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