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