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