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 "utilities/macros.hpp" 46 47 //============================================================================= 48 //------------------------------split_thru_phi--------------------------------- 49 // Split Node 'n' through merge point if there is enough win. 50 Node* PhaseIdealLoop::split_thru_phi(Node* n, Node* region, int policy) { 51 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) { 52 // ConvI2L may have type information on it which is unsafe to push up 53 // so disable this for now 54 return nullptr; 55 } 56 57 // Splitting range check CastIIs through a loop induction Phi can 58 // cause new Phis to be created that are left unrelated to the loop 59 // induction Phi and prevent optimizations (vectorization) 60 if (n->Opcode() == Op_CastII && region->is_CountedLoop() && 61 n->in(1) == region->as_CountedLoop()->phi()) { 62 return nullptr; 63 } 64 65 // Inline types should not be split through Phis because they cannot be merged 66 // through Phi nodes but each value input needs to be merged individually. 67 if (n->is_InlineType()) { 68 return NULL; 69 } 70 71 if (cannot_split_division(n, region)) { 72 return nullptr; 73 } 74 75 int wins = 0; 76 assert(!n->is_CFG(), ""); 77 assert(region->is_Region(), ""); 78 79 const Type* type = n->bottom_type(); 80 const TypeOopPtr* t_oop = _igvn.type(n)->isa_oopptr(); 81 Node* phi; 82 if (t_oop != nullptr && t_oop->is_known_instance_field()) { 83 int iid = t_oop->instance_id(); 84 int index = C->get_alias_index(t_oop); 85 int offset = t_oop->offset(); 86 phi = new PhiNode(region, type, nullptr, iid, index, offset); 87 } else { 88 phi = PhiNode::make_blank(region, n); 89 } 90 uint old_unique = C->unique(); 91 for (uint i = 1; i < region->req(); i++) { 92 Node* x; 93 Node* the_clone = nullptr; 94 if (region->in(i) == C->top()) { 95 x = C->top(); // Dead path? Use a dead data op 96 } else { 97 x = n->clone(); // Else clone up the data op 98 the_clone = x; // Remember for possible deletion. 99 // Alter data node to use pre-phi inputs 100 if (n->in(0) == region) 101 x->set_req( 0, region->in(i) ); 102 for (uint j = 1; j < n->req(); j++) { 103 Node* in = n->in(j); 104 if (in->is_Phi() && in->in(0) == region) 105 x->set_req(j, in->in(i)); // Use pre-Phi input for the clone 106 } 107 } 108 // Check for a 'win' on some paths 109 const Type* t = x->Value(&_igvn); 110 111 bool singleton = t->singleton(); 112 113 // A TOP singleton indicates that there are no possible values incoming 114 // along a particular edge. In most cases, this is OK, and the Phi will 115 // be eliminated later in an Ideal call. However, we can't allow this to 116 // happen if the singleton occurs on loop entry, as the elimination of 117 // the PhiNode may cause the resulting node to migrate back to a previous 118 // loop iteration. 119 if (singleton && t == Type::TOP) { 120 // Is_Loop() == false does not confirm the absence of a loop (e.g., an 121 // irreducible loop may not be indicated by an affirmative is_Loop()); 122 // therefore, the only top we can split thru a phi is on a backedge of 123 // a loop. 124 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 125 } 126 127 if (singleton) { 128 wins++; 129 x = ((PhaseGVN&)_igvn).makecon(t); 130 } else { 131 // We now call Identity to try to simplify the cloned node. 132 // Note that some Identity methods call phase->type(this). 133 // Make sure that the type array is big enough for 134 // our new node, even though we may throw the node away. 135 // (Note: This tweaking with igvn only works because x is a new node.) 136 _igvn.set_type(x, t); 137 // If x is a TypeNode, capture any more-precise type permanently into Node 138 // otherwise it will be not updated during igvn->transform since 139 // igvn->type(x) is set to x->Value() already. 140 x->raise_bottom_type(t); 141 Node* y = x->Identity(&_igvn); 142 if (y != x) { 143 wins++; 144 x = y; 145 } else { 146 y = _igvn.hash_find(x); 147 if (y) { 148 wins++; 149 x = y; 150 } else { 151 // Else x is a new node we are keeping 152 // We do not need register_new_node_with_optimizer 153 // because set_type has already been called. 154 _igvn._worklist.push(x); 155 } 156 } 157 } 158 if (x != the_clone && the_clone != nullptr) 159 _igvn.remove_dead_node(the_clone); 160 phi->set_req( i, x ); 161 } 162 // Too few wins? 163 if (wins <= policy) { 164 _igvn.remove_dead_node(phi); 165 return nullptr; 166 } 167 168 // Record Phi 169 register_new_node( phi, region ); 170 171 for (uint i2 = 1; i2 < phi->req(); i2++) { 172 Node *x = phi->in(i2); 173 // If we commoned up the cloned 'x' with another existing Node, 174 // the existing Node picks up a new use. We need to make the 175 // existing Node occur higher up so it dominates its uses. 176 Node *old_ctrl; 177 IdealLoopTree *old_loop; 178 179 if (x->is_Con()) { 180 // Constant's control is always root. 181 set_ctrl(x, C->root()); 182 continue; 183 } 184 // The occasional new node 185 if (x->_idx >= old_unique) { // Found a new, unplaced node? 186 old_ctrl = nullptr; 187 old_loop = nullptr; // Not in any prior loop 188 } else { 189 old_ctrl = get_ctrl(x); 190 old_loop = get_loop(old_ctrl); // Get prior loop 191 } 192 // New late point must dominate new use 193 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2)); 194 if (new_ctrl == old_ctrl) // Nothing is changed 195 continue; 196 197 IdealLoopTree *new_loop = get_loop(new_ctrl); 198 199 // Don't move x into a loop if its uses are 200 // outside of loop. Otherwise x will be cloned 201 // for each use outside of this loop. 202 IdealLoopTree *use_loop = get_loop(region); 203 if (!new_loop->is_member(use_loop) && 204 (old_loop == nullptr || !new_loop->is_member(old_loop))) { 205 // Take early control, later control will be recalculated 206 // during next iteration of loop optimizations. 207 new_ctrl = get_early_ctrl(x); 208 new_loop = get_loop(new_ctrl); 209 } 210 // Set new location 211 set_ctrl(x, new_ctrl); 212 // If changing loop bodies, see if we need to collect into new body 213 if (old_loop != new_loop) { 214 if (old_loop && !old_loop->_child) 215 old_loop->_body.yank(x); 216 if (!new_loop->_child) 217 new_loop->_body.push(x); // Collect body info 218 } 219 } 220 221 return phi; 222 } 223 224 // Return true if 'n' is a Div or Mod node (without zero check If node which was removed earlier) with a loop phi divisor 225 // of a trip-counted (integer or long) loop with a backedge input that could be zero (include zero in its type range). In 226 // this case, we cannot split the division to the backedge as it could freely float above the loop exit check resulting in 227 // a division by zero. This situation is possible because the type of an increment node of an iv phi (trip-counter) could 228 // include zero while the iv phi does not (see PhiNode::Value() for trip-counted loops where we improve types of iv phis). 229 // We also need to check other loop phis as they could have been created in the same split-if pass when applying 230 // PhaseIdealLoop::split_thru_phi() to split nodes through an iv phi. 231 bool PhaseIdealLoop::cannot_split_division(const Node* n, const Node* region) const { 232 const Type* zero; 233 switch (n->Opcode()) { 234 case Op_DivI: 235 case Op_ModI: 236 zero = TypeInt::ZERO; 237 break; 238 case Op_DivL: 239 case Op_ModL: 240 zero = TypeLong::ZERO; 241 break; 242 default: 243 return false; 244 } 245 246 assert(n->in(0) == nullptr, "divisions with zero check should already have bailed out earlier in split-if"); 247 Node* divisor = n->in(2); 248 return is_divisor_counted_loop_phi(divisor, region) && 249 loop_phi_backedge_type_contains_zero(divisor, zero); 250 } 251 252 bool PhaseIdealLoop::is_divisor_counted_loop_phi(const Node* divisor, const Node* loop) { 253 return loop->is_BaseCountedLoop() && divisor->is_Phi() && divisor->in(0) == loop; 254 } 255 256 bool PhaseIdealLoop::loop_phi_backedge_type_contains_zero(const Node* phi_divisor, const Type* zero) const { 257 return _igvn.type(phi_divisor->in(LoopNode::LoopBackControl))->filter_speculative(zero) != Type::TOP; 258 } 259 260 //------------------------------dominated_by------------------------------------ 261 // Replace the dominated test with an obvious true or false. Place it on the 262 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the 263 // live path up to the dominating control. 264 void PhaseIdealLoop::dominated_by(IfProjNode* prevdom, IfNode* iff, bool flip, bool exclude_loop_predicate) { 265 if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); } 266 267 // prevdom is the dominating projection of the dominating test. 268 assert(iff->Opcode() == Op_If || 269 iff->Opcode() == Op_CountedLoopEnd || 270 iff->Opcode() == Op_LongCountedLoopEnd || 271 iff->Opcode() == Op_RangeCheck, 272 "Check this code when new subtype is added"); 273 274 int pop = prevdom->Opcode(); 275 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" ); 276 if (flip) { 277 if (pop == Op_IfTrue) 278 pop = Op_IfFalse; 279 else 280 pop = Op_IfTrue; 281 } 282 // 'con' is set to true or false to kill the dominated test. 283 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO); 284 set_ctrl(con, C->root()); // Constant gets a new use 285 // Hack the dominated test 286 _igvn.replace_input_of(iff, 1, con); 287 288 // If I dont have a reachable TRUE and FALSE path following the IfNode then 289 // I can assume this path reaches an infinite loop. In this case it's not 290 // important to optimize the data Nodes - either the whole compilation will 291 // be tossed or this path (and all data Nodes) will go dead. 292 if (iff->outcnt() != 2) return; 293 294 // Make control-dependent data Nodes on the live path (path that will remain 295 // once the dominated IF is removed) become control-dependent on the 296 // dominating projection. 297 Node* dp = iff->proj_out_or_null(pop == Op_IfTrue); 298 299 // Loop predicates may have depending checks which should not 300 // be skipped. For example, range check predicate has two checks 301 // for lower and upper bounds. 302 if (dp == nullptr) 303 return; 304 305 ProjNode* dp_proj = dp->as_Proj(); 306 ProjNode* unc_proj = iff->proj_out(1 - dp_proj->_con)->as_Proj(); 307 if (exclude_loop_predicate && 308 (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) != nullptr || 309 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_profile_predicate) != nullptr || 310 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check) != nullptr)) { 311 // If this is a range check (IfNode::is_range_check), do not 312 // reorder because Compile::allow_range_check_smearing might have 313 // changed the check. 314 return; // Let IGVN transformation change control dependence. 315 } 316 317 IdealLoopTree* old_loop = get_loop(dp); 318 319 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { 320 Node* cd = dp->fast_out(i); // Control-dependent node 321 // Do not rewire Div and Mod nodes which could have a zero divisor to avoid skipping their zero check. 322 if (cd->depends_only_on_test() && _igvn.no_dependent_zero_check(cd)) { 323 assert(cd->in(0) == dp, ""); 324 _igvn.replace_input_of(cd, 0, prevdom); 325 set_early_ctrl(cd, false); 326 IdealLoopTree* new_loop = get_loop(get_ctrl(cd)); 327 if (old_loop != new_loop) { 328 if (!old_loop->_child) { 329 old_loop->_body.yank(cd); 330 } 331 if (!new_loop->_child) { 332 new_loop->_body.push(cd); 333 } 334 } 335 --i; 336 --imax; 337 } 338 } 339 } 340 341 //------------------------------has_local_phi_input---------------------------- 342 // Return TRUE if 'n' has Phi inputs from its local block and no other 343 // block-local inputs (all non-local-phi inputs come from earlier blocks) 344 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) { 345 Node *n_ctrl = get_ctrl(n); 346 // See if some inputs come from a Phi in this block, or from before 347 // this block. 348 uint i; 349 for( i = 1; i < n->req(); i++ ) { 350 Node *phi = n->in(i); 351 if( phi->is_Phi() && phi->in(0) == n_ctrl ) 352 break; 353 } 354 if( i >= n->req() ) 355 return nullptr; // No Phi inputs; nowhere to clone thru 356 357 // Check for inputs created between 'n' and the Phi input. These 358 // must split as well; they have already been given the chance 359 // (courtesy of a post-order visit) and since they did not we must 360 // recover the 'cost' of splitting them by being very profitable 361 // when splitting 'n'. Since this is unlikely we simply give up. 362 for( i = 1; i < n->req(); i++ ) { 363 Node *m = n->in(i); 364 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) { 365 // We allow the special case of AddP's with no local inputs. 366 // This allows us to split-up address expressions. 367 if (m->is_AddP() && 368 get_ctrl(m->in(AddPNode::Base)) != n_ctrl && 369 get_ctrl(m->in(AddPNode::Address)) != n_ctrl && 370 get_ctrl(m->in(AddPNode::Offset)) != n_ctrl) { 371 // Move the AddP up to the dominating point. That's fine because control of m's inputs 372 // must dominate get_ctrl(m) == n_ctrl and we just checked that the input controls are != n_ctrl. 373 Node* c = find_non_split_ctrl(idom(n_ctrl)); 374 if (c->is_OuterStripMinedLoop()) { 375 c->as_Loop()->verify_strip_mined(1); 376 c = c->in(LoopNode::EntryControl); 377 } 378 set_ctrl_and_loop(m, c); 379 continue; 380 } 381 return nullptr; 382 } 383 assert(n->is_Phi() || m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control"); 384 } 385 386 return n_ctrl; 387 } 388 389 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2). 390 Node* PhaseIdealLoop::remix_address_expressions_add_left_shift(Node* n, IdealLoopTree* n_loop, Node* n_ctrl, BasicType bt) { 391 assert(bt == T_INT || bt == T_LONG, "only for integers"); 392 int n_op = n->Opcode(); 393 394 if (n_op == Op_LShift(bt)) { 395 // Scale is loop invariant 396 Node* scale = n->in(2); 397 Node* scale_ctrl = get_ctrl(scale); 398 IdealLoopTree* scale_loop = get_loop(scale_ctrl); 399 if (n_loop == scale_loop || !scale_loop->is_member(n_loop)) { 400 return nullptr; 401 } 402 const TypeInt* scale_t = scale->bottom_type()->isa_int(); 403 if (scale_t != nullptr && scale_t->is_con() && scale_t->get_con() >= 16) { 404 return nullptr; // Dont bother with byte/short masking 405 } 406 // Add must vary with loop (else shift would be loop-invariant) 407 Node* add = n->in(1); 408 Node* add_ctrl = get_ctrl(add); 409 IdealLoopTree* add_loop = get_loop(add_ctrl); 410 if (n_loop != add_loop) { 411 return nullptr; // happens w/ evil ZKM loops 412 } 413 414 // Convert I-V into I+ (0-V); same for V-I 415 if (add->Opcode() == Op_Sub(bt) && 416 _igvn.type(add->in(1)) != TypeInteger::zero(bt)) { 417 assert(add->Opcode() == Op_SubI || add->Opcode() == Op_SubL, ""); 418 Node* zero = _igvn.integercon(0, bt); 419 set_ctrl(zero, C->root()); 420 Node* neg = SubNode::make(zero, add->in(2), bt); 421 register_new_node(neg, get_ctrl(add->in(2))); 422 add = AddNode::make(add->in(1), neg, bt); 423 register_new_node(add, add_ctrl); 424 } 425 if (add->Opcode() != Op_Add(bt)) return nullptr; 426 assert(add->Opcode() == Op_AddI || add->Opcode() == Op_AddL, ""); 427 // See if one add input is loop invariant 428 Node* add_var = add->in(1); 429 Node* add_var_ctrl = get_ctrl(add_var); 430 IdealLoopTree* add_var_loop = get_loop(add_var_ctrl); 431 Node* add_invar = add->in(2); 432 Node* add_invar_ctrl = get_ctrl(add_invar); 433 IdealLoopTree* add_invar_loop = get_loop(add_invar_ctrl); 434 if (add_invar_loop == n_loop) { 435 // Swap to find the invariant part 436 add_invar = add_var; 437 add_invar_ctrl = add_var_ctrl; 438 add_invar_loop = add_var_loop; 439 add_var = add->in(2); 440 } else if (add_var_loop != n_loop) { // Else neither input is loop invariant 441 return nullptr; 442 } 443 if (n_loop == add_invar_loop || !add_invar_loop->is_member(n_loop)) { 444 return nullptr; // No invariant part of the add? 445 } 446 447 // Yes! Reshape address expression! 448 Node* inv_scale = LShiftNode::make(add_invar, scale, bt); 449 Node* inv_scale_ctrl = 450 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ? 451 add_invar_ctrl : scale_ctrl; 452 register_new_node(inv_scale, inv_scale_ctrl); 453 Node* var_scale = LShiftNode::make(add_var, scale, bt); 454 register_new_node(var_scale, n_ctrl); 455 Node* var_add = AddNode::make(var_scale, inv_scale, bt); 456 register_new_node(var_add, n_ctrl); 457 _igvn.replace_node(n, var_add); 458 return var_add; 459 } 460 return nullptr; 461 } 462 463 //------------------------------remix_address_expressions---------------------- 464 // Rework addressing expressions to get the most loop-invariant stuff 465 // moved out. We'd like to do all associative operators, but it's especially 466 // important (common) to do address expressions. 467 Node* PhaseIdealLoop::remix_address_expressions(Node* n) { 468 if (!has_ctrl(n)) return nullptr; 469 Node* n_ctrl = get_ctrl(n); 470 IdealLoopTree* n_loop = get_loop(n_ctrl); 471 472 // See if 'n' mixes loop-varying and loop-invariant inputs and 473 // itself is loop-varying. 474 475 // Only interested in binary ops (and AddP) 476 if (n->req() < 3 || n->req() > 4) return nullptr; 477 478 Node* n1_ctrl = get_ctrl(n->in( 1)); 479 Node* n2_ctrl = get_ctrl(n->in( 2)); 480 Node* n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3)); 481 IdealLoopTree* n1_loop = get_loop(n1_ctrl); 482 IdealLoopTree* n2_loop = get_loop(n2_ctrl); 483 IdealLoopTree* n3_loop = get_loop(n3_ctrl); 484 485 // Does one of my inputs spin in a tighter loop than self? 486 if ((n_loop->is_member(n1_loop) && n_loop != n1_loop) || 487 (n_loop->is_member(n2_loop) && n_loop != n2_loop) || 488 (n_loop->is_member(n3_loop) && n_loop != n3_loop)) { 489 return nullptr; // Leave well enough alone 490 } 491 492 // Is at least one of my inputs loop-invariant? 493 if (n1_loop == n_loop && 494 n2_loop == n_loop && 495 n3_loop == n_loop) { 496 return nullptr; // No loop-invariant inputs 497 } 498 499 Node* res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_INT); 500 if (res != nullptr) { 501 return res; 502 } 503 res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_LONG); 504 if (res != nullptr) { 505 return res; 506 } 507 508 int n_op = n->Opcode(); 509 // Replace (I+V) with (V+I) 510 if (n_op == Op_AddI || 511 n_op == Op_AddL || 512 n_op == Op_AddF || 513 n_op == Op_AddD || 514 n_op == Op_MulI || 515 n_op == Op_MulL || 516 n_op == Op_MulF || 517 n_op == Op_MulD) { 518 if (n2_loop == n_loop) { 519 assert(n1_loop != n_loop, ""); 520 n->swap_edges(1, 2); 521 } 522 } 523 524 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V), 525 // but not if I2 is a constant. 526 if (n_op == Op_AddP) { 527 if (n2_loop == n_loop && n3_loop != n_loop) { 528 if (n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con()) { 529 Node* n22_ctrl = get_ctrl(n->in(2)->in(2)); 530 Node* n23_ctrl = get_ctrl(n->in(2)->in(3)); 531 IdealLoopTree* n22loop = get_loop(n22_ctrl); 532 IdealLoopTree* n23_loop = get_loop(n23_ctrl); 533 if (n22loop != n_loop && n22loop->is_member(n_loop) && 534 n23_loop == n_loop) { 535 Node* add1 = new AddPNode(n->in(1), n->in(2)->in(2), n->in(3)); 536 // Stuff new AddP in the loop preheader 537 register_new_node(add1, n_loop->_head->as_Loop()->skip_strip_mined(1)->in(LoopNode::EntryControl)); 538 Node* add2 = new AddPNode(n->in(1), add1, n->in(2)->in(3)); 539 register_new_node(add2, n_ctrl); 540 _igvn.replace_node(n, add2); 541 return add2; 542 } 543 } 544 } 545 546 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V) 547 if (n2_loop != n_loop && n3_loop == n_loop) { 548 if (n->in(3)->Opcode() == Op_AddX) { 549 Node* V = n->in(3)->in(1); 550 Node* I = n->in(3)->in(2); 551 if (is_member(n_loop,get_ctrl(V))) { 552 } else { 553 Node *tmp = V; V = I; I = tmp; 554 } 555 if (!is_member(n_loop,get_ctrl(I))) { 556 Node* add1 = new AddPNode(n->in(1), n->in(2), I); 557 // Stuff new AddP in the loop preheader 558 register_new_node(add1, n_loop->_head->as_Loop()->skip_strip_mined(1)->in(LoopNode::EntryControl)); 559 Node* add2 = new AddPNode(n->in(1), add1, V); 560 register_new_node(add2, n_ctrl); 561 _igvn.replace_node(n, add2); 562 return add2; 563 } 564 } 565 } 566 } 567 568 return nullptr; 569 } 570 571 // Optimize ((in1[2*i] * in2[2*i]) + (in1[2*i+1] * in2[2*i+1])) 572 Node *PhaseIdealLoop::convert_add_to_muladd(Node* n) { 573 assert(n->Opcode() == Op_AddI, "sanity"); 574 Node * nn = nullptr; 575 Node * in1 = n->in(1); 576 Node * in2 = n->in(2); 577 if (in1->Opcode() == Op_MulI && in2->Opcode() == Op_MulI) { 578 IdealLoopTree* loop_n = get_loop(get_ctrl(n)); 579 if (loop_n->is_counted() && 580 loop_n->_head->as_Loop()->is_valid_counted_loop(T_INT) && 581 Matcher::match_rule_supported(Op_MulAddVS2VI) && 582 Matcher::match_rule_supported(Op_MulAddS2I)) { 583 Node* mul_in1 = in1->in(1); 584 Node* mul_in2 = in1->in(2); 585 Node* mul_in3 = in2->in(1); 586 Node* mul_in4 = in2->in(2); 587 if (mul_in1->Opcode() == Op_LoadS && 588 mul_in2->Opcode() == Op_LoadS && 589 mul_in3->Opcode() == Op_LoadS && 590 mul_in4->Opcode() == Op_LoadS) { 591 IdealLoopTree* loop1 = get_loop(get_ctrl(mul_in1)); 592 IdealLoopTree* loop2 = get_loop(get_ctrl(mul_in2)); 593 IdealLoopTree* loop3 = get_loop(get_ctrl(mul_in3)); 594 IdealLoopTree* loop4 = get_loop(get_ctrl(mul_in4)); 595 IdealLoopTree* loop5 = get_loop(get_ctrl(in1)); 596 IdealLoopTree* loop6 = get_loop(get_ctrl(in2)); 597 // All nodes should be in the same counted loop. 598 if (loop_n == loop1 && loop_n == loop2 && loop_n == loop3 && 599 loop_n == loop4 && loop_n == loop5 && loop_n == loop6) { 600 Node* adr1 = mul_in1->in(MemNode::Address); 601 Node* adr2 = mul_in2->in(MemNode::Address); 602 Node* adr3 = mul_in3->in(MemNode::Address); 603 Node* adr4 = mul_in4->in(MemNode::Address); 604 if (adr1->is_AddP() && adr2->is_AddP() && adr3->is_AddP() && adr4->is_AddP()) { 605 if ((adr1->in(AddPNode::Base) == adr3->in(AddPNode::Base)) && 606 (adr2->in(AddPNode::Base) == adr4->in(AddPNode::Base))) { 607 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in3, mul_in4); 608 register_new_node(nn, get_ctrl(n)); 609 _igvn.replace_node(n, nn); 610 return nn; 611 } else if ((adr1->in(AddPNode::Base) == adr4->in(AddPNode::Base)) && 612 (adr2->in(AddPNode::Base) == adr3->in(AddPNode::Base))) { 613 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in4, mul_in3); 614 register_new_node(nn, get_ctrl(n)); 615 _igvn.replace_node(n, nn); 616 return nn; 617 } 618 } 619 } 620 } 621 } 622 } 623 return nn; 624 } 625 626 //------------------------------conditional_move------------------------------- 627 // Attempt to replace a Phi with a conditional move. We have some pretty 628 // strict profitability requirements. All Phis at the merge point must 629 // be converted, so we can remove the control flow. We need to limit the 630 // number of c-moves to a small handful. All code that was in the side-arms 631 // of the CFG diamond is now speculatively executed. This code has to be 632 // "cheap enough". We are pretty much limited to CFG diamonds that merge 633 // 1 or 2 items with a total of 1 or 2 ops executed speculatively. 634 Node *PhaseIdealLoop::conditional_move( Node *region ) { 635 636 assert(region->is_Region(), "sanity check"); 637 if (region->req() != 3) return nullptr; 638 639 // Check for CFG diamond 640 Node *lp = region->in(1); 641 Node *rp = region->in(2); 642 if (!lp || !rp) return nullptr; 643 Node *lp_c = lp->in(0); 644 if (lp_c == nullptr || lp_c != rp->in(0) || !lp_c->is_If()) return nullptr; 645 IfNode *iff = lp_c->as_If(); 646 647 // Check for ops pinned in an arm of the diamond. 648 // Can't remove the control flow in this case 649 if (lp->outcnt() > 1) return nullptr; 650 if (rp->outcnt() > 1) return nullptr; 651 652 IdealLoopTree* r_loop = get_loop(region); 653 assert(r_loop == get_loop(iff), "sanity"); 654 // Always convert to CMOVE if all results are used only outside this loop. 655 bool used_inside_loop = (r_loop == _ltree_root); 656 657 // Check profitability 658 int cost = 0; 659 int phis = 0; 660 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 661 Node *out = region->fast_out(i); 662 if (!out->is_Phi()) continue; // Ignore other control edges, etc 663 phis++; 664 PhiNode* phi = out->as_Phi(); 665 BasicType bt = phi->type()->basic_type(); 666 switch (bt) { 667 case T_DOUBLE: 668 case T_FLOAT: 669 if (C->use_cmove()) { 670 continue; //TODO: maybe we want to add some cost 671 } 672 cost += Matcher::float_cmove_cost(); // Could be very expensive 673 break; 674 case T_LONG: { 675 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's 676 } 677 case T_INT: // These all CMOV fine 678 case T_ADDRESS: { // (RawPtr) 679 cost++; 680 break; 681 } 682 case T_NARROWOOP: // Fall through 683 case T_OBJECT: { // Base oops are OK, but not derived oops 684 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr(); 685 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a 686 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus 687 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we 688 // have a Phi for the base here that we convert to a CMOVE all is well 689 // and good. But if the base is dead, we'll not make a CMOVE. Later 690 // the allocator will have to produce a base by creating a CMOVE of the 691 // relevant bases. This puts the allocator in the business of 692 // manufacturing expensive instructions, generally a bad plan. 693 // Just Say No to Conditionally-Moved Derived Pointers. 694 if (tp && tp->offset() != 0) 695 return nullptr; 696 cost++; 697 break; 698 } 699 default: 700 return nullptr; // In particular, can't do memory or I/O 701 } 702 // Add in cost any speculative ops 703 for (uint j = 1; j < region->req(); j++) { 704 Node *proj = region->in(j); 705 Node *inp = phi->in(j); 706 if (inp->isa_InlineType()) { 707 // TODO 8302217 This prevents PhiNode::push_inline_types_through 708 return NULL; 709 } 710 if (get_ctrl(inp) == proj) { // Found local op 711 cost++; 712 // Check for a chain of dependent ops; these will all become 713 // speculative in a CMOV. 714 for (uint k = 1; k < inp->req(); k++) 715 if (get_ctrl(inp->in(k)) == proj) 716 cost += ConditionalMoveLimit; // Too much speculative goo 717 } 718 } 719 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode. 720 // This will likely Split-If, a higher-payoff operation. 721 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) { 722 Node* use = phi->fast_out(k); 723 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr()) 724 cost += ConditionalMoveLimit; 725 // Is there a use inside the loop? 726 // Note: check only basic types since CMoveP is pinned. 727 if (!used_inside_loop && is_java_primitive(bt)) { 728 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use); 729 if (r_loop == u_loop || r_loop->is_member(u_loop)) { 730 used_inside_loop = true; 731 } 732 } 733 } 734 }//for 735 Node* bol = iff->in(1); 736 if (bol->Opcode() == Op_Opaque4) { 737 return nullptr; // Ignore loop predicate checks (the Opaque4 ensures they will go away) 738 } 739 assert(bol->Opcode() == Op_Bool, "Unexpected node"); 740 int cmp_op = bol->in(1)->Opcode(); 741 if (cmp_op == Op_SubTypeCheck) { // SubTypeCheck expansion expects an IfNode 742 return nullptr; 743 } 744 // It is expensive to generate flags from a float compare. 745 // Avoid duplicated float compare. 746 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return nullptr; 747 748 float infrequent_prob = PROB_UNLIKELY_MAG(3); 749 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop. 750 if (used_inside_loop) { 751 if (cost >= ConditionalMoveLimit) return nullptr; // Too much goo 752 753 // BlockLayoutByFrequency optimization moves infrequent branch 754 // from hot path. No point in CMOV'ing in such case (110 is used 755 // instead of 100 to take into account not exactness of float value). 756 if (BlockLayoutByFrequency) { 757 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f); 758 } 759 } 760 // Check for highly predictable branch. No point in CMOV'ing if 761 // we are going to predict accurately all the time. 762 if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) { 763 //keep going 764 } else if (iff->_prob < infrequent_prob || 765 iff->_prob > (1.0f - infrequent_prob)) 766 return nullptr; 767 768 // -------------- 769 // Now replace all Phis with CMOV's 770 Node *cmov_ctrl = iff->in(0); 771 uint flip = (lp->Opcode() == Op_IfTrue); 772 Node_List wq; 773 while (1) { 774 PhiNode* phi = nullptr; 775 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 776 Node *out = region->fast_out(i); 777 if (out->is_Phi()) { 778 phi = out->as_Phi(); 779 break; 780 } 781 } 782 if (phi == nullptr || _igvn.type(phi) == Type::TOP) { 783 break; 784 } 785 if (PrintOpto && VerifyLoopOptimizations) { tty->print_cr("CMOV"); } 786 // Move speculative ops 787 wq.push(phi); 788 while (wq.size() > 0) { 789 Node *n = wq.pop(); 790 for (uint j = 1; j < n->req(); j++) { 791 Node* m = n->in(j); 792 if (m != nullptr && !is_dominator(get_ctrl(m), cmov_ctrl)) { 793 #ifndef PRODUCT 794 if (PrintOpto && VerifyLoopOptimizations) { 795 tty->print(" speculate: "); 796 m->dump(); 797 } 798 #endif 799 set_ctrl(m, cmov_ctrl); 800 wq.push(m); 801 } 802 } 803 } 804 Node *cmov = CMoveNode::make(cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi)); 805 register_new_node( cmov, cmov_ctrl ); 806 _igvn.replace_node( phi, cmov ); 807 #ifndef PRODUCT 808 if (TraceLoopOpts) { 809 tty->print("CMOV "); 810 r_loop->dump_head(); 811 if (Verbose) { 812 bol->in(1)->dump(1); 813 cmov->dump(1); 814 } 815 } 816 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } ); 817 #endif 818 } 819 820 // The useless CFG diamond will fold up later; see the optimization in 821 // RegionNode::Ideal. 822 _igvn._worklist.push(region); 823 824 return iff->in(1); 825 } 826 827 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) { 828 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 829 Node* u = m->fast_out(i); 830 if (u->is_CFG()) { 831 if (u->is_NeverBranch()) { 832 u = u->as_NeverBranch()->proj_out(0); 833 enqueue_cfg_uses(u, wq); 834 } else { 835 wq.push(u); 836 } 837 } 838 } 839 } 840 841 // Try moving a store out of a loop, right before the loop 842 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) { 843 // Store has to be first in the loop body 844 IdealLoopTree *n_loop = get_loop(n_ctrl); 845 if (n->is_Store() && n_loop != _ltree_root && 846 n_loop->is_loop() && n_loop->_head->is_Loop() && 847 n->in(0) != nullptr) { 848 Node* address = n->in(MemNode::Address); 849 Node* value = n->in(MemNode::ValueIn); 850 Node* mem = n->in(MemNode::Memory); 851 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 852 IdealLoopTree* value_loop = get_loop(get_ctrl(value)); 853 854 // - address and value must be loop invariant 855 // - memory must be a memory Phi for the loop 856 // - Store must be the only store on this memory slice in the 857 // loop: if there's another store following this one then value 858 // written at iteration i by the second store could be overwritten 859 // at iteration i+n by the first store: it's not safe to move the 860 // first store out of the loop 861 // - nothing must observe the memory Phi: it guarantees no read 862 // before the store, we are also guaranteed the store post 863 // dominates the loop head (ignoring a possible early 864 // exit). Otherwise there would be extra Phi involved between the 865 // loop's Phi and the store. 866 // - there must be no early exit from the loop before the Store 867 // (such an exit most of the time would be an extra use of the 868 // memory Phi but sometimes is a bottom memory Phi that takes the 869 // store as input). 870 871 if (!n_loop->is_member(address_loop) && 872 !n_loop->is_member(value_loop) && 873 mem->is_Phi() && mem->in(0) == n_loop->_head && 874 mem->outcnt() == 1 && 875 mem->in(LoopNode::LoopBackControl) == n) { 876 877 assert(n_loop->_tail != nullptr, "need a tail"); 878 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop"); 879 880 // Verify that there's no early exit of the loop before the store. 881 bool ctrl_ok = false; 882 { 883 // Follow control from loop head until n, we exit the loop or 884 // we reach the tail 885 ResourceMark rm; 886 Unique_Node_List wq; 887 wq.push(n_loop->_head); 888 889 for (uint next = 0; next < wq.size(); ++next) { 890 Node *m = wq.at(next); 891 if (m == n->in(0)) { 892 ctrl_ok = true; 893 continue; 894 } 895 assert(!has_ctrl(m), "should be CFG"); 896 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) { 897 ctrl_ok = false; 898 break; 899 } 900 enqueue_cfg_uses(m, wq); 901 if (wq.size() > 10) { 902 ctrl_ok = false; 903 break; 904 } 905 } 906 } 907 if (ctrl_ok) { 908 // move the Store 909 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem); 910 _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl)); 911 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl)); 912 // Disconnect the phi now. An empty phi can confuse other 913 // optimizations in this pass of loop opts. 914 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl)); 915 n_loop->_body.yank(mem); 916 917 set_ctrl_and_loop(n, n->in(0)); 918 919 return n; 920 } 921 } 922 } 923 return nullptr; 924 } 925 926 // Try moving a store out of a loop, right after the loop 927 void PhaseIdealLoop::try_move_store_after_loop(Node* n) { 928 if (n->is_Store() && n->in(0) != nullptr) { 929 Node *n_ctrl = get_ctrl(n); 930 IdealLoopTree *n_loop = get_loop(n_ctrl); 931 // Store must be in a loop 932 if (n_loop != _ltree_root && !n_loop->_irreducible) { 933 Node* address = n->in(MemNode::Address); 934 Node* value = n->in(MemNode::ValueIn); 935 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 936 // address must be loop invariant 937 if (!n_loop->is_member(address_loop)) { 938 // Store must be last on this memory slice in the loop and 939 // nothing in the loop must observe it 940 Node* phi = nullptr; 941 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 942 Node* u = n->fast_out(i); 943 if (has_ctrl(u)) { // control use? 944 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 945 if (!n_loop->is_member(u_loop)) { 946 continue; 947 } 948 if (u->is_Phi() && u->in(0) == n_loop->_head) { 949 assert(_igvn.type(u) == Type::MEMORY, "bad phi"); 950 // multiple phis on the same slice are possible 951 if (phi != nullptr) { 952 return; 953 } 954 phi = u; 955 continue; 956 } 957 } 958 return; 959 } 960 if (phi != nullptr) { 961 // Nothing in the loop before the store (next iteration) 962 // must observe the stored value 963 bool mem_ok = true; 964 { 965 ResourceMark rm; 966 Unique_Node_List wq; 967 wq.push(phi); 968 for (uint next = 0; next < wq.size() && mem_ok; ++next) { 969 Node *m = wq.at(next); 970 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) { 971 Node* u = m->fast_out(i); 972 if (u->is_Store() || u->is_Phi()) { 973 if (u != n) { 974 wq.push(u); 975 mem_ok = (wq.size() <= 10); 976 } 977 } else { 978 mem_ok = false; 979 break; 980 } 981 } 982 } 983 } 984 if (mem_ok) { 985 // Move the store out of the loop if the LCA of all 986 // users (except for the phi) is outside the loop. 987 Node* hook = new Node(1); 988 hook->init_req(0, n_ctrl); // Add an input to prevent hook from being dead 989 _igvn.rehash_node_delayed(phi); 990 int count = phi->replace_edge(n, hook, &_igvn); 991 assert(count > 0, "inconsistent phi"); 992 993 // Compute latest point this store can go 994 Node* lca = get_late_ctrl(n, get_ctrl(n)); 995 if (lca->is_OuterStripMinedLoop()) { 996 lca = lca->in(LoopNode::EntryControl); 997 } 998 if (n_loop->is_member(get_loop(lca))) { 999 // LCA is in the loop - bail out 1000 _igvn.replace_node(hook, n); 1001 return; 1002 } 1003 #ifdef ASSERT 1004 if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) { 1005 assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined"); 1006 n_loop->_head->as_Loop()->verify_strip_mined(1); 1007 Node* outer = n_loop->_head->as_CountedLoop()->outer_loop(); 1008 IdealLoopTree* outer_loop = get_loop(outer); 1009 assert(n_loop->_parent == outer_loop, "broken loop tree"); 1010 assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state"); 1011 } 1012 #endif 1013 lca = place_outside_loop(lca, n_loop); 1014 assert(!n_loop->is_member(get_loop(lca)), "control must not be back in the loop"); 1015 assert(get_loop(lca)->_nest < n_loop->_nest || lca->in(0)->is_NeverBranch(), "must not be moved into inner loop"); 1016 1017 // Move store out of the loop 1018 _igvn.replace_node(hook, n->in(MemNode::Memory)); 1019 _igvn.replace_input_of(n, 0, lca); 1020 set_ctrl_and_loop(n, lca); 1021 1022 // Disconnect the phi now. An empty phi can confuse other 1023 // optimizations in this pass of loop opts.. 1024 if (phi->in(LoopNode::LoopBackControl) == phi) { 1025 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl)); 1026 n_loop->_body.yank(phi); 1027 } 1028 } 1029 } 1030 } 1031 } 1032 } 1033 } 1034 1035 // If UseArrayMarkWordCheck is enabled, we can't use immutable memory for the flat array check 1036 // because we are loading the mark word which is mutable. Although the bits we are interested in 1037 // are immutable (we check for markWord::unlocked_value), we need to use raw memory to not break 1038 // anti dependency analysis. Below code will attempt to still move flat array checks out of loops, 1039 // mainly to enable loop unswitching. 1040 void PhaseIdealLoop::move_flat_array_check_out_of_loop(Node* n) { 1041 // Skip checks for more than one array 1042 if (n->req() > 3) { 1043 return; 1044 } 1045 Node* mem = n->in(FlatArrayCheckNode::Memory); 1046 Node* array = n->in(FlatArrayCheckNode::ArrayOrKlass)->uncast(); 1047 IdealLoopTree* check_loop = get_loop(get_ctrl(n)); 1048 IdealLoopTree* ary_loop = get_loop(get_ctrl(array)); 1049 1050 // Check if array is loop invariant 1051 if (!check_loop->is_member(ary_loop)) { 1052 // Walk up memory graph from the check until we leave the loop 1053 VectorSet wq; 1054 wq.set(mem->_idx); 1055 while (check_loop->is_member(get_loop(ctrl_or_self(mem)))) { 1056 if (mem->is_Phi()) { 1057 mem = mem->in(1); 1058 } else if (mem->is_MergeMem()) { 1059 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw); 1060 } else if (mem->is_Proj()) { 1061 mem = mem->in(0); 1062 } else if (mem->is_MemBar() || mem->is_SafePoint()) { 1063 mem = mem->in(TypeFunc::Memory); 1064 } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) { 1065 mem = mem->in(MemNode::Memory); 1066 } else { 1067 #ifdef ASSERT 1068 mem->dump(); 1069 #endif 1070 ShouldNotReachHere(); 1071 } 1072 if (wq.test_set(mem->_idx)) { 1073 return; 1074 } 1075 } 1076 // Replace memory input and re-compute ctrl to move the check out of the loop 1077 _igvn.replace_input_of(n, 1, mem); 1078 set_ctrl_and_loop(n, get_early_ctrl(n)); 1079 Node* bol = n->unique_out(); 1080 set_ctrl_and_loop(bol, get_early_ctrl(bol)); 1081 } 1082 } 1083 1084 //------------------------------split_if_with_blocks_pre----------------------- 1085 // Do the real work in a non-recursive function. Data nodes want to be 1086 // cloned in the pre-order so they can feed each other nicely. 1087 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) { 1088 // Cloning these guys is unlikely to win 1089 int n_op = n->Opcode(); 1090 if (n_op == Op_MergeMem) { 1091 return n; 1092 } 1093 if (n->is_Proj()) { 1094 return n; 1095 } 1096 1097 if (UseArrayMarkWordCheck && n->isa_FlatArrayCheck()) { 1098 move_flat_array_check_out_of_loop(n); 1099 return n; 1100 } 1101 1102 // Do not clone-up CmpFXXX variations, as these are always 1103 // followed by a CmpI 1104 if (n->is_Cmp()) { 1105 return n; 1106 } 1107 // Attempt to use a conditional move instead of a phi/branch 1108 if (ConditionalMoveLimit > 0 && n_op == Op_Region) { 1109 Node *cmov = conditional_move( n ); 1110 if (cmov) { 1111 return cmov; 1112 } 1113 } 1114 if (n->is_CFG() || n->is_LoadStore()) { 1115 return n; 1116 } 1117 if (n->is_Opaque1()) { // Opaque nodes cannot be mod'd 1118 if (!C->major_progress()) { // If chance of no more loop opts... 1119 _igvn._worklist.push(n); // maybe we'll remove them 1120 } 1121 return n; 1122 } 1123 1124 if (n->is_Con()) { 1125 return n; // No cloning for Con nodes 1126 } 1127 1128 Node *n_ctrl = get_ctrl(n); 1129 if (!n_ctrl) { 1130 return n; // Dead node 1131 } 1132 1133 Node* res = try_move_store_before_loop(n, n_ctrl); 1134 if (res != nullptr) { 1135 return n; 1136 } 1137 1138 // Attempt to remix address expressions for loop invariants 1139 Node *m = remix_address_expressions( n ); 1140 if( m ) return m; 1141 1142 if (n_op == Op_AddI) { 1143 Node *nn = convert_add_to_muladd( n ); 1144 if ( nn ) return nn; 1145 } 1146 1147 if (n->is_ConstraintCast()) { 1148 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this); 1149 // ConstraintCastNode::dominating_cast() uses node control input to determine domination. 1150 // Node control inputs don't necessarily agree with loop control info (due to 1151 // transformations happened in between), thus additional dominance check is needed 1152 // to keep loop info valid. 1153 if (dom_cast != nullptr && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) { 1154 _igvn.replace_node(n, dom_cast); 1155 return dom_cast; 1156 } 1157 } 1158 1159 // Determine if the Node has inputs from some local Phi. 1160 // Returns the block to clone thru. 1161 Node *n_blk = has_local_phi_input( n ); 1162 if( !n_blk ) return n; 1163 1164 // Do not clone the trip counter through on a CountedLoop 1165 // (messes up the canonical shape). 1166 if (((n_blk->is_CountedLoop() || (n_blk->is_Loop() && n_blk->as_Loop()->is_loop_nest_inner_loop())) && n->Opcode() == Op_AddI) || 1167 (n_blk->is_LongCountedLoop() && n->Opcode() == Op_AddL)) { 1168 return n; 1169 } 1170 // Pushing a shift through the iv Phi can get in the way of addressing optimizations or range check elimination 1171 if (n_blk->is_BaseCountedLoop() && n->Opcode() == Op_LShift(n_blk->as_BaseCountedLoop()->bt()) && 1172 n->in(1) == n_blk->as_BaseCountedLoop()->phi()) { 1173 return n; 1174 } 1175 1176 // Check for having no control input; not pinned. Allow 1177 // dominating control. 1178 if (n->in(0)) { 1179 Node *dom = idom(n_blk); 1180 if (dom_lca(n->in(0), dom) != n->in(0)) { 1181 return n; 1182 } 1183 } 1184 // Policy: when is it profitable. You must get more wins than 1185 // policy before it is considered profitable. Policy is usually 0, 1186 // so 1 win is considered profitable. Big merges will require big 1187 // cloning, so get a larger policy. 1188 int policy = n_blk->req() >> 2; 1189 1190 // If the loop is a candidate for range check elimination, 1191 // delay splitting through it's phi until a later loop optimization 1192 if (n_blk->is_BaseCountedLoop()) { 1193 IdealLoopTree *lp = get_loop(n_blk); 1194 if (lp && lp->_rce_candidate) { 1195 return n; 1196 } 1197 } 1198 1199 if (must_throttle_split_if()) return n; 1200 1201 // Split 'n' through the merge point if it is profitable 1202 Node *phi = split_thru_phi( n, n_blk, policy ); 1203 if (!phi) return n; 1204 1205 // Found a Phi to split thru! 1206 // Replace 'n' with the new phi 1207 _igvn.replace_node( n, phi ); 1208 // Moved a load around the loop, 'en-registering' something. 1209 if (n_blk->is_Loop() && n->is_Load() && 1210 !phi->in(LoopNode::LoopBackControl)->is_Load()) 1211 C->set_major_progress(); 1212 1213 return phi; 1214 } 1215 1216 static bool merge_point_too_heavy(Compile* C, Node* region) { 1217 // Bail out if the region and its phis have too many users. 1218 int weight = 0; 1219 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1220 weight += region->fast_out(i)->outcnt(); 1221 } 1222 int nodes_left = C->max_node_limit() - C->live_nodes(); 1223 if (weight * 8 > nodes_left) { 1224 if (PrintOpto) { 1225 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight); 1226 } 1227 return true; 1228 } else { 1229 return false; 1230 } 1231 } 1232 1233 static bool merge_point_safe(Node* region) { 1234 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode 1235 // having a PhiNode input. This sidesteps the dangerous case where the split 1236 // ConvI2LNode may become TOP if the input Value() does not 1237 // overlap the ConvI2L range, leaving a node which may not dominate its 1238 // uses. 1239 // A better fix for this problem can be found in the BugTraq entry, but 1240 // expediency for Mantis demands this hack. 1241 #ifdef _LP64 1242 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1243 Node* n = region->fast_out(i); 1244 if (n->is_Phi()) { 1245 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1246 Node* m = n->fast_out(j); 1247 if (m->Opcode() == Op_ConvI2L) 1248 return false; 1249 if (m->is_CastII()) { 1250 return false; 1251 } 1252 } 1253 } 1254 } 1255 #endif 1256 return true; 1257 } 1258 1259 1260 //------------------------------place_outside_loop--------------------------------- 1261 // Place some computation outside of this loop on the path to the use passed as argument 1262 Node* PhaseIdealLoop::place_outside_loop(Node* useblock, IdealLoopTree* loop) const { 1263 Node* head = loop->_head; 1264 assert(!loop->is_member(get_loop(useblock)), "must be outside loop"); 1265 if (head->is_Loop() && head->as_Loop()->is_strip_mined()) { 1266 loop = loop->_parent; 1267 assert(loop->_head->is_OuterStripMinedLoop(), "malformed strip mined loop"); 1268 } 1269 1270 // Pick control right outside the loop 1271 for (;;) { 1272 Node* dom = idom(useblock); 1273 if (loop->is_member(get_loop(dom)) || 1274 // NeverBranch nodes are not assigned to the loop when constructed 1275 (dom->is_NeverBranch() && loop->is_member(get_loop(dom->in(0))))) { 1276 break; 1277 } 1278 useblock = dom; 1279 } 1280 assert(find_non_split_ctrl(useblock) == useblock, "should be non split control"); 1281 return useblock; 1282 } 1283 1284 1285 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) { 1286 if (!n->is_If() || n->is_BaseCountedLoopEnd()) { 1287 return false; 1288 } 1289 if (!n->in(0)->is_Region()) { 1290 return false; 1291 } 1292 1293 Node* region = n->in(0); 1294 Node* dom = idom(region); 1295 if (!dom->is_If() || dom->in(1) != n->in(1)) { 1296 return false; 1297 } 1298 IfNode* dom_if = dom->as_If(); 1299 Node* proj_true = dom_if->proj_out(1); 1300 Node* proj_false = dom_if->proj_out(0); 1301 1302 for (uint i = 1; i < region->req(); i++) { 1303 if (is_dominator(proj_true, region->in(i))) { 1304 continue; 1305 } 1306 if (is_dominator(proj_false, region->in(i))) { 1307 continue; 1308 } 1309 return false; 1310 } 1311 1312 return true; 1313 } 1314 1315 1316 bool PhaseIdealLoop::can_split_if(Node* n_ctrl) { 1317 if (must_throttle_split_if()) { 1318 return false; 1319 } 1320 1321 // Do not do 'split-if' if irreducible loops are present. 1322 if (_has_irreducible_loops) { 1323 return false; 1324 } 1325 1326 if (merge_point_too_heavy(C, n_ctrl)) { 1327 return false; 1328 } 1329 1330 // Do not do 'split-if' if some paths are dead. First do dead code 1331 // elimination and then see if its still profitable. 1332 for (uint i = 1; i < n_ctrl->req(); i++) { 1333 if (n_ctrl->in(i) == C->top()) { 1334 return false; 1335 } 1336 } 1337 1338 // If trying to do a 'Split-If' at the loop head, it is only 1339 // profitable if the cmp folds up on BOTH paths. Otherwise we 1340 // risk peeling a loop forever. 1341 1342 // CNC - Disabled for now. Requires careful handling of loop 1343 // body selection for the cloned code. Also, make sure we check 1344 // for any input path not being in the same loop as n_ctrl. For 1345 // irreducible loops we cannot check for 'n_ctrl->is_Loop()' 1346 // because the alternative loop entry points won't be converted 1347 // into LoopNodes. 1348 IdealLoopTree *n_loop = get_loop(n_ctrl); 1349 for (uint j = 1; j < n_ctrl->req(); j++) { 1350 if (get_loop(n_ctrl->in(j)) != n_loop) { 1351 return false; 1352 } 1353 } 1354 1355 // Check for safety of the merge point. 1356 if (!merge_point_safe(n_ctrl)) { 1357 return false; 1358 } 1359 1360 return true; 1361 } 1362 1363 // Detect if the node is the inner strip-mined loop 1364 // Return: null if it's not the case, or the exit of outer strip-mined loop 1365 static Node* is_inner_of_stripmined_loop(const Node* out) { 1366 Node* out_le = nullptr; 1367 1368 if (out->is_CountedLoopEnd()) { 1369 const CountedLoopNode* loop = out->as_CountedLoopEnd()->loopnode(); 1370 1371 if (loop != nullptr && loop->is_strip_mined()) { 1372 out_le = loop->in(LoopNode::EntryControl)->as_OuterStripMinedLoop()->outer_loop_exit(); 1373 } 1374 } 1375 1376 return out_le; 1377 } 1378 1379 bool PhaseIdealLoop::flatten_array_element_type_check(Node *n) { 1380 // If the CmpP is a subtype check for a value that has just been 1381 // loaded from an array, the subtype check guarantees the value 1382 // can't be stored in a flattened array and the load of the value 1383 // happens with a flattened array check then: push the type check 1384 // through the phi of the flattened array check. This needs special 1385 // logic because the subtype check's input is not a phi but a 1386 // LoadKlass that must first be cloned through the phi. 1387 if (n->Opcode() != Op_CmpP) { 1388 return false; 1389 } 1390 1391 Node* klassptr = n->in(1); 1392 Node* klasscon = n->in(2); 1393 1394 if (klassptr->is_DecodeNarrowPtr()) { 1395 klassptr = klassptr->in(1); 1396 } 1397 1398 if (klassptr->Opcode() != Op_LoadKlass && klassptr->Opcode() != Op_LoadNKlass) { 1399 return false; 1400 } 1401 1402 if (!klasscon->is_Con()) { 1403 return false; 1404 } 1405 1406 Node* addr = klassptr->in(MemNode::Address); 1407 1408 if (!addr->is_AddP()) { 1409 return false; 1410 } 1411 1412 intptr_t offset; 1413 Node* obj = AddPNode::Ideal_base_and_offset(addr, &_igvn, offset); 1414 1415 if (obj == NULL) { 1416 return false; 1417 } 1418 1419 assert(obj != NULL && addr->in(AddPNode::Base) == addr->in(AddPNode::Address), "malformed AddP?"); 1420 if (obj->Opcode() == Op_CastPP) { 1421 obj = obj->in(1); 1422 } 1423 1424 if (!obj->is_Phi()) { 1425 return false; 1426 } 1427 1428 Node* region = obj->in(0); 1429 1430 Node* phi = PhiNode::make_blank(region, n->in(1)); 1431 for (uint i = 1; i < region->req(); i++) { 1432 Node* in = obj->in(i); 1433 Node* ctrl = region->in(i); 1434 if (addr->in(AddPNode::Base) != obj) { 1435 Node* cast = addr->in(AddPNode::Base); 1436 assert(cast->Opcode() == Op_CastPP && cast->in(0) != NULL, "inconsistent subgraph"); 1437 Node* cast_clone = cast->clone(); 1438 cast_clone->set_req(0, ctrl); 1439 cast_clone->set_req(1, in); 1440 register_new_node(cast_clone, ctrl); 1441 const Type* tcast = cast_clone->Value(&_igvn); 1442 _igvn.set_type(cast_clone, tcast); 1443 cast_clone->as_Type()->set_type(tcast); 1444 in = cast_clone; 1445 } 1446 Node* addr_clone = addr->clone(); 1447 addr_clone->set_req(AddPNode::Base, in); 1448 addr_clone->set_req(AddPNode::Address, in); 1449 register_new_node(addr_clone, ctrl); 1450 _igvn.set_type(addr_clone, addr_clone->Value(&_igvn)); 1451 Node* klassptr_clone = klassptr->clone(); 1452 klassptr_clone->set_req(2, addr_clone); 1453 register_new_node(klassptr_clone, ctrl); 1454 _igvn.set_type(klassptr_clone, klassptr_clone->Value(&_igvn)); 1455 if (klassptr != n->in(1)) { 1456 Node* decode = n->in(1); 1457 assert(decode->is_DecodeNarrowPtr(), "inconsistent subgraph"); 1458 Node* decode_clone = decode->clone(); 1459 decode_clone->set_req(1, klassptr_clone); 1460 register_new_node(decode_clone, ctrl); 1461 _igvn.set_type(decode_clone, decode_clone->Value(&_igvn)); 1462 klassptr_clone = decode_clone; 1463 } 1464 phi->set_req(i, klassptr_clone); 1465 } 1466 register_new_node(phi, region); 1467 Node* orig = n->in(1); 1468 _igvn.replace_input_of(n, 1, phi); 1469 split_if_with_blocks_post(n); 1470 if (n->outcnt() != 0) { 1471 _igvn.replace_input_of(n, 1, orig); 1472 _igvn.remove_dead_node(phi); 1473 } 1474 return true; 1475 } 1476 1477 //------------------------------split_if_with_blocks_post---------------------- 1478 // Do the real work in a non-recursive function. CFG hackery wants to be 1479 // in the post-order, so it can dirty the I-DOM info and not use the dirtied 1480 // info. 1481 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) { 1482 1483 if (flatten_array_element_type_check(n)) { 1484 return; 1485 } 1486 1487 // Cloning Cmp through Phi's involves the split-if transform. 1488 // FastLock is not used by an If 1489 if (n->is_Cmp() && !n->is_FastLock()) { 1490 Node *n_ctrl = get_ctrl(n); 1491 // Determine if the Node has inputs from some local Phi. 1492 // Returns the block to clone thru. 1493 Node *n_blk = has_local_phi_input(n); 1494 if (n_blk != n_ctrl) { 1495 return; 1496 } 1497 1498 if (!can_split_if(n_ctrl)) { 1499 return; 1500 } 1501 1502 if (n->outcnt() != 1) { 1503 return; // Multiple bool's from 1 compare? 1504 } 1505 Node *bol = n->unique_out(); 1506 assert(bol->is_Bool(), "expect a bool here"); 1507 if (bol->outcnt() != 1) { 1508 return;// Multiple branches from 1 compare? 1509 } 1510 Node *iff = bol->unique_out(); 1511 1512 // Check some safety conditions 1513 if (iff->is_If()) { // Classic split-if? 1514 if (iff->in(0) != n_ctrl) { 1515 return; // Compare must be in same blk as if 1516 } 1517 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE 1518 // Can't split CMove with different control. 1519 if (get_ctrl(iff) != n_ctrl) { 1520 return; 1521 } 1522 if (get_ctrl(iff->in(2)) == n_ctrl || 1523 get_ctrl(iff->in(3)) == n_ctrl) { 1524 return; // Inputs not yet split-up 1525 } 1526 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) { 1527 return; // Loop-invar test gates loop-varying CMOVE 1528 } 1529 } else { 1530 return; // some other kind of node, such as an Allocate 1531 } 1532 1533 // When is split-if profitable? Every 'win' on means some control flow 1534 // goes dead, so it's almost always a win. 1535 int policy = 0; 1536 // Split compare 'n' through the merge point if it is profitable 1537 Node *phi = split_thru_phi( n, n_ctrl, policy); 1538 if (!phi) { 1539 return; 1540 } 1541 1542 // Found a Phi to split thru! 1543 // Replace 'n' with the new phi 1544 _igvn.replace_node(n, phi); 1545 1546 // Now split the bool up thru the phi 1547 Node *bolphi = split_thru_phi(bol, n_ctrl, -1); 1548 guarantee(bolphi != nullptr, "null boolean phi node"); 1549 1550 _igvn.replace_node(bol, bolphi); 1551 assert(iff->in(1) == bolphi, ""); 1552 1553 if (bolphi->Value(&_igvn)->singleton()) { 1554 return; 1555 } 1556 1557 // Conditional-move? Must split up now 1558 if (!iff->is_If()) { 1559 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1); 1560 _igvn.replace_node(iff, cmovphi); 1561 return; 1562 } 1563 1564 // Now split the IF 1565 do_split_if(iff); 1566 return; 1567 } 1568 1569 // Two identical ifs back to back can be merged 1570 if (try_merge_identical_ifs(n)) { 1571 return; 1572 } 1573 1574 // Check for an IF ready to split; one that has its 1575 // condition codes input coming from a Phi at the block start. 1576 int n_op = n->Opcode(); 1577 1578 // Check for an IF being dominated by another IF same test 1579 if (n_op == Op_If || 1580 n_op == Op_RangeCheck) { 1581 Node *bol = n->in(1); 1582 uint max = bol->outcnt(); 1583 // Check for same test used more than once? 1584 if (max > 1 && bol->is_Bool()) { 1585 // Search up IDOMs to see if this IF is dominated. 1586 Node *cutoff = get_ctrl(bol); 1587 1588 // Now search up IDOMs till cutoff, looking for a dominating test 1589 Node *prevdom = n; 1590 Node *dom = idom(prevdom); 1591 while (dom != cutoff) { 1592 if (dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom && 1593 safe_for_if_replacement(dom)) { 1594 // It's invalid to move control dependent data nodes in the inner 1595 // strip-mined loop, because: 1596 // 1) break validation of LoopNode::verify_strip_mined() 1597 // 2) move code with side-effect in strip-mined loop 1598 // Move to the exit of outer strip-mined loop in that case. 1599 Node* out_le = is_inner_of_stripmined_loop(dom); 1600 if (out_le != nullptr) { 1601 prevdom = out_le; 1602 } 1603 // Replace the dominated test with an obvious true or false. 1604 // Place it on the IGVN worklist for later cleanup. 1605 C->set_major_progress(); 1606 dominated_by(prevdom->as_IfProj(), n->as_If(), false, true); 1607 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } ); 1608 return; 1609 } 1610 prevdom = dom; 1611 dom = idom(prevdom); 1612 } 1613 } 1614 } 1615 1616 try_sink_out_of_loop(n); 1617 1618 try_move_store_after_loop(n); 1619 1620 // Remove multiple allocations of the same inline type 1621 if (n->is_InlineType()) { 1622 n->as_InlineType()->remove_redundant_allocations(this); 1623 } 1624 } 1625 1626 // Transform: 1627 // 1628 // if (some_condition) { 1629 // // body 1 1630 // } else { 1631 // // body 2 1632 // } 1633 // if (some_condition) { 1634 // // body 3 1635 // } else { 1636 // // body 4 1637 // } 1638 // 1639 // into: 1640 // 1641 // 1642 // if (some_condition) { 1643 // // body 1 1644 // // body 3 1645 // } else { 1646 // // body 2 1647 // // body 4 1648 // } 1649 bool PhaseIdealLoop::try_merge_identical_ifs(Node* n) { 1650 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) { 1651 Node *n_ctrl = n->in(0); 1652 IfNode* dom_if = idom(n_ctrl)->as_If(); 1653 ProjNode* dom_proj_true = dom_if->proj_out(1); 1654 ProjNode* dom_proj_false = dom_if->proj_out(0); 1655 1656 // Now split the IF 1657 RegionNode* new_false_region; 1658 RegionNode* new_true_region; 1659 do_split_if(n, &new_false_region, &new_true_region); 1660 assert(new_false_region->req() == new_true_region->req(), ""); 1661 #ifdef ASSERT 1662 for (uint i = 1; i < new_false_region->req(); ++i) { 1663 assert(new_false_region->in(i)->in(0) == new_true_region->in(i)->in(0), "unexpected shape following split if"); 1664 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"); 1665 } 1666 #endif 1667 assert(new_false_region->in(1)->in(0)->in(1) == dom_if->in(1), "dominating if and dominated if after split must share test"); 1668 1669 // We now have: 1670 // if (some_condition) { 1671 // // body 1 1672 // if (some_condition) { 1673 // body3: // new_true_region 1674 // // body3 1675 // } else { 1676 // goto body4; 1677 // } 1678 // } else { 1679 // // body 2 1680 // if (some_condition) { 1681 // goto body3; 1682 // } else { 1683 // body4: // new_false_region 1684 // // body4; 1685 // } 1686 // } 1687 // 1688 1689 // clone pinned nodes thru the resulting regions 1690 push_pinned_nodes_thru_region(dom_if, new_true_region); 1691 push_pinned_nodes_thru_region(dom_if, new_false_region); 1692 1693 // Optimize out the cloned ifs. Because pinned nodes were cloned, this also allows a CastPP that would be dependent 1694 // 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 1695 // unrelated control dependency. 1696 for (uint i = 1; i < new_false_region->req(); i++) { 1697 if (is_dominator(dom_proj_true, new_false_region->in(i))) { 1698 dominated_by(dom_proj_true->as_IfProj(), new_false_region->in(i)->in(0)->as_If(), false, false); 1699 } else { 1700 assert(is_dominator(dom_proj_false, new_false_region->in(i)), "bad if"); 1701 dominated_by(dom_proj_false->as_IfProj(), new_false_region->in(i)->in(0)->as_If(), false, false); 1702 } 1703 } 1704 return true; 1705 } 1706 return false; 1707 } 1708 1709 void PhaseIdealLoop::push_pinned_nodes_thru_region(IfNode* dom_if, Node* region) { 1710 for (DUIterator i = region->outs(); region->has_out(i); i++) { 1711 Node* u = region->out(i); 1712 if (!has_ctrl(u) || u->is_Phi() || !u->depends_only_on_test() || !_igvn.no_dependent_zero_check(u)) { 1713 continue; 1714 } 1715 assert(u->in(0) == region, "not a control dependent node?"); 1716 uint j = 1; 1717 for (; j < u->req(); ++j) { 1718 Node* in = u->in(j); 1719 if (!is_dominator(ctrl_or_self(in), dom_if)) { 1720 break; 1721 } 1722 } 1723 if (j == u->req()) { 1724 Node *phi = PhiNode::make_blank(region, u); 1725 for (uint k = 1; k < region->req(); ++k) { 1726 Node* clone = u->clone(); 1727 clone->set_req(0, region->in(k)); 1728 register_new_node(clone, region->in(k)); 1729 phi->init_req(k, clone); 1730 } 1731 register_new_node(phi, region); 1732 _igvn.replace_node(u, phi); 1733 --i; 1734 } 1735 } 1736 } 1737 1738 bool PhaseIdealLoop::safe_for_if_replacement(const Node* dom) const { 1739 if (!dom->is_CountedLoopEnd()) { 1740 return true; 1741 } 1742 CountedLoopEndNode* le = dom->as_CountedLoopEnd(); 1743 CountedLoopNode* cl = le->loopnode(); 1744 if (cl == nullptr) { 1745 return true; 1746 } 1747 if (!cl->is_main_loop()) { 1748 return true; 1749 } 1750 if (cl->is_canonical_loop_entry() == nullptr) { 1751 return true; 1752 } 1753 // Further unrolling is possible so loop exit condition might change 1754 return false; 1755 } 1756 1757 // See if a shared loop-varying computation has no loop-varying uses. 1758 // Happens if something is only used for JVM state in uncommon trap exits, 1759 // like various versions of induction variable+offset. Clone the 1760 // computation per usage to allow it to sink out of the loop. 1761 void PhaseIdealLoop::try_sink_out_of_loop(Node* n) { 1762 if (has_ctrl(n) && 1763 !n->is_Phi() && 1764 !n->is_Bool() && 1765 !n->is_Proj() && 1766 !n->is_MergeMem() && 1767 !n->is_CMove() && 1768 n->Opcode() != Op_Opaque4 && 1769 !n->is_Type()) { 1770 Node *n_ctrl = get_ctrl(n); 1771 IdealLoopTree *n_loop = get_loop(n_ctrl); 1772 1773 if (n->in(0) != nullptr) { 1774 IdealLoopTree* loop_ctrl = get_loop(n->in(0)); 1775 if (n_loop != loop_ctrl && n_loop->is_member(loop_ctrl)) { 1776 // n has a control input inside a loop but get_ctrl() is member of an outer loop. This could happen, for example, 1777 // for Div nodes inside a loop (control input inside loop) without a use except for an UCT (outside the loop). 1778 // Rewire control of n to right outside of the loop, regardless if its input(s) are later sunk or not. 1779 _igvn.replace_input_of(n, 0, place_outside_loop(n_ctrl, loop_ctrl)); 1780 } 1781 } 1782 if (n_loop != _ltree_root && n->outcnt() > 1) { 1783 // Compute early control: needed for anti-dependence analysis. It's also possible that as a result of 1784 // previous transformations in this loop opts round, the node can be hoisted now: early control will tell us. 1785 Node* early_ctrl = compute_early_ctrl(n, n_ctrl); 1786 if (n_loop->is_member(get_loop(early_ctrl)) && // check that this one can't be hoisted now 1787 ctrl_of_all_uses_out_of_loop(n, early_ctrl, n_loop)) { // All uses in outer loops! 1788 assert(!n->is_Store() && !n->is_LoadStore(), "no node with a side effect"); 1789 Node* outer_loop_clone = nullptr; 1790 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin;) { 1791 Node* u = n->last_out(j); // Clone private computation per use 1792 _igvn.rehash_node_delayed(u); 1793 Node* x = n->clone(); // Clone computation 1794 Node* x_ctrl = nullptr; 1795 if (u->is_Phi()) { 1796 // Replace all uses of normal nodes. Replace Phi uses 1797 // individually, so the separate Nodes can sink down 1798 // different paths. 1799 uint k = 1; 1800 while (u->in(k) != n) k++; 1801 u->set_req(k, x); 1802 // x goes next to Phi input path 1803 x_ctrl = u->in(0)->in(k); 1804 // Find control for 'x' next to use but not inside inner loops. 1805 x_ctrl = place_outside_loop(x_ctrl, n_loop); 1806 --j; 1807 } else { // Normal use 1808 if (has_ctrl(u)) { 1809 x_ctrl = get_ctrl(u); 1810 } else { 1811 x_ctrl = u->in(0); 1812 } 1813 // Find control for 'x' next to use but not inside inner loops. 1814 x_ctrl = place_outside_loop(x_ctrl, n_loop); 1815 // Replace all uses 1816 if (u->is_ConstraintCast() && u->bottom_type()->higher_equal(_igvn.type(n)) && u->in(0) == x_ctrl) { 1817 // If we're sinking a chain of data nodes, we might have inserted a cast to pin the use which is not necessary 1818 // anymore now that we're going to pin n as well 1819 _igvn.replace_node(u, x); 1820 --j; 1821 } else { 1822 int nb = u->replace_edge(n, x, &_igvn); 1823 j -= nb; 1824 } 1825 } 1826 1827 if (n->is_Load()) { 1828 // For loads, add a control edge to a CFG node outside of the loop 1829 // to force them to not combine and return back inside the loop 1830 // during GVN optimization (4641526). 1831 assert(x_ctrl == get_late_ctrl_with_anti_dep(x->as_Load(), early_ctrl, x_ctrl), "anti-dependences were already checked"); 1832 1833 IdealLoopTree* x_loop = get_loop(x_ctrl); 1834 Node* x_head = x_loop->_head; 1835 if (x_head->is_Loop() && x_head->is_OuterStripMinedLoop()) { 1836 // Do not add duplicate LoadNodes to the outer strip mined loop 1837 if (outer_loop_clone != nullptr) { 1838 _igvn.replace_node(x, outer_loop_clone); 1839 continue; 1840 } 1841 outer_loop_clone = x; 1842 } 1843 x->set_req(0, x_ctrl); 1844 } else if (n->in(0) != nullptr){ 1845 x->set_req(0, x_ctrl); 1846 } 1847 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone"); 1848 assert(!n_loop->is_member(get_loop(x_ctrl)), "should have moved out of loop"); 1849 register_new_node(x, x_ctrl); 1850 1851 // Chain of AddP: (AddP base (AddP base )) must keep the same base after sinking so: 1852 // 1- We don't add a CastPP here when the first one is sunk so if the second one is not, their bases remain 1853 // the same. 1854 // (see 2- below) 1855 assert(!x->is_AddP() || !x->in(AddPNode::Address)->is_AddP() || 1856 x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base) || 1857 !x->in(AddPNode::Address)->in(AddPNode::Base)->eqv_uncast(x->in(AddPNode::Base)), "unexpected AddP shape"); 1858 if (x->in(0) == nullptr && !x->is_DecodeNarrowPtr() && 1859 !(x->is_AddP() && x->in(AddPNode::Address)->is_AddP() && x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base))) { 1860 assert(!x->is_Load(), "load should be pinned"); 1861 // Use a cast node to pin clone out of loop 1862 Node* cast = nullptr; 1863 for (uint k = 0; k < x->req(); k++) { 1864 Node* in = x->in(k); 1865 if (in != nullptr && n_loop->is_member(get_loop(get_ctrl(in)))) { 1866 const Type* in_t = _igvn.type(in); 1867 cast = ConstraintCastNode::make_cast_for_type(x_ctrl, in, in_t, ConstraintCastNode::UnconditionalDependency); 1868 } 1869 if (cast != nullptr) { 1870 register_new_node(cast, x_ctrl); 1871 x->replace_edge(in, cast); 1872 // Chain of AddP: 1873 // 2- A CastPP of the base is only added now that both AddP nodes are sunk 1874 if (x->is_AddP() && k == AddPNode::Base) { 1875 for (DUIterator_Fast imax, i = x->fast_outs(imax); i < imax; i++) { 1876 Node* u = x->fast_out(i); 1877 if (u->is_AddP() && u->in(AddPNode::Base) == n->in(AddPNode::Base)) { 1878 _igvn.replace_input_of(u, AddPNode::Base, cast); 1879 assert(u->find_out_with(Op_AddP) == nullptr, "more than 2 chained AddP nodes?"); 1880 } 1881 } 1882 } 1883 break; 1884 } 1885 } 1886 assert(cast != nullptr, "must have added a cast to pin the node"); 1887 } 1888 } 1889 _igvn.remove_dead_node(n); 1890 } 1891 _dom_lca_tags_round = 0; 1892 } 1893 } 1894 } 1895 1896 // Compute the early control of a node by following its inputs until we reach 1897 // nodes that are pinned. Then compute the LCA of the control of all pinned nodes. 1898 Node* PhaseIdealLoop::compute_early_ctrl(Node* n, Node* n_ctrl) { 1899 Node* early_ctrl = nullptr; 1900 ResourceMark rm; 1901 Unique_Node_List wq; 1902 wq.push(n); 1903 for (uint i = 0; i < wq.size(); i++) { 1904 Node* m = wq.at(i); 1905 Node* c = nullptr; 1906 if (m->is_CFG()) { 1907 c = m; 1908 } else if (m->pinned()) { 1909 c = m->in(0); 1910 } else { 1911 for (uint j = 0; j < m->req(); j++) { 1912 Node* in = m->in(j); 1913 if (in != nullptr) { 1914 wq.push(in); 1915 } 1916 } 1917 } 1918 if (c != nullptr) { 1919 assert(is_dominator(c, n_ctrl), "control input must dominate current control"); 1920 if (early_ctrl == nullptr || is_dominator(early_ctrl, c)) { 1921 early_ctrl = c; 1922 } 1923 } 1924 } 1925 assert(is_dominator(early_ctrl, n_ctrl), "early control must dominate current control"); 1926 return early_ctrl; 1927 } 1928 1929 bool PhaseIdealLoop::ctrl_of_all_uses_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop) { 1930 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1931 Node* u = n->fast_out(i); 1932 if (u->is_Opaque1()) { 1933 return false; // Found loop limit, bugfix for 4677003 1934 } 1935 // We can't reuse tags in PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal() so make sure calls to 1936 // get_late_ctrl_with_anti_dep() use their own tag 1937 _dom_lca_tags_round++; 1938 assert(_dom_lca_tags_round != 0, "shouldn't wrap around"); 1939 1940 if (u->is_Phi()) { 1941 for (uint j = 1; j < u->req(); ++j) { 1942 if (u->in(j) == n && !ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, u->in(0)->in(j))) { 1943 return false; 1944 } 1945 } 1946 } else { 1947 Node* ctrl = has_ctrl(u) ? get_ctrl(u) : u->in(0); 1948 if (!ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, ctrl)) { 1949 return false; 1950 } 1951 } 1952 } 1953 return true; 1954 } 1955 1956 bool PhaseIdealLoop::ctrl_of_use_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop, Node* ctrl) { 1957 if (n->is_Load()) { 1958 ctrl = get_late_ctrl_with_anti_dep(n->as_Load(), n_ctrl, ctrl); 1959 } 1960 IdealLoopTree *u_loop = get_loop(ctrl); 1961 if (u_loop == n_loop) { 1962 return false; // Found loop-varying use 1963 } 1964 if (n_loop->is_member(u_loop)) { 1965 return false; // Found use in inner loop 1966 } 1967 return true; 1968 } 1969 1970 //------------------------------split_if_with_blocks--------------------------- 1971 // Check for aggressive application of 'split-if' optimization, 1972 // using basic block level info. 1973 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack) { 1974 Node* root = C->root(); 1975 visited.set(root->_idx); // first, mark root as visited 1976 // Do pre-visit work for root 1977 Node* n = split_if_with_blocks_pre(root); 1978 uint cnt = n->outcnt(); 1979 uint i = 0; 1980 1981 while (true) { 1982 // Visit all children 1983 if (i < cnt) { 1984 Node* use = n->raw_out(i); 1985 ++i; 1986 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) { 1987 // Now do pre-visit work for this use 1988 use = split_if_with_blocks_pre(use); 1989 nstack.push(n, i); // Save parent and next use's index. 1990 n = use; // Process all children of current use. 1991 cnt = use->outcnt(); 1992 i = 0; 1993 } 1994 } 1995 else { 1996 // All of n's children have been processed, complete post-processing. 1997 if (cnt != 0 && !n->is_Con()) { 1998 assert(has_node(n), "no dead nodes"); 1999 split_if_with_blocks_post(n); 2000 } 2001 if (must_throttle_split_if()) { 2002 nstack.clear(); 2003 } 2004 if (nstack.is_empty()) { 2005 // Finished all nodes on stack. 2006 break; 2007 } 2008 // Get saved parent node and next use's index. Visit the rest of uses. 2009 n = nstack.node(); 2010 cnt = n->outcnt(); 2011 i = nstack.index(); 2012 nstack.pop(); 2013 } 2014 } 2015 } 2016 2017 2018 //============================================================================= 2019 // 2020 // C L O N E A L O O P B O D Y 2021 // 2022 2023 //------------------------------clone_iff-------------------------------------- 2024 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 2025 // "Nearly" because all Nodes have been cloned from the original in the loop, 2026 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 2027 // through the Phi recursively, and return a Bool. 2028 Node* PhaseIdealLoop::clone_iff(PhiNode* phi) { 2029 2030 // Convert this Phi into a Phi merging Bools 2031 uint i; 2032 for (i = 1; i < phi->req(); i++) { 2033 Node *b = phi->in(i); 2034 if (b->is_Phi()) { 2035 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi())); 2036 } else { 2037 assert(b->is_Bool() || b->Opcode() == Op_Opaque4, ""); 2038 } 2039 } 2040 2041 Node* n = phi->in(1); 2042 Node* sample_opaque = nullptr; 2043 Node *sample_bool = nullptr; 2044 if (n->Opcode() == Op_Opaque4) { 2045 sample_opaque = n; 2046 sample_bool = n->in(1); 2047 assert(sample_bool->is_Bool(), "wrong type"); 2048 } else { 2049 sample_bool = n; 2050 } 2051 Node* sample_cmp = sample_bool->in(1); 2052 const Type* t = Type::TOP; 2053 const TypePtr* at = NULL; 2054 if (sample_cmp->is_FlatArrayCheck()) { 2055 // Left input of a FlatArrayCheckNode is memory, set the (adr) type of the phi accordingly 2056 assert(sample_cmp->in(1)->bottom_type() == Type::MEMORY, "unexpected input type"); 2057 t = Type::MEMORY; 2058 at = TypeRawPtr::BOTTOM; 2059 } 2060 2061 // Make Phis to merge the Cmp's inputs. 2062 PhiNode *phi1 = new PhiNode(phi->in(0), t, at); 2063 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP); 2064 for (i = 1; i < phi->req(); i++) { 2065 Node *n1 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1); 2066 Node *n2 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2); 2067 phi1->set_req(i, n1); 2068 phi2->set_req(i, n2); 2069 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 2070 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 2071 } 2072 // See if these Phis have been made before. 2073 // Register with optimizer 2074 Node *hit1 = _igvn.hash_find_insert(phi1); 2075 if (hit1) { // Hit, toss just made Phi 2076 _igvn.remove_dead_node(phi1); // Remove new phi 2077 assert(hit1->is_Phi(), "" ); 2078 phi1 = (PhiNode*)hit1; // Use existing phi 2079 } else { // Miss 2080 _igvn.register_new_node_with_optimizer(phi1); 2081 } 2082 Node *hit2 = _igvn.hash_find_insert(phi2); 2083 if (hit2) { // Hit, toss just made Phi 2084 _igvn.remove_dead_node(phi2); // Remove new phi 2085 assert(hit2->is_Phi(), "" ); 2086 phi2 = (PhiNode*)hit2; // Use existing phi 2087 } else { // Miss 2088 _igvn.register_new_node_with_optimizer(phi2); 2089 } 2090 // Register Phis with loop/block info 2091 set_ctrl(phi1, phi->in(0)); 2092 set_ctrl(phi2, phi->in(0)); 2093 // Make a new Cmp 2094 Node *cmp = sample_cmp->clone(); 2095 cmp->set_req(1, phi1); 2096 cmp->set_req(2, phi2); 2097 _igvn.register_new_node_with_optimizer(cmp); 2098 set_ctrl(cmp, phi->in(0)); 2099 2100 // Make a new Bool 2101 Node *b = sample_bool->clone(); 2102 b->set_req(1,cmp); 2103 _igvn.register_new_node_with_optimizer(b); 2104 set_ctrl(b, phi->in(0)); 2105 2106 if (sample_opaque != nullptr) { 2107 Node* opaque = sample_opaque->clone(); 2108 opaque->set_req(1, b); 2109 _igvn.register_new_node_with_optimizer(opaque); 2110 set_ctrl(opaque, phi->in(0)); 2111 return opaque; 2112 } 2113 2114 assert(b->is_Bool(), ""); 2115 return b; 2116 } 2117 2118 //------------------------------clone_bool------------------------------------- 2119 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 2120 // "Nearly" because all Nodes have been cloned from the original in the loop, 2121 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 2122 // through the Phi recursively, and return a Bool. 2123 CmpNode*PhaseIdealLoop::clone_bool(PhiNode* phi) { 2124 uint i; 2125 // Convert this Phi into a Phi merging Bools 2126 for( i = 1; i < phi->req(); i++ ) { 2127 Node *b = phi->in(i); 2128 if( b->is_Phi() ) { 2129 _igvn.replace_input_of(phi, i, clone_bool(b->as_Phi())); 2130 } else { 2131 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" ); 2132 } 2133 } 2134 2135 Node *sample_cmp = phi->in(1); 2136 2137 // Make Phis to merge the Cmp's inputs. 2138 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP ); 2139 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP ); 2140 for( uint j = 1; j < phi->req(); j++ ) { 2141 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP 2142 Node *n1, *n2; 2143 if( cmp_top->is_Cmp() ) { 2144 n1 = cmp_top->in(1); 2145 n2 = cmp_top->in(2); 2146 } else { 2147 n1 = n2 = cmp_top; 2148 } 2149 phi1->set_req( j, n1 ); 2150 phi2->set_req( j, n2 ); 2151 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 2152 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 2153 } 2154 2155 // See if these Phis have been made before. 2156 // Register with optimizer 2157 Node *hit1 = _igvn.hash_find_insert(phi1); 2158 if( hit1 ) { // Hit, toss just made Phi 2159 _igvn.remove_dead_node(phi1); // Remove new phi 2160 assert( hit1->is_Phi(), "" ); 2161 phi1 = (PhiNode*)hit1; // Use existing phi 2162 } else { // Miss 2163 _igvn.register_new_node_with_optimizer(phi1); 2164 } 2165 Node *hit2 = _igvn.hash_find_insert(phi2); 2166 if( hit2 ) { // Hit, toss just made Phi 2167 _igvn.remove_dead_node(phi2); // Remove new phi 2168 assert( hit2->is_Phi(), "" ); 2169 phi2 = (PhiNode*)hit2; // Use existing phi 2170 } else { // Miss 2171 _igvn.register_new_node_with_optimizer(phi2); 2172 } 2173 // Register Phis with loop/block info 2174 set_ctrl(phi1, phi->in(0)); 2175 set_ctrl(phi2, phi->in(0)); 2176 // Make a new Cmp 2177 Node *cmp = sample_cmp->clone(); 2178 cmp->set_req( 1, phi1 ); 2179 cmp->set_req( 2, phi2 ); 2180 _igvn.register_new_node_with_optimizer(cmp); 2181 set_ctrl(cmp, phi->in(0)); 2182 2183 assert( cmp->is_Cmp(), "" ); 2184 return (CmpNode*)cmp; 2185 } 2186 2187 //------------------------------sink_use--------------------------------------- 2188 // If 'use' was in the loop-exit block, it now needs to be sunk 2189 // below the post-loop merge point. 2190 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) { 2191 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) { 2192 set_ctrl(use, post_loop); 2193 for (DUIterator j = use->outs(); use->has_out(j); j++) 2194 sink_use(use->out(j), post_loop); 2195 } 2196 } 2197 2198 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new, 2199 IdealLoopTree* loop, IdealLoopTree* outer_loop, 2200 Node_List*& split_if_set, Node_List*& split_bool_set, 2201 Node_List*& split_cex_set, Node_List& worklist, 2202 uint new_counter, CloneLoopMode mode) { 2203 Node* nnn = old_new[old->_idx]; 2204 // Copy uses to a worklist, so I can munge the def-use info 2205 // with impunity. 2206 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 2207 worklist.push(old->fast_out(j)); 2208 2209 while( worklist.size() ) { 2210 Node *use = worklist.pop(); 2211 if (!has_node(use)) continue; // Ignore dead nodes 2212 if (use->in(0) == C->top()) continue; 2213 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 2214 // Check for data-use outside of loop - at least one of OLD or USE 2215 // must not be a CFG node. 2216 #ifdef ASSERT 2217 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == nullptr) { 2218 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint(); 2219 assert(mode != IgnoreStripMined, "incorrect cloning mode"); 2220 assert((mode == ControlAroundStripMined && use == sfpt) || !use->is_reachable_from_root(), "missed a node"); 2221 } 2222 #endif 2223 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) { 2224 2225 // If the Data use is an IF, that means we have an IF outside of the 2226 // loop that is switching on a condition that is set inside of the 2227 // loop. Happens if people set a loop-exit flag; then test the flag 2228 // in the loop to break the loop, then test is again outside of the 2229 // loop to determine which way the loop exited. 2230 // Loop predicate If node connects to Bool node through Opaque1 node. 2231 // 2232 // If the use is an AllocateArray through its ValidLengthTest input, 2233 // make sure the Bool/Cmp input is cloned down to avoid a Phi between 2234 // the AllocateArray node and its ValidLengthTest input that could cause 2235 // split if to break. 2236 if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use) || use->Opcode() == Op_Opaque4 || 2237 (use->Opcode() == Op_AllocateArray && use->in(AllocateNode::ValidLengthTest) == old)) { 2238 // Since this code is highly unlikely, we lazily build the worklist 2239 // of such Nodes to go split. 2240 if (!split_if_set) { 2241 split_if_set = new Node_List(); 2242 } 2243 split_if_set->push(use); 2244 } 2245 if (use->is_Bool()) { 2246 if (!split_bool_set) { 2247 split_bool_set = new Node_List(); 2248 } 2249 split_bool_set->push(use); 2250 } 2251 if (use->Opcode() == Op_CreateEx) { 2252 if (!split_cex_set) { 2253 split_cex_set = new Node_List(); 2254 } 2255 split_cex_set->push(use); 2256 } 2257 2258 2259 // Get "block" use is in 2260 uint idx = 0; 2261 while( use->in(idx) != old ) idx++; 2262 Node *prev = use->is_CFG() ? use : get_ctrl(use); 2263 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" ); 2264 Node *cfg = prev->_idx >= new_counter 2265 ? prev->in(2) 2266 : idom(prev); 2267 if( use->is_Phi() ) // Phi use is in prior block 2268 cfg = prev->in(idx); // NOT in block of Phi itself 2269 if (cfg->is_top()) { // Use is dead? 2270 _igvn.replace_input_of(use, idx, C->top()); 2271 continue; 2272 } 2273 2274 // If use is referenced through control edge... (idx == 0) 2275 if (mode == IgnoreStripMined && idx == 0) { 2276 LoopNode *head = loop->_head->as_Loop(); 2277 if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) { 2278 // That node is outside the inner loop, leave it outside the 2279 // outer loop as well to not confuse verification code. 2280 assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop"); 2281 _igvn.replace_input_of(use, 0, head->outer_loop_exit()); 2282 continue; 2283 } 2284 } 2285 2286 while(!outer_loop->is_member(get_loop(cfg))) { 2287 prev = cfg; 2288 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg); 2289 } 2290 // If the use occurs after merging several exits from the loop, then 2291 // old value must have dominated all those exits. Since the same old 2292 // value was used on all those exits we did not need a Phi at this 2293 // merge point. NOW we do need a Phi here. Each loop exit value 2294 // is now merged with the peeled body exit; each exit gets its own 2295 // private Phi and those Phis need to be merged here. 2296 Node *phi; 2297 if( prev->is_Region() ) { 2298 if( idx == 0 ) { // Updating control edge? 2299 phi = prev; // Just use existing control 2300 } else { // Else need a new Phi 2301 phi = PhiNode::make( prev, old ); 2302 // Now recursively fix up the new uses of old! 2303 for( uint i = 1; i < prev->req(); i++ ) { 2304 worklist.push(phi); // Onto worklist once for each 'old' input 2305 } 2306 } 2307 } else { 2308 // Get new RegionNode merging old and new loop exits 2309 prev = old_new[prev->_idx]; 2310 assert( prev, "just made this in step 7" ); 2311 if( idx == 0) { // Updating control edge? 2312 phi = prev; // Just use existing control 2313 } else { // Else need a new Phi 2314 // Make a new Phi merging data values properly 2315 phi = PhiNode::make( prev, old ); 2316 phi->set_req( 1, nnn ); 2317 } 2318 } 2319 // If inserting a new Phi, check for prior hits 2320 if( idx != 0 ) { 2321 Node *hit = _igvn.hash_find_insert(phi); 2322 if( hit == nullptr ) { 2323 _igvn.register_new_node_with_optimizer(phi); // Register new phi 2324 } else { // or 2325 // Remove the new phi from the graph and use the hit 2326 _igvn.remove_dead_node(phi); 2327 phi = hit; // Use existing phi 2328 } 2329 set_ctrl(phi, prev); 2330 } 2331 // Make 'use' use the Phi instead of the old loop body exit value 2332 assert(use->in(idx) == old, "old is still input of use"); 2333 // We notify all uses of old, including use, and the indirect uses, 2334 // that may now be optimized because we have replaced old with phi. 2335 _igvn.add_users_to_worklist(old); 2336 _igvn.replace_input_of(use, idx, phi); 2337 if( use->_idx >= new_counter ) { // If updating new phis 2338 // Not needed for correctness, but prevents a weak assert 2339 // in AddPNode from tripping (when we end up with different 2340 // base & derived Phis that will become the same after 2341 // IGVN does CSE). 2342 Node *hit = _igvn.hash_find_insert(use); 2343 if( hit ) // Go ahead and re-hash for hits. 2344 _igvn.replace_node( use, hit ); 2345 } 2346 2347 // If 'use' was in the loop-exit block, it now needs to be sunk 2348 // below the post-loop merge point. 2349 sink_use( use, prev ); 2350 } 2351 } 2352 } 2353 2354 static void collect_nodes_in_outer_loop_not_reachable_from_sfpt(Node* n, const IdealLoopTree *loop, const IdealLoopTree* outer_loop, 2355 const Node_List &old_new, Unique_Node_List& wq, PhaseIdealLoop* phase, 2356 bool check_old_new) { 2357 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2358 Node* u = n->fast_out(j); 2359 assert(check_old_new || old_new[u->_idx] == nullptr, "shouldn't have been cloned"); 2360 if (!u->is_CFG() && (!check_old_new || old_new[u->_idx] == nullptr)) { 2361 Node* c = phase->get_ctrl(u); 2362 IdealLoopTree* u_loop = phase->get_loop(c); 2363 assert(!loop->is_member(u_loop) || !loop->_body.contains(u), "can be in outer loop or out of both loops only"); 2364 if (!loop->is_member(u_loop)) { 2365 if (outer_loop->is_member(u_loop)) { 2366 wq.push(u); 2367 } else { 2368 // nodes pinned with control in the outer loop but not referenced from the safepoint must be moved out of 2369 // the outer loop too 2370 Node* u_c = u->in(0); 2371 if (u_c != nullptr) { 2372 IdealLoopTree* u_c_loop = phase->get_loop(u_c); 2373 if (outer_loop->is_member(u_c_loop) && !loop->is_member(u_c_loop)) { 2374 wq.push(u); 2375 } 2376 } 2377 } 2378 } 2379 } 2380 } 2381 } 2382 2383 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop, 2384 IdealLoopTree* outer_loop, int dd, Node_List &old_new, 2385 Node_List& extra_data_nodes) { 2386 if (head->is_strip_mined() && mode != IgnoreStripMined) { 2387 CountedLoopNode* cl = head->as_CountedLoop(); 2388 Node* l = cl->outer_loop(); 2389 Node* tail = cl->outer_loop_tail(); 2390 IfNode* le = cl->outer_loop_end(); 2391 Node* sfpt = cl->outer_safepoint(); 2392 CountedLoopEndNode* cle = cl->loopexit(); 2393 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop(); 2394 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null(); 2395 Node* cle_out = cle->proj_out(false); 2396 2397 Node* new_sfpt = nullptr; 2398 Node* new_cle_out = cle_out->clone(); 2399 old_new.map(cle_out->_idx, new_cle_out); 2400 if (mode == CloneIncludesStripMined) { 2401 // clone outer loop body 2402 Node* new_l = l->clone(); 2403 Node* new_tail = tail->clone(); 2404 IfNode* new_le = le->clone()->as_If(); 2405 new_sfpt = sfpt->clone(); 2406 2407 set_loop(new_l, outer_loop->_parent); 2408 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd); 2409 set_loop(new_cle_out, outer_loop->_parent); 2410 set_idom(new_cle_out, new_cle, dd); 2411 set_loop(new_sfpt, outer_loop->_parent); 2412 set_idom(new_sfpt, new_cle_out, dd); 2413 set_loop(new_le, outer_loop->_parent); 2414 set_idom(new_le, new_sfpt, dd); 2415 set_loop(new_tail, outer_loop->_parent); 2416 set_idom(new_tail, new_le, dd); 2417 set_idom(new_cl, new_l, dd); 2418 2419 old_new.map(l->_idx, new_l); 2420 old_new.map(tail->_idx, new_tail); 2421 old_new.map(le->_idx, new_le); 2422 old_new.map(sfpt->_idx, new_sfpt); 2423 2424 new_l->set_req(LoopNode::LoopBackControl, new_tail); 2425 new_l->set_req(0, new_l); 2426 new_tail->set_req(0, new_le); 2427 new_le->set_req(0, new_sfpt); 2428 new_sfpt->set_req(0, new_cle_out); 2429 new_cle_out->set_req(0, new_cle); 2430 new_cl->set_req(LoopNode::EntryControl, new_l); 2431 2432 _igvn.register_new_node_with_optimizer(new_l); 2433 _igvn.register_new_node_with_optimizer(new_tail); 2434 _igvn.register_new_node_with_optimizer(new_le); 2435 } else { 2436 Node *newhead = old_new[loop->_head->_idx]; 2437 newhead->as_Loop()->clear_strip_mined(); 2438 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl)); 2439 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 2440 } 2441 // Look at data node that were assigned a control in the outer 2442 // loop: they are kept in the outer loop by the safepoint so start 2443 // from the safepoint node's inputs. 2444 IdealLoopTree* outer_loop = get_loop(l); 2445 Node_Stack stack(2); 2446 stack.push(sfpt, 1); 2447 uint new_counter = C->unique(); 2448 while (stack.size() > 0) { 2449 Node* n = stack.node(); 2450 uint i = stack.index(); 2451 while (i < n->req() && 2452 (n->in(i) == nullptr || 2453 !has_ctrl(n->in(i)) || 2454 get_loop(get_ctrl(n->in(i))) != outer_loop || 2455 (old_new[n->in(i)->_idx] != nullptr && old_new[n->in(i)->_idx]->_idx >= new_counter))) { 2456 i++; 2457 } 2458 if (i < n->req()) { 2459 stack.set_index(i+1); 2460 stack.push(n->in(i), 0); 2461 } else { 2462 assert(old_new[n->_idx] == nullptr || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet"); 2463 Node* m = n == sfpt ? new_sfpt : n->clone(); 2464 if (m != nullptr) { 2465 for (uint i = 0; i < n->req(); i++) { 2466 if (m->in(i) != nullptr && old_new[m->in(i)->_idx] != nullptr) { 2467 m->set_req(i, old_new[m->in(i)->_idx]); 2468 } 2469 } 2470 } else { 2471 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?"); 2472 } 2473 if (n != sfpt) { 2474 extra_data_nodes.push(n); 2475 _igvn.register_new_node_with_optimizer(m); 2476 assert(get_ctrl(n) == cle_out, "what other control?"); 2477 set_ctrl(m, new_cle_out); 2478 old_new.map(n->_idx, m); 2479 } 2480 stack.pop(); 2481 } 2482 } 2483 if (mode == CloneIncludesStripMined) { 2484 _igvn.register_new_node_with_optimizer(new_sfpt); 2485 _igvn.register_new_node_with_optimizer(new_cle_out); 2486 } 2487 // Some other transformation may have pessimistically assigned some 2488 // data nodes to the outer loop. Set their control so they are out 2489 // of the outer loop. 2490 ResourceMark rm; 2491 Unique_Node_List wq; 2492 for (uint i = 0; i < extra_data_nodes.size(); i++) { 2493 Node* old = extra_data_nodes.at(i); 2494 collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true); 2495 } 2496 2497 for (uint i = 0; i < loop->_body.size(); i++) { 2498 Node* old = loop->_body.at(i); 2499 collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true); 2500 } 2501 2502 Node* inner_out = sfpt->in(0); 2503 if (inner_out->outcnt() > 1) { 2504 collect_nodes_in_outer_loop_not_reachable_from_sfpt(inner_out, loop, outer_loop, old_new, wq, this, true); 2505 } 2506 2507 Node* new_ctrl = cl->outer_loop_exit(); 2508 assert(get_loop(new_ctrl) != outer_loop, "must be out of the loop nest"); 2509 for (uint i = 0; i < wq.size(); i++) { 2510 Node* n = wq.at(i); 2511 set_ctrl(n, new_ctrl); 2512 if (n->in(0) != nullptr) { 2513 _igvn.replace_input_of(n, 0, new_ctrl); 2514 } 2515 collect_nodes_in_outer_loop_not_reachable_from_sfpt(n, loop, outer_loop, old_new, wq, this, false); 2516 } 2517 } else { 2518 Node *newhead = old_new[loop->_head->_idx]; 2519 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 2520 } 2521 } 2522 2523 //------------------------------clone_loop------------------------------------- 2524 // 2525 // C L O N E A L O O P B O D Y 2526 // 2527 // This is the basic building block of the loop optimizations. It clones an 2528 // entire loop body. It makes an old_new loop body mapping; with this mapping 2529 // you can find the new-loop equivalent to an old-loop node. All new-loop 2530 // nodes are exactly equal to their old-loop counterparts, all edges are the 2531 // same. All exits from the old-loop now have a RegionNode that merges the 2532 // equivalent new-loop path. This is true even for the normal "loop-exit" 2533 // condition. All uses of loop-invariant old-loop values now come from (one 2534 // or more) Phis that merge their new-loop equivalents. 2535 // 2536 // This operation leaves the graph in an illegal state: there are two valid 2537 // control edges coming from the loop pre-header to both loop bodies. I'll 2538 // definitely have to hack the graph after running this transform. 2539 // 2540 // From this building block I will further edit edges to perform loop peeling 2541 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc. 2542 // 2543 // Parameter side_by_size_idom: 2544 // When side_by_size_idom is null, the dominator tree is constructed for 2545 // the clone loop to dominate the original. Used in construction of 2546 // pre-main-post loop sequence. 2547 // When nonnull, the clone and original are side-by-side, both are 2548 // dominated by the side_by_side_idom node. Used in construction of 2549 // unswitched loops. 2550 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd, 2551 CloneLoopMode mode, Node* side_by_side_idom) { 2552 2553 LoopNode* head = loop->_head->as_Loop(); 2554 head->verify_strip_mined(1); 2555 2556 if (C->do_vector_loop() && PrintOpto) { 2557 const char* mname = C->method()->name()->as_quoted_ascii(); 2558 if (mname != nullptr) { 2559 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname); 2560 } 2561 } 2562 2563 CloneMap& cm = C->clone_map(); 2564 if (C->do_vector_loop()) { 2565 cm.set_clone_idx(cm.max_gen()+1); 2566 #ifndef PRODUCT 2567 if (PrintOpto) { 2568 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx()); 2569 loop->dump_head(); 2570 } 2571 #endif 2572 } 2573 2574 // Step 1: Clone the loop body. Make the old->new mapping. 2575 clone_loop_body(loop->_body, old_new, &cm); 2576 2577 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop; 2578 2579 // Step 2: Fix the edges in the new body. If the old input is outside the 2580 // loop use it. If the old input is INside the loop, use the corresponding 2581 // new node instead. 2582 fix_body_edges(loop->_body, loop, old_new, dd, outer_loop->_parent, false); 2583 2584 Node_List extra_data_nodes; // data nodes in the outer strip mined loop 2585 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes); 2586 2587 // Step 3: Now fix control uses. Loop varying control uses have already 2588 // been fixed up (as part of all input edges in Step 2). Loop invariant 2589 // control uses must be either an IfFalse or an IfTrue. Make a merge 2590 // point to merge the old and new IfFalse/IfTrue nodes; make the use 2591 // refer to this. 2592 Node_List worklist; 2593 uint new_counter = C->unique(); 2594 fix_ctrl_uses(loop->_body, loop, old_new, mode, side_by_side_idom, &cm, worklist); 2595 2596 // Step 4: If loop-invariant use is not control, it must be dominated by a 2597 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region 2598 // there if needed. Make a Phi there merging old and new used values. 2599 Node_List *split_if_set = nullptr; 2600 Node_List *split_bool_set = nullptr; 2601 Node_List *split_cex_set = nullptr; 2602 fix_data_uses(loop->_body, loop, mode, outer_loop, new_counter, old_new, worklist, split_if_set, split_bool_set, split_cex_set); 2603 2604 for (uint i = 0; i < extra_data_nodes.size(); i++) { 2605 Node* old = extra_data_nodes.at(i); 2606 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set, 2607 split_bool_set, split_cex_set, worklist, new_counter, 2608 mode); 2609 } 2610 2611 // Check for IFs that need splitting/cloning. Happens if an IF outside of 2612 // the loop uses a condition set in the loop. The original IF probably 2613 // takes control from one or more OLD Regions (which in turn get from NEW 2614 // Regions). In any case, there will be a set of Phis for each merge point 2615 // from the IF up to where the original BOOL def exists the loop. 2616 finish_clone_loop(split_if_set, split_bool_set, split_cex_set); 2617 2618 } 2619 2620 void PhaseIdealLoop::finish_clone_loop(Node_List* split_if_set, Node_List* split_bool_set, Node_List* split_cex_set) { 2621 if (split_if_set) { 2622 while (split_if_set->size()) { 2623 Node *iff = split_if_set->pop(); 2624 uint input = iff->Opcode() == Op_AllocateArray ? AllocateNode::ValidLengthTest : 1; 2625 if (iff->in(input)->is_Phi()) { 2626 Node *b = clone_iff(iff->in(input)->as_Phi()); 2627 _igvn.replace_input_of(iff, input, b); 2628 } 2629 } 2630 } 2631 if (split_bool_set) { 2632 while (split_bool_set->size()) { 2633 Node *b = split_bool_set->pop(); 2634 Node *phi = b->in(1); 2635 assert(phi->is_Phi(), ""); 2636 CmpNode *cmp = clone_bool((PhiNode*) phi); 2637 _igvn.replace_input_of(b, 1, cmp); 2638 } 2639 } 2640 if (split_cex_set) { 2641 while (split_cex_set->size()) { 2642 Node *b = split_cex_set->pop(); 2643 assert(b->in(0)->is_Region(), ""); 2644 assert(b->in(1)->is_Phi(), ""); 2645 assert(b->in(0)->in(0) == b->in(1)->in(0), ""); 2646 split_up(b, b->in(0), nullptr); 2647 } 2648 } 2649 } 2650 2651 void PhaseIdealLoop::fix_data_uses(Node_List& body, IdealLoopTree* loop, CloneLoopMode mode, IdealLoopTree* outer_loop, 2652 uint new_counter, Node_List &old_new, Node_List &worklist, Node_List*& split_if_set, 2653 Node_List*& split_bool_set, Node_List*& split_cex_set) { 2654 for(uint i = 0; i < body.size(); i++ ) { 2655 Node* old = body.at(i); 2656 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set, 2657 split_bool_set, split_cex_set, worklist, new_counter, 2658 mode); 2659 } 2660 } 2661 2662 void PhaseIdealLoop::fix_ctrl_uses(const Node_List& body, const IdealLoopTree* loop, Node_List &old_new, CloneLoopMode mode, 2663 Node* side_by_side_idom, CloneMap* cm, Node_List &worklist) { 2664 LoopNode* head = loop->_head->as_Loop(); 2665 for(uint i = 0; i < body.size(); i++ ) { 2666 Node* old = body.at(i); 2667 if( !old->is_CFG() ) continue; 2668 2669 // Copy uses to a worklist, so I can munge the def-use info 2670 // with impunity. 2671 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) { 2672 worklist.push(old->fast_out(j)); 2673 } 2674 2675 while (worklist.size()) { // Visit all uses 2676 Node *use = worklist.pop(); 2677 if (!has_node(use)) continue; // Ignore dead nodes 2678 IdealLoopTree *use_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use ); 2679 if (!loop->is_member(use_loop) && use->is_CFG()) { 2680 // Both OLD and USE are CFG nodes here. 2681 assert(use->is_Proj(), "" ); 2682 Node* nnn = old_new[old->_idx]; 2683 2684 Node* newuse = nullptr; 2685 if (head->is_strip_mined() && mode != IgnoreStripMined) { 2686 CountedLoopNode* cl = head->as_CountedLoop(); 2687 CountedLoopEndNode* cle = cl->loopexit(); 2688 Node* cle_out = cle->proj_out_or_null(false); 2689 if (use == cle_out) { 2690 IfNode* le = cl->outer_loop_end(); 2691 use = le->proj_out(false); 2692 use_loop = get_loop(use); 2693 if (mode == CloneIncludesStripMined) { 2694 nnn = old_new[le->_idx]; 2695 } else { 2696 newuse = old_new[cle_out->_idx]; 2697 } 2698 } 2699 } 2700 if (newuse == nullptr) { 2701 newuse = use->clone(); 2702 } 2703 2704 // Clone the loop exit control projection 2705 if (C->do_vector_loop() && cm != nullptr) { 2706 cm->verify_insert_and_clone(use, newuse, cm->clone_idx()); 2707 } 2708 newuse->set_req(0,nnn); 2709 _igvn.register_new_node_with_optimizer(newuse); 2710 set_loop(newuse, use_loop); 2711 set_idom(newuse, nnn, dom_depth(nnn) + 1 ); 2712 2713 // We need a Region to merge the exit from the peeled body and the 2714 // exit from the old loop body. 2715 RegionNode *r = new RegionNode(3); 2716 // Map the old use to the new merge point 2717 old_new.map( use->_idx, r ); 2718 uint dd_r = MIN2(dom_depth(newuse), dom_depth(use)); 2719 assert(dd_r >= dom_depth(dom_lca(newuse, use)), "" ); 2720 2721 // The original user of 'use' uses 'r' instead. 2722 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) { 2723 Node* useuse = use->last_out(l); 2724 _igvn.rehash_node_delayed(useuse); 2725 uint uses_found = 0; 2726 if (useuse->in(0) == use) { 2727 useuse->set_req(0, r); 2728 uses_found++; 2729 if (useuse->is_CFG()) { 2730 // This is not a dom_depth > dd_r because when new 2731 // control flow is constructed by a loop opt, a node and 2732 // its dominator can end up at the same dom_depth 2733 assert(dom_depth(useuse) >= dd_r, ""); 2734 set_idom(useuse, r, dom_depth(useuse)); 2735 } 2736 } 2737 for (uint k = 1; k < useuse->req(); k++) { 2738 if( useuse->in(k) == use ) { 2739 useuse->set_req(k, r); 2740 uses_found++; 2741 if (useuse->is_Loop() && k == LoopNode::EntryControl) { 2742 // This is not a dom_depth > dd_r because when new 2743 // control flow is constructed by a loop opt, a node 2744 // and its dominator can end up at the same dom_depth 2745 assert(dom_depth(useuse) >= dd_r , ""); 2746 set_idom(useuse, r, dom_depth(useuse)); 2747 } 2748 } 2749 } 2750 l -= uses_found; // we deleted 1 or more copies of this edge 2751 } 2752 2753 // Now finish up 'r' 2754 r->set_req(1, newuse); 2755 r->set_req(2, use); 2756 _igvn.register_new_node_with_optimizer(r); 2757 set_loop(r, use_loop); 2758 set_idom(r, (side_by_side_idom == nullptr) ? newuse->in(0) : side_by_side_idom, dd_r); 2759 } // End of if a loop-exit test 2760 } 2761 } 2762 } 2763 2764 void PhaseIdealLoop::fix_body_edges(const Node_List &body, IdealLoopTree* loop, const Node_List &old_new, int dd, 2765 IdealLoopTree* parent, bool partial) { 2766 for(uint i = 0; i < body.size(); i++ ) { 2767 Node *old = body.at(i); 2768 Node *nnn = old_new[old->_idx]; 2769 // Fix CFG/Loop controlling the new node 2770 if (has_ctrl(old)) { 2771 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]); 2772 } else { 2773 set_loop(nnn, parent); 2774 if (old->outcnt() > 0) { 2775 Node* dom = idom(old); 2776 if (old_new[dom->_idx] != nullptr) { 2777 dom = old_new[dom->_idx]; 2778 set_idom(nnn, dom, dd ); 2779 } 2780 } 2781 } 2782 // Correct edges to the new node 2783 for (uint j = 0; j < nnn->req(); j++) { 2784 Node *n = nnn->in(j); 2785 if (n != nullptr) { 2786 IdealLoopTree *old_in_loop = get_loop(has_ctrl(n) ? get_ctrl(n) : n); 2787 if (loop->is_member(old_in_loop)) { 2788 if (old_new[n->_idx] != nullptr) { 2789 nnn->set_req(j, old_new[n->_idx]); 2790 } else { 2791 assert(!body.contains(n), ""); 2792 assert(partial, "node not cloned"); 2793 } 2794 } 2795 } 2796 } 2797 _igvn.hash_find_insert(nnn); 2798 } 2799 } 2800 2801 void PhaseIdealLoop::clone_loop_body(const Node_List& body, Node_List &old_new, CloneMap* cm) { 2802 for (uint i = 0; i < body.size(); i++) { 2803 Node* old = body.at(i); 2804 Node* nnn = old->clone(); 2805 old_new.map(old->_idx, nnn); 2806 if (old->is_reduction()) { 2807 // Reduction flag is not copied by default. Copy it here when cloning the entire loop body. 2808 nnn->add_flag(Node::Flag_is_reduction); 2809 } 2810 if (C->do_vector_loop() && cm != nullptr) { 2811 cm->verify_insert_and_clone(old, nnn, cm->clone_idx()); 2812 } 2813 _igvn.register_new_node_with_optimizer(nnn); 2814 } 2815 } 2816 2817 2818 //---------------------- stride_of_possible_iv ------------------------------------- 2819 // Looks for an iff/bool/comp with one operand of the compare 2820 // being a cycle involving an add and a phi, 2821 // with an optional truncation (left-shift followed by a right-shift) 2822 // of the add. Returns zero if not an iv. 2823 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) { 2824 Node* trunc1 = nullptr; 2825 Node* trunc2 = nullptr; 2826 const TypeInteger* ttype = nullptr; 2827 if (!iff->is_If() || iff->in(1) == nullptr || !iff->in(1)->is_Bool()) { 2828 return 0; 2829 } 2830 BoolNode* bl = iff->in(1)->as_Bool(); 2831 Node* cmp = bl->in(1); 2832 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) { 2833 return 0; 2834 } 2835 // Must have an invariant operand 2836 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) { 2837 return 0; 2838 } 2839 Node* add2 = nullptr; 2840 Node* cmp1 = cmp->in(1); 2841 if (cmp1->is_Phi()) { 2842 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) ))) 2843 Node* phi = cmp1; 2844 for (uint i = 1; i < phi->req(); i++) { 2845 Node* in = phi->in(i); 2846 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in, 2847 &trunc1, &trunc2, &ttype, T_INT); 2848 if (add && add->in(1) == phi) { 2849 add2 = add->in(2); 2850 break; 2851 } 2852 } 2853 } else { 2854 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) ))) 2855 Node* addtrunc = cmp1; 2856 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc, 2857 &trunc1, &trunc2, &ttype, T_INT); 2858 if (add && add->in(1)->is_Phi()) { 2859 Node* phi = add->in(1); 2860 for (uint i = 1; i < phi->req(); i++) { 2861 if (phi->in(i) == addtrunc) { 2862 add2 = add->in(2); 2863 break; 2864 } 2865 } 2866 } 2867 } 2868 if (add2 != nullptr) { 2869 const TypeInt* add2t = _igvn.type(add2)->is_int(); 2870 if (add2t->is_con()) { 2871 return add2t->get_con(); 2872 } 2873 } 2874 return 0; 2875 } 2876 2877 2878 //---------------------- stay_in_loop ------------------------------------- 2879 // Return the (unique) control output node that's in the loop (if it exists.) 2880 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) { 2881 Node* unique = nullptr; 2882 if (!n) return nullptr; 2883 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2884 Node* use = n->fast_out(i); 2885 if (!has_ctrl(use) && loop->is_member(get_loop(use))) { 2886 if (unique != nullptr) { 2887 return nullptr; 2888 } 2889 unique = use; 2890 } 2891 } 2892 return unique; 2893 } 2894 2895 //------------------------------ register_node ------------------------------------- 2896 // Utility to register node "n" with PhaseIdealLoop 2897 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) { 2898 _igvn.register_new_node_with_optimizer(n); 2899 loop->_body.push(n); 2900 if (n->is_CFG()) { 2901 set_loop(n, loop); 2902 set_idom(n, pred, ddepth); 2903 } else { 2904 set_ctrl(n, pred); 2905 } 2906 } 2907 2908 //------------------------------ proj_clone ------------------------------------- 2909 // Utility to create an if-projection 2910 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) { 2911 ProjNode* c = p->clone()->as_Proj(); 2912 c->set_req(0, iff); 2913 return c; 2914 } 2915 2916 //------------------------------ short_circuit_if ------------------------------------- 2917 // Force the iff control output to be the live_proj 2918 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) { 2919 guarantee(live_proj != nullptr, "null projection"); 2920 int proj_con = live_proj->_con; 2921 assert(proj_con == 0 || proj_con == 1, "false or true projection"); 2922 Node *con = _igvn.intcon(proj_con); 2923 set_ctrl(con, C->root()); 2924 if (iff) { 2925 iff->set_req(1, con); 2926 } 2927 return con; 2928 } 2929 2930 //------------------------------ insert_if_before_proj ------------------------------------- 2931 // Insert a new if before an if projection (* - new node) 2932 // 2933 // before 2934 // if(test) 2935 // / \ 2936 // v v 2937 // other-proj proj (arg) 2938 // 2939 // after 2940 // if(test) 2941 // / \ 2942 // / v 2943 // | * proj-clone 2944 // v | 2945 // other-proj v 2946 // * new_if(relop(cmp[IU](left,right))) 2947 // / \ 2948 // v v 2949 // * new-proj proj 2950 // (returned) 2951 // 2952 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) { 2953 IfNode* iff = proj->in(0)->as_If(); 2954 IdealLoopTree *loop = get_loop(proj); 2955 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2956 int ddepth = dom_depth(proj); 2957 2958 _igvn.rehash_node_delayed(iff); 2959 _igvn.rehash_node_delayed(proj); 2960 2961 proj->set_req(0, nullptr); // temporary disconnect 2962 ProjNode* proj2 = proj_clone(proj, iff); 2963 register_node(proj2, loop, iff, ddepth); 2964 2965 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right); 2966 register_node(cmp, loop, proj2, ddepth); 2967 2968 BoolNode* bol = new BoolNode(cmp, relop); 2969 register_node(bol, loop, proj2, ddepth); 2970 2971 int opcode = iff->Opcode(); 2972 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode"); 2973 IfNode* new_if = (opcode == Op_If) ? new IfNode(proj2, bol, iff->_prob, iff->_fcnt): 2974 new RangeCheckNode(proj2, bol, iff->_prob, iff->_fcnt); 2975 register_node(new_if, loop, proj2, ddepth); 2976 2977 proj->set_req(0, new_if); // reattach 2978 set_idom(proj, new_if, ddepth); 2979 2980 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj(); 2981 guarantee(new_exit != nullptr, "null exit node"); 2982 register_node(new_exit, get_loop(other_proj), new_if, ddepth); 2983 2984 return new_exit; 2985 } 2986 2987 //------------------------------ insert_region_before_proj ------------------------------------- 2988 // Insert a region before an if projection (* - new node) 2989 // 2990 // before 2991 // if(test) 2992 // / | 2993 // v | 2994 // proj v 2995 // other-proj 2996 // 2997 // after 2998 // if(test) 2999 // / | 3000 // v | 3001 // * proj-clone v 3002 // | other-proj 3003 // v 3004 // * new-region 3005 // | 3006 // v 3007 // * dum_if 3008 // / \ 3009 // v \ 3010 // * dum-proj v 3011 // proj 3012 // 3013 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) { 3014 IfNode* iff = proj->in(0)->as_If(); 3015 IdealLoopTree *loop = get_loop(proj); 3016 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 3017 int ddepth = dom_depth(proj); 3018 3019 _igvn.rehash_node_delayed(iff); 3020 _igvn.rehash_node_delayed(proj); 3021 3022 proj->set_req(0, nullptr); // temporary disconnect 3023 ProjNode* proj2 = proj_clone(proj, iff); 3024 register_node(proj2, loop, iff, ddepth); 3025 3026 RegionNode* reg = new RegionNode(2); 3027 reg->set_req(1, proj2); 3028 register_node(reg, loop, iff, ddepth); 3029 3030 IfNode* dum_if = new IfNode(reg, short_circuit_if(nullptr, proj), iff->_prob, iff->_fcnt); 3031 register_node(dum_if, loop, reg, ddepth); 3032 3033 proj->set_req(0, dum_if); // reattach 3034 set_idom(proj, dum_if, ddepth); 3035 3036 ProjNode* dum_proj = proj_clone(other_proj, dum_if); 3037 register_node(dum_proj, loop, dum_if, ddepth); 3038 3039 return reg; 3040 } 3041 3042 //------------------------------ insert_cmpi_loop_exit ------------------------------------- 3043 // Clone a signed compare loop exit from an unsigned compare and 3044 // insert it before the unsigned cmp on the stay-in-loop path. 3045 // All new nodes inserted in the dominator tree between the original 3046 // if and it's projections. The original if test is replaced with 3047 // a constant to force the stay-in-loop path. 3048 // 3049 // This is done to make sure that the original if and it's projections 3050 // still dominate the same set of control nodes, that the ctrl() relation 3051 // from data nodes to them is preserved, and that their loop nesting is 3052 // preserved. 3053 // 3054 // before 3055 // if(i <u limit) unsigned compare loop exit 3056 // / | 3057 // v v 3058 // exit-proj stay-in-loop-proj 3059 // 3060 // after 3061 // if(stay-in-loop-const) original if 3062 // / | 3063 // / v 3064 // / if(i < limit) new signed test 3065 // / / | 3066 // / / v 3067 // / / if(i <u limit) new cloned unsigned test 3068 // / / / | 3069 // v v v | 3070 // region | 3071 // | | 3072 // dum-if | 3073 // / | | 3074 // ether | | 3075 // v v 3076 // exit-proj stay-in-loop-proj 3077 // 3078 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) { 3079 const bool Signed = true; 3080 const bool Unsigned = false; 3081 3082 BoolNode* bol = if_cmpu->in(1)->as_Bool(); 3083 if (bol->_test._test != BoolTest::lt) return nullptr; 3084 CmpNode* cmpu = bol->in(1)->as_Cmp(); 3085 if (cmpu->Opcode() != Op_CmpU) return nullptr; 3086 int stride = stride_of_possible_iv(if_cmpu); 3087 if (stride == 0) return nullptr; 3088 3089 Node* lp_proj = stay_in_loop(if_cmpu, loop); 3090 guarantee(lp_proj != nullptr, "null loop node"); 3091 3092 ProjNode* lp_continue = lp_proj->as_Proj(); 3093 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj(); 3094 if (!lp_exit->is_IfFalse()) { 3095 // The loop exit condition is (i <u limit) ==> (i >= 0 && i < limit). 3096 // We therefore can't add a single exit condition. 3097 return nullptr; 3098 } 3099 // The loop exit condition is !(i <u limit) ==> (i < 0 || i >= limit). 3100 // Split out the exit condition (i < 0) for stride < 0 or (i >= limit) for stride > 0. 3101 Node* limit = nullptr; 3102 if (stride > 0) { 3103 limit = cmpu->in(2); 3104 } else { 3105 limit = _igvn.makecon(TypeInt::ZERO); 3106 set_ctrl(limit, C->root()); 3107 } 3108 // Create a new region on the exit path 3109 RegionNode* reg = insert_region_before_proj(lp_exit); 3110 guarantee(reg != nullptr, "null region node"); 3111 3112 // Clone the if-cmpu-true-false using a signed compare 3113 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge; 3114 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue); 3115 reg->add_req(cmpi_exit); 3116 3117 // Clone the if-cmpu-true-false 3118 BoolTest::mask rel_u = bol->_test._test; 3119 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue); 3120 reg->add_req(cmpu_exit); 3121 3122 // Force original if to stay in loop. 3123 short_circuit_if(if_cmpu, lp_continue); 3124 3125 return cmpi_exit->in(0)->as_If(); 3126 } 3127 3128 //------------------------------ remove_cmpi_loop_exit ------------------------------------- 3129 // Remove a previously inserted signed compare loop exit. 3130 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) { 3131 Node* lp_proj = stay_in_loop(if_cmp, loop); 3132 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI && 3133 stay_in_loop(lp_proj, loop)->is_If() && 3134 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu"); 3135 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO); 3136 set_ctrl(con, C->root()); 3137 if_cmp->set_req(1, con); 3138 } 3139 3140 //------------------------------ scheduled_nodelist ------------------------------------- 3141 // Create a post order schedule of nodes that are in the 3142 // "member" set. The list is returned in "sched". 3143 // The first node in "sched" is the loop head, followed by 3144 // nodes which have no inputs in the "member" set, and then 3145 // followed by the nodes that have an immediate input dependence 3146 // on a node in "sched". 3147 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) { 3148 3149 assert(member.test(loop->_head->_idx), "loop head must be in member set"); 3150 VectorSet visited; 3151 Node_Stack nstack(loop->_body.size()); 3152 3153 Node* n = loop->_head; // top of stack is cached in "n" 3154 uint idx = 0; 3155 visited.set(n->_idx); 3156 3157 // Initially push all with no inputs from within member set 3158 for(uint i = 0; i < loop->_body.size(); i++ ) { 3159 Node *elt = loop->_body.at(i); 3160 if (member.test(elt->_idx)) { 3161 bool found = false; 3162 for (uint j = 0; j < elt->req(); j++) { 3163 Node* def = elt->in(j); 3164 if (def && member.test(def->_idx) && def != elt) { 3165 found = true; 3166 break; 3167 } 3168 } 3169 if (!found && elt != loop->_head) { 3170 nstack.push(n, idx); 3171 n = elt; 3172 assert(!visited.test(n->_idx), "not seen yet"); 3173 visited.set(n->_idx); 3174 } 3175 } 3176 } 3177 3178 // traverse out's that are in the member set 3179 while (true) { 3180 if (idx < n->outcnt()) { 3181 Node* use = n->raw_out(idx); 3182 idx++; 3183 if (!visited.test_set(use->_idx)) { 3184 if (member.test(use->_idx)) { 3185 nstack.push(n, idx); 3186 n = use; 3187 idx = 0; 3188 } 3189 } 3190 } else { 3191 // All outputs processed 3192 sched.push(n); 3193 if (nstack.is_empty()) break; 3194 n = nstack.node(); 3195 idx = nstack.index(); 3196 nstack.pop(); 3197 } 3198 } 3199 } 3200 3201 3202 //------------------------------ has_use_in_set ------------------------------------- 3203 // Has a use in the vector set 3204 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) { 3205 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3206 Node* use = n->fast_out(j); 3207 if (vset.test(use->_idx)) { 3208 return true; 3209 } 3210 } 3211 return false; 3212 } 3213 3214 3215 //------------------------------ has_use_internal_to_set ------------------------------------- 3216 // Has use internal to the vector set (ie. not in a phi at the loop head) 3217 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) { 3218 Node* head = loop->_head; 3219 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3220 Node* use = n->fast_out(j); 3221 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) { 3222 return true; 3223 } 3224 } 3225 return false; 3226 } 3227 3228 3229 //------------------------------ clone_for_use_outside_loop ------------------------------------- 3230 // clone "n" for uses that are outside of loop 3231 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) { 3232 int cloned = 0; 3233 assert(worklist.size() == 0, "should be empty"); 3234 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3235 Node* use = n->fast_out(j); 3236 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) { 3237 worklist.push(use); 3238 } 3239 } 3240 3241 if (C->check_node_count(worklist.size() + NodeLimitFudgeFactor, 3242 "Too many clones required in clone_for_use_outside_loop in partial peeling")) { 3243 return -1; 3244 } 3245 3246 while( worklist.size() ) { 3247 Node *use = worklist.pop(); 3248 if (!has_node(use) || use->in(0) == C->top()) continue; 3249 uint j; 3250 for (j = 0; j < use->req(); j++) { 3251 if (use->in(j) == n) break; 3252 } 3253 assert(j < use->req(), "must be there"); 3254 3255 // clone "n" and insert it between the inputs of "n" and the use outside the loop 3256 Node* n_clone = n->clone(); 3257 _igvn.replace_input_of(use, j, n_clone); 3258 cloned++; 3259 Node* use_c; 3260 if (!use->is_Phi()) { 3261 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0); 3262 } else { 3263 // Use in a phi is considered a use in the associated predecessor block 3264 use_c = use->in(0)->in(j); 3265 } 3266 set_ctrl(n_clone, use_c); 3267 assert(!loop->is_member(get_loop(use_c)), "should be outside loop"); 3268 get_loop(use_c)->_body.push(n_clone); 3269 _igvn.register_new_node_with_optimizer(n_clone); 3270 #ifndef PRODUCT 3271 if (TracePartialPeeling) { 3272 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx); 3273 } 3274 #endif 3275 } 3276 return cloned; 3277 } 3278 3279 3280 //------------------------------ clone_for_special_use_inside_loop ------------------------------------- 3281 // clone "n" for special uses that are in the not_peeled region. 3282 // If these def-uses occur in separate blocks, the code generator 3283 // marks the method as not compilable. For example, if a "BoolNode" 3284 // is in a different basic block than the "IfNode" that uses it, then 3285 // the compilation is aborted in the code generator. 3286 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 3287 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) { 3288 if (n->is_Phi() || n->is_Load()) { 3289 return; 3290 } 3291 assert(worklist.size() == 0, "should be empty"); 3292 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3293 Node* use = n->fast_out(j); 3294 if ( not_peel.test(use->_idx) && 3295 (use->is_If() || use->is_CMove() || use->is_Bool()) && 3296 use->in(1) == n) { 3297 worklist.push(use); 3298 } 3299 } 3300 if (worklist.size() > 0) { 3301 // clone "n" and insert it between inputs of "n" and the use 3302 Node* n_clone = n->clone(); 3303 loop->_body.push(n_clone); 3304 _igvn.register_new_node_with_optimizer(n_clone); 3305 set_ctrl(n_clone, get_ctrl(n)); 3306 sink_list.push(n_clone); 3307 not_peel.set(n_clone->_idx); 3308 #ifndef PRODUCT 3309 if (TracePartialPeeling) { 3310 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx); 3311 } 3312 #endif 3313 while( worklist.size() ) { 3314 Node *use = worklist.pop(); 3315 _igvn.rehash_node_delayed(use); 3316 for (uint j = 1; j < use->req(); j++) { 3317 if (use->in(j) == n) { 3318 use->set_req(j, n_clone); 3319 } 3320 } 3321 } 3322 } 3323 } 3324 3325 3326 //------------------------------ insert_phi_for_loop ------------------------------------- 3327 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 3328 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) { 3329 Node *phi = PhiNode::make(lp, back_edge_val); 3330 phi->set_req(LoopNode::EntryControl, lp_entry_val); 3331 // Use existing phi if it already exists 3332 Node *hit = _igvn.hash_find_insert(phi); 3333 if( hit == nullptr ) { 3334 _igvn.register_new_node_with_optimizer(phi); 3335 set_ctrl(phi, lp); 3336 } else { 3337 // Remove the new phi from the graph and use the hit 3338 _igvn.remove_dead_node(phi); 3339 phi = hit; 3340 } 3341 _igvn.replace_input_of(use, idx, phi); 3342 } 3343 3344 #ifdef ASSERT 3345 //------------------------------ is_valid_loop_partition ------------------------------------- 3346 // Validate the loop partition sets: peel and not_peel 3347 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, 3348 VectorSet& not_peel ) { 3349 uint i; 3350 // Check that peel_list entries are in the peel set 3351 for (i = 0; i < peel_list.size(); i++) { 3352 if (!peel.test(peel_list.at(i)->_idx)) { 3353 return false; 3354 } 3355 } 3356 // Check at loop members are in one of peel set or not_peel set 3357 for (i = 0; i < loop->_body.size(); i++ ) { 3358 Node *def = loop->_body.at(i); 3359 uint di = def->_idx; 3360 // Check that peel set elements are in peel_list 3361 if (peel.test(di)) { 3362 if (not_peel.test(di)) { 3363 return false; 3364 } 3365 // Must be in peel_list also 3366 bool found = false; 3367 for (uint j = 0; j < peel_list.size(); j++) { 3368 if (peel_list.at(j)->_idx == di) { 3369 found = true; 3370 break; 3371 } 3372 } 3373 if (!found) { 3374 return false; 3375 } 3376 } else if (not_peel.test(di)) { 3377 if (peel.test(di)) { 3378 return false; 3379 } 3380 } else { 3381 return false; 3382 } 3383 } 3384 return true; 3385 } 3386 3387 //------------------------------ is_valid_clone_loop_exit_use ------------------------------------- 3388 // Ensure a use outside of loop is of the right form 3389 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) { 3390 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 3391 return (use->is_Phi() && 3392 use_c->is_Region() && use_c->req() == 3 && 3393 (use_c->in(exit_idx)->Opcode() == Op_IfTrue || 3394 use_c->in(exit_idx)->Opcode() == Op_IfFalse || 3395 use_c->in(exit_idx)->Opcode() == Op_JumpProj) && 3396 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) ); 3397 } 3398 3399 //------------------------------ is_valid_clone_loop_form ------------------------------------- 3400 // Ensure that all uses outside of loop are of the right form 3401 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 3402 uint orig_exit_idx, uint clone_exit_idx) { 3403 uint len = peel_list.size(); 3404 for (uint i = 0; i < len; i++) { 3405 Node *def = peel_list.at(i); 3406 3407 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 3408 Node *use = def->fast_out(j); 3409 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 3410 if (!loop->is_member(get_loop(use_c))) { 3411 // use is not in the loop, check for correct structure 3412 if (use->in(0) == def) { 3413 // Okay 3414 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) { 3415 return false; 3416 } 3417 } 3418 } 3419 } 3420 return true; 3421 } 3422 #endif 3423 3424 //------------------------------ partial_peel ------------------------------------- 3425 // Partially peel (aka loop rotation) the top portion of a loop (called 3426 // the peel section below) by cloning it and placing one copy just before 3427 // the new loop head and the other copy at the bottom of the new loop. 3428 // 3429 // before after where it came from 3430 // 3431 // stmt1 stmt1 3432 // loop: stmt2 clone 3433 // stmt2 if condA goto exitA clone 3434 // if condA goto exitA new_loop: new 3435 // stmt3 stmt3 clone 3436 // if !condB goto loop if condB goto exitB clone 3437 // exitB: stmt2 orig 3438 // stmt4 if !condA goto new_loop orig 3439 // exitA: goto exitA 3440 // exitB: 3441 // stmt4 3442 // exitA: 3443 // 3444 // Step 1: find the cut point: an exit test on probable 3445 // induction variable. 3446 // Step 2: schedule (with cloning) operations in the peel 3447 // section that can be executed after the cut into 3448 // the section that is not peeled. This may need 3449 // to clone operations into exit blocks. For 3450 // instance, a reference to A[i] in the not-peel 3451 // section and a reference to B[i] in an exit block 3452 // may cause a left-shift of i by 2 to be placed 3453 // in the peel block. This step will clone the left 3454 // shift into the exit block and sink the left shift 3455 // from the peel to the not-peel section. 3456 // Step 3: clone the loop, retarget the control, and insert 3457 // phis for values that are live across the new loop 3458 // head. This is very dependent on the graph structure 3459 // from clone_loop. It creates region nodes for 3460 // exit control and associated phi nodes for values 3461 // flow out of the loop through that exit. The region 3462 // node is dominated by the clone's control projection. 3463 // So the clone's peel section is placed before the 3464 // new loop head, and the clone's not-peel section is 3465 // forms the top part of the new loop. The original 3466 // peel section forms the tail of the new loop. 3467 // Step 4: update the dominator tree and recompute the 3468 // dominator depth. 3469 // 3470 // orig 3471 // 3472 // stmt1 3473 // | 3474 // v 3475 // loop predicate 3476 // | 3477 // v 3478 // loop<----+ 3479 // | | 3480 // stmt2 | 3481 // | | 3482 // v | 3483 // ifA | 3484 // / | | 3485 // v v | 3486 // false true ^ <-- last_peel 3487 // / | | 3488 // / ===|==cut | 3489 // / stmt3 | <-- first_not_peel 3490 // / | | 3491 // | v | 3492 // v ifB | 3493 // exitA: / \ | 3494 // / \ | 3495 // v v | 3496 // false true | 3497 // / \ | 3498 // / ----+ 3499 // | 3500 // v 3501 // exitB: 3502 // stmt4 3503 // 3504 // 3505 // after clone loop 3506 // 3507 // stmt1 3508 // | 3509 // v 3510 // loop predicate 3511 // / \ 3512 // clone / \ orig 3513 // / \ 3514 // / \ 3515 // v v 3516 // +---->loop loop<----+ 3517 // | | | | 3518 // | stmt2 stmt2 | 3519 // | | | | 3520 // | v v | 3521 // | ifA ifA | 3522 // | | \ / | | 3523 // | v v v v | 3524 // ^ true false false true ^ <-- last_peel 3525 // | | ^ \ / | | 3526 // | cut==|== \ \ / ===|==cut | 3527 // | stmt3 \ \ / stmt3 | <-- first_not_peel 3528 // | | dom | | | | 3529 // | v \ 1v v2 v | 3530 // | ifB regionA ifB | 3531 // | / \ | / \ | 3532 // | / \ v / \ | 3533 // | v v exitA: v v | 3534 // | true false false true | 3535 // | / ^ \ / \ | 3536 // +---- \ \ / ----+ 3537 // dom \ / 3538 // \ 1v v2 3539 // regionB 3540 // | 3541 // v 3542 // exitB: 3543 // stmt4 3544 // 3545 // 3546 // after partial peel 3547 // 3548 // stmt1 3549 // | 3550 // v 3551 // loop predicate 3552 // / 3553 // clone / orig 3554 // / TOP 3555 // / \ 3556 // v v 3557 // TOP->loop loop----+ 3558 // | | | 3559 // stmt2 stmt2 | 3560 // | | | 3561 // v v | 3562 // ifA ifA | 3563 // | \ / | | 3564 // v v v v | 3565 // true false false true | <-- last_peel 3566 // | ^ \ / +------|---+ 3567 // +->newloop \ \ / === ==cut | | 3568 // | stmt3 \ \ / TOP | | 3569 // | | dom | | stmt3 | | <-- first_not_peel 3570 // | v \ 1v v2 v | | 3571 // | ifB regionA ifB ^ v 3572 // | / \ | / \ | | 3573 // | / \ v / \ | | 3574 // | v v exitA: v v | | 3575 // | true false false true | | 3576 // | / ^ \ / \ | | 3577 // | | \ \ / v | | 3578 // | | dom \ / TOP | | 3579 // | | \ 1v v2 | | 3580 // ^ v regionB | | 3581 // | | | | | 3582 // | | v ^ v 3583 // | | exitB: | | 3584 // | | stmt4 | | 3585 // | +------------>-----------------+ | 3586 // | | 3587 // +-----------------<---------------------+ 3588 // 3589 // 3590 // final graph 3591 // 3592 // stmt1 3593 // | 3594 // v 3595 // loop predicate 3596 // | 3597 // v 3598 // stmt2 clone 3599 // | 3600 // v 3601 // ........> ifA clone 3602 // : / | 3603 // dom / | 3604 // : v v 3605 // : false true 3606 // : | | 3607 // : | v 3608 // : | newloop<-----+ 3609 // : | | | 3610 // : | stmt3 clone | 3611 // : | | | 3612 // : | v | 3613 // : | ifB | 3614 // : | / \ | 3615 // : | v v | 3616 // : | false true | 3617 // : | | | | 3618 // : | v stmt2 | 3619 // : | exitB: | | 3620 // : | stmt4 v | 3621 // : | ifA orig | 3622 // : | / \ | 3623 // : | / \ | 3624 // : | v v | 3625 // : | false true | 3626 // : | / \ | 3627 // : v v -----+ 3628 // RegionA 3629 // | 3630 // v 3631 // exitA 3632 // 3633 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) { 3634 3635 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only"); 3636 if (!loop->_head->is_Loop()) { 3637 return false; 3638 } 3639 LoopNode *head = loop->_head->as_Loop(); 3640 3641 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) { 3642 return false; 3643 } 3644 3645 // Check for complex exit control 3646 for (uint ii = 0; ii < loop->_body.size(); ii++) { 3647 Node *n = loop->_body.at(ii); 3648 int opc = n->Opcode(); 3649 if (n->is_Call() || 3650 opc == Op_Catch || 3651 opc == Op_CatchProj || 3652 opc == Op_Jump || 3653 opc == Op_JumpProj) { 3654 #ifndef PRODUCT 3655 if (TracePartialPeeling) { 3656 tty->print_cr("\nExit control too complex: lp: %d", head->_idx); 3657 } 3658 #endif 3659 return false; 3660 } 3661 } 3662 3663 int dd = dom_depth(head); 3664 3665 // Step 1: find cut point 3666 3667 // Walk up dominators to loop head looking for first loop exit 3668 // which is executed on every path thru loop. 3669 IfNode *peel_if = nullptr; 3670 IfNode *peel_if_cmpu = nullptr; 3671 3672 Node *iff = loop->tail(); 3673 while (iff != head) { 3674 if (iff->is_If()) { 3675 Node *ctrl = get_ctrl(iff->in(1)); 3676 if (ctrl->is_top()) return false; // Dead test on live IF. 3677 // If loop-varying exit-test, check for induction variable 3678 if (loop->is_member(get_loop(ctrl)) && 3679 loop->is_loop_exit(iff) && 3680 is_possible_iv_test(iff)) { 3681 Node* cmp = iff->in(1)->in(1); 3682 if (cmp->Opcode() == Op_CmpI) { 3683 peel_if = iff->as_If(); 3684 } else { 3685 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU"); 3686 peel_if_cmpu = iff->as_If(); 3687 } 3688 } 3689 } 3690 iff = idom(iff); 3691 } 3692 3693 // Prefer signed compare over unsigned compare. 3694 IfNode* new_peel_if = nullptr; 3695 if (peel_if == nullptr) { 3696 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == nullptr) { 3697 return false; // No peel point found 3698 } 3699 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop); 3700 if (new_peel_if == nullptr) { 3701 return false; // No peel point found 3702 } 3703 peel_if = new_peel_if; 3704 } 3705 Node* last_peel = stay_in_loop(peel_if, loop); 3706 Node* first_not_peeled = stay_in_loop(last_peel, loop); 3707 if (first_not_peeled == nullptr || first_not_peeled == head) { 3708 return false; 3709 } 3710 3711 #ifndef PRODUCT 3712 if (TraceLoopOpts) { 3713 tty->print("PartialPeel "); 3714 loop->dump_head(); 3715 } 3716 3717 if (TracePartialPeeling) { 3718 tty->print_cr("before partial peel one iteration"); 3719 Node_List wl; 3720 Node* t = head->in(2); 3721 while (true) { 3722 wl.push(t); 3723 if (t == head) break; 3724 t = idom(t); 3725 } 3726 while (wl.size() > 0) { 3727 Node* tt = wl.pop(); 3728 tt->dump(); 3729 if (tt == last_peel) tty->print_cr("-- cut --"); 3730 } 3731 } 3732 #endif 3733 VectorSet peel; 3734 VectorSet not_peel; 3735 Node_List peel_list; 3736 Node_List worklist; 3737 Node_List sink_list; 3738 3739 uint estimate = loop->est_loop_clone_sz(1); 3740 if (exceeding_node_budget(estimate)) { 3741 return false; 3742 } 3743 3744 // Set of cfg nodes to peel are those that are executable from 3745 // the head through last_peel. 3746 assert(worklist.size() == 0, "should be empty"); 3747 worklist.push(head); 3748 peel.set(head->_idx); 3749 while (worklist.size() > 0) { 3750 Node *n = worklist.pop(); 3751 if (n != last_peel) { 3752 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3753 Node* use = n->fast_out(j); 3754 if (use->is_CFG() && 3755 loop->is_member(get_loop(use)) && 3756 !peel.test_set(use->_idx)) { 3757 worklist.push(use); 3758 } 3759 } 3760 } 3761 } 3762 3763 // Set of non-cfg nodes to peel are those that are control 3764 // dependent on the cfg nodes. 3765 for (uint i = 0; i < loop->_body.size(); i++) { 3766 Node *n = loop->_body.at(i); 3767 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n; 3768 if (peel.test(n_c->_idx)) { 3769 peel.set(n->_idx); 3770 } else { 3771 not_peel.set(n->_idx); 3772 } 3773 } 3774 3775 // Step 2: move operations from the peeled section down into the 3776 // not-peeled section 3777 3778 // Get a post order schedule of nodes in the peel region 3779 // Result in right-most operand. 3780 scheduled_nodelist(loop, peel, peel_list); 3781 3782 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 3783 3784 // For future check for too many new phis 3785 uint old_phi_cnt = 0; 3786 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 3787 Node* use = head->fast_out(j); 3788 if (use->is_Phi()) old_phi_cnt++; 3789 } 3790 3791 #ifndef PRODUCT 3792 if (TracePartialPeeling) { 3793 tty->print_cr("\npeeled list"); 3794 } 3795 #endif 3796 3797 // Evacuate nodes in peel region into the not_peeled region if possible 3798 bool too_many_clones = false; 3799 uint new_phi_cnt = 0; 3800 uint cloned_for_outside_use = 0; 3801 for (uint i = 0; i < peel_list.size();) { 3802 Node* n = peel_list.at(i); 3803 #ifndef PRODUCT 3804 if (TracePartialPeeling) n->dump(); 3805 #endif 3806 bool incr = true; 3807 if (!n->is_CFG()) { 3808 if (has_use_in_set(n, not_peel)) { 3809 // If not used internal to the peeled region, 3810 // move "n" from peeled to not_peeled region. 3811 if (!has_use_internal_to_set(n, peel, loop)) { 3812 // if not pinned and not a load (which maybe anti-dependent on a store) 3813 // and not a CMove (Matcher expects only bool->cmove). 3814 if (n->in(0) == nullptr && !n->is_Load() && !n->is_CMove()) { 3815 int new_clones = clone_for_use_outside_loop(loop, n, worklist); 3816 if (new_clones == -1) { 3817 too_many_clones = true; 3818 break; 3819 } 3820 cloned_for_outside_use += new_clones; 3821 sink_list.push(n); 3822 peel.remove(n->_idx); 3823 not_peel.set(n->_idx); 3824 peel_list.remove(i); 3825 incr = false; 3826 #ifndef PRODUCT 3827 if (TracePartialPeeling) { 3828 tty->print_cr("sink to not_peeled region: %d newbb: %d", 3829 n->_idx, get_ctrl(n)->_idx); 3830 } 3831 #endif 3832 } 3833 } else { 3834 // Otherwise check for special def-use cases that span 3835 // the peel/not_peel boundary such as bool->if 3836 clone_for_special_use_inside_loop(loop, n, not_peel, sink_list, worklist); 3837 new_phi_cnt++; 3838 } 3839 } 3840 } 3841 if (incr) i++; 3842 } 3843 3844 estimate += cloned_for_outside_use + new_phi_cnt; 3845 bool exceed_node_budget = !may_require_nodes(estimate); 3846 bool exceed_phi_limit = new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta; 3847 3848 if (too_many_clones || exceed_node_budget || exceed_phi_limit) { 3849 #ifndef PRODUCT 3850 if (TracePartialPeeling && exceed_phi_limit) { 3851 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c", 3852 new_phi_cnt, old_phi_cnt, new_peel_if != nullptr?'T':'F'); 3853 } 3854 #endif 3855 if (new_peel_if != nullptr) { 3856 remove_cmpi_loop_exit(new_peel_if, loop); 3857 } 3858 // Inhibit more partial peeling on this loop 3859 assert(!head->is_partial_peel_loop(), "not partial peeled"); 3860 head->mark_partial_peel_failed(); 3861 if (cloned_for_outside_use > 0) { 3862 // Terminate this round of loop opts because 3863 // the graph outside this loop was changed. 3864 C->set_major_progress(); 3865 return true; 3866 } 3867 return false; 3868 } 3869 3870 // Step 3: clone loop, retarget control, and insert new phis 3871 3872 // Create new loop head for new phis and to hang 3873 // the nodes being moved (sinked) from the peel region. 3874 LoopNode* new_head = new LoopNode(last_peel, last_peel); 3875 new_head->set_unswitch_count(head->unswitch_count()); // Preserve 3876 _igvn.register_new_node_with_optimizer(new_head); 3877 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled"); 3878 _igvn.replace_input_of(first_not_peeled, 0, new_head); 3879 set_loop(new_head, loop); 3880 loop->_body.push(new_head); 3881 not_peel.set(new_head->_idx); 3882 set_idom(new_head, last_peel, dom_depth(first_not_peeled)); 3883 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled)); 3884 3885 while (sink_list.size() > 0) { 3886 Node* n = sink_list.pop(); 3887 set_ctrl(n, new_head); 3888 } 3889 3890 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 3891 3892 clone_loop(loop, old_new, dd, IgnoreStripMined); 3893 3894 const uint clone_exit_idx = 1; 3895 const uint orig_exit_idx = 2; 3896 assert(is_valid_clone_loop_form(loop, peel_list, orig_exit_idx, clone_exit_idx), "bad clone loop"); 3897 3898 Node* head_clone = old_new[head->_idx]; 3899 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop(); 3900 Node* orig_tail_clone = head_clone->in(2); 3901 3902 // Add phi if "def" node is in peel set and "use" is not 3903 3904 for (uint i = 0; i < peel_list.size(); i++) { 3905 Node *def = peel_list.at(i); 3906 if (!def->is_CFG()) { 3907 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 3908 Node *use = def->fast_out(j); 3909 if (has_node(use) && use->in(0) != C->top() && 3910 (!peel.test(use->_idx) || 3911 (use->is_Phi() && use->in(0) == head)) ) { 3912 worklist.push(use); 3913 } 3914 } 3915 while( worklist.size() ) { 3916 Node *use = worklist.pop(); 3917 for (uint j = 1; j < use->req(); j++) { 3918 Node* n = use->in(j); 3919 if (n == def) { 3920 3921 // "def" is in peel set, "use" is not in peel set 3922 // or "use" is in the entry boundary (a phi) of the peel set 3923 3924 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use; 3925 3926 if ( loop->is_member(get_loop( use_c )) ) { 3927 // use is in loop 3928 if (old_new[use->_idx] != nullptr) { // null for dead code 3929 Node* use_clone = old_new[use->_idx]; 3930 _igvn.replace_input_of(use, j, C->top()); 3931 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone ); 3932 } 3933 } else { 3934 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format"); 3935 // use is not in the loop, check if the live range includes the cut 3936 Node* lp_if = use_c->in(orig_exit_idx)->in(0); 3937 if (not_peel.test(lp_if->_idx)) { 3938 assert(j == orig_exit_idx, "use from original loop"); 3939 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone ); 3940 } 3941 } 3942 } 3943 } 3944 } 3945 } 3946 } 3947 3948 // Step 3b: retarget control 3949 3950 // Redirect control to the new loop head if a cloned node in 3951 // the not_peeled region has control that points into the peeled region. 3952 // This necessary because the cloned peeled region will be outside 3953 // the loop. 3954 // from to 3955 // cloned-peeled <---+ 3956 // new_head_clone: | <--+ 3957 // cloned-not_peeled in(0) in(0) 3958 // orig-peeled 3959 3960 for (uint i = 0; i < loop->_body.size(); i++) { 3961 Node *n = loop->_body.at(i); 3962 if (!n->is_CFG() && n->in(0) != nullptr && 3963 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) { 3964 Node* n_clone = old_new[n->_idx]; 3965 _igvn.replace_input_of(n_clone, 0, new_head_clone); 3966 } 3967 } 3968 3969 // Backedge of the surviving new_head (the clone) is original last_peel 3970 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel); 3971 3972 // Cut first node in original not_peel set 3973 _igvn.rehash_node_delayed(new_head); // Multiple edge updates: 3974 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of 3975 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls 3976 3977 // Copy head_clone back-branch info to original head 3978 // and remove original head's loop entry and 3979 // clone head's back-branch 3980 _igvn.rehash_node_delayed(head); // Multiple edge updates 3981 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl)); 3982 head->set_req(LoopNode::LoopBackControl, C->top()); 3983 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top()); 3984 3985 // Similarly modify the phis 3986 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) { 3987 Node* use = head->fast_out(k); 3988 if (use->is_Phi() && use->outcnt() > 0) { 3989 Node* use_clone = old_new[use->_idx]; 3990 _igvn.rehash_node_delayed(use); // Multiple edge updates 3991 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl)); 3992 use->set_req(LoopNode::LoopBackControl, C->top()); 3993 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top()); 3994 } 3995 } 3996 3997 // Step 4: update dominator tree and dominator depth 3998 3999 set_idom(head, orig_tail_clone, dd); 4000 recompute_dom_depth(); 4001 4002 // Inhibit more partial peeling on this loop 4003 new_head_clone->set_partial_peel_loop(); 4004 C->set_major_progress(); 4005 loop->record_for_igvn(); 4006 4007 #ifndef PRODUCT 4008 if (TracePartialPeeling) { 4009 tty->print_cr("\nafter partial peel one iteration"); 4010 Node_List wl; 4011 Node* t = last_peel; 4012 while (true) { 4013 wl.push(t); 4014 if (t == head_clone) break; 4015 t = idom(t); 4016 } 4017 while (wl.size() > 0) { 4018 Node* tt = wl.pop(); 4019 if (tt == head) tty->print_cr("orig head"); 4020 else if (tt == new_head_clone) tty->print_cr("new head"); 4021 else if (tt == head_clone) tty->print_cr("clone head"); 4022 tt->dump(); 4023 } 4024 } 4025 #endif 4026 return true; 4027 } 4028 4029 // Transform: 4030 // 4031 // loop<-----------------+ 4032 // | | 4033 // stmt1 stmt2 .. stmtn | 4034 // | | | | 4035 // \ | / | 4036 // v v v | 4037 // region | 4038 // | | 4039 // shared_stmt | 4040 // | | 4041 // v | 4042 // if | 4043 // / \ | 4044 // | -----------+ 4045 // v 4046 // 4047 // into: 4048 // 4049 // loop<-------------------+ 4050 // | | 4051 // v | 4052 // +->loop | 4053 // | | | 4054 // | stmt1 stmt2 .. stmtn | 4055 // | | | | | 4056 // | | \ / | 4057 // | | v v | 4058 // | | region1 | 4059 // | | | | 4060 // | shared_stmt shared_stmt | 4061 // | | | | 4062 // | v v | 4063 // | if if | 4064 // | /\ / \ | 4065 // +-- | | -------+ 4066 // \ / 4067 // v v 4068 // region2 4069 // 4070 // (region2 is shown to merge mirrored projections of the loop exit 4071 // ifs to make the diagram clearer but they really merge the same 4072 // projection) 4073 // 4074 // Conditions for this transformation to trigger: 4075 // - the path through stmt1 is frequent enough 4076 // - the inner loop will be turned into a counted loop after transformation 4077 bool PhaseIdealLoop::duplicate_loop_backedge(IdealLoopTree *loop, Node_List &old_new) { 4078 if (!DuplicateBackedge) { 4079 return false; 4080 } 4081 assert(!loop->_head->is_CountedLoop() || StressDuplicateBackedge, "Non-counted loop only"); 4082 if (!loop->_head->is_Loop()) { 4083 return false; 4084 } 4085 4086 uint estimate = loop->est_loop_clone_sz(1); 4087 if (exceeding_node_budget(estimate)) { 4088 return false; 4089 } 4090 4091 LoopNode *head = loop->_head->as_Loop(); 4092 4093 Node* region = nullptr; 4094 IfNode* exit_test = nullptr; 4095 uint inner; 4096 float f; 4097 if (StressDuplicateBackedge) { 4098 if (head->is_strip_mined()) { 4099 return false; 4100 } 4101 Node* c = head->in(LoopNode::LoopBackControl); 4102 4103 while (c != head) { 4104 if (c->is_Region()) { 4105 region = c; 4106 } 4107 c = idom(c); 4108 } 4109 4110 if (region == nullptr) { 4111 return false; 4112 } 4113 4114 inner = 1; 4115 } else { 4116 // Is the shape of the loop that of a counted loop... 4117 Node* back_control = loop_exit_control(head, loop); 4118 if (back_control == nullptr) { 4119 return false; 4120 } 4121 4122 BoolTest::mask bt = BoolTest::illegal; 4123 float cl_prob = 0; 4124 Node* incr = nullptr; 4125 Node* limit = nullptr; 4126 Node* cmp = loop_exit_test(back_control, loop, incr, limit, bt, cl_prob); 4127 if (cmp == nullptr || cmp->Opcode() != Op_CmpI) { 4128 return false; 4129 } 4130 4131 // With an extra phi for the candidate iv? 4132 // Or the region node is the loop head 4133 if (!incr->is_Phi() || incr->in(0) == head) { 4134 return false; 4135 } 4136 4137 PathFrequency pf(head, this); 4138 region = incr->in(0); 4139 4140 // Go over all paths for the extra phi's region and see if that 4141 // path is frequent enough and would match the expected iv shape 4142 // if the extra phi is removed 4143 inner = 0; 4144 for (uint i = 1; i < incr->req(); ++i) { 4145 Node* in = incr->in(i); 4146 Node* trunc1 = nullptr; 4147 Node* trunc2 = nullptr; 4148 const TypeInteger* iv_trunc_t = nullptr; 4149 Node* orig_in = in; 4150 if (!(in = CountedLoopNode::match_incr_with_optional_truncation(in, &trunc1, &trunc2, &iv_trunc_t, T_INT))) { 4151 continue; 4152 } 4153 assert(in->Opcode() == Op_AddI, "wrong increment code"); 4154 Node* xphi = nullptr; 4155 Node* stride = loop_iv_stride(in, loop, xphi); 4156 4157 if (stride == nullptr) { 4158 continue; 4159 } 4160 4161 PhiNode* phi = loop_iv_phi(xphi, nullptr, head, loop); 4162 if (phi == nullptr || 4163 (trunc1 == nullptr && phi->in(LoopNode::LoopBackControl) != incr) || 4164 (trunc1 != nullptr && phi->in(LoopNode::LoopBackControl) != trunc1)) { 4165 return false; 4166 } 4167 4168 f = pf.to(region->in(i)); 4169 if (f > 0.5) { 4170 inner = i; 4171 break; 4172 } 4173 } 4174 4175 if (inner == 0) { 4176 return false; 4177 } 4178 4179 exit_test = back_control->in(0)->as_If(); 4180 } 4181 4182 if (idom(region)->is_Catch()) { 4183 return false; 4184 } 4185 4186 // Collect all control nodes that need to be cloned (shared_stmt in the diagram) 4187 Unique_Node_List wq; 4188 wq.push(head->in(LoopNode::LoopBackControl)); 4189 for (uint i = 0; i < wq.size(); i++) { 4190 Node* c = wq.at(i); 4191 assert(get_loop(c) == loop, "not in the right loop?"); 4192 if (c->is_Region()) { 4193 if (c != region) { 4194 for (uint j = 1; j < c->req(); ++j) { 4195 wq.push(c->in(j)); 4196 } 4197 } 4198 } else { 4199 wq.push(c->in(0)); 4200 } 4201 assert(!is_dominator(c, region) || c == region, "shouldn't go above region"); 4202 } 4203 4204 Node* region_dom = idom(region); 4205 4206 // Can't do the transformation if this would cause a membar pair to 4207 // be split 4208 for (uint i = 0; i < wq.size(); i++) { 4209 Node* c = wq.at(i); 4210 if (c->is_MemBar() && (c->as_MemBar()->trailing_store() || c->as_MemBar()->trailing_load_store())) { 4211 assert(c->as_MemBar()->leading_membar()->trailing_membar() == c, "bad membar pair"); 4212 if (!wq.member(c->as_MemBar()->leading_membar())) { 4213 return false; 4214 } 4215 } 4216 } 4217 4218 // Collect data nodes that need to be clones as well 4219 int dd = dom_depth(head); 4220 4221 for (uint i = 0; i < loop->_body.size(); ++i) { 4222 Node* n = loop->_body.at(i); 4223 if (has_ctrl(n)) { 4224 Node* c = get_ctrl(n); 4225 if (wq.member(c)) { 4226 wq.push(n); 4227 } 4228 } else { 4229 set_idom(n, idom(n), dd); 4230 } 4231 } 4232 4233 // clone shared_stmt 4234 clone_loop_body(wq, old_new, nullptr); 4235 4236 Node* region_clone = old_new[region->_idx]; 4237 region_clone->set_req(inner, C->top()); 4238 set_idom(region, region->in(inner), dd); 4239 4240 // Prepare the outer loop 4241 Node* outer_head = new LoopNode(head->in(LoopNode::EntryControl), old_new[head->in(LoopNode::LoopBackControl)->_idx]); 4242 register_control(outer_head, loop->_parent, outer_head->in(LoopNode::EntryControl)); 4243 _igvn.replace_input_of(head, LoopNode::EntryControl, outer_head); 4244 set_idom(head, outer_head, dd); 4245 4246 fix_body_edges(wq, loop, old_new, dd, loop->_parent, true); 4247 4248 // Make one of the shared_stmt copies only reachable from stmt1, the 4249 // other only from stmt2..stmtn. 4250 Node* dom = nullptr; 4251 for (uint i = 1; i < region->req(); ++i) { 4252 if (i != inner) { 4253 _igvn.replace_input_of(region, i, C->top()); 4254 } 4255 Node* in = region_clone->in(i); 4256 if (in->is_top()) { 4257 continue; 4258 } 4259 if (dom == nullptr) { 4260 dom = in; 4261 } else { 4262 dom = dom_lca(dom, in); 4263 } 4264 } 4265 4266 set_idom(region_clone, dom, dd); 4267 4268 // Set up the outer loop 4269 for (uint i = 0; i < head->outcnt(); i++) { 4270 Node* u = head->raw_out(i); 4271 if (u->is_Phi()) { 4272 Node* outer_phi = u->clone(); 4273 outer_phi->set_req(0, outer_head); 4274 Node* backedge = old_new[u->in(LoopNode::LoopBackControl)->_idx]; 4275 if (backedge == nullptr) { 4276 backedge = u->in(LoopNode::LoopBackControl); 4277 } 4278 outer_phi->set_req(LoopNode::LoopBackControl, backedge); 4279 register_new_node(outer_phi, outer_head); 4280 _igvn.replace_input_of(u, LoopNode::EntryControl, outer_phi); 4281 } 4282 } 4283 4284 // create control and data nodes for out of loop uses (including region2) 4285 Node_List worklist; 4286 uint new_counter = C->unique(); 4287 fix_ctrl_uses(wq, loop, old_new, ControlAroundStripMined, outer_head, nullptr, worklist); 4288 4289 Node_List *split_if_set = nullptr; 4290 Node_List *split_bool_set = nullptr; 4291 Node_List *split_cex_set = nullptr; 4292 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); 4293 4294 finish_clone_loop(split_if_set, split_bool_set, split_cex_set); 4295 4296 if (exit_test != nullptr) { 4297 float cnt = exit_test->_fcnt; 4298 if (cnt != COUNT_UNKNOWN) { 4299 exit_test->_fcnt = cnt * f; 4300 old_new[exit_test->_idx]->as_If()->_fcnt = cnt * (1 - f); 4301 } 4302 } 4303 4304 C->set_major_progress(); 4305 4306 return true; 4307 }