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