1 /* 2 * Copyright (c) 1999, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "memory/allocation.inline.hpp" 27 #include "opto/addnode.hpp" 28 #include "opto/connode.hpp" 29 #include "opto/divnode.hpp" 30 #include "opto/loopnode.hpp" 31 #include "opto/matcher.hpp" 32 #include "opto/mulnode.hpp" 33 #include "opto/movenode.hpp" 34 #include "opto/opaquenode.hpp" 35 #include "opto/rootnode.hpp" 36 #include "opto/subnode.hpp" 37 38 //============================================================================= 39 //------------------------------split_thru_phi--------------------------------- 40 // Split Node 'n' through merge point if there is enough win. 41 Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) { 42 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) { 43 // ConvI2L may have type information on it which is unsafe to push up 44 // so disable this for now 45 return NULL; 46 } 47 48 int wins = 0; 49 assert(!n->is_CFG(), ""); 50 assert(region->is_Region(), ""); 51 52 const Type* type = n->bottom_type(); 53 const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr(); 54 Node *phi; 55 if (t_oop != NULL && t_oop->is_known_instance_field()) { 56 int iid = t_oop->instance_id(); 57 int index = C->get_alias_index(t_oop); 58 int offset = t_oop->offset(); 59 phi = new PhiNode(region, type, NULL, iid, index, offset); 60 } else { 61 phi = PhiNode::make_blank(region, n); 62 } 63 uint old_unique = C->unique(); 64 for (uint i = 1; i < region->req(); i++) { 65 Node *x; 66 Node* the_clone = NULL; 67 if (region->in(i) == C->top()) { 68 x = C->top(); // Dead path? Use a dead data op 69 } else { 70 x = n->clone(); // Else clone up the data op 71 the_clone = x; // Remember for possible deletion. 72 // Alter data node to use pre-phi inputs 73 if (n->in(0) == region) 74 x->set_req( 0, region->in(i) ); 75 for (uint j = 1; j < n->req(); j++) { 76 Node *in = n->in(j); 77 if (in->is_Phi() && in->in(0) == region) 78 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone 79 } 80 } 81 // Check for a 'win' on some paths 82 const Type *t = x->Value(&_igvn); 83 84 bool singleton = t->singleton(); 85 86 // A TOP singleton indicates that there are no possible values incoming 87 // along a particular edge. In most cases, this is OK, and the Phi will 88 // be eliminated later in an Ideal call. However, we can't allow this to 89 // happen if the singleton occurs on loop entry, as the elimination of 90 // the PhiNode may cause the resulting node to migrate back to a previous 91 // loop iteration. 92 if (singleton && t == Type::TOP) { 93 // Is_Loop() == false does not confirm the absence of a loop (e.g., an 94 // irreducible loop may not be indicated by an affirmative is_Loop()); 95 // therefore, the only top we can split thru a phi is on a backedge of 96 // a loop. 97 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 98 } 99 100 if (singleton) { 101 wins++; 102 x = ((PhaseGVN&)_igvn).makecon(t); 103 } else { 104 // We now call Identity to try to simplify the cloned node. 105 // Note that some Identity methods call phase->type(this). 106 // Make sure that the type array is big enough for 107 // our new node, even though we may throw the node away. 108 // (Note: This tweaking with igvn only works because x is a new node.) 109 _igvn.set_type(x, t); 110 // If x is a TypeNode, capture any more-precise type permanently into Node 111 // otherwise it will be not updated during igvn->transform since 112 // igvn->type(x) is set to x->Value() already. 113 x->raise_bottom_type(t); 114 Node *y = x->Identity(&_igvn); 115 if (y != x) { 116 wins++; 117 x = y; 118 } else { 119 y = _igvn.hash_find(x); 120 if (y) { 121 wins++; 122 x = y; 123 } else { 124 // Else x is a new node we are keeping 125 // We do not need register_new_node_with_optimizer 126 // because set_type has already been called. 127 _igvn._worklist.push(x); 128 } 129 } 130 } 131 if (x != the_clone && the_clone != NULL) 132 _igvn.remove_dead_node(the_clone); 133 phi->set_req( i, x ); 134 } 135 // Too few wins? 136 if (wins <= policy) { 137 _igvn.remove_dead_node(phi); 138 return NULL; 139 } 140 141 // Record Phi 142 register_new_node( phi, region ); 143 144 for (uint i2 = 1; i2 < phi->req(); i2++) { 145 Node *x = phi->in(i2); 146 // If we commoned up the cloned 'x' with another existing Node, 147 // the existing Node picks up a new use. We need to make the 148 // existing Node occur higher up so it dominates its uses. 149 Node *old_ctrl; 150 IdealLoopTree *old_loop; 151 152 if (x->is_Con()) { 153 // Constant's control is always root. 154 set_ctrl(x, C->root()); 155 continue; 156 } 157 // The occasional new node 158 if (x->_idx >= old_unique) { // Found a new, unplaced node? 159 old_ctrl = NULL; 160 old_loop = NULL; // Not in any prior loop 161 } else { 162 old_ctrl = get_ctrl(x); 163 old_loop = get_loop(old_ctrl); // Get prior loop 164 } 165 // New late point must dominate new use 166 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2)); 167 if (new_ctrl == old_ctrl) // Nothing is changed 168 continue; 169 170 IdealLoopTree *new_loop = get_loop(new_ctrl); 171 172 // Don't move x into a loop if its uses are 173 // outside of loop. Otherwise x will be cloned 174 // for each use outside of this loop. 175 IdealLoopTree *use_loop = get_loop(region); 176 if (!new_loop->is_member(use_loop) && 177 (old_loop == NULL || !new_loop->is_member(old_loop))) { 178 // Take early control, later control will be recalculated 179 // during next iteration of loop optimizations. 180 new_ctrl = get_early_ctrl(x); 181 new_loop = get_loop(new_ctrl); 182 } 183 // Set new location 184 set_ctrl(x, new_ctrl); 185 // If changing loop bodies, see if we need to collect into new body 186 if (old_loop != new_loop) { 187 if (old_loop && !old_loop->_child) 188 old_loop->_body.yank(x); 189 if (!new_loop->_child) 190 new_loop->_body.push(x); // Collect body info 191 } 192 } 193 194 return phi; 195 } 196 197 //------------------------------dominated_by------------------------------------ 198 // Replace the dominated test with an obvious true or false. Place it on the 199 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the 200 // live path up to the dominating control. 201 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff, bool flip, bool exclude_loop_predicate ) { 202 #ifndef PRODUCT 203 if (VerifyLoopOptimizations && PrintOpto) tty->print_cr("dominating test"); 204 #endif 205 206 207 // prevdom is the dominating projection of the dominating test. 208 assert( iff->is_If(), "" ); 209 assert( iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added"); 210 int pop = prevdom->Opcode(); 211 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" ); 212 if (flip) { 213 if (pop == Op_IfTrue) 214 pop = Op_IfFalse; 215 else 216 pop = Op_IfTrue; 217 } 218 // 'con' is set to true or false to kill the dominated test. 219 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO); 220 set_ctrl(con, C->root()); // Constant gets a new use 221 // Hack the dominated test 222 _igvn.replace_input_of(iff, 1, con); 223 224 // If I dont have a reachable TRUE and FALSE path following the IfNode then 225 // I can assume this path reaches an infinite loop. In this case it's not 226 // important to optimize the data Nodes - either the whole compilation will 227 // be tossed or this path (and all data Nodes) will go dead. 228 if (iff->outcnt() != 2) return; 229 230 // Make control-dependent data Nodes on the live path (path that will remain 231 // once the dominated IF is removed) become control-dependent on the 232 // dominating projection. 233 Node* dp = iff->as_If()->proj_out(pop == Op_IfTrue); 234 235 // Loop predicates may have depending checks which should not 236 // be skipped. For example, range check predicate has two checks 237 // for lower and upper bounds. 238 if (dp == NULL) 239 return; 240 241 ProjNode* dp_proj = dp->as_Proj(); 242 ProjNode* unc_proj = iff->as_If()->proj_out(1 - dp_proj->_con)->as_Proj(); 243 if (exclude_loop_predicate && 244 (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) != NULL || 245 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check) != NULL)) { 246 // If this is a range check (IfNode::is_range_check), do not 247 // reorder because Compile::allow_range_check_smearing might have 248 // changed the check. 249 return; // Let IGVN transformation change control dependence. 250 } 251 252 IdealLoopTree *old_loop = get_loop(dp); 253 254 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { 255 Node* cd = dp->fast_out(i); // Control-dependent node 256 if (cd->depends_only_on_test()) { 257 assert(cd->in(0) == dp, ""); 258 _igvn.replace_input_of(cd, 0, prevdom); 259 set_early_ctrl(cd); 260 IdealLoopTree *new_loop = get_loop(get_ctrl(cd)); 261 if (old_loop != new_loop) { 262 if (!old_loop->_child) old_loop->_body.yank(cd); 263 if (!new_loop->_child) new_loop->_body.push(cd); 264 } 265 --i; 266 --imax; 267 } 268 } 269 } 270 271 //------------------------------has_local_phi_input---------------------------- 272 // Return TRUE if 'n' has Phi inputs from its local block and no other 273 // block-local inputs (all non-local-phi inputs come from earlier blocks) 274 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) { 275 Node *n_ctrl = get_ctrl(n); 276 // See if some inputs come from a Phi in this block, or from before 277 // this block. 278 uint i; 279 for( i = 1; i < n->req(); i++ ) { 280 Node *phi = n->in(i); 281 if( phi->is_Phi() && phi->in(0) == n_ctrl ) 282 break; 283 } 284 if( i >= n->req() ) 285 return NULL; // No Phi inputs; nowhere to clone thru 286 287 // Check for inputs created between 'n' and the Phi input. These 288 // must split as well; they have already been given the chance 289 // (courtesy of a post-order visit) and since they did not we must 290 // recover the 'cost' of splitting them by being very profitable 291 // when splitting 'n'. Since this is unlikely we simply give up. 292 for( i = 1; i < n->req(); i++ ) { 293 Node *m = n->in(i); 294 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) { 295 // We allow the special case of AddP's with no local inputs. 296 // This allows us to split-up address expressions. 297 if (m->is_AddP() && 298 get_ctrl(m->in(2)) != n_ctrl && 299 get_ctrl(m->in(3)) != n_ctrl) { 300 // Move the AddP up to dominating point 301 set_ctrl_and_loop(m, find_non_split_ctrl(idom(n_ctrl))); 302 continue; 303 } 304 return NULL; 305 } 306 } 307 308 return n_ctrl; 309 } 310 311 //------------------------------remix_address_expressions---------------------- 312 // Rework addressing expressions to get the most loop-invariant stuff 313 // moved out. We'd like to do all associative operators, but it's especially 314 // important (common) to do address expressions. 315 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) { 316 if (!has_ctrl(n)) return NULL; 317 Node *n_ctrl = get_ctrl(n); 318 IdealLoopTree *n_loop = get_loop(n_ctrl); 319 320 // See if 'n' mixes loop-varying and loop-invariant inputs and 321 // itself is loop-varying. 322 323 // Only interested in binary ops (and AddP) 324 if( n->req() < 3 || n->req() > 4 ) return NULL; 325 326 Node *n1_ctrl = get_ctrl(n->in( 1)); 327 Node *n2_ctrl = get_ctrl(n->in( 2)); 328 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3)); 329 IdealLoopTree *n1_loop = get_loop( n1_ctrl ); 330 IdealLoopTree *n2_loop = get_loop( n2_ctrl ); 331 IdealLoopTree *n3_loop = get_loop( n3_ctrl ); 332 333 // Does one of my inputs spin in a tighter loop than self? 334 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) || 335 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) || 336 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) ) 337 return NULL; // Leave well enough alone 338 339 // Is at least one of my inputs loop-invariant? 340 if( n1_loop == n_loop && 341 n2_loop == n_loop && 342 n3_loop == n_loop ) 343 return NULL; // No loop-invariant inputs 344 345 346 int n_op = n->Opcode(); 347 348 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2). 349 if( n_op == Op_LShiftI ) { 350 // Scale is loop invariant 351 Node *scale = n->in(2); 352 Node *scale_ctrl = get_ctrl(scale); 353 IdealLoopTree *scale_loop = get_loop(scale_ctrl ); 354 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) ) 355 return NULL; 356 const TypeInt *scale_t = scale->bottom_type()->isa_int(); 357 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 ) 358 return NULL; // Dont bother with byte/short masking 359 // Add must vary with loop (else shift would be loop-invariant) 360 Node *add = n->in(1); 361 Node *add_ctrl = get_ctrl(add); 362 IdealLoopTree *add_loop = get_loop(add_ctrl); 363 //assert( n_loop == add_loop, "" ); 364 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops 365 366 // Convert I-V into I+ (0-V); same for V-I 367 if( add->Opcode() == Op_SubI && 368 _igvn.type( add->in(1) ) != TypeInt::ZERO ) { 369 Node *zero = _igvn.intcon(0); 370 set_ctrl(zero, C->root()); 371 Node *neg = new SubINode( _igvn.intcon(0), add->in(2) ); 372 register_new_node( neg, get_ctrl(add->in(2) ) ); 373 add = new AddINode( add->in(1), neg ); 374 register_new_node( add, add_ctrl ); 375 } 376 if( add->Opcode() != Op_AddI ) return NULL; 377 // See if one add input is loop invariant 378 Node *add_var = add->in(1); 379 Node *add_var_ctrl = get_ctrl(add_var); 380 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 381 Node *add_invar = add->in(2); 382 Node *add_invar_ctrl = get_ctrl(add_invar); 383 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl ); 384 if( add_var_loop == n_loop ) { 385 } else if( add_invar_loop == n_loop ) { 386 // Swap to find the invariant part 387 add_invar = add_var; 388 add_invar_ctrl = add_var_ctrl; 389 add_invar_loop = add_var_loop; 390 add_var = add->in(2); 391 Node *add_var_ctrl = get_ctrl(add_var); 392 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 393 } else // Else neither input is loop invariant 394 return NULL; 395 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) ) 396 return NULL; // No invariant part of the add? 397 398 // Yes! Reshape address expression! 399 Node *inv_scale = new LShiftINode( add_invar, scale ); 400 Node *inv_scale_ctrl = 401 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ? 402 add_invar_ctrl : scale_ctrl; 403 register_new_node( inv_scale, inv_scale_ctrl ); 404 Node *var_scale = new LShiftINode( add_var, scale ); 405 register_new_node( var_scale, n_ctrl ); 406 Node *var_add = new AddINode( var_scale, inv_scale ); 407 register_new_node( var_add, n_ctrl ); 408 _igvn.replace_node( n, var_add ); 409 return var_add; 410 } 411 412 // Replace (I+V) with (V+I) 413 if( n_op == Op_AddI || 414 n_op == Op_AddL || 415 n_op == Op_AddF || 416 n_op == Op_AddD || 417 n_op == Op_MulI || 418 n_op == Op_MulL || 419 n_op == Op_MulF || 420 n_op == Op_MulD ) { 421 if( n2_loop == n_loop ) { 422 assert( n1_loop != n_loop, "" ); 423 n->swap_edges(1, 2); 424 } 425 } 426 427 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V), 428 // but not if I2 is a constant. 429 if( n_op == Op_AddP ) { 430 if( n2_loop == n_loop && n3_loop != n_loop ) { 431 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) { 432 Node *n22_ctrl = get_ctrl(n->in(2)->in(2)); 433 Node *n23_ctrl = get_ctrl(n->in(2)->in(3)); 434 IdealLoopTree *n22loop = get_loop( n22_ctrl ); 435 IdealLoopTree *n23_loop = get_loop( n23_ctrl ); 436 if( n22loop != n_loop && n22loop->is_member(n_loop) && 437 n23_loop == n_loop ) { 438 Node *add1 = new AddPNode( n->in(1), n->in(2)->in(2), n->in(3) ); 439 // Stuff new AddP in the loop preheader 440 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) ); 441 Node *add2 = new AddPNode( n->in(1), add1, n->in(2)->in(3) ); 442 register_new_node( add2, n_ctrl ); 443 _igvn.replace_node( n, add2 ); 444 return add2; 445 } 446 } 447 } 448 449 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V) 450 if (n2_loop != n_loop && n3_loop == n_loop) { 451 if (n->in(3)->Opcode() == Op_AddX) { 452 Node *V = n->in(3)->in(1); 453 Node *I = n->in(3)->in(2); 454 if (is_member(n_loop,get_ctrl(V))) { 455 } else { 456 Node *tmp = V; V = I; I = tmp; 457 } 458 if (!is_member(n_loop,get_ctrl(I))) { 459 Node *add1 = new AddPNode(n->in(1), n->in(2), I); 460 // Stuff new AddP in the loop preheader 461 register_new_node(add1, n_loop->_head->in(LoopNode::EntryControl)); 462 Node *add2 = new AddPNode(n->in(1), add1, V); 463 register_new_node(add2, n_ctrl); 464 _igvn.replace_node(n, add2); 465 return add2; 466 } 467 } 468 } 469 } 470 471 return NULL; 472 } 473 474 //------------------------------conditional_move------------------------------- 475 // Attempt to replace a Phi with a conditional move. We have some pretty 476 // strict profitability requirements. All Phis at the merge point must 477 // be converted, so we can remove the control flow. We need to limit the 478 // number of c-moves to a small handful. All code that was in the side-arms 479 // of the CFG diamond is now speculatively executed. This code has to be 480 // "cheap enough". We are pretty much limited to CFG diamonds that merge 481 // 1 or 2 items with a total of 1 or 2 ops executed speculatively. 482 Node *PhaseIdealLoop::conditional_move( Node *region ) { 483 484 assert(region->is_Region(), "sanity check"); 485 if (region->req() != 3) return NULL; 486 487 // Check for CFG diamond 488 Node *lp = region->in(1); 489 Node *rp = region->in(2); 490 if (!lp || !rp) return NULL; 491 Node *lp_c = lp->in(0); 492 if (lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If()) return NULL; 493 IfNode *iff = lp_c->as_If(); 494 495 // Check for ops pinned in an arm of the diamond. 496 // Can't remove the control flow in this case 497 if (lp->outcnt() > 1) return NULL; 498 if (rp->outcnt() > 1) return NULL; 499 500 IdealLoopTree* r_loop = get_loop(region); 501 assert(r_loop == get_loop(iff), "sanity"); 502 // Always convert to CMOVE if all results are used only outside this loop. 503 bool used_inside_loop = (r_loop == _ltree_root); 504 505 // Check profitability 506 int cost = 0; 507 int phis = 0; 508 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 509 Node *out = region->fast_out(i); 510 if (!out->is_Phi()) continue; // Ignore other control edges, etc 511 phis++; 512 PhiNode* phi = out->as_Phi(); 513 BasicType bt = phi->type()->basic_type(); 514 switch (bt) { 515 case T_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 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) { 657 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 658 Node* u = m->fast_out(i); 659 if (u->is_CFG()) { 660 if (u->Opcode() == Op_NeverBranch) { 661 u = ((NeverBranchNode*)u)->proj_out(0); 662 enqueue_cfg_uses(u, wq); 663 } else { 664 wq.push(u); 665 } 666 } 667 } 668 } 669 670 // Try moving a store out of a loop, right before the loop 671 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) { 672 // Store has to be first in the loop body 673 IdealLoopTree *n_loop = get_loop(n_ctrl); 674 if (n->is_Store() && n_loop != _ltree_root && n_loop->is_loop() && n->in(0) != NULL) { 675 Node* address = n->in(MemNode::Address); 676 Node* value = n->in(MemNode::ValueIn); 677 Node* mem = n->in(MemNode::Memory); 678 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 679 IdealLoopTree* value_loop = get_loop(get_ctrl(value)); 680 681 // - address and value must be loop invariant 682 // - memory must be a memory Phi for the loop 683 // - Store must be the only store on this memory slice in the 684 // loop: if there's another store following this one then value 685 // written at iteration i by the second store could be overwritten 686 // at iteration i+n by the first store: it's not safe to move the 687 // first store out of the loop 688 // - nothing must observe the memory Phi: it guarantees no read 689 // before the store, we are also guaranteed the store post 690 // dominates the loop head (ignoring a possible early 691 // exit). Otherwise there would be extra Phi involved between the 692 // loop's Phi and the store. 693 // - there must be no early exit from the loop before the Store 694 // (such an exit most of the time would be an extra use of the 695 // memory Phi but sometimes is a bottom memory Phi that takes the 696 // store as input). 697 698 if (!n_loop->is_member(address_loop) && 699 !n_loop->is_member(value_loop) && 700 mem->is_Phi() && mem->in(0) == n_loop->_head && 701 mem->outcnt() == 1 && 702 mem->in(LoopNode::LoopBackControl) == n) { 703 704 assert(n_loop->_tail != NULL, "need a tail"); 705 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop"); 706 707 // Verify that there's no early exit of the loop before the store. 708 bool ctrl_ok = false; 709 { 710 // Follow control from loop head until n, we exit the loop or 711 // we reach the tail 712 ResourceMark rm; 713 Unique_Node_List wq; 714 wq.push(n_loop->_head); 715 716 for (uint next = 0; next < wq.size(); ++next) { 717 Node *m = wq.at(next); 718 if (m == n->in(0)) { 719 ctrl_ok = true; 720 continue; 721 } 722 assert(!has_ctrl(m), "should be CFG"); 723 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) { 724 ctrl_ok = false; 725 break; 726 } 727 enqueue_cfg_uses(m, wq); 728 if (wq.size() > 10) { 729 ctrl_ok = false; 730 break; 731 } 732 } 733 } 734 if (ctrl_ok) { 735 // move the Store 736 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem); 737 _igvn.replace_input_of(n, 0, n_loop->_head->in(LoopNode::EntryControl)); 738 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl)); 739 // Disconnect the phi now. An empty phi can confuse other 740 // optimizations in this pass of loop opts. 741 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl)); 742 n_loop->_body.yank(mem); 743 744 IdealLoopTree* new_loop = get_loop(n->in(0)); 745 set_ctrl_and_loop(n, n->in(0)); 746 747 return n; 748 } 749 } 750 } 751 return NULL; 752 } 753 754 // Try moving a store out of a loop, right after the loop 755 void PhaseIdealLoop::try_move_store_after_loop(Node* n) { 756 if (n->is_Store() && n->in(0) != NULL) { 757 Node *n_ctrl = get_ctrl(n); 758 IdealLoopTree *n_loop = get_loop(n_ctrl); 759 // Store must be in a loop 760 if (n_loop != _ltree_root && !n_loop->_irreducible) { 761 Node* address = n->in(MemNode::Address); 762 Node* value = n->in(MemNode::ValueIn); 763 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 764 // address must be loop invariant 765 if (!n_loop->is_member(address_loop)) { 766 // Store must be last on this memory slice in the loop and 767 // nothing in the loop must observe it 768 Node* phi = NULL; 769 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 770 Node* u = n->fast_out(i); 771 if (has_ctrl(u)) { // control use? 772 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 773 if (!n_loop->is_member(u_loop)) { 774 continue; 775 } 776 if (u->is_Phi() && u->in(0) == n_loop->_head) { 777 assert(_igvn.type(u) == Type::MEMORY, "bad phi"); 778 // multiple phis on the same slice are possible 779 if (phi != NULL) { 780 return; 781 } 782 phi = u; 783 continue; 784 } 785 } 786 return; 787 } 788 if (phi != NULL) { 789 // Nothing in the loop before the store (next iteration) 790 // must observe the stored value 791 bool mem_ok = true; 792 { 793 ResourceMark rm; 794 Unique_Node_List wq; 795 wq.push(phi); 796 for (uint next = 0; next < wq.size() && mem_ok; ++next) { 797 Node *m = wq.at(next); 798 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) { 799 Node* u = m->fast_out(i); 800 if (u->is_Store() || u->is_Phi()) { 801 if (u != n) { 802 wq.push(u); 803 mem_ok = (wq.size() <= 10); 804 } 805 } else { 806 mem_ok = false; 807 break; 808 } 809 } 810 } 811 } 812 if (mem_ok) { 813 // Move the Store out of the loop creating clones along 814 // all paths out of the loop that observe the stored value 815 _igvn.rehash_node_delayed(phi); 816 int count = phi->replace_edge(n, n->in(MemNode::Memory)); 817 assert(count > 0, "inconsistent phi"); 818 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 819 Node* u = n->fast_out(i); 820 Node* c = get_ctrl(u); 821 822 if (u->is_Phi()) { 823 c = u->in(0)->in(u->find_edge(n)); 824 } 825 IdealLoopTree *u_loop = get_loop(c); 826 assert (!n_loop->is_member(u_loop), "only the phi should have been a use in the loop"); 827 while(true) { 828 Node* next_c = find_non_split_ctrl(idom(c)); 829 if (n_loop->is_member(get_loop(next_c))) { 830 break; 831 } 832 c = next_c; 833 } 834 835 Node* st = n->clone(); 836 st->set_req(0, c); 837 _igvn.register_new_node_with_optimizer(st); 838 839 set_ctrl(st, c); 840 IdealLoopTree* new_loop = get_loop(c); 841 assert(new_loop != n_loop, "should be moved out of loop"); 842 if (new_loop->_child == NULL) new_loop->_body.push(st); 843 844 _igvn.replace_input_of(u, u->find_edge(n), st); 845 --imax; 846 --i; 847 } 848 849 850 assert(n->outcnt() == 0, "all uses should be gone"); 851 _igvn.replace_input_of(n, MemNode::Memory, C->top()); 852 // Disconnect the phi now. An empty phi can confuse other 853 // optimizations in this pass of loop opts.. 854 if (phi->in(LoopNode::LoopBackControl) == phi) { 855 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl)); 856 n_loop->_body.yank(phi); 857 } 858 } 859 } 860 } 861 } 862 } 863 } 864 865 //------------------------------split_if_with_blocks_pre----------------------- 866 // Do the real work in a non-recursive function. Data nodes want to be 867 // cloned in the pre-order so they can feed each other nicely. 868 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) { 869 // Cloning these guys is unlikely to win 870 int n_op = n->Opcode(); 871 if( n_op == Op_MergeMem ) return n; 872 if( n->is_Proj() ) return n; 873 // Do not clone-up CmpFXXX variations, as these are always 874 // followed by a CmpI 875 if( n->is_Cmp() ) return n; 876 // Attempt to use a conditional move instead of a phi/branch 877 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) { 878 Node *cmov = conditional_move( n ); 879 if( cmov ) return cmov; 880 } 881 if( n->is_CFG() || n->is_LoadStore() ) 882 return n; 883 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd 884 n_op == Op_Opaque2 ) { 885 if( !C->major_progress() ) // If chance of no more loop opts... 886 _igvn._worklist.push(n); // maybe we'll remove them 887 return n; 888 } 889 890 if( n->is_Con() ) return n; // No cloning for Con nodes 891 892 Node *n_ctrl = get_ctrl(n); 893 if( !n_ctrl ) return n; // Dead node 894 895 Node* res = try_move_store_before_loop(n, n_ctrl); 896 if (res != NULL) { 897 return n; 898 } 899 900 // Attempt to remix address expressions for loop invariants 901 Node *m = remix_address_expressions( n ); 902 if( m ) return m; 903 904 // Determine if the Node has inputs from some local Phi. 905 // Returns the block to clone thru. 906 Node *n_blk = has_local_phi_input( n ); 907 if( !n_blk ) return n; 908 909 // Do not clone the trip counter through on a CountedLoop 910 // (messes up the canonical shape). 911 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n; 912 913 // Check for having no control input; not pinned. Allow 914 // dominating control. 915 if (n->in(0)) { 916 Node *dom = idom(n_blk); 917 if (dom_lca(n->in(0), dom) != n->in(0)) { 918 return n; 919 } 920 } 921 // Policy: when is it profitable. You must get more wins than 922 // policy before it is considered profitable. Policy is usually 0, 923 // so 1 win is considered profitable. Big merges will require big 924 // cloning, so get a larger policy. 925 int policy = n_blk->req() >> 2; 926 927 // If the loop is a candidate for range check elimination, 928 // delay splitting through it's phi until a later loop optimization 929 if (n_blk->is_CountedLoop()) { 930 IdealLoopTree *lp = get_loop(n_blk); 931 if (lp && lp->_rce_candidate) { 932 return n; 933 } 934 } 935 936 // Use same limit as split_if_with_blocks_post 937 if( C->live_nodes() > 35000 ) return n; // Method too big 938 939 // Split 'n' through the merge point if it is profitable 940 Node *phi = split_thru_phi( n, n_blk, policy ); 941 if (!phi) return n; 942 943 // Found a Phi to split thru! 944 // Replace 'n' with the new phi 945 _igvn.replace_node( n, phi ); 946 // Moved a load around the loop, 'en-registering' something. 947 if (n_blk->is_Loop() && n->is_Load() && 948 !phi->in(LoopNode::LoopBackControl)->is_Load()) 949 C->set_major_progress(); 950 951 return phi; 952 } 953 954 static bool merge_point_too_heavy(Compile* C, Node* region) { 955 // Bail out if the region and its phis have too many users. 956 int weight = 0; 957 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 958 weight += region->fast_out(i)->outcnt(); 959 } 960 int nodes_left = C->max_node_limit() - C->live_nodes(); 961 if (weight * 8 > nodes_left) { 962 #ifndef PRODUCT 963 if (PrintOpto) 964 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight); 965 #endif 966 return true; 967 } else { 968 return false; 969 } 970 } 971 972 static bool merge_point_safe(Node* region) { 973 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode 974 // having a PhiNode input. This sidesteps the dangerous case where the split 975 // ConvI2LNode may become TOP if the input Value() does not 976 // overlap the ConvI2L range, leaving a node which may not dominate its 977 // uses. 978 // A better fix for this problem can be found in the BugTraq entry, but 979 // expediency for Mantis demands this hack. 980 // 6855164: If the merge point has a FastLockNode with a PhiNode input, we stop 981 // split_if_with_blocks from splitting a block because we could not move around 982 // the FastLockNode. 983 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 984 Node* n = region->fast_out(i); 985 if (n->is_Phi()) { 986 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 987 Node* m = n->fast_out(j); 988 if (m->is_FastLock()) 989 return false; 990 #ifdef _LP64 991 if (m->Opcode() == Op_ConvI2L) 992 return false; 993 #endif 994 } 995 } 996 } 997 return true; 998 } 999 1000 1001 //------------------------------place_near_use--------------------------------- 1002 // Place some computation next to use but not inside inner loops. 1003 // For inner loop uses move it to the preheader area. 1004 Node *PhaseIdealLoop::place_near_use( Node *useblock ) const { 1005 IdealLoopTree *u_loop = get_loop( useblock ); 1006 return (u_loop->_irreducible || u_loop->_child) 1007 ? useblock 1008 : u_loop->_head->in(LoopNode::EntryControl); 1009 } 1010 1011 1012 //------------------------------split_if_with_blocks_post---------------------- 1013 // Do the real work in a non-recursive function. CFG hackery wants to be 1014 // in the post-order, so it can dirty the I-DOM info and not use the dirtied 1015 // info. 1016 void PhaseIdealLoop::split_if_with_blocks_post( Node *n ) { 1017 1018 // Cloning Cmp through Phi's involves the split-if transform. 1019 // FastLock is not used by an If 1020 if( n->is_Cmp() && !n->is_FastLock() ) { 1021 if( C->live_nodes() > 35000 ) return; // Method too big 1022 1023 // Do not do 'split-if' if irreducible loops are present. 1024 if( _has_irreducible_loops ) 1025 return; 1026 1027 Node *n_ctrl = get_ctrl(n); 1028 // Determine if the Node has inputs from some local Phi. 1029 // Returns the block to clone thru. 1030 Node *n_blk = has_local_phi_input( n ); 1031 if( n_blk != n_ctrl ) return; 1032 1033 if( merge_point_too_heavy(C, n_ctrl) ) 1034 return; 1035 1036 if( n->outcnt() != 1 ) return; // Multiple bool's from 1 compare? 1037 Node *bol = n->unique_out(); 1038 assert( bol->is_Bool(), "expect a bool here" ); 1039 if( bol->outcnt() != 1 ) return;// Multiple branches from 1 compare? 1040 Node *iff = bol->unique_out(); 1041 1042 // Check some safety conditions 1043 if( iff->is_If() ) { // Classic split-if? 1044 if( iff->in(0) != n_ctrl ) return; // Compare must be in same blk as if 1045 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE 1046 // Can't split CMove with different control edge. 1047 if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) return; 1048 if( get_ctrl(iff->in(2)) == n_ctrl || 1049 get_ctrl(iff->in(3)) == n_ctrl ) 1050 return; // Inputs not yet split-up 1051 if ( get_loop(n_ctrl) != get_loop(get_ctrl(iff)) ) { 1052 return; // Loop-invar test gates loop-varying CMOVE 1053 } 1054 } else { 1055 return; // some other kind of node, such as an Allocate 1056 } 1057 1058 // Do not do 'split-if' if some paths are dead. First do dead code 1059 // elimination and then see if its still profitable. 1060 for( uint i = 1; i < n_ctrl->req(); i++ ) 1061 if( n_ctrl->in(i) == C->top() ) 1062 return; 1063 1064 // When is split-if profitable? Every 'win' on means some control flow 1065 // goes dead, so it's almost always a win. 1066 int policy = 0; 1067 // If trying to do a 'Split-If' at the loop head, it is only 1068 // profitable if the cmp folds up on BOTH paths. Otherwise we 1069 // risk peeling a loop forever. 1070 1071 // CNC - Disabled for now. Requires careful handling of loop 1072 // body selection for the cloned code. Also, make sure we check 1073 // for any input path not being in the same loop as n_ctrl. For 1074 // irreducible loops we cannot check for 'n_ctrl->is_Loop()' 1075 // because the alternative loop entry points won't be converted 1076 // into LoopNodes. 1077 IdealLoopTree *n_loop = get_loop(n_ctrl); 1078 for( uint j = 1; j < n_ctrl->req(); j++ ) 1079 if( get_loop(n_ctrl->in(j)) != n_loop ) 1080 return; 1081 1082 // Check for safety of the merge point. 1083 if( !merge_point_safe(n_ctrl) ) { 1084 return; 1085 } 1086 1087 // Split compare 'n' through the merge point if it is profitable 1088 Node *phi = split_thru_phi( n, n_ctrl, policy ); 1089 if( !phi ) return; 1090 1091 // Found a Phi to split thru! 1092 // Replace 'n' with the new phi 1093 _igvn.replace_node( n, phi ); 1094 1095 // Now split the bool up thru the phi 1096 Node *bolphi = split_thru_phi( bol, n_ctrl, -1 ); 1097 guarantee(bolphi != NULL, "null boolean phi node"); 1098 1099 _igvn.replace_node( bol, bolphi ); 1100 assert( iff->in(1) == bolphi, "" ); 1101 1102 if( bolphi->Value(&_igvn)->singleton() ) 1103 return; 1104 1105 // Conditional-move? Must split up now 1106 if( !iff->is_If() ) { 1107 Node *cmovphi = split_thru_phi( iff, n_ctrl, -1 ); 1108 _igvn.replace_node( iff, cmovphi ); 1109 return; 1110 } 1111 1112 // Now split the IF 1113 do_split_if( iff ); 1114 return; 1115 } 1116 1117 // Check for an IF ready to split; one that has its 1118 // condition codes input coming from a Phi at the block start. 1119 int n_op = n->Opcode(); 1120 1121 // Check for an IF being dominated by another IF same test 1122 if (n_op == Op_If) { 1123 Node *bol = n->in(1); 1124 uint max = bol->outcnt(); 1125 // Check for same test used more than once? 1126 if (max > 1 && bol->is_Bool()) { 1127 // Search up IDOMs to see if this IF is dominated. 1128 Node *cutoff = get_ctrl(bol); 1129 1130 // Now search up IDOMs till cutoff, looking for a dominating test 1131 Node *prevdom = n; 1132 Node *dom = idom(prevdom); 1133 while (dom != cutoff) { 1134 if (dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom) { 1135 // Replace the dominated test with an obvious true or false. 1136 // Place it on the IGVN worklist for later cleanup. 1137 C->set_major_progress(); 1138 dominated_by(prevdom, n, false, true); 1139 #ifndef PRODUCT 1140 if( VerifyLoopOptimizations ) verify(); 1141 #endif 1142 return; 1143 } 1144 prevdom = dom; 1145 dom = idom(prevdom); 1146 } 1147 } 1148 } 1149 1150 // See if a shared loop-varying computation has no loop-varying uses. 1151 // Happens if something is only used for JVM state in uncommon trap exits, 1152 // like various versions of induction variable+offset. Clone the 1153 // computation per usage to allow it to sink out of the loop. 1154 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about) 1155 Node *n_ctrl = get_ctrl(n); 1156 IdealLoopTree *n_loop = get_loop(n_ctrl); 1157 if( n_loop != _ltree_root ) { 1158 DUIterator_Fast imax, i = n->fast_outs(imax); 1159 for (; i < imax; i++) { 1160 Node* u = n->fast_out(i); 1161 if( !has_ctrl(u) ) break; // Found control user 1162 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 1163 if( u_loop == n_loop ) break; // Found loop-varying use 1164 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop 1165 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003 1166 } 1167 bool did_break = (i < imax); // Did we break out of the previous loop? 1168 if (!did_break && n->outcnt() > 1) { // All uses in outer loops! 1169 Node *late_load_ctrl = NULL; 1170 if (n->is_Load()) { 1171 // If n is a load, get and save the result from get_late_ctrl(), 1172 // to be later used in calculating the control for n's clones. 1173 clear_dom_lca_tags(); 1174 late_load_ctrl = get_late_ctrl(n, n_ctrl); 1175 } 1176 // If n is a load, and the late control is the same as the current 1177 // control, then the cloning of n is a pointless exercise, because 1178 // GVN will ensure that we end up where we started. 1179 if (!n->is_Load() || late_load_ctrl != n_ctrl) { 1180 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) { 1181 Node *u = n->last_out(j); // Clone private computation per use 1182 _igvn.rehash_node_delayed(u); 1183 Node *x = n->clone(); // Clone computation 1184 Node *x_ctrl = NULL; 1185 if( u->is_Phi() ) { 1186 // Replace all uses of normal nodes. Replace Phi uses 1187 // individually, so the separate Nodes can sink down 1188 // different paths. 1189 uint k = 1; 1190 while( u->in(k) != n ) k++; 1191 u->set_req( k, x ); 1192 // x goes next to Phi input path 1193 x_ctrl = u->in(0)->in(k); 1194 --j; 1195 } else { // Normal use 1196 // Replace all uses 1197 for( uint k = 0; k < u->req(); k++ ) { 1198 if( u->in(k) == n ) { 1199 u->set_req( k, x ); 1200 --j; 1201 } 1202 } 1203 x_ctrl = get_ctrl(u); 1204 } 1205 1206 // Find control for 'x' next to use but not inside inner loops. 1207 // For inner loop uses get the preheader area. 1208 x_ctrl = place_near_use(x_ctrl); 1209 1210 if (n->is_Load()) { 1211 // For loads, add a control edge to a CFG node outside of the loop 1212 // to force them to not combine and return back inside the loop 1213 // during GVN optimization (4641526). 1214 // 1215 // Because we are setting the actual control input, factor in 1216 // the result from get_late_ctrl() so we respect any 1217 // anti-dependences. (6233005). 1218 x_ctrl = dom_lca(late_load_ctrl, x_ctrl); 1219 1220 // Don't allow the control input to be a CFG splitting node. 1221 // Such nodes should only have ProjNodes as outs, e.g. IfNode 1222 // should only have IfTrueNode and IfFalseNode (4985384). 1223 x_ctrl = find_non_split_ctrl(x_ctrl); 1224 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone"); 1225 1226 x->set_req(0, x_ctrl); 1227 } 1228 register_new_node(x, x_ctrl); 1229 1230 // Some institutional knowledge is needed here: 'x' is 1231 // yanked because if the optimizer runs GVN on it all the 1232 // cloned x's will common up and undo this optimization and 1233 // be forced back in the loop. This is annoying because it 1234 // makes +VerifyOpto report false-positives on progress. I 1235 // tried setting control edges on the x's to force them to 1236 // not combine, but the matching gets worried when it tries 1237 // to fold a StoreP and an AddP together (as part of an 1238 // address expression) and the AddP and StoreP have 1239 // different controls. 1240 if (!x->is_Load() && !x->is_DecodeNarrowPtr()) _igvn._worklist.yank(x); 1241 } 1242 _igvn.remove_dead_node(n); 1243 } 1244 } 1245 } 1246 } 1247 1248 try_move_store_after_loop(n); 1249 1250 // Check for Opaque2's who's loop has disappeared - who's input is in the 1251 // same loop nest as their output. Remove 'em, they are no longer useful. 1252 if( n_op == Op_Opaque2 && 1253 n->in(1) != NULL && 1254 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) { 1255 _igvn.replace_node( n, n->in(1) ); 1256 } 1257 } 1258 1259 //------------------------------split_if_with_blocks--------------------------- 1260 // Check for aggressive application of 'split-if' optimization, 1261 // using basic block level info. 1262 void PhaseIdealLoop::split_if_with_blocks( VectorSet &visited, Node_Stack &nstack ) { 1263 Node *n = C->root(); 1264 visited.set(n->_idx); // first, mark node as visited 1265 // Do pre-visit work for root 1266 n = split_if_with_blocks_pre( n ); 1267 uint cnt = n->outcnt(); 1268 uint i = 0; 1269 while (true) { 1270 // Visit all children 1271 if (i < cnt) { 1272 Node* use = n->raw_out(i); 1273 ++i; 1274 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) { 1275 // Now do pre-visit work for this use 1276 use = split_if_with_blocks_pre( use ); 1277 nstack.push(n, i); // Save parent and next use's index. 1278 n = use; // Process all children of current use. 1279 cnt = use->outcnt(); 1280 i = 0; 1281 } 1282 } 1283 else { 1284 // All of n's children have been processed, complete post-processing. 1285 if (cnt != 0 && !n->is_Con()) { 1286 assert(has_node(n), "no dead nodes"); 1287 split_if_with_blocks_post( n ); 1288 } 1289 if (nstack.is_empty()) { 1290 // Finished all nodes on stack. 1291 break; 1292 } 1293 // Get saved parent node and next use's index. Visit the rest of uses. 1294 n = nstack.node(); 1295 cnt = n->outcnt(); 1296 i = nstack.index(); 1297 nstack.pop(); 1298 } 1299 } 1300 } 1301 1302 1303 //============================================================================= 1304 // 1305 // C L O N E A L O O P B O D Y 1306 // 1307 1308 //------------------------------clone_iff-------------------------------------- 1309 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1310 // "Nearly" because all Nodes have been cloned from the original in the loop, 1311 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1312 // through the Phi recursively, and return a Bool. 1313 BoolNode *PhaseIdealLoop::clone_iff( PhiNode *phi, IdealLoopTree *loop ) { 1314 1315 // Convert this Phi into a Phi merging Bools 1316 uint i; 1317 for( i = 1; i < phi->req(); i++ ) { 1318 Node *b = phi->in(i); 1319 if( b->is_Phi() ) { 1320 _igvn.replace_input_of(phi, i, clone_iff( b->as_Phi(), loop )); 1321 } else { 1322 assert( b->is_Bool(), "" ); 1323 } 1324 } 1325 1326 Node *sample_bool = phi->in(1); 1327 Node *sample_cmp = sample_bool->in(1); 1328 1329 // Make Phis to merge the Cmp's inputs. 1330 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP ); 1331 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP ); 1332 for( i = 1; i < phi->req(); i++ ) { 1333 Node *n1 = phi->in(i)->in(1)->in(1); 1334 Node *n2 = phi->in(i)->in(1)->in(2); 1335 phi1->set_req( i, n1 ); 1336 phi2->set_req( i, n2 ); 1337 phi1->set_type( phi1->type()->meet_speculative(n1->bottom_type())); 1338 phi2->set_type( phi2->type()->meet_speculative(n2->bottom_type())); 1339 } 1340 // See if these Phis have been made before. 1341 // Register with optimizer 1342 Node *hit1 = _igvn.hash_find_insert(phi1); 1343 if( hit1 ) { // Hit, toss just made Phi 1344 _igvn.remove_dead_node(phi1); // Remove new phi 1345 assert( hit1->is_Phi(), "" ); 1346 phi1 = (PhiNode*)hit1; // Use existing phi 1347 } else { // Miss 1348 _igvn.register_new_node_with_optimizer(phi1); 1349 } 1350 Node *hit2 = _igvn.hash_find_insert(phi2); 1351 if( hit2 ) { // Hit, toss just made Phi 1352 _igvn.remove_dead_node(phi2); // Remove new phi 1353 assert( hit2->is_Phi(), "" ); 1354 phi2 = (PhiNode*)hit2; // Use existing phi 1355 } else { // Miss 1356 _igvn.register_new_node_with_optimizer(phi2); 1357 } 1358 // Register Phis with loop/block info 1359 set_ctrl(phi1, phi->in(0)); 1360 set_ctrl(phi2, phi->in(0)); 1361 // Make a new Cmp 1362 Node *cmp = sample_cmp->clone(); 1363 cmp->set_req( 1, phi1 ); 1364 cmp->set_req( 2, phi2 ); 1365 _igvn.register_new_node_with_optimizer(cmp); 1366 set_ctrl(cmp, phi->in(0)); 1367 1368 // Make a new Bool 1369 Node *b = sample_bool->clone(); 1370 b->set_req(1,cmp); 1371 _igvn.register_new_node_with_optimizer(b); 1372 set_ctrl(b, phi->in(0)); 1373 1374 assert( b->is_Bool(), "" ); 1375 return (BoolNode*)b; 1376 } 1377 1378 //------------------------------clone_bool------------------------------------- 1379 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1380 // "Nearly" because all Nodes have been cloned from the original in the loop, 1381 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1382 // through the Phi recursively, and return a Bool. 1383 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) { 1384 uint i; 1385 // Convert this Phi into a Phi merging Bools 1386 for( i = 1; i < phi->req(); i++ ) { 1387 Node *b = phi->in(i); 1388 if( b->is_Phi() ) { 1389 _igvn.replace_input_of(phi, i, clone_bool( b->as_Phi(), loop )); 1390 } else { 1391 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" ); 1392 } 1393 } 1394 1395 Node *sample_cmp = phi->in(1); 1396 1397 // Make Phis to merge the Cmp's inputs. 1398 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP ); 1399 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP ); 1400 for( uint j = 1; j < phi->req(); j++ ) { 1401 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP 1402 Node *n1, *n2; 1403 if( cmp_top->is_Cmp() ) { 1404 n1 = cmp_top->in(1); 1405 n2 = cmp_top->in(2); 1406 } else { 1407 n1 = n2 = cmp_top; 1408 } 1409 phi1->set_req( j, n1 ); 1410 phi2->set_req( j, n2 ); 1411 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 1412 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 1413 } 1414 1415 // See if these Phis have been made before. 1416 // Register with optimizer 1417 Node *hit1 = _igvn.hash_find_insert(phi1); 1418 if( hit1 ) { // Hit, toss just made Phi 1419 _igvn.remove_dead_node(phi1); // Remove new phi 1420 assert( hit1->is_Phi(), "" ); 1421 phi1 = (PhiNode*)hit1; // Use existing phi 1422 } else { // Miss 1423 _igvn.register_new_node_with_optimizer(phi1); 1424 } 1425 Node *hit2 = _igvn.hash_find_insert(phi2); 1426 if( hit2 ) { // Hit, toss just made Phi 1427 _igvn.remove_dead_node(phi2); // Remove new phi 1428 assert( hit2->is_Phi(), "" ); 1429 phi2 = (PhiNode*)hit2; // Use existing phi 1430 } else { // Miss 1431 _igvn.register_new_node_with_optimizer(phi2); 1432 } 1433 // Register Phis with loop/block info 1434 set_ctrl(phi1, phi->in(0)); 1435 set_ctrl(phi2, phi->in(0)); 1436 // Make a new Cmp 1437 Node *cmp = sample_cmp->clone(); 1438 cmp->set_req( 1, phi1 ); 1439 cmp->set_req( 2, phi2 ); 1440 _igvn.register_new_node_with_optimizer(cmp); 1441 set_ctrl(cmp, phi->in(0)); 1442 1443 assert( cmp->is_Cmp(), "" ); 1444 return (CmpNode*)cmp; 1445 } 1446 1447 //------------------------------sink_use--------------------------------------- 1448 // If 'use' was in the loop-exit block, it now needs to be sunk 1449 // below the post-loop merge point. 1450 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) { 1451 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) { 1452 set_ctrl(use, post_loop); 1453 for (DUIterator j = use->outs(); use->has_out(j); j++) 1454 sink_use(use->out(j), post_loop); 1455 } 1456 } 1457 1458 //------------------------------clone_loop------------------------------------- 1459 // 1460 // C L O N E A L O O P B O D Y 1461 // 1462 // This is the basic building block of the loop optimizations. It clones an 1463 // entire loop body. It makes an old_new loop body mapping; with this mapping 1464 // you can find the new-loop equivalent to an old-loop node. All new-loop 1465 // nodes are exactly equal to their old-loop counterparts, all edges are the 1466 // same. All exits from the old-loop now have a RegionNode that merges the 1467 // equivalent new-loop path. This is true even for the normal "loop-exit" 1468 // condition. All uses of loop-invariant old-loop values now come from (one 1469 // or more) Phis that merge their new-loop equivalents. 1470 // 1471 // This operation leaves the graph in an illegal state: there are two valid 1472 // control edges coming from the loop pre-header to both loop bodies. I'll 1473 // definitely have to hack the graph after running this transform. 1474 // 1475 // From this building block I will further edit edges to perform loop peeling 1476 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc. 1477 // 1478 // Parameter side_by_size_idom: 1479 // When side_by_size_idom is NULL, the dominator tree is constructed for 1480 // the clone loop to dominate the original. Used in construction of 1481 // pre-main-post loop sequence. 1482 // When nonnull, the clone and original are side-by-side, both are 1483 // dominated by the side_by_side_idom node. Used in construction of 1484 // unswitched loops. 1485 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd, 1486 Node* side_by_side_idom) { 1487 1488 #ifndef PRODUCT 1489 if (C->do_vector_loop() && PrintOpto) { 1490 const char* mname = C->method()->name()->as_quoted_ascii(); 1491 if (mname != NULL) { 1492 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname); 1493 } 1494 } 1495 #endif 1496 1497 CloneMap& cm = C->clone_map(); 1498 Dict* dict = cm.dict(); 1499 if (C->do_vector_loop()) { 1500 cm.set_clone_idx(cm.max_gen()+1); 1501 #ifndef PRODUCT 1502 if (PrintOpto) { 1503 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx()); 1504 loop->dump_head(); 1505 } 1506 #endif 1507 } 1508 1509 // Step 1: Clone the loop body. Make the old->new mapping. 1510 uint i; 1511 for( i = 0; i < loop->_body.size(); i++ ) { 1512 Node *old = loop->_body.at(i); 1513 Node *nnn = old->clone(); 1514 old_new.map( old->_idx, nnn ); 1515 if (C->do_vector_loop()) { 1516 cm.verify_insert_and_clone(old, nnn, cm.clone_idx()); 1517 } 1518 _igvn.register_new_node_with_optimizer(nnn); 1519 } 1520 1521 1522 // Step 2: Fix the edges in the new body. If the old input is outside the 1523 // loop use it. If the old input is INside the loop, use the corresponding 1524 // new node instead. 1525 for( i = 0; i < loop->_body.size(); i++ ) { 1526 Node *old = loop->_body.at(i); 1527 Node *nnn = old_new[old->_idx]; 1528 // Fix CFG/Loop controlling the new node 1529 if (has_ctrl(old)) { 1530 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]); 1531 } else { 1532 set_loop(nnn, loop->_parent); 1533 if (old->outcnt() > 0) { 1534 set_idom( nnn, old_new[idom(old)->_idx], dd ); 1535 } 1536 } 1537 // Correct edges to the new node 1538 for( uint j = 0; j < nnn->req(); j++ ) { 1539 Node *n = nnn->in(j); 1540 if( n ) { 1541 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n ); 1542 if( loop->is_member( old_in_loop ) ) 1543 nnn->set_req(j, old_new[n->_idx]); 1544 } 1545 } 1546 _igvn.hash_find_insert(nnn); 1547 } 1548 Node *newhead = old_new[loop->_head->_idx]; 1549 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 1550 1551 1552 // Step 3: Now fix control uses. Loop varying control uses have already 1553 // been fixed up (as part of all input edges in Step 2). Loop invariant 1554 // control uses must be either an IfFalse or an IfTrue. Make a merge 1555 // point to merge the old and new IfFalse/IfTrue nodes; make the use 1556 // refer to this. 1557 ResourceArea *area = Thread::current()->resource_area(); 1558 Node_List worklist(area); 1559 uint new_counter = C->unique(); 1560 for( i = 0; i < loop->_body.size(); i++ ) { 1561 Node* old = loop->_body.at(i); 1562 if( !old->is_CFG() ) continue; 1563 Node* nnn = old_new[old->_idx]; 1564 1565 // Copy uses to a worklist, so I can munge the def-use info 1566 // with impunity. 1567 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1568 worklist.push(old->fast_out(j)); 1569 1570 while( worklist.size() ) { // Visit all uses 1571 Node *use = worklist.pop(); 1572 if (!has_node(use)) continue; // Ignore dead nodes 1573 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1574 if( !loop->is_member( use_loop ) && use->is_CFG() ) { 1575 // Both OLD and USE are CFG nodes here. 1576 assert( use->is_Proj(), "" ); 1577 1578 // Clone the loop exit control projection 1579 Node *newuse = use->clone(); 1580 if (C->do_vector_loop()) { 1581 cm.verify_insert_and_clone(use, newuse, cm.clone_idx()); 1582 } 1583 newuse->set_req(0,nnn); 1584 _igvn.register_new_node_with_optimizer(newuse); 1585 set_loop(newuse, use_loop); 1586 set_idom(newuse, nnn, dom_depth(nnn) + 1 ); 1587 1588 // We need a Region to merge the exit from the peeled body and the 1589 // exit from the old loop body. 1590 RegionNode *r = new RegionNode(3); 1591 // Map the old use to the new merge point 1592 old_new.map( use->_idx, r ); 1593 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use)); 1594 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" ); 1595 1596 // The original user of 'use' uses 'r' instead. 1597 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) { 1598 Node* useuse = use->last_out(l); 1599 _igvn.rehash_node_delayed(useuse); 1600 uint uses_found = 0; 1601 if( useuse->in(0) == use ) { 1602 useuse->set_req(0, r); 1603 uses_found++; 1604 if( useuse->is_CFG() ) { 1605 assert( dom_depth(useuse) > dd_r, "" ); 1606 set_idom(useuse, r, dom_depth(useuse)); 1607 } 1608 } 1609 for( uint k = 1; k < useuse->req(); k++ ) { 1610 if( useuse->in(k) == use ) { 1611 useuse->set_req(k, r); 1612 uses_found++; 1613 } 1614 } 1615 l -= uses_found; // we deleted 1 or more copies of this edge 1616 } 1617 1618 // Now finish up 'r' 1619 r->set_req( 1, newuse ); 1620 r->set_req( 2, use ); 1621 _igvn.register_new_node_with_optimizer(r); 1622 set_loop(r, use_loop); 1623 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r); 1624 } // End of if a loop-exit test 1625 } 1626 } 1627 1628 // Step 4: If loop-invariant use is not control, it must be dominated by a 1629 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region 1630 // there if needed. Make a Phi there merging old and new used values. 1631 Node_List *split_if_set = NULL; 1632 Node_List *split_bool_set = NULL; 1633 Node_List *split_cex_set = NULL; 1634 for( i = 0; i < loop->_body.size(); i++ ) { 1635 Node* old = loop->_body.at(i); 1636 Node* nnn = old_new[old->_idx]; 1637 // Copy uses to a worklist, so I can munge the def-use info 1638 // with impunity. 1639 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1640 worklist.push(old->fast_out(j)); 1641 1642 while( worklist.size() ) { 1643 Node *use = worklist.pop(); 1644 if (!has_node(use)) continue; // Ignore dead nodes 1645 if (use->in(0) == C->top()) continue; 1646 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1647 // Check for data-use outside of loop - at least one of OLD or USE 1648 // must not be a CFG node. 1649 if( !loop->is_member( use_loop ) && (!old->is_CFG() || !use->is_CFG())) { 1650 1651 // If the Data use is an IF, that means we have an IF outside of the 1652 // loop that is switching on a condition that is set inside of the 1653 // loop. Happens if people set a loop-exit flag; then test the flag 1654 // in the loop to break the loop, then test is again outside of the 1655 // loop to determine which way the loop exited. 1656 // Loop predicate If node connects to Bool node through Opaque1 node. 1657 if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use)) { 1658 // Since this code is highly unlikely, we lazily build the worklist 1659 // of such Nodes to go split. 1660 if( !split_if_set ) 1661 split_if_set = new Node_List(area); 1662 split_if_set->push(use); 1663 } 1664 if( use->is_Bool() ) { 1665 if( !split_bool_set ) 1666 split_bool_set = new Node_List(area); 1667 split_bool_set->push(use); 1668 } 1669 if( use->Opcode() == Op_CreateEx ) { 1670 if( !split_cex_set ) 1671 split_cex_set = new Node_List(area); 1672 split_cex_set->push(use); 1673 } 1674 1675 1676 // Get "block" use is in 1677 uint idx = 0; 1678 while( use->in(idx) != old ) idx++; 1679 Node *prev = use->is_CFG() ? use : get_ctrl(use); 1680 assert( !loop->is_member( get_loop( prev ) ), "" ); 1681 Node *cfg = prev->_idx >= new_counter 1682 ? prev->in(2) 1683 : idom(prev); 1684 if( use->is_Phi() ) // Phi use is in prior block 1685 cfg = prev->in(idx); // NOT in block of Phi itself 1686 if (cfg->is_top()) { // Use is dead? 1687 _igvn.replace_input_of(use, idx, C->top()); 1688 continue; 1689 } 1690 1691 while( !loop->is_member( get_loop( cfg ) ) ) { 1692 prev = cfg; 1693 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg); 1694 } 1695 // If the use occurs after merging several exits from the loop, then 1696 // old value must have dominated all those exits. Since the same old 1697 // value was used on all those exits we did not need a Phi at this 1698 // merge point. NOW we do need a Phi here. Each loop exit value 1699 // is now merged with the peeled body exit; each exit gets its own 1700 // private Phi and those Phis need to be merged here. 1701 Node *phi; 1702 if( prev->is_Region() ) { 1703 if( idx == 0 ) { // Updating control edge? 1704 phi = prev; // Just use existing control 1705 } else { // Else need a new Phi 1706 phi = PhiNode::make( prev, old ); 1707 // Now recursively fix up the new uses of old! 1708 for( uint i = 1; i < prev->req(); i++ ) { 1709 worklist.push(phi); // Onto worklist once for each 'old' input 1710 } 1711 } 1712 } else { 1713 // Get new RegionNode merging old and new loop exits 1714 prev = old_new[prev->_idx]; 1715 assert( prev, "just made this in step 7" ); 1716 if( idx == 0 ) { // Updating control edge? 1717 phi = prev; // Just use existing control 1718 } else { // Else need a new Phi 1719 // Make a new Phi merging data values properly 1720 phi = PhiNode::make( prev, old ); 1721 phi->set_req( 1, nnn ); 1722 } 1723 } 1724 // If inserting a new Phi, check for prior hits 1725 if( idx != 0 ) { 1726 Node *hit = _igvn.hash_find_insert(phi); 1727 if( hit == NULL ) { 1728 _igvn.register_new_node_with_optimizer(phi); // Register new phi 1729 } else { // or 1730 // Remove the new phi from the graph and use the hit 1731 _igvn.remove_dead_node(phi); 1732 phi = hit; // Use existing phi 1733 } 1734 set_ctrl(phi, prev); 1735 } 1736 // Make 'use' use the Phi instead of the old loop body exit value 1737 _igvn.replace_input_of(use, idx, phi); 1738 if( use->_idx >= new_counter ) { // If updating new phis 1739 // Not needed for correctness, but prevents a weak assert 1740 // in AddPNode from tripping (when we end up with different 1741 // base & derived Phis that will become the same after 1742 // IGVN does CSE). 1743 Node *hit = _igvn.hash_find_insert(use); 1744 if( hit ) // Go ahead and re-hash for hits. 1745 _igvn.replace_node( use, hit ); 1746 } 1747 1748 // If 'use' was in the loop-exit block, it now needs to be sunk 1749 // below the post-loop merge point. 1750 sink_use( use, prev ); 1751 } 1752 } 1753 } 1754 1755 // Check for IFs that need splitting/cloning. Happens if an IF outside of 1756 // the loop uses a condition set in the loop. The original IF probably 1757 // takes control from one or more OLD Regions (which in turn get from NEW 1758 // Regions). In any case, there will be a set of Phis for each merge point 1759 // from the IF up to where the original BOOL def exists the loop. 1760 if( split_if_set ) { 1761 while( split_if_set->size() ) { 1762 Node *iff = split_if_set->pop(); 1763 if( iff->in(1)->is_Phi() ) { 1764 BoolNode *b = clone_iff( iff->in(1)->as_Phi(), loop ); 1765 _igvn.replace_input_of(iff, 1, b); 1766 } 1767 } 1768 } 1769 if( split_bool_set ) { 1770 while( split_bool_set->size() ) { 1771 Node *b = split_bool_set->pop(); 1772 Node *phi = b->in(1); 1773 assert( phi->is_Phi(), "" ); 1774 CmpNode *cmp = clone_bool( (PhiNode*)phi, loop ); 1775 _igvn.replace_input_of(b, 1, cmp); 1776 } 1777 } 1778 if( split_cex_set ) { 1779 while( split_cex_set->size() ) { 1780 Node *b = split_cex_set->pop(); 1781 assert( b->in(0)->is_Region(), "" ); 1782 assert( b->in(1)->is_Phi(), "" ); 1783 assert( b->in(0)->in(0) == b->in(1)->in(0), "" ); 1784 split_up( b, b->in(0), NULL ); 1785 } 1786 } 1787 1788 } 1789 1790 1791 //---------------------- stride_of_possible_iv ------------------------------------- 1792 // Looks for an iff/bool/comp with one operand of the compare 1793 // being a cycle involving an add and a phi, 1794 // with an optional truncation (left-shift followed by a right-shift) 1795 // of the add. Returns zero if not an iv. 1796 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) { 1797 Node* trunc1 = NULL; 1798 Node* trunc2 = NULL; 1799 const TypeInt* ttype = NULL; 1800 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) { 1801 return 0; 1802 } 1803 BoolNode* bl = iff->in(1)->as_Bool(); 1804 Node* cmp = bl->in(1); 1805 if (!cmp || cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU) { 1806 return 0; 1807 } 1808 // Must have an invariant operand 1809 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) { 1810 return 0; 1811 } 1812 Node* add2 = NULL; 1813 Node* cmp1 = cmp->in(1); 1814 if (cmp1->is_Phi()) { 1815 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) ))) 1816 Node* phi = cmp1; 1817 for (uint i = 1; i < phi->req(); i++) { 1818 Node* in = phi->in(i); 1819 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in, 1820 &trunc1, &trunc2, &ttype); 1821 if (add && add->in(1) == phi) { 1822 add2 = add->in(2); 1823 break; 1824 } 1825 } 1826 } else { 1827 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) ))) 1828 Node* addtrunc = cmp1; 1829 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc, 1830 &trunc1, &trunc2, &ttype); 1831 if (add && add->in(1)->is_Phi()) { 1832 Node* phi = add->in(1); 1833 for (uint i = 1; i < phi->req(); i++) { 1834 if (phi->in(i) == addtrunc) { 1835 add2 = add->in(2); 1836 break; 1837 } 1838 } 1839 } 1840 } 1841 if (add2 != NULL) { 1842 const TypeInt* add2t = _igvn.type(add2)->is_int(); 1843 if (add2t->is_con()) { 1844 return add2t->get_con(); 1845 } 1846 } 1847 return 0; 1848 } 1849 1850 1851 //---------------------- stay_in_loop ------------------------------------- 1852 // Return the (unique) control output node that's in the loop (if it exists.) 1853 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) { 1854 Node* unique = NULL; 1855 if (!n) return NULL; 1856 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1857 Node* use = n->fast_out(i); 1858 if (!has_ctrl(use) && loop->is_member(get_loop(use))) { 1859 if (unique != NULL) { 1860 return NULL; 1861 } 1862 unique = use; 1863 } 1864 } 1865 return unique; 1866 } 1867 1868 //------------------------------ register_node ------------------------------------- 1869 // Utility to register node "n" with PhaseIdealLoop 1870 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) { 1871 _igvn.register_new_node_with_optimizer(n); 1872 loop->_body.push(n); 1873 if (n->is_CFG()) { 1874 set_loop(n, loop); 1875 set_idom(n, pred, ddepth); 1876 } else { 1877 set_ctrl(n, pred); 1878 } 1879 } 1880 1881 //------------------------------ proj_clone ------------------------------------- 1882 // Utility to create an if-projection 1883 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) { 1884 ProjNode* c = p->clone()->as_Proj(); 1885 c->set_req(0, iff); 1886 return c; 1887 } 1888 1889 //------------------------------ short_circuit_if ------------------------------------- 1890 // Force the iff control output to be the live_proj 1891 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) { 1892 guarantee(live_proj != NULL, "null projection"); 1893 int proj_con = live_proj->_con; 1894 assert(proj_con == 0 || proj_con == 1, "false or true projection"); 1895 Node *con = _igvn.intcon(proj_con); 1896 set_ctrl(con, C->root()); 1897 if (iff) { 1898 iff->set_req(1, con); 1899 } 1900 return con; 1901 } 1902 1903 //------------------------------ insert_if_before_proj ------------------------------------- 1904 // Insert a new if before an if projection (* - new node) 1905 // 1906 // before 1907 // if(test) 1908 // / \ 1909 // v v 1910 // other-proj proj (arg) 1911 // 1912 // after 1913 // if(test) 1914 // / \ 1915 // / v 1916 // | * proj-clone 1917 // v | 1918 // other-proj v 1919 // * new_if(relop(cmp[IU](left,right))) 1920 // / \ 1921 // v v 1922 // * new-proj proj 1923 // (returned) 1924 // 1925 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) { 1926 IfNode* iff = proj->in(0)->as_If(); 1927 IdealLoopTree *loop = get_loop(proj); 1928 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 1929 int ddepth = dom_depth(proj); 1930 1931 _igvn.rehash_node_delayed(iff); 1932 _igvn.rehash_node_delayed(proj); 1933 1934 proj->set_req(0, NULL); // temporary disconnect 1935 ProjNode* proj2 = proj_clone(proj, iff); 1936 register_node(proj2, loop, iff, ddepth); 1937 1938 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right); 1939 register_node(cmp, loop, proj2, ddepth); 1940 1941 BoolNode* bol = new BoolNode(cmp, relop); 1942 register_node(bol, loop, proj2, ddepth); 1943 1944 IfNode* new_if = new IfNode(proj2, bol, iff->_prob, iff->_fcnt); 1945 register_node(new_if, loop, proj2, ddepth); 1946 1947 proj->set_req(0, new_if); // reattach 1948 set_idom(proj, new_if, ddepth); 1949 1950 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj(); 1951 guarantee(new_exit != NULL, "null exit node"); 1952 register_node(new_exit, get_loop(other_proj), new_if, ddepth); 1953 1954 return new_exit; 1955 } 1956 1957 //------------------------------ insert_region_before_proj ------------------------------------- 1958 // Insert a region before an if projection (* - new node) 1959 // 1960 // before 1961 // if(test) 1962 // / | 1963 // v | 1964 // proj v 1965 // other-proj 1966 // 1967 // after 1968 // if(test) 1969 // / | 1970 // v | 1971 // * proj-clone v 1972 // | other-proj 1973 // v 1974 // * new-region 1975 // | 1976 // v 1977 // * dum_if 1978 // / \ 1979 // v \ 1980 // * dum-proj v 1981 // proj 1982 // 1983 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) { 1984 IfNode* iff = proj->in(0)->as_If(); 1985 IdealLoopTree *loop = get_loop(proj); 1986 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 1987 int ddepth = dom_depth(proj); 1988 1989 _igvn.rehash_node_delayed(iff); 1990 _igvn.rehash_node_delayed(proj); 1991 1992 proj->set_req(0, NULL); // temporary disconnect 1993 ProjNode* proj2 = proj_clone(proj, iff); 1994 register_node(proj2, loop, iff, ddepth); 1995 1996 RegionNode* reg = new RegionNode(2); 1997 reg->set_req(1, proj2); 1998 register_node(reg, loop, iff, ddepth); 1999 2000 IfNode* dum_if = new IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt); 2001 register_node(dum_if, loop, reg, ddepth); 2002 2003 proj->set_req(0, dum_if); // reattach 2004 set_idom(proj, dum_if, ddepth); 2005 2006 ProjNode* dum_proj = proj_clone(other_proj, dum_if); 2007 register_node(dum_proj, loop, dum_if, ddepth); 2008 2009 return reg; 2010 } 2011 2012 //------------------------------ insert_cmpi_loop_exit ------------------------------------- 2013 // Clone a signed compare loop exit from an unsigned compare and 2014 // insert it before the unsigned cmp on the stay-in-loop path. 2015 // All new nodes inserted in the dominator tree between the original 2016 // if and it's projections. The original if test is replaced with 2017 // a constant to force the stay-in-loop path. 2018 // 2019 // This is done to make sure that the original if and it's projections 2020 // still dominate the same set of control nodes, that the ctrl() relation 2021 // from data nodes to them is preserved, and that their loop nesting is 2022 // preserved. 2023 // 2024 // before 2025 // if(i <u limit) unsigned compare loop exit 2026 // / | 2027 // v v 2028 // exit-proj stay-in-loop-proj 2029 // 2030 // after 2031 // if(stay-in-loop-const) original if 2032 // / | 2033 // / v 2034 // / if(i < limit) new signed test 2035 // / / | 2036 // / / v 2037 // / / if(i <u limit) new cloned unsigned test 2038 // / / / | 2039 // v v v | 2040 // region | 2041 // | | 2042 // dum-if | 2043 // / | | 2044 // ether | | 2045 // v v 2046 // exit-proj stay-in-loop-proj 2047 // 2048 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) { 2049 const bool Signed = true; 2050 const bool Unsigned = false; 2051 2052 BoolNode* bol = if_cmpu->in(1)->as_Bool(); 2053 if (bol->_test._test != BoolTest::lt) return NULL; 2054 CmpNode* cmpu = bol->in(1)->as_Cmp(); 2055 if (cmpu->Opcode() != Op_CmpU) return NULL; 2056 int stride = stride_of_possible_iv(if_cmpu); 2057 if (stride == 0) return NULL; 2058 2059 Node* lp_proj = stay_in_loop(if_cmpu, loop); 2060 guarantee(lp_proj != NULL, "null loop node"); 2061 2062 ProjNode* lp_continue = lp_proj->as_Proj(); 2063 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj(); 2064 2065 Node* limit = NULL; 2066 if (stride > 0) { 2067 limit = cmpu->in(2); 2068 } else { 2069 limit = _igvn.makecon(TypeInt::ZERO); 2070 set_ctrl(limit, C->root()); 2071 } 2072 // Create a new region on the exit path 2073 RegionNode* reg = insert_region_before_proj(lp_exit); 2074 guarantee(reg != NULL, "null region node"); 2075 2076 // Clone the if-cmpu-true-false using a signed compare 2077 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge; 2078 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue); 2079 reg->add_req(cmpi_exit); 2080 2081 // Clone the if-cmpu-true-false 2082 BoolTest::mask rel_u = bol->_test._test; 2083 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue); 2084 reg->add_req(cmpu_exit); 2085 2086 // Force original if to stay in loop. 2087 short_circuit_if(if_cmpu, lp_continue); 2088 2089 return cmpi_exit->in(0)->as_If(); 2090 } 2091 2092 //------------------------------ remove_cmpi_loop_exit ------------------------------------- 2093 // Remove a previously inserted signed compare loop exit. 2094 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) { 2095 Node* lp_proj = stay_in_loop(if_cmp, loop); 2096 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI && 2097 stay_in_loop(lp_proj, loop)->is_If() && 2098 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu"); 2099 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO); 2100 set_ctrl(con, C->root()); 2101 if_cmp->set_req(1, con); 2102 } 2103 2104 //------------------------------ scheduled_nodelist ------------------------------------- 2105 // Create a post order schedule of nodes that are in the 2106 // "member" set. The list is returned in "sched". 2107 // The first node in "sched" is the loop head, followed by 2108 // nodes which have no inputs in the "member" set, and then 2109 // followed by the nodes that have an immediate input dependence 2110 // on a node in "sched". 2111 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) { 2112 2113 assert(member.test(loop->_head->_idx), "loop head must be in member set"); 2114 Arena *a = Thread::current()->resource_area(); 2115 VectorSet visited(a); 2116 Node_Stack nstack(a, loop->_body.size()); 2117 2118 Node* n = loop->_head; // top of stack is cached in "n" 2119 uint idx = 0; 2120 visited.set(n->_idx); 2121 2122 // Initially push all with no inputs from within member set 2123 for(uint i = 0; i < loop->_body.size(); i++ ) { 2124 Node *elt = loop->_body.at(i); 2125 if (member.test(elt->_idx)) { 2126 bool found = false; 2127 for (uint j = 0; j < elt->req(); j++) { 2128 Node* def = elt->in(j); 2129 if (def && member.test(def->_idx) && def != elt) { 2130 found = true; 2131 break; 2132 } 2133 } 2134 if (!found && elt != loop->_head) { 2135 nstack.push(n, idx); 2136 n = elt; 2137 assert(!visited.test(n->_idx), "not seen yet"); 2138 visited.set(n->_idx); 2139 } 2140 } 2141 } 2142 2143 // traverse out's that are in the member set 2144 while (true) { 2145 if (idx < n->outcnt()) { 2146 Node* use = n->raw_out(idx); 2147 idx++; 2148 if (!visited.test_set(use->_idx)) { 2149 if (member.test(use->_idx)) { 2150 nstack.push(n, idx); 2151 n = use; 2152 idx = 0; 2153 } 2154 } 2155 } else { 2156 // All outputs processed 2157 sched.push(n); 2158 if (nstack.is_empty()) break; 2159 n = nstack.node(); 2160 idx = nstack.index(); 2161 nstack.pop(); 2162 } 2163 } 2164 } 2165 2166 2167 //------------------------------ has_use_in_set ------------------------------------- 2168 // Has a use in the vector set 2169 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) { 2170 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2171 Node* use = n->fast_out(j); 2172 if (vset.test(use->_idx)) { 2173 return true; 2174 } 2175 } 2176 return false; 2177 } 2178 2179 2180 //------------------------------ has_use_internal_to_set ------------------------------------- 2181 // Has use internal to the vector set (ie. not in a phi at the loop head) 2182 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) { 2183 Node* head = loop->_head; 2184 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2185 Node* use = n->fast_out(j); 2186 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) { 2187 return true; 2188 } 2189 } 2190 return false; 2191 } 2192 2193 2194 //------------------------------ clone_for_use_outside_loop ------------------------------------- 2195 // clone "n" for uses that are outside of loop 2196 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) { 2197 int cloned = 0; 2198 assert(worklist.size() == 0, "should be empty"); 2199 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2200 Node* use = n->fast_out(j); 2201 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) { 2202 worklist.push(use); 2203 } 2204 } 2205 while( worklist.size() ) { 2206 Node *use = worklist.pop(); 2207 if (!has_node(use) || use->in(0) == C->top()) continue; 2208 uint j; 2209 for (j = 0; j < use->req(); j++) { 2210 if (use->in(j) == n) break; 2211 } 2212 assert(j < use->req(), "must be there"); 2213 2214 // clone "n" and insert it between the inputs of "n" and the use outside the loop 2215 Node* n_clone = n->clone(); 2216 _igvn.replace_input_of(use, j, n_clone); 2217 cloned++; 2218 Node* use_c; 2219 if (!use->is_Phi()) { 2220 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0); 2221 } else { 2222 // Use in a phi is considered a use in the associated predecessor block 2223 use_c = use->in(0)->in(j); 2224 } 2225 set_ctrl(n_clone, use_c); 2226 assert(!loop->is_member(get_loop(use_c)), "should be outside loop"); 2227 get_loop(use_c)->_body.push(n_clone); 2228 _igvn.register_new_node_with_optimizer(n_clone); 2229 #if !defined(PRODUCT) 2230 if (TracePartialPeeling) { 2231 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx); 2232 } 2233 #endif 2234 } 2235 return cloned; 2236 } 2237 2238 2239 //------------------------------ clone_for_special_use_inside_loop ------------------------------------- 2240 // clone "n" for special uses that are in the not_peeled region. 2241 // If these def-uses occur in separate blocks, the code generator 2242 // marks the method as not compilable. For example, if a "BoolNode" 2243 // is in a different basic block than the "IfNode" that uses it, then 2244 // the compilation is aborted in the code generator. 2245 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 2246 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) { 2247 if (n->is_Phi() || n->is_Load()) { 2248 return; 2249 } 2250 assert(worklist.size() == 0, "should be empty"); 2251 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2252 Node* use = n->fast_out(j); 2253 if ( not_peel.test(use->_idx) && 2254 (use->is_If() || use->is_CMove() || use->is_Bool()) && 2255 use->in(1) == n) { 2256 worklist.push(use); 2257 } 2258 } 2259 if (worklist.size() > 0) { 2260 // clone "n" and insert it between inputs of "n" and the use 2261 Node* n_clone = n->clone(); 2262 loop->_body.push(n_clone); 2263 _igvn.register_new_node_with_optimizer(n_clone); 2264 set_ctrl(n_clone, get_ctrl(n)); 2265 sink_list.push(n_clone); 2266 not_peel <<= n_clone->_idx; // add n_clone to not_peel set. 2267 #if !defined(PRODUCT) 2268 if (TracePartialPeeling) { 2269 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx); 2270 } 2271 #endif 2272 while( worklist.size() ) { 2273 Node *use = worklist.pop(); 2274 _igvn.rehash_node_delayed(use); 2275 for (uint j = 1; j < use->req(); j++) { 2276 if (use->in(j) == n) { 2277 use->set_req(j, n_clone); 2278 } 2279 } 2280 } 2281 } 2282 } 2283 2284 2285 //------------------------------ insert_phi_for_loop ------------------------------------- 2286 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 2287 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) { 2288 Node *phi = PhiNode::make(lp, back_edge_val); 2289 phi->set_req(LoopNode::EntryControl, lp_entry_val); 2290 // Use existing phi if it already exists 2291 Node *hit = _igvn.hash_find_insert(phi); 2292 if( hit == NULL ) { 2293 _igvn.register_new_node_with_optimizer(phi); 2294 set_ctrl(phi, lp); 2295 } else { 2296 // Remove the new phi from the graph and use the hit 2297 _igvn.remove_dead_node(phi); 2298 phi = hit; 2299 } 2300 _igvn.replace_input_of(use, idx, phi); 2301 } 2302 2303 #ifdef ASSERT 2304 //------------------------------ is_valid_loop_partition ------------------------------------- 2305 // Validate the loop partition sets: peel and not_peel 2306 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, 2307 VectorSet& not_peel ) { 2308 uint i; 2309 // Check that peel_list entries are in the peel set 2310 for (i = 0; i < peel_list.size(); i++) { 2311 if (!peel.test(peel_list.at(i)->_idx)) { 2312 return false; 2313 } 2314 } 2315 // Check at loop members are in one of peel set or not_peel set 2316 for (i = 0; i < loop->_body.size(); i++ ) { 2317 Node *def = loop->_body.at(i); 2318 uint di = def->_idx; 2319 // Check that peel set elements are in peel_list 2320 if (peel.test(di)) { 2321 if (not_peel.test(di)) { 2322 return false; 2323 } 2324 // Must be in peel_list also 2325 bool found = false; 2326 for (uint j = 0; j < peel_list.size(); j++) { 2327 if (peel_list.at(j)->_idx == di) { 2328 found = true; 2329 break; 2330 } 2331 } 2332 if (!found) { 2333 return false; 2334 } 2335 } else if (not_peel.test(di)) { 2336 if (peel.test(di)) { 2337 return false; 2338 } 2339 } else { 2340 return false; 2341 } 2342 } 2343 return true; 2344 } 2345 2346 //------------------------------ is_valid_clone_loop_exit_use ------------------------------------- 2347 // Ensure a use outside of loop is of the right form 2348 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) { 2349 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2350 return (use->is_Phi() && 2351 use_c->is_Region() && use_c->req() == 3 && 2352 (use_c->in(exit_idx)->Opcode() == Op_IfTrue || 2353 use_c->in(exit_idx)->Opcode() == Op_IfFalse || 2354 use_c->in(exit_idx)->Opcode() == Op_JumpProj) && 2355 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) ); 2356 } 2357 2358 //------------------------------ is_valid_clone_loop_form ------------------------------------- 2359 // Ensure that all uses outside of loop are of the right form 2360 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 2361 uint orig_exit_idx, uint clone_exit_idx) { 2362 uint len = peel_list.size(); 2363 for (uint i = 0; i < len; i++) { 2364 Node *def = peel_list.at(i); 2365 2366 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 2367 Node *use = def->fast_out(j); 2368 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2369 if (!loop->is_member(get_loop(use_c))) { 2370 // use is not in the loop, check for correct structure 2371 if (use->in(0) == def) { 2372 // Okay 2373 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) { 2374 return false; 2375 } 2376 } 2377 } 2378 } 2379 return true; 2380 } 2381 #endif 2382 2383 //------------------------------ partial_peel ------------------------------------- 2384 // Partially peel (aka loop rotation) the top portion of a loop (called 2385 // the peel section below) by cloning it and placing one copy just before 2386 // the new loop head and the other copy at the bottom of the new loop. 2387 // 2388 // before after where it came from 2389 // 2390 // stmt1 stmt1 2391 // loop: stmt2 clone 2392 // stmt2 if condA goto exitA clone 2393 // if condA goto exitA new_loop: new 2394 // stmt3 stmt3 clone 2395 // if !condB goto loop if condB goto exitB clone 2396 // exitB: stmt2 orig 2397 // stmt4 if !condA goto new_loop orig 2398 // exitA: goto exitA 2399 // exitB: 2400 // stmt4 2401 // exitA: 2402 // 2403 // Step 1: find the cut point: an exit test on probable 2404 // induction variable. 2405 // Step 2: schedule (with cloning) operations in the peel 2406 // section that can be executed after the cut into 2407 // the section that is not peeled. This may need 2408 // to clone operations into exit blocks. For 2409 // instance, a reference to A[i] in the not-peel 2410 // section and a reference to B[i] in an exit block 2411 // may cause a left-shift of i by 2 to be placed 2412 // in the peel block. This step will clone the left 2413 // shift into the exit block and sink the left shift 2414 // from the peel to the not-peel section. 2415 // Step 3: clone the loop, retarget the control, and insert 2416 // phis for values that are live across the new loop 2417 // head. This is very dependent on the graph structure 2418 // from clone_loop. It creates region nodes for 2419 // exit control and associated phi nodes for values 2420 // flow out of the loop through that exit. The region 2421 // node is dominated by the clone's control projection. 2422 // So the clone's peel section is placed before the 2423 // new loop head, and the clone's not-peel section is 2424 // forms the top part of the new loop. The original 2425 // peel section forms the tail of the new loop. 2426 // Step 4: update the dominator tree and recompute the 2427 // dominator depth. 2428 // 2429 // orig 2430 // 2431 // stmt1 2432 // | 2433 // v 2434 // loop predicate 2435 // | 2436 // v 2437 // loop<----+ 2438 // | | 2439 // stmt2 | 2440 // | | 2441 // v | 2442 // ifA | 2443 // / | | 2444 // v v | 2445 // false true ^ <-- last_peel 2446 // / | | 2447 // / ===|==cut | 2448 // / stmt3 | <-- first_not_peel 2449 // / | | 2450 // | v | 2451 // v ifB | 2452 // exitA: / \ | 2453 // / \ | 2454 // v v | 2455 // false true | 2456 // / \ | 2457 // / ----+ 2458 // | 2459 // v 2460 // exitB: 2461 // stmt4 2462 // 2463 // 2464 // after clone loop 2465 // 2466 // stmt1 2467 // | 2468 // v 2469 // loop predicate 2470 // / \ 2471 // clone / \ orig 2472 // / \ 2473 // / \ 2474 // v v 2475 // +---->loop loop<----+ 2476 // | | | | 2477 // | stmt2 stmt2 | 2478 // | | | | 2479 // | v v | 2480 // | ifA ifA | 2481 // | | \ / | | 2482 // | v v v v | 2483 // ^ true false false true ^ <-- last_peel 2484 // | | ^ \ / | | 2485 // | cut==|== \ \ / ===|==cut | 2486 // | stmt3 \ \ / stmt3 | <-- first_not_peel 2487 // | | dom | | | | 2488 // | v \ 1v v2 v | 2489 // | ifB regionA ifB | 2490 // | / \ | / \ | 2491 // | / \ v / \ | 2492 // | v v exitA: v v | 2493 // | true false false true | 2494 // | / ^ \ / \ | 2495 // +---- \ \ / ----+ 2496 // dom \ / 2497 // \ 1v v2 2498 // regionB 2499 // | 2500 // v 2501 // exitB: 2502 // stmt4 2503 // 2504 // 2505 // after partial peel 2506 // 2507 // stmt1 2508 // | 2509 // v 2510 // loop predicate 2511 // / 2512 // clone / orig 2513 // / TOP 2514 // / \ 2515 // v v 2516 // TOP->loop loop----+ 2517 // | | | 2518 // stmt2 stmt2 | 2519 // | | | 2520 // v v | 2521 // ifA ifA | 2522 // | \ / | | 2523 // v v v v | 2524 // true false false true | <-- last_peel 2525 // | ^ \ / +------|---+ 2526 // +->newloop \ \ / === ==cut | | 2527 // | stmt3 \ \ / TOP | | 2528 // | | dom | | stmt3 | | <-- first_not_peel 2529 // | v \ 1v v2 v | | 2530 // | ifB regionA ifB ^ v 2531 // | / \ | / \ | | 2532 // | / \ v / \ | | 2533 // | v v exitA: v v | | 2534 // | true false false true | | 2535 // | / ^ \ / \ | | 2536 // | | \ \ / v | | 2537 // | | dom \ / TOP | | 2538 // | | \ 1v v2 | | 2539 // ^ v regionB | | 2540 // | | | | | 2541 // | | v ^ v 2542 // | | exitB: | | 2543 // | | stmt4 | | 2544 // | +------------>-----------------+ | 2545 // | | 2546 // +-----------------<---------------------+ 2547 // 2548 // 2549 // final graph 2550 // 2551 // stmt1 2552 // | 2553 // v 2554 // loop predicate 2555 // | 2556 // v 2557 // stmt2 clone 2558 // | 2559 // v 2560 // ........> ifA clone 2561 // : / | 2562 // dom / | 2563 // : v v 2564 // : false true 2565 // : | | 2566 // : | v 2567 // : | newloop<-----+ 2568 // : | | | 2569 // : | stmt3 clone | 2570 // : | | | 2571 // : | v | 2572 // : | ifB | 2573 // : | / \ | 2574 // : | v v | 2575 // : | false true | 2576 // : | | | | 2577 // : | v stmt2 | 2578 // : | exitB: | | 2579 // : | stmt4 v | 2580 // : | ifA orig | 2581 // : | / \ | 2582 // : | / \ | 2583 // : | v v | 2584 // : | false true | 2585 // : | / \ | 2586 // : v v -----+ 2587 // RegionA 2588 // | 2589 // v 2590 // exitA 2591 // 2592 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) { 2593 2594 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only"); 2595 if (!loop->_head->is_Loop()) { 2596 return false; } 2597 2598 LoopNode *head = loop->_head->as_Loop(); 2599 2600 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) { 2601 return false; 2602 } 2603 2604 // Check for complex exit control 2605 for(uint ii = 0; ii < loop->_body.size(); ii++ ) { 2606 Node *n = loop->_body.at(ii); 2607 int opc = n->Opcode(); 2608 if (n->is_Call() || 2609 opc == Op_Catch || 2610 opc == Op_CatchProj || 2611 opc == Op_Jump || 2612 opc == Op_JumpProj) { 2613 #if !defined(PRODUCT) 2614 if (TracePartialPeeling) { 2615 tty->print_cr("\nExit control too complex: lp: %d", head->_idx); 2616 } 2617 #endif 2618 return false; 2619 } 2620 } 2621 2622 int dd = dom_depth(head); 2623 2624 // Step 1: find cut point 2625 2626 // Walk up dominators to loop head looking for first loop exit 2627 // which is executed on every path thru loop. 2628 IfNode *peel_if = NULL; 2629 IfNode *peel_if_cmpu = NULL; 2630 2631 Node *iff = loop->tail(); 2632 while( iff != head ) { 2633 if( iff->is_If() ) { 2634 Node *ctrl = get_ctrl(iff->in(1)); 2635 if (ctrl->is_top()) return false; // Dead test on live IF. 2636 // If loop-varying exit-test, check for induction variable 2637 if( loop->is_member(get_loop(ctrl)) && 2638 loop->is_loop_exit(iff) && 2639 is_possible_iv_test(iff)) { 2640 Node* cmp = iff->in(1)->in(1); 2641 if (cmp->Opcode() == Op_CmpI) { 2642 peel_if = iff->as_If(); 2643 } else { 2644 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU"); 2645 peel_if_cmpu = iff->as_If(); 2646 } 2647 } 2648 } 2649 iff = idom(iff); 2650 } 2651 // Prefer signed compare over unsigned compare. 2652 IfNode* new_peel_if = NULL; 2653 if (peel_if == NULL) { 2654 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) { 2655 return false; // No peel point found 2656 } 2657 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop); 2658 if (new_peel_if == NULL) { 2659 return false; // No peel point found 2660 } 2661 peel_if = new_peel_if; 2662 } 2663 Node* last_peel = stay_in_loop(peel_if, loop); 2664 Node* first_not_peeled = stay_in_loop(last_peel, loop); 2665 if (first_not_peeled == NULL || first_not_peeled == head) { 2666 return false; 2667 } 2668 2669 #if !defined(PRODUCT) 2670 if (TraceLoopOpts) { 2671 tty->print("PartialPeel "); 2672 loop->dump_head(); 2673 } 2674 2675 if (TracePartialPeeling) { 2676 tty->print_cr("before partial peel one iteration"); 2677 Node_List wl; 2678 Node* t = head->in(2); 2679 while (true) { 2680 wl.push(t); 2681 if (t == head) break; 2682 t = idom(t); 2683 } 2684 while (wl.size() > 0) { 2685 Node* tt = wl.pop(); 2686 tt->dump(); 2687 if (tt == last_peel) tty->print_cr("-- cut --"); 2688 } 2689 } 2690 #endif 2691 ResourceArea *area = Thread::current()->resource_area(); 2692 VectorSet peel(area); 2693 VectorSet not_peel(area); 2694 Node_List peel_list(area); 2695 Node_List worklist(area); 2696 Node_List sink_list(area); 2697 2698 // Set of cfg nodes to peel are those that are executable from 2699 // the head through last_peel. 2700 assert(worklist.size() == 0, "should be empty"); 2701 worklist.push(head); 2702 peel.set(head->_idx); 2703 while (worklist.size() > 0) { 2704 Node *n = worklist.pop(); 2705 if (n != last_peel) { 2706 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2707 Node* use = n->fast_out(j); 2708 if (use->is_CFG() && 2709 loop->is_member(get_loop(use)) && 2710 !peel.test_set(use->_idx)) { 2711 worklist.push(use); 2712 } 2713 } 2714 } 2715 } 2716 2717 // Set of non-cfg nodes to peel are those that are control 2718 // dependent on the cfg nodes. 2719 uint i; 2720 for(i = 0; i < loop->_body.size(); i++ ) { 2721 Node *n = loop->_body.at(i); 2722 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n; 2723 if (peel.test(n_c->_idx)) { 2724 peel.set(n->_idx); 2725 } else { 2726 not_peel.set(n->_idx); 2727 } 2728 } 2729 2730 // Step 2: move operations from the peeled section down into the 2731 // not-peeled section 2732 2733 // Get a post order schedule of nodes in the peel region 2734 // Result in right-most operand. 2735 scheduled_nodelist(loop, peel, peel_list ); 2736 2737 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 2738 2739 // For future check for too many new phis 2740 uint old_phi_cnt = 0; 2741 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 2742 Node* use = head->fast_out(j); 2743 if (use->is_Phi()) old_phi_cnt++; 2744 } 2745 2746 #if !defined(PRODUCT) 2747 if (TracePartialPeeling) { 2748 tty->print_cr("\npeeled list"); 2749 } 2750 #endif 2751 2752 // Evacuate nodes in peel region into the not_peeled region if possible 2753 uint new_phi_cnt = 0; 2754 uint cloned_for_outside_use = 0; 2755 for (i = 0; i < peel_list.size();) { 2756 Node* n = peel_list.at(i); 2757 #if !defined(PRODUCT) 2758 if (TracePartialPeeling) n->dump(); 2759 #endif 2760 bool incr = true; 2761 if ( !n->is_CFG() ) { 2762 2763 if ( has_use_in_set(n, not_peel) ) { 2764 2765 // If not used internal to the peeled region, 2766 // move "n" from peeled to not_peeled region. 2767 2768 if ( !has_use_internal_to_set(n, peel, loop) ) { 2769 2770 // if not pinned and not a load (which maybe anti-dependent on a store) 2771 // and not a CMove (Matcher expects only bool->cmove). 2772 if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) { 2773 cloned_for_outside_use += clone_for_use_outside_loop( loop, n, worklist ); 2774 sink_list.push(n); 2775 peel >>= n->_idx; // delete n from peel set. 2776 not_peel <<= n->_idx; // add n to not_peel set. 2777 peel_list.remove(i); 2778 incr = false; 2779 #if !defined(PRODUCT) 2780 if (TracePartialPeeling) { 2781 tty->print_cr("sink to not_peeled region: %d newbb: %d", 2782 n->_idx, get_ctrl(n)->_idx); 2783 } 2784 #endif 2785 } 2786 } else { 2787 // Otherwise check for special def-use cases that span 2788 // the peel/not_peel boundary such as bool->if 2789 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist ); 2790 new_phi_cnt++; 2791 } 2792 } 2793 } 2794 if (incr) i++; 2795 } 2796 2797 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) { 2798 #if !defined(PRODUCT) 2799 if (TracePartialPeeling) { 2800 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c", 2801 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F'); 2802 } 2803 #endif 2804 if (new_peel_if != NULL) { 2805 remove_cmpi_loop_exit(new_peel_if, loop); 2806 } 2807 // Inhibit more partial peeling on this loop 2808 assert(!head->is_partial_peel_loop(), "not partial peeled"); 2809 head->mark_partial_peel_failed(); 2810 if (cloned_for_outside_use > 0) { 2811 // Terminate this round of loop opts because 2812 // the graph outside this loop was changed. 2813 C->set_major_progress(); 2814 return true; 2815 } 2816 return false; 2817 } 2818 2819 // Step 3: clone loop, retarget control, and insert new phis 2820 2821 // Create new loop head for new phis and to hang 2822 // the nodes being moved (sinked) from the peel region. 2823 LoopNode* new_head = new LoopNode(last_peel, last_peel); 2824 new_head->set_unswitch_count(head->unswitch_count()); // Preserve 2825 _igvn.register_new_node_with_optimizer(new_head); 2826 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled"); 2827 _igvn.replace_input_of(first_not_peeled, 0, new_head); 2828 set_loop(new_head, loop); 2829 loop->_body.push(new_head); 2830 not_peel.set(new_head->_idx); 2831 set_idom(new_head, last_peel, dom_depth(first_not_peeled)); 2832 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled)); 2833 2834 while (sink_list.size() > 0) { 2835 Node* n = sink_list.pop(); 2836 set_ctrl(n, new_head); 2837 } 2838 2839 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 2840 2841 clone_loop( loop, old_new, dd ); 2842 2843 const uint clone_exit_idx = 1; 2844 const uint orig_exit_idx = 2; 2845 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop"); 2846 2847 Node* head_clone = old_new[head->_idx]; 2848 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop(); 2849 Node* orig_tail_clone = head_clone->in(2); 2850 2851 // Add phi if "def" node is in peel set and "use" is not 2852 2853 for(i = 0; i < peel_list.size(); i++ ) { 2854 Node *def = peel_list.at(i); 2855 if (!def->is_CFG()) { 2856 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 2857 Node *use = def->fast_out(j); 2858 if (has_node(use) && use->in(0) != C->top() && 2859 (!peel.test(use->_idx) || 2860 (use->is_Phi() && use->in(0) == head)) ) { 2861 worklist.push(use); 2862 } 2863 } 2864 while( worklist.size() ) { 2865 Node *use = worklist.pop(); 2866 for (uint j = 1; j < use->req(); j++) { 2867 Node* n = use->in(j); 2868 if (n == def) { 2869 2870 // "def" is in peel set, "use" is not in peel set 2871 // or "use" is in the entry boundary (a phi) of the peel set 2872 2873 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use; 2874 2875 if ( loop->is_member(get_loop( use_c )) ) { 2876 // use is in loop 2877 if (old_new[use->_idx] != NULL) { // null for dead code 2878 Node* use_clone = old_new[use->_idx]; 2879 _igvn.replace_input_of(use, j, C->top()); 2880 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone ); 2881 } 2882 } else { 2883 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format"); 2884 // use is not in the loop, check if the live range includes the cut 2885 Node* lp_if = use_c->in(orig_exit_idx)->in(0); 2886 if (not_peel.test(lp_if->_idx)) { 2887 assert(j == orig_exit_idx, "use from original loop"); 2888 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone ); 2889 } 2890 } 2891 } 2892 } 2893 } 2894 } 2895 } 2896 2897 // Step 3b: retarget control 2898 2899 // Redirect control to the new loop head if a cloned node in 2900 // the not_peeled region has control that points into the peeled region. 2901 // This necessary because the cloned peeled region will be outside 2902 // the loop. 2903 // from to 2904 // cloned-peeled <---+ 2905 // new_head_clone: | <--+ 2906 // cloned-not_peeled in(0) in(0) 2907 // orig-peeled 2908 2909 for(i = 0; i < loop->_body.size(); i++ ) { 2910 Node *n = loop->_body.at(i); 2911 if (!n->is_CFG() && n->in(0) != NULL && 2912 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) { 2913 Node* n_clone = old_new[n->_idx]; 2914 _igvn.replace_input_of(n_clone, 0, new_head_clone); 2915 } 2916 } 2917 2918 // Backedge of the surviving new_head (the clone) is original last_peel 2919 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel); 2920 2921 // Cut first node in original not_peel set 2922 _igvn.rehash_node_delayed(new_head); // Multiple edge updates: 2923 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of 2924 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls 2925 2926 // Copy head_clone back-branch info to original head 2927 // and remove original head's loop entry and 2928 // clone head's back-branch 2929 _igvn.rehash_node_delayed(head); // Multiple edge updates 2930 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl)); 2931 head->set_req(LoopNode::LoopBackControl, C->top()); 2932 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top()); 2933 2934 // Similarly modify the phis 2935 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) { 2936 Node* use = head->fast_out(k); 2937 if (use->is_Phi() && use->outcnt() > 0) { 2938 Node* use_clone = old_new[use->_idx]; 2939 _igvn.rehash_node_delayed(use); // Multiple edge updates 2940 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl)); 2941 use->set_req(LoopNode::LoopBackControl, C->top()); 2942 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top()); 2943 } 2944 } 2945 2946 // Step 4: update dominator tree and dominator depth 2947 2948 set_idom(head, orig_tail_clone, dd); 2949 recompute_dom_depth(); 2950 2951 // Inhibit more partial peeling on this loop 2952 new_head_clone->set_partial_peel_loop(); 2953 C->set_major_progress(); 2954 loop->record_for_igvn(); 2955 2956 #if !defined(PRODUCT) 2957 if (TracePartialPeeling) { 2958 tty->print_cr("\nafter partial peel one iteration"); 2959 Node_List wl(area); 2960 Node* t = last_peel; 2961 while (true) { 2962 wl.push(t); 2963 if (t == head_clone) break; 2964 t = idom(t); 2965 } 2966 while (wl.size() > 0) { 2967 Node* tt = wl.pop(); 2968 if (tt == head) tty->print_cr("orig head"); 2969 else if (tt == new_head_clone) tty->print_cr("new head"); 2970 else if (tt == head_clone) tty->print_cr("clone head"); 2971 tt->dump(); 2972 } 2973 } 2974 #endif 2975 return true; 2976 } 2977 2978 //------------------------------reorg_offsets---------------------------------- 2979 // Reorganize offset computations to lower register pressure. Mostly 2980 // prevent loop-fallout uses of the pre-incremented trip counter (which are 2981 // then alive with the post-incremented trip counter forcing an extra 2982 // register move) 2983 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) { 2984 // Perform it only for canonical counted loops. 2985 // Loop's shape could be messed up by iteration_split_impl. 2986 if (!loop->_head->is_CountedLoop()) 2987 return; 2988 if (!loop->_head->as_Loop()->is_valid_counted_loop()) 2989 return; 2990 2991 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 2992 CountedLoopEndNode *cle = cl->loopexit(); 2993 Node *exit = cle->proj_out(false); 2994 Node *phi = cl->phi(); 2995 2996 // Check for the special case of folks using the pre-incremented 2997 // trip-counter on the fall-out path (forces the pre-incremented 2998 // and post-incremented trip counter to be live at the same time). 2999 // Fix this by adjusting to use the post-increment trip counter. 3000 3001 bool progress = true; 3002 while (progress) { 3003 progress = false; 3004 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) { 3005 Node* use = phi->fast_out(i); // User of trip-counter 3006 if (!has_ctrl(use)) continue; 3007 Node *u_ctrl = get_ctrl(use); 3008 if (use->is_Phi()) { 3009 u_ctrl = NULL; 3010 for (uint j = 1; j < use->req(); j++) 3011 if (use->in(j) == phi) 3012 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j)); 3013 } 3014 IdealLoopTree *u_loop = get_loop(u_ctrl); 3015 // Look for loop-invariant use 3016 if (u_loop == loop) continue; 3017 if (loop->is_member(u_loop)) continue; 3018 // Check that use is live out the bottom. Assuming the trip-counter 3019 // update is right at the bottom, uses of of the loop middle are ok. 3020 if (dom_lca(exit, u_ctrl) != exit) continue; 3021 // Hit! Refactor use to use the post-incremented tripcounter. 3022 // Compute a post-increment tripcounter. 3023 Node *opaq = new Opaque2Node( C, cle->incr() ); 3024 register_new_node(opaq, exit); 3025 Node *neg_stride = _igvn.intcon(-cle->stride_con()); 3026 set_ctrl(neg_stride, C->root()); 3027 Node *post = new AddINode( opaq, neg_stride); 3028 register_new_node(post, exit); 3029 _igvn.rehash_node_delayed(use); 3030 for (uint j = 1; j < use->req(); j++) { 3031 if (use->in(j) == phi) 3032 use->set_req(j, post); 3033 } 3034 // Since DU info changed, rerun loop 3035 progress = true; 3036 break; 3037 } 3038 } 3039 3040 }