1 /* 2 * Copyright (c) 1999, 2017, 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 "memory/allocation.inline.hpp" 27 #include "memory/resourceArea.hpp" 28 #include "opto/addnode.hpp" 29 #include "opto/castnode.hpp" 30 #include "opto/connode.hpp" 31 #include "opto/castnode.hpp" 32 #include "opto/divnode.hpp" 33 #include "opto/loopnode.hpp" 34 #include "opto/matcher.hpp" 35 #include "opto/mulnode.hpp" 36 #include "opto/movenode.hpp" 37 #include "opto/opaquenode.hpp" 38 #include "opto/rootnode.hpp" 39 #include "opto/subnode.hpp" 40 #include "opto/valuetypenode.hpp" 41 42 //============================================================================= 43 //------------------------------split_thru_phi--------------------------------- 44 // Split Node 'n' through merge point if there is enough win. 45 Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) { 46 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) { 47 // ConvI2L may have type information on it which is unsafe to push up 48 // so disable this for now 49 return NULL; 50 } 51 52 // Splitting range check CastIIs through a loop induction Phi can 53 // cause new Phis to be created that are left unrelated to the loop 54 // induction Phi and prevent optimizations (vectorization) 55 if (n->Opcode() == Op_CastII && n->as_CastII()->has_range_check() && 56 region->is_CountedLoop() && n->in(1) == region->as_CountedLoop()->phi()) { 57 return NULL; 58 } 59 60 // Value types should not be split through Phis because they cannot be merged 61 // through Phi nodes but each value input needs to be merged individually. 62 if (n->is_ValueType()) { 63 return NULL; 64 } 65 66 int wins = 0; 67 assert(!n->is_CFG(), ""); 68 assert(region->is_Region(), ""); 69 70 const Type* type = n->bottom_type(); 71 const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr(); 72 Node *phi; 73 if (t_oop != NULL && t_oop->is_known_instance_field()) { 74 int iid = t_oop->instance_id(); 75 int index = C->get_alias_index(t_oop); 76 int offset = t_oop->offset(); 77 phi = new PhiNode(region, type, NULL, iid, index, offset); 78 } else { 79 phi = PhiNode::make_blank(region, n); 80 } 81 uint old_unique = C->unique(); 82 for (uint i = 1; i < region->req(); i++) { 83 Node *x; 84 Node* the_clone = NULL; 85 if (region->in(i) == C->top()) { 86 x = C->top(); // Dead path? Use a dead data op 87 } else { 88 x = n->clone(); // Else clone up the data op 89 the_clone = x; // Remember for possible deletion. 90 // Alter data node to use pre-phi inputs 91 if (n->in(0) == region) 92 x->set_req( 0, region->in(i) ); 93 for (uint j = 1; j < n->req(); j++) { 94 Node *in = n->in(j); 95 if (in->is_Phi() && in->in(0) == region) 96 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone 97 } 98 } 99 // Check for a 'win' on some paths 100 const Type *t = x->Value(&_igvn); 101 102 bool singleton = t->singleton(); 103 104 // A TOP singleton indicates that there are no possible values incoming 105 // along a particular edge. In most cases, this is OK, and the Phi will 106 // be eliminated later in an Ideal call. However, we can't allow this to 107 // happen if the singleton occurs on loop entry, as the elimination of 108 // the PhiNode may cause the resulting node to migrate back to a previous 109 // loop iteration. 110 if (singleton && t == Type::TOP) { 111 // Is_Loop() == false does not confirm the absence of a loop (e.g., an 112 // irreducible loop may not be indicated by an affirmative is_Loop()); 113 // therefore, the only top we can split thru a phi is on a backedge of 114 // a loop. 115 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 116 } 117 118 if (singleton) { 119 wins++; 120 x = ((PhaseGVN&)_igvn).makecon(t); 121 } else { 122 // We now call Identity to try to simplify the cloned node. 123 // Note that some Identity methods call phase->type(this). 124 // Make sure that the type array is big enough for 125 // our new node, even though we may throw the node away. 126 // (Note: This tweaking with igvn only works because x is a new node.) 127 _igvn.set_type(x, t); 128 // If x is a TypeNode, capture any more-precise type permanently into Node 129 // otherwise it will be not updated during igvn->transform since 130 // igvn->type(x) is set to x->Value() already. 131 x->raise_bottom_type(t); 132 Node *y = x->Identity(&_igvn); 133 if (y != x) { 134 wins++; 135 x = y; 136 } else { 137 y = _igvn.hash_find(x); 138 if (y) { 139 wins++; 140 x = y; 141 } else { 142 // Else x is a new node we are keeping 143 // We do not need register_new_node_with_optimizer 144 // because set_type has already been called. 145 _igvn._worklist.push(x); 146 } 147 } 148 } 149 if (x != the_clone && the_clone != NULL) 150 _igvn.remove_dead_node(the_clone); 151 phi->set_req( i, x ); 152 } 153 // Too few wins? 154 if (wins <= policy) { 155 _igvn.remove_dead_node(phi); 156 return NULL; 157 } 158 159 // Record Phi 160 register_new_node( phi, region ); 161 162 for (uint i2 = 1; i2 < phi->req(); i2++) { 163 Node *x = phi->in(i2); 164 // If we commoned up the cloned 'x' with another existing Node, 165 // the existing Node picks up a new use. We need to make the 166 // existing Node occur higher up so it dominates its uses. 167 Node *old_ctrl; 168 IdealLoopTree *old_loop; 169 170 if (x->is_Con()) { 171 // Constant's control is always root. 172 set_ctrl(x, C->root()); 173 continue; 174 } 175 // The occasional new node 176 if (x->_idx >= old_unique) { // Found a new, unplaced node? 177 old_ctrl = NULL; 178 old_loop = NULL; // Not in any prior loop 179 } else { 180 old_ctrl = get_ctrl(x); 181 old_loop = get_loop(old_ctrl); // Get prior loop 182 } 183 // New late point must dominate new use 184 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2)); 185 if (new_ctrl == old_ctrl) // Nothing is changed 186 continue; 187 188 IdealLoopTree *new_loop = get_loop(new_ctrl); 189 190 // Don't move x into a loop if its uses are 191 // outside of loop. Otherwise x will be cloned 192 // for each use outside of this loop. 193 IdealLoopTree *use_loop = get_loop(region); 194 if (!new_loop->is_member(use_loop) && 195 (old_loop == NULL || !new_loop->is_member(old_loop))) { 196 // Take early control, later control will be recalculated 197 // during next iteration of loop optimizations. 198 new_ctrl = get_early_ctrl(x); 199 new_loop = get_loop(new_ctrl); 200 } 201 // Set new location 202 set_ctrl(x, new_ctrl); 203 // If changing loop bodies, see if we need to collect into new body 204 if (old_loop != new_loop) { 205 if (old_loop && !old_loop->_child) 206 old_loop->_body.yank(x); 207 if (!new_loop->_child) 208 new_loop->_body.push(x); // Collect body info 209 } 210 } 211 212 return phi; 213 } 214 215 //------------------------------dominated_by------------------------------------ 216 // Replace the dominated test with an obvious true or false. Place it on the 217 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the 218 // live path up to the dominating control. 219 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff, bool flip, bool exclude_loop_predicate ) { 220 if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); } 221 222 // prevdom is the dominating projection of the dominating test. 223 assert( iff->is_If(), "" ); 224 assert(iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd || iff->Opcode() == Op_RangeCheck, "Check this code when new subtype is added"); 225 int pop = prevdom->Opcode(); 226 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" ); 227 if (flip) { 228 if (pop == Op_IfTrue) 229 pop = Op_IfFalse; 230 else 231 pop = Op_IfTrue; 232 } 233 // 'con' is set to true or false to kill the dominated test. 234 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO); 235 set_ctrl(con, C->root()); // Constant gets a new use 236 // Hack the dominated test 237 _igvn.replace_input_of(iff, 1, con); 238 239 // If I dont have a reachable TRUE and FALSE path following the IfNode then 240 // I can assume this path reaches an infinite loop. In this case it's not 241 // important to optimize the data Nodes - either the whole compilation will 242 // be tossed or this path (and all data Nodes) will go dead. 243 if (iff->outcnt() != 2) return; 244 245 // Make control-dependent data Nodes on the live path (path that will remain 246 // once the dominated IF is removed) become control-dependent on the 247 // dominating projection. 248 Node* dp = iff->as_If()->proj_out(pop == Op_IfTrue); 249 250 // Loop predicates may have depending checks which should not 251 // be skipped. For example, range check predicate has two checks 252 // for lower and upper bounds. 253 if (dp == NULL) 254 return; 255 256 ProjNode* dp_proj = dp->as_Proj(); 257 ProjNode* unc_proj = iff->as_If()->proj_out(1 - dp_proj->_con)->as_Proj(); 258 if (exclude_loop_predicate && 259 (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) != NULL || 260 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check) != NULL)) { 261 // If this is a range check (IfNode::is_range_check), do not 262 // reorder because Compile::allow_range_check_smearing might have 263 // changed the check. 264 return; // Let IGVN transformation change control dependence. 265 } 266 267 IdealLoopTree *old_loop = get_loop(dp); 268 269 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { 270 Node* cd = dp->fast_out(i); // Control-dependent node 271 if (cd->depends_only_on_test()) { 272 assert(cd->in(0) == dp, ""); 273 _igvn.replace_input_of(cd, 0, prevdom); 274 set_early_ctrl(cd); 275 IdealLoopTree *new_loop = get_loop(get_ctrl(cd)); 276 if (old_loop != new_loop) { 277 if (!old_loop->_child) old_loop->_body.yank(cd); 278 if (!new_loop->_child) new_loop->_body.push(cd); 279 } 280 --i; 281 --imax; 282 } 283 } 284 } 285 286 //------------------------------has_local_phi_input---------------------------- 287 // Return TRUE if 'n' has Phi inputs from its local block and no other 288 // block-local inputs (all non-local-phi inputs come from earlier blocks) 289 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) { 290 Node *n_ctrl = get_ctrl(n); 291 // See if some inputs come from a Phi in this block, or from before 292 // this block. 293 uint i; 294 for( i = 1; i < n->req(); i++ ) { 295 Node *phi = n->in(i); 296 if( phi->is_Phi() && phi->in(0) == n_ctrl ) 297 break; 298 } 299 if( i >= n->req() ) 300 return NULL; // No Phi inputs; nowhere to clone thru 301 302 // Check for inputs created between 'n' and the Phi input. These 303 // must split as well; they have already been given the chance 304 // (courtesy of a post-order visit) and since they did not we must 305 // recover the 'cost' of splitting them by being very profitable 306 // when splitting 'n'. Since this is unlikely we simply give up. 307 for( i = 1; i < n->req(); i++ ) { 308 Node *m = n->in(i); 309 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) { 310 // We allow the special case of AddP's with no local inputs. 311 // This allows us to split-up address expressions. 312 if (m->is_AddP() && 313 get_ctrl(m->in(2)) != n_ctrl && 314 get_ctrl(m->in(3)) != n_ctrl) { 315 // Move the AddP up to dominating point 316 set_ctrl_and_loop(m, find_non_split_ctrl(idom(n_ctrl))); 317 continue; 318 } 319 return NULL; 320 } 321 } 322 323 return n_ctrl; 324 } 325 326 //------------------------------remix_address_expressions---------------------- 327 // Rework addressing expressions to get the most loop-invariant stuff 328 // moved out. We'd like to do all associative operators, but it's especially 329 // important (common) to do address expressions. 330 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) { 331 if (!has_ctrl(n)) return NULL; 332 Node *n_ctrl = get_ctrl(n); 333 IdealLoopTree *n_loop = get_loop(n_ctrl); 334 335 // See if 'n' mixes loop-varying and loop-invariant inputs and 336 // itself is loop-varying. 337 338 // Only interested in binary ops (and AddP) 339 if( n->req() < 3 || n->req() > 4 ) return NULL; 340 341 Node *n1_ctrl = get_ctrl(n->in( 1)); 342 Node *n2_ctrl = get_ctrl(n->in( 2)); 343 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3)); 344 IdealLoopTree *n1_loop = get_loop( n1_ctrl ); 345 IdealLoopTree *n2_loop = get_loop( n2_ctrl ); 346 IdealLoopTree *n3_loop = get_loop( n3_ctrl ); 347 348 // Does one of my inputs spin in a tighter loop than self? 349 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) || 350 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) || 351 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) ) 352 return NULL; // Leave well enough alone 353 354 // Is at least one of my inputs loop-invariant? 355 if( n1_loop == n_loop && 356 n2_loop == n_loop && 357 n3_loop == n_loop ) 358 return NULL; // No loop-invariant inputs 359 360 361 int n_op = n->Opcode(); 362 363 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2). 364 if( n_op == Op_LShiftI ) { 365 // Scale is loop invariant 366 Node *scale = n->in(2); 367 Node *scale_ctrl = get_ctrl(scale); 368 IdealLoopTree *scale_loop = get_loop(scale_ctrl ); 369 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) ) 370 return NULL; 371 const TypeInt *scale_t = scale->bottom_type()->isa_int(); 372 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 ) 373 return NULL; // Dont bother with byte/short masking 374 // Add must vary with loop (else shift would be loop-invariant) 375 Node *add = n->in(1); 376 Node *add_ctrl = get_ctrl(add); 377 IdealLoopTree *add_loop = get_loop(add_ctrl); 378 //assert( n_loop == add_loop, "" ); 379 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops 380 381 // Convert I-V into I+ (0-V); same for V-I 382 if( add->Opcode() == Op_SubI && 383 _igvn.type( add->in(1) ) != TypeInt::ZERO ) { 384 Node *zero = _igvn.intcon(0); 385 set_ctrl(zero, C->root()); 386 Node *neg = new SubINode( _igvn.intcon(0), add->in(2) ); 387 register_new_node( neg, get_ctrl(add->in(2) ) ); 388 add = new AddINode( add->in(1), neg ); 389 register_new_node( add, add_ctrl ); 390 } 391 if( add->Opcode() != Op_AddI ) return NULL; 392 // See if one add input is loop invariant 393 Node *add_var = add->in(1); 394 Node *add_var_ctrl = get_ctrl(add_var); 395 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 396 Node *add_invar = add->in(2); 397 Node *add_invar_ctrl = get_ctrl(add_invar); 398 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl ); 399 if( add_var_loop == n_loop ) { 400 } else if( add_invar_loop == n_loop ) { 401 // Swap to find the invariant part 402 add_invar = add_var; 403 add_invar_ctrl = add_var_ctrl; 404 add_invar_loop = add_var_loop; 405 add_var = add->in(2); 406 Node *add_var_ctrl = get_ctrl(add_var); 407 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 408 } else // Else neither input is loop invariant 409 return NULL; 410 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) ) 411 return NULL; // No invariant part of the add? 412 413 // Yes! Reshape address expression! 414 Node *inv_scale = new LShiftINode( add_invar, scale ); 415 Node *inv_scale_ctrl = 416 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ? 417 add_invar_ctrl : scale_ctrl; 418 register_new_node( inv_scale, inv_scale_ctrl ); 419 Node *var_scale = new LShiftINode( add_var, scale ); 420 register_new_node( var_scale, n_ctrl ); 421 Node *var_add = new AddINode( var_scale, inv_scale ); 422 register_new_node( var_add, n_ctrl ); 423 _igvn.replace_node( n, var_add ); 424 return var_add; 425 } 426 427 // Replace (I+V) with (V+I) 428 if( n_op == Op_AddI || 429 n_op == Op_AddL || 430 n_op == Op_AddF || 431 n_op == Op_AddD || 432 n_op == Op_MulI || 433 n_op == Op_MulL || 434 n_op == Op_MulF || 435 n_op == Op_MulD ) { 436 if( n2_loop == n_loop ) { 437 assert( n1_loop != n_loop, "" ); 438 n->swap_edges(1, 2); 439 } 440 } 441 442 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V), 443 // but not if I2 is a constant. 444 if( n_op == Op_AddP ) { 445 if( n2_loop == n_loop && n3_loop != n_loop ) { 446 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) { 447 Node *n22_ctrl = get_ctrl(n->in(2)->in(2)); 448 Node *n23_ctrl = get_ctrl(n->in(2)->in(3)); 449 IdealLoopTree *n22loop = get_loop( n22_ctrl ); 450 IdealLoopTree *n23_loop = get_loop( n23_ctrl ); 451 if( n22loop != n_loop && n22loop->is_member(n_loop) && 452 n23_loop == n_loop ) { 453 Node *add1 = new AddPNode( n->in(1), n->in(2)->in(2), n->in(3) ); 454 // Stuff new AddP in the loop preheader 455 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) ); 456 Node *add2 = new AddPNode( n->in(1), add1, n->in(2)->in(3) ); 457 register_new_node( add2, n_ctrl ); 458 _igvn.replace_node( n, add2 ); 459 return add2; 460 } 461 } 462 } 463 464 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V) 465 if (n2_loop != n_loop && n3_loop == n_loop) { 466 if (n->in(3)->Opcode() == Op_AddX) { 467 Node *V = n->in(3)->in(1); 468 Node *I = n->in(3)->in(2); 469 if (is_member(n_loop,get_ctrl(V))) { 470 } else { 471 Node *tmp = V; V = I; I = tmp; 472 } 473 if (!is_member(n_loop,get_ctrl(I))) { 474 Node *add1 = new AddPNode(n->in(1), n->in(2), I); 475 // Stuff new AddP in the loop preheader 476 register_new_node(add1, n_loop->_head->in(LoopNode::EntryControl)); 477 Node *add2 = new AddPNode(n->in(1), add1, V); 478 register_new_node(add2, n_ctrl); 479 _igvn.replace_node(n, add2); 480 return add2; 481 } 482 } 483 } 484 } 485 486 return NULL; 487 } 488 489 //------------------------------conditional_move------------------------------- 490 // Attempt to replace a Phi with a conditional move. We have some pretty 491 // strict profitability requirements. All Phis at the merge point must 492 // be converted, so we can remove the control flow. We need to limit the 493 // number of c-moves to a small handful. All code that was in the side-arms 494 // of the CFG diamond is now speculatively executed. This code has to be 495 // "cheap enough". We are pretty much limited to CFG diamonds that merge 496 // 1 or 2 items with a total of 1 or 2 ops executed speculatively. 497 Node *PhaseIdealLoop::conditional_move( Node *region ) { 498 499 assert(region->is_Region(), "sanity check"); 500 if (region->req() != 3) return NULL; 501 502 // Check for CFG diamond 503 Node *lp = region->in(1); 504 Node *rp = region->in(2); 505 if (!lp || !rp) return NULL; 506 Node *lp_c = lp->in(0); 507 if (lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If()) return NULL; 508 IfNode *iff = lp_c->as_If(); 509 510 // Check for ops pinned in an arm of the diamond. 511 // Can't remove the control flow in this case 512 if (lp->outcnt() > 1) return NULL; 513 if (rp->outcnt() > 1) return NULL; 514 515 IdealLoopTree* r_loop = get_loop(region); 516 assert(r_loop == get_loop(iff), "sanity"); 517 // Always convert to CMOVE if all results are used only outside this loop. 518 bool used_inside_loop = (r_loop == _ltree_root); 519 520 // Check profitability 521 int cost = 0; 522 int phis = 0; 523 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 524 Node *out = region->fast_out(i); 525 if (!out->is_Phi()) continue; // Ignore other control edges, etc 526 phis++; 527 PhiNode* phi = out->as_Phi(); 528 BasicType bt = phi->type()->basic_type(); 529 switch (bt) { 530 case T_DOUBLE: 531 if (C->use_cmove()) { 532 continue; //TODO: maybe we want to add some cost 533 } 534 case T_FLOAT: { 535 cost += Matcher::float_cmove_cost(); // Could be very expensive 536 break; 537 } 538 case T_LONG: { 539 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's 540 } 541 case T_INT: // These all CMOV fine 542 case T_ADDRESS: { // (RawPtr) 543 cost++; 544 break; 545 } 546 case T_NARROWOOP: // Fall through 547 case T_OBJECT: { // Base oops are OK, but not derived oops 548 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr(); 549 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a 550 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus 551 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we 552 // have a Phi for the base here that we convert to a CMOVE all is well 553 // and good. But if the base is dead, we'll not make a CMOVE. Later 554 // the allocator will have to produce a base by creating a CMOVE of the 555 // relevant bases. This puts the allocator in the business of 556 // manufacturing expensive instructions, generally a bad plan. 557 // Just Say No to Conditionally-Moved Derived Pointers. 558 if (tp && tp->offset() != 0) 559 return NULL; 560 cost++; 561 break; 562 } 563 default: 564 return NULL; // In particular, can't do memory or I/O 565 } 566 // Add in cost any speculative ops 567 for (uint j = 1; j < region->req(); j++) { 568 Node *proj = region->in(j); 569 Node *inp = phi->in(j); 570 if (get_ctrl(inp) == proj) { // Found local op 571 cost++; 572 // Check for a chain of dependent ops; these will all become 573 // speculative in a CMOV. 574 for (uint k = 1; k < inp->req(); k++) 575 if (get_ctrl(inp->in(k)) == proj) 576 cost += ConditionalMoveLimit; // Too much speculative goo 577 } 578 } 579 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode. 580 // This will likely Split-If, a higher-payoff operation. 581 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) { 582 Node* use = phi->fast_out(k); 583 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr()) 584 cost += ConditionalMoveLimit; 585 // Is there a use inside the loop? 586 // Note: check only basic types since CMoveP is pinned. 587 if (!used_inside_loop && is_java_primitive(bt)) { 588 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use); 589 if (r_loop == u_loop || r_loop->is_member(u_loop)) { 590 used_inside_loop = true; 591 } 592 } 593 } 594 }//for 595 Node* bol = iff->in(1); 596 assert(bol->Opcode() == Op_Bool, ""); 597 int cmp_op = bol->in(1)->Opcode(); 598 // It is expensive to generate flags from a float compare. 599 // Avoid duplicated float compare. 600 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL; 601 602 float infrequent_prob = PROB_UNLIKELY_MAG(3); 603 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop. 604 if (used_inside_loop) { 605 if (cost >= ConditionalMoveLimit) return NULL; // Too much goo 606 607 // BlockLayoutByFrequency optimization moves infrequent branch 608 // from hot path. No point in CMOV'ing in such case (110 is used 609 // instead of 100 to take into account not exactness of float value). 610 if (BlockLayoutByFrequency) { 611 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f); 612 } 613 } 614 // Check for highly predictable branch. No point in CMOV'ing if 615 // we are going to predict accurately all the time. 616 if (C->use_cmove() && cmp_op == Op_CmpD) ;//keep going 617 else if (iff->_prob < infrequent_prob || 618 iff->_prob > (1.0f - infrequent_prob)) 619 return NULL; 620 621 // -------------- 622 // Now replace all Phis with CMOV's 623 Node *cmov_ctrl = iff->in(0); 624 uint flip = (lp->Opcode() == Op_IfTrue); 625 while (1) { 626 PhiNode* phi = NULL; 627 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 628 Node *out = region->fast_out(i); 629 if (out->is_Phi()) { 630 phi = out->as_Phi(); 631 break; 632 } 633 } 634 if (phi == NULL) break; 635 if (PrintOpto && VerifyLoopOptimizations) { tty->print_cr("CMOV"); } 636 // Move speculative ops 637 for (uint j = 1; j < region->req(); j++) { 638 Node *proj = region->in(j); 639 Node *inp = phi->in(j); 640 if (get_ctrl(inp) == proj) { // Found local op 641 #ifndef PRODUCT 642 if (PrintOpto && VerifyLoopOptimizations) { 643 tty->print(" speculate: "); 644 inp->dump(); 645 } 646 #endif 647 set_ctrl(inp, cmov_ctrl); 648 } 649 } 650 Node *cmov = CMoveNode::make(cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi)); 651 register_new_node( cmov, cmov_ctrl ); 652 _igvn.replace_node( phi, cmov ); 653 #ifndef PRODUCT 654 if (TraceLoopOpts) { 655 tty->print("CMOV "); 656 r_loop->dump_head(); 657 if (Verbose) { 658 bol->in(1)->dump(1); 659 cmov->dump(1); 660 } 661 } 662 if (VerifyLoopOptimizations) verify(); 663 #endif 664 } 665 666 // The useless CFG diamond will fold up later; see the optimization in 667 // RegionNode::Ideal. 668 _igvn._worklist.push(region); 669 670 return iff->in(1); 671 } 672 673 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) { 674 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 675 Node* u = m->fast_out(i); 676 if (u->is_CFG()) { 677 if (u->Opcode() == Op_NeverBranch) { 678 u = ((NeverBranchNode*)u)->proj_out(0); 679 enqueue_cfg_uses(u, wq); 680 } else { 681 wq.push(u); 682 } 683 } 684 } 685 } 686 687 // Try moving a store out of a loop, right before the loop 688 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) { 689 // Store has to be first in the loop body 690 IdealLoopTree *n_loop = get_loop(n_ctrl); 691 if (n->is_Store() && n_loop != _ltree_root && n_loop->is_loop() && n->in(0) != NULL) { 692 Node* address = n->in(MemNode::Address); 693 Node* value = n->in(MemNode::ValueIn); 694 Node* mem = n->in(MemNode::Memory); 695 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 696 IdealLoopTree* value_loop = get_loop(get_ctrl(value)); 697 698 // - address and value must be loop invariant 699 // - memory must be a memory Phi for the loop 700 // - Store must be the only store on this memory slice in the 701 // loop: if there's another store following this one then value 702 // written at iteration i by the second store could be overwritten 703 // at iteration i+n by the first store: it's not safe to move the 704 // first store out of the loop 705 // - nothing must observe the memory Phi: it guarantees no read 706 // before the store, we are also guaranteed the store post 707 // dominates the loop head (ignoring a possible early 708 // exit). Otherwise there would be extra Phi involved between the 709 // loop's Phi and the store. 710 // - there must be no early exit from the loop before the Store 711 // (such an exit most of the time would be an extra use of the 712 // memory Phi but sometimes is a bottom memory Phi that takes the 713 // store as input). 714 715 if (!n_loop->is_member(address_loop) && 716 !n_loop->is_member(value_loop) && 717 mem->is_Phi() && mem->in(0) == n_loop->_head && 718 mem->outcnt() == 1 && 719 mem->in(LoopNode::LoopBackControl) == n) { 720 721 assert(n_loop->_tail != NULL, "need a tail"); 722 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop"); 723 724 // Verify that there's no early exit of the loop before the store. 725 bool ctrl_ok = false; 726 { 727 // Follow control from loop head until n, we exit the loop or 728 // we reach the tail 729 ResourceMark rm; 730 Unique_Node_List wq; 731 wq.push(n_loop->_head); 732 733 for (uint next = 0; next < wq.size(); ++next) { 734 Node *m = wq.at(next); 735 if (m == n->in(0)) { 736 ctrl_ok = true; 737 continue; 738 } 739 assert(!has_ctrl(m), "should be CFG"); 740 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) { 741 ctrl_ok = false; 742 break; 743 } 744 enqueue_cfg_uses(m, wq); 745 if (wq.size() > 10) { 746 ctrl_ok = false; 747 break; 748 } 749 } 750 } 751 if (ctrl_ok) { 752 // move the Store 753 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem); 754 _igvn.replace_input_of(n, 0, n_loop->_head->in(LoopNode::EntryControl)); 755 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl)); 756 // Disconnect the phi now. An empty phi can confuse other 757 // optimizations in this pass of loop opts. 758 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl)); 759 n_loop->_body.yank(mem); 760 761 IdealLoopTree* new_loop = get_loop(n->in(0)); 762 set_ctrl_and_loop(n, n->in(0)); 763 764 return n; 765 } 766 } 767 } 768 return NULL; 769 } 770 771 // Try moving a store out of a loop, right after the loop 772 void PhaseIdealLoop::try_move_store_after_loop(Node* n) { 773 if (n->is_Store() && n->in(0) != NULL) { 774 Node *n_ctrl = get_ctrl(n); 775 IdealLoopTree *n_loop = get_loop(n_ctrl); 776 // Store must be in a loop 777 if (n_loop != _ltree_root && !n_loop->_irreducible) { 778 Node* address = n->in(MemNode::Address); 779 Node* value = n->in(MemNode::ValueIn); 780 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 781 // address must be loop invariant 782 if (!n_loop->is_member(address_loop)) { 783 // Store must be last on this memory slice in the loop and 784 // nothing in the loop must observe it 785 Node* phi = NULL; 786 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 787 Node* u = n->fast_out(i); 788 if (has_ctrl(u)) { // control use? 789 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 790 if (!n_loop->is_member(u_loop)) { 791 continue; 792 } 793 if (u->is_Phi() && u->in(0) == n_loop->_head) { 794 assert(_igvn.type(u) == Type::MEMORY, "bad phi"); 795 // multiple phis on the same slice are possible 796 if (phi != NULL) { 797 return; 798 } 799 phi = u; 800 continue; 801 } 802 } 803 return; 804 } 805 if (phi != NULL) { 806 // Nothing in the loop before the store (next iteration) 807 // must observe the stored value 808 bool mem_ok = true; 809 { 810 ResourceMark rm; 811 Unique_Node_List wq; 812 wq.push(phi); 813 for (uint next = 0; next < wq.size() && mem_ok; ++next) { 814 Node *m = wq.at(next); 815 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) { 816 Node* u = m->fast_out(i); 817 if (u->is_Store() || u->is_Phi()) { 818 if (u != n) { 819 wq.push(u); 820 mem_ok = (wq.size() <= 10); 821 } 822 } else { 823 mem_ok = false; 824 break; 825 } 826 } 827 } 828 } 829 if (mem_ok) { 830 // Move the Store out of the loop creating clones along 831 // all paths out of the loop that observe the stored value 832 _igvn.rehash_node_delayed(phi); 833 int count = phi->replace_edge(n, n->in(MemNode::Memory)); 834 assert(count > 0, "inconsistent phi"); 835 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 836 Node* u = n->fast_out(i); 837 Node* c = get_ctrl(u); 838 839 if (u->is_Phi()) { 840 c = u->in(0)->in(u->find_edge(n)); 841 } 842 IdealLoopTree *u_loop = get_loop(c); 843 assert (!n_loop->is_member(u_loop), "only the phi should have been a use in the loop"); 844 while(true) { 845 Node* next_c = find_non_split_ctrl(idom(c)); 846 if (n_loop->is_member(get_loop(next_c))) { 847 break; 848 } 849 c = next_c; 850 } 851 852 Node* st = n->clone(); 853 st->set_req(0, c); 854 _igvn.register_new_node_with_optimizer(st); 855 856 set_ctrl(st, c); 857 IdealLoopTree* new_loop = get_loop(c); 858 assert(new_loop != n_loop, "should be moved out of loop"); 859 if (new_loop->_child == NULL) new_loop->_body.push(st); 860 861 _igvn.replace_input_of(u, u->find_edge(n), st); 862 --imax; 863 --i; 864 } 865 866 867 assert(n->outcnt() == 0, "all uses should be gone"); 868 _igvn.replace_input_of(n, MemNode::Memory, C->top()); 869 // Disconnect the phi now. An empty phi can confuse other 870 // optimizations in this pass of loop opts.. 871 if (phi->in(LoopNode::LoopBackControl) == phi) { 872 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl)); 873 n_loop->_body.yank(phi); 874 } 875 } 876 } 877 } 878 } 879 } 880 } 881 882 //------------------------------split_if_with_blocks_pre----------------------- 883 // Do the real work in a non-recursive function. Data nodes want to be 884 // cloned in the pre-order so they can feed each other nicely. 885 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) { 886 // Cloning these guys is unlikely to win 887 int n_op = n->Opcode(); 888 if( n_op == Op_MergeMem ) return n; 889 if( n->is_Proj() ) return n; 890 // Do not clone-up CmpFXXX variations, as these are always 891 // followed by a CmpI 892 if( n->is_Cmp() ) return n; 893 // Attempt to use a conditional move instead of a phi/branch 894 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) { 895 Node *cmov = conditional_move( n ); 896 if( cmov ) return cmov; 897 } 898 if( n->is_CFG() || n->is_LoadStore() ) 899 return n; 900 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd 901 n_op == Op_Opaque2 ) { 902 if( !C->major_progress() ) // If chance of no more loop opts... 903 _igvn._worklist.push(n); // maybe we'll remove them 904 return n; 905 } 906 907 if( n->is_Con() ) return n; // No cloning for Con nodes 908 909 Node *n_ctrl = get_ctrl(n); 910 if( !n_ctrl ) return n; // Dead node 911 912 Node* res = try_move_store_before_loop(n, n_ctrl); 913 if (res != NULL) { 914 return n; 915 } 916 917 // Attempt to remix address expressions for loop invariants 918 Node *m = remix_address_expressions( n ); 919 if( m ) return m; 920 921 if (n->is_ConstraintCast()) { 922 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this); 923 // ConstraintCastNode::dominating_cast() uses node control input to determine domination. 924 // Node control inputs don't necessarily agree with loop control info (due to 925 // transformations happened in between), thus additional dominance check is needed 926 // to keep loop info valid. 927 if (dom_cast != NULL && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) { 928 _igvn.replace_node(n, dom_cast); 929 return dom_cast; 930 } 931 } 932 933 // Determine if the Node has inputs from some local Phi. 934 // Returns the block to clone thru. 935 Node *n_blk = has_local_phi_input( n ); 936 if( !n_blk ) return n; 937 938 // Do not clone the trip counter through on a CountedLoop 939 // (messes up the canonical shape). 940 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n; 941 942 // Check for having no control input; not pinned. Allow 943 // dominating control. 944 if (n->in(0)) { 945 Node *dom = idom(n_blk); 946 if (dom_lca(n->in(0), dom) != n->in(0)) { 947 return n; 948 } 949 } 950 // Policy: when is it profitable. You must get more wins than 951 // policy before it is considered profitable. Policy is usually 0, 952 // so 1 win is considered profitable. Big merges will require big 953 // cloning, so get a larger policy. 954 int policy = n_blk->req() >> 2; 955 956 // If the loop is a candidate for range check elimination, 957 // delay splitting through it's phi until a later loop optimization 958 if (n_blk->is_CountedLoop()) { 959 IdealLoopTree *lp = get_loop(n_blk); 960 if (lp && lp->_rce_candidate) { 961 return n; 962 } 963 } 964 965 // Use same limit as split_if_with_blocks_post 966 if( C->live_nodes() > 35000 ) return n; // Method too big 967 968 // Split 'n' through the merge point if it is profitable 969 Node *phi = split_thru_phi( n, n_blk, policy ); 970 if (!phi) return n; 971 972 // Found a Phi to split thru! 973 // Replace 'n' with the new phi 974 _igvn.replace_node( n, phi ); 975 // Moved a load around the loop, 'en-registering' something. 976 if (n_blk->is_Loop() && n->is_Load() && 977 !phi->in(LoopNode::LoopBackControl)->is_Load()) 978 C->set_major_progress(); 979 980 return phi; 981 } 982 983 static bool merge_point_too_heavy(Compile* C, Node* region) { 984 // Bail out if the region and its phis have too many users. 985 int weight = 0; 986 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 987 weight += region->fast_out(i)->outcnt(); 988 } 989 int nodes_left = C->max_node_limit() - C->live_nodes(); 990 if (weight * 8 > nodes_left) { 991 if (PrintOpto) { 992 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight); 993 } 994 return true; 995 } else { 996 return false; 997 } 998 } 999 1000 static bool merge_point_safe(Node* region) { 1001 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode 1002 // having a PhiNode input. This sidesteps the dangerous case where the split 1003 // ConvI2LNode may become TOP if the input Value() does not 1004 // overlap the ConvI2L range, leaving a node which may not dominate its 1005 // uses. 1006 // A better fix for this problem can be found in the BugTraq entry, but 1007 // expediency for Mantis demands this hack. 1008 // 6855164: If the merge point has a FastLockNode with a PhiNode input, we stop 1009 // split_if_with_blocks from splitting a block because we could not move around 1010 // the FastLockNode. 1011 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1012 Node* n = region->fast_out(i); 1013 if (n->is_Phi()) { 1014 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1015 Node* m = n->fast_out(j); 1016 if (m->is_FastLock()) 1017 return false; 1018 if (m->is_ValueType()) { 1019 // TODO this breaks optimizations! 1020 // Value types should not be split through phis 1021 //return false; 1022 } 1023 #ifdef _LP64 1024 if (m->Opcode() == Op_ConvI2L) 1025 return false; 1026 if (m->is_CastII() && m->isa_CastII()->has_range_check()) { 1027 return false; 1028 } 1029 #endif 1030 } 1031 } 1032 } 1033 return true; 1034 } 1035 1036 1037 //------------------------------place_near_use--------------------------------- 1038 // Place some computation next to use but not inside inner loops. 1039 // For inner loop uses move it to the preheader area. 1040 Node *PhaseIdealLoop::place_near_use( Node *useblock ) const { 1041 IdealLoopTree *u_loop = get_loop( useblock ); 1042 return (u_loop->_irreducible || u_loop->_child) 1043 ? useblock 1044 : u_loop->_head->in(LoopNode::EntryControl); 1045 } 1046 1047 1048 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) { 1049 if (!n->is_If()) { 1050 return false; 1051 } 1052 if (!n->in(0)->is_Region()) { 1053 return false; 1054 } 1055 Node* region = n->in(0); 1056 Node* dom = idom(region); 1057 if (!dom->is_If() || dom->in(1) != n->in(1)) { 1058 return false; 1059 } 1060 IfNode* dom_if = dom->as_If(); 1061 Node* proj_true = dom_if->proj_out(1); 1062 Node* proj_false = dom_if->proj_out(0); 1063 1064 for (uint i = 1; i < region->req(); i++) { 1065 if (is_dominator(proj_true, region->in(i))) { 1066 continue; 1067 } 1068 if (is_dominator(proj_false, region->in(i))) { 1069 continue; 1070 } 1071 return false; 1072 } 1073 1074 return true; 1075 } 1076 1077 bool PhaseIdealLoop::can_split_if(Node *n_ctrl) { 1078 if (C->live_nodes() > 35000) { 1079 return false; // Method too big 1080 } 1081 1082 // Do not do 'split-if' if irreducible loops are present. 1083 if (_has_irreducible_loops) { 1084 return false; 1085 } 1086 1087 if (merge_point_too_heavy(C, n_ctrl)) { 1088 return false; 1089 } 1090 1091 // Do not do 'split-if' if some paths are dead. First do dead code 1092 // elimination and then see if its still profitable. 1093 for (uint i = 1; i < n_ctrl->req(); i++) { 1094 if (n_ctrl->in(i) == C->top()) { 1095 return false; 1096 } 1097 } 1098 1099 // If trying to do a 'Split-If' at the loop head, it is only 1100 // profitable if the cmp folds up on BOTH paths. Otherwise we 1101 // risk peeling a loop forever. 1102 1103 // CNC - Disabled for now. Requires careful handling of loop 1104 // body selection for the cloned code. Also, make sure we check 1105 // for any input path not being in the same loop as n_ctrl. For 1106 // irreducible loops we cannot check for 'n_ctrl->is_Loop()' 1107 // because the alternative loop entry points won't be converted 1108 // into LoopNodes. 1109 IdealLoopTree *n_loop = get_loop(n_ctrl); 1110 for (uint j = 1; j < n_ctrl->req(); j++) { 1111 if (get_loop(n_ctrl->in(j)) != n_loop) { 1112 return false; 1113 } 1114 } 1115 1116 // Check for safety of the merge point. 1117 if (!merge_point_safe(n_ctrl)) { 1118 return false; 1119 } 1120 1121 return true; 1122 } 1123 1124 //------------------------------split_if_with_blocks_post---------------------- 1125 // Do the real work in a non-recursive function. CFG hackery wants to be 1126 // in the post-order, so it can dirty the I-DOM info and not use the dirtied 1127 // info. 1128 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) { 1129 1130 // Cloning Cmp through Phi's involves the split-if transform. 1131 // FastLock is not used by an If 1132 if (n->is_Cmp() && !n->is_FastLock()) { 1133 Node *n_ctrl = get_ctrl(n); 1134 // Determine if the Node has inputs from some local Phi. 1135 // Returns the block to clone thru. 1136 Node *n_blk = has_local_phi_input(n); 1137 if (n_blk != n_ctrl) { 1138 return; 1139 } 1140 1141 if (!can_split_if(n_ctrl)) { 1142 return; 1143 } 1144 1145 if (n->outcnt() != 1) { 1146 return; // Multiple bool's from 1 compare? 1147 } 1148 Node *bol = n->unique_out(); 1149 assert(bol->is_Bool(), "expect a bool here"); 1150 if (bol->outcnt() != 1) { 1151 return;// Multiple branches from 1 compare? 1152 } 1153 Node *iff = bol->unique_out(); 1154 1155 // Check some safety conditions 1156 if (iff->is_If()) { // Classic split-if? 1157 if (iff->in(0) != n_ctrl) { 1158 return; // Compare must be in same blk as if 1159 } 1160 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE 1161 // Can't split CMove with different control edge. 1162 if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) { 1163 return; 1164 } 1165 if (get_ctrl(iff->in(2)) == n_ctrl || 1166 get_ctrl(iff->in(3)) == n_ctrl) { 1167 return; // Inputs not yet split-up 1168 } 1169 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) { 1170 return; // Loop-invar test gates loop-varying CMOVE 1171 } 1172 } else { 1173 return; // some other kind of node, such as an Allocate 1174 } 1175 1176 // When is split-if profitable? Every 'win' on means some control flow 1177 // goes dead, so it's almost always a win. 1178 int policy = 0; 1179 // Split compare 'n' through the merge point if it is profitable 1180 Node *phi = split_thru_phi( n, n_ctrl, policy); 1181 if (!phi) { 1182 return; 1183 } 1184 1185 // Found a Phi to split thru! 1186 // Replace 'n' with the new phi 1187 _igvn.replace_node(n, phi); 1188 1189 // Now split the bool up thru the phi 1190 Node *bolphi = split_thru_phi(bol, n_ctrl, -1); 1191 guarantee(bolphi != NULL, "null boolean phi node"); 1192 1193 _igvn.replace_node(bol, bolphi); 1194 assert(iff->in(1) == bolphi, ""); 1195 1196 if (bolphi->Value(&_igvn)->singleton()) { 1197 return; 1198 } 1199 1200 // Conditional-move? Must split up now 1201 if (!iff->is_If()) { 1202 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1); 1203 _igvn.replace_node(iff, cmovphi); 1204 return; 1205 } 1206 1207 // Now split the IF 1208 do_split_if(iff); 1209 return; 1210 } 1211 1212 // Two identical ifs back to back can be merged 1213 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) { 1214 Node *n_ctrl = n->in(0); 1215 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1)); 1216 IfNode* dom_if = idom(n_ctrl)->as_If(); 1217 Node* proj_true = dom_if->proj_out(1); 1218 Node* proj_false = dom_if->proj_out(0); 1219 Node* con_true = _igvn.makecon(TypeInt::ONE); 1220 Node* con_false = _igvn.makecon(TypeInt::ZERO); 1221 1222 for (uint i = 1; i < n_ctrl->req(); i++) { 1223 if (is_dominator(proj_true, n_ctrl->in(i))) { 1224 bolphi->init_req(i, con_true); 1225 } else { 1226 assert(is_dominator(proj_false, n_ctrl->in(i)), "bad if"); 1227 bolphi->init_req(i, con_false); 1228 } 1229 } 1230 register_new_node(bolphi, n_ctrl); 1231 _igvn.replace_input_of(n, 1, bolphi); 1232 1233 // Now split the IF 1234 do_split_if(n); 1235 return; 1236 } 1237 1238 // Check for an IF ready to split; one that has its 1239 // condition codes input coming from a Phi at the block start. 1240 int n_op = n->Opcode(); 1241 1242 // Check for an IF being dominated by another IF same test 1243 if (n_op == Op_If || 1244 n_op == Op_RangeCheck) { 1245 Node *bol = n->in(1); 1246 uint max = bol->outcnt(); 1247 // Check for same test used more than once? 1248 if (max > 1 && bol->is_Bool()) { 1249 // Search up IDOMs to see if this IF is dominated. 1250 Node *cutoff = get_ctrl(bol); 1251 1252 // Now search up IDOMs till cutoff, looking for a dominating test 1253 Node *prevdom = n; 1254 Node *dom = idom(prevdom); 1255 while (dom != cutoff) { 1256 if (dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom) { 1257 // Replace the dominated test with an obvious true or false. 1258 // Place it on the IGVN worklist for later cleanup. 1259 C->set_major_progress(); 1260 dominated_by(prevdom, n, false, true); 1261 #ifndef PRODUCT 1262 if( VerifyLoopOptimizations ) verify(); 1263 #endif 1264 return; 1265 } 1266 prevdom = dom; 1267 dom = idom(prevdom); 1268 } 1269 } 1270 } 1271 1272 // See if a shared loop-varying computation has no loop-varying uses. 1273 // Happens if something is only used for JVM state in uncommon trap exits, 1274 // like various versions of induction variable+offset. Clone the 1275 // computation per usage to allow it to sink out of the loop. 1276 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about) 1277 Node *n_ctrl = get_ctrl(n); 1278 IdealLoopTree *n_loop = get_loop(n_ctrl); 1279 if( n_loop != _ltree_root ) { 1280 DUIterator_Fast imax, i = n->fast_outs(imax); 1281 for (; i < imax; i++) { 1282 Node* u = n->fast_out(i); 1283 if( !has_ctrl(u) ) break; // Found control user 1284 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 1285 if( u_loop == n_loop ) break; // Found loop-varying use 1286 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop 1287 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003 1288 } 1289 bool did_break = (i < imax); // Did we break out of the previous loop? 1290 if (!did_break && n->outcnt() > 1) { // All uses in outer loops! 1291 Node *late_load_ctrl = NULL; 1292 if (n->is_Load()) { 1293 // If n is a load, get and save the result from get_late_ctrl(), 1294 // to be later used in calculating the control for n's clones. 1295 clear_dom_lca_tags(); 1296 late_load_ctrl = get_late_ctrl(n, n_ctrl); 1297 } 1298 // If n is a load, and the late control is the same as the current 1299 // control, then the cloning of n is a pointless exercise, because 1300 // GVN will ensure that we end up where we started. 1301 if (!n->is_Load() || late_load_ctrl != n_ctrl) { 1302 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) { 1303 Node *u = n->last_out(j); // Clone private computation per use 1304 _igvn.rehash_node_delayed(u); 1305 Node *x = n->clone(); // Clone computation 1306 Node *x_ctrl = NULL; 1307 if( u->is_Phi() ) { 1308 // Replace all uses of normal nodes. Replace Phi uses 1309 // individually, so the separate Nodes can sink down 1310 // different paths. 1311 uint k = 1; 1312 while( u->in(k) != n ) k++; 1313 u->set_req( k, x ); 1314 // x goes next to Phi input path 1315 x_ctrl = u->in(0)->in(k); 1316 --j; 1317 } else { // Normal use 1318 // Replace all uses 1319 for( uint k = 0; k < u->req(); k++ ) { 1320 if( u->in(k) == n ) { 1321 u->set_req( k, x ); 1322 --j; 1323 } 1324 } 1325 x_ctrl = get_ctrl(u); 1326 } 1327 1328 // Find control for 'x' next to use but not inside inner loops. 1329 // For inner loop uses get the preheader area. 1330 x_ctrl = place_near_use(x_ctrl); 1331 1332 if (n->is_Load()) { 1333 // For loads, add a control edge to a CFG node outside of the loop 1334 // to force them to not combine and return back inside the loop 1335 // during GVN optimization (4641526). 1336 // 1337 // Because we are setting the actual control input, factor in 1338 // the result from get_late_ctrl() so we respect any 1339 // anti-dependences. (6233005). 1340 x_ctrl = dom_lca(late_load_ctrl, x_ctrl); 1341 1342 // Don't allow the control input to be a CFG splitting node. 1343 // Such nodes should only have ProjNodes as outs, e.g. IfNode 1344 // should only have IfTrueNode and IfFalseNode (4985384). 1345 x_ctrl = find_non_split_ctrl(x_ctrl); 1346 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone"); 1347 1348 x->set_req(0, x_ctrl); 1349 } 1350 register_new_node(x, x_ctrl); 1351 1352 // Some institutional knowledge is needed here: 'x' is 1353 // yanked because if the optimizer runs GVN on it all the 1354 // cloned x's will common up and undo this optimization and 1355 // be forced back in the loop. This is annoying because it 1356 // makes +VerifyOpto report false-positives on progress. I 1357 // tried setting control edges on the x's to force them to 1358 // not combine, but the matching gets worried when it tries 1359 // to fold a StoreP and an AddP together (as part of an 1360 // address expression) and the AddP and StoreP have 1361 // different controls. 1362 if (!x->is_Load() && !x->is_DecodeNarrowPtr()) _igvn._worklist.yank(x); 1363 } 1364 _igvn.remove_dead_node(n); 1365 } 1366 } 1367 } 1368 } 1369 1370 try_move_store_after_loop(n); 1371 1372 // Remove multiple allocations of the same value type 1373 if (n->is_ValueType() && EliminateAllocations) { 1374 n->as_ValueType()->remove_redundant_allocations(&_igvn, this); 1375 } 1376 1377 // Check for Opaque2's who's loop has disappeared - who's input is in the 1378 // same loop nest as their output. Remove 'em, they are no longer useful. 1379 if( n_op == Op_Opaque2 && 1380 n->in(1) != NULL && 1381 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) { 1382 _igvn.replace_node( n, n->in(1) ); 1383 } 1384 } 1385 1386 //------------------------------split_if_with_blocks--------------------------- 1387 // Check for aggressive application of 'split-if' optimization, 1388 // using basic block level info. 1389 void PhaseIdealLoop::split_if_with_blocks( VectorSet &visited, Node_Stack &nstack ) { 1390 Node *n = C->root(); 1391 visited.set(n->_idx); // first, mark node as visited 1392 // Do pre-visit work for root 1393 n = split_if_with_blocks_pre( n ); 1394 uint cnt = n->outcnt(); 1395 uint i = 0; 1396 while (true) { 1397 // Visit all children 1398 if (i < cnt) { 1399 Node* use = n->raw_out(i); 1400 ++i; 1401 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) { 1402 // Now do pre-visit work for this use 1403 use = split_if_with_blocks_pre( use ); 1404 nstack.push(n, i); // Save parent and next use's index. 1405 n = use; // Process all children of current use. 1406 cnt = use->outcnt(); 1407 i = 0; 1408 } 1409 } 1410 else { 1411 // All of n's children have been processed, complete post-processing. 1412 if (cnt != 0 && !n->is_Con()) { 1413 assert(has_node(n), "no dead nodes"); 1414 split_if_with_blocks_post( n ); 1415 } 1416 if (nstack.is_empty()) { 1417 // Finished all nodes on stack. 1418 break; 1419 } 1420 // Get saved parent node and next use's index. Visit the rest of uses. 1421 n = nstack.node(); 1422 cnt = n->outcnt(); 1423 i = nstack.index(); 1424 nstack.pop(); 1425 } 1426 } 1427 } 1428 1429 1430 //============================================================================= 1431 // 1432 // C L O N E A L O O P B O D Y 1433 // 1434 1435 //------------------------------clone_iff-------------------------------------- 1436 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1437 // "Nearly" because all Nodes have been cloned from the original in the loop, 1438 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1439 // through the Phi recursively, and return a Bool. 1440 BoolNode *PhaseIdealLoop::clone_iff( PhiNode *phi, IdealLoopTree *loop ) { 1441 1442 // Convert this Phi into a Phi merging Bools 1443 uint i; 1444 for( i = 1; i < phi->req(); i++ ) { 1445 Node *b = phi->in(i); 1446 if( b->is_Phi() ) { 1447 _igvn.replace_input_of(phi, i, clone_iff( b->as_Phi(), loop )); 1448 } else { 1449 assert( b->is_Bool(), "" ); 1450 } 1451 } 1452 1453 Node *sample_bool = phi->in(1); 1454 Node *sample_cmp = sample_bool->in(1); 1455 1456 // Make Phis to merge the Cmp's inputs. 1457 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP ); 1458 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP ); 1459 for( i = 1; i < phi->req(); i++ ) { 1460 Node *n1 = phi->in(i)->in(1)->in(1); 1461 Node *n2 = phi->in(i)->in(1)->in(2); 1462 phi1->set_req( i, n1 ); 1463 phi2->set_req( i, n2 ); 1464 phi1->set_type( phi1->type()->meet_speculative(n1->bottom_type())); 1465 phi2->set_type( phi2->type()->meet_speculative(n2->bottom_type())); 1466 } 1467 // See if these Phis have been made before. 1468 // Register with optimizer 1469 Node *hit1 = _igvn.hash_find_insert(phi1); 1470 if( hit1 ) { // Hit, toss just made Phi 1471 _igvn.remove_dead_node(phi1); // Remove new phi 1472 assert( hit1->is_Phi(), "" ); 1473 phi1 = (PhiNode*)hit1; // Use existing phi 1474 } else { // Miss 1475 _igvn.register_new_node_with_optimizer(phi1); 1476 } 1477 Node *hit2 = _igvn.hash_find_insert(phi2); 1478 if( hit2 ) { // Hit, toss just made Phi 1479 _igvn.remove_dead_node(phi2); // Remove new phi 1480 assert( hit2->is_Phi(), "" ); 1481 phi2 = (PhiNode*)hit2; // Use existing phi 1482 } else { // Miss 1483 _igvn.register_new_node_with_optimizer(phi2); 1484 } 1485 // Register Phis with loop/block info 1486 set_ctrl(phi1, phi->in(0)); 1487 set_ctrl(phi2, phi->in(0)); 1488 // Make a new Cmp 1489 Node *cmp = sample_cmp->clone(); 1490 cmp->set_req( 1, phi1 ); 1491 cmp->set_req( 2, phi2 ); 1492 _igvn.register_new_node_with_optimizer(cmp); 1493 set_ctrl(cmp, phi->in(0)); 1494 1495 // Make a new Bool 1496 Node *b = sample_bool->clone(); 1497 b->set_req(1,cmp); 1498 _igvn.register_new_node_with_optimizer(b); 1499 set_ctrl(b, phi->in(0)); 1500 1501 assert( b->is_Bool(), "" ); 1502 return (BoolNode*)b; 1503 } 1504 1505 //------------------------------clone_bool------------------------------------- 1506 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1507 // "Nearly" because all Nodes have been cloned from the original in the loop, 1508 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1509 // through the Phi recursively, and return a Bool. 1510 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) { 1511 uint i; 1512 // Convert this Phi into a Phi merging Bools 1513 for( i = 1; i < phi->req(); i++ ) { 1514 Node *b = phi->in(i); 1515 if( b->is_Phi() ) { 1516 _igvn.replace_input_of(phi, i, clone_bool( b->as_Phi(), loop )); 1517 } else { 1518 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" ); 1519 } 1520 } 1521 1522 Node *sample_cmp = phi->in(1); 1523 1524 // Make Phis to merge the Cmp's inputs. 1525 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP ); 1526 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP ); 1527 for( uint j = 1; j < phi->req(); j++ ) { 1528 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP 1529 Node *n1, *n2; 1530 if( cmp_top->is_Cmp() ) { 1531 n1 = cmp_top->in(1); 1532 n2 = cmp_top->in(2); 1533 } else { 1534 n1 = n2 = cmp_top; 1535 } 1536 phi1->set_req( j, n1 ); 1537 phi2->set_req( j, n2 ); 1538 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 1539 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 1540 } 1541 1542 // See if these Phis have been made before. 1543 // Register with optimizer 1544 Node *hit1 = _igvn.hash_find_insert(phi1); 1545 if( hit1 ) { // Hit, toss just made Phi 1546 _igvn.remove_dead_node(phi1); // Remove new phi 1547 assert( hit1->is_Phi(), "" ); 1548 phi1 = (PhiNode*)hit1; // Use existing phi 1549 } else { // Miss 1550 _igvn.register_new_node_with_optimizer(phi1); 1551 } 1552 Node *hit2 = _igvn.hash_find_insert(phi2); 1553 if( hit2 ) { // Hit, toss just made Phi 1554 _igvn.remove_dead_node(phi2); // Remove new phi 1555 assert( hit2->is_Phi(), "" ); 1556 phi2 = (PhiNode*)hit2; // Use existing phi 1557 } else { // Miss 1558 _igvn.register_new_node_with_optimizer(phi2); 1559 } 1560 // Register Phis with loop/block info 1561 set_ctrl(phi1, phi->in(0)); 1562 set_ctrl(phi2, phi->in(0)); 1563 // Make a new Cmp 1564 Node *cmp = sample_cmp->clone(); 1565 cmp->set_req( 1, phi1 ); 1566 cmp->set_req( 2, phi2 ); 1567 _igvn.register_new_node_with_optimizer(cmp); 1568 set_ctrl(cmp, phi->in(0)); 1569 1570 assert( cmp->is_Cmp(), "" ); 1571 return (CmpNode*)cmp; 1572 } 1573 1574 //------------------------------sink_use--------------------------------------- 1575 // If 'use' was in the loop-exit block, it now needs to be sunk 1576 // below the post-loop merge point. 1577 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) { 1578 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) { 1579 set_ctrl(use, post_loop); 1580 for (DUIterator j = use->outs(); use->has_out(j); j++) 1581 sink_use(use->out(j), post_loop); 1582 } 1583 } 1584 1585 //------------------------------clone_loop------------------------------------- 1586 // 1587 // C L O N E A L O O P B O D Y 1588 // 1589 // This is the basic building block of the loop optimizations. It clones an 1590 // entire loop body. It makes an old_new loop body mapping; with this mapping 1591 // you can find the new-loop equivalent to an old-loop node. All new-loop 1592 // nodes are exactly equal to their old-loop counterparts, all edges are the 1593 // same. All exits from the old-loop now have a RegionNode that merges the 1594 // equivalent new-loop path. This is true even for the normal "loop-exit" 1595 // condition. All uses of loop-invariant old-loop values now come from (one 1596 // or more) Phis that merge their new-loop equivalents. 1597 // 1598 // This operation leaves the graph in an illegal state: there are two valid 1599 // control edges coming from the loop pre-header to both loop bodies. I'll 1600 // definitely have to hack the graph after running this transform. 1601 // 1602 // From this building block I will further edit edges to perform loop peeling 1603 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc. 1604 // 1605 // Parameter side_by_size_idom: 1606 // When side_by_size_idom is NULL, the dominator tree is constructed for 1607 // the clone loop to dominate the original. Used in construction of 1608 // pre-main-post loop sequence. 1609 // When nonnull, the clone and original are side-by-side, both are 1610 // dominated by the side_by_side_idom node. Used in construction of 1611 // unswitched loops. 1612 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd, 1613 Node* side_by_side_idom) { 1614 1615 if (C->do_vector_loop() && PrintOpto) { 1616 const char* mname = C->method()->name()->as_quoted_ascii(); 1617 if (mname != NULL) { 1618 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname); 1619 } 1620 } 1621 1622 CloneMap& cm = C->clone_map(); 1623 Dict* dict = cm.dict(); 1624 if (C->do_vector_loop()) { 1625 cm.set_clone_idx(cm.max_gen()+1); 1626 #ifndef PRODUCT 1627 if (PrintOpto) { 1628 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx()); 1629 loop->dump_head(); 1630 } 1631 #endif 1632 } 1633 1634 // Step 1: Clone the loop body. Make the old->new mapping. 1635 uint i; 1636 for( i = 0; i < loop->_body.size(); i++ ) { 1637 Node *old = loop->_body.at(i); 1638 Node *nnn = old->clone(); 1639 old_new.map( old->_idx, nnn ); 1640 if (C->do_vector_loop()) { 1641 cm.verify_insert_and_clone(old, nnn, cm.clone_idx()); 1642 } 1643 _igvn.register_new_node_with_optimizer(nnn); 1644 } 1645 1646 1647 // Step 2: Fix the edges in the new body. If the old input is outside the 1648 // loop use it. If the old input is INside the loop, use the corresponding 1649 // new node instead. 1650 for( i = 0; i < loop->_body.size(); i++ ) { 1651 Node *old = loop->_body.at(i); 1652 Node *nnn = old_new[old->_idx]; 1653 // Fix CFG/Loop controlling the new node 1654 if (has_ctrl(old)) { 1655 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]); 1656 } else { 1657 set_loop(nnn, loop->_parent); 1658 if (old->outcnt() > 0) { 1659 set_idom( nnn, old_new[idom(old)->_idx], dd ); 1660 } 1661 } 1662 // Correct edges to the new node 1663 for( uint j = 0; j < nnn->req(); j++ ) { 1664 Node *n = nnn->in(j); 1665 if( n ) { 1666 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n ); 1667 if( loop->is_member( old_in_loop ) ) 1668 nnn->set_req(j, old_new[n->_idx]); 1669 } 1670 } 1671 _igvn.hash_find_insert(nnn); 1672 } 1673 Node *newhead = old_new[loop->_head->_idx]; 1674 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 1675 1676 1677 // Step 3: Now fix control uses. Loop varying control uses have already 1678 // been fixed up (as part of all input edges in Step 2). Loop invariant 1679 // control uses must be either an IfFalse or an IfTrue. Make a merge 1680 // point to merge the old and new IfFalse/IfTrue nodes; make the use 1681 // refer to this. 1682 ResourceArea *area = Thread::current()->resource_area(); 1683 Node_List worklist(area); 1684 uint new_counter = C->unique(); 1685 for( i = 0; i < loop->_body.size(); i++ ) { 1686 Node* old = loop->_body.at(i); 1687 if( !old->is_CFG() ) continue; 1688 Node* nnn = old_new[old->_idx]; 1689 1690 // Copy uses to a worklist, so I can munge the def-use info 1691 // with impunity. 1692 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1693 worklist.push(old->fast_out(j)); 1694 1695 while( worklist.size() ) { // Visit all uses 1696 Node *use = worklist.pop(); 1697 if (!has_node(use)) continue; // Ignore dead nodes 1698 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1699 if( !loop->is_member( use_loop ) && use->is_CFG() ) { 1700 // Both OLD and USE are CFG nodes here. 1701 assert( use->is_Proj(), "" ); 1702 1703 // Clone the loop exit control projection 1704 Node *newuse = use->clone(); 1705 if (C->do_vector_loop()) { 1706 cm.verify_insert_and_clone(use, newuse, cm.clone_idx()); 1707 } 1708 newuse->set_req(0,nnn); 1709 _igvn.register_new_node_with_optimizer(newuse); 1710 set_loop(newuse, use_loop); 1711 set_idom(newuse, nnn, dom_depth(nnn) + 1 ); 1712 1713 // We need a Region to merge the exit from the peeled body and the 1714 // exit from the old loop body. 1715 RegionNode *r = new RegionNode(3); 1716 // Map the old use to the new merge point 1717 old_new.map( use->_idx, r ); 1718 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use)); 1719 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" ); 1720 1721 // The original user of 'use' uses 'r' instead. 1722 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) { 1723 Node* useuse = use->last_out(l); 1724 _igvn.rehash_node_delayed(useuse); 1725 uint uses_found = 0; 1726 if( useuse->in(0) == use ) { 1727 useuse->set_req(0, r); 1728 uses_found++; 1729 if( useuse->is_CFG() ) { 1730 assert( dom_depth(useuse) > dd_r, "" ); 1731 set_idom(useuse, r, dom_depth(useuse)); 1732 } 1733 } 1734 for( uint k = 1; k < useuse->req(); k++ ) { 1735 if( useuse->in(k) == use ) { 1736 useuse->set_req(k, r); 1737 uses_found++; 1738 } 1739 } 1740 l -= uses_found; // we deleted 1 or more copies of this edge 1741 } 1742 1743 // Now finish up 'r' 1744 r->set_req( 1, newuse ); 1745 r->set_req( 2, use ); 1746 _igvn.register_new_node_with_optimizer(r); 1747 set_loop(r, use_loop); 1748 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r); 1749 } // End of if a loop-exit test 1750 } 1751 } 1752 1753 // Step 4: If loop-invariant use is not control, it must be dominated by a 1754 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region 1755 // there if needed. Make a Phi there merging old and new used values. 1756 Node_List *split_if_set = NULL; 1757 Node_List *split_bool_set = NULL; 1758 Node_List *split_cex_set = NULL; 1759 for( i = 0; i < loop->_body.size(); i++ ) { 1760 Node* old = loop->_body.at(i); 1761 Node* nnn = old_new[old->_idx]; 1762 // Copy uses to a worklist, so I can munge the def-use info 1763 // with impunity. 1764 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1765 worklist.push(old->fast_out(j)); 1766 1767 while( worklist.size() ) { 1768 Node *use = worklist.pop(); 1769 if (!has_node(use)) continue; // Ignore dead nodes 1770 if (use->in(0) == C->top()) continue; 1771 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1772 // Check for data-use outside of loop - at least one of OLD or USE 1773 // must not be a CFG node. 1774 if( !loop->is_member( use_loop ) && (!old->is_CFG() || !use->is_CFG())) { 1775 1776 // If the Data use is an IF, that means we have an IF outside of the 1777 // loop that is switching on a condition that is set inside of the 1778 // loop. Happens if people set a loop-exit flag; then test the flag 1779 // in the loop to break the loop, then test is again outside of the 1780 // loop to determine which way the loop exited. 1781 // Loop predicate If node connects to Bool node through Opaque1 node. 1782 if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use)) { 1783 // Since this code is highly unlikely, we lazily build the worklist 1784 // of such Nodes to go split. 1785 if( !split_if_set ) 1786 split_if_set = new Node_List(area); 1787 split_if_set->push(use); 1788 } 1789 if( use->is_Bool() ) { 1790 if( !split_bool_set ) 1791 split_bool_set = new Node_List(area); 1792 split_bool_set->push(use); 1793 } 1794 if( use->Opcode() == Op_CreateEx ) { 1795 if( !split_cex_set ) 1796 split_cex_set = new Node_List(area); 1797 split_cex_set->push(use); 1798 } 1799 1800 1801 // Get "block" use is in 1802 uint idx = 0; 1803 while( use->in(idx) != old ) idx++; 1804 Node *prev = use->is_CFG() ? use : get_ctrl(use); 1805 assert( !loop->is_member( get_loop( prev ) ), "" ); 1806 Node *cfg = prev->_idx >= new_counter 1807 ? prev->in(2) 1808 : idom(prev); 1809 if( use->is_Phi() ) // Phi use is in prior block 1810 cfg = prev->in(idx); // NOT in block of Phi itself 1811 if (cfg->is_top()) { // Use is dead? 1812 _igvn.replace_input_of(use, idx, C->top()); 1813 continue; 1814 } 1815 1816 while( !loop->is_member( get_loop( cfg ) ) ) { 1817 prev = cfg; 1818 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg); 1819 } 1820 // If the use occurs after merging several exits from the loop, then 1821 // old value must have dominated all those exits. Since the same old 1822 // value was used on all those exits we did not need a Phi at this 1823 // merge point. NOW we do need a Phi here. Each loop exit value 1824 // is now merged with the peeled body exit; each exit gets its own 1825 // private Phi and those Phis need to be merged here. 1826 Node *phi; 1827 if( prev->is_Region() ) { 1828 if( idx == 0 ) { // Updating control edge? 1829 phi = prev; // Just use existing control 1830 } else { // Else need a new Phi 1831 phi = PhiNode::make( prev, old ); 1832 // Now recursively fix up the new uses of old! 1833 for( uint i = 1; i < prev->req(); i++ ) { 1834 worklist.push(phi); // Onto worklist once for each 'old' input 1835 } 1836 } 1837 } else { 1838 // Get new RegionNode merging old and new loop exits 1839 prev = old_new[prev->_idx]; 1840 assert( prev, "just made this in step 7" ); 1841 if( idx == 0 ) { // Updating control edge? 1842 phi = prev; // Just use existing control 1843 } else { // Else need a new Phi 1844 // Make a new Phi merging data values properly 1845 phi = PhiNode::make( prev, old ); 1846 phi->set_req( 1, nnn ); 1847 } 1848 } 1849 // If inserting a new Phi, check for prior hits 1850 if( idx != 0 ) { 1851 Node *hit = _igvn.hash_find_insert(phi); 1852 if( hit == NULL ) { 1853 _igvn.register_new_node_with_optimizer(phi); // Register new phi 1854 } else { // or 1855 // Remove the new phi from the graph and use the hit 1856 _igvn.remove_dead_node(phi); 1857 phi = hit; // Use existing phi 1858 } 1859 set_ctrl(phi, prev); 1860 } 1861 // Make 'use' use the Phi instead of the old loop body exit value 1862 _igvn.replace_input_of(use, idx, phi); 1863 if( use->_idx >= new_counter ) { // If updating new phis 1864 // Not needed for correctness, but prevents a weak assert 1865 // in AddPNode from tripping (when we end up with different 1866 // base & derived Phis that will become the same after 1867 // IGVN does CSE). 1868 Node *hit = _igvn.hash_find_insert(use); 1869 if( hit ) // Go ahead and re-hash for hits. 1870 _igvn.replace_node( use, hit ); 1871 } 1872 1873 // If 'use' was in the loop-exit block, it now needs to be sunk 1874 // below the post-loop merge point. 1875 sink_use( use, prev ); 1876 } 1877 } 1878 } 1879 1880 // Check for IFs that need splitting/cloning. Happens if an IF outside of 1881 // the loop uses a condition set in the loop. The original IF probably 1882 // takes control from one or more OLD Regions (which in turn get from NEW 1883 // Regions). In any case, there will be a set of Phis for each merge point 1884 // from the IF up to where the original BOOL def exists the loop. 1885 if( split_if_set ) { 1886 while( split_if_set->size() ) { 1887 Node *iff = split_if_set->pop(); 1888 if( iff->in(1)->is_Phi() ) { 1889 BoolNode *b = clone_iff( iff->in(1)->as_Phi(), loop ); 1890 _igvn.replace_input_of(iff, 1, b); 1891 } 1892 } 1893 } 1894 if( split_bool_set ) { 1895 while( split_bool_set->size() ) { 1896 Node *b = split_bool_set->pop(); 1897 Node *phi = b->in(1); 1898 assert( phi->is_Phi(), "" ); 1899 CmpNode *cmp = clone_bool( (PhiNode*)phi, loop ); 1900 _igvn.replace_input_of(b, 1, cmp); 1901 } 1902 } 1903 if( split_cex_set ) { 1904 while( split_cex_set->size() ) { 1905 Node *b = split_cex_set->pop(); 1906 assert( b->in(0)->is_Region(), "" ); 1907 assert( b->in(1)->is_Phi(), "" ); 1908 assert( b->in(0)->in(0) == b->in(1)->in(0), "" ); 1909 split_up( b, b->in(0), NULL ); 1910 } 1911 } 1912 1913 } 1914 1915 1916 //---------------------- stride_of_possible_iv ------------------------------------- 1917 // Looks for an iff/bool/comp with one operand of the compare 1918 // being a cycle involving an add and a phi, 1919 // with an optional truncation (left-shift followed by a right-shift) 1920 // of the add. Returns zero if not an iv. 1921 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) { 1922 Node* trunc1 = NULL; 1923 Node* trunc2 = NULL; 1924 const TypeInt* ttype = NULL; 1925 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) { 1926 return 0; 1927 } 1928 BoolNode* bl = iff->in(1)->as_Bool(); 1929 Node* cmp = bl->in(1); 1930 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) { 1931 return 0; 1932 } 1933 // Must have an invariant operand 1934 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) { 1935 return 0; 1936 } 1937 Node* add2 = NULL; 1938 Node* cmp1 = cmp->in(1); 1939 if (cmp1->is_Phi()) { 1940 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) ))) 1941 Node* phi = cmp1; 1942 for (uint i = 1; i < phi->req(); i++) { 1943 Node* in = phi->in(i); 1944 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in, 1945 &trunc1, &trunc2, &ttype); 1946 if (add && add->in(1) == phi) { 1947 add2 = add->in(2); 1948 break; 1949 } 1950 } 1951 } else { 1952 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) ))) 1953 Node* addtrunc = cmp1; 1954 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc, 1955 &trunc1, &trunc2, &ttype); 1956 if (add && add->in(1)->is_Phi()) { 1957 Node* phi = add->in(1); 1958 for (uint i = 1; i < phi->req(); i++) { 1959 if (phi->in(i) == addtrunc) { 1960 add2 = add->in(2); 1961 break; 1962 } 1963 } 1964 } 1965 } 1966 if (add2 != NULL) { 1967 const TypeInt* add2t = _igvn.type(add2)->is_int(); 1968 if (add2t->is_con()) { 1969 return add2t->get_con(); 1970 } 1971 } 1972 return 0; 1973 } 1974 1975 1976 //---------------------- stay_in_loop ------------------------------------- 1977 // Return the (unique) control output node that's in the loop (if it exists.) 1978 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) { 1979 Node* unique = NULL; 1980 if (!n) return NULL; 1981 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1982 Node* use = n->fast_out(i); 1983 if (!has_ctrl(use) && loop->is_member(get_loop(use))) { 1984 if (unique != NULL) { 1985 return NULL; 1986 } 1987 unique = use; 1988 } 1989 } 1990 return unique; 1991 } 1992 1993 //------------------------------ register_node ------------------------------------- 1994 // Utility to register node "n" with PhaseIdealLoop 1995 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) { 1996 _igvn.register_new_node_with_optimizer(n); 1997 loop->_body.push(n); 1998 if (n->is_CFG()) { 1999 set_loop(n, loop); 2000 set_idom(n, pred, ddepth); 2001 } else { 2002 set_ctrl(n, pred); 2003 } 2004 } 2005 2006 //------------------------------ proj_clone ------------------------------------- 2007 // Utility to create an if-projection 2008 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) { 2009 ProjNode* c = p->clone()->as_Proj(); 2010 c->set_req(0, iff); 2011 return c; 2012 } 2013 2014 //------------------------------ short_circuit_if ------------------------------------- 2015 // Force the iff control output to be the live_proj 2016 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) { 2017 guarantee(live_proj != NULL, "null projection"); 2018 int proj_con = live_proj->_con; 2019 assert(proj_con == 0 || proj_con == 1, "false or true projection"); 2020 Node *con = _igvn.intcon(proj_con); 2021 set_ctrl(con, C->root()); 2022 if (iff) { 2023 iff->set_req(1, con); 2024 } 2025 return con; 2026 } 2027 2028 //------------------------------ insert_if_before_proj ------------------------------------- 2029 // Insert a new if before an if projection (* - new node) 2030 // 2031 // before 2032 // if(test) 2033 // / \ 2034 // v v 2035 // other-proj proj (arg) 2036 // 2037 // after 2038 // if(test) 2039 // / \ 2040 // / v 2041 // | * proj-clone 2042 // v | 2043 // other-proj v 2044 // * new_if(relop(cmp[IU](left,right))) 2045 // / \ 2046 // v v 2047 // * new-proj proj 2048 // (returned) 2049 // 2050 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) { 2051 IfNode* iff = proj->in(0)->as_If(); 2052 IdealLoopTree *loop = get_loop(proj); 2053 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2054 int ddepth = dom_depth(proj); 2055 2056 _igvn.rehash_node_delayed(iff); 2057 _igvn.rehash_node_delayed(proj); 2058 2059 proj->set_req(0, NULL); // temporary disconnect 2060 ProjNode* proj2 = proj_clone(proj, iff); 2061 register_node(proj2, loop, iff, ddepth); 2062 2063 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right); 2064 register_node(cmp, loop, proj2, ddepth); 2065 2066 BoolNode* bol = new BoolNode(cmp, relop); 2067 register_node(bol, loop, proj2, ddepth); 2068 2069 int opcode = iff->Opcode(); 2070 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode"); 2071 IfNode* new_if = (opcode == Op_If) ? new IfNode(proj2, bol, iff->_prob, iff->_fcnt): 2072 new RangeCheckNode(proj2, bol, iff->_prob, iff->_fcnt); 2073 register_node(new_if, loop, proj2, ddepth); 2074 2075 proj->set_req(0, new_if); // reattach 2076 set_idom(proj, new_if, ddepth); 2077 2078 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj(); 2079 guarantee(new_exit != NULL, "null exit node"); 2080 register_node(new_exit, get_loop(other_proj), new_if, ddepth); 2081 2082 return new_exit; 2083 } 2084 2085 //------------------------------ insert_region_before_proj ------------------------------------- 2086 // Insert a region before an if projection (* - new node) 2087 // 2088 // before 2089 // if(test) 2090 // / | 2091 // v | 2092 // proj v 2093 // other-proj 2094 // 2095 // after 2096 // if(test) 2097 // / | 2098 // v | 2099 // * proj-clone v 2100 // | other-proj 2101 // v 2102 // * new-region 2103 // | 2104 // v 2105 // * dum_if 2106 // / \ 2107 // v \ 2108 // * dum-proj v 2109 // proj 2110 // 2111 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) { 2112 IfNode* iff = proj->in(0)->as_If(); 2113 IdealLoopTree *loop = get_loop(proj); 2114 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2115 int ddepth = dom_depth(proj); 2116 2117 _igvn.rehash_node_delayed(iff); 2118 _igvn.rehash_node_delayed(proj); 2119 2120 proj->set_req(0, NULL); // temporary disconnect 2121 ProjNode* proj2 = proj_clone(proj, iff); 2122 register_node(proj2, loop, iff, ddepth); 2123 2124 RegionNode* reg = new RegionNode(2); 2125 reg->set_req(1, proj2); 2126 register_node(reg, loop, iff, ddepth); 2127 2128 IfNode* dum_if = new IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt); 2129 register_node(dum_if, loop, reg, ddepth); 2130 2131 proj->set_req(0, dum_if); // reattach 2132 set_idom(proj, dum_if, ddepth); 2133 2134 ProjNode* dum_proj = proj_clone(other_proj, dum_if); 2135 register_node(dum_proj, loop, dum_if, ddepth); 2136 2137 return reg; 2138 } 2139 2140 //------------------------------ insert_cmpi_loop_exit ------------------------------------- 2141 // Clone a signed compare loop exit from an unsigned compare and 2142 // insert it before the unsigned cmp on the stay-in-loop path. 2143 // All new nodes inserted in the dominator tree between the original 2144 // if and it's projections. The original if test is replaced with 2145 // a constant to force the stay-in-loop path. 2146 // 2147 // This is done to make sure that the original if and it's projections 2148 // still dominate the same set of control nodes, that the ctrl() relation 2149 // from data nodes to them is preserved, and that their loop nesting is 2150 // preserved. 2151 // 2152 // before 2153 // if(i <u limit) unsigned compare loop exit 2154 // / | 2155 // v v 2156 // exit-proj stay-in-loop-proj 2157 // 2158 // after 2159 // if(stay-in-loop-const) original if 2160 // / | 2161 // / v 2162 // / if(i < limit) new signed test 2163 // / / | 2164 // / / v 2165 // / / if(i <u limit) new cloned unsigned test 2166 // / / / | 2167 // v v v | 2168 // region | 2169 // | | 2170 // dum-if | 2171 // / | | 2172 // ether | | 2173 // v v 2174 // exit-proj stay-in-loop-proj 2175 // 2176 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) { 2177 const bool Signed = true; 2178 const bool Unsigned = false; 2179 2180 BoolNode* bol = if_cmpu->in(1)->as_Bool(); 2181 if (bol->_test._test != BoolTest::lt) return NULL; 2182 CmpNode* cmpu = bol->in(1)->as_Cmp(); 2183 if (cmpu->Opcode() != Op_CmpU) return NULL; 2184 int stride = stride_of_possible_iv(if_cmpu); 2185 if (stride == 0) return NULL; 2186 2187 Node* lp_proj = stay_in_loop(if_cmpu, loop); 2188 guarantee(lp_proj != NULL, "null loop node"); 2189 2190 ProjNode* lp_continue = lp_proj->as_Proj(); 2191 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj(); 2192 2193 Node* limit = NULL; 2194 if (stride > 0) { 2195 limit = cmpu->in(2); 2196 } else { 2197 limit = _igvn.makecon(TypeInt::ZERO); 2198 set_ctrl(limit, C->root()); 2199 } 2200 // Create a new region on the exit path 2201 RegionNode* reg = insert_region_before_proj(lp_exit); 2202 guarantee(reg != NULL, "null region node"); 2203 2204 // Clone the if-cmpu-true-false using a signed compare 2205 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge; 2206 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue); 2207 reg->add_req(cmpi_exit); 2208 2209 // Clone the if-cmpu-true-false 2210 BoolTest::mask rel_u = bol->_test._test; 2211 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue); 2212 reg->add_req(cmpu_exit); 2213 2214 // Force original if to stay in loop. 2215 short_circuit_if(if_cmpu, lp_continue); 2216 2217 return cmpi_exit->in(0)->as_If(); 2218 } 2219 2220 //------------------------------ remove_cmpi_loop_exit ------------------------------------- 2221 // Remove a previously inserted signed compare loop exit. 2222 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) { 2223 Node* lp_proj = stay_in_loop(if_cmp, loop); 2224 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI && 2225 stay_in_loop(lp_proj, loop)->is_If() && 2226 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu"); 2227 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO); 2228 set_ctrl(con, C->root()); 2229 if_cmp->set_req(1, con); 2230 } 2231 2232 //------------------------------ scheduled_nodelist ------------------------------------- 2233 // Create a post order schedule of nodes that are in the 2234 // "member" set. The list is returned in "sched". 2235 // The first node in "sched" is the loop head, followed by 2236 // nodes which have no inputs in the "member" set, and then 2237 // followed by the nodes that have an immediate input dependence 2238 // on a node in "sched". 2239 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) { 2240 2241 assert(member.test(loop->_head->_idx), "loop head must be in member set"); 2242 Arena *a = Thread::current()->resource_area(); 2243 VectorSet visited(a); 2244 Node_Stack nstack(a, loop->_body.size()); 2245 2246 Node* n = loop->_head; // top of stack is cached in "n" 2247 uint idx = 0; 2248 visited.set(n->_idx); 2249 2250 // Initially push all with no inputs from within member set 2251 for(uint i = 0; i < loop->_body.size(); i++ ) { 2252 Node *elt = loop->_body.at(i); 2253 if (member.test(elt->_idx)) { 2254 bool found = false; 2255 for (uint j = 0; j < elt->req(); j++) { 2256 Node* def = elt->in(j); 2257 if (def && member.test(def->_idx) && def != elt) { 2258 found = true; 2259 break; 2260 } 2261 } 2262 if (!found && elt != loop->_head) { 2263 nstack.push(n, idx); 2264 n = elt; 2265 assert(!visited.test(n->_idx), "not seen yet"); 2266 visited.set(n->_idx); 2267 } 2268 } 2269 } 2270 2271 // traverse out's that are in the member set 2272 while (true) { 2273 if (idx < n->outcnt()) { 2274 Node* use = n->raw_out(idx); 2275 idx++; 2276 if (!visited.test_set(use->_idx)) { 2277 if (member.test(use->_idx)) { 2278 nstack.push(n, idx); 2279 n = use; 2280 idx = 0; 2281 } 2282 } 2283 } else { 2284 // All outputs processed 2285 sched.push(n); 2286 if (nstack.is_empty()) break; 2287 n = nstack.node(); 2288 idx = nstack.index(); 2289 nstack.pop(); 2290 } 2291 } 2292 } 2293 2294 2295 //------------------------------ has_use_in_set ------------------------------------- 2296 // Has a use in the vector set 2297 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) { 2298 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2299 Node* use = n->fast_out(j); 2300 if (vset.test(use->_idx)) { 2301 return true; 2302 } 2303 } 2304 return false; 2305 } 2306 2307 2308 //------------------------------ has_use_internal_to_set ------------------------------------- 2309 // Has use internal to the vector set (ie. not in a phi at the loop head) 2310 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) { 2311 Node* head = loop->_head; 2312 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2313 Node* use = n->fast_out(j); 2314 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) { 2315 return true; 2316 } 2317 } 2318 return false; 2319 } 2320 2321 2322 //------------------------------ clone_for_use_outside_loop ------------------------------------- 2323 // clone "n" for uses that are outside of loop 2324 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) { 2325 int cloned = 0; 2326 assert(worklist.size() == 0, "should be empty"); 2327 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2328 Node* use = n->fast_out(j); 2329 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) { 2330 worklist.push(use); 2331 } 2332 } 2333 while( worklist.size() ) { 2334 Node *use = worklist.pop(); 2335 if (!has_node(use) || use->in(0) == C->top()) continue; 2336 uint j; 2337 for (j = 0; j < use->req(); j++) { 2338 if (use->in(j) == n) break; 2339 } 2340 assert(j < use->req(), "must be there"); 2341 2342 // clone "n" and insert it between the inputs of "n" and the use outside the loop 2343 Node* n_clone = n->clone(); 2344 _igvn.replace_input_of(use, j, n_clone); 2345 cloned++; 2346 Node* use_c; 2347 if (!use->is_Phi()) { 2348 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0); 2349 } else { 2350 // Use in a phi is considered a use in the associated predecessor block 2351 use_c = use->in(0)->in(j); 2352 } 2353 set_ctrl(n_clone, use_c); 2354 assert(!loop->is_member(get_loop(use_c)), "should be outside loop"); 2355 get_loop(use_c)->_body.push(n_clone); 2356 _igvn.register_new_node_with_optimizer(n_clone); 2357 #if !defined(PRODUCT) 2358 if (TracePartialPeeling) { 2359 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx); 2360 } 2361 #endif 2362 } 2363 return cloned; 2364 } 2365 2366 2367 //------------------------------ clone_for_special_use_inside_loop ------------------------------------- 2368 // clone "n" for special uses that are in the not_peeled region. 2369 // If these def-uses occur in separate blocks, the code generator 2370 // marks the method as not compilable. For example, if a "BoolNode" 2371 // is in a different basic block than the "IfNode" that uses it, then 2372 // the compilation is aborted in the code generator. 2373 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 2374 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) { 2375 if (n->is_Phi() || n->is_Load()) { 2376 return; 2377 } 2378 assert(worklist.size() == 0, "should be empty"); 2379 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2380 Node* use = n->fast_out(j); 2381 if ( not_peel.test(use->_idx) && 2382 (use->is_If() || use->is_CMove() || use->is_Bool()) && 2383 use->in(1) == n) { 2384 worklist.push(use); 2385 } 2386 } 2387 if (worklist.size() > 0) { 2388 // clone "n" and insert it between inputs of "n" and the use 2389 Node* n_clone = n->clone(); 2390 loop->_body.push(n_clone); 2391 _igvn.register_new_node_with_optimizer(n_clone); 2392 set_ctrl(n_clone, get_ctrl(n)); 2393 sink_list.push(n_clone); 2394 not_peel <<= n_clone->_idx; // add n_clone to not_peel set. 2395 #if !defined(PRODUCT) 2396 if (TracePartialPeeling) { 2397 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx); 2398 } 2399 #endif 2400 while( worklist.size() ) { 2401 Node *use = worklist.pop(); 2402 _igvn.rehash_node_delayed(use); 2403 for (uint j = 1; j < use->req(); j++) { 2404 if (use->in(j) == n) { 2405 use->set_req(j, n_clone); 2406 } 2407 } 2408 } 2409 } 2410 } 2411 2412 2413 //------------------------------ insert_phi_for_loop ------------------------------------- 2414 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 2415 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) { 2416 Node *phi = PhiNode::make(lp, back_edge_val); 2417 phi->set_req(LoopNode::EntryControl, lp_entry_val); 2418 // Use existing phi if it already exists 2419 Node *hit = _igvn.hash_find_insert(phi); 2420 if( hit == NULL ) { 2421 _igvn.register_new_node_with_optimizer(phi); 2422 set_ctrl(phi, lp); 2423 } else { 2424 // Remove the new phi from the graph and use the hit 2425 _igvn.remove_dead_node(phi); 2426 phi = hit; 2427 } 2428 _igvn.replace_input_of(use, idx, phi); 2429 } 2430 2431 #ifdef ASSERT 2432 //------------------------------ is_valid_loop_partition ------------------------------------- 2433 // Validate the loop partition sets: peel and not_peel 2434 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, 2435 VectorSet& not_peel ) { 2436 uint i; 2437 // Check that peel_list entries are in the peel set 2438 for (i = 0; i < peel_list.size(); i++) { 2439 if (!peel.test(peel_list.at(i)->_idx)) { 2440 return false; 2441 } 2442 } 2443 // Check at loop members are in one of peel set or not_peel set 2444 for (i = 0; i < loop->_body.size(); i++ ) { 2445 Node *def = loop->_body.at(i); 2446 uint di = def->_idx; 2447 // Check that peel set elements are in peel_list 2448 if (peel.test(di)) { 2449 if (not_peel.test(di)) { 2450 return false; 2451 } 2452 // Must be in peel_list also 2453 bool found = false; 2454 for (uint j = 0; j < peel_list.size(); j++) { 2455 if (peel_list.at(j)->_idx == di) { 2456 found = true; 2457 break; 2458 } 2459 } 2460 if (!found) { 2461 return false; 2462 } 2463 } else if (not_peel.test(di)) { 2464 if (peel.test(di)) { 2465 return false; 2466 } 2467 } else { 2468 return false; 2469 } 2470 } 2471 return true; 2472 } 2473 2474 //------------------------------ is_valid_clone_loop_exit_use ------------------------------------- 2475 // Ensure a use outside of loop is of the right form 2476 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) { 2477 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2478 return (use->is_Phi() && 2479 use_c->is_Region() && use_c->req() == 3 && 2480 (use_c->in(exit_idx)->Opcode() == Op_IfTrue || 2481 use_c->in(exit_idx)->Opcode() == Op_IfFalse || 2482 use_c->in(exit_idx)->Opcode() == Op_JumpProj) && 2483 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) ); 2484 } 2485 2486 //------------------------------ is_valid_clone_loop_form ------------------------------------- 2487 // Ensure that all uses outside of loop are of the right form 2488 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 2489 uint orig_exit_idx, uint clone_exit_idx) { 2490 uint len = peel_list.size(); 2491 for (uint i = 0; i < len; i++) { 2492 Node *def = peel_list.at(i); 2493 2494 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 2495 Node *use = def->fast_out(j); 2496 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2497 if (!loop->is_member(get_loop(use_c))) { 2498 // use is not in the loop, check for correct structure 2499 if (use->in(0) == def) { 2500 // Okay 2501 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) { 2502 return false; 2503 } 2504 } 2505 } 2506 } 2507 return true; 2508 } 2509 #endif 2510 2511 //------------------------------ partial_peel ------------------------------------- 2512 // Partially peel (aka loop rotation) the top portion of a loop (called 2513 // the peel section below) by cloning it and placing one copy just before 2514 // the new loop head and the other copy at the bottom of the new loop. 2515 // 2516 // before after where it came from 2517 // 2518 // stmt1 stmt1 2519 // loop: stmt2 clone 2520 // stmt2 if condA goto exitA clone 2521 // if condA goto exitA new_loop: new 2522 // stmt3 stmt3 clone 2523 // if !condB goto loop if condB goto exitB clone 2524 // exitB: stmt2 orig 2525 // stmt4 if !condA goto new_loop orig 2526 // exitA: goto exitA 2527 // exitB: 2528 // stmt4 2529 // exitA: 2530 // 2531 // Step 1: find the cut point: an exit test on probable 2532 // induction variable. 2533 // Step 2: schedule (with cloning) operations in the peel 2534 // section that can be executed after the cut into 2535 // the section that is not peeled. This may need 2536 // to clone operations into exit blocks. For 2537 // instance, a reference to A[i] in the not-peel 2538 // section and a reference to B[i] in an exit block 2539 // may cause a left-shift of i by 2 to be placed 2540 // in the peel block. This step will clone the left 2541 // shift into the exit block and sink the left shift 2542 // from the peel to the not-peel section. 2543 // Step 3: clone the loop, retarget the control, and insert 2544 // phis for values that are live across the new loop 2545 // head. This is very dependent on the graph structure 2546 // from clone_loop. It creates region nodes for 2547 // exit control and associated phi nodes for values 2548 // flow out of the loop through that exit. The region 2549 // node is dominated by the clone's control projection. 2550 // So the clone's peel section is placed before the 2551 // new loop head, and the clone's not-peel section is 2552 // forms the top part of the new loop. The original 2553 // peel section forms the tail of the new loop. 2554 // Step 4: update the dominator tree and recompute the 2555 // dominator depth. 2556 // 2557 // orig 2558 // 2559 // stmt1 2560 // | 2561 // v 2562 // loop predicate 2563 // | 2564 // v 2565 // loop<----+ 2566 // | | 2567 // stmt2 | 2568 // | | 2569 // v | 2570 // ifA | 2571 // / | | 2572 // v v | 2573 // false true ^ <-- last_peel 2574 // / | | 2575 // / ===|==cut | 2576 // / stmt3 | <-- first_not_peel 2577 // / | | 2578 // | v | 2579 // v ifB | 2580 // exitA: / \ | 2581 // / \ | 2582 // v v | 2583 // false true | 2584 // / \ | 2585 // / ----+ 2586 // | 2587 // v 2588 // exitB: 2589 // stmt4 2590 // 2591 // 2592 // after clone loop 2593 // 2594 // stmt1 2595 // | 2596 // v 2597 // loop predicate 2598 // / \ 2599 // clone / \ orig 2600 // / \ 2601 // / \ 2602 // v v 2603 // +---->loop loop<----+ 2604 // | | | | 2605 // | stmt2 stmt2 | 2606 // | | | | 2607 // | v v | 2608 // | ifA ifA | 2609 // | | \ / | | 2610 // | v v v v | 2611 // ^ true false false true ^ <-- last_peel 2612 // | | ^ \ / | | 2613 // | cut==|== \ \ / ===|==cut | 2614 // | stmt3 \ \ / stmt3 | <-- first_not_peel 2615 // | | dom | | | | 2616 // | v \ 1v v2 v | 2617 // | ifB regionA ifB | 2618 // | / \ | / \ | 2619 // | / \ v / \ | 2620 // | v v exitA: v v | 2621 // | true false false true | 2622 // | / ^ \ / \ | 2623 // +---- \ \ / ----+ 2624 // dom \ / 2625 // \ 1v v2 2626 // regionB 2627 // | 2628 // v 2629 // exitB: 2630 // stmt4 2631 // 2632 // 2633 // after partial peel 2634 // 2635 // stmt1 2636 // | 2637 // v 2638 // loop predicate 2639 // / 2640 // clone / orig 2641 // / TOP 2642 // / \ 2643 // v v 2644 // TOP->loop loop----+ 2645 // | | | 2646 // stmt2 stmt2 | 2647 // | | | 2648 // v v | 2649 // ifA ifA | 2650 // | \ / | | 2651 // v v v v | 2652 // true false false true | <-- last_peel 2653 // | ^ \ / +------|---+ 2654 // +->newloop \ \ / === ==cut | | 2655 // | stmt3 \ \ / TOP | | 2656 // | | dom | | stmt3 | | <-- first_not_peel 2657 // | v \ 1v v2 v | | 2658 // | ifB regionA ifB ^ v 2659 // | / \ | / \ | | 2660 // | / \ v / \ | | 2661 // | v v exitA: v v | | 2662 // | true false false true | | 2663 // | / ^ \ / \ | | 2664 // | | \ \ / v | | 2665 // | | dom \ / TOP | | 2666 // | | \ 1v v2 | | 2667 // ^ v regionB | | 2668 // | | | | | 2669 // | | v ^ v 2670 // | | exitB: | | 2671 // | | stmt4 | | 2672 // | +------------>-----------------+ | 2673 // | | 2674 // +-----------------<---------------------+ 2675 // 2676 // 2677 // final graph 2678 // 2679 // stmt1 2680 // | 2681 // v 2682 // loop predicate 2683 // | 2684 // v 2685 // stmt2 clone 2686 // | 2687 // v 2688 // ........> ifA clone 2689 // : / | 2690 // dom / | 2691 // : v v 2692 // : false true 2693 // : | | 2694 // : | v 2695 // : | newloop<-----+ 2696 // : | | | 2697 // : | stmt3 clone | 2698 // : | | | 2699 // : | v | 2700 // : | ifB | 2701 // : | / \ | 2702 // : | v v | 2703 // : | false true | 2704 // : | | | | 2705 // : | v stmt2 | 2706 // : | exitB: | | 2707 // : | stmt4 v | 2708 // : | ifA orig | 2709 // : | / \ | 2710 // : | / \ | 2711 // : | v v | 2712 // : | false true | 2713 // : | / \ | 2714 // : v v -----+ 2715 // RegionA 2716 // | 2717 // v 2718 // exitA 2719 // 2720 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) { 2721 2722 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only"); 2723 if (!loop->_head->is_Loop()) { 2724 return false; } 2725 2726 LoopNode *head = loop->_head->as_Loop(); 2727 2728 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) { 2729 return false; 2730 } 2731 2732 // Check for complex exit control 2733 for(uint ii = 0; ii < loop->_body.size(); ii++ ) { 2734 Node *n = loop->_body.at(ii); 2735 int opc = n->Opcode(); 2736 if (n->is_Call() || 2737 opc == Op_Catch || 2738 opc == Op_CatchProj || 2739 opc == Op_Jump || 2740 opc == Op_JumpProj) { 2741 #if !defined(PRODUCT) 2742 if (TracePartialPeeling) { 2743 tty->print_cr("\nExit control too complex: lp: %d", head->_idx); 2744 } 2745 #endif 2746 return false; 2747 } 2748 } 2749 2750 int dd = dom_depth(head); 2751 2752 // Step 1: find cut point 2753 2754 // Walk up dominators to loop head looking for first loop exit 2755 // which is executed on every path thru loop. 2756 IfNode *peel_if = NULL; 2757 IfNode *peel_if_cmpu = NULL; 2758 2759 Node *iff = loop->tail(); 2760 while( iff != head ) { 2761 if( iff->is_If() ) { 2762 Node *ctrl = get_ctrl(iff->in(1)); 2763 if (ctrl->is_top()) return false; // Dead test on live IF. 2764 // If loop-varying exit-test, check for induction variable 2765 if( loop->is_member(get_loop(ctrl)) && 2766 loop->is_loop_exit(iff) && 2767 is_possible_iv_test(iff)) { 2768 Node* cmp = iff->in(1)->in(1); 2769 if (cmp->Opcode() == Op_CmpI) { 2770 peel_if = iff->as_If(); 2771 } else { 2772 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU"); 2773 peel_if_cmpu = iff->as_If(); 2774 } 2775 } 2776 } 2777 iff = idom(iff); 2778 } 2779 // Prefer signed compare over unsigned compare. 2780 IfNode* new_peel_if = NULL; 2781 if (peel_if == NULL) { 2782 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) { 2783 return false; // No peel point found 2784 } 2785 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop); 2786 if (new_peel_if == NULL) { 2787 return false; // No peel point found 2788 } 2789 peel_if = new_peel_if; 2790 } 2791 Node* last_peel = stay_in_loop(peel_if, loop); 2792 Node* first_not_peeled = stay_in_loop(last_peel, loop); 2793 if (first_not_peeled == NULL || first_not_peeled == head) { 2794 return false; 2795 } 2796 2797 #if !defined(PRODUCT) 2798 if (TraceLoopOpts) { 2799 tty->print("PartialPeel "); 2800 loop->dump_head(); 2801 } 2802 2803 if (TracePartialPeeling) { 2804 tty->print_cr("before partial peel one iteration"); 2805 Node_List wl; 2806 Node* t = head->in(2); 2807 while (true) { 2808 wl.push(t); 2809 if (t == head) break; 2810 t = idom(t); 2811 } 2812 while (wl.size() > 0) { 2813 Node* tt = wl.pop(); 2814 tt->dump(); 2815 if (tt == last_peel) tty->print_cr("-- cut --"); 2816 } 2817 } 2818 #endif 2819 ResourceArea *area = Thread::current()->resource_area(); 2820 VectorSet peel(area); 2821 VectorSet not_peel(area); 2822 Node_List peel_list(area); 2823 Node_List worklist(area); 2824 Node_List sink_list(area); 2825 2826 // Set of cfg nodes to peel are those that are executable from 2827 // the head through last_peel. 2828 assert(worklist.size() == 0, "should be empty"); 2829 worklist.push(head); 2830 peel.set(head->_idx); 2831 while (worklist.size() > 0) { 2832 Node *n = worklist.pop(); 2833 if (n != last_peel) { 2834 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2835 Node* use = n->fast_out(j); 2836 if (use->is_CFG() && 2837 loop->is_member(get_loop(use)) && 2838 !peel.test_set(use->_idx)) { 2839 worklist.push(use); 2840 } 2841 } 2842 } 2843 } 2844 2845 // Set of non-cfg nodes to peel are those that are control 2846 // dependent on the cfg nodes. 2847 uint i; 2848 for(i = 0; i < loop->_body.size(); i++ ) { 2849 Node *n = loop->_body.at(i); 2850 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n; 2851 if (peel.test(n_c->_idx)) { 2852 peel.set(n->_idx); 2853 } else { 2854 not_peel.set(n->_idx); 2855 } 2856 } 2857 2858 // Step 2: move operations from the peeled section down into the 2859 // not-peeled section 2860 2861 // Get a post order schedule of nodes in the peel region 2862 // Result in right-most operand. 2863 scheduled_nodelist(loop, peel, peel_list ); 2864 2865 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 2866 2867 // For future check for too many new phis 2868 uint old_phi_cnt = 0; 2869 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 2870 Node* use = head->fast_out(j); 2871 if (use->is_Phi()) old_phi_cnt++; 2872 } 2873 2874 #if !defined(PRODUCT) 2875 if (TracePartialPeeling) { 2876 tty->print_cr("\npeeled list"); 2877 } 2878 #endif 2879 2880 // Evacuate nodes in peel region into the not_peeled region if possible 2881 uint new_phi_cnt = 0; 2882 uint cloned_for_outside_use = 0; 2883 for (i = 0; i < peel_list.size();) { 2884 Node* n = peel_list.at(i); 2885 #if !defined(PRODUCT) 2886 if (TracePartialPeeling) n->dump(); 2887 #endif 2888 bool incr = true; 2889 if ( !n->is_CFG() ) { 2890 2891 if ( has_use_in_set(n, not_peel) ) { 2892 2893 // If not used internal to the peeled region, 2894 // move "n" from peeled to not_peeled region. 2895 2896 if ( !has_use_internal_to_set(n, peel, loop) ) { 2897 2898 // if not pinned and not a load (which maybe anti-dependent on a store) 2899 // and not a CMove (Matcher expects only bool->cmove). 2900 if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) { 2901 cloned_for_outside_use += clone_for_use_outside_loop( loop, n, worklist ); 2902 sink_list.push(n); 2903 peel >>= n->_idx; // delete n from peel set. 2904 not_peel <<= n->_idx; // add n to not_peel set. 2905 peel_list.remove(i); 2906 incr = false; 2907 #if !defined(PRODUCT) 2908 if (TracePartialPeeling) { 2909 tty->print_cr("sink to not_peeled region: %d newbb: %d", 2910 n->_idx, get_ctrl(n)->_idx); 2911 } 2912 #endif 2913 } 2914 } else { 2915 // Otherwise check for special def-use cases that span 2916 // the peel/not_peel boundary such as bool->if 2917 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist ); 2918 new_phi_cnt++; 2919 } 2920 } 2921 } 2922 if (incr) i++; 2923 } 2924 2925 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) { 2926 #if !defined(PRODUCT) 2927 if (TracePartialPeeling) { 2928 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c", 2929 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F'); 2930 } 2931 #endif 2932 if (new_peel_if != NULL) { 2933 remove_cmpi_loop_exit(new_peel_if, loop); 2934 } 2935 // Inhibit more partial peeling on this loop 2936 assert(!head->is_partial_peel_loop(), "not partial peeled"); 2937 head->mark_partial_peel_failed(); 2938 if (cloned_for_outside_use > 0) { 2939 // Terminate this round of loop opts because 2940 // the graph outside this loop was changed. 2941 C->set_major_progress(); 2942 return true; 2943 } 2944 return false; 2945 } 2946 2947 // Step 3: clone loop, retarget control, and insert new phis 2948 2949 // Create new loop head for new phis and to hang 2950 // the nodes being moved (sinked) from the peel region. 2951 LoopNode* new_head = new LoopNode(last_peel, last_peel); 2952 new_head->set_unswitch_count(head->unswitch_count()); // Preserve 2953 _igvn.register_new_node_with_optimizer(new_head); 2954 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled"); 2955 _igvn.replace_input_of(first_not_peeled, 0, new_head); 2956 set_loop(new_head, loop); 2957 loop->_body.push(new_head); 2958 not_peel.set(new_head->_idx); 2959 set_idom(new_head, last_peel, dom_depth(first_not_peeled)); 2960 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled)); 2961 2962 while (sink_list.size() > 0) { 2963 Node* n = sink_list.pop(); 2964 set_ctrl(n, new_head); 2965 } 2966 2967 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 2968 2969 clone_loop( loop, old_new, dd ); 2970 2971 const uint clone_exit_idx = 1; 2972 const uint orig_exit_idx = 2; 2973 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop"); 2974 2975 Node* head_clone = old_new[head->_idx]; 2976 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop(); 2977 Node* orig_tail_clone = head_clone->in(2); 2978 2979 // Add phi if "def" node is in peel set and "use" is not 2980 2981 for(i = 0; i < peel_list.size(); i++ ) { 2982 Node *def = peel_list.at(i); 2983 if (!def->is_CFG()) { 2984 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 2985 Node *use = def->fast_out(j); 2986 if (has_node(use) && use->in(0) != C->top() && 2987 (!peel.test(use->_idx) || 2988 (use->is_Phi() && use->in(0) == head)) ) { 2989 worklist.push(use); 2990 } 2991 } 2992 while( worklist.size() ) { 2993 Node *use = worklist.pop(); 2994 for (uint j = 1; j < use->req(); j++) { 2995 Node* n = use->in(j); 2996 if (n == def) { 2997 2998 // "def" is in peel set, "use" is not in peel set 2999 // or "use" is in the entry boundary (a phi) of the peel set 3000 3001 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use; 3002 3003 if ( loop->is_member(get_loop( use_c )) ) { 3004 // use is in loop 3005 if (old_new[use->_idx] != NULL) { // null for dead code 3006 Node* use_clone = old_new[use->_idx]; 3007 _igvn.replace_input_of(use, j, C->top()); 3008 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone ); 3009 } 3010 } else { 3011 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format"); 3012 // use is not in the loop, check if the live range includes the cut 3013 Node* lp_if = use_c->in(orig_exit_idx)->in(0); 3014 if (not_peel.test(lp_if->_idx)) { 3015 assert(j == orig_exit_idx, "use from original loop"); 3016 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone ); 3017 } 3018 } 3019 } 3020 } 3021 } 3022 } 3023 } 3024 3025 // Step 3b: retarget control 3026 3027 // Redirect control to the new loop head if a cloned node in 3028 // the not_peeled region has control that points into the peeled region. 3029 // This necessary because the cloned peeled region will be outside 3030 // the loop. 3031 // from to 3032 // cloned-peeled <---+ 3033 // new_head_clone: | <--+ 3034 // cloned-not_peeled in(0) in(0) 3035 // orig-peeled 3036 3037 for(i = 0; i < loop->_body.size(); i++ ) { 3038 Node *n = loop->_body.at(i); 3039 if (!n->is_CFG() && n->in(0) != NULL && 3040 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) { 3041 Node* n_clone = old_new[n->_idx]; 3042 _igvn.replace_input_of(n_clone, 0, new_head_clone); 3043 } 3044 } 3045 3046 // Backedge of the surviving new_head (the clone) is original last_peel 3047 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel); 3048 3049 // Cut first node in original not_peel set 3050 _igvn.rehash_node_delayed(new_head); // Multiple edge updates: 3051 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of 3052 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls 3053 3054 // Copy head_clone back-branch info to original head 3055 // and remove original head's loop entry and 3056 // clone head's back-branch 3057 _igvn.rehash_node_delayed(head); // Multiple edge updates 3058 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl)); 3059 head->set_req(LoopNode::LoopBackControl, C->top()); 3060 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top()); 3061 3062 // Similarly modify the phis 3063 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) { 3064 Node* use = head->fast_out(k); 3065 if (use->is_Phi() && use->outcnt() > 0) { 3066 Node* use_clone = old_new[use->_idx]; 3067 _igvn.rehash_node_delayed(use); // Multiple edge updates 3068 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl)); 3069 use->set_req(LoopNode::LoopBackControl, C->top()); 3070 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top()); 3071 } 3072 } 3073 3074 // Step 4: update dominator tree and dominator depth 3075 3076 set_idom(head, orig_tail_clone, dd); 3077 recompute_dom_depth(); 3078 3079 // Inhibit more partial peeling on this loop 3080 new_head_clone->set_partial_peel_loop(); 3081 C->set_major_progress(); 3082 loop->record_for_igvn(); 3083 3084 #if !defined(PRODUCT) 3085 if (TracePartialPeeling) { 3086 tty->print_cr("\nafter partial peel one iteration"); 3087 Node_List wl(area); 3088 Node* t = last_peel; 3089 while (true) { 3090 wl.push(t); 3091 if (t == head_clone) break; 3092 t = idom(t); 3093 } 3094 while (wl.size() > 0) { 3095 Node* tt = wl.pop(); 3096 if (tt == head) tty->print_cr("orig head"); 3097 else if (tt == new_head_clone) tty->print_cr("new head"); 3098 else if (tt == head_clone) tty->print_cr("clone head"); 3099 tt->dump(); 3100 } 3101 } 3102 #endif 3103 return true; 3104 } 3105 3106 //------------------------------reorg_offsets---------------------------------- 3107 // Reorganize offset computations to lower register pressure. Mostly 3108 // prevent loop-fallout uses of the pre-incremented trip counter (which are 3109 // then alive with the post-incremented trip counter forcing an extra 3110 // register move) 3111 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) { 3112 // Perform it only for canonical counted loops. 3113 // Loop's shape could be messed up by iteration_split_impl. 3114 if (!loop->_head->is_CountedLoop()) 3115 return; 3116 if (!loop->_head->as_Loop()->is_valid_counted_loop()) 3117 return; 3118 3119 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 3120 CountedLoopEndNode *cle = cl->loopexit(); 3121 Node *exit = cle->proj_out(false); 3122 Node *phi = cl->phi(); 3123 3124 // Check for the special case of folks using the pre-incremented 3125 // trip-counter on the fall-out path (forces the pre-incremented 3126 // and post-incremented trip counter to be live at the same time). 3127 // Fix this by adjusting to use the post-increment trip counter. 3128 3129 bool progress = true; 3130 while (progress) { 3131 progress = false; 3132 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) { 3133 Node* use = phi->fast_out(i); // User of trip-counter 3134 if (!has_ctrl(use)) continue; 3135 Node *u_ctrl = get_ctrl(use); 3136 if (use->is_Phi()) { 3137 u_ctrl = NULL; 3138 for (uint j = 1; j < use->req(); j++) 3139 if (use->in(j) == phi) 3140 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j)); 3141 } 3142 IdealLoopTree *u_loop = get_loop(u_ctrl); 3143 // Look for loop-invariant use 3144 if (u_loop == loop) continue; 3145 if (loop->is_member(u_loop)) continue; 3146 // Check that use is live out the bottom. Assuming the trip-counter 3147 // update is right at the bottom, uses of of the loop middle are ok. 3148 if (dom_lca(exit, u_ctrl) != exit) continue; 3149 // Hit! Refactor use to use the post-incremented tripcounter. 3150 // Compute a post-increment tripcounter. 3151 Node *opaq = new Opaque2Node( C, cle->incr() ); 3152 register_new_node(opaq, exit); 3153 Node *neg_stride = _igvn.intcon(-cle->stride_con()); 3154 set_ctrl(neg_stride, C->root()); 3155 Node *post = new AddINode( opaq, neg_stride); 3156 register_new_node(post, exit); 3157 _igvn.rehash_node_delayed(use); 3158 for (uint j = 1; j < use->req(); j++) { 3159 if (use->in(j) == phi) 3160 use->set_req(j, post); 3161 } 3162 // Since DU info changed, rerun loop 3163 progress = true; 3164 break; 3165 } 3166 } 3167 3168 }