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