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 "utilities/macros.hpp" 44 45 //============================================================================= 46 //------------------------------split_thru_phi--------------------------------- 47 // Split Node 'n' through merge point if there is enough win. 48 Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) { 49 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) { 50 // ConvI2L may have type information on it which is unsafe to push up 51 // so disable this for now 52 return NULL; 53 } 54 55 // Splitting range check CastIIs through a loop induction Phi can 56 // cause new Phis to be created that are left unrelated to the loop 57 // induction Phi and prevent optimizations (vectorization) 58 if (n->Opcode() == Op_CastII && n->as_CastII()->has_range_check() && 59 region->is_CountedLoop() && n->in(1) == region->as_CountedLoop()->phi()) { 60 return NULL; 61 } 62 63 int wins = 0; 64 assert(!n->is_CFG(), ""); 65 assert(region->is_Region(), ""); 66 67 const Type* type = n->bottom_type(); 68 const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr(); 69 Node *phi; 70 if (t_oop != NULL && t_oop->is_known_instance_field()) { 71 int iid = t_oop->instance_id(); 72 int index = C->get_alias_index(t_oop); 73 int offset = t_oop->offset(); 74 phi = new PhiNode(region, type, NULL, iid, index, offset); 75 } else { 76 phi = PhiNode::make_blank(region, n); 77 } 78 uint old_unique = C->unique(); 79 for (uint i = 1; i < region->req(); i++) { 80 Node *x; 81 Node* the_clone = NULL; 82 if (region->in(i) == C->top()) { 83 x = C->top(); // Dead path? Use a dead data op 84 } else { 85 x = n->clone(); // Else clone up the data op 86 the_clone = x; // Remember for possible deletion. 87 // Alter data node to use pre-phi inputs 88 if (n->in(0) == region) 89 x->set_req( 0, region->in(i) ); 90 for (uint j = 1; j < n->req(); j++) { 91 Node *in = n->in(j); 92 if (in->is_Phi() && in->in(0) == region) 93 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone 94 } 95 } 96 // Check for a 'win' on some paths 97 const Type *t = x->Value(&_igvn); 98 99 bool singleton = t->singleton(); 100 101 // A TOP singleton indicates that there are no possible values incoming 102 // along a particular edge. In most cases, this is OK, and the Phi will 103 // be eliminated later in an Ideal call. However, we can't allow this to 104 // happen if the singleton occurs on loop entry, as the elimination of 105 // the PhiNode may cause the resulting node to migrate back to a previous 106 // loop iteration. 107 if (singleton && t == Type::TOP) { 108 // Is_Loop() == false does not confirm the absence of a loop (e.g., an 109 // irreducible loop may not be indicated by an affirmative is_Loop()); 110 // therefore, the only top we can split thru a phi is on a backedge of 111 // a loop. 112 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 113 } 114 115 if (singleton) { 116 wins++; 117 x = ((PhaseGVN&)_igvn).makecon(t); 118 } else { 119 // We now call Identity to try to simplify the cloned node. 120 // Note that some Identity methods call phase->type(this). 121 // Make sure that the type array is big enough for 122 // our new node, even though we may throw the node away. 123 // (Note: This tweaking with igvn only works because x is a new node.) 124 _igvn.set_type(x, t); 125 // If x is a TypeNode, capture any more-precise type permanently into Node 126 // otherwise it will be not updated during igvn->transform since 127 // igvn->type(x) is set to x->Value() already. 128 x->raise_bottom_type(t); 129 Node *y = _igvn.apply_identity(x); 130 if (y != x) { 131 wins++; 132 x = y; 133 } else { 134 y = _igvn.hash_find(x); 135 if (y) { 136 wins++; 137 x = y; 138 } else { 139 // Else x is a new node we are keeping 140 // We do not need register_new_node_with_optimizer 141 // because set_type has already been called. 142 _igvn._worklist.push(x); 143 } 144 } 145 } 146 if (x != the_clone && the_clone != NULL) 147 _igvn.remove_dead_node(the_clone); 148 phi->set_req( i, x ); 149 } 150 // Too few wins? 151 if (wins <= policy) { 152 _igvn.remove_dead_node(phi); 153 return NULL; 154 } 155 156 // Record Phi 157 register_new_node( phi, region ); 158 159 for (uint i2 = 1; i2 < phi->req(); i2++) { 160 Node *x = phi->in(i2); 161 // If we commoned up the cloned 'x' with another existing Node, 162 // the existing Node picks up a new use. We need to make the 163 // existing Node occur higher up so it dominates its uses. 164 Node *old_ctrl; 165 IdealLoopTree *old_loop; 166 167 if (x->is_Con()) { 168 // Constant's control is always root. 169 set_ctrl(x, C->root()); 170 continue; 171 } 172 // The occasional new node 173 if (x->_idx >= old_unique) { // Found a new, unplaced node? 174 old_ctrl = NULL; 175 old_loop = NULL; // Not in any prior loop 176 } else { 177 old_ctrl = get_ctrl(x); 178 old_loop = get_loop(old_ctrl); // Get prior loop 179 } 180 // New late point must dominate new use 181 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2)); 182 if (new_ctrl == old_ctrl) // Nothing is changed 183 continue; 184 185 IdealLoopTree *new_loop = get_loop(new_ctrl); 186 187 // Don't move x into a loop if its uses are 188 // outside of loop. Otherwise x will be cloned 189 // for each use outside of this loop. 190 IdealLoopTree *use_loop = get_loop(region); 191 if (!new_loop->is_member(use_loop) && 192 (old_loop == NULL || !new_loop->is_member(old_loop))) { 193 // Take early control, later control will be recalculated 194 // during next iteration of loop optimizations. 195 new_ctrl = get_early_ctrl(x); 196 new_loop = get_loop(new_ctrl); 197 } 198 // Set new location 199 set_ctrl(x, new_ctrl); 200 // If changing loop bodies, see if we need to collect into new body 201 if (old_loop != new_loop) { 202 if (old_loop && !old_loop->_child) 203 old_loop->_body.yank(x); 204 if (!new_loop->_child) 205 new_loop->_body.push(x); // Collect body info 206 } 207 } 208 209 return phi; 210 } 211 212 //------------------------------dominated_by------------------------------------ 213 // Replace the dominated test with an obvious true or false. Place it on the 214 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the 215 // live path up to the dominating control. 216 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff, bool flip, bool exclude_loop_predicate ) { 217 if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); } 218 219 // prevdom is the dominating projection of the dominating test. 220 assert( iff->is_If(), "" ); 221 assert(iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd || iff->Opcode() == Op_RangeCheck, "Check this code when new subtype is added"); 222 int pop = prevdom->Opcode(); 223 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" ); 224 if (flip) { 225 if (pop == Op_IfTrue) 226 pop = Op_IfFalse; 227 else 228 pop = Op_IfTrue; 229 } 230 // 'con' is set to true or false to kill the dominated test. 231 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO); 232 set_ctrl(con, C->root()); // Constant gets a new use 233 // Hack the dominated test 234 _igvn.replace_input_of(iff, 1, con); 235 236 // If I dont have a reachable TRUE and FALSE path following the IfNode then 237 // I can assume this path reaches an infinite loop. In this case it's not 238 // important to optimize the data Nodes - either the whole compilation will 239 // be tossed or this path (and all data Nodes) will go dead. 240 if (iff->outcnt() != 2) return; 241 242 // Make control-dependent data Nodes on the live path (path that will remain 243 // once the dominated IF is removed) become control-dependent on the 244 // dominating projection. 245 Node* dp = iff->as_If()->proj_out_or_null(pop == Op_IfTrue); 246 247 // Loop predicates may have depending checks which should not 248 // be skipped. For example, range check predicate has two checks 249 // for lower and upper bounds. 250 if (dp == NULL) 251 return; 252 253 ProjNode* dp_proj = dp->as_Proj(); 254 ProjNode* unc_proj = iff->as_If()->proj_out(1 - dp_proj->_con)->as_Proj(); 255 if (exclude_loop_predicate && 256 (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) != NULL || 257 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_profile_predicate) != NULL || 258 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check) != NULL)) { 259 // If this is a range check (IfNode::is_range_check), do not 260 // reorder because Compile::allow_range_check_smearing might have 261 // changed the check. 262 return; // Let IGVN transformation change control dependence. 263 } 264 265 IdealLoopTree *old_loop = get_loop(dp); 266 267 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { 268 Node* cd = dp->fast_out(i); // Control-dependent node 269 if (cd->depends_only_on_test()) { 270 assert(cd->in(0) == dp, ""); 271 _igvn.replace_input_of(cd, 0, prevdom); 272 set_early_ctrl(cd); 273 IdealLoopTree *new_loop = get_loop(get_ctrl(cd)); 274 if (old_loop != new_loop) { 275 if (!old_loop->_child) old_loop->_body.yank(cd); 276 if (!new_loop->_child) new_loop->_body.push(cd); 277 } 278 --i; 279 --imax; 280 } 281 } 282 } 283 284 //------------------------------has_local_phi_input---------------------------- 285 // Return TRUE if 'n' has Phi inputs from its local block and no other 286 // block-local inputs (all non-local-phi inputs come from earlier blocks) 287 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) { 288 Node *n_ctrl = get_ctrl(n); 289 // See if some inputs come from a Phi in this block, or from before 290 // this block. 291 uint i; 292 for( i = 1; i < n->req(); i++ ) { 293 Node *phi = n->in(i); 294 if( phi->is_Phi() && phi->in(0) == n_ctrl ) 295 break; 296 } 297 if( i >= n->req() ) 298 return NULL; // No Phi inputs; nowhere to clone thru 299 300 // Check for inputs created between 'n' and the Phi input. These 301 // must split as well; they have already been given the chance 302 // (courtesy of a post-order visit) and since they did not we must 303 // recover the 'cost' of splitting them by being very profitable 304 // when splitting 'n'. Since this is unlikely we simply give up. 305 for( i = 1; i < n->req(); i++ ) { 306 Node *m = n->in(i); 307 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) { 308 // We allow the special case of AddP's with no local inputs. 309 // This allows us to split-up address expressions. 310 if (m->is_AddP() && 311 get_ctrl(m->in(2)) != n_ctrl && 312 get_ctrl(m->in(3)) != n_ctrl) { 313 // Move the AddP up to dominating point 314 Node* c = find_non_split_ctrl(idom(n_ctrl)); 315 if (c->is_OuterStripMinedLoop()) { 316 c->as_Loop()->verify_strip_mined(1); 317 c = c->in(LoopNode::EntryControl); 318 } 319 set_ctrl_and_loop(m, c); 320 continue; 321 } 322 return NULL; 323 } 324 assert(n->is_Phi() || m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control"); 325 } 326 327 return n_ctrl; 328 } 329 330 //------------------------------remix_address_expressions---------------------- 331 // Rework addressing expressions to get the most loop-invariant stuff 332 // moved out. We'd like to do all associative operators, but it's especially 333 // important (common) to do address expressions. 334 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) { 335 if (!has_ctrl(n)) return NULL; 336 Node *n_ctrl = get_ctrl(n); 337 IdealLoopTree *n_loop = get_loop(n_ctrl); 338 339 // See if 'n' mixes loop-varying and loop-invariant inputs and 340 // itself is loop-varying. 341 342 // Only interested in binary ops (and AddP) 343 if( n->req() < 3 || n->req() > 4 ) return NULL; 344 345 Node *n1_ctrl = get_ctrl(n->in( 1)); 346 Node *n2_ctrl = get_ctrl(n->in( 2)); 347 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3)); 348 IdealLoopTree *n1_loop = get_loop( n1_ctrl ); 349 IdealLoopTree *n2_loop = get_loop( n2_ctrl ); 350 IdealLoopTree *n3_loop = get_loop( n3_ctrl ); 351 352 // Does one of my inputs spin in a tighter loop than self? 353 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) || 354 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) || 355 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) ) 356 return NULL; // Leave well enough alone 357 358 // Is at least one of my inputs loop-invariant? 359 if( n1_loop == n_loop && 360 n2_loop == n_loop && 361 n3_loop == n_loop ) 362 return NULL; // No loop-invariant inputs 363 364 365 int n_op = n->Opcode(); 366 367 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2). 368 if( n_op == Op_LShiftI ) { 369 // Scale is loop invariant 370 Node *scale = n->in(2); 371 Node *scale_ctrl = get_ctrl(scale); 372 IdealLoopTree *scale_loop = get_loop(scale_ctrl ); 373 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) ) 374 return NULL; 375 const TypeInt *scale_t = scale->bottom_type()->isa_int(); 376 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 ) 377 return NULL; // Dont bother with byte/short masking 378 // Add must vary with loop (else shift would be loop-invariant) 379 Node *add = n->in(1); 380 Node *add_ctrl = get_ctrl(add); 381 IdealLoopTree *add_loop = get_loop(add_ctrl); 382 //assert( n_loop == add_loop, "" ); 383 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops 384 385 // Convert I-V into I+ (0-V); same for V-I 386 if( add->Opcode() == Op_SubI && 387 _igvn.type( add->in(1) ) != TypeInt::ZERO ) { 388 Node *zero = _igvn.intcon(0); 389 set_ctrl(zero, C->root()); 390 Node *neg = new SubINode( _igvn.intcon(0), add->in(2) ); 391 register_new_node( neg, get_ctrl(add->in(2) ) ); 392 add = new AddINode( add->in(1), neg ); 393 register_new_node( add, add_ctrl ); 394 } 395 if( add->Opcode() != Op_AddI ) return NULL; 396 // See if one add input is loop invariant 397 Node *add_var = add->in(1); 398 Node *add_var_ctrl = get_ctrl(add_var); 399 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 400 Node *add_invar = add->in(2); 401 Node *add_invar_ctrl = get_ctrl(add_invar); 402 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl ); 403 if( add_var_loop == n_loop ) { 404 } else if( add_invar_loop == n_loop ) { 405 // Swap to find the invariant part 406 add_invar = add_var; 407 add_invar_ctrl = add_var_ctrl; 408 add_invar_loop = add_var_loop; 409 add_var = add->in(2); 410 Node *add_var_ctrl = get_ctrl(add_var); 411 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 412 } else // Else neither input is loop invariant 413 return NULL; 414 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) ) 415 return NULL; // No invariant part of the add? 416 417 // Yes! Reshape address expression! 418 Node *inv_scale = new LShiftINode( add_invar, scale ); 419 Node *inv_scale_ctrl = 420 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ? 421 add_invar_ctrl : scale_ctrl; 422 register_new_node( inv_scale, inv_scale_ctrl ); 423 Node *var_scale = new LShiftINode( add_var, scale ); 424 register_new_node( var_scale, n_ctrl ); 425 Node *var_add = new AddINode( var_scale, inv_scale ); 426 register_new_node( var_add, n_ctrl ); 427 _igvn.replace_node( n, var_add ); 428 return var_add; 429 } 430 431 // Replace (I+V) with (V+I) 432 if( n_op == Op_AddI || 433 n_op == Op_AddL || 434 n_op == Op_AddF || 435 n_op == Op_AddD || 436 n_op == Op_MulI || 437 n_op == Op_MulL || 438 n_op == Op_MulF || 439 n_op == Op_MulD ) { 440 if( n2_loop == n_loop ) { 441 assert( n1_loop != n_loop, "" ); 442 n->swap_edges(1, 2); 443 } 444 } 445 446 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V), 447 // but not if I2 is a constant. 448 if( n_op == Op_AddP ) { 449 if( n2_loop == n_loop && n3_loop != n_loop ) { 450 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) { 451 Node *n22_ctrl = get_ctrl(n->in(2)->in(2)); 452 Node *n23_ctrl = get_ctrl(n->in(2)->in(3)); 453 IdealLoopTree *n22loop = get_loop( n22_ctrl ); 454 IdealLoopTree *n23_loop = get_loop( n23_ctrl ); 455 if( n22loop != n_loop && n22loop->is_member(n_loop) && 456 n23_loop == n_loop ) { 457 Node *add1 = new AddPNode( n->in(1), n->in(2)->in(2), n->in(3) ); 458 // Stuff new AddP in the loop preheader 459 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) ); 460 Node *add2 = new AddPNode( n->in(1), add1, n->in(2)->in(3) ); 461 register_new_node( add2, n_ctrl ); 462 _igvn.replace_node( n, add2 ); 463 return add2; 464 } 465 } 466 } 467 468 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V) 469 if (n2_loop != n_loop && n3_loop == n_loop) { 470 if (n->in(3)->Opcode() == Op_AddX) { 471 Node *V = n->in(3)->in(1); 472 Node *I = n->in(3)->in(2); 473 if (is_member(n_loop,get_ctrl(V))) { 474 } else { 475 Node *tmp = V; V = I; I = tmp; 476 } 477 if (!is_member(n_loop,get_ctrl(I))) { 478 Node *add1 = new AddPNode(n->in(1), n->in(2), I); 479 // Stuff new AddP in the loop preheader 480 register_new_node(add1, n_loop->_head->in(LoopNode::EntryControl)); 481 Node *add2 = new AddPNode(n->in(1), add1, V); 482 register_new_node(add2, n_ctrl); 483 _igvn.replace_node(n, add2); 484 return add2; 485 } 486 } 487 } 488 } 489 490 return NULL; 491 } 492 493 // Optimize ((in1[2*i] * in2[2*i]) + (in1[2*i+1] * in2[2*i+1])) 494 Node *PhaseIdealLoop::convert_add_to_muladd(Node* n) { 495 assert(n->Opcode() == Op_AddI, "sanity"); 496 Node * nn = NULL; 497 Node * in1 = n->in(1); 498 Node * in2 = n->in(2); 499 if (in1->Opcode() == Op_MulI && in2->Opcode() == Op_MulI) { 500 IdealLoopTree* loop_n = get_loop(get_ctrl(n)); 501 if (loop_n->_head->as_Loop()->is_valid_counted_loop() && 502 Matcher::match_rule_supported(Op_MulAddS2I) && 503 Matcher::match_rule_supported(Op_MulAddVS2VI)) { 504 Node* mul_in1 = in1->in(1); 505 Node* mul_in2 = in1->in(2); 506 Node* mul_in3 = in2->in(1); 507 Node* mul_in4 = in2->in(2); 508 if (mul_in1->Opcode() == Op_LoadS && 509 mul_in2->Opcode() == Op_LoadS && 510 mul_in3->Opcode() == Op_LoadS && 511 mul_in4->Opcode() == Op_LoadS) { 512 IdealLoopTree* loop1 = get_loop(get_ctrl(mul_in1)); 513 IdealLoopTree* loop2 = get_loop(get_ctrl(mul_in2)); 514 IdealLoopTree* loop3 = get_loop(get_ctrl(mul_in3)); 515 IdealLoopTree* loop4 = get_loop(get_ctrl(mul_in4)); 516 IdealLoopTree* loop5 = get_loop(get_ctrl(in1)); 517 IdealLoopTree* loop6 = get_loop(get_ctrl(in2)); 518 // All nodes should be in the same counted loop. 519 if (loop_n == loop1 && loop_n == loop2 && loop_n == loop3 && 520 loop_n == loop4 && loop_n == loop5 && loop_n == loop6) { 521 Node* adr1 = mul_in1->in(MemNode::Address); 522 Node* adr2 = mul_in2->in(MemNode::Address); 523 Node* adr3 = mul_in3->in(MemNode::Address); 524 Node* adr4 = mul_in4->in(MemNode::Address); 525 if (adr1->is_AddP() && adr2->is_AddP() && adr3->is_AddP() && adr4->is_AddP()) { 526 if ((adr1->in(AddPNode::Base) == adr3->in(AddPNode::Base)) && 527 (adr2->in(AddPNode::Base) == adr4->in(AddPNode::Base))) { 528 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in3, mul_in4); 529 register_new_node(nn, get_ctrl(n)); 530 _igvn.replace_node(n, nn); 531 return nn; 532 } else if ((adr1->in(AddPNode::Base) == adr4->in(AddPNode::Base)) && 533 (adr2->in(AddPNode::Base) == adr3->in(AddPNode::Base))) { 534 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in4, mul_in3); 535 register_new_node(nn, get_ctrl(n)); 536 _igvn.replace_node(n, nn); 537 return nn; 538 } 539 } 540 } 541 } 542 } 543 } 544 return nn; 545 } 546 547 //------------------------------conditional_move------------------------------- 548 // Attempt to replace a Phi with a conditional move. We have some pretty 549 // strict profitability requirements. All Phis at the merge point must 550 // be converted, so we can remove the control flow. We need to limit the 551 // number of c-moves to a small handful. All code that was in the side-arms 552 // of the CFG diamond is now speculatively executed. This code has to be 553 // "cheap enough". We are pretty much limited to CFG diamonds that merge 554 // 1 or 2 items with a total of 1 or 2 ops executed speculatively. 555 Node *PhaseIdealLoop::conditional_move( Node *region ) { 556 557 assert(region->is_Region(), "sanity check"); 558 if (region->req() != 3) return NULL; 559 560 // Check for CFG diamond 561 Node *lp = region->in(1); 562 Node *rp = region->in(2); 563 if (!lp || !rp) return NULL; 564 Node *lp_c = lp->in(0); 565 if (lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If()) return NULL; 566 IfNode *iff = lp_c->as_If(); 567 568 // Check for ops pinned in an arm of the diamond. 569 // Can't remove the control flow in this case 570 if (lp->outcnt() > 1) return NULL; 571 if (rp->outcnt() > 1) return NULL; 572 573 IdealLoopTree* r_loop = get_loop(region); 574 assert(r_loop == get_loop(iff), "sanity"); 575 // Always convert to CMOVE if all results are used only outside this loop. 576 bool used_inside_loop = (r_loop == _ltree_root); 577 578 // Check profitability 579 int cost = 0; 580 int phis = 0; 581 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 582 Node *out = region->fast_out(i); 583 if (!out->is_Phi()) continue; // Ignore other control edges, etc 584 phis++; 585 PhiNode* phi = out->as_Phi(); 586 BasicType bt = phi->type()->basic_type(); 587 switch (bt) { 588 case T_DOUBLE: 589 case T_FLOAT: 590 if (C->use_cmove()) { 591 continue; //TODO: maybe we want to add some cost 592 } 593 cost += Matcher::float_cmove_cost(); // Could be very expensive 594 break; 595 case T_LONG: { 596 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's 597 } 598 case T_INT: // These all CMOV fine 599 case T_ADDRESS: { // (RawPtr) 600 cost++; 601 break; 602 } 603 case T_NARROWOOP: // Fall through 604 case T_OBJECT: { // Base oops are OK, but not derived oops 605 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr(); 606 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a 607 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus 608 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we 609 // have a Phi for the base here that we convert to a CMOVE all is well 610 // and good. But if the base is dead, we'll not make a CMOVE. Later 611 // the allocator will have to produce a base by creating a CMOVE of the 612 // relevant bases. This puts the allocator in the business of 613 // manufacturing expensive instructions, generally a bad plan. 614 // Just Say No to Conditionally-Moved Derived Pointers. 615 if (tp && tp->offset() != 0) 616 return NULL; 617 cost++; 618 break; 619 } 620 default: 621 return NULL; // In particular, can't do memory or I/O 622 } 623 // Add in cost any speculative ops 624 for (uint j = 1; j < region->req(); j++) { 625 Node *proj = region->in(j); 626 Node *inp = phi->in(j); 627 if (get_ctrl(inp) == proj) { // Found local op 628 cost++; 629 // Check for a chain of dependent ops; these will all become 630 // speculative in a CMOV. 631 for (uint k = 1; k < inp->req(); k++) 632 if (get_ctrl(inp->in(k)) == proj) 633 cost += ConditionalMoveLimit; // Too much speculative goo 634 } 635 } 636 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode. 637 // This will likely Split-If, a higher-payoff operation. 638 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) { 639 Node* use = phi->fast_out(k); 640 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr()) 641 cost += ConditionalMoveLimit; 642 // Is there a use inside the loop? 643 // Note: check only basic types since CMoveP is pinned. 644 if (!used_inside_loop && is_java_primitive(bt)) { 645 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use); 646 if (r_loop == u_loop || r_loop->is_member(u_loop)) { 647 used_inside_loop = true; 648 } 649 } 650 } 651 }//for 652 Node* bol = iff->in(1); 653 if (bol->Opcode() == Op_Opaque4) { 654 return NULL; // Ignore loop predicate checks (the Opaque4 ensures they will go away) 655 } 656 assert(bol->Opcode() == Op_Bool, "Unexpected node"); 657 int cmp_op = bol->in(1)->Opcode(); 658 // It is expensive to generate flags from a float compare. 659 // Avoid duplicated float compare. 660 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL; 661 662 float infrequent_prob = PROB_UNLIKELY_MAG(3); 663 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop. 664 if (used_inside_loop) { 665 if (cost >= ConditionalMoveLimit) return NULL; // Too much goo 666 667 // BlockLayoutByFrequency optimization moves infrequent branch 668 // from hot path. No point in CMOV'ing in such case (110 is used 669 // instead of 100 to take into account not exactness of float value). 670 if (BlockLayoutByFrequency) { 671 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f); 672 } 673 } 674 // Check for highly predictable branch. No point in CMOV'ing if 675 // we are going to predict accurately all the time. 676 if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) { 677 //keep going 678 } else if (iff->_prob < infrequent_prob || 679 iff->_prob > (1.0f - infrequent_prob)) 680 return NULL; 681 682 // -------------- 683 // Now replace all Phis with CMOV's 684 Node *cmov_ctrl = iff->in(0); 685 uint flip = (lp->Opcode() == Op_IfTrue); 686 Node_List wq; 687 while (1) { 688 PhiNode* phi = NULL; 689 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 690 Node *out = region->fast_out(i); 691 if (out->is_Phi()) { 692 phi = out->as_Phi(); 693 break; 694 } 695 } 696 if (phi == NULL) break; 697 if (PrintOpto && VerifyLoopOptimizations) { tty->print_cr("CMOV"); } 698 // Move speculative ops 699 wq.push(phi); 700 while (wq.size() > 0) { 701 Node *n = wq.pop(); 702 for (uint j = 1; j < n->req(); j++) { 703 Node* m = n->in(j); 704 if (m != NULL && !is_dominator(get_ctrl(m), cmov_ctrl)) { 705 #ifndef PRODUCT 706 if (PrintOpto && VerifyLoopOptimizations) { 707 tty->print(" speculate: "); 708 m->dump(); 709 } 710 #endif 711 set_ctrl(m, cmov_ctrl); 712 wq.push(m); 713 } 714 } 715 } 716 Node *cmov = CMoveNode::make(cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi)); 717 register_new_node( cmov, cmov_ctrl ); 718 _igvn.replace_node( phi, cmov ); 719 #ifndef PRODUCT 720 if (TraceLoopOpts) { 721 tty->print("CMOV "); 722 r_loop->dump_head(); 723 if (Verbose) { 724 bol->in(1)->dump(1); 725 cmov->dump(1); 726 } 727 } 728 if (VerifyLoopOptimizations) verify(); 729 #endif 730 } 731 732 // The useless CFG diamond will fold up later; see the optimization in 733 // RegionNode::Ideal. 734 _igvn._worklist.push(region); 735 736 return iff->in(1); 737 } 738 739 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) { 740 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 741 Node* u = m->fast_out(i); 742 if (u->is_CFG()) { 743 if (u->Opcode() == Op_NeverBranch) { 744 u = ((NeverBranchNode*)u)->proj_out(0); 745 enqueue_cfg_uses(u, wq); 746 } else { 747 wq.push(u); 748 } 749 } 750 } 751 } 752 753 // Try moving a store out of a loop, right before the loop 754 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) { 755 // Store has to be first in the loop body 756 IdealLoopTree *n_loop = get_loop(n_ctrl); 757 if (n->is_Store() && n_loop != _ltree_root && 758 n_loop->is_loop() && n_loop->_head->is_Loop() && 759 n->in(0) != NULL) { 760 Node* address = n->in(MemNode::Address); 761 Node* value = n->in(MemNode::ValueIn); 762 Node* mem = n->in(MemNode::Memory); 763 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 764 IdealLoopTree* value_loop = get_loop(get_ctrl(value)); 765 766 // - address and value must be loop invariant 767 // - memory must be a memory Phi for the loop 768 // - Store must be the only store on this memory slice in the 769 // loop: if there's another store following this one then value 770 // written at iteration i by the second store could be overwritten 771 // at iteration i+n by the first store: it's not safe to move the 772 // first store out of the loop 773 // - nothing must observe the memory Phi: it guarantees no read 774 // before the store, we are also guaranteed the store post 775 // dominates the loop head (ignoring a possible early 776 // exit). Otherwise there would be extra Phi involved between the 777 // loop's Phi and the store. 778 // - there must be no early exit from the loop before the Store 779 // (such an exit most of the time would be an extra use of the 780 // memory Phi but sometimes is a bottom memory Phi that takes the 781 // store as input). 782 783 if (!n_loop->is_member(address_loop) && 784 !n_loop->is_member(value_loop) && 785 mem->is_Phi() && mem->in(0) == n_loop->_head && 786 mem->outcnt() == 1 && 787 mem->in(LoopNode::LoopBackControl) == n) { 788 789 assert(n_loop->_tail != NULL, "need a tail"); 790 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop"); 791 792 // Verify that there's no early exit of the loop before the store. 793 bool ctrl_ok = false; 794 { 795 // Follow control from loop head until n, we exit the loop or 796 // we reach the tail 797 ResourceMark rm; 798 Unique_Node_List wq; 799 wq.push(n_loop->_head); 800 801 for (uint next = 0; next < wq.size(); ++next) { 802 Node *m = wq.at(next); 803 if (m == n->in(0)) { 804 ctrl_ok = true; 805 continue; 806 } 807 assert(!has_ctrl(m), "should be CFG"); 808 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) { 809 ctrl_ok = false; 810 break; 811 } 812 enqueue_cfg_uses(m, wq); 813 if (wq.size() > 10) { 814 ctrl_ok = false; 815 break; 816 } 817 } 818 } 819 if (ctrl_ok) { 820 // move the Store 821 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem); 822 _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl)); 823 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl)); 824 // Disconnect the phi now. An empty phi can confuse other 825 // optimizations in this pass of loop opts. 826 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl)); 827 n_loop->_body.yank(mem); 828 829 set_ctrl_and_loop(n, n->in(0)); 830 831 return n; 832 } 833 } 834 } 835 return NULL; 836 } 837 838 // Try moving a store out of a loop, right after the loop 839 void PhaseIdealLoop::try_move_store_after_loop(Node* n) { 840 if (n->is_Store() && n->in(0) != NULL) { 841 Node *n_ctrl = get_ctrl(n); 842 IdealLoopTree *n_loop = get_loop(n_ctrl); 843 // Store must be in a loop 844 if (n_loop != _ltree_root && !n_loop->_irreducible) { 845 Node* address = n->in(MemNode::Address); 846 Node* value = n->in(MemNode::ValueIn); 847 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 848 // address must be loop invariant 849 if (!n_loop->is_member(address_loop)) { 850 // Store must be last on this memory slice in the loop and 851 // nothing in the loop must observe it 852 Node* phi = NULL; 853 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 854 Node* u = n->fast_out(i); 855 if (has_ctrl(u)) { // control use? 856 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 857 if (!n_loop->is_member(u_loop)) { 858 continue; 859 } 860 if (u->is_Phi() && u->in(0) == n_loop->_head) { 861 assert(_igvn.type(u) == Type::MEMORY, "bad phi"); 862 // multiple phis on the same slice are possible 863 if (phi != NULL) { 864 return; 865 } 866 phi = u; 867 continue; 868 } 869 } 870 return; 871 } 872 if (phi != NULL) { 873 // Nothing in the loop before the store (next iteration) 874 // must observe the stored value 875 bool mem_ok = true; 876 { 877 ResourceMark rm; 878 Unique_Node_List wq; 879 wq.push(phi); 880 for (uint next = 0; next < wq.size() && mem_ok; ++next) { 881 Node *m = wq.at(next); 882 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) { 883 Node* u = m->fast_out(i); 884 if (u->is_Store() || u->is_Phi()) { 885 if (u != n) { 886 wq.push(u); 887 mem_ok = (wq.size() <= 10); 888 } 889 } else { 890 mem_ok = false; 891 break; 892 } 893 } 894 } 895 } 896 if (mem_ok) { 897 // Move the store out of the loop if the LCA of all 898 // users (except for the phi) is outside the loop. 899 Node* hook = new Node(1); 900 _igvn.rehash_node_delayed(phi); 901 int count = phi->replace_edge(n, hook); 902 assert(count > 0, "inconsistent phi"); 903 904 // Compute latest point this store can go 905 Node* lca = get_late_ctrl(n, get_ctrl(n)); 906 if (n_loop->is_member(get_loop(lca))) { 907 // LCA is in the loop - bail out 908 _igvn.replace_node(hook, n); 909 return; 910 } 911 #ifdef ASSERT 912 if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) { 913 assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined"); 914 n_loop->_head->as_Loop()->verify_strip_mined(1); 915 Node* outer = n_loop->_head->as_CountedLoop()->outer_loop(); 916 IdealLoopTree* outer_loop = get_loop(outer); 917 assert(n_loop->_parent == outer_loop, "broken loop tree"); 918 assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state"); 919 } 920 #endif 921 922 // Move store out of the loop 923 _igvn.replace_node(hook, n->in(MemNode::Memory)); 924 _igvn.replace_input_of(n, 0, lca); 925 set_ctrl_and_loop(n, lca); 926 927 // Disconnect the phi now. An empty phi can confuse other 928 // optimizations in this pass of loop opts.. 929 if (phi->in(LoopNode::LoopBackControl) == phi) { 930 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl)); 931 n_loop->_body.yank(phi); 932 } 933 } 934 } 935 } 936 } 937 } 938 } 939 940 //------------------------------split_if_with_blocks_pre----------------------- 941 // Do the real work in a non-recursive function. Data nodes want to be 942 // cloned in the pre-order so they can feed each other nicely. 943 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) { 944 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 945 Node* bs_res = bs->split_if_pre(this, n); 946 if (bs_res != NULL) { 947 return bs_res; 948 } 949 // Cloning these guys is unlikely to win 950 int n_op = n->Opcode(); 951 if( n_op == Op_MergeMem ) return n; 952 if( n->is_Proj() ) return n; 953 // Do not clone-up CmpFXXX variations, as these are always 954 // followed by a CmpI 955 if( n->is_Cmp() ) return n; 956 // Attempt to use a conditional move instead of a phi/branch 957 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) { 958 Node *cmov = conditional_move( n ); 959 if( cmov ) return cmov; 960 } 961 if( n->is_CFG() || n->is_LoadStore() ) 962 return n; 963 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd 964 n_op == Op_Opaque2 ) { 965 if( !C->major_progress() ) // If chance of no more loop opts... 966 _igvn._worklist.push(n); // maybe we'll remove them 967 return n; 968 } 969 970 if( n->is_Con() ) return n; // No cloning for Con nodes 971 972 Node *n_ctrl = get_ctrl(n); 973 if( !n_ctrl ) return n; // Dead node 974 975 Node* res = try_move_store_before_loop(n, n_ctrl); 976 if (res != NULL) { 977 return n; 978 } 979 980 // Attempt to remix address expressions for loop invariants 981 Node *m = remix_address_expressions( n ); 982 if( m ) return m; 983 984 if (n_op == Op_AddI) { 985 Node *nn = convert_add_to_muladd( n ); 986 if ( nn ) return nn; 987 } 988 989 if (n->is_ConstraintCast()) { 990 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this); 991 // ConstraintCastNode::dominating_cast() uses node control input to determine domination. 992 // Node control inputs don't necessarily agree with loop control info (due to 993 // transformations happened in between), thus additional dominance check is needed 994 // to keep loop info valid. 995 if (dom_cast != NULL && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) { 996 _igvn.replace_node(n, dom_cast); 997 return dom_cast; 998 } 999 } 1000 1001 // Determine if the Node has inputs from some local Phi. 1002 // Returns the block to clone thru. 1003 Node *n_blk = has_local_phi_input( n ); 1004 if( !n_blk ) return n; 1005 1006 // Do not clone the trip counter through on a CountedLoop 1007 // (messes up the canonical shape). 1008 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n; 1009 1010 // Check for having no control input; not pinned. Allow 1011 // dominating control. 1012 if (n->in(0)) { 1013 Node *dom = idom(n_blk); 1014 if (dom_lca(n->in(0), dom) != n->in(0)) { 1015 return n; 1016 } 1017 } 1018 // Policy: when is it profitable. You must get more wins than 1019 // policy before it is considered profitable. Policy is usually 0, 1020 // so 1 win is considered profitable. Big merges will require big 1021 // cloning, so get a larger policy. 1022 int policy = n_blk->req() >> 2; 1023 1024 // If the loop is a candidate for range check elimination, 1025 // delay splitting through it's phi until a later loop optimization 1026 if (n_blk->is_CountedLoop()) { 1027 IdealLoopTree *lp = get_loop(n_blk); 1028 if (lp && lp->_rce_candidate) { 1029 return n; 1030 } 1031 } 1032 1033 if (must_throttle_split_if()) return n; 1034 1035 // Split 'n' through the merge point if it is profitable 1036 Node *phi = split_thru_phi( n, n_blk, policy ); 1037 if (!phi) return n; 1038 1039 // Found a Phi to split thru! 1040 // Replace 'n' with the new phi 1041 _igvn.replace_node( n, phi ); 1042 // Moved a load around the loop, 'en-registering' something. 1043 if (n_blk->is_Loop() && n->is_Load() && 1044 !phi->in(LoopNode::LoopBackControl)->is_Load()) 1045 C->set_major_progress(); 1046 1047 return phi; 1048 } 1049 1050 static bool merge_point_too_heavy(Compile* C, Node* region) { 1051 // Bail out if the region and its phis have too many users. 1052 int weight = 0; 1053 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1054 weight += region->fast_out(i)->outcnt(); 1055 } 1056 int nodes_left = C->max_node_limit() - C->live_nodes(); 1057 if (weight * 8 > nodes_left) { 1058 if (PrintOpto) { 1059 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight); 1060 } 1061 return true; 1062 } else { 1063 return false; 1064 } 1065 } 1066 1067 static bool merge_point_safe(Node* region) { 1068 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode 1069 // having a PhiNode input. This sidesteps the dangerous case where the split 1070 // ConvI2LNode may become TOP if the input Value() does not 1071 // overlap the ConvI2L range, leaving a node which may not dominate its 1072 // uses. 1073 // A better fix for this problem can be found in the BugTraq entry, but 1074 // expediency for Mantis demands this hack. 1075 // 6855164: If the merge point has a FastLockNode with a PhiNode input, we stop 1076 // split_if_with_blocks from splitting a block because we could not move around 1077 // the FastLockNode. 1078 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1079 Node* n = region->fast_out(i); 1080 if (n->is_Phi()) { 1081 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1082 Node* m = n->fast_out(j); 1083 if (m->is_FastLock()) 1084 return false; 1085 #ifdef _LP64 1086 if (m->Opcode() == Op_ConvI2L) 1087 return false; 1088 if (m->is_CastII() && m->isa_CastII()->has_range_check()) { 1089 return false; 1090 } 1091 #endif 1092 } 1093 } 1094 } 1095 return true; 1096 } 1097 1098 1099 //------------------------------place_near_use--------------------------------- 1100 // Place some computation next to use but not inside inner loops. 1101 // For inner loop uses move it to the preheader area. 1102 Node *PhaseIdealLoop::place_near_use(Node *useblock) const { 1103 IdealLoopTree *u_loop = get_loop( useblock ); 1104 if (u_loop->_irreducible) { 1105 return useblock; 1106 } 1107 if (u_loop->_child) { 1108 if (useblock == u_loop->_head && u_loop->_head->is_OuterStripMinedLoop()) { 1109 return u_loop->_head->in(LoopNode::EntryControl); 1110 } 1111 return useblock; 1112 } 1113 return u_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl); 1114 } 1115 1116 1117 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) { 1118 if (!n->is_If() || n->is_CountedLoopEnd()) { 1119 return false; 1120 } 1121 if (!n->in(0)->is_Region()) { 1122 return false; 1123 } 1124 Node* region = n->in(0); 1125 Node* dom = idom(region); 1126 if (!dom->is_If() || dom->in(1) != n->in(1)) { 1127 return false; 1128 } 1129 IfNode* dom_if = dom->as_If(); 1130 Node* proj_true = dom_if->proj_out(1); 1131 Node* proj_false = dom_if->proj_out(0); 1132 1133 for (uint i = 1; i < region->req(); i++) { 1134 if (is_dominator(proj_true, region->in(i))) { 1135 continue; 1136 } 1137 if (is_dominator(proj_false, region->in(i))) { 1138 continue; 1139 } 1140 return false; 1141 } 1142 1143 return true; 1144 } 1145 1146 1147 bool PhaseIdealLoop::can_split_if(Node* n_ctrl) { 1148 if (must_throttle_split_if()) { 1149 return false; 1150 } 1151 1152 // Do not do 'split-if' if irreducible loops are present. 1153 if (_has_irreducible_loops) { 1154 return false; 1155 } 1156 1157 if (merge_point_too_heavy(C, n_ctrl)) { 1158 return false; 1159 } 1160 1161 // Do not do 'split-if' if some paths are dead. First do dead code 1162 // elimination and then see if its still profitable. 1163 for (uint i = 1; i < n_ctrl->req(); i++) { 1164 if (n_ctrl->in(i) == C->top()) { 1165 return false; 1166 } 1167 } 1168 1169 // If trying to do a 'Split-If' at the loop head, it is only 1170 // profitable if the cmp folds up on BOTH paths. Otherwise we 1171 // risk peeling a loop forever. 1172 1173 // CNC - Disabled for now. Requires careful handling of loop 1174 // body selection for the cloned code. Also, make sure we check 1175 // for any input path not being in the same loop as n_ctrl. For 1176 // irreducible loops we cannot check for 'n_ctrl->is_Loop()' 1177 // because the alternative loop entry points won't be converted 1178 // into LoopNodes. 1179 IdealLoopTree *n_loop = get_loop(n_ctrl); 1180 for (uint j = 1; j < n_ctrl->req(); j++) { 1181 if (get_loop(n_ctrl->in(j)) != n_loop) { 1182 return false; 1183 } 1184 } 1185 1186 // Check for safety of the merge point. 1187 if (!merge_point_safe(n_ctrl)) { 1188 return false; 1189 } 1190 1191 return true; 1192 } 1193 1194 // Detect if the node is the inner strip-mined loop 1195 // Return: NULL if it's not the case, or the exit of outer strip-mined loop 1196 static Node* is_inner_of_stripmined_loop(const Node* out) { 1197 Node* out_le = NULL; 1198 1199 if (out->is_CountedLoopEnd()) { 1200 const CountedLoopNode* loop = out->as_CountedLoopEnd()->loopnode(); 1201 1202 if (loop != NULL && loop->is_strip_mined()) { 1203 out_le = loop->in(LoopNode::EntryControl)->as_OuterStripMinedLoop()->outer_loop_exit(); 1204 } 1205 } 1206 1207 return out_le; 1208 } 1209 1210 //------------------------------split_if_with_blocks_post---------------------- 1211 // Do the real work in a non-recursive function. CFG hackery wants to be 1212 // in the post-order, so it can dirty the I-DOM info and not use the dirtied 1213 // info. 1214 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) { 1215 1216 // Cloning Cmp through Phi's involves the split-if transform. 1217 // FastLock is not used by an If 1218 if (n->is_Cmp() && !n->is_FastLock()) { 1219 Node *n_ctrl = get_ctrl(n); 1220 // Determine if the Node has inputs from some local Phi. 1221 // Returns the block to clone thru. 1222 Node *n_blk = has_local_phi_input(n); 1223 if (n_blk != n_ctrl) { 1224 return; 1225 } 1226 1227 if (!can_split_if(n_ctrl)) { 1228 return; 1229 } 1230 1231 if (n->outcnt() != 1) { 1232 return; // Multiple bool's from 1 compare? 1233 } 1234 Node *bol = n->unique_out(); 1235 assert(bol->is_Bool(), "expect a bool here"); 1236 if (bol->outcnt() != 1) { 1237 return;// Multiple branches from 1 compare? 1238 } 1239 Node *iff = bol->unique_out(); 1240 1241 // Check some safety conditions 1242 if (iff->is_If()) { // Classic split-if? 1243 if (iff->in(0) != n_ctrl) { 1244 return; // Compare must be in same blk as if 1245 } 1246 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE 1247 // Can't split CMove with different control edge. 1248 if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) { 1249 return; 1250 } 1251 if (get_ctrl(iff->in(2)) == n_ctrl || 1252 get_ctrl(iff->in(3)) == n_ctrl) { 1253 return; // Inputs not yet split-up 1254 } 1255 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) { 1256 return; // Loop-invar test gates loop-varying CMOVE 1257 } 1258 } else { 1259 return; // some other kind of node, such as an Allocate 1260 } 1261 1262 // When is split-if profitable? Every 'win' on means some control flow 1263 // goes dead, so it's almost always a win. 1264 int policy = 0; 1265 // Split compare 'n' through the merge point if it is profitable 1266 Node *phi = split_thru_phi( n, n_ctrl, policy); 1267 if (!phi) { 1268 return; 1269 } 1270 1271 // Found a Phi to split thru! 1272 // Replace 'n' with the new phi 1273 _igvn.replace_node(n, phi); 1274 1275 // Now split the bool up thru the phi 1276 Node *bolphi = split_thru_phi(bol, n_ctrl, -1); 1277 guarantee(bolphi != NULL, "null boolean phi node"); 1278 1279 _igvn.replace_node(bol, bolphi); 1280 assert(iff->in(1) == bolphi, ""); 1281 1282 if (bolphi->Value(&_igvn)->singleton()) { 1283 return; 1284 } 1285 1286 // Conditional-move? Must split up now 1287 if (!iff->is_If()) { 1288 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1); 1289 _igvn.replace_node(iff, cmovphi); 1290 return; 1291 } 1292 1293 // Now split the IF 1294 do_split_if(iff); 1295 return; 1296 } 1297 1298 // Two identical ifs back to back can be merged 1299 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) { 1300 Node *n_ctrl = n->in(0); 1301 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1)); 1302 IfNode* dom_if = idom(n_ctrl)->as_If(); 1303 Node* proj_true = dom_if->proj_out(1); 1304 Node* proj_false = dom_if->proj_out(0); 1305 Node* con_true = _igvn.makecon(TypeInt::ONE); 1306 Node* con_false = _igvn.makecon(TypeInt::ZERO); 1307 1308 for (uint i = 1; i < n_ctrl->req(); i++) { 1309 if (is_dominator(proj_true, n_ctrl->in(i))) { 1310 bolphi->init_req(i, con_true); 1311 } else { 1312 assert(is_dominator(proj_false, n_ctrl->in(i)), "bad if"); 1313 bolphi->init_req(i, con_false); 1314 } 1315 } 1316 register_new_node(bolphi, n_ctrl); 1317 _igvn.replace_input_of(n, 1, bolphi); 1318 1319 // Now split the IF 1320 do_split_if(n); 1321 return; 1322 } 1323 1324 // Check for an IF ready to split; one that has its 1325 // condition codes input coming from a Phi at the block start. 1326 int n_op = n->Opcode(); 1327 1328 // Check for an IF being dominated by another IF same test 1329 if (n_op == Op_If || 1330 n_op == Op_RangeCheck) { 1331 Node *bol = n->in(1); 1332 uint max = bol->outcnt(); 1333 // Check for same test used more than once? 1334 if (max > 1 && bol->is_Bool()) { 1335 // Search up IDOMs to see if this IF is dominated. 1336 Node *cutoff = get_ctrl(bol); 1337 1338 // Now search up IDOMs till cutoff, looking for a dominating test 1339 Node *prevdom = n; 1340 Node *dom = idom(prevdom); 1341 while (dom != cutoff) { 1342 if (dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom) { 1343 // It's invalid to move control dependent data nodes in the inner 1344 // strip-mined loop, because: 1345 // 1) break validation of LoopNode::verify_strip_mined() 1346 // 2) move code with side-effect in strip-mined loop 1347 // Move to the exit of outer strip-mined loop in that case. 1348 Node* out_le = is_inner_of_stripmined_loop(dom); 1349 if (out_le != NULL) { 1350 prevdom = out_le; 1351 } 1352 // Replace the dominated test with an obvious true or false. 1353 // Place it on the IGVN worklist for later cleanup. 1354 C->set_major_progress(); 1355 dominated_by(prevdom, n, false, true); 1356 #ifndef PRODUCT 1357 if( VerifyLoopOptimizations ) verify(); 1358 #endif 1359 return; 1360 } 1361 prevdom = dom; 1362 dom = idom(prevdom); 1363 } 1364 } 1365 } 1366 1367 // See if a shared loop-varying computation has no loop-varying uses. 1368 // Happens if something is only used for JVM state in uncommon trap exits, 1369 // like various versions of induction variable+offset. Clone the 1370 // computation per usage to allow it to sink out of the loop. 1371 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about) 1372 Node *n_ctrl = get_ctrl(n); 1373 IdealLoopTree *n_loop = get_loop(n_ctrl); 1374 if( n_loop != _ltree_root ) { 1375 DUIterator_Fast imax, i = n->fast_outs(imax); 1376 for (; i < imax; i++) { 1377 Node* u = n->fast_out(i); 1378 if( !has_ctrl(u) ) break; // Found control user 1379 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 1380 if( u_loop == n_loop ) break; // Found loop-varying use 1381 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop 1382 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003 1383 } 1384 bool did_break = (i < imax); // Did we break out of the previous loop? 1385 if (!did_break && n->outcnt() > 1) { // All uses in outer loops! 1386 Node *late_load_ctrl = NULL; 1387 if (n->is_Load()) { 1388 // If n is a load, get and save the result from get_late_ctrl(), 1389 // to be later used in calculating the control for n's clones. 1390 clear_dom_lca_tags(); 1391 late_load_ctrl = get_late_ctrl(n, n_ctrl); 1392 } 1393 // If n is a load, and the late control is the same as the current 1394 // control, then the cloning of n is a pointless exercise, because 1395 // GVN will ensure that we end up where we started. 1396 if (!n->is_Load() || late_load_ctrl != n_ctrl) { 1397 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1398 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) { 1399 Node *u = n->last_out(j); // Clone private computation per use 1400 _igvn.rehash_node_delayed(u); 1401 Node *x = n->clone(); // Clone computation 1402 Node *x_ctrl = NULL; 1403 if( u->is_Phi() ) { 1404 // Replace all uses of normal nodes. Replace Phi uses 1405 // individually, so the separate Nodes can sink down 1406 // different paths. 1407 uint k = 1; 1408 while( u->in(k) != n ) k++; 1409 u->set_req( k, x ); 1410 // x goes next to Phi input path 1411 x_ctrl = u->in(0)->in(k); 1412 --j; 1413 } else { // Normal use 1414 // Replace all uses 1415 for( uint k = 0; k < u->req(); k++ ) { 1416 if( u->in(k) == n ) { 1417 u->set_req( k, x ); 1418 --j; 1419 } 1420 } 1421 x_ctrl = get_ctrl(u); 1422 } 1423 1424 // Find control for 'x' next to use but not inside inner loops. 1425 // For inner loop uses get the preheader area. 1426 x_ctrl = place_near_use(x_ctrl); 1427 1428 if (bs->sink_node(this, n, x, x_ctrl, n_ctrl)) { 1429 continue; 1430 } 1431 1432 if (n->is_Load()) { 1433 // For loads, add a control edge to a CFG node outside of the loop 1434 // to force them to not combine and return back inside the loop 1435 // during GVN optimization (4641526). 1436 // 1437 // Because we are setting the actual control input, factor in 1438 // the result from get_late_ctrl() so we respect any 1439 // anti-dependences. (6233005). 1440 x_ctrl = dom_lca(late_load_ctrl, x_ctrl); 1441 1442 // Don't allow the control input to be a CFG splitting node. 1443 // Such nodes should only have ProjNodes as outs, e.g. IfNode 1444 // should only have IfTrueNode and IfFalseNode (4985384). 1445 x_ctrl = find_non_split_ctrl(x_ctrl); 1446 1447 IdealLoopTree* x_loop = get_loop(x_ctrl); 1448 Node* x_head = x_loop->_head; 1449 if (x_head->is_Loop() && (x_head->is_OuterStripMinedLoop() || x_head->as_Loop()->is_strip_mined()) && is_dominator(n_ctrl, x_head)) { 1450 // Anti dependence analysis is sometimes too 1451 // conservative: a store in the outer strip mined loop 1452 // can prevent a load from floating out of the outer 1453 // strip mined loop but the load may not be referenced 1454 // from the safepoint: loop strip mining verification 1455 // code reports a problem in that case. Make sure the 1456 // load is not moved in the outer strip mined loop in 1457 // that case. 1458 x_ctrl = x_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl); 1459 } 1460 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone"); 1461 1462 x->set_req(0, x_ctrl); 1463 } 1464 register_new_node(x, x_ctrl); 1465 1466 // Some institutional knowledge is needed here: 'x' is 1467 // yanked because if the optimizer runs GVN on it all the 1468 // cloned x's will common up and undo this optimization and 1469 // be forced back in the loop. 1470 // I tried setting control edges on the x's to force them to 1471 // not combine, but the matching gets worried when it tries 1472 // to fold a StoreP and an AddP together (as part of an 1473 // address expression) and the AddP and StoreP have 1474 // different controls. 1475 if (!x->is_Load() && !x->is_DecodeNarrowPtr()) _igvn._worklist.yank(x); 1476 } 1477 _igvn.remove_dead_node(n); 1478 } 1479 } 1480 } 1481 } 1482 1483 try_move_store_after_loop(n); 1484 1485 // Check for Opaque2's who's loop has disappeared - who's input is in the 1486 // same loop nest as their output. Remove 'em, they are no longer useful. 1487 if( n_op == Op_Opaque2 && 1488 n->in(1) != NULL && 1489 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) { 1490 _igvn.replace_node( n, n->in(1) ); 1491 } 1492 } 1493 1494 //------------------------------split_if_with_blocks--------------------------- 1495 // Check for aggressive application of 'split-if' optimization, 1496 // using basic block level info. 1497 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack) { 1498 Node* root = C->root(); 1499 visited.set(root->_idx); // first, mark root as visited 1500 // Do pre-visit work for root 1501 Node* n = split_if_with_blocks_pre(root); 1502 uint cnt = n->outcnt(); 1503 uint i = 0; 1504 1505 while (true) { 1506 // Visit all children 1507 if (i < cnt) { 1508 Node* use = n->raw_out(i); 1509 ++i; 1510 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) { 1511 // Now do pre-visit work for this use 1512 use = split_if_with_blocks_pre(use); 1513 nstack.push(n, i); // Save parent and next use's index. 1514 n = use; // Process all children of current use. 1515 cnt = use->outcnt(); 1516 i = 0; 1517 } 1518 } 1519 else { 1520 // All of n's children have been processed, complete post-processing. 1521 if (cnt != 0 && !n->is_Con()) { 1522 assert(has_node(n), "no dead nodes"); 1523 split_if_with_blocks_post(n); 1524 } 1525 if (must_throttle_split_if()) { 1526 nstack.clear(); 1527 } 1528 if (nstack.is_empty()) { 1529 // Finished all nodes on stack. 1530 break; 1531 } 1532 // Get saved parent node and next use's index. Visit the rest of uses. 1533 n = nstack.node(); 1534 cnt = n->outcnt(); 1535 i = nstack.index(); 1536 nstack.pop(); 1537 } 1538 } 1539 } 1540 1541 1542 //============================================================================= 1543 // 1544 // C L O N E A L O O P B O D Y 1545 // 1546 1547 //------------------------------clone_iff-------------------------------------- 1548 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1549 // "Nearly" because all Nodes have been cloned from the original in the loop, 1550 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1551 // through the Phi recursively, and return a Bool. 1552 Node* PhaseIdealLoop::clone_iff(PhiNode *phi, IdealLoopTree *loop) { 1553 1554 // Convert this Phi into a Phi merging Bools 1555 uint i; 1556 for (i = 1; i < phi->req(); i++) { 1557 Node *b = phi->in(i); 1558 if (b->is_Phi()) { 1559 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi(), loop)); 1560 } else { 1561 assert(b->is_Bool() || b->Opcode() == Op_Opaque4, ""); 1562 } 1563 } 1564 1565 Node* n = phi->in(1); 1566 Node* sample_opaque = NULL; 1567 Node *sample_bool = NULL; 1568 if (n->Opcode() == Op_Opaque4) { 1569 sample_opaque = n; 1570 sample_bool = n->in(1); 1571 assert(sample_bool->is_Bool(), "wrong type"); 1572 } else { 1573 sample_bool = n; 1574 } 1575 Node *sample_cmp = sample_bool->in(1); 1576 1577 // Make Phis to merge the Cmp's inputs. 1578 PhiNode *phi1 = new PhiNode(phi->in(0), Type::TOP); 1579 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP); 1580 for (i = 1; i < phi->req(); i++) { 1581 Node *n1 = sample_opaque == NULL ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1); 1582 Node *n2 = sample_opaque == NULL ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2); 1583 phi1->set_req(i, n1); 1584 phi2->set_req(i, n2); 1585 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 1586 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 1587 } 1588 // See if these Phis have been made before. 1589 // Register with optimizer 1590 Node *hit1 = _igvn.hash_find_insert(phi1); 1591 if (hit1) { // Hit, toss just made Phi 1592 _igvn.remove_dead_node(phi1); // Remove new phi 1593 assert(hit1->is_Phi(), "" ); 1594 phi1 = (PhiNode*)hit1; // Use existing phi 1595 } else { // Miss 1596 _igvn.register_new_node_with_optimizer(phi1); 1597 } 1598 Node *hit2 = _igvn.hash_find_insert(phi2); 1599 if (hit2) { // Hit, toss just made Phi 1600 _igvn.remove_dead_node(phi2); // Remove new phi 1601 assert(hit2->is_Phi(), "" ); 1602 phi2 = (PhiNode*)hit2; // Use existing phi 1603 } else { // Miss 1604 _igvn.register_new_node_with_optimizer(phi2); 1605 } 1606 // Register Phis with loop/block info 1607 set_ctrl(phi1, phi->in(0)); 1608 set_ctrl(phi2, phi->in(0)); 1609 // Make a new Cmp 1610 Node *cmp = sample_cmp->clone(); 1611 cmp->set_req(1, phi1); 1612 cmp->set_req(2, phi2); 1613 _igvn.register_new_node_with_optimizer(cmp); 1614 set_ctrl(cmp, phi->in(0)); 1615 1616 // Make a new Bool 1617 Node *b = sample_bool->clone(); 1618 b->set_req(1,cmp); 1619 _igvn.register_new_node_with_optimizer(b); 1620 set_ctrl(b, phi->in(0)); 1621 1622 if (sample_opaque != NULL) { 1623 Node* opaque = sample_opaque->clone(); 1624 opaque->set_req(1, b); 1625 _igvn.register_new_node_with_optimizer(opaque); 1626 set_ctrl(opaque, phi->in(0)); 1627 return opaque; 1628 } 1629 1630 assert(b->is_Bool(), ""); 1631 return b; 1632 } 1633 1634 //------------------------------clone_bool------------------------------------- 1635 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1636 // "Nearly" because all Nodes have been cloned from the original in the loop, 1637 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1638 // through the Phi recursively, and return a Bool. 1639 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) { 1640 uint i; 1641 // Convert this Phi into a Phi merging Bools 1642 for( i = 1; i < phi->req(); i++ ) { 1643 Node *b = phi->in(i); 1644 if( b->is_Phi() ) { 1645 _igvn.replace_input_of(phi, i, clone_bool( b->as_Phi(), loop )); 1646 } else { 1647 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" ); 1648 } 1649 } 1650 1651 Node *sample_cmp = phi->in(1); 1652 1653 // Make Phis to merge the Cmp's inputs. 1654 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP ); 1655 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP ); 1656 for( uint j = 1; j < phi->req(); j++ ) { 1657 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP 1658 Node *n1, *n2; 1659 if( cmp_top->is_Cmp() ) { 1660 n1 = cmp_top->in(1); 1661 n2 = cmp_top->in(2); 1662 } else { 1663 n1 = n2 = cmp_top; 1664 } 1665 phi1->set_req( j, n1 ); 1666 phi2->set_req( j, n2 ); 1667 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 1668 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 1669 } 1670 1671 // See if these Phis have been made before. 1672 // Register with optimizer 1673 Node *hit1 = _igvn.hash_find_insert(phi1); 1674 if( hit1 ) { // Hit, toss just made Phi 1675 _igvn.remove_dead_node(phi1); // Remove new phi 1676 assert( hit1->is_Phi(), "" ); 1677 phi1 = (PhiNode*)hit1; // Use existing phi 1678 } else { // Miss 1679 _igvn.register_new_node_with_optimizer(phi1); 1680 } 1681 Node *hit2 = _igvn.hash_find_insert(phi2); 1682 if( hit2 ) { // Hit, toss just made Phi 1683 _igvn.remove_dead_node(phi2); // Remove new phi 1684 assert( hit2->is_Phi(), "" ); 1685 phi2 = (PhiNode*)hit2; // Use existing phi 1686 } else { // Miss 1687 _igvn.register_new_node_with_optimizer(phi2); 1688 } 1689 // Register Phis with loop/block info 1690 set_ctrl(phi1, phi->in(0)); 1691 set_ctrl(phi2, phi->in(0)); 1692 // Make a new Cmp 1693 Node *cmp = sample_cmp->clone(); 1694 cmp->set_req( 1, phi1 ); 1695 cmp->set_req( 2, phi2 ); 1696 _igvn.register_new_node_with_optimizer(cmp); 1697 set_ctrl(cmp, phi->in(0)); 1698 1699 assert( cmp->is_Cmp(), "" ); 1700 return (CmpNode*)cmp; 1701 } 1702 1703 //------------------------------sink_use--------------------------------------- 1704 // If 'use' was in the loop-exit block, it now needs to be sunk 1705 // below the post-loop merge point. 1706 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) { 1707 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) { 1708 set_ctrl(use, post_loop); 1709 for (DUIterator j = use->outs(); use->has_out(j); j++) 1710 sink_use(use->out(j), post_loop); 1711 } 1712 } 1713 1714 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new, 1715 IdealLoopTree* loop, IdealLoopTree* outer_loop, 1716 Node_List*& split_if_set, Node_List*& split_bool_set, 1717 Node_List*& split_cex_set, Node_List& worklist, 1718 uint new_counter, CloneLoopMode mode) { 1719 Node* nnn = old_new[old->_idx]; 1720 // Copy uses to a worklist, so I can munge the def-use info 1721 // with impunity. 1722 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1723 worklist.push(old->fast_out(j)); 1724 1725 while( worklist.size() ) { 1726 Node *use = worklist.pop(); 1727 if (!has_node(use)) continue; // Ignore dead nodes 1728 if (use->in(0) == C->top()) continue; 1729 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1730 // Check for data-use outside of loop - at least one of OLD or USE 1731 // must not be a CFG node. 1732 #ifdef ASSERT 1733 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == NULL) { 1734 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint(); 1735 assert(mode == ControlAroundStripMined && (use == sfpt || !use->is_reachable_from_root()), "missed a node"); 1736 } 1737 #endif 1738 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) { 1739 1740 // If the Data use is an IF, that means we have an IF outside of the 1741 // loop that is switching on a condition that is set inside of the 1742 // loop. Happens if people set a loop-exit flag; then test the flag 1743 // in the loop to break the loop, then test is again outside of the 1744 // loop to determine which way the loop exited. 1745 // Loop predicate If node connects to Bool node through Opaque1 node. 1746 if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use) || use->Opcode() == Op_Opaque4) { 1747 // Since this code is highly unlikely, we lazily build the worklist 1748 // of such Nodes to go split. 1749 if (!split_if_set) { 1750 ResourceArea *area = Thread::current()->resource_area(); 1751 split_if_set = new Node_List(area); 1752 } 1753 split_if_set->push(use); 1754 } 1755 if (use->is_Bool()) { 1756 if (!split_bool_set) { 1757 ResourceArea *area = Thread::current()->resource_area(); 1758 split_bool_set = new Node_List(area); 1759 } 1760 split_bool_set->push(use); 1761 } 1762 if (use->Opcode() == Op_CreateEx) { 1763 if (!split_cex_set) { 1764 ResourceArea *area = Thread::current()->resource_area(); 1765 split_cex_set = new Node_List(area); 1766 } 1767 split_cex_set->push(use); 1768 } 1769 1770 1771 // Get "block" use is in 1772 uint idx = 0; 1773 while( use->in(idx) != old ) idx++; 1774 Node *prev = use->is_CFG() ? use : get_ctrl(use); 1775 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" ); 1776 Node *cfg = prev->_idx >= new_counter 1777 ? prev->in(2) 1778 : idom(prev); 1779 if( use->is_Phi() ) // Phi use is in prior block 1780 cfg = prev->in(idx); // NOT in block of Phi itself 1781 if (cfg->is_top()) { // Use is dead? 1782 _igvn.replace_input_of(use, idx, C->top()); 1783 continue; 1784 } 1785 1786 // If use is referenced through control edge... (idx == 0) 1787 if (mode == IgnoreStripMined && idx == 0) { 1788 LoopNode *head = loop->_head->as_Loop(); 1789 if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) { 1790 // That node is outside the inner loop, leave it outside the 1791 // outer loop as well to not confuse verification code. 1792 assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop"); 1793 _igvn.replace_input_of(use, 0, head->outer_loop_exit()); 1794 continue; 1795 } 1796 } 1797 1798 while(!outer_loop->is_member(get_loop(cfg))) { 1799 prev = cfg; 1800 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg); 1801 } 1802 // If the use occurs after merging several exits from the loop, then 1803 // old value must have dominated all those exits. Since the same old 1804 // value was used on all those exits we did not need a Phi at this 1805 // merge point. NOW we do need a Phi here. Each loop exit value 1806 // is now merged with the peeled body exit; each exit gets its own 1807 // private Phi and those Phis need to be merged here. 1808 Node *phi; 1809 if( prev->is_Region() ) { 1810 if( idx == 0 ) { // Updating control edge? 1811 phi = prev; // Just use existing control 1812 } else { // Else need a new Phi 1813 phi = PhiNode::make( prev, old ); 1814 // Now recursively fix up the new uses of old! 1815 for( uint i = 1; i < prev->req(); i++ ) { 1816 worklist.push(phi); // Onto worklist once for each 'old' input 1817 } 1818 } 1819 } else { 1820 // Get new RegionNode merging old and new loop exits 1821 prev = old_new[prev->_idx]; 1822 assert( prev, "just made this in step 7" ); 1823 if( idx == 0) { // Updating control edge? 1824 phi = prev; // Just use existing control 1825 } else { // Else need a new Phi 1826 // Make a new Phi merging data values properly 1827 phi = PhiNode::make( prev, old ); 1828 phi->set_req( 1, nnn ); 1829 } 1830 } 1831 // If inserting a new Phi, check for prior hits 1832 if( idx != 0 ) { 1833 Node *hit = _igvn.hash_find_insert(phi); 1834 if( hit == NULL ) { 1835 _igvn.register_new_node_with_optimizer(phi); // Register new phi 1836 } else { // or 1837 // Remove the new phi from the graph and use the hit 1838 _igvn.remove_dead_node(phi); 1839 phi = hit; // Use existing phi 1840 } 1841 set_ctrl(phi, prev); 1842 } 1843 // Make 'use' use the Phi instead of the old loop body exit value 1844 _igvn.replace_input_of(use, idx, phi); 1845 if( use->_idx >= new_counter ) { // If updating new phis 1846 // Not needed for correctness, but prevents a weak assert 1847 // in AddPNode from tripping (when we end up with different 1848 // base & derived Phis that will become the same after 1849 // IGVN does CSE). 1850 Node *hit = _igvn.hash_find_insert(use); 1851 if( hit ) // Go ahead and re-hash for hits. 1852 _igvn.replace_node( use, hit ); 1853 } 1854 1855 // If 'use' was in the loop-exit block, it now needs to be sunk 1856 // below the post-loop merge point. 1857 sink_use( use, prev ); 1858 } 1859 } 1860 } 1861 1862 static void clone_outer_loop_helper(Node* n, const IdealLoopTree *loop, const IdealLoopTree* outer_loop, 1863 const Node_List &old_new, Unique_Node_List& wq, PhaseIdealLoop* phase, 1864 bool check_old_new) { 1865 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1866 Node* u = n->fast_out(j); 1867 assert(check_old_new || old_new[u->_idx] == NULL, "shouldn't have been cloned"); 1868 if (!u->is_CFG() && (!check_old_new || old_new[u->_idx] == NULL)) { 1869 Node* c = phase->get_ctrl(u); 1870 IdealLoopTree* u_loop = phase->get_loop(c); 1871 assert(!loop->is_member(u_loop), "can be in outer loop or out of both loops only"); 1872 if (outer_loop->is_member(u_loop)) { 1873 wq.push(u); 1874 } 1875 } 1876 } 1877 } 1878 1879 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop, 1880 IdealLoopTree* outer_loop, int dd, Node_List &old_new, 1881 Node_List& extra_data_nodes) { 1882 if (head->is_strip_mined() && mode != IgnoreStripMined) { 1883 CountedLoopNode* cl = head->as_CountedLoop(); 1884 Node* l = cl->outer_loop(); 1885 Node* tail = cl->outer_loop_tail(); 1886 IfNode* le = cl->outer_loop_end(); 1887 Node* sfpt = cl->outer_safepoint(); 1888 CountedLoopEndNode* cle = cl->loopexit(); 1889 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop(); 1890 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null(); 1891 Node* cle_out = cle->proj_out(false); 1892 1893 Node* new_sfpt = NULL; 1894 Node* new_cle_out = cle_out->clone(); 1895 old_new.map(cle_out->_idx, new_cle_out); 1896 if (mode == CloneIncludesStripMined) { 1897 // clone outer loop body 1898 Node* new_l = l->clone(); 1899 Node* new_tail = tail->clone(); 1900 IfNode* new_le = le->clone()->as_If(); 1901 new_sfpt = sfpt->clone(); 1902 1903 set_loop(new_l, outer_loop->_parent); 1904 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd); 1905 set_loop(new_cle_out, outer_loop->_parent); 1906 set_idom(new_cle_out, new_cle, dd); 1907 set_loop(new_sfpt, outer_loop->_parent); 1908 set_idom(new_sfpt, new_cle_out, dd); 1909 set_loop(new_le, outer_loop->_parent); 1910 set_idom(new_le, new_sfpt, dd); 1911 set_loop(new_tail, outer_loop->_parent); 1912 set_idom(new_tail, new_le, dd); 1913 set_idom(new_cl, new_l, dd); 1914 1915 old_new.map(l->_idx, new_l); 1916 old_new.map(tail->_idx, new_tail); 1917 old_new.map(le->_idx, new_le); 1918 old_new.map(sfpt->_idx, new_sfpt); 1919 1920 new_l->set_req(LoopNode::LoopBackControl, new_tail); 1921 new_l->set_req(0, new_l); 1922 new_tail->set_req(0, new_le); 1923 new_le->set_req(0, new_sfpt); 1924 new_sfpt->set_req(0, new_cle_out); 1925 new_cle_out->set_req(0, new_cle); 1926 new_cl->set_req(LoopNode::EntryControl, new_l); 1927 1928 _igvn.register_new_node_with_optimizer(new_l); 1929 _igvn.register_new_node_with_optimizer(new_tail); 1930 _igvn.register_new_node_with_optimizer(new_le); 1931 } else { 1932 Node *newhead = old_new[loop->_head->_idx]; 1933 newhead->as_Loop()->clear_strip_mined(); 1934 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl)); 1935 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 1936 } 1937 // Look at data node that were assigned a control in the outer 1938 // loop: they are kept in the outer loop by the safepoint so start 1939 // from the safepoint node's inputs. 1940 IdealLoopTree* outer_loop = get_loop(l); 1941 Node_Stack stack(2); 1942 stack.push(sfpt, 1); 1943 uint new_counter = C->unique(); 1944 while (stack.size() > 0) { 1945 Node* n = stack.node(); 1946 uint i = stack.index(); 1947 while (i < n->req() && 1948 (n->in(i) == NULL || 1949 !has_ctrl(n->in(i)) || 1950 get_loop(get_ctrl(n->in(i))) != outer_loop || 1951 (old_new[n->in(i)->_idx] != NULL && old_new[n->in(i)->_idx]->_idx >= new_counter))) { 1952 i++; 1953 } 1954 if (i < n->req()) { 1955 stack.set_index(i+1); 1956 stack.push(n->in(i), 0); 1957 } else { 1958 assert(old_new[n->_idx] == NULL || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet"); 1959 Node* m = n == sfpt ? new_sfpt : n->clone(); 1960 if (m != NULL) { 1961 for (uint i = 0; i < n->req(); i++) { 1962 if (m->in(i) != NULL && old_new[m->in(i)->_idx] != NULL) { 1963 m->set_req(i, old_new[m->in(i)->_idx]); 1964 } 1965 } 1966 } else { 1967 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?"); 1968 } 1969 if (n != sfpt) { 1970 extra_data_nodes.push(n); 1971 _igvn.register_new_node_with_optimizer(m); 1972 assert(get_ctrl(n) == cle_out, "what other control?"); 1973 set_ctrl(m, new_cle_out); 1974 old_new.map(n->_idx, m); 1975 } 1976 stack.pop(); 1977 } 1978 } 1979 if (mode == CloneIncludesStripMined) { 1980 _igvn.register_new_node_with_optimizer(new_sfpt); 1981 _igvn.register_new_node_with_optimizer(new_cle_out); 1982 } 1983 // Some other transformation may have pessimistically assign some 1984 // data nodes to the outer loop. Set their control so they are out 1985 // of the outer loop. 1986 ResourceMark rm; 1987 Unique_Node_List wq; 1988 for (uint i = 0; i < extra_data_nodes.size(); i++) { 1989 Node* old = extra_data_nodes.at(i); 1990 clone_outer_loop_helper(old, loop, outer_loop, old_new, wq, this, true); 1991 } 1992 Node* new_ctrl = cl->outer_loop_exit(); 1993 assert(get_loop(new_ctrl) != outer_loop, "must be out of the loop nest"); 1994 for (uint i = 0; i < wq.size(); i++) { 1995 Node* n = wq.at(i); 1996 set_ctrl(n, new_ctrl); 1997 clone_outer_loop_helper(n, loop, outer_loop, old_new, wq, this, false); 1998 } 1999 } else { 2000 Node *newhead = old_new[loop->_head->_idx]; 2001 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 2002 } 2003 } 2004 2005 //------------------------------clone_loop------------------------------------- 2006 // 2007 // C L O N E A L O O P B O D Y 2008 // 2009 // This is the basic building block of the loop optimizations. It clones an 2010 // entire loop body. It makes an old_new loop body mapping; with this mapping 2011 // you can find the new-loop equivalent to an old-loop node. All new-loop 2012 // nodes are exactly equal to their old-loop counterparts, all edges are the 2013 // same. All exits from the old-loop now have a RegionNode that merges the 2014 // equivalent new-loop path. This is true even for the normal "loop-exit" 2015 // condition. All uses of loop-invariant old-loop values now come from (one 2016 // or more) Phis that merge their new-loop equivalents. 2017 // 2018 // This operation leaves the graph in an illegal state: there are two valid 2019 // control edges coming from the loop pre-header to both loop bodies. I'll 2020 // definitely have to hack the graph after running this transform. 2021 // 2022 // From this building block I will further edit edges to perform loop peeling 2023 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc. 2024 // 2025 // Parameter side_by_size_idom: 2026 // When side_by_size_idom is NULL, the dominator tree is constructed for 2027 // the clone loop to dominate the original. Used in construction of 2028 // pre-main-post loop sequence. 2029 // When nonnull, the clone and original are side-by-side, both are 2030 // dominated by the side_by_side_idom node. Used in construction of 2031 // unswitched loops. 2032 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd, 2033 CloneLoopMode mode, Node* side_by_side_idom) { 2034 2035 LoopNode* head = loop->_head->as_Loop(); 2036 head->verify_strip_mined(1); 2037 2038 if (C->do_vector_loop() && PrintOpto) { 2039 const char* mname = C->method()->name()->as_quoted_ascii(); 2040 if (mname != NULL) { 2041 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname); 2042 } 2043 } 2044 2045 CloneMap& cm = C->clone_map(); 2046 Dict* dict = cm.dict(); 2047 if (C->do_vector_loop()) { 2048 cm.set_clone_idx(cm.max_gen()+1); 2049 #ifndef PRODUCT 2050 if (PrintOpto) { 2051 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx()); 2052 loop->dump_head(); 2053 } 2054 #endif 2055 } 2056 2057 // Step 1: Clone the loop body. Make the old->new mapping. 2058 uint i; 2059 for( i = 0; i < loop->_body.size(); i++ ) { 2060 Node *old = loop->_body.at(i); 2061 Node *nnn = old->clone(); 2062 old_new.map( old->_idx, nnn ); 2063 if (C->do_vector_loop()) { 2064 cm.verify_insert_and_clone(old, nnn, cm.clone_idx()); 2065 } 2066 _igvn.register_new_node_with_optimizer(nnn); 2067 } 2068 2069 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop; 2070 2071 // Step 2: Fix the edges in the new body. If the old input is outside the 2072 // loop use it. If the old input is INside the loop, use the corresponding 2073 // new node instead. 2074 for( i = 0; i < loop->_body.size(); i++ ) { 2075 Node *old = loop->_body.at(i); 2076 Node *nnn = old_new[old->_idx]; 2077 // Fix CFG/Loop controlling the new node 2078 if (has_ctrl(old)) { 2079 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]); 2080 } else { 2081 set_loop(nnn, outer_loop->_parent); 2082 if (old->outcnt() > 0) { 2083 set_idom( nnn, old_new[idom(old)->_idx], dd ); 2084 } 2085 } 2086 // Correct edges to the new node 2087 for( uint j = 0; j < nnn->req(); j++ ) { 2088 Node *n = nnn->in(j); 2089 if( n ) { 2090 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n ); 2091 if( loop->is_member( old_in_loop ) ) 2092 nnn->set_req(j, old_new[n->_idx]); 2093 } 2094 } 2095 _igvn.hash_find_insert(nnn); 2096 } 2097 2098 ResourceArea *area = Thread::current()->resource_area(); 2099 Node_List extra_data_nodes(area); // data nodes in the outer strip mined loop 2100 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes); 2101 2102 // Step 3: Now fix control uses. Loop varying control uses have already 2103 // been fixed up (as part of all input edges in Step 2). Loop invariant 2104 // control uses must be either an IfFalse or an IfTrue. Make a merge 2105 // point to merge the old and new IfFalse/IfTrue nodes; make the use 2106 // refer to this. 2107 Node_List worklist(area); 2108 uint new_counter = C->unique(); 2109 for( i = 0; i < loop->_body.size(); i++ ) { 2110 Node* old = loop->_body.at(i); 2111 if( !old->is_CFG() ) continue; 2112 2113 // Copy uses to a worklist, so I can munge the def-use info 2114 // with impunity. 2115 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 2116 worklist.push(old->fast_out(j)); 2117 2118 while( worklist.size() ) { // Visit all uses 2119 Node *use = worklist.pop(); 2120 if (!has_node(use)) continue; // Ignore dead nodes 2121 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 2122 if( !loop->is_member( use_loop ) && use->is_CFG() ) { 2123 // Both OLD and USE are CFG nodes here. 2124 assert( use->is_Proj(), "" ); 2125 Node* nnn = old_new[old->_idx]; 2126 2127 Node* newuse = NULL; 2128 if (head->is_strip_mined() && mode != IgnoreStripMined) { 2129 CountedLoopNode* cl = head->as_CountedLoop(); 2130 CountedLoopEndNode* cle = cl->loopexit(); 2131 Node* cle_out = cle->proj_out_or_null(false); 2132 if (use == cle_out) { 2133 IfNode* le = cl->outer_loop_end(); 2134 use = le->proj_out(false); 2135 use_loop = get_loop(use); 2136 if (mode == CloneIncludesStripMined) { 2137 nnn = old_new[le->_idx]; 2138 } else { 2139 newuse = old_new[cle_out->_idx]; 2140 } 2141 } 2142 } 2143 if (newuse == NULL) { 2144 newuse = use->clone(); 2145 } 2146 2147 // Clone the loop exit control projection 2148 if (C->do_vector_loop()) { 2149 cm.verify_insert_and_clone(use, newuse, cm.clone_idx()); 2150 } 2151 newuse->set_req(0,nnn); 2152 _igvn.register_new_node_with_optimizer(newuse); 2153 set_loop(newuse, use_loop); 2154 set_idom(newuse, nnn, dom_depth(nnn) + 1 ); 2155 2156 // We need a Region to merge the exit from the peeled body and the 2157 // exit from the old loop body. 2158 RegionNode *r = new RegionNode(3); 2159 // Map the old use to the new merge point 2160 old_new.map( use->_idx, r ); 2161 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use)); 2162 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" ); 2163 2164 // The original user of 'use' uses 'r' instead. 2165 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) { 2166 Node* useuse = use->last_out(l); 2167 _igvn.rehash_node_delayed(useuse); 2168 uint uses_found = 0; 2169 if( useuse->in(0) == use ) { 2170 useuse->set_req(0, r); 2171 uses_found++; 2172 if( useuse->is_CFG() ) { 2173 assert( dom_depth(useuse) > dd_r, "" ); 2174 set_idom(useuse, r, dom_depth(useuse)); 2175 } 2176 } 2177 for( uint k = 1; k < useuse->req(); k++ ) { 2178 if( useuse->in(k) == use ) { 2179 useuse->set_req(k, r); 2180 uses_found++; 2181 if (useuse->is_Loop() && k == LoopNode::EntryControl) { 2182 assert(dom_depth(useuse) > dd_r , ""); 2183 set_idom(useuse, r, dom_depth(useuse)); 2184 } 2185 } 2186 } 2187 l -= uses_found; // we deleted 1 or more copies of this edge 2188 } 2189 2190 // Now finish up 'r' 2191 r->set_req( 1, newuse ); 2192 r->set_req( 2, use ); 2193 _igvn.register_new_node_with_optimizer(r); 2194 set_loop(r, use_loop); 2195 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r); 2196 } // End of if a loop-exit test 2197 } 2198 } 2199 2200 // Step 4: If loop-invariant use is not control, it must be dominated by a 2201 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region 2202 // there if needed. Make a Phi there merging old and new used values. 2203 Node_List *split_if_set = NULL; 2204 Node_List *split_bool_set = NULL; 2205 Node_List *split_cex_set = NULL; 2206 for( i = 0; i < loop->_body.size(); i++ ) { 2207 Node* old = loop->_body.at(i); 2208 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set, 2209 split_bool_set, split_cex_set, worklist, new_counter, 2210 mode); 2211 } 2212 2213 for (i = 0; i < extra_data_nodes.size(); i++) { 2214 Node* old = extra_data_nodes.at(i); 2215 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set, 2216 split_bool_set, split_cex_set, worklist, new_counter, 2217 mode); 2218 } 2219 2220 // Check for IFs that need splitting/cloning. Happens if an IF outside of 2221 // the loop uses a condition set in the loop. The original IF probably 2222 // takes control from one or more OLD Regions (which in turn get from NEW 2223 // Regions). In any case, there will be a set of Phis for each merge point 2224 // from the IF up to where the original BOOL def exists the loop. 2225 if (split_if_set) { 2226 while (split_if_set->size()) { 2227 Node *iff = split_if_set->pop(); 2228 if (iff->in(1)->is_Phi()) { 2229 Node *b = clone_iff(iff->in(1)->as_Phi(), loop); 2230 _igvn.replace_input_of(iff, 1, b); 2231 } 2232 } 2233 } 2234 if (split_bool_set) { 2235 while (split_bool_set->size()) { 2236 Node *b = split_bool_set->pop(); 2237 Node *phi = b->in(1); 2238 assert(phi->is_Phi(), ""); 2239 CmpNode *cmp = clone_bool((PhiNode*)phi, loop); 2240 _igvn.replace_input_of(b, 1, cmp); 2241 } 2242 } 2243 if (split_cex_set) { 2244 while (split_cex_set->size()) { 2245 Node *b = split_cex_set->pop(); 2246 assert(b->in(0)->is_Region(), ""); 2247 assert(b->in(1)->is_Phi(), ""); 2248 assert(b->in(0)->in(0) == b->in(1)->in(0), ""); 2249 split_up(b, b->in(0), NULL); 2250 } 2251 } 2252 2253 } 2254 2255 2256 //---------------------- stride_of_possible_iv ------------------------------------- 2257 // Looks for an iff/bool/comp with one operand of the compare 2258 // being a cycle involving an add and a phi, 2259 // with an optional truncation (left-shift followed by a right-shift) 2260 // of the add. Returns zero if not an iv. 2261 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) { 2262 Node* trunc1 = NULL; 2263 Node* trunc2 = NULL; 2264 const TypeInt* ttype = NULL; 2265 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) { 2266 return 0; 2267 } 2268 BoolNode* bl = iff->in(1)->as_Bool(); 2269 Node* cmp = bl->in(1); 2270 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) { 2271 return 0; 2272 } 2273 // Must have an invariant operand 2274 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) { 2275 return 0; 2276 } 2277 Node* add2 = NULL; 2278 Node* cmp1 = cmp->in(1); 2279 if (cmp1->is_Phi()) { 2280 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) ))) 2281 Node* phi = cmp1; 2282 for (uint i = 1; i < phi->req(); i++) { 2283 Node* in = phi->in(i); 2284 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in, 2285 &trunc1, &trunc2, &ttype); 2286 if (add && add->in(1) == phi) { 2287 add2 = add->in(2); 2288 break; 2289 } 2290 } 2291 } else { 2292 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) ))) 2293 Node* addtrunc = cmp1; 2294 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc, 2295 &trunc1, &trunc2, &ttype); 2296 if (add && add->in(1)->is_Phi()) { 2297 Node* phi = add->in(1); 2298 for (uint i = 1; i < phi->req(); i++) { 2299 if (phi->in(i) == addtrunc) { 2300 add2 = add->in(2); 2301 break; 2302 } 2303 } 2304 } 2305 } 2306 if (add2 != NULL) { 2307 const TypeInt* add2t = _igvn.type(add2)->is_int(); 2308 if (add2t->is_con()) { 2309 return add2t->get_con(); 2310 } 2311 } 2312 return 0; 2313 } 2314 2315 2316 //---------------------- stay_in_loop ------------------------------------- 2317 // Return the (unique) control output node that's in the loop (if it exists.) 2318 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) { 2319 Node* unique = NULL; 2320 if (!n) return NULL; 2321 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2322 Node* use = n->fast_out(i); 2323 if (!has_ctrl(use) && loop->is_member(get_loop(use))) { 2324 if (unique != NULL) { 2325 return NULL; 2326 } 2327 unique = use; 2328 } 2329 } 2330 return unique; 2331 } 2332 2333 //------------------------------ register_node ------------------------------------- 2334 // Utility to register node "n" with PhaseIdealLoop 2335 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) { 2336 _igvn.register_new_node_with_optimizer(n); 2337 loop->_body.push(n); 2338 if (n->is_CFG()) { 2339 set_loop(n, loop); 2340 set_idom(n, pred, ddepth); 2341 } else { 2342 set_ctrl(n, pred); 2343 } 2344 } 2345 2346 //------------------------------ proj_clone ------------------------------------- 2347 // Utility to create an if-projection 2348 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) { 2349 ProjNode* c = p->clone()->as_Proj(); 2350 c->set_req(0, iff); 2351 return c; 2352 } 2353 2354 //------------------------------ short_circuit_if ------------------------------------- 2355 // Force the iff control output to be the live_proj 2356 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) { 2357 guarantee(live_proj != NULL, "null projection"); 2358 int proj_con = live_proj->_con; 2359 assert(proj_con == 0 || proj_con == 1, "false or true projection"); 2360 Node *con = _igvn.intcon(proj_con); 2361 set_ctrl(con, C->root()); 2362 if (iff) { 2363 iff->set_req(1, con); 2364 } 2365 return con; 2366 } 2367 2368 //------------------------------ insert_if_before_proj ------------------------------------- 2369 // Insert a new if before an if projection (* - new node) 2370 // 2371 // before 2372 // if(test) 2373 // / \ 2374 // v v 2375 // other-proj proj (arg) 2376 // 2377 // after 2378 // if(test) 2379 // / \ 2380 // / v 2381 // | * proj-clone 2382 // v | 2383 // other-proj v 2384 // * new_if(relop(cmp[IU](left,right))) 2385 // / \ 2386 // v v 2387 // * new-proj proj 2388 // (returned) 2389 // 2390 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) { 2391 IfNode* iff = proj->in(0)->as_If(); 2392 IdealLoopTree *loop = get_loop(proj); 2393 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2394 int ddepth = dom_depth(proj); 2395 2396 _igvn.rehash_node_delayed(iff); 2397 _igvn.rehash_node_delayed(proj); 2398 2399 proj->set_req(0, NULL); // temporary disconnect 2400 ProjNode* proj2 = proj_clone(proj, iff); 2401 register_node(proj2, loop, iff, ddepth); 2402 2403 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right); 2404 register_node(cmp, loop, proj2, ddepth); 2405 2406 BoolNode* bol = new BoolNode(cmp, relop); 2407 register_node(bol, loop, proj2, ddepth); 2408 2409 int opcode = iff->Opcode(); 2410 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode"); 2411 IfNode* new_if = (opcode == Op_If) ? new IfNode(proj2, bol, iff->_prob, iff->_fcnt): 2412 new RangeCheckNode(proj2, bol, iff->_prob, iff->_fcnt); 2413 register_node(new_if, loop, proj2, ddepth); 2414 2415 proj->set_req(0, new_if); // reattach 2416 set_idom(proj, new_if, ddepth); 2417 2418 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj(); 2419 guarantee(new_exit != NULL, "null exit node"); 2420 register_node(new_exit, get_loop(other_proj), new_if, ddepth); 2421 2422 return new_exit; 2423 } 2424 2425 //------------------------------ insert_region_before_proj ------------------------------------- 2426 // Insert a region before an if projection (* - new node) 2427 // 2428 // before 2429 // if(test) 2430 // / | 2431 // v | 2432 // proj v 2433 // other-proj 2434 // 2435 // after 2436 // if(test) 2437 // / | 2438 // v | 2439 // * proj-clone v 2440 // | other-proj 2441 // v 2442 // * new-region 2443 // | 2444 // v 2445 // * dum_if 2446 // / \ 2447 // v \ 2448 // * dum-proj v 2449 // proj 2450 // 2451 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) { 2452 IfNode* iff = proj->in(0)->as_If(); 2453 IdealLoopTree *loop = get_loop(proj); 2454 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2455 int ddepth = dom_depth(proj); 2456 2457 _igvn.rehash_node_delayed(iff); 2458 _igvn.rehash_node_delayed(proj); 2459 2460 proj->set_req(0, NULL); // temporary disconnect 2461 ProjNode* proj2 = proj_clone(proj, iff); 2462 register_node(proj2, loop, iff, ddepth); 2463 2464 RegionNode* reg = new RegionNode(2); 2465 reg->set_req(1, proj2); 2466 register_node(reg, loop, iff, ddepth); 2467 2468 IfNode* dum_if = new IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt); 2469 register_node(dum_if, loop, reg, ddepth); 2470 2471 proj->set_req(0, dum_if); // reattach 2472 set_idom(proj, dum_if, ddepth); 2473 2474 ProjNode* dum_proj = proj_clone(other_proj, dum_if); 2475 register_node(dum_proj, loop, dum_if, ddepth); 2476 2477 return reg; 2478 } 2479 2480 //------------------------------ insert_cmpi_loop_exit ------------------------------------- 2481 // Clone a signed compare loop exit from an unsigned compare and 2482 // insert it before the unsigned cmp on the stay-in-loop path. 2483 // All new nodes inserted in the dominator tree between the original 2484 // if and it's projections. The original if test is replaced with 2485 // a constant to force the stay-in-loop path. 2486 // 2487 // This is done to make sure that the original if and it's projections 2488 // still dominate the same set of control nodes, that the ctrl() relation 2489 // from data nodes to them is preserved, and that their loop nesting is 2490 // preserved. 2491 // 2492 // before 2493 // if(i <u limit) unsigned compare loop exit 2494 // / | 2495 // v v 2496 // exit-proj stay-in-loop-proj 2497 // 2498 // after 2499 // if(stay-in-loop-const) original if 2500 // / | 2501 // / v 2502 // / if(i < limit) new signed test 2503 // / / | 2504 // / / v 2505 // / / if(i <u limit) new cloned unsigned test 2506 // / / / | 2507 // v v v | 2508 // region | 2509 // | | 2510 // dum-if | 2511 // / | | 2512 // ether | | 2513 // v v 2514 // exit-proj stay-in-loop-proj 2515 // 2516 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) { 2517 const bool Signed = true; 2518 const bool Unsigned = false; 2519 2520 BoolNode* bol = if_cmpu->in(1)->as_Bool(); 2521 if (bol->_test._test != BoolTest::lt) return NULL; 2522 CmpNode* cmpu = bol->in(1)->as_Cmp(); 2523 if (cmpu->Opcode() != Op_CmpU) return NULL; 2524 int stride = stride_of_possible_iv(if_cmpu); 2525 if (stride == 0) return NULL; 2526 2527 Node* lp_proj = stay_in_loop(if_cmpu, loop); 2528 guarantee(lp_proj != NULL, "null loop node"); 2529 2530 ProjNode* lp_continue = lp_proj->as_Proj(); 2531 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj(); 2532 2533 Node* limit = NULL; 2534 if (stride > 0) { 2535 limit = cmpu->in(2); 2536 } else { 2537 limit = _igvn.makecon(TypeInt::ZERO); 2538 set_ctrl(limit, C->root()); 2539 } 2540 // Create a new region on the exit path 2541 RegionNode* reg = insert_region_before_proj(lp_exit); 2542 guarantee(reg != NULL, "null region node"); 2543 2544 // Clone the if-cmpu-true-false using a signed compare 2545 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge; 2546 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue); 2547 reg->add_req(cmpi_exit); 2548 2549 // Clone the if-cmpu-true-false 2550 BoolTest::mask rel_u = bol->_test._test; 2551 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue); 2552 reg->add_req(cmpu_exit); 2553 2554 // Force original if to stay in loop. 2555 short_circuit_if(if_cmpu, lp_continue); 2556 2557 return cmpi_exit->in(0)->as_If(); 2558 } 2559 2560 //------------------------------ remove_cmpi_loop_exit ------------------------------------- 2561 // Remove a previously inserted signed compare loop exit. 2562 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) { 2563 Node* lp_proj = stay_in_loop(if_cmp, loop); 2564 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI && 2565 stay_in_loop(lp_proj, loop)->is_If() && 2566 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu"); 2567 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO); 2568 set_ctrl(con, C->root()); 2569 if_cmp->set_req(1, con); 2570 } 2571 2572 //------------------------------ scheduled_nodelist ------------------------------------- 2573 // Create a post order schedule of nodes that are in the 2574 // "member" set. The list is returned in "sched". 2575 // The first node in "sched" is the loop head, followed by 2576 // nodes which have no inputs in the "member" set, and then 2577 // followed by the nodes that have an immediate input dependence 2578 // on a node in "sched". 2579 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) { 2580 2581 assert(member.test(loop->_head->_idx), "loop head must be in member set"); 2582 Arena *a = Thread::current()->resource_area(); 2583 VectorSet visited(a); 2584 Node_Stack nstack(a, loop->_body.size()); 2585 2586 Node* n = loop->_head; // top of stack is cached in "n" 2587 uint idx = 0; 2588 visited.set(n->_idx); 2589 2590 // Initially push all with no inputs from within member set 2591 for(uint i = 0; i < loop->_body.size(); i++ ) { 2592 Node *elt = loop->_body.at(i); 2593 if (member.test(elt->_idx)) { 2594 bool found = false; 2595 for (uint j = 0; j < elt->req(); j++) { 2596 Node* def = elt->in(j); 2597 if (def && member.test(def->_idx) && def != elt) { 2598 found = true; 2599 break; 2600 } 2601 } 2602 if (!found && elt != loop->_head) { 2603 nstack.push(n, idx); 2604 n = elt; 2605 assert(!visited.test(n->_idx), "not seen yet"); 2606 visited.set(n->_idx); 2607 } 2608 } 2609 } 2610 2611 // traverse out's that are in the member set 2612 while (true) { 2613 if (idx < n->outcnt()) { 2614 Node* use = n->raw_out(idx); 2615 idx++; 2616 if (!visited.test_set(use->_idx)) { 2617 if (member.test(use->_idx)) { 2618 nstack.push(n, idx); 2619 n = use; 2620 idx = 0; 2621 } 2622 } 2623 } else { 2624 // All outputs processed 2625 sched.push(n); 2626 if (nstack.is_empty()) break; 2627 n = nstack.node(); 2628 idx = nstack.index(); 2629 nstack.pop(); 2630 } 2631 } 2632 } 2633 2634 2635 //------------------------------ has_use_in_set ------------------------------------- 2636 // Has a use in the vector set 2637 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) { 2638 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2639 Node* use = n->fast_out(j); 2640 if (vset.test(use->_idx)) { 2641 return true; 2642 } 2643 } 2644 return false; 2645 } 2646 2647 2648 //------------------------------ has_use_internal_to_set ------------------------------------- 2649 // Has use internal to the vector set (ie. not in a phi at the loop head) 2650 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) { 2651 Node* head = loop->_head; 2652 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2653 Node* use = n->fast_out(j); 2654 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) { 2655 return true; 2656 } 2657 } 2658 return false; 2659 } 2660 2661 2662 //------------------------------ clone_for_use_outside_loop ------------------------------------- 2663 // clone "n" for uses that are outside of loop 2664 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) { 2665 int cloned = 0; 2666 assert(worklist.size() == 0, "should be empty"); 2667 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2668 Node* use = n->fast_out(j); 2669 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) { 2670 worklist.push(use); 2671 } 2672 } 2673 while( worklist.size() ) { 2674 Node *use = worklist.pop(); 2675 if (!has_node(use) || use->in(0) == C->top()) continue; 2676 uint j; 2677 for (j = 0; j < use->req(); j++) { 2678 if (use->in(j) == n) break; 2679 } 2680 assert(j < use->req(), "must be there"); 2681 2682 // clone "n" and insert it between the inputs of "n" and the use outside the loop 2683 Node* n_clone = n->clone(); 2684 _igvn.replace_input_of(use, j, n_clone); 2685 cloned++; 2686 Node* use_c; 2687 if (!use->is_Phi()) { 2688 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0); 2689 } else { 2690 // Use in a phi is considered a use in the associated predecessor block 2691 use_c = use->in(0)->in(j); 2692 } 2693 set_ctrl(n_clone, use_c); 2694 assert(!loop->is_member(get_loop(use_c)), "should be outside loop"); 2695 get_loop(use_c)->_body.push(n_clone); 2696 _igvn.register_new_node_with_optimizer(n_clone); 2697 #ifndef PRODUCT 2698 if (TracePartialPeeling) { 2699 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx); 2700 } 2701 #endif 2702 } 2703 return cloned; 2704 } 2705 2706 2707 //------------------------------ clone_for_special_use_inside_loop ------------------------------------- 2708 // clone "n" for special uses that are in the not_peeled region. 2709 // If these def-uses occur in separate blocks, the code generator 2710 // marks the method as not compilable. For example, if a "BoolNode" 2711 // is in a different basic block than the "IfNode" that uses it, then 2712 // the compilation is aborted in the code generator. 2713 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 2714 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) { 2715 if (n->is_Phi() || n->is_Load()) { 2716 return; 2717 } 2718 assert(worklist.size() == 0, "should be empty"); 2719 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2720 Node* use = n->fast_out(j); 2721 if ( not_peel.test(use->_idx) && 2722 (use->is_If() || use->is_CMove() || use->is_Bool()) && 2723 use->in(1) == n) { 2724 worklist.push(use); 2725 } 2726 } 2727 if (worklist.size() > 0) { 2728 // clone "n" and insert it between inputs of "n" and the use 2729 Node* n_clone = n->clone(); 2730 loop->_body.push(n_clone); 2731 _igvn.register_new_node_with_optimizer(n_clone); 2732 set_ctrl(n_clone, get_ctrl(n)); 2733 sink_list.push(n_clone); 2734 not_peel <<= n_clone->_idx; // add n_clone to not_peel set. 2735 #ifndef PRODUCT 2736 if (TracePartialPeeling) { 2737 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx); 2738 } 2739 #endif 2740 while( worklist.size() ) { 2741 Node *use = worklist.pop(); 2742 _igvn.rehash_node_delayed(use); 2743 for (uint j = 1; j < use->req(); j++) { 2744 if (use->in(j) == n) { 2745 use->set_req(j, n_clone); 2746 } 2747 } 2748 } 2749 } 2750 } 2751 2752 2753 //------------------------------ insert_phi_for_loop ------------------------------------- 2754 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 2755 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) { 2756 Node *phi = PhiNode::make(lp, back_edge_val); 2757 phi->set_req(LoopNode::EntryControl, lp_entry_val); 2758 // Use existing phi if it already exists 2759 Node *hit = _igvn.hash_find_insert(phi); 2760 if( hit == NULL ) { 2761 _igvn.register_new_node_with_optimizer(phi); 2762 set_ctrl(phi, lp); 2763 } else { 2764 // Remove the new phi from the graph and use the hit 2765 _igvn.remove_dead_node(phi); 2766 phi = hit; 2767 } 2768 _igvn.replace_input_of(use, idx, phi); 2769 } 2770 2771 #ifdef ASSERT 2772 //------------------------------ is_valid_loop_partition ------------------------------------- 2773 // Validate the loop partition sets: peel and not_peel 2774 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, 2775 VectorSet& not_peel ) { 2776 uint i; 2777 // Check that peel_list entries are in the peel set 2778 for (i = 0; i < peel_list.size(); i++) { 2779 if (!peel.test(peel_list.at(i)->_idx)) { 2780 return false; 2781 } 2782 } 2783 // Check at loop members are in one of peel set or not_peel set 2784 for (i = 0; i < loop->_body.size(); i++ ) { 2785 Node *def = loop->_body.at(i); 2786 uint di = def->_idx; 2787 // Check that peel set elements are in peel_list 2788 if (peel.test(di)) { 2789 if (not_peel.test(di)) { 2790 return false; 2791 } 2792 // Must be in peel_list also 2793 bool found = false; 2794 for (uint j = 0; j < peel_list.size(); j++) { 2795 if (peel_list.at(j)->_idx == di) { 2796 found = true; 2797 break; 2798 } 2799 } 2800 if (!found) { 2801 return false; 2802 } 2803 } else if (not_peel.test(di)) { 2804 if (peel.test(di)) { 2805 return false; 2806 } 2807 } else { 2808 return false; 2809 } 2810 } 2811 return true; 2812 } 2813 2814 //------------------------------ is_valid_clone_loop_exit_use ------------------------------------- 2815 // Ensure a use outside of loop is of the right form 2816 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) { 2817 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2818 return (use->is_Phi() && 2819 use_c->is_Region() && use_c->req() == 3 && 2820 (use_c->in(exit_idx)->Opcode() == Op_IfTrue || 2821 use_c->in(exit_idx)->Opcode() == Op_IfFalse || 2822 use_c->in(exit_idx)->Opcode() == Op_JumpProj) && 2823 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) ); 2824 } 2825 2826 //------------------------------ is_valid_clone_loop_form ------------------------------------- 2827 // Ensure that all uses outside of loop are of the right form 2828 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 2829 uint orig_exit_idx, uint clone_exit_idx) { 2830 uint len = peel_list.size(); 2831 for (uint i = 0; i < len; i++) { 2832 Node *def = peel_list.at(i); 2833 2834 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 2835 Node *use = def->fast_out(j); 2836 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2837 if (!loop->is_member(get_loop(use_c))) { 2838 // use is not in the loop, check for correct structure 2839 if (use->in(0) == def) { 2840 // Okay 2841 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) { 2842 return false; 2843 } 2844 } 2845 } 2846 } 2847 return true; 2848 } 2849 #endif 2850 2851 //------------------------------ partial_peel ------------------------------------- 2852 // Partially peel (aka loop rotation) the top portion of a loop (called 2853 // the peel section below) by cloning it and placing one copy just before 2854 // the new loop head and the other copy at the bottom of the new loop. 2855 // 2856 // before after where it came from 2857 // 2858 // stmt1 stmt1 2859 // loop: stmt2 clone 2860 // stmt2 if condA goto exitA clone 2861 // if condA goto exitA new_loop: new 2862 // stmt3 stmt3 clone 2863 // if !condB goto loop if condB goto exitB clone 2864 // exitB: stmt2 orig 2865 // stmt4 if !condA goto new_loop orig 2866 // exitA: goto exitA 2867 // exitB: 2868 // stmt4 2869 // exitA: 2870 // 2871 // Step 1: find the cut point: an exit test on probable 2872 // induction variable. 2873 // Step 2: schedule (with cloning) operations in the peel 2874 // section that can be executed after the cut into 2875 // the section that is not peeled. This may need 2876 // to clone operations into exit blocks. For 2877 // instance, a reference to A[i] in the not-peel 2878 // section and a reference to B[i] in an exit block 2879 // may cause a left-shift of i by 2 to be placed 2880 // in the peel block. This step will clone the left 2881 // shift into the exit block and sink the left shift 2882 // from the peel to the not-peel section. 2883 // Step 3: clone the loop, retarget the control, and insert 2884 // phis for values that are live across the new loop 2885 // head. This is very dependent on the graph structure 2886 // from clone_loop. It creates region nodes for 2887 // exit control and associated phi nodes for values 2888 // flow out of the loop through that exit. The region 2889 // node is dominated by the clone's control projection. 2890 // So the clone's peel section is placed before the 2891 // new loop head, and the clone's not-peel section is 2892 // forms the top part of the new loop. The original 2893 // peel section forms the tail of the new loop. 2894 // Step 4: update the dominator tree and recompute the 2895 // dominator depth. 2896 // 2897 // orig 2898 // 2899 // stmt1 2900 // | 2901 // v 2902 // loop predicate 2903 // | 2904 // v 2905 // loop<----+ 2906 // | | 2907 // stmt2 | 2908 // | | 2909 // v | 2910 // ifA | 2911 // / | | 2912 // v v | 2913 // false true ^ <-- last_peel 2914 // / | | 2915 // / ===|==cut | 2916 // / stmt3 | <-- first_not_peel 2917 // / | | 2918 // | v | 2919 // v ifB | 2920 // exitA: / \ | 2921 // / \ | 2922 // v v | 2923 // false true | 2924 // / \ | 2925 // / ----+ 2926 // | 2927 // v 2928 // exitB: 2929 // stmt4 2930 // 2931 // 2932 // after clone loop 2933 // 2934 // stmt1 2935 // | 2936 // v 2937 // loop predicate 2938 // / \ 2939 // clone / \ orig 2940 // / \ 2941 // / \ 2942 // v v 2943 // +---->loop loop<----+ 2944 // | | | | 2945 // | stmt2 stmt2 | 2946 // | | | | 2947 // | v v | 2948 // | ifA ifA | 2949 // | | \ / | | 2950 // | v v v v | 2951 // ^ true false false true ^ <-- last_peel 2952 // | | ^ \ / | | 2953 // | cut==|== \ \ / ===|==cut | 2954 // | stmt3 \ \ / stmt3 | <-- first_not_peel 2955 // | | dom | | | | 2956 // | v \ 1v v2 v | 2957 // | ifB regionA ifB | 2958 // | / \ | / \ | 2959 // | / \ v / \ | 2960 // | v v exitA: v v | 2961 // | true false false true | 2962 // | / ^ \ / \ | 2963 // +---- \ \ / ----+ 2964 // dom \ / 2965 // \ 1v v2 2966 // regionB 2967 // | 2968 // v 2969 // exitB: 2970 // stmt4 2971 // 2972 // 2973 // after partial peel 2974 // 2975 // stmt1 2976 // | 2977 // v 2978 // loop predicate 2979 // / 2980 // clone / orig 2981 // / TOP 2982 // / \ 2983 // v v 2984 // TOP->loop loop----+ 2985 // | | | 2986 // stmt2 stmt2 | 2987 // | | | 2988 // v v | 2989 // ifA ifA | 2990 // | \ / | | 2991 // v v v v | 2992 // true false false true | <-- last_peel 2993 // | ^ \ / +------|---+ 2994 // +->newloop \ \ / === ==cut | | 2995 // | stmt3 \ \ / TOP | | 2996 // | | dom | | stmt3 | | <-- first_not_peel 2997 // | v \ 1v v2 v | | 2998 // | ifB regionA ifB ^ v 2999 // | / \ | / \ | | 3000 // | / \ v / \ | | 3001 // | v v exitA: v v | | 3002 // | true false false true | | 3003 // | / ^ \ / \ | | 3004 // | | \ \ / v | | 3005 // | | dom \ / TOP | | 3006 // | | \ 1v v2 | | 3007 // ^ v regionB | | 3008 // | | | | | 3009 // | | v ^ v 3010 // | | exitB: | | 3011 // | | stmt4 | | 3012 // | +------------>-----------------+ | 3013 // | | 3014 // +-----------------<---------------------+ 3015 // 3016 // 3017 // final graph 3018 // 3019 // stmt1 3020 // | 3021 // v 3022 // loop predicate 3023 // | 3024 // v 3025 // stmt2 clone 3026 // | 3027 // v 3028 // ........> ifA clone 3029 // : / | 3030 // dom / | 3031 // : v v 3032 // : false true 3033 // : | | 3034 // : | v 3035 // : | newloop<-----+ 3036 // : | | | 3037 // : | stmt3 clone | 3038 // : | | | 3039 // : | v | 3040 // : | ifB | 3041 // : | / \ | 3042 // : | v v | 3043 // : | false true | 3044 // : | | | | 3045 // : | v stmt2 | 3046 // : | exitB: | | 3047 // : | stmt4 v | 3048 // : | ifA orig | 3049 // : | / \ | 3050 // : | / \ | 3051 // : | v v | 3052 // : | false true | 3053 // : | / \ | 3054 // : v v -----+ 3055 // RegionA 3056 // | 3057 // v 3058 // exitA 3059 // 3060 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) { 3061 3062 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only"); 3063 if (!loop->_head->is_Loop()) { 3064 return false; 3065 } 3066 LoopNode *head = loop->_head->as_Loop(); 3067 3068 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) { 3069 return false; 3070 } 3071 3072 // Check for complex exit control 3073 for (uint ii = 0; ii < loop->_body.size(); ii++) { 3074 Node *n = loop->_body.at(ii); 3075 int opc = n->Opcode(); 3076 if (n->is_Call() || 3077 opc == Op_Catch || 3078 opc == Op_CatchProj || 3079 opc == Op_Jump || 3080 opc == Op_JumpProj) { 3081 #ifndef PRODUCT 3082 if (TracePartialPeeling) { 3083 tty->print_cr("\nExit control too complex: lp: %d", head->_idx); 3084 } 3085 #endif 3086 return false; 3087 } 3088 } 3089 3090 int dd = dom_depth(head); 3091 3092 // Step 1: find cut point 3093 3094 // Walk up dominators to loop head looking for first loop exit 3095 // which is executed on every path thru loop. 3096 IfNode *peel_if = NULL; 3097 IfNode *peel_if_cmpu = NULL; 3098 3099 Node *iff = loop->tail(); 3100 while (iff != head) { 3101 if (iff->is_If()) { 3102 Node *ctrl = get_ctrl(iff->in(1)); 3103 if (ctrl->is_top()) return false; // Dead test on live IF. 3104 // If loop-varying exit-test, check for induction variable 3105 if (loop->is_member(get_loop(ctrl)) && 3106 loop->is_loop_exit(iff) && 3107 is_possible_iv_test(iff)) { 3108 Node* cmp = iff->in(1)->in(1); 3109 if (cmp->Opcode() == Op_CmpI) { 3110 peel_if = iff->as_If(); 3111 } else { 3112 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU"); 3113 peel_if_cmpu = iff->as_If(); 3114 } 3115 } 3116 } 3117 iff = idom(iff); 3118 } 3119 3120 // Prefer signed compare over unsigned compare. 3121 IfNode* new_peel_if = NULL; 3122 if (peel_if == NULL) { 3123 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) { 3124 return false; // No peel point found 3125 } 3126 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop); 3127 if (new_peel_if == NULL) { 3128 return false; // No peel point found 3129 } 3130 peel_if = new_peel_if; 3131 } 3132 Node* last_peel = stay_in_loop(peel_if, loop); 3133 Node* first_not_peeled = stay_in_loop(last_peel, loop); 3134 if (first_not_peeled == NULL || first_not_peeled == head) { 3135 return false; 3136 } 3137 3138 #ifndef PRODUCT 3139 if (TraceLoopOpts) { 3140 tty->print("PartialPeel "); 3141 loop->dump_head(); 3142 } 3143 3144 if (TracePartialPeeling) { 3145 tty->print_cr("before partial peel one iteration"); 3146 Node_List wl; 3147 Node* t = head->in(2); 3148 while (true) { 3149 wl.push(t); 3150 if (t == head) break; 3151 t = idom(t); 3152 } 3153 while (wl.size() > 0) { 3154 Node* tt = wl.pop(); 3155 tt->dump(); 3156 if (tt == last_peel) tty->print_cr("-- cut --"); 3157 } 3158 } 3159 #endif 3160 ResourceArea *area = Thread::current()->resource_area(); 3161 VectorSet peel(area); 3162 VectorSet not_peel(area); 3163 Node_List peel_list(area); 3164 Node_List worklist(area); 3165 Node_List sink_list(area); 3166 3167 if (!may_require_nodes(loop->est_loop_clone_sz(2))) { 3168 return false; 3169 } 3170 3171 // Set of cfg nodes to peel are those that are executable from 3172 // the head through last_peel. 3173 assert(worklist.size() == 0, "should be empty"); 3174 worklist.push(head); 3175 peel.set(head->_idx); 3176 while (worklist.size() > 0) { 3177 Node *n = worklist.pop(); 3178 if (n != last_peel) { 3179 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3180 Node* use = n->fast_out(j); 3181 if (use->is_CFG() && 3182 loop->is_member(get_loop(use)) && 3183 !peel.test_set(use->_idx)) { 3184 worklist.push(use); 3185 } 3186 } 3187 } 3188 } 3189 3190 // Set of non-cfg nodes to peel are those that are control 3191 // dependent on the cfg nodes. 3192 uint i; 3193 for(i = 0; i < loop->_body.size(); i++ ) { 3194 Node *n = loop->_body.at(i); 3195 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n; 3196 if (peel.test(n_c->_idx)) { 3197 peel.set(n->_idx); 3198 } else { 3199 not_peel.set(n->_idx); 3200 } 3201 } 3202 3203 // Step 2: move operations from the peeled section down into the 3204 // not-peeled section 3205 3206 // Get a post order schedule of nodes in the peel region 3207 // Result in right-most operand. 3208 scheduled_nodelist(loop, peel, peel_list ); 3209 3210 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 3211 3212 // For future check for too many new phis 3213 uint old_phi_cnt = 0; 3214 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 3215 Node* use = head->fast_out(j); 3216 if (use->is_Phi()) old_phi_cnt++; 3217 } 3218 3219 #ifndef PRODUCT 3220 if (TracePartialPeeling) { 3221 tty->print_cr("\npeeled list"); 3222 } 3223 #endif 3224 3225 // Evacuate nodes in peel region into the not_peeled region if possible 3226 uint new_phi_cnt = 0; 3227 uint cloned_for_outside_use = 0; 3228 for (i = 0; i < peel_list.size();) { 3229 Node* n = peel_list.at(i); 3230 #ifndef PRODUCT 3231 if (TracePartialPeeling) n->dump(); 3232 #endif 3233 bool incr = true; 3234 if ( !n->is_CFG() ) { 3235 3236 if ( has_use_in_set(n, not_peel) ) { 3237 3238 // If not used internal to the peeled region, 3239 // move "n" from peeled to not_peeled region. 3240 3241 if ( !has_use_internal_to_set(n, peel, loop) ) { 3242 3243 // if not pinned and not a load (which maybe anti-dependent on a store) 3244 // and not a CMove (Matcher expects only bool->cmove). 3245 if (n->in(0) == NULL && !n->is_Load() && !n->is_CMove()) { 3246 cloned_for_outside_use += clone_for_use_outside_loop( loop, n, worklist ); 3247 sink_list.push(n); 3248 peel >>= n->_idx; // delete n from peel set. 3249 not_peel <<= n->_idx; // add n to not_peel set. 3250 peel_list.remove(i); 3251 incr = false; 3252 #ifndef PRODUCT 3253 if (TracePartialPeeling) { 3254 tty->print_cr("sink to not_peeled region: %d newbb: %d", 3255 n->_idx, get_ctrl(n)->_idx); 3256 } 3257 #endif 3258 } 3259 } else { 3260 // Otherwise check for special def-use cases that span 3261 // the peel/not_peel boundary such as bool->if 3262 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist ); 3263 new_phi_cnt++; 3264 } 3265 } 3266 } 3267 if (incr) i++; 3268 } 3269 3270 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) { 3271 #ifndef PRODUCT 3272 if (TracePartialPeeling) { 3273 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c", 3274 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F'); 3275 } 3276 #endif 3277 if (new_peel_if != NULL) { 3278 remove_cmpi_loop_exit(new_peel_if, loop); 3279 } 3280 // Inhibit more partial peeling on this loop 3281 assert(!head->is_partial_peel_loop(), "not partial peeled"); 3282 head->mark_partial_peel_failed(); 3283 if (cloned_for_outside_use > 0) { 3284 // Terminate this round of loop opts because 3285 // the graph outside this loop was changed. 3286 C->set_major_progress(); 3287 return true; 3288 } 3289 return false; 3290 } 3291 3292 // Step 3: clone loop, retarget control, and insert new phis 3293 3294 // Create new loop head for new phis and to hang 3295 // the nodes being moved (sinked) from the peel region. 3296 LoopNode* new_head = new LoopNode(last_peel, last_peel); 3297 new_head->set_unswitch_count(head->unswitch_count()); // Preserve 3298 _igvn.register_new_node_with_optimizer(new_head); 3299 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled"); 3300 _igvn.replace_input_of(first_not_peeled, 0, new_head); 3301 set_loop(new_head, loop); 3302 loop->_body.push(new_head); 3303 not_peel.set(new_head->_idx); 3304 set_idom(new_head, last_peel, dom_depth(first_not_peeled)); 3305 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled)); 3306 3307 while (sink_list.size() > 0) { 3308 Node* n = sink_list.pop(); 3309 set_ctrl(n, new_head); 3310 } 3311 3312 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 3313 3314 clone_loop(loop, old_new, dd, IgnoreStripMined); 3315 3316 const uint clone_exit_idx = 1; 3317 const uint orig_exit_idx = 2; 3318 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop"); 3319 3320 Node* head_clone = old_new[head->_idx]; 3321 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop(); 3322 Node* orig_tail_clone = head_clone->in(2); 3323 3324 // Add phi if "def" node is in peel set and "use" is not 3325 3326 for(i = 0; i < peel_list.size(); i++ ) { 3327 Node *def = peel_list.at(i); 3328 if (!def->is_CFG()) { 3329 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 3330 Node *use = def->fast_out(j); 3331 if (has_node(use) && use->in(0) != C->top() && 3332 (!peel.test(use->_idx) || 3333 (use->is_Phi() && use->in(0) == head)) ) { 3334 worklist.push(use); 3335 } 3336 } 3337 while( worklist.size() ) { 3338 Node *use = worklist.pop(); 3339 for (uint j = 1; j < use->req(); j++) { 3340 Node* n = use->in(j); 3341 if (n == def) { 3342 3343 // "def" is in peel set, "use" is not in peel set 3344 // or "use" is in the entry boundary (a phi) of the peel set 3345 3346 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use; 3347 3348 if ( loop->is_member(get_loop( use_c )) ) { 3349 // use is in loop 3350 if (old_new[use->_idx] != NULL) { // null for dead code 3351 Node* use_clone = old_new[use->_idx]; 3352 _igvn.replace_input_of(use, j, C->top()); 3353 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone ); 3354 } 3355 } else { 3356 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format"); 3357 // use is not in the loop, check if the live range includes the cut 3358 Node* lp_if = use_c->in(orig_exit_idx)->in(0); 3359 if (not_peel.test(lp_if->_idx)) { 3360 assert(j == orig_exit_idx, "use from original loop"); 3361 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone ); 3362 } 3363 } 3364 } 3365 } 3366 } 3367 } 3368 } 3369 3370 // Step 3b: retarget control 3371 3372 // Redirect control to the new loop head if a cloned node in 3373 // the not_peeled region has control that points into the peeled region. 3374 // This necessary because the cloned peeled region will be outside 3375 // the loop. 3376 // from to 3377 // cloned-peeled <---+ 3378 // new_head_clone: | <--+ 3379 // cloned-not_peeled in(0) in(0) 3380 // orig-peeled 3381 3382 for(i = 0; i < loop->_body.size(); i++ ) { 3383 Node *n = loop->_body.at(i); 3384 if (!n->is_CFG() && n->in(0) != NULL && 3385 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) { 3386 Node* n_clone = old_new[n->_idx]; 3387 _igvn.replace_input_of(n_clone, 0, new_head_clone); 3388 } 3389 } 3390 3391 // Backedge of the surviving new_head (the clone) is original last_peel 3392 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel); 3393 3394 // Cut first node in original not_peel set 3395 _igvn.rehash_node_delayed(new_head); // Multiple edge updates: 3396 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of 3397 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls 3398 3399 // Copy head_clone back-branch info to original head 3400 // and remove original head's loop entry and 3401 // clone head's back-branch 3402 _igvn.rehash_node_delayed(head); // Multiple edge updates 3403 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl)); 3404 head->set_req(LoopNode::LoopBackControl, C->top()); 3405 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top()); 3406 3407 // Similarly modify the phis 3408 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) { 3409 Node* use = head->fast_out(k); 3410 if (use->is_Phi() && use->outcnt() > 0) { 3411 Node* use_clone = old_new[use->_idx]; 3412 _igvn.rehash_node_delayed(use); // Multiple edge updates 3413 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl)); 3414 use->set_req(LoopNode::LoopBackControl, C->top()); 3415 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top()); 3416 } 3417 } 3418 3419 // Step 4: update dominator tree and dominator depth 3420 3421 set_idom(head, orig_tail_clone, dd); 3422 recompute_dom_depth(); 3423 3424 // Inhibit more partial peeling on this loop 3425 new_head_clone->set_partial_peel_loop(); 3426 C->set_major_progress(); 3427 loop->record_for_igvn(); 3428 3429 #ifndef PRODUCT 3430 if (TracePartialPeeling) { 3431 tty->print_cr("\nafter partial peel one iteration"); 3432 Node_List wl(area); 3433 Node* t = last_peel; 3434 while (true) { 3435 wl.push(t); 3436 if (t == head_clone) break; 3437 t = idom(t); 3438 } 3439 while (wl.size() > 0) { 3440 Node* tt = wl.pop(); 3441 if (tt == head) tty->print_cr("orig head"); 3442 else if (tt == new_head_clone) tty->print_cr("new head"); 3443 else if (tt == head_clone) tty->print_cr("clone head"); 3444 tt->dump(); 3445 } 3446 } 3447 #endif 3448 return true; 3449 } 3450 3451 //------------------------------reorg_offsets---------------------------------- 3452 // Reorganize offset computations to lower register pressure. Mostly 3453 // prevent loop-fallout uses of the pre-incremented trip counter (which are 3454 // then alive with the post-incremented trip counter forcing an extra 3455 // register move) 3456 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) { 3457 // Perform it only for canonical counted loops. 3458 // Loop's shape could be messed up by iteration_split_impl. 3459 if (!loop->_head->is_CountedLoop()) 3460 return; 3461 if (!loop->_head->as_Loop()->is_valid_counted_loop()) 3462 return; 3463 3464 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 3465 CountedLoopEndNode *cle = cl->loopexit(); 3466 Node *exit = cle->proj_out(false); 3467 Node *phi = cl->phi(); 3468 3469 // Check for the special case when using the pre-incremented trip-counter on 3470 // the fall-out path (forces the pre-incremented and post-incremented trip 3471 // counter to be live at the same time). Fix this by adjusting to use the 3472 // post-increment trip counter. 3473 3474 bool progress = true; 3475 while (progress) { 3476 progress = false; 3477 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) { 3478 Node* use = phi->fast_out(i); // User of trip-counter 3479 if (!has_ctrl(use)) continue; 3480 Node *u_ctrl = get_ctrl(use); 3481 if (use->is_Phi()) { 3482 u_ctrl = NULL; 3483 for (uint j = 1; j < use->req(); j++) 3484 if (use->in(j) == phi) 3485 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j)); 3486 } 3487 IdealLoopTree *u_loop = get_loop(u_ctrl); 3488 // Look for loop-invariant use 3489 if (u_loop == loop) continue; 3490 if (loop->is_member(u_loop)) continue; 3491 // Check that use is live out the bottom. Assuming the trip-counter 3492 // update is right at the bottom, uses of of the loop middle are ok. 3493 if (dom_lca(exit, u_ctrl) != exit) continue; 3494 // Hit! Refactor use to use the post-incremented tripcounter. 3495 // Compute a post-increment tripcounter. 3496 Node* c = exit; 3497 if (cl->is_strip_mined()) { 3498 IdealLoopTree* outer_loop = get_loop(cl->outer_loop()); 3499 if (!outer_loop->is_member(u_loop)) { 3500 c = cl->outer_loop_exit(); 3501 } 3502 } 3503 Node *opaq = new Opaque2Node(C, cle->incr()); 3504 register_new_node(opaq, c); 3505 Node *neg_stride = _igvn.intcon(-cle->stride_con()); 3506 set_ctrl(neg_stride, C->root()); 3507 Node *post = new AddINode(opaq, neg_stride); 3508 register_new_node(post, c); 3509 _igvn.rehash_node_delayed(use); 3510 for (uint j = 1; j < use->req(); j++) { 3511 if (use->in(j) == phi) 3512 use->set_req(j, post); 3513 } 3514 // Since DU info changed, rerun loop 3515 progress = true; 3516 break; 3517 } 3518 } 3519 3520 }