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/subtypenode.hpp" 44 #include "utilities/macros.hpp" 45 46 //============================================================================= 47 //------------------------------split_thru_phi--------------------------------- 48 // Split Node 'n' through merge point if there is enough win. 49 Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) { 50 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) { 51 // ConvI2L may have type information on it which is unsafe to push up 52 // so disable this for now 53 return NULL; 54 } 55 56 // Splitting range check CastIIs through a loop induction Phi can 57 // cause new Phis to be created that are left unrelated to the loop 58 // induction Phi and prevent optimizations (vectorization) 59 if (n->Opcode() == Op_CastII && n->as_CastII()->has_range_check() && 60 region->is_CountedLoop() && n->in(1) == region->as_CountedLoop()->phi()) { 61 return NULL; 62 } 63 64 int wins = 0; 65 assert(!n->is_CFG(), ""); 66 assert(region->is_Region(), ""); 67 68 const Type* type = n->bottom_type(); 69 const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr(); 70 Node *phi; 71 if (t_oop != NULL && t_oop->is_known_instance_field()) { 72 int iid = t_oop->instance_id(); 73 int index = C->get_alias_index(t_oop); 74 int offset = t_oop->offset(); 75 phi = new PhiNode(region, type, NULL, iid, index, offset); 76 } else { 77 phi = PhiNode::make_blank(region, n); 78 } 79 uint old_unique = C->unique(); 80 for (uint i = 1; i < region->req(); i++) { 81 Node *x; 82 Node* the_clone = NULL; 83 if (region->in(i) == C->top()) { 84 x = C->top(); // Dead path? Use a dead data op 85 } else { 86 x = n->clone(); // Else clone up the data op 87 the_clone = x; // Remember for possible deletion. 88 // Alter data node to use pre-phi inputs 89 if (n->in(0) == region) 90 x->set_req( 0, region->in(i) ); 91 for (uint j = 1; j < n->req(); j++) { 92 Node *in = n->in(j); 93 if (in->is_Phi() && in->in(0) == region) 94 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone 95 } 96 } 97 // Check for a 'win' on some paths 98 const Type *t = x->Value(&_igvn); 99 100 bool singleton = t->singleton(); 101 102 // A TOP singleton indicates that there are no possible values incoming 103 // along a particular edge. In most cases, this is OK, and the Phi will 104 // be eliminated later in an Ideal call. However, we can't allow this to 105 // happen if the singleton occurs on loop entry, as the elimination of 106 // the PhiNode may cause the resulting node to migrate back to a previous 107 // loop iteration. 108 if (singleton && t == Type::TOP) { 109 // Is_Loop() == false does not confirm the absence of a loop (e.g., an 110 // irreducible loop may not be indicated by an affirmative is_Loop()); 111 // therefore, the only top we can split thru a phi is on a backedge of 112 // a loop. 113 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 114 } 115 116 if (singleton) { 117 wins++; 118 x = ((PhaseGVN&)_igvn).makecon(t); 119 } else { 120 // We now call Identity to try to simplify the cloned node. 121 // Note that some Identity methods call phase->type(this). 122 // Make sure that the type array is big enough for 123 // our new node, even though we may throw the node away. 124 // (Note: This tweaking with igvn only works because x is a new node.) 125 _igvn.set_type(x, t); 126 // If x is a TypeNode, capture any more-precise type permanently into Node 127 // otherwise it will be not updated during igvn->transform since 128 // igvn->type(x) is set to x->Value() already. 129 x->raise_bottom_type(t); 130 Node* y = x->Identity(&_igvn); 131 if (y != x) { 132 wins++; 133 x = y; 134 } else { 135 y = _igvn.hash_find(x); 136 if (y) { 137 wins++; 138 x = y; 139 } else { 140 // Else x is a new node we are keeping 141 // We do not need register_new_node_with_optimizer 142 // because set_type has already been called. 143 _igvn._worklist.push(x); 144 } 145 } 146 } 147 if (x != the_clone && the_clone != NULL) 148 _igvn.remove_dead_node(the_clone); 149 phi->set_req( i, x ); 150 } 151 // Too few wins? 152 if (wins <= policy) { 153 _igvn.remove_dead_node(phi); 154 return NULL; 155 } 156 157 // Record Phi 158 register_new_node( phi, region ); 159 160 for (uint i2 = 1; i2 < phi->req(); i2++) { 161 Node *x = phi->in(i2); 162 // If we commoned up the cloned 'x' with another existing Node, 163 // the existing Node picks up a new use. We need to make the 164 // existing Node occur higher up so it dominates its uses. 165 Node *old_ctrl; 166 IdealLoopTree *old_loop; 167 168 if (x->is_Con()) { 169 // Constant's control is always root. 170 set_ctrl(x, C->root()); 171 continue; 172 } 173 // The occasional new node 174 if (x->_idx >= old_unique) { // Found a new, unplaced node? 175 old_ctrl = NULL; 176 old_loop = NULL; // Not in any prior loop 177 } else { 178 old_ctrl = get_ctrl(x); 179 old_loop = get_loop(old_ctrl); // Get prior loop 180 } 181 // New late point must dominate new use 182 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2)); 183 if (new_ctrl == old_ctrl) // Nothing is changed 184 continue; 185 186 IdealLoopTree *new_loop = get_loop(new_ctrl); 187 188 // Don't move x into a loop if its uses are 189 // outside of loop. Otherwise x will be cloned 190 // for each use outside of this loop. 191 IdealLoopTree *use_loop = get_loop(region); 192 if (!new_loop->is_member(use_loop) && 193 (old_loop == NULL || !new_loop->is_member(old_loop))) { 194 // Take early control, later control will be recalculated 195 // during next iteration of loop optimizations. 196 new_ctrl = get_early_ctrl(x); 197 new_loop = get_loop(new_ctrl); 198 } 199 // Set new location 200 set_ctrl(x, new_ctrl); 201 // If changing loop bodies, see if we need to collect into new body 202 if (old_loop != new_loop) { 203 if (old_loop && !old_loop->_child) 204 old_loop->_body.yank(x); 205 if (!new_loop->_child) 206 new_loop->_body.push(x); // Collect body info 207 } 208 } 209 210 return phi; 211 } 212 213 //------------------------------dominated_by------------------------------------ 214 // Replace the dominated test with an obvious true or false. Place it on the 215 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the 216 // live path up to the dominating control. 217 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff, bool flip, bool exclude_loop_predicate ) { 218 if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); } 219 220 // prevdom is the dominating projection of the dominating test. 221 assert( iff->is_If(), "" ); 222 assert(iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd || iff->Opcode() == Op_RangeCheck, "Check this code when new subtype is added"); 223 int pop = prevdom->Opcode(); 224 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" ); 225 if (flip) { 226 if (pop == Op_IfTrue) 227 pop = Op_IfFalse; 228 else 229 pop = Op_IfTrue; 230 } 231 // 'con' is set to true or false to kill the dominated test. 232 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO); 233 set_ctrl(con, C->root()); // Constant gets a new use 234 // Hack the dominated test 235 _igvn.replace_input_of(iff, 1, con); 236 237 // If I dont have a reachable TRUE and FALSE path following the IfNode then 238 // I can assume this path reaches an infinite loop. In this case it's not 239 // important to optimize the data Nodes - either the whole compilation will 240 // be tossed or this path (and all data Nodes) will go dead. 241 if (iff->outcnt() != 2) return; 242 243 // Make control-dependent data Nodes on the live path (path that will remain 244 // once the dominated IF is removed) become control-dependent on the 245 // dominating projection. 246 Node* dp = iff->as_If()->proj_out_or_null(pop == Op_IfTrue); 247 248 // Loop predicates may have depending checks which should not 249 // be skipped. For example, range check predicate has two checks 250 // for lower and upper bounds. 251 if (dp == NULL) 252 return; 253 254 ProjNode* dp_proj = dp->as_Proj(); 255 ProjNode* unc_proj = iff->as_If()->proj_out(1 - dp_proj->_con)->as_Proj(); 256 if (exclude_loop_predicate && 257 (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) != NULL || 258 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_profile_predicate) != NULL || 259 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check) != NULL)) { 260 // If this is a range check (IfNode::is_range_check), do not 261 // reorder because Compile::allow_range_check_smearing might have 262 // changed the check. 263 return; // Let IGVN transformation change control dependence. 264 } 265 266 IdealLoopTree *old_loop = get_loop(dp); 267 268 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { 269 Node* cd = dp->fast_out(i); // Control-dependent node 270 if (cd->depends_only_on_test()) { 271 assert(cd->in(0) == dp, ""); 272 _igvn.replace_input_of(cd, 0, prevdom); 273 set_early_ctrl(cd); 274 IdealLoopTree *new_loop = get_loop(get_ctrl(cd)); 275 if (old_loop != new_loop) { 276 if (!old_loop->_child) old_loop->_body.yank(cd); 277 if (!new_loop->_child) new_loop->_body.push(cd); 278 } 279 --i; 280 --imax; 281 } 282 } 283 } 284 285 //------------------------------has_local_phi_input---------------------------- 286 // Return TRUE if 'n' has Phi inputs from its local block and no other 287 // block-local inputs (all non-local-phi inputs come from earlier blocks) 288 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) { 289 Node *n_ctrl = get_ctrl(n); 290 // See if some inputs come from a Phi in this block, or from before 291 // this block. 292 uint i; 293 for( i = 1; i < n->req(); i++ ) { 294 Node *phi = n->in(i); 295 if( phi->is_Phi() && phi->in(0) == n_ctrl ) 296 break; 297 } 298 if( i >= n->req() ) 299 return NULL; // No Phi inputs; nowhere to clone thru 300 301 // Check for inputs created between 'n' and the Phi input. These 302 // must split as well; they have already been given the chance 303 // (courtesy of a post-order visit) and since they did not we must 304 // recover the 'cost' of splitting them by being very profitable 305 // when splitting 'n'. Since this is unlikely we simply give up. 306 for( i = 1; i < n->req(); i++ ) { 307 Node *m = n->in(i); 308 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) { 309 // We allow the special case of AddP's with no local inputs. 310 // This allows us to split-up address expressions. 311 if (m->is_AddP() && 312 get_ctrl(m->in(2)) != n_ctrl && 313 get_ctrl(m->in(3)) != n_ctrl) { 314 // Move the AddP up to dominating point 315 Node* c = find_non_split_ctrl(idom(n_ctrl)); 316 if (c->is_OuterStripMinedLoop()) { 317 c->as_Loop()->verify_strip_mined(1); 318 c = c->in(LoopNode::EntryControl); 319 } 320 set_ctrl_and_loop(m, c); 321 continue; 322 } 323 return NULL; 324 } 325 assert(n->is_Phi() || m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control"); 326 } 327 328 return n_ctrl; 329 } 330 331 //------------------------------remix_address_expressions---------------------- 332 // Rework addressing expressions to get the most loop-invariant stuff 333 // moved out. We'd like to do all associative operators, but it's especially 334 // important (common) to do address expressions. 335 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) { 336 if (!has_ctrl(n)) return NULL; 337 Node *n_ctrl = get_ctrl(n); 338 IdealLoopTree *n_loop = get_loop(n_ctrl); 339 340 // See if 'n' mixes loop-varying and loop-invariant inputs and 341 // itself is loop-varying. 342 343 // Only interested in binary ops (and AddP) 344 if( n->req() < 3 || n->req() > 4 ) return NULL; 345 346 Node *n1_ctrl = get_ctrl(n->in( 1)); 347 Node *n2_ctrl = get_ctrl(n->in( 2)); 348 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3)); 349 IdealLoopTree *n1_loop = get_loop( n1_ctrl ); 350 IdealLoopTree *n2_loop = get_loop( n2_ctrl ); 351 IdealLoopTree *n3_loop = get_loop( n3_ctrl ); 352 353 // Does one of my inputs spin in a tighter loop than self? 354 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) || 355 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) || 356 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) ) 357 return NULL; // Leave well enough alone 358 359 // Is at least one of my inputs loop-invariant? 360 if( n1_loop == n_loop && 361 n2_loop == n_loop && 362 n3_loop == n_loop ) 363 return NULL; // No loop-invariant inputs 364 365 366 int n_op = n->Opcode(); 367 368 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2). 369 if( n_op == Op_LShiftI ) { 370 // Scale is loop invariant 371 Node *scale = n->in(2); 372 Node *scale_ctrl = get_ctrl(scale); 373 IdealLoopTree *scale_loop = get_loop(scale_ctrl ); 374 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) ) 375 return NULL; 376 const TypeInt *scale_t = scale->bottom_type()->isa_int(); 377 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 ) 378 return NULL; // Dont bother with byte/short masking 379 // Add must vary with loop (else shift would be loop-invariant) 380 Node *add = n->in(1); 381 Node *add_ctrl = get_ctrl(add); 382 IdealLoopTree *add_loop = get_loop(add_ctrl); 383 //assert( n_loop == add_loop, "" ); 384 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops 385 386 // Convert I-V into I+ (0-V); same for V-I 387 if( add->Opcode() == Op_SubI && 388 _igvn.type( add->in(1) ) != TypeInt::ZERO ) { 389 Node *zero = _igvn.intcon(0); 390 set_ctrl(zero, C->root()); 391 Node *neg = new SubINode( _igvn.intcon(0), add->in(2) ); 392 register_new_node( neg, get_ctrl(add->in(2) ) ); 393 add = new AddINode( add->in(1), neg ); 394 register_new_node( add, add_ctrl ); 395 } 396 if( add->Opcode() != Op_AddI ) return NULL; 397 // See if one add input is loop invariant 398 Node *add_var = add->in(1); 399 Node *add_var_ctrl = get_ctrl(add_var); 400 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 401 Node *add_invar = add->in(2); 402 Node *add_invar_ctrl = get_ctrl(add_invar); 403 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl ); 404 if( add_var_loop == n_loop ) { 405 } else if( add_invar_loop == n_loop ) { 406 // Swap to find the invariant part 407 add_invar = add_var; 408 add_invar_ctrl = add_var_ctrl; 409 add_invar_loop = add_var_loop; 410 add_var = add->in(2); 411 Node *add_var_ctrl = get_ctrl(add_var); 412 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 413 } else // Else neither input is loop invariant 414 return NULL; 415 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) ) 416 return NULL; // No invariant part of the add? 417 418 // Yes! Reshape address expression! 419 Node *inv_scale = new LShiftINode( add_invar, scale ); 420 Node *inv_scale_ctrl = 421 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ? 422 add_invar_ctrl : scale_ctrl; 423 register_new_node( inv_scale, inv_scale_ctrl ); 424 Node *var_scale = new LShiftINode( add_var, scale ); 425 register_new_node( var_scale, n_ctrl ); 426 Node *var_add = new AddINode( var_scale, inv_scale ); 427 register_new_node( var_add, n_ctrl ); 428 _igvn.replace_node( n, var_add ); 429 return var_add; 430 } 431 432 // Replace (I+V) with (V+I) 433 if( n_op == Op_AddI || 434 n_op == Op_AddL || 435 n_op == Op_AddF || 436 n_op == Op_AddD || 437 n_op == Op_MulI || 438 n_op == Op_MulL || 439 n_op == Op_MulF || 440 n_op == Op_MulD ) { 441 if( n2_loop == n_loop ) { 442 assert( n1_loop != n_loop, "" ); 443 n->swap_edges(1, 2); 444 } 445 } 446 447 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V), 448 // but not if I2 is a constant. 449 if( n_op == Op_AddP ) { 450 if( n2_loop == n_loop && n3_loop != n_loop ) { 451 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) { 452 Node *n22_ctrl = get_ctrl(n->in(2)->in(2)); 453 Node *n23_ctrl = get_ctrl(n->in(2)->in(3)); 454 IdealLoopTree *n22loop = get_loop( n22_ctrl ); 455 IdealLoopTree *n23_loop = get_loop( n23_ctrl ); 456 if( n22loop != n_loop && n22loop->is_member(n_loop) && 457 n23_loop == n_loop ) { 458 Node *add1 = new AddPNode( n->in(1), n->in(2)->in(2), n->in(3) ); 459 // Stuff new AddP in the loop preheader 460 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) ); 461 Node *add2 = new AddPNode( n->in(1), add1, n->in(2)->in(3) ); 462 register_new_node( add2, n_ctrl ); 463 _igvn.replace_node( n, add2 ); 464 return add2; 465 } 466 } 467 } 468 469 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V) 470 if (n2_loop != n_loop && n3_loop == n_loop) { 471 if (n->in(3)->Opcode() == Op_AddX) { 472 Node *V = n->in(3)->in(1); 473 Node *I = n->in(3)->in(2); 474 if (is_member(n_loop,get_ctrl(V))) { 475 } else { 476 Node *tmp = V; V = I; I = tmp; 477 } 478 if (!is_member(n_loop,get_ctrl(I))) { 479 Node *add1 = new AddPNode(n->in(1), n->in(2), I); 480 // Stuff new AddP in the loop preheader 481 register_new_node(add1, n_loop->_head->in(LoopNode::EntryControl)); 482 Node *add2 = new AddPNode(n->in(1), add1, V); 483 register_new_node(add2, n_ctrl); 484 _igvn.replace_node(n, add2); 485 return add2; 486 } 487 } 488 } 489 } 490 491 return NULL; 492 } 493 494 // Optimize ((in1[2*i] * in2[2*i]) + (in1[2*i+1] * in2[2*i+1])) 495 Node *PhaseIdealLoop::convert_add_to_muladd(Node* n) { 496 assert(n->Opcode() == Op_AddI, "sanity"); 497 Node * nn = NULL; 498 Node * in1 = n->in(1); 499 Node * in2 = n->in(2); 500 if (in1->Opcode() == Op_MulI && in2->Opcode() == Op_MulI) { 501 IdealLoopTree* loop_n = get_loop(get_ctrl(n)); 502 if (loop_n->is_counted() && 503 loop_n->_head->as_Loop()->is_valid_counted_loop() && 504 Matcher::match_rule_supported(Op_MulAddVS2VI) && 505 Matcher::match_rule_supported(Op_MulAddS2I)) { 506 Node* mul_in1 = in1->in(1); 507 Node* mul_in2 = in1->in(2); 508 Node* mul_in3 = in2->in(1); 509 Node* mul_in4 = in2->in(2); 510 if (mul_in1->Opcode() == Op_LoadS && 511 mul_in2->Opcode() == Op_LoadS && 512 mul_in3->Opcode() == Op_LoadS && 513 mul_in4->Opcode() == Op_LoadS) { 514 IdealLoopTree* loop1 = get_loop(get_ctrl(mul_in1)); 515 IdealLoopTree* loop2 = get_loop(get_ctrl(mul_in2)); 516 IdealLoopTree* loop3 = get_loop(get_ctrl(mul_in3)); 517 IdealLoopTree* loop4 = get_loop(get_ctrl(mul_in4)); 518 IdealLoopTree* loop5 = get_loop(get_ctrl(in1)); 519 IdealLoopTree* loop6 = get_loop(get_ctrl(in2)); 520 // All nodes should be in the same counted loop. 521 if (loop_n == loop1 && loop_n == loop2 && loop_n == loop3 && 522 loop_n == loop4 && loop_n == loop5 && loop_n == loop6) { 523 Node* adr1 = mul_in1->in(MemNode::Address); 524 Node* adr2 = mul_in2->in(MemNode::Address); 525 Node* adr3 = mul_in3->in(MemNode::Address); 526 Node* adr4 = mul_in4->in(MemNode::Address); 527 if (adr1->is_AddP() && adr2->is_AddP() && adr3->is_AddP() && adr4->is_AddP()) { 528 if ((adr1->in(AddPNode::Base) == adr3->in(AddPNode::Base)) && 529 (adr2->in(AddPNode::Base) == adr4->in(AddPNode::Base))) { 530 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in3, mul_in4); 531 register_new_node(nn, get_ctrl(n)); 532 _igvn.replace_node(n, nn); 533 return nn; 534 } else if ((adr1->in(AddPNode::Base) == adr4->in(AddPNode::Base)) && 535 (adr2->in(AddPNode::Base) == adr3->in(AddPNode::Base))) { 536 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in4, mul_in3); 537 register_new_node(nn, get_ctrl(n)); 538 _igvn.replace_node(n, nn); 539 return nn; 540 } 541 } 542 } 543 } 544 } 545 } 546 return nn; 547 } 548 549 //------------------------------conditional_move------------------------------- 550 // Attempt to replace a Phi with a conditional move. We have some pretty 551 // strict profitability requirements. All Phis at the merge point must 552 // be converted, so we can remove the control flow. We need to limit the 553 // number of c-moves to a small handful. All code that was in the side-arms 554 // of the CFG diamond is now speculatively executed. This code has to be 555 // "cheap enough". We are pretty much limited to CFG diamonds that merge 556 // 1 or 2 items with a total of 1 or 2 ops executed speculatively. 557 Node *PhaseIdealLoop::conditional_move( Node *region ) { 558 559 assert(region->is_Region(), "sanity check"); 560 if (region->req() != 3) return NULL; 561 562 // Check for CFG diamond 563 Node *lp = region->in(1); 564 Node *rp = region->in(2); 565 if (!lp || !rp) return NULL; 566 Node *lp_c = lp->in(0); 567 if (lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If()) return NULL; 568 IfNode *iff = lp_c->as_If(); 569 570 // Check for ops pinned in an arm of the diamond. 571 // Can't remove the control flow in this case 572 if (lp->outcnt() > 1) return NULL; 573 if (rp->outcnt() > 1) return NULL; 574 575 IdealLoopTree* r_loop = get_loop(region); 576 assert(r_loop == get_loop(iff), "sanity"); 577 // Always convert to CMOVE if all results are used only outside this loop. 578 bool used_inside_loop = (r_loop == _ltree_root); 579 580 // Check profitability 581 int cost = 0; 582 int phis = 0; 583 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 584 Node *out = region->fast_out(i); 585 if (!out->is_Phi()) continue; // Ignore other control edges, etc 586 phis++; 587 PhiNode* phi = out->as_Phi(); 588 BasicType bt = phi->type()->basic_type(); 589 switch (bt) { 590 case T_DOUBLE: 591 case T_FLOAT: 592 if (C->use_cmove()) { 593 continue; //TODO: maybe we want to add some cost 594 } 595 cost += Matcher::float_cmove_cost(); // Could be very expensive 596 break; 597 case T_LONG: { 598 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's 599 } 600 case T_INT: // These all CMOV fine 601 case T_ADDRESS: { // (RawPtr) 602 cost++; 603 break; 604 } 605 case T_NARROWOOP: // Fall through 606 case T_OBJECT: { // Base oops are OK, but not derived oops 607 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr(); 608 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a 609 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus 610 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we 611 // have a Phi for the base here that we convert to a CMOVE all is well 612 // and good. But if the base is dead, we'll not make a CMOVE. Later 613 // the allocator will have to produce a base by creating a CMOVE of the 614 // relevant bases. This puts the allocator in the business of 615 // manufacturing expensive instructions, generally a bad plan. 616 // Just Say No to Conditionally-Moved Derived Pointers. 617 if (tp && tp->offset() != 0) 618 return NULL; 619 cost++; 620 break; 621 } 622 default: 623 return NULL; // In particular, can't do memory or I/O 624 } 625 // Add in cost any speculative ops 626 for (uint j = 1; j < region->req(); j++) { 627 Node *proj = region->in(j); 628 Node *inp = phi->in(j); 629 if (get_ctrl(inp) == proj) { // Found local op 630 cost++; 631 // Check for a chain of dependent ops; these will all become 632 // speculative in a CMOV. 633 for (uint k = 1; k < inp->req(); k++) 634 if (get_ctrl(inp->in(k)) == proj) 635 cost += ConditionalMoveLimit; // Too much speculative goo 636 } 637 } 638 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode. 639 // This will likely Split-If, a higher-payoff operation. 640 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) { 641 Node* use = phi->fast_out(k); 642 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr()) 643 cost += ConditionalMoveLimit; 644 // Is there a use inside the loop? 645 // Note: check only basic types since CMoveP is pinned. 646 if (!used_inside_loop && is_java_primitive(bt)) { 647 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use); 648 if (r_loop == u_loop || r_loop->is_member(u_loop)) { 649 used_inside_loop = true; 650 } 651 } 652 } 653 }//for 654 Node* bol = iff->in(1); 655 if (bol->Opcode() == Op_Opaque4) { 656 return NULL; // Ignore loop predicate checks (the Opaque4 ensures they will go away) 657 } 658 assert(bol->Opcode() == Op_Bool, "Unexpected node"); 659 int cmp_op = bol->in(1)->Opcode(); 660 if (cmp_op == Op_SubTypeCheck) { // SubTypeCheck expansion expects an IfNode 661 return NULL; 662 } 663 // It is expensive to generate flags from a float compare. 664 // Avoid duplicated float compare. 665 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL; 666 667 float infrequent_prob = PROB_UNLIKELY_MAG(3); 668 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop. 669 if (used_inside_loop) { 670 if (cost >= ConditionalMoveLimit) return NULL; // Too much goo 671 672 // BlockLayoutByFrequency optimization moves infrequent branch 673 // from hot path. No point in CMOV'ing in such case (110 is used 674 // instead of 100 to take into account not exactness of float value). 675 if (BlockLayoutByFrequency) { 676 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f); 677 } 678 } 679 // Check for highly predictable branch. No point in CMOV'ing if 680 // we are going to predict accurately all the time. 681 if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) { 682 //keep going 683 } else if (iff->_prob < infrequent_prob || 684 iff->_prob > (1.0f - infrequent_prob)) 685 return NULL; 686 687 // -------------- 688 // Now replace all Phis with CMOV's 689 Node *cmov_ctrl = iff->in(0); 690 uint flip = (lp->Opcode() == Op_IfTrue); 691 Node_List wq; 692 while (1) { 693 PhiNode* phi = NULL; 694 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 695 Node *out = region->fast_out(i); 696 if (out->is_Phi()) { 697 phi = out->as_Phi(); 698 break; 699 } 700 } 701 if (phi == NULL) break; 702 if (PrintOpto && VerifyLoopOptimizations) { tty->print_cr("CMOV"); } 703 // Move speculative ops 704 wq.push(phi); 705 while (wq.size() > 0) { 706 Node *n = wq.pop(); 707 for (uint j = 1; j < n->req(); j++) { 708 Node* m = n->in(j); 709 if (m != NULL && !is_dominator(get_ctrl(m), cmov_ctrl)) { 710 #ifndef PRODUCT 711 if (PrintOpto && VerifyLoopOptimizations) { 712 tty->print(" speculate: "); 713 m->dump(); 714 } 715 #endif 716 set_ctrl(m, cmov_ctrl); 717 wq.push(m); 718 } 719 } 720 } 721 Node *cmov = CMoveNode::make(cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi)); 722 register_new_node( cmov, cmov_ctrl ); 723 _igvn.replace_node( phi, cmov ); 724 #ifndef PRODUCT 725 if (TraceLoopOpts) { 726 tty->print("CMOV "); 727 r_loop->dump_head(); 728 if (Verbose) { 729 bol->in(1)->dump(1); 730 cmov->dump(1); 731 } 732 } 733 if (VerifyLoopOptimizations) verify(); 734 #endif 735 } 736 737 // The useless CFG diamond will fold up later; see the optimization in 738 // RegionNode::Ideal. 739 _igvn._worklist.push(region); 740 741 return iff->in(1); 742 } 743 744 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) { 745 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 746 Node* u = m->fast_out(i); 747 if (u->is_CFG()) { 748 if (u->Opcode() == Op_NeverBranch) { 749 u = ((NeverBranchNode*)u)->proj_out(0); 750 enqueue_cfg_uses(u, wq); 751 } else { 752 wq.push(u); 753 } 754 } 755 } 756 } 757 758 // Try moving a store out of a loop, right before the loop 759 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) { 760 // Store has to be first in the loop body 761 IdealLoopTree *n_loop = get_loop(n_ctrl); 762 if (n->is_Store() && n_loop != _ltree_root && 763 n_loop->is_loop() && n_loop->_head->is_Loop() && 764 n->in(0) != NULL) { 765 Node* address = n->in(MemNode::Address); 766 Node* value = n->in(MemNode::ValueIn); 767 Node* mem = n->in(MemNode::Memory); 768 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 769 IdealLoopTree* value_loop = get_loop(get_ctrl(value)); 770 771 // - address and value must be loop invariant 772 // - memory must be a memory Phi for the loop 773 // - Store must be the only store on this memory slice in the 774 // loop: if there's another store following this one then value 775 // written at iteration i by the second store could be overwritten 776 // at iteration i+n by the first store: it's not safe to move the 777 // first store out of the loop 778 // - nothing must observe the memory Phi: it guarantees no read 779 // before the store, we are also guaranteed the store post 780 // dominates the loop head (ignoring a possible early 781 // exit). Otherwise there would be extra Phi involved between the 782 // loop's Phi and the store. 783 // - there must be no early exit from the loop before the Store 784 // (such an exit most of the time would be an extra use of the 785 // memory Phi but sometimes is a bottom memory Phi that takes the 786 // store as input). 787 788 if (!n_loop->is_member(address_loop) && 789 !n_loop->is_member(value_loop) && 790 mem->is_Phi() && mem->in(0) == n_loop->_head && 791 mem->outcnt() == 1 && 792 mem->in(LoopNode::LoopBackControl) == n) { 793 794 assert(n_loop->_tail != NULL, "need a tail"); 795 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop"); 796 797 // Verify that there's no early exit of the loop before the store. 798 bool ctrl_ok = false; 799 { 800 // Follow control from loop head until n, we exit the loop or 801 // we reach the tail 802 ResourceMark rm; 803 Unique_Node_List wq; 804 wq.push(n_loop->_head); 805 806 for (uint next = 0; next < wq.size(); ++next) { 807 Node *m = wq.at(next); 808 if (m == n->in(0)) { 809 ctrl_ok = true; 810 continue; 811 } 812 assert(!has_ctrl(m), "should be CFG"); 813 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) { 814 ctrl_ok = false; 815 break; 816 } 817 enqueue_cfg_uses(m, wq); 818 if (wq.size() > 10) { 819 ctrl_ok = false; 820 break; 821 } 822 } 823 } 824 if (ctrl_ok) { 825 // move the Store 826 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem); 827 _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl)); 828 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl)); 829 // Disconnect the phi now. An empty phi can confuse other 830 // optimizations in this pass of loop opts. 831 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl)); 832 n_loop->_body.yank(mem); 833 834 set_ctrl_and_loop(n, n->in(0)); 835 836 return n; 837 } 838 } 839 } 840 return NULL; 841 } 842 843 // Try moving a store out of a loop, right after the loop 844 void PhaseIdealLoop::try_move_store_after_loop(Node* n) { 845 if (n->is_Store() && n->in(0) != NULL) { 846 Node *n_ctrl = get_ctrl(n); 847 IdealLoopTree *n_loop = get_loop(n_ctrl); 848 // Store must be in a loop 849 if (n_loop != _ltree_root && !n_loop->_irreducible) { 850 Node* address = n->in(MemNode::Address); 851 Node* value = n->in(MemNode::ValueIn); 852 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 853 // address must be loop invariant 854 if (!n_loop->is_member(address_loop)) { 855 // Store must be last on this memory slice in the loop and 856 // nothing in the loop must observe it 857 Node* phi = NULL; 858 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 859 Node* u = n->fast_out(i); 860 if (has_ctrl(u)) { // control use? 861 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 862 if (!n_loop->is_member(u_loop)) { 863 continue; 864 } 865 if (u->is_Phi() && u->in(0) == n_loop->_head) { 866 assert(_igvn.type(u) == Type::MEMORY, "bad phi"); 867 // multiple phis on the same slice are possible 868 if (phi != NULL) { 869 return; 870 } 871 phi = u; 872 continue; 873 } 874 } 875 return; 876 } 877 if (phi != NULL) { 878 // Nothing in the loop before the store (next iteration) 879 // must observe the stored value 880 bool mem_ok = true; 881 { 882 ResourceMark rm; 883 Unique_Node_List wq; 884 wq.push(phi); 885 for (uint next = 0; next < wq.size() && mem_ok; ++next) { 886 Node *m = wq.at(next); 887 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) { 888 Node* u = m->fast_out(i); 889 if (u->is_Store() || u->is_Phi()) { 890 if (u != n) { 891 wq.push(u); 892 mem_ok = (wq.size() <= 10); 893 } 894 } else { 895 mem_ok = false; 896 break; 897 } 898 } 899 } 900 } 901 if (mem_ok) { 902 // Move the store out of the loop if the LCA of all 903 // users (except for the phi) is outside the loop. 904 Node* hook = new Node(1); 905 hook->init_req(0, n_ctrl); // Add an input to prevent hook from being dead 906 _igvn.rehash_node_delayed(phi); 907 int count = phi->replace_edge(n, hook); 908 assert(count > 0, "inconsistent phi"); 909 910 // Compute latest point this store can go 911 Node* lca = get_late_ctrl(n, get_ctrl(n)); 912 if (lca->is_OuterStripMinedLoop()) { 913 lca = lca->in(LoopNode::EntryControl); 914 } 915 if (n_loop->is_member(get_loop(lca))) { 916 // LCA is in the loop - bail out 917 _igvn.replace_node(hook, n); 918 return; 919 } 920 #ifdef ASSERT 921 if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) { 922 assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined"); 923 n_loop->_head->as_Loop()->verify_strip_mined(1); 924 Node* outer = n_loop->_head->as_CountedLoop()->outer_loop(); 925 IdealLoopTree* outer_loop = get_loop(outer); 926 assert(n_loop->_parent == outer_loop, "broken loop tree"); 927 assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state"); 928 } 929 #endif 930 931 // Move store out of the loop 932 _igvn.replace_node(hook, n->in(MemNode::Memory)); 933 _igvn.replace_input_of(n, 0, lca); 934 set_ctrl_and_loop(n, lca); 935 936 // Disconnect the phi now. An empty phi can confuse other 937 // optimizations in this pass of loop opts.. 938 if (phi->in(LoopNode::LoopBackControl) == phi) { 939 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl)); 940 n_loop->_body.yank(phi); 941 } 942 } 943 } 944 } 945 } 946 } 947 } 948 949 //------------------------------split_if_with_blocks_pre----------------------- 950 // Do the real work in a non-recursive function. Data nodes want to be 951 // cloned in the pre-order so they can feed each other nicely. 952 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) { 953 // Cloning these guys is unlikely to win 954 int n_op = n->Opcode(); 955 if( n_op == Op_MergeMem ) return n; 956 if( n->is_Proj() ) return n; 957 // Do not clone-up CmpFXXX variations, as these are always 958 // followed by a CmpI 959 if( n->is_Cmp() ) return n; 960 // Attempt to use a conditional move instead of a phi/branch 961 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) { 962 Node *cmov = conditional_move( n ); 963 if( cmov ) return cmov; 964 } 965 if( n->is_CFG() || n->is_LoadStore() ) 966 return n; 967 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd 968 n_op == Op_Opaque2 ) { 969 if( !C->major_progress() ) // If chance of no more loop opts... 970 _igvn._worklist.push(n); // maybe we'll remove them 971 return n; 972 } 973 974 if( n->is_Con() ) return n; // No cloning for Con nodes 975 976 Node *n_ctrl = get_ctrl(n); 977 if( !n_ctrl ) return n; // Dead node 978 979 Node* res = try_move_store_before_loop(n, n_ctrl); 980 if (res != NULL) { 981 return n; 982 } 983 984 // Attempt to remix address expressions for loop invariants 985 Node *m = remix_address_expressions( n ); 986 if( m ) return m; 987 988 if (n_op == Op_AddI) { 989 Node *nn = convert_add_to_muladd( n ); 990 if ( nn ) return nn; 991 } 992 993 if (n->is_ConstraintCast()) { 994 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this); 995 // ConstraintCastNode::dominating_cast() uses node control input to determine domination. 996 // Node control inputs don't necessarily agree with loop control info (due to 997 // transformations happened in between), thus additional dominance check is needed 998 // to keep loop info valid. 999 if (dom_cast != NULL && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) { 1000 _igvn.replace_node(n, dom_cast); 1001 return dom_cast; 1002 } 1003 } 1004 1005 // Determine if the Node has inputs from some local Phi. 1006 // Returns the block to clone thru. 1007 Node *n_blk = has_local_phi_input( n ); 1008 if( !n_blk ) return n; 1009 1010 // Do not clone the trip counter through on a CountedLoop 1011 // (messes up the canonical shape). 1012 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n; 1013 1014 // Check for having no control input; not pinned. Allow 1015 // dominating control. 1016 if (n->in(0)) { 1017 Node *dom = idom(n_blk); 1018 if (dom_lca(n->in(0), dom) != n->in(0)) { 1019 return n; 1020 } 1021 } 1022 // Policy: when is it profitable. You must get more wins than 1023 // policy before it is considered profitable. Policy is usually 0, 1024 // so 1 win is considered profitable. Big merges will require big 1025 // cloning, so get a larger policy. 1026 int policy = n_blk->req() >> 2; 1027 1028 // If the loop is a candidate for range check elimination, 1029 // delay splitting through it's phi until a later loop optimization 1030 if (n_blk->is_CountedLoop()) { 1031 IdealLoopTree *lp = get_loop(n_blk); 1032 if (lp && lp->_rce_candidate) { 1033 return n; 1034 } 1035 } 1036 1037 if (must_throttle_split_if()) return n; 1038 1039 // Split 'n' through the merge point if it is profitable 1040 Node *phi = split_thru_phi( n, n_blk, policy ); 1041 if (!phi) return n; 1042 1043 // Found a Phi to split thru! 1044 // Replace 'n' with the new phi 1045 _igvn.replace_node( n, phi ); 1046 // Moved a load around the loop, 'en-registering' something. 1047 if (n_blk->is_Loop() && n->is_Load() && 1048 !phi->in(LoopNode::LoopBackControl)->is_Load()) 1049 C->set_major_progress(); 1050 1051 return phi; 1052 } 1053 1054 static bool merge_point_too_heavy(Compile* C, Node* region) { 1055 // Bail out if the region and its phis have too many users. 1056 int weight = 0; 1057 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1058 weight += region->fast_out(i)->outcnt(); 1059 } 1060 int nodes_left = C->max_node_limit() - C->live_nodes(); 1061 if (weight * 8 > nodes_left) { 1062 if (PrintOpto) { 1063 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight); 1064 } 1065 return true; 1066 } else { 1067 return false; 1068 } 1069 } 1070 1071 static bool merge_point_safe(Node* region) { 1072 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode 1073 // having a PhiNode input. This sidesteps the dangerous case where the split 1074 // ConvI2LNode may become TOP if the input Value() does not 1075 // overlap the ConvI2L range, leaving a node which may not dominate its 1076 // uses. 1077 // A better fix for this problem can be found in the BugTraq entry, but 1078 // expediency for Mantis demands this hack. 1079 #ifdef _LP64 1080 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1081 Node* n = region->fast_out(i); 1082 if (n->is_Phi()) { 1083 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1084 Node* m = n->fast_out(j); 1085 if (m->Opcode() == Op_ConvI2L) 1086 return false; 1087 if (m->is_CastII() && m->isa_CastII()->has_range_check()) { 1088 return false; 1089 } 1090 } 1091 } 1092 } 1093 #endif 1094 return true; 1095 } 1096 1097 1098 //------------------------------place_near_use--------------------------------- 1099 // Place some computation next to use but not inside inner loops. 1100 // For inner loop uses move it to the preheader area. 1101 Node *PhaseIdealLoop::place_near_use(Node *useblock) const { 1102 IdealLoopTree *u_loop = get_loop( useblock ); 1103 if (u_loop->_irreducible) { 1104 return useblock; 1105 } 1106 if (u_loop->_child) { 1107 if (useblock == u_loop->_head && u_loop->_head->is_OuterStripMinedLoop()) { 1108 return u_loop->_head->in(LoopNode::EntryControl); 1109 } 1110 return useblock; 1111 } 1112 return u_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl); 1113 } 1114 1115 1116 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) { 1117 if (!n->is_If() || n->is_CountedLoopEnd()) { 1118 return false; 1119 } 1120 if (!n->in(0)->is_Region()) { 1121 return false; 1122 } 1123 Node* region = n->in(0); 1124 Node* dom = idom(region); 1125 if (!dom->is_If() || dom->in(1) != n->in(1)) { 1126 return false; 1127 } 1128 IfNode* dom_if = dom->as_If(); 1129 Node* proj_true = dom_if->proj_out(1); 1130 Node* proj_false = dom_if->proj_out(0); 1131 1132 for (uint i = 1; i < region->req(); i++) { 1133 if (is_dominator(proj_true, region->in(i))) { 1134 continue; 1135 } 1136 if (is_dominator(proj_false, region->in(i))) { 1137 continue; 1138 } 1139 return false; 1140 } 1141 1142 return true; 1143 } 1144 1145 1146 bool PhaseIdealLoop::can_split_if(Node* n_ctrl) { 1147 if (must_throttle_split_if()) { 1148 return false; 1149 } 1150 1151 // Do not do 'split-if' if irreducible loops are present. 1152 if (_has_irreducible_loops) { 1153 return false; 1154 } 1155 1156 if (merge_point_too_heavy(C, n_ctrl)) { 1157 return false; 1158 } 1159 1160 // Do not do 'split-if' if some paths are dead. First do dead code 1161 // elimination and then see if its still profitable. 1162 for (uint i = 1; i < n_ctrl->req(); i++) { 1163 if (n_ctrl->in(i) == C->top()) { 1164 return false; 1165 } 1166 } 1167 1168 // If trying to do a 'Split-If' at the loop head, it is only 1169 // profitable if the cmp folds up on BOTH paths. Otherwise we 1170 // risk peeling a loop forever. 1171 1172 // CNC - Disabled for now. Requires careful handling of loop 1173 // body selection for the cloned code. Also, make sure we check 1174 // for any input path not being in the same loop as n_ctrl. For 1175 // irreducible loops we cannot check for 'n_ctrl->is_Loop()' 1176 // because the alternative loop entry points won't be converted 1177 // into LoopNodes. 1178 IdealLoopTree *n_loop = get_loop(n_ctrl); 1179 for (uint j = 1; j < n_ctrl->req(); j++) { 1180 if (get_loop(n_ctrl->in(j)) != n_loop) { 1181 return false; 1182 } 1183 } 1184 1185 // Check for safety of the merge point. 1186 if (!merge_point_safe(n_ctrl)) { 1187 return false; 1188 } 1189 1190 return true; 1191 } 1192 1193 // Detect if the node is the inner strip-mined loop 1194 // Return: NULL if it's not the case, or the exit of outer strip-mined loop 1195 static Node* is_inner_of_stripmined_loop(const Node* out) { 1196 Node* out_le = NULL; 1197 1198 if (out->is_CountedLoopEnd()) { 1199 const CountedLoopNode* loop = out->as_CountedLoopEnd()->loopnode(); 1200 1201 if (loop != NULL && loop->is_strip_mined()) { 1202 out_le = loop->in(LoopNode::EntryControl)->as_OuterStripMinedLoop()->outer_loop_exit(); 1203 } 1204 } 1205 1206 return out_le; 1207 } 1208 1209 //------------------------------split_if_with_blocks_post---------------------- 1210 // Do the real work in a non-recursive function. CFG hackery wants to be 1211 // in the post-order, so it can dirty the I-DOM info and not use the dirtied 1212 // info. 1213 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) { 1214 1215 // Cloning Cmp through Phi's involves the split-if transform. 1216 // FastLock is not used by an If 1217 if (n->is_Cmp() && !n->is_FastLock()) { 1218 Node *n_ctrl = get_ctrl(n); 1219 // Determine if the Node has inputs from some local Phi. 1220 // Returns the block to clone thru. 1221 Node *n_blk = has_local_phi_input(n); 1222 if (n_blk != n_ctrl) { 1223 return; 1224 } 1225 1226 if (!can_split_if(n_ctrl)) { 1227 return; 1228 } 1229 1230 if (n->outcnt() != 1) { 1231 return; // Multiple bool's from 1 compare? 1232 } 1233 Node *bol = n->unique_out(); 1234 assert(bol->is_Bool(), "expect a bool here"); 1235 if (bol->outcnt() != 1) { 1236 return;// Multiple branches from 1 compare? 1237 } 1238 Node *iff = bol->unique_out(); 1239 1240 // Check some safety conditions 1241 if (iff->is_If()) { // Classic split-if? 1242 if (iff->in(0) != n_ctrl) { 1243 return; // Compare must be in same blk as if 1244 } 1245 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE 1246 // Can't split CMove with different control edge. 1247 if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) { 1248 return; 1249 } 1250 if (get_ctrl(iff->in(2)) == n_ctrl || 1251 get_ctrl(iff->in(3)) == n_ctrl) { 1252 return; // Inputs not yet split-up 1253 } 1254 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) { 1255 return; // Loop-invar test gates loop-varying CMOVE 1256 } 1257 } else { 1258 return; // some other kind of node, such as an Allocate 1259 } 1260 1261 // When is split-if profitable? Every 'win' on means some control flow 1262 // goes dead, so it's almost always a win. 1263 int policy = 0; 1264 // Split compare 'n' through the merge point if it is profitable 1265 Node *phi = split_thru_phi( n, n_ctrl, policy); 1266 if (!phi) { 1267 return; 1268 } 1269 1270 // Found a Phi to split thru! 1271 // Replace 'n' with the new phi 1272 _igvn.replace_node(n, phi); 1273 1274 // Now split the bool up thru the phi 1275 Node *bolphi = split_thru_phi(bol, n_ctrl, -1); 1276 guarantee(bolphi != NULL, "null boolean phi node"); 1277 1278 _igvn.replace_node(bol, bolphi); 1279 assert(iff->in(1) == bolphi, ""); 1280 1281 if (bolphi->Value(&_igvn)->singleton()) { 1282 return; 1283 } 1284 1285 // Conditional-move? Must split up now 1286 if (!iff->is_If()) { 1287 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1); 1288 _igvn.replace_node(iff, cmovphi); 1289 return; 1290 } 1291 1292 // Now split the IF 1293 do_split_if(iff); 1294 return; 1295 } 1296 1297 // Two identical ifs back to back can be merged 1298 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) { 1299 Node *n_ctrl = n->in(0); 1300 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1)); 1301 IfNode* dom_if = idom(n_ctrl)->as_If(); 1302 Node* proj_true = dom_if->proj_out(1); 1303 Node* proj_false = dom_if->proj_out(0); 1304 Node* con_true = _igvn.makecon(TypeInt::ONE); 1305 Node* con_false = _igvn.makecon(TypeInt::ZERO); 1306 1307 for (uint i = 1; i < n_ctrl->req(); i++) { 1308 if (is_dominator(proj_true, n_ctrl->in(i))) { 1309 bolphi->init_req(i, con_true); 1310 } else { 1311 assert(is_dominator(proj_false, n_ctrl->in(i)), "bad if"); 1312 bolphi->init_req(i, con_false); 1313 } 1314 } 1315 register_new_node(bolphi, n_ctrl); 1316 _igvn.replace_input_of(n, 1, bolphi); 1317 1318 // Now split the IF 1319 do_split_if(n); 1320 return; 1321 } 1322 1323 // Check for an IF ready to split; one that has its 1324 // condition codes input coming from a Phi at the block start. 1325 int n_op = n->Opcode(); 1326 1327 // Check for an IF being dominated by another IF same test 1328 if (n_op == Op_If || 1329 n_op == Op_RangeCheck) { 1330 Node *bol = n->in(1); 1331 uint max = bol->outcnt(); 1332 // Check for same test used more than once? 1333 if (max > 1 && bol->is_Bool()) { 1334 // Search up IDOMs to see if this IF is dominated. 1335 Node *cutoff = get_ctrl(bol); 1336 1337 // Now search up IDOMs till cutoff, looking for a dominating test 1338 Node *prevdom = n; 1339 Node *dom = idom(prevdom); 1340 while (dom != cutoff) { 1341 if (dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom) { 1342 // It's invalid to move control dependent data nodes in the inner 1343 // strip-mined loop, because: 1344 // 1) break validation of LoopNode::verify_strip_mined() 1345 // 2) move code with side-effect in strip-mined loop 1346 // Move to the exit of outer strip-mined loop in that case. 1347 Node* out_le = is_inner_of_stripmined_loop(dom); 1348 if (out_le != NULL) { 1349 prevdom = out_le; 1350 } 1351 // Replace the dominated test with an obvious true or false. 1352 // Place it on the IGVN worklist for later cleanup. 1353 C->set_major_progress(); 1354 dominated_by(prevdom, n, false, true); 1355 #ifndef PRODUCT 1356 if( VerifyLoopOptimizations ) verify(); 1357 #endif 1358 return; 1359 } 1360 prevdom = dom; 1361 dom = idom(prevdom); 1362 } 1363 } 1364 } 1365 1366 // See if a shared loop-varying computation has no loop-varying uses. 1367 // Happens if something is only used for JVM state in uncommon trap exits, 1368 // like various versions of induction variable+offset. Clone the 1369 // computation per usage to allow it to sink out of the loop. 1370 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about) 1371 Node *n_ctrl = get_ctrl(n); 1372 IdealLoopTree *n_loop = get_loop(n_ctrl); 1373 if( n_loop != _ltree_root ) { 1374 DUIterator_Fast imax, i = n->fast_outs(imax); 1375 for (; i < imax; i++) { 1376 Node* u = n->fast_out(i); 1377 if( !has_ctrl(u) ) break; // Found control user 1378 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 1379 if( u_loop == n_loop ) break; // Found loop-varying use 1380 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop 1381 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003 1382 } 1383 bool did_break = (i < imax); // Did we break out of the previous loop? 1384 if (!did_break && n->outcnt() > 1) { // All uses in outer loops! 1385 Node *late_load_ctrl = NULL; 1386 if (n->is_Load()) { 1387 // If n is a load, get and save the result from get_late_ctrl(), 1388 // to be later used in calculating the control for n's clones. 1389 clear_dom_lca_tags(); 1390 late_load_ctrl = get_late_ctrl(n, n_ctrl); 1391 } 1392 // If n is a load, and the late control is the same as the current 1393 // control, then the cloning of n is a pointless exercise, because 1394 // GVN will ensure that we end up where we started. 1395 if (!n->is_Load() || late_load_ctrl != n_ctrl) { 1396 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) { 1397 Node *u = n->last_out(j); // Clone private computation per use 1398 _igvn.rehash_node_delayed(u); 1399 Node *x = n->clone(); // Clone computation 1400 Node *x_ctrl = NULL; 1401 if( u->is_Phi() ) { 1402 // Replace all uses of normal nodes. Replace Phi uses 1403 // individually, so the separate Nodes can sink down 1404 // different paths. 1405 uint k = 1; 1406 while( u->in(k) != n ) k++; 1407 u->set_req( k, x ); 1408 // x goes next to Phi input path 1409 x_ctrl = u->in(0)->in(k); 1410 --j; 1411 } else { // Normal use 1412 // Replace all uses 1413 for( uint k = 0; k < u->req(); k++ ) { 1414 if( u->in(k) == n ) { 1415 u->set_req( k, x ); 1416 --j; 1417 } 1418 } 1419 x_ctrl = get_ctrl(u); 1420 } 1421 1422 // Find control for 'x' next to use but not inside inner loops. 1423 // For inner loop uses get the preheader area. 1424 x_ctrl = place_near_use(x_ctrl); 1425 1426 if (n->is_Load()) { 1427 // For loads, add a control edge to a CFG node outside of the loop 1428 // to force them to not combine and return back inside the loop 1429 // during GVN optimization (4641526). 1430 // 1431 // Because we are setting the actual control input, factor in 1432 // the result from get_late_ctrl() so we respect any 1433 // anti-dependences. (6233005). 1434 x_ctrl = dom_lca(late_load_ctrl, x_ctrl); 1435 1436 // Don't allow the control input to be a CFG splitting node. 1437 // Such nodes should only have ProjNodes as outs, e.g. IfNode 1438 // should only have IfTrueNode and IfFalseNode (4985384). 1439 x_ctrl = find_non_split_ctrl(x_ctrl); 1440 1441 IdealLoopTree* x_loop = get_loop(x_ctrl); 1442 Node* x_head = x_loop->_head; 1443 if (x_head->is_Loop() && (x_head->is_OuterStripMinedLoop() || x_head->as_Loop()->is_strip_mined()) && is_dominator(n_ctrl, x_head)) { 1444 // Anti dependence analysis is sometimes too 1445 // conservative: a store in the outer strip mined loop 1446 // can prevent a load from floating out of the outer 1447 // strip mined loop but the load may not be referenced 1448 // from the safepoint: loop strip mining verification 1449 // code reports a problem in that case. Make sure the 1450 // load is not moved in the outer strip mined loop in 1451 // that case. 1452 x_ctrl = x_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl); 1453 } 1454 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone"); 1455 1456 x->set_req(0, x_ctrl); 1457 } 1458 register_new_node(x, x_ctrl); 1459 1460 // Some institutional knowledge is needed here: 'x' is 1461 // yanked because if the optimizer runs GVN on it all the 1462 // cloned x's will common up and undo this optimization and 1463 // be forced back in the loop. 1464 // I tried setting control edges on the x's to force them to 1465 // not combine, but the matching gets worried when it tries 1466 // to fold a StoreP and an AddP together (as part of an 1467 // address expression) and the AddP and StoreP have 1468 // different controls. 1469 if (!x->is_Load() && !x->is_DecodeNarrowPtr()) _igvn._worklist.yank(x); 1470 } 1471 _igvn.remove_dead_node(n); 1472 } 1473 } 1474 } 1475 } 1476 1477 try_move_store_after_loop(n); 1478 1479 // Check for Opaque2's who's loop has disappeared - who's input is in the 1480 // same loop nest as their output. Remove 'em, they are no longer useful. 1481 if( n_op == Op_Opaque2 && 1482 n->in(1) != NULL && 1483 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) { 1484 _igvn.replace_node( n, n->in(1) ); 1485 } 1486 } 1487 1488 //------------------------------split_if_with_blocks--------------------------- 1489 // Check for aggressive application of 'split-if' optimization, 1490 // using basic block level info. 1491 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack) { 1492 Node* root = C->root(); 1493 visited.set(root->_idx); // first, mark root as visited 1494 // Do pre-visit work for root 1495 Node* n = split_if_with_blocks_pre(root); 1496 uint cnt = n->outcnt(); 1497 uint i = 0; 1498 1499 while (true) { 1500 // Visit all children 1501 if (i < cnt) { 1502 Node* use = n->raw_out(i); 1503 ++i; 1504 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) { 1505 // Now do pre-visit work for this use 1506 use = split_if_with_blocks_pre(use); 1507 nstack.push(n, i); // Save parent and next use's index. 1508 n = use; // Process all children of current use. 1509 cnt = use->outcnt(); 1510 i = 0; 1511 } 1512 } 1513 else { 1514 // All of n's children have been processed, complete post-processing. 1515 if (cnt != 0 && !n->is_Con()) { 1516 assert(has_node(n), "no dead nodes"); 1517 split_if_with_blocks_post(n); 1518 } 1519 if (must_throttle_split_if()) { 1520 nstack.clear(); 1521 } 1522 if (nstack.is_empty()) { 1523 // Finished all nodes on stack. 1524 break; 1525 } 1526 // Get saved parent node and next use's index. Visit the rest of uses. 1527 n = nstack.node(); 1528 cnt = n->outcnt(); 1529 i = nstack.index(); 1530 nstack.pop(); 1531 } 1532 } 1533 } 1534 1535 1536 //============================================================================= 1537 // 1538 // C L O N E A L O O P B O D Y 1539 // 1540 1541 //------------------------------clone_iff-------------------------------------- 1542 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1543 // "Nearly" because all Nodes have been cloned from the original in the loop, 1544 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1545 // through the Phi recursively, and return a Bool. 1546 Node* PhaseIdealLoop::clone_iff(PhiNode *phi, IdealLoopTree *loop) { 1547 1548 // Convert this Phi into a Phi merging Bools 1549 uint i; 1550 for (i = 1; i < phi->req(); i++) { 1551 Node *b = phi->in(i); 1552 if (b->is_Phi()) { 1553 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi(), loop)); 1554 } else { 1555 assert(b->is_Bool() || b->Opcode() == Op_Opaque4, ""); 1556 } 1557 } 1558 1559 Node* n = phi->in(1); 1560 Node* sample_opaque = NULL; 1561 Node *sample_bool = NULL; 1562 if (n->Opcode() == Op_Opaque4) { 1563 sample_opaque = n; 1564 sample_bool = n->in(1); 1565 assert(sample_bool->is_Bool(), "wrong type"); 1566 } else { 1567 sample_bool = n; 1568 } 1569 Node *sample_cmp = sample_bool->in(1); 1570 1571 // Make Phis to merge the Cmp's inputs. 1572 PhiNode *phi1 = new PhiNode(phi->in(0), Type::TOP); 1573 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP); 1574 for (i = 1; i < phi->req(); i++) { 1575 Node *n1 = sample_opaque == NULL ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1); 1576 Node *n2 = sample_opaque == NULL ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2); 1577 phi1->set_req(i, n1); 1578 phi2->set_req(i, n2); 1579 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 1580 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 1581 } 1582 // See if these Phis have been made before. 1583 // Register with optimizer 1584 Node *hit1 = _igvn.hash_find_insert(phi1); 1585 if (hit1) { // Hit, toss just made Phi 1586 _igvn.remove_dead_node(phi1); // Remove new phi 1587 assert(hit1->is_Phi(), "" ); 1588 phi1 = (PhiNode*)hit1; // Use existing phi 1589 } else { // Miss 1590 _igvn.register_new_node_with_optimizer(phi1); 1591 } 1592 Node *hit2 = _igvn.hash_find_insert(phi2); 1593 if (hit2) { // Hit, toss just made Phi 1594 _igvn.remove_dead_node(phi2); // Remove new phi 1595 assert(hit2->is_Phi(), "" ); 1596 phi2 = (PhiNode*)hit2; // Use existing phi 1597 } else { // Miss 1598 _igvn.register_new_node_with_optimizer(phi2); 1599 } 1600 // Register Phis with loop/block info 1601 set_ctrl(phi1, phi->in(0)); 1602 set_ctrl(phi2, phi->in(0)); 1603 // Make a new Cmp 1604 Node *cmp = sample_cmp->clone(); 1605 cmp->set_req(1, phi1); 1606 cmp->set_req(2, phi2); 1607 _igvn.register_new_node_with_optimizer(cmp); 1608 set_ctrl(cmp, phi->in(0)); 1609 1610 // Make a new Bool 1611 Node *b = sample_bool->clone(); 1612 b->set_req(1,cmp); 1613 _igvn.register_new_node_with_optimizer(b); 1614 set_ctrl(b, phi->in(0)); 1615 1616 if (sample_opaque != NULL) { 1617 Node* opaque = sample_opaque->clone(); 1618 opaque->set_req(1, b); 1619 _igvn.register_new_node_with_optimizer(opaque); 1620 set_ctrl(opaque, phi->in(0)); 1621 return opaque; 1622 } 1623 1624 assert(b->is_Bool(), ""); 1625 return b; 1626 } 1627 1628 //------------------------------clone_bool------------------------------------- 1629 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1630 // "Nearly" because all Nodes have been cloned from the original in the loop, 1631 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1632 // through the Phi recursively, and return a Bool. 1633 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) { 1634 uint i; 1635 // Convert this Phi into a Phi merging Bools 1636 for( i = 1; i < phi->req(); i++ ) { 1637 Node *b = phi->in(i); 1638 if( b->is_Phi() ) { 1639 _igvn.replace_input_of(phi, i, clone_bool( b->as_Phi(), loop )); 1640 } else { 1641 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" ); 1642 } 1643 } 1644 1645 Node *sample_cmp = phi->in(1); 1646 1647 // Make Phis to merge the Cmp's inputs. 1648 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP ); 1649 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP ); 1650 for( uint j = 1; j < phi->req(); j++ ) { 1651 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP 1652 Node *n1, *n2; 1653 if( cmp_top->is_Cmp() ) { 1654 n1 = cmp_top->in(1); 1655 n2 = cmp_top->in(2); 1656 } else { 1657 n1 = n2 = cmp_top; 1658 } 1659 phi1->set_req( j, n1 ); 1660 phi2->set_req( j, n2 ); 1661 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 1662 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 1663 } 1664 1665 // See if these Phis have been made before. 1666 // Register with optimizer 1667 Node *hit1 = _igvn.hash_find_insert(phi1); 1668 if( hit1 ) { // Hit, toss just made Phi 1669 _igvn.remove_dead_node(phi1); // Remove new phi 1670 assert( hit1->is_Phi(), "" ); 1671 phi1 = (PhiNode*)hit1; // Use existing phi 1672 } else { // Miss 1673 _igvn.register_new_node_with_optimizer(phi1); 1674 } 1675 Node *hit2 = _igvn.hash_find_insert(phi2); 1676 if( hit2 ) { // Hit, toss just made Phi 1677 _igvn.remove_dead_node(phi2); // Remove new phi 1678 assert( hit2->is_Phi(), "" ); 1679 phi2 = (PhiNode*)hit2; // Use existing phi 1680 } else { // Miss 1681 _igvn.register_new_node_with_optimizer(phi2); 1682 } 1683 // Register Phis with loop/block info 1684 set_ctrl(phi1, phi->in(0)); 1685 set_ctrl(phi2, phi->in(0)); 1686 // Make a new Cmp 1687 Node *cmp = sample_cmp->clone(); 1688 cmp->set_req( 1, phi1 ); 1689 cmp->set_req( 2, phi2 ); 1690 _igvn.register_new_node_with_optimizer(cmp); 1691 set_ctrl(cmp, phi->in(0)); 1692 1693 assert( cmp->is_Cmp(), "" ); 1694 return (CmpNode*)cmp; 1695 } 1696 1697 //------------------------------sink_use--------------------------------------- 1698 // If 'use' was in the loop-exit block, it now needs to be sunk 1699 // below the post-loop merge point. 1700 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) { 1701 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) { 1702 set_ctrl(use, post_loop); 1703 for (DUIterator j = use->outs(); use->has_out(j); j++) 1704 sink_use(use->out(j), post_loop); 1705 } 1706 } 1707 1708 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new, 1709 IdealLoopTree* loop, IdealLoopTree* outer_loop, 1710 Node_List*& split_if_set, Node_List*& split_bool_set, 1711 Node_List*& split_cex_set, Node_List& worklist, 1712 uint new_counter, CloneLoopMode mode) { 1713 Node* nnn = old_new[old->_idx]; 1714 // Copy uses to a worklist, so I can munge the def-use info 1715 // with impunity. 1716 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1717 worklist.push(old->fast_out(j)); 1718 1719 while( worklist.size() ) { 1720 Node *use = worklist.pop(); 1721 if (!has_node(use)) continue; // Ignore dead nodes 1722 if (use->in(0) == C->top()) continue; 1723 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1724 // Check for data-use outside of loop - at least one of OLD or USE 1725 // must not be a CFG node. 1726 #ifdef ASSERT 1727 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == NULL) { 1728 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint(); 1729 assert(mode != IgnoreStripMined, "incorrect cloning mode"); 1730 assert((mode == ControlAroundStripMined && use == sfpt) || !use->is_reachable_from_root(), "missed a node"); 1731 } 1732 #endif 1733 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) { 1734 1735 // If the Data use is an IF, that means we have an IF outside of the 1736 // loop that is switching on a condition that is set inside of the 1737 // loop. Happens if people set a loop-exit flag; then test the flag 1738 // in the loop to break the loop, then test is again outside of the 1739 // loop to determine which way the loop exited. 1740 // Loop predicate If node connects to Bool node through Opaque1 node. 1741 if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use) || use->Opcode() == Op_Opaque4) { 1742 // Since this code is highly unlikely, we lazily build the worklist 1743 // of such Nodes to go split. 1744 if (!split_if_set) { 1745 split_if_set = new Node_List(); 1746 } 1747 split_if_set->push(use); 1748 } 1749 if (use->is_Bool()) { 1750 if (!split_bool_set) { 1751 split_bool_set = new Node_List(); 1752 } 1753 split_bool_set->push(use); 1754 } 1755 if (use->Opcode() == Op_CreateEx) { 1756 if (!split_cex_set) { 1757 split_cex_set = new Node_List(); 1758 } 1759 split_cex_set->push(use); 1760 } 1761 1762 1763 // Get "block" use is in 1764 uint idx = 0; 1765 while( use->in(idx) != old ) idx++; 1766 Node *prev = use->is_CFG() ? use : get_ctrl(use); 1767 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" ); 1768 Node *cfg = prev->_idx >= new_counter 1769 ? prev->in(2) 1770 : idom(prev); 1771 if( use->is_Phi() ) // Phi use is in prior block 1772 cfg = prev->in(idx); // NOT in block of Phi itself 1773 if (cfg->is_top()) { // Use is dead? 1774 _igvn.replace_input_of(use, idx, C->top()); 1775 continue; 1776 } 1777 1778 // If use is referenced through control edge... (idx == 0) 1779 if (mode == IgnoreStripMined && idx == 0) { 1780 LoopNode *head = loop->_head->as_Loop(); 1781 if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) { 1782 // That node is outside the inner loop, leave it outside the 1783 // outer loop as well to not confuse verification code. 1784 assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop"); 1785 _igvn.replace_input_of(use, 0, head->outer_loop_exit()); 1786 continue; 1787 } 1788 } 1789 1790 while(!outer_loop->is_member(get_loop(cfg))) { 1791 prev = cfg; 1792 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg); 1793 } 1794 // If the use occurs after merging several exits from the loop, then 1795 // old value must have dominated all those exits. Since the same old 1796 // value was used on all those exits we did not need a Phi at this 1797 // merge point. NOW we do need a Phi here. Each loop exit value 1798 // is now merged with the peeled body exit; each exit gets its own 1799 // private Phi and those Phis need to be merged here. 1800 Node *phi; 1801 if( prev->is_Region() ) { 1802 if( idx == 0 ) { // Updating control edge? 1803 phi = prev; // Just use existing control 1804 } else { // Else need a new Phi 1805 phi = PhiNode::make( prev, old ); 1806 // Now recursively fix up the new uses of old! 1807 for( uint i = 1; i < prev->req(); i++ ) { 1808 worklist.push(phi); // Onto worklist once for each 'old' input 1809 } 1810 } 1811 } else { 1812 // Get new RegionNode merging old and new loop exits 1813 prev = old_new[prev->_idx]; 1814 assert( prev, "just made this in step 7" ); 1815 if( idx == 0) { // Updating control edge? 1816 phi = prev; // Just use existing control 1817 } else { // Else need a new Phi 1818 // Make a new Phi merging data values properly 1819 phi = PhiNode::make( prev, old ); 1820 phi->set_req( 1, nnn ); 1821 } 1822 } 1823 // If inserting a new Phi, check for prior hits 1824 if( idx != 0 ) { 1825 Node *hit = _igvn.hash_find_insert(phi); 1826 if( hit == NULL ) { 1827 _igvn.register_new_node_with_optimizer(phi); // Register new phi 1828 } else { // or 1829 // Remove the new phi from the graph and use the hit 1830 _igvn.remove_dead_node(phi); 1831 phi = hit; // Use existing phi 1832 } 1833 set_ctrl(phi, prev); 1834 } 1835 // Make 'use' use the Phi instead of the old loop body exit value 1836 _igvn.replace_input_of(use, idx, phi); 1837 if( use->_idx >= new_counter ) { // If updating new phis 1838 // Not needed for correctness, but prevents a weak assert 1839 // in AddPNode from tripping (when we end up with different 1840 // base & derived Phis that will become the same after 1841 // IGVN does CSE). 1842 Node *hit = _igvn.hash_find_insert(use); 1843 if( hit ) // Go ahead and re-hash for hits. 1844 _igvn.replace_node( use, hit ); 1845 } 1846 1847 // If 'use' was in the loop-exit block, it now needs to be sunk 1848 // below the post-loop merge point. 1849 sink_use( use, prev ); 1850 } 1851 } 1852 } 1853 1854 static void clone_outer_loop_helper(Node* n, const IdealLoopTree *loop, const IdealLoopTree* outer_loop, 1855 const Node_List &old_new, Unique_Node_List& wq, PhaseIdealLoop* phase, 1856 bool check_old_new) { 1857 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1858 Node* u = n->fast_out(j); 1859 assert(check_old_new || old_new[u->_idx] == NULL, "shouldn't have been cloned"); 1860 if (!u->is_CFG() && (!check_old_new || old_new[u->_idx] == NULL)) { 1861 Node* c = phase->get_ctrl(u); 1862 IdealLoopTree* u_loop = phase->get_loop(c); 1863 assert(!loop->is_member(u_loop), "can be in outer loop or out of both loops only"); 1864 if (outer_loop->is_member(u_loop)) { 1865 wq.push(u); 1866 } 1867 } 1868 } 1869 } 1870 1871 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop, 1872 IdealLoopTree* outer_loop, int dd, Node_List &old_new, 1873 Node_List& extra_data_nodes) { 1874 if (head->is_strip_mined() && mode != IgnoreStripMined) { 1875 CountedLoopNode* cl = head->as_CountedLoop(); 1876 Node* l = cl->outer_loop(); 1877 Node* tail = cl->outer_loop_tail(); 1878 IfNode* le = cl->outer_loop_end(); 1879 Node* sfpt = cl->outer_safepoint(); 1880 CountedLoopEndNode* cle = cl->loopexit(); 1881 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop(); 1882 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null(); 1883 Node* cle_out = cle->proj_out(false); 1884 1885 Node* new_sfpt = NULL; 1886 Node* new_cle_out = cle_out->clone(); 1887 old_new.map(cle_out->_idx, new_cle_out); 1888 if (mode == CloneIncludesStripMined) { 1889 // clone outer loop body 1890 Node* new_l = l->clone(); 1891 Node* new_tail = tail->clone(); 1892 IfNode* new_le = le->clone()->as_If(); 1893 new_sfpt = sfpt->clone(); 1894 1895 set_loop(new_l, outer_loop->_parent); 1896 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd); 1897 set_loop(new_cle_out, outer_loop->_parent); 1898 set_idom(new_cle_out, new_cle, dd); 1899 set_loop(new_sfpt, outer_loop->_parent); 1900 set_idom(new_sfpt, new_cle_out, dd); 1901 set_loop(new_le, outer_loop->_parent); 1902 set_idom(new_le, new_sfpt, dd); 1903 set_loop(new_tail, outer_loop->_parent); 1904 set_idom(new_tail, new_le, dd); 1905 set_idom(new_cl, new_l, dd); 1906 1907 old_new.map(l->_idx, new_l); 1908 old_new.map(tail->_idx, new_tail); 1909 old_new.map(le->_idx, new_le); 1910 old_new.map(sfpt->_idx, new_sfpt); 1911 1912 new_l->set_req(LoopNode::LoopBackControl, new_tail); 1913 new_l->set_req(0, new_l); 1914 new_tail->set_req(0, new_le); 1915 new_le->set_req(0, new_sfpt); 1916 new_sfpt->set_req(0, new_cle_out); 1917 new_cle_out->set_req(0, new_cle); 1918 new_cl->set_req(LoopNode::EntryControl, new_l); 1919 1920 _igvn.register_new_node_with_optimizer(new_l); 1921 _igvn.register_new_node_with_optimizer(new_tail); 1922 _igvn.register_new_node_with_optimizer(new_le); 1923 } else { 1924 Node *newhead = old_new[loop->_head->_idx]; 1925 newhead->as_Loop()->clear_strip_mined(); 1926 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl)); 1927 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 1928 } 1929 // Look at data node that were assigned a control in the outer 1930 // loop: they are kept in the outer loop by the safepoint so start 1931 // from the safepoint node's inputs. 1932 IdealLoopTree* outer_loop = get_loop(l); 1933 Node_Stack stack(2); 1934 stack.push(sfpt, 1); 1935 uint new_counter = C->unique(); 1936 while (stack.size() > 0) { 1937 Node* n = stack.node(); 1938 uint i = stack.index(); 1939 while (i < n->req() && 1940 (n->in(i) == NULL || 1941 !has_ctrl(n->in(i)) || 1942 get_loop(get_ctrl(n->in(i))) != outer_loop || 1943 (old_new[n->in(i)->_idx] != NULL && old_new[n->in(i)->_idx]->_idx >= new_counter))) { 1944 i++; 1945 } 1946 if (i < n->req()) { 1947 stack.set_index(i+1); 1948 stack.push(n->in(i), 0); 1949 } else { 1950 assert(old_new[n->_idx] == NULL || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet"); 1951 Node* m = n == sfpt ? new_sfpt : n->clone(); 1952 if (m != NULL) { 1953 for (uint i = 0; i < n->req(); i++) { 1954 if (m->in(i) != NULL && old_new[m->in(i)->_idx] != NULL) { 1955 m->set_req(i, old_new[m->in(i)->_idx]); 1956 } 1957 } 1958 } else { 1959 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?"); 1960 } 1961 if (n != sfpt) { 1962 extra_data_nodes.push(n); 1963 _igvn.register_new_node_with_optimizer(m); 1964 assert(get_ctrl(n) == cle_out, "what other control?"); 1965 set_ctrl(m, new_cle_out); 1966 old_new.map(n->_idx, m); 1967 } 1968 stack.pop(); 1969 } 1970 } 1971 if (mode == CloneIncludesStripMined) { 1972 _igvn.register_new_node_with_optimizer(new_sfpt); 1973 _igvn.register_new_node_with_optimizer(new_cle_out); 1974 } 1975 // Some other transformation may have pessimistically assign some 1976 // data nodes to the outer loop. Set their control so they are out 1977 // of the outer loop. 1978 ResourceMark rm; 1979 Unique_Node_List wq; 1980 for (uint i = 0; i < extra_data_nodes.size(); i++) { 1981 Node* old = extra_data_nodes.at(i); 1982 clone_outer_loop_helper(old, loop, outer_loop, old_new, wq, this, true); 1983 } 1984 Node* new_ctrl = cl->outer_loop_exit(); 1985 assert(get_loop(new_ctrl) != outer_loop, "must be out of the loop nest"); 1986 for (uint i = 0; i < wq.size(); i++) { 1987 Node* n = wq.at(i); 1988 set_ctrl(n, new_ctrl); 1989 clone_outer_loop_helper(n, loop, outer_loop, old_new, wq, this, false); 1990 } 1991 } else { 1992 Node *newhead = old_new[loop->_head->_idx]; 1993 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 1994 } 1995 } 1996 1997 //------------------------------clone_loop------------------------------------- 1998 // 1999 // C L O N E A L O O P B O D Y 2000 // 2001 // This is the basic building block of the loop optimizations. It clones an 2002 // entire loop body. It makes an old_new loop body mapping; with this mapping 2003 // you can find the new-loop equivalent to an old-loop node. All new-loop 2004 // nodes are exactly equal to their old-loop counterparts, all edges are the 2005 // same. All exits from the old-loop now have a RegionNode that merges the 2006 // equivalent new-loop path. This is true even for the normal "loop-exit" 2007 // condition. All uses of loop-invariant old-loop values now come from (one 2008 // or more) Phis that merge their new-loop equivalents. 2009 // 2010 // This operation leaves the graph in an illegal state: there are two valid 2011 // control edges coming from the loop pre-header to both loop bodies. I'll 2012 // definitely have to hack the graph after running this transform. 2013 // 2014 // From this building block I will further edit edges to perform loop peeling 2015 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc. 2016 // 2017 // Parameter side_by_size_idom: 2018 // When side_by_size_idom is NULL, the dominator tree is constructed for 2019 // the clone loop to dominate the original. Used in construction of 2020 // pre-main-post loop sequence. 2021 // When nonnull, the clone and original are side-by-side, both are 2022 // dominated by the side_by_side_idom node. Used in construction of 2023 // unswitched loops. 2024 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd, 2025 CloneLoopMode mode, Node* side_by_side_idom) { 2026 2027 LoopNode* head = loop->_head->as_Loop(); 2028 head->verify_strip_mined(1); 2029 2030 if (C->do_vector_loop() && PrintOpto) { 2031 const char* mname = C->method()->name()->as_quoted_ascii(); 2032 if (mname != NULL) { 2033 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname); 2034 } 2035 } 2036 2037 CloneMap& cm = C->clone_map(); 2038 Dict* dict = cm.dict(); 2039 if (C->do_vector_loop()) { 2040 cm.set_clone_idx(cm.max_gen()+1); 2041 #ifndef PRODUCT 2042 if (PrintOpto) { 2043 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx()); 2044 loop->dump_head(); 2045 } 2046 #endif 2047 } 2048 2049 // Step 1: Clone the loop body. Make the old->new mapping. 2050 uint i; 2051 for( i = 0; i < loop->_body.size(); i++ ) { 2052 Node *old = loop->_body.at(i); 2053 Node *nnn = old->clone(); 2054 old_new.map( old->_idx, nnn ); 2055 if (C->do_vector_loop()) { 2056 cm.verify_insert_and_clone(old, nnn, cm.clone_idx()); 2057 } 2058 _igvn.register_new_node_with_optimizer(nnn); 2059 } 2060 2061 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop; 2062 2063 // Step 2: Fix the edges in the new body. If the old input is outside the 2064 // loop use it. If the old input is INside the loop, use the corresponding 2065 // new node instead. 2066 for( i = 0; i < loop->_body.size(); i++ ) { 2067 Node *old = loop->_body.at(i); 2068 Node *nnn = old_new[old->_idx]; 2069 // Fix CFG/Loop controlling the new node 2070 if (has_ctrl(old)) { 2071 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]); 2072 } else { 2073 set_loop(nnn, outer_loop->_parent); 2074 if (old->outcnt() > 0) { 2075 set_idom( nnn, old_new[idom(old)->_idx], dd ); 2076 } 2077 } 2078 // Correct edges to the new node 2079 for( uint j = 0; j < nnn->req(); j++ ) { 2080 Node *n = nnn->in(j); 2081 if( n ) { 2082 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n ); 2083 if( loop->is_member( old_in_loop ) ) 2084 nnn->set_req(j, old_new[n->_idx]); 2085 } 2086 } 2087 _igvn.hash_find_insert(nnn); 2088 } 2089 2090 Node_List extra_data_nodes; // data nodes in the outer strip mined loop 2091 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes); 2092 2093 // Step 3: Now fix control uses. Loop varying control uses have already 2094 // been fixed up (as part of all input edges in Step 2). Loop invariant 2095 // control uses must be either an IfFalse or an IfTrue. Make a merge 2096 // point to merge the old and new IfFalse/IfTrue nodes; make the use 2097 // refer to this. 2098 Node_List worklist; 2099 uint new_counter = C->unique(); 2100 for( i = 0; i < loop->_body.size(); i++ ) { 2101 Node* old = loop->_body.at(i); 2102 if( !old->is_CFG() ) continue; 2103 2104 // Copy uses to a worklist, so I can munge the def-use info 2105 // with impunity. 2106 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 2107 worklist.push(old->fast_out(j)); 2108 2109 while( worklist.size() ) { // Visit all uses 2110 Node *use = worklist.pop(); 2111 if (!has_node(use)) continue; // Ignore dead nodes 2112 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 2113 if( !loop->is_member( use_loop ) && use->is_CFG() ) { 2114 // Both OLD and USE are CFG nodes here. 2115 assert( use->is_Proj(), "" ); 2116 Node* nnn = old_new[old->_idx]; 2117 2118 Node* newuse = NULL; 2119 if (head->is_strip_mined() && mode != IgnoreStripMined) { 2120 CountedLoopNode* cl = head->as_CountedLoop(); 2121 CountedLoopEndNode* cle = cl->loopexit(); 2122 Node* cle_out = cle->proj_out_or_null(false); 2123 if (use == cle_out) { 2124 IfNode* le = cl->outer_loop_end(); 2125 use = le->proj_out(false); 2126 use_loop = get_loop(use); 2127 if (mode == CloneIncludesStripMined) { 2128 nnn = old_new[le->_idx]; 2129 } else { 2130 newuse = old_new[cle_out->_idx]; 2131 } 2132 } 2133 } 2134 if (newuse == NULL) { 2135 newuse = use->clone(); 2136 } 2137 2138 // Clone the loop exit control projection 2139 if (C->do_vector_loop()) { 2140 cm.verify_insert_and_clone(use, newuse, cm.clone_idx()); 2141 } 2142 newuse->set_req(0,nnn); 2143 _igvn.register_new_node_with_optimizer(newuse); 2144 set_loop(newuse, use_loop); 2145 set_idom(newuse, nnn, dom_depth(nnn) + 1 ); 2146 2147 // We need a Region to merge the exit from the peeled body and the 2148 // exit from the old loop body. 2149 RegionNode *r = new RegionNode(3); 2150 // Map the old use to the new merge point 2151 old_new.map( use->_idx, r ); 2152 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use)); 2153 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" ); 2154 2155 // The original user of 'use' uses 'r' instead. 2156 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) { 2157 Node* useuse = use->last_out(l); 2158 _igvn.rehash_node_delayed(useuse); 2159 uint uses_found = 0; 2160 if( useuse->in(0) == use ) { 2161 useuse->set_req(0, r); 2162 uses_found++; 2163 if( useuse->is_CFG() ) { 2164 // This is not a dom_depth > dd_r because when new 2165 // control flow is constructed by a loop opt, a node and 2166 // its dominator can end up at the same dom_depth 2167 assert(dom_depth(useuse) >= dd_r, ""); 2168 set_idom(useuse, r, dom_depth(useuse)); 2169 } 2170 } 2171 for( uint k = 1; k < useuse->req(); k++ ) { 2172 if( useuse->in(k) == use ) { 2173 useuse->set_req(k, r); 2174 uses_found++; 2175 if (useuse->is_Loop() && k == LoopNode::EntryControl) { 2176 // This is not a dom_depth > dd_r because when new 2177 // control flow is constructed by a loop opt, a node 2178 // and its dominator can end up at the same dom_depth 2179 assert(dom_depth(useuse) >= dd_r , ""); 2180 set_idom(useuse, r, dom_depth(useuse)); 2181 } 2182 } 2183 } 2184 l -= uses_found; // we deleted 1 or more copies of this edge 2185 } 2186 2187 // Now finish up 'r' 2188 r->set_req( 1, newuse ); 2189 r->set_req( 2, use ); 2190 _igvn.register_new_node_with_optimizer(r); 2191 set_loop(r, use_loop); 2192 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r); 2193 } // End of if a loop-exit test 2194 } 2195 } 2196 2197 // Step 4: If loop-invariant use is not control, it must be dominated by a 2198 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region 2199 // there if needed. Make a Phi there merging old and new used values. 2200 Node_List *split_if_set = NULL; 2201 Node_List *split_bool_set = NULL; 2202 Node_List *split_cex_set = NULL; 2203 for( i = 0; i < loop->_body.size(); i++ ) { 2204 Node* old = loop->_body.at(i); 2205 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set, 2206 split_bool_set, split_cex_set, worklist, new_counter, 2207 mode); 2208 } 2209 2210 for (i = 0; i < extra_data_nodes.size(); i++) { 2211 Node* old = extra_data_nodes.at(i); 2212 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set, 2213 split_bool_set, split_cex_set, worklist, new_counter, 2214 mode); 2215 } 2216 2217 // Check for IFs that need splitting/cloning. Happens if an IF outside of 2218 // the loop uses a condition set in the loop. The original IF probably 2219 // takes control from one or more OLD Regions (which in turn get from NEW 2220 // Regions). In any case, there will be a set of Phis for each merge point 2221 // from the IF up to where the original BOOL def exists the loop. 2222 if (split_if_set) { 2223 while (split_if_set->size()) { 2224 Node *iff = split_if_set->pop(); 2225 if (iff->in(1)->is_Phi()) { 2226 Node *b = clone_iff(iff->in(1)->as_Phi(), loop); 2227 _igvn.replace_input_of(iff, 1, b); 2228 } 2229 } 2230 } 2231 if (split_bool_set) { 2232 while (split_bool_set->size()) { 2233 Node *b = split_bool_set->pop(); 2234 Node *phi = b->in(1); 2235 assert(phi->is_Phi(), ""); 2236 CmpNode *cmp = clone_bool((PhiNode*)phi, loop); 2237 _igvn.replace_input_of(b, 1, cmp); 2238 } 2239 } 2240 if (split_cex_set) { 2241 while (split_cex_set->size()) { 2242 Node *b = split_cex_set->pop(); 2243 assert(b->in(0)->is_Region(), ""); 2244 assert(b->in(1)->is_Phi(), ""); 2245 assert(b->in(0)->in(0) == b->in(1)->in(0), ""); 2246 split_up(b, b->in(0), NULL); 2247 } 2248 } 2249 2250 } 2251 2252 2253 //---------------------- stride_of_possible_iv ------------------------------------- 2254 // Looks for an iff/bool/comp with one operand of the compare 2255 // being a cycle involving an add and a phi, 2256 // with an optional truncation (left-shift followed by a right-shift) 2257 // of the add. Returns zero if not an iv. 2258 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) { 2259 Node* trunc1 = NULL; 2260 Node* trunc2 = NULL; 2261 const TypeInt* ttype = NULL; 2262 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) { 2263 return 0; 2264 } 2265 BoolNode* bl = iff->in(1)->as_Bool(); 2266 Node* cmp = bl->in(1); 2267 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) { 2268 return 0; 2269 } 2270 // Must have an invariant operand 2271 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) { 2272 return 0; 2273 } 2274 Node* add2 = NULL; 2275 Node* cmp1 = cmp->in(1); 2276 if (cmp1->is_Phi()) { 2277 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) ))) 2278 Node* phi = cmp1; 2279 for (uint i = 1; i < phi->req(); i++) { 2280 Node* in = phi->in(i); 2281 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in, 2282 &trunc1, &trunc2, &ttype); 2283 if (add && add->in(1) == phi) { 2284 add2 = add->in(2); 2285 break; 2286 } 2287 } 2288 } else { 2289 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) ))) 2290 Node* addtrunc = cmp1; 2291 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc, 2292 &trunc1, &trunc2, &ttype); 2293 if (add && add->in(1)->is_Phi()) { 2294 Node* phi = add->in(1); 2295 for (uint i = 1; i < phi->req(); i++) { 2296 if (phi->in(i) == addtrunc) { 2297 add2 = add->in(2); 2298 break; 2299 } 2300 } 2301 } 2302 } 2303 if (add2 != NULL) { 2304 const TypeInt* add2t = _igvn.type(add2)->is_int(); 2305 if (add2t->is_con()) { 2306 return add2t->get_con(); 2307 } 2308 } 2309 return 0; 2310 } 2311 2312 2313 //---------------------- stay_in_loop ------------------------------------- 2314 // Return the (unique) control output node that's in the loop (if it exists.) 2315 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) { 2316 Node* unique = NULL; 2317 if (!n) return NULL; 2318 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2319 Node* use = n->fast_out(i); 2320 if (!has_ctrl(use) && loop->is_member(get_loop(use))) { 2321 if (unique != NULL) { 2322 return NULL; 2323 } 2324 unique = use; 2325 } 2326 } 2327 return unique; 2328 } 2329 2330 //------------------------------ register_node ------------------------------------- 2331 // Utility to register node "n" with PhaseIdealLoop 2332 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) { 2333 _igvn.register_new_node_with_optimizer(n); 2334 loop->_body.push(n); 2335 if (n->is_CFG()) { 2336 set_loop(n, loop); 2337 set_idom(n, pred, ddepth); 2338 } else { 2339 set_ctrl(n, pred); 2340 } 2341 } 2342 2343 //------------------------------ proj_clone ------------------------------------- 2344 // Utility to create an if-projection 2345 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) { 2346 ProjNode* c = p->clone()->as_Proj(); 2347 c->set_req(0, iff); 2348 return c; 2349 } 2350 2351 //------------------------------ short_circuit_if ------------------------------------- 2352 // Force the iff control output to be the live_proj 2353 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) { 2354 guarantee(live_proj != NULL, "null projection"); 2355 int proj_con = live_proj->_con; 2356 assert(proj_con == 0 || proj_con == 1, "false or true projection"); 2357 Node *con = _igvn.intcon(proj_con); 2358 set_ctrl(con, C->root()); 2359 if (iff) { 2360 iff->set_req(1, con); 2361 } 2362 return con; 2363 } 2364 2365 //------------------------------ insert_if_before_proj ------------------------------------- 2366 // Insert a new if before an if projection (* - new node) 2367 // 2368 // before 2369 // if(test) 2370 // / \ 2371 // v v 2372 // other-proj proj (arg) 2373 // 2374 // after 2375 // if(test) 2376 // / \ 2377 // / v 2378 // | * proj-clone 2379 // v | 2380 // other-proj v 2381 // * new_if(relop(cmp[IU](left,right))) 2382 // / \ 2383 // v v 2384 // * new-proj proj 2385 // (returned) 2386 // 2387 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) { 2388 IfNode* iff = proj->in(0)->as_If(); 2389 IdealLoopTree *loop = get_loop(proj); 2390 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2391 int ddepth = dom_depth(proj); 2392 2393 _igvn.rehash_node_delayed(iff); 2394 _igvn.rehash_node_delayed(proj); 2395 2396 proj->set_req(0, NULL); // temporary disconnect 2397 ProjNode* proj2 = proj_clone(proj, iff); 2398 register_node(proj2, loop, iff, ddepth); 2399 2400 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right); 2401 register_node(cmp, loop, proj2, ddepth); 2402 2403 BoolNode* bol = new BoolNode(cmp, relop); 2404 register_node(bol, loop, proj2, ddepth); 2405 2406 int opcode = iff->Opcode(); 2407 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode"); 2408 IfNode* new_if = (opcode == Op_If) ? new IfNode(proj2, bol, iff->_prob, iff->_fcnt): 2409 new RangeCheckNode(proj2, bol, iff->_prob, iff->_fcnt); 2410 register_node(new_if, loop, proj2, ddepth); 2411 2412 proj->set_req(0, new_if); // reattach 2413 set_idom(proj, new_if, ddepth); 2414 2415 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj(); 2416 guarantee(new_exit != NULL, "null exit node"); 2417 register_node(new_exit, get_loop(other_proj), new_if, ddepth); 2418 2419 return new_exit; 2420 } 2421 2422 //------------------------------ insert_region_before_proj ------------------------------------- 2423 // Insert a region before an if projection (* - new node) 2424 // 2425 // before 2426 // if(test) 2427 // / | 2428 // v | 2429 // proj v 2430 // other-proj 2431 // 2432 // after 2433 // if(test) 2434 // / | 2435 // v | 2436 // * proj-clone v 2437 // | other-proj 2438 // v 2439 // * new-region 2440 // | 2441 // v 2442 // * dum_if 2443 // / \ 2444 // v \ 2445 // * dum-proj v 2446 // proj 2447 // 2448 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) { 2449 IfNode* iff = proj->in(0)->as_If(); 2450 IdealLoopTree *loop = get_loop(proj); 2451 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2452 int ddepth = dom_depth(proj); 2453 2454 _igvn.rehash_node_delayed(iff); 2455 _igvn.rehash_node_delayed(proj); 2456 2457 proj->set_req(0, NULL); // temporary disconnect 2458 ProjNode* proj2 = proj_clone(proj, iff); 2459 register_node(proj2, loop, iff, ddepth); 2460 2461 RegionNode* reg = new RegionNode(2); 2462 reg->set_req(1, proj2); 2463 register_node(reg, loop, iff, ddepth); 2464 2465 IfNode* dum_if = new IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt); 2466 register_node(dum_if, loop, reg, ddepth); 2467 2468 proj->set_req(0, dum_if); // reattach 2469 set_idom(proj, dum_if, ddepth); 2470 2471 ProjNode* dum_proj = proj_clone(other_proj, dum_if); 2472 register_node(dum_proj, loop, dum_if, ddepth); 2473 2474 return reg; 2475 } 2476 2477 //------------------------------ insert_cmpi_loop_exit ------------------------------------- 2478 // Clone a signed compare loop exit from an unsigned compare and 2479 // insert it before the unsigned cmp on the stay-in-loop path. 2480 // All new nodes inserted in the dominator tree between the original 2481 // if and it's projections. The original if test is replaced with 2482 // a constant to force the stay-in-loop path. 2483 // 2484 // This is done to make sure that the original if and it's projections 2485 // still dominate the same set of control nodes, that the ctrl() relation 2486 // from data nodes to them is preserved, and that their loop nesting is 2487 // preserved. 2488 // 2489 // before 2490 // if(i <u limit) unsigned compare loop exit 2491 // / | 2492 // v v 2493 // exit-proj stay-in-loop-proj 2494 // 2495 // after 2496 // if(stay-in-loop-const) original if 2497 // / | 2498 // / v 2499 // / if(i < limit) new signed test 2500 // / / | 2501 // / / v 2502 // / / if(i <u limit) new cloned unsigned test 2503 // / / / | 2504 // v v v | 2505 // region | 2506 // | | 2507 // dum-if | 2508 // / | | 2509 // ether | | 2510 // v v 2511 // exit-proj stay-in-loop-proj 2512 // 2513 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) { 2514 const bool Signed = true; 2515 const bool Unsigned = false; 2516 2517 BoolNode* bol = if_cmpu->in(1)->as_Bool(); 2518 if (bol->_test._test != BoolTest::lt) return NULL; 2519 CmpNode* cmpu = bol->in(1)->as_Cmp(); 2520 if (cmpu->Opcode() != Op_CmpU) return NULL; 2521 int stride = stride_of_possible_iv(if_cmpu); 2522 if (stride == 0) return NULL; 2523 2524 Node* lp_proj = stay_in_loop(if_cmpu, loop); 2525 guarantee(lp_proj != NULL, "null loop node"); 2526 2527 ProjNode* lp_continue = lp_proj->as_Proj(); 2528 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj(); 2529 2530 Node* limit = NULL; 2531 if (stride > 0) { 2532 limit = cmpu->in(2); 2533 } else { 2534 limit = _igvn.makecon(TypeInt::ZERO); 2535 set_ctrl(limit, C->root()); 2536 } 2537 // Create a new region on the exit path 2538 RegionNode* reg = insert_region_before_proj(lp_exit); 2539 guarantee(reg != NULL, "null region node"); 2540 2541 // Clone the if-cmpu-true-false using a signed compare 2542 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge; 2543 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue); 2544 reg->add_req(cmpi_exit); 2545 2546 // Clone the if-cmpu-true-false 2547 BoolTest::mask rel_u = bol->_test._test; 2548 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue); 2549 reg->add_req(cmpu_exit); 2550 2551 // Force original if to stay in loop. 2552 short_circuit_if(if_cmpu, lp_continue); 2553 2554 return cmpi_exit->in(0)->as_If(); 2555 } 2556 2557 //------------------------------ remove_cmpi_loop_exit ------------------------------------- 2558 // Remove a previously inserted signed compare loop exit. 2559 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) { 2560 Node* lp_proj = stay_in_loop(if_cmp, loop); 2561 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI && 2562 stay_in_loop(lp_proj, loop)->is_If() && 2563 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu"); 2564 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO); 2565 set_ctrl(con, C->root()); 2566 if_cmp->set_req(1, con); 2567 } 2568 2569 //------------------------------ scheduled_nodelist ------------------------------------- 2570 // Create a post order schedule of nodes that are in the 2571 // "member" set. The list is returned in "sched". 2572 // The first node in "sched" is the loop head, followed by 2573 // nodes which have no inputs in the "member" set, and then 2574 // followed by the nodes that have an immediate input dependence 2575 // on a node in "sched". 2576 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) { 2577 2578 assert(member.test(loop->_head->_idx), "loop head must be in member set"); 2579 VectorSet visited; 2580 Node_Stack nstack(loop->_body.size()); 2581 2582 Node* n = loop->_head; // top of stack is cached in "n" 2583 uint idx = 0; 2584 visited.set(n->_idx); 2585 2586 // Initially push all with no inputs from within member set 2587 for(uint i = 0; i < loop->_body.size(); i++ ) { 2588 Node *elt = loop->_body.at(i); 2589 if (member.test(elt->_idx)) { 2590 bool found = false; 2591 for (uint j = 0; j < elt->req(); j++) { 2592 Node* def = elt->in(j); 2593 if (def && member.test(def->_idx) && def != elt) { 2594 found = true; 2595 break; 2596 } 2597 } 2598 if (!found && elt != loop->_head) { 2599 nstack.push(n, idx); 2600 n = elt; 2601 assert(!visited.test(n->_idx), "not seen yet"); 2602 visited.set(n->_idx); 2603 } 2604 } 2605 } 2606 2607 // traverse out's that are in the member set 2608 while (true) { 2609 if (idx < n->outcnt()) { 2610 Node* use = n->raw_out(idx); 2611 idx++; 2612 if (!visited.test_set(use->_idx)) { 2613 if (member.test(use->_idx)) { 2614 nstack.push(n, idx); 2615 n = use; 2616 idx = 0; 2617 } 2618 } 2619 } else { 2620 // All outputs processed 2621 sched.push(n); 2622 if (nstack.is_empty()) break; 2623 n = nstack.node(); 2624 idx = nstack.index(); 2625 nstack.pop(); 2626 } 2627 } 2628 } 2629 2630 2631 //------------------------------ has_use_in_set ------------------------------------- 2632 // Has a use in the vector set 2633 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) { 2634 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2635 Node* use = n->fast_out(j); 2636 if (vset.test(use->_idx)) { 2637 return true; 2638 } 2639 } 2640 return false; 2641 } 2642 2643 2644 //------------------------------ has_use_internal_to_set ------------------------------------- 2645 // Has use internal to the vector set (ie. not in a phi at the loop head) 2646 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) { 2647 Node* head = loop->_head; 2648 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2649 Node* use = n->fast_out(j); 2650 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) { 2651 return true; 2652 } 2653 } 2654 return false; 2655 } 2656 2657 2658 //------------------------------ clone_for_use_outside_loop ------------------------------------- 2659 // clone "n" for uses that are outside of loop 2660 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) { 2661 int cloned = 0; 2662 assert(worklist.size() == 0, "should be empty"); 2663 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2664 Node* use = n->fast_out(j); 2665 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) { 2666 worklist.push(use); 2667 } 2668 } 2669 while( worklist.size() ) { 2670 Node *use = worklist.pop(); 2671 if (!has_node(use) || use->in(0) == C->top()) continue; 2672 uint j; 2673 for (j = 0; j < use->req(); j++) { 2674 if (use->in(j) == n) break; 2675 } 2676 assert(j < use->req(), "must be there"); 2677 2678 // clone "n" and insert it between the inputs of "n" and the use outside the loop 2679 Node* n_clone = n->clone(); 2680 _igvn.replace_input_of(use, j, n_clone); 2681 cloned++; 2682 Node* use_c; 2683 if (!use->is_Phi()) { 2684 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0); 2685 } else { 2686 // Use in a phi is considered a use in the associated predecessor block 2687 use_c = use->in(0)->in(j); 2688 } 2689 set_ctrl(n_clone, use_c); 2690 assert(!loop->is_member(get_loop(use_c)), "should be outside loop"); 2691 get_loop(use_c)->_body.push(n_clone); 2692 _igvn.register_new_node_with_optimizer(n_clone); 2693 #ifndef PRODUCT 2694 if (TracePartialPeeling) { 2695 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx); 2696 } 2697 #endif 2698 } 2699 return cloned; 2700 } 2701 2702 2703 //------------------------------ clone_for_special_use_inside_loop ------------------------------------- 2704 // clone "n" for special uses that are in the not_peeled region. 2705 // If these def-uses occur in separate blocks, the code generator 2706 // marks the method as not compilable. For example, if a "BoolNode" 2707 // is in a different basic block than the "IfNode" that uses it, then 2708 // the compilation is aborted in the code generator. 2709 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 2710 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) { 2711 if (n->is_Phi() || n->is_Load()) { 2712 return; 2713 } 2714 assert(worklist.size() == 0, "should be empty"); 2715 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2716 Node* use = n->fast_out(j); 2717 if ( not_peel.test(use->_idx) && 2718 (use->is_If() || use->is_CMove() || use->is_Bool()) && 2719 use->in(1) == n) { 2720 worklist.push(use); 2721 } 2722 } 2723 if (worklist.size() > 0) { 2724 // clone "n" and insert it between inputs of "n" and the use 2725 Node* n_clone = n->clone(); 2726 loop->_body.push(n_clone); 2727 _igvn.register_new_node_with_optimizer(n_clone); 2728 set_ctrl(n_clone, get_ctrl(n)); 2729 sink_list.push(n_clone); 2730 not_peel.set(n_clone->_idx); 2731 #ifndef PRODUCT 2732 if (TracePartialPeeling) { 2733 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx); 2734 } 2735 #endif 2736 while( worklist.size() ) { 2737 Node *use = worklist.pop(); 2738 _igvn.rehash_node_delayed(use); 2739 for (uint j = 1; j < use->req(); j++) { 2740 if (use->in(j) == n) { 2741 use->set_req(j, n_clone); 2742 } 2743 } 2744 } 2745 } 2746 } 2747 2748 2749 //------------------------------ insert_phi_for_loop ------------------------------------- 2750 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 2751 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) { 2752 Node *phi = PhiNode::make(lp, back_edge_val); 2753 phi->set_req(LoopNode::EntryControl, lp_entry_val); 2754 // Use existing phi if it already exists 2755 Node *hit = _igvn.hash_find_insert(phi); 2756 if( hit == NULL ) { 2757 _igvn.register_new_node_with_optimizer(phi); 2758 set_ctrl(phi, lp); 2759 } else { 2760 // Remove the new phi from the graph and use the hit 2761 _igvn.remove_dead_node(phi); 2762 phi = hit; 2763 } 2764 _igvn.replace_input_of(use, idx, phi); 2765 } 2766 2767 #ifdef ASSERT 2768 //------------------------------ is_valid_loop_partition ------------------------------------- 2769 // Validate the loop partition sets: peel and not_peel 2770 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, 2771 VectorSet& not_peel ) { 2772 uint i; 2773 // Check that peel_list entries are in the peel set 2774 for (i = 0; i < peel_list.size(); i++) { 2775 if (!peel.test(peel_list.at(i)->_idx)) { 2776 return false; 2777 } 2778 } 2779 // Check at loop members are in one of peel set or not_peel set 2780 for (i = 0; i < loop->_body.size(); i++ ) { 2781 Node *def = loop->_body.at(i); 2782 uint di = def->_idx; 2783 // Check that peel set elements are in peel_list 2784 if (peel.test(di)) { 2785 if (not_peel.test(di)) { 2786 return false; 2787 } 2788 // Must be in peel_list also 2789 bool found = false; 2790 for (uint j = 0; j < peel_list.size(); j++) { 2791 if (peel_list.at(j)->_idx == di) { 2792 found = true; 2793 break; 2794 } 2795 } 2796 if (!found) { 2797 return false; 2798 } 2799 } else if (not_peel.test(di)) { 2800 if (peel.test(di)) { 2801 return false; 2802 } 2803 } else { 2804 return false; 2805 } 2806 } 2807 return true; 2808 } 2809 2810 //------------------------------ is_valid_clone_loop_exit_use ------------------------------------- 2811 // Ensure a use outside of loop is of the right form 2812 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) { 2813 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2814 return (use->is_Phi() && 2815 use_c->is_Region() && use_c->req() == 3 && 2816 (use_c->in(exit_idx)->Opcode() == Op_IfTrue || 2817 use_c->in(exit_idx)->Opcode() == Op_IfFalse || 2818 use_c->in(exit_idx)->Opcode() == Op_JumpProj) && 2819 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) ); 2820 } 2821 2822 //------------------------------ is_valid_clone_loop_form ------------------------------------- 2823 // Ensure that all uses outside of loop are of the right form 2824 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 2825 uint orig_exit_idx, uint clone_exit_idx) { 2826 uint len = peel_list.size(); 2827 for (uint i = 0; i < len; i++) { 2828 Node *def = peel_list.at(i); 2829 2830 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 2831 Node *use = def->fast_out(j); 2832 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2833 if (!loop->is_member(get_loop(use_c))) { 2834 // use is not in the loop, check for correct structure 2835 if (use->in(0) == def) { 2836 // Okay 2837 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) { 2838 return false; 2839 } 2840 } 2841 } 2842 } 2843 return true; 2844 } 2845 #endif 2846 2847 //------------------------------ partial_peel ------------------------------------- 2848 // Partially peel (aka loop rotation) the top portion of a loop (called 2849 // the peel section below) by cloning it and placing one copy just before 2850 // the new loop head and the other copy at the bottom of the new loop. 2851 // 2852 // before after where it came from 2853 // 2854 // stmt1 stmt1 2855 // loop: stmt2 clone 2856 // stmt2 if condA goto exitA clone 2857 // if condA goto exitA new_loop: new 2858 // stmt3 stmt3 clone 2859 // if !condB goto loop if condB goto exitB clone 2860 // exitB: stmt2 orig 2861 // stmt4 if !condA goto new_loop orig 2862 // exitA: goto exitA 2863 // exitB: 2864 // stmt4 2865 // exitA: 2866 // 2867 // Step 1: find the cut point: an exit test on probable 2868 // induction variable. 2869 // Step 2: schedule (with cloning) operations in the peel 2870 // section that can be executed after the cut into 2871 // the section that is not peeled. This may need 2872 // to clone operations into exit blocks. For 2873 // instance, a reference to A[i] in the not-peel 2874 // section and a reference to B[i] in an exit block 2875 // may cause a left-shift of i by 2 to be placed 2876 // in the peel block. This step will clone the left 2877 // shift into the exit block and sink the left shift 2878 // from the peel to the not-peel section. 2879 // Step 3: clone the loop, retarget the control, and insert 2880 // phis for values that are live across the new loop 2881 // head. This is very dependent on the graph structure 2882 // from clone_loop. It creates region nodes for 2883 // exit control and associated phi nodes for values 2884 // flow out of the loop through that exit. The region 2885 // node is dominated by the clone's control projection. 2886 // So the clone's peel section is placed before the 2887 // new loop head, and the clone's not-peel section is 2888 // forms the top part of the new loop. The original 2889 // peel section forms the tail of the new loop. 2890 // Step 4: update the dominator tree and recompute the 2891 // dominator depth. 2892 // 2893 // orig 2894 // 2895 // stmt1 2896 // | 2897 // v 2898 // loop predicate 2899 // | 2900 // v 2901 // loop<----+ 2902 // | | 2903 // stmt2 | 2904 // | | 2905 // v | 2906 // ifA | 2907 // / | | 2908 // v v | 2909 // false true ^ <-- last_peel 2910 // / | | 2911 // / ===|==cut | 2912 // / stmt3 | <-- first_not_peel 2913 // / | | 2914 // | v | 2915 // v ifB | 2916 // exitA: / \ | 2917 // / \ | 2918 // v v | 2919 // false true | 2920 // / \ | 2921 // / ----+ 2922 // | 2923 // v 2924 // exitB: 2925 // stmt4 2926 // 2927 // 2928 // after clone loop 2929 // 2930 // stmt1 2931 // | 2932 // v 2933 // loop predicate 2934 // / \ 2935 // clone / \ orig 2936 // / \ 2937 // / \ 2938 // v v 2939 // +---->loop loop<----+ 2940 // | | | | 2941 // | stmt2 stmt2 | 2942 // | | | | 2943 // | v v | 2944 // | ifA ifA | 2945 // | | \ / | | 2946 // | v v v v | 2947 // ^ true false false true ^ <-- last_peel 2948 // | | ^ \ / | | 2949 // | cut==|== \ \ / ===|==cut | 2950 // | stmt3 \ \ / stmt3 | <-- first_not_peel 2951 // | | dom | | | | 2952 // | v \ 1v v2 v | 2953 // | ifB regionA ifB | 2954 // | / \ | / \ | 2955 // | / \ v / \ | 2956 // | v v exitA: v v | 2957 // | true false false true | 2958 // | / ^ \ / \ | 2959 // +---- \ \ / ----+ 2960 // dom \ / 2961 // \ 1v v2 2962 // regionB 2963 // | 2964 // v 2965 // exitB: 2966 // stmt4 2967 // 2968 // 2969 // after partial peel 2970 // 2971 // stmt1 2972 // | 2973 // v 2974 // loop predicate 2975 // / 2976 // clone / orig 2977 // / TOP 2978 // / \ 2979 // v v 2980 // TOP->loop loop----+ 2981 // | | | 2982 // stmt2 stmt2 | 2983 // | | | 2984 // v v | 2985 // ifA ifA | 2986 // | \ / | | 2987 // v v v v | 2988 // true false false true | <-- last_peel 2989 // | ^ \ / +------|---+ 2990 // +->newloop \ \ / === ==cut | | 2991 // | stmt3 \ \ / TOP | | 2992 // | | dom | | stmt3 | | <-- first_not_peel 2993 // | v \ 1v v2 v | | 2994 // | ifB regionA ifB ^ v 2995 // | / \ | / \ | | 2996 // | / \ v / \ | | 2997 // | v v exitA: v v | | 2998 // | true false false true | | 2999 // | / ^ \ / \ | | 3000 // | | \ \ / v | | 3001 // | | dom \ / TOP | | 3002 // | | \ 1v v2 | | 3003 // ^ v regionB | | 3004 // | | | | | 3005 // | | v ^ v 3006 // | | exitB: | | 3007 // | | stmt4 | | 3008 // | +------------>-----------------+ | 3009 // | | 3010 // +-----------------<---------------------+ 3011 // 3012 // 3013 // final graph 3014 // 3015 // stmt1 3016 // | 3017 // v 3018 // loop predicate 3019 // | 3020 // v 3021 // stmt2 clone 3022 // | 3023 // v 3024 // ........> ifA clone 3025 // : / | 3026 // dom / | 3027 // : v v 3028 // : false true 3029 // : | | 3030 // : | v 3031 // : | newloop<-----+ 3032 // : | | | 3033 // : | stmt3 clone | 3034 // : | | | 3035 // : | v | 3036 // : | ifB | 3037 // : | / \ | 3038 // : | v v | 3039 // : | false true | 3040 // : | | | | 3041 // : | v stmt2 | 3042 // : | exitB: | | 3043 // : | stmt4 v | 3044 // : | ifA orig | 3045 // : | / \ | 3046 // : | / \ | 3047 // : | v v | 3048 // : | false true | 3049 // : | / \ | 3050 // : v v -----+ 3051 // RegionA 3052 // | 3053 // v 3054 // exitA 3055 // 3056 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) { 3057 3058 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only"); 3059 if (!loop->_head->is_Loop()) { 3060 return false; 3061 } 3062 LoopNode *head = loop->_head->as_Loop(); 3063 3064 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) { 3065 return false; 3066 } 3067 3068 // Check for complex exit control 3069 for (uint ii = 0; ii < loop->_body.size(); ii++) { 3070 Node *n = loop->_body.at(ii); 3071 int opc = n->Opcode(); 3072 if (n->is_Call() || 3073 opc == Op_Catch || 3074 opc == Op_CatchProj || 3075 opc == Op_Jump || 3076 opc == Op_JumpProj) { 3077 #ifndef PRODUCT 3078 if (TracePartialPeeling) { 3079 tty->print_cr("\nExit control too complex: lp: %d", head->_idx); 3080 } 3081 #endif 3082 return false; 3083 } 3084 } 3085 3086 int dd = dom_depth(head); 3087 3088 // Step 1: find cut point 3089 3090 // Walk up dominators to loop head looking for first loop exit 3091 // which is executed on every path thru loop. 3092 IfNode *peel_if = NULL; 3093 IfNode *peel_if_cmpu = NULL; 3094 3095 Node *iff = loop->tail(); 3096 while (iff != head) { 3097 if (iff->is_If()) { 3098 Node *ctrl = get_ctrl(iff->in(1)); 3099 if (ctrl->is_top()) return false; // Dead test on live IF. 3100 // If loop-varying exit-test, check for induction variable 3101 if (loop->is_member(get_loop(ctrl)) && 3102 loop->is_loop_exit(iff) && 3103 is_possible_iv_test(iff)) { 3104 Node* cmp = iff->in(1)->in(1); 3105 if (cmp->Opcode() == Op_CmpI) { 3106 peel_if = iff->as_If(); 3107 } else { 3108 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU"); 3109 peel_if_cmpu = iff->as_If(); 3110 } 3111 } 3112 } 3113 iff = idom(iff); 3114 } 3115 3116 // Prefer signed compare over unsigned compare. 3117 IfNode* new_peel_if = NULL; 3118 if (peel_if == NULL) { 3119 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) { 3120 return false; // No peel point found 3121 } 3122 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop); 3123 if (new_peel_if == NULL) { 3124 return false; // No peel point found 3125 } 3126 peel_if = new_peel_if; 3127 } 3128 Node* last_peel = stay_in_loop(peel_if, loop); 3129 Node* first_not_peeled = stay_in_loop(last_peel, loop); 3130 if (first_not_peeled == NULL || first_not_peeled == head) { 3131 return false; 3132 } 3133 3134 #ifndef PRODUCT 3135 if (TraceLoopOpts) { 3136 tty->print("PartialPeel "); 3137 loop->dump_head(); 3138 } 3139 3140 if (TracePartialPeeling) { 3141 tty->print_cr("before partial peel one iteration"); 3142 Node_List wl; 3143 Node* t = head->in(2); 3144 while (true) { 3145 wl.push(t); 3146 if (t == head) break; 3147 t = idom(t); 3148 } 3149 while (wl.size() > 0) { 3150 Node* tt = wl.pop(); 3151 tt->dump(); 3152 if (tt == last_peel) tty->print_cr("-- cut --"); 3153 } 3154 } 3155 #endif 3156 VectorSet peel; 3157 VectorSet not_peel; 3158 Node_List peel_list; 3159 Node_List worklist; 3160 Node_List sink_list; 3161 3162 uint estimate = loop->est_loop_clone_sz(1); 3163 if (exceeding_node_budget(estimate)) { 3164 return false; 3165 } 3166 3167 // Set of cfg nodes to peel are those that are executable from 3168 // the head through last_peel. 3169 assert(worklist.size() == 0, "should be empty"); 3170 worklist.push(head); 3171 peel.set(head->_idx); 3172 while (worklist.size() > 0) { 3173 Node *n = worklist.pop(); 3174 if (n != last_peel) { 3175 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3176 Node* use = n->fast_out(j); 3177 if (use->is_CFG() && 3178 loop->is_member(get_loop(use)) && 3179 !peel.test_set(use->_idx)) { 3180 worklist.push(use); 3181 } 3182 } 3183 } 3184 } 3185 3186 // Set of non-cfg nodes to peel are those that are control 3187 // dependent on the cfg nodes. 3188 for (uint i = 0; i < loop->_body.size(); i++) { 3189 Node *n = loop->_body.at(i); 3190 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n; 3191 if (peel.test(n_c->_idx)) { 3192 peel.set(n->_idx); 3193 } else { 3194 not_peel.set(n->_idx); 3195 } 3196 } 3197 3198 // Step 2: move operations from the peeled section down into the 3199 // not-peeled section 3200 3201 // Get a post order schedule of nodes in the peel region 3202 // Result in right-most operand. 3203 scheduled_nodelist(loop, peel, peel_list); 3204 3205 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 3206 3207 // For future check for too many new phis 3208 uint old_phi_cnt = 0; 3209 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 3210 Node* use = head->fast_out(j); 3211 if (use->is_Phi()) old_phi_cnt++; 3212 } 3213 3214 #ifndef PRODUCT 3215 if (TracePartialPeeling) { 3216 tty->print_cr("\npeeled list"); 3217 } 3218 #endif 3219 3220 // Evacuate nodes in peel region into the not_peeled region if possible 3221 uint new_phi_cnt = 0; 3222 uint cloned_for_outside_use = 0; 3223 for (uint i = 0; i < peel_list.size();) { 3224 Node* n = peel_list.at(i); 3225 #ifndef PRODUCT 3226 if (TracePartialPeeling) n->dump(); 3227 #endif 3228 bool incr = true; 3229 if (!n->is_CFG()) { 3230 if (has_use_in_set(n, not_peel)) { 3231 // If not used internal to the peeled region, 3232 // move "n" from peeled to not_peeled region. 3233 if (!has_use_internal_to_set(n, peel, loop)) { 3234 // if not pinned and not a load (which maybe anti-dependent on a store) 3235 // and not a CMove (Matcher expects only bool->cmove). 3236 if (n->in(0) == NULL && !n->is_Load() && !n->is_CMove()) { 3237 cloned_for_outside_use += clone_for_use_outside_loop(loop, n, worklist); 3238 sink_list.push(n); 3239 peel.remove(n->_idx); 3240 not_peel.set(n->_idx); 3241 peel_list.remove(i); 3242 incr = false; 3243 #ifndef PRODUCT 3244 if (TracePartialPeeling) { 3245 tty->print_cr("sink to not_peeled region: %d newbb: %d", 3246 n->_idx, get_ctrl(n)->_idx); 3247 } 3248 #endif 3249 } 3250 } else { 3251 // Otherwise check for special def-use cases that span 3252 // the peel/not_peel boundary such as bool->if 3253 clone_for_special_use_inside_loop(loop, n, not_peel, sink_list, worklist); 3254 new_phi_cnt++; 3255 } 3256 } 3257 } 3258 if (incr) i++; 3259 } 3260 3261 estimate += cloned_for_outside_use + new_phi_cnt; 3262 bool exceed_node_budget = !may_require_nodes(estimate); 3263 bool exceed_phi_limit = new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta; 3264 3265 if (exceed_node_budget || exceed_phi_limit) { 3266 #ifndef PRODUCT 3267 if (TracePartialPeeling) { 3268 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c", 3269 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F'); 3270 } 3271 #endif 3272 if (new_peel_if != NULL) { 3273 remove_cmpi_loop_exit(new_peel_if, loop); 3274 } 3275 // Inhibit more partial peeling on this loop 3276 assert(!head->is_partial_peel_loop(), "not partial peeled"); 3277 head->mark_partial_peel_failed(); 3278 if (cloned_for_outside_use > 0) { 3279 // Terminate this round of loop opts because 3280 // the graph outside this loop was changed. 3281 C->set_major_progress(); 3282 return true; 3283 } 3284 return false; 3285 } 3286 3287 // Step 3: clone loop, retarget control, and insert new phis 3288 3289 // Create new loop head for new phis and to hang 3290 // the nodes being moved (sinked) from the peel region. 3291 LoopNode* new_head = new LoopNode(last_peel, last_peel); 3292 new_head->set_unswitch_count(head->unswitch_count()); // Preserve 3293 _igvn.register_new_node_with_optimizer(new_head); 3294 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled"); 3295 _igvn.replace_input_of(first_not_peeled, 0, new_head); 3296 set_loop(new_head, loop); 3297 loop->_body.push(new_head); 3298 not_peel.set(new_head->_idx); 3299 set_idom(new_head, last_peel, dom_depth(first_not_peeled)); 3300 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled)); 3301 3302 while (sink_list.size() > 0) { 3303 Node* n = sink_list.pop(); 3304 set_ctrl(n, new_head); 3305 } 3306 3307 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 3308 3309 clone_loop(loop, old_new, dd, IgnoreStripMined); 3310 3311 const uint clone_exit_idx = 1; 3312 const uint orig_exit_idx = 2; 3313 assert(is_valid_clone_loop_form(loop, peel_list, orig_exit_idx, clone_exit_idx), "bad clone loop"); 3314 3315 Node* head_clone = old_new[head->_idx]; 3316 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop(); 3317 Node* orig_tail_clone = head_clone->in(2); 3318 3319 // Add phi if "def" node is in peel set and "use" is not 3320 3321 for (uint i = 0; i < peel_list.size(); i++) { 3322 Node *def = peel_list.at(i); 3323 if (!def->is_CFG()) { 3324 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 3325 Node *use = def->fast_out(j); 3326 if (has_node(use) && use->in(0) != C->top() && 3327 (!peel.test(use->_idx) || 3328 (use->is_Phi() && use->in(0) == head)) ) { 3329 worklist.push(use); 3330 } 3331 } 3332 while( worklist.size() ) { 3333 Node *use = worklist.pop(); 3334 for (uint j = 1; j < use->req(); j++) { 3335 Node* n = use->in(j); 3336 if (n == def) { 3337 3338 // "def" is in peel set, "use" is not in peel set 3339 // or "use" is in the entry boundary (a phi) of the peel set 3340 3341 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use; 3342 3343 if ( loop->is_member(get_loop( use_c )) ) { 3344 // use is in loop 3345 if (old_new[use->_idx] != NULL) { // null for dead code 3346 Node* use_clone = old_new[use->_idx]; 3347 _igvn.replace_input_of(use, j, C->top()); 3348 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone ); 3349 } 3350 } else { 3351 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format"); 3352 // use is not in the loop, check if the live range includes the cut 3353 Node* lp_if = use_c->in(orig_exit_idx)->in(0); 3354 if (not_peel.test(lp_if->_idx)) { 3355 assert(j == orig_exit_idx, "use from original loop"); 3356 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone ); 3357 } 3358 } 3359 } 3360 } 3361 } 3362 } 3363 } 3364 3365 // Step 3b: retarget control 3366 3367 // Redirect control to the new loop head if a cloned node in 3368 // the not_peeled region has control that points into the peeled region. 3369 // This necessary because the cloned peeled region will be outside 3370 // the loop. 3371 // from to 3372 // cloned-peeled <---+ 3373 // new_head_clone: | <--+ 3374 // cloned-not_peeled in(0) in(0) 3375 // orig-peeled 3376 3377 for (uint i = 0; i < loop->_body.size(); i++) { 3378 Node *n = loop->_body.at(i); 3379 if (!n->is_CFG() && n->in(0) != NULL && 3380 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) { 3381 Node* n_clone = old_new[n->_idx]; 3382 _igvn.replace_input_of(n_clone, 0, new_head_clone); 3383 } 3384 } 3385 3386 // Backedge of the surviving new_head (the clone) is original last_peel 3387 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel); 3388 3389 // Cut first node in original not_peel set 3390 _igvn.rehash_node_delayed(new_head); // Multiple edge updates: 3391 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of 3392 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls 3393 3394 // Copy head_clone back-branch info to original head 3395 // and remove original head's loop entry and 3396 // clone head's back-branch 3397 _igvn.rehash_node_delayed(head); // Multiple edge updates 3398 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl)); 3399 head->set_req(LoopNode::LoopBackControl, C->top()); 3400 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top()); 3401 3402 // Similarly modify the phis 3403 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) { 3404 Node* use = head->fast_out(k); 3405 if (use->is_Phi() && use->outcnt() > 0) { 3406 Node* use_clone = old_new[use->_idx]; 3407 _igvn.rehash_node_delayed(use); // Multiple edge updates 3408 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl)); 3409 use->set_req(LoopNode::LoopBackControl, C->top()); 3410 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top()); 3411 } 3412 } 3413 3414 // Step 4: update dominator tree and dominator depth 3415 3416 set_idom(head, orig_tail_clone, dd); 3417 recompute_dom_depth(); 3418 3419 // Inhibit more partial peeling on this loop 3420 new_head_clone->set_partial_peel_loop(); 3421 C->set_major_progress(); 3422 loop->record_for_igvn(); 3423 3424 #ifndef PRODUCT 3425 if (TracePartialPeeling) { 3426 tty->print_cr("\nafter partial peel one iteration"); 3427 Node_List wl; 3428 Node* t = last_peel; 3429 while (true) { 3430 wl.push(t); 3431 if (t == head_clone) break; 3432 t = idom(t); 3433 } 3434 while (wl.size() > 0) { 3435 Node* tt = wl.pop(); 3436 if (tt == head) tty->print_cr("orig head"); 3437 else if (tt == new_head_clone) tty->print_cr("new head"); 3438 else if (tt == head_clone) tty->print_cr("clone head"); 3439 tt->dump(); 3440 } 3441 } 3442 #endif 3443 return true; 3444 } 3445 3446 //------------------------------reorg_offsets---------------------------------- 3447 // Reorganize offset computations to lower register pressure. Mostly 3448 // prevent loop-fallout uses of the pre-incremented trip counter (which are 3449 // then alive with the post-incremented trip counter forcing an extra 3450 // register move) 3451 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) { 3452 // Perform it only for canonical counted loops. 3453 // Loop's shape could be messed up by iteration_split_impl. 3454 if (!loop->_head->is_CountedLoop()) 3455 return; 3456 if (!loop->_head->as_Loop()->is_valid_counted_loop()) 3457 return; 3458 3459 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 3460 CountedLoopEndNode *cle = cl->loopexit(); 3461 Node *exit = cle->proj_out(false); 3462 Node *phi = cl->phi(); 3463 3464 // Check for the special case when using the pre-incremented trip-counter on 3465 // the fall-out path (forces the pre-incremented and post-incremented trip 3466 // counter to be live at the same time). Fix this by adjusting to use the 3467 // post-increment trip counter. 3468 3469 bool progress = true; 3470 while (progress) { 3471 progress = false; 3472 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) { 3473 Node* use = phi->fast_out(i); // User of trip-counter 3474 if (!has_ctrl(use)) continue; 3475 Node *u_ctrl = get_ctrl(use); 3476 if (use->is_Phi()) { 3477 u_ctrl = NULL; 3478 for (uint j = 1; j < use->req(); j++) 3479 if (use->in(j) == phi) 3480 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j)); 3481 } 3482 IdealLoopTree *u_loop = get_loop(u_ctrl); 3483 // Look for loop-invariant use 3484 if (u_loop == loop) continue; 3485 if (loop->is_member(u_loop)) continue; 3486 // Check that use is live out the bottom. Assuming the trip-counter 3487 // update is right at the bottom, uses of of the loop middle are ok. 3488 if (dom_lca(exit, u_ctrl) != exit) continue; 3489 // Hit! Refactor use to use the post-incremented tripcounter. 3490 // Compute a post-increment tripcounter. 3491 Node* c = exit; 3492 if (cl->is_strip_mined()) { 3493 IdealLoopTree* outer_loop = get_loop(cl->outer_loop()); 3494 if (!outer_loop->is_member(u_loop)) { 3495 c = cl->outer_loop_exit(); 3496 } 3497 } 3498 Node *opaq = new Opaque2Node(C, cle->incr()); 3499 register_new_node(opaq, c); 3500 Node *neg_stride = _igvn.intcon(-cle->stride_con()); 3501 set_ctrl(neg_stride, C->root()); 3502 Node *post = new AddINode(opaq, neg_stride); 3503 register_new_node(post, c); 3504 _igvn.rehash_node_delayed(use); 3505 for (uint j = 1; j < use->req(); j++) { 3506 if (use->in(j) == phi) 3507 use->set_req(j, post); 3508 } 3509 // Since DU info changed, rerun loop 3510 progress = true; 3511 break; 3512 } 3513 } 3514 3515 }