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 #ifdef _LP64 1019 if (m->Opcode() == Op_ConvI2L) 1020 return false; 1021 if (m->is_CastII() && m->isa_CastII()->has_range_check()) { 1022 return false; 1023 } 1024 #endif 1025 } 1026 } 1027 } 1028 return true; 1029 } 1030 1031 1032 //------------------------------place_near_use--------------------------------- 1033 // Place some computation next to use but not inside inner loops. 1034 // For inner loop uses move it to the preheader area. 1035 Node *PhaseIdealLoop::place_near_use( Node *useblock ) const { 1036 IdealLoopTree *u_loop = get_loop( useblock ); 1037 return (u_loop->_irreducible || u_loop->_child) 1038 ? useblock 1039 : u_loop->_head->in(LoopNode::EntryControl); 1040 } 1041 1042 1043 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) { 1044 if (!n->is_If()) { 1045 return false; 1046 } 1047 if (!n->in(0)->is_Region()) { 1048 return false; 1049 } 1050 Node* region = n->in(0); 1051 Node* dom = idom(region); 1052 if (!dom->is_If() || dom->in(1) != n->in(1)) { 1053 return false; 1054 } 1055 IfNode* dom_if = dom->as_If(); 1056 Node* proj_true = dom_if->proj_out(1); 1057 Node* proj_false = dom_if->proj_out(0); 1058 1059 for (uint i = 1; i < region->req(); i++) { 1060 if (is_dominator(proj_true, region->in(i))) { 1061 continue; 1062 } 1063 if (is_dominator(proj_false, region->in(i))) { 1064 continue; 1065 } 1066 return false; 1067 } 1068 1069 return true; 1070 } 1071 1072 bool PhaseIdealLoop::can_split_if(Node *n_ctrl) { 1073 if (C->live_nodes() > 35000) { 1074 return false; // Method too big 1075 } 1076 1077 // Do not do 'split-if' if irreducible loops are present. 1078 if (_has_irreducible_loops) { 1079 return false; 1080 } 1081 1082 if (merge_point_too_heavy(C, n_ctrl)) { 1083 return false; 1084 } 1085 1086 // Do not do 'split-if' if some paths are dead. First do dead code 1087 // elimination and then see if its still profitable. 1088 for (uint i = 1; i < n_ctrl->req(); i++) { 1089 if (n_ctrl->in(i) == C->top()) { 1090 return false; 1091 } 1092 } 1093 1094 // If trying to do a 'Split-If' at the loop head, it is only 1095 // profitable if the cmp folds up on BOTH paths. Otherwise we 1096 // risk peeling a loop forever. 1097 1098 // CNC - Disabled for now. Requires careful handling of loop 1099 // body selection for the cloned code. Also, make sure we check 1100 // for any input path not being in the same loop as n_ctrl. For 1101 // irreducible loops we cannot check for 'n_ctrl->is_Loop()' 1102 // because the alternative loop entry points won't be converted 1103 // into LoopNodes. 1104 IdealLoopTree *n_loop = get_loop(n_ctrl); 1105 for (uint j = 1; j < n_ctrl->req(); j++) { 1106 if (get_loop(n_ctrl->in(j)) != n_loop) { 1107 return false; 1108 } 1109 } 1110 1111 // Check for safety of the merge point. 1112 if (!merge_point_safe(n_ctrl)) { 1113 return false; 1114 } 1115 1116 return true; 1117 } 1118 1119 //------------------------------split_if_with_blocks_post---------------------- 1120 // Do the real work in a non-recursive function. CFG hackery wants to be 1121 // in the post-order, so it can dirty the I-DOM info and not use the dirtied 1122 // info. 1123 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) { 1124 1125 // Cloning Cmp through Phi's involves the split-if transform. 1126 // FastLock is not used by an If 1127 if (n->is_Cmp() && !n->is_FastLock()) { 1128 Node *n_ctrl = get_ctrl(n); 1129 // Determine if the Node has inputs from some local Phi. 1130 // Returns the block to clone thru. 1131 Node *n_blk = has_local_phi_input(n); 1132 if (n_blk != n_ctrl) { 1133 return; 1134 } 1135 1136 if (!can_split_if(n_ctrl)) { 1137 return; 1138 } 1139 1140 if (n->outcnt() != 1) { 1141 return; // Multiple bool's from 1 compare? 1142 } 1143 Node *bol = n->unique_out(); 1144 assert(bol->is_Bool(), "expect a bool here"); 1145 if (bol->outcnt() != 1) { 1146 return;// Multiple branches from 1 compare? 1147 } 1148 Node *iff = bol->unique_out(); 1149 1150 // Check some safety conditions 1151 if (iff->is_If()) { // Classic split-if? 1152 if (iff->in(0) != n_ctrl) { 1153 return; // Compare must be in same blk as if 1154 } 1155 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE 1156 // Can't split CMove with different control edge. 1157 if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) { 1158 return; 1159 } 1160 if (get_ctrl(iff->in(2)) == n_ctrl || 1161 get_ctrl(iff->in(3)) == n_ctrl) { 1162 return; // Inputs not yet split-up 1163 } 1164 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) { 1165 return; // Loop-invar test gates loop-varying CMOVE 1166 } 1167 } else { 1168 return; // some other kind of node, such as an Allocate 1169 } 1170 1171 // When is split-if profitable? Every 'win' on means some control flow 1172 // goes dead, so it's almost always a win. 1173 int policy = 0; 1174 // Split compare 'n' through the merge point if it is profitable 1175 Node *phi = split_thru_phi( n, n_ctrl, policy); 1176 if (!phi) { 1177 return; 1178 } 1179 1180 // Found a Phi to split thru! 1181 // Replace 'n' with the new phi 1182 _igvn.replace_node(n, phi); 1183 1184 // Now split the bool up thru the phi 1185 Node *bolphi = split_thru_phi(bol, n_ctrl, -1); 1186 guarantee(bolphi != NULL, "null boolean phi node"); 1187 1188 _igvn.replace_node(bol, bolphi); 1189 assert(iff->in(1) == bolphi, ""); 1190 1191 if (bolphi->Value(&_igvn)->singleton()) { 1192 return; 1193 } 1194 1195 // Conditional-move? Must split up now 1196 if (!iff->is_If()) { 1197 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1); 1198 _igvn.replace_node(iff, cmovphi); 1199 return; 1200 } 1201 1202 // Now split the IF 1203 do_split_if(iff); 1204 return; 1205 } 1206 1207 // Two identical ifs back to back can be merged 1208 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) { 1209 Node *n_ctrl = n->in(0); 1210 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1)); 1211 IfNode* dom_if = idom(n_ctrl)->as_If(); 1212 Node* proj_true = dom_if->proj_out(1); 1213 Node* proj_false = dom_if->proj_out(0); 1214 Node* con_true = _igvn.makecon(TypeInt::ONE); 1215 Node* con_false = _igvn.makecon(TypeInt::ZERO); 1216 1217 for (uint i = 1; i < n_ctrl->req(); i++) { 1218 if (is_dominator(proj_true, n_ctrl->in(i))) { 1219 bolphi->init_req(i, con_true); 1220 } else { 1221 assert(is_dominator(proj_false, n_ctrl->in(i)), "bad if"); 1222 bolphi->init_req(i, con_false); 1223 } 1224 } 1225 register_new_node(bolphi, n_ctrl); 1226 _igvn.replace_input_of(n, 1, bolphi); 1227 1228 // Now split the IF 1229 do_split_if(n); 1230 return; 1231 } 1232 1233 // Check for an IF ready to split; one that has its 1234 // condition codes input coming from a Phi at the block start. 1235 int n_op = n->Opcode(); 1236 1237 // Check for an IF being dominated by another IF same test 1238 if (n_op == Op_If || 1239 n_op == Op_RangeCheck) { 1240 Node *bol = n->in(1); 1241 uint max = bol->outcnt(); 1242 // Check for same test used more than once? 1243 if (max > 1 && bol->is_Bool()) { 1244 // Search up IDOMs to see if this IF is dominated. 1245 Node *cutoff = get_ctrl(bol); 1246 1247 // Now search up IDOMs till cutoff, looking for a dominating test 1248 Node *prevdom = n; 1249 Node *dom = idom(prevdom); 1250 while (dom != cutoff) { 1251 if (dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom) { 1252 // Replace the dominated test with an obvious true or false. 1253 // Place it on the IGVN worklist for later cleanup. 1254 C->set_major_progress(); 1255 dominated_by(prevdom, n, false, true); 1256 #ifndef PRODUCT 1257 if( VerifyLoopOptimizations ) verify(); 1258 #endif 1259 return; 1260 } 1261 prevdom = dom; 1262 dom = idom(prevdom); 1263 } 1264 } 1265 } 1266 1267 // See if a shared loop-varying computation has no loop-varying uses. 1268 // Happens if something is only used for JVM state in uncommon trap exits, 1269 // like various versions of induction variable+offset. Clone the 1270 // computation per usage to allow it to sink out of the loop. 1271 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about) 1272 Node *n_ctrl = get_ctrl(n); 1273 IdealLoopTree *n_loop = get_loop(n_ctrl); 1274 if( n_loop != _ltree_root ) { 1275 DUIterator_Fast imax, i = n->fast_outs(imax); 1276 for (; i < imax; i++) { 1277 Node* u = n->fast_out(i); 1278 if( !has_ctrl(u) ) break; // Found control user 1279 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 1280 if( u_loop == n_loop ) break; // Found loop-varying use 1281 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop 1282 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003 1283 } 1284 bool did_break = (i < imax); // Did we break out of the previous loop? 1285 if (!did_break && n->outcnt() > 1) { // All uses in outer loops! 1286 Node *late_load_ctrl = NULL; 1287 if (n->is_Load()) { 1288 // If n is a load, get and save the result from get_late_ctrl(), 1289 // to be later used in calculating the control for n's clones. 1290 clear_dom_lca_tags(); 1291 late_load_ctrl = get_late_ctrl(n, n_ctrl); 1292 } 1293 // If n is a load, and the late control is the same as the current 1294 // control, then the cloning of n is a pointless exercise, because 1295 // GVN will ensure that we end up where we started. 1296 if (!n->is_Load() || late_load_ctrl != n_ctrl) { 1297 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) { 1298 Node *u = n->last_out(j); // Clone private computation per use 1299 _igvn.rehash_node_delayed(u); 1300 Node *x = n->clone(); // Clone computation 1301 Node *x_ctrl = NULL; 1302 if( u->is_Phi() ) { 1303 // Replace all uses of normal nodes. Replace Phi uses 1304 // individually, so the separate Nodes can sink down 1305 // different paths. 1306 uint k = 1; 1307 while( u->in(k) != n ) k++; 1308 u->set_req( k, x ); 1309 // x goes next to Phi input path 1310 x_ctrl = u->in(0)->in(k); 1311 --j; 1312 } else { // Normal use 1313 // Replace all uses 1314 for( uint k = 0; k < u->req(); k++ ) { 1315 if( u->in(k) == n ) { 1316 u->set_req( k, x ); 1317 --j; 1318 } 1319 } 1320 x_ctrl = get_ctrl(u); 1321 } 1322 1323 // Find control for 'x' next to use but not inside inner loops. 1324 // For inner loop uses get the preheader area. 1325 x_ctrl = place_near_use(x_ctrl); 1326 1327 if (n->is_Load()) { 1328 // For loads, add a control edge to a CFG node outside of the loop 1329 // to force them to not combine and return back inside the loop 1330 // during GVN optimization (4641526). 1331 // 1332 // Because we are setting the actual control input, factor in 1333 // the result from get_late_ctrl() so we respect any 1334 // anti-dependences. (6233005). 1335 x_ctrl = dom_lca(late_load_ctrl, x_ctrl); 1336 1337 // Don't allow the control input to be a CFG splitting node. 1338 // Such nodes should only have ProjNodes as outs, e.g. IfNode 1339 // should only have IfTrueNode and IfFalseNode (4985384). 1340 x_ctrl = find_non_split_ctrl(x_ctrl); 1341 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone"); 1342 1343 x->set_req(0, x_ctrl); 1344 } 1345 register_new_node(x, x_ctrl); 1346 1347 // Some institutional knowledge is needed here: 'x' is 1348 // yanked because if the optimizer runs GVN on it all the 1349 // cloned x's will common up and undo this optimization and 1350 // be forced back in the loop. This is annoying because it 1351 // makes +VerifyOpto report false-positives on progress. I 1352 // tried setting control edges on the x's to force them to 1353 // not combine, but the matching gets worried when it tries 1354 // to fold a StoreP and an AddP together (as part of an 1355 // address expression) and the AddP and StoreP have 1356 // different controls. 1357 if (!x->is_Load() && !x->is_DecodeNarrowPtr()) _igvn._worklist.yank(x); 1358 } 1359 _igvn.remove_dead_node(n); 1360 } 1361 } 1362 } 1363 } 1364 1365 try_move_store_after_loop(n); 1366 1367 // Remove multiple allocations of the same value type 1368 if (n->is_ValueType() && EliminateAllocations) { 1369 n->as_ValueType()->remove_redundant_allocations(&_igvn, this); 1370 } 1371 1372 // Check for Opaque2's who's loop has disappeared - who's input is in the 1373 // same loop nest as their output. Remove 'em, they are no longer useful. 1374 if( n_op == Op_Opaque2 && 1375 n->in(1) != NULL && 1376 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) { 1377 _igvn.replace_node( n, n->in(1) ); 1378 } 1379 } 1380 1381 //------------------------------split_if_with_blocks--------------------------- 1382 // Check for aggressive application of 'split-if' optimization, 1383 // using basic block level info. 1384 void PhaseIdealLoop::split_if_with_blocks( VectorSet &visited, Node_Stack &nstack ) { 1385 Node *n = C->root(); 1386 visited.set(n->_idx); // first, mark node as visited 1387 // Do pre-visit work for root 1388 n = split_if_with_blocks_pre( n ); 1389 uint cnt = n->outcnt(); 1390 uint i = 0; 1391 while (true) { 1392 // Visit all children 1393 if (i < cnt) { 1394 Node* use = n->raw_out(i); 1395 ++i; 1396 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) { 1397 // Now do pre-visit work for this use 1398 use = split_if_with_blocks_pre( use ); 1399 nstack.push(n, i); // Save parent and next use's index. 1400 n = use; // Process all children of current use. 1401 cnt = use->outcnt(); 1402 i = 0; 1403 } 1404 } 1405 else { 1406 // All of n's children have been processed, complete post-processing. 1407 if (cnt != 0 && !n->is_Con()) { 1408 assert(has_node(n), "no dead nodes"); 1409 split_if_with_blocks_post( n ); 1410 } 1411 if (nstack.is_empty()) { 1412 // Finished all nodes on stack. 1413 break; 1414 } 1415 // Get saved parent node and next use's index. Visit the rest of uses. 1416 n = nstack.node(); 1417 cnt = n->outcnt(); 1418 i = nstack.index(); 1419 nstack.pop(); 1420 } 1421 } 1422 } 1423 1424 1425 //============================================================================= 1426 // 1427 // C L O N E A L O O P B O D Y 1428 // 1429 1430 //------------------------------clone_iff-------------------------------------- 1431 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1432 // "Nearly" because all Nodes have been cloned from the original in the loop, 1433 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1434 // through the Phi recursively, and return a Bool. 1435 BoolNode *PhaseIdealLoop::clone_iff( PhiNode *phi, IdealLoopTree *loop ) { 1436 1437 // Convert this Phi into a Phi merging Bools 1438 uint i; 1439 for( i = 1; i < phi->req(); i++ ) { 1440 Node *b = phi->in(i); 1441 if( b->is_Phi() ) { 1442 _igvn.replace_input_of(phi, i, clone_iff( b->as_Phi(), loop )); 1443 } else { 1444 assert( b->is_Bool(), "" ); 1445 } 1446 } 1447 1448 Node *sample_bool = phi->in(1); 1449 Node *sample_cmp = sample_bool->in(1); 1450 1451 // Make Phis to merge the Cmp's inputs. 1452 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP ); 1453 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP ); 1454 for( i = 1; i < phi->req(); i++ ) { 1455 Node *n1 = phi->in(i)->in(1)->in(1); 1456 Node *n2 = phi->in(i)->in(1)->in(2); 1457 phi1->set_req( i, n1 ); 1458 phi2->set_req( i, n2 ); 1459 phi1->set_type( phi1->type()->meet_speculative(n1->bottom_type())); 1460 phi2->set_type( phi2->type()->meet_speculative(n2->bottom_type())); 1461 } 1462 // See if these Phis have been made before. 1463 // Register with optimizer 1464 Node *hit1 = _igvn.hash_find_insert(phi1); 1465 if( hit1 ) { // Hit, toss just made Phi 1466 _igvn.remove_dead_node(phi1); // Remove new phi 1467 assert( hit1->is_Phi(), "" ); 1468 phi1 = (PhiNode*)hit1; // Use existing phi 1469 } else { // Miss 1470 _igvn.register_new_node_with_optimizer(phi1); 1471 } 1472 Node *hit2 = _igvn.hash_find_insert(phi2); 1473 if( hit2 ) { // Hit, toss just made Phi 1474 _igvn.remove_dead_node(phi2); // Remove new phi 1475 assert( hit2->is_Phi(), "" ); 1476 phi2 = (PhiNode*)hit2; // Use existing phi 1477 } else { // Miss 1478 _igvn.register_new_node_with_optimizer(phi2); 1479 } 1480 // Register Phis with loop/block info 1481 set_ctrl(phi1, phi->in(0)); 1482 set_ctrl(phi2, phi->in(0)); 1483 // Make a new Cmp 1484 Node *cmp = sample_cmp->clone(); 1485 cmp->set_req( 1, phi1 ); 1486 cmp->set_req( 2, phi2 ); 1487 _igvn.register_new_node_with_optimizer(cmp); 1488 set_ctrl(cmp, phi->in(0)); 1489 1490 // Make a new Bool 1491 Node *b = sample_bool->clone(); 1492 b->set_req(1,cmp); 1493 _igvn.register_new_node_with_optimizer(b); 1494 set_ctrl(b, phi->in(0)); 1495 1496 assert( b->is_Bool(), "" ); 1497 return (BoolNode*)b; 1498 } 1499 1500 //------------------------------clone_bool------------------------------------- 1501 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1502 // "Nearly" because all Nodes have been cloned from the original in the loop, 1503 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1504 // through the Phi recursively, and return a Bool. 1505 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) { 1506 uint i; 1507 // Convert this Phi into a Phi merging Bools 1508 for( i = 1; i < phi->req(); i++ ) { 1509 Node *b = phi->in(i); 1510 if( b->is_Phi() ) { 1511 _igvn.replace_input_of(phi, i, clone_bool( b->as_Phi(), loop )); 1512 } else { 1513 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" ); 1514 } 1515 } 1516 1517 Node *sample_cmp = phi->in(1); 1518 1519 // Make Phis to merge the Cmp's inputs. 1520 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP ); 1521 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP ); 1522 for( uint j = 1; j < phi->req(); j++ ) { 1523 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP 1524 Node *n1, *n2; 1525 if( cmp_top->is_Cmp() ) { 1526 n1 = cmp_top->in(1); 1527 n2 = cmp_top->in(2); 1528 } else { 1529 n1 = n2 = cmp_top; 1530 } 1531 phi1->set_req( j, n1 ); 1532 phi2->set_req( j, n2 ); 1533 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 1534 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 1535 } 1536 1537 // See if these Phis have been made before. 1538 // Register with optimizer 1539 Node *hit1 = _igvn.hash_find_insert(phi1); 1540 if( hit1 ) { // Hit, toss just made Phi 1541 _igvn.remove_dead_node(phi1); // Remove new phi 1542 assert( hit1->is_Phi(), "" ); 1543 phi1 = (PhiNode*)hit1; // Use existing phi 1544 } else { // Miss 1545 _igvn.register_new_node_with_optimizer(phi1); 1546 } 1547 Node *hit2 = _igvn.hash_find_insert(phi2); 1548 if( hit2 ) { // Hit, toss just made Phi 1549 _igvn.remove_dead_node(phi2); // Remove new phi 1550 assert( hit2->is_Phi(), "" ); 1551 phi2 = (PhiNode*)hit2; // Use existing phi 1552 } else { // Miss 1553 _igvn.register_new_node_with_optimizer(phi2); 1554 } 1555 // Register Phis with loop/block info 1556 set_ctrl(phi1, phi->in(0)); 1557 set_ctrl(phi2, phi->in(0)); 1558 // Make a new Cmp 1559 Node *cmp = sample_cmp->clone(); 1560 cmp->set_req( 1, phi1 ); 1561 cmp->set_req( 2, phi2 ); 1562 _igvn.register_new_node_with_optimizer(cmp); 1563 set_ctrl(cmp, phi->in(0)); 1564 1565 assert( cmp->is_Cmp(), "" ); 1566 return (CmpNode*)cmp; 1567 } 1568 1569 //------------------------------sink_use--------------------------------------- 1570 // If 'use' was in the loop-exit block, it now needs to be sunk 1571 // below the post-loop merge point. 1572 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) { 1573 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) { 1574 set_ctrl(use, post_loop); 1575 for (DUIterator j = use->outs(); use->has_out(j); j++) 1576 sink_use(use->out(j), post_loop); 1577 } 1578 } 1579 1580 //------------------------------clone_loop------------------------------------- 1581 // 1582 // C L O N E A L O O P B O D Y 1583 // 1584 // This is the basic building block of the loop optimizations. It clones an 1585 // entire loop body. It makes an old_new loop body mapping; with this mapping 1586 // you can find the new-loop equivalent to an old-loop node. All new-loop 1587 // nodes are exactly equal to their old-loop counterparts, all edges are the 1588 // same. All exits from the old-loop now have a RegionNode that merges the 1589 // equivalent new-loop path. This is true even for the normal "loop-exit" 1590 // condition. All uses of loop-invariant old-loop values now come from (one 1591 // or more) Phis that merge their new-loop equivalents. 1592 // 1593 // This operation leaves the graph in an illegal state: there are two valid 1594 // control edges coming from the loop pre-header to both loop bodies. I'll 1595 // definitely have to hack the graph after running this transform. 1596 // 1597 // From this building block I will further edit edges to perform loop peeling 1598 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc. 1599 // 1600 // Parameter side_by_size_idom: 1601 // When side_by_size_idom is NULL, the dominator tree is constructed for 1602 // the clone loop to dominate the original. Used in construction of 1603 // pre-main-post loop sequence. 1604 // When nonnull, the clone and original are side-by-side, both are 1605 // dominated by the side_by_side_idom node. Used in construction of 1606 // unswitched loops. 1607 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd, 1608 Node* side_by_side_idom) { 1609 1610 if (C->do_vector_loop() && PrintOpto) { 1611 const char* mname = C->method()->name()->as_quoted_ascii(); 1612 if (mname != NULL) { 1613 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname); 1614 } 1615 } 1616 1617 CloneMap& cm = C->clone_map(); 1618 Dict* dict = cm.dict(); 1619 if (C->do_vector_loop()) { 1620 cm.set_clone_idx(cm.max_gen()+1); 1621 #ifndef PRODUCT 1622 if (PrintOpto) { 1623 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx()); 1624 loop->dump_head(); 1625 } 1626 #endif 1627 } 1628 1629 // Step 1: Clone the loop body. Make the old->new mapping. 1630 uint i; 1631 for( i = 0; i < loop->_body.size(); i++ ) { 1632 Node *old = loop->_body.at(i); 1633 Node *nnn = old->clone(); 1634 old_new.map( old->_idx, nnn ); 1635 if (C->do_vector_loop()) { 1636 cm.verify_insert_and_clone(old, nnn, cm.clone_idx()); 1637 } 1638 _igvn.register_new_node_with_optimizer(nnn); 1639 } 1640 1641 1642 // Step 2: Fix the edges in the new body. If the old input is outside the 1643 // loop use it. If the old input is INside the loop, use the corresponding 1644 // new node instead. 1645 for( i = 0; i < loop->_body.size(); i++ ) { 1646 Node *old = loop->_body.at(i); 1647 Node *nnn = old_new[old->_idx]; 1648 // Fix CFG/Loop controlling the new node 1649 if (has_ctrl(old)) { 1650 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]); 1651 } else { 1652 set_loop(nnn, loop->_parent); 1653 if (old->outcnt() > 0) { 1654 set_idom( nnn, old_new[idom(old)->_idx], dd ); 1655 } 1656 } 1657 // Correct edges to the new node 1658 for( uint j = 0; j < nnn->req(); j++ ) { 1659 Node *n = nnn->in(j); 1660 if( n ) { 1661 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n ); 1662 if( loop->is_member( old_in_loop ) ) 1663 nnn->set_req(j, old_new[n->_idx]); 1664 } 1665 } 1666 _igvn.hash_find_insert(nnn); 1667 } 1668 Node *newhead = old_new[loop->_head->_idx]; 1669 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 1670 1671 1672 // Step 3: Now fix control uses. Loop varying control uses have already 1673 // been fixed up (as part of all input edges in Step 2). Loop invariant 1674 // control uses must be either an IfFalse or an IfTrue. Make a merge 1675 // point to merge the old and new IfFalse/IfTrue nodes; make the use 1676 // refer to this. 1677 ResourceArea *area = Thread::current()->resource_area(); 1678 Node_List worklist(area); 1679 uint new_counter = C->unique(); 1680 for( i = 0; i < loop->_body.size(); i++ ) { 1681 Node* old = loop->_body.at(i); 1682 if( !old->is_CFG() ) continue; 1683 Node* nnn = old_new[old->_idx]; 1684 1685 // Copy uses to a worklist, so I can munge the def-use info 1686 // with impunity. 1687 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1688 worklist.push(old->fast_out(j)); 1689 1690 while( worklist.size() ) { // Visit all uses 1691 Node *use = worklist.pop(); 1692 if (!has_node(use)) continue; // Ignore dead nodes 1693 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1694 if( !loop->is_member( use_loop ) && use->is_CFG() ) { 1695 // Both OLD and USE are CFG nodes here. 1696 assert( use->is_Proj(), "" ); 1697 1698 // Clone the loop exit control projection 1699 Node *newuse = use->clone(); 1700 if (C->do_vector_loop()) { 1701 cm.verify_insert_and_clone(use, newuse, cm.clone_idx()); 1702 } 1703 newuse->set_req(0,nnn); 1704 _igvn.register_new_node_with_optimizer(newuse); 1705 set_loop(newuse, use_loop); 1706 set_idom(newuse, nnn, dom_depth(nnn) + 1 ); 1707 1708 // We need a Region to merge the exit from the peeled body and the 1709 // exit from the old loop body. 1710 RegionNode *r = new RegionNode(3); 1711 // Map the old use to the new merge point 1712 old_new.map( use->_idx, r ); 1713 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use)); 1714 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" ); 1715 1716 // The original user of 'use' uses 'r' instead. 1717 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) { 1718 Node* useuse = use->last_out(l); 1719 _igvn.rehash_node_delayed(useuse); 1720 uint uses_found = 0; 1721 if( useuse->in(0) == use ) { 1722 useuse->set_req(0, r); 1723 uses_found++; 1724 if( useuse->is_CFG() ) { 1725 assert( dom_depth(useuse) > dd_r, "" ); 1726 set_idom(useuse, r, dom_depth(useuse)); 1727 } 1728 } 1729 for( uint k = 1; k < useuse->req(); k++ ) { 1730 if( useuse->in(k) == use ) { 1731 useuse->set_req(k, r); 1732 uses_found++; 1733 } 1734 } 1735 l -= uses_found; // we deleted 1 or more copies of this edge 1736 } 1737 1738 // Now finish up 'r' 1739 r->set_req( 1, newuse ); 1740 r->set_req( 2, use ); 1741 _igvn.register_new_node_with_optimizer(r); 1742 set_loop(r, use_loop); 1743 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r); 1744 } // End of if a loop-exit test 1745 } 1746 } 1747 1748 // Step 4: If loop-invariant use is not control, it must be dominated by a 1749 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region 1750 // there if needed. Make a Phi there merging old and new used values. 1751 Node_List *split_if_set = NULL; 1752 Node_List *split_bool_set = NULL; 1753 Node_List *split_cex_set = NULL; 1754 for( i = 0; i < loop->_body.size(); i++ ) { 1755 Node* old = loop->_body.at(i); 1756 Node* nnn = old_new[old->_idx]; 1757 // Copy uses to a worklist, so I can munge the def-use info 1758 // with impunity. 1759 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1760 worklist.push(old->fast_out(j)); 1761 1762 while( worklist.size() ) { 1763 Node *use = worklist.pop(); 1764 if (!has_node(use)) continue; // Ignore dead nodes 1765 if (use->in(0) == C->top()) continue; 1766 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1767 // Check for data-use outside of loop - at least one of OLD or USE 1768 // must not be a CFG node. 1769 if( !loop->is_member( use_loop ) && (!old->is_CFG() || !use->is_CFG())) { 1770 1771 // If the Data use is an IF, that means we have an IF outside of the 1772 // loop that is switching on a condition that is set inside of the 1773 // loop. Happens if people set a loop-exit flag; then test the flag 1774 // in the loop to break the loop, then test is again outside of the 1775 // loop to determine which way the loop exited. 1776 // Loop predicate If node connects to Bool node through Opaque1 node. 1777 if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use)) { 1778 // Since this code is highly unlikely, we lazily build the worklist 1779 // of such Nodes to go split. 1780 if( !split_if_set ) 1781 split_if_set = new Node_List(area); 1782 split_if_set->push(use); 1783 } 1784 if( use->is_Bool() ) { 1785 if( !split_bool_set ) 1786 split_bool_set = new Node_List(area); 1787 split_bool_set->push(use); 1788 } 1789 if( use->Opcode() == Op_CreateEx ) { 1790 if( !split_cex_set ) 1791 split_cex_set = new Node_List(area); 1792 split_cex_set->push(use); 1793 } 1794 1795 1796 // Get "block" use is in 1797 uint idx = 0; 1798 while( use->in(idx) != old ) idx++; 1799 Node *prev = use->is_CFG() ? use : get_ctrl(use); 1800 assert( !loop->is_member( get_loop( prev ) ), "" ); 1801 Node *cfg = prev->_idx >= new_counter 1802 ? prev->in(2) 1803 : idom(prev); 1804 if( use->is_Phi() ) // Phi use is in prior block 1805 cfg = prev->in(idx); // NOT in block of Phi itself 1806 if (cfg->is_top()) { // Use is dead? 1807 _igvn.replace_input_of(use, idx, C->top()); 1808 continue; 1809 } 1810 1811 while( !loop->is_member( get_loop( cfg ) ) ) { 1812 prev = cfg; 1813 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg); 1814 } 1815 // If the use occurs after merging several exits from the loop, then 1816 // old value must have dominated all those exits. Since the same old 1817 // value was used on all those exits we did not need a Phi at this 1818 // merge point. NOW we do need a Phi here. Each loop exit value 1819 // is now merged with the peeled body exit; each exit gets its own 1820 // private Phi and those Phis need to be merged here. 1821 Node *phi; 1822 if( prev->is_Region() ) { 1823 if( idx == 0 ) { // Updating control edge? 1824 phi = prev; // Just use existing control 1825 } else { // Else need a new Phi 1826 phi = PhiNode::make( prev, old ); 1827 // Now recursively fix up the new uses of old! 1828 for( uint i = 1; i < prev->req(); i++ ) { 1829 worklist.push(phi); // Onto worklist once for each 'old' input 1830 } 1831 } 1832 } else { 1833 // Get new RegionNode merging old and new loop exits 1834 prev = old_new[prev->_idx]; 1835 assert( prev, "just made this in step 7" ); 1836 if( idx == 0 ) { // Updating control edge? 1837 phi = prev; // Just use existing control 1838 } else { // Else need a new Phi 1839 // Make a new Phi merging data values properly 1840 phi = PhiNode::make( prev, old ); 1841 phi->set_req( 1, nnn ); 1842 } 1843 } 1844 // If inserting a new Phi, check for prior hits 1845 if( idx != 0 ) { 1846 Node *hit = _igvn.hash_find_insert(phi); 1847 if( hit == NULL ) { 1848 _igvn.register_new_node_with_optimizer(phi); // Register new phi 1849 } else { // or 1850 // Remove the new phi from the graph and use the hit 1851 _igvn.remove_dead_node(phi); 1852 phi = hit; // Use existing phi 1853 } 1854 set_ctrl(phi, prev); 1855 } 1856 // Make 'use' use the Phi instead of the old loop body exit value 1857 _igvn.replace_input_of(use, idx, phi); 1858 if( use->_idx >= new_counter ) { // If updating new phis 1859 // Not needed for correctness, but prevents a weak assert 1860 // in AddPNode from tripping (when we end up with different 1861 // base & derived Phis that will become the same after 1862 // IGVN does CSE). 1863 Node *hit = _igvn.hash_find_insert(use); 1864 if( hit ) // Go ahead and re-hash for hits. 1865 _igvn.replace_node( use, hit ); 1866 } 1867 1868 // If 'use' was in the loop-exit block, it now needs to be sunk 1869 // below the post-loop merge point. 1870 sink_use( use, prev ); 1871 } 1872 } 1873 } 1874 1875 // Check for IFs that need splitting/cloning. Happens if an IF outside of 1876 // the loop uses a condition set in the loop. The original IF probably 1877 // takes control from one or more OLD Regions (which in turn get from NEW 1878 // Regions). In any case, there will be a set of Phis for each merge point 1879 // from the IF up to where the original BOOL def exists the loop. 1880 if( split_if_set ) { 1881 while( split_if_set->size() ) { 1882 Node *iff = split_if_set->pop(); 1883 if( iff->in(1)->is_Phi() ) { 1884 BoolNode *b = clone_iff( iff->in(1)->as_Phi(), loop ); 1885 _igvn.replace_input_of(iff, 1, b); 1886 } 1887 } 1888 } 1889 if( split_bool_set ) { 1890 while( split_bool_set->size() ) { 1891 Node *b = split_bool_set->pop(); 1892 Node *phi = b->in(1); 1893 assert( phi->is_Phi(), "" ); 1894 CmpNode *cmp = clone_bool( (PhiNode*)phi, loop ); 1895 _igvn.replace_input_of(b, 1, cmp); 1896 } 1897 } 1898 if( split_cex_set ) { 1899 while( split_cex_set->size() ) { 1900 Node *b = split_cex_set->pop(); 1901 assert( b->in(0)->is_Region(), "" ); 1902 assert( b->in(1)->is_Phi(), "" ); 1903 assert( b->in(0)->in(0) == b->in(1)->in(0), "" ); 1904 split_up( b, b->in(0), NULL ); 1905 } 1906 } 1907 1908 } 1909 1910 1911 //---------------------- stride_of_possible_iv ------------------------------------- 1912 // Looks for an iff/bool/comp with one operand of the compare 1913 // being a cycle involving an add and a phi, 1914 // with an optional truncation (left-shift followed by a right-shift) 1915 // of the add. Returns zero if not an iv. 1916 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) { 1917 Node* trunc1 = NULL; 1918 Node* trunc2 = NULL; 1919 const TypeInt* ttype = NULL; 1920 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) { 1921 return 0; 1922 } 1923 BoolNode* bl = iff->in(1)->as_Bool(); 1924 Node* cmp = bl->in(1); 1925 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) { 1926 return 0; 1927 } 1928 // Must have an invariant operand 1929 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) { 1930 return 0; 1931 } 1932 Node* add2 = NULL; 1933 Node* cmp1 = cmp->in(1); 1934 if (cmp1->is_Phi()) { 1935 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) ))) 1936 Node* phi = cmp1; 1937 for (uint i = 1; i < phi->req(); i++) { 1938 Node* in = phi->in(i); 1939 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in, 1940 &trunc1, &trunc2, &ttype); 1941 if (add && add->in(1) == phi) { 1942 add2 = add->in(2); 1943 break; 1944 } 1945 } 1946 } else { 1947 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) ))) 1948 Node* addtrunc = cmp1; 1949 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc, 1950 &trunc1, &trunc2, &ttype); 1951 if (add && add->in(1)->is_Phi()) { 1952 Node* phi = add->in(1); 1953 for (uint i = 1; i < phi->req(); i++) { 1954 if (phi->in(i) == addtrunc) { 1955 add2 = add->in(2); 1956 break; 1957 } 1958 } 1959 } 1960 } 1961 if (add2 != NULL) { 1962 const TypeInt* add2t = _igvn.type(add2)->is_int(); 1963 if (add2t->is_con()) { 1964 return add2t->get_con(); 1965 } 1966 } 1967 return 0; 1968 } 1969 1970 1971 //---------------------- stay_in_loop ------------------------------------- 1972 // Return the (unique) control output node that's in the loop (if it exists.) 1973 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) { 1974 Node* unique = NULL; 1975 if (!n) return NULL; 1976 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1977 Node* use = n->fast_out(i); 1978 if (!has_ctrl(use) && loop->is_member(get_loop(use))) { 1979 if (unique != NULL) { 1980 return NULL; 1981 } 1982 unique = use; 1983 } 1984 } 1985 return unique; 1986 } 1987 1988 //------------------------------ register_node ------------------------------------- 1989 // Utility to register node "n" with PhaseIdealLoop 1990 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) { 1991 _igvn.register_new_node_with_optimizer(n); 1992 loop->_body.push(n); 1993 if (n->is_CFG()) { 1994 set_loop(n, loop); 1995 set_idom(n, pred, ddepth); 1996 } else { 1997 set_ctrl(n, pred); 1998 } 1999 } 2000 2001 //------------------------------ proj_clone ------------------------------------- 2002 // Utility to create an if-projection 2003 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) { 2004 ProjNode* c = p->clone()->as_Proj(); 2005 c->set_req(0, iff); 2006 return c; 2007 } 2008 2009 //------------------------------ short_circuit_if ------------------------------------- 2010 // Force the iff control output to be the live_proj 2011 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) { 2012 guarantee(live_proj != NULL, "null projection"); 2013 int proj_con = live_proj->_con; 2014 assert(proj_con == 0 || proj_con == 1, "false or true projection"); 2015 Node *con = _igvn.intcon(proj_con); 2016 set_ctrl(con, C->root()); 2017 if (iff) { 2018 iff->set_req(1, con); 2019 } 2020 return con; 2021 } 2022 2023 //------------------------------ insert_if_before_proj ------------------------------------- 2024 // Insert a new if before an if projection (* - new node) 2025 // 2026 // before 2027 // if(test) 2028 // / \ 2029 // v v 2030 // other-proj proj (arg) 2031 // 2032 // after 2033 // if(test) 2034 // / \ 2035 // / v 2036 // | * proj-clone 2037 // v | 2038 // other-proj v 2039 // * new_if(relop(cmp[IU](left,right))) 2040 // / \ 2041 // v v 2042 // * new-proj proj 2043 // (returned) 2044 // 2045 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) { 2046 IfNode* iff = proj->in(0)->as_If(); 2047 IdealLoopTree *loop = get_loop(proj); 2048 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2049 int ddepth = dom_depth(proj); 2050 2051 _igvn.rehash_node_delayed(iff); 2052 _igvn.rehash_node_delayed(proj); 2053 2054 proj->set_req(0, NULL); // temporary disconnect 2055 ProjNode* proj2 = proj_clone(proj, iff); 2056 register_node(proj2, loop, iff, ddepth); 2057 2058 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right); 2059 register_node(cmp, loop, proj2, ddepth); 2060 2061 BoolNode* bol = new BoolNode(cmp, relop); 2062 register_node(bol, loop, proj2, ddepth); 2063 2064 int opcode = iff->Opcode(); 2065 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode"); 2066 IfNode* new_if = (opcode == Op_If) ? new IfNode(proj2, bol, iff->_prob, iff->_fcnt): 2067 new RangeCheckNode(proj2, bol, iff->_prob, iff->_fcnt); 2068 register_node(new_if, loop, proj2, ddepth); 2069 2070 proj->set_req(0, new_if); // reattach 2071 set_idom(proj, new_if, ddepth); 2072 2073 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj(); 2074 guarantee(new_exit != NULL, "null exit node"); 2075 register_node(new_exit, get_loop(other_proj), new_if, ddepth); 2076 2077 return new_exit; 2078 } 2079 2080 //------------------------------ insert_region_before_proj ------------------------------------- 2081 // Insert a region before an if projection (* - new node) 2082 // 2083 // before 2084 // if(test) 2085 // / | 2086 // v | 2087 // proj v 2088 // other-proj 2089 // 2090 // after 2091 // if(test) 2092 // / | 2093 // v | 2094 // * proj-clone v 2095 // | other-proj 2096 // v 2097 // * new-region 2098 // | 2099 // v 2100 // * dum_if 2101 // / \ 2102 // v \ 2103 // * dum-proj v 2104 // proj 2105 // 2106 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) { 2107 IfNode* iff = proj->in(0)->as_If(); 2108 IdealLoopTree *loop = get_loop(proj); 2109 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2110 int ddepth = dom_depth(proj); 2111 2112 _igvn.rehash_node_delayed(iff); 2113 _igvn.rehash_node_delayed(proj); 2114 2115 proj->set_req(0, NULL); // temporary disconnect 2116 ProjNode* proj2 = proj_clone(proj, iff); 2117 register_node(proj2, loop, iff, ddepth); 2118 2119 RegionNode* reg = new RegionNode(2); 2120 reg->set_req(1, proj2); 2121 register_node(reg, loop, iff, ddepth); 2122 2123 IfNode* dum_if = new IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt); 2124 register_node(dum_if, loop, reg, ddepth); 2125 2126 proj->set_req(0, dum_if); // reattach 2127 set_idom(proj, dum_if, ddepth); 2128 2129 ProjNode* dum_proj = proj_clone(other_proj, dum_if); 2130 register_node(dum_proj, loop, dum_if, ddepth); 2131 2132 return reg; 2133 } 2134 2135 //------------------------------ insert_cmpi_loop_exit ------------------------------------- 2136 // Clone a signed compare loop exit from an unsigned compare and 2137 // insert it before the unsigned cmp on the stay-in-loop path. 2138 // All new nodes inserted in the dominator tree between the original 2139 // if and it's projections. The original if test is replaced with 2140 // a constant to force the stay-in-loop path. 2141 // 2142 // This is done to make sure that the original if and it's projections 2143 // still dominate the same set of control nodes, that the ctrl() relation 2144 // from data nodes to them is preserved, and that their loop nesting is 2145 // preserved. 2146 // 2147 // before 2148 // if(i <u limit) unsigned compare loop exit 2149 // / | 2150 // v v 2151 // exit-proj stay-in-loop-proj 2152 // 2153 // after 2154 // if(stay-in-loop-const) original if 2155 // / | 2156 // / v 2157 // / if(i < limit) new signed test 2158 // / / | 2159 // / / v 2160 // / / if(i <u limit) new cloned unsigned test 2161 // / / / | 2162 // v v v | 2163 // region | 2164 // | | 2165 // dum-if | 2166 // / | | 2167 // ether | | 2168 // v v 2169 // exit-proj stay-in-loop-proj 2170 // 2171 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) { 2172 const bool Signed = true; 2173 const bool Unsigned = false; 2174 2175 BoolNode* bol = if_cmpu->in(1)->as_Bool(); 2176 if (bol->_test._test != BoolTest::lt) return NULL; 2177 CmpNode* cmpu = bol->in(1)->as_Cmp(); 2178 if (cmpu->Opcode() != Op_CmpU) return NULL; 2179 int stride = stride_of_possible_iv(if_cmpu); 2180 if (stride == 0) return NULL; 2181 2182 Node* lp_proj = stay_in_loop(if_cmpu, loop); 2183 guarantee(lp_proj != NULL, "null loop node"); 2184 2185 ProjNode* lp_continue = lp_proj->as_Proj(); 2186 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj(); 2187 2188 Node* limit = NULL; 2189 if (stride > 0) { 2190 limit = cmpu->in(2); 2191 } else { 2192 limit = _igvn.makecon(TypeInt::ZERO); 2193 set_ctrl(limit, C->root()); 2194 } 2195 // Create a new region on the exit path 2196 RegionNode* reg = insert_region_before_proj(lp_exit); 2197 guarantee(reg != NULL, "null region node"); 2198 2199 // Clone the if-cmpu-true-false using a signed compare 2200 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge; 2201 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue); 2202 reg->add_req(cmpi_exit); 2203 2204 // Clone the if-cmpu-true-false 2205 BoolTest::mask rel_u = bol->_test._test; 2206 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue); 2207 reg->add_req(cmpu_exit); 2208 2209 // Force original if to stay in loop. 2210 short_circuit_if(if_cmpu, lp_continue); 2211 2212 return cmpi_exit->in(0)->as_If(); 2213 } 2214 2215 //------------------------------ remove_cmpi_loop_exit ------------------------------------- 2216 // Remove a previously inserted signed compare loop exit. 2217 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) { 2218 Node* lp_proj = stay_in_loop(if_cmp, loop); 2219 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI && 2220 stay_in_loop(lp_proj, loop)->is_If() && 2221 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu"); 2222 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO); 2223 set_ctrl(con, C->root()); 2224 if_cmp->set_req(1, con); 2225 } 2226 2227 //------------------------------ scheduled_nodelist ------------------------------------- 2228 // Create a post order schedule of nodes that are in the 2229 // "member" set. The list is returned in "sched". 2230 // The first node in "sched" is the loop head, followed by 2231 // nodes which have no inputs in the "member" set, and then 2232 // followed by the nodes that have an immediate input dependence 2233 // on a node in "sched". 2234 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) { 2235 2236 assert(member.test(loop->_head->_idx), "loop head must be in member set"); 2237 Arena *a = Thread::current()->resource_area(); 2238 VectorSet visited(a); 2239 Node_Stack nstack(a, loop->_body.size()); 2240 2241 Node* n = loop->_head; // top of stack is cached in "n" 2242 uint idx = 0; 2243 visited.set(n->_idx); 2244 2245 // Initially push all with no inputs from within member set 2246 for(uint i = 0; i < loop->_body.size(); i++ ) { 2247 Node *elt = loop->_body.at(i); 2248 if (member.test(elt->_idx)) { 2249 bool found = false; 2250 for (uint j = 0; j < elt->req(); j++) { 2251 Node* def = elt->in(j); 2252 if (def && member.test(def->_idx) && def != elt) { 2253 found = true; 2254 break; 2255 } 2256 } 2257 if (!found && elt != loop->_head) { 2258 nstack.push(n, idx); 2259 n = elt; 2260 assert(!visited.test(n->_idx), "not seen yet"); 2261 visited.set(n->_idx); 2262 } 2263 } 2264 } 2265 2266 // traverse out's that are in the member set 2267 while (true) { 2268 if (idx < n->outcnt()) { 2269 Node* use = n->raw_out(idx); 2270 idx++; 2271 if (!visited.test_set(use->_idx)) { 2272 if (member.test(use->_idx)) { 2273 nstack.push(n, idx); 2274 n = use; 2275 idx = 0; 2276 } 2277 } 2278 } else { 2279 // All outputs processed 2280 sched.push(n); 2281 if (nstack.is_empty()) break; 2282 n = nstack.node(); 2283 idx = nstack.index(); 2284 nstack.pop(); 2285 } 2286 } 2287 } 2288 2289 2290 //------------------------------ has_use_in_set ------------------------------------- 2291 // Has a use in the vector set 2292 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) { 2293 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2294 Node* use = n->fast_out(j); 2295 if (vset.test(use->_idx)) { 2296 return true; 2297 } 2298 } 2299 return false; 2300 } 2301 2302 2303 //------------------------------ has_use_internal_to_set ------------------------------------- 2304 // Has use internal to the vector set (ie. not in a phi at the loop head) 2305 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) { 2306 Node* head = loop->_head; 2307 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2308 Node* use = n->fast_out(j); 2309 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) { 2310 return true; 2311 } 2312 } 2313 return false; 2314 } 2315 2316 2317 //------------------------------ clone_for_use_outside_loop ------------------------------------- 2318 // clone "n" for uses that are outside of loop 2319 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) { 2320 int cloned = 0; 2321 assert(worklist.size() == 0, "should be empty"); 2322 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2323 Node* use = n->fast_out(j); 2324 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) { 2325 worklist.push(use); 2326 } 2327 } 2328 while( worklist.size() ) { 2329 Node *use = worklist.pop(); 2330 if (!has_node(use) || use->in(0) == C->top()) continue; 2331 uint j; 2332 for (j = 0; j < use->req(); j++) { 2333 if (use->in(j) == n) break; 2334 } 2335 assert(j < use->req(), "must be there"); 2336 2337 // clone "n" and insert it between the inputs of "n" and the use outside the loop 2338 Node* n_clone = n->clone(); 2339 _igvn.replace_input_of(use, j, n_clone); 2340 cloned++; 2341 Node* use_c; 2342 if (!use->is_Phi()) { 2343 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0); 2344 } else { 2345 // Use in a phi is considered a use in the associated predecessor block 2346 use_c = use->in(0)->in(j); 2347 } 2348 set_ctrl(n_clone, use_c); 2349 assert(!loop->is_member(get_loop(use_c)), "should be outside loop"); 2350 get_loop(use_c)->_body.push(n_clone); 2351 _igvn.register_new_node_with_optimizer(n_clone); 2352 #if !defined(PRODUCT) 2353 if (TracePartialPeeling) { 2354 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx); 2355 } 2356 #endif 2357 } 2358 return cloned; 2359 } 2360 2361 2362 //------------------------------ clone_for_special_use_inside_loop ------------------------------------- 2363 // clone "n" for special uses that are in the not_peeled region. 2364 // If these def-uses occur in separate blocks, the code generator 2365 // marks the method as not compilable. For example, if a "BoolNode" 2366 // is in a different basic block than the "IfNode" that uses it, then 2367 // the compilation is aborted in the code generator. 2368 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 2369 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) { 2370 if (n->is_Phi() || n->is_Load()) { 2371 return; 2372 } 2373 assert(worklist.size() == 0, "should be empty"); 2374 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2375 Node* use = n->fast_out(j); 2376 if ( not_peel.test(use->_idx) && 2377 (use->is_If() || use->is_CMove() || use->is_Bool()) && 2378 use->in(1) == n) { 2379 worklist.push(use); 2380 } 2381 } 2382 if (worklist.size() > 0) { 2383 // clone "n" and insert it between inputs of "n" and the use 2384 Node* n_clone = n->clone(); 2385 loop->_body.push(n_clone); 2386 _igvn.register_new_node_with_optimizer(n_clone); 2387 set_ctrl(n_clone, get_ctrl(n)); 2388 sink_list.push(n_clone); 2389 not_peel <<= n_clone->_idx; // add n_clone to not_peel set. 2390 #if !defined(PRODUCT) 2391 if (TracePartialPeeling) { 2392 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx); 2393 } 2394 #endif 2395 while( worklist.size() ) { 2396 Node *use = worklist.pop(); 2397 _igvn.rehash_node_delayed(use); 2398 for (uint j = 1; j < use->req(); j++) { 2399 if (use->in(j) == n) { 2400 use->set_req(j, n_clone); 2401 } 2402 } 2403 } 2404 } 2405 } 2406 2407 2408 //------------------------------ insert_phi_for_loop ------------------------------------- 2409 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 2410 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) { 2411 Node *phi = PhiNode::make(lp, back_edge_val); 2412 phi->set_req(LoopNode::EntryControl, lp_entry_val); 2413 // Use existing phi if it already exists 2414 Node *hit = _igvn.hash_find_insert(phi); 2415 if( hit == NULL ) { 2416 _igvn.register_new_node_with_optimizer(phi); 2417 set_ctrl(phi, lp); 2418 } else { 2419 // Remove the new phi from the graph and use the hit 2420 _igvn.remove_dead_node(phi); 2421 phi = hit; 2422 } 2423 _igvn.replace_input_of(use, idx, phi); 2424 } 2425 2426 #ifdef ASSERT 2427 //------------------------------ is_valid_loop_partition ------------------------------------- 2428 // Validate the loop partition sets: peel and not_peel 2429 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, 2430 VectorSet& not_peel ) { 2431 uint i; 2432 // Check that peel_list entries are in the peel set 2433 for (i = 0; i < peel_list.size(); i++) { 2434 if (!peel.test(peel_list.at(i)->_idx)) { 2435 return false; 2436 } 2437 } 2438 // Check at loop members are in one of peel set or not_peel set 2439 for (i = 0; i < loop->_body.size(); i++ ) { 2440 Node *def = loop->_body.at(i); 2441 uint di = def->_idx; 2442 // Check that peel set elements are in peel_list 2443 if (peel.test(di)) { 2444 if (not_peel.test(di)) { 2445 return false; 2446 } 2447 // Must be in peel_list also 2448 bool found = false; 2449 for (uint j = 0; j < peel_list.size(); j++) { 2450 if (peel_list.at(j)->_idx == di) { 2451 found = true; 2452 break; 2453 } 2454 } 2455 if (!found) { 2456 return false; 2457 } 2458 } else if (not_peel.test(di)) { 2459 if (peel.test(di)) { 2460 return false; 2461 } 2462 } else { 2463 return false; 2464 } 2465 } 2466 return true; 2467 } 2468 2469 //------------------------------ is_valid_clone_loop_exit_use ------------------------------------- 2470 // Ensure a use outside of loop is of the right form 2471 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) { 2472 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2473 return (use->is_Phi() && 2474 use_c->is_Region() && use_c->req() == 3 && 2475 (use_c->in(exit_idx)->Opcode() == Op_IfTrue || 2476 use_c->in(exit_idx)->Opcode() == Op_IfFalse || 2477 use_c->in(exit_idx)->Opcode() == Op_JumpProj) && 2478 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) ); 2479 } 2480 2481 //------------------------------ is_valid_clone_loop_form ------------------------------------- 2482 // Ensure that all uses outside of loop are of the right form 2483 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 2484 uint orig_exit_idx, uint clone_exit_idx) { 2485 uint len = peel_list.size(); 2486 for (uint i = 0; i < len; i++) { 2487 Node *def = peel_list.at(i); 2488 2489 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 2490 Node *use = def->fast_out(j); 2491 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2492 if (!loop->is_member(get_loop(use_c))) { 2493 // use is not in the loop, check for correct structure 2494 if (use->in(0) == def) { 2495 // Okay 2496 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) { 2497 return false; 2498 } 2499 } 2500 } 2501 } 2502 return true; 2503 } 2504 #endif 2505 2506 //------------------------------ partial_peel ------------------------------------- 2507 // Partially peel (aka loop rotation) the top portion of a loop (called 2508 // the peel section below) by cloning it and placing one copy just before 2509 // the new loop head and the other copy at the bottom of the new loop. 2510 // 2511 // before after where it came from 2512 // 2513 // stmt1 stmt1 2514 // loop: stmt2 clone 2515 // stmt2 if condA goto exitA clone 2516 // if condA goto exitA new_loop: new 2517 // stmt3 stmt3 clone 2518 // if !condB goto loop if condB goto exitB clone 2519 // exitB: stmt2 orig 2520 // stmt4 if !condA goto new_loop orig 2521 // exitA: goto exitA 2522 // exitB: 2523 // stmt4 2524 // exitA: 2525 // 2526 // Step 1: find the cut point: an exit test on probable 2527 // induction variable. 2528 // Step 2: schedule (with cloning) operations in the peel 2529 // section that can be executed after the cut into 2530 // the section that is not peeled. This may need 2531 // to clone operations into exit blocks. For 2532 // instance, a reference to A[i] in the not-peel 2533 // section and a reference to B[i] in an exit block 2534 // may cause a left-shift of i by 2 to be placed 2535 // in the peel block. This step will clone the left 2536 // shift into the exit block and sink the left shift 2537 // from the peel to the not-peel section. 2538 // Step 3: clone the loop, retarget the control, and insert 2539 // phis for values that are live across the new loop 2540 // head. This is very dependent on the graph structure 2541 // from clone_loop. It creates region nodes for 2542 // exit control and associated phi nodes for values 2543 // flow out of the loop through that exit. The region 2544 // node is dominated by the clone's control projection. 2545 // So the clone's peel section is placed before the 2546 // new loop head, and the clone's not-peel section is 2547 // forms the top part of the new loop. The original 2548 // peel section forms the tail of the new loop. 2549 // Step 4: update the dominator tree and recompute the 2550 // dominator depth. 2551 // 2552 // orig 2553 // 2554 // stmt1 2555 // | 2556 // v 2557 // loop predicate 2558 // | 2559 // v 2560 // loop<----+ 2561 // | | 2562 // stmt2 | 2563 // | | 2564 // v | 2565 // ifA | 2566 // / | | 2567 // v v | 2568 // false true ^ <-- last_peel 2569 // / | | 2570 // / ===|==cut | 2571 // / stmt3 | <-- first_not_peel 2572 // / | | 2573 // | v | 2574 // v ifB | 2575 // exitA: / \ | 2576 // / \ | 2577 // v v | 2578 // false true | 2579 // / \ | 2580 // / ----+ 2581 // | 2582 // v 2583 // exitB: 2584 // stmt4 2585 // 2586 // 2587 // after clone loop 2588 // 2589 // stmt1 2590 // | 2591 // v 2592 // loop predicate 2593 // / \ 2594 // clone / \ orig 2595 // / \ 2596 // / \ 2597 // v v 2598 // +---->loop loop<----+ 2599 // | | | | 2600 // | stmt2 stmt2 | 2601 // | | | | 2602 // | v v | 2603 // | ifA ifA | 2604 // | | \ / | | 2605 // | v v v v | 2606 // ^ true false false true ^ <-- last_peel 2607 // | | ^ \ / | | 2608 // | cut==|== \ \ / ===|==cut | 2609 // | stmt3 \ \ / stmt3 | <-- first_not_peel 2610 // | | dom | | | | 2611 // | v \ 1v v2 v | 2612 // | ifB regionA ifB | 2613 // | / \ | / \ | 2614 // | / \ v / \ | 2615 // | v v exitA: v v | 2616 // | true false false true | 2617 // | / ^ \ / \ | 2618 // +---- \ \ / ----+ 2619 // dom \ / 2620 // \ 1v v2 2621 // regionB 2622 // | 2623 // v 2624 // exitB: 2625 // stmt4 2626 // 2627 // 2628 // after partial peel 2629 // 2630 // stmt1 2631 // | 2632 // v 2633 // loop predicate 2634 // / 2635 // clone / orig 2636 // / TOP 2637 // / \ 2638 // v v 2639 // TOP->loop loop----+ 2640 // | | | 2641 // stmt2 stmt2 | 2642 // | | | 2643 // v v | 2644 // ifA ifA | 2645 // | \ / | | 2646 // v v v v | 2647 // true false false true | <-- last_peel 2648 // | ^ \ / +------|---+ 2649 // +->newloop \ \ / === ==cut | | 2650 // | stmt3 \ \ / TOP | | 2651 // | | dom | | stmt3 | | <-- first_not_peel 2652 // | v \ 1v v2 v | | 2653 // | ifB regionA ifB ^ v 2654 // | / \ | / \ | | 2655 // | / \ v / \ | | 2656 // | v v exitA: v v | | 2657 // | true false false true | | 2658 // | / ^ \ / \ | | 2659 // | | \ \ / v | | 2660 // | | dom \ / TOP | | 2661 // | | \ 1v v2 | | 2662 // ^ v regionB | | 2663 // | | | | | 2664 // | | v ^ v 2665 // | | exitB: | | 2666 // | | stmt4 | | 2667 // | +------------>-----------------+ | 2668 // | | 2669 // +-----------------<---------------------+ 2670 // 2671 // 2672 // final graph 2673 // 2674 // stmt1 2675 // | 2676 // v 2677 // loop predicate 2678 // | 2679 // v 2680 // stmt2 clone 2681 // | 2682 // v 2683 // ........> ifA clone 2684 // : / | 2685 // dom / | 2686 // : v v 2687 // : false true 2688 // : | | 2689 // : | v 2690 // : | newloop<-----+ 2691 // : | | | 2692 // : | stmt3 clone | 2693 // : | | | 2694 // : | v | 2695 // : | ifB | 2696 // : | / \ | 2697 // : | v v | 2698 // : | false true | 2699 // : | | | | 2700 // : | v stmt2 | 2701 // : | exitB: | | 2702 // : | stmt4 v | 2703 // : | ifA orig | 2704 // : | / \ | 2705 // : | / \ | 2706 // : | v v | 2707 // : | false true | 2708 // : | / \ | 2709 // : v v -----+ 2710 // RegionA 2711 // | 2712 // v 2713 // exitA 2714 // 2715 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) { 2716 2717 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only"); 2718 if (!loop->_head->is_Loop()) { 2719 return false; } 2720 2721 LoopNode *head = loop->_head->as_Loop(); 2722 2723 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) { 2724 return false; 2725 } 2726 2727 // Check for complex exit control 2728 for(uint ii = 0; ii < loop->_body.size(); ii++ ) { 2729 Node *n = loop->_body.at(ii); 2730 int opc = n->Opcode(); 2731 if (n->is_Call() || 2732 opc == Op_Catch || 2733 opc == Op_CatchProj || 2734 opc == Op_Jump || 2735 opc == Op_JumpProj) { 2736 #if !defined(PRODUCT) 2737 if (TracePartialPeeling) { 2738 tty->print_cr("\nExit control too complex: lp: %d", head->_idx); 2739 } 2740 #endif 2741 return false; 2742 } 2743 } 2744 2745 int dd = dom_depth(head); 2746 2747 // Step 1: find cut point 2748 2749 // Walk up dominators to loop head looking for first loop exit 2750 // which is executed on every path thru loop. 2751 IfNode *peel_if = NULL; 2752 IfNode *peel_if_cmpu = NULL; 2753 2754 Node *iff = loop->tail(); 2755 while( iff != head ) { 2756 if( iff->is_If() ) { 2757 Node *ctrl = get_ctrl(iff->in(1)); 2758 if (ctrl->is_top()) return false; // Dead test on live IF. 2759 // If loop-varying exit-test, check for induction variable 2760 if( loop->is_member(get_loop(ctrl)) && 2761 loop->is_loop_exit(iff) && 2762 is_possible_iv_test(iff)) { 2763 Node* cmp = iff->in(1)->in(1); 2764 if (cmp->Opcode() == Op_CmpI) { 2765 peel_if = iff->as_If(); 2766 } else { 2767 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU"); 2768 peel_if_cmpu = iff->as_If(); 2769 } 2770 } 2771 } 2772 iff = idom(iff); 2773 } 2774 // Prefer signed compare over unsigned compare. 2775 IfNode* new_peel_if = NULL; 2776 if (peel_if == NULL) { 2777 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) { 2778 return false; // No peel point found 2779 } 2780 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop); 2781 if (new_peel_if == NULL) { 2782 return false; // No peel point found 2783 } 2784 peel_if = new_peel_if; 2785 } 2786 Node* last_peel = stay_in_loop(peel_if, loop); 2787 Node* first_not_peeled = stay_in_loop(last_peel, loop); 2788 if (first_not_peeled == NULL || first_not_peeled == head) { 2789 return false; 2790 } 2791 2792 #if !defined(PRODUCT) 2793 if (TraceLoopOpts) { 2794 tty->print("PartialPeel "); 2795 loop->dump_head(); 2796 } 2797 2798 if (TracePartialPeeling) { 2799 tty->print_cr("before partial peel one iteration"); 2800 Node_List wl; 2801 Node* t = head->in(2); 2802 while (true) { 2803 wl.push(t); 2804 if (t == head) break; 2805 t = idom(t); 2806 } 2807 while (wl.size() > 0) { 2808 Node* tt = wl.pop(); 2809 tt->dump(); 2810 if (tt == last_peel) tty->print_cr("-- cut --"); 2811 } 2812 } 2813 #endif 2814 ResourceArea *area = Thread::current()->resource_area(); 2815 VectorSet peel(area); 2816 VectorSet not_peel(area); 2817 Node_List peel_list(area); 2818 Node_List worklist(area); 2819 Node_List sink_list(area); 2820 2821 // Set of cfg nodes to peel are those that are executable from 2822 // the head through last_peel. 2823 assert(worklist.size() == 0, "should be empty"); 2824 worklist.push(head); 2825 peel.set(head->_idx); 2826 while (worklist.size() > 0) { 2827 Node *n = worklist.pop(); 2828 if (n != last_peel) { 2829 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2830 Node* use = n->fast_out(j); 2831 if (use->is_CFG() && 2832 loop->is_member(get_loop(use)) && 2833 !peel.test_set(use->_idx)) { 2834 worklist.push(use); 2835 } 2836 } 2837 } 2838 } 2839 2840 // Set of non-cfg nodes to peel are those that are control 2841 // dependent on the cfg nodes. 2842 uint i; 2843 for(i = 0; i < loop->_body.size(); i++ ) { 2844 Node *n = loop->_body.at(i); 2845 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n; 2846 if (peel.test(n_c->_idx)) { 2847 peel.set(n->_idx); 2848 } else { 2849 not_peel.set(n->_idx); 2850 } 2851 } 2852 2853 // Step 2: move operations from the peeled section down into the 2854 // not-peeled section 2855 2856 // Get a post order schedule of nodes in the peel region 2857 // Result in right-most operand. 2858 scheduled_nodelist(loop, peel, peel_list ); 2859 2860 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 2861 2862 // For future check for too many new phis 2863 uint old_phi_cnt = 0; 2864 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 2865 Node* use = head->fast_out(j); 2866 if (use->is_Phi()) old_phi_cnt++; 2867 } 2868 2869 #if !defined(PRODUCT) 2870 if (TracePartialPeeling) { 2871 tty->print_cr("\npeeled list"); 2872 } 2873 #endif 2874 2875 // Evacuate nodes in peel region into the not_peeled region if possible 2876 uint new_phi_cnt = 0; 2877 uint cloned_for_outside_use = 0; 2878 for (i = 0; i < peel_list.size();) { 2879 Node* n = peel_list.at(i); 2880 #if !defined(PRODUCT) 2881 if (TracePartialPeeling) n->dump(); 2882 #endif 2883 bool incr = true; 2884 if ( !n->is_CFG() ) { 2885 2886 if ( has_use_in_set(n, not_peel) ) { 2887 2888 // If not used internal to the peeled region, 2889 // move "n" from peeled to not_peeled region. 2890 2891 if ( !has_use_internal_to_set(n, peel, loop) ) { 2892 2893 // if not pinned and not a load (which maybe anti-dependent on a store) 2894 // and not a CMove (Matcher expects only bool->cmove). 2895 if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) { 2896 cloned_for_outside_use += clone_for_use_outside_loop( loop, n, worklist ); 2897 sink_list.push(n); 2898 peel >>= n->_idx; // delete n from peel set. 2899 not_peel <<= n->_idx; // add n to not_peel set. 2900 peel_list.remove(i); 2901 incr = false; 2902 #if !defined(PRODUCT) 2903 if (TracePartialPeeling) { 2904 tty->print_cr("sink to not_peeled region: %d newbb: %d", 2905 n->_idx, get_ctrl(n)->_idx); 2906 } 2907 #endif 2908 } 2909 } else { 2910 // Otherwise check for special def-use cases that span 2911 // the peel/not_peel boundary such as bool->if 2912 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist ); 2913 new_phi_cnt++; 2914 } 2915 } 2916 } 2917 if (incr) i++; 2918 } 2919 2920 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) { 2921 #if !defined(PRODUCT) 2922 if (TracePartialPeeling) { 2923 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c", 2924 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F'); 2925 } 2926 #endif 2927 if (new_peel_if != NULL) { 2928 remove_cmpi_loop_exit(new_peel_if, loop); 2929 } 2930 // Inhibit more partial peeling on this loop 2931 assert(!head->is_partial_peel_loop(), "not partial peeled"); 2932 head->mark_partial_peel_failed(); 2933 if (cloned_for_outside_use > 0) { 2934 // Terminate this round of loop opts because 2935 // the graph outside this loop was changed. 2936 C->set_major_progress(); 2937 return true; 2938 } 2939 return false; 2940 } 2941 2942 // Step 3: clone loop, retarget control, and insert new phis 2943 2944 // Create new loop head for new phis and to hang 2945 // the nodes being moved (sinked) from the peel region. 2946 LoopNode* new_head = new LoopNode(last_peel, last_peel); 2947 new_head->set_unswitch_count(head->unswitch_count()); // Preserve 2948 _igvn.register_new_node_with_optimizer(new_head); 2949 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled"); 2950 _igvn.replace_input_of(first_not_peeled, 0, new_head); 2951 set_loop(new_head, loop); 2952 loop->_body.push(new_head); 2953 not_peel.set(new_head->_idx); 2954 set_idom(new_head, last_peel, dom_depth(first_not_peeled)); 2955 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled)); 2956 2957 while (sink_list.size() > 0) { 2958 Node* n = sink_list.pop(); 2959 set_ctrl(n, new_head); 2960 } 2961 2962 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 2963 2964 clone_loop( loop, old_new, dd ); 2965 2966 const uint clone_exit_idx = 1; 2967 const uint orig_exit_idx = 2; 2968 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop"); 2969 2970 Node* head_clone = old_new[head->_idx]; 2971 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop(); 2972 Node* orig_tail_clone = head_clone->in(2); 2973 2974 // Add phi if "def" node is in peel set and "use" is not 2975 2976 for(i = 0; i < peel_list.size(); i++ ) { 2977 Node *def = peel_list.at(i); 2978 if (!def->is_CFG()) { 2979 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 2980 Node *use = def->fast_out(j); 2981 if (has_node(use) && use->in(0) != C->top() && 2982 (!peel.test(use->_idx) || 2983 (use->is_Phi() && use->in(0) == head)) ) { 2984 worklist.push(use); 2985 } 2986 } 2987 while( worklist.size() ) { 2988 Node *use = worklist.pop(); 2989 for (uint j = 1; j < use->req(); j++) { 2990 Node* n = use->in(j); 2991 if (n == def) { 2992 2993 // "def" is in peel set, "use" is not in peel set 2994 // or "use" is in the entry boundary (a phi) of the peel set 2995 2996 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use; 2997 2998 if ( loop->is_member(get_loop( use_c )) ) { 2999 // use is in loop 3000 if (old_new[use->_idx] != NULL) { // null for dead code 3001 Node* use_clone = old_new[use->_idx]; 3002 _igvn.replace_input_of(use, j, C->top()); 3003 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone ); 3004 } 3005 } else { 3006 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format"); 3007 // use is not in the loop, check if the live range includes the cut 3008 Node* lp_if = use_c->in(orig_exit_idx)->in(0); 3009 if (not_peel.test(lp_if->_idx)) { 3010 assert(j == orig_exit_idx, "use from original loop"); 3011 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone ); 3012 } 3013 } 3014 } 3015 } 3016 } 3017 } 3018 } 3019 3020 // Step 3b: retarget control 3021 3022 // Redirect control to the new loop head if a cloned node in 3023 // the not_peeled region has control that points into the peeled region. 3024 // This necessary because the cloned peeled region will be outside 3025 // the loop. 3026 // from to 3027 // cloned-peeled <---+ 3028 // new_head_clone: | <--+ 3029 // cloned-not_peeled in(0) in(0) 3030 // orig-peeled 3031 3032 for(i = 0; i < loop->_body.size(); i++ ) { 3033 Node *n = loop->_body.at(i); 3034 if (!n->is_CFG() && n->in(0) != NULL && 3035 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) { 3036 Node* n_clone = old_new[n->_idx]; 3037 _igvn.replace_input_of(n_clone, 0, new_head_clone); 3038 } 3039 } 3040 3041 // Backedge of the surviving new_head (the clone) is original last_peel 3042 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel); 3043 3044 // Cut first node in original not_peel set 3045 _igvn.rehash_node_delayed(new_head); // Multiple edge updates: 3046 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of 3047 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls 3048 3049 // Copy head_clone back-branch info to original head 3050 // and remove original head's loop entry and 3051 // clone head's back-branch 3052 _igvn.rehash_node_delayed(head); // Multiple edge updates 3053 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl)); 3054 head->set_req(LoopNode::LoopBackControl, C->top()); 3055 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top()); 3056 3057 // Similarly modify the phis 3058 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) { 3059 Node* use = head->fast_out(k); 3060 if (use->is_Phi() && use->outcnt() > 0) { 3061 Node* use_clone = old_new[use->_idx]; 3062 _igvn.rehash_node_delayed(use); // Multiple edge updates 3063 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl)); 3064 use->set_req(LoopNode::LoopBackControl, C->top()); 3065 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top()); 3066 } 3067 } 3068 3069 // Step 4: update dominator tree and dominator depth 3070 3071 set_idom(head, orig_tail_clone, dd); 3072 recompute_dom_depth(); 3073 3074 // Inhibit more partial peeling on this loop 3075 new_head_clone->set_partial_peel_loop(); 3076 C->set_major_progress(); 3077 loop->record_for_igvn(); 3078 3079 #if !defined(PRODUCT) 3080 if (TracePartialPeeling) { 3081 tty->print_cr("\nafter partial peel one iteration"); 3082 Node_List wl(area); 3083 Node* t = last_peel; 3084 while (true) { 3085 wl.push(t); 3086 if (t == head_clone) break; 3087 t = idom(t); 3088 } 3089 while (wl.size() > 0) { 3090 Node* tt = wl.pop(); 3091 if (tt == head) tty->print_cr("orig head"); 3092 else if (tt == new_head_clone) tty->print_cr("new head"); 3093 else if (tt == head_clone) tty->print_cr("clone head"); 3094 tt->dump(); 3095 } 3096 } 3097 #endif 3098 return true; 3099 } 3100 3101 //------------------------------reorg_offsets---------------------------------- 3102 // Reorganize offset computations to lower register pressure. Mostly 3103 // prevent loop-fallout uses of the pre-incremented trip counter (which are 3104 // then alive with the post-incremented trip counter forcing an extra 3105 // register move) 3106 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) { 3107 // Perform it only for canonical counted loops. 3108 // Loop's shape could be messed up by iteration_split_impl. 3109 if (!loop->_head->is_CountedLoop()) 3110 return; 3111 if (!loop->_head->as_Loop()->is_valid_counted_loop()) 3112 return; 3113 3114 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 3115 CountedLoopEndNode *cle = cl->loopexit(); 3116 Node *exit = cle->proj_out(false); 3117 Node *phi = cl->phi(); 3118 3119 // Check for the special case of folks using the pre-incremented 3120 // trip-counter on the fall-out path (forces the pre-incremented 3121 // and post-incremented trip counter to be live at the same time). 3122 // Fix this by adjusting to use the post-increment trip counter. 3123 3124 bool progress = true; 3125 while (progress) { 3126 progress = false; 3127 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) { 3128 Node* use = phi->fast_out(i); // User of trip-counter 3129 if (!has_ctrl(use)) continue; 3130 Node *u_ctrl = get_ctrl(use); 3131 if (use->is_Phi()) { 3132 u_ctrl = NULL; 3133 for (uint j = 1; j < use->req(); j++) 3134 if (use->in(j) == phi) 3135 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j)); 3136 } 3137 IdealLoopTree *u_loop = get_loop(u_ctrl); 3138 // Look for loop-invariant use 3139 if (u_loop == loop) continue; 3140 if (loop->is_member(u_loop)) continue; 3141 // Check that use is live out the bottom. Assuming the trip-counter 3142 // update is right at the bottom, uses of of the loop middle are ok. 3143 if (dom_lca(exit, u_ctrl) != exit) continue; 3144 // Hit! Refactor use to use the post-incremented tripcounter. 3145 // Compute a post-increment tripcounter. 3146 Node *opaq = new Opaque2Node( C, cle->incr() ); 3147 register_new_node(opaq, exit); 3148 Node *neg_stride = _igvn.intcon(-cle->stride_con()); 3149 set_ctrl(neg_stride, C->root()); 3150 Node *post = new AddINode( opaq, neg_stride); 3151 register_new_node(post, exit); 3152 _igvn.rehash_node_delayed(use); 3153 for (uint j = 1; j < use->req(); j++) { 3154 if (use->in(j) == phi) 3155 use->set_req(j, post); 3156 } 3157 // Since DU info changed, rerun loop 3158 progress = true; 3159 break; 3160 } 3161 } 3162 3163 }