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/loopnode.hpp" 37 #include "opto/matcher.hpp" 38 #include "opto/mulnode.hpp" 39 #include "opto/movenode.hpp" 40 #include "opto/opaquenode.hpp" 41 #include "opto/rootnode.hpp" 42 #include "opto/subnode.hpp" 43 #include "opto/subtypenode.hpp" 44 #include "utilities/macros.hpp" 45 46 //============================================================================= 47 //------------------------------split_thru_phi--------------------------------- 48 // Split Node 'n' through merge point if there is enough win. 49 Node* PhaseIdealLoop::split_thru_phi(Node* n, Node* region, int policy) { 50 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) { 51 // ConvI2L may have type information on it which is unsafe to push up 52 // so disable this for now 53 return NULL; 54 } 55 56 // Splitting range check CastIIs through a loop induction Phi can 57 // cause new Phis to be created that are left unrelated to the loop 58 // induction Phi and prevent optimizations (vectorization) 59 if (n->Opcode() == Op_CastII && region->is_CountedLoop() && 60 n->in(1) == region->as_CountedLoop()->phi()) { 61 return NULL; 62 } 63 64 if (cannot_split_division(n, region)) { 65 return NULL; 66 } 67 68 int wins = 0; 69 assert(!n->is_CFG(), ""); 70 assert(region->is_Region(), ""); 71 72 const Type* type = n->bottom_type(); 73 const TypeOopPtr* t_oop = _igvn.type(n)->isa_oopptr(); 74 Node* phi; 75 if (t_oop != NULL && t_oop->is_known_instance_field()) { 76 int iid = t_oop->instance_id(); 77 int index = C->get_alias_index(t_oop); 78 int offset = t_oop->offset(); 79 phi = new PhiNode(region, type, NULL, iid, index, offset); 80 } else { 81 phi = PhiNode::make_blank(region, n); 82 } 83 uint old_unique = C->unique(); 84 for (uint i = 1; i < region->req(); i++) { 85 Node* x; 86 Node* the_clone = NULL; 87 if (region->in(i) == C->top()) { 88 x = C->top(); // Dead path? Use a dead data op 89 } else { 90 x = n->clone(); // Else clone up the data op 91 the_clone = x; // Remember for possible deletion. 92 // Alter data node to use pre-phi inputs 93 if (n->in(0) == region) 94 x->set_req( 0, region->in(i) ); 95 for (uint j = 1; j < n->req(); j++) { 96 Node* in = n->in(j); 97 if (in->is_Phi() && in->in(0) == region) 98 x->set_req(j, in->in(i)); // Use pre-Phi input for the clone 99 } 100 } 101 // Check for a 'win' on some paths 102 const Type* t = x->Value(&_igvn); 103 104 bool singleton = t->singleton(); 105 106 // A TOP singleton indicates that there are no possible values incoming 107 // along a particular edge. In most cases, this is OK, and the Phi will 108 // be eliminated later in an Ideal call. However, we can't allow this to 109 // happen if the singleton occurs on loop entry, as the elimination of 110 // the PhiNode may cause the resulting node to migrate back to a previous 111 // loop iteration. 112 if (singleton && t == Type::TOP) { 113 // Is_Loop() == false does not confirm the absence of a loop (e.g., an 114 // irreducible loop may not be indicated by an affirmative is_Loop()); 115 // therefore, the only top we can split thru a phi is on a backedge of 116 // a loop. 117 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 118 } 119 120 if (singleton) { 121 wins++; 122 x = ((PhaseGVN&)_igvn).makecon(t); 123 } else { 124 // We now call Identity to try to simplify the cloned node. 125 // Note that some Identity methods call phase->type(this). 126 // Make sure that the type array is big enough for 127 // our new node, even though we may throw the node away. 128 // (Note: This tweaking with igvn only works because x is a new node.) 129 _igvn.set_type(x, t); 130 // If x is a TypeNode, capture any more-precise type permanently into Node 131 // otherwise it will be not updated during igvn->transform since 132 // igvn->type(x) is set to x->Value() already. 133 x->raise_bottom_type(t); 134 Node* y = x->Identity(&_igvn); 135 if (y != x) { 136 wins++; 137 x = y; 138 } else { 139 y = _igvn.hash_find(x); 140 if (y) { 141 wins++; 142 x = y; 143 } else { 144 // Else x is a new node we are keeping 145 // We do not need register_new_node_with_optimizer 146 // because set_type has already been called. 147 _igvn._worklist.push(x); 148 } 149 } 150 } 151 if (x != the_clone && the_clone != NULL) 152 _igvn.remove_dead_node(the_clone); 153 phi->set_req( i, x ); 154 } 155 // Too few wins? 156 if (wins <= policy) { 157 _igvn.remove_dead_node(phi); 158 return NULL; 159 } 160 161 // Record Phi 162 register_new_node( phi, region ); 163 164 for (uint i2 = 1; i2 < phi->req(); i2++) { 165 Node *x = phi->in(i2); 166 // If we commoned up the cloned 'x' with another existing Node, 167 // the existing Node picks up a new use. We need to make the 168 // existing Node occur higher up so it dominates its uses. 169 Node *old_ctrl; 170 IdealLoopTree *old_loop; 171 172 if (x->is_Con()) { 173 // Constant's control is always root. 174 set_ctrl(x, C->root()); 175 continue; 176 } 177 // The occasional new node 178 if (x->_idx >= old_unique) { // Found a new, unplaced node? 179 old_ctrl = NULL; 180 old_loop = NULL; // Not in any prior loop 181 } else { 182 old_ctrl = get_ctrl(x); 183 old_loop = get_loop(old_ctrl); // Get prior loop 184 } 185 // New late point must dominate new use 186 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2)); 187 if (new_ctrl == old_ctrl) // Nothing is changed 188 continue; 189 190 IdealLoopTree *new_loop = get_loop(new_ctrl); 191 192 // Don't move x into a loop if its uses are 193 // outside of loop. Otherwise x will be cloned 194 // for each use outside of this loop. 195 IdealLoopTree *use_loop = get_loop(region); 196 if (!new_loop->is_member(use_loop) && 197 (old_loop == NULL || !new_loop->is_member(old_loop))) { 198 // Take early control, later control will be recalculated 199 // during next iteration of loop optimizations. 200 new_ctrl = get_early_ctrl(x); 201 new_loop = get_loop(new_ctrl); 202 } 203 // Set new location 204 set_ctrl(x, new_ctrl); 205 // If changing loop bodies, see if we need to collect into new body 206 if (old_loop != new_loop) { 207 if (old_loop && !old_loop->_child) 208 old_loop->_body.yank(x); 209 if (!new_loop->_child) 210 new_loop->_body.push(x); // Collect body info 211 } 212 } 213 214 return phi; 215 } 216 217 // Return true if 'n' is a Div or Mod node (without zero check If node which was removed earlier) with a loop phi divisor 218 // of a trip-counted (integer or long) loop with a backedge input that could be zero (include zero in its type range). In 219 // this case, we cannot split the division to the backedge as it could freely float above the loop exit check resulting in 220 // a division by zero. This situation is possible because the type of an increment node of an iv phi (trip-counter) could 221 // include zero while the iv phi does not (see PhiNode::Value() for trip-counted loops where we improve types of iv phis). 222 // We also need to check other loop phis as they could have been created in the same split-if pass when applying 223 // PhaseIdealLoop::split_thru_phi() to split nodes through an iv phi. 224 bool PhaseIdealLoop::cannot_split_division(const Node* n, const Node* region) const { 225 const Type* zero; 226 switch (n->Opcode()) { 227 case Op_DivI: 228 case Op_ModI: 229 zero = TypeInt::ZERO; 230 break; 231 case Op_DivL: 232 case Op_ModL: 233 zero = TypeLong::ZERO; 234 break; 235 default: 236 return false; 237 } 238 239 assert(n->in(0) == NULL, "divisions with zero check should already have bailed out earlier in split-if"); 240 Node* divisor = n->in(2); 241 return is_divisor_counted_loop_phi(divisor, region) && 242 loop_phi_backedge_type_contains_zero(divisor, zero); 243 } 244 245 bool PhaseIdealLoop::is_divisor_counted_loop_phi(const Node* divisor, const Node* loop) { 246 return loop->is_BaseCountedLoop() && divisor->is_Phi() && divisor->in(0) == loop; 247 } 248 249 bool PhaseIdealLoop::loop_phi_backedge_type_contains_zero(const Node* phi_divisor, const Type* zero) const { 250 return _igvn.type(phi_divisor->in(LoopNode::LoopBackControl))->filter_speculative(zero) != Type::TOP; 251 } 252 253 //------------------------------dominated_by------------------------------------ 254 // Replace the dominated test with an obvious true or false. Place it on the 255 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the 256 // live path up to the dominating control. 257 void PhaseIdealLoop::dominated_by(IfProjNode* prevdom, IfNode* iff, bool flip, bool exclude_loop_predicate) { 258 if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); } 259 260 // prevdom is the dominating projection of the dominating test. 261 assert(iff->Opcode() == Op_If || 262 iff->Opcode() == Op_CountedLoopEnd || 263 iff->Opcode() == Op_LongCountedLoopEnd || 264 iff->Opcode() == Op_RangeCheck, 265 "Check this code when new subtype is added"); 266 267 int pop = prevdom->Opcode(); 268 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" ); 269 if (flip) { 270 if (pop == Op_IfTrue) 271 pop = Op_IfFalse; 272 else 273 pop = Op_IfTrue; 274 } 275 // 'con' is set to true or false to kill the dominated test. 276 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO); 277 set_ctrl(con, C->root()); // Constant gets a new use 278 // Hack the dominated test 279 _igvn.replace_input_of(iff, 1, con); 280 281 // If I dont have a reachable TRUE and FALSE path following the IfNode then 282 // I can assume this path reaches an infinite loop. In this case it's not 283 // important to optimize the data Nodes - either the whole compilation will 284 // be tossed or this path (and all data Nodes) will go dead. 285 if (iff->outcnt() != 2) return; 286 287 // Make control-dependent data Nodes on the live path (path that will remain 288 // once the dominated IF is removed) become control-dependent on the 289 // dominating projection. 290 Node* dp = iff->proj_out_or_null(pop == Op_IfTrue); 291 292 // Loop predicates may have depending checks which should not 293 // be skipped. For example, range check predicate has two checks 294 // for lower and upper bounds. 295 if (dp == NULL) 296 return; 297 298 ProjNode* dp_proj = dp->as_Proj(); 299 ProjNode* unc_proj = iff->proj_out(1 - dp_proj->_con)->as_Proj(); 300 if (exclude_loop_predicate && 301 (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) != NULL || 302 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_profile_predicate) != NULL || 303 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check) != NULL)) { 304 // If this is a range check (IfNode::is_range_check), do not 305 // reorder because Compile::allow_range_check_smearing might have 306 // changed the check. 307 return; // Let IGVN transformation change control dependence. 308 } 309 310 IdealLoopTree* old_loop = get_loop(dp); 311 312 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { 313 Node* cd = dp->fast_out(i); // Control-dependent node 314 // Do not rewire Div and Mod nodes which could have a zero divisor to avoid skipping their zero check. 315 if (cd->depends_only_on_test() && _igvn.no_dependent_zero_check(cd)) { 316 assert(cd->in(0) == dp, ""); 317 _igvn.replace_input_of(cd, 0, prevdom); 318 set_early_ctrl(cd, false); 319 IdealLoopTree* new_loop = get_loop(get_ctrl(cd)); 320 if (old_loop != new_loop) { 321 if (!old_loop->_child) { 322 old_loop->_body.yank(cd); 323 } 324 if (!new_loop->_child) { 325 new_loop->_body.push(cd); 326 } 327 } 328 --i; 329 --imax; 330 } 331 } 332 } 333 334 //------------------------------has_local_phi_input---------------------------- 335 // Return TRUE if 'n' has Phi inputs from its local block and no other 336 // block-local inputs (all non-local-phi inputs come from earlier blocks) 337 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) { 338 Node *n_ctrl = get_ctrl(n); 339 // See if some inputs come from a Phi in this block, or from before 340 // this block. 341 uint i; 342 for( i = 1; i < n->req(); i++ ) { 343 Node *phi = n->in(i); 344 if( phi->is_Phi() && phi->in(0) == n_ctrl ) 345 break; 346 } 347 if( i >= n->req() ) 348 return NULL; // No Phi inputs; nowhere to clone thru 349 350 // Check for inputs created between 'n' and the Phi input. These 351 // must split as well; they have already been given the chance 352 // (courtesy of a post-order visit) and since they did not we must 353 // recover the 'cost' of splitting them by being very profitable 354 // when splitting 'n'. Since this is unlikely we simply give up. 355 for( i = 1; i < n->req(); i++ ) { 356 Node *m = n->in(i); 357 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) { 358 // We allow the special case of AddP's with no local inputs. 359 // This allows us to split-up address expressions. 360 if (m->is_AddP() && 361 get_ctrl(m->in(AddPNode::Base)) != n_ctrl && 362 get_ctrl(m->in(AddPNode::Address)) != n_ctrl && 363 get_ctrl(m->in(AddPNode::Offset)) != n_ctrl) { 364 // Move the AddP up to the dominating point. That's fine because control of m's inputs 365 // must dominate get_ctrl(m) == n_ctrl and we just checked that the input controls are != n_ctrl. 366 Node* c = find_non_split_ctrl(idom(n_ctrl)); 367 if (c->is_OuterStripMinedLoop()) { 368 c->as_Loop()->verify_strip_mined(1); 369 c = c->in(LoopNode::EntryControl); 370 } 371 set_ctrl_and_loop(m, c); 372 continue; 373 } 374 return NULL; 375 } 376 assert(n->is_Phi() || m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control"); 377 } 378 379 return n_ctrl; 380 } 381 382 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2). 383 Node* PhaseIdealLoop::remix_address_expressions_add_left_shift(Node* n, IdealLoopTree* n_loop, Node* n_ctrl, BasicType bt) { 384 assert(bt == T_INT || bt == T_LONG, "only for integers"); 385 int n_op = n->Opcode(); 386 387 if (n_op == Op_LShift(bt)) { 388 // Scale is loop invariant 389 Node* scale = n->in(2); 390 Node* scale_ctrl = get_ctrl(scale); 391 IdealLoopTree* scale_loop = get_loop(scale_ctrl); 392 if (n_loop == scale_loop || !scale_loop->is_member(n_loop)) { 393 return NULL; 394 } 395 const TypeInt* scale_t = scale->bottom_type()->isa_int(); 396 if (scale_t != NULL && scale_t->is_con() && scale_t->get_con() >= 16) { 397 return NULL; // Dont bother with byte/short masking 398 } 399 // Add must vary with loop (else shift would be loop-invariant) 400 Node* add = n->in(1); 401 Node* add_ctrl = get_ctrl(add); 402 IdealLoopTree* add_loop = get_loop(add_ctrl); 403 if (n_loop != add_loop) { 404 return NULL; // happens w/ evil ZKM loops 405 } 406 407 // Convert I-V into I+ (0-V); same for V-I 408 if (add->Opcode() == Op_Sub(bt) && 409 _igvn.type(add->in(1)) != TypeInteger::zero(bt)) { 410 assert(add->Opcode() == Op_SubI || add->Opcode() == Op_SubL, ""); 411 Node* zero = _igvn.integercon(0, bt); 412 set_ctrl(zero, C->root()); 413 Node* neg = SubNode::make(zero, add->in(2), bt); 414 register_new_node(neg, get_ctrl(add->in(2))); 415 add = AddNode::make(add->in(1), neg, bt); 416 register_new_node(add, add_ctrl); 417 } 418 if (add->Opcode() != Op_Add(bt)) return NULL; 419 assert(add->Opcode() == Op_AddI || add->Opcode() == Op_AddL, ""); 420 // See if one add input is loop invariant 421 Node* add_var = add->in(1); 422 Node* add_var_ctrl = get_ctrl(add_var); 423 IdealLoopTree* add_var_loop = get_loop(add_var_ctrl); 424 Node* add_invar = add->in(2); 425 Node* add_invar_ctrl = get_ctrl(add_invar); 426 IdealLoopTree* add_invar_loop = get_loop(add_invar_ctrl); 427 if (add_invar_loop == n_loop) { 428 // Swap to find the invariant part 429 add_invar = add_var; 430 add_invar_ctrl = add_var_ctrl; 431 add_invar_loop = add_var_loop; 432 add_var = add->in(2); 433 } else if (add_var_loop != n_loop) { // Else neither input is loop invariant 434 return NULL; 435 } 436 if (n_loop == add_invar_loop || !add_invar_loop->is_member(n_loop)) { 437 return NULL; // No invariant part of the add? 438 } 439 440 // Yes! Reshape address expression! 441 Node* inv_scale = LShiftNode::make(add_invar, scale, bt); 442 Node* inv_scale_ctrl = 443 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ? 444 add_invar_ctrl : scale_ctrl; 445 register_new_node(inv_scale, inv_scale_ctrl); 446 Node* var_scale = LShiftNode::make(add_var, scale, bt); 447 register_new_node(var_scale, n_ctrl); 448 Node* var_add = AddNode::make(var_scale, inv_scale, bt); 449 register_new_node(var_add, n_ctrl); 450 _igvn.replace_node(n, var_add); 451 return var_add; 452 } 453 return NULL; 454 } 455 456 //------------------------------remix_address_expressions---------------------- 457 // Rework addressing expressions to get the most loop-invariant stuff 458 // moved out. We'd like to do all associative operators, but it's especially 459 // important (common) to do address expressions. 460 Node* PhaseIdealLoop::remix_address_expressions(Node* n) { 461 if (!has_ctrl(n)) return NULL; 462 Node* n_ctrl = get_ctrl(n); 463 IdealLoopTree* n_loop = get_loop(n_ctrl); 464 465 // See if 'n' mixes loop-varying and loop-invariant inputs and 466 // itself is loop-varying. 467 468 // Only interested in binary ops (and AddP) 469 if (n->req() < 3 || n->req() > 4) return NULL; 470 471 Node* n1_ctrl = get_ctrl(n->in( 1)); 472 Node* n2_ctrl = get_ctrl(n->in( 2)); 473 Node* n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3)); 474 IdealLoopTree* n1_loop = get_loop(n1_ctrl); 475 IdealLoopTree* n2_loop = get_loop(n2_ctrl); 476 IdealLoopTree* n3_loop = get_loop(n3_ctrl); 477 478 // Does one of my inputs spin in a tighter loop than self? 479 if ((n_loop->is_member(n1_loop) && n_loop != n1_loop) || 480 (n_loop->is_member(n2_loop) && n_loop != n2_loop) || 481 (n_loop->is_member(n3_loop) && n_loop != n3_loop)) { 482 return NULL; // Leave well enough alone 483 } 484 485 // Is at least one of my inputs loop-invariant? 486 if (n1_loop == n_loop && 487 n2_loop == n_loop && 488 n3_loop == n_loop) { 489 return NULL; // No loop-invariant inputs 490 } 491 492 Node* res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_INT); 493 if (res != NULL) { 494 return res; 495 } 496 res = remix_address_expressions_add_left_shift(n, n_loop, n_ctrl, T_LONG); 497 if (res != NULL) { 498 return res; 499 } 500 501 int n_op = n->Opcode(); 502 // Replace (I+V) with (V+I) 503 if (n_op == Op_AddI || 504 n_op == Op_AddL || 505 n_op == Op_AddF || 506 n_op == Op_AddD || 507 n_op == Op_MulI || 508 n_op == Op_MulL || 509 n_op == Op_MulF || 510 n_op == Op_MulD) { 511 if (n2_loop == n_loop) { 512 assert(n1_loop != n_loop, ""); 513 n->swap_edges(1, 2); 514 } 515 } 516 517 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V), 518 // but not if I2 is a constant. 519 if (n_op == Op_AddP) { 520 if (n2_loop == n_loop && n3_loop != n_loop) { 521 if (n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con()) { 522 Node* n22_ctrl = get_ctrl(n->in(2)->in(2)); 523 Node* n23_ctrl = get_ctrl(n->in(2)->in(3)); 524 IdealLoopTree* n22loop = get_loop(n22_ctrl); 525 IdealLoopTree* n23_loop = get_loop(n23_ctrl); 526 if (n22loop != n_loop && n22loop->is_member(n_loop) && 527 n23_loop == n_loop) { 528 Node* add1 = new AddPNode(n->in(1), n->in(2)->in(2), n->in(3)); 529 // Stuff new AddP in the loop preheader 530 register_new_node(add1, n_loop->_head->in(LoopNode::EntryControl)); 531 Node* add2 = new AddPNode(n->in(1), add1, n->in(2)->in(3)); 532 register_new_node(add2, n_ctrl); 533 _igvn.replace_node(n, add2); 534 return add2; 535 } 536 } 537 } 538 539 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V) 540 if (n2_loop != n_loop && n3_loop == n_loop) { 541 if (n->in(3)->Opcode() == Op_AddX) { 542 Node* V = n->in(3)->in(1); 543 Node* I = n->in(3)->in(2); 544 if (is_member(n_loop,get_ctrl(V))) { 545 } else { 546 Node *tmp = V; V = I; I = tmp; 547 } 548 if (!is_member(n_loop,get_ctrl(I))) { 549 Node* add1 = new AddPNode(n->in(1), n->in(2), I); 550 // Stuff new AddP in the loop preheader 551 register_new_node(add1, n_loop->_head->in(LoopNode::EntryControl)); 552 Node* add2 = new AddPNode(n->in(1), add1, V); 553 register_new_node(add2, n_ctrl); 554 _igvn.replace_node(n, add2); 555 return add2; 556 } 557 } 558 } 559 } 560 561 return NULL; 562 } 563 564 // Optimize ((in1[2*i] * in2[2*i]) + (in1[2*i+1] * in2[2*i+1])) 565 Node *PhaseIdealLoop::convert_add_to_muladd(Node* n) { 566 assert(n->Opcode() == Op_AddI, "sanity"); 567 Node * nn = NULL; 568 Node * in1 = n->in(1); 569 Node * in2 = n->in(2); 570 if (in1->Opcode() == Op_MulI && in2->Opcode() == Op_MulI) { 571 IdealLoopTree* loop_n = get_loop(get_ctrl(n)); 572 if (loop_n->is_counted() && 573 loop_n->_head->as_Loop()->is_valid_counted_loop(T_INT) && 574 Matcher::match_rule_supported(Op_MulAddVS2VI) && 575 Matcher::match_rule_supported(Op_MulAddS2I)) { 576 Node* mul_in1 = in1->in(1); 577 Node* mul_in2 = in1->in(2); 578 Node* mul_in3 = in2->in(1); 579 Node* mul_in4 = in2->in(2); 580 if (mul_in1->Opcode() == Op_LoadS && 581 mul_in2->Opcode() == Op_LoadS && 582 mul_in3->Opcode() == Op_LoadS && 583 mul_in4->Opcode() == Op_LoadS) { 584 IdealLoopTree* loop1 = get_loop(get_ctrl(mul_in1)); 585 IdealLoopTree* loop2 = get_loop(get_ctrl(mul_in2)); 586 IdealLoopTree* loop3 = get_loop(get_ctrl(mul_in3)); 587 IdealLoopTree* loop4 = get_loop(get_ctrl(mul_in4)); 588 IdealLoopTree* loop5 = get_loop(get_ctrl(in1)); 589 IdealLoopTree* loop6 = get_loop(get_ctrl(in2)); 590 // All nodes should be in the same counted loop. 591 if (loop_n == loop1 && loop_n == loop2 && loop_n == loop3 && 592 loop_n == loop4 && loop_n == loop5 && loop_n == loop6) { 593 Node* adr1 = mul_in1->in(MemNode::Address); 594 Node* adr2 = mul_in2->in(MemNode::Address); 595 Node* adr3 = mul_in3->in(MemNode::Address); 596 Node* adr4 = mul_in4->in(MemNode::Address); 597 if (adr1->is_AddP() && adr2->is_AddP() && adr3->is_AddP() && adr4->is_AddP()) { 598 if ((adr1->in(AddPNode::Base) == adr3->in(AddPNode::Base)) && 599 (adr2->in(AddPNode::Base) == adr4->in(AddPNode::Base))) { 600 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in3, mul_in4); 601 register_new_node(nn, get_ctrl(n)); 602 _igvn.replace_node(n, nn); 603 return nn; 604 } else if ((adr1->in(AddPNode::Base) == adr4->in(AddPNode::Base)) && 605 (adr2->in(AddPNode::Base) == adr3->in(AddPNode::Base))) { 606 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in4, mul_in3); 607 register_new_node(nn, get_ctrl(n)); 608 _igvn.replace_node(n, nn); 609 return nn; 610 } 611 } 612 } 613 } 614 } 615 } 616 return nn; 617 } 618 619 //------------------------------conditional_move------------------------------- 620 // Attempt to replace a Phi with a conditional move. We have some pretty 621 // strict profitability requirements. All Phis at the merge point must 622 // be converted, so we can remove the control flow. We need to limit the 623 // number of c-moves to a small handful. All code that was in the side-arms 624 // of the CFG diamond is now speculatively executed. This code has to be 625 // "cheap enough". We are pretty much limited to CFG diamonds that merge 626 // 1 or 2 items with a total of 1 or 2 ops executed speculatively. 627 Node *PhaseIdealLoop::conditional_move( Node *region ) { 628 629 assert(region->is_Region(), "sanity check"); 630 if (region->req() != 3) return NULL; 631 632 // Check for CFG diamond 633 Node *lp = region->in(1); 634 Node *rp = region->in(2); 635 if (!lp || !rp) return NULL; 636 Node *lp_c = lp->in(0); 637 if (lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If()) return NULL; 638 IfNode *iff = lp_c->as_If(); 639 640 // Check for ops pinned in an arm of the diamond. 641 // Can't remove the control flow in this case 642 if (lp->outcnt() > 1) return NULL; 643 if (rp->outcnt() > 1) return NULL; 644 645 IdealLoopTree* r_loop = get_loop(region); 646 assert(r_loop == get_loop(iff), "sanity"); 647 // Always convert to CMOVE if all results are used only outside this loop. 648 bool used_inside_loop = (r_loop == _ltree_root); 649 650 // Check profitability 651 int cost = 0; 652 int phis = 0; 653 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 654 Node *out = region->fast_out(i); 655 if (!out->is_Phi()) continue; // Ignore other control edges, etc 656 phis++; 657 PhiNode* phi = out->as_Phi(); 658 BasicType bt = phi->type()->basic_type(); 659 switch (bt) { 660 case T_DOUBLE: 661 case T_FLOAT: 662 if (C->use_cmove()) { 663 continue; //TODO: maybe we want to add some cost 664 } 665 cost += Matcher::float_cmove_cost(); // Could be very expensive 666 break; 667 case T_LONG: { 668 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's 669 } 670 case T_INT: // These all CMOV fine 671 case T_ADDRESS: { // (RawPtr) 672 cost++; 673 break; 674 } 675 case T_NARROWOOP: // Fall through 676 case T_OBJECT: { // Base oops are OK, but not derived oops 677 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr(); 678 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a 679 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus 680 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we 681 // have a Phi for the base here that we convert to a CMOVE all is well 682 // and good. But if the base is dead, we'll not make a CMOVE. Later 683 // the allocator will have to produce a base by creating a CMOVE of the 684 // relevant bases. This puts the allocator in the business of 685 // manufacturing expensive instructions, generally a bad plan. 686 // Just Say No to Conditionally-Moved Derived Pointers. 687 if (tp && tp->offset() != 0) 688 return NULL; 689 cost++; 690 break; 691 } 692 default: 693 return NULL; // In particular, can't do memory or I/O 694 } 695 // Add in cost any speculative ops 696 for (uint j = 1; j < region->req(); j++) { 697 Node *proj = region->in(j); 698 Node *inp = phi->in(j); 699 if (get_ctrl(inp) == proj) { // Found local op 700 cost++; 701 // Check for a chain of dependent ops; these will all become 702 // speculative in a CMOV. 703 for (uint k = 1; k < inp->req(); k++) 704 if (get_ctrl(inp->in(k)) == proj) 705 cost += ConditionalMoveLimit; // Too much speculative goo 706 } 707 } 708 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode. 709 // This will likely Split-If, a higher-payoff operation. 710 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) { 711 Node* use = phi->fast_out(k); 712 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr()) 713 cost += ConditionalMoveLimit; 714 // Is there a use inside the loop? 715 // Note: check only basic types since CMoveP is pinned. 716 if (!used_inside_loop && is_java_primitive(bt)) { 717 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use); 718 if (r_loop == u_loop || r_loop->is_member(u_loop)) { 719 used_inside_loop = true; 720 } 721 } 722 } 723 }//for 724 Node* bol = iff->in(1); 725 if (bol->Opcode() == Op_Opaque4) { 726 return NULL; // Ignore loop predicate checks (the Opaque4 ensures they will go away) 727 } 728 assert(bol->Opcode() == Op_Bool, "Unexpected node"); 729 int cmp_op = bol->in(1)->Opcode(); 730 if (cmp_op == Op_SubTypeCheck) { // SubTypeCheck expansion expects an IfNode 731 return NULL; 732 } 733 // It is expensive to generate flags from a float compare. 734 // Avoid duplicated float compare. 735 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL; 736 737 float infrequent_prob = PROB_UNLIKELY_MAG(3); 738 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop. 739 if (used_inside_loop) { 740 if (cost >= ConditionalMoveLimit) return NULL; // Too much goo 741 742 // BlockLayoutByFrequency optimization moves infrequent branch 743 // from hot path. No point in CMOV'ing in such case (110 is used 744 // instead of 100 to take into account not exactness of float value). 745 if (BlockLayoutByFrequency) { 746 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f); 747 } 748 } 749 // Check for highly predictable branch. No point in CMOV'ing if 750 // we are going to predict accurately all the time. 751 if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) { 752 //keep going 753 } else if (iff->_prob < infrequent_prob || 754 iff->_prob > (1.0f - infrequent_prob)) 755 return NULL; 756 757 // -------------- 758 // Now replace all Phis with CMOV's 759 Node *cmov_ctrl = iff->in(0); 760 uint flip = (lp->Opcode() == Op_IfTrue); 761 Node_List wq; 762 while (1) { 763 PhiNode* phi = NULL; 764 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 765 Node *out = region->fast_out(i); 766 if (out->is_Phi()) { 767 phi = out->as_Phi(); 768 break; 769 } 770 } 771 if (phi == NULL || _igvn.type(phi) == Type::TOP) { 772 break; 773 } 774 if (PrintOpto && VerifyLoopOptimizations) { tty->print_cr("CMOV"); } 775 // Move speculative ops 776 wq.push(phi); 777 while (wq.size() > 0) { 778 Node *n = wq.pop(); 779 for (uint j = 1; j < n->req(); j++) { 780 Node* m = n->in(j); 781 if (m != NULL && !is_dominator(get_ctrl(m), cmov_ctrl)) { 782 #ifndef PRODUCT 783 if (PrintOpto && VerifyLoopOptimizations) { 784 tty->print(" speculate: "); 785 m->dump(); 786 } 787 #endif 788 set_ctrl(m, cmov_ctrl); 789 wq.push(m); 790 } 791 } 792 } 793 Node *cmov = CMoveNode::make(cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi)); 794 register_new_node( cmov, cmov_ctrl ); 795 _igvn.replace_node( phi, cmov ); 796 #ifndef PRODUCT 797 if (TraceLoopOpts) { 798 tty->print("CMOV "); 799 r_loop->dump_head(); 800 if (Verbose) { 801 bol->in(1)->dump(1); 802 cmov->dump(1); 803 } 804 } 805 if (VerifyLoopOptimizations) verify(); 806 #endif 807 } 808 809 // The useless CFG diamond will fold up later; see the optimization in 810 // RegionNode::Ideal. 811 _igvn._worklist.push(region); 812 813 return iff->in(1); 814 } 815 816 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) { 817 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 818 Node* u = m->fast_out(i); 819 if (u->is_CFG()) { 820 if (u->is_NeverBranch()) { 821 u = u->as_NeverBranch()->proj_out(0); 822 enqueue_cfg_uses(u, wq); 823 } else { 824 wq.push(u); 825 } 826 } 827 } 828 } 829 830 // Try moving a store out of a loop, right before the loop 831 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) { 832 // Store has to be first in the loop body 833 IdealLoopTree *n_loop = get_loop(n_ctrl); 834 if (n->is_Store() && n_loop != _ltree_root && 835 n_loop->is_loop() && n_loop->_head->is_Loop() && 836 n->in(0) != NULL) { 837 Node* address = n->in(MemNode::Address); 838 Node* value = n->in(MemNode::ValueIn); 839 Node* mem = n->in(MemNode::Memory); 840 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 841 IdealLoopTree* value_loop = get_loop(get_ctrl(value)); 842 843 // - address and value must be loop invariant 844 // - memory must be a memory Phi for the loop 845 // - Store must be the only store on this memory slice in the 846 // loop: if there's another store following this one then value 847 // written at iteration i by the second store could be overwritten 848 // at iteration i+n by the first store: it's not safe to move the 849 // first store out of the loop 850 // - nothing must observe the memory Phi: it guarantees no read 851 // before the store, we are also guaranteed the store post 852 // dominates the loop head (ignoring a possible early 853 // exit). Otherwise there would be extra Phi involved between the 854 // loop's Phi and the store. 855 // - there must be no early exit from the loop before the Store 856 // (such an exit most of the time would be an extra use of the 857 // memory Phi but sometimes is a bottom memory Phi that takes the 858 // store as input). 859 860 if (!n_loop->is_member(address_loop) && 861 !n_loop->is_member(value_loop) && 862 mem->is_Phi() && mem->in(0) == n_loop->_head && 863 mem->outcnt() == 1 && 864 mem->in(LoopNode::LoopBackControl) == n) { 865 866 assert(n_loop->_tail != NULL, "need a tail"); 867 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop"); 868 869 // Verify that there's no early exit of the loop before the store. 870 bool ctrl_ok = false; 871 { 872 // Follow control from loop head until n, we exit the loop or 873 // we reach the tail 874 ResourceMark rm; 875 Unique_Node_List wq; 876 wq.push(n_loop->_head); 877 878 for (uint next = 0; next < wq.size(); ++next) { 879 Node *m = wq.at(next); 880 if (m == n->in(0)) { 881 ctrl_ok = true; 882 continue; 883 } 884 assert(!has_ctrl(m), "should be CFG"); 885 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) { 886 ctrl_ok = false; 887 break; 888 } 889 enqueue_cfg_uses(m, wq); 890 if (wq.size() > 10) { 891 ctrl_ok = false; 892 break; 893 } 894 } 895 } 896 if (ctrl_ok) { 897 // move the Store 898 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem); 899 _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl)); 900 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl)); 901 // Disconnect the phi now. An empty phi can confuse other 902 // optimizations in this pass of loop opts. 903 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl)); 904 n_loop->_body.yank(mem); 905 906 set_ctrl_and_loop(n, n->in(0)); 907 908 return n; 909 } 910 } 911 } 912 return NULL; 913 } 914 915 // Try moving a store out of a loop, right after the loop 916 void PhaseIdealLoop::try_move_store_after_loop(Node* n) { 917 if (n->is_Store() && n->in(0) != NULL) { 918 Node *n_ctrl = get_ctrl(n); 919 IdealLoopTree *n_loop = get_loop(n_ctrl); 920 // Store must be in a loop 921 if (n_loop != _ltree_root && !n_loop->_irreducible) { 922 Node* address = n->in(MemNode::Address); 923 Node* value = n->in(MemNode::ValueIn); 924 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 925 // address must be loop invariant 926 if (!n_loop->is_member(address_loop)) { 927 // Store must be last on this memory slice in the loop and 928 // nothing in the loop must observe it 929 Node* phi = NULL; 930 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 931 Node* u = n->fast_out(i); 932 if (has_ctrl(u)) { // control use? 933 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 934 if (!n_loop->is_member(u_loop)) { 935 continue; 936 } 937 if (u->is_Phi() && u->in(0) == n_loop->_head) { 938 assert(_igvn.type(u) == Type::MEMORY, "bad phi"); 939 // multiple phis on the same slice are possible 940 if (phi != NULL) { 941 return; 942 } 943 phi = u; 944 continue; 945 } 946 } 947 return; 948 } 949 if (phi != NULL) { 950 // Nothing in the loop before the store (next iteration) 951 // must observe the stored value 952 bool mem_ok = true; 953 { 954 ResourceMark rm; 955 Unique_Node_List wq; 956 wq.push(phi); 957 for (uint next = 0; next < wq.size() && mem_ok; ++next) { 958 Node *m = wq.at(next); 959 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) { 960 Node* u = m->fast_out(i); 961 if (u->is_Store() || u->is_Phi()) { 962 if (u != n) { 963 wq.push(u); 964 mem_ok = (wq.size() <= 10); 965 } 966 } else { 967 mem_ok = false; 968 break; 969 } 970 } 971 } 972 } 973 if (mem_ok) { 974 // Move the store out of the loop if the LCA of all 975 // users (except for the phi) is outside the loop. 976 Node* hook = new Node(1); 977 hook->init_req(0, n_ctrl); // Add an input to prevent hook from being dead 978 _igvn.rehash_node_delayed(phi); 979 int count = phi->replace_edge(n, hook, &_igvn); 980 assert(count > 0, "inconsistent phi"); 981 982 // Compute latest point this store can go 983 Node* lca = get_late_ctrl(n, get_ctrl(n)); 984 if (lca->is_OuterStripMinedLoop()) { 985 lca = lca->in(LoopNode::EntryControl); 986 } 987 if (n_loop->is_member(get_loop(lca))) { 988 // LCA is in the loop - bail out 989 _igvn.replace_node(hook, n); 990 return; 991 } 992 #ifdef ASSERT 993 if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) { 994 assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined"); 995 n_loop->_head->as_Loop()->verify_strip_mined(1); 996 Node* outer = n_loop->_head->as_CountedLoop()->outer_loop(); 997 IdealLoopTree* outer_loop = get_loop(outer); 998 assert(n_loop->_parent == outer_loop, "broken loop tree"); 999 assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state"); 1000 } 1001 #endif 1002 lca = place_outside_loop(lca, n_loop); 1003 assert(!n_loop->is_member(get_loop(lca)), "control must not be back in the loop"); 1004 assert(get_loop(lca)->_nest < n_loop->_nest || lca->in(0)->is_NeverBranch(), "must not be moved into inner loop"); 1005 1006 // Move store out of the loop 1007 _igvn.replace_node(hook, n->in(MemNode::Memory)); 1008 _igvn.replace_input_of(n, 0, lca); 1009 set_ctrl_and_loop(n, lca); 1010 1011 // Disconnect the phi now. An empty phi can confuse other 1012 // optimizations in this pass of loop opts.. 1013 if (phi->in(LoopNode::LoopBackControl) == phi) { 1014 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl)); 1015 n_loop->_body.yank(phi); 1016 } 1017 } 1018 } 1019 } 1020 } 1021 } 1022 } 1023 1024 //------------------------------split_if_with_blocks_pre----------------------- 1025 // Do the real work in a non-recursive function. Data nodes want to be 1026 // cloned in the pre-order so they can feed each other nicely. 1027 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) { 1028 // Cloning these guys is unlikely to win 1029 int n_op = n->Opcode(); 1030 if (n_op == Op_MergeMem) { 1031 return n; 1032 } 1033 if (n->is_Proj()) { 1034 return n; 1035 } 1036 // Do not clone-up CmpFXXX variations, as these are always 1037 // followed by a CmpI 1038 if (n->is_Cmp()) { 1039 return n; 1040 } 1041 // Attempt to use a conditional move instead of a phi/branch 1042 if (ConditionalMoveLimit > 0 && n_op == Op_Region) { 1043 Node *cmov = conditional_move( n ); 1044 if (cmov) { 1045 return cmov; 1046 } 1047 } 1048 if (n->is_CFG() || n->is_LoadStore()) { 1049 return n; 1050 } 1051 if (n->is_Opaque1()) { // Opaque nodes cannot be mod'd 1052 if (!C->major_progress()) { // If chance of no more loop opts... 1053 _igvn._worklist.push(n); // maybe we'll remove them 1054 } 1055 return n; 1056 } 1057 1058 if (n->is_Con()) { 1059 return n; // No cloning for Con nodes 1060 } 1061 1062 Node *n_ctrl = get_ctrl(n); 1063 if (!n_ctrl) { 1064 return n; // Dead node 1065 } 1066 1067 Node* res = try_move_store_before_loop(n, n_ctrl); 1068 if (res != NULL) { 1069 return n; 1070 } 1071 1072 // Attempt to remix address expressions for loop invariants 1073 Node *m = remix_address_expressions( n ); 1074 if( m ) return m; 1075 1076 if (n_op == Op_AddI) { 1077 Node *nn = convert_add_to_muladd( n ); 1078 if ( nn ) return nn; 1079 } 1080 1081 if (n->is_ConstraintCast()) { 1082 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this); 1083 // ConstraintCastNode::dominating_cast() uses node control input to determine domination. 1084 // Node control inputs don't necessarily agree with loop control info (due to 1085 // transformations happened in between), thus additional dominance check is needed 1086 // to keep loop info valid. 1087 if (dom_cast != NULL && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) { 1088 _igvn.replace_node(n, dom_cast); 1089 return dom_cast; 1090 } 1091 } 1092 1093 // Determine if the Node has inputs from some local Phi. 1094 // Returns the block to clone thru. 1095 Node *n_blk = has_local_phi_input( n ); 1096 if( !n_blk ) return n; 1097 1098 // Do not clone the trip counter through on a CountedLoop 1099 // (messes up the canonical shape). 1100 if (((n_blk->is_CountedLoop() || (n_blk->is_Loop() && n_blk->as_Loop()->is_loop_nest_inner_loop())) && n->Opcode() == Op_AddI) || 1101 (n_blk->is_LongCountedLoop() && n->Opcode() == Op_AddL)) { 1102 return n; 1103 } 1104 // Pushing a shift through the iv Phi can get in the way of addressing optimizations or range check elimination 1105 if (n_blk->is_BaseCountedLoop() && n->Opcode() == Op_LShift(n_blk->as_BaseCountedLoop()->bt()) && 1106 n->in(1) == n_blk->as_BaseCountedLoop()->phi()) { 1107 return n; 1108 } 1109 1110 // Check for having no control input; not pinned. Allow 1111 // dominating control. 1112 if (n->in(0)) { 1113 Node *dom = idom(n_blk); 1114 if (dom_lca(n->in(0), dom) != n->in(0)) { 1115 return n; 1116 } 1117 } 1118 // Policy: when is it profitable. You must get more wins than 1119 // policy before it is considered profitable. Policy is usually 0, 1120 // so 1 win is considered profitable. Big merges will require big 1121 // cloning, so get a larger policy. 1122 int policy = n_blk->req() >> 2; 1123 1124 // If the loop is a candidate for range check elimination, 1125 // delay splitting through it's phi until a later loop optimization 1126 if (n_blk->is_BaseCountedLoop()) { 1127 IdealLoopTree *lp = get_loop(n_blk); 1128 if (lp && lp->_rce_candidate) { 1129 return n; 1130 } 1131 } 1132 1133 if (must_throttle_split_if()) return n; 1134 1135 // Split 'n' through the merge point if it is profitable 1136 Node *phi = split_thru_phi( n, n_blk, policy ); 1137 if (!phi) return n; 1138 1139 // Found a Phi to split thru! 1140 // Replace 'n' with the new phi 1141 _igvn.replace_node( n, phi ); 1142 // Moved a load around the loop, 'en-registering' something. 1143 if (n_blk->is_Loop() && n->is_Load() && 1144 !phi->in(LoopNode::LoopBackControl)->is_Load()) 1145 C->set_major_progress(); 1146 1147 return phi; 1148 } 1149 1150 static bool merge_point_too_heavy(Compile* C, Node* region) { 1151 // Bail out if the region and its phis have too many users. 1152 int weight = 0; 1153 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1154 weight += region->fast_out(i)->outcnt(); 1155 } 1156 int nodes_left = C->max_node_limit() - C->live_nodes(); 1157 if (weight * 8 > nodes_left) { 1158 if (PrintOpto) { 1159 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight); 1160 } 1161 return true; 1162 } else { 1163 return false; 1164 } 1165 } 1166 1167 static bool merge_point_safe(Node* region) { 1168 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode 1169 // having a PhiNode input. This sidesteps the dangerous case where the split 1170 // ConvI2LNode may become TOP if the input Value() does not 1171 // overlap the ConvI2L range, leaving a node which may not dominate its 1172 // uses. 1173 // A better fix for this problem can be found in the BugTraq entry, but 1174 // expediency for Mantis demands this hack. 1175 #ifdef _LP64 1176 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1177 Node* n = region->fast_out(i); 1178 if (n->is_Phi()) { 1179 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1180 Node* m = n->fast_out(j); 1181 if (m->Opcode() == Op_ConvI2L) 1182 return false; 1183 if (m->is_CastII()) { 1184 return false; 1185 } 1186 } 1187 } 1188 } 1189 #endif 1190 return true; 1191 } 1192 1193 1194 //------------------------------place_outside_loop--------------------------------- 1195 // Place some computation outside of this loop on the path to the use passed as argument 1196 Node* PhaseIdealLoop::place_outside_loop(Node* useblock, IdealLoopTree* loop) const { 1197 Node* head = loop->_head; 1198 assert(!loop->is_member(get_loop(useblock)), "must be outside loop"); 1199 if (head->is_Loop() && head->as_Loop()->is_strip_mined()) { 1200 loop = loop->_parent; 1201 assert(loop->_head->is_OuterStripMinedLoop(), "malformed strip mined loop"); 1202 } 1203 1204 // Pick control right outside the loop 1205 for (;;) { 1206 Node* dom = idom(useblock); 1207 if (loop->is_member(get_loop(dom)) || 1208 // NeverBranch nodes are not assigned to the loop when constructed 1209 (dom->is_NeverBranch() && loop->is_member(get_loop(dom->in(0))))) { 1210 break; 1211 } 1212 useblock = dom; 1213 } 1214 assert(find_non_split_ctrl(useblock) == useblock, "should be non split control"); 1215 return useblock; 1216 } 1217 1218 1219 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) { 1220 if (!n->is_If() || n->is_BaseCountedLoopEnd()) { 1221 return false; 1222 } 1223 if (!n->in(0)->is_Region()) { 1224 return false; 1225 } 1226 1227 Node* region = n->in(0); 1228 Node* dom = idom(region); 1229 if (!dom->is_If() || dom->in(1) != n->in(1)) { 1230 return false; 1231 } 1232 IfNode* dom_if = dom->as_If(); 1233 Node* proj_true = dom_if->proj_out(1); 1234 Node* proj_false = dom_if->proj_out(0); 1235 1236 for (uint i = 1; i < region->req(); i++) { 1237 if (is_dominator(proj_true, region->in(i))) { 1238 continue; 1239 } 1240 if (is_dominator(proj_false, region->in(i))) { 1241 continue; 1242 } 1243 return false; 1244 } 1245 1246 return true; 1247 } 1248 1249 1250 bool PhaseIdealLoop::can_split_if(Node* n_ctrl) { 1251 if (must_throttle_split_if()) { 1252 return false; 1253 } 1254 1255 // Do not do 'split-if' if irreducible loops are present. 1256 if (_has_irreducible_loops) { 1257 return false; 1258 } 1259 1260 if (merge_point_too_heavy(C, n_ctrl)) { 1261 return false; 1262 } 1263 1264 // Do not do 'split-if' if some paths are dead. First do dead code 1265 // elimination and then see if its still profitable. 1266 for (uint i = 1; i < n_ctrl->req(); i++) { 1267 if (n_ctrl->in(i) == C->top()) { 1268 return false; 1269 } 1270 } 1271 1272 // If trying to do a 'Split-If' at the loop head, it is only 1273 // profitable if the cmp folds up on BOTH paths. Otherwise we 1274 // risk peeling a loop forever. 1275 1276 // CNC - Disabled for now. Requires careful handling of loop 1277 // body selection for the cloned code. Also, make sure we check 1278 // for any input path not being in the same loop as n_ctrl. For 1279 // irreducible loops we cannot check for 'n_ctrl->is_Loop()' 1280 // because the alternative loop entry points won't be converted 1281 // into LoopNodes. 1282 IdealLoopTree *n_loop = get_loop(n_ctrl); 1283 for (uint j = 1; j < n_ctrl->req(); j++) { 1284 if (get_loop(n_ctrl->in(j)) != n_loop) { 1285 return false; 1286 } 1287 } 1288 1289 // Check for safety of the merge point. 1290 if (!merge_point_safe(n_ctrl)) { 1291 return false; 1292 } 1293 1294 return true; 1295 } 1296 1297 // Detect if the node is the inner strip-mined loop 1298 // Return: NULL if it's not the case, or the exit of outer strip-mined loop 1299 static Node* is_inner_of_stripmined_loop(const Node* out) { 1300 Node* out_le = NULL; 1301 1302 if (out->is_CountedLoopEnd()) { 1303 const CountedLoopNode* loop = out->as_CountedLoopEnd()->loopnode(); 1304 1305 if (loop != NULL && loop->is_strip_mined()) { 1306 out_le = loop->in(LoopNode::EntryControl)->as_OuterStripMinedLoop()->outer_loop_exit(); 1307 } 1308 } 1309 1310 return out_le; 1311 } 1312 1313 //------------------------------split_if_with_blocks_post---------------------- 1314 // Do the real work in a non-recursive function. CFG hackery wants to be 1315 // in the post-order, so it can dirty the I-DOM info and not use the dirtied 1316 // info. 1317 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) { 1318 1319 // Cloning Cmp through Phi's involves the split-if transform. 1320 // FastLock is not used by an If 1321 if (n->is_Cmp() && !n->is_FastLock()) { 1322 Node *n_ctrl = get_ctrl(n); 1323 // Determine if the Node has inputs from some local Phi. 1324 // Returns the block to clone thru. 1325 Node *n_blk = has_local_phi_input(n); 1326 if (n_blk != n_ctrl) { 1327 return; 1328 } 1329 1330 if (!can_split_if(n_ctrl)) { 1331 return; 1332 } 1333 1334 if (n->outcnt() != 1) { 1335 return; // Multiple bool's from 1 compare? 1336 } 1337 Node *bol = n->unique_out(); 1338 assert(bol->is_Bool(), "expect a bool here"); 1339 if (bol->outcnt() != 1) { 1340 return;// Multiple branches from 1 compare? 1341 } 1342 Node *iff = bol->unique_out(); 1343 1344 // Check some safety conditions 1345 if (iff->is_If()) { // Classic split-if? 1346 if (iff->in(0) != n_ctrl) { 1347 return; // Compare must be in same blk as if 1348 } 1349 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE 1350 // Can't split CMove with different control edge. 1351 if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) { 1352 return; 1353 } 1354 if (get_ctrl(iff->in(2)) == n_ctrl || 1355 get_ctrl(iff->in(3)) == n_ctrl) { 1356 return; // Inputs not yet split-up 1357 } 1358 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) { 1359 return; // Loop-invar test gates loop-varying CMOVE 1360 } 1361 } else { 1362 return; // some other kind of node, such as an Allocate 1363 } 1364 1365 // When is split-if profitable? Every 'win' on means some control flow 1366 // goes dead, so it's almost always a win. 1367 int policy = 0; 1368 // Split compare 'n' through the merge point if it is profitable 1369 Node *phi = split_thru_phi( n, n_ctrl, policy); 1370 if (!phi) { 1371 return; 1372 } 1373 1374 // Found a Phi to split thru! 1375 // Replace 'n' with the new phi 1376 _igvn.replace_node(n, phi); 1377 1378 // Now split the bool up thru the phi 1379 Node *bolphi = split_thru_phi(bol, n_ctrl, -1); 1380 guarantee(bolphi != NULL, "null boolean phi node"); 1381 1382 _igvn.replace_node(bol, bolphi); 1383 assert(iff->in(1) == bolphi, ""); 1384 1385 if (bolphi->Value(&_igvn)->singleton()) { 1386 return; 1387 } 1388 1389 // Conditional-move? Must split up now 1390 if (!iff->is_If()) { 1391 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1); 1392 _igvn.replace_node(iff, cmovphi); 1393 return; 1394 } 1395 1396 // Now split the IF 1397 do_split_if(iff); 1398 return; 1399 } 1400 1401 // Two identical ifs back to back can be merged 1402 if (try_merge_identical_ifs(n)) { 1403 return; 1404 } 1405 1406 // Check for an IF ready to split; one that has its 1407 // condition codes input coming from a Phi at the block start. 1408 int n_op = n->Opcode(); 1409 1410 // Check for an IF being dominated by another IF same test 1411 if (n_op == Op_If || 1412 n_op == Op_RangeCheck) { 1413 Node *bol = n->in(1); 1414 uint max = bol->outcnt(); 1415 // Check for same test used more than once? 1416 if (max > 1 && bol->is_Bool()) { 1417 // Search up IDOMs to see if this IF is dominated. 1418 Node *cutoff = get_ctrl(bol); 1419 1420 // Now search up IDOMs till cutoff, looking for a dominating test 1421 Node *prevdom = n; 1422 Node *dom = idom(prevdom); 1423 while (dom != cutoff) { 1424 if (dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom && 1425 safe_for_if_replacement(dom)) { 1426 // It's invalid to move control dependent data nodes in the inner 1427 // strip-mined loop, because: 1428 // 1) break validation of LoopNode::verify_strip_mined() 1429 // 2) move code with side-effect in strip-mined loop 1430 // Move to the exit of outer strip-mined loop in that case. 1431 Node* out_le = is_inner_of_stripmined_loop(dom); 1432 if (out_le != NULL) { 1433 prevdom = out_le; 1434 } 1435 // Replace the dominated test with an obvious true or false. 1436 // Place it on the IGVN worklist for later cleanup. 1437 C->set_major_progress(); 1438 dominated_by(prevdom->as_IfProj(), n->as_If(), false, true); 1439 #ifndef PRODUCT 1440 if( VerifyLoopOptimizations ) verify(); 1441 #endif 1442 return; 1443 } 1444 prevdom = dom; 1445 dom = idom(prevdom); 1446 } 1447 } 1448 } 1449 1450 try_sink_out_of_loop(n); 1451 1452 try_move_store_after_loop(n); 1453 } 1454 1455 // Transform: 1456 // 1457 // if (some_condition) { 1458 // // body 1 1459 // } else { 1460 // // body 2 1461 // } 1462 // if (some_condition) { 1463 // // body 3 1464 // } else { 1465 // // body 4 1466 // } 1467 // 1468 // into: 1469 // 1470 // 1471 // if (some_condition) { 1472 // // body 1 1473 // // body 3 1474 // } else { 1475 // // body 2 1476 // // body 4 1477 // } 1478 bool PhaseIdealLoop::try_merge_identical_ifs(Node* n) { 1479 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) { 1480 Node *n_ctrl = n->in(0); 1481 IfNode* dom_if = idom(n_ctrl)->as_If(); 1482 ProjNode* dom_proj_true = dom_if->proj_out(1); 1483 ProjNode* dom_proj_false = dom_if->proj_out(0); 1484 1485 // Now split the IF 1486 RegionNode* new_false_region; 1487 RegionNode* new_true_region; 1488 do_split_if(n, &new_false_region, &new_true_region); 1489 assert(new_false_region->req() == new_true_region->req(), ""); 1490 #ifdef ASSERT 1491 for (uint i = 1; i < new_false_region->req(); ++i) { 1492 assert(new_false_region->in(i)->in(0) == new_true_region->in(i)->in(0), "unexpected shape following split if"); 1493 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"); 1494 } 1495 #endif 1496 assert(new_false_region->in(1)->in(0)->in(1) == dom_if->in(1), "dominating if and dominated if after split must share test"); 1497 1498 // We now have: 1499 // if (some_condition) { 1500 // // body 1 1501 // if (some_condition) { 1502 // body3: // new_true_region 1503 // // body3 1504 // } else { 1505 // goto body4; 1506 // } 1507 // } else { 1508 // // body 2 1509 // if (some_condition) { 1510 // goto body3; 1511 // } else { 1512 // body4: // new_false_region 1513 // // body4; 1514 // } 1515 // } 1516 // 1517 1518 // clone pinned nodes thru the resulting regions 1519 push_pinned_nodes_thru_region(dom_if, new_true_region); 1520 push_pinned_nodes_thru_region(dom_if, new_false_region); 1521 1522 // Optimize out the cloned ifs. Because pinned nodes were cloned, this also allows a CastPP that would be dependent 1523 // 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 1524 // unrelated control dependency. 1525 for (uint i = 1; i < new_false_region->req(); i++) { 1526 if (is_dominator(dom_proj_true, new_false_region->in(i))) { 1527 dominated_by(dom_proj_true->as_IfProj(), new_false_region->in(i)->in(0)->as_If(), false, false); 1528 } else { 1529 assert(is_dominator(dom_proj_false, new_false_region->in(i)), "bad if"); 1530 dominated_by(dom_proj_false->as_IfProj(), new_false_region->in(i)->in(0)->as_If(), false, false); 1531 } 1532 } 1533 return true; 1534 } 1535 return false; 1536 } 1537 1538 void PhaseIdealLoop::push_pinned_nodes_thru_region(IfNode* dom_if, Node* region) { 1539 for (DUIterator i = region->outs(); region->has_out(i); i++) { 1540 Node* u = region->out(i); 1541 if (!has_ctrl(u) || u->is_Phi() || !u->depends_only_on_test() || !_igvn.no_dependent_zero_check(u)) { 1542 continue; 1543 } 1544 assert(u->in(0) == region, "not a control dependent node?"); 1545 uint j = 1; 1546 for (; j < u->req(); ++j) { 1547 Node* in = u->in(j); 1548 if (!is_dominator(ctrl_or_self(in), dom_if)) { 1549 break; 1550 } 1551 } 1552 if (j == u->req()) { 1553 Node *phi = PhiNode::make_blank(region, u); 1554 for (uint k = 1; k < region->req(); ++k) { 1555 Node* clone = u->clone(); 1556 clone->set_req(0, region->in(k)); 1557 register_new_node(clone, region->in(k)); 1558 phi->init_req(k, clone); 1559 } 1560 register_new_node(phi, region); 1561 _igvn.replace_node(u, phi); 1562 --i; 1563 } 1564 } 1565 } 1566 1567 bool PhaseIdealLoop::safe_for_if_replacement(const Node* dom) const { 1568 if (!dom->is_CountedLoopEnd()) { 1569 return true; 1570 } 1571 CountedLoopEndNode* le = dom->as_CountedLoopEnd(); 1572 CountedLoopNode* cl = le->loopnode(); 1573 if (cl == NULL) { 1574 return true; 1575 } 1576 if (!cl->is_main_loop()) { 1577 return true; 1578 } 1579 if (cl->is_canonical_loop_entry() == NULL) { 1580 return true; 1581 } 1582 // Further unrolling is possible so loop exit condition might change 1583 return false; 1584 } 1585 1586 // See if a shared loop-varying computation has no loop-varying uses. 1587 // Happens if something is only used for JVM state in uncommon trap exits, 1588 // like various versions of induction variable+offset. Clone the 1589 // computation per usage to allow it to sink out of the loop. 1590 void PhaseIdealLoop::try_sink_out_of_loop(Node* n) { 1591 if (has_ctrl(n) && 1592 !n->is_Phi() && 1593 !n->is_Bool() && 1594 !n->is_Proj() && 1595 !n->is_MergeMem() && 1596 !n->is_CMove() && 1597 n->Opcode() != Op_Opaque4 && 1598 !n->is_Type()) { 1599 Node *n_ctrl = get_ctrl(n); 1600 IdealLoopTree *n_loop = get_loop(n_ctrl); 1601 1602 if (n->in(0) != NULL) { 1603 IdealLoopTree* loop_ctrl = get_loop(n->in(0)); 1604 if (n_loop != loop_ctrl && n_loop->is_member(loop_ctrl)) { 1605 // n has a control input inside a loop but get_ctrl() is member of an outer loop. This could happen, for example, 1606 // for Div nodes inside a loop (control input inside loop) without a use except for an UCT (outside the loop). 1607 // Rewire control of n to right outside of the loop, regardless if its input(s) are later sunk or not. 1608 _igvn.replace_input_of(n, 0, place_outside_loop(n_ctrl, loop_ctrl)); 1609 } 1610 } 1611 if (n_loop != _ltree_root && n->outcnt() > 1) { 1612 // Compute early control: needed for anti-dependence analysis. It's also possible that as a result of 1613 // previous transformations in this loop opts round, the node can be hoisted now: early control will tell us. 1614 Node* early_ctrl = compute_early_ctrl(n, n_ctrl); 1615 if (n_loop->is_member(get_loop(early_ctrl)) && // check that this one can't be hoisted now 1616 ctrl_of_all_uses_out_of_loop(n, early_ctrl, n_loop)) { // All uses in outer loops! 1617 assert(!n->is_Store() && !n->is_LoadStore(), "no node with a side effect"); 1618 Node* outer_loop_clone = NULL; 1619 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin;) { 1620 Node* u = n->last_out(j); // Clone private computation per use 1621 _igvn.rehash_node_delayed(u); 1622 Node* x = n->clone(); // Clone computation 1623 Node* x_ctrl = NULL; 1624 if (u->is_Phi()) { 1625 // Replace all uses of normal nodes. Replace Phi uses 1626 // individually, so the separate Nodes can sink down 1627 // different paths. 1628 uint k = 1; 1629 while (u->in(k) != n) k++; 1630 u->set_req(k, x); 1631 // x goes next to Phi input path 1632 x_ctrl = u->in(0)->in(k); 1633 // Find control for 'x' next to use but not inside inner loops. 1634 x_ctrl = place_outside_loop(x_ctrl, n_loop); 1635 --j; 1636 } else { // Normal use 1637 if (has_ctrl(u)) { 1638 x_ctrl = get_ctrl(u); 1639 } else { 1640 x_ctrl = u->in(0); 1641 } 1642 // Find control for 'x' next to use but not inside inner loops. 1643 x_ctrl = place_outside_loop(x_ctrl, n_loop); 1644 // Replace all uses 1645 if (u->is_ConstraintCast() && u->bottom_type()->higher_equal(_igvn.type(n)) && u->in(0) == x_ctrl) { 1646 // If we're sinking a chain of data nodes, we might have inserted a cast to pin the use which is not necessary 1647 // anymore now that we're going to pin n as well 1648 _igvn.replace_node(u, x); 1649 --j; 1650 } else { 1651 int nb = u->replace_edge(n, x, &_igvn); 1652 j -= nb; 1653 } 1654 } 1655 1656 if (n->is_Load()) { 1657 // For loads, add a control edge to a CFG node outside of the loop 1658 // to force them to not combine and return back inside the loop 1659 // during GVN optimization (4641526). 1660 assert(x_ctrl == get_late_ctrl_with_anti_dep(x->as_Load(), early_ctrl, x_ctrl), "anti-dependences were already checked"); 1661 1662 IdealLoopTree* x_loop = get_loop(x_ctrl); 1663 Node* x_head = x_loop->_head; 1664 if (x_head->is_Loop() && x_head->is_OuterStripMinedLoop()) { 1665 // Do not add duplicate LoadNodes to the outer strip mined loop 1666 if (outer_loop_clone != NULL) { 1667 _igvn.replace_node(x, outer_loop_clone); 1668 continue; 1669 } 1670 outer_loop_clone = x; 1671 } 1672 x->set_req(0, x_ctrl); 1673 } else if (n->in(0) != NULL){ 1674 x->set_req(0, x_ctrl); 1675 } 1676 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone"); 1677 assert(!n_loop->is_member(get_loop(x_ctrl)), "should have moved out of loop"); 1678 register_new_node(x, x_ctrl); 1679 1680 // Chain of AddP: (AddP base (AddP base )) must keep the same base after sinking so: 1681 // 1- We don't add a CastPP here when the first one is sunk so if the second one is not, their bases remain 1682 // the same. 1683 // (see 2- below) 1684 assert(!x->is_AddP() || !x->in(AddPNode::Address)->is_AddP() || 1685 x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base) || 1686 !x->in(AddPNode::Address)->in(AddPNode::Base)->eqv_uncast(x->in(AddPNode::Base)), "unexpected AddP shape"); 1687 if (x->in(0) == NULL && !x->is_DecodeNarrowPtr() && 1688 !(x->is_AddP() && x->in(AddPNode::Address)->is_AddP() && x->in(AddPNode::Address)->in(AddPNode::Base) == x->in(AddPNode::Base))) { 1689 assert(!x->is_Load(), "load should be pinned"); 1690 // Use a cast node to pin clone out of loop 1691 Node* cast = NULL; 1692 for (uint k = 0; k < x->req(); k++) { 1693 Node* in = x->in(k); 1694 if (in != NULL && n_loop->is_member(get_loop(get_ctrl(in)))) { 1695 const Type* in_t = _igvn.type(in); 1696 cast = ConstraintCastNode::make_cast_for_type(x_ctrl, in, in_t, ConstraintCastNode::UnconditionalDependency); 1697 } 1698 if (cast != NULL) { 1699 register_new_node(cast, x_ctrl); 1700 x->replace_edge(in, cast); 1701 // Chain of AddP: 1702 // 2- A CastPP of the base is only added now that both AddP nodes are sunk 1703 if (x->is_AddP() && k == AddPNode::Base) { 1704 for (DUIterator_Fast imax, i = x->fast_outs(imax); i < imax; i++) { 1705 Node* u = x->fast_out(i); 1706 if (u->is_AddP() && u->in(AddPNode::Base) == n->in(AddPNode::Base)) { 1707 _igvn.replace_input_of(u, AddPNode::Base, cast); 1708 assert(u->find_out_with(Op_AddP) == NULL, "more than 2 chained AddP nodes?"); 1709 } 1710 } 1711 } 1712 break; 1713 } 1714 } 1715 assert(cast != NULL, "must have added a cast to pin the node"); 1716 } 1717 } 1718 _igvn.remove_dead_node(n); 1719 } 1720 _dom_lca_tags_round = 0; 1721 } 1722 } 1723 } 1724 1725 // Compute the early control of a node by following its inputs until we reach 1726 // nodes that are pinned. Then compute the LCA of the control of all pinned nodes. 1727 Node* PhaseIdealLoop::compute_early_ctrl(Node* n, Node* n_ctrl) { 1728 Node* early_ctrl = NULL; 1729 ResourceMark rm; 1730 Unique_Node_List wq; 1731 wq.push(n); 1732 for (uint i = 0; i < wq.size(); i++) { 1733 Node* m = wq.at(i); 1734 Node* c = NULL; 1735 if (m->is_CFG()) { 1736 c = m; 1737 } else if (m->pinned()) { 1738 c = m->in(0); 1739 } else { 1740 for (uint j = 0; j < m->req(); j++) { 1741 Node* in = m->in(j); 1742 if (in != NULL) { 1743 wq.push(in); 1744 } 1745 } 1746 } 1747 if (c != NULL) { 1748 assert(is_dominator(c, n_ctrl), "control input must dominate current control"); 1749 if (early_ctrl == NULL || is_dominator(early_ctrl, c)) { 1750 early_ctrl = c; 1751 } 1752 } 1753 } 1754 assert(is_dominator(early_ctrl, n_ctrl), "early control must dominate current control"); 1755 return early_ctrl; 1756 } 1757 1758 bool PhaseIdealLoop::ctrl_of_all_uses_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop) { 1759 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1760 Node* u = n->fast_out(i); 1761 if (u->is_Opaque1()) { 1762 return false; // Found loop limit, bugfix for 4677003 1763 } 1764 // We can't reuse tags in PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal() so make sure calls to 1765 // get_late_ctrl_with_anti_dep() use their own tag 1766 _dom_lca_tags_round++; 1767 assert(_dom_lca_tags_round != 0, "shouldn't wrap around"); 1768 1769 if (u->is_Phi()) { 1770 for (uint j = 1; j < u->req(); ++j) { 1771 if (u->in(j) == n && !ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, u->in(0)->in(j))) { 1772 return false; 1773 } 1774 } 1775 } else { 1776 Node* ctrl = has_ctrl(u) ? get_ctrl(u) : u->in(0); 1777 if (!ctrl_of_use_out_of_loop(n, n_ctrl, n_loop, ctrl)) { 1778 return false; 1779 } 1780 } 1781 } 1782 return true; 1783 } 1784 1785 bool PhaseIdealLoop::ctrl_of_use_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop, Node* ctrl) { 1786 if (n->is_Load()) { 1787 ctrl = get_late_ctrl_with_anti_dep(n->as_Load(), n_ctrl, ctrl); 1788 } 1789 IdealLoopTree *u_loop = get_loop(ctrl); 1790 if (u_loop == n_loop) { 1791 return false; // Found loop-varying use 1792 } 1793 if (n_loop->is_member(u_loop)) { 1794 return false; // Found use in inner loop 1795 } 1796 return true; 1797 } 1798 1799 //------------------------------split_if_with_blocks--------------------------- 1800 // Check for aggressive application of 'split-if' optimization, 1801 // using basic block level info. 1802 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack) { 1803 Node* root = C->root(); 1804 visited.set(root->_idx); // first, mark root as visited 1805 // Do pre-visit work for root 1806 Node* n = split_if_with_blocks_pre(root); 1807 uint cnt = n->outcnt(); 1808 uint i = 0; 1809 1810 while (true) { 1811 // Visit all children 1812 if (i < cnt) { 1813 Node* use = n->raw_out(i); 1814 ++i; 1815 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) { 1816 // Now do pre-visit work for this use 1817 use = split_if_with_blocks_pre(use); 1818 nstack.push(n, i); // Save parent and next use's index. 1819 n = use; // Process all children of current use. 1820 cnt = use->outcnt(); 1821 i = 0; 1822 } 1823 } 1824 else { 1825 // All of n's children have been processed, complete post-processing. 1826 if (cnt != 0 && !n->is_Con()) { 1827 assert(has_node(n), "no dead nodes"); 1828 split_if_with_blocks_post(n); 1829 } 1830 if (must_throttle_split_if()) { 1831 nstack.clear(); 1832 } 1833 if (nstack.is_empty()) { 1834 // Finished all nodes on stack. 1835 break; 1836 } 1837 // Get saved parent node and next use's index. Visit the rest of uses. 1838 n = nstack.node(); 1839 cnt = n->outcnt(); 1840 i = nstack.index(); 1841 nstack.pop(); 1842 } 1843 } 1844 } 1845 1846 1847 //============================================================================= 1848 // 1849 // C L O N E A L O O P B O D Y 1850 // 1851 1852 //------------------------------clone_iff-------------------------------------- 1853 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1854 // "Nearly" because all Nodes have been cloned from the original in the loop, 1855 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1856 // through the Phi recursively, and return a Bool. 1857 Node* PhaseIdealLoop::clone_iff(PhiNode* phi) { 1858 1859 // Convert this Phi into a Phi merging Bools 1860 uint i; 1861 for (i = 1; i < phi->req(); i++) { 1862 Node *b = phi->in(i); 1863 if (b->is_Phi()) { 1864 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi())); 1865 } else { 1866 assert(b->is_Bool() || b->Opcode() == Op_Opaque4, ""); 1867 } 1868 } 1869 1870 Node* n = phi->in(1); 1871 Node* sample_opaque = NULL; 1872 Node *sample_bool = NULL; 1873 if (n->Opcode() == Op_Opaque4) { 1874 sample_opaque = n; 1875 sample_bool = n->in(1); 1876 assert(sample_bool->is_Bool(), "wrong type"); 1877 } else { 1878 sample_bool = n; 1879 } 1880 Node *sample_cmp = sample_bool->in(1); 1881 1882 // Make Phis to merge the Cmp's inputs. 1883 PhiNode *phi1 = new PhiNode(phi->in(0), Type::TOP); 1884 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP); 1885 for (i = 1; i < phi->req(); i++) { 1886 Node *n1 = sample_opaque == NULL ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1); 1887 Node *n2 = sample_opaque == NULL ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2); 1888 phi1->set_req(i, n1); 1889 phi2->set_req(i, n2); 1890 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 1891 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 1892 } 1893 // See if these Phis have been made before. 1894 // Register with optimizer 1895 Node *hit1 = _igvn.hash_find_insert(phi1); 1896 if (hit1) { // Hit, toss just made Phi 1897 _igvn.remove_dead_node(phi1); // Remove new phi 1898 assert(hit1->is_Phi(), "" ); 1899 phi1 = (PhiNode*)hit1; // Use existing phi 1900 } else { // Miss 1901 _igvn.register_new_node_with_optimizer(phi1); 1902 } 1903 Node *hit2 = _igvn.hash_find_insert(phi2); 1904 if (hit2) { // Hit, toss just made Phi 1905 _igvn.remove_dead_node(phi2); // Remove new phi 1906 assert(hit2->is_Phi(), "" ); 1907 phi2 = (PhiNode*)hit2; // Use existing phi 1908 } else { // Miss 1909 _igvn.register_new_node_with_optimizer(phi2); 1910 } 1911 // Register Phis with loop/block info 1912 set_ctrl(phi1, phi->in(0)); 1913 set_ctrl(phi2, phi->in(0)); 1914 // Make a new Cmp 1915 Node *cmp = sample_cmp->clone(); 1916 cmp->set_req(1, phi1); 1917 cmp->set_req(2, phi2); 1918 _igvn.register_new_node_with_optimizer(cmp); 1919 set_ctrl(cmp, phi->in(0)); 1920 1921 // Make a new Bool 1922 Node *b = sample_bool->clone(); 1923 b->set_req(1,cmp); 1924 _igvn.register_new_node_with_optimizer(b); 1925 set_ctrl(b, phi->in(0)); 1926 1927 if (sample_opaque != NULL) { 1928 Node* opaque = sample_opaque->clone(); 1929 opaque->set_req(1, b); 1930 _igvn.register_new_node_with_optimizer(opaque); 1931 set_ctrl(opaque, phi->in(0)); 1932 return opaque; 1933 } 1934 1935 assert(b->is_Bool(), ""); 1936 return b; 1937 } 1938 1939 //------------------------------clone_bool------------------------------------- 1940 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1941 // "Nearly" because all Nodes have been cloned from the original in the loop, 1942 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1943 // through the Phi recursively, and return a Bool. 1944 CmpNode*PhaseIdealLoop::clone_bool(PhiNode* phi) { 1945 uint i; 1946 // Convert this Phi into a Phi merging Bools 1947 for( i = 1; i < phi->req(); i++ ) { 1948 Node *b = phi->in(i); 1949 if( b->is_Phi() ) { 1950 _igvn.replace_input_of(phi, i, clone_bool(b->as_Phi())); 1951 } else { 1952 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" ); 1953 } 1954 } 1955 1956 Node *sample_cmp = phi->in(1); 1957 1958 // Make Phis to merge the Cmp's inputs. 1959 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP ); 1960 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP ); 1961 for( uint j = 1; j < phi->req(); j++ ) { 1962 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP 1963 Node *n1, *n2; 1964 if( cmp_top->is_Cmp() ) { 1965 n1 = cmp_top->in(1); 1966 n2 = cmp_top->in(2); 1967 } else { 1968 n1 = n2 = cmp_top; 1969 } 1970 phi1->set_req( j, n1 ); 1971 phi2->set_req( j, n2 ); 1972 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 1973 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 1974 } 1975 1976 // See if these Phis have been made before. 1977 // Register with optimizer 1978 Node *hit1 = _igvn.hash_find_insert(phi1); 1979 if( hit1 ) { // Hit, toss just made Phi 1980 _igvn.remove_dead_node(phi1); // Remove new phi 1981 assert( hit1->is_Phi(), "" ); 1982 phi1 = (PhiNode*)hit1; // Use existing phi 1983 } else { // Miss 1984 _igvn.register_new_node_with_optimizer(phi1); 1985 } 1986 Node *hit2 = _igvn.hash_find_insert(phi2); 1987 if( hit2 ) { // Hit, toss just made Phi 1988 _igvn.remove_dead_node(phi2); // Remove new phi 1989 assert( hit2->is_Phi(), "" ); 1990 phi2 = (PhiNode*)hit2; // Use existing phi 1991 } else { // Miss 1992 _igvn.register_new_node_with_optimizer(phi2); 1993 } 1994 // Register Phis with loop/block info 1995 set_ctrl(phi1, phi->in(0)); 1996 set_ctrl(phi2, phi->in(0)); 1997 // Make a new Cmp 1998 Node *cmp = sample_cmp->clone(); 1999 cmp->set_req( 1, phi1 ); 2000 cmp->set_req( 2, phi2 ); 2001 _igvn.register_new_node_with_optimizer(cmp); 2002 set_ctrl(cmp, phi->in(0)); 2003 2004 assert( cmp->is_Cmp(), "" ); 2005 return (CmpNode*)cmp; 2006 } 2007 2008 //------------------------------sink_use--------------------------------------- 2009 // If 'use' was in the loop-exit block, it now needs to be sunk 2010 // below the post-loop merge point. 2011 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) { 2012 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) { 2013 set_ctrl(use, post_loop); 2014 for (DUIterator j = use->outs(); use->has_out(j); j++) 2015 sink_use(use->out(j), post_loop); 2016 } 2017 } 2018 2019 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new, 2020 IdealLoopTree* loop, IdealLoopTree* outer_loop, 2021 Node_List*& split_if_set, Node_List*& split_bool_set, 2022 Node_List*& split_cex_set, Node_List& worklist, 2023 uint new_counter, CloneLoopMode mode) { 2024 Node* nnn = old_new[old->_idx]; 2025 // Copy uses to a worklist, so I can munge the def-use info 2026 // with impunity. 2027 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 2028 worklist.push(old->fast_out(j)); 2029 2030 while( worklist.size() ) { 2031 Node *use = worklist.pop(); 2032 if (!has_node(use)) continue; // Ignore dead nodes 2033 if (use->in(0) == C->top()) continue; 2034 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 2035 // Check for data-use outside of loop - at least one of OLD or USE 2036 // must not be a CFG node. 2037 #ifdef ASSERT 2038 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == NULL) { 2039 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint(); 2040 assert(mode != IgnoreStripMined, "incorrect cloning mode"); 2041 assert((mode == ControlAroundStripMined && use == sfpt) || !use->is_reachable_from_root(), "missed a node"); 2042 } 2043 #endif 2044 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) { 2045 2046 // If the Data use is an IF, that means we have an IF outside of the 2047 // loop that is switching on a condition that is set inside of the 2048 // loop. Happens if people set a loop-exit flag; then test the flag 2049 // in the loop to break the loop, then test is again outside of the 2050 // loop to determine which way the loop exited. 2051 // Loop predicate If node connects to Bool node through Opaque1 node. 2052 // 2053 // If the use is an AllocateArray through its ValidLengthTest input, 2054 // make sure the Bool/Cmp input is cloned down to avoid a Phi between 2055 // the AllocateArray node and its ValidLengthTest input that could cause 2056 // split if to break. 2057 if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use) || use->Opcode() == Op_Opaque4 || 2058 (use->Opcode() == Op_AllocateArray && use->in(AllocateNode::ValidLengthTest) == old)) { 2059 // Since this code is highly unlikely, we lazily build the worklist 2060 // of such Nodes to go split. 2061 if (!split_if_set) { 2062 split_if_set = new Node_List(); 2063 } 2064 split_if_set->push(use); 2065 } 2066 if (use->is_Bool()) { 2067 if (!split_bool_set) { 2068 split_bool_set = new Node_List(); 2069 } 2070 split_bool_set->push(use); 2071 } 2072 if (use->Opcode() == Op_CreateEx) { 2073 if (!split_cex_set) { 2074 split_cex_set = new Node_List(); 2075 } 2076 split_cex_set->push(use); 2077 } 2078 2079 2080 // Get "block" use is in 2081 uint idx = 0; 2082 while( use->in(idx) != old ) idx++; 2083 Node *prev = use->is_CFG() ? use : get_ctrl(use); 2084 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" ); 2085 Node *cfg = prev->_idx >= new_counter 2086 ? prev->in(2) 2087 : idom(prev); 2088 if( use->is_Phi() ) // Phi use is in prior block 2089 cfg = prev->in(idx); // NOT in block of Phi itself 2090 if (cfg->is_top()) { // Use is dead? 2091 _igvn.replace_input_of(use, idx, C->top()); 2092 continue; 2093 } 2094 2095 // If use is referenced through control edge... (idx == 0) 2096 if (mode == IgnoreStripMined && idx == 0) { 2097 LoopNode *head = loop->_head->as_Loop(); 2098 if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) { 2099 // That node is outside the inner loop, leave it outside the 2100 // outer loop as well to not confuse verification code. 2101 assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop"); 2102 _igvn.replace_input_of(use, 0, head->outer_loop_exit()); 2103 continue; 2104 } 2105 } 2106 2107 while(!outer_loop->is_member(get_loop(cfg))) { 2108 prev = cfg; 2109 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg); 2110 } 2111 // If the use occurs after merging several exits from the loop, then 2112 // old value must have dominated all those exits. Since the same old 2113 // value was used on all those exits we did not need a Phi at this 2114 // merge point. NOW we do need a Phi here. Each loop exit value 2115 // is now merged with the peeled body exit; each exit gets its own 2116 // private Phi and those Phis need to be merged here. 2117 Node *phi; 2118 if( prev->is_Region() ) { 2119 if( idx == 0 ) { // Updating control edge? 2120 phi = prev; // Just use existing control 2121 } else { // Else need a new Phi 2122 phi = PhiNode::make( prev, old ); 2123 // Now recursively fix up the new uses of old! 2124 for( uint i = 1; i < prev->req(); i++ ) { 2125 worklist.push(phi); // Onto worklist once for each 'old' input 2126 } 2127 } 2128 } else { 2129 // Get new RegionNode merging old and new loop exits 2130 prev = old_new[prev->_idx]; 2131 assert( prev, "just made this in step 7" ); 2132 if( idx == 0) { // Updating control edge? 2133 phi = prev; // Just use existing control 2134 } else { // Else need a new Phi 2135 // Make a new Phi merging data values properly 2136 phi = PhiNode::make( prev, old ); 2137 phi->set_req( 1, nnn ); 2138 } 2139 } 2140 // If inserting a new Phi, check for prior hits 2141 if( idx != 0 ) { 2142 Node *hit = _igvn.hash_find_insert(phi); 2143 if( hit == NULL ) { 2144 _igvn.register_new_node_with_optimizer(phi); // Register new phi 2145 } else { // or 2146 // Remove the new phi from the graph and use the hit 2147 _igvn.remove_dead_node(phi); 2148 phi = hit; // Use existing phi 2149 } 2150 set_ctrl(phi, prev); 2151 } 2152 // Make 'use' use the Phi instead of the old loop body exit value 2153 _igvn.replace_input_of(use, idx, phi); 2154 if( use->_idx >= new_counter ) { // If updating new phis 2155 // Not needed for correctness, but prevents a weak assert 2156 // in AddPNode from tripping (when we end up with different 2157 // base & derived Phis that will become the same after 2158 // IGVN does CSE). 2159 Node *hit = _igvn.hash_find_insert(use); 2160 if( hit ) // Go ahead and re-hash for hits. 2161 _igvn.replace_node( use, hit ); 2162 } 2163 2164 // If 'use' was in the loop-exit block, it now needs to be sunk 2165 // below the post-loop merge point. 2166 sink_use( use, prev ); 2167 } 2168 } 2169 } 2170 2171 static void collect_nodes_in_outer_loop_not_reachable_from_sfpt(Node* n, const IdealLoopTree *loop, const IdealLoopTree* outer_loop, 2172 const Node_List &old_new, Unique_Node_List& wq, PhaseIdealLoop* phase, 2173 bool check_old_new) { 2174 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2175 Node* u = n->fast_out(j); 2176 assert(check_old_new || old_new[u->_idx] == NULL, "shouldn't have been cloned"); 2177 if (!u->is_CFG() && (!check_old_new || old_new[u->_idx] == NULL)) { 2178 Node* c = phase->get_ctrl(u); 2179 IdealLoopTree* u_loop = phase->get_loop(c); 2180 assert(!loop->is_member(u_loop) || !loop->_body.contains(u), "can be in outer loop or out of both loops only"); 2181 if (!loop->is_member(u_loop)) { 2182 if (outer_loop->is_member(u_loop)) { 2183 wq.push(u); 2184 } else { 2185 // nodes pinned with control in the outer loop but not referenced from the safepoint must be moved out of 2186 // the outer loop too 2187 Node* u_c = u->in(0); 2188 if (u_c != NULL) { 2189 IdealLoopTree* u_c_loop = phase->get_loop(u_c); 2190 if (outer_loop->is_member(u_c_loop) && !loop->is_member(u_c_loop)) { 2191 wq.push(u); 2192 } 2193 } 2194 } 2195 } 2196 } 2197 } 2198 } 2199 2200 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop, 2201 IdealLoopTree* outer_loop, int dd, Node_List &old_new, 2202 Node_List& extra_data_nodes) { 2203 if (head->is_strip_mined() && mode != IgnoreStripMined) { 2204 CountedLoopNode* cl = head->as_CountedLoop(); 2205 Node* l = cl->outer_loop(); 2206 Node* tail = cl->outer_loop_tail(); 2207 IfNode* le = cl->outer_loop_end(); 2208 Node* sfpt = cl->outer_safepoint(); 2209 CountedLoopEndNode* cle = cl->loopexit(); 2210 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop(); 2211 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null(); 2212 Node* cle_out = cle->proj_out(false); 2213 2214 Node* new_sfpt = NULL; 2215 Node* new_cle_out = cle_out->clone(); 2216 old_new.map(cle_out->_idx, new_cle_out); 2217 if (mode == CloneIncludesStripMined) { 2218 // clone outer loop body 2219 Node* new_l = l->clone(); 2220 Node* new_tail = tail->clone(); 2221 IfNode* new_le = le->clone()->as_If(); 2222 new_sfpt = sfpt->clone(); 2223 2224 set_loop(new_l, outer_loop->_parent); 2225 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd); 2226 set_loop(new_cle_out, outer_loop->_parent); 2227 set_idom(new_cle_out, new_cle, dd); 2228 set_loop(new_sfpt, outer_loop->_parent); 2229 set_idom(new_sfpt, new_cle_out, dd); 2230 set_loop(new_le, outer_loop->_parent); 2231 set_idom(new_le, new_sfpt, dd); 2232 set_loop(new_tail, outer_loop->_parent); 2233 set_idom(new_tail, new_le, dd); 2234 set_idom(new_cl, new_l, dd); 2235 2236 old_new.map(l->_idx, new_l); 2237 old_new.map(tail->_idx, new_tail); 2238 old_new.map(le->_idx, new_le); 2239 old_new.map(sfpt->_idx, new_sfpt); 2240 2241 new_l->set_req(LoopNode::LoopBackControl, new_tail); 2242 new_l->set_req(0, new_l); 2243 new_tail->set_req(0, new_le); 2244 new_le->set_req(0, new_sfpt); 2245 new_sfpt->set_req(0, new_cle_out); 2246 new_cle_out->set_req(0, new_cle); 2247 new_cl->set_req(LoopNode::EntryControl, new_l); 2248 2249 _igvn.register_new_node_with_optimizer(new_l); 2250 _igvn.register_new_node_with_optimizer(new_tail); 2251 _igvn.register_new_node_with_optimizer(new_le); 2252 } else { 2253 Node *newhead = old_new[loop->_head->_idx]; 2254 newhead->as_Loop()->clear_strip_mined(); 2255 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl)); 2256 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 2257 } 2258 // Look at data node that were assigned a control in the outer 2259 // loop: they are kept in the outer loop by the safepoint so start 2260 // from the safepoint node's inputs. 2261 IdealLoopTree* outer_loop = get_loop(l); 2262 Node_Stack stack(2); 2263 stack.push(sfpt, 1); 2264 uint new_counter = C->unique(); 2265 while (stack.size() > 0) { 2266 Node* n = stack.node(); 2267 uint i = stack.index(); 2268 while (i < n->req() && 2269 (n->in(i) == NULL || 2270 !has_ctrl(n->in(i)) || 2271 get_loop(get_ctrl(n->in(i))) != outer_loop || 2272 (old_new[n->in(i)->_idx] != NULL && old_new[n->in(i)->_idx]->_idx >= new_counter))) { 2273 i++; 2274 } 2275 if (i < n->req()) { 2276 stack.set_index(i+1); 2277 stack.push(n->in(i), 0); 2278 } else { 2279 assert(old_new[n->_idx] == NULL || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet"); 2280 Node* m = n == sfpt ? new_sfpt : n->clone(); 2281 if (m != NULL) { 2282 for (uint i = 0; i < n->req(); i++) { 2283 if (m->in(i) != NULL && old_new[m->in(i)->_idx] != NULL) { 2284 m->set_req(i, old_new[m->in(i)->_idx]); 2285 } 2286 } 2287 } else { 2288 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?"); 2289 } 2290 if (n != sfpt) { 2291 extra_data_nodes.push(n); 2292 _igvn.register_new_node_with_optimizer(m); 2293 assert(get_ctrl(n) == cle_out, "what other control?"); 2294 set_ctrl(m, new_cle_out); 2295 old_new.map(n->_idx, m); 2296 } 2297 stack.pop(); 2298 } 2299 } 2300 if (mode == CloneIncludesStripMined) { 2301 _igvn.register_new_node_with_optimizer(new_sfpt); 2302 _igvn.register_new_node_with_optimizer(new_cle_out); 2303 } 2304 // Some other transformation may have pessimistically assigned some 2305 // data nodes to the outer loop. Set their control so they are out 2306 // of the outer loop. 2307 ResourceMark rm; 2308 Unique_Node_List wq; 2309 for (uint i = 0; i < extra_data_nodes.size(); i++) { 2310 Node* old = extra_data_nodes.at(i); 2311 collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true); 2312 } 2313 2314 for (uint i = 0; i < loop->_body.size(); i++) { 2315 Node* old = loop->_body.at(i); 2316 collect_nodes_in_outer_loop_not_reachable_from_sfpt(old, loop, outer_loop, old_new, wq, this, true); 2317 } 2318 2319 Node* inner_out = sfpt->in(0); 2320 if (inner_out->outcnt() > 1) { 2321 collect_nodes_in_outer_loop_not_reachable_from_sfpt(inner_out, loop, outer_loop, old_new, wq, this, true); 2322 } 2323 2324 Node* new_ctrl = cl->outer_loop_exit(); 2325 assert(get_loop(new_ctrl) != outer_loop, "must be out of the loop nest"); 2326 for (uint i = 0; i < wq.size(); i++) { 2327 Node* n = wq.at(i); 2328 set_ctrl(n, new_ctrl); 2329 if (n->in(0) != NULL) { 2330 _igvn.replace_input_of(n, 0, new_ctrl); 2331 } 2332 collect_nodes_in_outer_loop_not_reachable_from_sfpt(n, loop, outer_loop, old_new, wq, this, false); 2333 } 2334 } else { 2335 Node *newhead = old_new[loop->_head->_idx]; 2336 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 2337 } 2338 } 2339 2340 //------------------------------clone_loop------------------------------------- 2341 // 2342 // C L O N E A L O O P B O D Y 2343 // 2344 // This is the basic building block of the loop optimizations. It clones an 2345 // entire loop body. It makes an old_new loop body mapping; with this mapping 2346 // you can find the new-loop equivalent to an old-loop node. All new-loop 2347 // nodes are exactly equal to their old-loop counterparts, all edges are the 2348 // same. All exits from the old-loop now have a RegionNode that merges the 2349 // equivalent new-loop path. This is true even for the normal "loop-exit" 2350 // condition. All uses of loop-invariant old-loop values now come from (one 2351 // or more) Phis that merge their new-loop equivalents. 2352 // 2353 // This operation leaves the graph in an illegal state: there are two valid 2354 // control edges coming from the loop pre-header to both loop bodies. I'll 2355 // definitely have to hack the graph after running this transform. 2356 // 2357 // From this building block I will further edit edges to perform loop peeling 2358 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc. 2359 // 2360 // Parameter side_by_size_idom: 2361 // When side_by_size_idom is NULL, the dominator tree is constructed for 2362 // the clone loop to dominate the original. Used in construction of 2363 // pre-main-post loop sequence. 2364 // When nonnull, the clone and original are side-by-side, both are 2365 // dominated by the side_by_side_idom node. Used in construction of 2366 // unswitched loops. 2367 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd, 2368 CloneLoopMode mode, Node* side_by_side_idom) { 2369 2370 LoopNode* head = loop->_head->as_Loop(); 2371 head->verify_strip_mined(1); 2372 2373 if (C->do_vector_loop() && PrintOpto) { 2374 const char* mname = C->method()->name()->as_quoted_ascii(); 2375 if (mname != NULL) { 2376 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname); 2377 } 2378 } 2379 2380 CloneMap& cm = C->clone_map(); 2381 if (C->do_vector_loop()) { 2382 cm.set_clone_idx(cm.max_gen()+1); 2383 #ifndef PRODUCT 2384 if (PrintOpto) { 2385 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx()); 2386 loop->dump_head(); 2387 } 2388 #endif 2389 } 2390 2391 // Step 1: Clone the loop body. Make the old->new mapping. 2392 clone_loop_body(loop->_body, old_new, &cm); 2393 2394 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop; 2395 2396 // Step 2: Fix the edges in the new body. If the old input is outside the 2397 // loop use it. If the old input is INside the loop, use the corresponding 2398 // new node instead. 2399 fix_body_edges(loop->_body, loop, old_new, dd, outer_loop->_parent, false); 2400 2401 Node_List extra_data_nodes; // data nodes in the outer strip mined loop 2402 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes); 2403 2404 // Step 3: Now fix control uses. Loop varying control uses have already 2405 // been fixed up (as part of all input edges in Step 2). Loop invariant 2406 // control uses must be either an IfFalse or an IfTrue. Make a merge 2407 // point to merge the old and new IfFalse/IfTrue nodes; make the use 2408 // refer to this. 2409 Node_List worklist; 2410 uint new_counter = C->unique(); 2411 fix_ctrl_uses(loop->_body, loop, old_new, mode, side_by_side_idom, &cm, worklist); 2412 2413 // Step 4: If loop-invariant use is not control, it must be dominated by a 2414 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region 2415 // there if needed. Make a Phi there merging old and new used values. 2416 Node_List *split_if_set = NULL; 2417 Node_List *split_bool_set = NULL; 2418 Node_List *split_cex_set = NULL; 2419 fix_data_uses(loop->_body, loop, mode, outer_loop, new_counter, old_new, worklist, split_if_set, split_bool_set, split_cex_set); 2420 2421 for (uint i = 0; i < extra_data_nodes.size(); i++) { 2422 Node* old = extra_data_nodes.at(i); 2423 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set, 2424 split_bool_set, split_cex_set, worklist, new_counter, 2425 mode); 2426 } 2427 2428 // Check for IFs that need splitting/cloning. Happens if an IF outside of 2429 // the loop uses a condition set in the loop. The original IF probably 2430 // takes control from one or more OLD Regions (which in turn get from NEW 2431 // Regions). In any case, there will be a set of Phis for each merge point 2432 // from the IF up to where the original BOOL def exists the loop. 2433 finish_clone_loop(split_if_set, split_bool_set, split_cex_set); 2434 2435 } 2436 2437 void PhaseIdealLoop::finish_clone_loop(Node_List* split_if_set, Node_List* split_bool_set, Node_List* split_cex_set) { 2438 if (split_if_set) { 2439 while (split_if_set->size()) { 2440 Node *iff = split_if_set->pop(); 2441 uint input = iff->Opcode() == Op_AllocateArray ? AllocateNode::ValidLengthTest : 1; 2442 if (iff->in(input)->is_Phi()) { 2443 Node *b = clone_iff(iff->in(input)->as_Phi()); 2444 _igvn.replace_input_of(iff, input, b); 2445 } 2446 } 2447 } 2448 if (split_bool_set) { 2449 while (split_bool_set->size()) { 2450 Node *b = split_bool_set->pop(); 2451 Node *phi = b->in(1); 2452 assert(phi->is_Phi(), ""); 2453 CmpNode *cmp = clone_bool((PhiNode*) phi); 2454 _igvn.replace_input_of(b, 1, cmp); 2455 } 2456 } 2457 if (split_cex_set) { 2458 while (split_cex_set->size()) { 2459 Node *b = split_cex_set->pop(); 2460 assert(b->in(0)->is_Region(), ""); 2461 assert(b->in(1)->is_Phi(), ""); 2462 assert(b->in(0)->in(0) == b->in(1)->in(0), ""); 2463 split_up(b, b->in(0), NULL); 2464 } 2465 } 2466 } 2467 2468 void PhaseIdealLoop::fix_data_uses(Node_List& body, IdealLoopTree* loop, CloneLoopMode mode, IdealLoopTree* outer_loop, 2469 uint new_counter, Node_List &old_new, Node_List &worklist, Node_List*& split_if_set, 2470 Node_List*& split_bool_set, Node_List*& split_cex_set) { 2471 for(uint i = 0; i < body.size(); i++ ) { 2472 Node* old = body.at(i); 2473 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set, 2474 split_bool_set, split_cex_set, worklist, new_counter, 2475 mode); 2476 } 2477 } 2478 2479 void PhaseIdealLoop::fix_ctrl_uses(const Node_List& body, const IdealLoopTree* loop, Node_List &old_new, CloneLoopMode mode, 2480 Node* side_by_side_idom, CloneMap* cm, Node_List &worklist) { 2481 LoopNode* head = loop->_head->as_Loop(); 2482 for(uint i = 0; i < body.size(); i++ ) { 2483 Node* old = body.at(i); 2484 if( !old->is_CFG() ) continue; 2485 2486 // Copy uses to a worklist, so I can munge the def-use info 2487 // with impunity. 2488 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) { 2489 worklist.push(old->fast_out(j)); 2490 } 2491 2492 while (worklist.size()) { // Visit all uses 2493 Node *use = worklist.pop(); 2494 if (!has_node(use)) continue; // Ignore dead nodes 2495 IdealLoopTree *use_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use ); 2496 if (!loop->is_member(use_loop) && use->is_CFG()) { 2497 // Both OLD and USE are CFG nodes here. 2498 assert(use->is_Proj(), "" ); 2499 Node* nnn = old_new[old->_idx]; 2500 2501 Node* newuse = NULL; 2502 if (head->is_strip_mined() && mode != IgnoreStripMined) { 2503 CountedLoopNode* cl = head->as_CountedLoop(); 2504 CountedLoopEndNode* cle = cl->loopexit(); 2505 Node* cle_out = cle->proj_out_or_null(false); 2506 if (use == cle_out) { 2507 IfNode* le = cl->outer_loop_end(); 2508 use = le->proj_out(false); 2509 use_loop = get_loop(use); 2510 if (mode == CloneIncludesStripMined) { 2511 nnn = old_new[le->_idx]; 2512 } else { 2513 newuse = old_new[cle_out->_idx]; 2514 } 2515 } 2516 } 2517 if (newuse == NULL) { 2518 newuse = use->clone(); 2519 } 2520 2521 // Clone the loop exit control projection 2522 if (C->do_vector_loop() && cm != NULL) { 2523 cm->verify_insert_and_clone(use, newuse, cm->clone_idx()); 2524 } 2525 newuse->set_req(0,nnn); 2526 _igvn.register_new_node_with_optimizer(newuse); 2527 set_loop(newuse, use_loop); 2528 set_idom(newuse, nnn, dom_depth(nnn) + 1 ); 2529 2530 // We need a Region to merge the exit from the peeled body and the 2531 // exit from the old loop body. 2532 RegionNode *r = new RegionNode(3); 2533 // Map the old use to the new merge point 2534 old_new.map( use->_idx, r ); 2535 uint dd_r = MIN2(dom_depth(newuse), dom_depth(use)); 2536 assert(dd_r >= dom_depth(dom_lca(newuse, use)), "" ); 2537 2538 // The original user of 'use' uses 'r' instead. 2539 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) { 2540 Node* useuse = use->last_out(l); 2541 _igvn.rehash_node_delayed(useuse); 2542 uint uses_found = 0; 2543 if (useuse->in(0) == use) { 2544 useuse->set_req(0, r); 2545 uses_found++; 2546 if (useuse->is_CFG()) { 2547 // This is not a dom_depth > dd_r because when new 2548 // control flow is constructed by a loop opt, a node and 2549 // its dominator can end up at the same dom_depth 2550 assert(dom_depth(useuse) >= dd_r, ""); 2551 set_idom(useuse, r, dom_depth(useuse)); 2552 } 2553 } 2554 for (uint k = 1; k < useuse->req(); k++) { 2555 if( useuse->in(k) == use ) { 2556 useuse->set_req(k, r); 2557 uses_found++; 2558 if (useuse->is_Loop() && k == LoopNode::EntryControl) { 2559 // This is not a dom_depth > dd_r because when new 2560 // control flow is constructed by a loop opt, a node 2561 // and its dominator can end up at the same dom_depth 2562 assert(dom_depth(useuse) >= dd_r , ""); 2563 set_idom(useuse, r, dom_depth(useuse)); 2564 } 2565 } 2566 } 2567 l -= uses_found; // we deleted 1 or more copies of this edge 2568 } 2569 2570 // Now finish up 'r' 2571 r->set_req(1, newuse); 2572 r->set_req(2, use); 2573 _igvn.register_new_node_with_optimizer(r); 2574 set_loop(r, use_loop); 2575 set_idom(r, (side_by_side_idom == NULL) ? newuse->in(0) : side_by_side_idom, dd_r); 2576 } // End of if a loop-exit test 2577 } 2578 } 2579 } 2580 2581 void PhaseIdealLoop::fix_body_edges(const Node_List &body, IdealLoopTree* loop, const Node_List &old_new, int dd, 2582 IdealLoopTree* parent, bool partial) { 2583 for(uint i = 0; i < body.size(); i++ ) { 2584 Node *old = body.at(i); 2585 Node *nnn = old_new[old->_idx]; 2586 // Fix CFG/Loop controlling the new node 2587 if (has_ctrl(old)) { 2588 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]); 2589 } else { 2590 set_loop(nnn, parent); 2591 if (old->outcnt() > 0) { 2592 Node* dom = idom(old); 2593 if (old_new[dom->_idx] != NULL) { 2594 dom = old_new[dom->_idx]; 2595 set_idom(nnn, dom, dd ); 2596 } 2597 } 2598 } 2599 // Correct edges to the new node 2600 for (uint j = 0; j < nnn->req(); j++) { 2601 Node *n = nnn->in(j); 2602 if (n != NULL) { 2603 IdealLoopTree *old_in_loop = get_loop(has_ctrl(n) ? get_ctrl(n) : n); 2604 if (loop->is_member(old_in_loop)) { 2605 if (old_new[n->_idx] != NULL) { 2606 nnn->set_req(j, old_new[n->_idx]); 2607 } else { 2608 assert(!body.contains(n), ""); 2609 assert(partial, "node not cloned"); 2610 } 2611 } 2612 } 2613 } 2614 _igvn.hash_find_insert(nnn); 2615 } 2616 } 2617 2618 void PhaseIdealLoop::clone_loop_body(const Node_List& body, Node_List &old_new, CloneMap* cm) { 2619 for (uint i = 0; i < body.size(); i++) { 2620 Node* old = body.at(i); 2621 Node* nnn = old->clone(); 2622 old_new.map(old->_idx, nnn); 2623 if (old->is_reduction()) { 2624 // Reduction flag is not copied by default. Copy it here when cloning the entire loop body. 2625 nnn->add_flag(Node::Flag_is_reduction); 2626 } 2627 if (C->do_vector_loop() && cm != NULL) { 2628 cm->verify_insert_and_clone(old, nnn, cm->clone_idx()); 2629 } 2630 _igvn.register_new_node_with_optimizer(nnn); 2631 } 2632 } 2633 2634 2635 //---------------------- stride_of_possible_iv ------------------------------------- 2636 // Looks for an iff/bool/comp with one operand of the compare 2637 // being a cycle involving an add and a phi, 2638 // with an optional truncation (left-shift followed by a right-shift) 2639 // of the add. Returns zero if not an iv. 2640 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) { 2641 Node* trunc1 = NULL; 2642 Node* trunc2 = NULL; 2643 const TypeInteger* ttype = NULL; 2644 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) { 2645 return 0; 2646 } 2647 BoolNode* bl = iff->in(1)->as_Bool(); 2648 Node* cmp = bl->in(1); 2649 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) { 2650 return 0; 2651 } 2652 // Must have an invariant operand 2653 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) { 2654 return 0; 2655 } 2656 Node* add2 = NULL; 2657 Node* cmp1 = cmp->in(1); 2658 if (cmp1->is_Phi()) { 2659 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) ))) 2660 Node* phi = cmp1; 2661 for (uint i = 1; i < phi->req(); i++) { 2662 Node* in = phi->in(i); 2663 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in, 2664 &trunc1, &trunc2, &ttype, T_INT); 2665 if (add && add->in(1) == phi) { 2666 add2 = add->in(2); 2667 break; 2668 } 2669 } 2670 } else { 2671 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) ))) 2672 Node* addtrunc = cmp1; 2673 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc, 2674 &trunc1, &trunc2, &ttype, T_INT); 2675 if (add && add->in(1)->is_Phi()) { 2676 Node* phi = add->in(1); 2677 for (uint i = 1; i < phi->req(); i++) { 2678 if (phi->in(i) == addtrunc) { 2679 add2 = add->in(2); 2680 break; 2681 } 2682 } 2683 } 2684 } 2685 if (add2 != NULL) { 2686 const TypeInt* add2t = _igvn.type(add2)->is_int(); 2687 if (add2t->is_con()) { 2688 return add2t->get_con(); 2689 } 2690 } 2691 return 0; 2692 } 2693 2694 2695 //---------------------- stay_in_loop ------------------------------------- 2696 // Return the (unique) control output node that's in the loop (if it exists.) 2697 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) { 2698 Node* unique = NULL; 2699 if (!n) return NULL; 2700 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2701 Node* use = n->fast_out(i); 2702 if (!has_ctrl(use) && loop->is_member(get_loop(use))) { 2703 if (unique != NULL) { 2704 return NULL; 2705 } 2706 unique = use; 2707 } 2708 } 2709 return unique; 2710 } 2711 2712 //------------------------------ register_node ------------------------------------- 2713 // Utility to register node "n" with PhaseIdealLoop 2714 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) { 2715 _igvn.register_new_node_with_optimizer(n); 2716 loop->_body.push(n); 2717 if (n->is_CFG()) { 2718 set_loop(n, loop); 2719 set_idom(n, pred, ddepth); 2720 } else { 2721 set_ctrl(n, pred); 2722 } 2723 } 2724 2725 //------------------------------ proj_clone ------------------------------------- 2726 // Utility to create an if-projection 2727 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) { 2728 ProjNode* c = p->clone()->as_Proj(); 2729 c->set_req(0, iff); 2730 return c; 2731 } 2732 2733 //------------------------------ short_circuit_if ------------------------------------- 2734 // Force the iff control output to be the live_proj 2735 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) { 2736 guarantee(live_proj != NULL, "null projection"); 2737 int proj_con = live_proj->_con; 2738 assert(proj_con == 0 || proj_con == 1, "false or true projection"); 2739 Node *con = _igvn.intcon(proj_con); 2740 set_ctrl(con, C->root()); 2741 if (iff) { 2742 iff->set_req(1, con); 2743 } 2744 return con; 2745 } 2746 2747 //------------------------------ insert_if_before_proj ------------------------------------- 2748 // Insert a new if before an if projection (* - new node) 2749 // 2750 // before 2751 // if(test) 2752 // / \ 2753 // v v 2754 // other-proj proj (arg) 2755 // 2756 // after 2757 // if(test) 2758 // / \ 2759 // / v 2760 // | * proj-clone 2761 // v | 2762 // other-proj v 2763 // * new_if(relop(cmp[IU](left,right))) 2764 // / \ 2765 // v v 2766 // * new-proj proj 2767 // (returned) 2768 // 2769 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) { 2770 IfNode* iff = proj->in(0)->as_If(); 2771 IdealLoopTree *loop = get_loop(proj); 2772 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2773 int ddepth = dom_depth(proj); 2774 2775 _igvn.rehash_node_delayed(iff); 2776 _igvn.rehash_node_delayed(proj); 2777 2778 proj->set_req(0, NULL); // temporary disconnect 2779 ProjNode* proj2 = proj_clone(proj, iff); 2780 register_node(proj2, loop, iff, ddepth); 2781 2782 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right); 2783 register_node(cmp, loop, proj2, ddepth); 2784 2785 BoolNode* bol = new BoolNode(cmp, relop); 2786 register_node(bol, loop, proj2, ddepth); 2787 2788 int opcode = iff->Opcode(); 2789 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode"); 2790 IfNode* new_if = (opcode == Op_If) ? new IfNode(proj2, bol, iff->_prob, iff->_fcnt): 2791 new RangeCheckNode(proj2, bol, iff->_prob, iff->_fcnt); 2792 register_node(new_if, loop, proj2, ddepth); 2793 2794 proj->set_req(0, new_if); // reattach 2795 set_idom(proj, new_if, ddepth); 2796 2797 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj(); 2798 guarantee(new_exit != NULL, "null exit node"); 2799 register_node(new_exit, get_loop(other_proj), new_if, ddepth); 2800 2801 return new_exit; 2802 } 2803 2804 //------------------------------ insert_region_before_proj ------------------------------------- 2805 // Insert a region before an if projection (* - new node) 2806 // 2807 // before 2808 // if(test) 2809 // / | 2810 // v | 2811 // proj v 2812 // other-proj 2813 // 2814 // after 2815 // if(test) 2816 // / | 2817 // v | 2818 // * proj-clone v 2819 // | other-proj 2820 // v 2821 // * new-region 2822 // | 2823 // v 2824 // * dum_if 2825 // / \ 2826 // v \ 2827 // * dum-proj v 2828 // proj 2829 // 2830 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) { 2831 IfNode* iff = proj->in(0)->as_If(); 2832 IdealLoopTree *loop = get_loop(proj); 2833 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2834 int ddepth = dom_depth(proj); 2835 2836 _igvn.rehash_node_delayed(iff); 2837 _igvn.rehash_node_delayed(proj); 2838 2839 proj->set_req(0, NULL); // temporary disconnect 2840 ProjNode* proj2 = proj_clone(proj, iff); 2841 register_node(proj2, loop, iff, ddepth); 2842 2843 RegionNode* reg = new RegionNode(2); 2844 reg->set_req(1, proj2); 2845 register_node(reg, loop, iff, ddepth); 2846 2847 IfNode* dum_if = new IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt); 2848 register_node(dum_if, loop, reg, ddepth); 2849 2850 proj->set_req(0, dum_if); // reattach 2851 set_idom(proj, dum_if, ddepth); 2852 2853 ProjNode* dum_proj = proj_clone(other_proj, dum_if); 2854 register_node(dum_proj, loop, dum_if, ddepth); 2855 2856 return reg; 2857 } 2858 2859 //------------------------------ insert_cmpi_loop_exit ------------------------------------- 2860 // Clone a signed compare loop exit from an unsigned compare and 2861 // insert it before the unsigned cmp on the stay-in-loop path. 2862 // All new nodes inserted in the dominator tree between the original 2863 // if and it's projections. The original if test is replaced with 2864 // a constant to force the stay-in-loop path. 2865 // 2866 // This is done to make sure that the original if and it's projections 2867 // still dominate the same set of control nodes, that the ctrl() relation 2868 // from data nodes to them is preserved, and that their loop nesting is 2869 // preserved. 2870 // 2871 // before 2872 // if(i <u limit) unsigned compare loop exit 2873 // / | 2874 // v v 2875 // exit-proj stay-in-loop-proj 2876 // 2877 // after 2878 // if(stay-in-loop-const) original if 2879 // / | 2880 // / v 2881 // / if(i < limit) new signed test 2882 // / / | 2883 // / / v 2884 // / / if(i <u limit) new cloned unsigned test 2885 // / / / | 2886 // v v v | 2887 // region | 2888 // | | 2889 // dum-if | 2890 // / | | 2891 // ether | | 2892 // v v 2893 // exit-proj stay-in-loop-proj 2894 // 2895 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) { 2896 const bool Signed = true; 2897 const bool Unsigned = false; 2898 2899 BoolNode* bol = if_cmpu->in(1)->as_Bool(); 2900 if (bol->_test._test != BoolTest::lt) return NULL; 2901 CmpNode* cmpu = bol->in(1)->as_Cmp(); 2902 if (cmpu->Opcode() != Op_CmpU) return NULL; 2903 int stride = stride_of_possible_iv(if_cmpu); 2904 if (stride == 0) return NULL; 2905 2906 Node* lp_proj = stay_in_loop(if_cmpu, loop); 2907 guarantee(lp_proj != NULL, "null loop node"); 2908 2909 ProjNode* lp_continue = lp_proj->as_Proj(); 2910 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj(); 2911 if (!lp_exit->is_IfFalse()) { 2912 // The loop exit condition is (i <u limit) ==> (i >= 0 && i < limit). 2913 // We therefore can't add a single exit condition. 2914 return NULL; 2915 } 2916 // The loop exit condition is !(i <u limit) ==> (i < 0 || i >= limit). 2917 // Split out the exit condition (i < 0) for stride < 0 or (i >= limit) for stride > 0. 2918 Node* limit = NULL; 2919 if (stride > 0) { 2920 limit = cmpu->in(2); 2921 } else { 2922 limit = _igvn.makecon(TypeInt::ZERO); 2923 set_ctrl(limit, C->root()); 2924 } 2925 // Create a new region on the exit path 2926 RegionNode* reg = insert_region_before_proj(lp_exit); 2927 guarantee(reg != NULL, "null region node"); 2928 2929 // Clone the if-cmpu-true-false using a signed compare 2930 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge; 2931 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue); 2932 reg->add_req(cmpi_exit); 2933 2934 // Clone the if-cmpu-true-false 2935 BoolTest::mask rel_u = bol->_test._test; 2936 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue); 2937 reg->add_req(cmpu_exit); 2938 2939 // Force original if to stay in loop. 2940 short_circuit_if(if_cmpu, lp_continue); 2941 2942 return cmpi_exit->in(0)->as_If(); 2943 } 2944 2945 //------------------------------ remove_cmpi_loop_exit ------------------------------------- 2946 // Remove a previously inserted signed compare loop exit. 2947 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) { 2948 Node* lp_proj = stay_in_loop(if_cmp, loop); 2949 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI && 2950 stay_in_loop(lp_proj, loop)->is_If() && 2951 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu"); 2952 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO); 2953 set_ctrl(con, C->root()); 2954 if_cmp->set_req(1, con); 2955 } 2956 2957 //------------------------------ scheduled_nodelist ------------------------------------- 2958 // Create a post order schedule of nodes that are in the 2959 // "member" set. The list is returned in "sched". 2960 // The first node in "sched" is the loop head, followed by 2961 // nodes which have no inputs in the "member" set, and then 2962 // followed by the nodes that have an immediate input dependence 2963 // on a node in "sched". 2964 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) { 2965 2966 assert(member.test(loop->_head->_idx), "loop head must be in member set"); 2967 VectorSet visited; 2968 Node_Stack nstack(loop->_body.size()); 2969 2970 Node* n = loop->_head; // top of stack is cached in "n" 2971 uint idx = 0; 2972 visited.set(n->_idx); 2973 2974 // Initially push all with no inputs from within member set 2975 for(uint i = 0; i < loop->_body.size(); i++ ) { 2976 Node *elt = loop->_body.at(i); 2977 if (member.test(elt->_idx)) { 2978 bool found = false; 2979 for (uint j = 0; j < elt->req(); j++) { 2980 Node* def = elt->in(j); 2981 if (def && member.test(def->_idx) && def != elt) { 2982 found = true; 2983 break; 2984 } 2985 } 2986 if (!found && elt != loop->_head) { 2987 nstack.push(n, idx); 2988 n = elt; 2989 assert(!visited.test(n->_idx), "not seen yet"); 2990 visited.set(n->_idx); 2991 } 2992 } 2993 } 2994 2995 // traverse out's that are in the member set 2996 while (true) { 2997 if (idx < n->outcnt()) { 2998 Node* use = n->raw_out(idx); 2999 idx++; 3000 if (!visited.test_set(use->_idx)) { 3001 if (member.test(use->_idx)) { 3002 nstack.push(n, idx); 3003 n = use; 3004 idx = 0; 3005 } 3006 } 3007 } else { 3008 // All outputs processed 3009 sched.push(n); 3010 if (nstack.is_empty()) break; 3011 n = nstack.node(); 3012 idx = nstack.index(); 3013 nstack.pop(); 3014 } 3015 } 3016 } 3017 3018 3019 //------------------------------ has_use_in_set ------------------------------------- 3020 // Has a use in the vector set 3021 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) { 3022 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3023 Node* use = n->fast_out(j); 3024 if (vset.test(use->_idx)) { 3025 return true; 3026 } 3027 } 3028 return false; 3029 } 3030 3031 3032 //------------------------------ has_use_internal_to_set ------------------------------------- 3033 // Has use internal to the vector set (ie. not in a phi at the loop head) 3034 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) { 3035 Node* head = loop->_head; 3036 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3037 Node* use = n->fast_out(j); 3038 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) { 3039 return true; 3040 } 3041 } 3042 return false; 3043 } 3044 3045 3046 //------------------------------ clone_for_use_outside_loop ------------------------------------- 3047 // clone "n" for uses that are outside of loop 3048 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) { 3049 int cloned = 0; 3050 assert(worklist.size() == 0, "should be empty"); 3051 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3052 Node* use = n->fast_out(j); 3053 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) { 3054 worklist.push(use); 3055 } 3056 } 3057 3058 if (C->check_node_count(worklist.size() + NodeLimitFudgeFactor, 3059 "Too many clones required in clone_for_use_outside_loop in partial peeling")) { 3060 return -1; 3061 } 3062 3063 while( worklist.size() ) { 3064 Node *use = worklist.pop(); 3065 if (!has_node(use) || use->in(0) == C->top()) continue; 3066 uint j; 3067 for (j = 0; j < use->req(); j++) { 3068 if (use->in(j) == n) break; 3069 } 3070 assert(j < use->req(), "must be there"); 3071 3072 // clone "n" and insert it between the inputs of "n" and the use outside the loop 3073 Node* n_clone = n->clone(); 3074 _igvn.replace_input_of(use, j, n_clone); 3075 cloned++; 3076 Node* use_c; 3077 if (!use->is_Phi()) { 3078 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0); 3079 } else { 3080 // Use in a phi is considered a use in the associated predecessor block 3081 use_c = use->in(0)->in(j); 3082 } 3083 set_ctrl(n_clone, use_c); 3084 assert(!loop->is_member(get_loop(use_c)), "should be outside loop"); 3085 get_loop(use_c)->_body.push(n_clone); 3086 _igvn.register_new_node_with_optimizer(n_clone); 3087 #ifndef PRODUCT 3088 if (TracePartialPeeling) { 3089 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx); 3090 } 3091 #endif 3092 } 3093 return cloned; 3094 } 3095 3096 3097 //------------------------------ clone_for_special_use_inside_loop ------------------------------------- 3098 // clone "n" for special uses that are in the not_peeled region. 3099 // If these def-uses occur in separate blocks, the code generator 3100 // marks the method as not compilable. For example, if a "BoolNode" 3101 // is in a different basic block than the "IfNode" that uses it, then 3102 // the compilation is aborted in the code generator. 3103 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 3104 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) { 3105 if (n->is_Phi() || n->is_Load()) { 3106 return; 3107 } 3108 assert(worklist.size() == 0, "should be empty"); 3109 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3110 Node* use = n->fast_out(j); 3111 if ( not_peel.test(use->_idx) && 3112 (use->is_If() || use->is_CMove() || use->is_Bool()) && 3113 use->in(1) == n) { 3114 worklist.push(use); 3115 } 3116 } 3117 if (worklist.size() > 0) { 3118 // clone "n" and insert it between inputs of "n" and the use 3119 Node* n_clone = n->clone(); 3120 loop->_body.push(n_clone); 3121 _igvn.register_new_node_with_optimizer(n_clone); 3122 set_ctrl(n_clone, get_ctrl(n)); 3123 sink_list.push(n_clone); 3124 not_peel.set(n_clone->_idx); 3125 #ifndef PRODUCT 3126 if (TracePartialPeeling) { 3127 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx); 3128 } 3129 #endif 3130 while( worklist.size() ) { 3131 Node *use = worklist.pop(); 3132 _igvn.rehash_node_delayed(use); 3133 for (uint j = 1; j < use->req(); j++) { 3134 if (use->in(j) == n) { 3135 use->set_req(j, n_clone); 3136 } 3137 } 3138 } 3139 } 3140 } 3141 3142 3143 //------------------------------ insert_phi_for_loop ------------------------------------- 3144 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 3145 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) { 3146 Node *phi = PhiNode::make(lp, back_edge_val); 3147 phi->set_req(LoopNode::EntryControl, lp_entry_val); 3148 // Use existing phi if it already exists 3149 Node *hit = _igvn.hash_find_insert(phi); 3150 if( hit == NULL ) { 3151 _igvn.register_new_node_with_optimizer(phi); 3152 set_ctrl(phi, lp); 3153 } else { 3154 // Remove the new phi from the graph and use the hit 3155 _igvn.remove_dead_node(phi); 3156 phi = hit; 3157 } 3158 _igvn.replace_input_of(use, idx, phi); 3159 } 3160 3161 #ifdef ASSERT 3162 //------------------------------ is_valid_loop_partition ------------------------------------- 3163 // Validate the loop partition sets: peel and not_peel 3164 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, 3165 VectorSet& not_peel ) { 3166 uint i; 3167 // Check that peel_list entries are in the peel set 3168 for (i = 0; i < peel_list.size(); i++) { 3169 if (!peel.test(peel_list.at(i)->_idx)) { 3170 return false; 3171 } 3172 } 3173 // Check at loop members are in one of peel set or not_peel set 3174 for (i = 0; i < loop->_body.size(); i++ ) { 3175 Node *def = loop->_body.at(i); 3176 uint di = def->_idx; 3177 // Check that peel set elements are in peel_list 3178 if (peel.test(di)) { 3179 if (not_peel.test(di)) { 3180 return false; 3181 } 3182 // Must be in peel_list also 3183 bool found = false; 3184 for (uint j = 0; j < peel_list.size(); j++) { 3185 if (peel_list.at(j)->_idx == di) { 3186 found = true; 3187 break; 3188 } 3189 } 3190 if (!found) { 3191 return false; 3192 } 3193 } else if (not_peel.test(di)) { 3194 if (peel.test(di)) { 3195 return false; 3196 } 3197 } else { 3198 return false; 3199 } 3200 } 3201 return true; 3202 } 3203 3204 //------------------------------ is_valid_clone_loop_exit_use ------------------------------------- 3205 // Ensure a use outside of loop is of the right form 3206 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) { 3207 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 3208 return (use->is_Phi() && 3209 use_c->is_Region() && use_c->req() == 3 && 3210 (use_c->in(exit_idx)->Opcode() == Op_IfTrue || 3211 use_c->in(exit_idx)->Opcode() == Op_IfFalse || 3212 use_c->in(exit_idx)->Opcode() == Op_JumpProj) && 3213 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) ); 3214 } 3215 3216 //------------------------------ is_valid_clone_loop_form ------------------------------------- 3217 // Ensure that all uses outside of loop are of the right form 3218 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 3219 uint orig_exit_idx, uint clone_exit_idx) { 3220 uint len = peel_list.size(); 3221 for (uint i = 0; i < len; i++) { 3222 Node *def = peel_list.at(i); 3223 3224 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 3225 Node *use = def->fast_out(j); 3226 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 3227 if (!loop->is_member(get_loop(use_c))) { 3228 // use is not in the loop, check for correct structure 3229 if (use->in(0) == def) { 3230 // Okay 3231 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) { 3232 return false; 3233 } 3234 } 3235 } 3236 } 3237 return true; 3238 } 3239 #endif 3240 3241 //------------------------------ partial_peel ------------------------------------- 3242 // Partially peel (aka loop rotation) the top portion of a loop (called 3243 // the peel section below) by cloning it and placing one copy just before 3244 // the new loop head and the other copy at the bottom of the new loop. 3245 // 3246 // before after where it came from 3247 // 3248 // stmt1 stmt1 3249 // loop: stmt2 clone 3250 // stmt2 if condA goto exitA clone 3251 // if condA goto exitA new_loop: new 3252 // stmt3 stmt3 clone 3253 // if !condB goto loop if condB goto exitB clone 3254 // exitB: stmt2 orig 3255 // stmt4 if !condA goto new_loop orig 3256 // exitA: goto exitA 3257 // exitB: 3258 // stmt4 3259 // exitA: 3260 // 3261 // Step 1: find the cut point: an exit test on probable 3262 // induction variable. 3263 // Step 2: schedule (with cloning) operations in the peel 3264 // section that can be executed after the cut into 3265 // the section that is not peeled. This may need 3266 // to clone operations into exit blocks. For 3267 // instance, a reference to A[i] in the not-peel 3268 // section and a reference to B[i] in an exit block 3269 // may cause a left-shift of i by 2 to be placed 3270 // in the peel block. This step will clone the left 3271 // shift into the exit block and sink the left shift 3272 // from the peel to the not-peel section. 3273 // Step 3: clone the loop, retarget the control, and insert 3274 // phis for values that are live across the new loop 3275 // head. This is very dependent on the graph structure 3276 // from clone_loop. It creates region nodes for 3277 // exit control and associated phi nodes for values 3278 // flow out of the loop through that exit. The region 3279 // node is dominated by the clone's control projection. 3280 // So the clone's peel section is placed before the 3281 // new loop head, and the clone's not-peel section is 3282 // forms the top part of the new loop. The original 3283 // peel section forms the tail of the new loop. 3284 // Step 4: update the dominator tree and recompute the 3285 // dominator depth. 3286 // 3287 // orig 3288 // 3289 // stmt1 3290 // | 3291 // v 3292 // loop predicate 3293 // | 3294 // v 3295 // loop<----+ 3296 // | | 3297 // stmt2 | 3298 // | | 3299 // v | 3300 // ifA | 3301 // / | | 3302 // v v | 3303 // false true ^ <-- last_peel 3304 // / | | 3305 // / ===|==cut | 3306 // / stmt3 | <-- first_not_peel 3307 // / | | 3308 // | v | 3309 // v ifB | 3310 // exitA: / \ | 3311 // / \ | 3312 // v v | 3313 // false true | 3314 // / \ | 3315 // / ----+ 3316 // | 3317 // v 3318 // exitB: 3319 // stmt4 3320 // 3321 // 3322 // after clone loop 3323 // 3324 // stmt1 3325 // | 3326 // v 3327 // loop predicate 3328 // / \ 3329 // clone / \ orig 3330 // / \ 3331 // / \ 3332 // v v 3333 // +---->loop loop<----+ 3334 // | | | | 3335 // | stmt2 stmt2 | 3336 // | | | | 3337 // | v v | 3338 // | ifA ifA | 3339 // | | \ / | | 3340 // | v v v v | 3341 // ^ true false false true ^ <-- last_peel 3342 // | | ^ \ / | | 3343 // | cut==|== \ \ / ===|==cut | 3344 // | stmt3 \ \ / stmt3 | <-- first_not_peel 3345 // | | dom | | | | 3346 // | v \ 1v v2 v | 3347 // | ifB regionA ifB | 3348 // | / \ | / \ | 3349 // | / \ v / \ | 3350 // | v v exitA: v v | 3351 // | true false false true | 3352 // | / ^ \ / \ | 3353 // +---- \ \ / ----+ 3354 // dom \ / 3355 // \ 1v v2 3356 // regionB 3357 // | 3358 // v 3359 // exitB: 3360 // stmt4 3361 // 3362 // 3363 // after partial peel 3364 // 3365 // stmt1 3366 // | 3367 // v 3368 // loop predicate 3369 // / 3370 // clone / orig 3371 // / TOP 3372 // / \ 3373 // v v 3374 // TOP->loop loop----+ 3375 // | | | 3376 // stmt2 stmt2 | 3377 // | | | 3378 // v v | 3379 // ifA ifA | 3380 // | \ / | | 3381 // v v v v | 3382 // true false false true | <-- last_peel 3383 // | ^ \ / +------|---+ 3384 // +->newloop \ \ / === ==cut | | 3385 // | stmt3 \ \ / TOP | | 3386 // | | dom | | stmt3 | | <-- first_not_peel 3387 // | v \ 1v v2 v | | 3388 // | ifB regionA ifB ^ v 3389 // | / \ | / \ | | 3390 // | / \ v / \ | | 3391 // | v v exitA: v v | | 3392 // | true false false true | | 3393 // | / ^ \ / \ | | 3394 // | | \ \ / v | | 3395 // | | dom \ / TOP | | 3396 // | | \ 1v v2 | | 3397 // ^ v regionB | | 3398 // | | | | | 3399 // | | v ^ v 3400 // | | exitB: | | 3401 // | | stmt4 | | 3402 // | +------------>-----------------+ | 3403 // | | 3404 // +-----------------<---------------------+ 3405 // 3406 // 3407 // final graph 3408 // 3409 // stmt1 3410 // | 3411 // v 3412 // loop predicate 3413 // | 3414 // v 3415 // stmt2 clone 3416 // | 3417 // v 3418 // ........> ifA clone 3419 // : / | 3420 // dom / | 3421 // : v v 3422 // : false true 3423 // : | | 3424 // : | v 3425 // : | newloop<-----+ 3426 // : | | | 3427 // : | stmt3 clone | 3428 // : | | | 3429 // : | v | 3430 // : | ifB | 3431 // : | / \ | 3432 // : | v v | 3433 // : | false true | 3434 // : | | | | 3435 // : | v stmt2 | 3436 // : | exitB: | | 3437 // : | stmt4 v | 3438 // : | ifA orig | 3439 // : | / \ | 3440 // : | / \ | 3441 // : | v v | 3442 // : | false true | 3443 // : | / \ | 3444 // : v v -----+ 3445 // RegionA 3446 // | 3447 // v 3448 // exitA 3449 // 3450 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) { 3451 3452 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only"); 3453 if (!loop->_head->is_Loop()) { 3454 return false; 3455 } 3456 LoopNode *head = loop->_head->as_Loop(); 3457 3458 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) { 3459 return false; 3460 } 3461 3462 // Check for complex exit control 3463 for (uint ii = 0; ii < loop->_body.size(); ii++) { 3464 Node *n = loop->_body.at(ii); 3465 int opc = n->Opcode(); 3466 if (n->is_Call() || 3467 opc == Op_Catch || 3468 opc == Op_CatchProj || 3469 opc == Op_Jump || 3470 opc == Op_JumpProj) { 3471 #ifndef PRODUCT 3472 if (TracePartialPeeling) { 3473 tty->print_cr("\nExit control too complex: lp: %d", head->_idx); 3474 } 3475 #endif 3476 return false; 3477 } 3478 } 3479 3480 int dd = dom_depth(head); 3481 3482 // Step 1: find cut point 3483 3484 // Walk up dominators to loop head looking for first loop exit 3485 // which is executed on every path thru loop. 3486 IfNode *peel_if = NULL; 3487 IfNode *peel_if_cmpu = NULL; 3488 3489 Node *iff = loop->tail(); 3490 while (iff != head) { 3491 if (iff->is_If()) { 3492 Node *ctrl = get_ctrl(iff->in(1)); 3493 if (ctrl->is_top()) return false; // Dead test on live IF. 3494 // If loop-varying exit-test, check for induction variable 3495 if (loop->is_member(get_loop(ctrl)) && 3496 loop->is_loop_exit(iff) && 3497 is_possible_iv_test(iff)) { 3498 Node* cmp = iff->in(1)->in(1); 3499 if (cmp->Opcode() == Op_CmpI) { 3500 peel_if = iff->as_If(); 3501 } else { 3502 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU"); 3503 peel_if_cmpu = iff->as_If(); 3504 } 3505 } 3506 } 3507 iff = idom(iff); 3508 } 3509 3510 // Prefer signed compare over unsigned compare. 3511 IfNode* new_peel_if = NULL; 3512 if (peel_if == NULL) { 3513 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) { 3514 return false; // No peel point found 3515 } 3516 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop); 3517 if (new_peel_if == NULL) { 3518 return false; // No peel point found 3519 } 3520 peel_if = new_peel_if; 3521 } 3522 Node* last_peel = stay_in_loop(peel_if, loop); 3523 Node* first_not_peeled = stay_in_loop(last_peel, loop); 3524 if (first_not_peeled == NULL || first_not_peeled == head) { 3525 return false; 3526 } 3527 3528 #ifndef PRODUCT 3529 if (TraceLoopOpts) { 3530 tty->print("PartialPeel "); 3531 loop->dump_head(); 3532 } 3533 3534 if (TracePartialPeeling) { 3535 tty->print_cr("before partial peel one iteration"); 3536 Node_List wl; 3537 Node* t = head->in(2); 3538 while (true) { 3539 wl.push(t); 3540 if (t == head) break; 3541 t = idom(t); 3542 } 3543 while (wl.size() > 0) { 3544 Node* tt = wl.pop(); 3545 tt->dump(); 3546 if (tt == last_peel) tty->print_cr("-- cut --"); 3547 } 3548 } 3549 #endif 3550 VectorSet peel; 3551 VectorSet not_peel; 3552 Node_List peel_list; 3553 Node_List worklist; 3554 Node_List sink_list; 3555 3556 uint estimate = loop->est_loop_clone_sz(1); 3557 if (exceeding_node_budget(estimate)) { 3558 return false; 3559 } 3560 3561 // Set of cfg nodes to peel are those that are executable from 3562 // the head through last_peel. 3563 assert(worklist.size() == 0, "should be empty"); 3564 worklist.push(head); 3565 peel.set(head->_idx); 3566 while (worklist.size() > 0) { 3567 Node *n = worklist.pop(); 3568 if (n != last_peel) { 3569 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3570 Node* use = n->fast_out(j); 3571 if (use->is_CFG() && 3572 loop->is_member(get_loop(use)) && 3573 !peel.test_set(use->_idx)) { 3574 worklist.push(use); 3575 } 3576 } 3577 } 3578 } 3579 3580 // Set of non-cfg nodes to peel are those that are control 3581 // dependent on the cfg nodes. 3582 for (uint i = 0; i < loop->_body.size(); i++) { 3583 Node *n = loop->_body.at(i); 3584 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n; 3585 if (peel.test(n_c->_idx)) { 3586 peel.set(n->_idx); 3587 } else { 3588 not_peel.set(n->_idx); 3589 } 3590 } 3591 3592 // Step 2: move operations from the peeled section down into the 3593 // not-peeled section 3594 3595 // Get a post order schedule of nodes in the peel region 3596 // Result in right-most operand. 3597 scheduled_nodelist(loop, peel, peel_list); 3598 3599 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 3600 3601 // For future check for too many new phis 3602 uint old_phi_cnt = 0; 3603 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 3604 Node* use = head->fast_out(j); 3605 if (use->is_Phi()) old_phi_cnt++; 3606 } 3607 3608 #ifndef PRODUCT 3609 if (TracePartialPeeling) { 3610 tty->print_cr("\npeeled list"); 3611 } 3612 #endif 3613 3614 // Evacuate nodes in peel region into the not_peeled region if possible 3615 bool too_many_clones = false; 3616 uint new_phi_cnt = 0; 3617 uint cloned_for_outside_use = 0; 3618 for (uint i = 0; i < peel_list.size();) { 3619 Node* n = peel_list.at(i); 3620 #ifndef PRODUCT 3621 if (TracePartialPeeling) n->dump(); 3622 #endif 3623 bool incr = true; 3624 if (!n->is_CFG()) { 3625 if (has_use_in_set(n, not_peel)) { 3626 // If not used internal to the peeled region, 3627 // move "n" from peeled to not_peeled region. 3628 if (!has_use_internal_to_set(n, peel, loop)) { 3629 // if not pinned and not a load (which maybe anti-dependent on a store) 3630 // and not a CMove (Matcher expects only bool->cmove). 3631 if (n->in(0) == NULL && !n->is_Load() && !n->is_CMove()) { 3632 int new_clones = clone_for_use_outside_loop(loop, n, worklist); 3633 if (new_clones == -1) { 3634 too_many_clones = true; 3635 break; 3636 } 3637 cloned_for_outside_use += new_clones; 3638 sink_list.push(n); 3639 peel.remove(n->_idx); 3640 not_peel.set(n->_idx); 3641 peel_list.remove(i); 3642 incr = false; 3643 #ifndef PRODUCT 3644 if (TracePartialPeeling) { 3645 tty->print_cr("sink to not_peeled region: %d newbb: %d", 3646 n->_idx, get_ctrl(n)->_idx); 3647 } 3648 #endif 3649 } 3650 } else { 3651 // Otherwise check for special def-use cases that span 3652 // the peel/not_peel boundary such as bool->if 3653 clone_for_special_use_inside_loop(loop, n, not_peel, sink_list, worklist); 3654 new_phi_cnt++; 3655 } 3656 } 3657 } 3658 if (incr) i++; 3659 } 3660 3661 estimate += cloned_for_outside_use + new_phi_cnt; 3662 bool exceed_node_budget = !may_require_nodes(estimate); 3663 bool exceed_phi_limit = new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta; 3664 3665 if (too_many_clones || exceed_node_budget || exceed_phi_limit) { 3666 #ifndef PRODUCT 3667 if (TracePartialPeeling && exceed_phi_limit) { 3668 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c", 3669 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F'); 3670 } 3671 #endif 3672 if (new_peel_if != NULL) { 3673 remove_cmpi_loop_exit(new_peel_if, loop); 3674 } 3675 // Inhibit more partial peeling on this loop 3676 assert(!head->is_partial_peel_loop(), "not partial peeled"); 3677 head->mark_partial_peel_failed(); 3678 if (cloned_for_outside_use > 0) { 3679 // Terminate this round of loop opts because 3680 // the graph outside this loop was changed. 3681 C->set_major_progress(); 3682 return true; 3683 } 3684 return false; 3685 } 3686 3687 // Step 3: clone loop, retarget control, and insert new phis 3688 3689 // Create new loop head for new phis and to hang 3690 // the nodes being moved (sinked) from the peel region. 3691 LoopNode* new_head = new LoopNode(last_peel, last_peel); 3692 new_head->set_unswitch_count(head->unswitch_count()); // Preserve 3693 _igvn.register_new_node_with_optimizer(new_head); 3694 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled"); 3695 _igvn.replace_input_of(first_not_peeled, 0, new_head); 3696 set_loop(new_head, loop); 3697 loop->_body.push(new_head); 3698 not_peel.set(new_head->_idx); 3699 set_idom(new_head, last_peel, dom_depth(first_not_peeled)); 3700 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled)); 3701 3702 while (sink_list.size() > 0) { 3703 Node* n = sink_list.pop(); 3704 set_ctrl(n, new_head); 3705 } 3706 3707 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 3708 3709 clone_loop(loop, old_new, dd, IgnoreStripMined); 3710 3711 const uint clone_exit_idx = 1; 3712 const uint orig_exit_idx = 2; 3713 assert(is_valid_clone_loop_form(loop, peel_list, orig_exit_idx, clone_exit_idx), "bad clone loop"); 3714 3715 Node* head_clone = old_new[head->_idx]; 3716 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop(); 3717 Node* orig_tail_clone = head_clone->in(2); 3718 3719 // Add phi if "def" node is in peel set and "use" is not 3720 3721 for (uint i = 0; i < peel_list.size(); i++) { 3722 Node *def = peel_list.at(i); 3723 if (!def->is_CFG()) { 3724 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 3725 Node *use = def->fast_out(j); 3726 if (has_node(use) && use->in(0) != C->top() && 3727 (!peel.test(use->_idx) || 3728 (use->is_Phi() && use->in(0) == head)) ) { 3729 worklist.push(use); 3730 } 3731 } 3732 while( worklist.size() ) { 3733 Node *use = worklist.pop(); 3734 for (uint j = 1; j < use->req(); j++) { 3735 Node* n = use->in(j); 3736 if (n == def) { 3737 3738 // "def" is in peel set, "use" is not in peel set 3739 // or "use" is in the entry boundary (a phi) of the peel set 3740 3741 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use; 3742 3743 if ( loop->is_member(get_loop( use_c )) ) { 3744 // use is in loop 3745 if (old_new[use->_idx] != NULL) { // null for dead code 3746 Node* use_clone = old_new[use->_idx]; 3747 _igvn.replace_input_of(use, j, C->top()); 3748 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone ); 3749 } 3750 } else { 3751 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format"); 3752 // use is not in the loop, check if the live range includes the cut 3753 Node* lp_if = use_c->in(orig_exit_idx)->in(0); 3754 if (not_peel.test(lp_if->_idx)) { 3755 assert(j == orig_exit_idx, "use from original loop"); 3756 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone ); 3757 } 3758 } 3759 } 3760 } 3761 } 3762 } 3763 } 3764 3765 // Step 3b: retarget control 3766 3767 // Redirect control to the new loop head if a cloned node in 3768 // the not_peeled region has control that points into the peeled region. 3769 // This necessary because the cloned peeled region will be outside 3770 // the loop. 3771 // from to 3772 // cloned-peeled <---+ 3773 // new_head_clone: | <--+ 3774 // cloned-not_peeled in(0) in(0) 3775 // orig-peeled 3776 3777 for (uint i = 0; i < loop->_body.size(); i++) { 3778 Node *n = loop->_body.at(i); 3779 if (!n->is_CFG() && n->in(0) != NULL && 3780 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) { 3781 Node* n_clone = old_new[n->_idx]; 3782 _igvn.replace_input_of(n_clone, 0, new_head_clone); 3783 } 3784 } 3785 3786 // Backedge of the surviving new_head (the clone) is original last_peel 3787 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel); 3788 3789 // Cut first node in original not_peel set 3790 _igvn.rehash_node_delayed(new_head); // Multiple edge updates: 3791 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of 3792 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls 3793 3794 // Copy head_clone back-branch info to original head 3795 // and remove original head's loop entry and 3796 // clone head's back-branch 3797 _igvn.rehash_node_delayed(head); // Multiple edge updates 3798 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl)); 3799 head->set_req(LoopNode::LoopBackControl, C->top()); 3800 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top()); 3801 3802 // Similarly modify the phis 3803 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) { 3804 Node* use = head->fast_out(k); 3805 if (use->is_Phi() && use->outcnt() > 0) { 3806 Node* use_clone = old_new[use->_idx]; 3807 _igvn.rehash_node_delayed(use); // Multiple edge updates 3808 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl)); 3809 use->set_req(LoopNode::LoopBackControl, C->top()); 3810 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top()); 3811 } 3812 } 3813 3814 // Step 4: update dominator tree and dominator depth 3815 3816 set_idom(head, orig_tail_clone, dd); 3817 recompute_dom_depth(); 3818 3819 // Inhibit more partial peeling on this loop 3820 new_head_clone->set_partial_peel_loop(); 3821 C->set_major_progress(); 3822 loop->record_for_igvn(); 3823 3824 #ifndef PRODUCT 3825 if (TracePartialPeeling) { 3826 tty->print_cr("\nafter partial peel one iteration"); 3827 Node_List wl; 3828 Node* t = last_peel; 3829 while (true) { 3830 wl.push(t); 3831 if (t == head_clone) break; 3832 t = idom(t); 3833 } 3834 while (wl.size() > 0) { 3835 Node* tt = wl.pop(); 3836 if (tt == head) tty->print_cr("orig head"); 3837 else if (tt == new_head_clone) tty->print_cr("new head"); 3838 else if (tt == head_clone) tty->print_cr("clone head"); 3839 tt->dump(); 3840 } 3841 } 3842 #endif 3843 return true; 3844 } 3845 3846 // Transform: 3847 // 3848 // loop<-----------------+ 3849 // | | 3850 // stmt1 stmt2 .. stmtn | 3851 // | | | | 3852 // \ | / | 3853 // v v v | 3854 // region | 3855 // | | 3856 // shared_stmt | 3857 // | | 3858 // v | 3859 // if | 3860 // / \ | 3861 // | -----------+ 3862 // v 3863 // 3864 // into: 3865 // 3866 // loop<-------------------+ 3867 // | | 3868 // v | 3869 // +->loop | 3870 // | | | 3871 // | stmt1 stmt2 .. stmtn | 3872 // | | | | | 3873 // | | \ / | 3874 // | | v v | 3875 // | | region1 | 3876 // | | | | 3877 // | shared_stmt shared_stmt | 3878 // | | | | 3879 // | v v | 3880 // | if if | 3881 // | /\ / \ | 3882 // +-- | | -------+ 3883 // \ / 3884 // v v 3885 // region2 3886 // 3887 // (region2 is shown to merge mirrored projections of the loop exit 3888 // ifs to make the diagram clearer but they really merge the same 3889 // projection) 3890 // 3891 // Conditions for this transformation to trigger: 3892 // - the path through stmt1 is frequent enough 3893 // - the inner loop will be turned into a counted loop after transformation 3894 bool PhaseIdealLoop::duplicate_loop_backedge(IdealLoopTree *loop, Node_List &old_new) { 3895 if (!DuplicateBackedge) { 3896 return false; 3897 } 3898 assert(!loop->_head->is_CountedLoop() || StressDuplicateBackedge, "Non-counted loop only"); 3899 if (!loop->_head->is_Loop()) { 3900 return false; 3901 } 3902 3903 uint estimate = loop->est_loop_clone_sz(1); 3904 if (exceeding_node_budget(estimate)) { 3905 return false; 3906 } 3907 3908 LoopNode *head = loop->_head->as_Loop(); 3909 3910 Node* region = NULL; 3911 IfNode* exit_test = NULL; 3912 uint inner; 3913 float f; 3914 if (StressDuplicateBackedge) { 3915 if (head->is_strip_mined()) { 3916 return false; 3917 } 3918 Node* c = head->in(LoopNode::LoopBackControl); 3919 3920 while (c != head) { 3921 if (c->is_Region()) { 3922 region = c; 3923 } 3924 c = idom(c); 3925 } 3926 3927 if (region == NULL) { 3928 return false; 3929 } 3930 3931 inner = 1; 3932 } else { 3933 // Is the shape of the loop that of a counted loop... 3934 Node* back_control = loop_exit_control(head, loop); 3935 if (back_control == NULL) { 3936 return false; 3937 } 3938 3939 BoolTest::mask bt = BoolTest::illegal; 3940 float cl_prob = 0; 3941 Node* incr = NULL; 3942 Node* limit = NULL; 3943 Node* cmp = loop_exit_test(back_control, loop, incr, limit, bt, cl_prob); 3944 if (cmp == NULL || cmp->Opcode() != Op_CmpI) { 3945 return false; 3946 } 3947 3948 // With an extra phi for the candidate iv? 3949 // Or the region node is the loop head 3950 if (!incr->is_Phi() || incr->in(0) == head) { 3951 return false; 3952 } 3953 3954 PathFrequency pf(head, this); 3955 region = incr->in(0); 3956 3957 // Go over all paths for the extra phi's region and see if that 3958 // path is frequent enough and would match the expected iv shape 3959 // if the extra phi is removed 3960 inner = 0; 3961 for (uint i = 1; i < incr->req(); ++i) { 3962 Node* in = incr->in(i); 3963 Node* trunc1 = NULL; 3964 Node* trunc2 = NULL; 3965 const TypeInteger* iv_trunc_t = NULL; 3966 Node* orig_in = in; 3967 if (!(in = CountedLoopNode::match_incr_with_optional_truncation(in, &trunc1, &trunc2, &iv_trunc_t, T_INT))) { 3968 continue; 3969 } 3970 assert(in->Opcode() == Op_AddI, "wrong increment code"); 3971 Node* xphi = NULL; 3972 Node* stride = loop_iv_stride(in, loop, xphi); 3973 3974 if (stride == NULL) { 3975 continue; 3976 } 3977 3978 PhiNode* phi = loop_iv_phi(xphi, NULL, head, loop); 3979 if (phi == NULL || 3980 (trunc1 == NULL && phi->in(LoopNode::LoopBackControl) != incr) || 3981 (trunc1 != NULL && phi->in(LoopNode::LoopBackControl) != trunc1)) { 3982 return false; 3983 } 3984 3985 f = pf.to(region->in(i)); 3986 if (f > 0.5) { 3987 inner = i; 3988 break; 3989 } 3990 } 3991 3992 if (inner == 0) { 3993 return false; 3994 } 3995 3996 exit_test = back_control->in(0)->as_If(); 3997 } 3998 3999 if (idom(region)->is_Catch()) { 4000 return false; 4001 } 4002 4003 // Collect all control nodes that need to be cloned (shared_stmt in the diagram) 4004 Unique_Node_List wq; 4005 wq.push(head->in(LoopNode::LoopBackControl)); 4006 for (uint i = 0; i < wq.size(); i++) { 4007 Node* c = wq.at(i); 4008 assert(get_loop(c) == loop, "not in the right loop?"); 4009 if (c->is_Region()) { 4010 if (c != region) { 4011 for (uint j = 1; j < c->req(); ++j) { 4012 wq.push(c->in(j)); 4013 } 4014 } 4015 } else { 4016 wq.push(c->in(0)); 4017 } 4018 assert(!is_dominator(c, region) || c == region, "shouldn't go above region"); 4019 } 4020 4021 Node* region_dom = idom(region); 4022 4023 // Can't do the transformation if this would cause a membar pair to 4024 // be split 4025 for (uint i = 0; i < wq.size(); i++) { 4026 Node* c = wq.at(i); 4027 if (c->is_MemBar() && (c->as_MemBar()->trailing_store() || c->as_MemBar()->trailing_load_store())) { 4028 assert(c->as_MemBar()->leading_membar()->trailing_membar() == c, "bad membar pair"); 4029 if (!wq.member(c->as_MemBar()->leading_membar())) { 4030 return false; 4031 } 4032 } 4033 } 4034 4035 // Collect data nodes that need to be clones as well 4036 int dd = dom_depth(head); 4037 4038 for (uint i = 0; i < loop->_body.size(); ++i) { 4039 Node* n = loop->_body.at(i); 4040 if (has_ctrl(n)) { 4041 Node* c = get_ctrl(n); 4042 if (wq.member(c)) { 4043 wq.push(n); 4044 } 4045 } else { 4046 set_idom(n, idom(n), dd); 4047 } 4048 } 4049 4050 // clone shared_stmt 4051 clone_loop_body(wq, old_new, NULL); 4052 4053 Node* region_clone = old_new[region->_idx]; 4054 region_clone->set_req(inner, C->top()); 4055 set_idom(region, region->in(inner), dd); 4056 4057 // Prepare the outer loop 4058 Node* outer_head = new LoopNode(head->in(LoopNode::EntryControl), old_new[head->in(LoopNode::LoopBackControl)->_idx]); 4059 register_control(outer_head, loop->_parent, outer_head->in(LoopNode::EntryControl)); 4060 _igvn.replace_input_of(head, LoopNode::EntryControl, outer_head); 4061 set_idom(head, outer_head, dd); 4062 4063 fix_body_edges(wq, loop, old_new, dd, loop->_parent, true); 4064 4065 // Make one of the shared_stmt copies only reachable from stmt1, the 4066 // other only from stmt2..stmtn. 4067 Node* dom = NULL; 4068 for (uint i = 1; i < region->req(); ++i) { 4069 if (i != inner) { 4070 _igvn.replace_input_of(region, i, C->top()); 4071 } 4072 Node* in = region_clone->in(i); 4073 if (in->is_top()) { 4074 continue; 4075 } 4076 if (dom == NULL) { 4077 dom = in; 4078 } else { 4079 dom = dom_lca(dom, in); 4080 } 4081 } 4082 4083 set_idom(region_clone, dom, dd); 4084 4085 // Set up the outer loop 4086 for (uint i = 0; i < head->outcnt(); i++) { 4087 Node* u = head->raw_out(i); 4088 if (u->is_Phi()) { 4089 Node* outer_phi = u->clone(); 4090 outer_phi->set_req(0, outer_head); 4091 Node* backedge = old_new[u->in(LoopNode::LoopBackControl)->_idx]; 4092 if (backedge == NULL) { 4093 backedge = u->in(LoopNode::LoopBackControl); 4094 } 4095 outer_phi->set_req(LoopNode::LoopBackControl, backedge); 4096 register_new_node(outer_phi, outer_head); 4097 _igvn.replace_input_of(u, LoopNode::EntryControl, outer_phi); 4098 } 4099 } 4100 4101 // create control and data nodes for out of loop uses (including region2) 4102 Node_List worklist; 4103 uint new_counter = C->unique(); 4104 fix_ctrl_uses(wq, loop, old_new, ControlAroundStripMined, outer_head, NULL, worklist); 4105 4106 Node_List *split_if_set = NULL; 4107 Node_List *split_bool_set = NULL; 4108 Node_List *split_cex_set = NULL; 4109 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); 4110 4111 finish_clone_loop(split_if_set, split_bool_set, split_cex_set); 4112 4113 if (exit_test != NULL) { 4114 float cnt = exit_test->_fcnt; 4115 if (cnt != COUNT_UNKNOWN) { 4116 exit_test->_fcnt = cnt * f; 4117 old_new[exit_test->_idx]->as_If()->_fcnt = cnt * (1 - f); 4118 } 4119 } 4120 4121 C->set_major_progress(); 4122 4123 return true; 4124 }