loopopts.cpp revision 196:d1605aabd0a1
1/* 2 * Copyright 1999-2008 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, 20 * CA 95054 USA or visit www.sun.com if you need additional information or 21 * have any questions. 22 * 23 */ 24 25#include "incls/_precompiled.incl" 26#include "incls/_loopopts.cpp.incl" 27 28//============================================================================= 29//------------------------------split_thru_phi--------------------------------- 30// Split Node 'n' through merge point if there is enough win. 31Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) { 32 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) { 33 // ConvI2L may have type information on it which is unsafe to push up 34 // so disable this for now 35 return NULL; 36 } 37 int wins = 0; 38 assert( !n->is_CFG(), "" ); 39 assert( region->is_Region(), "" ); 40 41 const Type* type = n->bottom_type(); 42 const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr(); 43 Node *phi; 44 if( t_oop != NULL && t_oop->is_instance_field() ) { 45 int iid = t_oop->instance_id(); 46 int index = C->get_alias_index(t_oop); 47 int offset = t_oop->offset(); 48 phi = new (C,region->req()) PhiNode(region, type, NULL, iid, index, offset); 49 } else { 50 phi = new (C,region->req()) PhiNode(region, type); 51 } 52 uint old_unique = C->unique(); 53 for( uint i = 1; i < region->req(); i++ ) { 54 Node *x; 55 Node* the_clone = NULL; 56 if( region->in(i) == C->top() ) { 57 x = C->top(); // Dead path? Use a dead data op 58 } else { 59 x = n->clone(); // Else clone up the data op 60 the_clone = x; // Remember for possible deletion. 61 // Alter data node to use pre-phi inputs 62 if( n->in(0) == region ) 63 x->set_req( 0, region->in(i) ); 64 for( uint j = 1; j < n->req(); j++ ) { 65 Node *in = n->in(j); 66 if( in->is_Phi() && in->in(0) == region ) 67 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone 68 } 69 } 70 // Check for a 'win' on some paths 71 const Type *t = x->Value(&_igvn); 72 73 bool singleton = t->singleton(); 74 75 // A TOP singleton indicates that there are no possible values incoming 76 // along a particular edge. In most cases, this is OK, and the Phi will 77 // be eliminated later in an Ideal call. However, we can't allow this to 78 // happen if the singleton occurs on loop entry, as the elimination of 79 // the PhiNode may cause the resulting node to migrate back to a previous 80 // loop iteration. 81 if( singleton && t == Type::TOP ) { 82 // Is_Loop() == false does not confirm the absence of a loop (e.g., an 83 // irreducible loop may not be indicated by an affirmative is_Loop()); 84 // therefore, the only top we can split thru a phi is on a backedge of 85 // a loop. 86 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 87 } 88 89 if( singleton ) { 90 wins++; 91 x = ((PhaseGVN&)_igvn).makecon(t); 92 } else { 93 // We now call Identity to try to simplify the cloned node. 94 // Note that some Identity methods call phase->type(this). 95 // Make sure that the type array is big enough for 96 // our new node, even though we may throw the node away. 97 // (Note: This tweaking with igvn only works because x is a new node.) 98 _igvn.set_type(x, t); 99 Node *y = x->Identity(&_igvn); 100 if( y != x ) { 101 wins++; 102 x = y; 103 } else { 104 y = _igvn.hash_find(x); 105 if( y ) { 106 wins++; 107 x = y; 108 } else { 109 // Else x is a new node we are keeping 110 // We do not need register_new_node_with_optimizer 111 // because set_type has already been called. 112 _igvn._worklist.push(x); 113 } 114 } 115 } 116 if (x != the_clone && the_clone != NULL) 117 _igvn.remove_dead_node(the_clone); 118 phi->set_req( i, x ); 119 } 120 // Too few wins? 121 if( wins <= policy ) { 122 _igvn.remove_dead_node(phi); 123 return NULL; 124 } 125 126 // Record Phi 127 register_new_node( phi, region ); 128 129 for( uint i2 = 1; i2 < phi->req(); i2++ ) { 130 Node *x = phi->in(i2); 131 // If we commoned up the cloned 'x' with another existing Node, 132 // the existing Node picks up a new use. We need to make the 133 // existing Node occur higher up so it dominates its uses. 134 Node *old_ctrl; 135 IdealLoopTree *old_loop; 136 137 // The occasional new node 138 if( x->_idx >= old_unique ) { // Found a new, unplaced node? 139 old_ctrl = x->is_Con() ? C->root() : NULL; 140 old_loop = NULL; // Not in any prior loop 141 } else { 142 old_ctrl = x->is_Con() ? C->root() : get_ctrl(x); 143 old_loop = get_loop(old_ctrl); // Get prior loop 144 } 145 // New late point must dominate new use 146 Node *new_ctrl = dom_lca( old_ctrl, region->in(i2) ); 147 // Set new location 148 set_ctrl(x, new_ctrl); 149 IdealLoopTree *new_loop = get_loop( new_ctrl ); 150 // If changing loop bodies, see if we need to collect into new body 151 if( old_loop != new_loop ) { 152 if( old_loop && !old_loop->_child ) 153 old_loop->_body.yank(x); 154 if( !new_loop->_child ) 155 new_loop->_body.push(x); // Collect body info 156 } 157 } 158 159 return phi; 160} 161 162//------------------------------dominated_by------------------------------------ 163// Replace the dominated test with an obvious true or false. Place it on the 164// IGVN worklist for later cleanup. Move control-dependent data Nodes on the 165// live path up to the dominating control. 166void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff ) { 167#ifndef PRODUCT 168 if( VerifyLoopOptimizations && PrintOpto ) tty->print_cr("dominating test"); 169#endif 170 171 172 // prevdom is the dominating projection of the dominating test. 173 assert( iff->is_If(), "" ); 174 assert( iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added"); 175 int pop = prevdom->Opcode(); 176 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" ); 177 // 'con' is set to true or false to kill the dominated test. 178 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO); 179 set_ctrl(con, C->root()); // Constant gets a new use 180 // Hack the dominated test 181 _igvn.hash_delete(iff); 182 iff->set_req(1, con); 183 _igvn._worklist.push(iff); 184 185 // If I dont have a reachable TRUE and FALSE path following the IfNode then 186 // I can assume this path reaches an infinite loop. In this case it's not 187 // important to optimize the data Nodes - either the whole compilation will 188 // be tossed or this path (and all data Nodes) will go dead. 189 if( iff->outcnt() != 2 ) return; 190 191 // Make control-dependent data Nodes on the live path (path that will remain 192 // once the dominated IF is removed) become control-dependent on the 193 // dominating projection. 194 Node* dp = ((IfNode*)iff)->proj_out(pop == Op_IfTrue); 195 IdealLoopTree *old_loop = get_loop(dp); 196 197 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { 198 Node* cd = dp->fast_out(i); // Control-dependent node 199 if( cd->depends_only_on_test() ) { 200 assert( cd->in(0) == dp, "" ); 201 _igvn.hash_delete( cd ); 202 cd->set_req(0, prevdom); 203 set_early_ctrl( cd ); 204 _igvn._worklist.push(cd); 205 IdealLoopTree *new_loop = get_loop(get_ctrl(cd)); 206 if( old_loop != new_loop ) { 207 if( !old_loop->_child ) old_loop->_body.yank(cd); 208 if( !new_loop->_child ) new_loop->_body.push(cd); 209 } 210 --i; 211 --imax; 212 } 213 } 214} 215 216//------------------------------has_local_phi_input---------------------------- 217// Return TRUE if 'n' has Phi inputs from its local block and no other 218// block-local inputs (all non-local-phi inputs come from earlier blocks) 219Node *PhaseIdealLoop::has_local_phi_input( Node *n ) { 220 Node *n_ctrl = get_ctrl(n); 221 // See if some inputs come from a Phi in this block, or from before 222 // this block. 223 uint i; 224 for( i = 1; i < n->req(); i++ ) { 225 Node *phi = n->in(i); 226 if( phi->is_Phi() && phi->in(0) == n_ctrl ) 227 break; 228 } 229 if( i >= n->req() ) 230 return NULL; // No Phi inputs; nowhere to clone thru 231 232 // Check for inputs created between 'n' and the Phi input. These 233 // must split as well; they have already been given the chance 234 // (courtesy of a post-order visit) and since they did not we must 235 // recover the 'cost' of splitting them by being very profitable 236 // when splitting 'n'. Since this is unlikely we simply give up. 237 for( i = 1; i < n->req(); i++ ) { 238 Node *m = n->in(i); 239 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) { 240 // We allow the special case of AddP's with no local inputs. 241 // This allows us to split-up address expressions. 242 if (m->is_AddP() && 243 get_ctrl(m->in(2)) != n_ctrl && 244 get_ctrl(m->in(3)) != n_ctrl) { 245 // Move the AddP up to dominating point 246 set_ctrl_and_loop(m, find_non_split_ctrl(idom(n_ctrl))); 247 continue; 248 } 249 return NULL; 250 } 251 } 252 253 return n_ctrl; 254} 255 256//------------------------------remix_address_expressions---------------------- 257// Rework addressing expressions to get the most loop-invariant stuff 258// moved out. We'd like to do all associative operators, but it's especially 259// important (common) to do address expressions. 260Node *PhaseIdealLoop::remix_address_expressions( Node *n ) { 261 if (!has_ctrl(n)) return NULL; 262 Node *n_ctrl = get_ctrl(n); 263 IdealLoopTree *n_loop = get_loop(n_ctrl); 264 265 // See if 'n' mixes loop-varying and loop-invariant inputs and 266 // itself is loop-varying. 267 268 // Only interested in binary ops (and AddP) 269 if( n->req() < 3 || n->req() > 4 ) return NULL; 270 271 Node *n1_ctrl = get_ctrl(n->in( 1)); 272 Node *n2_ctrl = get_ctrl(n->in( 2)); 273 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3)); 274 IdealLoopTree *n1_loop = get_loop( n1_ctrl ); 275 IdealLoopTree *n2_loop = get_loop( n2_ctrl ); 276 IdealLoopTree *n3_loop = get_loop( n3_ctrl ); 277 278 // Does one of my inputs spin in a tighter loop than self? 279 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) || 280 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) || 281 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) ) 282 return NULL; // Leave well enough alone 283 284 // Is at least one of my inputs loop-invariant? 285 if( n1_loop == n_loop && 286 n2_loop == n_loop && 287 n3_loop == n_loop ) 288 return NULL; // No loop-invariant inputs 289 290 291 int n_op = n->Opcode(); 292 293 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2). 294 if( n_op == Op_LShiftI ) { 295 // Scale is loop invariant 296 Node *scale = n->in(2); 297 Node *scale_ctrl = get_ctrl(scale); 298 IdealLoopTree *scale_loop = get_loop(scale_ctrl ); 299 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) ) 300 return NULL; 301 const TypeInt *scale_t = scale->bottom_type()->isa_int(); 302 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 ) 303 return NULL; // Dont bother with byte/short masking 304 // Add must vary with loop (else shift would be loop-invariant) 305 Node *add = n->in(1); 306 Node *add_ctrl = get_ctrl(add); 307 IdealLoopTree *add_loop = get_loop(add_ctrl); 308 //assert( n_loop == add_loop, "" ); 309 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops 310 311 // Convert I-V into I+ (0-V); same for V-I 312 if( add->Opcode() == Op_SubI && 313 _igvn.type( add->in(1) ) != TypeInt::ZERO ) { 314 Node *zero = _igvn.intcon(0); 315 set_ctrl(zero, C->root()); 316 Node *neg = new (C, 3) SubINode( _igvn.intcon(0), add->in(2) ); 317 register_new_node( neg, get_ctrl(add->in(2) ) ); 318 add = new (C, 3) AddINode( add->in(1), neg ); 319 register_new_node( add, add_ctrl ); 320 } 321 if( add->Opcode() != Op_AddI ) return NULL; 322 // See if one add input is loop invariant 323 Node *add_var = add->in(1); 324 Node *add_var_ctrl = get_ctrl(add_var); 325 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 326 Node *add_invar = add->in(2); 327 Node *add_invar_ctrl = get_ctrl(add_invar); 328 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl ); 329 if( add_var_loop == n_loop ) { 330 } else if( add_invar_loop == n_loop ) { 331 // Swap to find the invariant part 332 add_invar = add_var; 333 add_invar_ctrl = add_var_ctrl; 334 add_invar_loop = add_var_loop; 335 add_var = add->in(2); 336 Node *add_var_ctrl = get_ctrl(add_var); 337 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 338 } else // Else neither input is loop invariant 339 return NULL; 340 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) ) 341 return NULL; // No invariant part of the add? 342 343 // Yes! Reshape address expression! 344 Node *inv_scale = new (C, 3) LShiftINode( add_invar, scale ); 345 register_new_node( inv_scale, add_invar_ctrl ); 346 Node *var_scale = new (C, 3) LShiftINode( add_var, scale ); 347 register_new_node( var_scale, n_ctrl ); 348 Node *var_add = new (C, 3) AddINode( var_scale, inv_scale ); 349 register_new_node( var_add, n_ctrl ); 350 _igvn.hash_delete( n ); 351 _igvn.subsume_node( n, var_add ); 352 return var_add; 353 } 354 355 // Replace (I+V) with (V+I) 356 if( n_op == Op_AddI || 357 n_op == Op_AddL || 358 n_op == Op_AddF || 359 n_op == Op_AddD || 360 n_op == Op_MulI || 361 n_op == Op_MulL || 362 n_op == Op_MulF || 363 n_op == Op_MulD ) { 364 if( n2_loop == n_loop ) { 365 assert( n1_loop != n_loop, "" ); 366 n->swap_edges(1, 2); 367 } 368 } 369 370 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V), 371 // but not if I2 is a constant. 372 if( n_op == Op_AddP ) { 373 if( n2_loop == n_loop && n3_loop != n_loop ) { 374 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) { 375 Node *n22_ctrl = get_ctrl(n->in(2)->in(2)); 376 Node *n23_ctrl = get_ctrl(n->in(2)->in(3)); 377 IdealLoopTree *n22loop = get_loop( n22_ctrl ); 378 IdealLoopTree *n23_loop = get_loop( n23_ctrl ); 379 if( n22loop != n_loop && n22loop->is_member(n_loop) && 380 n23_loop == n_loop ) { 381 Node *add1 = new (C, 4) AddPNode( n->in(1), n->in(2)->in(2), n->in(3) ); 382 // Stuff new AddP in the loop preheader 383 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) ); 384 Node *add2 = new (C, 4) AddPNode( n->in(1), add1, n->in(2)->in(3) ); 385 register_new_node( add2, n_ctrl ); 386 _igvn.hash_delete( n ); 387 _igvn.subsume_node( n, add2 ); 388 return add2; 389 } 390 } 391 } 392 393 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V) 394 if( n2_loop != n_loop && n3_loop == n_loop ) { 395 if( n->in(3)->Opcode() == Op_AddI ) { 396 Node *V = n->in(3)->in(1); 397 Node *I = n->in(3)->in(2); 398 if( is_member(n_loop,get_ctrl(V)) ) { 399 } else { 400 Node *tmp = V; V = I; I = tmp; 401 } 402 if( !is_member(n_loop,get_ctrl(I)) ) { 403 Node *add1 = new (C, 4) AddPNode( n->in(1), n->in(2), I ); 404 // Stuff new AddP in the loop preheader 405 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) ); 406 Node *add2 = new (C, 4) AddPNode( n->in(1), add1, V ); 407 register_new_node( add2, n_ctrl ); 408 _igvn.hash_delete( n ); 409 _igvn.subsume_node( n, add2 ); 410 return add2; 411 } 412 } 413 } 414 } 415 416 return NULL; 417} 418 419//------------------------------conditional_move------------------------------- 420// Attempt to replace a Phi with a conditional move. We have some pretty 421// strict profitability requirements. All Phis at the merge point must 422// be converted, so we can remove the control flow. We need to limit the 423// number of c-moves to a small handful. All code that was in the side-arms 424// of the CFG diamond is now speculatively executed. This code has to be 425// "cheap enough". We are pretty much limited to CFG diamonds that merge 426// 1 or 2 items with a total of 1 or 2 ops executed speculatively. 427Node *PhaseIdealLoop::conditional_move( Node *region ) { 428 429 assert( region->is_Region(), "sanity check" ); 430 if( region->req() != 3 ) return NULL; 431 432 // Check for CFG diamond 433 Node *lp = region->in(1); 434 Node *rp = region->in(2); 435 if( !lp || !rp ) return NULL; 436 Node *lp_c = lp->in(0); 437 if( lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If() ) return NULL; 438 IfNode *iff = lp_c->as_If(); 439 440 // Check for highly predictable branch. No point in CMOV'ing if 441 // we are going to predict accurately all the time. 442 // %%% This hides patterns produced by utility methods like Math.min. 443 if( iff->_prob < PROB_UNLIKELY_MAG(3) || 444 iff->_prob > PROB_LIKELY_MAG(3) ) 445 return NULL; 446 447 // Check for ops pinned in an arm of the diamond. 448 // Can't remove the control flow in this case 449 if( lp->outcnt() > 1 ) return NULL; 450 if( rp->outcnt() > 1 ) return NULL; 451 452 // Check profitability 453 int cost = 0; 454 int phis = 0; 455 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 456 Node *out = region->fast_out(i); 457 if( !out->is_Phi() ) continue; // Ignore other control edges, etc 458 phis++; 459 PhiNode* phi = out->as_Phi(); 460 switch (phi->type()->basic_type()) { 461 case T_LONG: 462 cost++; // Probably encodes as 2 CMOV's 463 case T_INT: // These all CMOV fine 464 case T_FLOAT: 465 case T_DOUBLE: 466 case T_ADDRESS: // (RawPtr) 467 case T_NARROWOOP: 468 cost++; 469 break; 470 case T_OBJECT: { // Base oops are OK, but not derived oops 471 const TypeOopPtr *tp = phi->type()->isa_oopptr(); 472 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a 473 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus 474 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we 475 // have a Phi for the base here that we convert to a CMOVE all is well 476 // and good. But if the base is dead, we'll not make a CMOVE. Later 477 // the allocator will have to produce a base by creating a CMOVE of the 478 // relevant bases. This puts the allocator in the business of 479 // manufacturing expensive instructions, generally a bad plan. 480 // Just Say No to Conditionally-Moved Derived Pointers. 481 if( tp && tp->offset() != 0 ) 482 return NULL; 483 cost++; 484 break; 485 } 486 default: 487 return NULL; // In particular, can't do memory or I/O 488 } 489 // Add in cost any speculative ops 490 for( uint j = 1; j < region->req(); j++ ) { 491 Node *proj = region->in(j); 492 Node *inp = phi->in(j); 493 if (get_ctrl(inp) == proj) { // Found local op 494 cost++; 495 // Check for a chain of dependent ops; these will all become 496 // speculative in a CMOV. 497 for( uint k = 1; k < inp->req(); k++ ) 498 if (get_ctrl(inp->in(k)) == proj) 499 return NULL; // Too much speculative goo 500 } 501 } 502 // See if the Phi is used by a Cmp. This will likely Split-If, a 503 // higher-payoff operation. 504 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) { 505 Node* use = phi->fast_out(k); 506 if( use->is_Cmp() ) 507 return NULL; 508 } 509 } 510 if( cost >= ConditionalMoveLimit ) return NULL; // Too much goo 511 Node* bol = iff->in(1); 512 assert( bol->Opcode() == Op_Bool, "" ); 513 int cmp_op = bol->in(1)->Opcode(); 514 // It is expensive to generate flags from a float compare. 515 // Avoid duplicated float compare. 516 if( phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL; 517 518 // -------------- 519 // Now replace all Phis with CMOV's 520 Node *cmov_ctrl = iff->in(0); 521 uint flip = (lp->Opcode() == Op_IfTrue); 522 while( 1 ) { 523 PhiNode* phi = NULL; 524 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 525 Node *out = region->fast_out(i); 526 if (out->is_Phi()) { 527 phi = out->as_Phi(); 528 break; 529 } 530 } 531 if (phi == NULL) break; 532#ifndef PRODUCT 533 if( PrintOpto && VerifyLoopOptimizations ) tty->print_cr("CMOV"); 534#endif 535 // Move speculative ops 536 for( uint j = 1; j < region->req(); j++ ) { 537 Node *proj = region->in(j); 538 Node *inp = phi->in(j); 539 if (get_ctrl(inp) == proj) { // Found local op 540#ifndef PRODUCT 541 if( PrintOpto && VerifyLoopOptimizations ) { 542 tty->print(" speculate: "); 543 inp->dump(); 544 } 545#endif 546 set_ctrl(inp, cmov_ctrl); 547 } 548 } 549 Node *cmov = CMoveNode::make( C, cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi) ); 550 register_new_node( cmov, cmov_ctrl ); 551 _igvn.hash_delete(phi); 552 _igvn.subsume_node( phi, cmov ); 553#ifndef PRODUCT 554 if( VerifyLoopOptimizations ) verify(); 555#endif 556 } 557 558 // The useless CFG diamond will fold up later; see the optimization in 559 // RegionNode::Ideal. 560 _igvn._worklist.push(region); 561 562 return iff->in(1); 563} 564 565//------------------------------split_if_with_blocks_pre----------------------- 566// Do the real work in a non-recursive function. Data nodes want to be 567// cloned in the pre-order so they can feed each other nicely. 568Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) { 569 // Cloning these guys is unlikely to win 570 int n_op = n->Opcode(); 571 if( n_op == Op_MergeMem ) return n; 572 if( n->is_Proj() ) return n; 573 // Do not clone-up CmpFXXX variations, as these are always 574 // followed by a CmpI 575 if( n->is_Cmp() ) return n; 576 // Attempt to use a conditional move instead of a phi/branch 577 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) { 578 Node *cmov = conditional_move( n ); 579 if( cmov ) return cmov; 580 } 581 if( n->is_CFG() || n_op == Op_StorePConditional || n_op == Op_StoreLConditional || n_op == Op_CompareAndSwapI || n_op == Op_CompareAndSwapL ||n_op == Op_CompareAndSwapP) return n; 582 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd 583 n_op == Op_Opaque2 ) { 584 if( !C->major_progress() ) // If chance of no more loop opts... 585 _igvn._worklist.push(n); // maybe we'll remove them 586 return n; 587 } 588 589 if( n->is_Con() ) return n; // No cloning for Con nodes 590 591 Node *n_ctrl = get_ctrl(n); 592 if( !n_ctrl ) return n; // Dead node 593 594 // Attempt to remix address expressions for loop invariants 595 Node *m = remix_address_expressions( n ); 596 if( m ) return m; 597 598 // Determine if the Node has inputs from some local Phi. 599 // Returns the block to clone thru. 600 Node *n_blk = has_local_phi_input( n ); 601 if( !n_blk ) return n; 602 // Do not clone the trip counter through on a CountedLoop 603 // (messes up the canonical shape). 604 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n; 605 606 // Check for having no control input; not pinned. Allow 607 // dominating control. 608 if( n->in(0) ) { 609 Node *dom = idom(n_blk); 610 if( dom_lca( n->in(0), dom ) != n->in(0) ) 611 return n; 612 } 613 // Policy: when is it profitable. You must get more wins than 614 // policy before it is considered profitable. Policy is usually 0, 615 // so 1 win is considered profitable. Big merges will require big 616 // cloning, so get a larger policy. 617 int policy = n_blk->req() >> 2; 618 619 // If the loop is a candidate for range check elimination, 620 // delay splitting through it's phi until a later loop optimization 621 if (n_blk->is_CountedLoop()) { 622 IdealLoopTree *lp = get_loop(n_blk); 623 if (lp && lp->_rce_candidate) { 624 return n; 625 } 626 } 627 628 // Use same limit as split_if_with_blocks_post 629 if( C->unique() > 35000 ) return n; // Method too big 630 631 // Split 'n' through the merge point if it is profitable 632 Node *phi = split_thru_phi( n, n_blk, policy ); 633 if( !phi ) return n; 634 635 // Found a Phi to split thru! 636 // Replace 'n' with the new phi 637 _igvn.hash_delete(n); 638 _igvn.subsume_node( n, phi ); 639 // Moved a load around the loop, 'en-registering' something. 640 if( n_blk->Opcode() == Op_Loop && n->is_Load() && 641 !phi->in(LoopNode::LoopBackControl)->is_Load() ) 642 C->set_major_progress(); 643 644 return phi; 645} 646 647static bool merge_point_too_heavy(Compile* C, Node* region) { 648 // Bail out if the region and its phis have too many users. 649 int weight = 0; 650 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 651 weight += region->fast_out(i)->outcnt(); 652 } 653 int nodes_left = MaxNodeLimit - C->unique(); 654 if (weight * 8 > nodes_left) { 655#ifndef PRODUCT 656 if (PrintOpto) 657 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight); 658#endif 659 return true; 660 } else { 661 return false; 662 } 663} 664 665#ifdef _LP64 666static bool merge_point_safe(Node* region) { 667 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode 668 // having a PhiNode input. This sidesteps the dangerous case where the split 669 // ConvI2LNode may become TOP if the input Value() does not 670 // overlap the ConvI2L range, leaving a node which may not dominate its 671 // uses. 672 // A better fix for this problem can be found in the BugTraq entry, but 673 // expediency for Mantis demands this hack. 674 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 675 Node* n = region->fast_out(i); 676 if (n->is_Phi()) { 677 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 678 Node* m = n->fast_out(j); 679 if (m->Opcode() == Op_ConvI2L) { 680 return false; 681 } 682 } 683 } 684 } 685 return true; 686} 687#endif 688 689 690//------------------------------place_near_use--------------------------------- 691// Place some computation next to use but not inside inner loops. 692// For inner loop uses move it to the preheader area. 693Node *PhaseIdealLoop::place_near_use( Node *useblock ) const { 694 IdealLoopTree *u_loop = get_loop( useblock ); 695 return (u_loop->_irreducible || u_loop->_child) 696 ? useblock 697 : u_loop->_head->in(LoopNode::EntryControl); 698} 699 700 701//------------------------------split_if_with_blocks_post---------------------- 702// Do the real work in a non-recursive function. CFG hackery wants to be 703// in the post-order, so it can dirty the I-DOM info and not use the dirtied 704// info. 705void PhaseIdealLoop::split_if_with_blocks_post( Node *n ) { 706 707 // Cloning Cmp through Phi's involves the split-if transform. 708 // FastLock is not used by an If 709 if( n->is_Cmp() && !n->is_FastLock() ) { 710 if( C->unique() > 35000 ) return; // Method too big 711 712 // Do not do 'split-if' if irreducible loops are present. 713 if( _has_irreducible_loops ) 714 return; 715 716 Node *n_ctrl = get_ctrl(n); 717 // Determine if the Node has inputs from some local Phi. 718 // Returns the block to clone thru. 719 Node *n_blk = has_local_phi_input( n ); 720 if( n_blk != n_ctrl ) return; 721 722 if( merge_point_too_heavy(C, n_ctrl) ) 723 return; 724 725 if( n->outcnt() != 1 ) return; // Multiple bool's from 1 compare? 726 Node *bol = n->unique_out(); 727 assert( bol->is_Bool(), "expect a bool here" ); 728 if( bol->outcnt() != 1 ) return;// Multiple branches from 1 compare? 729 Node *iff = bol->unique_out(); 730 731 // Check some safety conditions 732 if( iff->is_If() ) { // Classic split-if? 733 if( iff->in(0) != n_ctrl ) return; // Compare must be in same blk as if 734 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE 735 if( get_ctrl(iff->in(2)) == n_ctrl || 736 get_ctrl(iff->in(3)) == n_ctrl ) 737 return; // Inputs not yet split-up 738 if ( get_loop(n_ctrl) != get_loop(get_ctrl(iff)) ) { 739 return; // Loop-invar test gates loop-varying CMOVE 740 } 741 } else { 742 return; // some other kind of node, such as an Allocate 743 } 744 745 // Do not do 'split-if' if some paths are dead. First do dead code 746 // elimination and then see if its still profitable. 747 for( uint i = 1; i < n_ctrl->req(); i++ ) 748 if( n_ctrl->in(i) == C->top() ) 749 return; 750 751 // When is split-if profitable? Every 'win' on means some control flow 752 // goes dead, so it's almost always a win. 753 int policy = 0; 754 // If trying to do a 'Split-If' at the loop head, it is only 755 // profitable if the cmp folds up on BOTH paths. Otherwise we 756 // risk peeling a loop forever. 757 758 // CNC - Disabled for now. Requires careful handling of loop 759 // body selection for the cloned code. Also, make sure we check 760 // for any input path not being in the same loop as n_ctrl. For 761 // irreducible loops we cannot check for 'n_ctrl->is_Loop()' 762 // because the alternative loop entry points won't be converted 763 // into LoopNodes. 764 IdealLoopTree *n_loop = get_loop(n_ctrl); 765 for( uint j = 1; j < n_ctrl->req(); j++ ) 766 if( get_loop(n_ctrl->in(j)) != n_loop ) 767 return; 768 769#ifdef _LP64 770 // Check for safety of the merge point. 771 if( !merge_point_safe(n_ctrl) ) { 772 return; 773 } 774#endif 775 776 // Split compare 'n' through the merge point if it is profitable 777 Node *phi = split_thru_phi( n, n_ctrl, policy ); 778 if( !phi ) return; 779 780 // Found a Phi to split thru! 781 // Replace 'n' with the new phi 782 _igvn.hash_delete(n); 783 _igvn.subsume_node( n, phi ); 784 785 // Now split the bool up thru the phi 786 Node *bolphi = split_thru_phi( bol, n_ctrl, -1 ); 787 _igvn.hash_delete(bol); 788 _igvn.subsume_node( bol, bolphi ); 789 assert( iff->in(1) == bolphi, "" ); 790 if( bolphi->Value(&_igvn)->singleton() ) 791 return; 792 793 // Conditional-move? Must split up now 794 if( !iff->is_If() ) { 795 Node *cmovphi = split_thru_phi( iff, n_ctrl, -1 ); 796 _igvn.hash_delete(iff); 797 _igvn.subsume_node( iff, cmovphi ); 798 return; 799 } 800 801 // Now split the IF 802 do_split_if( iff ); 803 return; 804 } 805 806 // Check for an IF ready to split; one that has its 807 // condition codes input coming from a Phi at the block start. 808 int n_op = n->Opcode(); 809 810 // Check for an IF being dominated by another IF same test 811 if( n_op == Op_If ) { 812 Node *bol = n->in(1); 813 uint max = bol->outcnt(); 814 // Check for same test used more than once? 815 if( n_op == Op_If && max > 1 && bol->is_Bool() ) { 816 // Search up IDOMs to see if this IF is dominated. 817 Node *cutoff = get_ctrl(bol); 818 819 // Now search up IDOMs till cutoff, looking for a dominating test 820 Node *prevdom = n; 821 Node *dom = idom(prevdom); 822 while( dom != cutoff ) { 823 if( dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom ) { 824 // Replace the dominated test with an obvious true or false. 825 // Place it on the IGVN worklist for later cleanup. 826 C->set_major_progress(); 827 dominated_by( prevdom, n ); 828#ifndef PRODUCT 829 if( VerifyLoopOptimizations ) verify(); 830#endif 831 return; 832 } 833 prevdom = dom; 834 dom = idom(prevdom); 835 } 836 } 837 } 838 839 // See if a shared loop-varying computation has no loop-varying uses. 840 // Happens if something is only used for JVM state in uncommon trap exits, 841 // like various versions of induction variable+offset. Clone the 842 // computation per usage to allow it to sink out of the loop. 843 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about) 844 Node *n_ctrl = get_ctrl(n); 845 IdealLoopTree *n_loop = get_loop(n_ctrl); 846 if( n_loop != _ltree_root ) { 847 DUIterator_Fast imax, i = n->fast_outs(imax); 848 for (; i < imax; i++) { 849 Node* u = n->fast_out(i); 850 if( !has_ctrl(u) ) break; // Found control user 851 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 852 if( u_loop == n_loop ) break; // Found loop-varying use 853 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop 854 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003 855 } 856 bool did_break = (i < imax); // Did we break out of the previous loop? 857 if (!did_break && n->outcnt() > 1) { // All uses in outer loops! 858 Node *late_load_ctrl; 859 if (n->is_Load()) { 860 // If n is a load, get and save the result from get_late_ctrl(), 861 // to be later used in calculating the control for n's clones. 862 clear_dom_lca_tags(); 863 late_load_ctrl = get_late_ctrl(n, n_ctrl); 864 } 865 // If n is a load, and the late control is the same as the current 866 // control, then the cloning of n is a pointless exercise, because 867 // GVN will ensure that we end up where we started. 868 if (!n->is_Load() || late_load_ctrl != n_ctrl) { 869 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) { 870 Node *u = n->last_out(j); // Clone private computation per use 871 _igvn.hash_delete(u); 872 _igvn._worklist.push(u); 873 Node *x = n->clone(); // Clone computation 874 Node *x_ctrl = NULL; 875 if( u->is_Phi() ) { 876 // Replace all uses of normal nodes. Replace Phi uses 877 // individually, so the seperate Nodes can sink down 878 // different paths. 879 uint k = 1; 880 while( u->in(k) != n ) k++; 881 u->set_req( k, x ); 882 // x goes next to Phi input path 883 x_ctrl = u->in(0)->in(k); 884 --j; 885 } else { // Normal use 886 // Replace all uses 887 for( uint k = 0; k < u->req(); k++ ) { 888 if( u->in(k) == n ) { 889 u->set_req( k, x ); 890 --j; 891 } 892 } 893 x_ctrl = get_ctrl(u); 894 } 895 896 // Find control for 'x' next to use but not inside inner loops. 897 // For inner loop uses get the preheader area. 898 x_ctrl = place_near_use(x_ctrl); 899 900 if (n->is_Load()) { 901 // For loads, add a control edge to a CFG node outside of the loop 902 // to force them to not combine and return back inside the loop 903 // during GVN optimization (4641526). 904 // 905 // Because we are setting the actual control input, factor in 906 // the result from get_late_ctrl() so we respect any 907 // anti-dependences. (6233005). 908 x_ctrl = dom_lca(late_load_ctrl, x_ctrl); 909 910 // Don't allow the control input to be a CFG splitting node. 911 // Such nodes should only have ProjNodes as outs, e.g. IfNode 912 // should only have IfTrueNode and IfFalseNode (4985384). 913 x_ctrl = find_non_split_ctrl(x_ctrl); 914 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone"); 915 916 x->set_req(0, x_ctrl); 917 } 918 register_new_node(x, x_ctrl); 919 920 // Some institutional knowledge is needed here: 'x' is 921 // yanked because if the optimizer runs GVN on it all the 922 // cloned x's will common up and undo this optimization and 923 // be forced back in the loop. This is annoying because it 924 // makes +VerifyOpto report false-positives on progress. I 925 // tried setting control edges on the x's to force them to 926 // not combine, but the matching gets worried when it tries 927 // to fold a StoreP and an AddP together (as part of an 928 // address expression) and the AddP and StoreP have 929 // different controls. 930 if( !x->is_Load() ) _igvn._worklist.yank(x); 931 } 932 _igvn.remove_dead_node(n); 933 } 934 } 935 } 936 } 937 938 // Check for Opaque2's who's loop has disappeared - who's input is in the 939 // same loop nest as their output. Remove 'em, they are no longer useful. 940 if( n_op == Op_Opaque2 && 941 n->in(1) != NULL && 942 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) { 943 _igvn.add_users_to_worklist(n); 944 _igvn.hash_delete(n); 945 _igvn.subsume_node( n, n->in(1) ); 946 } 947} 948 949//------------------------------split_if_with_blocks--------------------------- 950// Check for aggressive application of 'split-if' optimization, 951// using basic block level info. 952void PhaseIdealLoop::split_if_with_blocks( VectorSet &visited, Node_Stack &nstack ) { 953 Node *n = C->root(); 954 visited.set(n->_idx); // first, mark node as visited 955 // Do pre-visit work for root 956 n = split_if_with_blocks_pre( n ); 957 uint cnt = n->outcnt(); 958 uint i = 0; 959 while (true) { 960 // Visit all children 961 if (i < cnt) { 962 Node* use = n->raw_out(i); 963 ++i; 964 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) { 965 // Now do pre-visit work for this use 966 use = split_if_with_blocks_pre( use ); 967 nstack.push(n, i); // Save parent and next use's index. 968 n = use; // Process all children of current use. 969 cnt = use->outcnt(); 970 i = 0; 971 } 972 } 973 else { 974 // All of n's children have been processed, complete post-processing. 975 if (cnt != 0 && !n->is_Con()) { 976 assert(has_node(n), "no dead nodes"); 977 split_if_with_blocks_post( n ); 978 } 979 if (nstack.is_empty()) { 980 // Finished all nodes on stack. 981 break; 982 } 983 // Get saved parent node and next use's index. Visit the rest of uses. 984 n = nstack.node(); 985 cnt = n->outcnt(); 986 i = nstack.index(); 987 nstack.pop(); 988 } 989 } 990} 991 992 993//============================================================================= 994// 995// C L O N E A L O O P B O D Y 996// 997 998//------------------------------clone_iff-------------------------------------- 999// Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1000// "Nearly" because all Nodes have been cloned from the original in the loop, 1001// but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1002// through the Phi recursively, and return a Bool. 1003BoolNode *PhaseIdealLoop::clone_iff( PhiNode *phi, IdealLoopTree *loop ) { 1004 1005 // Convert this Phi into a Phi merging Bools 1006 uint i; 1007 for( i = 1; i < phi->req(); i++ ) { 1008 Node *b = phi->in(i); 1009 if( b->is_Phi() ) { 1010 _igvn.hash_delete(phi); 1011 _igvn._worklist.push(phi); 1012 phi->set_req(i, clone_iff( b->as_Phi(), loop )); 1013 } else { 1014 assert( b->is_Bool(), "" ); 1015 } 1016 } 1017 1018 Node *sample_bool = phi->in(1); 1019 Node *sample_cmp = sample_bool->in(1); 1020 1021 // Make Phis to merge the Cmp's inputs. 1022 int size = phi->in(0)->req(); 1023 PhiNode *phi1 = new (C, size) PhiNode( phi->in(0), Type::TOP ); 1024 PhiNode *phi2 = new (C, size) PhiNode( phi->in(0), Type::TOP ); 1025 for( i = 1; i < phi->req(); i++ ) { 1026 Node *n1 = phi->in(i)->in(1)->in(1); 1027 Node *n2 = phi->in(i)->in(1)->in(2); 1028 phi1->set_req( i, n1 ); 1029 phi2->set_req( i, n2 ); 1030 phi1->set_type( phi1->type()->meet(n1->bottom_type()) ); 1031 phi2->set_type( phi2->type()->meet(n2->bottom_type()) ); 1032 } 1033 // See if these Phis have been made before. 1034 // Register with optimizer 1035 Node *hit1 = _igvn.hash_find_insert(phi1); 1036 if( hit1 ) { // Hit, toss just made Phi 1037 _igvn.remove_dead_node(phi1); // Remove new phi 1038 assert( hit1->is_Phi(), "" ); 1039 phi1 = (PhiNode*)hit1; // Use existing phi 1040 } else { // Miss 1041 _igvn.register_new_node_with_optimizer(phi1); 1042 } 1043 Node *hit2 = _igvn.hash_find_insert(phi2); 1044 if( hit2 ) { // Hit, toss just made Phi 1045 _igvn.remove_dead_node(phi2); // Remove new phi 1046 assert( hit2->is_Phi(), "" ); 1047 phi2 = (PhiNode*)hit2; // Use existing phi 1048 } else { // Miss 1049 _igvn.register_new_node_with_optimizer(phi2); 1050 } 1051 // Register Phis with loop/block info 1052 set_ctrl(phi1, phi->in(0)); 1053 set_ctrl(phi2, phi->in(0)); 1054 // Make a new Cmp 1055 Node *cmp = sample_cmp->clone(); 1056 cmp->set_req( 1, phi1 ); 1057 cmp->set_req( 2, phi2 ); 1058 _igvn.register_new_node_with_optimizer(cmp); 1059 set_ctrl(cmp, phi->in(0)); 1060 1061 // Make a new Bool 1062 Node *b = sample_bool->clone(); 1063 b->set_req(1,cmp); 1064 _igvn.register_new_node_with_optimizer(b); 1065 set_ctrl(b, phi->in(0)); 1066 1067 assert( b->is_Bool(), "" ); 1068 return (BoolNode*)b; 1069} 1070 1071//------------------------------clone_bool------------------------------------- 1072// Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1073// "Nearly" because all Nodes have been cloned from the original in the loop, 1074// but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1075// through the Phi recursively, and return a Bool. 1076CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) { 1077 uint i; 1078 // Convert this Phi into a Phi merging Bools 1079 for( i = 1; i < phi->req(); i++ ) { 1080 Node *b = phi->in(i); 1081 if( b->is_Phi() ) { 1082 _igvn.hash_delete(phi); 1083 _igvn._worklist.push(phi); 1084 phi->set_req(i, clone_bool( b->as_Phi(), loop )); 1085 } else { 1086 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" ); 1087 } 1088 } 1089 1090 Node *sample_cmp = phi->in(1); 1091 1092 // Make Phis to merge the Cmp's inputs. 1093 int size = phi->in(0)->req(); 1094 PhiNode *phi1 = new (C, size) PhiNode( phi->in(0), Type::TOP ); 1095 PhiNode *phi2 = new (C, size) PhiNode( phi->in(0), Type::TOP ); 1096 for( uint j = 1; j < phi->req(); j++ ) { 1097 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP 1098 Node *n1, *n2; 1099 if( cmp_top->is_Cmp() ) { 1100 n1 = cmp_top->in(1); 1101 n2 = cmp_top->in(2); 1102 } else { 1103 n1 = n2 = cmp_top; 1104 } 1105 phi1->set_req( j, n1 ); 1106 phi2->set_req( j, n2 ); 1107 phi1->set_type( phi1->type()->meet(n1->bottom_type()) ); 1108 phi2->set_type( phi2->type()->meet(n2->bottom_type()) ); 1109 } 1110 1111 // See if these Phis have been made before. 1112 // Register with optimizer 1113 Node *hit1 = _igvn.hash_find_insert(phi1); 1114 if( hit1 ) { // Hit, toss just made Phi 1115 _igvn.remove_dead_node(phi1); // Remove new phi 1116 assert( hit1->is_Phi(), "" ); 1117 phi1 = (PhiNode*)hit1; // Use existing phi 1118 } else { // Miss 1119 _igvn.register_new_node_with_optimizer(phi1); 1120 } 1121 Node *hit2 = _igvn.hash_find_insert(phi2); 1122 if( hit2 ) { // Hit, toss just made Phi 1123 _igvn.remove_dead_node(phi2); // Remove new phi 1124 assert( hit2->is_Phi(), "" ); 1125 phi2 = (PhiNode*)hit2; // Use existing phi 1126 } else { // Miss 1127 _igvn.register_new_node_with_optimizer(phi2); 1128 } 1129 // Register Phis with loop/block info 1130 set_ctrl(phi1, phi->in(0)); 1131 set_ctrl(phi2, phi->in(0)); 1132 // Make a new Cmp 1133 Node *cmp = sample_cmp->clone(); 1134 cmp->set_req( 1, phi1 ); 1135 cmp->set_req( 2, phi2 ); 1136 _igvn.register_new_node_with_optimizer(cmp); 1137 set_ctrl(cmp, phi->in(0)); 1138 1139 assert( cmp->is_Cmp(), "" ); 1140 return (CmpNode*)cmp; 1141} 1142 1143//------------------------------sink_use--------------------------------------- 1144// If 'use' was in the loop-exit block, it now needs to be sunk 1145// below the post-loop merge point. 1146void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) { 1147 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) { 1148 set_ctrl(use, post_loop); 1149 for (DUIterator j = use->outs(); use->has_out(j); j++) 1150 sink_use(use->out(j), post_loop); 1151 } 1152} 1153 1154//------------------------------clone_loop------------------------------------- 1155// 1156// C L O N E A L O O P B O D Y 1157// 1158// This is the basic building block of the loop optimizations. It clones an 1159// entire loop body. It makes an old_new loop body mapping; with this mapping 1160// you can find the new-loop equivalent to an old-loop node. All new-loop 1161// nodes are exactly equal to their old-loop counterparts, all edges are the 1162// same. All exits from the old-loop now have a RegionNode that merges the 1163// equivalent new-loop path. This is true even for the normal "loop-exit" 1164// condition. All uses of loop-invariant old-loop values now come from (one 1165// or more) Phis that merge their new-loop equivalents. 1166// 1167// This operation leaves the graph in an illegal state: there are two valid 1168// control edges coming from the loop pre-header to both loop bodies. I'll 1169// definitely have to hack the graph after running this transform. 1170// 1171// From this building block I will further edit edges to perform loop peeling 1172// or loop unrolling or iteration splitting (Range-Check-Elimination), etc. 1173// 1174// Parameter side_by_size_idom: 1175// When side_by_size_idom is NULL, the dominator tree is constructed for 1176// the clone loop to dominate the original. Used in construction of 1177// pre-main-post loop sequence. 1178// When nonnull, the clone and original are side-by-side, both are 1179// dominated by the side_by_side_idom node. Used in construction of 1180// unswitched loops. 1181void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd, 1182 Node* side_by_side_idom) { 1183 1184 // Step 1: Clone the loop body. Make the old->new mapping. 1185 uint i; 1186 for( i = 0; i < loop->_body.size(); i++ ) { 1187 Node *old = loop->_body.at(i); 1188 Node *nnn = old->clone(); 1189 old_new.map( old->_idx, nnn ); 1190 _igvn.register_new_node_with_optimizer(nnn); 1191 } 1192 1193 1194 // Step 2: Fix the edges in the new body. If the old input is outside the 1195 // loop use it. If the old input is INside the loop, use the corresponding 1196 // new node instead. 1197 for( i = 0; i < loop->_body.size(); i++ ) { 1198 Node *old = loop->_body.at(i); 1199 Node *nnn = old_new[old->_idx]; 1200 // Fix CFG/Loop controlling the new node 1201 if (has_ctrl(old)) { 1202 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]); 1203 } else { 1204 set_loop(nnn, loop->_parent); 1205 if (old->outcnt() > 0) { 1206 set_idom( nnn, old_new[idom(old)->_idx], dd ); 1207 } 1208 } 1209 // Correct edges to the new node 1210 for( uint j = 0; j < nnn->req(); j++ ) { 1211 Node *n = nnn->in(j); 1212 if( n ) { 1213 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n ); 1214 if( loop->is_member( old_in_loop ) ) 1215 nnn->set_req(j, old_new[n->_idx]); 1216 } 1217 } 1218 _igvn.hash_find_insert(nnn); 1219 } 1220 Node *newhead = old_new[loop->_head->_idx]; 1221 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 1222 1223 1224 // Step 3: Now fix control uses. Loop varying control uses have already 1225 // been fixed up (as part of all input edges in Step 2). Loop invariant 1226 // control uses must be either an IfFalse or an IfTrue. Make a merge 1227 // point to merge the old and new IfFalse/IfTrue nodes; make the use 1228 // refer to this. 1229 ResourceArea *area = Thread::current()->resource_area(); 1230 Node_List worklist(area); 1231 uint new_counter = C->unique(); 1232 for( i = 0; i < loop->_body.size(); i++ ) { 1233 Node* old = loop->_body.at(i); 1234 if( !old->is_CFG() ) continue; 1235 Node* nnn = old_new[old->_idx]; 1236 1237 // Copy uses to a worklist, so I can munge the def-use info 1238 // with impunity. 1239 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1240 worklist.push(old->fast_out(j)); 1241 1242 while( worklist.size() ) { // Visit all uses 1243 Node *use = worklist.pop(); 1244 if (!has_node(use)) continue; // Ignore dead nodes 1245 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1246 if( !loop->is_member( use_loop ) && use->is_CFG() ) { 1247 // Both OLD and USE are CFG nodes here. 1248 assert( use->is_Proj(), "" ); 1249 1250 // Clone the loop exit control projection 1251 Node *newuse = use->clone(); 1252 newuse->set_req(0,nnn); 1253 _igvn.register_new_node_with_optimizer(newuse); 1254 set_loop(newuse, use_loop); 1255 set_idom(newuse, nnn, dom_depth(nnn) + 1 ); 1256 1257 // We need a Region to merge the exit from the peeled body and the 1258 // exit from the old loop body. 1259 RegionNode *r = new (C, 3) RegionNode(3); 1260 // Map the old use to the new merge point 1261 old_new.map( use->_idx, r ); 1262 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use)); 1263 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" ); 1264 1265 // The original user of 'use' uses 'r' instead. 1266 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) { 1267 Node* useuse = use->last_out(l); 1268 _igvn.hash_delete(useuse); 1269 _igvn._worklist.push(useuse); 1270 uint uses_found = 0; 1271 if( useuse->in(0) == use ) { 1272 useuse->set_req(0, r); 1273 uses_found++; 1274 if( useuse->is_CFG() ) { 1275 assert( dom_depth(useuse) > dd_r, "" ); 1276 set_idom(useuse, r, dom_depth(useuse)); 1277 } 1278 } 1279 for( uint k = 1; k < useuse->req(); k++ ) { 1280 if( useuse->in(k) == use ) { 1281 useuse->set_req(k, r); 1282 uses_found++; 1283 } 1284 } 1285 l -= uses_found; // we deleted 1 or more copies of this edge 1286 } 1287 1288 // Now finish up 'r' 1289 r->set_req( 1, newuse ); 1290 r->set_req( 2, use ); 1291 _igvn.register_new_node_with_optimizer(r); 1292 set_loop(r, use_loop); 1293 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r); 1294 } // End of if a loop-exit test 1295 } 1296 } 1297 1298 // Step 4: If loop-invariant use is not control, it must be dominated by a 1299 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region 1300 // there if needed. Make a Phi there merging old and new used values. 1301 Node_List *split_if_set = NULL; 1302 Node_List *split_bool_set = NULL; 1303 Node_List *split_cex_set = NULL; 1304 for( i = 0; i < loop->_body.size(); i++ ) { 1305 Node* old = loop->_body.at(i); 1306 Node* nnn = old_new[old->_idx]; 1307 // Copy uses to a worklist, so I can munge the def-use info 1308 // with impunity. 1309 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1310 worklist.push(old->fast_out(j)); 1311 1312 while( worklist.size() ) { 1313 Node *use = worklist.pop(); 1314 if (!has_node(use)) continue; // Ignore dead nodes 1315 if (use->in(0) == C->top()) continue; 1316 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1317 // Check for data-use outside of loop - at least one of OLD or USE 1318 // must not be a CFG node. 1319 if( !loop->is_member( use_loop ) && (!old->is_CFG() || !use->is_CFG())) { 1320 1321 // If the Data use is an IF, that means we have an IF outside of the 1322 // loop that is switching on a condition that is set inside of the 1323 // loop. Happens if people set a loop-exit flag; then test the flag 1324 // in the loop to break the loop, then test is again outside of the 1325 // loop to determine which way the loop exited. 1326 if( use->is_If() || use->is_CMove() ) { 1327 // Since this code is highly unlikely, we lazily build the worklist 1328 // of such Nodes to go split. 1329 if( !split_if_set ) 1330 split_if_set = new Node_List(area); 1331 split_if_set->push(use); 1332 } 1333 if( use->is_Bool() ) { 1334 if( !split_bool_set ) 1335 split_bool_set = new Node_List(area); 1336 split_bool_set->push(use); 1337 } 1338 if( use->Opcode() == Op_CreateEx ) { 1339 if( !split_cex_set ) 1340 split_cex_set = new Node_List(area); 1341 split_cex_set->push(use); 1342 } 1343 1344 1345 // Get "block" use is in 1346 uint idx = 0; 1347 while( use->in(idx) != old ) idx++; 1348 Node *prev = use->is_CFG() ? use : get_ctrl(use); 1349 assert( !loop->is_member( get_loop( prev ) ), "" ); 1350 Node *cfg = prev->_idx >= new_counter 1351 ? prev->in(2) 1352 : idom(prev); 1353 if( use->is_Phi() ) // Phi use is in prior block 1354 cfg = prev->in(idx); // NOT in block of Phi itself 1355 if (cfg->is_top()) { // Use is dead? 1356 _igvn.hash_delete(use); 1357 _igvn._worklist.push(use); 1358 use->set_req(idx, C->top()); 1359 continue; 1360 } 1361 1362 while( !loop->is_member( get_loop( cfg ) ) ) { 1363 prev = cfg; 1364 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg); 1365 } 1366 // If the use occurs after merging several exits from the loop, then 1367 // old value must have dominated all those exits. Since the same old 1368 // value was used on all those exits we did not need a Phi at this 1369 // merge point. NOW we do need a Phi here. Each loop exit value 1370 // is now merged with the peeled body exit; each exit gets its own 1371 // private Phi and those Phis need to be merged here. 1372 Node *phi; 1373 if( prev->is_Region() ) { 1374 if( idx == 0 ) { // Updating control edge? 1375 phi = prev; // Just use existing control 1376 } else { // Else need a new Phi 1377 phi = PhiNode::make( prev, old ); 1378 // Now recursively fix up the new uses of old! 1379 for( uint i = 1; i < prev->req(); i++ ) { 1380 worklist.push(phi); // Onto worklist once for each 'old' input 1381 } 1382 } 1383 } else { 1384 // Get new RegionNode merging old and new loop exits 1385 prev = old_new[prev->_idx]; 1386 assert( prev, "just made this in step 7" ); 1387 if( idx == 0 ) { // Updating control edge? 1388 phi = prev; // Just use existing control 1389 } else { // Else need a new Phi 1390 // Make a new Phi merging data values properly 1391 phi = PhiNode::make( prev, old ); 1392 phi->set_req( 1, nnn ); 1393 } 1394 } 1395 // If inserting a new Phi, check for prior hits 1396 if( idx != 0 ) { 1397 Node *hit = _igvn.hash_find_insert(phi); 1398 if( hit == NULL ) { 1399 _igvn.register_new_node_with_optimizer(phi); // Register new phi 1400 } else { // or 1401 // Remove the new phi from the graph and use the hit 1402 _igvn.remove_dead_node(phi); 1403 phi = hit; // Use existing phi 1404 } 1405 set_ctrl(phi, prev); 1406 } 1407 // Make 'use' use the Phi instead of the old loop body exit value 1408 _igvn.hash_delete(use); 1409 _igvn._worklist.push(use); 1410 use->set_req(idx, phi); 1411 if( use->_idx >= new_counter ) { // If updating new phis 1412 // Not needed for correctness, but prevents a weak assert 1413 // in AddPNode from tripping (when we end up with different 1414 // base & derived Phis that will become the same after 1415 // IGVN does CSE). 1416 Node *hit = _igvn.hash_find_insert(use); 1417 if( hit ) // Go ahead and re-hash for hits. 1418 _igvn.subsume_node( use, hit ); 1419 } 1420 1421 // If 'use' was in the loop-exit block, it now needs to be sunk 1422 // below the post-loop merge point. 1423 sink_use( use, prev ); 1424 } 1425 } 1426 } 1427 1428 // Check for IFs that need splitting/cloning. Happens if an IF outside of 1429 // the loop uses a condition set in the loop. The original IF probably 1430 // takes control from one or more OLD Regions (which in turn get from NEW 1431 // Regions). In any case, there will be a set of Phis for each merge point 1432 // from the IF up to where the original BOOL def exists the loop. 1433 if( split_if_set ) { 1434 while( split_if_set->size() ) { 1435 Node *iff = split_if_set->pop(); 1436 if( iff->in(1)->is_Phi() ) { 1437 BoolNode *b = clone_iff( iff->in(1)->as_Phi(), loop ); 1438 _igvn.hash_delete(iff); 1439 _igvn._worklist.push(iff); 1440 iff->set_req(1, b); 1441 } 1442 } 1443 } 1444 if( split_bool_set ) { 1445 while( split_bool_set->size() ) { 1446 Node *b = split_bool_set->pop(); 1447 Node *phi = b->in(1); 1448 assert( phi->is_Phi(), "" ); 1449 CmpNode *cmp = clone_bool( (PhiNode*)phi, loop ); 1450 _igvn.hash_delete(b); 1451 _igvn._worklist.push(b); 1452 b->set_req(1, cmp); 1453 } 1454 } 1455 if( split_cex_set ) { 1456 while( split_cex_set->size() ) { 1457 Node *b = split_cex_set->pop(); 1458 assert( b->in(0)->is_Region(), "" ); 1459 assert( b->in(1)->is_Phi(), "" ); 1460 assert( b->in(0)->in(0) == b->in(1)->in(0), "" ); 1461 split_up( b, b->in(0), NULL ); 1462 } 1463 } 1464 1465} 1466 1467 1468//---------------------- stride_of_possible_iv ------------------------------------- 1469// Looks for an iff/bool/comp with one operand of the compare 1470// being a cycle involving an add and a phi, 1471// with an optional truncation (left-shift followed by a right-shift) 1472// of the add. Returns zero if not an iv. 1473int PhaseIdealLoop::stride_of_possible_iv(Node* iff) { 1474 Node* trunc1 = NULL; 1475 Node* trunc2 = NULL; 1476 const TypeInt* ttype = NULL; 1477 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) { 1478 return 0; 1479 } 1480 BoolNode* bl = iff->in(1)->as_Bool(); 1481 Node* cmp = bl->in(1); 1482 if (!cmp || cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU) { 1483 return 0; 1484 } 1485 // Must have an invariant operand 1486 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) { 1487 return 0; 1488 } 1489 Node* add2 = NULL; 1490 Node* cmp1 = cmp->in(1); 1491 if (cmp1->is_Phi()) { 1492 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) ))) 1493 Node* phi = cmp1; 1494 for (uint i = 1; i < phi->req(); i++) { 1495 Node* in = phi->in(i); 1496 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in, 1497 &trunc1, &trunc2, &ttype); 1498 if (add && add->in(1) == phi) { 1499 add2 = add->in(2); 1500 break; 1501 } 1502 } 1503 } else { 1504 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) ))) 1505 Node* addtrunc = cmp1; 1506 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc, 1507 &trunc1, &trunc2, &ttype); 1508 if (add && add->in(1)->is_Phi()) { 1509 Node* phi = add->in(1); 1510 for (uint i = 1; i < phi->req(); i++) { 1511 if (phi->in(i) == addtrunc) { 1512 add2 = add->in(2); 1513 break; 1514 } 1515 } 1516 } 1517 } 1518 if (add2 != NULL) { 1519 const TypeInt* add2t = _igvn.type(add2)->is_int(); 1520 if (add2t->is_con()) { 1521 return add2t->get_con(); 1522 } 1523 } 1524 return 0; 1525} 1526 1527 1528//---------------------- stay_in_loop ------------------------------------- 1529// Return the (unique) control output node that's in the loop (if it exists.) 1530Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) { 1531 Node* unique = NULL; 1532 if (!n) return NULL; 1533 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1534 Node* use = n->fast_out(i); 1535 if (!has_ctrl(use) && loop->is_member(get_loop(use))) { 1536 if (unique != NULL) { 1537 return NULL; 1538 } 1539 unique = use; 1540 } 1541 } 1542 return unique; 1543} 1544 1545//------------------------------ register_node ------------------------------------- 1546// Utility to register node "n" with PhaseIdealLoop 1547void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) { 1548 _igvn.register_new_node_with_optimizer(n); 1549 loop->_body.push(n); 1550 if (n->is_CFG()) { 1551 set_loop(n, loop); 1552 set_idom(n, pred, ddepth); 1553 } else { 1554 set_ctrl(n, pred); 1555 } 1556} 1557 1558//------------------------------ proj_clone ------------------------------------- 1559// Utility to create an if-projection 1560ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) { 1561 ProjNode* c = p->clone()->as_Proj(); 1562 c->set_req(0, iff); 1563 return c; 1564} 1565 1566//------------------------------ short_circuit_if ------------------------------------- 1567// Force the iff control output to be the live_proj 1568Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) { 1569 int proj_con = live_proj->_con; 1570 assert(proj_con == 0 || proj_con == 1, "false or true projection"); 1571 Node *con = _igvn.intcon(proj_con); 1572 set_ctrl(con, C->root()); 1573 if (iff) { 1574 iff->set_req(1, con); 1575 } 1576 return con; 1577} 1578 1579//------------------------------ insert_if_before_proj ------------------------------------- 1580// Insert a new if before an if projection (* - new node) 1581// 1582// before 1583// if(test) 1584// / \ 1585// v v 1586// other-proj proj (arg) 1587// 1588// after 1589// if(test) 1590// / \ 1591// / v 1592// | * proj-clone 1593// v | 1594// other-proj v 1595// * new_if(relop(cmp[IU](left,right))) 1596// / \ 1597// v v 1598// * new-proj proj 1599// (returned) 1600// 1601ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) { 1602 IfNode* iff = proj->in(0)->as_If(); 1603 IdealLoopTree *loop = get_loop(proj); 1604 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 1605 int ddepth = dom_depth(proj); 1606 1607 _igvn.hash_delete(iff); 1608 _igvn._worklist.push(iff); 1609 _igvn.hash_delete(proj); 1610 _igvn._worklist.push(proj); 1611 1612 proj->set_req(0, NULL); // temporary disconnect 1613 ProjNode* proj2 = proj_clone(proj, iff); 1614 register_node(proj2, loop, iff, ddepth); 1615 1616 Node* cmp = Signed ? (Node*) new (C,3)CmpINode(left, right) : (Node*) new (C,3)CmpUNode(left, right); 1617 register_node(cmp, loop, proj2, ddepth); 1618 1619 BoolNode* bol = new (C,2)BoolNode(cmp, relop); 1620 register_node(bol, loop, proj2, ddepth); 1621 1622 IfNode* new_if = new (C,2)IfNode(proj2, bol, iff->_prob, iff->_fcnt); 1623 register_node(new_if, loop, proj2, ddepth); 1624 1625 proj->set_req(0, new_if); // reattach 1626 set_idom(proj, new_if, ddepth); 1627 1628 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj(); 1629 register_node(new_exit, get_loop(other_proj), new_if, ddepth); 1630 1631 return new_exit; 1632} 1633 1634//------------------------------ insert_region_before_proj ------------------------------------- 1635// Insert a region before an if projection (* - new node) 1636// 1637// before 1638// if(test) 1639// / | 1640// v | 1641// proj v 1642// other-proj 1643// 1644// after 1645// if(test) 1646// / | 1647// v | 1648// * proj-clone v 1649// | other-proj 1650// v 1651// * new-region 1652// | 1653// v 1654// * dum_if 1655// / \ 1656// v \ 1657// * dum-proj v 1658// proj 1659// 1660RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) { 1661 IfNode* iff = proj->in(0)->as_If(); 1662 IdealLoopTree *loop = get_loop(proj); 1663 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 1664 int ddepth = dom_depth(proj); 1665 1666 _igvn.hash_delete(iff); 1667 _igvn._worklist.push(iff); 1668 _igvn.hash_delete(proj); 1669 _igvn._worklist.push(proj); 1670 1671 proj->set_req(0, NULL); // temporary disconnect 1672 ProjNode* proj2 = proj_clone(proj, iff); 1673 register_node(proj2, loop, iff, ddepth); 1674 1675 RegionNode* reg = new (C,2)RegionNode(2); 1676 reg->set_req(1, proj2); 1677 register_node(reg, loop, iff, ddepth); 1678 1679 IfNode* dum_if = new (C,2)IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt); 1680 register_node(dum_if, loop, reg, ddepth); 1681 1682 proj->set_req(0, dum_if); // reattach 1683 set_idom(proj, dum_if, ddepth); 1684 1685 ProjNode* dum_proj = proj_clone(other_proj, dum_if); 1686 register_node(dum_proj, loop, dum_if, ddepth); 1687 1688 return reg; 1689} 1690 1691//------------------------------ insert_cmpi_loop_exit ------------------------------------- 1692// Clone a signed compare loop exit from an unsigned compare and 1693// insert it before the unsigned cmp on the stay-in-loop path. 1694// All new nodes inserted in the dominator tree between the original 1695// if and it's projections. The original if test is replaced with 1696// a constant to force the stay-in-loop path. 1697// 1698// This is done to make sure that the original if and it's projections 1699// still dominate the same set of control nodes, that the ctrl() relation 1700// from data nodes to them is preserved, and that their loop nesting is 1701// preserved. 1702// 1703// before 1704// if(i <u limit) unsigned compare loop exit 1705// / | 1706// v v 1707// exit-proj stay-in-loop-proj 1708// 1709// after 1710// if(stay-in-loop-const) original if 1711// / | 1712// / v 1713// / if(i < limit) new signed test 1714// / / | 1715// / / v 1716// / / if(i <u limit) new cloned unsigned test 1717// / / / | 1718// v v v | 1719// region | 1720// | | 1721// dum-if | 1722// / | | 1723// ether | | 1724// v v 1725// exit-proj stay-in-loop-proj 1726// 1727IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) { 1728 const bool Signed = true; 1729 const bool Unsigned = false; 1730 1731 BoolNode* bol = if_cmpu->in(1)->as_Bool(); 1732 if (bol->_test._test != BoolTest::lt) return NULL; 1733 CmpNode* cmpu = bol->in(1)->as_Cmp(); 1734 if (cmpu->Opcode() != Op_CmpU) return NULL; 1735 int stride = stride_of_possible_iv(if_cmpu); 1736 if (stride == 0) return NULL; 1737 1738 ProjNode* lp_continue = stay_in_loop(if_cmpu, loop)->as_Proj(); 1739 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj(); 1740 1741 Node* limit = NULL; 1742 if (stride > 0) { 1743 limit = cmpu->in(2); 1744 } else { 1745 limit = _igvn.makecon(TypeInt::ZERO); 1746 set_ctrl(limit, C->root()); 1747 } 1748 // Create a new region on the exit path 1749 RegionNode* reg = insert_region_before_proj(lp_exit); 1750 1751 // Clone the if-cmpu-true-false using a signed compare 1752 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge; 1753 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue); 1754 reg->add_req(cmpi_exit); 1755 1756 // Clone the if-cmpu-true-false 1757 BoolTest::mask rel_u = bol->_test._test; 1758 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue); 1759 reg->add_req(cmpu_exit); 1760 1761 // Force original if to stay in loop. 1762 short_circuit_if(if_cmpu, lp_continue); 1763 1764 return cmpi_exit->in(0)->as_If(); 1765} 1766 1767//------------------------------ remove_cmpi_loop_exit ------------------------------------- 1768// Remove a previously inserted signed compare loop exit. 1769void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) { 1770 Node* lp_proj = stay_in_loop(if_cmp, loop); 1771 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI && 1772 stay_in_loop(lp_proj, loop)->is_If() && 1773 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu"); 1774 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO); 1775 set_ctrl(con, C->root()); 1776 if_cmp->set_req(1, con); 1777} 1778 1779//------------------------------ scheduled_nodelist ------------------------------------- 1780// Create a post order schedule of nodes that are in the 1781// "member" set. The list is returned in "sched". 1782// The first node in "sched" is the loop head, followed by 1783// nodes which have no inputs in the "member" set, and then 1784// followed by the nodes that have an immediate input dependence 1785// on a node in "sched". 1786void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) { 1787 1788 assert(member.test(loop->_head->_idx), "loop head must be in member set"); 1789 Arena *a = Thread::current()->resource_area(); 1790 VectorSet visited(a); 1791 Node_Stack nstack(a, loop->_body.size()); 1792 1793 Node* n = loop->_head; // top of stack is cached in "n" 1794 uint idx = 0; 1795 visited.set(n->_idx); 1796 1797 // Initially push all with no inputs from within member set 1798 for(uint i = 0; i < loop->_body.size(); i++ ) { 1799 Node *elt = loop->_body.at(i); 1800 if (member.test(elt->_idx)) { 1801 bool found = false; 1802 for (uint j = 0; j < elt->req(); j++) { 1803 Node* def = elt->in(j); 1804 if (def && member.test(def->_idx) && def != elt) { 1805 found = true; 1806 break; 1807 } 1808 } 1809 if (!found && elt != loop->_head) { 1810 nstack.push(n, idx); 1811 n = elt; 1812 assert(!visited.test(n->_idx), "not seen yet"); 1813 visited.set(n->_idx); 1814 } 1815 } 1816 } 1817 1818 // traverse out's that are in the member set 1819 while (true) { 1820 if (idx < n->outcnt()) { 1821 Node* use = n->raw_out(idx); 1822 idx++; 1823 if (!visited.test_set(use->_idx)) { 1824 if (member.test(use->_idx)) { 1825 nstack.push(n, idx); 1826 n = use; 1827 idx = 0; 1828 } 1829 } 1830 } else { 1831 // All outputs processed 1832 sched.push(n); 1833 if (nstack.is_empty()) break; 1834 n = nstack.node(); 1835 idx = nstack.index(); 1836 nstack.pop(); 1837 } 1838 } 1839} 1840 1841 1842//------------------------------ has_use_in_set ------------------------------------- 1843// Has a use in the vector set 1844bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) { 1845 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1846 Node* use = n->fast_out(j); 1847 if (vset.test(use->_idx)) { 1848 return true; 1849 } 1850 } 1851 return false; 1852} 1853 1854 1855//------------------------------ has_use_internal_to_set ------------------------------------- 1856// Has use internal to the vector set (ie. not in a phi at the loop head) 1857bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) { 1858 Node* head = loop->_head; 1859 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1860 Node* use = n->fast_out(j); 1861 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) { 1862 return true; 1863 } 1864 } 1865 return false; 1866} 1867 1868 1869//------------------------------ clone_for_use_outside_loop ------------------------------------- 1870// clone "n" for uses that are outside of loop 1871void PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) { 1872 1873 assert(worklist.size() == 0, "should be empty"); 1874 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1875 Node* use = n->fast_out(j); 1876 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) { 1877 worklist.push(use); 1878 } 1879 } 1880 while( worklist.size() ) { 1881 Node *use = worklist.pop(); 1882 if (!has_node(use) || use->in(0) == C->top()) continue; 1883 uint j; 1884 for (j = 0; j < use->req(); j++) { 1885 if (use->in(j) == n) break; 1886 } 1887 assert(j < use->req(), "must be there"); 1888 1889 // clone "n" and insert it between the inputs of "n" and the use outside the loop 1890 Node* n_clone = n->clone(); 1891 _igvn.hash_delete(use); 1892 use->set_req(j, n_clone); 1893 _igvn._worklist.push(use); 1894 if (!use->is_Phi()) { 1895 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0); 1896 set_ctrl(n_clone, use_c); 1897 assert(!loop->is_member(get_loop(use_c)), "should be outside loop"); 1898 get_loop(use_c)->_body.push(n_clone); 1899 } else { 1900 // Use in a phi is considered a use in the associated predecessor block 1901 Node *prevbb = use->in(0)->in(j); 1902 set_ctrl(n_clone, prevbb); 1903 assert(!loop->is_member(get_loop(prevbb)), "should be outside loop"); 1904 get_loop(prevbb)->_body.push(n_clone); 1905 } 1906 _igvn.register_new_node_with_optimizer(n_clone); 1907#if !defined(PRODUCT) 1908 if (TracePartialPeeling) { 1909 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx); 1910 } 1911#endif 1912 } 1913} 1914 1915 1916//------------------------------ clone_for_special_use_inside_loop ------------------------------------- 1917// clone "n" for special uses that are in the not_peeled region. 1918// If these def-uses occur in separate blocks, the code generator 1919// marks the method as not compilable. For example, if a "BoolNode" 1920// is in a different basic block than the "IfNode" that uses it, then 1921// the compilation is aborted in the code generator. 1922void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 1923 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) { 1924 if (n->is_Phi() || n->is_Load()) { 1925 return; 1926 } 1927 assert(worklist.size() == 0, "should be empty"); 1928 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1929 Node* use = n->fast_out(j); 1930 if ( not_peel.test(use->_idx) && 1931 (use->is_If() || use->is_CMove() || use->is_Bool()) && 1932 use->in(1) == n) { 1933 worklist.push(use); 1934 } 1935 } 1936 if (worklist.size() > 0) { 1937 // clone "n" and insert it between inputs of "n" and the use 1938 Node* n_clone = n->clone(); 1939 loop->_body.push(n_clone); 1940 _igvn.register_new_node_with_optimizer(n_clone); 1941 set_ctrl(n_clone, get_ctrl(n)); 1942 sink_list.push(n_clone); 1943 not_peel <<= n_clone->_idx; // add n_clone to not_peel set. 1944#if !defined(PRODUCT) 1945 if (TracePartialPeeling) { 1946 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx); 1947 } 1948#endif 1949 while( worklist.size() ) { 1950 Node *use = worklist.pop(); 1951 _igvn.hash_delete(use); 1952 _igvn._worklist.push(use); 1953 for (uint j = 1; j < use->req(); j++) { 1954 if (use->in(j) == n) { 1955 use->set_req(j, n_clone); 1956 } 1957 } 1958 } 1959 } 1960} 1961 1962 1963//------------------------------ insert_phi_for_loop ------------------------------------- 1964// Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 1965void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) { 1966 Node *phi = PhiNode::make(lp, back_edge_val); 1967 phi->set_req(LoopNode::EntryControl, lp_entry_val); 1968 // Use existing phi if it already exists 1969 Node *hit = _igvn.hash_find_insert(phi); 1970 if( hit == NULL ) { 1971 _igvn.register_new_node_with_optimizer(phi); 1972 set_ctrl(phi, lp); 1973 } else { 1974 // Remove the new phi from the graph and use the hit 1975 _igvn.remove_dead_node(phi); 1976 phi = hit; 1977 } 1978 _igvn.hash_delete(use); 1979 _igvn._worklist.push(use); 1980 use->set_req(idx, phi); 1981} 1982 1983#ifdef ASSERT 1984//------------------------------ is_valid_loop_partition ------------------------------------- 1985// Validate the loop partition sets: peel and not_peel 1986bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, 1987 VectorSet& not_peel ) { 1988 uint i; 1989 // Check that peel_list entries are in the peel set 1990 for (i = 0; i < peel_list.size(); i++) { 1991 if (!peel.test(peel_list.at(i)->_idx)) { 1992 return false; 1993 } 1994 } 1995 // Check at loop members are in one of peel set or not_peel set 1996 for (i = 0; i < loop->_body.size(); i++ ) { 1997 Node *def = loop->_body.at(i); 1998 uint di = def->_idx; 1999 // Check that peel set elements are in peel_list 2000 if (peel.test(di)) { 2001 if (not_peel.test(di)) { 2002 return false; 2003 } 2004 // Must be in peel_list also 2005 bool found = false; 2006 for (uint j = 0; j < peel_list.size(); j++) { 2007 if (peel_list.at(j)->_idx == di) { 2008 found = true; 2009 break; 2010 } 2011 } 2012 if (!found) { 2013 return false; 2014 } 2015 } else if (not_peel.test(di)) { 2016 if (peel.test(di)) { 2017 return false; 2018 } 2019 } else { 2020 return false; 2021 } 2022 } 2023 return true; 2024} 2025 2026//------------------------------ is_valid_clone_loop_exit_use ------------------------------------- 2027// Ensure a use outside of loop is of the right form 2028bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) { 2029 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2030 return (use->is_Phi() && 2031 use_c->is_Region() && use_c->req() == 3 && 2032 (use_c->in(exit_idx)->Opcode() == Op_IfTrue || 2033 use_c->in(exit_idx)->Opcode() == Op_IfFalse || 2034 use_c->in(exit_idx)->Opcode() == Op_JumpProj) && 2035 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) ); 2036} 2037 2038//------------------------------ is_valid_clone_loop_form ------------------------------------- 2039// Ensure that all uses outside of loop are of the right form 2040bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 2041 uint orig_exit_idx, uint clone_exit_idx) { 2042 uint len = peel_list.size(); 2043 for (uint i = 0; i < len; i++) { 2044 Node *def = peel_list.at(i); 2045 2046 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 2047 Node *use = def->fast_out(j); 2048 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2049 if (!loop->is_member(get_loop(use_c))) { 2050 // use is not in the loop, check for correct structure 2051 if (use->in(0) == def) { 2052 // Okay 2053 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) { 2054 return false; 2055 } 2056 } 2057 } 2058 } 2059 return true; 2060} 2061#endif 2062 2063//------------------------------ partial_peel ------------------------------------- 2064// Partially peel (aka loop rotation) the top portion of a loop (called 2065// the peel section below) by cloning it and placing one copy just before 2066// the new loop head and the other copy at the bottom of the new loop. 2067// 2068// before after where it came from 2069// 2070// stmt1 stmt1 2071// loop: stmt2 clone 2072// stmt2 if condA goto exitA clone 2073// if condA goto exitA new_loop: new 2074// stmt3 stmt3 clone 2075// if !condB goto loop if condB goto exitB clone 2076// exitB: stmt2 orig 2077// stmt4 if !condA goto new_loop orig 2078// exitA: goto exitA 2079// exitB: 2080// stmt4 2081// exitA: 2082// 2083// Step 1: find the cut point: an exit test on probable 2084// induction variable. 2085// Step 2: schedule (with cloning) operations in the peel 2086// section that can be executed after the cut into 2087// the section that is not peeled. This may need 2088// to clone operations into exit blocks. For 2089// instance, a reference to A[i] in the not-peel 2090// section and a reference to B[i] in an exit block 2091// may cause a left-shift of i by 2 to be placed 2092// in the peel block. This step will clone the left 2093// shift into the exit block and sink the left shift 2094// from the peel to the not-peel section. 2095// Step 3: clone the loop, retarget the control, and insert 2096// phis for values that are live across the new loop 2097// head. This is very dependent on the graph structure 2098// from clone_loop. It creates region nodes for 2099// exit control and associated phi nodes for values 2100// flow out of the loop through that exit. The region 2101// node is dominated by the clone's control projection. 2102// So the clone's peel section is placed before the 2103// new loop head, and the clone's not-peel section is 2104// forms the top part of the new loop. The original 2105// peel section forms the tail of the new loop. 2106// Step 4: update the dominator tree and recompute the 2107// dominator depth. 2108// 2109// orig 2110// 2111// stmt1 2112// | 2113// v 2114// loop<----+ 2115// | | 2116// stmt2 | 2117// | | 2118// v | 2119// ifA | 2120// / | | 2121// v v | 2122// false true ^ <-- last_peel 2123// / | | 2124// / ===|==cut | 2125// / stmt3 | <-- first_not_peel 2126// / | | 2127// | v | 2128// v ifB | 2129// exitA: / \ | 2130// / \ | 2131// v v | 2132// false true | 2133// / \ | 2134// / ----+ 2135// | 2136// v 2137// exitB: 2138// stmt4 2139// 2140// 2141// after clone loop 2142// 2143// stmt1 2144// / \ 2145// clone / \ orig 2146// / \ 2147// / \ 2148// v v 2149// +---->loop loop<----+ 2150// | | | | 2151// | stmt2 stmt2 | 2152// | | | | 2153// | v v | 2154// | ifA ifA | 2155// | | \ / | | 2156// | v v v v | 2157// ^ true false false true ^ <-- last_peel 2158// | | ^ \ / | | 2159// | cut==|== \ \ / ===|==cut | 2160// | stmt3 \ \ / stmt3 | <-- first_not_peel 2161// | | dom | | | | 2162// | v \ 1v v2 v | 2163// | ifB regionA ifB | 2164// | / \ | / \ | 2165// | / \ v / \ | 2166// | v v exitA: v v | 2167// | true false false true | 2168// | / ^ \ / \ | 2169// +---- \ \ / ----+ 2170// dom \ / 2171// \ 1v v2 2172// regionB 2173// | 2174// v 2175// exitB: 2176// stmt4 2177// 2178// 2179// after partial peel 2180// 2181// stmt1 2182// / 2183// clone / orig 2184// / TOP 2185// / \ 2186// v v 2187// TOP->region region----+ 2188// | | | 2189// stmt2 stmt2 | 2190// | | | 2191// v v | 2192// ifA ifA | 2193// | \ / | | 2194// v v v v | 2195// true false false true | <-- last_peel 2196// | ^ \ / +------|---+ 2197// +->newloop \ \ / === ==cut | | 2198// | stmt3 \ \ / TOP | | 2199// | | dom | | stmt3 | | <-- first_not_peel 2200// | v \ 1v v2 v | | 2201// | ifB regionA ifB ^ v 2202// | / \ | / \ | | 2203// | / \ v / \ | | 2204// | v v exitA: v v | | 2205// | true false false true | | 2206// | / ^ \ / \ | | 2207// | | \ \ / v | | 2208// | | dom \ / TOP | | 2209// | | \ 1v v2 | | 2210// ^ v regionB | | 2211// | | | | | 2212// | | v ^ v 2213// | | exitB: | | 2214// | | stmt4 | | 2215// | +------------>-----------------+ | 2216// | | 2217// +-----------------<---------------------+ 2218// 2219// 2220// final graph 2221// 2222// stmt1 2223// | 2224// v 2225// ........> ifA clone 2226// : / | 2227// dom / | 2228// : v v 2229// : false true 2230// : | | 2231// : | stmt2 clone 2232// : | | 2233// : | v 2234// : | newloop<-----+ 2235// : | | | 2236// : | stmt3 clone | 2237// : | | | 2238// : | v | 2239// : | ifB | 2240// : | / \ | 2241// : | v v | 2242// : | false true | 2243// : | | | | 2244// : | v stmt2 | 2245// : | exitB: | | 2246// : | stmt4 v | 2247// : | ifA orig | 2248// : | / \ | 2249// : | / \ | 2250// : | v v | 2251// : | false true | 2252// : | / \ | 2253// : v v -----+ 2254// RegionA 2255// | 2256// v 2257// exitA 2258// 2259bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) { 2260 2261 if (!loop->_head->is_Loop()) { 2262 return false; } 2263 2264 LoopNode *head = loop->_head->as_Loop(); 2265 2266 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) { 2267 return false; 2268 } 2269 2270 // Check for complex exit control 2271 for(uint ii = 0; ii < loop->_body.size(); ii++ ) { 2272 Node *n = loop->_body.at(ii); 2273 int opc = n->Opcode(); 2274 if (n->is_Call() || 2275 opc == Op_Catch || 2276 opc == Op_CatchProj || 2277 opc == Op_Jump || 2278 opc == Op_JumpProj) { 2279#if !defined(PRODUCT) 2280 if (TracePartialPeeling) { 2281 tty->print_cr("\nExit control too complex: lp: %d", head->_idx); 2282 } 2283#endif 2284 return false; 2285 } 2286 } 2287 2288 int dd = dom_depth(head); 2289 2290 // Step 1: find cut point 2291 2292 // Walk up dominators to loop head looking for first loop exit 2293 // which is executed on every path thru loop. 2294 IfNode *peel_if = NULL; 2295 IfNode *peel_if_cmpu = NULL; 2296 2297 Node *iff = loop->tail(); 2298 while( iff != head ) { 2299 if( iff->is_If() ) { 2300 Node *ctrl = get_ctrl(iff->in(1)); 2301 if (ctrl->is_top()) return false; // Dead test on live IF. 2302 // If loop-varying exit-test, check for induction variable 2303 if( loop->is_member(get_loop(ctrl)) && 2304 loop->is_loop_exit(iff) && 2305 is_possible_iv_test(iff)) { 2306 Node* cmp = iff->in(1)->in(1); 2307 if (cmp->Opcode() == Op_CmpI) { 2308 peel_if = iff->as_If(); 2309 } else { 2310 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU"); 2311 peel_if_cmpu = iff->as_If(); 2312 } 2313 } 2314 } 2315 iff = idom(iff); 2316 } 2317 // Prefer signed compare over unsigned compare. 2318 IfNode* new_peel_if = NULL; 2319 if (peel_if == NULL) { 2320 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) { 2321 return false; // No peel point found 2322 } 2323 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop); 2324 if (new_peel_if == NULL) { 2325 return false; // No peel point found 2326 } 2327 peel_if = new_peel_if; 2328 } 2329 Node* last_peel = stay_in_loop(peel_if, loop); 2330 Node* first_not_peeled = stay_in_loop(last_peel, loop); 2331 if (first_not_peeled == NULL || first_not_peeled == head) { 2332 return false; 2333 } 2334 2335#if !defined(PRODUCT) 2336 if (TracePartialPeeling) { 2337 tty->print_cr("before partial peel one iteration"); 2338 Node_List wl; 2339 Node* t = head->in(2); 2340 while (true) { 2341 wl.push(t); 2342 if (t == head) break; 2343 t = idom(t); 2344 } 2345 while (wl.size() > 0) { 2346 Node* tt = wl.pop(); 2347 tt->dump(); 2348 if (tt == last_peel) tty->print_cr("-- cut --"); 2349 } 2350 } 2351#endif 2352 ResourceArea *area = Thread::current()->resource_area(); 2353 VectorSet peel(area); 2354 VectorSet not_peel(area); 2355 Node_List peel_list(area); 2356 Node_List worklist(area); 2357 Node_List sink_list(area); 2358 2359 // Set of cfg nodes to peel are those that are executable from 2360 // the head through last_peel. 2361 assert(worklist.size() == 0, "should be empty"); 2362 worklist.push(head); 2363 peel.set(head->_idx); 2364 while (worklist.size() > 0) { 2365 Node *n = worklist.pop(); 2366 if (n != last_peel) { 2367 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2368 Node* use = n->fast_out(j); 2369 if (use->is_CFG() && 2370 loop->is_member(get_loop(use)) && 2371 !peel.test_set(use->_idx)) { 2372 worklist.push(use); 2373 } 2374 } 2375 } 2376 } 2377 2378 // Set of non-cfg nodes to peel are those that are control 2379 // dependent on the cfg nodes. 2380 uint i; 2381 for(i = 0; i < loop->_body.size(); i++ ) { 2382 Node *n = loop->_body.at(i); 2383 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n; 2384 if (peel.test(n_c->_idx)) { 2385 peel.set(n->_idx); 2386 } else { 2387 not_peel.set(n->_idx); 2388 } 2389 } 2390 2391 // Step 2: move operations from the peeled section down into the 2392 // not-peeled section 2393 2394 // Get a post order schedule of nodes in the peel region 2395 // Result in right-most operand. 2396 scheduled_nodelist(loop, peel, peel_list ); 2397 2398 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 2399 2400 // For future check for too many new phis 2401 uint old_phi_cnt = 0; 2402 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 2403 Node* use = head->fast_out(j); 2404 if (use->is_Phi()) old_phi_cnt++; 2405 } 2406 2407#if !defined(PRODUCT) 2408 if (TracePartialPeeling) { 2409 tty->print_cr("\npeeled list"); 2410 } 2411#endif 2412 2413 // Evacuate nodes in peel region into the not_peeled region if possible 2414 uint new_phi_cnt = 0; 2415 for (i = 0; i < peel_list.size();) { 2416 Node* n = peel_list.at(i); 2417#if !defined(PRODUCT) 2418 if (TracePartialPeeling) n->dump(); 2419#endif 2420 bool incr = true; 2421 if ( !n->is_CFG() ) { 2422 2423 if ( has_use_in_set(n, not_peel) ) { 2424 2425 // If not used internal to the peeled region, 2426 // move "n" from peeled to not_peeled region. 2427 2428 if ( !has_use_internal_to_set(n, peel, loop) ) { 2429 2430 // if not pinned and not a load (which maybe anti-dependent on a store) 2431 // and not a CMove (Matcher expects only bool->cmove). 2432 if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) { 2433 clone_for_use_outside_loop( loop, n, worklist ); 2434 2435 sink_list.push(n); 2436 peel >>= n->_idx; // delete n from peel set. 2437 not_peel <<= n->_idx; // add n to not_peel set. 2438 peel_list.remove(i); 2439 incr = false; 2440#if !defined(PRODUCT) 2441 if (TracePartialPeeling) { 2442 tty->print_cr("sink to not_peeled region: %d newbb: %d", 2443 n->_idx, get_ctrl(n)->_idx); 2444 } 2445#endif 2446 } 2447 } else { 2448 // Otherwise check for special def-use cases that span 2449 // the peel/not_peel boundary such as bool->if 2450 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist ); 2451 new_phi_cnt++; 2452 } 2453 } 2454 } 2455 if (incr) i++; 2456 } 2457 2458 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) { 2459#if !defined(PRODUCT) 2460 if (TracePartialPeeling) { 2461 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c", 2462 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F'); 2463 } 2464#endif 2465 if (new_peel_if != NULL) { 2466 remove_cmpi_loop_exit(new_peel_if, loop); 2467 } 2468 // Inhibit more partial peeling on this loop 2469 assert(!head->is_partial_peel_loop(), "not partial peeled"); 2470 head->mark_partial_peel_failed(); 2471 return false; 2472 } 2473 2474 // Step 3: clone loop, retarget control, and insert new phis 2475 2476 // Create new loop head for new phis and to hang 2477 // the nodes being moved (sinked) from the peel region. 2478 LoopNode* new_head = new (C, 3) LoopNode(last_peel, last_peel); 2479 _igvn.register_new_node_with_optimizer(new_head); 2480 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled"); 2481 first_not_peeled->set_req(0, new_head); 2482 set_loop(new_head, loop); 2483 loop->_body.push(new_head); 2484 not_peel.set(new_head->_idx); 2485 set_idom(new_head, last_peel, dom_depth(first_not_peeled)); 2486 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled)); 2487 2488 while (sink_list.size() > 0) { 2489 Node* n = sink_list.pop(); 2490 set_ctrl(n, new_head); 2491 } 2492 2493 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 2494 2495 clone_loop( loop, old_new, dd ); 2496 2497 const uint clone_exit_idx = 1; 2498 const uint orig_exit_idx = 2; 2499 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop"); 2500 2501 Node* head_clone = old_new[head->_idx]; 2502 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop(); 2503 Node* orig_tail_clone = head_clone->in(2); 2504 2505 // Add phi if "def" node is in peel set and "use" is not 2506 2507 for(i = 0; i < peel_list.size(); i++ ) { 2508 Node *def = peel_list.at(i); 2509 if (!def->is_CFG()) { 2510 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 2511 Node *use = def->fast_out(j); 2512 if (has_node(use) && use->in(0) != C->top() && 2513 (!peel.test(use->_idx) || 2514 (use->is_Phi() && use->in(0) == head)) ) { 2515 worklist.push(use); 2516 } 2517 } 2518 while( worklist.size() ) { 2519 Node *use = worklist.pop(); 2520 for (uint j = 1; j < use->req(); j++) { 2521 Node* n = use->in(j); 2522 if (n == def) { 2523 2524 // "def" is in peel set, "use" is not in peel set 2525 // or "use" is in the entry boundary (a phi) of the peel set 2526 2527 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use; 2528 2529 if ( loop->is_member(get_loop( use_c )) ) { 2530 // use is in loop 2531 if (old_new[use->_idx] != NULL) { // null for dead code 2532 Node* use_clone = old_new[use->_idx]; 2533 _igvn.hash_delete(use); 2534 use->set_req(j, C->top()); 2535 _igvn._worklist.push(use); 2536 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone ); 2537 } 2538 } else { 2539 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format"); 2540 // use is not in the loop, check if the live range includes the cut 2541 Node* lp_if = use_c->in(orig_exit_idx)->in(0); 2542 if (not_peel.test(lp_if->_idx)) { 2543 assert(j == orig_exit_idx, "use from original loop"); 2544 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone ); 2545 } 2546 } 2547 } 2548 } 2549 } 2550 } 2551 } 2552 2553 // Step 3b: retarget control 2554 2555 // Redirect control to the new loop head if a cloned node in 2556 // the not_peeled region has control that points into the peeled region. 2557 // This necessary because the cloned peeled region will be outside 2558 // the loop. 2559 // from to 2560 // cloned-peeled <---+ 2561 // new_head_clone: | <--+ 2562 // cloned-not_peeled in(0) in(0) 2563 // orig-peeled 2564 2565 for(i = 0; i < loop->_body.size(); i++ ) { 2566 Node *n = loop->_body.at(i); 2567 if (!n->is_CFG() && n->in(0) != NULL && 2568 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) { 2569 Node* n_clone = old_new[n->_idx]; 2570 _igvn.hash_delete(n_clone); 2571 n_clone->set_req(0, new_head_clone); 2572 _igvn._worklist.push(n_clone); 2573 } 2574 } 2575 2576 // Backedge of the surviving new_head (the clone) is original last_peel 2577 _igvn.hash_delete(new_head_clone); 2578 new_head_clone->set_req(LoopNode::LoopBackControl, last_peel); 2579 _igvn._worklist.push(new_head_clone); 2580 2581 // Cut first node in original not_peel set 2582 _igvn.hash_delete(new_head); 2583 new_head->set_req(LoopNode::EntryControl, C->top()); 2584 new_head->set_req(LoopNode::LoopBackControl, C->top()); 2585 _igvn._worklist.push(new_head); 2586 2587 // Copy head_clone back-branch info to original head 2588 // and remove original head's loop entry and 2589 // clone head's back-branch 2590 _igvn.hash_delete(head); 2591 _igvn.hash_delete(head_clone); 2592 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl)); 2593 head->set_req(LoopNode::LoopBackControl, C->top()); 2594 head_clone->set_req(LoopNode::LoopBackControl, C->top()); 2595 _igvn._worklist.push(head); 2596 _igvn._worklist.push(head_clone); 2597 2598 // Similarly modify the phis 2599 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) { 2600 Node* use = head->fast_out(k); 2601 if (use->is_Phi() && use->outcnt() > 0) { 2602 Node* use_clone = old_new[use->_idx]; 2603 _igvn.hash_delete(use); 2604 _igvn.hash_delete(use_clone); 2605 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl)); 2606 use->set_req(LoopNode::LoopBackControl, C->top()); 2607 use_clone->set_req(LoopNode::LoopBackControl, C->top()); 2608 _igvn._worklist.push(use); 2609 _igvn._worklist.push(use_clone); 2610 } 2611 } 2612 2613 // Step 4: update dominator tree and dominator depth 2614 2615 set_idom(head, orig_tail_clone, dd); 2616 recompute_dom_depth(); 2617 2618 // Inhibit more partial peeling on this loop 2619 new_head_clone->set_partial_peel_loop(); 2620 C->set_major_progress(); 2621 2622#if !defined(PRODUCT) 2623 if (TracePartialPeeling) { 2624 tty->print_cr("\nafter partial peel one iteration"); 2625 Node_List wl(area); 2626 Node* t = last_peel; 2627 while (true) { 2628 wl.push(t); 2629 if (t == head_clone) break; 2630 t = idom(t); 2631 } 2632 while (wl.size() > 0) { 2633 Node* tt = wl.pop(); 2634 if (tt == head) tty->print_cr("orig head"); 2635 else if (tt == new_head_clone) tty->print_cr("new head"); 2636 else if (tt == head_clone) tty->print_cr("clone head"); 2637 tt->dump(); 2638 } 2639 } 2640#endif 2641 return true; 2642} 2643 2644//------------------------------reorg_offsets---------------------------------- 2645// Reorganize offset computations to lower register pressure. Mostly 2646// prevent loop-fallout uses of the pre-incremented trip counter (which are 2647// then alive with the post-incremented trip counter forcing an extra 2648// register move) 2649void PhaseIdealLoop::reorg_offsets( IdealLoopTree *loop ) { 2650 2651 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 2652 CountedLoopEndNode *cle = cl->loopexit(); 2653 if( !cle ) return; // The occasional dead loop 2654 // Find loop exit control 2655 Node *exit = cle->proj_out(false); 2656 assert( exit->Opcode() == Op_IfFalse, "" ); 2657 2658 // Check for the special case of folks using the pre-incremented 2659 // trip-counter on the fall-out path (forces the pre-incremented 2660 // and post-incremented trip counter to be live at the same time). 2661 // Fix this by adjusting to use the post-increment trip counter. 2662 Node *phi = cl->phi(); 2663 if( !phi ) return; // Dead infinite loop 2664 bool progress = true; 2665 while (progress) { 2666 progress = false; 2667 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) { 2668 Node* use = phi->fast_out(i); // User of trip-counter 2669 if (!has_ctrl(use)) continue; 2670 Node *u_ctrl = get_ctrl(use); 2671 if( use->is_Phi() ) { 2672 u_ctrl = NULL; 2673 for( uint j = 1; j < use->req(); j++ ) 2674 if( use->in(j) == phi ) 2675 u_ctrl = dom_lca( u_ctrl, use->in(0)->in(j) ); 2676 } 2677 IdealLoopTree *u_loop = get_loop(u_ctrl); 2678 // Look for loop-invariant use 2679 if( u_loop == loop ) continue; 2680 if( loop->is_member( u_loop ) ) continue; 2681 // Check that use is live out the bottom. Assuming the trip-counter 2682 // update is right at the bottom, uses of of the loop middle are ok. 2683 if( dom_lca( exit, u_ctrl ) != exit ) continue; 2684 // protect against stride not being a constant 2685 if( !cle->stride_is_con() ) continue; 2686 // Hit! Refactor use to use the post-incremented tripcounter. 2687 // Compute a post-increment tripcounter. 2688 Node *opaq = new (C, 2) Opaque2Node( cle->incr() ); 2689 register_new_node( opaq, u_ctrl ); 2690 Node *neg_stride = _igvn.intcon(-cle->stride_con()); 2691 set_ctrl(neg_stride, C->root()); 2692 Node *post = new (C, 3) AddINode( opaq, neg_stride); 2693 register_new_node( post, u_ctrl ); 2694 _igvn.hash_delete(use); 2695 _igvn._worklist.push(use); 2696 for( uint j = 1; j < use->req(); j++ ) 2697 if( use->in(j) == phi ) 2698 use->set_req(j, post); 2699 // Since DU info changed, rerun loop 2700 progress = true; 2701 break; 2702 } 2703 } 2704 2705} 2706