1 /* 2 * Copyright (c) 1999, 2019, 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/valuetypenode.hpp" 44 #include "utilities/macros.hpp" 45 #if INCLUDE_ZGC 46 #include "gc/z/c2/zBarrierSetC2.hpp" 47 #endif 48 49 //============================================================================= 50 //------------------------------split_thru_phi--------------------------------- 51 // Split Node 'n' through merge point if there is enough win. 52 Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) { 53 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) { 54 // ConvI2L may have type information on it which is unsafe to push up 55 // so disable this for now 56 return NULL; 57 } 58 59 // Splitting range check CastIIs through a loop induction Phi can 60 // cause new Phis to be created that are left unrelated to the loop 61 // induction Phi and prevent optimizations (vectorization) 62 if (n->Opcode() == Op_CastII && n->as_CastII()->has_range_check() && 63 region->is_CountedLoop() && n->in(1) == region->as_CountedLoop()->phi()) { 64 return NULL; 65 } 66 67 // Value types should not be split through Phis because they cannot be merged 68 // through Phi nodes but each value input needs to be merged individually. 69 if (n->is_ValueType()) { 70 return NULL; 71 } 72 73 int wins = 0; 74 assert(!n->is_CFG(), ""); 75 assert(region->is_Region(), ""); 76 77 const Type* type = n->bottom_type(); 78 const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr(); 79 Node *phi; 80 if (t_oop != NULL && t_oop->is_known_instance_field()) { 81 int iid = t_oop->instance_id(); 82 int index = C->get_alias_index(t_oop); 83 int offset = t_oop->offset(); 84 phi = new PhiNode(region, type, NULL, iid, index, offset); 85 } else { 86 phi = PhiNode::make_blank(region, n); 87 } 88 uint old_unique = C->unique(); 89 for (uint i = 1; i < region->req(); i++) { 90 Node *x; 91 Node* the_clone = NULL; 92 if (region->in(i) == C->top()) { 93 x = C->top(); // Dead path? Use a dead data op 94 } else { 95 x = n->clone(); // Else clone up the data op 96 the_clone = x; // Remember for possible deletion. 97 // Alter data node to use pre-phi inputs 98 if (n->in(0) == region) 99 x->set_req( 0, region->in(i) ); 100 for (uint j = 1; j < n->req(); j++) { 101 Node *in = n->in(j); 102 if (in->is_Phi() && in->in(0) == region) 103 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone 104 } 105 } 106 // Check for a 'win' on some paths 107 const Type *t = x->Value(&_igvn); 108 109 bool singleton = t->singleton(); 110 111 // A TOP singleton indicates that there are no possible values incoming 112 // along a particular edge. In most cases, this is OK, and the Phi will 113 // be eliminated later in an Ideal call. However, we can't allow this to 114 // happen if the singleton occurs on loop entry, as the elimination of 115 // the PhiNode may cause the resulting node to migrate back to a previous 116 // loop iteration. 117 if (singleton && t == Type::TOP) { 118 // Is_Loop() == false does not confirm the absence of a loop (e.g., an 119 // irreducible loop may not be indicated by an affirmative is_Loop()); 120 // therefore, the only top we can split thru a phi is on a backedge of 121 // a loop. 122 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 123 } 124 125 if (singleton) { 126 wins++; 127 x = ((PhaseGVN&)_igvn).makecon(t); 128 } else { 129 // We now call Identity to try to simplify the cloned node. 130 // Note that some Identity methods call phase->type(this). 131 // Make sure that the type array is big enough for 132 // our new node, even though we may throw the node away. 133 // (Note: This tweaking with igvn only works because x is a new node.) 134 _igvn.set_type(x, t); 135 // If x is a TypeNode, capture any more-precise type permanently into Node 136 // otherwise it will be not updated during igvn->transform since 137 // igvn->type(x) is set to x->Value() already. 138 x->raise_bottom_type(t); 139 Node *y = _igvn.apply_identity(x); 140 if (y != x) { 141 wins++; 142 x = y; 143 } else { 144 y = _igvn.hash_find(x); 145 if (y) { 146 wins++; 147 x = y; 148 } else { 149 // Else x is a new node we are keeping 150 // We do not need register_new_node_with_optimizer 151 // because set_type has already been called. 152 _igvn._worklist.push(x); 153 } 154 } 155 } 156 if (x != the_clone && the_clone != NULL) 157 _igvn.remove_dead_node(the_clone); 158 phi->set_req( i, x ); 159 } 160 // Too few wins? 161 if (wins <= policy) { 162 _igvn.remove_dead_node(phi); 163 return NULL; 164 } 165 166 // Record Phi 167 register_new_node( phi, region ); 168 169 for (uint i2 = 1; i2 < phi->req(); i2++) { 170 Node *x = phi->in(i2); 171 // If we commoned up the cloned 'x' with another existing Node, 172 // the existing Node picks up a new use. We need to make the 173 // existing Node occur higher up so it dominates its uses. 174 Node *old_ctrl; 175 IdealLoopTree *old_loop; 176 177 if (x->is_Con()) { 178 // Constant's control is always root. 179 set_ctrl(x, C->root()); 180 continue; 181 } 182 // The occasional new node 183 if (x->_idx >= old_unique) { // Found a new, unplaced node? 184 old_ctrl = NULL; 185 old_loop = NULL; // Not in any prior loop 186 } else { 187 old_ctrl = get_ctrl(x); 188 old_loop = get_loop(old_ctrl); // Get prior loop 189 } 190 // New late point must dominate new use 191 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2)); 192 if (new_ctrl == old_ctrl) // Nothing is changed 193 continue; 194 195 IdealLoopTree *new_loop = get_loop(new_ctrl); 196 197 // Don't move x into a loop if its uses are 198 // outside of loop. Otherwise x will be cloned 199 // for each use outside of this loop. 200 IdealLoopTree *use_loop = get_loop(region); 201 if (!new_loop->is_member(use_loop) && 202 (old_loop == NULL || !new_loop->is_member(old_loop))) { 203 // Take early control, later control will be recalculated 204 // during next iteration of loop optimizations. 205 new_ctrl = get_early_ctrl(x); 206 new_loop = get_loop(new_ctrl); 207 } 208 // Set new location 209 set_ctrl(x, new_ctrl); 210 // If changing loop bodies, see if we need to collect into new body 211 if (old_loop != new_loop) { 212 if (old_loop && !old_loop->_child) 213 old_loop->_body.yank(x); 214 if (!new_loop->_child) 215 new_loop->_body.push(x); // Collect body info 216 } 217 } 218 219 return phi; 220 } 221 222 //------------------------------dominated_by------------------------------------ 223 // Replace the dominated test with an obvious true or false. Place it on the 224 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the 225 // live path up to the dominating control. 226 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff, bool flip, bool exclude_loop_predicate ) { 227 if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); } 228 229 // prevdom is the dominating projection of the dominating test. 230 assert( iff->is_If(), "" ); 231 assert(iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd || iff->Opcode() == Op_RangeCheck, "Check this code when new subtype is added"); 232 int pop = prevdom->Opcode(); 233 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" ); 234 if (flip) { 235 if (pop == Op_IfTrue) 236 pop = Op_IfFalse; 237 else 238 pop = Op_IfTrue; 239 } 240 // 'con' is set to true or false to kill the dominated test. 241 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO); 242 set_ctrl(con, C->root()); // Constant gets a new use 243 // Hack the dominated test 244 _igvn.replace_input_of(iff, 1, con); 245 246 // If I dont have a reachable TRUE and FALSE path following the IfNode then 247 // I can assume this path reaches an infinite loop. In this case it's not 248 // important to optimize the data Nodes - either the whole compilation will 249 // be tossed or this path (and all data Nodes) will go dead. 250 if (iff->outcnt() != 2) return; 251 252 // Make control-dependent data Nodes on the live path (path that will remain 253 // once the dominated IF is removed) become control-dependent on the 254 // dominating projection. 255 Node* dp = iff->as_If()->proj_out_or_null(pop == Op_IfTrue); 256 257 // Loop predicates may have depending checks which should not 258 // be skipped. For example, range check predicate has two checks 259 // for lower and upper bounds. 260 if (dp == NULL) 261 return; 262 263 ProjNode* dp_proj = dp->as_Proj(); 264 ProjNode* unc_proj = iff->as_If()->proj_out(1 - dp_proj->_con)->as_Proj(); 265 if (exclude_loop_predicate && 266 (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) != NULL || 267 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_profile_predicate) != NULL || 268 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check) != NULL)) { 269 // If this is a range check (IfNode::is_range_check), do not 270 // reorder because Compile::allow_range_check_smearing might have 271 // changed the check. 272 return; // Let IGVN transformation change control dependence. 273 } 274 275 IdealLoopTree *old_loop = get_loop(dp); 276 277 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { 278 Node* cd = dp->fast_out(i); // Control-dependent node 279 if (cd->depends_only_on_test()) { 280 assert(cd->in(0) == dp, ""); 281 _igvn.replace_input_of(cd, 0, prevdom); 282 set_early_ctrl(cd); 283 IdealLoopTree *new_loop = get_loop(get_ctrl(cd)); 284 if (old_loop != new_loop) { 285 if (!old_loop->_child) old_loop->_body.yank(cd); 286 if (!new_loop->_child) new_loop->_body.push(cd); 287 } 288 --i; 289 --imax; 290 } 291 } 292 } 293 294 //------------------------------has_local_phi_input---------------------------- 295 // Return TRUE if 'n' has Phi inputs from its local block and no other 296 // block-local inputs (all non-local-phi inputs come from earlier blocks) 297 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) { 298 Node *n_ctrl = get_ctrl(n); 299 // See if some inputs come from a Phi in this block, or from before 300 // this block. 301 uint i; 302 for( i = 1; i < n->req(); i++ ) { 303 Node *phi = n->in(i); 304 if( phi->is_Phi() && phi->in(0) == n_ctrl ) 305 break; 306 } 307 if( i >= n->req() ) 308 return NULL; // No Phi inputs; nowhere to clone thru 309 310 // Check for inputs created between 'n' and the Phi input. These 311 // must split as well; they have already been given the chance 312 // (courtesy of a post-order visit) and since they did not we must 313 // recover the 'cost' of splitting them by being very profitable 314 // when splitting 'n'. Since this is unlikely we simply give up. 315 for( i = 1; i < n->req(); i++ ) { 316 Node *m = n->in(i); 317 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) { 318 // We allow the special case of AddP's with no local inputs. 319 // This allows us to split-up address expressions. 320 if (m->is_AddP() && 321 get_ctrl(m->in(2)) != n_ctrl && 322 get_ctrl(m->in(3)) != n_ctrl) { 323 // Move the AddP up to dominating point 324 Node* c = find_non_split_ctrl(idom(n_ctrl)); 325 if (c->is_OuterStripMinedLoop()) { 326 c->as_Loop()->verify_strip_mined(1); 327 c = c->in(LoopNode::EntryControl); 328 } 329 set_ctrl_and_loop(m, c); 330 continue; 331 } 332 return NULL; 333 } 334 assert(n->is_Phi() || m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control"); 335 } 336 337 return n_ctrl; 338 } 339 340 //------------------------------remix_address_expressions---------------------- 341 // Rework addressing expressions to get the most loop-invariant stuff 342 // moved out. We'd like to do all associative operators, but it's especially 343 // important (common) to do address expressions. 344 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) { 345 if (!has_ctrl(n)) return NULL; 346 Node *n_ctrl = get_ctrl(n); 347 IdealLoopTree *n_loop = get_loop(n_ctrl); 348 349 // See if 'n' mixes loop-varying and loop-invariant inputs and 350 // itself is loop-varying. 351 352 // Only interested in binary ops (and AddP) 353 if( n->req() < 3 || n->req() > 4 ) return NULL; 354 355 Node *n1_ctrl = get_ctrl(n->in( 1)); 356 Node *n2_ctrl = get_ctrl(n->in( 2)); 357 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3)); 358 IdealLoopTree *n1_loop = get_loop( n1_ctrl ); 359 IdealLoopTree *n2_loop = get_loop( n2_ctrl ); 360 IdealLoopTree *n3_loop = get_loop( n3_ctrl ); 361 362 // Does one of my inputs spin in a tighter loop than self? 363 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) || 364 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) || 365 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) ) 366 return NULL; // Leave well enough alone 367 368 // Is at least one of my inputs loop-invariant? 369 if( n1_loop == n_loop && 370 n2_loop == n_loop && 371 n3_loop == n_loop ) 372 return NULL; // No loop-invariant inputs 373 374 375 int n_op = n->Opcode(); 376 377 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2). 378 if( n_op == Op_LShiftI ) { 379 // Scale is loop invariant 380 Node *scale = n->in(2); 381 Node *scale_ctrl = get_ctrl(scale); 382 IdealLoopTree *scale_loop = get_loop(scale_ctrl ); 383 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) ) 384 return NULL; 385 const TypeInt *scale_t = scale->bottom_type()->isa_int(); 386 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 ) 387 return NULL; // Dont bother with byte/short masking 388 // Add must vary with loop (else shift would be loop-invariant) 389 Node *add = n->in(1); 390 Node *add_ctrl = get_ctrl(add); 391 IdealLoopTree *add_loop = get_loop(add_ctrl); 392 //assert( n_loop == add_loop, "" ); 393 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops 394 395 // Convert I-V into I+ (0-V); same for V-I 396 if( add->Opcode() == Op_SubI && 397 _igvn.type( add->in(1) ) != TypeInt::ZERO ) { 398 Node *zero = _igvn.intcon(0); 399 set_ctrl(zero, C->root()); 400 Node *neg = new SubINode( _igvn.intcon(0), add->in(2) ); 401 register_new_node( neg, get_ctrl(add->in(2) ) ); 402 add = new AddINode( add->in(1), neg ); 403 register_new_node( add, add_ctrl ); 404 } 405 if( add->Opcode() != Op_AddI ) return NULL; 406 // See if one add input is loop invariant 407 Node *add_var = add->in(1); 408 Node *add_var_ctrl = get_ctrl(add_var); 409 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 410 Node *add_invar = add->in(2); 411 Node *add_invar_ctrl = get_ctrl(add_invar); 412 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl ); 413 if( add_var_loop == n_loop ) { 414 } else if( add_invar_loop == n_loop ) { 415 // Swap to find the invariant part 416 add_invar = add_var; 417 add_invar_ctrl = add_var_ctrl; 418 add_invar_loop = add_var_loop; 419 add_var = add->in(2); 420 Node *add_var_ctrl = get_ctrl(add_var); 421 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 422 } else // Else neither input is loop invariant 423 return NULL; 424 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) ) 425 return NULL; // No invariant part of the add? 426 427 // Yes! Reshape address expression! 428 Node *inv_scale = new LShiftINode( add_invar, scale ); 429 Node *inv_scale_ctrl = 430 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ? 431 add_invar_ctrl : scale_ctrl; 432 register_new_node( inv_scale, inv_scale_ctrl ); 433 Node *var_scale = new LShiftINode( add_var, scale ); 434 register_new_node( var_scale, n_ctrl ); 435 Node *var_add = new AddINode( var_scale, inv_scale ); 436 register_new_node( var_add, n_ctrl ); 437 _igvn.replace_node( n, var_add ); 438 return var_add; 439 } 440 441 // Replace (I+V) with (V+I) 442 if( n_op == Op_AddI || 443 n_op == Op_AddL || 444 n_op == Op_AddF || 445 n_op == Op_AddD || 446 n_op == Op_MulI || 447 n_op == Op_MulL || 448 n_op == Op_MulF || 449 n_op == Op_MulD ) { 450 if( n2_loop == n_loop ) { 451 assert( n1_loop != n_loop, "" ); 452 n->swap_edges(1, 2); 453 } 454 } 455 456 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V), 457 // but not if I2 is a constant. 458 if( n_op == Op_AddP ) { 459 if( n2_loop == n_loop && n3_loop != n_loop ) { 460 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) { 461 Node *n22_ctrl = get_ctrl(n->in(2)->in(2)); 462 Node *n23_ctrl = get_ctrl(n->in(2)->in(3)); 463 IdealLoopTree *n22loop = get_loop( n22_ctrl ); 464 IdealLoopTree *n23_loop = get_loop( n23_ctrl ); 465 if( n22loop != n_loop && n22loop->is_member(n_loop) && 466 n23_loop == n_loop ) { 467 Node *add1 = new AddPNode( n->in(1), n->in(2)->in(2), n->in(3) ); 468 // Stuff new AddP in the loop preheader 469 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) ); 470 Node *add2 = new AddPNode( n->in(1), add1, n->in(2)->in(3) ); 471 register_new_node( add2, n_ctrl ); 472 _igvn.replace_node( n, add2 ); 473 return add2; 474 } 475 } 476 } 477 478 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V) 479 if (n2_loop != n_loop && n3_loop == n_loop) { 480 if (n->in(3)->Opcode() == Op_AddX) { 481 Node *V = n->in(3)->in(1); 482 Node *I = n->in(3)->in(2); 483 if (is_member(n_loop,get_ctrl(V))) { 484 } else { 485 Node *tmp = V; V = I; I = tmp; 486 } 487 if (!is_member(n_loop,get_ctrl(I))) { 488 Node *add1 = new AddPNode(n->in(1), n->in(2), I); 489 // Stuff new AddP in the loop preheader 490 register_new_node(add1, n_loop->_head->in(LoopNode::EntryControl)); 491 Node *add2 = new AddPNode(n->in(1), add1, V); 492 register_new_node(add2, n_ctrl); 493 _igvn.replace_node(n, add2); 494 return add2; 495 } 496 } 497 } 498 } 499 500 return NULL; 501 } 502 503 // Optimize ((in1[2*i] * in2[2*i]) + (in1[2*i+1] * in2[2*i+1])) 504 Node *PhaseIdealLoop::convert_add_to_muladd(Node* n) { 505 assert(n->Opcode() == Op_AddI, "sanity"); 506 Node * nn = NULL; 507 Node * in1 = n->in(1); 508 Node * in2 = n->in(2); 509 if (in1->Opcode() == Op_MulI && in2->Opcode() == Op_MulI) { 510 IdealLoopTree* loop_n = get_loop(get_ctrl(n)); 511 if (loop_n->_head->as_Loop()->is_valid_counted_loop() && 512 Matcher::match_rule_supported(Op_MulAddS2I) && 513 Matcher::match_rule_supported(Op_MulAddVS2VI)) { 514 Node* mul_in1 = in1->in(1); 515 Node* mul_in2 = in1->in(2); 516 Node* mul_in3 = in2->in(1); 517 Node* mul_in4 = in2->in(2); 518 if (mul_in1->Opcode() == Op_LoadS && 519 mul_in2->Opcode() == Op_LoadS && 520 mul_in3->Opcode() == Op_LoadS && 521 mul_in4->Opcode() == Op_LoadS) { 522 IdealLoopTree* loop1 = get_loop(get_ctrl(mul_in1)); 523 IdealLoopTree* loop2 = get_loop(get_ctrl(mul_in2)); 524 IdealLoopTree* loop3 = get_loop(get_ctrl(mul_in3)); 525 IdealLoopTree* loop4 = get_loop(get_ctrl(mul_in4)); 526 IdealLoopTree* loop5 = get_loop(get_ctrl(in1)); 527 IdealLoopTree* loop6 = get_loop(get_ctrl(in2)); 528 // All nodes should be in the same counted loop. 529 if (loop_n == loop1 && loop_n == loop2 && loop_n == loop3 && 530 loop_n == loop4 && loop_n == loop5 && loop_n == loop6) { 531 Node* adr1 = mul_in1->in(MemNode::Address); 532 Node* adr2 = mul_in2->in(MemNode::Address); 533 Node* adr3 = mul_in3->in(MemNode::Address); 534 Node* adr4 = mul_in4->in(MemNode::Address); 535 if (adr1->is_AddP() && adr2->is_AddP() && adr3->is_AddP() && adr4->is_AddP()) { 536 if ((adr1->in(AddPNode::Base) == adr3->in(AddPNode::Base)) && 537 (adr2->in(AddPNode::Base) == adr4->in(AddPNode::Base))) { 538 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in3, mul_in4); 539 register_new_node(nn, get_ctrl(n)); 540 _igvn.replace_node(n, nn); 541 return nn; 542 } else if ((adr1->in(AddPNode::Base) == adr4->in(AddPNode::Base)) && 543 (adr2->in(AddPNode::Base) == adr3->in(AddPNode::Base))) { 544 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in4, mul_in3); 545 register_new_node(nn, get_ctrl(n)); 546 _igvn.replace_node(n, nn); 547 return nn; 548 } 549 } 550 } 551 } 552 } 553 } 554 return nn; 555 } 556 557 //------------------------------conditional_move------------------------------- 558 // Attempt to replace a Phi with a conditional move. We have some pretty 559 // strict profitability requirements. All Phis at the merge point must 560 // be converted, so we can remove the control flow. We need to limit the 561 // number of c-moves to a small handful. All code that was in the side-arms 562 // of the CFG diamond is now speculatively executed. This code has to be 563 // "cheap enough". We are pretty much limited to CFG diamonds that merge 564 // 1 or 2 items with a total of 1 or 2 ops executed speculatively. 565 Node *PhaseIdealLoop::conditional_move( Node *region ) { 566 567 assert(region->is_Region(), "sanity check"); 568 if (region->req() != 3) return NULL; 569 570 // Check for CFG diamond 571 Node *lp = region->in(1); 572 Node *rp = region->in(2); 573 if (!lp || !rp) return NULL; 574 Node *lp_c = lp->in(0); 575 if (lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If()) return NULL; 576 IfNode *iff = lp_c->as_If(); 577 578 // Check for ops pinned in an arm of the diamond. 579 // Can't remove the control flow in this case 580 if (lp->outcnt() > 1) return NULL; 581 if (rp->outcnt() > 1) return NULL; 582 583 IdealLoopTree* r_loop = get_loop(region); 584 assert(r_loop == get_loop(iff), "sanity"); 585 // Always convert to CMOVE if all results are used only outside this loop. 586 bool used_inside_loop = (r_loop == _ltree_root); 587 588 // Check profitability 589 int cost = 0; 590 int phis = 0; 591 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 592 Node *out = region->fast_out(i); 593 if (!out->is_Phi()) continue; // Ignore other control edges, etc 594 phis++; 595 PhiNode* phi = out->as_Phi(); 596 BasicType bt = phi->type()->basic_type(); 597 switch (bt) { 598 case T_DOUBLE: 599 case T_FLOAT: 600 if (C->use_cmove()) { 601 continue; //TODO: maybe we want to add some cost 602 } 603 cost += Matcher::float_cmove_cost(); // Could be very expensive 604 break; 605 case T_LONG: { 606 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's 607 } 608 case T_INT: // These all CMOV fine 609 case T_ADDRESS: { // (RawPtr) 610 cost++; 611 break; 612 } 613 case T_NARROWOOP: // Fall through 614 case T_OBJECT: { // Base oops are OK, but not derived oops 615 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr(); 616 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a 617 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus 618 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we 619 // have a Phi for the base here that we convert to a CMOVE all is well 620 // and good. But if the base is dead, we'll not make a CMOVE. Later 621 // the allocator will have to produce a base by creating a CMOVE of the 622 // relevant bases. This puts the allocator in the business of 623 // manufacturing expensive instructions, generally a bad plan. 624 // Just Say No to Conditionally-Moved Derived Pointers. 625 if (tp && tp->offset() != 0) 626 return NULL; 627 cost++; 628 break; 629 } 630 default: 631 return NULL; // In particular, can't do memory or I/O 632 } 633 // Add in cost any speculative ops 634 for (uint j = 1; j < region->req(); j++) { 635 Node *proj = region->in(j); 636 Node *inp = phi->in(j); 637 if (get_ctrl(inp) == proj) { // Found local op 638 cost++; 639 // Check for a chain of dependent ops; these will all become 640 // speculative in a CMOV. 641 for (uint k = 1; k < inp->req(); k++) 642 if (get_ctrl(inp->in(k)) == proj) 643 cost += ConditionalMoveLimit; // Too much speculative goo 644 } 645 } 646 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode. 647 // This will likely Split-If, a higher-payoff operation. 648 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) { 649 Node* use = phi->fast_out(k); 650 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr()) 651 cost += ConditionalMoveLimit; 652 // Is there a use inside the loop? 653 // Note: check only basic types since CMoveP is pinned. 654 if (!used_inside_loop && is_java_primitive(bt)) { 655 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use); 656 if (r_loop == u_loop || r_loop->is_member(u_loop)) { 657 used_inside_loop = true; 658 } 659 } 660 } 661 }//for 662 Node* bol = iff->in(1); 663 assert(bol->Opcode() == Op_Bool, ""); 664 int cmp_op = bol->in(1)->Opcode(); 665 // It is expensive to generate flags from a float compare. 666 // Avoid duplicated float compare. 667 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL; 668 669 float infrequent_prob = PROB_UNLIKELY_MAG(3); 670 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop. 671 if (used_inside_loop) { 672 if (cost >= ConditionalMoveLimit) return NULL; // Too much goo 673 674 // BlockLayoutByFrequency optimization moves infrequent branch 675 // from hot path. No point in CMOV'ing in such case (110 is used 676 // instead of 100 to take into account not exactness of float value). 677 if (BlockLayoutByFrequency) { 678 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f); 679 } 680 } 681 // Check for highly predictable branch. No point in CMOV'ing if 682 // we are going to predict accurately all the time. 683 if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) { 684 //keep going 685 } else if (iff->_prob < infrequent_prob || 686 iff->_prob > (1.0f - infrequent_prob)) 687 return NULL; 688 689 // -------------- 690 // Now replace all Phis with CMOV's 691 Node *cmov_ctrl = iff->in(0); 692 uint flip = (lp->Opcode() == Op_IfTrue); 693 Node_List wq; 694 while (1) { 695 PhiNode* phi = NULL; 696 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 697 Node *out = region->fast_out(i); 698 if (out->is_Phi()) { 699 phi = out->as_Phi(); 700 break; 701 } 702 } 703 if (phi == NULL) break; 704 if (PrintOpto && VerifyLoopOptimizations) { tty->print_cr("CMOV"); } 705 // Move speculative ops 706 wq.push(phi); 707 while (wq.size() > 0) { 708 Node *n = wq.pop(); 709 for (uint j = 1; j < n->req(); j++) { 710 Node* m = n->in(j); 711 if (m != NULL && !is_dominator(get_ctrl(m), cmov_ctrl)) { 712 #ifndef PRODUCT 713 if (PrintOpto && VerifyLoopOptimizations) { 714 tty->print(" speculate: "); 715 m->dump(); 716 } 717 #endif 718 set_ctrl(m, cmov_ctrl); 719 wq.push(m); 720 } 721 } 722 } 723 Node *cmov = CMoveNode::make(cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi)); 724 register_new_node( cmov, cmov_ctrl ); 725 _igvn.replace_node( phi, cmov ); 726 #ifndef PRODUCT 727 if (TraceLoopOpts) { 728 tty->print("CMOV "); 729 r_loop->dump_head(); 730 if (Verbose) { 731 bol->in(1)->dump(1); 732 cmov->dump(1); 733 } 734 } 735 if (VerifyLoopOptimizations) verify(); 736 #endif 737 } 738 739 // The useless CFG diamond will fold up later; see the optimization in 740 // RegionNode::Ideal. 741 _igvn._worklist.push(region); 742 743 return iff->in(1); 744 } 745 746 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) { 747 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 748 Node* u = m->fast_out(i); 749 if (u->is_CFG()) { 750 if (u->Opcode() == Op_NeverBranch) { 751 u = ((NeverBranchNode*)u)->proj_out(0); 752 enqueue_cfg_uses(u, wq); 753 } else { 754 wq.push(u); 755 } 756 } 757 } 758 } 759 760 // Try moving a store out of a loop, right before the loop 761 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) { 762 // Store has to be first in the loop body 763 IdealLoopTree *n_loop = get_loop(n_ctrl); 764 if (n->is_Store() && n_loop != _ltree_root && 765 n_loop->is_loop() && n_loop->_head->is_Loop() && 766 n->in(0) != NULL) { 767 Node* address = n->in(MemNode::Address); 768 Node* value = n->in(MemNode::ValueIn); 769 Node* mem = n->in(MemNode::Memory); 770 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 771 IdealLoopTree* value_loop = get_loop(get_ctrl(value)); 772 773 // - address and value must be loop invariant 774 // - memory must be a memory Phi for the loop 775 // - Store must be the only store on this memory slice in the 776 // loop: if there's another store following this one then value 777 // written at iteration i by the second store could be overwritten 778 // at iteration i+n by the first store: it's not safe to move the 779 // first store out of the loop 780 // - nothing must observe the memory Phi: it guarantees no read 781 // before the store, we are also guaranteed the store post 782 // dominates the loop head (ignoring a possible early 783 // exit). Otherwise there would be extra Phi involved between the 784 // loop's Phi and the store. 785 // - there must be no early exit from the loop before the Store 786 // (such an exit most of the time would be an extra use of the 787 // memory Phi but sometimes is a bottom memory Phi that takes the 788 // store as input). 789 790 if (!n_loop->is_member(address_loop) && 791 !n_loop->is_member(value_loop) && 792 mem->is_Phi() && mem->in(0) == n_loop->_head && 793 mem->outcnt() == 1 && 794 mem->in(LoopNode::LoopBackControl) == n) { 795 796 assert(n_loop->_tail != NULL, "need a tail"); 797 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop"); 798 799 // Verify that there's no early exit of the loop before the store. 800 bool ctrl_ok = false; 801 { 802 // Follow control from loop head until n, we exit the loop or 803 // we reach the tail 804 ResourceMark rm; 805 Unique_Node_List wq; 806 wq.push(n_loop->_head); 807 808 for (uint next = 0; next < wq.size(); ++next) { 809 Node *m = wq.at(next); 810 if (m == n->in(0)) { 811 ctrl_ok = true; 812 continue; 813 } 814 assert(!has_ctrl(m), "should be CFG"); 815 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) { 816 ctrl_ok = false; 817 break; 818 } 819 enqueue_cfg_uses(m, wq); 820 if (wq.size() > 10) { 821 ctrl_ok = false; 822 break; 823 } 824 } 825 } 826 if (ctrl_ok) { 827 // move the Store 828 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem); 829 _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl)); 830 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl)); 831 // Disconnect the phi now. An empty phi can confuse other 832 // optimizations in this pass of loop opts. 833 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl)); 834 n_loop->_body.yank(mem); 835 836 set_ctrl_and_loop(n, n->in(0)); 837 838 return n; 839 } 840 } 841 } 842 return NULL; 843 } 844 845 // Try moving a store out of a loop, right after the loop 846 void PhaseIdealLoop::try_move_store_after_loop(Node* n) { 847 if (n->is_Store() && n->in(0) != NULL) { 848 Node *n_ctrl = get_ctrl(n); 849 IdealLoopTree *n_loop = get_loop(n_ctrl); 850 // Store must be in a loop 851 if (n_loop != _ltree_root && !n_loop->_irreducible) { 852 Node* address = n->in(MemNode::Address); 853 Node* value = n->in(MemNode::ValueIn); 854 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 855 // address must be loop invariant 856 if (!n_loop->is_member(address_loop)) { 857 // Store must be last on this memory slice in the loop and 858 // nothing in the loop must observe it 859 Node* phi = NULL; 860 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 861 Node* u = n->fast_out(i); 862 if (has_ctrl(u)) { // control use? 863 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 864 if (!n_loop->is_member(u_loop)) { 865 continue; 866 } 867 if (u->is_Phi() && u->in(0) == n_loop->_head) { 868 assert(_igvn.type(u) == Type::MEMORY, "bad phi"); 869 // multiple phis on the same slice are possible 870 if (phi != NULL) { 871 return; 872 } 873 phi = u; 874 continue; 875 } 876 } 877 return; 878 } 879 if (phi != NULL) { 880 // Nothing in the loop before the store (next iteration) 881 // must observe the stored value 882 bool mem_ok = true; 883 { 884 ResourceMark rm; 885 Unique_Node_List wq; 886 wq.push(phi); 887 for (uint next = 0; next < wq.size() && mem_ok; ++next) { 888 Node *m = wq.at(next); 889 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) { 890 Node* u = m->fast_out(i); 891 if (u->is_Store() || u->is_Phi()) { 892 if (u != n) { 893 wq.push(u); 894 mem_ok = (wq.size() <= 10); 895 } 896 } else { 897 mem_ok = false; 898 break; 899 } 900 } 901 } 902 } 903 if (mem_ok) { 904 // Move the store out of the loop if the LCA of all 905 // users (except for the phi) is outside the loop. 906 Node* hook = new Node(1); 907 _igvn.rehash_node_delayed(phi); 908 int count = phi->replace_edge(n, hook); 909 assert(count > 0, "inconsistent phi"); 910 911 // Compute latest point this store can go 912 Node* lca = get_late_ctrl(n, get_ctrl(n)); 913 if (n_loop->is_member(get_loop(lca))) { 914 // LCA is in the loop - bail out 915 _igvn.replace_node(hook, n); 916 return; 917 } 918 #ifdef ASSERT 919 if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) { 920 assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined"); 921 n_loop->_head->as_Loop()->verify_strip_mined(1); 922 Node* outer = n_loop->_head->as_CountedLoop()->outer_loop(); 923 IdealLoopTree* outer_loop = get_loop(outer); 924 assert(n_loop->_parent == outer_loop, "broken loop tree"); 925 assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state"); 926 } 927 #endif 928 929 // Move store out of the loop 930 _igvn.replace_node(hook, n->in(MemNode::Memory)); 931 _igvn.replace_input_of(n, 0, lca); 932 set_ctrl_and_loop(n, lca); 933 934 // Disconnect the phi now. An empty phi can confuse other 935 // optimizations in this pass of loop opts.. 936 if (phi->in(LoopNode::LoopBackControl) == phi) { 937 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl)); 938 n_loop->_body.yank(phi); 939 } 940 } 941 } 942 } 943 } 944 } 945 } 946 947 //------------------------------split_if_with_blocks_pre----------------------- 948 // Do the real work in a non-recursive function. Data nodes want to be 949 // cloned in the pre-order so they can feed each other nicely. 950 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) { 951 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 952 Node* bs_res = bs->split_if_pre(this, n); 953 if (bs_res != NULL) { 954 return bs_res; 955 } 956 // Cloning these guys is unlikely to win 957 int n_op = n->Opcode(); 958 if( n_op == Op_MergeMem ) return n; 959 if( n->is_Proj() ) return n; 960 // Do not clone-up CmpFXXX variations, as these are always 961 // followed by a CmpI 962 if( n->is_Cmp() ) return n; 963 // Attempt to use a conditional move instead of a phi/branch 964 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) { 965 Node *cmov = conditional_move( n ); 966 if( cmov ) return cmov; 967 } 968 if( n->is_CFG() || n->is_LoadStore() ) 969 return n; 970 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd 971 n_op == Op_Opaque2 ) { 972 if( !C->major_progress() ) // If chance of no more loop opts... 973 _igvn._worklist.push(n); // maybe we'll remove them 974 return n; 975 } 976 977 if( n->is_Con() ) return n; // No cloning for Con nodes 978 979 Node *n_ctrl = get_ctrl(n); 980 if( !n_ctrl ) return n; // Dead node 981 982 Node* res = try_move_store_before_loop(n, n_ctrl); 983 if (res != NULL) { 984 return n; 985 } 986 987 // Attempt to remix address expressions for loop invariants 988 Node *m = remix_address_expressions( n ); 989 if( m ) return m; 990 991 if (n_op == Op_AddI) { 992 Node *nn = convert_add_to_muladd( n ); 993 if ( nn ) return nn; 994 } 995 996 if (n->is_ConstraintCast()) { 997 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this); 998 // ConstraintCastNode::dominating_cast() uses node control input to determine domination. 999 // Node control inputs don't necessarily agree with loop control info (due to 1000 // transformations happened in between), thus additional dominance check is needed 1001 // to keep loop info valid. 1002 if (dom_cast != NULL && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) { 1003 _igvn.replace_node(n, dom_cast); 1004 return dom_cast; 1005 } 1006 } 1007 1008 // Determine if the Node has inputs from some local Phi. 1009 // Returns the block to clone thru. 1010 Node *n_blk = has_local_phi_input( n ); 1011 if( !n_blk ) return n; 1012 1013 // Do not clone the trip counter through on a CountedLoop 1014 // (messes up the canonical shape). 1015 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n; 1016 1017 // Check for having no control input; not pinned. Allow 1018 // dominating control. 1019 if (n->in(0)) { 1020 Node *dom = idom(n_blk); 1021 if (dom_lca(n->in(0), dom) != n->in(0)) { 1022 return n; 1023 } 1024 } 1025 // Policy: when is it profitable. You must get more wins than 1026 // policy before it is considered profitable. Policy is usually 0, 1027 // so 1 win is considered profitable. Big merges will require big 1028 // cloning, so get a larger policy. 1029 int policy = n_blk->req() >> 2; 1030 1031 // If the loop is a candidate for range check elimination, 1032 // delay splitting through it's phi until a later loop optimization 1033 if (n_blk->is_CountedLoop()) { 1034 IdealLoopTree *lp = get_loop(n_blk); 1035 if (lp && lp->_rce_candidate) { 1036 return n; 1037 } 1038 } 1039 1040 if (must_throttle_split_if()) return n; 1041 1042 // Split 'n' through the merge point if it is profitable 1043 Node *phi = split_thru_phi( n, n_blk, policy ); 1044 if (!phi) return n; 1045 1046 // Found a Phi to split thru! 1047 // Replace 'n' with the new phi 1048 _igvn.replace_node( n, phi ); 1049 // Moved a load around the loop, 'en-registering' something. 1050 if (n_blk->is_Loop() && n->is_Load() && 1051 !phi->in(LoopNode::LoopBackControl)->is_Load()) 1052 C->set_major_progress(); 1053 1054 return phi; 1055 } 1056 1057 static bool merge_point_too_heavy(Compile* C, Node* region) { 1058 // Bail out if the region and its phis have too many users. 1059 int weight = 0; 1060 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1061 weight += region->fast_out(i)->outcnt(); 1062 } 1063 int nodes_left = C->max_node_limit() - C->live_nodes(); 1064 if (weight * 8 > nodes_left) { 1065 if (PrintOpto) { 1066 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight); 1067 } 1068 return true; 1069 } else { 1070 return false; 1071 } 1072 } 1073 1074 static bool merge_point_safe(Node* region) { 1075 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode 1076 // having a PhiNode input. This sidesteps the dangerous case where the split 1077 // ConvI2LNode may become TOP if the input Value() does not 1078 // overlap the ConvI2L range, leaving a node which may not dominate its 1079 // uses. 1080 // A better fix for this problem can be found in the BugTraq entry, but 1081 // expediency for Mantis demands this hack. 1082 // 6855164: If the merge point has a FastLockNode with a PhiNode input, we stop 1083 // split_if_with_blocks from splitting a block because we could not move around 1084 // the FastLockNode. 1085 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1086 Node* n = region->fast_out(i); 1087 if (n->is_Phi()) { 1088 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1089 Node* m = n->fast_out(j); 1090 if (m->is_FastLock()) 1091 return false; 1092 #ifdef _LP64 1093 if (m->Opcode() == Op_ConvI2L) 1094 return false; 1095 if (m->is_CastII() && m->isa_CastII()->has_range_check()) { 1096 return false; 1097 } 1098 #endif 1099 } 1100 } 1101 } 1102 return true; 1103 } 1104 1105 1106 //------------------------------place_near_use--------------------------------- 1107 // Place some computation next to use but not inside inner loops. 1108 // For inner loop uses move it to the preheader area. 1109 Node *PhaseIdealLoop::place_near_use(Node *useblock) const { 1110 IdealLoopTree *u_loop = get_loop( useblock ); 1111 if (u_loop->_irreducible) { 1112 return useblock; 1113 } 1114 if (u_loop->_child) { 1115 if (useblock == u_loop->_head && u_loop->_head->is_OuterStripMinedLoop()) { 1116 return u_loop->_head->in(LoopNode::EntryControl); 1117 } 1118 return useblock; 1119 } 1120 return u_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl); 1121 } 1122 1123 1124 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) { 1125 if (!n->is_If() || n->is_CountedLoopEnd()) { 1126 return false; 1127 } 1128 if (!n->in(0)->is_Region()) { 1129 return false; 1130 } 1131 Node* region = n->in(0); 1132 Node* dom = idom(region); 1133 if (!dom->is_If() || dom->in(1) != n->in(1)) { 1134 return false; 1135 } 1136 IfNode* dom_if = dom->as_If(); 1137 Node* proj_true = dom_if->proj_out(1); 1138 Node* proj_false = dom_if->proj_out(0); 1139 1140 for (uint i = 1; i < region->req(); i++) { 1141 if (is_dominator(proj_true, region->in(i))) { 1142 continue; 1143 } 1144 if (is_dominator(proj_false, region->in(i))) { 1145 continue; 1146 } 1147 return false; 1148 } 1149 1150 return true; 1151 } 1152 1153 1154 bool PhaseIdealLoop::can_split_if(Node* n_ctrl) { 1155 if (must_throttle_split_if()) { 1156 return false; 1157 } 1158 1159 // Do not do 'split-if' if irreducible loops are present. 1160 if (_has_irreducible_loops) { 1161 return false; 1162 } 1163 1164 if (merge_point_too_heavy(C, n_ctrl)) { 1165 return false; 1166 } 1167 1168 // Do not do 'split-if' if some paths are dead. First do dead code 1169 // elimination and then see if its still profitable. 1170 for (uint i = 1; i < n_ctrl->req(); i++) { 1171 if (n_ctrl->in(i) == C->top()) { 1172 return false; 1173 } 1174 } 1175 1176 // If trying to do a 'Split-If' at the loop head, it is only 1177 // profitable if the cmp folds up on BOTH paths. Otherwise we 1178 // risk peeling a loop forever. 1179 1180 // CNC - Disabled for now. Requires careful handling of loop 1181 // body selection for the cloned code. Also, make sure we check 1182 // for any input path not being in the same loop as n_ctrl. For 1183 // irreducible loops we cannot check for 'n_ctrl->is_Loop()' 1184 // because the alternative loop entry points won't be converted 1185 // into LoopNodes. 1186 IdealLoopTree *n_loop = get_loop(n_ctrl); 1187 for (uint j = 1; j < n_ctrl->req(); j++) { 1188 if (get_loop(n_ctrl->in(j)) != n_loop) { 1189 return false; 1190 } 1191 } 1192 1193 // Check for safety of the merge point. 1194 if (!merge_point_safe(n_ctrl)) { 1195 return false; 1196 } 1197 1198 return true; 1199 } 1200 1201 // Detect if the node is the inner strip-mined loop 1202 // Return: NULL if it's not the case, or the exit of outer strip-mined loop 1203 static Node* is_inner_of_stripmined_loop(const Node* out) { 1204 Node* out_le = NULL; 1205 1206 if (out->is_CountedLoopEnd()) { 1207 const CountedLoopNode* loop = out->as_CountedLoopEnd()->loopnode(); 1208 1209 if (loop != NULL && loop->is_strip_mined()) { 1210 out_le = loop->in(LoopNode::EntryControl)->as_OuterStripMinedLoop()->outer_loop_exit(); 1211 } 1212 } 1213 1214 return out_le; 1215 } 1216 1217 //------------------------------split_if_with_blocks_post---------------------- 1218 // Do the real work in a non-recursive function. CFG hackery wants to be 1219 // in the post-order, so it can dirty the I-DOM info and not use the dirtied 1220 // info. 1221 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) { 1222 1223 // Cloning Cmp through Phi's involves the split-if transform. 1224 // FastLock is not used by an If 1225 if (n->is_Cmp() && !n->is_FastLock()) { 1226 Node *n_ctrl = get_ctrl(n); 1227 // Determine if the Node has inputs from some local Phi. 1228 // Returns the block to clone thru. 1229 Node *n_blk = has_local_phi_input(n); 1230 if (n_blk != n_ctrl) { 1231 return; 1232 } 1233 1234 if (!can_split_if(n_ctrl)) { 1235 return; 1236 } 1237 1238 if (n->outcnt() != 1) { 1239 return; // Multiple bool's from 1 compare? 1240 } 1241 Node *bol = n->unique_out(); 1242 assert(bol->is_Bool(), "expect a bool here"); 1243 if (bol->outcnt() != 1) { 1244 return;// Multiple branches from 1 compare? 1245 } 1246 Node *iff = bol->unique_out(); 1247 1248 // Check some safety conditions 1249 if (iff->is_If()) { // Classic split-if? 1250 if (iff->in(0) != n_ctrl) { 1251 return; // Compare must be in same blk as if 1252 } 1253 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE 1254 // Can't split CMove with different control edge. 1255 if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) { 1256 return; 1257 } 1258 if (get_ctrl(iff->in(2)) == n_ctrl || 1259 get_ctrl(iff->in(3)) == n_ctrl) { 1260 return; // Inputs not yet split-up 1261 } 1262 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) { 1263 return; // Loop-invar test gates loop-varying CMOVE 1264 } 1265 } else { 1266 return; // some other kind of node, such as an Allocate 1267 } 1268 1269 // When is split-if profitable? Every 'win' on means some control flow 1270 // goes dead, so it's almost always a win. 1271 int policy = 0; 1272 // Split compare 'n' through the merge point if it is profitable 1273 Node *phi = split_thru_phi( n, n_ctrl, policy); 1274 if (!phi) { 1275 return; 1276 } 1277 1278 // Found a Phi to split thru! 1279 // Replace 'n' with the new phi 1280 _igvn.replace_node(n, phi); 1281 1282 // Now split the bool up thru the phi 1283 Node *bolphi = split_thru_phi(bol, n_ctrl, -1); 1284 guarantee(bolphi != NULL, "null boolean phi node"); 1285 1286 _igvn.replace_node(bol, bolphi); 1287 assert(iff->in(1) == bolphi, ""); 1288 1289 if (bolphi->Value(&_igvn)->singleton()) { 1290 return; 1291 } 1292 1293 // Conditional-move? Must split up now 1294 if (!iff->is_If()) { 1295 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1); 1296 _igvn.replace_node(iff, cmovphi); 1297 return; 1298 } 1299 1300 // Now split the IF 1301 do_split_if(iff); 1302 return; 1303 } 1304 1305 // Two identical ifs back to back can be merged 1306 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) { 1307 Node *n_ctrl = n->in(0); 1308 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1)); 1309 IfNode* dom_if = idom(n_ctrl)->as_If(); 1310 Node* proj_true = dom_if->proj_out(1); 1311 Node* proj_false = dom_if->proj_out(0); 1312 Node* con_true = _igvn.makecon(TypeInt::ONE); 1313 Node* con_false = _igvn.makecon(TypeInt::ZERO); 1314 1315 for (uint i = 1; i < n_ctrl->req(); i++) { 1316 if (is_dominator(proj_true, n_ctrl->in(i))) { 1317 bolphi->init_req(i, con_true); 1318 } else { 1319 assert(is_dominator(proj_false, n_ctrl->in(i)), "bad if"); 1320 bolphi->init_req(i, con_false); 1321 } 1322 } 1323 register_new_node(bolphi, n_ctrl); 1324 _igvn.replace_input_of(n, 1, bolphi); 1325 1326 // Now split the IF 1327 do_split_if(n); 1328 return; 1329 } 1330 1331 // Check for an IF ready to split; one that has its 1332 // condition codes input coming from a Phi at the block start. 1333 int n_op = n->Opcode(); 1334 1335 // Check for an IF being dominated by another IF same test 1336 if (n_op == Op_If || 1337 n_op == Op_RangeCheck) { 1338 Node *bol = n->in(1); 1339 uint max = bol->outcnt(); 1340 // Check for same test used more than once? 1341 if (max > 1 && bol->is_Bool()) { 1342 // Search up IDOMs to see if this IF is dominated. 1343 Node *cutoff = get_ctrl(bol); 1344 1345 // Now search up IDOMs till cutoff, looking for a dominating test 1346 Node *prevdom = n; 1347 Node *dom = idom(prevdom); 1348 while (dom != cutoff) { 1349 if (dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom) { 1350 // It's invalid to move control dependent data nodes in the inner 1351 // strip-mined loop, because: 1352 // 1) break validation of LoopNode::verify_strip_mined() 1353 // 2) move code with side-effect in strip-mined loop 1354 // Move to the exit of outer strip-mined loop in that case. 1355 Node* out_le = is_inner_of_stripmined_loop(dom); 1356 if (out_le != NULL) { 1357 prevdom = out_le; 1358 } 1359 // Replace the dominated test with an obvious true or false. 1360 // Place it on the IGVN worklist for later cleanup. 1361 C->set_major_progress(); 1362 dominated_by(prevdom, n, false, true); 1363 #ifndef PRODUCT 1364 if( VerifyLoopOptimizations ) verify(); 1365 #endif 1366 return; 1367 } 1368 prevdom = dom; 1369 dom = idom(prevdom); 1370 } 1371 } 1372 } 1373 1374 // See if a shared loop-varying computation has no loop-varying uses. 1375 // Happens if something is only used for JVM state in uncommon trap exits, 1376 // like various versions of induction variable+offset. Clone the 1377 // computation per usage to allow it to sink out of the loop. 1378 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about) 1379 Node *n_ctrl = get_ctrl(n); 1380 IdealLoopTree *n_loop = get_loop(n_ctrl); 1381 if( n_loop != _ltree_root ) { 1382 DUIterator_Fast imax, i = n->fast_outs(imax); 1383 for (; i < imax; i++) { 1384 Node* u = n->fast_out(i); 1385 if( !has_ctrl(u) ) break; // Found control user 1386 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 1387 if( u_loop == n_loop ) break; // Found loop-varying use 1388 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop 1389 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003 1390 } 1391 bool did_break = (i < imax); // Did we break out of the previous loop? 1392 if (!did_break && n->outcnt() > 1) { // All uses in outer loops! 1393 Node *late_load_ctrl = NULL; 1394 if (n->is_Load()) { 1395 // If n is a load, get and save the result from get_late_ctrl(), 1396 // to be later used in calculating the control for n's clones. 1397 clear_dom_lca_tags(); 1398 late_load_ctrl = get_late_ctrl(n, n_ctrl); 1399 } 1400 // If n is a load, and the late control is the same as the current 1401 // control, then the cloning of n is a pointless exercise, because 1402 // GVN will ensure that we end up where we started. 1403 if (!n->is_Load() || late_load_ctrl != n_ctrl) { 1404 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1405 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) { 1406 Node *u = n->last_out(j); // Clone private computation per use 1407 _igvn.rehash_node_delayed(u); 1408 Node *x = n->clone(); // Clone computation 1409 Node *x_ctrl = NULL; 1410 if( u->is_Phi() ) { 1411 // Replace all uses of normal nodes. Replace Phi uses 1412 // individually, so the separate Nodes can sink down 1413 // different paths. 1414 uint k = 1; 1415 while( u->in(k) != n ) k++; 1416 u->set_req( k, x ); 1417 // x goes next to Phi input path 1418 x_ctrl = u->in(0)->in(k); 1419 --j; 1420 } else { // Normal use 1421 // Replace all uses 1422 for( uint k = 0; k < u->req(); k++ ) { 1423 if( u->in(k) == n ) { 1424 u->set_req( k, x ); 1425 --j; 1426 } 1427 } 1428 x_ctrl = get_ctrl(u); 1429 } 1430 1431 // Find control for 'x' next to use but not inside inner loops. 1432 // For inner loop uses get the preheader area. 1433 x_ctrl = place_near_use(x_ctrl); 1434 1435 if (bs->sink_node(this, n, x, x_ctrl, n_ctrl)) { 1436 continue; 1437 } 1438 1439 if (n->is_Load()) { 1440 // For loads, add a control edge to a CFG node outside of the loop 1441 // to force them to not combine and return back inside the loop 1442 // during GVN optimization (4641526). 1443 // 1444 // Because we are setting the actual control input, factor in 1445 // the result from get_late_ctrl() so we respect any 1446 // anti-dependences. (6233005). 1447 x_ctrl = dom_lca(late_load_ctrl, x_ctrl); 1448 1449 // Don't allow the control input to be a CFG splitting node. 1450 // Such nodes should only have ProjNodes as outs, e.g. IfNode 1451 // should only have IfTrueNode and IfFalseNode (4985384). 1452 x_ctrl = find_non_split_ctrl(x_ctrl); 1453 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone"); 1454 1455 x->set_req(0, x_ctrl); 1456 } 1457 register_new_node(x, x_ctrl); 1458 1459 // Some institutional knowledge is needed here: 'x' is 1460 // yanked because if the optimizer runs GVN on it all the 1461 // cloned x's will common up and undo this optimization and 1462 // be forced back in the loop. 1463 // I tried setting control edges on the x's to force them to 1464 // not combine, but the matching gets worried when it tries 1465 // to fold a StoreP and an AddP together (as part of an 1466 // address expression) and the AddP and StoreP have 1467 // different controls. 1468 if (!x->is_Load() && !x->is_DecodeNarrowPtr()) _igvn._worklist.yank(x); 1469 } 1470 _igvn.remove_dead_node(n); 1471 } 1472 } 1473 } 1474 } 1475 1476 try_move_store_after_loop(n); 1477 1478 // Remove multiple allocations of the same value type 1479 if (n->is_ValueType()) { 1480 n->as_ValueType()->remove_redundant_allocations(&_igvn, this); 1481 return; // n is now dead 1482 } 1483 1484 // Check for Opaque2's who's loop has disappeared - who's input is in the 1485 // same loop nest as their output. Remove 'em, they are no longer useful. 1486 if( n_op == Op_Opaque2 && 1487 n->in(1) != NULL && 1488 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) { 1489 _igvn.replace_node( n, n->in(1) ); 1490 } 1491 } 1492 1493 //------------------------------split_if_with_blocks--------------------------- 1494 // Check for aggressive application of 'split-if' optimization, 1495 // using basic block level info. 1496 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack) { 1497 Node* root = C->root(); 1498 visited.set(root->_idx); // first, mark root as visited 1499 // Do pre-visit work for root 1500 Node* n = split_if_with_blocks_pre(root); 1501 uint cnt = n->outcnt(); 1502 uint i = 0; 1503 1504 while (true) { 1505 // Visit all children 1506 if (i < cnt) { 1507 Node* use = n->raw_out(i); 1508 ++i; 1509 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) { 1510 // Now do pre-visit work for this use 1511 use = split_if_with_blocks_pre(use); 1512 nstack.push(n, i); // Save parent and next use's index. 1513 n = use; // Process all children of current use. 1514 cnt = use->outcnt(); 1515 i = 0; 1516 } 1517 } 1518 else { 1519 // All of n's children have been processed, complete post-processing. 1520 if (cnt != 0 && !n->is_Con()) { 1521 assert(has_node(n), "no dead nodes"); 1522 split_if_with_blocks_post(n); 1523 } 1524 if (must_throttle_split_if()) { 1525 nstack.clear(); 1526 } 1527 if (nstack.is_empty()) { 1528 // Finished all nodes on stack. 1529 break; 1530 } 1531 // Get saved parent node and next use's index. Visit the rest of uses. 1532 n = nstack.node(); 1533 cnt = n->outcnt(); 1534 i = nstack.index(); 1535 nstack.pop(); 1536 } 1537 } 1538 } 1539 1540 1541 //============================================================================= 1542 // 1543 // C L O N E A L O O P B O D Y 1544 // 1545 1546 //------------------------------clone_iff-------------------------------------- 1547 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1548 // "Nearly" because all Nodes have been cloned from the original in the loop, 1549 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1550 // through the Phi recursively, and return a Bool. 1551 Node* PhaseIdealLoop::clone_iff(PhiNode *phi, IdealLoopTree *loop) { 1552 1553 // Convert this Phi into a Phi merging Bools 1554 uint i; 1555 for (i = 1; i < phi->req(); i++) { 1556 Node *b = phi->in(i); 1557 if (b->is_Phi()) { 1558 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi(), loop)); 1559 } else { 1560 assert(b->is_Bool() || b->Opcode() == Op_Opaque4, ""); 1561 } 1562 } 1563 1564 Node* n = phi->in(1); 1565 Node* sample_opaque = NULL; 1566 Node *sample_bool = NULL; 1567 if (n->Opcode() == Op_Opaque4) { 1568 sample_opaque = n; 1569 sample_bool = n->in(1); 1570 assert(sample_bool->is_Bool(), "wrong type"); 1571 } else { 1572 sample_bool = n; 1573 } 1574 Node *sample_cmp = sample_bool->in(1); 1575 1576 // Make Phis to merge the Cmp's inputs. 1577 PhiNode *phi1 = new PhiNode(phi->in(0), Type::TOP); 1578 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP); 1579 for (i = 1; i < phi->req(); i++) { 1580 Node *n1 = sample_opaque == NULL ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1); 1581 Node *n2 = sample_opaque == NULL ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2); 1582 phi1->set_req(i, n1); 1583 phi2->set_req(i, n2); 1584 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 1585 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 1586 } 1587 // See if these Phis have been made before. 1588 // Register with optimizer 1589 Node *hit1 = _igvn.hash_find_insert(phi1); 1590 if (hit1) { // Hit, toss just made Phi 1591 _igvn.remove_dead_node(phi1); // Remove new phi 1592 assert(hit1->is_Phi(), "" ); 1593 phi1 = (PhiNode*)hit1; // Use existing phi 1594 } else { // Miss 1595 _igvn.register_new_node_with_optimizer(phi1); 1596 } 1597 Node *hit2 = _igvn.hash_find_insert(phi2); 1598 if (hit2) { // Hit, toss just made Phi 1599 _igvn.remove_dead_node(phi2); // Remove new phi 1600 assert(hit2->is_Phi(), "" ); 1601 phi2 = (PhiNode*)hit2; // Use existing phi 1602 } else { // Miss 1603 _igvn.register_new_node_with_optimizer(phi2); 1604 } 1605 // Register Phis with loop/block info 1606 set_ctrl(phi1, phi->in(0)); 1607 set_ctrl(phi2, phi->in(0)); 1608 // Make a new Cmp 1609 Node *cmp = sample_cmp->clone(); 1610 cmp->set_req(1, phi1); 1611 cmp->set_req(2, phi2); 1612 _igvn.register_new_node_with_optimizer(cmp); 1613 set_ctrl(cmp, phi->in(0)); 1614 1615 // Make a new Bool 1616 Node *b = sample_bool->clone(); 1617 b->set_req(1,cmp); 1618 _igvn.register_new_node_with_optimizer(b); 1619 set_ctrl(b, phi->in(0)); 1620 1621 if (sample_opaque != NULL) { 1622 Node* opaque = sample_opaque->clone(); 1623 opaque->set_req(1, b); 1624 _igvn.register_new_node_with_optimizer(opaque); 1625 set_ctrl(opaque, phi->in(0)); 1626 return opaque; 1627 } 1628 1629 assert(b->is_Bool(), ""); 1630 return b; 1631 } 1632 1633 //------------------------------clone_bool------------------------------------- 1634 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1635 // "Nearly" because all Nodes have been cloned from the original in the loop, 1636 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1637 // through the Phi recursively, and return a Bool. 1638 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) { 1639 uint i; 1640 // Convert this Phi into a Phi merging Bools 1641 for( i = 1; i < phi->req(); i++ ) { 1642 Node *b = phi->in(i); 1643 if( b->is_Phi() ) { 1644 _igvn.replace_input_of(phi, i, clone_bool( b->as_Phi(), loop )); 1645 } else { 1646 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" ); 1647 } 1648 } 1649 1650 Node *sample_cmp = phi->in(1); 1651 1652 // Make Phis to merge the Cmp's inputs. 1653 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP ); 1654 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP ); 1655 for( uint j = 1; j < phi->req(); j++ ) { 1656 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP 1657 Node *n1, *n2; 1658 if( cmp_top->is_Cmp() ) { 1659 n1 = cmp_top->in(1); 1660 n2 = cmp_top->in(2); 1661 } else { 1662 n1 = n2 = cmp_top; 1663 } 1664 phi1->set_req( j, n1 ); 1665 phi2->set_req( j, n2 ); 1666 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 1667 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 1668 } 1669 1670 // See if these Phis have been made before. 1671 // Register with optimizer 1672 Node *hit1 = _igvn.hash_find_insert(phi1); 1673 if( hit1 ) { // Hit, toss just made Phi 1674 _igvn.remove_dead_node(phi1); // Remove new phi 1675 assert( hit1->is_Phi(), "" ); 1676 phi1 = (PhiNode*)hit1; // Use existing phi 1677 } else { // Miss 1678 _igvn.register_new_node_with_optimizer(phi1); 1679 } 1680 Node *hit2 = _igvn.hash_find_insert(phi2); 1681 if( hit2 ) { // Hit, toss just made Phi 1682 _igvn.remove_dead_node(phi2); // Remove new phi 1683 assert( hit2->is_Phi(), "" ); 1684 phi2 = (PhiNode*)hit2; // Use existing phi 1685 } else { // Miss 1686 _igvn.register_new_node_with_optimizer(phi2); 1687 } 1688 // Register Phis with loop/block info 1689 set_ctrl(phi1, phi->in(0)); 1690 set_ctrl(phi2, phi->in(0)); 1691 // Make a new Cmp 1692 Node *cmp = sample_cmp->clone(); 1693 cmp->set_req( 1, phi1 ); 1694 cmp->set_req( 2, phi2 ); 1695 _igvn.register_new_node_with_optimizer(cmp); 1696 set_ctrl(cmp, phi->in(0)); 1697 1698 assert( cmp->is_Cmp(), "" ); 1699 return (CmpNode*)cmp; 1700 } 1701 1702 //------------------------------sink_use--------------------------------------- 1703 // If 'use' was in the loop-exit block, it now needs to be sunk 1704 // below the post-loop merge point. 1705 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) { 1706 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) { 1707 set_ctrl(use, post_loop); 1708 for (DUIterator j = use->outs(); use->has_out(j); j++) 1709 sink_use(use->out(j), post_loop); 1710 } 1711 } 1712 1713 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new, 1714 IdealLoopTree* loop, IdealLoopTree* outer_loop, 1715 Node_List*& split_if_set, Node_List*& split_bool_set, 1716 Node_List*& split_cex_set, Node_List& worklist, 1717 uint new_counter, CloneLoopMode mode) { 1718 Node* nnn = old_new[old->_idx]; 1719 // Copy uses to a worklist, so I can munge the def-use info 1720 // with impunity. 1721 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1722 worklist.push(old->fast_out(j)); 1723 1724 while( worklist.size() ) { 1725 Node *use = worklist.pop(); 1726 if (!has_node(use)) continue; // Ignore dead nodes 1727 if (use->in(0) == C->top()) continue; 1728 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1729 // Check for data-use outside of loop - at least one of OLD or USE 1730 // must not be a CFG node. 1731 #ifdef ASSERT 1732 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == NULL) { 1733 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint(); 1734 assert(mode == ControlAroundStripMined && use == sfpt, "missed a node"); 1735 } 1736 #endif 1737 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) { 1738 1739 // If the Data use is an IF, that means we have an IF outside of the 1740 // loop that is switching on a condition that is set inside of the 1741 // loop. Happens if people set a loop-exit flag; then test the flag 1742 // in the loop to break the loop, then test is again outside of the 1743 // loop to determine which way the loop exited. 1744 // Loop predicate If node connects to Bool node through Opaque1 node. 1745 if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use) || use->Opcode() == Op_Opaque4) { 1746 // Since this code is highly unlikely, we lazily build the worklist 1747 // of such Nodes to go split. 1748 if (!split_if_set) { 1749 ResourceArea *area = Thread::current()->resource_area(); 1750 split_if_set = new Node_List(area); 1751 } 1752 split_if_set->push(use); 1753 } 1754 if (use->is_Bool()) { 1755 if (!split_bool_set) { 1756 ResourceArea *area = Thread::current()->resource_area(); 1757 split_bool_set = new Node_List(area); 1758 } 1759 split_bool_set->push(use); 1760 } 1761 if (use->Opcode() == Op_CreateEx) { 1762 if (!split_cex_set) { 1763 ResourceArea *area = Thread::current()->resource_area(); 1764 split_cex_set = new Node_List(area); 1765 } 1766 split_cex_set->push(use); 1767 } 1768 1769 1770 // Get "block" use is in 1771 uint idx = 0; 1772 while( use->in(idx) != old ) idx++; 1773 Node *prev = use->is_CFG() ? use : get_ctrl(use); 1774 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" ); 1775 Node *cfg = prev->_idx >= new_counter 1776 ? prev->in(2) 1777 : idom(prev); 1778 if( use->is_Phi() ) // Phi use is in prior block 1779 cfg = prev->in(idx); // NOT in block of Phi itself 1780 if (cfg->is_top()) { // Use is dead? 1781 _igvn.replace_input_of(use, idx, C->top()); 1782 continue; 1783 } 1784 1785 // If use is referenced through control edge... (idx == 0) 1786 if (mode == IgnoreStripMined && idx == 0) { 1787 LoopNode *head = loop->_head->as_Loop(); 1788 if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) { 1789 // That node is outside the inner loop, leave it outside the 1790 // outer loop as well to not confuse verification code. 1791 assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop"); 1792 _igvn.replace_input_of(use, 0, head->outer_loop_exit()); 1793 continue; 1794 } 1795 } 1796 1797 while(!outer_loop->is_member(get_loop(cfg))) { 1798 prev = cfg; 1799 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg); 1800 } 1801 // If the use occurs after merging several exits from the loop, then 1802 // old value must have dominated all those exits. Since the same old 1803 // value was used on all those exits we did not need a Phi at this 1804 // merge point. NOW we do need a Phi here. Each loop exit value 1805 // is now merged with the peeled body exit; each exit gets its own 1806 // private Phi and those Phis need to be merged here. 1807 Node *phi; 1808 if( prev->is_Region() ) { 1809 if( idx == 0 ) { // Updating control edge? 1810 phi = prev; // Just use existing control 1811 } else { // Else need a new Phi 1812 phi = PhiNode::make( prev, old ); 1813 // Now recursively fix up the new uses of old! 1814 for( uint i = 1; i < prev->req(); i++ ) { 1815 worklist.push(phi); // Onto worklist once for each 'old' input 1816 } 1817 } 1818 } else { 1819 // Get new RegionNode merging old and new loop exits 1820 prev = old_new[prev->_idx]; 1821 assert( prev, "just made this in step 7" ); 1822 if( idx == 0) { // Updating control edge? 1823 phi = prev; // Just use existing control 1824 } else { // Else need a new Phi 1825 // Make a new Phi merging data values properly 1826 phi = PhiNode::make( prev, old ); 1827 phi->set_req( 1, nnn ); 1828 } 1829 } 1830 // If inserting a new Phi, check for prior hits 1831 if( idx != 0 ) { 1832 Node *hit = _igvn.hash_find_insert(phi); 1833 if( hit == NULL ) { 1834 _igvn.register_new_node_with_optimizer(phi); // Register new phi 1835 } else { // or 1836 // Remove the new phi from the graph and use the hit 1837 _igvn.remove_dead_node(phi); 1838 phi = hit; // Use existing phi 1839 } 1840 set_ctrl(phi, prev); 1841 } 1842 // Make 'use' use the Phi instead of the old loop body exit value 1843 _igvn.replace_input_of(use, idx, phi); 1844 if( use->_idx >= new_counter ) { // If updating new phis 1845 // Not needed for correctness, but prevents a weak assert 1846 // in AddPNode from tripping (when we end up with different 1847 // base & derived Phis that will become the same after 1848 // IGVN does CSE). 1849 Node *hit = _igvn.hash_find_insert(use); 1850 if( hit ) // Go ahead and re-hash for hits. 1851 _igvn.replace_node( use, hit ); 1852 } 1853 1854 // If 'use' was in the loop-exit block, it now needs to be sunk 1855 // below the post-loop merge point. 1856 sink_use( use, prev ); 1857 } 1858 } 1859 } 1860 1861 static void clone_outer_loop_helper(Node* n, const IdealLoopTree *loop, const IdealLoopTree* outer_loop, 1862 const Node_List &old_new, Unique_Node_List& wq, PhaseIdealLoop* phase, 1863 bool check_old_new) { 1864 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1865 Node* u = n->fast_out(j); 1866 assert(check_old_new || old_new[u->_idx] == NULL, "shouldn't have been cloned"); 1867 if (!u->is_CFG() && (!check_old_new || old_new[u->_idx] == NULL)) { 1868 Node* c = phase->get_ctrl(u); 1869 IdealLoopTree* u_loop = phase->get_loop(c); 1870 assert(!loop->is_member(u_loop), "can be in outer loop or out of both loops only"); 1871 if (outer_loop->is_member(u_loop)) { 1872 wq.push(u); 1873 } 1874 } 1875 } 1876 } 1877 1878 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop, 1879 IdealLoopTree* outer_loop, int dd, Node_List &old_new, 1880 Node_List& extra_data_nodes) { 1881 if (head->is_strip_mined() && mode != IgnoreStripMined) { 1882 CountedLoopNode* cl = head->as_CountedLoop(); 1883 Node* l = cl->outer_loop(); 1884 Node* tail = cl->outer_loop_tail(); 1885 IfNode* le = cl->outer_loop_end(); 1886 Node* sfpt = cl->outer_safepoint(); 1887 CountedLoopEndNode* cle = cl->loopexit(); 1888 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop(); 1889 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null(); 1890 Node* cle_out = cle->proj_out(false); 1891 1892 Node* new_sfpt = NULL; 1893 Node* new_cle_out = cle_out->clone(); 1894 old_new.map(cle_out->_idx, new_cle_out); 1895 if (mode == CloneIncludesStripMined) { 1896 // clone outer loop body 1897 Node* new_l = l->clone(); 1898 Node* new_tail = tail->clone(); 1899 IfNode* new_le = le->clone()->as_If(); 1900 new_sfpt = sfpt->clone(); 1901 1902 set_loop(new_l, outer_loop->_parent); 1903 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd); 1904 set_loop(new_cle_out, outer_loop->_parent); 1905 set_idom(new_cle_out, new_cle, dd); 1906 set_loop(new_sfpt, outer_loop->_parent); 1907 set_idom(new_sfpt, new_cle_out, dd); 1908 set_loop(new_le, outer_loop->_parent); 1909 set_idom(new_le, new_sfpt, dd); 1910 set_loop(new_tail, outer_loop->_parent); 1911 set_idom(new_tail, new_le, dd); 1912 set_idom(new_cl, new_l, dd); 1913 1914 old_new.map(l->_idx, new_l); 1915 old_new.map(tail->_idx, new_tail); 1916 old_new.map(le->_idx, new_le); 1917 old_new.map(sfpt->_idx, new_sfpt); 1918 1919 new_l->set_req(LoopNode::LoopBackControl, new_tail); 1920 new_l->set_req(0, new_l); 1921 new_tail->set_req(0, new_le); 1922 new_le->set_req(0, new_sfpt); 1923 new_sfpt->set_req(0, new_cle_out); 1924 new_cle_out->set_req(0, new_cle); 1925 new_cl->set_req(LoopNode::EntryControl, new_l); 1926 1927 _igvn.register_new_node_with_optimizer(new_l); 1928 _igvn.register_new_node_with_optimizer(new_tail); 1929 _igvn.register_new_node_with_optimizer(new_le); 1930 } else { 1931 Node *newhead = old_new[loop->_head->_idx]; 1932 newhead->as_Loop()->clear_strip_mined(); 1933 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl)); 1934 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 1935 } 1936 // Look at data node that were assigned a control in the outer 1937 // loop: they are kept in the outer loop by the safepoint so start 1938 // from the safepoint node's inputs. 1939 IdealLoopTree* outer_loop = get_loop(l); 1940 Node_Stack stack(2); 1941 stack.push(sfpt, 1); 1942 uint new_counter = C->unique(); 1943 while (stack.size() > 0) { 1944 Node* n = stack.node(); 1945 uint i = stack.index(); 1946 while (i < n->req() && 1947 (n->in(i) == NULL || 1948 !has_ctrl(n->in(i)) || 1949 get_loop(get_ctrl(n->in(i))) != outer_loop || 1950 (old_new[n->in(i)->_idx] != NULL && old_new[n->in(i)->_idx]->_idx >= new_counter))) { 1951 i++; 1952 } 1953 if (i < n->req()) { 1954 stack.set_index(i+1); 1955 stack.push(n->in(i), 0); 1956 } else { 1957 assert(old_new[n->_idx] == NULL || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet"); 1958 Node* m = n == sfpt ? new_sfpt : n->clone(); 1959 if (m != NULL) { 1960 for (uint i = 0; i < n->req(); i++) { 1961 if (m->in(i) != NULL && old_new[m->in(i)->_idx] != NULL) { 1962 m->set_req(i, old_new[m->in(i)->_idx]); 1963 } 1964 } 1965 } else { 1966 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?"); 1967 } 1968 if (n != sfpt) { 1969 extra_data_nodes.push(n); 1970 _igvn.register_new_node_with_optimizer(m); 1971 assert(get_ctrl(n) == cle_out, "what other control?"); 1972 set_ctrl(m, new_cle_out); 1973 old_new.map(n->_idx, m); 1974 } 1975 stack.pop(); 1976 } 1977 } 1978 if (mode == CloneIncludesStripMined) { 1979 _igvn.register_new_node_with_optimizer(new_sfpt); 1980 _igvn.register_new_node_with_optimizer(new_cle_out); 1981 } 1982 // Some other transformation may have pessimistically assign some 1983 // data nodes to the outer loop. Set their control so they are out 1984 // of the outer loop. 1985 ResourceMark rm; 1986 Unique_Node_List wq; 1987 for (uint i = 0; i < extra_data_nodes.size(); i++) { 1988 Node* old = extra_data_nodes.at(i); 1989 clone_outer_loop_helper(old, loop, outer_loop, old_new, wq, this, true); 1990 } 1991 Node* new_ctrl = cl->outer_loop_exit(); 1992 assert(get_loop(new_ctrl) != outer_loop, "must be out of the loop nest"); 1993 for (uint i = 0; i < wq.size(); i++) { 1994 Node* n = wq.at(i); 1995 set_ctrl(n, new_ctrl); 1996 clone_outer_loop_helper(n, loop, outer_loop, old_new, wq, this, false); 1997 } 1998 } else { 1999 Node *newhead = old_new[loop->_head->_idx]; 2000 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 2001 } 2002 } 2003 2004 //------------------------------clone_loop------------------------------------- 2005 // 2006 // C L O N E A L O O P B O D Y 2007 // 2008 // This is the basic building block of the loop optimizations. It clones an 2009 // entire loop body. It makes an old_new loop body mapping; with this mapping 2010 // you can find the new-loop equivalent to an old-loop node. All new-loop 2011 // nodes are exactly equal to their old-loop counterparts, all edges are the 2012 // same. All exits from the old-loop now have a RegionNode that merges the 2013 // equivalent new-loop path. This is true even for the normal "loop-exit" 2014 // condition. All uses of loop-invariant old-loop values now come from (one 2015 // or more) Phis that merge their new-loop equivalents. 2016 // 2017 // This operation leaves the graph in an illegal state: there are two valid 2018 // control edges coming from the loop pre-header to both loop bodies. I'll 2019 // definitely have to hack the graph after running this transform. 2020 // 2021 // From this building block I will further edit edges to perform loop peeling 2022 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc. 2023 // 2024 // Parameter side_by_size_idom: 2025 // When side_by_size_idom is NULL, the dominator tree is constructed for 2026 // the clone loop to dominate the original. Used in construction of 2027 // pre-main-post loop sequence. 2028 // When nonnull, the clone and original are side-by-side, both are 2029 // dominated by the side_by_side_idom node. Used in construction of 2030 // unswitched loops. 2031 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd, 2032 CloneLoopMode mode, Node* side_by_side_idom) { 2033 2034 LoopNode* head = loop->_head->as_Loop(); 2035 head->verify_strip_mined(1); 2036 2037 if (C->do_vector_loop() && PrintOpto) { 2038 const char* mname = C->method()->name()->as_quoted_ascii(); 2039 if (mname != NULL) { 2040 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname); 2041 } 2042 } 2043 2044 CloneMap& cm = C->clone_map(); 2045 Dict* dict = cm.dict(); 2046 if (C->do_vector_loop()) { 2047 cm.set_clone_idx(cm.max_gen()+1); 2048 #ifndef PRODUCT 2049 if (PrintOpto) { 2050 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx()); 2051 loop->dump_head(); 2052 } 2053 #endif 2054 } 2055 2056 // Step 1: Clone the loop body. Make the old->new mapping. 2057 uint i; 2058 for( i = 0; i < loop->_body.size(); i++ ) { 2059 Node *old = loop->_body.at(i); 2060 Node *nnn = old->clone(); 2061 old_new.map( old->_idx, nnn ); 2062 if (C->do_vector_loop()) { 2063 cm.verify_insert_and_clone(old, nnn, cm.clone_idx()); 2064 } 2065 _igvn.register_new_node_with_optimizer(nnn); 2066 } 2067 2068 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop; 2069 2070 // Step 2: Fix the edges in the new body. If the old input is outside the 2071 // loop use it. If the old input is INside the loop, use the corresponding 2072 // new node instead. 2073 for( i = 0; i < loop->_body.size(); i++ ) { 2074 Node *old = loop->_body.at(i); 2075 Node *nnn = old_new[old->_idx]; 2076 // Fix CFG/Loop controlling the new node 2077 if (has_ctrl(old)) { 2078 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]); 2079 } else { 2080 set_loop(nnn, outer_loop->_parent); 2081 if (old->outcnt() > 0) { 2082 set_idom( nnn, old_new[idom(old)->_idx], dd ); 2083 } 2084 } 2085 // Correct edges to the new node 2086 for( uint j = 0; j < nnn->req(); j++ ) { 2087 Node *n = nnn->in(j); 2088 if( n ) { 2089 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n ); 2090 if( loop->is_member( old_in_loop ) ) 2091 nnn->set_req(j, old_new[n->_idx]); 2092 } 2093 } 2094 _igvn.hash_find_insert(nnn); 2095 } 2096 2097 ResourceArea *area = Thread::current()->resource_area(); 2098 Node_List extra_data_nodes(area); // data nodes in the outer strip mined loop 2099 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes); 2100 2101 // Step 3: Now fix control uses. Loop varying control uses have already 2102 // been fixed up (as part of all input edges in Step 2). Loop invariant 2103 // control uses must be either an IfFalse or an IfTrue. Make a merge 2104 // point to merge the old and new IfFalse/IfTrue nodes; make the use 2105 // refer to this. 2106 Node_List worklist(area); 2107 uint new_counter = C->unique(); 2108 for( i = 0; i < loop->_body.size(); i++ ) { 2109 Node* old = loop->_body.at(i); 2110 if( !old->is_CFG() ) continue; 2111 2112 // Copy uses to a worklist, so I can munge the def-use info 2113 // with impunity. 2114 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 2115 worklist.push(old->fast_out(j)); 2116 2117 while( worklist.size() ) { // Visit all uses 2118 Node *use = worklist.pop(); 2119 if (!has_node(use)) continue; // Ignore dead nodes 2120 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 2121 if( !loop->is_member( use_loop ) && use->is_CFG() ) { 2122 // Both OLD and USE are CFG nodes here. 2123 assert( use->is_Proj(), "" ); 2124 Node* nnn = old_new[old->_idx]; 2125 2126 Node* newuse = NULL; 2127 if (head->is_strip_mined() && mode != IgnoreStripMined) { 2128 CountedLoopNode* cl = head->as_CountedLoop(); 2129 CountedLoopEndNode* cle = cl->loopexit(); 2130 Node* cle_out = cle->proj_out_or_null(false); 2131 if (use == cle_out) { 2132 IfNode* le = cl->outer_loop_end(); 2133 use = le->proj_out(false); 2134 use_loop = get_loop(use); 2135 if (mode == CloneIncludesStripMined) { 2136 nnn = old_new[le->_idx]; 2137 } else { 2138 newuse = old_new[cle_out->_idx]; 2139 } 2140 } 2141 } 2142 if (newuse == NULL) { 2143 newuse = use->clone(); 2144 } 2145 2146 // Clone the loop exit control projection 2147 if (C->do_vector_loop()) { 2148 cm.verify_insert_and_clone(use, newuse, cm.clone_idx()); 2149 } 2150 newuse->set_req(0,nnn); 2151 _igvn.register_new_node_with_optimizer(newuse); 2152 set_loop(newuse, use_loop); 2153 set_idom(newuse, nnn, dom_depth(nnn) + 1 ); 2154 2155 // We need a Region to merge the exit from the peeled body and the 2156 // exit from the old loop body. 2157 RegionNode *r = new RegionNode(3); 2158 // Map the old use to the new merge point 2159 old_new.map( use->_idx, r ); 2160 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use)); 2161 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" ); 2162 2163 // The original user of 'use' uses 'r' instead. 2164 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) { 2165 Node* useuse = use->last_out(l); 2166 _igvn.rehash_node_delayed(useuse); 2167 uint uses_found = 0; 2168 if( useuse->in(0) == use ) { 2169 useuse->set_req(0, r); 2170 uses_found++; 2171 if( useuse->is_CFG() ) { 2172 assert( dom_depth(useuse) > dd_r, "" ); 2173 set_idom(useuse, r, dom_depth(useuse)); 2174 } 2175 } 2176 for( uint k = 1; k < useuse->req(); k++ ) { 2177 if( useuse->in(k) == use ) { 2178 useuse->set_req(k, r); 2179 uses_found++; 2180 if (useuse->is_Loop() && k == LoopNode::EntryControl) { 2181 assert(dom_depth(useuse) > dd_r , ""); 2182 set_idom(useuse, r, dom_depth(useuse)); 2183 } 2184 } 2185 } 2186 l -= uses_found; // we deleted 1 or more copies of this edge 2187 } 2188 2189 // Now finish up 'r' 2190 r->set_req( 1, newuse ); 2191 r->set_req( 2, use ); 2192 _igvn.register_new_node_with_optimizer(r); 2193 set_loop(r, use_loop); 2194 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r); 2195 } // End of if a loop-exit test 2196 } 2197 } 2198 2199 // Step 4: If loop-invariant use is not control, it must be dominated by a 2200 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region 2201 // there if needed. Make a Phi there merging old and new used values. 2202 Node_List *split_if_set = NULL; 2203 Node_List *split_bool_set = NULL; 2204 Node_List *split_cex_set = NULL; 2205 for( i = 0; i < loop->_body.size(); i++ ) { 2206 Node* old = loop->_body.at(i); 2207 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set, 2208 split_bool_set, split_cex_set, worklist, new_counter, 2209 mode); 2210 } 2211 2212 for (i = 0; i < extra_data_nodes.size(); i++) { 2213 Node* old = extra_data_nodes.at(i); 2214 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set, 2215 split_bool_set, split_cex_set, worklist, new_counter, 2216 mode); 2217 } 2218 2219 // Check for IFs that need splitting/cloning. Happens if an IF outside of 2220 // the loop uses a condition set in the loop. The original IF probably 2221 // takes control from one or more OLD Regions (which in turn get from NEW 2222 // Regions). In any case, there will be a set of Phis for each merge point 2223 // from the IF up to where the original BOOL def exists the loop. 2224 if (split_if_set) { 2225 while (split_if_set->size()) { 2226 Node *iff = split_if_set->pop(); 2227 if (iff->in(1)->is_Phi()) { 2228 Node *b = clone_iff(iff->in(1)->as_Phi(), loop); 2229 _igvn.replace_input_of(iff, 1, b); 2230 } 2231 } 2232 } 2233 if (split_bool_set) { 2234 while (split_bool_set->size()) { 2235 Node *b = split_bool_set->pop(); 2236 Node *phi = b->in(1); 2237 assert(phi->is_Phi(), ""); 2238 CmpNode *cmp = clone_bool((PhiNode*)phi, loop); 2239 _igvn.replace_input_of(b, 1, cmp); 2240 } 2241 } 2242 if (split_cex_set) { 2243 while (split_cex_set->size()) { 2244 Node *b = split_cex_set->pop(); 2245 assert(b->in(0)->is_Region(), ""); 2246 assert(b->in(1)->is_Phi(), ""); 2247 assert(b->in(0)->in(0) == b->in(1)->in(0), ""); 2248 split_up(b, b->in(0), NULL); 2249 } 2250 } 2251 2252 } 2253 2254 2255 //---------------------- stride_of_possible_iv ------------------------------------- 2256 // Looks for an iff/bool/comp with one operand of the compare 2257 // being a cycle involving an add and a phi, 2258 // with an optional truncation (left-shift followed by a right-shift) 2259 // of the add. Returns zero if not an iv. 2260 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) { 2261 Node* trunc1 = NULL; 2262 Node* trunc2 = NULL; 2263 const TypeInt* ttype = NULL; 2264 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) { 2265 return 0; 2266 } 2267 BoolNode* bl = iff->in(1)->as_Bool(); 2268 Node* cmp = bl->in(1); 2269 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) { 2270 return 0; 2271 } 2272 // Must have an invariant operand 2273 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) { 2274 return 0; 2275 } 2276 Node* add2 = NULL; 2277 Node* cmp1 = cmp->in(1); 2278 if (cmp1->is_Phi()) { 2279 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) ))) 2280 Node* phi = cmp1; 2281 for (uint i = 1; i < phi->req(); i++) { 2282 Node* in = phi->in(i); 2283 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in, 2284 &trunc1, &trunc2, &ttype); 2285 if (add && add->in(1) == phi) { 2286 add2 = add->in(2); 2287 break; 2288 } 2289 } 2290 } else { 2291 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) ))) 2292 Node* addtrunc = cmp1; 2293 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc, 2294 &trunc1, &trunc2, &ttype); 2295 if (add && add->in(1)->is_Phi()) { 2296 Node* phi = add->in(1); 2297 for (uint i = 1; i < phi->req(); i++) { 2298 if (phi->in(i) == addtrunc) { 2299 add2 = add->in(2); 2300 break; 2301 } 2302 } 2303 } 2304 } 2305 if (add2 != NULL) { 2306 const TypeInt* add2t = _igvn.type(add2)->is_int(); 2307 if (add2t->is_con()) { 2308 return add2t->get_con(); 2309 } 2310 } 2311 return 0; 2312 } 2313 2314 2315 //---------------------- stay_in_loop ------------------------------------- 2316 // Return the (unique) control output node that's in the loop (if it exists.) 2317 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) { 2318 Node* unique = NULL; 2319 if (!n) return NULL; 2320 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2321 Node* use = n->fast_out(i); 2322 if (!has_ctrl(use) && loop->is_member(get_loop(use))) { 2323 if (unique != NULL) { 2324 return NULL; 2325 } 2326 unique = use; 2327 } 2328 } 2329 return unique; 2330 } 2331 2332 //------------------------------ register_node ------------------------------------- 2333 // Utility to register node "n" with PhaseIdealLoop 2334 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) { 2335 _igvn.register_new_node_with_optimizer(n); 2336 loop->_body.push(n); 2337 if (n->is_CFG()) { 2338 set_loop(n, loop); 2339 set_idom(n, pred, ddepth); 2340 } else { 2341 set_ctrl(n, pred); 2342 } 2343 } 2344 2345 //------------------------------ proj_clone ------------------------------------- 2346 // Utility to create an if-projection 2347 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) { 2348 ProjNode* c = p->clone()->as_Proj(); 2349 c->set_req(0, iff); 2350 return c; 2351 } 2352 2353 //------------------------------ short_circuit_if ------------------------------------- 2354 // Force the iff control output to be the live_proj 2355 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) { 2356 guarantee(live_proj != NULL, "null projection"); 2357 int proj_con = live_proj->_con; 2358 assert(proj_con == 0 || proj_con == 1, "false or true projection"); 2359 Node *con = _igvn.intcon(proj_con); 2360 set_ctrl(con, C->root()); 2361 if (iff) { 2362 iff->set_req(1, con); 2363 } 2364 return con; 2365 } 2366 2367 //------------------------------ insert_if_before_proj ------------------------------------- 2368 // Insert a new if before an if projection (* - new node) 2369 // 2370 // before 2371 // if(test) 2372 // / \ 2373 // v v 2374 // other-proj proj (arg) 2375 // 2376 // after 2377 // if(test) 2378 // / \ 2379 // / v 2380 // | * proj-clone 2381 // v | 2382 // other-proj v 2383 // * new_if(relop(cmp[IU](left,right))) 2384 // / \ 2385 // v v 2386 // * new-proj proj 2387 // (returned) 2388 // 2389 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) { 2390 IfNode* iff = proj->in(0)->as_If(); 2391 IdealLoopTree *loop = get_loop(proj); 2392 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2393 int ddepth = dom_depth(proj); 2394 2395 _igvn.rehash_node_delayed(iff); 2396 _igvn.rehash_node_delayed(proj); 2397 2398 proj->set_req(0, NULL); // temporary disconnect 2399 ProjNode* proj2 = proj_clone(proj, iff); 2400 register_node(proj2, loop, iff, ddepth); 2401 2402 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right); 2403 register_node(cmp, loop, proj2, ddepth); 2404 2405 BoolNode* bol = new BoolNode(cmp, relop); 2406 register_node(bol, loop, proj2, ddepth); 2407 2408 int opcode = iff->Opcode(); 2409 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode"); 2410 IfNode* new_if = (opcode == Op_If) ? new IfNode(proj2, bol, iff->_prob, iff->_fcnt): 2411 new RangeCheckNode(proj2, bol, iff->_prob, iff->_fcnt); 2412 register_node(new_if, loop, proj2, ddepth); 2413 2414 proj->set_req(0, new_if); // reattach 2415 set_idom(proj, new_if, ddepth); 2416 2417 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj(); 2418 guarantee(new_exit != NULL, "null exit node"); 2419 register_node(new_exit, get_loop(other_proj), new_if, ddepth); 2420 2421 return new_exit; 2422 } 2423 2424 //------------------------------ insert_region_before_proj ------------------------------------- 2425 // Insert a region before an if projection (* - new node) 2426 // 2427 // before 2428 // if(test) 2429 // / | 2430 // v | 2431 // proj v 2432 // other-proj 2433 // 2434 // after 2435 // if(test) 2436 // / | 2437 // v | 2438 // * proj-clone v 2439 // | other-proj 2440 // v 2441 // * new-region 2442 // | 2443 // v 2444 // * dum_if 2445 // / \ 2446 // v \ 2447 // * dum-proj v 2448 // proj 2449 // 2450 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) { 2451 IfNode* iff = proj->in(0)->as_If(); 2452 IdealLoopTree *loop = get_loop(proj); 2453 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2454 int ddepth = dom_depth(proj); 2455 2456 _igvn.rehash_node_delayed(iff); 2457 _igvn.rehash_node_delayed(proj); 2458 2459 proj->set_req(0, NULL); // temporary disconnect 2460 ProjNode* proj2 = proj_clone(proj, iff); 2461 register_node(proj2, loop, iff, ddepth); 2462 2463 RegionNode* reg = new RegionNode(2); 2464 reg->set_req(1, proj2); 2465 register_node(reg, loop, iff, ddepth); 2466 2467 IfNode* dum_if = new IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt); 2468 register_node(dum_if, loop, reg, ddepth); 2469 2470 proj->set_req(0, dum_if); // reattach 2471 set_idom(proj, dum_if, ddepth); 2472 2473 ProjNode* dum_proj = proj_clone(other_proj, dum_if); 2474 register_node(dum_proj, loop, dum_if, ddepth); 2475 2476 return reg; 2477 } 2478 2479 //------------------------------ insert_cmpi_loop_exit ------------------------------------- 2480 // Clone a signed compare loop exit from an unsigned compare and 2481 // insert it before the unsigned cmp on the stay-in-loop path. 2482 // All new nodes inserted in the dominator tree between the original 2483 // if and it's projections. The original if test is replaced with 2484 // a constant to force the stay-in-loop path. 2485 // 2486 // This is done to make sure that the original if and it's projections 2487 // still dominate the same set of control nodes, that the ctrl() relation 2488 // from data nodes to them is preserved, and that their loop nesting is 2489 // preserved. 2490 // 2491 // before 2492 // if(i <u limit) unsigned compare loop exit 2493 // / | 2494 // v v 2495 // exit-proj stay-in-loop-proj 2496 // 2497 // after 2498 // if(stay-in-loop-const) original if 2499 // / | 2500 // / v 2501 // / if(i < limit) new signed test 2502 // / / | 2503 // / / v 2504 // / / if(i <u limit) new cloned unsigned test 2505 // / / / | 2506 // v v v | 2507 // region | 2508 // | | 2509 // dum-if | 2510 // / | | 2511 // ether | | 2512 // v v 2513 // exit-proj stay-in-loop-proj 2514 // 2515 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) { 2516 const bool Signed = true; 2517 const bool Unsigned = false; 2518 2519 BoolNode* bol = if_cmpu->in(1)->as_Bool(); 2520 if (bol->_test._test != BoolTest::lt) return NULL; 2521 CmpNode* cmpu = bol->in(1)->as_Cmp(); 2522 if (cmpu->Opcode() != Op_CmpU) return NULL; 2523 int stride = stride_of_possible_iv(if_cmpu); 2524 if (stride == 0) return NULL; 2525 2526 Node* lp_proj = stay_in_loop(if_cmpu, loop); 2527 guarantee(lp_proj != NULL, "null loop node"); 2528 2529 ProjNode* lp_continue = lp_proj->as_Proj(); 2530 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj(); 2531 2532 Node* limit = NULL; 2533 if (stride > 0) { 2534 limit = cmpu->in(2); 2535 } else { 2536 limit = _igvn.makecon(TypeInt::ZERO); 2537 set_ctrl(limit, C->root()); 2538 } 2539 // Create a new region on the exit path 2540 RegionNode* reg = insert_region_before_proj(lp_exit); 2541 guarantee(reg != NULL, "null region node"); 2542 2543 // Clone the if-cmpu-true-false using a signed compare 2544 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge; 2545 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue); 2546 reg->add_req(cmpi_exit); 2547 2548 // Clone the if-cmpu-true-false 2549 BoolTest::mask rel_u = bol->_test._test; 2550 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue); 2551 reg->add_req(cmpu_exit); 2552 2553 // Force original if to stay in loop. 2554 short_circuit_if(if_cmpu, lp_continue); 2555 2556 return cmpi_exit->in(0)->as_If(); 2557 } 2558 2559 //------------------------------ remove_cmpi_loop_exit ------------------------------------- 2560 // Remove a previously inserted signed compare loop exit. 2561 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) { 2562 Node* lp_proj = stay_in_loop(if_cmp, loop); 2563 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI && 2564 stay_in_loop(lp_proj, loop)->is_If() && 2565 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu"); 2566 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO); 2567 set_ctrl(con, C->root()); 2568 if_cmp->set_req(1, con); 2569 } 2570 2571 //------------------------------ scheduled_nodelist ------------------------------------- 2572 // Create a post order schedule of nodes that are in the 2573 // "member" set. The list is returned in "sched". 2574 // The first node in "sched" is the loop head, followed by 2575 // nodes which have no inputs in the "member" set, and then 2576 // followed by the nodes that have an immediate input dependence 2577 // on a node in "sched". 2578 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) { 2579 2580 assert(member.test(loop->_head->_idx), "loop head must be in member set"); 2581 Arena *a = Thread::current()->resource_area(); 2582 VectorSet visited(a); 2583 Node_Stack nstack(a, loop->_body.size()); 2584 2585 Node* n = loop->_head; // top of stack is cached in "n" 2586 uint idx = 0; 2587 visited.set(n->_idx); 2588 2589 // Initially push all with no inputs from within member set 2590 for(uint i = 0; i < loop->_body.size(); i++ ) { 2591 Node *elt = loop->_body.at(i); 2592 if (member.test(elt->_idx)) { 2593 bool found = false; 2594 for (uint j = 0; j < elt->req(); j++) { 2595 Node* def = elt->in(j); 2596 if (def && member.test(def->_idx) && def != elt) { 2597 found = true; 2598 break; 2599 } 2600 } 2601 if (!found && elt != loop->_head) { 2602 nstack.push(n, idx); 2603 n = elt; 2604 assert(!visited.test(n->_idx), "not seen yet"); 2605 visited.set(n->_idx); 2606 } 2607 } 2608 } 2609 2610 // traverse out's that are in the member set 2611 while (true) { 2612 if (idx < n->outcnt()) { 2613 Node* use = n->raw_out(idx); 2614 idx++; 2615 if (!visited.test_set(use->_idx)) { 2616 if (member.test(use->_idx)) { 2617 nstack.push(n, idx); 2618 n = use; 2619 idx = 0; 2620 } 2621 } 2622 } else { 2623 // All outputs processed 2624 sched.push(n); 2625 if (nstack.is_empty()) break; 2626 n = nstack.node(); 2627 idx = nstack.index(); 2628 nstack.pop(); 2629 } 2630 } 2631 } 2632 2633 2634 //------------------------------ has_use_in_set ------------------------------------- 2635 // Has a use in the vector set 2636 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) { 2637 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2638 Node* use = n->fast_out(j); 2639 if (vset.test(use->_idx)) { 2640 return true; 2641 } 2642 } 2643 return false; 2644 } 2645 2646 2647 //------------------------------ has_use_internal_to_set ------------------------------------- 2648 // Has use internal to the vector set (ie. not in a phi at the loop head) 2649 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) { 2650 Node* head = loop->_head; 2651 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2652 Node* use = n->fast_out(j); 2653 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) { 2654 return true; 2655 } 2656 } 2657 return false; 2658 } 2659 2660 2661 //------------------------------ clone_for_use_outside_loop ------------------------------------- 2662 // clone "n" for uses that are outside of loop 2663 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) { 2664 int cloned = 0; 2665 assert(worklist.size() == 0, "should be empty"); 2666 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2667 Node* use = n->fast_out(j); 2668 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) { 2669 worklist.push(use); 2670 } 2671 } 2672 while( worklist.size() ) { 2673 Node *use = worklist.pop(); 2674 if (!has_node(use) || use->in(0) == C->top()) continue; 2675 uint j; 2676 for (j = 0; j < use->req(); j++) { 2677 if (use->in(j) == n) break; 2678 } 2679 assert(j < use->req(), "must be there"); 2680 2681 // clone "n" and insert it between the inputs of "n" and the use outside the loop 2682 Node* n_clone = n->clone(); 2683 _igvn.replace_input_of(use, j, n_clone); 2684 cloned++; 2685 Node* use_c; 2686 if (!use->is_Phi()) { 2687 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0); 2688 } else { 2689 // Use in a phi is considered a use in the associated predecessor block 2690 use_c = use->in(0)->in(j); 2691 } 2692 set_ctrl(n_clone, use_c); 2693 assert(!loop->is_member(get_loop(use_c)), "should be outside loop"); 2694 get_loop(use_c)->_body.push(n_clone); 2695 _igvn.register_new_node_with_optimizer(n_clone); 2696 #ifndef PRODUCT 2697 if (TracePartialPeeling) { 2698 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx); 2699 } 2700 #endif 2701 } 2702 return cloned; 2703 } 2704 2705 2706 //------------------------------ clone_for_special_use_inside_loop ------------------------------------- 2707 // clone "n" for special uses that are in the not_peeled region. 2708 // If these def-uses occur in separate blocks, the code generator 2709 // marks the method as not compilable. For example, if a "BoolNode" 2710 // is in a different basic block than the "IfNode" that uses it, then 2711 // the compilation is aborted in the code generator. 2712 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 2713 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) { 2714 if (n->is_Phi() || n->is_Load()) { 2715 return; 2716 } 2717 assert(worklist.size() == 0, "should be empty"); 2718 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2719 Node* use = n->fast_out(j); 2720 if ( not_peel.test(use->_idx) && 2721 (use->is_If() || use->is_CMove() || use->is_Bool()) && 2722 use->in(1) == n) { 2723 worklist.push(use); 2724 } 2725 } 2726 if (worklist.size() > 0) { 2727 // clone "n" and insert it between inputs of "n" and the use 2728 Node* n_clone = n->clone(); 2729 loop->_body.push(n_clone); 2730 _igvn.register_new_node_with_optimizer(n_clone); 2731 set_ctrl(n_clone, get_ctrl(n)); 2732 sink_list.push(n_clone); 2733 not_peel <<= n_clone->_idx; // add n_clone to not_peel set. 2734 #ifndef PRODUCT 2735 if (TracePartialPeeling) { 2736 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx); 2737 } 2738 #endif 2739 while( worklist.size() ) { 2740 Node *use = worklist.pop(); 2741 _igvn.rehash_node_delayed(use); 2742 for (uint j = 1; j < use->req(); j++) { 2743 if (use->in(j) == n) { 2744 use->set_req(j, n_clone); 2745 } 2746 } 2747 } 2748 } 2749 } 2750 2751 2752 //------------------------------ insert_phi_for_loop ------------------------------------- 2753 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 2754 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) { 2755 Node *phi = PhiNode::make(lp, back_edge_val); 2756 phi->set_req(LoopNode::EntryControl, lp_entry_val); 2757 // Use existing phi if it already exists 2758 Node *hit = _igvn.hash_find_insert(phi); 2759 if( hit == NULL ) { 2760 _igvn.register_new_node_with_optimizer(phi); 2761 set_ctrl(phi, lp); 2762 } else { 2763 // Remove the new phi from the graph and use the hit 2764 _igvn.remove_dead_node(phi); 2765 phi = hit; 2766 } 2767 _igvn.replace_input_of(use, idx, phi); 2768 } 2769 2770 #ifdef ASSERT 2771 //------------------------------ is_valid_loop_partition ------------------------------------- 2772 // Validate the loop partition sets: peel and not_peel 2773 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, 2774 VectorSet& not_peel ) { 2775 uint i; 2776 // Check that peel_list entries are in the peel set 2777 for (i = 0; i < peel_list.size(); i++) { 2778 if (!peel.test(peel_list.at(i)->_idx)) { 2779 return false; 2780 } 2781 } 2782 // Check at loop members are in one of peel set or not_peel set 2783 for (i = 0; i < loop->_body.size(); i++ ) { 2784 Node *def = loop->_body.at(i); 2785 uint di = def->_idx; 2786 // Check that peel set elements are in peel_list 2787 if (peel.test(di)) { 2788 if (not_peel.test(di)) { 2789 return false; 2790 } 2791 // Must be in peel_list also 2792 bool found = false; 2793 for (uint j = 0; j < peel_list.size(); j++) { 2794 if (peel_list.at(j)->_idx == di) { 2795 found = true; 2796 break; 2797 } 2798 } 2799 if (!found) { 2800 return false; 2801 } 2802 } else if (not_peel.test(di)) { 2803 if (peel.test(di)) { 2804 return false; 2805 } 2806 } else { 2807 return false; 2808 } 2809 } 2810 return true; 2811 } 2812 2813 //------------------------------ is_valid_clone_loop_exit_use ------------------------------------- 2814 // Ensure a use outside of loop is of the right form 2815 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) { 2816 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2817 return (use->is_Phi() && 2818 use_c->is_Region() && use_c->req() == 3 && 2819 (use_c->in(exit_idx)->Opcode() == Op_IfTrue || 2820 use_c->in(exit_idx)->Opcode() == Op_IfFalse || 2821 use_c->in(exit_idx)->Opcode() == Op_JumpProj) && 2822 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) ); 2823 } 2824 2825 //------------------------------ is_valid_clone_loop_form ------------------------------------- 2826 // Ensure that all uses outside of loop are of the right form 2827 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 2828 uint orig_exit_idx, uint clone_exit_idx) { 2829 uint len = peel_list.size(); 2830 for (uint i = 0; i < len; i++) { 2831 Node *def = peel_list.at(i); 2832 2833 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 2834 Node *use = def->fast_out(j); 2835 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2836 if (!loop->is_member(get_loop(use_c))) { 2837 // use is not in the loop, check for correct structure 2838 if (use->in(0) == def) { 2839 // Okay 2840 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) { 2841 return false; 2842 } 2843 } 2844 } 2845 } 2846 return true; 2847 } 2848 #endif 2849 2850 //------------------------------ partial_peel ------------------------------------- 2851 // Partially peel (aka loop rotation) the top portion of a loop (called 2852 // the peel section below) by cloning it and placing one copy just before 2853 // the new loop head and the other copy at the bottom of the new loop. 2854 // 2855 // before after where it came from 2856 // 2857 // stmt1 stmt1 2858 // loop: stmt2 clone 2859 // stmt2 if condA goto exitA clone 2860 // if condA goto exitA new_loop: new 2861 // stmt3 stmt3 clone 2862 // if !condB goto loop if condB goto exitB clone 2863 // exitB: stmt2 orig 2864 // stmt4 if !condA goto new_loop orig 2865 // exitA: goto exitA 2866 // exitB: 2867 // stmt4 2868 // exitA: 2869 // 2870 // Step 1: find the cut point: an exit test on probable 2871 // induction variable. 2872 // Step 2: schedule (with cloning) operations in the peel 2873 // section that can be executed after the cut into 2874 // the section that is not peeled. This may need 2875 // to clone operations into exit blocks. For 2876 // instance, a reference to A[i] in the not-peel 2877 // section and a reference to B[i] in an exit block 2878 // may cause a left-shift of i by 2 to be placed 2879 // in the peel block. This step will clone the left 2880 // shift into the exit block and sink the left shift 2881 // from the peel to the not-peel section. 2882 // Step 3: clone the loop, retarget the control, and insert 2883 // phis for values that are live across the new loop 2884 // head. This is very dependent on the graph structure 2885 // from clone_loop. It creates region nodes for 2886 // exit control and associated phi nodes for values 2887 // flow out of the loop through that exit. The region 2888 // node is dominated by the clone's control projection. 2889 // So the clone's peel section is placed before the 2890 // new loop head, and the clone's not-peel section is 2891 // forms the top part of the new loop. The original 2892 // peel section forms the tail of the new loop. 2893 // Step 4: update the dominator tree and recompute the 2894 // dominator depth. 2895 // 2896 // orig 2897 // 2898 // stmt1 2899 // | 2900 // v 2901 // loop predicate 2902 // | 2903 // v 2904 // loop<----+ 2905 // | | 2906 // stmt2 | 2907 // | | 2908 // v | 2909 // ifA | 2910 // / | | 2911 // v v | 2912 // false true ^ <-- last_peel 2913 // / | | 2914 // / ===|==cut | 2915 // / stmt3 | <-- first_not_peel 2916 // / | | 2917 // | v | 2918 // v ifB | 2919 // exitA: / \ | 2920 // / \ | 2921 // v v | 2922 // false true | 2923 // / \ | 2924 // / ----+ 2925 // | 2926 // v 2927 // exitB: 2928 // stmt4 2929 // 2930 // 2931 // after clone loop 2932 // 2933 // stmt1 2934 // | 2935 // v 2936 // loop predicate 2937 // / \ 2938 // clone / \ orig 2939 // / \ 2940 // / \ 2941 // v v 2942 // +---->loop loop<----+ 2943 // | | | | 2944 // | stmt2 stmt2 | 2945 // | | | | 2946 // | v v | 2947 // | ifA ifA | 2948 // | | \ / | | 2949 // | v v v v | 2950 // ^ true false false true ^ <-- last_peel 2951 // | | ^ \ / | | 2952 // | cut==|== \ \ / ===|==cut | 2953 // | stmt3 \ \ / stmt3 | <-- first_not_peel 2954 // | | dom | | | | 2955 // | v \ 1v v2 v | 2956 // | ifB regionA ifB | 2957 // | / \ | / \ | 2958 // | / \ v / \ | 2959 // | v v exitA: v v | 2960 // | true false false true | 2961 // | / ^ \ / \ | 2962 // +---- \ \ / ----+ 2963 // dom \ / 2964 // \ 1v v2 2965 // regionB 2966 // | 2967 // v 2968 // exitB: 2969 // stmt4 2970 // 2971 // 2972 // after partial peel 2973 // 2974 // stmt1 2975 // | 2976 // v 2977 // loop predicate 2978 // / 2979 // clone / orig 2980 // / TOP 2981 // / \ 2982 // v v 2983 // TOP->loop loop----+ 2984 // | | | 2985 // stmt2 stmt2 | 2986 // | | | 2987 // v v | 2988 // ifA ifA | 2989 // | \ / | | 2990 // v v v v | 2991 // true false false true | <-- last_peel 2992 // | ^ \ / +------|---+ 2993 // +->newloop \ \ / === ==cut | | 2994 // | stmt3 \ \ / TOP | | 2995 // | | dom | | stmt3 | | <-- first_not_peel 2996 // | v \ 1v v2 v | | 2997 // | ifB regionA ifB ^ v 2998 // | / \ | / \ | | 2999 // | / \ v / \ | | 3000 // | v v exitA: v v | | 3001 // | true false false true | | 3002 // | / ^ \ / \ | | 3003 // | | \ \ / v | | 3004 // | | dom \ / TOP | | 3005 // | | \ 1v v2 | | 3006 // ^ v regionB | | 3007 // | | | | | 3008 // | | v ^ v 3009 // | | exitB: | | 3010 // | | stmt4 | | 3011 // | +------------>-----------------+ | 3012 // | | 3013 // +-----------------<---------------------+ 3014 // 3015 // 3016 // final graph 3017 // 3018 // stmt1 3019 // | 3020 // v 3021 // loop predicate 3022 // | 3023 // v 3024 // stmt2 clone 3025 // | 3026 // v 3027 // ........> ifA clone 3028 // : / | 3029 // dom / | 3030 // : v v 3031 // : false true 3032 // : | | 3033 // : | v 3034 // : | newloop<-----+ 3035 // : | | | 3036 // : | stmt3 clone | 3037 // : | | | 3038 // : | v | 3039 // : | ifB | 3040 // : | / \ | 3041 // : | v v | 3042 // : | false true | 3043 // : | | | | 3044 // : | v stmt2 | 3045 // : | exitB: | | 3046 // : | stmt4 v | 3047 // : | ifA orig | 3048 // : | / \ | 3049 // : | / \ | 3050 // : | v v | 3051 // : | false true | 3052 // : | / \ | 3053 // : v v -----+ 3054 // RegionA 3055 // | 3056 // v 3057 // exitA 3058 // 3059 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) { 3060 3061 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only"); 3062 if (!loop->_head->is_Loop()) { 3063 return false; 3064 } 3065 LoopNode *head = loop->_head->as_Loop(); 3066 3067 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) { 3068 return false; 3069 } 3070 3071 // Check for complex exit control 3072 for (uint ii = 0; ii < loop->_body.size(); ii++) { 3073 Node *n = loop->_body.at(ii); 3074 int opc = n->Opcode(); 3075 if (n->is_Call() || 3076 opc == Op_Catch || 3077 opc == Op_CatchProj || 3078 opc == Op_Jump || 3079 opc == Op_JumpProj) { 3080 #ifndef PRODUCT 3081 if (TracePartialPeeling) { 3082 tty->print_cr("\nExit control too complex: lp: %d", head->_idx); 3083 } 3084 #endif 3085 return false; 3086 } 3087 } 3088 3089 int dd = dom_depth(head); 3090 3091 // Step 1: find cut point 3092 3093 // Walk up dominators to loop head looking for first loop exit 3094 // which is executed on every path thru loop. 3095 IfNode *peel_if = NULL; 3096 IfNode *peel_if_cmpu = NULL; 3097 3098 Node *iff = loop->tail(); 3099 while (iff != head) { 3100 if (iff->is_If()) { 3101 Node *ctrl = get_ctrl(iff->in(1)); 3102 if (ctrl->is_top()) return false; // Dead test on live IF. 3103 // If loop-varying exit-test, check for induction variable 3104 if (loop->is_member(get_loop(ctrl)) && 3105 loop->is_loop_exit(iff) && 3106 is_possible_iv_test(iff)) { 3107 Node* cmp = iff->in(1)->in(1); 3108 if (cmp->Opcode() == Op_CmpI) { 3109 peel_if = iff->as_If(); 3110 } else { 3111 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU"); 3112 peel_if_cmpu = iff->as_If(); 3113 } 3114 } 3115 } 3116 iff = idom(iff); 3117 } 3118 3119 // Prefer signed compare over unsigned compare. 3120 IfNode* new_peel_if = NULL; 3121 if (peel_if == NULL) { 3122 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) { 3123 return false; // No peel point found 3124 } 3125 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop); 3126 if (new_peel_if == NULL) { 3127 return false; // No peel point found 3128 } 3129 peel_if = new_peel_if; 3130 } 3131 Node* last_peel = stay_in_loop(peel_if, loop); 3132 Node* first_not_peeled = stay_in_loop(last_peel, loop); 3133 if (first_not_peeled == NULL || first_not_peeled == head) { 3134 return false; 3135 } 3136 3137 #ifndef PRODUCT 3138 if (TraceLoopOpts) { 3139 tty->print("PartialPeel "); 3140 loop->dump_head(); 3141 } 3142 3143 if (TracePartialPeeling) { 3144 tty->print_cr("before partial peel one iteration"); 3145 Node_List wl; 3146 Node* t = head->in(2); 3147 while (true) { 3148 wl.push(t); 3149 if (t == head) break; 3150 t = idom(t); 3151 } 3152 while (wl.size() > 0) { 3153 Node* tt = wl.pop(); 3154 tt->dump(); 3155 if (tt == last_peel) tty->print_cr("-- cut --"); 3156 } 3157 } 3158 #endif 3159 ResourceArea *area = Thread::current()->resource_area(); 3160 VectorSet peel(area); 3161 VectorSet not_peel(area); 3162 Node_List peel_list(area); 3163 Node_List worklist(area); 3164 Node_List sink_list(area); 3165 3166 if (!may_require_nodes(loop->est_loop_clone_sz(2))) { 3167 return false; 3168 } 3169 3170 // Set of cfg nodes to peel are those that are executable from 3171 // the head through last_peel. 3172 assert(worklist.size() == 0, "should be empty"); 3173 worklist.push(head); 3174 peel.set(head->_idx); 3175 while (worklist.size() > 0) { 3176 Node *n = worklist.pop(); 3177 if (n != last_peel) { 3178 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3179 Node* use = n->fast_out(j); 3180 if (use->is_CFG() && 3181 loop->is_member(get_loop(use)) && 3182 !peel.test_set(use->_idx)) { 3183 worklist.push(use); 3184 } 3185 } 3186 } 3187 } 3188 3189 // Set of non-cfg nodes to peel are those that are control 3190 // dependent on the cfg nodes. 3191 uint i; 3192 for(i = 0; i < loop->_body.size(); i++ ) { 3193 Node *n = loop->_body.at(i); 3194 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n; 3195 if (peel.test(n_c->_idx)) { 3196 peel.set(n->_idx); 3197 } else { 3198 not_peel.set(n->_idx); 3199 } 3200 } 3201 3202 // Step 2: move operations from the peeled section down into the 3203 // not-peeled section 3204 3205 // Get a post order schedule of nodes in the peel region 3206 // Result in right-most operand. 3207 scheduled_nodelist(loop, peel, peel_list ); 3208 3209 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 3210 3211 // For future check for too many new phis 3212 uint old_phi_cnt = 0; 3213 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 3214 Node* use = head->fast_out(j); 3215 if (use->is_Phi()) old_phi_cnt++; 3216 } 3217 3218 #ifndef PRODUCT 3219 if (TracePartialPeeling) { 3220 tty->print_cr("\npeeled list"); 3221 } 3222 #endif 3223 3224 // Evacuate nodes in peel region into the not_peeled region if possible 3225 uint new_phi_cnt = 0; 3226 uint cloned_for_outside_use = 0; 3227 for (i = 0; i < peel_list.size();) { 3228 Node* n = peel_list.at(i); 3229 #ifndef PRODUCT 3230 if (TracePartialPeeling) n->dump(); 3231 #endif 3232 bool incr = true; 3233 if ( !n->is_CFG() ) { 3234 3235 if ( has_use_in_set(n, not_peel) ) { 3236 3237 // If not used internal to the peeled region, 3238 // move "n" from peeled to not_peeled region. 3239 3240 if ( !has_use_internal_to_set(n, peel, loop) ) { 3241 3242 // if not pinned and not a load (which maybe anti-dependent on a store) 3243 // and not a CMove (Matcher expects only bool->cmove). 3244 if (n->in(0) == NULL && !n->is_Load() && !n->is_CMove()) { 3245 cloned_for_outside_use += clone_for_use_outside_loop( loop, n, worklist ); 3246 sink_list.push(n); 3247 peel >>= n->_idx; // delete n from peel set. 3248 not_peel <<= n->_idx; // add n to not_peel set. 3249 peel_list.remove(i); 3250 incr = false; 3251 #ifndef PRODUCT 3252 if (TracePartialPeeling) { 3253 tty->print_cr("sink to not_peeled region: %d newbb: %d", 3254 n->_idx, get_ctrl(n)->_idx); 3255 } 3256 #endif 3257 } 3258 } else { 3259 // Otherwise check for special def-use cases that span 3260 // the peel/not_peel boundary such as bool->if 3261 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist ); 3262 new_phi_cnt++; 3263 } 3264 } 3265 } 3266 if (incr) i++; 3267 } 3268 3269 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) { 3270 #ifndef PRODUCT 3271 if (TracePartialPeeling) { 3272 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c", 3273 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F'); 3274 } 3275 #endif 3276 if (new_peel_if != NULL) { 3277 remove_cmpi_loop_exit(new_peel_if, loop); 3278 } 3279 // Inhibit more partial peeling on this loop 3280 assert(!head->is_partial_peel_loop(), "not partial peeled"); 3281 head->mark_partial_peel_failed(); 3282 if (cloned_for_outside_use > 0) { 3283 // Terminate this round of loop opts because 3284 // the graph outside this loop was changed. 3285 C->set_major_progress(); 3286 return true; 3287 } 3288 return false; 3289 } 3290 3291 // Step 3: clone loop, retarget control, and insert new phis 3292 3293 // Create new loop head for new phis and to hang 3294 // the nodes being moved (sinked) from the peel region. 3295 LoopNode* new_head = new LoopNode(last_peel, last_peel); 3296 new_head->set_unswitch_count(head->unswitch_count()); // Preserve 3297 _igvn.register_new_node_with_optimizer(new_head); 3298 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled"); 3299 _igvn.replace_input_of(first_not_peeled, 0, new_head); 3300 set_loop(new_head, loop); 3301 loop->_body.push(new_head); 3302 not_peel.set(new_head->_idx); 3303 set_idom(new_head, last_peel, dom_depth(first_not_peeled)); 3304 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled)); 3305 3306 while (sink_list.size() > 0) { 3307 Node* n = sink_list.pop(); 3308 set_ctrl(n, new_head); 3309 } 3310 3311 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 3312 3313 clone_loop(loop, old_new, dd, IgnoreStripMined); 3314 3315 const uint clone_exit_idx = 1; 3316 const uint orig_exit_idx = 2; 3317 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop"); 3318 3319 Node* head_clone = old_new[head->_idx]; 3320 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop(); 3321 Node* orig_tail_clone = head_clone->in(2); 3322 3323 // Add phi if "def" node is in peel set and "use" is not 3324 3325 for(i = 0; i < peel_list.size(); i++ ) { 3326 Node *def = peel_list.at(i); 3327 if (!def->is_CFG()) { 3328 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 3329 Node *use = def->fast_out(j); 3330 if (has_node(use) && use->in(0) != C->top() && 3331 (!peel.test(use->_idx) || 3332 (use->is_Phi() && use->in(0) == head)) ) { 3333 worklist.push(use); 3334 } 3335 } 3336 while( worklist.size() ) { 3337 Node *use = worklist.pop(); 3338 for (uint j = 1; j < use->req(); j++) { 3339 Node* n = use->in(j); 3340 if (n == def) { 3341 3342 // "def" is in peel set, "use" is not in peel set 3343 // or "use" is in the entry boundary (a phi) of the peel set 3344 3345 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use; 3346 3347 if ( loop->is_member(get_loop( use_c )) ) { 3348 // use is in loop 3349 if (old_new[use->_idx] != NULL) { // null for dead code 3350 Node* use_clone = old_new[use->_idx]; 3351 _igvn.replace_input_of(use, j, C->top()); 3352 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone ); 3353 } 3354 } else { 3355 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format"); 3356 // use is not in the loop, check if the live range includes the cut 3357 Node* lp_if = use_c->in(orig_exit_idx)->in(0); 3358 if (not_peel.test(lp_if->_idx)) { 3359 assert(j == orig_exit_idx, "use from original loop"); 3360 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone ); 3361 } 3362 } 3363 } 3364 } 3365 } 3366 } 3367 } 3368 3369 // Step 3b: retarget control 3370 3371 // Redirect control to the new loop head if a cloned node in 3372 // the not_peeled region has control that points into the peeled region. 3373 // This necessary because the cloned peeled region will be outside 3374 // the loop. 3375 // from to 3376 // cloned-peeled <---+ 3377 // new_head_clone: | <--+ 3378 // cloned-not_peeled in(0) in(0) 3379 // orig-peeled 3380 3381 for(i = 0; i < loop->_body.size(); i++ ) { 3382 Node *n = loop->_body.at(i); 3383 if (!n->is_CFG() && n->in(0) != NULL && 3384 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) { 3385 Node* n_clone = old_new[n->_idx]; 3386 _igvn.replace_input_of(n_clone, 0, new_head_clone); 3387 } 3388 } 3389 3390 // Backedge of the surviving new_head (the clone) is original last_peel 3391 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel); 3392 3393 // Cut first node in original not_peel set 3394 _igvn.rehash_node_delayed(new_head); // Multiple edge updates: 3395 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of 3396 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls 3397 3398 // Copy head_clone back-branch info to original head 3399 // and remove original head's loop entry and 3400 // clone head's back-branch 3401 _igvn.rehash_node_delayed(head); // Multiple edge updates 3402 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl)); 3403 head->set_req(LoopNode::LoopBackControl, C->top()); 3404 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top()); 3405 3406 // Similarly modify the phis 3407 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) { 3408 Node* use = head->fast_out(k); 3409 if (use->is_Phi() && use->outcnt() > 0) { 3410 Node* use_clone = old_new[use->_idx]; 3411 _igvn.rehash_node_delayed(use); // Multiple edge updates 3412 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl)); 3413 use->set_req(LoopNode::LoopBackControl, C->top()); 3414 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top()); 3415 } 3416 } 3417 3418 // Step 4: update dominator tree and dominator depth 3419 3420 set_idom(head, orig_tail_clone, dd); 3421 recompute_dom_depth(); 3422 3423 // Inhibit more partial peeling on this loop 3424 new_head_clone->set_partial_peel_loop(); 3425 C->set_major_progress(); 3426 loop->record_for_igvn(); 3427 3428 #ifndef PRODUCT 3429 if (TracePartialPeeling) { 3430 tty->print_cr("\nafter partial peel one iteration"); 3431 Node_List wl(area); 3432 Node* t = last_peel; 3433 while (true) { 3434 wl.push(t); 3435 if (t == head_clone) break; 3436 t = idom(t); 3437 } 3438 while (wl.size() > 0) { 3439 Node* tt = wl.pop(); 3440 if (tt == head) tty->print_cr("orig head"); 3441 else if (tt == new_head_clone) tty->print_cr("new head"); 3442 else if (tt == head_clone) tty->print_cr("clone head"); 3443 tt->dump(); 3444 } 3445 } 3446 #endif 3447 return true; 3448 } 3449 3450 //------------------------------reorg_offsets---------------------------------- 3451 // Reorganize offset computations to lower register pressure. Mostly 3452 // prevent loop-fallout uses of the pre-incremented trip counter (which are 3453 // then alive with the post-incremented trip counter forcing an extra 3454 // register move) 3455 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) { 3456 // Perform it only for canonical counted loops. 3457 // Loop's shape could be messed up by iteration_split_impl. 3458 if (!loop->_head->is_CountedLoop()) 3459 return; 3460 if (!loop->_head->as_Loop()->is_valid_counted_loop()) 3461 return; 3462 3463 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 3464 CountedLoopEndNode *cle = cl->loopexit(); 3465 Node *exit = cle->proj_out(false); 3466 Node *phi = cl->phi(); 3467 3468 // Check for the special case when using the pre-incremented trip-counter on 3469 // the fall-out path (forces the pre-incremented and post-incremented trip 3470 // counter to be live at the same time). Fix this by adjusting to use the 3471 // post-increment trip counter. 3472 3473 bool progress = true; 3474 while (progress) { 3475 progress = false; 3476 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) { 3477 Node* use = phi->fast_out(i); // User of trip-counter 3478 if (!has_ctrl(use)) continue; 3479 Node *u_ctrl = get_ctrl(use); 3480 if (use->is_Phi()) { 3481 u_ctrl = NULL; 3482 for (uint j = 1; j < use->req(); j++) 3483 if (use->in(j) == phi) 3484 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j)); 3485 } 3486 IdealLoopTree *u_loop = get_loop(u_ctrl); 3487 // Look for loop-invariant use 3488 if (u_loop == loop) continue; 3489 if (loop->is_member(u_loop)) continue; 3490 // Check that use is live out the bottom. Assuming the trip-counter 3491 // update is right at the bottom, uses of of the loop middle are ok. 3492 if (dom_lca(exit, u_ctrl) != exit) continue; 3493 // Hit! Refactor use to use the post-incremented tripcounter. 3494 // Compute a post-increment tripcounter. 3495 Node* c = exit; 3496 if (cl->is_strip_mined()) { 3497 IdealLoopTree* outer_loop = get_loop(cl->outer_loop()); 3498 if (!outer_loop->is_member(u_loop)) { 3499 c = cl->outer_loop_exit(); 3500 } 3501 } 3502 Node *opaq = new Opaque2Node(C, cle->incr()); 3503 register_new_node(opaq, c); 3504 Node *neg_stride = _igvn.intcon(-cle->stride_con()); 3505 set_ctrl(neg_stride, C->root()); 3506 Node *post = new AddINode(opaq, neg_stride); 3507 register_new_node(post, c); 3508 _igvn.rehash_node_delayed(use); 3509 for (uint j = 1; j < use->req(); j++) { 3510 if (use->in(j) == phi) 3511 use->set_req(j, post); 3512 } 3513 // Since DU info changed, rerun loop 3514 progress = true; 3515 break; 3516 } 3517 } 3518 3519 }