1 /* 2 * Copyright (c) 1999, 2016, 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(this); 923 if (dom_cast != NULL) { 924 _igvn.replace_node(n, dom_cast); 925 return dom_cast; 926 } 927 } 928 929 // Determine if the Node has inputs from some local Phi. 930 // Returns the block to clone thru. 931 Node *n_blk = has_local_phi_input( n ); 932 if( !n_blk ) return n; 933 934 // Do not clone the trip counter through on a CountedLoop 935 // (messes up the canonical shape). 936 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n; 937 938 // Check for having no control input; not pinned. Allow 939 // dominating control. 940 if (n->in(0)) { 941 Node *dom = idom(n_blk); 942 if (dom_lca(n->in(0), dom) != n->in(0)) { 943 return n; 944 } 945 } 946 // Policy: when is it profitable. You must get more wins than 947 // policy before it is considered profitable. Policy is usually 0, 948 // so 1 win is considered profitable. Big merges will require big 949 // cloning, so get a larger policy. 950 int policy = n_blk->req() >> 2; 951 952 // If the loop is a candidate for range check elimination, 953 // delay splitting through it's phi until a later loop optimization 954 if (n_blk->is_CountedLoop()) { 955 IdealLoopTree *lp = get_loop(n_blk); 956 if (lp && lp->_rce_candidate) { 957 return n; 958 } 959 } 960 961 // Use same limit as split_if_with_blocks_post 962 if( C->live_nodes() > 35000 ) return n; // Method too big 963 964 // Split 'n' through the merge point if it is profitable 965 Node *phi = split_thru_phi( n, n_blk, policy ); 966 if (!phi) return n; 967 968 // Found a Phi to split thru! 969 // Replace 'n' with the new phi 970 _igvn.replace_node( n, phi ); 971 // Moved a load around the loop, 'en-registering' something. 972 if (n_blk->is_Loop() && n->is_Load() && 973 !phi->in(LoopNode::LoopBackControl)->is_Load()) 974 C->set_major_progress(); 975 976 return phi; 977 } 978 979 static bool merge_point_too_heavy(Compile* C, Node* region) { 980 // Bail out if the region and its phis have too many users. 981 int weight = 0; 982 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 983 weight += region->fast_out(i)->outcnt(); 984 } 985 int nodes_left = C->max_node_limit() - C->live_nodes(); 986 if (weight * 8 > nodes_left) { 987 if (PrintOpto) { 988 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight); 989 } 990 return true; 991 } else { 992 return false; 993 } 994 } 995 996 static bool merge_point_safe(Node* region) { 997 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode 998 // having a PhiNode input. This sidesteps the dangerous case where the split 999 // ConvI2LNode may become TOP if the input Value() does not 1000 // overlap the ConvI2L range, leaving a node which may not dominate its 1001 // uses. 1002 // A better fix for this problem can be found in the BugTraq entry, but 1003 // expediency for Mantis demands this hack. 1004 // 6855164: If the merge point has a FastLockNode with a PhiNode input, we stop 1005 // split_if_with_blocks from splitting a block because we could not move around 1006 // the FastLockNode. 1007 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1008 Node* n = region->fast_out(i); 1009 if (n->is_Phi()) { 1010 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1011 Node* m = n->fast_out(j); 1012 if (m->is_FastLock()) 1013 return false; 1014 #ifdef _LP64 1015 if (m->Opcode() == Op_ConvI2L) 1016 return false; 1017 if (m->is_CastII() && m->isa_CastII()->has_range_check()) { 1018 return false; 1019 } 1020 #endif 1021 } 1022 } 1023 } 1024 return true; 1025 } 1026 1027 1028 //------------------------------place_near_use--------------------------------- 1029 // Place some computation next to use but not inside inner loops. 1030 // For inner loop uses move it to the preheader area. 1031 Node *PhaseIdealLoop::place_near_use( Node *useblock ) const { 1032 IdealLoopTree *u_loop = get_loop( useblock ); 1033 return (u_loop->_irreducible || u_loop->_child) 1034 ? useblock 1035 : u_loop->_head->in(LoopNode::EntryControl); 1036 } 1037 1038 1039 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) { 1040 if (!n->is_If()) { 1041 return false; 1042 } 1043 if (!n->in(0)->is_Region()) { 1044 return false; 1045 } 1046 Node* region = n->in(0); 1047 Node* dom = idom(region); 1048 if (!dom->is_If() || dom->in(1) != n->in(1)) { 1049 return false; 1050 } 1051 IfNode* dom_if = dom->as_If(); 1052 Node* proj_true = dom_if->proj_out(1); 1053 Node* proj_false = dom_if->proj_out(0); 1054 1055 for (uint i = 1; i < region->req(); i++) { 1056 if (is_dominator(proj_true, region->in(i))) { 1057 continue; 1058 } 1059 if (is_dominator(proj_false, region->in(i))) { 1060 continue; 1061 } 1062 return false; 1063 } 1064 1065 return true; 1066 } 1067 1068 bool PhaseIdealLoop::can_split_if(Node *n_ctrl) { 1069 if (C->live_nodes() > 35000) { 1070 return false; // Method too big 1071 } 1072 1073 // Do not do 'split-if' if irreducible loops are present. 1074 if (_has_irreducible_loops) { 1075 return false; 1076 } 1077 1078 if (merge_point_too_heavy(C, n_ctrl)) { 1079 return false; 1080 } 1081 1082 // Do not do 'split-if' if some paths are dead. First do dead code 1083 // elimination and then see if its still profitable. 1084 for (uint i = 1; i < n_ctrl->req(); i++) { 1085 if (n_ctrl->in(i) == C->top()) { 1086 return false; 1087 } 1088 } 1089 1090 // If trying to do a 'Split-If' at the loop head, it is only 1091 // profitable if the cmp folds up on BOTH paths. Otherwise we 1092 // risk peeling a loop forever. 1093 1094 // CNC - Disabled for now. Requires careful handling of loop 1095 // body selection for the cloned code. Also, make sure we check 1096 // for any input path not being in the same loop as n_ctrl. For 1097 // irreducible loops we cannot check for 'n_ctrl->is_Loop()' 1098 // because the alternative loop entry points won't be converted 1099 // into LoopNodes. 1100 IdealLoopTree *n_loop = get_loop(n_ctrl); 1101 for (uint j = 1; j < n_ctrl->req(); j++) { 1102 if (get_loop(n_ctrl->in(j)) != n_loop) { 1103 return false; 1104 } 1105 } 1106 1107 // Check for safety of the merge point. 1108 if (!merge_point_safe(n_ctrl)) { 1109 return false; 1110 } 1111 1112 return true; 1113 } 1114 1115 //------------------------------split_if_with_blocks_post---------------------- 1116 // Do the real work in a non-recursive function. CFG hackery wants to be 1117 // in the post-order, so it can dirty the I-DOM info and not use the dirtied 1118 // info. 1119 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) { 1120 1121 // Cloning Cmp through Phi's involves the split-if transform. 1122 // FastLock is not used by an If 1123 if (n->is_Cmp() && !n->is_FastLock()) { 1124 Node *n_ctrl = get_ctrl(n); 1125 // Determine if the Node has inputs from some local Phi. 1126 // Returns the block to clone thru. 1127 Node *n_blk = has_local_phi_input(n); 1128 if (n_blk != n_ctrl) { 1129 return; 1130 } 1131 1132 if (!can_split_if(n_ctrl)) { 1133 return; 1134 } 1135 1136 if (n->outcnt() != 1) { 1137 return; // Multiple bool's from 1 compare? 1138 } 1139 Node *bol = n->unique_out(); 1140 assert(bol->is_Bool(), "expect a bool here"); 1141 if (bol->outcnt() != 1) { 1142 return;// Multiple branches from 1 compare? 1143 } 1144 Node *iff = bol->unique_out(); 1145 1146 // Check some safety conditions 1147 if (iff->is_If()) { // Classic split-if? 1148 if (iff->in(0) != n_ctrl) { 1149 return; // Compare must be in same blk as if 1150 } 1151 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE 1152 // Can't split CMove with different control edge. 1153 if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) { 1154 return; 1155 } 1156 if (get_ctrl(iff->in(2)) == n_ctrl || 1157 get_ctrl(iff->in(3)) == n_ctrl) { 1158 return; // Inputs not yet split-up 1159 } 1160 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) { 1161 return; // Loop-invar test gates loop-varying CMOVE 1162 } 1163 } else { 1164 return; // some other kind of node, such as an Allocate 1165 } 1166 1167 // When is split-if profitable? Every 'win' on means some control flow 1168 // goes dead, so it's almost always a win. 1169 int policy = 0; 1170 // Split compare 'n' through the merge point if it is profitable 1171 Node *phi = split_thru_phi( n, n_ctrl, policy); 1172 if (!phi) { 1173 return; 1174 } 1175 1176 // Found a Phi to split thru! 1177 // Replace 'n' with the new phi 1178 _igvn.replace_node(n, phi); 1179 1180 // Now split the bool up thru the phi 1181 Node *bolphi = split_thru_phi(bol, n_ctrl, -1); 1182 guarantee(bolphi != NULL, "null boolean phi node"); 1183 1184 _igvn.replace_node(bol, bolphi); 1185 assert(iff->in(1) == bolphi, ""); 1186 1187 if (bolphi->Value(&_igvn)->singleton()) { 1188 return; 1189 } 1190 1191 // Conditional-move? Must split up now 1192 if (!iff->is_If()) { 1193 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1); 1194 _igvn.replace_node(iff, cmovphi); 1195 return; 1196 } 1197 1198 // Now split the IF 1199 do_split_if(iff); 1200 return; 1201 } 1202 1203 // Two identical ifs back to back can be merged 1204 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) { 1205 Node *n_ctrl = n->in(0); 1206 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1)); 1207 IfNode* dom_if = idom(n_ctrl)->as_If(); 1208 Node* proj_true = dom_if->proj_out(1); 1209 Node* proj_false = dom_if->proj_out(0); 1210 Node* con_true = _igvn.makecon(TypeInt::ONE); 1211 Node* con_false = _igvn.makecon(TypeInt::ZERO); 1212 1213 for (uint i = 1; i < n_ctrl->req(); i++) { 1214 if (is_dominator(proj_true, n_ctrl->in(i))) { 1215 bolphi->init_req(i, con_true); 1216 } else { 1217 assert(is_dominator(proj_false, n_ctrl->in(i)), "bad if"); 1218 bolphi->init_req(i, con_false); 1219 } 1220 } 1221 register_new_node(bolphi, n_ctrl); 1222 _igvn.replace_input_of(n, 1, bolphi); 1223 1224 // Now split the IF 1225 do_split_if(n); 1226 return; 1227 } 1228 1229 // Check for an IF ready to split; one that has its 1230 // condition codes input coming from a Phi at the block start. 1231 int n_op = n->Opcode(); 1232 1233 // Check for an IF being dominated by another IF same test 1234 if (n_op == Op_If || 1235 n_op == Op_RangeCheck) { 1236 Node *bol = n->in(1); 1237 uint max = bol->outcnt(); 1238 // Check for same test used more than once? 1239 if (max > 1 && bol->is_Bool()) { 1240 // Search up IDOMs to see if this IF is dominated. 1241 Node *cutoff = get_ctrl(bol); 1242 1243 // Now search up IDOMs till cutoff, looking for a dominating test 1244 Node *prevdom = n; 1245 Node *dom = idom(prevdom); 1246 while (dom != cutoff) { 1247 if (dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom) { 1248 // Replace the dominated test with an obvious true or false. 1249 // Place it on the IGVN worklist for later cleanup. 1250 C->set_major_progress(); 1251 dominated_by(prevdom, n, false, true); 1252 #ifndef PRODUCT 1253 if( VerifyLoopOptimizations ) verify(); 1254 #endif 1255 return; 1256 } 1257 prevdom = dom; 1258 dom = idom(prevdom); 1259 } 1260 } 1261 } 1262 1263 // See if a shared loop-varying computation has no loop-varying uses. 1264 // Happens if something is only used for JVM state in uncommon trap exits, 1265 // like various versions of induction variable+offset. Clone the 1266 // computation per usage to allow it to sink out of the loop. 1267 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about) 1268 Node *n_ctrl = get_ctrl(n); 1269 IdealLoopTree *n_loop = get_loop(n_ctrl); 1270 if( n_loop != _ltree_root ) { 1271 DUIterator_Fast imax, i = n->fast_outs(imax); 1272 for (; i < imax; i++) { 1273 Node* u = n->fast_out(i); 1274 if( !has_ctrl(u) ) break; // Found control user 1275 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 1276 if( u_loop == n_loop ) break; // Found loop-varying use 1277 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop 1278 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003 1279 } 1280 bool did_break = (i < imax); // Did we break out of the previous loop? 1281 if (!did_break && n->outcnt() > 1) { // All uses in outer loops! 1282 Node *late_load_ctrl = NULL; 1283 if (n->is_Load()) { 1284 // If n is a load, get and save the result from get_late_ctrl(), 1285 // to be later used in calculating the control for n's clones. 1286 clear_dom_lca_tags(); 1287 late_load_ctrl = get_late_ctrl(n, n_ctrl); 1288 } 1289 // If n is a load, and the late control is the same as the current 1290 // control, then the cloning of n is a pointless exercise, because 1291 // GVN will ensure that we end up where we started. 1292 if (!n->is_Load() || late_load_ctrl != n_ctrl) { 1293 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) { 1294 Node *u = n->last_out(j); // Clone private computation per use 1295 _igvn.rehash_node_delayed(u); 1296 Node *x = n->clone(); // Clone computation 1297 Node *x_ctrl = NULL; 1298 if( u->is_Phi() ) { 1299 // Replace all uses of normal nodes. Replace Phi uses 1300 // individually, so the separate Nodes can sink down 1301 // different paths. 1302 uint k = 1; 1303 while( u->in(k) != n ) k++; 1304 u->set_req( k, x ); 1305 // x goes next to Phi input path 1306 x_ctrl = u->in(0)->in(k); 1307 --j; 1308 } else { // Normal use 1309 // Replace all uses 1310 for( uint k = 0; k < u->req(); k++ ) { 1311 if( u->in(k) == n ) { 1312 u->set_req( k, x ); 1313 --j; 1314 } 1315 } 1316 x_ctrl = get_ctrl(u); 1317 } 1318 1319 // Find control for 'x' next to use but not inside inner loops. 1320 // For inner loop uses get the preheader area. 1321 x_ctrl = place_near_use(x_ctrl); 1322 1323 if (n->is_Load()) { 1324 // For loads, add a control edge to a CFG node outside of the loop 1325 // to force them to not combine and return back inside the loop 1326 // during GVN optimization (4641526). 1327 // 1328 // Because we are setting the actual control input, factor in 1329 // the result from get_late_ctrl() so we respect any 1330 // anti-dependences. (6233005). 1331 x_ctrl = dom_lca(late_load_ctrl, x_ctrl); 1332 1333 // Don't allow the control input to be a CFG splitting node. 1334 // Such nodes should only have ProjNodes as outs, e.g. IfNode 1335 // should only have IfTrueNode and IfFalseNode (4985384). 1336 x_ctrl = find_non_split_ctrl(x_ctrl); 1337 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone"); 1338 1339 x->set_req(0, x_ctrl); 1340 } 1341 register_new_node(x, x_ctrl); 1342 1343 // Some institutional knowledge is needed here: 'x' is 1344 // yanked because if the optimizer runs GVN on it all the 1345 // cloned x's will common up and undo this optimization and 1346 // be forced back in the loop. This is annoying because it 1347 // makes +VerifyOpto report false-positives on progress. I 1348 // tried setting control edges on the x's to force them to 1349 // not combine, but the matching gets worried when it tries 1350 // to fold a StoreP and an AddP together (as part of an 1351 // address expression) and the AddP and StoreP have 1352 // different controls. 1353 if (!x->is_Load() && !x->is_DecodeNarrowPtr()) _igvn._worklist.yank(x); 1354 } 1355 _igvn.remove_dead_node(n); 1356 } 1357 } 1358 } 1359 } 1360 1361 try_move_store_after_loop(n); 1362 1363 // Remove multiple allocations of the same value type 1364 if (n->is_ValueType() && EliminateAllocations) { 1365 n->as_ValueType()->remove_redundant_allocations(&_igvn, this); 1366 } 1367 1368 // Check for Opaque2's who's loop has disappeared - who's input is in the 1369 // same loop nest as their output. Remove 'em, they are no longer useful. 1370 if( n_op == Op_Opaque2 && 1371 n->in(1) != NULL && 1372 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) { 1373 _igvn.replace_node( n, n->in(1) ); 1374 } 1375 } 1376 1377 //------------------------------split_if_with_blocks--------------------------- 1378 // Check for aggressive application of 'split-if' optimization, 1379 // using basic block level info. 1380 void PhaseIdealLoop::split_if_with_blocks( VectorSet &visited, Node_Stack &nstack ) { 1381 Node *n = C->root(); 1382 visited.set(n->_idx); // first, mark node as visited 1383 // Do pre-visit work for root 1384 n = split_if_with_blocks_pre( n ); 1385 uint cnt = n->outcnt(); 1386 uint i = 0; 1387 while (true) { 1388 // Visit all children 1389 if (i < cnt) { 1390 Node* use = n->raw_out(i); 1391 ++i; 1392 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) { 1393 // Now do pre-visit work for this use 1394 use = split_if_with_blocks_pre( use ); 1395 nstack.push(n, i); // Save parent and next use's index. 1396 n = use; // Process all children of current use. 1397 cnt = use->outcnt(); 1398 i = 0; 1399 } 1400 } 1401 else { 1402 // All of n's children have been processed, complete post-processing. 1403 if (cnt != 0 && !n->is_Con()) { 1404 assert(has_node(n), "no dead nodes"); 1405 split_if_with_blocks_post( n ); 1406 } 1407 if (nstack.is_empty()) { 1408 // Finished all nodes on stack. 1409 break; 1410 } 1411 // Get saved parent node and next use's index. Visit the rest of uses. 1412 n = nstack.node(); 1413 cnt = n->outcnt(); 1414 i = nstack.index(); 1415 nstack.pop(); 1416 } 1417 } 1418 } 1419 1420 1421 //============================================================================= 1422 // 1423 // C L O N E A L O O P B O D Y 1424 // 1425 1426 //------------------------------clone_iff-------------------------------------- 1427 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1428 // "Nearly" because all Nodes have been cloned from the original in the loop, 1429 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1430 // through the Phi recursively, and return a Bool. 1431 BoolNode *PhaseIdealLoop::clone_iff( PhiNode *phi, IdealLoopTree *loop ) { 1432 1433 // Convert this Phi into a Phi merging Bools 1434 uint i; 1435 for( i = 1; i < phi->req(); i++ ) { 1436 Node *b = phi->in(i); 1437 if( b->is_Phi() ) { 1438 _igvn.replace_input_of(phi, i, clone_iff( b->as_Phi(), loop )); 1439 } else { 1440 assert( b->is_Bool(), "" ); 1441 } 1442 } 1443 1444 Node *sample_bool = phi->in(1); 1445 Node *sample_cmp = sample_bool->in(1); 1446 1447 // Make Phis to merge the Cmp's inputs. 1448 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP ); 1449 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP ); 1450 for( i = 1; i < phi->req(); i++ ) { 1451 Node *n1 = phi->in(i)->in(1)->in(1); 1452 Node *n2 = phi->in(i)->in(1)->in(2); 1453 phi1->set_req( i, n1 ); 1454 phi2->set_req( i, n2 ); 1455 phi1->set_type( phi1->type()->meet_speculative(n1->bottom_type())); 1456 phi2->set_type( phi2->type()->meet_speculative(n2->bottom_type())); 1457 } 1458 // See if these Phis have been made before. 1459 // Register with optimizer 1460 Node *hit1 = _igvn.hash_find_insert(phi1); 1461 if( hit1 ) { // Hit, toss just made Phi 1462 _igvn.remove_dead_node(phi1); // Remove new phi 1463 assert( hit1->is_Phi(), "" ); 1464 phi1 = (PhiNode*)hit1; // Use existing phi 1465 } else { // Miss 1466 _igvn.register_new_node_with_optimizer(phi1); 1467 } 1468 Node *hit2 = _igvn.hash_find_insert(phi2); 1469 if( hit2 ) { // Hit, toss just made Phi 1470 _igvn.remove_dead_node(phi2); // Remove new phi 1471 assert( hit2->is_Phi(), "" ); 1472 phi2 = (PhiNode*)hit2; // Use existing phi 1473 } else { // Miss 1474 _igvn.register_new_node_with_optimizer(phi2); 1475 } 1476 // Register Phis with loop/block info 1477 set_ctrl(phi1, phi->in(0)); 1478 set_ctrl(phi2, phi->in(0)); 1479 // Make a new Cmp 1480 Node *cmp = sample_cmp->clone(); 1481 cmp->set_req( 1, phi1 ); 1482 cmp->set_req( 2, phi2 ); 1483 _igvn.register_new_node_with_optimizer(cmp); 1484 set_ctrl(cmp, phi->in(0)); 1485 1486 // Make a new Bool 1487 Node *b = sample_bool->clone(); 1488 b->set_req(1,cmp); 1489 _igvn.register_new_node_with_optimizer(b); 1490 set_ctrl(b, phi->in(0)); 1491 1492 assert( b->is_Bool(), "" ); 1493 return (BoolNode*)b; 1494 } 1495 1496 //------------------------------clone_bool------------------------------------- 1497 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1498 // "Nearly" because all Nodes have been cloned from the original in the loop, 1499 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1500 // through the Phi recursively, and return a Bool. 1501 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) { 1502 uint i; 1503 // Convert this Phi into a Phi merging Bools 1504 for( i = 1; i < phi->req(); i++ ) { 1505 Node *b = phi->in(i); 1506 if( b->is_Phi() ) { 1507 _igvn.replace_input_of(phi, i, clone_bool( b->as_Phi(), loop )); 1508 } else { 1509 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" ); 1510 } 1511 } 1512 1513 Node *sample_cmp = phi->in(1); 1514 1515 // Make Phis to merge the Cmp's inputs. 1516 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP ); 1517 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP ); 1518 for( uint j = 1; j < phi->req(); j++ ) { 1519 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP 1520 Node *n1, *n2; 1521 if( cmp_top->is_Cmp() ) { 1522 n1 = cmp_top->in(1); 1523 n2 = cmp_top->in(2); 1524 } else { 1525 n1 = n2 = cmp_top; 1526 } 1527 phi1->set_req( j, n1 ); 1528 phi2->set_req( j, n2 ); 1529 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 1530 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 1531 } 1532 1533 // See if these Phis have been made before. 1534 // Register with optimizer 1535 Node *hit1 = _igvn.hash_find_insert(phi1); 1536 if( hit1 ) { // Hit, toss just made Phi 1537 _igvn.remove_dead_node(phi1); // Remove new phi 1538 assert( hit1->is_Phi(), "" ); 1539 phi1 = (PhiNode*)hit1; // Use existing phi 1540 } else { // Miss 1541 _igvn.register_new_node_with_optimizer(phi1); 1542 } 1543 Node *hit2 = _igvn.hash_find_insert(phi2); 1544 if( hit2 ) { // Hit, toss just made Phi 1545 _igvn.remove_dead_node(phi2); // Remove new phi 1546 assert( hit2->is_Phi(), "" ); 1547 phi2 = (PhiNode*)hit2; // Use existing phi 1548 } else { // Miss 1549 _igvn.register_new_node_with_optimizer(phi2); 1550 } 1551 // Register Phis with loop/block info 1552 set_ctrl(phi1, phi->in(0)); 1553 set_ctrl(phi2, phi->in(0)); 1554 // Make a new Cmp 1555 Node *cmp = sample_cmp->clone(); 1556 cmp->set_req( 1, phi1 ); 1557 cmp->set_req( 2, phi2 ); 1558 _igvn.register_new_node_with_optimizer(cmp); 1559 set_ctrl(cmp, phi->in(0)); 1560 1561 assert( cmp->is_Cmp(), "" ); 1562 return (CmpNode*)cmp; 1563 } 1564 1565 //------------------------------sink_use--------------------------------------- 1566 // If 'use' was in the loop-exit block, it now needs to be sunk 1567 // below the post-loop merge point. 1568 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) { 1569 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) { 1570 set_ctrl(use, post_loop); 1571 for (DUIterator j = use->outs(); use->has_out(j); j++) 1572 sink_use(use->out(j), post_loop); 1573 } 1574 } 1575 1576 //------------------------------clone_loop------------------------------------- 1577 // 1578 // C L O N E A L O O P B O D Y 1579 // 1580 // This is the basic building block of the loop optimizations. It clones an 1581 // entire loop body. It makes an old_new loop body mapping; with this mapping 1582 // you can find the new-loop equivalent to an old-loop node. All new-loop 1583 // nodes are exactly equal to their old-loop counterparts, all edges are the 1584 // same. All exits from the old-loop now have a RegionNode that merges the 1585 // equivalent new-loop path. This is true even for the normal "loop-exit" 1586 // condition. All uses of loop-invariant old-loop values now come from (one 1587 // or more) Phis that merge their new-loop equivalents. 1588 // 1589 // This operation leaves the graph in an illegal state: there are two valid 1590 // control edges coming from the loop pre-header to both loop bodies. I'll 1591 // definitely have to hack the graph after running this transform. 1592 // 1593 // From this building block I will further edit edges to perform loop peeling 1594 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc. 1595 // 1596 // Parameter side_by_size_idom: 1597 // When side_by_size_idom is NULL, the dominator tree is constructed for 1598 // the clone loop to dominate the original. Used in construction of 1599 // pre-main-post loop sequence. 1600 // When nonnull, the clone and original are side-by-side, both are 1601 // dominated by the side_by_side_idom node. Used in construction of 1602 // unswitched loops. 1603 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd, 1604 Node* side_by_side_idom) { 1605 1606 if (C->do_vector_loop() && PrintOpto) { 1607 const char* mname = C->method()->name()->as_quoted_ascii(); 1608 if (mname != NULL) { 1609 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname); 1610 } 1611 } 1612 1613 CloneMap& cm = C->clone_map(); 1614 Dict* dict = cm.dict(); 1615 if (C->do_vector_loop()) { 1616 cm.set_clone_idx(cm.max_gen()+1); 1617 #ifndef PRODUCT 1618 if (PrintOpto) { 1619 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx()); 1620 loop->dump_head(); 1621 } 1622 #endif 1623 } 1624 1625 // Step 1: Clone the loop body. Make the old->new mapping. 1626 uint i; 1627 for( i = 0; i < loop->_body.size(); i++ ) { 1628 Node *old = loop->_body.at(i); 1629 Node *nnn = old->clone(); 1630 old_new.map( old->_idx, nnn ); 1631 if (C->do_vector_loop()) { 1632 cm.verify_insert_and_clone(old, nnn, cm.clone_idx()); 1633 } 1634 _igvn.register_new_node_with_optimizer(nnn); 1635 } 1636 1637 1638 // Step 2: Fix the edges in the new body. If the old input is outside the 1639 // loop use it. If the old input is INside the loop, use the corresponding 1640 // new node instead. 1641 for( i = 0; i < loop->_body.size(); i++ ) { 1642 Node *old = loop->_body.at(i); 1643 Node *nnn = old_new[old->_idx]; 1644 // Fix CFG/Loop controlling the new node 1645 if (has_ctrl(old)) { 1646 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]); 1647 } else { 1648 set_loop(nnn, loop->_parent); 1649 if (old->outcnt() > 0) { 1650 set_idom( nnn, old_new[idom(old)->_idx], dd ); 1651 } 1652 } 1653 // Correct edges to the new node 1654 for( uint j = 0; j < nnn->req(); j++ ) { 1655 Node *n = nnn->in(j); 1656 if( n ) { 1657 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n ); 1658 if( loop->is_member( old_in_loop ) ) 1659 nnn->set_req(j, old_new[n->_idx]); 1660 } 1661 } 1662 _igvn.hash_find_insert(nnn); 1663 } 1664 Node *newhead = old_new[loop->_head->_idx]; 1665 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 1666 1667 1668 // Step 3: Now fix control uses. Loop varying control uses have already 1669 // been fixed up (as part of all input edges in Step 2). Loop invariant 1670 // control uses must be either an IfFalse or an IfTrue. Make a merge 1671 // point to merge the old and new IfFalse/IfTrue nodes; make the use 1672 // refer to this. 1673 ResourceArea *area = Thread::current()->resource_area(); 1674 Node_List worklist(area); 1675 uint new_counter = C->unique(); 1676 for( i = 0; i < loop->_body.size(); i++ ) { 1677 Node* old = loop->_body.at(i); 1678 if( !old->is_CFG() ) continue; 1679 Node* nnn = old_new[old->_idx]; 1680 1681 // Copy uses to a worklist, so I can munge the def-use info 1682 // with impunity. 1683 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1684 worklist.push(old->fast_out(j)); 1685 1686 while( worklist.size() ) { // Visit all uses 1687 Node *use = worklist.pop(); 1688 if (!has_node(use)) continue; // Ignore dead nodes 1689 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1690 if( !loop->is_member( use_loop ) && use->is_CFG() ) { 1691 // Both OLD and USE are CFG nodes here. 1692 assert( use->is_Proj(), "" ); 1693 1694 // Clone the loop exit control projection 1695 Node *newuse = use->clone(); 1696 if (C->do_vector_loop()) { 1697 cm.verify_insert_and_clone(use, newuse, cm.clone_idx()); 1698 } 1699 newuse->set_req(0,nnn); 1700 _igvn.register_new_node_with_optimizer(newuse); 1701 set_loop(newuse, use_loop); 1702 set_idom(newuse, nnn, dom_depth(nnn) + 1 ); 1703 1704 // We need a Region to merge the exit from the peeled body and the 1705 // exit from the old loop body. 1706 RegionNode *r = new RegionNode(3); 1707 // Map the old use to the new merge point 1708 old_new.map( use->_idx, r ); 1709 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use)); 1710 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" ); 1711 1712 // The original user of 'use' uses 'r' instead. 1713 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) { 1714 Node* useuse = use->last_out(l); 1715 _igvn.rehash_node_delayed(useuse); 1716 uint uses_found = 0; 1717 if( useuse->in(0) == use ) { 1718 useuse->set_req(0, r); 1719 uses_found++; 1720 if( useuse->is_CFG() ) { 1721 assert( dom_depth(useuse) > dd_r, "" ); 1722 set_idom(useuse, r, dom_depth(useuse)); 1723 } 1724 } 1725 for( uint k = 1; k < useuse->req(); k++ ) { 1726 if( useuse->in(k) == use ) { 1727 useuse->set_req(k, r); 1728 uses_found++; 1729 } 1730 } 1731 l -= uses_found; // we deleted 1 or more copies of this edge 1732 } 1733 1734 // Now finish up 'r' 1735 r->set_req( 1, newuse ); 1736 r->set_req( 2, use ); 1737 _igvn.register_new_node_with_optimizer(r); 1738 set_loop(r, use_loop); 1739 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r); 1740 } // End of if a loop-exit test 1741 } 1742 } 1743 1744 // Step 4: If loop-invariant use is not control, it must be dominated by a 1745 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region 1746 // there if needed. Make a Phi there merging old and new used values. 1747 Node_List *split_if_set = NULL; 1748 Node_List *split_bool_set = NULL; 1749 Node_List *split_cex_set = NULL; 1750 for( i = 0; i < loop->_body.size(); i++ ) { 1751 Node* old = loop->_body.at(i); 1752 Node* nnn = old_new[old->_idx]; 1753 // Copy uses to a worklist, so I can munge the def-use info 1754 // with impunity. 1755 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1756 worklist.push(old->fast_out(j)); 1757 1758 while( worklist.size() ) { 1759 Node *use = worklist.pop(); 1760 if (!has_node(use)) continue; // Ignore dead nodes 1761 if (use->in(0) == C->top()) continue; 1762 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1763 // Check for data-use outside of loop - at least one of OLD or USE 1764 // must not be a CFG node. 1765 if( !loop->is_member( use_loop ) && (!old->is_CFG() || !use->is_CFG())) { 1766 1767 // If the Data use is an IF, that means we have an IF outside of the 1768 // loop that is switching on a condition that is set inside of the 1769 // loop. Happens if people set a loop-exit flag; then test the flag 1770 // in the loop to break the loop, then test is again outside of the 1771 // loop to determine which way the loop exited. 1772 // Loop predicate If node connects to Bool node through Opaque1 node. 1773 if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use)) { 1774 // Since this code is highly unlikely, we lazily build the worklist 1775 // of such Nodes to go split. 1776 if( !split_if_set ) 1777 split_if_set = new Node_List(area); 1778 split_if_set->push(use); 1779 } 1780 if( use->is_Bool() ) { 1781 if( !split_bool_set ) 1782 split_bool_set = new Node_List(area); 1783 split_bool_set->push(use); 1784 } 1785 if( use->Opcode() == Op_CreateEx ) { 1786 if( !split_cex_set ) 1787 split_cex_set = new Node_List(area); 1788 split_cex_set->push(use); 1789 } 1790 1791 1792 // Get "block" use is in 1793 uint idx = 0; 1794 while( use->in(idx) != old ) idx++; 1795 Node *prev = use->is_CFG() ? use : get_ctrl(use); 1796 assert( !loop->is_member( get_loop( prev ) ), "" ); 1797 Node *cfg = prev->_idx >= new_counter 1798 ? prev->in(2) 1799 : idom(prev); 1800 if( use->is_Phi() ) // Phi use is in prior block 1801 cfg = prev->in(idx); // NOT in block of Phi itself 1802 if (cfg->is_top()) { // Use is dead? 1803 _igvn.replace_input_of(use, idx, C->top()); 1804 continue; 1805 } 1806 1807 while( !loop->is_member( get_loop( cfg ) ) ) { 1808 prev = cfg; 1809 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg); 1810 } 1811 // If the use occurs after merging several exits from the loop, then 1812 // old value must have dominated all those exits. Since the same old 1813 // value was used on all those exits we did not need a Phi at this 1814 // merge point. NOW we do need a Phi here. Each loop exit value 1815 // is now merged with the peeled body exit; each exit gets its own 1816 // private Phi and those Phis need to be merged here. 1817 Node *phi; 1818 if( prev->is_Region() ) { 1819 if( idx == 0 ) { // Updating control edge? 1820 phi = prev; // Just use existing control 1821 } else { // Else need a new Phi 1822 phi = PhiNode::make( prev, old ); 1823 // Now recursively fix up the new uses of old! 1824 for( uint i = 1; i < prev->req(); i++ ) { 1825 worklist.push(phi); // Onto worklist once for each 'old' input 1826 } 1827 } 1828 } else { 1829 // Get new RegionNode merging old and new loop exits 1830 prev = old_new[prev->_idx]; 1831 assert( prev, "just made this in step 7" ); 1832 if( idx == 0 ) { // Updating control edge? 1833 phi = prev; // Just use existing control 1834 } else { // Else need a new Phi 1835 // Make a new Phi merging data values properly 1836 phi = PhiNode::make( prev, old ); 1837 phi->set_req( 1, nnn ); 1838 } 1839 } 1840 // If inserting a new Phi, check for prior hits 1841 if( idx != 0 ) { 1842 Node *hit = _igvn.hash_find_insert(phi); 1843 if( hit == NULL ) { 1844 _igvn.register_new_node_with_optimizer(phi); // Register new phi 1845 } else { // or 1846 // Remove the new phi from the graph and use the hit 1847 _igvn.remove_dead_node(phi); 1848 phi = hit; // Use existing phi 1849 } 1850 set_ctrl(phi, prev); 1851 } 1852 // Make 'use' use the Phi instead of the old loop body exit value 1853 _igvn.replace_input_of(use, idx, phi); 1854 if( use->_idx >= new_counter ) { // If updating new phis 1855 // Not needed for correctness, but prevents a weak assert 1856 // in AddPNode from tripping (when we end up with different 1857 // base & derived Phis that will become the same after 1858 // IGVN does CSE). 1859 Node *hit = _igvn.hash_find_insert(use); 1860 if( hit ) // Go ahead and re-hash for hits. 1861 _igvn.replace_node( use, hit ); 1862 } 1863 1864 // If 'use' was in the loop-exit block, it now needs to be sunk 1865 // below the post-loop merge point. 1866 sink_use( use, prev ); 1867 } 1868 } 1869 } 1870 1871 // Check for IFs that need splitting/cloning. Happens if an IF outside of 1872 // the loop uses a condition set in the loop. The original IF probably 1873 // takes control from one or more OLD Regions (which in turn get from NEW 1874 // Regions). In any case, there will be a set of Phis for each merge point 1875 // from the IF up to where the original BOOL def exists the loop. 1876 if( split_if_set ) { 1877 while( split_if_set->size() ) { 1878 Node *iff = split_if_set->pop(); 1879 if( iff->in(1)->is_Phi() ) { 1880 BoolNode *b = clone_iff( iff->in(1)->as_Phi(), loop ); 1881 _igvn.replace_input_of(iff, 1, b); 1882 } 1883 } 1884 } 1885 if( split_bool_set ) { 1886 while( split_bool_set->size() ) { 1887 Node *b = split_bool_set->pop(); 1888 Node *phi = b->in(1); 1889 assert( phi->is_Phi(), "" ); 1890 CmpNode *cmp = clone_bool( (PhiNode*)phi, loop ); 1891 _igvn.replace_input_of(b, 1, cmp); 1892 } 1893 } 1894 if( split_cex_set ) { 1895 while( split_cex_set->size() ) { 1896 Node *b = split_cex_set->pop(); 1897 assert( b->in(0)->is_Region(), "" ); 1898 assert( b->in(1)->is_Phi(), "" ); 1899 assert( b->in(0)->in(0) == b->in(1)->in(0), "" ); 1900 split_up( b, b->in(0), NULL ); 1901 } 1902 } 1903 1904 } 1905 1906 1907 //---------------------- stride_of_possible_iv ------------------------------------- 1908 // Looks for an iff/bool/comp with one operand of the compare 1909 // being a cycle involving an add and a phi, 1910 // with an optional truncation (left-shift followed by a right-shift) 1911 // of the add. Returns zero if not an iv. 1912 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) { 1913 Node* trunc1 = NULL; 1914 Node* trunc2 = NULL; 1915 const TypeInt* ttype = NULL; 1916 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) { 1917 return 0; 1918 } 1919 BoolNode* bl = iff->in(1)->as_Bool(); 1920 Node* cmp = bl->in(1); 1921 if (!cmp || cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU) { 1922 return 0; 1923 } 1924 // Must have an invariant operand 1925 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) { 1926 return 0; 1927 } 1928 Node* add2 = NULL; 1929 Node* cmp1 = cmp->in(1); 1930 if (cmp1->is_Phi()) { 1931 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) ))) 1932 Node* phi = cmp1; 1933 for (uint i = 1; i < phi->req(); i++) { 1934 Node* in = phi->in(i); 1935 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in, 1936 &trunc1, &trunc2, &ttype); 1937 if (add && add->in(1) == phi) { 1938 add2 = add->in(2); 1939 break; 1940 } 1941 } 1942 } else { 1943 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) ))) 1944 Node* addtrunc = cmp1; 1945 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc, 1946 &trunc1, &trunc2, &ttype); 1947 if (add && add->in(1)->is_Phi()) { 1948 Node* phi = add->in(1); 1949 for (uint i = 1; i < phi->req(); i++) { 1950 if (phi->in(i) == addtrunc) { 1951 add2 = add->in(2); 1952 break; 1953 } 1954 } 1955 } 1956 } 1957 if (add2 != NULL) { 1958 const TypeInt* add2t = _igvn.type(add2)->is_int(); 1959 if (add2t->is_con()) { 1960 return add2t->get_con(); 1961 } 1962 } 1963 return 0; 1964 } 1965 1966 1967 //---------------------- stay_in_loop ------------------------------------- 1968 // Return the (unique) control output node that's in the loop (if it exists.) 1969 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) { 1970 Node* unique = NULL; 1971 if (!n) return NULL; 1972 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1973 Node* use = n->fast_out(i); 1974 if (!has_ctrl(use) && loop->is_member(get_loop(use))) { 1975 if (unique != NULL) { 1976 return NULL; 1977 } 1978 unique = use; 1979 } 1980 } 1981 return unique; 1982 } 1983 1984 //------------------------------ register_node ------------------------------------- 1985 // Utility to register node "n" with PhaseIdealLoop 1986 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) { 1987 _igvn.register_new_node_with_optimizer(n); 1988 loop->_body.push(n); 1989 if (n->is_CFG()) { 1990 set_loop(n, loop); 1991 set_idom(n, pred, ddepth); 1992 } else { 1993 set_ctrl(n, pred); 1994 } 1995 } 1996 1997 //------------------------------ proj_clone ------------------------------------- 1998 // Utility to create an if-projection 1999 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) { 2000 ProjNode* c = p->clone()->as_Proj(); 2001 c->set_req(0, iff); 2002 return c; 2003 } 2004 2005 //------------------------------ short_circuit_if ------------------------------------- 2006 // Force the iff control output to be the live_proj 2007 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) { 2008 guarantee(live_proj != NULL, "null projection"); 2009 int proj_con = live_proj->_con; 2010 assert(proj_con == 0 || proj_con == 1, "false or true projection"); 2011 Node *con = _igvn.intcon(proj_con); 2012 set_ctrl(con, C->root()); 2013 if (iff) { 2014 iff->set_req(1, con); 2015 } 2016 return con; 2017 } 2018 2019 //------------------------------ insert_if_before_proj ------------------------------------- 2020 // Insert a new if before an if projection (* - new node) 2021 // 2022 // before 2023 // if(test) 2024 // / \ 2025 // v v 2026 // other-proj proj (arg) 2027 // 2028 // after 2029 // if(test) 2030 // / \ 2031 // / v 2032 // | * proj-clone 2033 // v | 2034 // other-proj v 2035 // * new_if(relop(cmp[IU](left,right))) 2036 // / \ 2037 // v v 2038 // * new-proj proj 2039 // (returned) 2040 // 2041 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) { 2042 IfNode* iff = proj->in(0)->as_If(); 2043 IdealLoopTree *loop = get_loop(proj); 2044 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2045 int ddepth = dom_depth(proj); 2046 2047 _igvn.rehash_node_delayed(iff); 2048 _igvn.rehash_node_delayed(proj); 2049 2050 proj->set_req(0, NULL); // temporary disconnect 2051 ProjNode* proj2 = proj_clone(proj, iff); 2052 register_node(proj2, loop, iff, ddepth); 2053 2054 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right); 2055 register_node(cmp, loop, proj2, ddepth); 2056 2057 BoolNode* bol = new BoolNode(cmp, relop); 2058 register_node(bol, loop, proj2, ddepth); 2059 2060 int opcode = iff->Opcode(); 2061 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode"); 2062 IfNode* new_if = (opcode == Op_If) ? new IfNode(proj2, bol, iff->_prob, iff->_fcnt): 2063 new RangeCheckNode(proj2, bol, iff->_prob, iff->_fcnt); 2064 register_node(new_if, loop, proj2, ddepth); 2065 2066 proj->set_req(0, new_if); // reattach 2067 set_idom(proj, new_if, ddepth); 2068 2069 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj(); 2070 guarantee(new_exit != NULL, "null exit node"); 2071 register_node(new_exit, get_loop(other_proj), new_if, ddepth); 2072 2073 return new_exit; 2074 } 2075 2076 //------------------------------ insert_region_before_proj ------------------------------------- 2077 // Insert a region before an if projection (* - new node) 2078 // 2079 // before 2080 // if(test) 2081 // / | 2082 // v | 2083 // proj v 2084 // other-proj 2085 // 2086 // after 2087 // if(test) 2088 // / | 2089 // v | 2090 // * proj-clone v 2091 // | other-proj 2092 // v 2093 // * new-region 2094 // | 2095 // v 2096 // * dum_if 2097 // / \ 2098 // v \ 2099 // * dum-proj v 2100 // proj 2101 // 2102 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) { 2103 IfNode* iff = proj->in(0)->as_If(); 2104 IdealLoopTree *loop = get_loop(proj); 2105 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2106 int ddepth = dom_depth(proj); 2107 2108 _igvn.rehash_node_delayed(iff); 2109 _igvn.rehash_node_delayed(proj); 2110 2111 proj->set_req(0, NULL); // temporary disconnect 2112 ProjNode* proj2 = proj_clone(proj, iff); 2113 register_node(proj2, loop, iff, ddepth); 2114 2115 RegionNode* reg = new RegionNode(2); 2116 reg->set_req(1, proj2); 2117 register_node(reg, loop, iff, ddepth); 2118 2119 IfNode* dum_if = new IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt); 2120 register_node(dum_if, loop, reg, ddepth); 2121 2122 proj->set_req(0, dum_if); // reattach 2123 set_idom(proj, dum_if, ddepth); 2124 2125 ProjNode* dum_proj = proj_clone(other_proj, dum_if); 2126 register_node(dum_proj, loop, dum_if, ddepth); 2127 2128 return reg; 2129 } 2130 2131 //------------------------------ insert_cmpi_loop_exit ------------------------------------- 2132 // Clone a signed compare loop exit from an unsigned compare and 2133 // insert it before the unsigned cmp on the stay-in-loop path. 2134 // All new nodes inserted in the dominator tree between the original 2135 // if and it's projections. The original if test is replaced with 2136 // a constant to force the stay-in-loop path. 2137 // 2138 // This is done to make sure that the original if and it's projections 2139 // still dominate the same set of control nodes, that the ctrl() relation 2140 // from data nodes to them is preserved, and that their loop nesting is 2141 // preserved. 2142 // 2143 // before 2144 // if(i <u limit) unsigned compare loop exit 2145 // / | 2146 // v v 2147 // exit-proj stay-in-loop-proj 2148 // 2149 // after 2150 // if(stay-in-loop-const) original if 2151 // / | 2152 // / v 2153 // / if(i < limit) new signed test 2154 // / / | 2155 // / / v 2156 // / / if(i <u limit) new cloned unsigned test 2157 // / / / | 2158 // v v v | 2159 // region | 2160 // | | 2161 // dum-if | 2162 // / | | 2163 // ether | | 2164 // v v 2165 // exit-proj stay-in-loop-proj 2166 // 2167 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) { 2168 const bool Signed = true; 2169 const bool Unsigned = false; 2170 2171 BoolNode* bol = if_cmpu->in(1)->as_Bool(); 2172 if (bol->_test._test != BoolTest::lt) return NULL; 2173 CmpNode* cmpu = bol->in(1)->as_Cmp(); 2174 if (cmpu->Opcode() != Op_CmpU) return NULL; 2175 int stride = stride_of_possible_iv(if_cmpu); 2176 if (stride == 0) return NULL; 2177 2178 Node* lp_proj = stay_in_loop(if_cmpu, loop); 2179 guarantee(lp_proj != NULL, "null loop node"); 2180 2181 ProjNode* lp_continue = lp_proj->as_Proj(); 2182 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj(); 2183 2184 Node* limit = NULL; 2185 if (stride > 0) { 2186 limit = cmpu->in(2); 2187 } else { 2188 limit = _igvn.makecon(TypeInt::ZERO); 2189 set_ctrl(limit, C->root()); 2190 } 2191 // Create a new region on the exit path 2192 RegionNode* reg = insert_region_before_proj(lp_exit); 2193 guarantee(reg != NULL, "null region node"); 2194 2195 // Clone the if-cmpu-true-false using a signed compare 2196 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge; 2197 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue); 2198 reg->add_req(cmpi_exit); 2199 2200 // Clone the if-cmpu-true-false 2201 BoolTest::mask rel_u = bol->_test._test; 2202 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue); 2203 reg->add_req(cmpu_exit); 2204 2205 // Force original if to stay in loop. 2206 short_circuit_if(if_cmpu, lp_continue); 2207 2208 return cmpi_exit->in(0)->as_If(); 2209 } 2210 2211 //------------------------------ remove_cmpi_loop_exit ------------------------------------- 2212 // Remove a previously inserted signed compare loop exit. 2213 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) { 2214 Node* lp_proj = stay_in_loop(if_cmp, loop); 2215 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI && 2216 stay_in_loop(lp_proj, loop)->is_If() && 2217 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu"); 2218 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO); 2219 set_ctrl(con, C->root()); 2220 if_cmp->set_req(1, con); 2221 } 2222 2223 //------------------------------ scheduled_nodelist ------------------------------------- 2224 // Create a post order schedule of nodes that are in the 2225 // "member" set. The list is returned in "sched". 2226 // The first node in "sched" is the loop head, followed by 2227 // nodes which have no inputs in the "member" set, and then 2228 // followed by the nodes that have an immediate input dependence 2229 // on a node in "sched". 2230 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) { 2231 2232 assert(member.test(loop->_head->_idx), "loop head must be in member set"); 2233 Arena *a = Thread::current()->resource_area(); 2234 VectorSet visited(a); 2235 Node_Stack nstack(a, loop->_body.size()); 2236 2237 Node* n = loop->_head; // top of stack is cached in "n" 2238 uint idx = 0; 2239 visited.set(n->_idx); 2240 2241 // Initially push all with no inputs from within member set 2242 for(uint i = 0; i < loop->_body.size(); i++ ) { 2243 Node *elt = loop->_body.at(i); 2244 if (member.test(elt->_idx)) { 2245 bool found = false; 2246 for (uint j = 0; j < elt->req(); j++) { 2247 Node* def = elt->in(j); 2248 if (def && member.test(def->_idx) && def != elt) { 2249 found = true; 2250 break; 2251 } 2252 } 2253 if (!found && elt != loop->_head) { 2254 nstack.push(n, idx); 2255 n = elt; 2256 assert(!visited.test(n->_idx), "not seen yet"); 2257 visited.set(n->_idx); 2258 } 2259 } 2260 } 2261 2262 // traverse out's that are in the member set 2263 while (true) { 2264 if (idx < n->outcnt()) { 2265 Node* use = n->raw_out(idx); 2266 idx++; 2267 if (!visited.test_set(use->_idx)) { 2268 if (member.test(use->_idx)) { 2269 nstack.push(n, idx); 2270 n = use; 2271 idx = 0; 2272 } 2273 } 2274 } else { 2275 // All outputs processed 2276 sched.push(n); 2277 if (nstack.is_empty()) break; 2278 n = nstack.node(); 2279 idx = nstack.index(); 2280 nstack.pop(); 2281 } 2282 } 2283 } 2284 2285 2286 //------------------------------ has_use_in_set ------------------------------------- 2287 // Has a use in the vector set 2288 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) { 2289 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2290 Node* use = n->fast_out(j); 2291 if (vset.test(use->_idx)) { 2292 return true; 2293 } 2294 } 2295 return false; 2296 } 2297 2298 2299 //------------------------------ has_use_internal_to_set ------------------------------------- 2300 // Has use internal to the vector set (ie. not in a phi at the loop head) 2301 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) { 2302 Node* head = loop->_head; 2303 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2304 Node* use = n->fast_out(j); 2305 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) { 2306 return true; 2307 } 2308 } 2309 return false; 2310 } 2311 2312 2313 //------------------------------ clone_for_use_outside_loop ------------------------------------- 2314 // clone "n" for uses that are outside of loop 2315 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) { 2316 int cloned = 0; 2317 assert(worklist.size() == 0, "should be empty"); 2318 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2319 Node* use = n->fast_out(j); 2320 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) { 2321 worklist.push(use); 2322 } 2323 } 2324 while( worklist.size() ) { 2325 Node *use = worklist.pop(); 2326 if (!has_node(use) || use->in(0) == C->top()) continue; 2327 uint j; 2328 for (j = 0; j < use->req(); j++) { 2329 if (use->in(j) == n) break; 2330 } 2331 assert(j < use->req(), "must be there"); 2332 2333 // clone "n" and insert it between the inputs of "n" and the use outside the loop 2334 Node* n_clone = n->clone(); 2335 _igvn.replace_input_of(use, j, n_clone); 2336 cloned++; 2337 Node* use_c; 2338 if (!use->is_Phi()) { 2339 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0); 2340 } else { 2341 // Use in a phi is considered a use in the associated predecessor block 2342 use_c = use->in(0)->in(j); 2343 } 2344 set_ctrl(n_clone, use_c); 2345 assert(!loop->is_member(get_loop(use_c)), "should be outside loop"); 2346 get_loop(use_c)->_body.push(n_clone); 2347 _igvn.register_new_node_with_optimizer(n_clone); 2348 #if !defined(PRODUCT) 2349 if (TracePartialPeeling) { 2350 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx); 2351 } 2352 #endif 2353 } 2354 return cloned; 2355 } 2356 2357 2358 //------------------------------ clone_for_special_use_inside_loop ------------------------------------- 2359 // clone "n" for special uses that are in the not_peeled region. 2360 // If these def-uses occur in separate blocks, the code generator 2361 // marks the method as not compilable. For example, if a "BoolNode" 2362 // is in a different basic block than the "IfNode" that uses it, then 2363 // the compilation is aborted in the code generator. 2364 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 2365 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) { 2366 if (n->is_Phi() || n->is_Load()) { 2367 return; 2368 } 2369 assert(worklist.size() == 0, "should be empty"); 2370 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2371 Node* use = n->fast_out(j); 2372 if ( not_peel.test(use->_idx) && 2373 (use->is_If() || use->is_CMove() || use->is_Bool()) && 2374 use->in(1) == n) { 2375 worklist.push(use); 2376 } 2377 } 2378 if (worklist.size() > 0) { 2379 // clone "n" and insert it between inputs of "n" and the use 2380 Node* n_clone = n->clone(); 2381 loop->_body.push(n_clone); 2382 _igvn.register_new_node_with_optimizer(n_clone); 2383 set_ctrl(n_clone, get_ctrl(n)); 2384 sink_list.push(n_clone); 2385 not_peel <<= n_clone->_idx; // add n_clone to not_peel set. 2386 #if !defined(PRODUCT) 2387 if (TracePartialPeeling) { 2388 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx); 2389 } 2390 #endif 2391 while( worklist.size() ) { 2392 Node *use = worklist.pop(); 2393 _igvn.rehash_node_delayed(use); 2394 for (uint j = 1; j < use->req(); j++) { 2395 if (use->in(j) == n) { 2396 use->set_req(j, n_clone); 2397 } 2398 } 2399 } 2400 } 2401 } 2402 2403 2404 //------------------------------ insert_phi_for_loop ------------------------------------- 2405 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 2406 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) { 2407 Node *phi = PhiNode::make(lp, back_edge_val); 2408 phi->set_req(LoopNode::EntryControl, lp_entry_val); 2409 // Use existing phi if it already exists 2410 Node *hit = _igvn.hash_find_insert(phi); 2411 if( hit == NULL ) { 2412 _igvn.register_new_node_with_optimizer(phi); 2413 set_ctrl(phi, lp); 2414 } else { 2415 // Remove the new phi from the graph and use the hit 2416 _igvn.remove_dead_node(phi); 2417 phi = hit; 2418 } 2419 _igvn.replace_input_of(use, idx, phi); 2420 } 2421 2422 #ifdef ASSERT 2423 //------------------------------ is_valid_loop_partition ------------------------------------- 2424 // Validate the loop partition sets: peel and not_peel 2425 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, 2426 VectorSet& not_peel ) { 2427 uint i; 2428 // Check that peel_list entries are in the peel set 2429 for (i = 0; i < peel_list.size(); i++) { 2430 if (!peel.test(peel_list.at(i)->_idx)) { 2431 return false; 2432 } 2433 } 2434 // Check at loop members are in one of peel set or not_peel set 2435 for (i = 0; i < loop->_body.size(); i++ ) { 2436 Node *def = loop->_body.at(i); 2437 uint di = def->_idx; 2438 // Check that peel set elements are in peel_list 2439 if (peel.test(di)) { 2440 if (not_peel.test(di)) { 2441 return false; 2442 } 2443 // Must be in peel_list also 2444 bool found = false; 2445 for (uint j = 0; j < peel_list.size(); j++) { 2446 if (peel_list.at(j)->_idx == di) { 2447 found = true; 2448 break; 2449 } 2450 } 2451 if (!found) { 2452 return false; 2453 } 2454 } else if (not_peel.test(di)) { 2455 if (peel.test(di)) { 2456 return false; 2457 } 2458 } else { 2459 return false; 2460 } 2461 } 2462 return true; 2463 } 2464 2465 //------------------------------ is_valid_clone_loop_exit_use ------------------------------------- 2466 // Ensure a use outside of loop is of the right form 2467 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) { 2468 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2469 return (use->is_Phi() && 2470 use_c->is_Region() && use_c->req() == 3 && 2471 (use_c->in(exit_idx)->Opcode() == Op_IfTrue || 2472 use_c->in(exit_idx)->Opcode() == Op_IfFalse || 2473 use_c->in(exit_idx)->Opcode() == Op_JumpProj) && 2474 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) ); 2475 } 2476 2477 //------------------------------ is_valid_clone_loop_form ------------------------------------- 2478 // Ensure that all uses outside of loop are of the right form 2479 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 2480 uint orig_exit_idx, uint clone_exit_idx) { 2481 uint len = peel_list.size(); 2482 for (uint i = 0; i < len; i++) { 2483 Node *def = peel_list.at(i); 2484 2485 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 2486 Node *use = def->fast_out(j); 2487 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2488 if (!loop->is_member(get_loop(use_c))) { 2489 // use is not in the loop, check for correct structure 2490 if (use->in(0) == def) { 2491 // Okay 2492 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) { 2493 return false; 2494 } 2495 } 2496 } 2497 } 2498 return true; 2499 } 2500 #endif 2501 2502 //------------------------------ partial_peel ------------------------------------- 2503 // Partially peel (aka loop rotation) the top portion of a loop (called 2504 // the peel section below) by cloning it and placing one copy just before 2505 // the new loop head and the other copy at the bottom of the new loop. 2506 // 2507 // before after where it came from 2508 // 2509 // stmt1 stmt1 2510 // loop: stmt2 clone 2511 // stmt2 if condA goto exitA clone 2512 // if condA goto exitA new_loop: new 2513 // stmt3 stmt3 clone 2514 // if !condB goto loop if condB goto exitB clone 2515 // exitB: stmt2 orig 2516 // stmt4 if !condA goto new_loop orig 2517 // exitA: goto exitA 2518 // exitB: 2519 // stmt4 2520 // exitA: 2521 // 2522 // Step 1: find the cut point: an exit test on probable 2523 // induction variable. 2524 // Step 2: schedule (with cloning) operations in the peel 2525 // section that can be executed after the cut into 2526 // the section that is not peeled. This may need 2527 // to clone operations into exit blocks. For 2528 // instance, a reference to A[i] in the not-peel 2529 // section and a reference to B[i] in an exit block 2530 // may cause a left-shift of i by 2 to be placed 2531 // in the peel block. This step will clone the left 2532 // shift into the exit block and sink the left shift 2533 // from the peel to the not-peel section. 2534 // Step 3: clone the loop, retarget the control, and insert 2535 // phis for values that are live across the new loop 2536 // head. This is very dependent on the graph structure 2537 // from clone_loop. It creates region nodes for 2538 // exit control and associated phi nodes for values 2539 // flow out of the loop through that exit. The region 2540 // node is dominated by the clone's control projection. 2541 // So the clone's peel section is placed before the 2542 // new loop head, and the clone's not-peel section is 2543 // forms the top part of the new loop. The original 2544 // peel section forms the tail of the new loop. 2545 // Step 4: update the dominator tree and recompute the 2546 // dominator depth. 2547 // 2548 // orig 2549 // 2550 // stmt1 2551 // | 2552 // v 2553 // loop predicate 2554 // | 2555 // v 2556 // loop<----+ 2557 // | | 2558 // stmt2 | 2559 // | | 2560 // v | 2561 // ifA | 2562 // / | | 2563 // v v | 2564 // false true ^ <-- last_peel 2565 // / | | 2566 // / ===|==cut | 2567 // / stmt3 | <-- first_not_peel 2568 // / | | 2569 // | v | 2570 // v ifB | 2571 // exitA: / \ | 2572 // / \ | 2573 // v v | 2574 // false true | 2575 // / \ | 2576 // / ----+ 2577 // | 2578 // v 2579 // exitB: 2580 // stmt4 2581 // 2582 // 2583 // after clone loop 2584 // 2585 // stmt1 2586 // | 2587 // v 2588 // loop predicate 2589 // / \ 2590 // clone / \ orig 2591 // / \ 2592 // / \ 2593 // v v 2594 // +---->loop loop<----+ 2595 // | | | | 2596 // | stmt2 stmt2 | 2597 // | | | | 2598 // | v v | 2599 // | ifA ifA | 2600 // | | \ / | | 2601 // | v v v v | 2602 // ^ true false false true ^ <-- last_peel 2603 // | | ^ \ / | | 2604 // | cut==|== \ \ / ===|==cut | 2605 // | stmt3 \ \ / stmt3 | <-- first_not_peel 2606 // | | dom | | | | 2607 // | v \ 1v v2 v | 2608 // | ifB regionA ifB | 2609 // | / \ | / \ | 2610 // | / \ v / \ | 2611 // | v v exitA: v v | 2612 // | true false false true | 2613 // | / ^ \ / \ | 2614 // +---- \ \ / ----+ 2615 // dom \ / 2616 // \ 1v v2 2617 // regionB 2618 // | 2619 // v 2620 // exitB: 2621 // stmt4 2622 // 2623 // 2624 // after partial peel 2625 // 2626 // stmt1 2627 // | 2628 // v 2629 // loop predicate 2630 // / 2631 // clone / orig 2632 // / TOP 2633 // / \ 2634 // v v 2635 // TOP->loop loop----+ 2636 // | | | 2637 // stmt2 stmt2 | 2638 // | | | 2639 // v v | 2640 // ifA ifA | 2641 // | \ / | | 2642 // v v v v | 2643 // true false false true | <-- last_peel 2644 // | ^ \ / +------|---+ 2645 // +->newloop \ \ / === ==cut | | 2646 // | stmt3 \ \ / TOP | | 2647 // | | dom | | stmt3 | | <-- first_not_peel 2648 // | v \ 1v v2 v | | 2649 // | ifB regionA ifB ^ v 2650 // | / \ | / \ | | 2651 // | / \ v / \ | | 2652 // | v v exitA: v v | | 2653 // | true false false true | | 2654 // | / ^ \ / \ | | 2655 // | | \ \ / v | | 2656 // | | dom \ / TOP | | 2657 // | | \ 1v v2 | | 2658 // ^ v regionB | | 2659 // | | | | | 2660 // | | v ^ v 2661 // | | exitB: | | 2662 // | | stmt4 | | 2663 // | +------------>-----------------+ | 2664 // | | 2665 // +-----------------<---------------------+ 2666 // 2667 // 2668 // final graph 2669 // 2670 // stmt1 2671 // | 2672 // v 2673 // loop predicate 2674 // | 2675 // v 2676 // stmt2 clone 2677 // | 2678 // v 2679 // ........> ifA clone 2680 // : / | 2681 // dom / | 2682 // : v v 2683 // : false true 2684 // : | | 2685 // : | v 2686 // : | newloop<-----+ 2687 // : | | | 2688 // : | stmt3 clone | 2689 // : | | | 2690 // : | v | 2691 // : | ifB | 2692 // : | / \ | 2693 // : | v v | 2694 // : | false true | 2695 // : | | | | 2696 // : | v stmt2 | 2697 // : | exitB: | | 2698 // : | stmt4 v | 2699 // : | ifA orig | 2700 // : | / \ | 2701 // : | / \ | 2702 // : | v v | 2703 // : | false true | 2704 // : | / \ | 2705 // : v v -----+ 2706 // RegionA 2707 // | 2708 // v 2709 // exitA 2710 // 2711 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) { 2712 2713 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only"); 2714 if (!loop->_head->is_Loop()) { 2715 return false; } 2716 2717 LoopNode *head = loop->_head->as_Loop(); 2718 2719 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) { 2720 return false; 2721 } 2722 2723 // Check for complex exit control 2724 for(uint ii = 0; ii < loop->_body.size(); ii++ ) { 2725 Node *n = loop->_body.at(ii); 2726 int opc = n->Opcode(); 2727 if (n->is_Call() || 2728 opc == Op_Catch || 2729 opc == Op_CatchProj || 2730 opc == Op_Jump || 2731 opc == Op_JumpProj) { 2732 #if !defined(PRODUCT) 2733 if (TracePartialPeeling) { 2734 tty->print_cr("\nExit control too complex: lp: %d", head->_idx); 2735 } 2736 #endif 2737 return false; 2738 } 2739 } 2740 2741 int dd = dom_depth(head); 2742 2743 // Step 1: find cut point 2744 2745 // Walk up dominators to loop head looking for first loop exit 2746 // which is executed on every path thru loop. 2747 IfNode *peel_if = NULL; 2748 IfNode *peel_if_cmpu = NULL; 2749 2750 Node *iff = loop->tail(); 2751 while( iff != head ) { 2752 if( iff->is_If() ) { 2753 Node *ctrl = get_ctrl(iff->in(1)); 2754 if (ctrl->is_top()) return false; // Dead test on live IF. 2755 // If loop-varying exit-test, check for induction variable 2756 if( loop->is_member(get_loop(ctrl)) && 2757 loop->is_loop_exit(iff) && 2758 is_possible_iv_test(iff)) { 2759 Node* cmp = iff->in(1)->in(1); 2760 if (cmp->Opcode() == Op_CmpI) { 2761 peel_if = iff->as_If(); 2762 } else { 2763 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU"); 2764 peel_if_cmpu = iff->as_If(); 2765 } 2766 } 2767 } 2768 iff = idom(iff); 2769 } 2770 // Prefer signed compare over unsigned compare. 2771 IfNode* new_peel_if = NULL; 2772 if (peel_if == NULL) { 2773 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) { 2774 return false; // No peel point found 2775 } 2776 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop); 2777 if (new_peel_if == NULL) { 2778 return false; // No peel point found 2779 } 2780 peel_if = new_peel_if; 2781 } 2782 Node* last_peel = stay_in_loop(peel_if, loop); 2783 Node* first_not_peeled = stay_in_loop(last_peel, loop); 2784 if (first_not_peeled == NULL || first_not_peeled == head) { 2785 return false; 2786 } 2787 2788 #if !defined(PRODUCT) 2789 if (TraceLoopOpts) { 2790 tty->print("PartialPeel "); 2791 loop->dump_head(); 2792 } 2793 2794 if (TracePartialPeeling) { 2795 tty->print_cr("before partial peel one iteration"); 2796 Node_List wl; 2797 Node* t = head->in(2); 2798 while (true) { 2799 wl.push(t); 2800 if (t == head) break; 2801 t = idom(t); 2802 } 2803 while (wl.size() > 0) { 2804 Node* tt = wl.pop(); 2805 tt->dump(); 2806 if (tt == last_peel) tty->print_cr("-- cut --"); 2807 } 2808 } 2809 #endif 2810 ResourceArea *area = Thread::current()->resource_area(); 2811 VectorSet peel(area); 2812 VectorSet not_peel(area); 2813 Node_List peel_list(area); 2814 Node_List worklist(area); 2815 Node_List sink_list(area); 2816 2817 // Set of cfg nodes to peel are those that are executable from 2818 // the head through last_peel. 2819 assert(worklist.size() == 0, "should be empty"); 2820 worklist.push(head); 2821 peel.set(head->_idx); 2822 while (worklist.size() > 0) { 2823 Node *n = worklist.pop(); 2824 if (n != last_peel) { 2825 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2826 Node* use = n->fast_out(j); 2827 if (use->is_CFG() && 2828 loop->is_member(get_loop(use)) && 2829 !peel.test_set(use->_idx)) { 2830 worklist.push(use); 2831 } 2832 } 2833 } 2834 } 2835 2836 // Set of non-cfg nodes to peel are those that are control 2837 // dependent on the cfg nodes. 2838 uint i; 2839 for(i = 0; i < loop->_body.size(); i++ ) { 2840 Node *n = loop->_body.at(i); 2841 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n; 2842 if (peel.test(n_c->_idx)) { 2843 peel.set(n->_idx); 2844 } else { 2845 not_peel.set(n->_idx); 2846 } 2847 } 2848 2849 // Step 2: move operations from the peeled section down into the 2850 // not-peeled section 2851 2852 // Get a post order schedule of nodes in the peel region 2853 // Result in right-most operand. 2854 scheduled_nodelist(loop, peel, peel_list ); 2855 2856 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 2857 2858 // For future check for too many new phis 2859 uint old_phi_cnt = 0; 2860 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 2861 Node* use = head->fast_out(j); 2862 if (use->is_Phi()) old_phi_cnt++; 2863 } 2864 2865 #if !defined(PRODUCT) 2866 if (TracePartialPeeling) { 2867 tty->print_cr("\npeeled list"); 2868 } 2869 #endif 2870 2871 // Evacuate nodes in peel region into the not_peeled region if possible 2872 uint new_phi_cnt = 0; 2873 uint cloned_for_outside_use = 0; 2874 for (i = 0; i < peel_list.size();) { 2875 Node* n = peel_list.at(i); 2876 #if !defined(PRODUCT) 2877 if (TracePartialPeeling) n->dump(); 2878 #endif 2879 bool incr = true; 2880 if ( !n->is_CFG() ) { 2881 2882 if ( has_use_in_set(n, not_peel) ) { 2883 2884 // If not used internal to the peeled region, 2885 // move "n" from peeled to not_peeled region. 2886 2887 if ( !has_use_internal_to_set(n, peel, loop) ) { 2888 2889 // if not pinned and not a load (which maybe anti-dependent on a store) 2890 // and not a CMove (Matcher expects only bool->cmove). 2891 if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) { 2892 cloned_for_outside_use += clone_for_use_outside_loop( loop, n, worklist ); 2893 sink_list.push(n); 2894 peel >>= n->_idx; // delete n from peel set. 2895 not_peel <<= n->_idx; // add n to not_peel set. 2896 peel_list.remove(i); 2897 incr = false; 2898 #if !defined(PRODUCT) 2899 if (TracePartialPeeling) { 2900 tty->print_cr("sink to not_peeled region: %d newbb: %d", 2901 n->_idx, get_ctrl(n)->_idx); 2902 } 2903 #endif 2904 } 2905 } else { 2906 // Otherwise check for special def-use cases that span 2907 // the peel/not_peel boundary such as bool->if 2908 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist ); 2909 new_phi_cnt++; 2910 } 2911 } 2912 } 2913 if (incr) i++; 2914 } 2915 2916 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) { 2917 #if !defined(PRODUCT) 2918 if (TracePartialPeeling) { 2919 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c", 2920 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F'); 2921 } 2922 #endif 2923 if (new_peel_if != NULL) { 2924 remove_cmpi_loop_exit(new_peel_if, loop); 2925 } 2926 // Inhibit more partial peeling on this loop 2927 assert(!head->is_partial_peel_loop(), "not partial peeled"); 2928 head->mark_partial_peel_failed(); 2929 if (cloned_for_outside_use > 0) { 2930 // Terminate this round of loop opts because 2931 // the graph outside this loop was changed. 2932 C->set_major_progress(); 2933 return true; 2934 } 2935 return false; 2936 } 2937 2938 // Step 3: clone loop, retarget control, and insert new phis 2939 2940 // Create new loop head for new phis and to hang 2941 // the nodes being moved (sinked) from the peel region. 2942 LoopNode* new_head = new LoopNode(last_peel, last_peel); 2943 new_head->set_unswitch_count(head->unswitch_count()); // Preserve 2944 _igvn.register_new_node_with_optimizer(new_head); 2945 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled"); 2946 _igvn.replace_input_of(first_not_peeled, 0, new_head); 2947 set_loop(new_head, loop); 2948 loop->_body.push(new_head); 2949 not_peel.set(new_head->_idx); 2950 set_idom(new_head, last_peel, dom_depth(first_not_peeled)); 2951 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled)); 2952 2953 while (sink_list.size() > 0) { 2954 Node* n = sink_list.pop(); 2955 set_ctrl(n, new_head); 2956 } 2957 2958 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 2959 2960 clone_loop( loop, old_new, dd ); 2961 2962 const uint clone_exit_idx = 1; 2963 const uint orig_exit_idx = 2; 2964 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop"); 2965 2966 Node* head_clone = old_new[head->_idx]; 2967 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop(); 2968 Node* orig_tail_clone = head_clone->in(2); 2969 2970 // Add phi if "def" node is in peel set and "use" is not 2971 2972 for(i = 0; i < peel_list.size(); i++ ) { 2973 Node *def = peel_list.at(i); 2974 if (!def->is_CFG()) { 2975 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 2976 Node *use = def->fast_out(j); 2977 if (has_node(use) && use->in(0) != C->top() && 2978 (!peel.test(use->_idx) || 2979 (use->is_Phi() && use->in(0) == head)) ) { 2980 worklist.push(use); 2981 } 2982 } 2983 while( worklist.size() ) { 2984 Node *use = worklist.pop(); 2985 for (uint j = 1; j < use->req(); j++) { 2986 Node* n = use->in(j); 2987 if (n == def) { 2988 2989 // "def" is in peel set, "use" is not in peel set 2990 // or "use" is in the entry boundary (a phi) of the peel set 2991 2992 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use; 2993 2994 if ( loop->is_member(get_loop( use_c )) ) { 2995 // use is in loop 2996 if (old_new[use->_idx] != NULL) { // null for dead code 2997 Node* use_clone = old_new[use->_idx]; 2998 _igvn.replace_input_of(use, j, C->top()); 2999 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone ); 3000 } 3001 } else { 3002 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format"); 3003 // use is not in the loop, check if the live range includes the cut 3004 Node* lp_if = use_c->in(orig_exit_idx)->in(0); 3005 if (not_peel.test(lp_if->_idx)) { 3006 assert(j == orig_exit_idx, "use from original loop"); 3007 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone ); 3008 } 3009 } 3010 } 3011 } 3012 } 3013 } 3014 } 3015 3016 // Step 3b: retarget control 3017 3018 // Redirect control to the new loop head if a cloned node in 3019 // the not_peeled region has control that points into the peeled region. 3020 // This necessary because the cloned peeled region will be outside 3021 // the loop. 3022 // from to 3023 // cloned-peeled <---+ 3024 // new_head_clone: | <--+ 3025 // cloned-not_peeled in(0) in(0) 3026 // orig-peeled 3027 3028 for(i = 0; i < loop->_body.size(); i++ ) { 3029 Node *n = loop->_body.at(i); 3030 if (!n->is_CFG() && n->in(0) != NULL && 3031 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) { 3032 Node* n_clone = old_new[n->_idx]; 3033 _igvn.replace_input_of(n_clone, 0, new_head_clone); 3034 } 3035 } 3036 3037 // Backedge of the surviving new_head (the clone) is original last_peel 3038 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel); 3039 3040 // Cut first node in original not_peel set 3041 _igvn.rehash_node_delayed(new_head); // Multiple edge updates: 3042 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of 3043 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls 3044 3045 // Copy head_clone back-branch info to original head 3046 // and remove original head's loop entry and 3047 // clone head's back-branch 3048 _igvn.rehash_node_delayed(head); // Multiple edge updates 3049 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl)); 3050 head->set_req(LoopNode::LoopBackControl, C->top()); 3051 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top()); 3052 3053 // Similarly modify the phis 3054 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) { 3055 Node* use = head->fast_out(k); 3056 if (use->is_Phi() && use->outcnt() > 0) { 3057 Node* use_clone = old_new[use->_idx]; 3058 _igvn.rehash_node_delayed(use); // Multiple edge updates 3059 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl)); 3060 use->set_req(LoopNode::LoopBackControl, C->top()); 3061 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top()); 3062 } 3063 } 3064 3065 // Step 4: update dominator tree and dominator depth 3066 3067 set_idom(head, orig_tail_clone, dd); 3068 recompute_dom_depth(); 3069 3070 // Inhibit more partial peeling on this loop 3071 new_head_clone->set_partial_peel_loop(); 3072 C->set_major_progress(); 3073 loop->record_for_igvn(); 3074 3075 #if !defined(PRODUCT) 3076 if (TracePartialPeeling) { 3077 tty->print_cr("\nafter partial peel one iteration"); 3078 Node_List wl(area); 3079 Node* t = last_peel; 3080 while (true) { 3081 wl.push(t); 3082 if (t == head_clone) break; 3083 t = idom(t); 3084 } 3085 while (wl.size() > 0) { 3086 Node* tt = wl.pop(); 3087 if (tt == head) tty->print_cr("orig head"); 3088 else if (tt == new_head_clone) tty->print_cr("new head"); 3089 else if (tt == head_clone) tty->print_cr("clone head"); 3090 tt->dump(); 3091 } 3092 } 3093 #endif 3094 return true; 3095 } 3096 3097 //------------------------------reorg_offsets---------------------------------- 3098 // Reorganize offset computations to lower register pressure. Mostly 3099 // prevent loop-fallout uses of the pre-incremented trip counter (which are 3100 // then alive with the post-incremented trip counter forcing an extra 3101 // register move) 3102 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) { 3103 // Perform it only for canonical counted loops. 3104 // Loop's shape could be messed up by iteration_split_impl. 3105 if (!loop->_head->is_CountedLoop()) 3106 return; 3107 if (!loop->_head->as_Loop()->is_valid_counted_loop()) 3108 return; 3109 3110 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 3111 CountedLoopEndNode *cle = cl->loopexit(); 3112 Node *exit = cle->proj_out(false); 3113 Node *phi = cl->phi(); 3114 3115 // Check for the special case of folks using the pre-incremented 3116 // trip-counter on the fall-out path (forces the pre-incremented 3117 // and post-incremented trip counter to be live at the same time). 3118 // Fix this by adjusting to use the post-increment trip counter. 3119 3120 bool progress = true; 3121 while (progress) { 3122 progress = false; 3123 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) { 3124 Node* use = phi->fast_out(i); // User of trip-counter 3125 if (!has_ctrl(use)) continue; 3126 Node *u_ctrl = get_ctrl(use); 3127 if (use->is_Phi()) { 3128 u_ctrl = NULL; 3129 for (uint j = 1; j < use->req(); j++) 3130 if (use->in(j) == phi) 3131 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j)); 3132 } 3133 IdealLoopTree *u_loop = get_loop(u_ctrl); 3134 // Look for loop-invariant use 3135 if (u_loop == loop) continue; 3136 if (loop->is_member(u_loop)) continue; 3137 // Check that use is live out the bottom. Assuming the trip-counter 3138 // update is right at the bottom, uses of of the loop middle are ok. 3139 if (dom_lca(exit, u_ctrl) != exit) continue; 3140 // Hit! Refactor use to use the post-incremented tripcounter. 3141 // Compute a post-increment tripcounter. 3142 Node *opaq = new Opaque2Node( C, cle->incr() ); 3143 register_new_node(opaq, exit); 3144 Node *neg_stride = _igvn.intcon(-cle->stride_con()); 3145 set_ctrl(neg_stride, C->root()); 3146 Node *post = new AddINode( opaq, neg_stride); 3147 register_new_node(post, exit); 3148 _igvn.rehash_node_delayed(use); 3149 for (uint j = 1; j < use->req(); j++) { 3150 if (use->in(j) == phi) 3151 use->set_req(j, post); 3152 } 3153 // Since DU info changed, rerun loop 3154 progress = true; 3155 break; 3156 } 3157 } 3158 3159 }