1 /* 2 * Copyright (c) 1999, 2019, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "gc/shared/barrierSet.hpp" 27 #include "gc/shared/c2/barrierSetC2.hpp" 28 #include "memory/allocation.inline.hpp" 29 #include "memory/resourceArea.hpp" 30 #include "opto/addnode.hpp" 31 #include "opto/callnode.hpp" 32 #include "opto/castnode.hpp" 33 #include "opto/connode.hpp" 34 #include "opto/castnode.hpp" 35 #include "opto/divnode.hpp" 36 #include "opto/loopnode.hpp" 37 #include "opto/matcher.hpp" 38 #include "opto/mulnode.hpp" 39 #include "opto/movenode.hpp" 40 #include "opto/opaquenode.hpp" 41 #include "opto/rootnode.hpp" 42 #include "opto/subnode.hpp" 43 #include "opto/valuetypenode.hpp" 44 #include "utilities/macros.hpp" 45 #if INCLUDE_ZGC 46 #include "gc/z/c2/zBarrierSetC2.hpp" 47 #endif 48 49 //============================================================================= 50 //------------------------------split_thru_phi--------------------------------- 51 // Split Node 'n' through merge point if there is enough win. 52 Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) { 53 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) { 54 // ConvI2L may have type information on it which is unsafe to push up 55 // so disable this for now 56 return NULL; 57 } 58 59 // Splitting range check CastIIs through a loop induction Phi can 60 // cause new Phis to be created that are left unrelated to the loop 61 // induction Phi and prevent optimizations (vectorization) 62 if (n->Opcode() == Op_CastII && n->as_CastII()->has_range_check() && 63 region->is_CountedLoop() && n->in(1) == region->as_CountedLoop()->phi()) { 64 return NULL; 65 } 66 67 // Value types should not be split through Phis because they cannot be merged 68 // through Phi nodes but each value input needs to be merged individually. 69 if (n->is_ValueType()) { 70 return NULL; 71 } 72 73 int wins = 0; 74 assert(!n->is_CFG(), ""); 75 assert(region->is_Region(), ""); 76 77 const Type* type = n->bottom_type(); 78 const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr(); 79 Node *phi; 80 if (t_oop != NULL && t_oop->is_known_instance_field()) { 81 int iid = t_oop->instance_id(); 82 int index = C->get_alias_index(t_oop); 83 int offset = t_oop->offset(); 84 phi = new PhiNode(region, type, NULL, iid, index, offset); 85 } else { 86 phi = PhiNode::make_blank(region, n); 87 } 88 uint old_unique = C->unique(); 89 for (uint i = 1; i < region->req(); i++) { 90 Node *x; 91 Node* the_clone = NULL; 92 if (region->in(i) == C->top()) { 93 x = C->top(); // Dead path? Use a dead data op 94 } else { 95 x = n->clone(); // Else clone up the data op 96 the_clone = x; // Remember for possible deletion. 97 // Alter data node to use pre-phi inputs 98 if (n->in(0) == region) 99 x->set_req( 0, region->in(i) ); 100 for (uint j = 1; j < n->req(); j++) { 101 Node *in = n->in(j); 102 if (in->is_Phi() && in->in(0) == region) 103 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone 104 } 105 } 106 // Check for a 'win' on some paths 107 const Type *t = x->Value(&_igvn); 108 109 bool singleton = t->singleton(); 110 111 // A TOP singleton indicates that there are no possible values incoming 112 // along a particular edge. In most cases, this is OK, and the Phi will 113 // be eliminated later in an Ideal call. However, we can't allow this to 114 // happen if the singleton occurs on loop entry, as the elimination of 115 // the PhiNode may cause the resulting node to migrate back to a previous 116 // loop iteration. 117 if (singleton && t == Type::TOP) { 118 // Is_Loop() == false does not confirm the absence of a loop (e.g., an 119 // irreducible loop may not be indicated by an affirmative is_Loop()); 120 // therefore, the only top we can split thru a phi is on a backedge of 121 // a loop. 122 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 123 } 124 125 if (singleton) { 126 wins++; 127 x = ((PhaseGVN&)_igvn).makecon(t); 128 } else { 129 // We now call Identity to try to simplify the cloned node. 130 // Note that some Identity methods call phase->type(this). 131 // Make sure that the type array is big enough for 132 // our new node, even though we may throw the node away. 133 // (Note: This tweaking with igvn only works because x is a new node.) 134 _igvn.set_type(x, t); 135 // If x is a TypeNode, capture any more-precise type permanently into Node 136 // otherwise it will be not updated during igvn->transform since 137 // igvn->type(x) is set to x->Value() already. 138 x->raise_bottom_type(t); 139 Node *y = _igvn.apply_identity(x); 140 if (y != x) { 141 wins++; 142 x = y; 143 } else { 144 y = _igvn.hash_find(x); 145 if (y) { 146 wins++; 147 x = y; 148 } else { 149 // Else x is a new node we are keeping 150 // We do not need register_new_node_with_optimizer 151 // because set_type has already been called. 152 _igvn._worklist.push(x); 153 } 154 } 155 } 156 if (x != the_clone && the_clone != NULL) 157 _igvn.remove_dead_node(the_clone); 158 phi->set_req( i, x ); 159 } 160 // Too few wins? 161 if (wins <= policy) { 162 _igvn.remove_dead_node(phi); 163 return NULL; 164 } 165 166 // Record Phi 167 register_new_node( phi, region ); 168 169 for (uint i2 = 1; i2 < phi->req(); i2++) { 170 Node *x = phi->in(i2); 171 // If we commoned up the cloned 'x' with another existing Node, 172 // the existing Node picks up a new use. We need to make the 173 // existing Node occur higher up so it dominates its uses. 174 Node *old_ctrl; 175 IdealLoopTree *old_loop; 176 177 if (x->is_Con()) { 178 // Constant's control is always root. 179 set_ctrl(x, C->root()); 180 continue; 181 } 182 // The occasional new node 183 if (x->_idx >= old_unique) { // Found a new, unplaced node? 184 old_ctrl = NULL; 185 old_loop = NULL; // Not in any prior loop 186 } else { 187 old_ctrl = get_ctrl(x); 188 old_loop = get_loop(old_ctrl); // Get prior loop 189 } 190 // New late point must dominate new use 191 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2)); 192 if (new_ctrl == old_ctrl) // Nothing is changed 193 continue; 194 195 IdealLoopTree *new_loop = get_loop(new_ctrl); 196 197 // Don't move x into a loop if its uses are 198 // outside of loop. Otherwise x will be cloned 199 // for each use outside of this loop. 200 IdealLoopTree *use_loop = get_loop(region); 201 if (!new_loop->is_member(use_loop) && 202 (old_loop == NULL || !new_loop->is_member(old_loop))) { 203 // Take early control, later control will be recalculated 204 // during next iteration of loop optimizations. 205 new_ctrl = get_early_ctrl(x); 206 new_loop = get_loop(new_ctrl); 207 } 208 // Set new location 209 set_ctrl(x, new_ctrl); 210 // If changing loop bodies, see if we need to collect into new body 211 if (old_loop != new_loop) { 212 if (old_loop && !old_loop->_child) 213 old_loop->_body.yank(x); 214 if (!new_loop->_child) 215 new_loop->_body.push(x); // Collect body info 216 } 217 } 218 219 return phi; 220 } 221 222 //------------------------------dominated_by------------------------------------ 223 // Replace the dominated test with an obvious true or false. Place it on the 224 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the 225 // live path up to the dominating control. 226 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff, bool flip, bool exclude_loop_predicate ) { 227 if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); } 228 229 // prevdom is the dominating projection of the dominating test. 230 assert( iff->is_If(), "" ); 231 assert(iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd || iff->Opcode() == Op_RangeCheck, "Check this code when new subtype is added"); 232 int pop = prevdom->Opcode(); 233 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" ); 234 if (flip) { 235 if (pop == Op_IfTrue) 236 pop = Op_IfFalse; 237 else 238 pop = Op_IfTrue; 239 } 240 // 'con' is set to true or false to kill the dominated test. 241 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO); 242 set_ctrl(con, C->root()); // Constant gets a new use 243 // Hack the dominated test 244 _igvn.replace_input_of(iff, 1, con); 245 246 // If I dont have a reachable TRUE and FALSE path following the IfNode then 247 // I can assume this path reaches an infinite loop. In this case it's not 248 // important to optimize the data Nodes - either the whole compilation will 249 // be tossed or this path (and all data Nodes) will go dead. 250 if (iff->outcnt() != 2) return; 251 252 // Make control-dependent data Nodes on the live path (path that will remain 253 // once the dominated IF is removed) become control-dependent on the 254 // dominating projection. 255 Node* dp = iff->as_If()->proj_out_or_null(pop == Op_IfTrue); 256 257 // Loop predicates may have depending checks which should not 258 // be skipped. For example, range check predicate has two checks 259 // for lower and upper bounds. 260 if (dp == NULL) 261 return; 262 263 ProjNode* dp_proj = dp->as_Proj(); 264 ProjNode* unc_proj = iff->as_If()->proj_out(1 - dp_proj->_con)->as_Proj(); 265 if (exclude_loop_predicate && 266 (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) != NULL || 267 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_profile_predicate) != NULL || 268 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check) != NULL)) { 269 // If this is a range check (IfNode::is_range_check), do not 270 // reorder because Compile::allow_range_check_smearing might have 271 // changed the check. 272 return; // Let IGVN transformation change control dependence. 273 } 274 275 IdealLoopTree *old_loop = get_loop(dp); 276 277 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { 278 Node* cd = dp->fast_out(i); // Control-dependent node 279 if (cd->depends_only_on_test()) { 280 assert(cd->in(0) == dp, ""); 281 _igvn.replace_input_of(cd, 0, prevdom); 282 set_early_ctrl(cd); 283 IdealLoopTree *new_loop = get_loop(get_ctrl(cd)); 284 if (old_loop != new_loop) { 285 if (!old_loop->_child) old_loop->_body.yank(cd); 286 if (!new_loop->_child) new_loop->_body.push(cd); 287 } 288 --i; 289 --imax; 290 } 291 } 292 } 293 294 //------------------------------has_local_phi_input---------------------------- 295 // Return TRUE if 'n' has Phi inputs from its local block and no other 296 // block-local inputs (all non-local-phi inputs come from earlier blocks) 297 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) { 298 Node *n_ctrl = get_ctrl(n); 299 // See if some inputs come from a Phi in this block, or from before 300 // this block. 301 uint i; 302 for( i = 1; i < n->req(); i++ ) { 303 Node *phi = n->in(i); 304 if( phi->is_Phi() && phi->in(0) == n_ctrl ) 305 break; 306 } 307 if( i >= n->req() ) 308 return NULL; // No Phi inputs; nowhere to clone thru 309 310 // Check for inputs created between 'n' and the Phi input. These 311 // must split as well; they have already been given the chance 312 // (courtesy of a post-order visit) and since they did not we must 313 // recover the 'cost' of splitting them by being very profitable 314 // when splitting 'n'. Since this is unlikely we simply give up. 315 for( i = 1; i < n->req(); i++ ) { 316 Node *m = n->in(i); 317 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) { 318 // We allow the special case of AddP's with no local inputs. 319 // This allows us to split-up address expressions. 320 if (m->is_AddP() && 321 get_ctrl(m->in(2)) != n_ctrl && 322 get_ctrl(m->in(3)) != n_ctrl) { 323 // Move the AddP up to dominating point 324 Node* c = find_non_split_ctrl(idom(n_ctrl)); 325 if (c->is_OuterStripMinedLoop()) { 326 c->as_Loop()->verify_strip_mined(1); 327 c = c->in(LoopNode::EntryControl); 328 } 329 set_ctrl_and_loop(m, c); 330 continue; 331 } 332 return NULL; 333 } 334 assert(n->is_Phi() || m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control"); 335 } 336 337 return n_ctrl; 338 } 339 340 //------------------------------remix_address_expressions---------------------- 341 // Rework addressing expressions to get the most loop-invariant stuff 342 // moved out. We'd like to do all associative operators, but it's especially 343 // important (common) to do address expressions. 344 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) { 345 if (!has_ctrl(n)) return NULL; 346 Node *n_ctrl = get_ctrl(n); 347 IdealLoopTree *n_loop = get_loop(n_ctrl); 348 349 // See if 'n' mixes loop-varying and loop-invariant inputs and 350 // itself is loop-varying. 351 352 // Only interested in binary ops (and AddP) 353 if( n->req() < 3 || n->req() > 4 ) return NULL; 354 355 Node *n1_ctrl = get_ctrl(n->in( 1)); 356 Node *n2_ctrl = get_ctrl(n->in( 2)); 357 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3)); 358 IdealLoopTree *n1_loop = get_loop( n1_ctrl ); 359 IdealLoopTree *n2_loop = get_loop( n2_ctrl ); 360 IdealLoopTree *n3_loop = get_loop( n3_ctrl ); 361 362 // Does one of my inputs spin in a tighter loop than self? 363 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) || 364 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) || 365 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) ) 366 return NULL; // Leave well enough alone 367 368 // Is at least one of my inputs loop-invariant? 369 if( n1_loop == n_loop && 370 n2_loop == n_loop && 371 n3_loop == n_loop ) 372 return NULL; // No loop-invariant inputs 373 374 375 int n_op = n->Opcode(); 376 377 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2). 378 if( n_op == Op_LShiftI ) { 379 // Scale is loop invariant 380 Node *scale = n->in(2); 381 Node *scale_ctrl = get_ctrl(scale); 382 IdealLoopTree *scale_loop = get_loop(scale_ctrl ); 383 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) ) 384 return NULL; 385 const TypeInt *scale_t = scale->bottom_type()->isa_int(); 386 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 ) 387 return NULL; // Dont bother with byte/short masking 388 // Add must vary with loop (else shift would be loop-invariant) 389 Node *add = n->in(1); 390 Node *add_ctrl = get_ctrl(add); 391 IdealLoopTree *add_loop = get_loop(add_ctrl); 392 //assert( n_loop == add_loop, "" ); 393 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops 394 395 // Convert I-V into I+ (0-V); same for V-I 396 if( add->Opcode() == Op_SubI && 397 _igvn.type( add->in(1) ) != TypeInt::ZERO ) { 398 Node *zero = _igvn.intcon(0); 399 set_ctrl(zero, C->root()); 400 Node *neg = new SubINode( _igvn.intcon(0), add->in(2) ); 401 register_new_node( neg, get_ctrl(add->in(2) ) ); 402 add = new AddINode( add->in(1), neg ); 403 register_new_node( add, add_ctrl ); 404 } 405 if( add->Opcode() != Op_AddI ) return NULL; 406 // See if one add input is loop invariant 407 Node *add_var = add->in(1); 408 Node *add_var_ctrl = get_ctrl(add_var); 409 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 410 Node *add_invar = add->in(2); 411 Node *add_invar_ctrl = get_ctrl(add_invar); 412 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl ); 413 if( add_var_loop == n_loop ) { 414 } else if( add_invar_loop == n_loop ) { 415 // Swap to find the invariant part 416 add_invar = add_var; 417 add_invar_ctrl = add_var_ctrl; 418 add_invar_loop = add_var_loop; 419 add_var = add->in(2); 420 Node *add_var_ctrl = get_ctrl(add_var); 421 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 422 } else // Else neither input is loop invariant 423 return NULL; 424 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) ) 425 return NULL; // No invariant part of the add? 426 427 // Yes! Reshape address expression! 428 Node *inv_scale = new LShiftINode( add_invar, scale ); 429 Node *inv_scale_ctrl = 430 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ? 431 add_invar_ctrl : scale_ctrl; 432 register_new_node( inv_scale, inv_scale_ctrl ); 433 Node *var_scale = new LShiftINode( add_var, scale ); 434 register_new_node( var_scale, n_ctrl ); 435 Node *var_add = new AddINode( var_scale, inv_scale ); 436 register_new_node( var_add, n_ctrl ); 437 _igvn.replace_node( n, var_add ); 438 return var_add; 439 } 440 441 // Replace (I+V) with (V+I) 442 if( n_op == Op_AddI || 443 n_op == Op_AddL || 444 n_op == Op_AddF || 445 n_op == Op_AddD || 446 n_op == Op_MulI || 447 n_op == Op_MulL || 448 n_op == Op_MulF || 449 n_op == Op_MulD ) { 450 if( n2_loop == n_loop ) { 451 assert( n1_loop != n_loop, "" ); 452 n->swap_edges(1, 2); 453 } 454 } 455 456 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V), 457 // but not if I2 is a constant. 458 if( n_op == Op_AddP ) { 459 if( n2_loop == n_loop && n3_loop != n_loop ) { 460 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) { 461 Node *n22_ctrl = get_ctrl(n->in(2)->in(2)); 462 Node *n23_ctrl = get_ctrl(n->in(2)->in(3)); 463 IdealLoopTree *n22loop = get_loop( n22_ctrl ); 464 IdealLoopTree *n23_loop = get_loop( n23_ctrl ); 465 if( n22loop != n_loop && n22loop->is_member(n_loop) && 466 n23_loop == n_loop ) { 467 Node *add1 = new AddPNode( n->in(1), n->in(2)->in(2), n->in(3) ); 468 // Stuff new AddP in the loop preheader 469 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) ); 470 Node *add2 = new AddPNode( n->in(1), add1, n->in(2)->in(3) ); 471 register_new_node( add2, n_ctrl ); 472 _igvn.replace_node( n, add2 ); 473 return add2; 474 } 475 } 476 } 477 478 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V) 479 if (n2_loop != n_loop && n3_loop == n_loop) { 480 if (n->in(3)->Opcode() == Op_AddX) { 481 Node *V = n->in(3)->in(1); 482 Node *I = n->in(3)->in(2); 483 if (is_member(n_loop,get_ctrl(V))) { 484 } else { 485 Node *tmp = V; V = I; I = tmp; 486 } 487 if (!is_member(n_loop,get_ctrl(I))) { 488 Node *add1 = new AddPNode(n->in(1), n->in(2), I); 489 // Stuff new AddP in the loop preheader 490 register_new_node(add1, n_loop->_head->in(LoopNode::EntryControl)); 491 Node *add2 = new AddPNode(n->in(1), add1, V); 492 register_new_node(add2, n_ctrl); 493 _igvn.replace_node(n, add2); 494 return add2; 495 } 496 } 497 } 498 } 499 500 return NULL; 501 } 502 503 // Optimize ((in1[2*i] * in2[2*i]) + (in1[2*i+1] * in2[2*i+1])) 504 Node *PhaseIdealLoop::convert_add_to_muladd(Node* n) { 505 assert(n->Opcode() == Op_AddI, "sanity"); 506 Node * nn = NULL; 507 Node * in1 = n->in(1); 508 Node * in2 = n->in(2); 509 if (in1->Opcode() == Op_MulI && in2->Opcode() == Op_MulI) { 510 IdealLoopTree* loop_n = get_loop(get_ctrl(n)); 511 if (loop_n->_head->as_Loop()->is_valid_counted_loop() && 512 Matcher::match_rule_supported(Op_MulAddS2I) && 513 Matcher::match_rule_supported(Op_MulAddVS2VI)) { 514 Node* mul_in1 = in1->in(1); 515 Node* mul_in2 = in1->in(2); 516 Node* mul_in3 = in2->in(1); 517 Node* mul_in4 = in2->in(2); 518 if (mul_in1->Opcode() == Op_LoadS && 519 mul_in2->Opcode() == Op_LoadS && 520 mul_in3->Opcode() == Op_LoadS && 521 mul_in4->Opcode() == Op_LoadS) { 522 IdealLoopTree* loop1 = get_loop(get_ctrl(mul_in1)); 523 IdealLoopTree* loop2 = get_loop(get_ctrl(mul_in2)); 524 IdealLoopTree* loop3 = get_loop(get_ctrl(mul_in3)); 525 IdealLoopTree* loop4 = get_loop(get_ctrl(mul_in4)); 526 IdealLoopTree* loop5 = get_loop(get_ctrl(in1)); 527 IdealLoopTree* loop6 = get_loop(get_ctrl(in2)); 528 // All nodes should be in the same counted loop. 529 if (loop_n == loop1 && loop_n == loop2 && loop_n == loop3 && 530 loop_n == loop4 && loop_n == loop5 && loop_n == loop6) { 531 Node* adr1 = mul_in1->in(MemNode::Address); 532 Node* adr2 = mul_in2->in(MemNode::Address); 533 Node* adr3 = mul_in3->in(MemNode::Address); 534 Node* adr4 = mul_in4->in(MemNode::Address); 535 if (adr1->is_AddP() && adr2->is_AddP() && adr3->is_AddP() && adr4->is_AddP()) { 536 if ((adr1->in(AddPNode::Base) == adr3->in(AddPNode::Base)) && 537 (adr2->in(AddPNode::Base) == adr4->in(AddPNode::Base))) { 538 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in3, mul_in4); 539 register_new_node(nn, get_ctrl(n)); 540 _igvn.replace_node(n, nn); 541 return nn; 542 } else if ((adr1->in(AddPNode::Base) == adr4->in(AddPNode::Base)) && 543 (adr2->in(AddPNode::Base) == adr3->in(AddPNode::Base))) { 544 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in4, mul_in3); 545 register_new_node(nn, get_ctrl(n)); 546 _igvn.replace_node(n, nn); 547 return nn; 548 } 549 } 550 } 551 } 552 } 553 } 554 return nn; 555 } 556 557 //------------------------------conditional_move------------------------------- 558 // Attempt to replace a Phi with a conditional move. We have some pretty 559 // strict profitability requirements. All Phis at the merge point must 560 // be converted, so we can remove the control flow. We need to limit the 561 // number of c-moves to a small handful. All code that was in the side-arms 562 // of the CFG diamond is now speculatively executed. This code has to be 563 // "cheap enough". We are pretty much limited to CFG diamonds that merge 564 // 1 or 2 items with a total of 1 or 2 ops executed speculatively. 565 Node *PhaseIdealLoop::conditional_move( Node *region ) { 566 567 assert(region->is_Region(), "sanity check"); 568 if (region->req() != 3) return NULL; 569 570 // Check for CFG diamond 571 Node *lp = region->in(1); 572 Node *rp = region->in(2); 573 if (!lp || !rp) return NULL; 574 Node *lp_c = lp->in(0); 575 if (lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If()) return NULL; 576 IfNode *iff = lp_c->as_If(); 577 578 // Check for ops pinned in an arm of the diamond. 579 // Can't remove the control flow in this case 580 if (lp->outcnt() > 1) return NULL; 581 if (rp->outcnt() > 1) return NULL; 582 583 IdealLoopTree* r_loop = get_loop(region); 584 assert(r_loop == get_loop(iff), "sanity"); 585 // Always convert to CMOVE if all results are used only outside this loop. 586 bool used_inside_loop = (r_loop == _ltree_root); 587 588 // Check profitability 589 int cost = 0; 590 int phis = 0; 591 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 592 Node *out = region->fast_out(i); 593 if (!out->is_Phi()) continue; // Ignore other control edges, etc 594 phis++; 595 PhiNode* phi = out->as_Phi(); 596 BasicType bt = phi->type()->basic_type(); 597 switch (bt) { 598 case T_DOUBLE: 599 case T_FLOAT: 600 if (C->use_cmove()) { 601 continue; //TODO: maybe we want to add some cost 602 } 603 cost += Matcher::float_cmove_cost(); // Could be very expensive 604 break; 605 case T_LONG: { 606 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's 607 } 608 case T_INT: // These all CMOV fine 609 case T_ADDRESS: { // (RawPtr) 610 cost++; 611 break; 612 } 613 case T_NARROWOOP: // Fall through 614 case T_OBJECT: { // Base oops are OK, but not derived oops 615 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr(); 616 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a 617 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus 618 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we 619 // have a Phi for the base here that we convert to a CMOVE all is well 620 // and good. But if the base is dead, we'll not make a CMOVE. Later 621 // the allocator will have to produce a base by creating a CMOVE of the 622 // relevant bases. This puts the allocator in the business of 623 // manufacturing expensive instructions, generally a bad plan. 624 // Just Say No to Conditionally-Moved Derived Pointers. 625 if (tp && tp->offset() != 0) 626 return NULL; 627 cost++; 628 break; 629 } 630 default: 631 return NULL; // In particular, can't do memory or I/O 632 } 633 // Add in cost any speculative ops 634 for (uint j = 1; j < region->req(); j++) { 635 Node *proj = region->in(j); 636 Node *inp = phi->in(j); 637 if (get_ctrl(inp) == proj) { // Found local op 638 cost++; 639 // Check for a chain of dependent ops; these will all become 640 // speculative in a CMOV. 641 for (uint k = 1; k < inp->req(); k++) 642 if (get_ctrl(inp->in(k)) == proj) 643 cost += ConditionalMoveLimit; // Too much speculative goo 644 } 645 } 646 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode. 647 // This will likely Split-If, a higher-payoff operation. 648 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) { 649 Node* use = phi->fast_out(k); 650 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr()) 651 cost += ConditionalMoveLimit; 652 // Is there a use inside the loop? 653 // Note: check only basic types since CMoveP is pinned. 654 if (!used_inside_loop && is_java_primitive(bt)) { 655 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use); 656 if (r_loop == u_loop || r_loop->is_member(u_loop)) { 657 used_inside_loop = true; 658 } 659 } 660 } 661 }//for 662 Node* bol = iff->in(1); 663 assert(bol->Opcode() == Op_Bool, ""); 664 int cmp_op = bol->in(1)->Opcode(); 665 // It is expensive to generate flags from a float compare. 666 // Avoid duplicated float compare. 667 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL; 668 669 float infrequent_prob = PROB_UNLIKELY_MAG(3); 670 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop. 671 if (used_inside_loop) { 672 if (cost >= ConditionalMoveLimit) return NULL; // Too much goo 673 674 // BlockLayoutByFrequency optimization moves infrequent branch 675 // from hot path. No point in CMOV'ing in such case (110 is used 676 // instead of 100 to take into account not exactness of float value). 677 if (BlockLayoutByFrequency) { 678 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f); 679 } 680 } 681 // Check for highly predictable branch. No point in CMOV'ing if 682 // we are going to predict accurately all the time. 683 if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) { 684 //keep going 685 } else if (iff->_prob < infrequent_prob || 686 iff->_prob > (1.0f - infrequent_prob)) 687 return NULL; 688 689 // -------------- 690 // Now replace all Phis with CMOV's 691 Node *cmov_ctrl = iff->in(0); 692 uint flip = (lp->Opcode() == Op_IfTrue); 693 Node_List wq; 694 while (1) { 695 PhiNode* phi = NULL; 696 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 697 Node *out = region->fast_out(i); 698 if (out->is_Phi()) { 699 phi = out->as_Phi(); 700 break; 701 } 702 } 703 if (phi == NULL) break; 704 if (PrintOpto && VerifyLoopOptimizations) { tty->print_cr("CMOV"); } 705 // Move speculative ops 706 wq.push(phi); 707 while (wq.size() > 0) { 708 Node *n = wq.pop(); 709 for (uint j = 1; j < n->req(); j++) { 710 Node* m = n->in(j); 711 if (m != NULL && !is_dominator(get_ctrl(m), cmov_ctrl)) { 712 #ifndef PRODUCT 713 if (PrintOpto && VerifyLoopOptimizations) { 714 tty->print(" speculate: "); 715 m->dump(); 716 } 717 #endif 718 set_ctrl(m, cmov_ctrl); 719 wq.push(m); 720 } 721 } 722 } 723 Node *cmov = CMoveNode::make(cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi)); 724 register_new_node( cmov, cmov_ctrl ); 725 _igvn.replace_node( phi, cmov ); 726 #ifndef PRODUCT 727 if (TraceLoopOpts) { 728 tty->print("CMOV "); 729 r_loop->dump_head(); 730 if (Verbose) { 731 bol->in(1)->dump(1); 732 cmov->dump(1); 733 } 734 } 735 if (VerifyLoopOptimizations) verify(); 736 #endif 737 } 738 739 // The useless CFG diamond will fold up later; see the optimization in 740 // RegionNode::Ideal. 741 _igvn._worklist.push(region); 742 743 return iff->in(1); 744 } 745 746 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) { 747 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 748 Node* u = m->fast_out(i); 749 if (u->is_CFG()) { 750 if (u->Opcode() == Op_NeverBranch) { 751 u = ((NeverBranchNode*)u)->proj_out(0); 752 enqueue_cfg_uses(u, wq); 753 } else { 754 wq.push(u); 755 } 756 } 757 } 758 } 759 760 // Try moving a store out of a loop, right before the loop 761 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) { 762 // Store has to be first in the loop body 763 IdealLoopTree *n_loop = get_loop(n_ctrl); 764 if (n->is_Store() && n_loop != _ltree_root && 765 n_loop->is_loop() && n_loop->_head->is_Loop() && 766 n->in(0) != NULL) { 767 Node* address = n->in(MemNode::Address); 768 Node* value = n->in(MemNode::ValueIn); 769 Node* mem = n->in(MemNode::Memory); 770 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 771 IdealLoopTree* value_loop = get_loop(get_ctrl(value)); 772 773 // - address and value must be loop invariant 774 // - memory must be a memory Phi for the loop 775 // - Store must be the only store on this memory slice in the 776 // loop: if there's another store following this one then value 777 // written at iteration i by the second store could be overwritten 778 // at iteration i+n by the first store: it's not safe to move the 779 // first store out of the loop 780 // - nothing must observe the memory Phi: it guarantees no read 781 // before the store, we are also guaranteed the store post 782 // dominates the loop head (ignoring a possible early 783 // exit). Otherwise there would be extra Phi involved between the 784 // loop's Phi and the store. 785 // - there must be no early exit from the loop before the Store 786 // (such an exit most of the time would be an extra use of the 787 // memory Phi but sometimes is a bottom memory Phi that takes the 788 // store as input). 789 790 if (!n_loop->is_member(address_loop) && 791 !n_loop->is_member(value_loop) && 792 mem->is_Phi() && mem->in(0) == n_loop->_head && 793 mem->outcnt() == 1 && 794 mem->in(LoopNode::LoopBackControl) == n) { 795 796 assert(n_loop->_tail != NULL, "need a tail"); 797 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop"); 798 799 // Verify that there's no early exit of the loop before the store. 800 bool ctrl_ok = false; 801 { 802 // Follow control from loop head until n, we exit the loop or 803 // we reach the tail 804 ResourceMark rm; 805 Unique_Node_List wq; 806 wq.push(n_loop->_head); 807 808 for (uint next = 0; next < wq.size(); ++next) { 809 Node *m = wq.at(next); 810 if (m == n->in(0)) { 811 ctrl_ok = true; 812 continue; 813 } 814 assert(!has_ctrl(m), "should be CFG"); 815 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) { 816 ctrl_ok = false; 817 break; 818 } 819 enqueue_cfg_uses(m, wq); 820 if (wq.size() > 10) { 821 ctrl_ok = false; 822 break; 823 } 824 } 825 } 826 if (ctrl_ok) { 827 // move the Store 828 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem); 829 _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl)); 830 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl)); 831 // Disconnect the phi now. An empty phi can confuse other 832 // optimizations in this pass of loop opts. 833 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl)); 834 n_loop->_body.yank(mem); 835 836 set_ctrl_and_loop(n, n->in(0)); 837 838 return n; 839 } 840 } 841 } 842 return NULL; 843 } 844 845 // Try moving a store out of a loop, right after the loop 846 void PhaseIdealLoop::try_move_store_after_loop(Node* n) { 847 if (n->is_Store() && n->in(0) != NULL) { 848 Node *n_ctrl = get_ctrl(n); 849 IdealLoopTree *n_loop = get_loop(n_ctrl); 850 // Store must be in a loop 851 if (n_loop != _ltree_root && !n_loop->_irreducible) { 852 Node* address = n->in(MemNode::Address); 853 Node* value = n->in(MemNode::ValueIn); 854 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 855 // address must be loop invariant 856 if (!n_loop->is_member(address_loop)) { 857 // Store must be last on this memory slice in the loop and 858 // nothing in the loop must observe it 859 Node* phi = NULL; 860 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 861 Node* u = n->fast_out(i); 862 if (has_ctrl(u)) { // control use? 863 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 864 if (!n_loop->is_member(u_loop)) { 865 continue; 866 } 867 if (u->is_Phi() && u->in(0) == n_loop->_head) { 868 assert(_igvn.type(u) == Type::MEMORY, "bad phi"); 869 // multiple phis on the same slice are possible 870 if (phi != NULL) { 871 return; 872 } 873 phi = u; 874 continue; 875 } 876 } 877 return; 878 } 879 if (phi != NULL) { 880 // Nothing in the loop before the store (next iteration) 881 // must observe the stored value 882 bool mem_ok = true; 883 { 884 ResourceMark rm; 885 Unique_Node_List wq; 886 wq.push(phi); 887 for (uint next = 0; next < wq.size() && mem_ok; ++next) { 888 Node *m = wq.at(next); 889 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) { 890 Node* u = m->fast_out(i); 891 if (u->is_Store() || u->is_Phi()) { 892 if (u != n) { 893 wq.push(u); 894 mem_ok = (wq.size() <= 10); 895 } 896 } else { 897 mem_ok = false; 898 break; 899 } 900 } 901 } 902 } 903 if (mem_ok) { 904 // Move the store out of the loop if the LCA of all 905 // users (except for the phi) is outside the loop. 906 Node* hook = new Node(1); 907 _igvn.rehash_node_delayed(phi); 908 int count = phi->replace_edge(n, hook); 909 assert(count > 0, "inconsistent phi"); 910 911 // Compute latest point this store can go 912 Node* lca = get_late_ctrl(n, get_ctrl(n)); 913 if (n_loop->is_member(get_loop(lca))) { 914 // LCA is in the loop - bail out 915 _igvn.replace_node(hook, n); 916 return; 917 } 918 #ifdef ASSERT 919 if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) { 920 assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined"); 921 n_loop->_head->as_Loop()->verify_strip_mined(1); 922 Node* outer = n_loop->_head->as_CountedLoop()->outer_loop(); 923 IdealLoopTree* outer_loop = get_loop(outer); 924 assert(n_loop->_parent == outer_loop, "broken loop tree"); 925 assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state"); 926 } 927 #endif 928 929 // Move store out of the loop 930 _igvn.replace_node(hook, n->in(MemNode::Memory)); 931 _igvn.replace_input_of(n, 0, lca); 932 set_ctrl_and_loop(n, lca); 933 934 // Disconnect the phi now. An empty phi can confuse other 935 // optimizations in this pass of loop opts.. 936 if (phi->in(LoopNode::LoopBackControl) == phi) { 937 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl)); 938 n_loop->_body.yank(phi); 939 } 940 } 941 } 942 } 943 } 944 } 945 } 946 947 //------------------------------split_if_with_blocks_pre----------------------- 948 // Do the real work in a non-recursive function. Data nodes want to be 949 // cloned in the pre-order so they can feed each other nicely. 950 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) { 951 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 952 Node* bs_res = bs->split_if_pre(this, n); 953 if (bs_res != NULL) { 954 return bs_res; 955 } 956 // Cloning these guys is unlikely to win 957 int n_op = n->Opcode(); 958 if( n_op == Op_MergeMem ) return n; 959 if( n->is_Proj() ) return n; 960 // Do not clone-up CmpFXXX variations, as these are always 961 // followed by a CmpI 962 if( n->is_Cmp() ) return n; 963 // Attempt to use a conditional move instead of a phi/branch 964 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) { 965 Node *cmov = conditional_move( n ); 966 if( cmov ) return cmov; 967 } 968 if( n->is_CFG() || n->is_LoadStore() ) 969 return n; 970 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd 971 n_op == Op_Opaque2 ) { 972 if( !C->major_progress() ) // If chance of no more loop opts... 973 _igvn._worklist.push(n); // maybe we'll remove them 974 return n; 975 } 976 977 if( n->is_Con() ) return n; // No cloning for Con nodes 978 979 Node *n_ctrl = get_ctrl(n); 980 if( !n_ctrl ) return n; // Dead node 981 982 Node* res = try_move_store_before_loop(n, n_ctrl); 983 if (res != NULL) { 984 return n; 985 } 986 987 // Attempt to remix address expressions for loop invariants 988 Node *m = remix_address_expressions( n ); 989 if( m ) return m; 990 991 if (n_op == Op_AddI) { 992 Node *nn = convert_add_to_muladd( n ); 993 if ( nn ) return nn; 994 } 995 996 if (n->is_ConstraintCast()) { 997 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this); 998 // ConstraintCastNode::dominating_cast() uses node control input to determine domination. 999 // Node control inputs don't necessarily agree with loop control info (due to 1000 // transformations happened in between), thus additional dominance check is needed 1001 // to keep loop info valid. 1002 if (dom_cast != NULL && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) { 1003 _igvn.replace_node(n, dom_cast); 1004 return dom_cast; 1005 } 1006 } 1007 1008 // Determine if the Node has inputs from some local Phi. 1009 // Returns the block to clone thru. 1010 Node *n_blk = has_local_phi_input( n ); 1011 if( !n_blk ) return n; 1012 1013 // Do not clone the trip counter through on a CountedLoop 1014 // (messes up the canonical shape). 1015 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n; 1016 1017 // Check for having no control input; not pinned. Allow 1018 // dominating control. 1019 if (n->in(0)) { 1020 Node *dom = idom(n_blk); 1021 if (dom_lca(n->in(0), dom) != n->in(0)) { 1022 return n; 1023 } 1024 } 1025 // Policy: when is it profitable. You must get more wins than 1026 // policy before it is considered profitable. Policy is usually 0, 1027 // so 1 win is considered profitable. Big merges will require big 1028 // cloning, so get a larger policy. 1029 int policy = n_blk->req() >> 2; 1030 1031 // If the loop is a candidate for range check elimination, 1032 // delay splitting through it's phi until a later loop optimization 1033 if (n_blk->is_CountedLoop()) { 1034 IdealLoopTree *lp = get_loop(n_blk); 1035 if (lp && lp->_rce_candidate) { 1036 return n; 1037 } 1038 } 1039 1040 if (must_throttle_split_if()) return n; 1041 1042 // Split 'n' through the merge point if it is profitable 1043 Node *phi = split_thru_phi( n, n_blk, policy ); 1044 if (!phi) return n; 1045 1046 // Found a Phi to split thru! 1047 // Replace 'n' with the new phi 1048 _igvn.replace_node( n, phi ); 1049 // Moved a load around the loop, 'en-registering' something. 1050 if (n_blk->is_Loop() && n->is_Load() && 1051 !phi->in(LoopNode::LoopBackControl)->is_Load()) 1052 C->set_major_progress(); 1053 1054 return phi; 1055 } 1056 1057 static bool merge_point_too_heavy(Compile* C, Node* region) { 1058 // Bail out if the region and its phis have too many users. 1059 int weight = 0; 1060 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1061 weight += region->fast_out(i)->outcnt(); 1062 } 1063 int nodes_left = C->max_node_limit() - C->live_nodes(); 1064 if (weight * 8 > nodes_left) { 1065 if (PrintOpto) { 1066 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight); 1067 } 1068 return true; 1069 } else { 1070 return false; 1071 } 1072 } 1073 1074 static bool merge_point_safe(Node* region) { 1075 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode 1076 // having a PhiNode input. This sidesteps the dangerous case where the split 1077 // ConvI2LNode may become TOP if the input Value() does not 1078 // overlap the ConvI2L range, leaving a node which may not dominate its 1079 // uses. 1080 // A better fix for this problem can be found in the BugTraq entry, but 1081 // expediency for Mantis demands this hack. 1082 // 6855164: If the merge point has a FastLockNode with a PhiNode input, we stop 1083 // split_if_with_blocks from splitting a block because we could not move around 1084 // the FastLockNode. 1085 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1086 Node* n = region->fast_out(i); 1087 if (n->is_Phi()) { 1088 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1089 Node* m = n->fast_out(j); 1090 if (m->is_FastLock()) 1091 return false; 1092 #ifdef _LP64 1093 if (m->Opcode() == Op_ConvI2L) 1094 return false; 1095 if (m->is_CastII() && m->isa_CastII()->has_range_check()) { 1096 return false; 1097 } 1098 #endif 1099 } 1100 } 1101 } 1102 return true; 1103 } 1104 1105 1106 //------------------------------place_near_use--------------------------------- 1107 // Place some computation next to use but not inside inner loops. 1108 // For inner loop uses move it to the preheader area. 1109 Node *PhaseIdealLoop::place_near_use(Node *useblock) const { 1110 IdealLoopTree *u_loop = get_loop( useblock ); 1111 if (u_loop->_irreducible) { 1112 return useblock; 1113 } 1114 if (u_loop->_child) { 1115 if (useblock == u_loop->_head && u_loop->_head->is_OuterStripMinedLoop()) { 1116 return u_loop->_head->in(LoopNode::EntryControl); 1117 } 1118 return useblock; 1119 } 1120 return u_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl); 1121 } 1122 1123 1124 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) { 1125 if (!n->is_If() || n->is_CountedLoopEnd()) { 1126 return false; 1127 } 1128 if (!n->in(0)->is_Region()) { 1129 return false; 1130 } 1131 Node* region = n->in(0); 1132 Node* dom = idom(region); 1133 if (!dom->is_If() || dom->in(1) != n->in(1)) { 1134 return false; 1135 } 1136 IfNode* dom_if = dom->as_If(); 1137 Node* proj_true = dom_if->proj_out(1); 1138 Node* proj_false = dom_if->proj_out(0); 1139 1140 for (uint i = 1; i < region->req(); i++) { 1141 if (is_dominator(proj_true, region->in(i))) { 1142 continue; 1143 } 1144 if (is_dominator(proj_false, region->in(i))) { 1145 continue; 1146 } 1147 return false; 1148 } 1149 1150 return true; 1151 } 1152 1153 1154 bool PhaseIdealLoop::can_split_if(Node* n_ctrl) { 1155 if (must_throttle_split_if()) { 1156 return false; 1157 } 1158 1159 // Do not do 'split-if' if irreducible loops are present. 1160 if (_has_irreducible_loops) { 1161 return false; 1162 } 1163 1164 if (merge_point_too_heavy(C, n_ctrl)) { 1165 return false; 1166 } 1167 1168 // Do not do 'split-if' if some paths are dead. First do dead code 1169 // elimination and then see if its still profitable. 1170 for (uint i = 1; i < n_ctrl->req(); i++) { 1171 if (n_ctrl->in(i) == C->top()) { 1172 return false; 1173 } 1174 } 1175 1176 // If trying to do a 'Split-If' at the loop head, it is only 1177 // profitable if the cmp folds up on BOTH paths. Otherwise we 1178 // risk peeling a loop forever. 1179 1180 // CNC - Disabled for now. Requires careful handling of loop 1181 // body selection for the cloned code. Also, make sure we check 1182 // for any input path not being in the same loop as n_ctrl. For 1183 // irreducible loops we cannot check for 'n_ctrl->is_Loop()' 1184 // because the alternative loop entry points won't be converted 1185 // into LoopNodes. 1186 IdealLoopTree *n_loop = get_loop(n_ctrl); 1187 for (uint j = 1; j < n_ctrl->req(); j++) { 1188 if (get_loop(n_ctrl->in(j)) != n_loop) { 1189 return false; 1190 } 1191 } 1192 1193 // Check for safety of the merge point. 1194 if (!merge_point_safe(n_ctrl)) { 1195 return false; 1196 } 1197 1198 return true; 1199 } 1200 1201 // Detect if the node is the inner strip-mined loop 1202 // Return: NULL if it's not the case, or the exit of outer strip-mined loop 1203 static Node* is_inner_of_stripmined_loop(const Node* out) { 1204 Node* out_le = NULL; 1205 1206 if (out->is_CountedLoopEnd()) { 1207 const CountedLoopNode* loop = out->as_CountedLoopEnd()->loopnode(); 1208 1209 if (loop != NULL && loop->is_strip_mined()) { 1210 out_le = loop->in(LoopNode::EntryControl)->as_OuterStripMinedLoop()->outer_loop_exit(); 1211 } 1212 } 1213 1214 return out_le; 1215 } 1216 1217 bool PhaseIdealLoop::flatten_array_element_type_check(Node *n) { 1218 // If the CmpP is a subtype check for a value that has just been 1219 // loaded from an array, the subtype checks guarantees the value 1220 // can't be stored in a flattened array and the load of the value 1221 // happens with a flattened array check then: push the type check 1222 // through the phi of the flattened array check. This needs special 1223 // logic because the subtype check's input is not a phi but a 1224 // LoadKlass that must first be cloned through the phi. 1225 if (n->Opcode() != Op_CmpP) { 1226 return false; 1227 } 1228 1229 Node* klassptr = n->in(1); 1230 Node* klasscon = n->in(2); 1231 1232 if (klassptr->is_DecodeNarrowPtr()) { 1233 klassptr = klassptr->in(1); 1234 } 1235 1236 if (klassptr->Opcode() != Op_LoadKlass && klassptr->Opcode() != Op_LoadNKlass) { 1237 return false; 1238 } 1239 1240 if (!klasscon->is_Con()) { 1241 return false; 1242 } 1243 1244 Node* addr = klassptr->in(MemNode::Address); 1245 1246 if (!addr->is_AddP()) { 1247 return false; 1248 } 1249 1250 intptr_t offset; 1251 Node* obj = AddPNode::Ideal_base_and_offset(addr, &_igvn, offset); 1252 assert(obj != NULL && addr->in(AddPNode::Base) == addr->in(AddPNode::Address), "malformed AddP?"); 1253 if (obj->Opcode() == Op_CastPP) { 1254 obj = obj->in(1); 1255 } 1256 1257 if (!obj->is_Phi()) { 1258 return false; 1259 } 1260 1261 Node* region = obj->in(0); 1262 1263 Node* phi = PhiNode::make_blank(region, n->in(1)); 1264 for (uint i = 1; i < region->req(); i++) { 1265 Node* in = obj->in(i); 1266 Node* ctrl = get_ctrl(in); 1267 if (addr->in(AddPNode::Base) != obj) { 1268 Node* cast = addr->in(AddPNode::Base); 1269 assert(cast->Opcode() == Op_CastPP && cast->in(0) != NULL, "inconsistent subgraph"); 1270 Node* cast_clone = cast->clone(); 1271 cast_clone->set_req(0, region->in(i)); 1272 cast_clone->set_req(1, in); 1273 register_new_node(cast_clone, region->in(i)); 1274 _igvn.set_type(cast_clone, cast_clone->Value(&_igvn)); 1275 in = cast_clone; 1276 } 1277 Node* addr_clone = addr->clone(); 1278 addr_clone->set_req(AddPNode::Base, in); 1279 addr_clone->set_req(AddPNode::Address, in); 1280 register_new_node(addr_clone, ctrl); 1281 _igvn.set_type(addr_clone, addr_clone->Value(&_igvn)); 1282 Node* klassptr_clone = klassptr->clone(); 1283 klassptr_clone->set_req(2, addr_clone); 1284 register_new_node(klassptr_clone, ctrl); 1285 _igvn.set_type(klassptr_clone, klassptr_clone->Value(&_igvn)); 1286 if (klassptr != n->in(1)) { 1287 Node* decode = n->in(1); 1288 assert(decode->is_DecodeNarrowPtr(), "inconcistent subgraph"); 1289 Node* decode_clone = decode->clone(); 1290 decode_clone->set_req(1, klassptr_clone); 1291 register_new_node(decode_clone, ctrl); 1292 _igvn.set_type(decode_clone, decode_clone->Value(&_igvn)); 1293 klassptr_clone = decode_clone; 1294 } 1295 phi->set_req(i, klassptr_clone); 1296 } 1297 register_new_node(phi, region); 1298 Node* orig = n->in(1); 1299 _igvn.replace_input_of(n, 1, phi); 1300 split_if_with_blocks_post(n); 1301 if (n->outcnt() != 0) { 1302 _igvn.replace_input_of(n, 1, orig); 1303 _igvn._worklist.push(phi); 1304 } 1305 return true; 1306 } 1307 1308 //------------------------------split_if_with_blocks_post---------------------- 1309 // Do the real work in a non-recursive function. CFG hackery wants to be 1310 // in the post-order, so it can dirty the I-DOM info and not use the dirtied 1311 // info. 1312 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) { 1313 1314 if (flatten_array_element_type_check(n)) { 1315 return; 1316 } 1317 1318 // Cloning Cmp through Phi's involves the split-if transform. 1319 // FastLock is not used by an If 1320 if (n->is_Cmp() && !n->is_FastLock()) { 1321 Node *n_ctrl = get_ctrl(n); 1322 // Determine if the Node has inputs from some local Phi. 1323 // Returns the block to clone thru. 1324 Node *n_blk = has_local_phi_input(n); 1325 if (n_blk != n_ctrl) { 1326 return; 1327 } 1328 1329 if (!can_split_if(n_ctrl)) { 1330 return; 1331 } 1332 1333 if (n->outcnt() != 1) { 1334 return; // Multiple bool's from 1 compare? 1335 } 1336 Node *bol = n->unique_out(); 1337 assert(bol->is_Bool(), "expect a bool here"); 1338 if (bol->outcnt() != 1) { 1339 return;// Multiple branches from 1 compare? 1340 } 1341 Node *iff = bol->unique_out(); 1342 1343 // Check some safety conditions 1344 if (iff->is_If()) { // Classic split-if? 1345 if (iff->in(0) != n_ctrl) { 1346 return; // Compare must be in same blk as if 1347 } 1348 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE 1349 // Can't split CMove with different control edge. 1350 if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) { 1351 return; 1352 } 1353 if (get_ctrl(iff->in(2)) == n_ctrl || 1354 get_ctrl(iff->in(3)) == n_ctrl) { 1355 return; // Inputs not yet split-up 1356 } 1357 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) { 1358 return; // Loop-invar test gates loop-varying CMOVE 1359 } 1360 } else { 1361 return; // some other kind of node, such as an Allocate 1362 } 1363 1364 // When is split-if profitable? Every 'win' on means some control flow 1365 // goes dead, so it's almost always a win. 1366 int policy = 0; 1367 // Split compare 'n' through the merge point if it is profitable 1368 Node *phi = split_thru_phi( n, n_ctrl, policy); 1369 if (!phi) { 1370 return; 1371 } 1372 1373 // Found a Phi to split thru! 1374 // Replace 'n' with the new phi 1375 _igvn.replace_node(n, phi); 1376 1377 // Now split the bool up thru the phi 1378 Node *bolphi = split_thru_phi(bol, n_ctrl, -1); 1379 guarantee(bolphi != NULL, "null boolean phi node"); 1380 1381 _igvn.replace_node(bol, bolphi); 1382 assert(iff->in(1) == bolphi, ""); 1383 1384 if (bolphi->Value(&_igvn)->singleton()) { 1385 return; 1386 } 1387 1388 // Conditional-move? Must split up now 1389 if (!iff->is_If()) { 1390 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1); 1391 _igvn.replace_node(iff, cmovphi); 1392 return; 1393 } 1394 1395 // Now split the IF 1396 do_split_if(iff); 1397 return; 1398 } 1399 1400 // Two identical ifs back to back can be merged 1401 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) { 1402 Node *n_ctrl = n->in(0); 1403 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1)); 1404 IfNode* dom_if = idom(n_ctrl)->as_If(); 1405 Node* proj_true = dom_if->proj_out(1); 1406 Node* proj_false = dom_if->proj_out(0); 1407 Node* con_true = _igvn.makecon(TypeInt::ONE); 1408 Node* con_false = _igvn.makecon(TypeInt::ZERO); 1409 1410 for (uint i = 1; i < n_ctrl->req(); i++) { 1411 if (is_dominator(proj_true, n_ctrl->in(i))) { 1412 bolphi->init_req(i, con_true); 1413 } else { 1414 assert(is_dominator(proj_false, n_ctrl->in(i)), "bad if"); 1415 bolphi->init_req(i, con_false); 1416 } 1417 } 1418 register_new_node(bolphi, n_ctrl); 1419 _igvn.replace_input_of(n, 1, bolphi); 1420 1421 // Now split the IF 1422 do_split_if(n); 1423 return; 1424 } 1425 1426 // Check for an IF ready to split; one that has its 1427 // condition codes input coming from a Phi at the block start. 1428 int n_op = n->Opcode(); 1429 1430 // Check for an IF being dominated by another IF same test 1431 if (n_op == Op_If || 1432 n_op == Op_RangeCheck) { 1433 Node *bol = n->in(1); 1434 uint max = bol->outcnt(); 1435 // Check for same test used more than once? 1436 if (max > 1 && bol->is_Bool()) { 1437 // Search up IDOMs to see if this IF is dominated. 1438 Node *cutoff = get_ctrl(bol); 1439 1440 // Now search up IDOMs till cutoff, looking for a dominating test 1441 Node *prevdom = n; 1442 Node *dom = idom(prevdom); 1443 while (dom != cutoff) { 1444 if (dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom) { 1445 // It's invalid to move control dependent data nodes in the inner 1446 // strip-mined loop, because: 1447 // 1) break validation of LoopNode::verify_strip_mined() 1448 // 2) move code with side-effect in strip-mined loop 1449 // Move to the exit of outer strip-mined loop in that case. 1450 Node* out_le = is_inner_of_stripmined_loop(dom); 1451 if (out_le != NULL) { 1452 prevdom = out_le; 1453 } 1454 // Replace the dominated test with an obvious true or false. 1455 // Place it on the IGVN worklist for later cleanup. 1456 C->set_major_progress(); 1457 dominated_by(prevdom, n, false, true); 1458 #ifndef PRODUCT 1459 if( VerifyLoopOptimizations ) verify(); 1460 #endif 1461 return; 1462 } 1463 prevdom = dom; 1464 dom = idom(prevdom); 1465 } 1466 } 1467 } 1468 1469 // See if a shared loop-varying computation has no loop-varying uses. 1470 // Happens if something is only used for JVM state in uncommon trap exits, 1471 // like various versions of induction variable+offset. Clone the 1472 // computation per usage to allow it to sink out of the loop. 1473 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about) 1474 Node *n_ctrl = get_ctrl(n); 1475 IdealLoopTree *n_loop = get_loop(n_ctrl); 1476 if( n_loop != _ltree_root ) { 1477 DUIterator_Fast imax, i = n->fast_outs(imax); 1478 for (; i < imax; i++) { 1479 Node* u = n->fast_out(i); 1480 if( !has_ctrl(u) ) break; // Found control user 1481 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 1482 if( u_loop == n_loop ) break; // Found loop-varying use 1483 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop 1484 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003 1485 } 1486 bool did_break = (i < imax); // Did we break out of the previous loop? 1487 if (!did_break && n->outcnt() > 1) { // All uses in outer loops! 1488 Node *late_load_ctrl = NULL; 1489 if (n->is_Load()) { 1490 // If n is a load, get and save the result from get_late_ctrl(), 1491 // to be later used in calculating the control for n's clones. 1492 clear_dom_lca_tags(); 1493 late_load_ctrl = get_late_ctrl(n, n_ctrl); 1494 } 1495 // If n is a load, and the late control is the same as the current 1496 // control, then the cloning of n is a pointless exercise, because 1497 // GVN will ensure that we end up where we started. 1498 if (!n->is_Load() || late_load_ctrl != n_ctrl) { 1499 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1500 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) { 1501 Node *u = n->last_out(j); // Clone private computation per use 1502 _igvn.rehash_node_delayed(u); 1503 Node *x = n->clone(); // Clone computation 1504 Node *x_ctrl = NULL; 1505 if( u->is_Phi() ) { 1506 // Replace all uses of normal nodes. Replace Phi uses 1507 // individually, so the separate Nodes can sink down 1508 // different paths. 1509 uint k = 1; 1510 while( u->in(k) != n ) k++; 1511 u->set_req( k, x ); 1512 // x goes next to Phi input path 1513 x_ctrl = u->in(0)->in(k); 1514 --j; 1515 } else { // Normal use 1516 // Replace all uses 1517 for( uint k = 0; k < u->req(); k++ ) { 1518 if( u->in(k) == n ) { 1519 u->set_req( k, x ); 1520 --j; 1521 } 1522 } 1523 x_ctrl = get_ctrl(u); 1524 } 1525 1526 // Find control for 'x' next to use but not inside inner loops. 1527 // For inner loop uses get the preheader area. 1528 x_ctrl = place_near_use(x_ctrl); 1529 1530 if (bs->sink_node(this, n, x, x_ctrl, n_ctrl)) { 1531 continue; 1532 } 1533 1534 if (n->is_Load()) { 1535 // For loads, add a control edge to a CFG node outside of the loop 1536 // to force them to not combine and return back inside the loop 1537 // during GVN optimization (4641526). 1538 // 1539 // Because we are setting the actual control input, factor in 1540 // the result from get_late_ctrl() so we respect any 1541 // anti-dependences. (6233005). 1542 x_ctrl = dom_lca(late_load_ctrl, x_ctrl); 1543 1544 // Don't allow the control input to be a CFG splitting node. 1545 // Such nodes should only have ProjNodes as outs, e.g. IfNode 1546 // should only have IfTrueNode and IfFalseNode (4985384). 1547 x_ctrl = find_non_split_ctrl(x_ctrl); 1548 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone"); 1549 1550 x->set_req(0, x_ctrl); 1551 } 1552 register_new_node(x, x_ctrl); 1553 1554 // Some institutional knowledge is needed here: 'x' is 1555 // yanked because if the optimizer runs GVN on it all the 1556 // cloned x's will common up and undo this optimization and 1557 // be forced back in the loop. 1558 // I tried setting control edges on the x's to force them to 1559 // not combine, but the matching gets worried when it tries 1560 // to fold a StoreP and an AddP together (as part of an 1561 // address expression) and the AddP and StoreP have 1562 // different controls. 1563 if (!x->is_Load() && !x->is_DecodeNarrowPtr()) _igvn._worklist.yank(x); 1564 } 1565 _igvn.remove_dead_node(n); 1566 } 1567 } 1568 } 1569 } 1570 1571 try_move_store_after_loop(n); 1572 1573 // Remove multiple allocations of the same value type 1574 if (n->is_ValueType()) { 1575 n->as_ValueType()->remove_redundant_allocations(&_igvn, this); 1576 return; // n is now dead 1577 } 1578 1579 // Check for Opaque2's who's loop has disappeared - who's input is in the 1580 // same loop nest as their output. Remove 'em, they are no longer useful. 1581 if( n_op == Op_Opaque2 && 1582 n->in(1) != NULL && 1583 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) { 1584 _igvn.replace_node( n, n->in(1) ); 1585 } 1586 } 1587 1588 //------------------------------split_if_with_blocks--------------------------- 1589 // Check for aggressive application of 'split-if' optimization, 1590 // using basic block level info. 1591 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack) { 1592 Node* root = C->root(); 1593 visited.set(root->_idx); // first, mark root as visited 1594 // Do pre-visit work for root 1595 Node* n = split_if_with_blocks_pre(root); 1596 uint cnt = n->outcnt(); 1597 uint i = 0; 1598 1599 while (true) { 1600 // Visit all children 1601 if (i < cnt) { 1602 Node* use = n->raw_out(i); 1603 ++i; 1604 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) { 1605 // Now do pre-visit work for this use 1606 use = split_if_with_blocks_pre(use); 1607 nstack.push(n, i); // Save parent and next use's index. 1608 n = use; // Process all children of current use. 1609 cnt = use->outcnt(); 1610 i = 0; 1611 } 1612 } 1613 else { 1614 // All of n's children have been processed, complete post-processing. 1615 if (cnt != 0 && !n->is_Con()) { 1616 assert(has_node(n), "no dead nodes"); 1617 split_if_with_blocks_post(n); 1618 } 1619 if (must_throttle_split_if()) { 1620 nstack.clear(); 1621 } 1622 if (nstack.is_empty()) { 1623 // Finished all nodes on stack. 1624 break; 1625 } 1626 // Get saved parent node and next use's index. Visit the rest of uses. 1627 n = nstack.node(); 1628 cnt = n->outcnt(); 1629 i = nstack.index(); 1630 nstack.pop(); 1631 } 1632 } 1633 } 1634 1635 1636 //============================================================================= 1637 // 1638 // C L O N E A L O O P B O D Y 1639 // 1640 1641 //------------------------------clone_iff-------------------------------------- 1642 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1643 // "Nearly" because all Nodes have been cloned from the original in the loop, 1644 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1645 // through the Phi recursively, and return a Bool. 1646 Node* PhaseIdealLoop::clone_iff(PhiNode *phi, IdealLoopTree *loop) { 1647 1648 // Convert this Phi into a Phi merging Bools 1649 uint i; 1650 for (i = 1; i < phi->req(); i++) { 1651 Node *b = phi->in(i); 1652 if (b->is_Phi()) { 1653 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi(), loop)); 1654 } else { 1655 assert(b->is_Bool() || b->Opcode() == Op_Opaque4, ""); 1656 } 1657 } 1658 1659 Node* n = phi->in(1); 1660 Node* sample_opaque = NULL; 1661 Node *sample_bool = NULL; 1662 if (n->Opcode() == Op_Opaque4) { 1663 sample_opaque = n; 1664 sample_bool = n->in(1); 1665 assert(sample_bool->is_Bool(), "wrong type"); 1666 } else { 1667 sample_bool = n; 1668 } 1669 Node *sample_cmp = sample_bool->in(1); 1670 1671 // Make Phis to merge the Cmp's inputs. 1672 PhiNode *phi1 = new PhiNode(phi->in(0), Type::TOP); 1673 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP); 1674 for (i = 1; i < phi->req(); i++) { 1675 Node *n1 = sample_opaque == NULL ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1); 1676 Node *n2 = sample_opaque == NULL ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2); 1677 phi1->set_req(i, n1); 1678 phi2->set_req(i, n2); 1679 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 1680 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 1681 } 1682 // See if these Phis have been made before. 1683 // Register with optimizer 1684 Node *hit1 = _igvn.hash_find_insert(phi1); 1685 if (hit1) { // Hit, toss just made Phi 1686 _igvn.remove_dead_node(phi1); // Remove new phi 1687 assert(hit1->is_Phi(), "" ); 1688 phi1 = (PhiNode*)hit1; // Use existing phi 1689 } else { // Miss 1690 _igvn.register_new_node_with_optimizer(phi1); 1691 } 1692 Node *hit2 = _igvn.hash_find_insert(phi2); 1693 if (hit2) { // Hit, toss just made Phi 1694 _igvn.remove_dead_node(phi2); // Remove new phi 1695 assert(hit2->is_Phi(), "" ); 1696 phi2 = (PhiNode*)hit2; // Use existing phi 1697 } else { // Miss 1698 _igvn.register_new_node_with_optimizer(phi2); 1699 } 1700 // Register Phis with loop/block info 1701 set_ctrl(phi1, phi->in(0)); 1702 set_ctrl(phi2, phi->in(0)); 1703 // Make a new Cmp 1704 Node *cmp = sample_cmp->clone(); 1705 cmp->set_req(1, phi1); 1706 cmp->set_req(2, phi2); 1707 _igvn.register_new_node_with_optimizer(cmp); 1708 set_ctrl(cmp, phi->in(0)); 1709 1710 // Make a new Bool 1711 Node *b = sample_bool->clone(); 1712 b->set_req(1,cmp); 1713 _igvn.register_new_node_with_optimizer(b); 1714 set_ctrl(b, phi->in(0)); 1715 1716 if (sample_opaque != NULL) { 1717 Node* opaque = sample_opaque->clone(); 1718 opaque->set_req(1, b); 1719 _igvn.register_new_node_with_optimizer(opaque); 1720 set_ctrl(opaque, phi->in(0)); 1721 return opaque; 1722 } 1723 1724 assert(b->is_Bool(), ""); 1725 return b; 1726 } 1727 1728 //------------------------------clone_bool------------------------------------- 1729 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1730 // "Nearly" because all Nodes have been cloned from the original in the loop, 1731 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1732 // through the Phi recursively, and return a Bool. 1733 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) { 1734 uint i; 1735 // Convert this Phi into a Phi merging Bools 1736 for( i = 1; i < phi->req(); i++ ) { 1737 Node *b = phi->in(i); 1738 if( b->is_Phi() ) { 1739 _igvn.replace_input_of(phi, i, clone_bool( b->as_Phi(), loop )); 1740 } else { 1741 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" ); 1742 } 1743 } 1744 1745 Node *sample_cmp = phi->in(1); 1746 1747 // Make Phis to merge the Cmp's inputs. 1748 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP ); 1749 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP ); 1750 for( uint j = 1; j < phi->req(); j++ ) { 1751 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP 1752 Node *n1, *n2; 1753 if( cmp_top->is_Cmp() ) { 1754 n1 = cmp_top->in(1); 1755 n2 = cmp_top->in(2); 1756 } else { 1757 n1 = n2 = cmp_top; 1758 } 1759 phi1->set_req( j, n1 ); 1760 phi2->set_req( j, n2 ); 1761 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 1762 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 1763 } 1764 1765 // See if these Phis have been made before. 1766 // Register with optimizer 1767 Node *hit1 = _igvn.hash_find_insert(phi1); 1768 if( hit1 ) { // Hit, toss just made Phi 1769 _igvn.remove_dead_node(phi1); // Remove new phi 1770 assert( hit1->is_Phi(), "" ); 1771 phi1 = (PhiNode*)hit1; // Use existing phi 1772 } else { // Miss 1773 _igvn.register_new_node_with_optimizer(phi1); 1774 } 1775 Node *hit2 = _igvn.hash_find_insert(phi2); 1776 if( hit2 ) { // Hit, toss just made Phi 1777 _igvn.remove_dead_node(phi2); // Remove new phi 1778 assert( hit2->is_Phi(), "" ); 1779 phi2 = (PhiNode*)hit2; // Use existing phi 1780 } else { // Miss 1781 _igvn.register_new_node_with_optimizer(phi2); 1782 } 1783 // Register Phis with loop/block info 1784 set_ctrl(phi1, phi->in(0)); 1785 set_ctrl(phi2, phi->in(0)); 1786 // Make a new Cmp 1787 Node *cmp = sample_cmp->clone(); 1788 cmp->set_req( 1, phi1 ); 1789 cmp->set_req( 2, phi2 ); 1790 _igvn.register_new_node_with_optimizer(cmp); 1791 set_ctrl(cmp, phi->in(0)); 1792 1793 assert( cmp->is_Cmp(), "" ); 1794 return (CmpNode*)cmp; 1795 } 1796 1797 //------------------------------sink_use--------------------------------------- 1798 // If 'use' was in the loop-exit block, it now needs to be sunk 1799 // below the post-loop merge point. 1800 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) { 1801 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) { 1802 set_ctrl(use, post_loop); 1803 for (DUIterator j = use->outs(); use->has_out(j); j++) 1804 sink_use(use->out(j), post_loop); 1805 } 1806 } 1807 1808 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new, 1809 IdealLoopTree* loop, IdealLoopTree* outer_loop, 1810 Node_List*& split_if_set, Node_List*& split_bool_set, 1811 Node_List*& split_cex_set, Node_List& worklist, 1812 uint new_counter, CloneLoopMode mode) { 1813 Node* nnn = old_new[old->_idx]; 1814 // Copy uses to a worklist, so I can munge the def-use info 1815 // with impunity. 1816 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1817 worklist.push(old->fast_out(j)); 1818 1819 while( worklist.size() ) { 1820 Node *use = worklist.pop(); 1821 if (!has_node(use)) continue; // Ignore dead nodes 1822 if (use->in(0) == C->top()) continue; 1823 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1824 // Check for data-use outside of loop - at least one of OLD or USE 1825 // must not be a CFG node. 1826 #ifdef ASSERT 1827 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == NULL) { 1828 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint(); 1829 assert(mode == ControlAroundStripMined && use == sfpt, "missed a node"); 1830 } 1831 #endif 1832 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) { 1833 1834 // If the Data use is an IF, that means we have an IF outside of the 1835 // loop that is switching on a condition that is set inside of the 1836 // loop. Happens if people set a loop-exit flag; then test the flag 1837 // in the loop to break the loop, then test is again outside of the 1838 // loop to determine which way the loop exited. 1839 // Loop predicate If node connects to Bool node through Opaque1 node. 1840 if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use) || use->Opcode() == Op_Opaque4) { 1841 // Since this code is highly unlikely, we lazily build the worklist 1842 // of such Nodes to go split. 1843 if (!split_if_set) { 1844 ResourceArea *area = Thread::current()->resource_area(); 1845 split_if_set = new Node_List(area); 1846 } 1847 split_if_set->push(use); 1848 } 1849 if (use->is_Bool()) { 1850 if (!split_bool_set) { 1851 ResourceArea *area = Thread::current()->resource_area(); 1852 split_bool_set = new Node_List(area); 1853 } 1854 split_bool_set->push(use); 1855 } 1856 if (use->Opcode() == Op_CreateEx) { 1857 if (!split_cex_set) { 1858 ResourceArea *area = Thread::current()->resource_area(); 1859 split_cex_set = new Node_List(area); 1860 } 1861 split_cex_set->push(use); 1862 } 1863 1864 1865 // Get "block" use is in 1866 uint idx = 0; 1867 while( use->in(idx) != old ) idx++; 1868 Node *prev = use->is_CFG() ? use : get_ctrl(use); 1869 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" ); 1870 Node *cfg = prev->_idx >= new_counter 1871 ? prev->in(2) 1872 : idom(prev); 1873 if( use->is_Phi() ) // Phi use is in prior block 1874 cfg = prev->in(idx); // NOT in block of Phi itself 1875 if (cfg->is_top()) { // Use is dead? 1876 _igvn.replace_input_of(use, idx, C->top()); 1877 continue; 1878 } 1879 1880 // If use is referenced through control edge... (idx == 0) 1881 if (mode == IgnoreStripMined && idx == 0) { 1882 LoopNode *head = loop->_head->as_Loop(); 1883 if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) { 1884 // That node is outside the inner loop, leave it outside the 1885 // outer loop as well to not confuse verification code. 1886 assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop"); 1887 _igvn.replace_input_of(use, 0, head->outer_loop_exit()); 1888 continue; 1889 } 1890 } 1891 1892 while(!outer_loop->is_member(get_loop(cfg))) { 1893 prev = cfg; 1894 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg); 1895 } 1896 // If the use occurs after merging several exits from the loop, then 1897 // old value must have dominated all those exits. Since the same old 1898 // value was used on all those exits we did not need a Phi at this 1899 // merge point. NOW we do need a Phi here. Each loop exit value 1900 // is now merged with the peeled body exit; each exit gets its own 1901 // private Phi and those Phis need to be merged here. 1902 Node *phi; 1903 if( prev->is_Region() ) { 1904 if( idx == 0 ) { // Updating control edge? 1905 phi = prev; // Just use existing control 1906 } else { // Else need a new Phi 1907 phi = PhiNode::make( prev, old ); 1908 // Now recursively fix up the new uses of old! 1909 for( uint i = 1; i < prev->req(); i++ ) { 1910 worklist.push(phi); // Onto worklist once for each 'old' input 1911 } 1912 } 1913 } else { 1914 // Get new RegionNode merging old and new loop exits 1915 prev = old_new[prev->_idx]; 1916 assert( prev, "just made this in step 7" ); 1917 if( idx == 0) { // Updating control edge? 1918 phi = prev; // Just use existing control 1919 } else { // Else need a new Phi 1920 // Make a new Phi merging data values properly 1921 phi = PhiNode::make( prev, old ); 1922 phi->set_req( 1, nnn ); 1923 } 1924 } 1925 // If inserting a new Phi, check for prior hits 1926 if( idx != 0 ) { 1927 Node *hit = _igvn.hash_find_insert(phi); 1928 if( hit == NULL ) { 1929 _igvn.register_new_node_with_optimizer(phi); // Register new phi 1930 } else { // or 1931 // Remove the new phi from the graph and use the hit 1932 _igvn.remove_dead_node(phi); 1933 phi = hit; // Use existing phi 1934 } 1935 set_ctrl(phi, prev); 1936 } 1937 // Make 'use' use the Phi instead of the old loop body exit value 1938 _igvn.replace_input_of(use, idx, phi); 1939 if( use->_idx >= new_counter ) { // If updating new phis 1940 // Not needed for correctness, but prevents a weak assert 1941 // in AddPNode from tripping (when we end up with different 1942 // base & derived Phis that will become the same after 1943 // IGVN does CSE). 1944 Node *hit = _igvn.hash_find_insert(use); 1945 if( hit ) // Go ahead and re-hash for hits. 1946 _igvn.replace_node( use, hit ); 1947 } 1948 1949 // If 'use' was in the loop-exit block, it now needs to be sunk 1950 // below the post-loop merge point. 1951 sink_use( use, prev ); 1952 } 1953 } 1954 } 1955 1956 static void clone_outer_loop_helper(Node* n, const IdealLoopTree *loop, const IdealLoopTree* outer_loop, 1957 const Node_List &old_new, Unique_Node_List& wq, PhaseIdealLoop* phase, 1958 bool check_old_new) { 1959 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1960 Node* u = n->fast_out(j); 1961 assert(check_old_new || old_new[u->_idx] == NULL, "shouldn't have been cloned"); 1962 if (!u->is_CFG() && (!check_old_new || old_new[u->_idx] == NULL)) { 1963 Node* c = phase->get_ctrl(u); 1964 IdealLoopTree* u_loop = phase->get_loop(c); 1965 assert(!loop->is_member(u_loop), "can be in outer loop or out of both loops only"); 1966 if (outer_loop->is_member(u_loop)) { 1967 wq.push(u); 1968 } 1969 } 1970 } 1971 } 1972 1973 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop, 1974 IdealLoopTree* outer_loop, int dd, Node_List &old_new, 1975 Node_List& extra_data_nodes) { 1976 if (head->is_strip_mined() && mode != IgnoreStripMined) { 1977 CountedLoopNode* cl = head->as_CountedLoop(); 1978 Node* l = cl->outer_loop(); 1979 Node* tail = cl->outer_loop_tail(); 1980 IfNode* le = cl->outer_loop_end(); 1981 Node* sfpt = cl->outer_safepoint(); 1982 CountedLoopEndNode* cle = cl->loopexit(); 1983 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop(); 1984 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null(); 1985 Node* cle_out = cle->proj_out(false); 1986 1987 Node* new_sfpt = NULL; 1988 Node* new_cle_out = cle_out->clone(); 1989 old_new.map(cle_out->_idx, new_cle_out); 1990 if (mode == CloneIncludesStripMined) { 1991 // clone outer loop body 1992 Node* new_l = l->clone(); 1993 Node* new_tail = tail->clone(); 1994 IfNode* new_le = le->clone()->as_If(); 1995 new_sfpt = sfpt->clone(); 1996 1997 set_loop(new_l, outer_loop->_parent); 1998 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd); 1999 set_loop(new_cle_out, outer_loop->_parent); 2000 set_idom(new_cle_out, new_cle, dd); 2001 set_loop(new_sfpt, outer_loop->_parent); 2002 set_idom(new_sfpt, new_cle_out, dd); 2003 set_loop(new_le, outer_loop->_parent); 2004 set_idom(new_le, new_sfpt, dd); 2005 set_loop(new_tail, outer_loop->_parent); 2006 set_idom(new_tail, new_le, dd); 2007 set_idom(new_cl, new_l, dd); 2008 2009 old_new.map(l->_idx, new_l); 2010 old_new.map(tail->_idx, new_tail); 2011 old_new.map(le->_idx, new_le); 2012 old_new.map(sfpt->_idx, new_sfpt); 2013 2014 new_l->set_req(LoopNode::LoopBackControl, new_tail); 2015 new_l->set_req(0, new_l); 2016 new_tail->set_req(0, new_le); 2017 new_le->set_req(0, new_sfpt); 2018 new_sfpt->set_req(0, new_cle_out); 2019 new_cle_out->set_req(0, new_cle); 2020 new_cl->set_req(LoopNode::EntryControl, new_l); 2021 2022 _igvn.register_new_node_with_optimizer(new_l); 2023 _igvn.register_new_node_with_optimizer(new_tail); 2024 _igvn.register_new_node_with_optimizer(new_le); 2025 } else { 2026 Node *newhead = old_new[loop->_head->_idx]; 2027 newhead->as_Loop()->clear_strip_mined(); 2028 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl)); 2029 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 2030 } 2031 // Look at data node that were assigned a control in the outer 2032 // loop: they are kept in the outer loop by the safepoint so start 2033 // from the safepoint node's inputs. 2034 IdealLoopTree* outer_loop = get_loop(l); 2035 Node_Stack stack(2); 2036 stack.push(sfpt, 1); 2037 uint new_counter = C->unique(); 2038 while (stack.size() > 0) { 2039 Node* n = stack.node(); 2040 uint i = stack.index(); 2041 while (i < n->req() && 2042 (n->in(i) == NULL || 2043 !has_ctrl(n->in(i)) || 2044 get_loop(get_ctrl(n->in(i))) != outer_loop || 2045 (old_new[n->in(i)->_idx] != NULL && old_new[n->in(i)->_idx]->_idx >= new_counter))) { 2046 i++; 2047 } 2048 if (i < n->req()) { 2049 stack.set_index(i+1); 2050 stack.push(n->in(i), 0); 2051 } else { 2052 assert(old_new[n->_idx] == NULL || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet"); 2053 Node* m = n == sfpt ? new_sfpt : n->clone(); 2054 if (m != NULL) { 2055 for (uint i = 0; i < n->req(); i++) { 2056 if (m->in(i) != NULL && old_new[m->in(i)->_idx] != NULL) { 2057 m->set_req(i, old_new[m->in(i)->_idx]); 2058 } 2059 } 2060 } else { 2061 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?"); 2062 } 2063 if (n != sfpt) { 2064 extra_data_nodes.push(n); 2065 _igvn.register_new_node_with_optimizer(m); 2066 assert(get_ctrl(n) == cle_out, "what other control?"); 2067 set_ctrl(m, new_cle_out); 2068 old_new.map(n->_idx, m); 2069 } 2070 stack.pop(); 2071 } 2072 } 2073 if (mode == CloneIncludesStripMined) { 2074 _igvn.register_new_node_with_optimizer(new_sfpt); 2075 _igvn.register_new_node_with_optimizer(new_cle_out); 2076 } 2077 // Some other transformation may have pessimistically assign some 2078 // data nodes to the outer loop. Set their control so they are out 2079 // of the outer loop. 2080 ResourceMark rm; 2081 Unique_Node_List wq; 2082 for (uint i = 0; i < extra_data_nodes.size(); i++) { 2083 Node* old = extra_data_nodes.at(i); 2084 clone_outer_loop_helper(old, loop, outer_loop, old_new, wq, this, true); 2085 } 2086 Node* new_ctrl = cl->outer_loop_exit(); 2087 assert(get_loop(new_ctrl) != outer_loop, "must be out of the loop nest"); 2088 for (uint i = 0; i < wq.size(); i++) { 2089 Node* n = wq.at(i); 2090 set_ctrl(n, new_ctrl); 2091 clone_outer_loop_helper(n, loop, outer_loop, old_new, wq, this, false); 2092 } 2093 } else { 2094 Node *newhead = old_new[loop->_head->_idx]; 2095 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 2096 } 2097 } 2098 2099 //------------------------------clone_loop------------------------------------- 2100 // 2101 // C L O N E A L O O P B O D Y 2102 // 2103 // This is the basic building block of the loop optimizations. It clones an 2104 // entire loop body. It makes an old_new loop body mapping; with this mapping 2105 // you can find the new-loop equivalent to an old-loop node. All new-loop 2106 // nodes are exactly equal to their old-loop counterparts, all edges are the 2107 // same. All exits from the old-loop now have a RegionNode that merges the 2108 // equivalent new-loop path. This is true even for the normal "loop-exit" 2109 // condition. All uses of loop-invariant old-loop values now come from (one 2110 // or more) Phis that merge their new-loop equivalents. 2111 // 2112 // This operation leaves the graph in an illegal state: there are two valid 2113 // control edges coming from the loop pre-header to both loop bodies. I'll 2114 // definitely have to hack the graph after running this transform. 2115 // 2116 // From this building block I will further edit edges to perform loop peeling 2117 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc. 2118 // 2119 // Parameter side_by_size_idom: 2120 // When side_by_size_idom is NULL, the dominator tree is constructed for 2121 // the clone loop to dominate the original. Used in construction of 2122 // pre-main-post loop sequence. 2123 // When nonnull, the clone and original are side-by-side, both are 2124 // dominated by the side_by_side_idom node. Used in construction of 2125 // unswitched loops. 2126 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd, 2127 CloneLoopMode mode, Node* side_by_side_idom) { 2128 2129 LoopNode* head = loop->_head->as_Loop(); 2130 head->verify_strip_mined(1); 2131 2132 if (C->do_vector_loop() && PrintOpto) { 2133 const char* mname = C->method()->name()->as_quoted_ascii(); 2134 if (mname != NULL) { 2135 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname); 2136 } 2137 } 2138 2139 CloneMap& cm = C->clone_map(); 2140 Dict* dict = cm.dict(); 2141 if (C->do_vector_loop()) { 2142 cm.set_clone_idx(cm.max_gen()+1); 2143 #ifndef PRODUCT 2144 if (PrintOpto) { 2145 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx()); 2146 loop->dump_head(); 2147 } 2148 #endif 2149 } 2150 2151 // Step 1: Clone the loop body. Make the old->new mapping. 2152 uint i; 2153 for( i = 0; i < loop->_body.size(); i++ ) { 2154 Node *old = loop->_body.at(i); 2155 Node *nnn = old->clone(); 2156 old_new.map( old->_idx, nnn ); 2157 if (C->do_vector_loop()) { 2158 cm.verify_insert_and_clone(old, nnn, cm.clone_idx()); 2159 } 2160 _igvn.register_new_node_with_optimizer(nnn); 2161 } 2162 2163 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop; 2164 2165 // Step 2: Fix the edges in the new body. If the old input is outside the 2166 // loop use it. If the old input is INside the loop, use the corresponding 2167 // new node instead. 2168 for( i = 0; i < loop->_body.size(); i++ ) { 2169 Node *old = loop->_body.at(i); 2170 Node *nnn = old_new[old->_idx]; 2171 // Fix CFG/Loop controlling the new node 2172 if (has_ctrl(old)) { 2173 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]); 2174 } else { 2175 set_loop(nnn, outer_loop->_parent); 2176 if (old->outcnt() > 0) { 2177 set_idom( nnn, old_new[idom(old)->_idx], dd ); 2178 } 2179 } 2180 // Correct edges to the new node 2181 for( uint j = 0; j < nnn->req(); j++ ) { 2182 Node *n = nnn->in(j); 2183 if( n ) { 2184 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n ); 2185 if( loop->is_member( old_in_loop ) ) 2186 nnn->set_req(j, old_new[n->_idx]); 2187 } 2188 } 2189 _igvn.hash_find_insert(nnn); 2190 } 2191 2192 ResourceArea *area = Thread::current()->resource_area(); 2193 Node_List extra_data_nodes(area); // data nodes in the outer strip mined loop 2194 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes); 2195 2196 // Step 3: Now fix control uses. Loop varying control uses have already 2197 // been fixed up (as part of all input edges in Step 2). Loop invariant 2198 // control uses must be either an IfFalse or an IfTrue. Make a merge 2199 // point to merge the old and new IfFalse/IfTrue nodes; make the use 2200 // refer to this. 2201 Node_List worklist(area); 2202 uint new_counter = C->unique(); 2203 for( i = 0; i < loop->_body.size(); i++ ) { 2204 Node* old = loop->_body.at(i); 2205 if( !old->is_CFG() ) continue; 2206 2207 // Copy uses to a worklist, so I can munge the def-use info 2208 // with impunity. 2209 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 2210 worklist.push(old->fast_out(j)); 2211 2212 while( worklist.size() ) { // Visit all uses 2213 Node *use = worklist.pop(); 2214 if (!has_node(use)) continue; // Ignore dead nodes 2215 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 2216 if( !loop->is_member( use_loop ) && use->is_CFG() ) { 2217 // Both OLD and USE are CFG nodes here. 2218 assert( use->is_Proj(), "" ); 2219 Node* nnn = old_new[old->_idx]; 2220 2221 Node* newuse = NULL; 2222 if (head->is_strip_mined() && mode != IgnoreStripMined) { 2223 CountedLoopNode* cl = head->as_CountedLoop(); 2224 CountedLoopEndNode* cle = cl->loopexit(); 2225 Node* cle_out = cle->proj_out_or_null(false); 2226 if (use == cle_out) { 2227 IfNode* le = cl->outer_loop_end(); 2228 use = le->proj_out(false); 2229 use_loop = get_loop(use); 2230 if (mode == CloneIncludesStripMined) { 2231 nnn = old_new[le->_idx]; 2232 } else { 2233 newuse = old_new[cle_out->_idx]; 2234 } 2235 } 2236 } 2237 if (newuse == NULL) { 2238 newuse = use->clone(); 2239 } 2240 2241 // Clone the loop exit control projection 2242 if (C->do_vector_loop()) { 2243 cm.verify_insert_and_clone(use, newuse, cm.clone_idx()); 2244 } 2245 newuse->set_req(0,nnn); 2246 _igvn.register_new_node_with_optimizer(newuse); 2247 set_loop(newuse, use_loop); 2248 set_idom(newuse, nnn, dom_depth(nnn) + 1 ); 2249 2250 // We need a Region to merge the exit from the peeled body and the 2251 // exit from the old loop body. 2252 RegionNode *r = new RegionNode(3); 2253 // Map the old use to the new merge point 2254 old_new.map( use->_idx, r ); 2255 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use)); 2256 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" ); 2257 2258 // The original user of 'use' uses 'r' instead. 2259 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) { 2260 Node* useuse = use->last_out(l); 2261 _igvn.rehash_node_delayed(useuse); 2262 uint uses_found = 0; 2263 if( useuse->in(0) == use ) { 2264 useuse->set_req(0, r); 2265 uses_found++; 2266 if( useuse->is_CFG() ) { 2267 assert( dom_depth(useuse) > dd_r, "" ); 2268 set_idom(useuse, r, dom_depth(useuse)); 2269 } 2270 } 2271 for( uint k = 1; k < useuse->req(); k++ ) { 2272 if( useuse->in(k) == use ) { 2273 useuse->set_req(k, r); 2274 uses_found++; 2275 if (useuse->is_Loop() && k == LoopNode::EntryControl) { 2276 assert(dom_depth(useuse) > dd_r , ""); 2277 set_idom(useuse, r, dom_depth(useuse)); 2278 } 2279 } 2280 } 2281 l -= uses_found; // we deleted 1 or more copies of this edge 2282 } 2283 2284 // Now finish up 'r' 2285 r->set_req( 1, newuse ); 2286 r->set_req( 2, use ); 2287 _igvn.register_new_node_with_optimizer(r); 2288 set_loop(r, use_loop); 2289 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r); 2290 } // End of if a loop-exit test 2291 } 2292 } 2293 2294 // Step 4: If loop-invariant use is not control, it must be dominated by a 2295 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region 2296 // there if needed. Make a Phi there merging old and new used values. 2297 Node_List *split_if_set = NULL; 2298 Node_List *split_bool_set = NULL; 2299 Node_List *split_cex_set = NULL; 2300 for( i = 0; i < loop->_body.size(); i++ ) { 2301 Node* old = loop->_body.at(i); 2302 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set, 2303 split_bool_set, split_cex_set, worklist, new_counter, 2304 mode); 2305 } 2306 2307 for (i = 0; i < extra_data_nodes.size(); i++) { 2308 Node* old = extra_data_nodes.at(i); 2309 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set, 2310 split_bool_set, split_cex_set, worklist, new_counter, 2311 mode); 2312 } 2313 2314 // Check for IFs that need splitting/cloning. Happens if an IF outside of 2315 // the loop uses a condition set in the loop. The original IF probably 2316 // takes control from one or more OLD Regions (which in turn get from NEW 2317 // Regions). In any case, there will be a set of Phis for each merge point 2318 // from the IF up to where the original BOOL def exists the loop. 2319 if (split_if_set) { 2320 while (split_if_set->size()) { 2321 Node *iff = split_if_set->pop(); 2322 if (iff->in(1)->is_Phi()) { 2323 Node *b = clone_iff(iff->in(1)->as_Phi(), loop); 2324 _igvn.replace_input_of(iff, 1, b); 2325 } 2326 } 2327 } 2328 if (split_bool_set) { 2329 while (split_bool_set->size()) { 2330 Node *b = split_bool_set->pop(); 2331 Node *phi = b->in(1); 2332 assert(phi->is_Phi(), ""); 2333 CmpNode *cmp = clone_bool((PhiNode*)phi, loop); 2334 _igvn.replace_input_of(b, 1, cmp); 2335 } 2336 } 2337 if (split_cex_set) { 2338 while (split_cex_set->size()) { 2339 Node *b = split_cex_set->pop(); 2340 assert(b->in(0)->is_Region(), ""); 2341 assert(b->in(1)->is_Phi(), ""); 2342 assert(b->in(0)->in(0) == b->in(1)->in(0), ""); 2343 split_up(b, b->in(0), NULL); 2344 } 2345 } 2346 2347 } 2348 2349 2350 //---------------------- stride_of_possible_iv ------------------------------------- 2351 // Looks for an iff/bool/comp with one operand of the compare 2352 // being a cycle involving an add and a phi, 2353 // with an optional truncation (left-shift followed by a right-shift) 2354 // of the add. Returns zero if not an iv. 2355 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) { 2356 Node* trunc1 = NULL; 2357 Node* trunc2 = NULL; 2358 const TypeInt* ttype = NULL; 2359 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) { 2360 return 0; 2361 } 2362 BoolNode* bl = iff->in(1)->as_Bool(); 2363 Node* cmp = bl->in(1); 2364 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) { 2365 return 0; 2366 } 2367 // Must have an invariant operand 2368 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) { 2369 return 0; 2370 } 2371 Node* add2 = NULL; 2372 Node* cmp1 = cmp->in(1); 2373 if (cmp1->is_Phi()) { 2374 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) ))) 2375 Node* phi = cmp1; 2376 for (uint i = 1; i < phi->req(); i++) { 2377 Node* in = phi->in(i); 2378 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in, 2379 &trunc1, &trunc2, &ttype); 2380 if (add && add->in(1) == phi) { 2381 add2 = add->in(2); 2382 break; 2383 } 2384 } 2385 } else { 2386 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) ))) 2387 Node* addtrunc = cmp1; 2388 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc, 2389 &trunc1, &trunc2, &ttype); 2390 if (add && add->in(1)->is_Phi()) { 2391 Node* phi = add->in(1); 2392 for (uint i = 1; i < phi->req(); i++) { 2393 if (phi->in(i) == addtrunc) { 2394 add2 = add->in(2); 2395 break; 2396 } 2397 } 2398 } 2399 } 2400 if (add2 != NULL) { 2401 const TypeInt* add2t = _igvn.type(add2)->is_int(); 2402 if (add2t->is_con()) { 2403 return add2t->get_con(); 2404 } 2405 } 2406 return 0; 2407 } 2408 2409 2410 //---------------------- stay_in_loop ------------------------------------- 2411 // Return the (unique) control output node that's in the loop (if it exists.) 2412 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) { 2413 Node* unique = NULL; 2414 if (!n) return NULL; 2415 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2416 Node* use = n->fast_out(i); 2417 if (!has_ctrl(use) && loop->is_member(get_loop(use))) { 2418 if (unique != NULL) { 2419 return NULL; 2420 } 2421 unique = use; 2422 } 2423 } 2424 return unique; 2425 } 2426 2427 //------------------------------ register_node ------------------------------------- 2428 // Utility to register node "n" with PhaseIdealLoop 2429 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) { 2430 _igvn.register_new_node_with_optimizer(n); 2431 loop->_body.push(n); 2432 if (n->is_CFG()) { 2433 set_loop(n, loop); 2434 set_idom(n, pred, ddepth); 2435 } else { 2436 set_ctrl(n, pred); 2437 } 2438 } 2439 2440 //------------------------------ proj_clone ------------------------------------- 2441 // Utility to create an if-projection 2442 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) { 2443 ProjNode* c = p->clone()->as_Proj(); 2444 c->set_req(0, iff); 2445 return c; 2446 } 2447 2448 //------------------------------ short_circuit_if ------------------------------------- 2449 // Force the iff control output to be the live_proj 2450 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) { 2451 guarantee(live_proj != NULL, "null projection"); 2452 int proj_con = live_proj->_con; 2453 assert(proj_con == 0 || proj_con == 1, "false or true projection"); 2454 Node *con = _igvn.intcon(proj_con); 2455 set_ctrl(con, C->root()); 2456 if (iff) { 2457 iff->set_req(1, con); 2458 } 2459 return con; 2460 } 2461 2462 //------------------------------ insert_if_before_proj ------------------------------------- 2463 // Insert a new if before an if projection (* - new node) 2464 // 2465 // before 2466 // if(test) 2467 // / \ 2468 // v v 2469 // other-proj proj (arg) 2470 // 2471 // after 2472 // if(test) 2473 // / \ 2474 // / v 2475 // | * proj-clone 2476 // v | 2477 // other-proj v 2478 // * new_if(relop(cmp[IU](left,right))) 2479 // / \ 2480 // v v 2481 // * new-proj proj 2482 // (returned) 2483 // 2484 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) { 2485 IfNode* iff = proj->in(0)->as_If(); 2486 IdealLoopTree *loop = get_loop(proj); 2487 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2488 int ddepth = dom_depth(proj); 2489 2490 _igvn.rehash_node_delayed(iff); 2491 _igvn.rehash_node_delayed(proj); 2492 2493 proj->set_req(0, NULL); // temporary disconnect 2494 ProjNode* proj2 = proj_clone(proj, iff); 2495 register_node(proj2, loop, iff, ddepth); 2496 2497 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right); 2498 register_node(cmp, loop, proj2, ddepth); 2499 2500 BoolNode* bol = new BoolNode(cmp, relop); 2501 register_node(bol, loop, proj2, ddepth); 2502 2503 int opcode = iff->Opcode(); 2504 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode"); 2505 IfNode* new_if = (opcode == Op_If) ? new IfNode(proj2, bol, iff->_prob, iff->_fcnt): 2506 new RangeCheckNode(proj2, bol, iff->_prob, iff->_fcnt); 2507 register_node(new_if, loop, proj2, ddepth); 2508 2509 proj->set_req(0, new_if); // reattach 2510 set_idom(proj, new_if, ddepth); 2511 2512 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj(); 2513 guarantee(new_exit != NULL, "null exit node"); 2514 register_node(new_exit, get_loop(other_proj), new_if, ddepth); 2515 2516 return new_exit; 2517 } 2518 2519 //------------------------------ insert_region_before_proj ------------------------------------- 2520 // Insert a region before an if projection (* - new node) 2521 // 2522 // before 2523 // if(test) 2524 // / | 2525 // v | 2526 // proj v 2527 // other-proj 2528 // 2529 // after 2530 // if(test) 2531 // / | 2532 // v | 2533 // * proj-clone v 2534 // | other-proj 2535 // v 2536 // * new-region 2537 // | 2538 // v 2539 // * dum_if 2540 // / \ 2541 // v \ 2542 // * dum-proj v 2543 // proj 2544 // 2545 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) { 2546 IfNode* iff = proj->in(0)->as_If(); 2547 IdealLoopTree *loop = get_loop(proj); 2548 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2549 int ddepth = dom_depth(proj); 2550 2551 _igvn.rehash_node_delayed(iff); 2552 _igvn.rehash_node_delayed(proj); 2553 2554 proj->set_req(0, NULL); // temporary disconnect 2555 ProjNode* proj2 = proj_clone(proj, iff); 2556 register_node(proj2, loop, iff, ddepth); 2557 2558 RegionNode* reg = new RegionNode(2); 2559 reg->set_req(1, proj2); 2560 register_node(reg, loop, iff, ddepth); 2561 2562 IfNode* dum_if = new IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt); 2563 register_node(dum_if, loop, reg, ddepth); 2564 2565 proj->set_req(0, dum_if); // reattach 2566 set_idom(proj, dum_if, ddepth); 2567 2568 ProjNode* dum_proj = proj_clone(other_proj, dum_if); 2569 register_node(dum_proj, loop, dum_if, ddepth); 2570 2571 return reg; 2572 } 2573 2574 //------------------------------ insert_cmpi_loop_exit ------------------------------------- 2575 // Clone a signed compare loop exit from an unsigned compare and 2576 // insert it before the unsigned cmp on the stay-in-loop path. 2577 // All new nodes inserted in the dominator tree between the original 2578 // if and it's projections. The original if test is replaced with 2579 // a constant to force the stay-in-loop path. 2580 // 2581 // This is done to make sure that the original if and it's projections 2582 // still dominate the same set of control nodes, that the ctrl() relation 2583 // from data nodes to them is preserved, and that their loop nesting is 2584 // preserved. 2585 // 2586 // before 2587 // if(i <u limit) unsigned compare loop exit 2588 // / | 2589 // v v 2590 // exit-proj stay-in-loop-proj 2591 // 2592 // after 2593 // if(stay-in-loop-const) original if 2594 // / | 2595 // / v 2596 // / if(i < limit) new signed test 2597 // / / | 2598 // / / v 2599 // / / if(i <u limit) new cloned unsigned test 2600 // / / / | 2601 // v v v | 2602 // region | 2603 // | | 2604 // dum-if | 2605 // / | | 2606 // ether | | 2607 // v v 2608 // exit-proj stay-in-loop-proj 2609 // 2610 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) { 2611 const bool Signed = true; 2612 const bool Unsigned = false; 2613 2614 BoolNode* bol = if_cmpu->in(1)->as_Bool(); 2615 if (bol->_test._test != BoolTest::lt) return NULL; 2616 CmpNode* cmpu = bol->in(1)->as_Cmp(); 2617 if (cmpu->Opcode() != Op_CmpU) return NULL; 2618 int stride = stride_of_possible_iv(if_cmpu); 2619 if (stride == 0) return NULL; 2620 2621 Node* lp_proj = stay_in_loop(if_cmpu, loop); 2622 guarantee(lp_proj != NULL, "null loop node"); 2623 2624 ProjNode* lp_continue = lp_proj->as_Proj(); 2625 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj(); 2626 2627 Node* limit = NULL; 2628 if (stride > 0) { 2629 limit = cmpu->in(2); 2630 } else { 2631 limit = _igvn.makecon(TypeInt::ZERO); 2632 set_ctrl(limit, C->root()); 2633 } 2634 // Create a new region on the exit path 2635 RegionNode* reg = insert_region_before_proj(lp_exit); 2636 guarantee(reg != NULL, "null region node"); 2637 2638 // Clone the if-cmpu-true-false using a signed compare 2639 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge; 2640 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue); 2641 reg->add_req(cmpi_exit); 2642 2643 // Clone the if-cmpu-true-false 2644 BoolTest::mask rel_u = bol->_test._test; 2645 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue); 2646 reg->add_req(cmpu_exit); 2647 2648 // Force original if to stay in loop. 2649 short_circuit_if(if_cmpu, lp_continue); 2650 2651 return cmpi_exit->in(0)->as_If(); 2652 } 2653 2654 //------------------------------ remove_cmpi_loop_exit ------------------------------------- 2655 // Remove a previously inserted signed compare loop exit. 2656 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) { 2657 Node* lp_proj = stay_in_loop(if_cmp, loop); 2658 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI && 2659 stay_in_loop(lp_proj, loop)->is_If() && 2660 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu"); 2661 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO); 2662 set_ctrl(con, C->root()); 2663 if_cmp->set_req(1, con); 2664 } 2665 2666 //------------------------------ scheduled_nodelist ------------------------------------- 2667 // Create a post order schedule of nodes that are in the 2668 // "member" set. The list is returned in "sched". 2669 // The first node in "sched" is the loop head, followed by 2670 // nodes which have no inputs in the "member" set, and then 2671 // followed by the nodes that have an immediate input dependence 2672 // on a node in "sched". 2673 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) { 2674 2675 assert(member.test(loop->_head->_idx), "loop head must be in member set"); 2676 Arena *a = Thread::current()->resource_area(); 2677 VectorSet visited(a); 2678 Node_Stack nstack(a, loop->_body.size()); 2679 2680 Node* n = loop->_head; // top of stack is cached in "n" 2681 uint idx = 0; 2682 visited.set(n->_idx); 2683 2684 // Initially push all with no inputs from within member set 2685 for(uint i = 0; i < loop->_body.size(); i++ ) { 2686 Node *elt = loop->_body.at(i); 2687 if (member.test(elt->_idx)) { 2688 bool found = false; 2689 for (uint j = 0; j < elt->req(); j++) { 2690 Node* def = elt->in(j); 2691 if (def && member.test(def->_idx) && def != elt) { 2692 found = true; 2693 break; 2694 } 2695 } 2696 if (!found && elt != loop->_head) { 2697 nstack.push(n, idx); 2698 n = elt; 2699 assert(!visited.test(n->_idx), "not seen yet"); 2700 visited.set(n->_idx); 2701 } 2702 } 2703 } 2704 2705 // traverse out's that are in the member set 2706 while (true) { 2707 if (idx < n->outcnt()) { 2708 Node* use = n->raw_out(idx); 2709 idx++; 2710 if (!visited.test_set(use->_idx)) { 2711 if (member.test(use->_idx)) { 2712 nstack.push(n, idx); 2713 n = use; 2714 idx = 0; 2715 } 2716 } 2717 } else { 2718 // All outputs processed 2719 sched.push(n); 2720 if (nstack.is_empty()) break; 2721 n = nstack.node(); 2722 idx = nstack.index(); 2723 nstack.pop(); 2724 } 2725 } 2726 } 2727 2728 2729 //------------------------------ has_use_in_set ------------------------------------- 2730 // Has a use in the vector set 2731 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) { 2732 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2733 Node* use = n->fast_out(j); 2734 if (vset.test(use->_idx)) { 2735 return true; 2736 } 2737 } 2738 return false; 2739 } 2740 2741 2742 //------------------------------ has_use_internal_to_set ------------------------------------- 2743 // Has use internal to the vector set (ie. not in a phi at the loop head) 2744 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) { 2745 Node* head = loop->_head; 2746 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2747 Node* use = n->fast_out(j); 2748 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) { 2749 return true; 2750 } 2751 } 2752 return false; 2753 } 2754 2755 2756 //------------------------------ clone_for_use_outside_loop ------------------------------------- 2757 // clone "n" for uses that are outside of loop 2758 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) { 2759 int cloned = 0; 2760 assert(worklist.size() == 0, "should be empty"); 2761 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2762 Node* use = n->fast_out(j); 2763 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) { 2764 worklist.push(use); 2765 } 2766 } 2767 while( worklist.size() ) { 2768 Node *use = worklist.pop(); 2769 if (!has_node(use) || use->in(0) == C->top()) continue; 2770 uint j; 2771 for (j = 0; j < use->req(); j++) { 2772 if (use->in(j) == n) break; 2773 } 2774 assert(j < use->req(), "must be there"); 2775 2776 // clone "n" and insert it between the inputs of "n" and the use outside the loop 2777 Node* n_clone = n->clone(); 2778 _igvn.replace_input_of(use, j, n_clone); 2779 cloned++; 2780 Node* use_c; 2781 if (!use->is_Phi()) { 2782 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0); 2783 } else { 2784 // Use in a phi is considered a use in the associated predecessor block 2785 use_c = use->in(0)->in(j); 2786 } 2787 set_ctrl(n_clone, use_c); 2788 assert(!loop->is_member(get_loop(use_c)), "should be outside loop"); 2789 get_loop(use_c)->_body.push(n_clone); 2790 _igvn.register_new_node_with_optimizer(n_clone); 2791 #ifndef PRODUCT 2792 if (TracePartialPeeling) { 2793 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx); 2794 } 2795 #endif 2796 } 2797 return cloned; 2798 } 2799 2800 2801 //------------------------------ clone_for_special_use_inside_loop ------------------------------------- 2802 // clone "n" for special uses that are in the not_peeled region. 2803 // If these def-uses occur in separate blocks, the code generator 2804 // marks the method as not compilable. For example, if a "BoolNode" 2805 // is in a different basic block than the "IfNode" that uses it, then 2806 // the compilation is aborted in the code generator. 2807 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 2808 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) { 2809 if (n->is_Phi() || n->is_Load()) { 2810 return; 2811 } 2812 assert(worklist.size() == 0, "should be empty"); 2813 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2814 Node* use = n->fast_out(j); 2815 if ( not_peel.test(use->_idx) && 2816 (use->is_If() || use->is_CMove() || use->is_Bool()) && 2817 use->in(1) == n) { 2818 worklist.push(use); 2819 } 2820 } 2821 if (worklist.size() > 0) { 2822 // clone "n" and insert it between inputs of "n" and the use 2823 Node* n_clone = n->clone(); 2824 loop->_body.push(n_clone); 2825 _igvn.register_new_node_with_optimizer(n_clone); 2826 set_ctrl(n_clone, get_ctrl(n)); 2827 sink_list.push(n_clone); 2828 not_peel <<= n_clone->_idx; // add n_clone to not_peel set. 2829 #ifndef PRODUCT 2830 if (TracePartialPeeling) { 2831 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx); 2832 } 2833 #endif 2834 while( worklist.size() ) { 2835 Node *use = worklist.pop(); 2836 _igvn.rehash_node_delayed(use); 2837 for (uint j = 1; j < use->req(); j++) { 2838 if (use->in(j) == n) { 2839 use->set_req(j, n_clone); 2840 } 2841 } 2842 } 2843 } 2844 } 2845 2846 2847 //------------------------------ insert_phi_for_loop ------------------------------------- 2848 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 2849 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) { 2850 Node *phi = PhiNode::make(lp, back_edge_val); 2851 phi->set_req(LoopNode::EntryControl, lp_entry_val); 2852 // Use existing phi if it already exists 2853 Node *hit = _igvn.hash_find_insert(phi); 2854 if( hit == NULL ) { 2855 _igvn.register_new_node_with_optimizer(phi); 2856 set_ctrl(phi, lp); 2857 } else { 2858 // Remove the new phi from the graph and use the hit 2859 _igvn.remove_dead_node(phi); 2860 phi = hit; 2861 } 2862 _igvn.replace_input_of(use, idx, phi); 2863 } 2864 2865 #ifdef ASSERT 2866 //------------------------------ is_valid_loop_partition ------------------------------------- 2867 // Validate the loop partition sets: peel and not_peel 2868 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, 2869 VectorSet& not_peel ) { 2870 uint i; 2871 // Check that peel_list entries are in the peel set 2872 for (i = 0; i < peel_list.size(); i++) { 2873 if (!peel.test(peel_list.at(i)->_idx)) { 2874 return false; 2875 } 2876 } 2877 // Check at loop members are in one of peel set or not_peel set 2878 for (i = 0; i < loop->_body.size(); i++ ) { 2879 Node *def = loop->_body.at(i); 2880 uint di = def->_idx; 2881 // Check that peel set elements are in peel_list 2882 if (peel.test(di)) { 2883 if (not_peel.test(di)) { 2884 return false; 2885 } 2886 // Must be in peel_list also 2887 bool found = false; 2888 for (uint j = 0; j < peel_list.size(); j++) { 2889 if (peel_list.at(j)->_idx == di) { 2890 found = true; 2891 break; 2892 } 2893 } 2894 if (!found) { 2895 return false; 2896 } 2897 } else if (not_peel.test(di)) { 2898 if (peel.test(di)) { 2899 return false; 2900 } 2901 } else { 2902 return false; 2903 } 2904 } 2905 return true; 2906 } 2907 2908 //------------------------------ is_valid_clone_loop_exit_use ------------------------------------- 2909 // Ensure a use outside of loop is of the right form 2910 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) { 2911 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2912 return (use->is_Phi() && 2913 use_c->is_Region() && use_c->req() == 3 && 2914 (use_c->in(exit_idx)->Opcode() == Op_IfTrue || 2915 use_c->in(exit_idx)->Opcode() == Op_IfFalse || 2916 use_c->in(exit_idx)->Opcode() == Op_JumpProj) && 2917 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) ); 2918 } 2919 2920 //------------------------------ is_valid_clone_loop_form ------------------------------------- 2921 // Ensure that all uses outside of loop are of the right form 2922 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 2923 uint orig_exit_idx, uint clone_exit_idx) { 2924 uint len = peel_list.size(); 2925 for (uint i = 0; i < len; i++) { 2926 Node *def = peel_list.at(i); 2927 2928 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 2929 Node *use = def->fast_out(j); 2930 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2931 if (!loop->is_member(get_loop(use_c))) { 2932 // use is not in the loop, check for correct structure 2933 if (use->in(0) == def) { 2934 // Okay 2935 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) { 2936 return false; 2937 } 2938 } 2939 } 2940 } 2941 return true; 2942 } 2943 #endif 2944 2945 //------------------------------ partial_peel ------------------------------------- 2946 // Partially peel (aka loop rotation) the top portion of a loop (called 2947 // the peel section below) by cloning it and placing one copy just before 2948 // the new loop head and the other copy at the bottom of the new loop. 2949 // 2950 // before after where it came from 2951 // 2952 // stmt1 stmt1 2953 // loop: stmt2 clone 2954 // stmt2 if condA goto exitA clone 2955 // if condA goto exitA new_loop: new 2956 // stmt3 stmt3 clone 2957 // if !condB goto loop if condB goto exitB clone 2958 // exitB: stmt2 orig 2959 // stmt4 if !condA goto new_loop orig 2960 // exitA: goto exitA 2961 // exitB: 2962 // stmt4 2963 // exitA: 2964 // 2965 // Step 1: find the cut point: an exit test on probable 2966 // induction variable. 2967 // Step 2: schedule (with cloning) operations in the peel 2968 // section that can be executed after the cut into 2969 // the section that is not peeled. This may need 2970 // to clone operations into exit blocks. For 2971 // instance, a reference to A[i] in the not-peel 2972 // section and a reference to B[i] in an exit block 2973 // may cause a left-shift of i by 2 to be placed 2974 // in the peel block. This step will clone the left 2975 // shift into the exit block and sink the left shift 2976 // from the peel to the not-peel section. 2977 // Step 3: clone the loop, retarget the control, and insert 2978 // phis for values that are live across the new loop 2979 // head. This is very dependent on the graph structure 2980 // from clone_loop. It creates region nodes for 2981 // exit control and associated phi nodes for values 2982 // flow out of the loop through that exit. The region 2983 // node is dominated by the clone's control projection. 2984 // So the clone's peel section is placed before the 2985 // new loop head, and the clone's not-peel section is 2986 // forms the top part of the new loop. The original 2987 // peel section forms the tail of the new loop. 2988 // Step 4: update the dominator tree and recompute the 2989 // dominator depth. 2990 // 2991 // orig 2992 // 2993 // stmt1 2994 // | 2995 // v 2996 // loop predicate 2997 // | 2998 // v 2999 // loop<----+ 3000 // | | 3001 // stmt2 | 3002 // | | 3003 // v | 3004 // ifA | 3005 // / | | 3006 // v v | 3007 // false true ^ <-- last_peel 3008 // / | | 3009 // / ===|==cut | 3010 // / stmt3 | <-- first_not_peel 3011 // / | | 3012 // | v | 3013 // v ifB | 3014 // exitA: / \ | 3015 // / \ | 3016 // v v | 3017 // false true | 3018 // / \ | 3019 // / ----+ 3020 // | 3021 // v 3022 // exitB: 3023 // stmt4 3024 // 3025 // 3026 // after clone loop 3027 // 3028 // stmt1 3029 // | 3030 // v 3031 // loop predicate 3032 // / \ 3033 // clone / \ orig 3034 // / \ 3035 // / \ 3036 // v v 3037 // +---->loop loop<----+ 3038 // | | | | 3039 // | stmt2 stmt2 | 3040 // | | | | 3041 // | v v | 3042 // | ifA ifA | 3043 // | | \ / | | 3044 // | v v v v | 3045 // ^ true false false true ^ <-- last_peel 3046 // | | ^ \ / | | 3047 // | cut==|== \ \ / ===|==cut | 3048 // | stmt3 \ \ / stmt3 | <-- first_not_peel 3049 // | | dom | | | | 3050 // | v \ 1v v2 v | 3051 // | ifB regionA ifB | 3052 // | / \ | / \ | 3053 // | / \ v / \ | 3054 // | v v exitA: v v | 3055 // | true false false true | 3056 // | / ^ \ / \ | 3057 // +---- \ \ / ----+ 3058 // dom \ / 3059 // \ 1v v2 3060 // regionB 3061 // | 3062 // v 3063 // exitB: 3064 // stmt4 3065 // 3066 // 3067 // after partial peel 3068 // 3069 // stmt1 3070 // | 3071 // v 3072 // loop predicate 3073 // / 3074 // clone / orig 3075 // / TOP 3076 // / \ 3077 // v v 3078 // TOP->loop loop----+ 3079 // | | | 3080 // stmt2 stmt2 | 3081 // | | | 3082 // v v | 3083 // ifA ifA | 3084 // | \ / | | 3085 // v v v v | 3086 // true false false true | <-- last_peel 3087 // | ^ \ / +------|---+ 3088 // +->newloop \ \ / === ==cut | | 3089 // | stmt3 \ \ / TOP | | 3090 // | | dom | | stmt3 | | <-- first_not_peel 3091 // | v \ 1v v2 v | | 3092 // | ifB regionA ifB ^ v 3093 // | / \ | / \ | | 3094 // | / \ v / \ | | 3095 // | v v exitA: v v | | 3096 // | true false false true | | 3097 // | / ^ \ / \ | | 3098 // | | \ \ / v | | 3099 // | | dom \ / TOP | | 3100 // | | \ 1v v2 | | 3101 // ^ v regionB | | 3102 // | | | | | 3103 // | | v ^ v 3104 // | | exitB: | | 3105 // | | stmt4 | | 3106 // | +------------>-----------------+ | 3107 // | | 3108 // +-----------------<---------------------+ 3109 // 3110 // 3111 // final graph 3112 // 3113 // stmt1 3114 // | 3115 // v 3116 // loop predicate 3117 // | 3118 // v 3119 // stmt2 clone 3120 // | 3121 // v 3122 // ........> ifA clone 3123 // : / | 3124 // dom / | 3125 // : v v 3126 // : false true 3127 // : | | 3128 // : | v 3129 // : | newloop<-----+ 3130 // : | | | 3131 // : | stmt3 clone | 3132 // : | | | 3133 // : | v | 3134 // : | ifB | 3135 // : | / \ | 3136 // : | v v | 3137 // : | false true | 3138 // : | | | | 3139 // : | v stmt2 | 3140 // : | exitB: | | 3141 // : | stmt4 v | 3142 // : | ifA orig | 3143 // : | / \ | 3144 // : | / \ | 3145 // : | v v | 3146 // : | false true | 3147 // : | / \ | 3148 // : v v -----+ 3149 // RegionA 3150 // | 3151 // v 3152 // exitA 3153 // 3154 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) { 3155 3156 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only"); 3157 if (!loop->_head->is_Loop()) { 3158 return false; 3159 } 3160 LoopNode *head = loop->_head->as_Loop(); 3161 3162 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) { 3163 return false; 3164 } 3165 3166 // Check for complex exit control 3167 for (uint ii = 0; ii < loop->_body.size(); ii++) { 3168 Node *n = loop->_body.at(ii); 3169 int opc = n->Opcode(); 3170 if (n->is_Call() || 3171 opc == Op_Catch || 3172 opc == Op_CatchProj || 3173 opc == Op_Jump || 3174 opc == Op_JumpProj) { 3175 #ifndef PRODUCT 3176 if (TracePartialPeeling) { 3177 tty->print_cr("\nExit control too complex: lp: %d", head->_idx); 3178 } 3179 #endif 3180 return false; 3181 } 3182 } 3183 3184 int dd = dom_depth(head); 3185 3186 // Step 1: find cut point 3187 3188 // Walk up dominators to loop head looking for first loop exit 3189 // which is executed on every path thru loop. 3190 IfNode *peel_if = NULL; 3191 IfNode *peel_if_cmpu = NULL; 3192 3193 Node *iff = loop->tail(); 3194 while (iff != head) { 3195 if (iff->is_If()) { 3196 Node *ctrl = get_ctrl(iff->in(1)); 3197 if (ctrl->is_top()) return false; // Dead test on live IF. 3198 // If loop-varying exit-test, check for induction variable 3199 if (loop->is_member(get_loop(ctrl)) && 3200 loop->is_loop_exit(iff) && 3201 is_possible_iv_test(iff)) { 3202 Node* cmp = iff->in(1)->in(1); 3203 if (cmp->Opcode() == Op_CmpI) { 3204 peel_if = iff->as_If(); 3205 } else { 3206 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU"); 3207 peel_if_cmpu = iff->as_If(); 3208 } 3209 } 3210 } 3211 iff = idom(iff); 3212 } 3213 3214 // Prefer signed compare over unsigned compare. 3215 IfNode* new_peel_if = NULL; 3216 if (peel_if == NULL) { 3217 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) { 3218 return false; // No peel point found 3219 } 3220 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop); 3221 if (new_peel_if == NULL) { 3222 return false; // No peel point found 3223 } 3224 peel_if = new_peel_if; 3225 } 3226 Node* last_peel = stay_in_loop(peel_if, loop); 3227 Node* first_not_peeled = stay_in_loop(last_peel, loop); 3228 if (first_not_peeled == NULL || first_not_peeled == head) { 3229 return false; 3230 } 3231 3232 #ifndef PRODUCT 3233 if (TraceLoopOpts) { 3234 tty->print("PartialPeel "); 3235 loop->dump_head(); 3236 } 3237 3238 if (TracePartialPeeling) { 3239 tty->print_cr("before partial peel one iteration"); 3240 Node_List wl; 3241 Node* t = head->in(2); 3242 while (true) { 3243 wl.push(t); 3244 if (t == head) break; 3245 t = idom(t); 3246 } 3247 while (wl.size() > 0) { 3248 Node* tt = wl.pop(); 3249 tt->dump(); 3250 if (tt == last_peel) tty->print_cr("-- cut --"); 3251 } 3252 } 3253 #endif 3254 ResourceArea *area = Thread::current()->resource_area(); 3255 VectorSet peel(area); 3256 VectorSet not_peel(area); 3257 Node_List peel_list(area); 3258 Node_List worklist(area); 3259 Node_List sink_list(area); 3260 3261 if (!may_require_nodes(loop->est_loop_clone_sz(2))) { 3262 return false; 3263 } 3264 3265 // Set of cfg nodes to peel are those that are executable from 3266 // the head through last_peel. 3267 assert(worklist.size() == 0, "should be empty"); 3268 worklist.push(head); 3269 peel.set(head->_idx); 3270 while (worklist.size() > 0) { 3271 Node *n = worklist.pop(); 3272 if (n != last_peel) { 3273 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3274 Node* use = n->fast_out(j); 3275 if (use->is_CFG() && 3276 loop->is_member(get_loop(use)) && 3277 !peel.test_set(use->_idx)) { 3278 worklist.push(use); 3279 } 3280 } 3281 } 3282 } 3283 3284 // Set of non-cfg nodes to peel are those that are control 3285 // dependent on the cfg nodes. 3286 uint i; 3287 for(i = 0; i < loop->_body.size(); i++ ) { 3288 Node *n = loop->_body.at(i); 3289 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n; 3290 if (peel.test(n_c->_idx)) { 3291 peel.set(n->_idx); 3292 } else { 3293 not_peel.set(n->_idx); 3294 } 3295 } 3296 3297 // Step 2: move operations from the peeled section down into the 3298 // not-peeled section 3299 3300 // Get a post order schedule of nodes in the peel region 3301 // Result in right-most operand. 3302 scheduled_nodelist(loop, peel, peel_list ); 3303 3304 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 3305 3306 // For future check for too many new phis 3307 uint old_phi_cnt = 0; 3308 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 3309 Node* use = head->fast_out(j); 3310 if (use->is_Phi()) old_phi_cnt++; 3311 } 3312 3313 #ifndef PRODUCT 3314 if (TracePartialPeeling) { 3315 tty->print_cr("\npeeled list"); 3316 } 3317 #endif 3318 3319 // Evacuate nodes in peel region into the not_peeled region if possible 3320 uint new_phi_cnt = 0; 3321 uint cloned_for_outside_use = 0; 3322 for (i = 0; i < peel_list.size();) { 3323 Node* n = peel_list.at(i); 3324 #ifndef PRODUCT 3325 if (TracePartialPeeling) n->dump(); 3326 #endif 3327 bool incr = true; 3328 if ( !n->is_CFG() ) { 3329 3330 if ( has_use_in_set(n, not_peel) ) { 3331 3332 // If not used internal to the peeled region, 3333 // move "n" from peeled to not_peeled region. 3334 3335 if ( !has_use_internal_to_set(n, peel, loop) ) { 3336 3337 // if not pinned and not a load (which maybe anti-dependent on a store) 3338 // and not a CMove (Matcher expects only bool->cmove). 3339 if (n->in(0) == NULL && !n->is_Load() && !n->is_CMove()) { 3340 cloned_for_outside_use += clone_for_use_outside_loop( loop, n, worklist ); 3341 sink_list.push(n); 3342 peel >>= n->_idx; // delete n from peel set. 3343 not_peel <<= n->_idx; // add n to not_peel set. 3344 peel_list.remove(i); 3345 incr = false; 3346 #ifndef PRODUCT 3347 if (TracePartialPeeling) { 3348 tty->print_cr("sink to not_peeled region: %d newbb: %d", 3349 n->_idx, get_ctrl(n)->_idx); 3350 } 3351 #endif 3352 } 3353 } else { 3354 // Otherwise check for special def-use cases that span 3355 // the peel/not_peel boundary such as bool->if 3356 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist ); 3357 new_phi_cnt++; 3358 } 3359 } 3360 } 3361 if (incr) i++; 3362 } 3363 3364 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) { 3365 #ifndef PRODUCT 3366 if (TracePartialPeeling) { 3367 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c", 3368 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F'); 3369 } 3370 #endif 3371 if (new_peel_if != NULL) { 3372 remove_cmpi_loop_exit(new_peel_if, loop); 3373 } 3374 // Inhibit more partial peeling on this loop 3375 assert(!head->is_partial_peel_loop(), "not partial peeled"); 3376 head->mark_partial_peel_failed(); 3377 if (cloned_for_outside_use > 0) { 3378 // Terminate this round of loop opts because 3379 // the graph outside this loop was changed. 3380 C->set_major_progress(); 3381 return true; 3382 } 3383 return false; 3384 } 3385 3386 // Step 3: clone loop, retarget control, and insert new phis 3387 3388 // Create new loop head for new phis and to hang 3389 // the nodes being moved (sinked) from the peel region. 3390 LoopNode* new_head = new LoopNode(last_peel, last_peel); 3391 new_head->set_unswitch_count(head->unswitch_count()); // Preserve 3392 _igvn.register_new_node_with_optimizer(new_head); 3393 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled"); 3394 _igvn.replace_input_of(first_not_peeled, 0, new_head); 3395 set_loop(new_head, loop); 3396 loop->_body.push(new_head); 3397 not_peel.set(new_head->_idx); 3398 set_idom(new_head, last_peel, dom_depth(first_not_peeled)); 3399 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled)); 3400 3401 while (sink_list.size() > 0) { 3402 Node* n = sink_list.pop(); 3403 set_ctrl(n, new_head); 3404 } 3405 3406 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 3407 3408 clone_loop(loop, old_new, dd, IgnoreStripMined); 3409 3410 const uint clone_exit_idx = 1; 3411 const uint orig_exit_idx = 2; 3412 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop"); 3413 3414 Node* head_clone = old_new[head->_idx]; 3415 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop(); 3416 Node* orig_tail_clone = head_clone->in(2); 3417 3418 // Add phi if "def" node is in peel set and "use" is not 3419 3420 for(i = 0; i < peel_list.size(); i++ ) { 3421 Node *def = peel_list.at(i); 3422 if (!def->is_CFG()) { 3423 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 3424 Node *use = def->fast_out(j); 3425 if (has_node(use) && use->in(0) != C->top() && 3426 (!peel.test(use->_idx) || 3427 (use->is_Phi() && use->in(0) == head)) ) { 3428 worklist.push(use); 3429 } 3430 } 3431 while( worklist.size() ) { 3432 Node *use = worklist.pop(); 3433 for (uint j = 1; j < use->req(); j++) { 3434 Node* n = use->in(j); 3435 if (n == def) { 3436 3437 // "def" is in peel set, "use" is not in peel set 3438 // or "use" is in the entry boundary (a phi) of the peel set 3439 3440 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use; 3441 3442 if ( loop->is_member(get_loop( use_c )) ) { 3443 // use is in loop 3444 if (old_new[use->_idx] != NULL) { // null for dead code 3445 Node* use_clone = old_new[use->_idx]; 3446 _igvn.replace_input_of(use, j, C->top()); 3447 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone ); 3448 } 3449 } else { 3450 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format"); 3451 // use is not in the loop, check if the live range includes the cut 3452 Node* lp_if = use_c->in(orig_exit_idx)->in(0); 3453 if (not_peel.test(lp_if->_idx)) { 3454 assert(j == orig_exit_idx, "use from original loop"); 3455 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone ); 3456 } 3457 } 3458 } 3459 } 3460 } 3461 } 3462 } 3463 3464 // Step 3b: retarget control 3465 3466 // Redirect control to the new loop head if a cloned node in 3467 // the not_peeled region has control that points into the peeled region. 3468 // This necessary because the cloned peeled region will be outside 3469 // the loop. 3470 // from to 3471 // cloned-peeled <---+ 3472 // new_head_clone: | <--+ 3473 // cloned-not_peeled in(0) in(0) 3474 // orig-peeled 3475 3476 for(i = 0; i < loop->_body.size(); i++ ) { 3477 Node *n = loop->_body.at(i); 3478 if (!n->is_CFG() && n->in(0) != NULL && 3479 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) { 3480 Node* n_clone = old_new[n->_idx]; 3481 _igvn.replace_input_of(n_clone, 0, new_head_clone); 3482 } 3483 } 3484 3485 // Backedge of the surviving new_head (the clone) is original last_peel 3486 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel); 3487 3488 // Cut first node in original not_peel set 3489 _igvn.rehash_node_delayed(new_head); // Multiple edge updates: 3490 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of 3491 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls 3492 3493 // Copy head_clone back-branch info to original head 3494 // and remove original head's loop entry and 3495 // clone head's back-branch 3496 _igvn.rehash_node_delayed(head); // Multiple edge updates 3497 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl)); 3498 head->set_req(LoopNode::LoopBackControl, C->top()); 3499 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top()); 3500 3501 // Similarly modify the phis 3502 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) { 3503 Node* use = head->fast_out(k); 3504 if (use->is_Phi() && use->outcnt() > 0) { 3505 Node* use_clone = old_new[use->_idx]; 3506 _igvn.rehash_node_delayed(use); // Multiple edge updates 3507 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl)); 3508 use->set_req(LoopNode::LoopBackControl, C->top()); 3509 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top()); 3510 } 3511 } 3512 3513 // Step 4: update dominator tree and dominator depth 3514 3515 set_idom(head, orig_tail_clone, dd); 3516 recompute_dom_depth(); 3517 3518 // Inhibit more partial peeling on this loop 3519 new_head_clone->set_partial_peel_loop(); 3520 C->set_major_progress(); 3521 loop->record_for_igvn(); 3522 3523 #ifndef PRODUCT 3524 if (TracePartialPeeling) { 3525 tty->print_cr("\nafter partial peel one iteration"); 3526 Node_List wl(area); 3527 Node* t = last_peel; 3528 while (true) { 3529 wl.push(t); 3530 if (t == head_clone) break; 3531 t = idom(t); 3532 } 3533 while (wl.size() > 0) { 3534 Node* tt = wl.pop(); 3535 if (tt == head) tty->print_cr("orig head"); 3536 else if (tt == new_head_clone) tty->print_cr("new head"); 3537 else if (tt == head_clone) tty->print_cr("clone head"); 3538 tt->dump(); 3539 } 3540 } 3541 #endif 3542 return true; 3543 } 3544 3545 //------------------------------reorg_offsets---------------------------------- 3546 // Reorganize offset computations to lower register pressure. Mostly 3547 // prevent loop-fallout uses of the pre-incremented trip counter (which are 3548 // then alive with the post-incremented trip counter forcing an extra 3549 // register move) 3550 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) { 3551 // Perform it only for canonical counted loops. 3552 // Loop's shape could be messed up by iteration_split_impl. 3553 if (!loop->_head->is_CountedLoop()) 3554 return; 3555 if (!loop->_head->as_Loop()->is_valid_counted_loop()) 3556 return; 3557 3558 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 3559 CountedLoopEndNode *cle = cl->loopexit(); 3560 Node *exit = cle->proj_out(false); 3561 Node *phi = cl->phi(); 3562 3563 // Check for the special case when using the pre-incremented trip-counter on 3564 // the fall-out path (forces the pre-incremented and post-incremented trip 3565 // counter to be live at the same time). Fix this by adjusting to use the 3566 // post-increment trip counter. 3567 3568 bool progress = true; 3569 while (progress) { 3570 progress = false; 3571 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) { 3572 Node* use = phi->fast_out(i); // User of trip-counter 3573 if (!has_ctrl(use)) continue; 3574 Node *u_ctrl = get_ctrl(use); 3575 if (use->is_Phi()) { 3576 u_ctrl = NULL; 3577 for (uint j = 1; j < use->req(); j++) 3578 if (use->in(j) == phi) 3579 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j)); 3580 } 3581 IdealLoopTree *u_loop = get_loop(u_ctrl); 3582 // Look for loop-invariant use 3583 if (u_loop == loop) continue; 3584 if (loop->is_member(u_loop)) continue; 3585 // Check that use is live out the bottom. Assuming the trip-counter 3586 // update is right at the bottom, uses of of the loop middle are ok. 3587 if (dom_lca(exit, u_ctrl) != exit) continue; 3588 // Hit! Refactor use to use the post-incremented tripcounter. 3589 // Compute a post-increment tripcounter. 3590 Node* c = exit; 3591 if (cl->is_strip_mined()) { 3592 IdealLoopTree* outer_loop = get_loop(cl->outer_loop()); 3593 if (!outer_loop->is_member(u_loop)) { 3594 c = cl->outer_loop_exit(); 3595 } 3596 } 3597 Node *opaq = new Opaque2Node(C, cle->incr()); 3598 register_new_node(opaq, c); 3599 Node *neg_stride = _igvn.intcon(-cle->stride_con()); 3600 set_ctrl(neg_stride, C->root()); 3601 Node *post = new AddINode(opaq, neg_stride); 3602 register_new_node(post, c); 3603 _igvn.rehash_node_delayed(use); 3604 for (uint j = 1; j < use->req(); j++) { 3605 if (use->in(j) == phi) 3606 use->set_req(j, post); 3607 } 3608 // Since DU info changed, rerun loop 3609 progress = true; 3610 break; 3611 } 3612 } 3613 3614 }