_ctrl(n); 959 IdealLoopTree *n_loop = get_loop(n_ctrl); 960 // Store must be in a loop 961 if (n_loop != _ltree_root && !n_loop->_irreducible) { 962 Node* address = n->in(MemNode::Address); 963 Node* value = n->in(MemNode::ValueIn); 964 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 965 // address must be loop invariant 966 if (!n_loop->is_member(address_loop)) { 967 // Store must be last on this memory slice in the loop and 968 // nothing in the loop must observe it 969 Node* phi = nullptr; 970 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 971 Node* u = n->fast_out(i); 972 if (has_ctrl(u)) { // control use? 973 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 974 if (!n_loop->is_member(u_loop)) { 975 continue; 976 } 977 if (u->is_Phi() && u->in(0) == n_loop->_head) { 978 assert(_igvn.type(u) == Type::MEMORY, "bad phi"); 979 // multiple phis on the same slice are possible 980 if (phi != nullptr) { 981 return; 982 } 983 phi = u; 984 continue; 985 } 986 } 987 return; 988 } 989 if (phi != nullptr) { 990 // Nothing in the loop before the store (next iteration) 991 // must observe the stored value 992 bool mem_ok = true; 993 { 994 ResourceMark rm; 995 Unique_Node_List wq; 996 wq.push(phi); 997 for (uint next = 0; next < wq.size() && mem_ok; ++next) { 998 Node *m = wq.at(next); 999 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) { 1000 Node* u = m->fast_out(i); 1001 if (u->is_Store() || u->is_Phi()) { 1002 if (u != n) { 1003 wq.push(u); 1004 mem_ok = (wq.size() <= 10); 1005 } 1006 } else { 1007 mem_ok = false; 1008 break; 1009 } 1010 } 1011 } 1012 } 1013 if (mem_ok) { 1014 // Move the store out of the loop if the LCA of all 1015 // users (except for the phi) is outside the loop. 1016 Node* hook = new Node(1); 1017 hook->init_req(0, n_ctrl); // Add an input to prevent hook from being dead 1018 _igvn.rehash_node_delayed(phi); 1019 int count = phi->replace_edge(n, hook, &_igvn); 1020 assert(count > 0, "inconsistent phi"); 1021 1022 // Compute latest point this store can go 1023 Node* lca = get_late_ctrl(n, get_ctrl(n)); 1024 if (lca->is_OuterStripMinedLoop()) { 1025 lca = lca->in(LoopNode::EntryControl); 1026 } 1027 if (n_loop->is_member(get_loop(lca))) { 1028 // LCA is in the loop - bail out 1029 _igvn.replace_node(hook, n); 1030 return; 1031 } 1032 #ifdef ASSERT 1033 if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) { 1034 assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined"); 1035 n_loop->_head->as_Loop()->verify_strip_mined(1); 1036 Node* outer = n_loop->_head->as_CountedLoop()->outer_loop(); 1037 IdealLoopTree* outer_loop = get_loop(outer); 1038 assert(n_loop->_parent == outer_loop, "broken loop tree"); 1039 assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state"); 1040 } 1041 #endif 1042 lca = place_outside_loop(lca, n_loop); 1043 assert(!n_loop->is_member(get_loop(lca)), "control must not be back in the loop"); 1044 assert(get_loop(lca)->_nest < n_loop->_nest || lca->in(0)->is_NeverBranch(), "must not be moved into inner loop"); 1045 1046 // Move store out of the loop 1047 _igvn.replace_node(hook, n->in(MemNode::Memory)); 1048 _igvn.replace_input_of(n, 0, lca); 1049 set_ctrl_and_loop(n, lca); 1050 1051 // Disconnect the phi now. An empty phi can confuse other 1052 // optimizations in this pass of loop opts.. 1053 if (phi->in(LoopNode::LoopBackControl) == phi) { 1054 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl)); 1055 n_loop->_body.yank(phi); 1056 } 1057 } 1058 } 1059 } 1060 } 1061 } 1062 } 1063 1064 // If UseArrayMarkWordCheck is enabled, we can't use immutable memory for the flat array check 1065 // because we are loading the mark word which is mutable. Although the bits we are interested in 1066 // are immutable (we check for markWord::unlocked_value), we need to use raw memory to not break 1067 // anti dependency analysis. Below code will attempt to still move flat array checks out of loops, 1068 // mainly to enable loop unswitching. 1069 void PhaseIdealLoop::move_flat_array_check_out_of_loop(Node* n) { 1070 // Skip checks for more than one array 1071 if (n->req() > 3) { 1072 return; 1073 } 1074 Node* mem = n->in(FlatArrayCheckNode::Memory); 1075 Node* array = n->in(FlatArrayCheckNode::ArrayOrKlass)->uncast(); 1076 IdealLoopTree* check_loop = get_loop(get_ctrl(n)); 1077 IdealLoopTree* ary_loop = get_loop(get_ctrl(array)); 1078 1079 // Check if array is loop invariant 1080 if (!check_loop->is_member(ary_loop)) { 1081 // Walk up memory graph from the check until we leave the loop 1082 VectorSet wq; 1083 wq.set(mem->_idx); 1084 while (check_loop->is_member(get_loop(ctrl_or_self(mem)))) { 1085 if (mem->is_Phi()) { 1086 mem = mem->in(1); 1087 } else if (mem->is_MergeMem()) { 1088 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw); 1089 } else if (mem->is_Proj()) { 1090 mem = mem->in(0); 1091 } else if (mem->is_MemBar() || mem->is_SafePoint()) { 1092 mem = mem->in(TypeFunc::Memory); 1093 } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) { 1094 mem = mem->in(MemNode::Memory); 1095 } else { 1096 #ifdef ASSERT 1097 mem->dump(); 1098 #endif 1099 ShouldNotReachHere(); 1100 } 1101 if (wq.test_set(mem->_idx)) { 1102 return; 1103 } 1104 } 1105 // Replace memory input and re-compute ctrl to move the check out of the loop 1106 _igvn.replace_input_of(n, 1, mem); 1107 set_ctrl_and_loop(n, get_early_ctrl(n)); 1108 Node* bol = n->unique_out(); 1109 set_ctrl_and_loop(bol, get_early_ctrl(bol)); 1110 } 1111 } 1112 1113 //------------------------------split_if_with_blocks_pre----------------------- 1114 // Do the real work in a non-recursive function. Data nodes want to be 1115 // cloned in the pre-order so they can feed each other nicely. 1116 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) { 1117 // Cloning these guys is unlikely to win 1118 int n_op = n->Opcode(); 1119 if (n_op == Op_MergeMem) { 1120 return n; 1121 } 1122 if (n->is_Proj()) { 1123 return n; 1124 } 1125 1126 if (UseArrayMarkWordCheck && n->isa_FlatArrayCheck()) { 1127 move_flat_array_check_out_of_loop(n); 1128 return n; 1129 } 1130 1131 // Do not clone-up CmpFXXX variations, as these are always 1132 // followed by a CmpI 1133 if (n->is_Cmp()) { 1134 return n; 1135 } 1136 // Attempt to use a conditional move instead of a phi/branch 1137 if (ConditionalMoveLimit > 0 && n_op == Op_Region) { 1138 Node *cmov = conditional_move( n ); 1139 if (cmov) { 1140 return cmov; 1141 } 1142 } 1143 if (n->is_CFG() || n->is_LoadStore()) { 1144 return n; 1145 } 1146 if (n->is_Opaque1()) { // Opaque nodes cannot be mod'd 1147 if (!C->major_progress()) { // If chance of no more loop opts... 1148 _igvn._worklist.push(n); // maybe we'll remove them 1149 } 1150 return n; 1151 } 1152 1153 if (n->is_Con()) { 1154 return n; // No cloning for Con nodes 1155 } 1156 1157 Node *n_ctrl = get_ctrl(n); 1158 if (!n_ctrl) { 1159 return n; // Dead node 1160 } 1161 1162 Node* res = try_move_store_before_loop(n, n_ctrl); 1163 if (res != nullptr) { 1164 return n; 1165 } 1166 1167 // Attempt to remix address expressions for loop invariants 1168 Node *m = remix_address_expressions( n ); 1169 if( m ) return m; 1170 1171 if (n_op == Op_AddI) { 1172 Node *nn = convert_add_to_muladd( n ); 1173 if ( nn ) return nn; 1174 } 1175 1176 if (n->is_ConstraintCast()) { 1177 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this); 1178 // ConstraintCastNode::dominating_cast() uses node control input to determine domination. 1179 // Node control inputs don't necessarily agree with loop control info (due to 1180 // transformations happened in between), thus additional dominance check is needed 1181 // to keep loop info valid. 1182 if (dom_cast != nullptr && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) { 1183 _igvn.replace_node(n, dom_cast); 1184 return dom_cast; 1185 } 1186 } 1187 1188 // Determine if the Node has inputs from some local Phi. 1189 // Returns the block to clone thru. 1190 Node *n_blk = has_local_phi_input( n ); 1191 if( !n_blk ) return n; 1192 1193 // Do not clone the trip counter through on a CountedLoop 1194 // (messes up the canonical shape). 1195 if (((n_blk->is_CountedLoop() || (n_blk->is_Loop() && n_blk->as_Loop()->is_loop_nest_inner_loop())) && n->Opcode() == Op_AddI) || 1196 (n_blk->is_LongCountedLoop() && n->Opcode() == Op_AddL)) { 1197 return n; 1198 } 1199 // Pushing a shift through the iv Phi can get in the way of addressing optimizations or range check elimination 1200 if (n_blk->is_BaseCountedLoop() && n->Opcode() == Op_LShift(n_blk->as_BaseCountedLoop()->bt()) && 1201 n->in(1) == n_blk->as_BaseCountedLoop()->phi()) { 1202 return n; 1203 } 1204 1205 // Check for having no control input; not pinned. Allow 1206 // dominating control. 1207 if (n->in(0)) { 1208 Node *dom = idom(n_blk); 1209 if (dom_lca(n->in(0), dom) != n->in(0)) { 1210 return n; 1211 } 1212 } 1213 // Policy: when is it profitable. You must get more wins than 1214 // policy before it is considered profitable. Policy is usually 0, 1215 // so 1 win is considered profitable. Big merges will require big 1216 // cloning, so get a larger policy. 1217 int policy = n_blk->req() >> 2; 1218 1219 // If the loop is a candidate for range check elimination, 1220 // delay splitting through it's phi until a later loop optimization 1221 if (n_blk->is_BaseCountedLoop()) { 1222 IdealLoopTree *lp = get_loop(n_blk); 1223 if (lp && lp->_rce_candidate) { 1224 return n; 1225 } 1226 } 1227 1228 if (must_throttle_split_if()) return n; 1229 1230 // Split 'n' through the merge point if it is profitable 1231 Node *phi = split_thru_phi( n, n_blk, policy ); 1232 if (!phi) return n; 1233 1234 // Found a Phi to split thru! 1235 // Replace 'n' with the new phi 1236 _igvn.replace_node( n, phi ); 1237 // Moved a load around the loop, 'en-registering' something. 1238 if (n_blk->is_Loop() && n->is_Load() && 1239 !phi->in(LoopNode::LoopBackControl)->is_Load()) 1240 C->set_major_progress(); 1241 1242 return phi; 1243 } 1244 1245 static bool merge_point_too_heavy(Compile* C, Node* region) { 1246 // Bail out if the region and its phis have too many users. 1247 int weight = 0; 1248 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1249 weight += region->fast_out(i)->outcnt(); 1250 } 1251 int nodes_left = C->max_node_limit() - C->live_nodes(); 1252 if (weight * 8 > nodes_left) { 1253 if (PrintOpto) { 1254 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight); 1255 } 1256 return true; 1257 } else { 1258 return false; 1259 } 1260 } 1261 1262 static bool merge_point_safe(Node* region) { 1263 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode 1264 // having a PhiNode input. This sidesteps the dangerous case where the split 1265 // ConvI2LNode may become TOP if the input Value() does not 1266 // overlap the ConvI2L range, leaving a node which may not dominate its 1267 // uses. 1268 // A better fix for this problem can be found in the BugTraq entry, but 1269 // expediency for Mantis demands this hack. 1270 #ifdef _LP64 1271 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1272 Node* n = region->fast_out(i); 1273 if (n->is_Phi()) { 1274 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1275 Node* m = n->fast_out(j); 1276 if (m->Opcode() == Op_ConvI2L) 1277 return false; 1278 if (m->is_CastII()) { 1279 return false; 1280 } 1281 } 1282 } 1283 } 1284 #endif 1285 return true; 1286 } 1287 1288 1289 //------------------------------place_outside_loop--------------------------------- 1290 // Place some computation outside of this loop on the path to the use passed as argument 1291 Node* PhaseIdealLoop::place_outside_loop(Node* useblock, IdealLoopTree* loop) const { 1292 Node* head = loop->_head; 1293 assert(!loop->is_member(get_loop(useblock)), "must be outside loop"); 1294 if (head->is_Loop() && head->as_Loop()->is_strip_mined()) { 1295 loop = loop->_parent; 1296 assert(loop->_head->is_OuterStripMinedLoop(), "malformed strip mined loop"); 1297 } 1298 1299 // Pick control right outside the loop 1300 for (;;) { 1301 Node* dom = idom(useblock); 1302 if (loop->is_member(get_loop(dom)) || 1303 // NeverBranch nodes are not assigned to the loop when constructed 1304 (dom->is_NeverBranch() && loop->is_member(get_loop(dom->in(0))))) { 1305 break; 1306 } 1307 useblock = dom; 1308 } 1309 assert(find_non_split_ctrl(useblock) == useblock, "should be non split control"); 1310 return useblock; 1311 } 1312 1313 1314 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) { 1315 if (!n->is_If() || n->is_BaseCountedLoopEnd()) { 1316 return false; 1317 } 1318 if (!n->in(0)->is_Region()) { 1319 return false; 1320 } 1321 1322 Node* region = n->in(0); 1323 Node* dom = idom(region); 1324 if (!dom->is_If() || !n->as_If()->same_condition(dom, &_igvn)) { 1325 return false; 1326 } 1327 IfNode* dom_if = dom->as_If(); 1328 Node* proj_true = dom_if->proj_out(1); 1329 Node* proj_false = dom_if->proj_out(0); 1330 1331 for (uint i = 1; i < region->req(); i++) { 1332 if (is_dominator(proj_true, region->in(i))) { 1333 continue; 1334 } 1335 if (is_dominator(proj_false, region->in(i))) { 1336 continue; 1337 } 1338 return false; 1339 } 1340 1341 return true; 1342 } 1343 1344 1345 bool PhaseIdealLoop::can_split_if(Node* n_ctrl) { 1346 if (must_throttle_split_if()) { 1347 return false; 1348 } 1349 1350 // Do not do 'split-if' if irreducible loops are present. 1351 if (_has_irreducible_loops) { 1352 return false; 1353 } 1354 1355 if (merge_point_too_heavy(C, n_ctrl)) { 1356 return false; 1357 } 1358 1359 // Do not do 'split-if' if some paths are dead. First do dead code 1360 // elimination and then see if its still profitable. 1361 for (uint i = 1; i < n_ctrl->req(); i++) { 1362 if (n_ctrl->in(i) == C->top()) { 1363 return false; 1364 } 1365 } 1366 1367 // If trying to do a 'Split-If' at the loop head, it is only 1368 // profitable if the cmp folds up on BOTH paths. Otherwise we 1369 // risk peeling a loop forever. 1370 1371 // CNC - Disabled for now. Requires careful handling of loop 1372 // body selection for the cloned code. Also, make sure we check 1373 // for any input path not being in the same loop as n_ctrl. For 1374 // irreducible loops we cannot check for 'n_ctrl->is_Loop()' 1375 // because the alternative loop entry points won't be converted 1376 // into LoopNodes. 1377 IdealLoopTree *n_loop = get_loop(n_ctrl); 1378 for (uint j = 1; j < n_ctrl->req(); j++) { 1379 if (get_loop(n_ctrl->in(j)) != n_loop) { 1380 return false; 1381 } 1382 } 1383 1384 // Check for safety of the merge point. 1385 if (!merge_point_safe(n_ctrl)) { 1386 return false; 1387 } 1388 1389 return true; 1390 } 1391 1392 // Detect if the node is the inner strip-mined loop 1393 // Return: null if it's not the case, or the exit of outer strip-mined loop 1394 static Node* is_inner_of_stripmined_loop(const Node* out) { 1395 Node* out_le = nullptr; 1396 1397 if (out->is_CountedLoopEnd()) { 1398 const CountedLoopNode* loop = out->as_CountedLoopEnd()->loopnode(); 1399 1400 if (loop != nullptr && loop->is_strip_mined()) { 1401 out_le = loop->in(LoopNode::EntryControl)->as_OuterStripMinedLoop()->outer_loop_exit(); 1402 } 1403 } 1404 1405 return out_le; 1406 } 1407 1408 bool PhaseIdealLoop::flat_array_element_type_check(Node *n) { 1409 // If the CmpP is a subtype check for a value that has just been 1410 // loaded from an array, the subtype check guarantees the value 1411 // can't be stored in a flat array and the load of the value 1412 // happens with a flat array check then: push the type check 1413 // through the phi of the flat array check. This needs special 1414 // logic because the subtype check's input is not a phi but a 1415 // LoadKlass that must first be cloned through the phi. 1416 if (n->Opcode() != Op_CmpP) { 1417 return false; 1418 } 1419 1420 Node* klassptr = n->in(1); 1421 Node* klasscon = n->in(2); 1422 1423 if (klassptr->is_DecodeNarrowPtr()) { 1424 klassptr = klassptr->in(1); 1425 } 1426 1427 if (klassptr->Opcode() != Op_LoadKlass && klassptr->Opcode() != Op_LoadNKlass) { 1428 return false; 1429 } 1430 1431 if (!klasscon->is_Con()) { 1432 return false; 1433 } 1434 1435 Node* addr = klassptr->in(MemNode::Address); 1436 1437 if (!addr->is_AddP()) { 1438 return false; 1439 } 1440 1441 intptr_t offset; 1442 Node* obj = AddPNode::Ideal_base_and_offset(addr, &_igvn, offset); 1443 1444 if (obj == nullptr) { 1445 return false; 1446 } 1447 1448 assert(obj != nullptr && addr->in(AddPNode::Base) == addr->in(AddPNode::Address), "malformed AddP?"); 1449 if (obj->Opcode() == Op_CastPP) { 1450 obj = obj->in(1); 1451 } 1452 1453 if (!obj->is_Phi()) { 1454 return false; 1455 } 1456 1457 Node* region = obj->in(0); 1458 1459 Node* phi = PhiNode::make_blank(region, n->in(1)); 1460 for (uint i = 1; i < region->req(); i++) { 1461 Node* in = obj->in(i); 1462 Node* ctrl = region->in(i); 1463 if (addr->in(AddPNode::Base) != obj) { 1464 Node* cast = addr->in(AddPNode::Base); 1465 assert(cast->Opcode() == Op_CastPP && cast->in(0) != nullptr, "inconsistent subgraph"); 1466 Node* cast_clone = cast->clone(); 1467 cast_clone->set_req(0, ctrl); 1468 cast_clone->set_req(1, in); 1469 register_new_node(cast_clone, ctrl); 1470 const Type* tcast = cast_clone->Value(&_igvn); 1471 _igvn.set_type(cast_clone, tcast); 1472 cast_clone->as_Type()->set_type(tcast); 1473 in = cast_clone; 1474 } 1475 Node* addr_clone = addr->clone(); 1476 addr_clone->set_req(AddPNode::Base, in); 1477 addr_clone->set_req(AddPNode::Address, in); 1478 register_new_node(addr_clone, ctrl); 1479 _igvn.set_type(addr_clone, addr_clone->Value(&_igvn)); 1480 Node* klassptr_clone = klassptr->clone(); 1481 klassptr_clone->set_req(2, addr_clone); 1482 register_new_node(klassptr_clone, ctrl); 1483 _igvn.set_type(klassptr_clone, klassptr_clone->Value(&_igvn)); 1484 if (klassptr != n->in(1)) { 1485 Node* decode = n->in(1); 1486 assert(decode->is_DecodeNarrowPtr(), "inconsistent subgraph"); 1487 Node* decode_clone = decode->clone(); 1488 decode_clone->set_req(1, klassptr_clone); 1489 register_new_node(decode_clone, ctrl); 1490 _igvn.set_type(decode_clone, decode_clone->Value(&_igvn)); 1491 klassptr_clone = decode_clone; 1492 } 1493 phi->set_req(i, klassptr_clone); 1494 } 1495 register_new_node(phi, region); 1496 Node* orig = n->in(1); 1497 _igvn.replace_input_of(n, 1, phi); 1498 split_if_with_blocks_post(n); 1499 if (n->outcnt() != 0) { 1500 _igvn.replace_input_of(n, 1, orig); 1501 _igvn.remove_dead_node(phi); 1502 } 1503 return true; 1504 } 1505 1506 //------------------------------split_if_with_blocks_post---------------------- 1507 // Do the real work in a non-recursive function. CFG hackery wants to be 1508 // in the post-order, so it can dirty the I-DOM info and not use the dirtied 1509 // info. 1510 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) { 1511 1512 if (flat_array_element_type_check(n)) { 1513 return; 1514 } 1515 1516 // Cloning Cmp through Phi's involves the split-if transform. 1517 // FastLock is not used by an If 1518 if (n->is_Cmp() && !n->is_FastLock()) { 1519 Node *n_ctrl = get_ctrl(n); 1520 // Determine if the Node has inputs from some local Phi. 1521 // Returns the block to clone thru. 1522 Node *n_blk = has_local_phi_input(n); 1523 if (n_blk != n_ctrl) { 1524 return; 1525 } 1526 1527 if (!can_split_if(n_ctrl)) { 1528 return; 1529 } 1530 1531 if (n->outcnt() != 1) { 1532 return; // Multiple bool's from 1 compare? 1533 } 1534 Node *bol = n->unique_out(); 1535 assert(bol->is_Bool(), "expect a bool here"); 1536 if (bol->outcnt() != 1) { 1537 return;// Multiple branches from 1 compare? 1538 } 1539 Node *iff = bol->unique_out(); 1540 1541 // Check some safety conditions 1542 if (iff->is_If()) { // Classic split-if? 1543 if (iff->in(0) != n_ctrl) { 1544 return; // Compare must be in same blk as if 1545 } 1546 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE 1547 // Can't split CMove with different control. 1548 if (get_ctrl(iff) != n_ctrl) { 1549 return; 1550 } 1551 if (get_ctrl(iff->in(2)) == n_ctrl || 1552 get_ctrl(iff->in(3)) == n_ctrl) { 1553 return; // Inputs not yet split-up 1554 } 1555 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) { 1556 return; // Loop-invar test gates loop-varying CMOVE 1557 } 1558 } else { 1559 return; // some other kind of node, such as an Allocate 1560 } 1561 1562 // When is split-if profitable? Every 'win' on means some control flow 1563 // goes dead, so it's almost always a win. 1564 int policy = 0; 1565 // Split compare 'n' through the merge point if it is profitable 1566 Node *phi = split_thru_phi( n, n_ctrl, policy); 1567 if (!phi) { 1568 return; 1569 } 1570 1571 // Found a Phi to split thru! 1572 // Replace 'n' with the new phi 1573 _igvn.replace_node(n, phi); 1574 1575 // Now split the bool up thru the phi 1576 Node *bolphi = split_thru_phi(bol, n_ctrl, -1); 1577 guarantee(bolphi != nullptr, "null boolean phi node"); 1578 1579 _igvn.replace_node(bol, bolphi); 1580 assert(iff->in(1) == bolphi, ""); 1581 1582 if (bolphi->Value(&_igvn)->singleton()) { 1583 return; 1584 } 1585 1586 // Conditional-move? Must split up now 1587 if (!iff->is_If()) { 1588 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1); 1589 _igvn.replace_node(iff, cmovphi); 1590 return; 1591 } 1592 1593 // Now split the IF 1594 do_split_if(iff); 1595 return; 1596 } 1597 1598 // Two identical ifs back to back can be merged 1599 if (try_merge_identical_ifs(n)) { 1600 return; 1601 } 1602 1603 // Check for an IF ready to split; one that has its 1604 // condition codes input coming from a Phi at the block start. 1605 int n_op = n->Opcode(); 1606 1607 // Check for an IF being dominated by another IF same test 1608 if (n_op == Op_If || 1609 n_op == Op_RangeCheck) { 1610 Node *bol = n->in(1); 1611 uint max = bol->outcnt(); 1612 // Check for same test used more than once? 1613 if (bol->is_Bool() && (max > 1 || bol->in(1)->is_SubTypeCheck())) { 1614 // Search up IDOMs to see if this IF is dominated. 1615 Node* cmp = bol->in(1); 1616 Node *cutoff = cmp->is_SubTypeCheck() ? dom_lca(get_ctrl(cmp->in(1)), get_ctrl(cmp->in(2))) : get_ctrl(bol); 1617 1618 // Now search up IDOMs till cutoff, looking for a dominating test 1619 Node *prevdom = n; 1620 Node *dom = idom(prevdom); 1621 while (dom != cutoff) { 1622 if (dom->req() > 1 && n->as_If()->same_condition(dom, &_igvn) && prevdom->in(0) == dom && 1623 safe_for_if_replacement(dom)) { 1624 // It's invalid to move control dependent data nodes in the inner 1625 // strip-mined loop, because: 1626 // 1) break validation of LoopNode::verify_strip_mined() 1627 // 2) move code with side-effect in strip-mined loop 1628 // Move to the exit of outer strip-mined loop in that case. 1629 Node* out_le = is_inner_of_stripmined_loop(dom); 1630 if (out_le != nullptr) { 1631 prevdom = out_le; 1632 } 1633 // Replace the dominated test with an obvious true or false. 1634 // Place it on the IGVN worklist for later cleanup. 1635 C->set_major_progress(); 1636 dominated_by(prevdom->as_IfProj(), n->as_If(), false, true); 1637 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } ); 1638 return; 1639 } 1640 prevdom = dom; 1641 dom = idom(prevdom); 1642 } 1643 } 1644 } 1645 1646 try_sink_out_of_loop(n); 1647 1648 try_move_store_after_loop(n); 1649 1650 // Remove multiple allocations of the same inline type 1651 if (n->is_InlineType()) { 1652 n->as_InlineType()->remove_redundant_allocations(this); 1653 } 1654 } 1655 1656 // Transform: 1657 // 1658 // if (some_condition) { 1659 // // body 1 1660 // } else { 1661 // // body 2 1662 // } 1663 // if (some_condition) { 1664 // // body 3 1665 // } else { 1666 // // body 4 1667 // } 1668 // 1669 // into: 1670 // 1671 // 1672 // if (some_condition) { 1673 // // body 1 1674 // // body 3 1675 // } else { 1676 // // body 2 1677 // // body 4 1678 // } 1679 bool PhaseIdealLoop::try_merge_identical_ifs(Node* n) { 1680 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) { 1681 Node *n_ctrl = n->in(0); 1682 IfNode* dom_if = idom(n_ctrl)->as_If(); 1683 if (n->in(1) != dom_if->in(1)) { 1684 assert(n->in(1)->in(1)->is_SubTypeCheck() && 1685 (n->in(1)->in(1)->as_SubTypeCheck()->method() != nullptr || 1686 dom_if->in(1)->in(1)->as_SubTypeCheck()->method() != nullptr), "only for subtype checks with profile data attached"); 1687 _igvn.replace_input_of(n, 1, dom_if->in(1)); 1688 } 1689 ProjNode* dom_proj_true = dom_if->proj_out(1); 1690 ProjNode* dom_proj_false = dom_if->proj_out(0); 1691 1692 // Now split the IF 1693 RegionNode* new_false_region; 1694 RegionNode* new_true_region; 1695 do_split_if(n, &new_false_region, &new_true_region); 1696 assert(new_false_region->req() == new_true_region->req(), ""); 1697 #ifdef ASSERT 1698 for (uint i = 1; i < new_false_region->req(); ++i) { 1699 assert(new_false_region->in(i)->in(0) == new_true_region->in(i)->in(0), "unexpected shape following split if"); 1700 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"); 1701 } 1702 #endif 1703 assert(new_false_region->in(1)->in(0)->in(1) == dom_if->in(1), "dominating if and dominated if after split must share test"); 1704 1705 // We now have: 1706 // if (some_condition) { 1707 // // body 1 1708 // if (some_condition) { 1709 // body3: // new_true_region 1710 // // body3 1711 // } else { 1712 // goto body4; 1713 // } 1714 // } else { 1715 // // body 2 1716 // if (some_condition) { 1717 // goto body3; 1718 // } else { 1719 // body4: // new_false_region 1720 // // body4; 1721 // } 1722 // } 1723 // 1724 1725 // clone pinned nodes thru the resulting regions 1726 push_pinned_nodes_thru_region(dom_if, new_true_region); 1727 push_pinned_nodes_thru_region(dom_if, new_false_region); 1728 1729 // Optimize out the cloned ifs. Because pinned nodes were cloned, this also allows a CastPP that would be dependent 1730 // 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 1731 // unrelated control dependency. 1732 for (uint i = 1; i < new_false_region->req(); i++) { 1733 if (is_dominator(dom_proj_true, new_false_region->in(i))) { 1734 dominated_by(dom_proj_true->as_IfProj(), new_false_region->in(i)->in(0)->as_If(), false, false); 1735 } else { 1736 assert(is_dominator(dom_proj_false, new_false_region->in(i)), "bad if"); 1737 dominated_by(dom_proj_false->as_IfProj(), new_false_region->in(i)->in(0)->as_If(), false, false); 1738 } 1739 } 1740 return true; 1741 } 1742 return false; 1743 } 1744 1745 void PhaseIdealLoop::push_pinned_nodes_thru_region(IfNode* dom_if, Node* region) { 1746 for (DUIterator i = region->outs(); region->has_out(i); i++) { 1747 Node* u = region->out(i); 1748 if (!has_ctrl(u) || u->is_Phi() || !u->depends_only_on_test() || !_igvn.no_dependent_zero_check(u)) { 1749 continue; 1750 } 1751 assert(u->in(0) == region, "not a control dependent node?"); 1752 uint j = 1; 1753 for (; j < u->req(); ++j) { 1754 Node* in = u->in(j); 1755 if (!is_dominator(ctrl_or_self(in), dom_if)) { 1756 break; 1757 } 1758 } 1759 if (j == u->req()) { 1760 Node *phi = PhiNode::make_blank(region, u); 1761 for (uint k = 1; k < region->req(); ++k) { 1762 Node* clone = u->clone(); 1763 clone->set_req(0, region->in(k)); 1764 register_new_node(clone, region->in(k)); 1765 phi->init_req(k, clone); 1766 } 1767 register_new_node(phi, region); 1768 _igvn.replace_node(u, phi); 1769 --i; 1770 } 1771 } 1772 } 1773 1774 bool PhaseIdealLoop::safe_for_if_replacement(const Node* dom) const { 1775 if (!dom->is_CountedLoopEnd()) { 1776 return true; 1777 } 1778 CountedLoopEndNode* le = dom->as_CountedLoopEnd(); 1779 CountedLoopNode* cl = le->loopnode(); 1780 if (cl == nullptr) { 1781 return true; 1782 } 1783 if (!cl->is_main_loop()) { 1784 return true; 1785 } 1786 if (cl->is_canonical_loop_entry() == nullptr) { 1787 return true; 1788 } 1789 // Further unrolling is possible so loop exit condition might change 1790 return false; 1791 } 1792 1793 // See if a shared loop-varying computation has no loop-varying uses. 1794 // Happens if something is only used for JVM state in uncommon trap exits, 1795 // like various versions of induction variable+offset. Clone the 1796 // computation per usage to allow it to sink out of the loop. 1797 void PhaseIdealLoop::try_sink_out_of_loop(Node* n) { 1798 if (has_ctrl(n) && 1799 !n->is_Phi() && 1800 !n->is_Bool() && 1801 !n->is_Proj() && 1802 !n->is_MergeMem() && 1803 !n->is_CMove() && 1804 n->Opcode() != Op_Opaque4 && 1805 !n->is_Type()) { 1806 Node *n_ctrl = get_ctrl(n); 1807 IdealLoopTree *n_loop = get_loop(n_ctrl); 1808 1809 if (n->in(0) != nullptr) { 1810 IdealLoopTree* loop_ctrl = get_loop(n->in(0)); 1811 if (n_loop != loop_ctrl && n_loop->is_member(loop_ctrl)) { 1812 // n has a control input inside a loop but get_ctrl() is member of an outer loop. This could happen, for example, 1813 // for Div nodes inside a loop (control input inside loop) without a use except for an UCT (outside the loop). 1814 // Rewire control of n to right outside of the loop, regardless if its input(s) are later sunk or not. 1815 _igvn.replace_input_of(n, 0, place_outside_loop(n_ctrl, loop_ctrl)); 1816 } 1817 } 1818 if (n_loop != _ltree_root && n->outcnt() > 1) { 1819 // Compute early control: needed for anti-dependence analysis. It's also possible that as a result of 1820 // previous transformations in this loop opts round, the node can be hoisted now: early control will tell us. 1821 Node* early_ctrl = compute_early_ctrl(n, n_ctrl); 1822 if (n_loop->is_member(get_loop(early_ctrl)) && // check that this one can't be hoisted now 1823 ctrl_of_all_uses_out_of_loop(n, early_ctrl, n_loop)) { // All uses in outer loops! 1824 assert(!n->is_Store() && !n->is_LoadStore(), "no node with a side effect"); 1825 Node* outer_loop_clone = nullptr; 1826 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin;) { 1827 Node* u = n->last_out(j); // Clone private computation per use 1828 _igvn.rehash_node_delayed(u); 1829 Node* x = n->clone(); // Clone computation 1830 Node* x_ctrl = nullptr; 1831 if (u->is_Phi()) { 1832 // Replace all uses of normal nodes. Replace Phi uses 1833 // individually, so the separate Nodes can sink down 1834 // different paths. 1835 uint k = 1; 1836 while (u->in(k) != n) k++; 1837 u->set_req(k, x); 1838 // x goes next to Phi input path 1839 x_ctrl = u->in(0)->in(k); 1840 // Find control for 'x' next to use but not inside inner loops. 1841 x_ctrl = place_outside_loop(x_ctrl, n_loop); 1842 --j; 1843 } else { // Normal use 1844 if (has_ctrl(u)) { 1845 x_ctrl = get_ctrl(u); 1846 } else { 1847 x_ctrl = u->in(0); 1848 } 1849 // Find control for 'x' next to use but not inside inner loops. 1850 x_ctrl = place_outside_loop(x_ctrl, n_loop); 1851 // Replace all uses 1852 if (u->is_ConstraintCast() && _igvn.type(n)->higher_equal(u->bottom_type()) && u->in(0) == x_ctrl) { 1853 // If we're sinking a chain of data nodes, we might have inserted a cast to pin the use which is not necessary 1854 // anymore now that we're going to pin n as well 1855 _igvn.replace_node(u, x); 1856 --j; 1857 } else { 1858 int nb = u->replace_edge(n, x, &_igvn); 1859 j -= nb; 1860 } 1861 } 1862 1863 if (n->is_Load()) { 1864 // For loads, add a control edge to a CFG node outside of the loop 1865 // to force them to not combine and return back inside the loop 1866 // during GVN optimization (4641526). 1867 assert(x_ctrl == get_late_ctrl_with_anti_dep(x->as_Load(), early_ctrl, x_ctrl), "anti-dependences were already checked"); 1868 1869 IdealLoopTree* x_loop = get_loop(x_ctrl); 1870 Node* x_head = x_loop->_head; 1871 if (x_head->is_Loop() && x_head->is_OuterStripMinedLoop()) { 1872 // Do not add duplicate LoadNodes to the outer strip mined loop 1873 if (outer_loop_clone != nullptr) { 1874 _igvn.replace_node(x, outer_loop_clone); 1875 continue; 1876 } 1877 outer_loop_clone = x; 1878 } 1879 x->set_req(0, x_ctrl); 1880 } else if (n->in(0) != nullptr){ 1881 x->set_req(0, x_ctrl); 1882 } 1883 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone"); 1884 assert(!n_loop->is_member(get_loop(x_ctrl)), "should have moved out of loop"); 1885 register_new_node(x, x_ctrl); 1886 1887 // Chain of AddP nodes: (AddP base (AddP base (AddP base ))) 1888 // All AddP nodes must keep the same base after sinking so: 1889 // 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, 1890 // their bases remain the same. 1891 // (see 2- below) 1892 assert(!x->is_AddP() || !x->in(AddPNode::Address)->is_AddP() || 1893 x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base) || 1894 !x->in(AddPNode::Address)->in(AddPNode::Base)->eqv_uncast(x->in(AddPNode::Base)), "unexpected AddP shape"); 1895 if (x->in(0) == nullptr && !x->is_DecodeNarrowPtr() && 1896 !(x->is_AddP() && x->in(AddPNode::Address)->is_AddP() && x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base))) { 1897 assert(!x->is_Load(), "load should be pinned"); 1898 // Use a cast node to pin clone out of loop 1899 Node* cast = nullptr; 1900 for (uint k = 0; k < x->req(); k++) { 1901 Node* in = x->in(k); 1902 if (in != nullptr && n_loop->is_member(get_loop(get_ctrl(in)))) { 1903 const Type* in_t = _igvn.type(in); 1904 cast = ConstraintCastNode::make_cast_for_type(x_ctrl, in, in_t, ConstraintCastNode::UnconditionalDependency); 1905 } 1906 if (cast != nullptr) { 1907 Node* prev = _igvn.hash_find_insert(cast); 1908 if (prev != nullptr && get_ctrl(prev) == x_ctrl) { 1909 cast->destruct(&_igvn); 1910 cast = prev; 1911 } else { 1912 register_new_node(cast, x_ctrl); 1913 } 1914 x->replace_edge(in, cast); 1915 // Chain of AddP nodes: 1916 // 2- A CastPP of the base is only added now that all AddP nodes are sunk 1917 if (x->is_AddP() && k == AddPNode::Base) { 1918 update_addp_chain_base(x, n->in(AddPNode::Base), cast); 1919 } 1920 break; 1921 } 1922 } 1923 assert(cast != nullptr, "must have added a cast to pin the node"); 1924 } 1925 } 1926 _igvn.remove_dead_node(n); 1927 } 1928 _dom_lca_tags_round = 0; 1929 } 1930 } 1931 } 1932 1933 void PhaseIdealLoop::update_addp_chain_base(Node* x, Node* old_base, Node* new_base) { 1934 ResourceMark rm; 1935 Node_List wq; 1936 wq.push(x); 1937 while (wq.size() != 0) { 1938 Node* n = wq.pop(); 1939 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1940 Node* u = n->fast_out(i); 1941 if (u->is_AddP() && u->in(AddPNode::Base) == old_base) { 1942 _igvn.replace_input_of(u, AddPNode::Base, new_base); 1943 wq.push(u); 1944 } 1945 } 1946 } 1947 } 1948 1949 // Compute the early control of a node by following its inputs until we reach 1950 // nodes that are pinned. Then compute the LCA of the control of all pinned nodes. 1951 Node* PhaseIdealLoop::compute_early_ctrl(Node* n, Node* n_ctrl) { 1952 Node* early_ctrl = nullptr; 1953 ResourceMark rm; 1954 Unique_Node_List wq; 1955 wq.push(n); 1956 for (uint i = 0; i < wq.size(); i++) { 1957 Node* m = wq.at(i); 1958 Node* c = nullptr; 1959 if (m->is_CFG()) { 1960 c = m; 1961 } else if (m->pinned()) { 1962 c = m->in(0); 1963 } else { 1964 for (uint j = 0; j < m->req(); j++) { 1965 Node* in = m->in(j); 1966 if (in != nullptr) { 1967 wq.push(in); 1968 } 1969 } 1970 } 1971 if (c != nullptr) { 1972 assert(is_dominator(c, n_ctrl), "control input must dominate current control"); 1973 if (early_ctrl == nullptr || is_dominator(early_ctrl, c)) { 1974 early_ctrl = c; 1975 } 1976 } 1977 } 1978 assert(is_dominator(early_ctrl, n_ctrl), "early control must dominate current control"); 1979 return early_ctrl; 1980 } 1981 1982 bool PhaseIdealLoop::ctrl_of_all_uses_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop) { 1983 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1984 Node* u = n->fast_out(i); 1985 if (u->is_Opaque1()) { 1986 return false; // Found loop limit, bugfix for 4677003 1987 } 1988 // We can't reuse tags in PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal() so make sure calls to 1989 // get_late_ctrl_with_anti_dep() use their own tag 1990 _dom_lca_tags_round++; 1991 assert(_dom_lca_tags_round != 0, "shouldn't wrap around"); 1992 1993 if (u->is_Phi()) { 1994 for (uint j = 1; j < u->req(); ++j) { 1995 if (u->in(j) == n && !ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, u->in(0)->in(j))) { 1996 return false; 1997 } 1998 } 1999 } else { 2000 Node* ctrl = has_ctrl(u) ? get_ctrl(u) : u->in(0); 2001 if (!ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, ctrl)) { 2002 return false; 2003 } 2004 } 2005 } 2006 return true; 2007 } 2008 2009 bool PhaseIdealLoop::ctrl_of_use_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop, Node* ctrl) { 2010 if (n->is_Load()) { 2011 ctrl = get_late_ctrl_with_anti_dep(n->as_Load(), n_ctrl, ctrl); 2012 } 2013 IdealLoopTree *u_loop = get_loop(ctrl); 2014 if (u_loop == n_loop) { 2015 return false; // Found loop-varying use 2016 } 2017 if (n_loop->is_member(u_loop)) { 2018 return false; // Found use in inner loop 2019 } 2020 // 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 2021 // to a check that's eliminated by range check elimination, it becomes input to an expression that feeds into the exit 2022 // test of the pre loop above the point in the graph where it's pinned. 2023 if (n_loop->_head->is_CountedLoop() && n_loop->_head->as_CountedLoop()->is_pre_loop() && 2024 u_loop->_head->is_CountedLoop() && u_loop->_head->as_CountedLoop()->is_main_loop() && 2025 n_loop->_next == get_loop(u_loop->_head->as_CountedLoop()->skip_strip_mined())) { 2026 return false; 2027 } 2028 return true; 2029 } 2030 2031 //------------------------------split_if_with_blocks--------------------------- 2032 // Check for aggressive application of 'split-if' optimization, 2033 // using basic block level info. 2034 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack) { 2035 Node* root = C->root(); 2036 visited.set(root->_idx); // first, mark root as visited 2037 // Do pre-visit work for root 2038 Node* n = split_if_with_blocks_pre(root); 2039 uint cnt = n->outcnt(); 2040 uint i = 0; 2041 2042 while (true) { 2043 // Visit all children 2044 if (i < cnt) { 2045 Node* use = n->raw_out(i); 2046 ++i; 2047 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) { 2048 // Now do pre-visit work for this use 2049 use = split_if_with_blocks_pre(use); 2050 nstack.push(n, i); // Save parent and next use's index. 2051 n = use; // Process all children of current use. 2052 cnt = use->outcnt(); 2053 i = 0; 2054 } 2055 } 2056 else { 2057 // All of n's children have been processed, complete post-processing. 2058 if (cnt != 0 && !n->is_Con()) { 2059 assert(has_node(n), "no dead nodes"); 2060 split_if_with_blocks_post(n); 2061 } 2062 if (must_throttle_split_if()) { 2063 nstack.clear(); 2064 } 2065 if (nstack.is_empty()) { 2066 // Finished all nodes on stack. 2067 break; 2068 } 2069 // Get saved parent node and next use's index. Visit the rest of uses. 2070 n = nstack.node(); 2071 cnt = n->outcnt(); 2072 i = nstack.index(); 2073 nstack.pop(); 2074 } 2075 } 2076 } 2077 2078 2079 //============================================================================= 2080 // 2081 // C L O N E A L O O P B O D Y 2082 // 2083 2084 //------------------------------clone_iff-------------------------------------- 2085 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 2086 // "Nearly" because all Nodes have been cloned from the original in the loop, 2087 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 2088 // through the Phi recursively, and return a Bool. 2089 Node* PhaseIdealLoop::clone_iff(PhiNode* phi) { 2090 2091 // Convert this Phi into a Phi merging Bools 2092 uint i; 2093 for (i = 1; i < phi->req(); i++) { 2094 Node *b = phi->in(i); 2095 if (b->is_Phi()) { 2096 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi())); 2097 } else { 2098 assert(b->is_Bool() || b->Opcode() == Op_Opaque4, ""); 2099 } 2100 } 2101 2102 Node* n = phi->in(1); 2103 Node* sample_opaque = nullptr; 2104 Node *sample_bool = nullptr; 2105 if (n->Opcode() == Op_Opaque4) { 2106 sample_opaque = n; 2107 sample_bool = n->in(1); 2108 assert(sample_bool->is_Bool(), "wrong type"); 2109 } else { 2110 sample_bool = n; 2111 } 2112 Node* sample_cmp = sample_bool->in(1); 2113 const Type* t = Type::TOP; 2114 const TypePtr* at = nullptr; 2115 if (sample_cmp->is_FlatArrayCheck()) { 2116 // Left input of a FlatArrayCheckNode is memory, set the (adr) type of the phi accordingly 2117 assert(sample_cmp->in(1)->bottom_type() == Type::MEMORY, "unexpected input type"); 2118 t = Type::MEMORY; 2119 at = TypeRawPtr::BOTTOM; 2120 } 2121 2122 // Make Phis to merge the Cmp's inputs. 2123 PhiNode *phi1 = new PhiNode(phi->in(0), t, at); 2124 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP); 2125 for (i = 1; i < phi->req(); i++) { 2126 Node *n1 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1); 2127 Node *n2 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2); 2128 phi1->set_req(i, n1); 2129 phi2->set_req(i, n2); 2130 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 2131 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 2132 } 2133 // See if these Phis have been made before. 2134 // Register with optimizer 2135 Node *hit1 = _igvn.hash_find_insert(phi1); 2136 if (hit1) { // Hit, toss just made Phi 2137 _igvn.remove_dead_node(phi1); // Remove new phi 2138 assert(hit1->is_Phi(), "" ); 2139 phi1 = (PhiNode*)hit1; // Use existing phi 2140 } else { // Miss 2141 _igvn.register_new_node_with_optimizer(phi1); 2142 } 2143 Node *hit2 = _igvn.hash_find_insert(phi2); 2144 if (hit2) { // Hit, toss just made Phi 2145 _igvn.remove_dead_node(phi2); // Remove new phi 2146 assert(hit2->is_Phi(), "" ); 2147 phi2 = (PhiNode*)hit2; // Use existing phi 2148 } else { // Miss 2149 _igvn.register_new_node_with_optimizer(phi2); 2150 } 2151 // Register Phis with loop/block info 2152 set_ctrl(phi1, phi->in(0)); 2153 set_ctrl(phi2, phi->in(0)); 2154 // Make a new Cmp 2155 Node *cmp = sample_cmp->clone(); 2156 cmp->set_req(1, phi1); 2157 cmp->set_req(2, phi2); 2158 _igvn.register_new_node_with_optimizer(cmp); 2159 set_ctrl(cmp, phi->in(0)); 2160 2161 // Make a new Bool 2162 Node *b = sample_bool->clone(); 2163 b->set_req(1,cmp); 2164 _igvn.register_new_node_with_optimizer(b); 2165 set_ctrl(b, phi->in(0)); 2166 2167 if (sample_opaque != nullptr) { 2168 Node* opaque = sample_opaque->clone(); 2169 opaque->set_req(1, b); 2170 _igvn.register_new_node_with_optimizer(opaque); 2171 set_ctrl(opaque, phi->in(0)); 2172 return opaque; 2173 } 2174 2175 assert(b->is_Bool(), ""); 2176 return b; 2177 } 2178 2179 //------------------------------clone_bool------------------------------------- 2180 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 2181 // "Nearly" because all Nodes have been cloned from the original in the loop, 2182 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 2183 // through the Phi recursively, and return a Bool. 2184 CmpNode*PhaseIdealLoop::clone_bool(PhiNode* phi) { 2185 uint i; 2186 // Convert this Phi into a Phi merging Bools 2187 for( i = 1; i < phi->req(); i++ ) { 2188 Node *b = phi->in(i); 2189 if( b->is_Phi() ) { 2190 _igvn.replace_input_of(phi, i, clone_bool(b->as_Phi())); 2191 } else { 2192 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" ); 2193 } 2194 } 2195 2196 Node *sample_cmp = phi->in(1); 2197 2198 // Make Phis to merge the Cmp's inputs. 2199 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP ); 2200 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP ); 2201 for( uint j = 1; j < phi->req(); j++ ) { 2202 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP 2203 Node *n1, *n2; 2204 if( cmp_top->is_Cmp() ) { 2205 n1 = cmp_top->in(1); 2206 n2 = cmp_top->in(2); 2207 } else { 2208 n1 = n2 = cmp_top; 2209 } 2210 phi1->set_req( j, n1 ); 2211 phi2->set_req( j, n2 ); 2212 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 2213 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 2214 } 2215 2216 // See if these Phis have been made before. 2217 // Register with optimizer 2218 Node *hit1 = _igvn.hash_find_insert(phi1); 2219 if( hit1 ) { // Hit, toss just made Phi 2220 _igvn.remove_dead_node(phi1); // Remove new phi 2221 assert( hit1->is_Phi(), "" ); 2222 phi1 = (PhiNode*)hit1; // Use existing phi 2223 } else { // Miss 2224 _igvn.register_new_node_with_optimizer(phi1); 2225 } 2226 Node *hit2 = _igvn.hash_find_insert(phi2); 2227 if( hit2 ) { // Hit, toss just made Phi 2228 _igvn.remove_dead_node(phi2); // Remove new phi 2229 assert( hit2->is_Phi(), "" ); 2230 phi2 = (PhiNode*)hit2; // Use existing phi 2231 } else { // Miss 2232 _igvn.register_new_node_with_optimizer(phi2); 2233 } 2234 // Register Phis with loop/block info 2235 set_ctrl(phi1, phi->in(0)); 2236 set_ctrl(phi2, phi->in(0)); 2237 // Make a new Cmp 2238 Node *cmp = sample_cmp->clone(); 2239 cmp->set_req( 1, phi1 ); 2240 cmp->set_req( 2, phi2 ); 2241 _igvn.register_new_node_with_optimizer(cmp); 2242 set_ctrl(cmp, phi->in(0)); 2243 2244 assert( cmp->is_Cmp(), "" ); 2245 return (CmpNode*)cmp; 2246 } 2247 2248 //------------------------------sink_use--------------------------------------- 2249 // If 'use' was in the loop-exit block, it now needs to be sunk 2250 // below the post-loop merge point. 2251 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) { 2252 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) { 2253 set_ctrl(use, post_loop); 2254 for (DUIterator j = use->outs(); use->has_out(j); j++) 2255 sink_use(use->out(j), post_loop); 2256 } 2257 } 2258 2259 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new, 2260 IdealLoopTree* loop, IdealLoopTree* outer_loop, 2261 Node_List*& split_if_set, Node_List*& split_bool_set, 2262 Node_List*& split_cex_set, Node_List& worklist, 2263 uint new_counter, CloneLoopMode mode) { 2264 Node* nnn = old_new[old->_idx]; 2265 // Copy uses to a worklist, so I can munge the def-use info 2266 // with impunity. 2267 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 2268 worklist.push(old->fast_out(j)); 2269 2270 while( worklist.size() ) { 2271 Node *use = worklist.pop(); 2272 if (!has_node(use)) continue; // Ignore dead nodes 2273 if (use->in(0) == C->top()) continue; 2274 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 2275 // Check for data-use outside of loop - at least one of OLD or USE 2276 // must not be a CFG node. 2277 #ifdef ASSERT 2278 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == nullptr) { 2279 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint(); 2280 assert(mode != IgnoreStripMined, "incorrect cloning mode"); 2281 assert((mode == ControlAroundStripMined && use == sfpt) || !use->is_reachable_from_root(), "missed a node"); 2282 } 2283 #endif 2284 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) { 2285 2286 // If the Data use is an IF, that means we have an IF outside of the 2287 // loop that is switching on a condition that is set inside of the 2288 // loop. Happens if people set a loop-exit flag; then test the flag 2289 // in the loop to break the loop, then test is again outside of the 2290 // loop to determine which way the loop exited. 2291 // Loop predicate If node connects to Bool node through Opaque1 node. 2292 // 2293 // If the use is an AllocateArray through its ValidLengthTest input, 2294 // make sure the Bool/Cmp input is cloned down to avoid a Phi between 2295 // the AllocateArray node and its ValidLengthTest input that could cause 2296 // split if to break. 2297 if (use->is_If() || use->is_CMove() || use->Opcode() == Op_Opaque4 || 2298 (use->Opcode() == Op_AllocateArray && use->in(AllocateNode::ValidLengthTest) == old)) { 2299 // Since this code is highly unlikely, we lazily build the worklist 2300 // of such Nodes to go split. 2301 if (!split_if_set) { 2302 split_if_set = new Node_List(); 2303 } 2304 split_if_set->push(use); 2305 } 2306 if (use->is_Bool()) { 2307 if (!split_bool_set) { 2308 split_bool_set = new Node_List(); 2309 } 2310 split_bool_set->push(use); 2311 } 2312 if (use->Opcode() == Op_CreateEx) { 2313 if (!split_cex_set) { 2314 split_cex_set = new Node_List(); 2315 } 2316 split_cex_set->push(use); 2317 } 2318 2319 2320 // Get "block" use is in 2321 uint idx = 0; 2322 while( use->in(idx) != old ) idx++; 2323 Node *prev = use->is_CFG() ? use : get_ctrl(use); 2324 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" ); 2325 Node *cfg = prev->_idx >= new_counter 2326 ? prev->in(2) 2327 : idom(prev); 2328 if( use->is_Phi() ) // Phi use is in prior block 2329 cfg = prev->in(idx); // NOT in block of Phi itself 2330 if (cfg->is_top()) { // Use is dead? 2331 _igvn.replace_input_of(use, idx, C->top()); 2332 continue; 2333 } 2334 2335 // If use is referenced through control edge... (idx == 0) 2336 if (mode == IgnoreStripMined && idx == 0) { 2337 LoopNode *head = loop->_head->as_Loop(); 2338 if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) { 2339 // That node is outside the inner loop, leave it outside the 2340 // outer loop as well to not confuse verification code. 2341 assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop"); 2342 _igvn.replace_input_of(use, 0, head->outer_loop_exit()); 2343 continue; 2344 } 2345 } 2346 2347 while(!outer_loop->is_member(get_loop(cfg))) { 2348 prev = cfg; 2349 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg); 2350 } 2351 // If the use occurs after merging several exits from the loop, then 2352 // old value must have dominated all those exits. Since the same old 2353 // value was used on all those exits we did not need a Phi at this 2354 // merge point. NOW we do need a Phi here. Each loop exit value 2355 // is now merged with the peeled body exit; each exit gets its own 2356 // private Phi and those Phis need to be merged here. 2357 Node *phi; 2358 if( prev->is_Region() ) { 2359 if( idx == 0 ) { // Updating control edge? 2360 phi = prev; // Just use existing control 2361 } else { // Else need a new Phi 2362 phi = PhiNode::make( prev, old ); 2363 // Now recursively fix up the new uses of old! 2364 for( uint i = 1; i < prev->req(); i++ ) { 2365 worklist.push(phi); // Onto worklist once for each 'old' input 2366 } 2367 } 2368 } else { 2369 // Get new RegionNode merging old and new loop exits 2370 prev = old_new[prev->_idx]; 2371 assert( prev, "just made this in step 7" ); 2372 if( idx == 0) { // Updating control edge? 2373 phi = prev; // Just use existing control 2374 } else { // Else need a new Phi 2375 // Make a new Phi merging data values properly 2376 phi = PhiNode::make( prev, old ); 2377 phi->set_req( 1, nnn ); 2378 } 2379 } 2380 // If inserting a new Phi, check for prior hits 2381 if( idx != 0 ) { 2382 Node *hit = _igvn.hash_find_insert(phi); 2383 if( hit == nullptr ) { 2384 _igvn.register_new_node_with_optimizer(phi); // Register new phi 2385 } else { // or 2386 // Remove the new phi from the graph and use the hit 2387 _igvn.remove_dead_node(phi); 2388 phi = hit; // Use existing phi 2389 } 2390 set_ctrl(phi, prev); 2391 } 2392 // Make 'use' use the Phi instead of the old loop body exit value 2393 assert(use->in(idx) == old, "old is still input of use"); 2394 // We notify all uses of old, including use, and the indirect uses, 2395 // that may now be optimized because we have replaced old with phi. 2396 _igvn.add_users_to_worklist(old); 2397 _igvn.replace_input_of(use, idx, phi); 2398 if( use->_idx >= new_counter ) { // If updating new phis 2399 // Not needed for correctness, but prevents a weak assert 2400 // in AddPNode from tripping (when we end up with different 2401 // base & derived Phis that will become the same after 2402 // IGVN does CSE). 2403 Node *hit = _igvn.hash_find_insert(use); 2404 if( hit ) // Go ahead and re-hash for hits. 2405 _igvn.replace_node( use, hit ); 2406 } 2407 2408 // If 'use' was in the loop-exit block, it now needs to be sunk 2409 // below the post-loop merge point. 2410 sink_use( use, prev ); 2411 } 2412 } 2413 } 2414 2415 static void collect_nodes_in_outer_loop_not_reachable_from_sfpt(Node* n, const IdealLoopTree *loop, const IdealLoopTree* outer_loop, 2416 const Node_List &old_new, Unique_Node_List& wq, PhaseIdealLoop* phase, 2417 bool check_old_new) { 2418 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2419 Node* u = n->fast_out(j); 2420 assert(check_old_new || old_new[u->_idx] == nullptr, "shouldn't have been cloned"); 2421 if (!u->is_CFG() && (!check_old_new || old_new[u->_idx] == nullptr)) { 2422 Node* c = phase->get_ctrl(u); 2423 IdealLoopTree* u_loop = phase->get_loop(c); 2424 assert(!loop->is_member(u_loop) || !loop->_body.contains(u), "can be in outer loop or out of both loops only"); 2425 if (!loop->is_member(u_loop)) { 2426 if (outer_loop->is_member(u_loop)) { 2427 wq.push(u); 2428 } else { 2429 // nodes pinned with control in the outer loop but not referenced from the safepoint must be moved out of 2430 // the outer loop too 2431 Node* u_c = u->in(0); 2432 if (u_c != nullptr) { 2433 IdealLoopTree* u_c_loop = phase->get_loop(u_c); 2434 if (outer_loop->is_member(u_c_loop) && !loop->is_member(u_c_loop)) { 2435 wq.push(u); 2436 } 2437 } 2438 } 2439 } 2440 } 2441 } 2442 } 2443 2444 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop, 2445 IdealLoopTree* outer_loop, int dd, Node_List &old_new, 2446 Node_List& extra_data_nodes) { 2447 if (head->is_strip_mined() && mode != IgnoreStripMined) { 2448 CountedLoopNode* cl = head->as_CountedLoop(); 2449 Node* l = cl->outer_loop(); 2450 Node* tail = cl->outer_loop_tail(); 2451 IfNode* le = cl->outer_loop_end(); 2452 Node* sfpt = cl->outer_safepoint(); 2453 CountedLoopEndNode* cle = cl->loopexit(); 2454 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop(); 2455 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null(); 2456 Node* cle_out = cle->proj_out(false); 2457 2458 Node* new_sfpt = nullptr; 2459 Node* new_cle_out = cle_out->clone(); 2460 old_new.map(cle_out->_idx, new_cle_out); 2461 if (mode == CloneIncludesStripMined) { 2462 // clone outer loop body 2463 Node* new_l = l->clone(); 2464 Node* new_tail = tail->clone(); 2465 IfNode* new_le = le->clone()->as_If(); 2466 new_sfpt = sfpt->clone(); 2467 2468 set_loop(new_l, outer_loop->_parent); 2469 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd); 2470 set_loop(new_cle_out, outer_loop->_parent); 2471 set_idom(new_cle_out, new_cle, dd); 2472 set_loop(new_sfpt, outer_loop->_parent); 2473 set_idom(new_sfpt, new_cle_out, dd); 2474 set_loop(new_le, outer_loop->_parent); 2475 set_idom(new_le, new_sfpt, dd); 2476 set_loop(new_tail, outer_loop->_parent); 2477 set_idom(new_tail, new_le, dd); 2478 set_idom(new_cl, new_l, dd); 2479 2480 old_new.map(l->_idx, new_l); 2481 old_new.map(tail->_idx, new_tail); 2482 old_new.map(le->_idx, new_le); 2483 old_new.map(sfpt->_idx, new_sfpt); 2484 2485 new_l->set_req(LoopNode::LoopBackControl, new_tail); 2486 new_l->set_req(0, new_l); 2487 new_tail->set_req(0, new_le); 2488 new_le->set_req(0, new_sfpt); 2489 new_sfpt->set_req(0, new_cle_out); 2490 new_cle_out->set_req(0, new_cle); 2491 new_cl->set_req(LoopNode::EntryControl, new_l); 2492 2493 _igvn.register_new_node_with_optimizer(new_l); 2494 _igvn.register_new_node_with_optimizer(new_tail); 2495 _igvn.register_new_node_with_optimizer(new_le); 2496 } else { 2497 Node *newhead = old_new[loop->_head->_idx]; 2498 newhead->as_Loop()->clear_strip_mined(); 2499 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl)); 2500 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 2501 } 2502 // Look at data node that were assigned a control in the outer 2503 // loop: they are kept in the outer loop by the safepoint so start 2504 // from the safepoint node's inputs. 2505 IdealLoopTree* outer_loop = get_loop(l); 2506 Node_Stack stack(2); 2507 stack.push(sfpt, 1); 2508 uint new_counter = C->unique(); 2509 while (stack.size() > 0) { 2510 Node* n = stack.node(); 2511 uint i = stack.index(); 2512 while (i < n->req() && 2513 (n->in(i) == nullptr || 2514 !has_ctrl(n->in(i)) || 2515 get_loop(get_ctrl(n->in(i))) != outer_loop || 2516 (old_new[n->in(i)->_idx] != nullptr && old_new[n->in(i)->_idx]->_idx >= new_counter))) { 2517 i++; 2518 } 2519 if (i < n->req()) { 2520 stack.set_index(i+1); 2521 stack.push(n->in(i), 0); 2522 } else { 2523 assert(old_new[n->_idx] == nullptr || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet"); 2524 Node* m = n == sfpt ? new_sfpt : n->clone(); 2525 if (m != nullptr) { 2526 for (uint i = 0; i < n->req(); i++) { 2527 if (m->in(i) != nullptr && old_new[m->in(i)->_idx] != nullptr) { 2528 m->set_req(i, old_new[m->in(i)->_idx]); 2529 } 2530 } 2531 } else { 2532 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?"); 2533 } 2534 if (n != sfpt) { 2535 extra_data_nodes.push(n); 2536 _igvn.register_new_node_with_optimizer(m); 2537 assert(get_ctrl(n) == cle_out, "what other control?"); 2538 set_ctrl(m, new_cle_out); 2539 old_new.map(n->_idx, m); 2540 } 2541 stack.pop(); 2542 } 2543 } 2544 if (mode == CloneIncludesStripMined) { 2545 _igvn.register_new_node_with_optimizer(new_sfpt); 2546 _igvn.register_new_node_with_optimizer(new_cle_out); 2547 } 2548 // Some other transformation may have pessimistically assigned some 2549 // data nodes to the outer loop. Set their control so they are out 2550 // of the outer loop. 2551 ResourceMark rm; 2552 Unique_Node_List wq; 2553 for (uint i = 0; i < extra_data_nodes.size(); i++) { 2554 Node* old = extra_data_nodes.at(i); 2555 collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true); 2556 } 2557 2558 for (uint i = 0; i < loop->_body.size(); i++) { 2559 Node* old = loop->_body.at(i); 2560 collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true); 2561 } 2562 2563 Node* inner_out = sfpt->in(0); 2564 if (inner_out->outcnt() > 1) { 2565 collect_nodes_in_outer_loop_not_reachable_from_sfpt(inner_out, loop, outer_loop, old_new, wq, this, true); 2566 } 2567 2568 Node* new_ctrl = cl->outer_loop_exit(); 2569 assert(get_loop(new_ctrl) != outer_loop, "must be out of the loop nest"); 2570 for (uint i = 0; i < wq.size(); i++) { 2571 Node* n = wq.at(i); 2572 set_ctrl(n, new_ctrl); 2573 if (n->in(0) != nullptr) { 2574 _igvn.replace_input_of(n, 0, new_ctrl); 2575 } 2576 collect_nodes_in_outer_loop_not_reachable_from_sfpt(n, loop, outer_loop, old_new, wq, this, false); 2577 } 2578 } else { 2579 Node *newhead = old_new[loop->_head->_idx]; 2580 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 2581 } 2582 } 2583 2584 //------------------------------clone_loop------------------------------------- 2585 // 2586 // C L O N E A L O O P B O D Y 2587 // 2588 // This is the basic building block of the loop optimizations. It clones an 2589 // entire loop body. It makes an old_new loop body mapping; with this mapping 2590 // you can find the new-loop equivalent to an old-loop node. All new-loop 2591 // nodes are exactly equal to their old-loop counterparts, all edges are the 2592 // same. All exits from the old-loop now have a RegionNode that merges the 2593 // equivalent new-loop path. This is true even for the normal "loop-exit" 2594 // condition. All uses of loop-invariant old-loop values now come from (one 2595 // or more) Phis that merge their new-loop equivalents. 2596 // 2597 // This operation leaves the graph in an illegal state: there are two valid 2598 // control edges coming from the loop pre-header to both loop bodies. I'll 2599 // definitely have to hack the graph after running this transform. 2600 // 2601 // From this building block I will further edit edges to perform loop peeling 2602 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc. 2603 // 2604 // Parameter side_by_size_idom: 2605 // When side_by_size_idom is null, the dominator tree is constructed for 2606 // the clone loop to dominate the original. Used in construction of 2607 // pre-main-post loop sequence. 2608 // When nonnull, the clone and original are side-by-side, both are 2609 // dominated by the side_by_side_idom node. Used in construction of 2610 // unswitched loops. 2611 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd, 2612 CloneLoopMode mode, Node* side_by_side_idom) { 2613 2614 LoopNode* head = loop->_head->as_Loop(); 2615 head->verify_strip_mined(1); 2616 2617 if (C->do_vector_loop() && PrintOpto) { 2618 const char* mname = C->method()->name()->as_quoted_ascii(); 2619 if (mname != nullptr) { 2620 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname); 2621 } 2622 } 2623 2624 CloneMap& cm = C->clone_map(); 2625 if (C->do_vector_loop()) { 2626 cm.set_clone_idx(cm.max_gen()+1); 2627 #ifndef PRODUCT 2628 if (PrintOpto) { 2629 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx()); 2630 loop->dump_head(); 2631 } 2632 #endif 2633 } 2634 2635 // Step 1: Clone the loop body. Make the old->new mapping. 2636 clone_loop_body(loop->_body, old_new, &cm); 2637 2638 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop; 2639 2640 // Step 2: Fix the edges in the new body. If the old input is outside the 2641 // loop use it. If the old input is INside the loop, use the corresponding 2642 // new node instead. 2643 fix_body_edges(loop->_body, loop, old_new, dd, outer_loop->_parent, false); 2644 2645 Node_List extra_data_nodes; // data nodes in the outer strip mined loop 2646 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes); 2647 2648 // Step 3: Now fix control uses. Loop varying control uses have already 2649 // been fixed up (as part of all input edges in Step 2). Loop invariant 2650 // control uses must be either an IfFalse or an IfTrue. Make a merge 2651 // point to merge the old and new IfFalse/IfTrue nodes; make the use 2652 // refer to this. 2653 Node_List worklist; 2654 uint new_counter = C->unique(); 2655 fix_ctrl_uses(loop->_body, loop, old_new, mode, side_by_side_idom, &cm, worklist); 2656 2657 // Step 4: If loop-invariant use is not control, it must be dominated by a 2658 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region 2659 // there if needed. Make a Phi there merging old and new used values. 2660 Node_List *split_if_set = nullptr; 2661 Node_List *split_bool_set = nullptr; 2662 Node_List *split_cex_set = nullptr; 2663 fix_data_uses(loop->_body, loop, mode, outer_loop, new_counter, old_new, worklist, split_if_set, split_bool_set, split_cex_set); 2664 2665 for (uint i = 0; i < extra_data_nodes.size(); i++) { 2666 Node* old = extra_data_nodes.at(i); 2667 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set, 2668 split_bool_set, split_cex_set, worklist, new_counter, 2669 mode); 2670 } 2671 2672 // Check for IFs that need splitting/cloning. Happens if an IF outside of 2673 // the loop uses a condition set in the loop. The original IF probably 2674 // takes control from one or more OLD Regions (which in turn get from NEW 2675 // Regions). In any case, there will be a set of Phis for each merge point 2676 // from the IF up to where the original BOOL def exists the loop. 2677 finish_clone_loop(split_if_set, split_bool_set, split_cex_set); 2678 2679 } 2680 2681 void PhaseIdealLoop::finish_clone_loop(Node_List* split_if_set, Node_List* split_bool_set, Node_List* split_cex_set) { 2682 if (split_if_set) { 2683 while (split_if_set->size()) { 2684 Node *iff = split_if_set->pop(); 2685 uint input = iff->Opcode() == Op_AllocateArray ? AllocateNode::ValidLengthTest : 1; 2686 if (iff->in(input)->is_Phi()) { 2687 Node *b = clone_iff(iff->in(input)->as_Phi()); 2688 _igvn.replace_input_of(iff, input, b); 2689 } 2690 } 2691 } 2692 if (split_bool_set) { 2693 while (split_bool_set->size()) { 2694 Node *b = split_bool_set->pop(); 2695 Node *phi = b->in(1); 2696 assert(phi->is_Phi(), ""); 2697 CmpNode *cmp = clone_bool((PhiNode*) phi); 2698 _igvn.replace_input_of(b, 1, cmp); 2699 } 2700 } 2701 if (split_cex_set) { 2702 while (split_cex_set->size()) { 2703 Node *b = split_cex_set->pop(); 2704 assert(b->in(0)->is_Region(), ""); 2705 assert(b->in(1)->is_Phi(), ""); 2706 assert(b->in(0)->in(0) == b->in(1)->in(0), ""); 2707 split_up(b, b->in(0), nullptr); 2708 } 2709 } 2710 } 2711 2712 void PhaseIdealLoop::fix_data_uses(Node_List& body, IdealLoopTree* loop, CloneLoopMode mode, IdealLoopTree* outer_loop, 2713 uint new_counter, Node_List &old_new, Node_List &worklist, Node_List*& split_if_set, 2714 Node_List*& split_bool_set, Node_List*& split_cex_set) { 2715 for(uint i = 0; i < body.size(); i++ ) { 2716 Node* old = body.at(i); 2717 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set, 2718 split_bool_set, split_cex_set, worklist, new_counter, 2719 mode); 2720 } 2721 } 2722 2723 void PhaseIdealLoop::fix_ctrl_uses(const Node_List& body, const IdealLoopTree* loop, Node_List &old_new, CloneLoopMode mode, 2724 Node* side_by_side_idom, CloneMap* cm, Node_List &worklist) { 2725 LoopNode* head = loop->_head->as_Loop(); 2726 for(uint i = 0; i < body.size(); i++ ) { 2727 Node* old = body.at(i); 2728 if( !old->is_CFG() ) continue; 2729 2730 // Copy uses to a worklist, so I can munge the def-use info 2731 // with impunity. 2732 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) { 2733 worklist.push(old->fast_out(j)); 2734 } 2735 2736 while (worklist.size()) { // Visit all uses 2737 Node *use = worklist.pop(); 2738 if (!has_node(use)) continue; // Ignore dead nodes 2739 IdealLoopTree *use_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use ); 2740 if (!loop->is_member(use_loop) && use->is_CFG()) { 2741 // Both OLD and USE are CFG nodes here. 2742 assert(use->is_Proj(), "" ); 2743 Node* nnn = old_new[old->_idx]; 2744 2745 Node* newuse = nullptr; 2746 if (head->is_strip_mined() && mode != IgnoreStripMined) { 2747 CountedLoopNode* cl = head->as_CountedLoop(); 2748 CountedLoopEndNode* cle = cl->loopexit(); 2749 Node* cle_out = cle->proj_out_or_null(false); 2750 if (use == cle_out) { 2751 IfNode* le = cl->outer_loop_end(); 2752 use = le->proj_out(false); 2753 use_loop = get_loop(use); 2754 if (mode == CloneIncludesStripMined) { 2755 nnn = old_new[le->_idx]; 2756 } else { 2757 newuse = old_new[cle_out->_idx]; 2758 } 2759 } 2760 } 2761 if (newuse == nullptr) { 2762 newuse = use->clone(); 2763 } 2764 2765 // Clone the loop exit control projection 2766 if (C->do_vector_loop() && cm != nullptr) { 2767 cm->verify_insert_and_clone(use, newuse, cm->clone_idx()); 2768 } 2769 newuse->set_req(0,nnn); 2770 _igvn.register_new_node_with_optimizer(newuse); 2771 set_loop(newuse, use_loop); 2772 set_idom(newuse, nnn, dom_depth(nnn) + 1 ); 2773 2774 // We need a Region to merge the exit from the peeled body and the 2775 // exit from the old loop body. 2776 RegionNode *r = new RegionNode(3); 2777 // Map the old use to the new merge point 2778 old_new.map( use->_idx, r ); 2779 uint dd_r = MIN2(dom_depth(newuse), dom_depth(use)); 2780 assert(dd_r >= dom_depth(dom_lca(newuse, use)), "" ); 2781 2782 // The original user of 'use' uses 'r' instead. 2783 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) { 2784 Node* useuse = use->last_out(l); 2785 _igvn.rehash_node_delayed(useuse); 2786 uint uses_found = 0; 2787 if (useuse->in(0) == use) { 2788 useuse->set_req(0, r); 2789 uses_found++; 2790 if (useuse->is_CFG()) { 2791 // This is not a dom_depth > dd_r because when new 2792 // control flow is constructed by a loop opt, a node and 2793 // its dominator can end up at the same dom_depth 2794 assert(dom_depth(useuse) >= dd_r, ""); 2795 set_idom(useuse, r, dom_depth(useuse)); 2796 } 2797 } 2798 for (uint k = 1; k < useuse->req(); k++) { 2799 if( useuse->in(k) == use ) { 2800 useuse->set_req(k, r); 2801 uses_found++; 2802 if (useuse->is_Loop() && k == LoopNode::EntryControl) { 2803 // This is not a dom_depth > dd_r because when new 2804 // control flow is constructed by a loop opt, a node 2805 // and its dominator can end up at the same dom_depth 2806 assert(dom_depth(useuse) >= dd_r , ""); 2807 set_idom(useuse, r, dom_depth(useuse)); 2808 } 2809 } 2810 } 2811 l -= uses_found; // we deleted 1 or more copies of this edge 2812 } 2813 2814 // Now finish up 'r' 2815 r->set_req(1, newuse); 2816 r->set_req(2, use); 2817 _igvn.register_new_node_with_optimizer(r); 2818 set_loop(r, use_loop); 2819 set_idom(r, (side_by_side_idom == nullptr) ? newuse->in(0) : side_by_side_idom, dd_r); 2820 } // End of if a loop-exit test 2821 } 2822 } 2823 } 2824 2825 void PhaseIdealLoop::fix_body_edges(const Node_List &body, IdealLoopTree* loop, const Node_List &old_new, int dd, 2826 IdealLoopTree* parent, bool partial) { 2827 for(uint i = 0; i < body.size(); i++ ) { 2828 Node *old = body.at(i); 2829 Node *nnn = old_new[old->_idx]; 2830 // Fix CFG/Loop controlling the new node 2831 if (has_ctrl(old)) { 2832 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]); 2833 } else { 2834 set_loop(nnn, parent); 2835 if (old->outcnt() > 0) { 2836 Node* dom = idom(old); 2837 if (old_new[dom->_idx] != nullptr) { 2838 dom = old_new[dom->_idx]; 2839 set_idom(nnn, dom, dd ); 2840 } 2841 } 2842 } 2843 // Correct edges to the new node 2844 for (uint j = 0; j < nnn->req(); j++) { 2845 Node *n = nnn->in(j); 2846 if (n != nullptr) { 2847 IdealLoopTree *old_in_loop = get_loop(has_ctrl(n) ? get_ctrl(n) : n); 2848 if (loop->is_member(old_in_loop)) { 2849 if (old_new[n->_idx] != nullptr) { 2850 nnn->set_req(j, old_new[n->_idx]); 2851 } else { 2852 assert(!body.contains(n), ""); 2853 assert(partial, "node not cloned"); 2854 } 2855 } 2856 } 2857 } 2858 _igvn.hash_find_insert(nnn); 2859 } 2860 } 2861 2862 void PhaseIdealLoop::clone_loop_body(const Node_List& body, Node_List &old_new, CloneMap* cm) { 2863 for (uint i = 0; i < body.size(); i++) { 2864 Node* old = body.at(i); 2865 Node* nnn = old->clone(); 2866 old_new.map(old->_idx, nnn); 2867 if (C->do_vector_loop() && cm != nullptr) { 2868 cm->verify_insert_and_clone(old, nnn, cm->clone_idx()); 2869 } 2870 _igvn.register_new_node_with_optimizer(nnn); 2871 } 2872 } 2873 2874 2875 //---------------------- stride_of_possible_iv ------------------------------------- 2876 // Looks for an iff/bool/comp with one operand of the compare 2877 // being a cycle involving an add and a phi, 2878 // with an optional truncation (left-shift followed by a right-shift) 2879 // of the add. Returns zero if not an iv. 2880 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) { 2881 Node* trunc1 = nullptr; 2882 Node* trunc2 = nullptr; 2883 const TypeInteger* ttype = nullptr; 2884 if (!iff->is_If() || iff->in(1) == nullptr || !iff->in(1)->is_Bool()) { 2885 return 0; 2886 } 2887 BoolNode* bl = iff->in(1)->as_Bool(); 2888 Node* cmp = bl->in(1); 2889 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) { 2890 return 0; 2891 } 2892 // Must have an invariant operand 2893 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) { 2894 return 0; 2895 } 2896 Node* add2 = nullptr; 2897 Node* cmp1 = cmp->in(1); 2898 if (cmp1->is_Phi()) { 2899 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) ))) 2900 Node* phi = cmp1; 2901 for (uint i = 1; i < phi->req(); i++) { 2902 Node* in = phi->in(i); 2903 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in, 2904 &trunc1, &trunc2, &ttype, T_INT); 2905 if (add && add->in(1) == phi) { 2906 add2 = add->in(2); 2907 break; 2908 } 2909 } 2910 } else { 2911 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) ))) 2912 Node* addtrunc = cmp1; 2913 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc, 2914 &trunc1, &trunc2, &ttype, T_INT); 2915 if (add && add->in(1)->is_Phi()) { 2916 Node* phi = add->in(1); 2917 for (uint i = 1; i < phi->req(); i++) { 2918 if (phi->in(i) == addtrunc) { 2919 add2 = add->in(2); 2920 break; 2921 } 2922 } 2923 } 2924 } 2925 if (add2 != nullptr) { 2926 const TypeInt* add2t = _igvn.type(add2)->is_int(); 2927 if (add2t->is_con()) { 2928 return add2t->get_con(); 2929 } 2930 } 2931 return 0; 2932 } 2933 2934 2935 //---------------------- stay_in_loop ------------------------------------- 2936 // Return the (unique) control output node that's in the loop (if it exists.) 2937 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) { 2938 Node* unique = nullptr; 2939 if (!n) return nullptr; 2940 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2941 Node* use = n->fast_out(i); 2942 if (!has_ctrl(use) && loop->is_member(get_loop(use))) { 2943 if (unique != nullptr) { 2944 return nullptr; 2945 } 2946 unique = use; 2947 } 2948 } 2949 return unique; 2950 } 2951 2952 //------------------------------ register_node ------------------------------------- 2953 // Utility to register node "n" with PhaseIdealLoop 2954 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree* loop, Node* pred, uint ddepth) { 2955 _igvn.register_new_node_with_optimizer(n); 2956 loop->_body.push(n); 2957 if (n->is_CFG()) { 2958 set_loop(n, loop); 2959 set_idom(n, pred, ddepth); 2960 } else { 2961 set_ctrl(n, pred); 2962 } 2963 } 2964 2965 //------------------------------ proj_clone ------------------------------------- 2966 // Utility to create an if-projection 2967 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) { 2968 ProjNode* c = p->clone()->as_Proj(); 2969 c->set_req(0, iff); 2970 return c; 2971 } 2972 2973 //------------------------------ short_circuit_if ------------------------------------- 2974 // Force the iff control output to be the live_proj 2975 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) { 2976 guarantee(live_proj != nullptr, "null projection"); 2977 int proj_con = live_proj->_con; 2978 assert(proj_con == 0 || proj_con == 1, "false or true projection"); 2979 Node *con = _igvn.intcon(proj_con); 2980 set_ctrl(con, C->root()); 2981 if (iff) { 2982 iff->set_req(1, con); 2983 } 2984 return con; 2985 } 2986 2987 //------------------------------ insert_if_before_proj ------------------------------------- 2988 // Insert a new if before an if projection (* - new node) 2989 // 2990 // before 2991 // if(test) 2992 // / \ 2993 // v v 2994 // other-proj proj (arg) 2995 // 2996 // after 2997 // if(test) 2998 // / \ 2999 // / v 3000 // | * proj-clone 3001 // v | 3002 // other-proj v 3003 // * new_if(relop(cmp[IU](left,right))) 3004 // / \ 3005 // v v 3006 // * new-proj proj 3007 // (returned) 3008 // 3009 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) { 3010 IfNode* iff = proj->in(0)->as_If(); 3011 IdealLoopTree *loop = get_loop(proj); 3012 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 3013 uint ddepth = dom_depth(proj); 3014 3015 _igvn.rehash_node_delayed(iff); 3016 _igvn.rehash_node_delayed(proj); 3017 3018 proj->set_req(0, nullptr); // temporary disconnect 3019 ProjNode* proj2 = proj_clone(proj, iff); 3020 register_node(proj2, loop, iff, ddepth); 3021 3022 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right); 3023 register_node(cmp, loop, proj2, ddepth); 3024 3025 BoolNode* bol = new BoolNode(cmp, relop); 3026 register_node(bol, loop, proj2, ddepth); 3027 3028 int opcode = iff->Opcode(); 3029 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode"); 3030 IfNode* new_if = (opcode == Op_If) ? new IfNode(proj2, bol, iff->_prob, iff->_fcnt): 3031 new RangeCheckNode(proj2, bol, iff->_prob, iff->_fcnt); 3032 register_node(new_if, loop, proj2, ddepth); 3033 3034 proj->set_req(0, new_if); // reattach 3035 set_idom(proj, new_if, ddepth); 3036 3037 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj(); 3038 guarantee(new_exit != nullptr, "null exit node"); 3039 register_node(new_exit, get_loop(other_proj), new_if, ddepth); 3040 3041 return new_exit; 3042 } 3043 3044 //------------------------------ insert_region_before_proj ------------------------------------- 3045 // Insert a region before an if projection (* - new node) 3046 // 3047 // before 3048 // if(test) 3049 // / | 3050 // v | 3051 // proj v 3052 // other-proj 3053 // 3054 // after 3055 // if(test) 3056 // / | 3057 // v | 3058 // * proj-clone v 3059 // | other-proj 3060 // v 3061 // * new-region 3062 // | 3063 // v 3064 // * dum_if 3065 // / \ 3066 // v \ 3067 // * dum-proj v 3068 // proj 3069 // 3070 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) { 3071 IfNode* iff = proj->in(0)->as_If(); 3072 IdealLoopTree *loop = get_loop(proj); 3073 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 3074 uint ddepth = dom_depth(proj); 3075 3076 _igvn.rehash_node_delayed(iff); 3077 _igvn.rehash_node_delayed(proj); 3078 3079 proj->set_req(0, nullptr); // temporary disconnect 3080 ProjNode* proj2 = proj_clone(proj, iff); 3081 register_node(proj2, loop, iff, ddepth); 3082 3083 RegionNode* reg = new RegionNode(2); 3084 reg->set_req(1, proj2); 3085 register_node(reg, loop, iff, ddepth); 3086 3087 IfNode* dum_if = new IfNode(reg, short_circuit_if(nullptr, proj), iff->_prob, iff->_fcnt); 3088 register_node(dum_if, loop, reg, ddepth); 3089 3090 proj->set_req(0, dum_if); // reattach 3091 set_idom(proj, dum_if, ddepth); 3092 3093 ProjNode* dum_proj = proj_clone(other_proj, dum_if); 3094 register_node(dum_proj, loop, dum_if, ddepth); 3095 3096 return reg; 3097 } 3098 3099 //------------------------------ insert_cmpi_loop_exit ------------------------------------- 3100 // Clone a signed compare loop exit from an unsigned compare and 3101 // insert it before the unsigned cmp on the stay-in-loop path. 3102 // All new nodes inserted in the dominator tree between the original 3103 // if and it's projections. The original if test is replaced with 3104 // a constant to force the stay-in-loop path. 3105 // 3106 // This is done to make sure that the original if and it's projections 3107 // still dominate the same set of control nodes, that the ctrl() relation 3108 // from data nodes to them is preserved, and that their loop nesting is 3109 // preserved. 3110 // 3111 // before 3112 // if(i <u limit) unsigned compare loop exit 3113 // / | 3114 // v v 3115 // exit-proj stay-in-loop-proj 3116 // 3117 // after 3118 // if(stay-in-loop-const) original if 3119 // / | 3120 // / v 3121 // / if(i < limit) new signed test 3122 // / / | 3123 // / / v 3124 // / / if(i <u limit) new cloned unsigned test 3125 // / / / | 3126 // v v v | 3127 // region | 3128 // | | 3129 // dum-if | 3130 // / | | 3131 // ether | | 3132 // v v 3133 // exit-proj stay-in-loop-proj 3134 // 3135 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) { 3136 const bool Signed = true; 3137 const bool Unsigned = false; 3138 3139 BoolNode* bol = if_cmpu->in(1)->as_Bool(); 3140 if (bol->_test._test != BoolTest::lt) return nullptr; 3141 CmpNode* cmpu = bol->in(1)->as_Cmp(); 3142 if (cmpu->Opcode() != Op_CmpU) return nullptr; 3143 int stride = stride_of_possible_iv(if_cmpu); 3144 if (stride == 0) return nullptr; 3145 3146 Node* lp_proj = stay_in_loop(if_cmpu, loop); 3147 guarantee(lp_proj != nullptr, "null loop node"); 3148 3149 ProjNode* lp_continue = lp_proj->as_Proj(); 3150 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj(); 3151 if (!lp_exit->is_IfFalse()) { 3152 // The loop exit condition is (i <u limit) ==> (i >= 0 && i < limit). 3153 // We therefore can't add a single exit condition. 3154 return nullptr; 3155 } 3156 // The loop exit condition is !(i <u limit) ==> (i < 0 || i >= limit). 3157 // Split out the exit condition (i < 0) for stride < 0 or (i >= limit) for stride > 0. 3158 Node* limit = nullptr; 3159 if (stride > 0) { 3160 limit = cmpu->in(2); 3161 } else { 3162 limit = _igvn.makecon(TypeInt::ZERO); 3163 set_ctrl(limit, C->root()); 3164 } 3165 // Create a new region on the exit path 3166 RegionNode* reg = insert_region_before_proj(lp_exit); 3167 guarantee(reg != nullptr, "null region node"); 3168 3169 // Clone the if-cmpu-true-false using a signed compare 3170 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge; 3171 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue); 3172 reg->add_req(cmpi_exit); 3173 3174 // Clone the if-cmpu-true-false 3175 BoolTest::mask rel_u = bol->_test._test; 3176 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue); 3177 reg->add_req(cmpu_exit); 3178 3179 // Force original if to stay in loop. 3180 short_circuit_if(if_cmpu, lp_continue); 3181 3182 return cmpi_exit->in(0)->as_If(); 3183 } 3184 3185 //------------------------------ remove_cmpi_loop_exit ------------------------------------- 3186 // Remove a previously inserted signed compare loop exit. 3187 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) { 3188 Node* lp_proj = stay_in_loop(if_cmp, loop); 3189 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI && 3190 stay_in_loop(lp_proj, loop)->is_If() && 3191 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu"); 3192 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO); 3193 set_ctrl(con, C->root()); 3194 if_cmp->set_req(1, con); 3195 } 3196 3197 //------------------------------ scheduled_nodelist ------------------------------------- 3198 // Create a post order schedule of nodes that are in the 3199 // "member" set. The list is returned in "sched". 3200 // The first node in "sched" is the loop head, followed by 3201 // nodes which have no inputs in the "member" set, and then 3202 // followed by the nodes that have an immediate input dependence 3203 // on a node in "sched". 3204 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) { 3205 3206 assert(member.test(loop->_head->_idx), "loop head must be in member set"); 3207 VectorSet visited; 3208 Node_Stack nstack(loop->_body.size()); 3209 3210 Node* n = loop->_head; // top of stack is cached in "n" 3211 uint idx = 0; 3212 visited.set(n->_idx); 3213 3214 // Initially push all with no inputs from within member set 3215 for(uint i = 0; i < loop->_body.size(); i++ ) { 3216 Node *elt = loop->_body.at(i); 3217 if (member.test(elt->_idx)) { 3218 bool found = false; 3219 for (uint j = 0; j < elt->req(); j++) { 3220 Node* def = elt->in(j); 3221 if (def && member.test(def->_idx) && def != elt) { 3222 found = true; 3223 break; 3224 } 3225 } 3226 if (!found && elt != loop->_head) { 3227 nstack.push(n, idx); 3228 n = elt; 3229 assert(!visited.test(n->_idx), "not seen yet"); 3230 visited.set(n->_idx); 3231 } 3232 } 3233 } 3234 3235 // traverse out's that are in the member set 3236 while (true) { 3237 if (idx < n->outcnt()) { 3238 Node* use = n->raw_out(idx); 3239 idx++; 3240 if (!visited.test_set(use->_idx)) { 3241 if (member.test(use->_idx)) { 3242 nstack.push(n, idx); 3243 n = use; 3244 idx = 0; 3245 } 3246 } 3247 } else { 3248 // All outputs processed 3249 sched.push(n); 3250 if (nstack.is_empty()) break; 3251 n = nstack.node(); 3252 idx = nstack.index(); 3253 nstack.pop(); 3254 } 3255 } 3256 } 3257 3258 3259 //------------------------------ has_use_in_set ------------------------------------- 3260 // Has a use in the vector set 3261 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) { 3262 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3263 Node* use = n->fast_out(j); 3264 if (vset.test(use->_idx)) { 3265 return true; 3266 } 3267 } 3268 return false; 3269 } 3270 3271 3272 //------------------------------ has_use_internal_to_set ------------------------------------- 3273 // Has use internal to the vector set (ie. not in a phi at the loop head) 3274 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) { 3275 Node* head = loop->_head; 3276 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3277 Node* use = n->fast_out(j); 3278 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) { 3279 return true; 3280 } 3281 } 3282 return false; 3283 } 3284 3285 3286 //------------------------------ clone_for_use_outside_loop ------------------------------------- 3287 // clone "n" for uses that are outside of loop 3288 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) { 3289 int cloned = 0; 3290 assert(worklist.size() == 0, "should be empty"); 3291 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3292 Node* use = n->fast_out(j); 3293 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) { 3294 worklist.push(use); 3295 } 3296 } 3297 3298 if (C->check_node_count(worklist.size() + NodeLimitFudgeFactor, 3299 "Too many clones required in clone_for_use_outside_loop in partial peeling")) { 3300 return -1; 3301 } 3302 3303 while( worklist.size() ) { 3304 Node *use = worklist.pop(); 3305 if (!has_node(use) || use->in(0) == C->top()) continue; 3306 uint j; 3307 for (j = 0; j < use->req(); j++) { 3308 if (use->in(j) == n) break; 3309 } 3310 assert(j < use->req(), "must be there"); 3311 3312 // clone "n" and insert it between the inputs of "n" and the use outside the loop 3313 Node* n_clone = n->clone(); 3314 _igvn.replace_input_of(use, j, n_clone); 3315 cloned++; 3316 Node* use_c; 3317 if (!use->is_Phi()) { 3318 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0); 3319 } else { 3320 // Use in a phi is considered a use in the associated predecessor block 3321 use_c = use->in(0)->in(j); 3322 } 3323 set_ctrl(n_clone, use_c); 3324 assert(!loop->is_member(get_loop(use_c)), "should be outside loop"); 3325 get_loop(use_c)->_body.push(n_clone); 3326 _igvn.register_new_node_with_optimizer(n_clone); 3327 #ifndef PRODUCT 3328 if (TracePartialPeeling) { 3329 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx); 3330 } 3331 #endif 3332 } 3333 return cloned; 3334 } 3335 3336 3337 //------------------------------ clone_for_special_use_inside_loop ------------------------------------- 3338 // clone "n" for special uses that are in the not_peeled region. 3339 // If these def-uses occur in separate blocks, the code generator 3340 // marks the method as not compilable. For example, if a "BoolNode" 3341 // is in a different basic block than the "IfNode" that uses it, then 3342 // the compilation is aborted in the code generator. 3343 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 3344 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) { 3345 if (n->is_Phi() || n->is_Load()) { 3346 return; 3347 } 3348 assert(worklist.size() == 0, "should be empty"); 3349 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3350 Node* use = n->fast_out(j); 3351 if ( not_peel.test(use->_idx) && 3352 (use->is_If() || use->is_CMove() || use->is_Bool()) && 3353 use->in(1) == n) { 3354 worklist.push(use); 3355 } 3356 } 3357 if (worklist.size() > 0) { 3358 // clone "n" and insert it between inputs of "n" and the use 3359 Node* n_clone = n->clone(); 3360 loop->_body.push(n_clone); 3361 _igvn.register_new_node_with_optimizer(n_clone); 3362 set_ctrl(n_clone, get_ctrl(n)); 3363 sink_list.push(n_clone); 3364 not_peel.set(n_clone->_idx); 3365 #ifndef PRODUCT 3366 if (TracePartialPeeling) { 3367 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx); 3368 } 3369 #endif 3370 while( worklist.size() ) { 3371 Node *use = worklist.pop(); 3372 _igvn.rehash_node_delayed(use); 3373 for (uint j = 1; j < use->req(); j++) { 3374 if (use->in(j) == n) { 3375 use->set_req(j, n_clone); 3376 } 3377 } 3378 } 3379 } 3380 } 3381 3382 3383 //------------------------------ insert_phi_for_loop ------------------------------------- 3384 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 3385 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) { 3386 Node *phi = PhiNode::make(lp, back_edge_val); 3387 phi->set_req(LoopNode::EntryControl, lp_entry_val); 3388 // Use existing phi if it already exists 3389 Node *hit = _igvn.hash_find_insert(phi); 3390 if( hit == nullptr ) { 3391 _igvn.register_new_node_with_optimizer(phi); 3392 set_ctrl(phi, lp); 3393 } else { 3394 // Remove the new phi from the graph and use the hit 3395 _igvn.remove_dead_node(phi); 3396 phi = hit; 3397 } 3398 _igvn.replace_input_of(use, idx, phi); 3399 } 3400 3401 #ifdef ASSERT 3402 //------------------------------ is_valid_loop_partition ------------------------------------- 3403 // Validate the loop partition sets: peel and not_peel 3404 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, 3405 VectorSet& not_peel ) { 3406 uint i; 3407 // Check that peel_list entries are in the peel set 3408 for (i = 0; i < peel_list.size(); i++) { 3409 if (!peel.test(peel_list.at(i)->_idx)) { 3410 return false; 3411 } 3412 } 3413 // Check at loop members are in one of peel set or not_peel set 3414 for (i = 0; i < loop->_body.size(); i++ ) { 3415 Node *def = loop->_body.at(i); 3416 uint di = def->_idx; 3417 // Check that peel set elements are in peel_list 3418 if (peel.test(di)) { 3419 if (not_peel.test(di)) { 3420 return false; 3421 } 3422 // Must be in peel_list also 3423 bool found = false; 3424 for (uint j = 0; j < peel_list.size(); j++) { 3425 if (peel_list.at(j)->_idx == di) { 3426 found = true; 3427 break; 3428 } 3429 } 3430 if (!found) { 3431 return false; 3432 } 3433 } else if (not_peel.test(di)) { 3434 if (peel.test(di)) { 3435 return false; 3436 } 3437 } else { 3438 return false; 3439 } 3440 } 3441 return true; 3442 } 3443 3444 //------------------------------ is_valid_clone_loop_exit_use ------------------------------------- 3445 // Ensure a use outside of loop is of the right form 3446 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) { 3447 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 3448 return (use->is_Phi() && 3449 use_c->is_Region() && use_c->req() == 3 && 3450 (use_c->in(exit_idx)->Opcode() == Op_IfTrue || 3451 use_c->in(exit_idx)->Opcode() == Op_IfFalse || 3452 use_c->in(exit_idx)->Opcode() == Op_JumpProj) && 3453 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) ); 3454 } 3455 3456 //------------------------------ is_valid_clone_loop_form ------------------------------------- 3457 // Ensure that all uses outside of loop are of the right form 3458 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 3459 uint orig_exit_idx, uint clone_exit_idx) { 3460 uint len = peel_list.size(); 3461 for (uint i = 0; i < len; i++) { 3462 Node *def = peel_list.at(i); 3463 3464 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 3465 Node *use = def->fast_out(j); 3466 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 3467 if (!loop->is_member(get_loop(use_c))) { 3468 // use is not in the loop, check for correct structure 3469 if (use->in(0) == def) { 3470 // Okay 3471 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) { 3472 return false; 3473 } 3474 } 3475 } 3476 } 3477 return true; 3478 } 3479 #endif 3480 3481 //------------------------------ partial_peel ------------------------------------- 3482 // Partially peel (aka loop rotation) the top portion of a loop (called 3483 // the peel section below) by cloning it and placing one copy just before 3484 // the new loop head and the other copy at the bottom of the new loop. 3485 // 3486 // before after where it came from 3487 // 3488 // stmt1 stmt1 3489 // loop: stmt2 clone 3490 // stmt2 if condA goto exitA clone 3491 // if condA goto exitA new_loop: new 3492 // stmt3 stmt3 clone 3493 // if !condB goto loop if condB goto exitB clone 3494 // exitB: stmt2 orig 3495 // stmt4 if !condA goto new_loop orig 3496 // exitA: goto exitA 3497 // exitB: 3498 // stmt4 3499 // exitA: 3500 // 3501 // Step 1: find the cut point: an exit test on probable 3502 // induction variable. 3503 // Step 2: schedule (with cloning) operations in the peel 3504 // section that can be executed after the cut into 3505 // the section that is not peeled. This may need 3506 // to clone operations into exit blocks. For 3507 // instance, a reference to A[i] in the not-peel 3508 // section and a reference to B[i] in an exit block 3509 // may cause a left-shift of i by 2 to be placed 3510 // in the peel block. This step will clone the left 3511 // shift into the exit block and sink the left shift 3512 // from the peel to the not-peel section. 3513 // Step 3: clone the loop, retarget the control, and insert 3514 // phis for values that are live across the new loop 3515 // head. This is very dependent on the graph structure 3516 // from clone_loop. It creates region nodes for 3517 // exit control and associated phi nodes for values 3518 // flow out of the loop through that exit. The region 3519 // node is dominated by the clone's control projection. 3520 // So the clone's peel section is placed before the 3521 // new loop head, and the clone's not-peel section is 3522 // forms the top part of the new loop. The original 3523 // peel section forms the tail of the new loop. 3524 // Step 4: update the dominator tree and recompute the 3525 // dominator depth. 3526 // 3527 // orig 3528 // 3529 // stmt1 3530 // | 3531 // v 3532 // predicates 3533 // | 3534 // v 3535 // loop<----+ 3536 // | | 3537 // stmt2 | 3538 // | | 3539 // v | 3540 // ifA | 3541 // / | | 3542 // v v | 3543 // false true ^ <-- last_peel 3544 // / | | 3545 // / ===|==cut | 3546 // / stmt3 | <-- first_not_peel 3547 // / | | 3548 // | v | 3549 // v ifB | 3550 // exitA: / \ | 3551 // / \ | 3552 // v v | 3553 // false true | 3554 // / \ | 3555 // / ----+ 3556 // | 3557 // v 3558 // exitB: 3559 // stmt4 3560 // 3561 // 3562 // after clone loop 3563 // 3564 // stmt1 3565 // | 3566 // v 3567 // predicates 3568 // / \ 3569 // clone / \ orig 3570 // / \ 3571 // / \ 3572 // v v 3573 // +---->loop loop<----+ 3574 // | | | | 3575 // | stmt2 stmt2 | 3576 // | | | | 3577 // | v v | 3578 // | ifA ifA | 3579 // | | \ / | | 3580 // | v v v v | 3581 // ^ true false false true ^ <-- last_peel 3582 // | | ^ \ / | | 3583 // | cut==|== \ \ / ===|==cut | 3584 // | stmt3 \ \ / stmt3 | <-- first_not_peel 3585 // | | dom | | | | 3586 // | v \ 1v v2 v | 3587 // | ifB regionA ifB | 3588 // | / \ | / \ | 3589 // | / \ v / \ | 3590 // | v v exitA: v v | 3591 // | true false false true | 3592 // | / ^ \ / \ | 3593 // +---- \ \ / ----+ 3594 // dom \ / 3595 // \ 1v v2 3596 // regionB 3597 // | 3598 // v 3599 // exitB: 3600 // stmt4 3601 // 3602 // 3603 // after partial peel 3604 // 3605 // stmt1 3606 // | 3607 // v 3608 // predicates 3609 // / 3610 // clone / orig 3611 // / TOP 3612 // / \ 3613 // v v 3614 // TOP->loop loop----+ 3615 // | | | 3616 // stmt2 stmt2 | 3617 // | | | 3618 // v v | 3619 // ifA ifA | 3620 // | \ / | | 3621 // v v v v | 3622 // true false false true | <-- last_peel 3623 // | ^ \ / +------|---+ 3624 // +->newloop \ \ / === ==cut | | 3625 // | stmt3 \ \ / TOP | | 3626 // | | dom | | stmt3 | | <-- first_not_peel 3627 // | v \ 1v v2 v | | 3628 // | ifB regionA ifB ^ v 3629 // | / \ | / \ | | 3630 // | / \ v / \ | | 3631 // | v v exitA: v v | | 3632 // | true false false true | | 3633 // | / ^ \ / \ | | 3634 // | | \ \ / v | | 3635 // | | dom \ / TOP | | 3636 // | | \ 1v v2 | | 3637 // ^ v regionB | | 3638 // | | | | | 3639 // | | v ^ v 3640 // | | exitB: | | 3641 // | | stmt4 | | 3642 // | +------------>-----------------+ | 3643 // | | 3644 // +-----------------<---------------------+ 3645 // 3646 // 3647 // final graph 3648 // 3649 // stmt1 3650 // | 3651 // v 3652 // predicates 3653 // | 3654 // v 3655 // stmt2 clone 3656 // | 3657 // v 3658 // ........> ifA clone 3659 // : / | 3660 // dom / | 3661 // : v v 3662 // : false true 3663 // : | | 3664 // : | v 3665 // : | newloop<-----+ 3666 // : | | | 3667 // : | stmt3 clone | 3668 // : | | | 3669 // : | v | 3670 // : | ifB | 3671 // : | / \ | 3672 // : | v v | 3673 // : | false true | 3674 // : | | | | 3675 // : | v stmt2 | 3676 // : | exitB: | | 3677 // : | stmt4 v | 3678 // : | ifA orig | 3679 // : | / \ | 3680 // : | / \ | 3681 // : | v v | 3682 // : | false true | 3683 // : | / \ | 3684 // : v v -----+ 3685 // RegionA 3686 // | 3687 // v 3688 // exitA 3689 // 3690 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) { 3691 3692 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only"); 3693 if (!loop->_head->is_Loop()) { 3694 return false; 3695 } 3696 LoopNode *head = loop->_head->as_Loop(); 3697 3698 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) { 3699 return false; 3700 } 3701 3702 // Check for complex exit control 3703 for (uint ii = 0; ii < loop->_body.size(); ii++) { 3704 Node *n = loop->_body.at(ii); 3705 int opc = n->Opcode(); 3706 if (n->is_Call() || 3707 opc == Op_Catch || 3708 opc == Op_CatchProj || 3709 opc == Op_Jump || 3710 opc == Op_JumpProj) { 3711 #ifndef PRODUCT 3712 if (TracePartialPeeling) { 3713 tty->print_cr("\nExit control too complex: lp: %d", head->_idx); 3714 } 3715 #endif 3716 return false; 3717 } 3718 } 3719 3720 int dd = dom_depth(head); 3721 3722 // Step 1: find cut point 3723 3724 // Walk up dominators to loop head looking for first loop exit 3725 // which is executed on every path thru loop. 3726 IfNode *peel_if = nullptr; 3727 IfNode *peel_if_cmpu = nullptr; 3728 3729 Node *iff = loop->tail(); 3730 while (iff != head) { 3731 if (iff->is_If()) { 3732 Node *ctrl = get_ctrl(iff->in(1)); 3733 if (ctrl->is_top()) return false; // Dead test on live IF. 3734 // If loop-varying exit-test, check for induction variable 3735 if (loop->is_member(get_loop(ctrl)) && 3736 loop->is_loop_exit(iff) && 3737 is_possible_iv_test(iff)) { 3738 Node* cmp = iff->in(1)->in(1); 3739 if (cmp->Opcode() == Op_CmpI) { 3740 peel_if = iff->as_If(); 3741 } else { 3742 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU"); 3743 peel_if_cmpu = iff->as_If(); 3744 } 3745 } 3746 } 3747 iff = idom(iff); 3748 } 3749 3750 // Prefer signed compare over unsigned compare. 3751 IfNode* new_peel_if = nullptr; 3752 if (peel_if == nullptr) { 3753 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == nullptr) { 3754 return false; // No peel point found 3755 } 3756 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop); 3757 if (new_peel_if == nullptr) { 3758 return false; // No peel point found 3759 } 3760 peel_if = new_peel_if; 3761 } 3762 Node* last_peel = stay_in_loop(peel_if, loop); 3763 Node* first_not_peeled = stay_in_loop(last_peel, loop); 3764 if (first_not_peeled == nullptr || first_not_peeled == head) { 3765 return false; 3766 } 3767 3768 #ifndef PRODUCT 3769 if (TraceLoopOpts) { 3770 tty->print("PartialPeel "); 3771 loop->dump_head(); 3772 } 3773 3774 if (TracePartialPeeling) { 3775 tty->print_cr("before partial peel one iteration"); 3776 Node_List wl; 3777 Node* t = head->in(2); 3778 while (true) { 3779 wl.push(t); 3780 if (t == head) break; 3781 t = idom(t); 3782 } 3783 while (wl.size() > 0) { 3784 Node* tt = wl.pop(); 3785 tt->dump(); 3786 if (tt == last_peel) tty->print_cr("-- cut --"); 3787 } 3788 } 3789 #endif 3790 VectorSet peel; 3791 VectorSet not_peel; 3792 Node_List peel_list; 3793 Node_List worklist; 3794 Node_List sink_list; 3795 3796 uint estimate = loop->est_loop_clone_sz(1); 3797 if (exceeding_node_budget(estimate)) { 3798 return false; 3799 } 3800 3801 // Set of cfg nodes to peel are those that are executable from 3802 // the head through last_peel. 3803 assert(worklist.size() == 0, "should be empty"); 3804 worklist.push(head); 3805 peel.set(head->_idx); 3806 while (worklist.size() > 0) { 3807 Node *n = worklist.pop(); 3808 if (n != last_peel) { 3809 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3810 Node* use = n->fast_out(j); 3811 if (use->is_CFG() && 3812 loop->is_member(get_loop(use)) && 3813 !peel.test_set(use->_idx)) { 3814 worklist.push(use); 3815 } 3816 } 3817 } 3818 } 3819 3820 // Set of non-cfg nodes to peel are those that are control 3821 // dependent on the cfg nodes. 3822 for (uint i = 0; i < loop->_body.size(); i++) { 3823 Node *n = loop->_body.at(i); 3824 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n; 3825 if (peel.test(n_c->_idx)) { 3826 peel.set(n->_idx); 3827 } else { 3828 not_peel.set(n->_idx); 3829 } 3830 } 3831 3832 // Step 2: move operations from the peeled section down into the 3833 // not-peeled section 3834 3835 // Get a post order schedule of nodes in the peel region 3836 // Result in right-most operand. 3837 scheduled_nodelist(loop, peel, peel_list); 3838 3839 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 3840 3841 // For future check for too many new phis 3842 uint old_phi_cnt = 0; 3843 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 3844 Node* use = head->fast_out(j); 3845 if (use->is_Phi()) old_phi_cnt++; 3846 } 3847 3848 #ifndef PRODUCT 3849 if (TracePartialPeeling) { 3850 tty->print_cr("\npeeled list"); 3851 } 3852 #endif 3853 3854 // Evacuate nodes in peel region into the not_peeled region if possible 3855 bool too_many_clones = false; 3856 uint new_phi_cnt = 0; 3857 uint cloned_for_outside_use = 0; 3858 for (uint i = 0; i < peel_list.size();) { 3859 Node* n = peel_list.at(i); 3860 #ifndef PRODUCT 3861 if (TracePartialPeeling) n->dump(); 3862 #endif 3863 bool incr = true; 3864 if (!n->is_CFG()) { 3865 if (has_use_in_set(n, not_peel)) { 3866 // If not used internal to the peeled region, 3867 // move "n" from peeled to not_peeled region. 3868 if (!has_use_internal_to_set(n, peel, loop)) { 3869 // if not pinned and not a load (which maybe anti-dependent on a store) 3870 // and not a CMove (Matcher expects only bool->cmove). 3871 if (n->in(0) == nullptr && !n->is_Load() && !n->is_CMove()) { 3872 int new_clones = clone_for_use_outside_loop(loop, n, worklist); 3873 if (new_clones == -1) { 3874 too_many_clones = true; 3875 break; 3876 } 3877 cloned_for_outside_use += new_clones; 3878 sink_list.push(n); 3879 peel.remove(n->_idx); 3880 not_peel.set(n->_idx); 3881 peel_list.remove(i); 3882 incr = false; 3883 #ifndef PRODUCT 3884 if (TracePartialPeeling) { 3885 tty->print_cr("sink to not_peeled region: %d newbb: %d", 3886 n->_idx, get_ctrl(n)->_idx); 3887 } 3888 #endif 3889 } 3890 } else { 3891 // Otherwise check for special def-use cases that span 3892 // the peel/not_peel boundary such as bool->if 3893 clone_for_special_use_inside_loop(loop, n, not_peel, sink_list, worklist); 3894 new_phi_cnt++; 3895 } 3896 } 3897 } 3898 if (incr) i++; 3899 } 3900 3901 estimate += cloned_for_outside_use + new_phi_cnt; 3902 bool exceed_node_budget = !may_require_nodes(estimate); 3903 bool exceed_phi_limit = new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta; 3904 3905 if (too_many_clones || exceed_node_budget || exceed_phi_limit) { 3906 #ifndef PRODUCT 3907 if (TracePartialPeeling && exceed_phi_limit) { 3908 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c", 3909 new_phi_cnt, old_phi_cnt, new_peel_if != nullptr?'T':'F'); 3910 } 3911 #endif 3912 if (new_peel_if != nullptr) { 3913 remove_cmpi_loop_exit(new_peel_if, loop); 3914 } 3915 // Inhibit more partial peeling on this loop 3916 assert(!head->is_partial_peel_loop(), "not partial peeled"); 3917 head->mark_partial_peel_failed(); 3918 if (cloned_for_outside_use > 0) { 3919 // Terminate this round of loop opts because 3920 // the graph outside this loop was changed. 3921 C->set_major_progress(); 3922 return true; 3923 } 3924 return false; 3925 } 3926 3927 // Step 3: clone loop, retarget control, and insert new phis 3928 3929 // Create new loop head for new phis and to hang 3930 // the nodes being moved (sinked) from the peel region. 3931 LoopNode* new_head = new LoopNode(last_peel, last_peel); 3932 new_head->set_unswitch_count(head->unswitch_count()); // Preserve 3933 _igvn.register_new_node_with_optimizer(new_head); 3934 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled"); 3935 _igvn.replace_input_of(first_not_peeled, 0, new_head); 3936 set_loop(new_head, loop); 3937 loop->_body.push(new_head); 3938 not_peel.set(new_head->_idx); 3939 set_idom(new_head, last_peel, dom_depth(first_not_peeled)); 3940 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled)); 3941 3942 while (sink_list.size() > 0) { 3943 Node* n = sink_list.pop(); 3944 set_ctrl(n, new_head); 3945 } 3946 3947 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 3948 3949 clone_loop(loop, old_new, dd, IgnoreStripMined); 3950 3951 const uint clone_exit_idx = 1; 3952 const uint orig_exit_idx = 2; 3953 assert(is_valid_clone_loop_form(loop, peel_list, orig_exit_idx, clone_exit_idx), "bad clone loop"); 3954 3955 Node* head_clone = old_new[head->_idx]; 3956 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop(); 3957 Node* orig_tail_clone = head_clone->in(2); 3958 3959 // Add phi if "def" node is in peel set and "use" is not 3960 3961 for (uint i = 0; i < peel_list.size(); i++) { 3962 Node *def = peel_list.at(i); 3963 if (!def->is_CFG()) { 3964 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 3965 Node *use = def->fast_out(j); 3966 if (has_node(use) && use->in(0) != C->top() && 3967 (!peel.test(use->_idx) || 3968 (use->is_Phi() && use->in(0) == head)) ) { 3969 worklist.push(use); 3970 } 3971 } 3972 while( worklist.size() ) { 3973 Node *use = worklist.pop(); 3974 for (uint j = 1; j < use->req(); j++) { 3975 Node* n = use->in(j); 3976 if (n == def) { 3977 3978 // "def" is in peel set, "use" is not in peel set 3979 // or "use" is in the entry boundary (a phi) of the peel set 3980 3981 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use; 3982 3983 if ( loop->is_member(get_loop( use_c )) ) { 3984 // use is in loop 3985 if (old_new[use->_idx] != nullptr) { // null for dead code 3986 Node* use_clone = old_new[use->_idx]; 3987 _igvn.replace_input_of(use, j, C->top()); 3988 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone ); 3989 } 3990 } else { 3991 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format"); 3992 // use is not in the loop, check if the live range includes the cut 3993 Node* lp_if = use_c->in(orig_exit_idx)->in(0); 3994 if (not_peel.test(lp_if->_idx)) { 3995 assert(j == orig_exit_idx, "use from original loop"); 3996 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone ); 3997 } 3998 } 3999 } 4000 } 4001 } 4002 } 4003 } 4004 4005 // Step 3b: retarget control 4006 4007 // Redirect control to the new loop head if a cloned node in 4008 // the not_peeled region has control that points into the peeled region. 4009 // This necessary because the cloned peeled region will be outside 4010 // the loop. 4011 // from to 4012 // cloned-peeled <---+ 4013 // new_head_clone: | <--+ 4014 // cloned-not_peeled in(0) in(0) 4015 // orig-peeled 4016 4017 for (uint i = 0; i < loop->_body.size(); i++) { 4018 Node *n = loop->_body.at(i); 4019 if (!n->is_CFG() && n->in(0) != nullptr && 4020 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) { 4021 Node* n_clone = old_new[n->_idx]; 4022 _igvn.replace_input_of(n_clone, 0, new_head_clone); 4023 } 4024 } 4025 4026 // Backedge of the surviving new_head (the clone) is original last_peel 4027 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel); 4028 4029 // Cut first node in original not_peel set 4030 _igvn.rehash_node_delayed(new_head); // Multiple edge updates: 4031 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of 4032 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls 4033 4034 // Copy head_clone back-branch info to original head 4035 // and remove original head's loop entry and 4036 // clone head's back-branch 4037 _igvn.rehash_node_delayed(head); // Multiple edge updates 4038 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl)); 4039 head->set_req(LoopNode::LoopBackControl, C->top()); 4040 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top()); 4041 4042 // Similarly modify the phis 4043 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) { 4044 Node* use = head->fast_out(k); 4045 if (use->is_Phi() && use->outcnt() > 0) { 4046 Node* use_clone = old_new[use->_idx]; 4047 _igvn.rehash_node_delayed(use); // Multiple edge updates 4048 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl)); 4049 use->set_req(LoopNode::LoopBackControl, C->top()); 4050 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top()); 4051 } 4052 } 4053 4054 // Step 4: update dominator tree and dominator depth 4055 4056 set_idom(head, orig_tail_clone, dd); 4057 recompute_dom_depth(); 4058 4059 // Inhibit more partial peeling on this loop 4060 new_head_clone->set_partial_peel_loop(); 4061 C->set_major_progress(); 4062 loop->record_for_igvn(); 4063 4064 #ifndef PRODUCT 4065 if (TracePartialPeeling) { 4066 tty->print_cr("\nafter partial peel one iteration"); 4067 Node_List wl; 4068 Node* t = last_peel; 4069 while (true) { 4070 wl.push(t); 4071 if (t == head_clone) break; 4072 t = idom(t); 4073 } 4074 while (wl.size() > 0) { 4075 Node* tt = wl.pop(); 4076 if (tt == head) tty->print_cr("orig head"); 4077 else if (tt == new_head_clone) tty->print_cr("new head"); 4078 else if (tt == head_clone) tty->print_cr("clone head"); 4079 tt->dump(); 4080 } 4081 } 4082 #endif 4083 return true; 4084 } 4085 4086 // Transform: 4087 // 4088 // loop<-----------------+ 4089 // | | 4090 // stmt1 stmt2 .. stmtn | 4091 // | | | | 4092 // \ | / | 4093 // v v v | 4094 // region | 4095 // | | 4096 // shared_stmt | 4097 // | | 4098 // v | 4099 // if | 4100 // / \ | 4101 // | -----------+ 4102 // v 4103 // 4104 // into: 4105 // 4106 // loop<-------------------+ 4107 // | | 4108 // v | 4109 // +->loop | 4110 // | | | 4111 // | stmt1 stmt2 .. stmtn | 4112 // | | | | | 4113 // | | \ / | 4114 // | | v v | 4115 // | | region1 | 4116 // | | | | 4117 // | shared_stmt shared_stmt | 4118 // | | | | 4119 // | v v | 4120 // | if if | 4121 // | /\ / \ | 4122 // +-- | | -------+ 4123 // \ / 4124 // v v 4125 // region2 4126 // 4127 // (region2 is shown to merge mirrored projections of the loop exit 4128 // ifs to make the diagram clearer but they really merge the same 4129 // projection) 4130 // 4131 // Conditions for this transformation to trigger: 4132 // - the path through stmt1 is frequent enough 4133 // - the inner loop will be turned into a counted loop after transformation 4134 bool PhaseIdealLoop::duplicate_loop_backedge(IdealLoopTree *loop, Node_List &old_new) { 4135 if (!DuplicateBackedge) { 4136 return false; 4137 } 4138 assert(!loop->_head->is_CountedLoop() || StressDuplicateBackedge, "Non-counted loop only"); 4139 if (!loop->_head->is_Loop()) { 4140 return false; 4141 } 4142 4143 uint estimate = loop->est_loop_clone_sz(1); 4144 if (exceeding_node_budget(estimate)) { 4145 return false; 4146 } 4147 4148 LoopNode *head = loop->_head->as_Loop(); 4149 4150 Node* region = nullptr; 4151 IfNode* exit_test = nullptr; 4152 uint inner; 4153 float f; 4154 if (StressDuplicateBackedge) { 4155 if (head->is_strip_mined()) { 4156 return false; 4157 } 4158 Node* c = head->in(LoopNode::LoopBackControl); 4159 4160 while (c != head) { 4161 if (c->is_Region()) { 4162 region = c; 4163 } 4164 c = idom(c); 4165 } 4166 4167 if (region == nullptr) { 4168 return false; 4169 } 4170 4171 inner = 1; 4172 } else { 4173 // Is the shape of the loop that of a counted loop... 4174 Node* back_control = loop_exit_control(head, loop); 4175 if (back_control == nullptr) { 4176 return false; 4177 } 4178 4179 BoolTest::mask bt = BoolTest::illegal; 4180 float cl_prob = 0; 4181 Node* incr = nullptr; 4182 Node* limit = nullptr; 4183 Node* cmp = loop_exit_test(back_control, loop, incr, limit, bt, cl_prob); 4184 if (cmp == nullptr || cmp->Opcode() != Op_CmpI) { 4185 return false; 4186 } 4187 4188 // With an extra phi for the candidate iv? 4189 // Or the region node is the loop head 4190 if (!incr->is_Phi() || incr->in(0) == head) { 4191 return false; 4192 } 4193 4194 PathFrequency pf(head, this); 4195 region = incr->in(0); 4196 4197 // Go over all paths for the extra phi's region and see if that 4198 // path is frequent enough and would match the expected iv shape 4199 // if the extra phi is removed 4200 inner = 0; 4201 for (uint i = 1; i < incr->req(); ++i) { 4202 Node* in = incr->in(i); 4203 Node* trunc1 = nullptr; 4204 Node* trunc2 = nullptr; 4205 const TypeInteger* iv_trunc_t = nullptr; 4206 Node* orig_in = in; 4207 if (!(in = CountedLoopNode::match_incr_with_optional_truncation(in, &trunc1, &trunc2, &iv_trunc_t, T_INT))) { 4208 continue; 4209 } 4210 assert(in->Opcode() == Op_AddI, "wrong increment code"); 4211 Node* xphi = nullptr; 4212 Node* stride = loop_iv_stride(in, loop, xphi); 4213 4214 if (stride == nullptr) { 4215 continue; 4216 } 4217 4218 PhiNode* phi = loop_iv_phi(xphi, nullptr, head, loop); 4219 if (phi == nullptr || 4220 (trunc1 == nullptr && phi->in(LoopNode::LoopBackControl) != incr) || 4221 (trunc1 != nullptr && phi->in(LoopNode::LoopBackControl) != trunc1)) { 4222 return false; 4223 } 4224 4225 f = pf.to(region->in(i)); 4226 if (f > 0.5) { 4227 inner = i; 4228 break; 4229 } 4230 } 4231 4232 if (inner == 0) { 4233 return false; 4234 } 4235 4236 exit_test = back_control->in(0)->as_If(); 4237 } 4238 4239 if (idom(region)->is_Catch()) { 4240 return false; 4241 } 4242 4243 // Collect all control nodes that need to be cloned (shared_stmt in the diagram) 4244 Unique_Node_List wq; 4245 wq.push(head->in(LoopNode::LoopBackControl)); 4246 for (uint i = 0; i < wq.size(); i++) { 4247 Node* c = wq.at(i); 4248 assert(get_loop(c) == loop, "not in the right loop?"); 4249 if (c->is_Region()) { 4250 if (c != region) { 4251 for (uint j = 1; j < c->req(); ++j) { 4252 wq.push(c->in(j)); 4253 } 4254 } 4255 } else { 4256 wq.push(c->in(0)); 4257 } 4258 assert(!is_dominator(c, region) || c == region, "shouldn't go above region"); 4259 } 4260 4261 Node* region_dom = idom(region); 4262 4263 // Can't do the transformation if this would cause a membar pair to 4264 // be split 4265 for (uint i = 0; i < wq.size(); i++) { 4266 Node* c = wq.at(i); 4267 if (c->is_MemBar() && (c->as_MemBar()->trailing_store() || c->as_MemBar()->trailing_load_store())) { 4268 assert(c->as_MemBar()->leading_membar()->trailing_membar() == c, "bad membar pair"); 4269 if (!wq.member(c->as_MemBar()->leading_membar())) { 4270 return false; 4271 } 4272 } 4273 } 4274 4275 // Collect data nodes that need to be clones as well 4276 int dd = dom_depth(head); 4277 4278 for (uint i = 0; i < loop->_body.size(); ++i) { 4279 Node* n = loop->_body.at(i); 4280 if (has_ctrl(n)) { 4281 Node* c = get_ctrl(n); 4282 if (wq.member(c)) { 4283 wq.push(n); 4284 } 4285 } else { 4286 set_idom(n, idom(n), dd); 4287 } 4288 } 4289 4290 // clone shared_stmt 4291 clone_loop_body(wq, old_new, nullptr); 4292 4293 Node* region_clone = old_new[region->_idx]; 4294 region_clone->set_req(inner, C->top()); 4295 set_idom(region, region->in(inner), dd); 4296 4297 // Prepare the outer loop 4298 Node* outer_head = new LoopNode(head->in(LoopNode::EntryControl), old_new[head->in(LoopNode::LoopBackControl)->_idx]); 4299 register_control(outer_head, loop->_parent, outer_head->in(LoopNode::EntryControl)); 4300 _igvn.replace_input_of(head, LoopNode::EntryControl, outer_head); 4301 set_idom(head, outer_head, dd); 4302 4303 fix_body_edges(wq, loop, old_new, dd, loop->_parent, true); 4304 4305 // Make one of the shared_stmt copies only reachable from stmt1, the 4306 // other only from stmt2..stmtn. 4307 Node* dom = nullptr; 4308 for (uint i = 1; i < region->req(); ++i) { 4309 if (i != inner) { 4310 _igvn.replace_input_of(region, i, C->top()); 4311 } 4312 Node* in = region_clone->in(i); 4313 if (in->is_top()) { 4314 continue; 4315 } 4316 if (dom == nullptr) { 4317 dom = in; 4318 } else { 4319 dom = dom_lca(dom, in); 4320 } 4321 } 4322 4323 set_idom(region_clone, dom, dd); 4324 4325 // Set up the outer loop 4326 for (uint i = 0; i < head->outcnt(); i++) { 4327 Node* u = head->raw_out(i); 4328 if (u->is_Phi()) { 4329 Node* outer_phi = u->clone(); 4330 outer_phi->set_req(0, outer_head); 4331 Node* backedge = old_new[u->in(LoopNode::LoopBackControl)->_idx]; 4332 if (backedge == nullptr) { 4333 backedge = u->in(LoopNode::LoopBackControl); 4334 } 4335 outer_phi->set_req(LoopNode::LoopBackControl, backedge); 4336 register_new_node(outer_phi, outer_head); 4337 _igvn.replace_input_of(u, LoopNode::EntryControl, outer_phi); 4338 } 4339 } 4340 4341 // create control and data nodes for out of loop uses (including region2) 4342 Node_List worklist; 4343 uint new_counter = C->unique(); 4344 fix_ctrl_uses(wq, loop, old_new, ControlAroundStripMined, outer_head, nullptr, worklist); 4345 4346 Node_List *split_if_set = nullptr; 4347 Node_List *split_bool_set = nullptr; 4348 Node_List *split_cex_set = nullptr; 4349 fix_data_uses(wq, loop, ControlAroundStripMined, head->is_strip_mined() ? loop->_parent : loop, new_counter, old_new, worklist, split_if_set, split_bool_set, split_cex_set); 4350 4351 finish_clone_loop(split_if_set, split_bool_set, split_cex_set); 4352 4353 if (exit_test != nullptr) { 4354 float cnt = exit_test->_fcnt; 4355 if (cnt != COUNT_UNKNOWN) { 4356 exit_test->_fcnt = cnt * f; 4357 old_new[exit_test->_idx]->as_If()->_fcnt = cnt * (1 - f); 4358 } 4359 } 4360 4361 C->set_major_progress(); 4362 4363 return true; 4364 } 4365 4366 // Having ReductionNodes in the loop is expensive. They need to recursively 4367 // fold together the vector values, for every vectorized loop iteration. If 4368 // we encounter the following pattern, we can vector accumulate the values 4369 // inside the loop, and only have a single UnorderedReduction after the loop. 4370 // 4371 // CountedLoop init 4372 // | | 4373 // +------+ | +-----------------------+ 4374 // | | | | 4375 // PhiNode (s) | 4376 // | | 4377 // | Vector | 4378 // | | | 4379 // UnorderedReduction (first_ur) | 4380 // | | 4381 // ... Vector | 4382 // | | | 4383 // UnorderedReduction (last_ur) | 4384 // | | 4385 // +---------------------+ 4386 // 4387 // We patch the graph to look like this: 4388 // 4389 // CountedLoop identity_vector 4390 // | | 4391 // +-------+ | +---------------+ 4392 // | | | | 4393 // PhiNode (v) | 4394 // | | 4395 // | Vector | 4396 // | | | 4397 // VectorAccumulator | 4398 // | | 4399 // ... Vector | 4400 // | | | 4401 // init VectorAccumulator | 4402 // | | | | 4403 // UnorderedReduction +-----------+ 4404 // 4405 // We turned the scalar (s) Phi into a vectorized one (v). In the loop, we 4406 // use vector_accumulators, which do the same reductions, but only element 4407 // wise. This is a single operation per vector_accumulator, rather than many 4408 // for a UnorderedReduction. We can then reduce the last vector_accumulator 4409 // after the loop, and also reduce the init value into it. 4410 // We can not do this with all reductions. Some reductions do not allow the 4411 // reordering of operations (for example float addition). 4412 void PhaseIdealLoop::move_unordered_reduction_out_of_loop(IdealLoopTree* loop) { 4413 assert(!C->major_progress() && loop->is_counted() && loop->is_innermost(), "sanity"); 4414 4415 // Find all Phi nodes with UnorderedReduction on backedge. 4416 CountedLoopNode* cl = loop->_head->as_CountedLoop(); 4417 for (DUIterator_Fast jmax, j = cl->fast_outs(jmax); j < jmax; j++) { 4418 Node* phi = cl->fast_out(j); 4419 // We have a phi with a single use, and a UnorderedReduction on the backedge. 4420 if (!phi->is_Phi() || phi->outcnt() != 1 || !phi->in(2)->is_UnorderedReduction()) { 4421 continue; 4422 } 4423 4424 UnorderedReductionNode* last_ur = phi->in(2)->as_UnorderedReduction(); 4425 4426 // Determine types 4427 const TypeVect* vec_t = last_ur->vect_type(); 4428 uint vector_length = vec_t->length(); 4429 BasicType bt = vec_t->element_basic_type(); 4430 const Type* bt_t = Type::get_const_basic_type(bt); 4431 4432 // Convert opcode from vector-reduction -> scalar -> normal-vector-op 4433 const int sopc = VectorNode::scalar_opcode(last_ur->Opcode(), bt); 4434 const int vopc = VectorNode::opcode(sopc, bt); 4435 if (!Matcher::match_rule_supported_vector(vopc, vector_length, bt)) { 4436 DEBUG_ONLY( last_ur->dump(); ) 4437 assert(false, "do not have normal vector op for this reduction"); 4438 continue; // not implemented -> fails 4439 } 4440 4441 // Traverse up the chain of UnorderedReductions, checking that it loops back to 4442 // the phi. Check that all UnorderedReductions only have a single use, except for 4443 // the last (last_ur), which only has phi as a use in the loop, and all other uses 4444 // are outside the loop. 4445 UnorderedReductionNode* current = last_ur; 4446 UnorderedReductionNode* first_ur = nullptr; 4447 while (true) { 4448 assert(current->is_UnorderedReduction(), "sanity"); 4449 4450 // Expect no ctrl and a vector_input from within the loop. 4451 Node* ctrl = current->in(0); 4452 Node* vector_input = current->in(2); 4453 if (ctrl != nullptr || get_ctrl(vector_input) != cl) { 4454 DEBUG_ONLY( current->dump(1); ) 4455 assert(false, "reduction has ctrl or bad vector_input"); 4456 break; // Chain traversal fails. 4457 } 4458 4459 assert(current->vect_type() != nullptr, "must have vector type"); 4460 if (current->vect_type() != last_ur->vect_type()) { 4461 // Reductions do not have the same vector type (length and element type). 4462 break; // Chain traversal fails. 4463 } 4464 4465 // Expect single use of UnorderedReduction, except for last_ur. 4466 if (current == last_ur) { 4467 // Expect all uses to be outside the loop, except phi. 4468 for (DUIterator_Fast kmax, k = current->fast_outs(kmax); k < kmax; k++) { 4469 Node* use = current->fast_out(k); 4470 if (use != phi && ctrl_or_self(use) == cl) { 4471 DEBUG_ONLY( current->dump(-1); ) 4472 assert(false, "reduction has use inside loop"); 4473 // Should not be allowed by SuperWord::mark_reductions 4474 return; // bail out of optimization 4475 } 4476 } 4477 } else { 4478 if (current->outcnt() != 1) { 4479 break; // Chain traversal fails. 4480 } 4481 } 4482 4483 // Expect another UnorderedReduction or phi as the scalar input. 4484 Node* scalar_input = current->in(1); 4485 if (scalar_input->is_UnorderedReduction() && 4486 scalar_input->Opcode() == current->Opcode()) { 4487 // Move up the UnorderedReduction chain. 4488 current = scalar_input->as_UnorderedReduction(); 4489 } else if (scalar_input == phi) { 4490 // Chain terminates at phi. 4491 first_ur = current; 4492 current = nullptr; 4493 break; // Success. 4494 } else { 4495 // scalar_input is neither phi nor a matching reduction 4496 // Can for example be scalar reduction when we have 4497 // partial vectorization. 4498 break; // Chain traversal fails. 4499 } 4500 } 4501 if (current != nullptr) { 4502 // Chain traversal was not successful. 4503 continue; 4504 } 4505 assert(first_ur != nullptr, "must have successfully terminated chain traversal"); 4506 4507 Node* identity_scalar = ReductionNode::make_identity_con_scalar(_igvn, sopc, bt); 4508 set_ctrl(identity_scalar, C->root()); 4509 VectorNode* identity_vector = VectorNode::scalar2vector(identity_scalar, vector_length, bt_t); 4510 register_new_node(identity_vector, C->root()); 4511 assert(vec_t == identity_vector->vect_type(), "matching vector type"); 4512 VectorNode::trace_new_vector(identity_vector, "UnorderedReduction"); 4513 4514 // Turn the scalar phi into a vector phi. 4515 _igvn.rehash_node_delayed(phi); 4516 Node* init = phi->in(1); // Remember init before replacing it. 4517 phi->set_req_X(1, identity_vector, &_igvn); 4518 phi->as_Type()->set_type(vec_t); 4519 _igvn.set_type(phi, vec_t); 4520 4521 // Traverse down the chain of UnorderedReductions, and replace them with vector_accumulators. 4522 current = first_ur; 4523 while (true) { 4524 // Create vector_accumulator to replace current. 4525 Node* last_vector_accumulator = current->in(1); 4526 Node* vector_input = current->in(2); 4527 VectorNode* vector_accumulator = VectorNode::make(vopc, last_vector_accumulator, vector_input, vec_t); 4528 register_new_node(vector_accumulator, cl); 4529 _igvn.replace_node(current, vector_accumulator); 4530 VectorNode::trace_new_vector(vector_accumulator, "UnorderedReduction"); 4531 if (current == last_ur) { 4532 break; 4533 } 4534 current = vector_accumulator->unique_out()->as_UnorderedReduction(); 4535 } 4536 4537 // Create post-loop reduction. 4538 Node* last_accumulator = phi->in(2); 4539 Node* post_loop_reduction = ReductionNode::make(sopc, nullptr, init, last_accumulator, bt); 4540 4541 // Take over uses of last_accumulator that are not in the loop. 4542 for (DUIterator i = last_accumulator->outs(); last_accumulator->has_out(i); i++) { 4543 Node* use = last_accumulator->out(i); 4544 if (use != phi && use != post_loop_reduction) { 4545 assert(ctrl_or_self(use) != cl, "use must be outside loop"); 4546 use->replace_edge(last_accumulator, post_loop_reduction, &_igvn); 4547 --i; 4548 } 4549 } 4550 register_new_node(post_loop_reduction, get_late_ctrl(post_loop_reduction, cl)); 4551 VectorNode::trace_new_vector(post_loop_reduction, "UnorderedReduction"); 4552 4553 assert(last_accumulator->outcnt() == 2, "last_accumulator has 2 uses: phi and post_loop_reduction"); 4554 assert(post_loop_reduction->outcnt() > 0, "should have taken over all non loop uses of last_accumulator"); 4555 assert(phi->outcnt() == 1, "accumulator is the only use of phi"); 4556 } 4557 }