1 #ifdef USE_PRAGMA_IDENT_SRC 2 #pragma ident "@(#)loopopts.cpp 1.222 08/11/24 12:23:09 JVM" 3 #endif 4 /* 5 * Copyright 1999-2008 Sun Microsystems, Inc. All Rights Reserved. 6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 7 * 8 * This code is free software; you can redistribute it and/or modify it 9 * under the terms of the GNU General Public License version 2 only, as 10 * published by the Free Software Foundation. 11 * 12 * This code is distributed in the hope that it will be useful, but WITHOUT 13 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 14 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 15 * version 2 for more details (a copy is included in the LICENSE file that 16 * accompanied this code). 17 * 18 * You should have received a copy of the GNU General Public License version 19 * 2 along with this work; if not, write to the Free Software Foundation, 20 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 21 * 22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, 23 * CA 95054 USA or visit www.sun.com if you need additional information or 24 * have any questions. 25 * 26 */ 27 28 #include "incls/_precompiled.incl" 29 #include "incls/_loopopts.cpp.incl" 30 31 //============================================================================= 32 //------------------------------split_thru_phi--------------------------------- 33 // Split Node 'n' through merge point if there is enough win. 34 Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) { 35 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) { 36 // ConvI2L may have type information on it which is unsafe to push up 37 // so disable this for now 38 return NULL; 39 } 40 int wins = 0; 41 assert( !n->is_CFG(), "" ); 42 assert( region->is_Region(), "" ); 43 44 const Type* type = n->bottom_type(); 45 const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr(); 46 Node *phi; 47 if( t_oop != NULL && t_oop->is_known_instance_field() ) { 48 int iid = t_oop->instance_id(); 49 int index = C->get_alias_index(t_oop); 50 int offset = t_oop->offset(); 51 phi = new (C,region->req()) PhiNode(region, type, NULL, iid, index, offset); 52 } else { 53 phi = new (C,region->req()) PhiNode(region, type); 54 } 55 uint old_unique = C->unique(); 56 for( uint i = 1; i < region->req(); i++ ) { 57 Node *x; 58 Node* the_clone = NULL; 59 if( region->in(i) == C->top() ) { 60 x = C->top(); // Dead path? Use a dead data op 61 } else { 62 x = n->clone(); // Else clone up the data op 63 the_clone = x; // Remember for possible deletion. 64 // Alter data node to use pre-phi inputs 65 if( n->in(0) == region ) 66 x->set_req( 0, region->in(i) ); 67 for( uint j = 1; j < n->req(); j++ ) { 68 Node *in = n->in(j); 69 if( in->is_Phi() && in->in(0) == region ) 70 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone 71 } 72 } 73 // Check for a 'win' on some paths 74 const Type *t = x->Value(&_igvn); 75 76 bool singleton = t->singleton(); 77 78 // A TOP singleton indicates that there are no possible values incoming 79 // along a particular edge. In most cases, this is OK, and the Phi will 80 // be eliminated later in an Ideal call. However, we can't allow this to 81 // happen if the singleton occurs on loop entry, as the elimination of 82 // the PhiNode may cause the resulting node to migrate back to a previous 83 // loop iteration. 84 if( singleton && t == Type::TOP ) { 85 // Is_Loop() == false does not confirm the absence of a loop (e.g., an 86 // irreducible loop may not be indicated by an affirmative is_Loop()); 87 // therefore, the only top we can split thru a phi is on a backedge of 88 // a loop. 89 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 90 } 91 92 if( singleton ) { 93 wins++; 94 x = ((PhaseGVN&)_igvn).makecon(t); 95 } else { 96 // We now call Identity to try to simplify the cloned node. 97 // Note that some Identity methods call phase->type(this). 98 // Make sure that the type array is big enough for 99 // our new node, even though we may throw the node away. 100 // (Note: This tweaking with igvn only works because x is a new node.) 101 _igvn.set_type(x, t); 102 // If x is a TypeNode, capture any more-precise type permanently into Node 103 // othewise it will be not updated during igvn->transform since 104 // igvn->type(x) is set to x->Value() already. 105 x->raise_bottom_type(t); 106 Node *y = x->Identity(&_igvn); 107 if( y != x ) { 108 wins++; 109 x = y; 110 } else { 111 y = _igvn.hash_find(x); 112 if( y ) { 113 wins++; 114 x = y; 115 } else { 116 // Else x is a new node we are keeping 117 // We do not need register_new_node_with_optimizer 118 // because set_type has already been called. 119 _igvn._worklist.push(x); 120 } 121 } 122 } 123 if (x != the_clone && the_clone != NULL) 124 _igvn.remove_dead_node(the_clone); 125 phi->set_req( i, x ); 126 } 127 // Too few wins? 128 if( wins <= policy ) { 129 _igvn.remove_dead_node(phi); 130 return NULL; 131 } 132 133 // Record Phi 134 register_new_node( phi, region ); 135 136 for( uint i2 = 1; i2 < phi->req(); i2++ ) { 137 Node *x = phi->in(i2); 138 // If we commoned up the cloned 'x' with another existing Node, 139 // the existing Node picks up a new use. We need to make the 140 // existing Node occur higher up so it dominates its uses. 141 Node *old_ctrl; 142 IdealLoopTree *old_loop; 143 144 // The occasional new node 145 if( x->_idx >= old_unique ) { // Found a new, unplaced node? 146 old_ctrl = x->is_Con() ? C->root() : NULL; 147 old_loop = NULL; // Not in any prior loop 148 } else { 149 old_ctrl = x->is_Con() ? C->root() : get_ctrl(x); 150 old_loop = get_loop(old_ctrl); // Get prior loop 151 } 152 // New late point must dominate new use 153 Node *new_ctrl = dom_lca( old_ctrl, region->in(i2) ); 154 // Set new location 155 set_ctrl(x, new_ctrl); 156 IdealLoopTree *new_loop = get_loop( new_ctrl ); 157 // If changing loop bodies, see if we need to collect into new body 158 if( old_loop != new_loop ) { 159 if( old_loop && !old_loop->_child ) 160 old_loop->_body.yank(x); 161 if( !new_loop->_child ) 162 new_loop->_body.push(x); // Collect body info 163 } 164 } 165 166 return phi; 167 } 168 169 //------------------------------dominated_by------------------------------------ 170 // Replace the dominated test with an obvious true or false. Place it on the 171 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the 172 // live path up to the dominating control. 173 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff ) { 174 #ifndef PRODUCT 175 if( VerifyLoopOptimizations && PrintOpto ) tty->print_cr("dominating test"); 176 #endif 177 178 179 // prevdom is the dominating projection of the dominating test. 180 assert( iff->is_If(), "" ); 181 assert( iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added"); 182 int pop = prevdom->Opcode(); 183 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" ); 184 // 'con' is set to true or false to kill the dominated test. 185 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO); 186 set_ctrl(con, C->root()); // Constant gets a new use 187 // Hack the dominated test 188 _igvn.hash_delete(iff); 189 iff->set_req(1, con); 190 _igvn._worklist.push(iff); 191 192 // If I dont have a reachable TRUE and FALSE path following the IfNode then 193 // I can assume this path reaches an infinite loop. In this case it's not 194 // important to optimize the data Nodes - either the whole compilation will 195 // be tossed or this path (and all data Nodes) will go dead. 196 if( iff->outcnt() != 2 ) return; 197 198 // Make control-dependent data Nodes on the live path (path that will remain 199 // once the dominated IF is removed) become control-dependent on the 200 // dominating projection. 201 Node* dp = ((IfNode*)iff)->proj_out(pop == Op_IfTrue); 202 IdealLoopTree *old_loop = get_loop(dp); 203 204 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { 205 Node* cd = dp->fast_out(i); // Control-dependent node 206 if( cd->depends_only_on_test() ) { 207 assert( cd->in(0) == dp, "" ); 208 _igvn.hash_delete( cd ); 209 cd->set_req(0, prevdom); 210 set_early_ctrl( cd ); 211 _igvn._worklist.push(cd); 212 IdealLoopTree *new_loop = get_loop(get_ctrl(cd)); 213 if( old_loop != new_loop ) { 214 if( !old_loop->_child ) old_loop->_body.yank(cd); 215 if( !new_loop->_child ) new_loop->_body.push(cd); 216 } 217 --i; 218 --imax; 219 } 220 } 221 } 222 223 //------------------------------has_local_phi_input---------------------------- 224 // Return TRUE if 'n' has Phi inputs from its local block and no other 225 // block-local inputs (all non-local-phi inputs come from earlier blocks) 226 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) { 227 Node *n_ctrl = get_ctrl(n); 228 // See if some inputs come from a Phi in this block, or from before 229 // this block. 230 uint i; 231 for( i = 1; i < n->req(); i++ ) { 232 Node *phi = n->in(i); 233 if( phi->is_Phi() && phi->in(0) == n_ctrl ) 234 break; 235 } 236 if( i >= n->req() ) 237 return NULL; // No Phi inputs; nowhere to clone thru 238 239 // Check for inputs created between 'n' and the Phi input. These 240 // must split as well; they have already been given the chance 241 // (courtesy of a post-order visit) and since they did not we must 242 // recover the 'cost' of splitting them by being very profitable 243 // when splitting 'n'. Since this is unlikely we simply give up. 244 for( i = 1; i < n->req(); i++ ) { 245 Node *m = n->in(i); 246 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) { 247 // We allow the special case of AddP's with no local inputs. 248 // This allows us to split-up address expressions. 249 if (m->is_AddP() && 250 get_ctrl(m->in(2)) != n_ctrl && 251 get_ctrl(m->in(3)) != n_ctrl) { 252 // Move the AddP up to dominating point 253 set_ctrl_and_loop(m, find_non_split_ctrl(idom(n_ctrl))); 254 continue; 255 } 256 return NULL; 257 } 258 } 259 260 return n_ctrl; 261 } 262 263 //------------------------------remix_address_expressions---------------------- 264 // Rework addressing expressions to get the most loop-invariant stuff 265 // moved out. We'd like to do all associative operators, but it's especially 266 // important (common) to do address expressions. 267 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) { 268 if (!has_ctrl(n)) return NULL; 269 Node *n_ctrl = get_ctrl(n); 270 IdealLoopTree *n_loop = get_loop(n_ctrl); 271 272 // See if 'n' mixes loop-varying and loop-invariant inputs and 273 // itself is loop-varying. 274 275 // Only interested in binary ops (and AddP) 276 if( n->req() < 3 || n->req() > 4 ) return NULL; 277 278 Node *n1_ctrl = get_ctrl(n->in( 1)); 279 Node *n2_ctrl = get_ctrl(n->in( 2)); 280 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3)); 281 IdealLoopTree *n1_loop = get_loop( n1_ctrl ); 282 IdealLoopTree *n2_loop = get_loop( n2_ctrl ); 283 IdealLoopTree *n3_loop = get_loop( n3_ctrl ); 284 285 // Does one of my inputs spin in a tighter loop than self? 286 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) || 287 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) || 288 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) ) 289 return NULL; // Leave well enough alone 290 291 // Is at least one of my inputs loop-invariant? 292 if( n1_loop == n_loop && 293 n2_loop == n_loop && 294 n3_loop == n_loop ) 295 return NULL; // No loop-invariant inputs 296 297 298 int n_op = n->Opcode(); 299 300 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2). 301 if( n_op == Op_LShiftI ) { 302 // Scale is loop invariant 303 Node *scale = n->in(2); 304 Node *scale_ctrl = get_ctrl(scale); 305 IdealLoopTree *scale_loop = get_loop(scale_ctrl ); 306 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) ) 307 return NULL; 308 const TypeInt *scale_t = scale->bottom_type()->isa_int(); 309 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 ) 310 return NULL; // Dont bother with byte/short masking 311 // Add must vary with loop (else shift would be loop-invariant) 312 Node *add = n->in(1); 313 Node *add_ctrl = get_ctrl(add); 314 IdealLoopTree *add_loop = get_loop(add_ctrl); 315 //assert( n_loop == add_loop, "" ); 316 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops 317 318 // Convert I-V into I+ (0-V); same for V-I 319 if( add->Opcode() == Op_SubI && 320 _igvn.type( add->in(1) ) != TypeInt::ZERO ) { 321 Node *zero = _igvn.intcon(0); 322 set_ctrl(zero, C->root()); 323 Node *neg = new (C, 3) SubINode( _igvn.intcon(0), add->in(2) ); 324 register_new_node( neg, get_ctrl(add->in(2) ) ); 325 add = new (C, 3) AddINode( add->in(1), neg ); 326 register_new_node( add, add_ctrl ); 327 } 328 if( add->Opcode() != Op_AddI ) return NULL; 329 // See if one add input is loop invariant 330 Node *add_var = add->in(1); 331 Node *add_var_ctrl = get_ctrl(add_var); 332 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 333 Node *add_invar = add->in(2); 334 Node *add_invar_ctrl = get_ctrl(add_invar); 335 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl ); 336 if( add_var_loop == n_loop ) { 337 } else if( add_invar_loop == n_loop ) { 338 // Swap to find the invariant part 339 add_invar = add_var; 340 add_invar_ctrl = add_var_ctrl; 341 add_invar_loop = add_var_loop; 342 add_var = add->in(2); 343 Node *add_var_ctrl = get_ctrl(add_var); 344 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 345 } else // Else neither input is loop invariant 346 return NULL; 347 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) ) 348 return NULL; // No invariant part of the add? 349 350 // Yes! Reshape address expression! 351 Node *inv_scale = new (C, 3) LShiftINode( add_invar, scale ); 352 register_new_node( inv_scale, add_invar_ctrl ); 353 Node *var_scale = new (C, 3) LShiftINode( add_var, scale ); 354 register_new_node( var_scale, n_ctrl ); 355 Node *var_add = new (C, 3) AddINode( var_scale, inv_scale ); 356 register_new_node( var_add, n_ctrl ); 357 _igvn.hash_delete( n ); 358 _igvn.subsume_node( n, var_add ); 359 return var_add; 360 } 361 362 // Replace (I+V) with (V+I) 363 if( n_op == Op_AddI || 364 n_op == Op_AddL || 365 n_op == Op_AddF || 366 n_op == Op_AddD || 367 n_op == Op_MulI || 368 n_op == Op_MulL || 369 n_op == Op_MulF || 370 n_op == Op_MulD ) { 371 if( n2_loop == n_loop ) { 372 assert( n1_loop != n_loop, "" ); 373 n->swap_edges(1, 2); 374 } 375 } 376 377 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V), 378 // but not if I2 is a constant. 379 if( n_op == Op_AddP ) { 380 if( n2_loop == n_loop && n3_loop != n_loop ) { 381 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) { 382 Node *n22_ctrl = get_ctrl(n->in(2)->in(2)); 383 Node *n23_ctrl = get_ctrl(n->in(2)->in(3)); 384 IdealLoopTree *n22loop = get_loop( n22_ctrl ); 385 IdealLoopTree *n23_loop = get_loop( n23_ctrl ); 386 if( n22loop != n_loop && n22loop->is_member(n_loop) && 387 n23_loop == n_loop ) { 388 Node *add1 = new (C, 4) AddPNode( n->in(1), n->in(2)->in(2), n->in(3) ); 389 // Stuff new AddP in the loop preheader 390 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) ); 391 Node *add2 = new (C, 4) AddPNode( n->in(1), add1, n->in(2)->in(3) ); 392 register_new_node( add2, n_ctrl ); 393 _igvn.hash_delete( n ); 394 _igvn.subsume_node( n, add2 ); 395 return add2; 396 } 397 } 398 } 399 400 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V) 401 if( n2_loop != n_loop && n3_loop == n_loop ) { 402 if( n->in(3)->Opcode() == Op_AddI ) { 403 Node *V = n->in(3)->in(1); 404 Node *I = n->in(3)->in(2); 405 if( is_member(n_loop,get_ctrl(V)) ) { 406 } else { 407 Node *tmp = V; V = I; I = tmp; 408 } 409 if( !is_member(n_loop,get_ctrl(I)) ) { 410 Node *add1 = new (C, 4) AddPNode( n->in(1), n->in(2), I ); 411 // Stuff new AddP in the loop preheader 412 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) ); 413 Node *add2 = new (C, 4) AddPNode( n->in(1), add1, V ); 414 register_new_node( add2, n_ctrl ); 415 _igvn.hash_delete( n ); 416 _igvn.subsume_node( n, add2 ); 417 return add2; 418 } 419 } 420 } 421 } 422 423 return NULL; 424 } 425 426 //------------------------------conditional_move------------------------------- 427 // Attempt to replace a Phi with a conditional move. We have some pretty 428 // strict profitability requirements. All Phis at the merge point must 429 // be converted, so we can remove the control flow. We need to limit the 430 // number of c-moves to a small handful. All code that was in the side-arms 431 // of the CFG diamond is now speculatively executed. This code has to be 432 // "cheap enough". We are pretty much limited to CFG diamonds that merge 433 // 1 or 2 items with a total of 1 or 2 ops executed speculatively. 434 Node *PhaseIdealLoop::conditional_move( Node *region ) { 435 436 assert( region->is_Region(), "sanity check" ); 437 if( region->req() != 3 ) return NULL; 438 439 // Check for CFG diamond 440 Node *lp = region->in(1); 441 Node *rp = region->in(2); 442 if( !lp || !rp ) return NULL; 443 Node *lp_c = lp->in(0); 444 if( lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If() ) return NULL; 445 IfNode *iff = lp_c->as_If(); 446 447 // Check for highly predictable branch. No point in CMOV'ing if 448 // we are going to predict accurately all the time. 449 // %%% This hides patterns produced by utility methods like Math.min. 450 if( iff->_prob < PROB_UNLIKELY_MAG(3) || 451 iff->_prob > PROB_LIKELY_MAG(3) ) 452 return NULL; 453 454 // Check for ops pinned in an arm of the diamond. 455 // Can't remove the control flow in this case 456 if( lp->outcnt() > 1 ) return NULL; 457 if( rp->outcnt() > 1 ) return NULL; 458 459 // Check profitability 460 int cost = 0; 461 int phis = 0; 462 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 463 Node *out = region->fast_out(i); 464 if( !out->is_Phi() ) continue; // Ignore other control edges, etc 465 phis++; 466 PhiNode* phi = out->as_Phi(); 467 switch (phi->type()->basic_type()) { 468 case T_LONG: 469 cost++; // Probably encodes as 2 CMOV's 470 case T_INT: // These all CMOV fine 471 case T_FLOAT: 472 case T_DOUBLE: 473 case T_ADDRESS: // (RawPtr) 474 cost++; 475 break; 476 case T_NARROWOOP: // Fall through 477 case T_OBJECT: { // Base oops are OK, but not derived oops 478 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr(); 479 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a 480 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus 481 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we 482 // have a Phi for the base here that we convert to a CMOVE all is well 483 // and good. But if the base is dead, we'll not make a CMOVE. Later 484 // the allocator will have to produce a base by creating a CMOVE of the 485 // relevant bases. This puts the allocator in the business of 486 // manufacturing expensive instructions, generally a bad plan. 487 // Just Say No to Conditionally-Moved Derived Pointers. 488 if( tp && tp->offset() != 0 ) 489 return NULL; 490 cost++; 491 break; 492 } 493 default: 494 return NULL; // In particular, can't do memory or I/O 495 } 496 // Add in cost any speculative ops 497 for( uint j = 1; j < region->req(); j++ ) { 498 Node *proj = region->in(j); 499 Node *inp = phi->in(j); 500 if (get_ctrl(inp) == proj) { // Found local op 501 cost++; 502 // Check for a chain of dependent ops; these will all become 503 // speculative in a CMOV. 504 for( uint k = 1; k < inp->req(); k++ ) 505 if (get_ctrl(inp->in(k)) == proj) 506 return NULL; // Too much speculative goo 507 } 508 } 509 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode. 510 // This will likely Split-If, a higher-payoff operation. 511 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) { 512 Node* use = phi->fast_out(k); 513 if( use->is_Cmp() || use->is_DecodeN() || use->is_EncodeP() ) 514 return NULL; 515 } 516 } 517 if( cost >= ConditionalMoveLimit ) return NULL; // Too much goo 518 Node* bol = iff->in(1); 519 assert( bol->Opcode() == Op_Bool, "" ); 520 int cmp_op = bol->in(1)->Opcode(); 521 // It is expensive to generate flags from a float compare. 522 // Avoid duplicated float compare. 523 if( phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL; 524 525 // -------------- 526 // Now replace all Phis with CMOV's 527 Node *cmov_ctrl = iff->in(0); 528 uint flip = (lp->Opcode() == Op_IfTrue); 529 while( 1 ) { 530 PhiNode* phi = NULL; 531 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 532 Node *out = region->fast_out(i); 533 if (out->is_Phi()) { 534 phi = out->as_Phi(); 535 break; 536 } 537 } 538 if (phi == NULL) break; 539 #ifndef PRODUCT 540 if( PrintOpto && VerifyLoopOptimizations ) tty->print_cr("CMOV"); 541 #endif 542 // Move speculative ops 543 for( uint j = 1; j < region->req(); j++ ) { 544 Node *proj = region->in(j); 545 Node *inp = phi->in(j); 546 if (get_ctrl(inp) == proj) { // Found local op 547 #ifndef PRODUCT 548 if( PrintOpto && VerifyLoopOptimizations ) { 549 tty->print(" speculate: "); 550 inp->dump(); 551 } 552 #endif 553 set_ctrl(inp, cmov_ctrl); 554 } 555 } 556 Node *cmov = CMoveNode::make( C, cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi) ); 557 register_new_node( cmov, cmov_ctrl ); 558 _igvn.hash_delete(phi); 559 _igvn.subsume_node( phi, cmov ); 560 #ifndef PRODUCT 561 if( VerifyLoopOptimizations ) verify(); 562 #endif 563 } 564 565 // The useless CFG diamond will fold up later; see the optimization in 566 // RegionNode::Ideal. 567 _igvn._worklist.push(region); 568 569 return iff->in(1); 570 } 571 572 //------------------------------split_if_with_blocks_pre----------------------- 573 // Do the real work in a non-recursive function. Data nodes want to be 574 // cloned in the pre-order so they can feed each other nicely. 575 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) { 576 // Cloning these guys is unlikely to win 577 int n_op = n->Opcode(); 578 if( n_op == Op_MergeMem ) return n; 579 if( n->is_Proj() ) return n; 580 // Do not clone-up CmpFXXX variations, as these are always 581 // followed by a CmpI 582 if( n->is_Cmp() ) return n; 583 // Attempt to use a conditional move instead of a phi/branch 584 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) { 585 Node *cmov = conditional_move( n ); 586 if( cmov ) return cmov; 587 } 588 if( n->is_CFG() || n->is_LoadStore() ) 589 return n; 590 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd 591 n_op == Op_Opaque2 ) { 592 if( !C->major_progress() ) // If chance of no more loop opts... 593 _igvn._worklist.push(n); // maybe we'll remove them 594 return n; 595 } 596 597 if( n->is_Con() ) return n; // No cloning for Con nodes 598 599 Node *n_ctrl = get_ctrl(n); 600 if( !n_ctrl ) return n; // Dead node 601 602 // Attempt to remix address expressions for loop invariants 603 Node *m = remix_address_expressions( n ); 604 if( m ) return m; 605 606 // Determine if the Node has inputs from some local Phi. 607 // Returns the block to clone thru. 608 Node *n_blk = has_local_phi_input( n ); 609 if( !n_blk ) return n; 610 // Do not clone the trip counter through on a CountedLoop 611 // (messes up the canonical shape). 612 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n; 613 614 // Check for having no control input; not pinned. Allow 615 // dominating control. 616 if( n->in(0) ) { 617 Node *dom = idom(n_blk); 618 if( dom_lca( n->in(0), dom ) != n->in(0) ) 619 return n; 620 } 621 // Policy: when is it profitable. You must get more wins than 622 // policy before it is considered profitable. Policy is usually 0, 623 // so 1 win is considered profitable. Big merges will require big 624 // cloning, so get a larger policy. 625 int policy = n_blk->req() >> 2; 626 627 // If the loop is a candidate for range check elimination, 628 // delay splitting through it's phi until a later loop optimization 629 if (n_blk->is_CountedLoop()) { 630 IdealLoopTree *lp = get_loop(n_blk); 631 if (lp && lp->_rce_candidate) { 632 return n; 633 } 634 } 635 636 // Use same limit as split_if_with_blocks_post 637 if( C->unique() > 35000 ) return n; // Method too big 638 639 // Split 'n' through the merge point if it is profitable 640 Node *phi = split_thru_phi( n, n_blk, policy ); 641 if( !phi ) return n; 642 643 // Found a Phi to split thru! 644 // Replace 'n' with the new phi 645 _igvn.hash_delete(n); 646 _igvn.subsume_node( n, phi ); 647 // Moved a load around the loop, 'en-registering' something. 648 if( n_blk->Opcode() == Op_Loop && n->is_Load() && 649 !phi->in(LoopNode::LoopBackControl)->is_Load() ) 650 C->set_major_progress(); 651 652 return phi; 653 } 654 655 static bool merge_point_too_heavy(Compile* C, Node* region) { 656 // Bail out if the region and its phis have too many users. 657 int weight = 0; 658 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 659 weight += region->fast_out(i)->outcnt(); 660 } 661 int nodes_left = MaxNodeLimit - C->unique(); 662 if (weight * 8 > nodes_left) { 663 #ifndef PRODUCT 664 if (PrintOpto) 665 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight); 666 #endif 667 return true; 668 } else { 669 return false; 670 } 671 } 672 673 #ifdef _LP64 674 static bool merge_point_safe(Node* region) { 675 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode 676 // having a PhiNode input. This sidesteps the dangerous case where the split 677 // ConvI2LNode may become TOP if the input Value() does not 678 // overlap the ConvI2L range, leaving a node which may not dominate its 679 // uses. 680 // A better fix for this problem can be found in the BugTraq entry, but 681 // expediency for Mantis demands this hack. 682 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 683 Node* n = region->fast_out(i); 684 if (n->is_Phi()) { 685 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 686 Node* m = n->fast_out(j); 687 if (m->Opcode() == Op_ConvI2L) { 688 return false; 689 } 690 } 691 } 692 } 693 return true; 694 } 695 #endif 696 697 698 //------------------------------place_near_use--------------------------------- 699 // Place some computation next to use but not inside inner loops. 700 // For inner loop uses move it to the preheader area. 701 Node *PhaseIdealLoop::place_near_use( Node *useblock ) const { 702 IdealLoopTree *u_loop = get_loop( useblock ); 703 return (u_loop->_irreducible || u_loop->_child) 704 ? useblock 705 : u_loop->_head->in(LoopNode::EntryControl); 706 } 707 708 709 //------------------------------split_if_with_blocks_post---------------------- 710 // Do the real work in a non-recursive function. CFG hackery wants to be 711 // in the post-order, so it can dirty the I-DOM info and not use the dirtied 712 // info. 713 void PhaseIdealLoop::split_if_with_blocks_post( Node *n ) { 714 715 // Cloning Cmp through Phi's involves the split-if transform. 716 // FastLock is not used by an If 717 if( n->is_Cmp() && !n->is_FastLock() ) { 718 if( C->unique() > 35000 ) return; // Method too big 719 720 // Do not do 'split-if' if irreducible loops are present. 721 if( _has_irreducible_loops ) 722 return; 723 724 Node *n_ctrl = get_ctrl(n); 725 // Determine if the Node has inputs from some local Phi. 726 // Returns the block to clone thru. 727 Node *n_blk = has_local_phi_input( n ); 728 if( n_blk != n_ctrl ) return; 729 730 if( merge_point_too_heavy(C, n_ctrl) ) 731 return; 732 733 if( n->outcnt() != 1 ) return; // Multiple bool's from 1 compare? 734 Node *bol = n->unique_out(); 735 assert( bol->is_Bool(), "expect a bool here" ); 736 if( bol->outcnt() != 1 ) return;// Multiple branches from 1 compare? 737 Node *iff = bol->unique_out(); 738 739 // Check some safety conditions 740 if( iff->is_If() ) { // Classic split-if? 741 if( iff->in(0) != n_ctrl ) return; // Compare must be in same blk as if 742 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE 743 if( get_ctrl(iff->in(2)) == n_ctrl || 744 get_ctrl(iff->in(3)) == n_ctrl ) 745 return; // Inputs not yet split-up 746 if ( get_loop(n_ctrl) != get_loop(get_ctrl(iff)) ) { 747 return; // Loop-invar test gates loop-varying CMOVE 748 } 749 } else { 750 return; // some other kind of node, such as an Allocate 751 } 752 753 // Do not do 'split-if' if some paths are dead. First do dead code 754 // elimination and then see if its still profitable. 755 for( uint i = 1; i < n_ctrl->req(); i++ ) 756 if( n_ctrl->in(i) == C->top() ) 757 return; 758 759 // When is split-if profitable? Every 'win' on means some control flow 760 // goes dead, so it's almost always a win. 761 int policy = 0; 762 // If trying to do a 'Split-If' at the loop head, it is only 763 // profitable if the cmp folds up on BOTH paths. Otherwise we 764 // risk peeling a loop forever. 765 766 // CNC - Disabled for now. Requires careful handling of loop 767 // body selection for the cloned code. Also, make sure we check 768 // for any input path not being in the same loop as n_ctrl. For 769 // irreducible loops we cannot check for 'n_ctrl->is_Loop()' 770 // because the alternative loop entry points won't be converted 771 // into LoopNodes. 772 IdealLoopTree *n_loop = get_loop(n_ctrl); 773 for( uint j = 1; j < n_ctrl->req(); j++ ) 774 if( get_loop(n_ctrl->in(j)) != n_loop ) 775 return; 776 777 #ifdef _LP64 778 // Check for safety of the merge point. 779 if( !merge_point_safe(n_ctrl) ) { 780 return; 781 } 782 #endif 783 784 // Split compare 'n' through the merge point if it is profitable 785 Node *phi = split_thru_phi( n, n_ctrl, policy ); 786 if( !phi ) return; 787 788 // Found a Phi to split thru! 789 // Replace 'n' with the new phi 790 _igvn.hash_delete(n); 791 _igvn.subsume_node( n, phi ); 792 793 // Now split the bool up thru the phi 794 Node *bolphi = split_thru_phi( bol, n_ctrl, -1 ); 795 _igvn.hash_delete(bol); 796 _igvn.subsume_node( bol, bolphi ); 797 assert( iff->in(1) == bolphi, "" ); 798 if( bolphi->Value(&_igvn)->singleton() ) 799 return; 800 801 // Conditional-move? Must split up now 802 if( !iff->is_If() ) { 803 Node *cmovphi = split_thru_phi( iff, n_ctrl, -1 ); 804 _igvn.hash_delete(iff); 805 _igvn.subsume_node( iff, cmovphi ); 806 return; 807 } 808 809 // Now split the IF 810 do_split_if( iff ); 811 return; 812 } 813 814 // Check for an IF ready to split; one that has its 815 // condition codes input coming from a Phi at the block start. 816 int n_op = n->Opcode(); 817 818 // Check for an IF being dominated by another IF same test 819 if( n_op == Op_If ) { 820 Node *bol = n->in(1); 821 uint max = bol->outcnt(); 822 // Check for same test used more than once? 823 if( n_op == Op_If && max > 1 && bol->is_Bool() ) { 824 // Search up IDOMs to see if this IF is dominated. 825 Node *cutoff = get_ctrl(bol); 826 827 // Now search up IDOMs till cutoff, looking for a dominating test 828 Node *prevdom = n; 829 Node *dom = idom(prevdom); 830 while( dom != cutoff ) { 831 if( dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom ) { 832 // Replace the dominated test with an obvious true or false. 833 // Place it on the IGVN worklist for later cleanup. 834 C->set_major_progress(); 835 dominated_by( prevdom, n ); 836 #ifndef PRODUCT 837 if( VerifyLoopOptimizations ) verify(); 838 #endif 839 return; 840 } 841 prevdom = dom; 842 dom = idom(prevdom); 843 } 844 } 845 } 846 847 // See if a shared loop-varying computation has no loop-varying uses. 848 // Happens if something is only used for JVM state in uncommon trap exits, 849 // like various versions of induction variable+offset. Clone the 850 // computation per usage to allow it to sink out of the loop. 851 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about) 852 Node *n_ctrl = get_ctrl(n); 853 IdealLoopTree *n_loop = get_loop(n_ctrl); 854 if( n_loop != _ltree_root ) { 855 DUIterator_Fast imax, i = n->fast_outs(imax); 856 for (; i < imax; i++) { 857 Node* u = n->fast_out(i); 858 if( !has_ctrl(u) ) break; // Found control user 859 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 860 if( u_loop == n_loop ) break; // Found loop-varying use 861 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop 862 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003 863 } 864 bool did_break = (i < imax); // Did we break out of the previous loop? 865 if (!did_break && n->outcnt() > 1) { // All uses in outer loops! 866 Node *late_load_ctrl; 867 if (n->is_Load()) { 868 // If n is a load, get and save the result from get_late_ctrl(), 869 // to be later used in calculating the control for n's clones. 870 clear_dom_lca_tags(); 871 late_load_ctrl = get_late_ctrl(n, n_ctrl); 872 } 873 // If n is a load, and the late control is the same as the current 874 // control, then the cloning of n is a pointless exercise, because 875 // GVN will ensure that we end up where we started. 876 if (!n->is_Load() || late_load_ctrl != n_ctrl) { 877 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) { 878 Node *u = n->last_out(j); // Clone private computation per use 879 _igvn.hash_delete(u); 880 _igvn._worklist.push(u); 881 Node *x = n->clone(); // Clone computation 882 Node *x_ctrl = NULL; 883 if( u->is_Phi() ) { 884 // Replace all uses of normal nodes. Replace Phi uses 885 // individually, so the seperate Nodes can sink down 886 // different paths. 887 uint k = 1; 888 while( u->in(k) != n ) k++; 889 u->set_req( k, x ); 890 // x goes next to Phi input path 891 x_ctrl = u->in(0)->in(k); 892 --j; 893 } else { // Normal use 894 // Replace all uses 895 for( uint k = 0; k < u->req(); k++ ) { 896 if( u->in(k) == n ) { 897 u->set_req( k, x ); 898 --j; 899 } 900 } 901 x_ctrl = get_ctrl(u); 902 } 903 904 // Find control for 'x' next to use but not inside inner loops. 905 // For inner loop uses get the preheader area. 906 x_ctrl = place_near_use(x_ctrl); 907 908 if (n->is_Load()) { 909 // For loads, add a control edge to a CFG node outside of the loop 910 // to force them to not combine and return back inside the loop 911 // during GVN optimization (4641526). 912 // 913 // Because we are setting the actual control input, factor in 914 // the result from get_late_ctrl() so we respect any 915 // anti-dependences. (6233005). 916 x_ctrl = dom_lca(late_load_ctrl, x_ctrl); 917 918 // Don't allow the control input to be a CFG splitting node. 919 // Such nodes should only have ProjNodes as outs, e.g. IfNode 920 // should only have IfTrueNode and IfFalseNode (4985384). 921 x_ctrl = find_non_split_ctrl(x_ctrl); 922 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone"); 923 924 x->set_req(0, x_ctrl); 925 } 926 register_new_node(x, x_ctrl); 927 928 // Some institutional knowledge is needed here: 'x' is 929 // yanked because if the optimizer runs GVN on it all the 930 // cloned x's will common up and undo this optimization and 931 // be forced back in the loop. This is annoying because it 932 // makes +VerifyOpto report false-positives on progress. I 933 // tried setting control edges on the x's to force them to 934 // not combine, but the matching gets worried when it tries 935 // to fold a StoreP and an AddP together (as part of an 936 // address expression) and the AddP and StoreP have 937 // different controls. 938 if( !x->is_Load() && !x->is_DecodeN() ) _igvn._worklist.yank(x); 939 } 940 _igvn.remove_dead_node(n); 941 } 942 } 943 } 944 } 945 946 // Check for Opaque2's who's loop has disappeared - who's input is in the 947 // same loop nest as their output. Remove 'em, they are no longer useful. 948 if( n_op == Op_Opaque2 && 949 n->in(1) != NULL && 950 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) { 951 _igvn.add_users_to_worklist(n); 952 _igvn.hash_delete(n); 953 _igvn.subsume_node( n, n->in(1) ); 954 } 955 } 956 957 //------------------------------split_if_with_blocks--------------------------- 958 // Check for aggressive application of 'split-if' optimization, 959 // using basic block level info. 960 void PhaseIdealLoop::split_if_with_blocks( VectorSet &visited, Node_Stack &nstack ) { 961 Node *n = C->root(); 962 visited.set(n->_idx); // first, mark node as visited 963 // Do pre-visit work for root 964 n = split_if_with_blocks_pre( n ); 965 uint cnt = n->outcnt(); 966 uint i = 0; 967 while (true) { 968 // Visit all children 969 if (i < cnt) { 970 Node* use = n->raw_out(i); 971 ++i; 972 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) { 973 // Now do pre-visit work for this use 974 use = split_if_with_blocks_pre( use ); 975 nstack.push(n, i); // Save parent and next use's index. 976 n = use; // Process all children of current use. 977 cnt = use->outcnt(); 978 i = 0; 979 } 980 } 981 else { 982 // All of n's children have been processed, complete post-processing. 983 if (cnt != 0 && !n->is_Con()) { 984 assert(has_node(n), "no dead nodes"); 985 split_if_with_blocks_post( n ); 986 } 987 if (nstack.is_empty()) { 988 // Finished all nodes on stack. 989 break; 990 } 991 // Get saved parent node and next use's index. Visit the rest of uses. 992 n = nstack.node(); 993 cnt = n->outcnt(); 994 i = nstack.index(); 995 nstack.pop(); 996 } 997 } 998 } 999 1000 1001 //============================================================================= 1002 // 1003 // C L O N E A L O O P B O D Y 1004 // 1005 1006 //------------------------------clone_iff-------------------------------------- 1007 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1008 // "Nearly" because all Nodes have been cloned from the original in the loop, 1009 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1010 // through the Phi recursively, and return a Bool. 1011 BoolNode *PhaseIdealLoop::clone_iff( PhiNode *phi, IdealLoopTree *loop ) { 1012 1013 // Convert this Phi into a Phi merging Bools 1014 uint i; 1015 for( i = 1; i < phi->req(); i++ ) { 1016 Node *b = phi->in(i); 1017 if( b->is_Phi() ) { 1018 _igvn.hash_delete(phi); 1019 _igvn._worklist.push(phi); 1020 phi->set_req(i, clone_iff( b->as_Phi(), loop )); 1021 } else { 1022 assert( b->is_Bool(), "" ); 1023 } 1024 } 1025 1026 Node *sample_bool = phi->in(1); 1027 Node *sample_cmp = sample_bool->in(1); 1028 1029 // Make Phis to merge the Cmp's inputs. 1030 int size = phi->in(0)->req(); 1031 PhiNode *phi1 = new (C, size) PhiNode( phi->in(0), Type::TOP ); 1032 PhiNode *phi2 = new (C, size) PhiNode( phi->in(0), Type::TOP ); 1033 for( i = 1; i < phi->req(); i++ ) { 1034 Node *n1 = phi->in(i)->in(1)->in(1); 1035 Node *n2 = phi->in(i)->in(1)->in(2); 1036 phi1->set_req( i, n1 ); 1037 phi2->set_req( i, n2 ); 1038 phi1->set_type( phi1->type()->meet(n1->bottom_type()) ); 1039 phi2->set_type( phi2->type()->meet(n2->bottom_type()) ); 1040 } 1041 // See if these Phis have been made before. 1042 // Register with optimizer 1043 Node *hit1 = _igvn.hash_find_insert(phi1); 1044 if( hit1 ) { // Hit, toss just made Phi 1045 _igvn.remove_dead_node(phi1); // Remove new phi 1046 assert( hit1->is_Phi(), "" ); 1047 phi1 = (PhiNode*)hit1; // Use existing phi 1048 } else { // Miss 1049 _igvn.register_new_node_with_optimizer(phi1); 1050 } 1051 Node *hit2 = _igvn.hash_find_insert(phi2); 1052 if( hit2 ) { // Hit, toss just made Phi 1053 _igvn.remove_dead_node(phi2); // Remove new phi 1054 assert( hit2->is_Phi(), "" ); 1055 phi2 = (PhiNode*)hit2; // Use existing phi 1056 } else { // Miss 1057 _igvn.register_new_node_with_optimizer(phi2); 1058 } 1059 // Register Phis with loop/block info 1060 set_ctrl(phi1, phi->in(0)); 1061 set_ctrl(phi2, phi->in(0)); 1062 // Make a new Cmp 1063 Node *cmp = sample_cmp->clone(); 1064 cmp->set_req( 1, phi1 ); 1065 cmp->set_req( 2, phi2 ); 1066 _igvn.register_new_node_with_optimizer(cmp); 1067 set_ctrl(cmp, phi->in(0)); 1068 1069 // Make a new Bool 1070 Node *b = sample_bool->clone(); 1071 b->set_req(1,cmp); 1072 _igvn.register_new_node_with_optimizer(b); 1073 set_ctrl(b, phi->in(0)); 1074 1075 assert( b->is_Bool(), "" ); 1076 return (BoolNode*)b; 1077 } 1078 1079 //------------------------------clone_bool------------------------------------- 1080 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1081 // "Nearly" because all Nodes have been cloned from the original in the loop, 1082 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1083 // through the Phi recursively, and return a Bool. 1084 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) { 1085 uint i; 1086 // Convert this Phi into a Phi merging Bools 1087 for( i = 1; i < phi->req(); i++ ) { 1088 Node *b = phi->in(i); 1089 if( b->is_Phi() ) { 1090 _igvn.hash_delete(phi); 1091 _igvn._worklist.push(phi); 1092 phi->set_req(i, clone_bool( b->as_Phi(), loop )); 1093 } else { 1094 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" ); 1095 } 1096 } 1097 1098 Node *sample_cmp = phi->in(1); 1099 1100 // Make Phis to merge the Cmp's inputs. 1101 int size = phi->in(0)->req(); 1102 PhiNode *phi1 = new (C, size) PhiNode( phi->in(0), Type::TOP ); 1103 PhiNode *phi2 = new (C, size) PhiNode( phi->in(0), Type::TOP ); 1104 for( uint j = 1; j < phi->req(); j++ ) { 1105 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP 1106 Node *n1, *n2; 1107 if( cmp_top->is_Cmp() ) { 1108 n1 = cmp_top->in(1); 1109 n2 = cmp_top->in(2); 1110 } else { 1111 n1 = n2 = cmp_top; 1112 } 1113 phi1->set_req( j, n1 ); 1114 phi2->set_req( j, n2 ); 1115 phi1->set_type( phi1->type()->meet(n1->bottom_type()) ); 1116 phi2->set_type( phi2->type()->meet(n2->bottom_type()) ); 1117 } 1118 1119 // See if these Phis have been made before. 1120 // Register with optimizer 1121 Node *hit1 = _igvn.hash_find_insert(phi1); 1122 if( hit1 ) { // Hit, toss just made Phi 1123 _igvn.remove_dead_node(phi1); // Remove new phi 1124 assert( hit1->is_Phi(), "" ); 1125 phi1 = (PhiNode*)hit1; // Use existing phi 1126 } else { // Miss 1127 _igvn.register_new_node_with_optimizer(phi1); 1128 } 1129 Node *hit2 = _igvn.hash_find_insert(phi2); 1130 if( hit2 ) { // Hit, toss just made Phi 1131 _igvn.remove_dead_node(phi2); // Remove new phi 1132 assert( hit2->is_Phi(), "" ); 1133 phi2 = (PhiNode*)hit2; // Use existing phi 1134 } else { // Miss 1135 _igvn.register_new_node_with_optimizer(phi2); 1136 } 1137 // Register Phis with loop/block info 1138 set_ctrl(phi1, phi->in(0)); 1139 set_ctrl(phi2, phi->in(0)); 1140 // Make a new Cmp 1141 Node *cmp = sample_cmp->clone(); 1142 cmp->set_req( 1, phi1 ); 1143 cmp->set_req( 2, phi2 ); 1144 _igvn.register_new_node_with_optimizer(cmp); 1145 set_ctrl(cmp, phi->in(0)); 1146 1147 assert( cmp->is_Cmp(), "" ); 1148 return (CmpNode*)cmp; 1149 } 1150 1151 //------------------------------sink_use--------------------------------------- 1152 // If 'use' was in the loop-exit block, it now needs to be sunk 1153 // below the post-loop merge point. 1154 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) { 1155 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) { 1156 set_ctrl(use, post_loop); 1157 for (DUIterator j = use->outs(); use->has_out(j); j++) 1158 sink_use(use->out(j), post_loop); 1159 } 1160 } 1161 1162 //------------------------------clone_loop------------------------------------- 1163 // 1164 // C L O N E A L O O P B O D Y 1165 // 1166 // This is the basic building block of the loop optimizations. It clones an 1167 // entire loop body. It makes an old_new loop body mapping; with this mapping 1168 // you can find the new-loop equivalent to an old-loop node. All new-loop 1169 // nodes are exactly equal to their old-loop counterparts, all edges are the 1170 // same. All exits from the old-loop now have a RegionNode that merges the 1171 // equivalent new-loop path. This is true even for the normal "loop-exit" 1172 // condition. All uses of loop-invariant old-loop values now come from (one 1173 // or more) Phis that merge their new-loop equivalents. 1174 // 1175 // This operation leaves the graph in an illegal state: there are two valid 1176 // control edges coming from the loop pre-header to both loop bodies. I'll 1177 // definitely have to hack the graph after running this transform. 1178 // 1179 // From this building block I will further edit edges to perform loop peeling 1180 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc. 1181 // 1182 // Parameter side_by_size_idom: 1183 // When side_by_size_idom is NULL, the dominator tree is constructed for 1184 // the clone loop to dominate the original. Used in construction of 1185 // pre-main-post loop sequence. 1186 // When nonnull, the clone and original are side-by-side, both are 1187 // dominated by the side_by_side_idom node. Used in construction of 1188 // unswitched loops. 1189 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd, 1190 Node* side_by_side_idom) { 1191 1192 // Step 1: Clone the loop body. Make the old->new mapping. 1193 uint i; 1194 for( i = 0; i < loop->_body.size(); i++ ) { 1195 Node *old = loop->_body.at(i); 1196 Node *nnn = old->clone(); 1197 old_new.map( old->_idx, nnn ); 1198 _igvn.register_new_node_with_optimizer(nnn); 1199 } 1200 1201 1202 // Step 2: Fix the edges in the new body. If the old input is outside the 1203 // loop use it. If the old input is INside the loop, use the corresponding 1204 // new node instead. 1205 for( i = 0; i < loop->_body.size(); i++ ) { 1206 Node *old = loop->_body.at(i); 1207 Node *nnn = old_new[old->_idx]; 1208 // Fix CFG/Loop controlling the new node 1209 if (has_ctrl(old)) { 1210 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]); 1211 } else { 1212 set_loop(nnn, loop->_parent); 1213 if (old->outcnt() > 0) { 1214 set_idom( nnn, old_new[idom(old)->_idx], dd ); 1215 } 1216 } 1217 // Correct edges to the new node 1218 for( uint j = 0; j < nnn->req(); j++ ) { 1219 Node *n = nnn->in(j); 1220 if( n ) { 1221 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n ); 1222 if( loop->is_member( old_in_loop ) ) 1223 nnn->set_req(j, old_new[n->_idx]); 1224 } 1225 } 1226 _igvn.hash_find_insert(nnn); 1227 } 1228 Node *newhead = old_new[loop->_head->_idx]; 1229 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 1230 1231 1232 // Step 3: Now fix control uses. Loop varying control uses have already 1233 // been fixed up (as part of all input edges in Step 2). Loop invariant 1234 // control uses must be either an IfFalse or an IfTrue. Make a merge 1235 // point to merge the old and new IfFalse/IfTrue nodes; make the use 1236 // refer to this. 1237 ResourceArea *area = Thread::current()->resource_area(); 1238 Node_List worklist(area); 1239 uint new_counter = C->unique(); 1240 for( i = 0; i < loop->_body.size(); i++ ) { 1241 Node* old = loop->_body.at(i); 1242 if( !old->is_CFG() ) continue; 1243 Node* nnn = old_new[old->_idx]; 1244 1245 // Copy uses to a worklist, so I can munge the def-use info 1246 // with impunity. 1247 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1248 worklist.push(old->fast_out(j)); 1249 1250 while( worklist.size() ) { // Visit all uses 1251 Node *use = worklist.pop(); 1252 if (!has_node(use)) continue; // Ignore dead nodes 1253 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1254 if( !loop->is_member( use_loop ) && use->is_CFG() ) { 1255 // Both OLD and USE are CFG nodes here. 1256 assert( use->is_Proj(), "" ); 1257 1258 // Clone the loop exit control projection 1259 Node *newuse = use->clone(); 1260 newuse->set_req(0,nnn); 1261 _igvn.register_new_node_with_optimizer(newuse); 1262 set_loop(newuse, use_loop); 1263 set_idom(newuse, nnn, dom_depth(nnn) + 1 ); 1264 1265 // We need a Region to merge the exit from the peeled body and the 1266 // exit from the old loop body. 1267 RegionNode *r = new (C, 3) RegionNode(3); 1268 // Map the old use to the new merge point 1269 old_new.map( use->_idx, r ); 1270 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use)); 1271 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" ); 1272 1273 // The original user of 'use' uses 'r' instead. 1274 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) { 1275 Node* useuse = use->last_out(l); 1276 _igvn.hash_delete(useuse); 1277 _igvn._worklist.push(useuse); 1278 uint uses_found = 0; 1279 if( useuse->in(0) == use ) { 1280 useuse->set_req(0, r); 1281 uses_found++; 1282 if( useuse->is_CFG() ) { 1283 assert( dom_depth(useuse) > dd_r, "" ); 1284 set_idom(useuse, r, dom_depth(useuse)); 1285 } 1286 } 1287 for( uint k = 1; k < useuse->req(); k++ ) { 1288 if( useuse->in(k) == use ) { 1289 useuse->set_req(k, r); 1290 uses_found++; 1291 } 1292 } 1293 l -= uses_found; // we deleted 1 or more copies of this edge 1294 } 1295 1296 // Now finish up 'r' 1297 r->set_req( 1, newuse ); 1298 r->set_req( 2, use ); 1299 _igvn.register_new_node_with_optimizer(r); 1300 set_loop(r, use_loop); 1301 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r); 1302 } // End of if a loop-exit test 1303 } 1304 } 1305 1306 // Step 4: If loop-invariant use is not control, it must be dominated by a 1307 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region 1308 // there if needed. Make a Phi there merging old and new used values. 1309 Node_List *split_if_set = NULL; 1310 Node_List *split_bool_set = NULL; 1311 Node_List *split_cex_set = NULL; 1312 for( i = 0; i < loop->_body.size(); i++ ) { 1313 Node* old = loop->_body.at(i); 1314 Node* nnn = old_new[old->_idx]; 1315 // Copy uses to a worklist, so I can munge the def-use info 1316 // with impunity. 1317 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1318 worklist.push(old->fast_out(j)); 1319 1320 while( worklist.size() ) { 1321 Node *use = worklist.pop(); 1322 if (!has_node(use)) continue; // Ignore dead nodes 1323 if (use->in(0) == C->top()) continue; 1324 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1325 // Check for data-use outside of loop - at least one of OLD or USE 1326 // must not be a CFG node. 1327 if( !loop->is_member( use_loop ) && (!old->is_CFG() || !use->is_CFG())) { 1328 1329 // If the Data use is an IF, that means we have an IF outside of the 1330 // loop that is switching on a condition that is set inside of the 1331 // loop. Happens if people set a loop-exit flag; then test the flag 1332 // in the loop to break the loop, then test is again outside of the 1333 // loop to determine which way the loop exited. 1334 if( use->is_If() || use->is_CMove() ) { 1335 // Since this code is highly unlikely, we lazily build the worklist 1336 // of such Nodes to go split. 1337 if( !split_if_set ) 1338 split_if_set = new Node_List(area); 1339 split_if_set->push(use); 1340 } 1341 if( use->is_Bool() ) { 1342 if( !split_bool_set ) 1343 split_bool_set = new Node_List(area); 1344 split_bool_set->push(use); 1345 } 1346 if( use->Opcode() == Op_CreateEx ) { 1347 if( !split_cex_set ) 1348 split_cex_set = new Node_List(area); 1349 split_cex_set->push(use); 1350 } 1351 1352 1353 // Get "block" use is in 1354 uint idx = 0; 1355 while( use->in(idx) != old ) idx++; 1356 Node *prev = use->is_CFG() ? use : get_ctrl(use); 1357 assert( !loop->is_member( get_loop( prev ) ), "" ); 1358 Node *cfg = prev->_idx >= new_counter 1359 ? prev->in(2) 1360 : idom(prev); 1361 if( use->is_Phi() ) // Phi use is in prior block 1362 cfg = prev->in(idx); // NOT in block of Phi itself 1363 if (cfg->is_top()) { // Use is dead? 1364 _igvn.hash_delete(use); 1365 _igvn._worklist.push(use); 1366 use->set_req(idx, C->top()); 1367 continue; 1368 } 1369 1370 while( !loop->is_member( get_loop( cfg ) ) ) { 1371 prev = cfg; 1372 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg); 1373 } 1374 // If the use occurs after merging several exits from the loop, then 1375 // old value must have dominated all those exits. Since the same old 1376 // value was used on all those exits we did not need a Phi at this 1377 // merge point. NOW we do need a Phi here. Each loop exit value 1378 // is now merged with the peeled body exit; each exit gets its own 1379 // private Phi and those Phis need to be merged here. 1380 Node *phi; 1381 if( prev->is_Region() ) { 1382 if( idx == 0 ) { // Updating control edge? 1383 phi = prev; // Just use existing control 1384 } else { // Else need a new Phi 1385 phi = PhiNode::make( prev, old ); 1386 // Now recursively fix up the new uses of old! 1387 for( uint i = 1; i < prev->req(); i++ ) { 1388 worklist.push(phi); // Onto worklist once for each 'old' input 1389 } 1390 } 1391 } else { 1392 // Get new RegionNode merging old and new loop exits 1393 prev = old_new[prev->_idx]; 1394 assert( prev, "just made this in step 7" ); 1395 if( idx == 0 ) { // Updating control edge? 1396 phi = prev; // Just use existing control 1397 } else { // Else need a new Phi 1398 // Make a new Phi merging data values properly 1399 phi = PhiNode::make( prev, old ); 1400 phi->set_req( 1, nnn ); 1401 } 1402 } 1403 // If inserting a new Phi, check for prior hits 1404 if( idx != 0 ) { 1405 Node *hit = _igvn.hash_find_insert(phi); 1406 if( hit == NULL ) { 1407 _igvn.register_new_node_with_optimizer(phi); // Register new phi 1408 } else { // or 1409 // Remove the new phi from the graph and use the hit 1410 _igvn.remove_dead_node(phi); 1411 phi = hit; // Use existing phi 1412 } 1413 set_ctrl(phi, prev); 1414 } 1415 // Make 'use' use the Phi instead of the old loop body exit value 1416 _igvn.hash_delete(use); 1417 _igvn._worklist.push(use); 1418 use->set_req(idx, phi); 1419 if( use->_idx >= new_counter ) { // If updating new phis 1420 // Not needed for correctness, but prevents a weak assert 1421 // in AddPNode from tripping (when we end up with different 1422 // base & derived Phis that will become the same after 1423 // IGVN does CSE). 1424 Node *hit = _igvn.hash_find_insert(use); 1425 if( hit ) // Go ahead and re-hash for hits. 1426 _igvn.subsume_node( use, hit ); 1427 } 1428 1429 // If 'use' was in the loop-exit block, it now needs to be sunk 1430 // below the post-loop merge point. 1431 sink_use( use, prev ); 1432 } 1433 } 1434 } 1435 1436 // Check for IFs that need splitting/cloning. Happens if an IF outside of 1437 // the loop uses a condition set in the loop. The original IF probably 1438 // takes control from one or more OLD Regions (which in turn get from NEW 1439 // Regions). In any case, there will be a set of Phis for each merge point 1440 // from the IF up to where the original BOOL def exists the loop. 1441 if( split_if_set ) { 1442 while( split_if_set->size() ) { 1443 Node *iff = split_if_set->pop(); 1444 if( iff->in(1)->is_Phi() ) { 1445 BoolNode *b = clone_iff( iff->in(1)->as_Phi(), loop ); 1446 _igvn.hash_delete(iff); 1447 _igvn._worklist.push(iff); 1448 iff->set_req(1, b); 1449 } 1450 } 1451 } 1452 if( split_bool_set ) { 1453 while( split_bool_set->size() ) { 1454 Node *b = split_bool_set->pop(); 1455 Node *phi = b->in(1); 1456 assert( phi->is_Phi(), "" ); 1457 CmpNode *cmp = clone_bool( (PhiNode*)phi, loop ); 1458 _igvn.hash_delete(b); 1459 _igvn._worklist.push(b); 1460 b->set_req(1, cmp); 1461 } 1462 } 1463 if( split_cex_set ) { 1464 while( split_cex_set->size() ) { 1465 Node *b = split_cex_set->pop(); 1466 assert( b->in(0)->is_Region(), "" ); 1467 assert( b->in(1)->is_Phi(), "" ); 1468 assert( b->in(0)->in(0) == b->in(1)->in(0), "" ); 1469 split_up( b, b->in(0), NULL ); 1470 } 1471 } 1472 1473 } 1474 1475 1476 //---------------------- stride_of_possible_iv ------------------------------------- 1477 // Looks for an iff/bool/comp with one operand of the compare 1478 // being a cycle involving an add and a phi, 1479 // with an optional truncation (left-shift followed by a right-shift) 1480 // of the add. Returns zero if not an iv. 1481 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) { 1482 Node* trunc1 = NULL; 1483 Node* trunc2 = NULL; 1484 const TypeInt* ttype = NULL; 1485 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) { 1486 return 0; 1487 } 1488 BoolNode* bl = iff->in(1)->as_Bool(); 1489 Node* cmp = bl->in(1); 1490 if (!cmp || cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU) { 1491 return 0; 1492 } 1493 // Must have an invariant operand 1494 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) { 1495 return 0; 1496 } 1497 Node* add2 = NULL; 1498 Node* cmp1 = cmp->in(1); 1499 if (cmp1->is_Phi()) { 1500 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) ))) 1501 Node* phi = cmp1; 1502 for (uint i = 1; i < phi->req(); i++) { 1503 Node* in = phi->in(i); 1504 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in, 1505 &trunc1, &trunc2, &ttype); 1506 if (add && add->in(1) == phi) { 1507 add2 = add->in(2); 1508 break; 1509 } 1510 } 1511 } else { 1512 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) ))) 1513 Node* addtrunc = cmp1; 1514 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc, 1515 &trunc1, &trunc2, &ttype); 1516 if (add && add->in(1)->is_Phi()) { 1517 Node* phi = add->in(1); 1518 for (uint i = 1; i < phi->req(); i++) { 1519 if (phi->in(i) == addtrunc) { 1520 add2 = add->in(2); 1521 break; 1522 } 1523 } 1524 } 1525 } 1526 if (add2 != NULL) { 1527 const TypeInt* add2t = _igvn.type(add2)->is_int(); 1528 if (add2t->is_con()) { 1529 return add2t->get_con(); 1530 } 1531 } 1532 return 0; 1533 } 1534 1535 1536 //---------------------- stay_in_loop ------------------------------------- 1537 // Return the (unique) control output node that's in the loop (if it exists.) 1538 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) { 1539 Node* unique = NULL; 1540 if (!n) return NULL; 1541 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1542 Node* use = n->fast_out(i); 1543 if (!has_ctrl(use) && loop->is_member(get_loop(use))) { 1544 if (unique != NULL) { 1545 return NULL; 1546 } 1547 unique = use; 1548 } 1549 } 1550 return unique; 1551 } 1552 1553 //------------------------------ register_node ------------------------------------- 1554 // Utility to register node "n" with PhaseIdealLoop 1555 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) { 1556 _igvn.register_new_node_with_optimizer(n); 1557 loop->_body.push(n); 1558 if (n->is_CFG()) { 1559 set_loop(n, loop); 1560 set_idom(n, pred, ddepth); 1561 } else { 1562 set_ctrl(n, pred); 1563 } 1564 } 1565 1566 //------------------------------ proj_clone ------------------------------------- 1567 // Utility to create an if-projection 1568 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) { 1569 ProjNode* c = p->clone()->as_Proj(); 1570 c->set_req(0, iff); 1571 return c; 1572 } 1573 1574 //------------------------------ short_circuit_if ------------------------------------- 1575 // Force the iff control output to be the live_proj 1576 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) { 1577 int proj_con = live_proj->_con; 1578 assert(proj_con == 0 || proj_con == 1, "false or true projection"); 1579 Node *con = _igvn.intcon(proj_con); 1580 set_ctrl(con, C->root()); 1581 if (iff) { 1582 iff->set_req(1, con); 1583 } 1584 return con; 1585 } 1586 1587 //------------------------------ insert_if_before_proj ------------------------------------- 1588 // Insert a new if before an if projection (* - new node) 1589 // 1590 // before 1591 // if(test) 1592 // / \ 1593 // v v 1594 // other-proj proj (arg) 1595 // 1596 // after 1597 // if(test) 1598 // / \ 1599 // / v 1600 // | * proj-clone 1601 // v | 1602 // other-proj v 1603 // * new_if(relop(cmp[IU](left,right))) 1604 // / \ 1605 // v v 1606 // * new-proj proj 1607 // (returned) 1608 // 1609 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) { 1610 IfNode* iff = proj->in(0)->as_If(); 1611 IdealLoopTree *loop = get_loop(proj); 1612 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 1613 int ddepth = dom_depth(proj); 1614 1615 _igvn.hash_delete(iff); 1616 _igvn._worklist.push(iff); 1617 _igvn.hash_delete(proj); 1618 _igvn._worklist.push(proj); 1619 1620 proj->set_req(0, NULL); // temporary disconnect 1621 ProjNode* proj2 = proj_clone(proj, iff); 1622 register_node(proj2, loop, iff, ddepth); 1623 1624 Node* cmp = Signed ? (Node*) new (C,3)CmpINode(left, right) : (Node*) new (C,3)CmpUNode(left, right); 1625 register_node(cmp, loop, proj2, ddepth); 1626 1627 BoolNode* bol = new (C,2)BoolNode(cmp, relop); 1628 register_node(bol, loop, proj2, ddepth); 1629 1630 IfNode* new_if = new (C,2)IfNode(proj2, bol, iff->_prob, iff->_fcnt); 1631 register_node(new_if, loop, proj2, ddepth); 1632 1633 proj->set_req(0, new_if); // reattach 1634 set_idom(proj, new_if, ddepth); 1635 1636 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj(); 1637 register_node(new_exit, get_loop(other_proj), new_if, ddepth); 1638 1639 return new_exit; 1640 } 1641 1642 //------------------------------ insert_region_before_proj ------------------------------------- 1643 // Insert a region before an if projection (* - new node) 1644 // 1645 // before 1646 // if(test) 1647 // / | 1648 // v | 1649 // proj v 1650 // other-proj 1651 // 1652 // after 1653 // if(test) 1654 // / | 1655 // v | 1656 // * proj-clone v 1657 // | other-proj 1658 // v 1659 // * new-region 1660 // | 1661 // v 1662 // * dum_if 1663 // / \ 1664 // v \ 1665 // * dum-proj v 1666 // proj 1667 // 1668 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) { 1669 IfNode* iff = proj->in(0)->as_If(); 1670 IdealLoopTree *loop = get_loop(proj); 1671 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 1672 int ddepth = dom_depth(proj); 1673 1674 _igvn.hash_delete(iff); 1675 _igvn._worklist.push(iff); 1676 _igvn.hash_delete(proj); 1677 _igvn._worklist.push(proj); 1678 1679 proj->set_req(0, NULL); // temporary disconnect 1680 ProjNode* proj2 = proj_clone(proj, iff); 1681 register_node(proj2, loop, iff, ddepth); 1682 1683 RegionNode* reg = new (C,2)RegionNode(2); 1684 reg->set_req(1, proj2); 1685 register_node(reg, loop, iff, ddepth); 1686 1687 IfNode* dum_if = new (C,2)IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt); 1688 register_node(dum_if, loop, reg, ddepth); 1689 1690 proj->set_req(0, dum_if); // reattach 1691 set_idom(proj, dum_if, ddepth); 1692 1693 ProjNode* dum_proj = proj_clone(other_proj, dum_if); 1694 register_node(dum_proj, loop, dum_if, ddepth); 1695 1696 return reg; 1697 } 1698 1699 //------------------------------ insert_cmpi_loop_exit ------------------------------------- 1700 // Clone a signed compare loop exit from an unsigned compare and 1701 // insert it before the unsigned cmp on the stay-in-loop path. 1702 // All new nodes inserted in the dominator tree between the original 1703 // if and it's projections. The original if test is replaced with 1704 // a constant to force the stay-in-loop path. 1705 // 1706 // This is done to make sure that the original if and it's projections 1707 // still dominate the same set of control nodes, that the ctrl() relation 1708 // from data nodes to them is preserved, and that their loop nesting is 1709 // preserved. 1710 // 1711 // before 1712 // if(i <u limit) unsigned compare loop exit 1713 // / | 1714 // v v 1715 // exit-proj stay-in-loop-proj 1716 // 1717 // after 1718 // if(stay-in-loop-const) original if 1719 // / | 1720 // / v 1721 // / if(i < limit) new signed test 1722 // / / | 1723 // / / v 1724 // / / if(i <u limit) new cloned unsigned test 1725 // / / / | 1726 // v v v | 1727 // region | 1728 // | | 1729 // dum-if | 1730 // / | | 1731 // ether | | 1732 // v v 1733 // exit-proj stay-in-loop-proj 1734 // 1735 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) { 1736 const bool Signed = true; 1737 const bool Unsigned = false; 1738 1739 BoolNode* bol = if_cmpu->in(1)->as_Bool(); 1740 if (bol->_test._test != BoolTest::lt) return NULL; 1741 CmpNode* cmpu = bol->in(1)->as_Cmp(); 1742 if (cmpu->Opcode() != Op_CmpU) return NULL; 1743 int stride = stride_of_possible_iv(if_cmpu); 1744 if (stride == 0) return NULL; 1745 1746 ProjNode* lp_continue = stay_in_loop(if_cmpu, loop)->as_Proj(); 1747 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj(); 1748 1749 Node* limit = NULL; 1750 if (stride > 0) { 1751 limit = cmpu->in(2); 1752 } else { 1753 limit = _igvn.makecon(TypeInt::ZERO); 1754 set_ctrl(limit, C->root()); 1755 } 1756 // Create a new region on the exit path 1757 RegionNode* reg = insert_region_before_proj(lp_exit); 1758 1759 // Clone the if-cmpu-true-false using a signed compare 1760 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge; 1761 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue); 1762 reg->add_req(cmpi_exit); 1763 1764 // Clone the if-cmpu-true-false 1765 BoolTest::mask rel_u = bol->_test._test; 1766 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue); 1767 reg->add_req(cmpu_exit); 1768 1769 // Force original if to stay in loop. 1770 short_circuit_if(if_cmpu, lp_continue); 1771 1772 return cmpi_exit->in(0)->as_If(); 1773 } 1774 1775 //------------------------------ remove_cmpi_loop_exit ------------------------------------- 1776 // Remove a previously inserted signed compare loop exit. 1777 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) { 1778 Node* lp_proj = stay_in_loop(if_cmp, loop); 1779 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI && 1780 stay_in_loop(lp_proj, loop)->is_If() && 1781 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu"); 1782 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO); 1783 set_ctrl(con, C->root()); 1784 if_cmp->set_req(1, con); 1785 } 1786 1787 //------------------------------ scheduled_nodelist ------------------------------------- 1788 // Create a post order schedule of nodes that are in the 1789 // "member" set. The list is returned in "sched". 1790 // The first node in "sched" is the loop head, followed by 1791 // nodes which have no inputs in the "member" set, and then 1792 // followed by the nodes that have an immediate input dependence 1793 // on a node in "sched". 1794 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) { 1795 1796 assert(member.test(loop->_head->_idx), "loop head must be in member set"); 1797 Arena *a = Thread::current()->resource_area(); 1798 VectorSet visited(a); 1799 Node_Stack nstack(a, loop->_body.size()); 1800 1801 Node* n = loop->_head; // top of stack is cached in "n" 1802 uint idx = 0; 1803 visited.set(n->_idx); 1804 1805 // Initially push all with no inputs from within member set 1806 for(uint i = 0; i < loop->_body.size(); i++ ) { 1807 Node *elt = loop->_body.at(i); 1808 if (member.test(elt->_idx)) { 1809 bool found = false; 1810 for (uint j = 0; j < elt->req(); j++) { 1811 Node* def = elt->in(j); 1812 if (def && member.test(def->_idx) && def != elt) { 1813 found = true; 1814 break; 1815 } 1816 } 1817 if (!found && elt != loop->_head) { 1818 nstack.push(n, idx); 1819 n = elt; 1820 assert(!visited.test(n->_idx), "not seen yet"); 1821 visited.set(n->_idx); 1822 } 1823 } 1824 } 1825 1826 // traverse out's that are in the member set 1827 while (true) { 1828 if (idx < n->outcnt()) { 1829 Node* use = n->raw_out(idx); 1830 idx++; 1831 if (!visited.test_set(use->_idx)) { 1832 if (member.test(use->_idx)) { 1833 nstack.push(n, idx); 1834 n = use; 1835 idx = 0; 1836 } 1837 } 1838 } else { 1839 // All outputs processed 1840 sched.push(n); 1841 if (nstack.is_empty()) break; 1842 n = nstack.node(); 1843 idx = nstack.index(); 1844 nstack.pop(); 1845 } 1846 } 1847 } 1848 1849 1850 //------------------------------ has_use_in_set ------------------------------------- 1851 // Has a use in the vector set 1852 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) { 1853 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1854 Node* use = n->fast_out(j); 1855 if (vset.test(use->_idx)) { 1856 return true; 1857 } 1858 } 1859 return false; 1860 } 1861 1862 1863 //------------------------------ has_use_internal_to_set ------------------------------------- 1864 // Has use internal to the vector set (ie. not in a phi at the loop head) 1865 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) { 1866 Node* head = loop->_head; 1867 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1868 Node* use = n->fast_out(j); 1869 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) { 1870 return true; 1871 } 1872 } 1873 return false; 1874 } 1875 1876 1877 //------------------------------ clone_for_use_outside_loop ------------------------------------- 1878 // clone "n" for uses that are outside of loop 1879 void PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) { 1880 1881 assert(worklist.size() == 0, "should be empty"); 1882 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1883 Node* use = n->fast_out(j); 1884 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) { 1885 worklist.push(use); 1886 } 1887 } 1888 while( worklist.size() ) { 1889 Node *use = worklist.pop(); 1890 if (!has_node(use) || use->in(0) == C->top()) continue; 1891 uint j; 1892 for (j = 0; j < use->req(); j++) { 1893 if (use->in(j) == n) break; 1894 } 1895 assert(j < use->req(), "must be there"); 1896 1897 // clone "n" and insert it between the inputs of "n" and the use outside the loop 1898 Node* n_clone = n->clone(); 1899 _igvn.hash_delete(use); 1900 use->set_req(j, n_clone); 1901 _igvn._worklist.push(use); 1902 Node* use_c; 1903 if (!use->is_Phi()) { 1904 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0); 1905 } else { 1906 // Use in a phi is considered a use in the associated predecessor block 1907 use_c = use->in(0)->in(j); 1908 } 1909 set_ctrl(n_clone, use_c); 1910 assert(!loop->is_member(get_loop(use_c)), "should be outside loop"); 1911 get_loop(use_c)->_body.push(n_clone); 1912 _igvn.register_new_node_with_optimizer(n_clone); 1913 #if !defined(PRODUCT) 1914 if (TracePartialPeeling) { 1915 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx); 1916 } 1917 #endif 1918 } 1919 } 1920 1921 1922 //------------------------------ clone_for_special_use_inside_loop ------------------------------------- 1923 // clone "n" for special uses that are in the not_peeled region. 1924 // If these def-uses occur in separate blocks, the code generator 1925 // marks the method as not compilable. For example, if a "BoolNode" 1926 // is in a different basic block than the "IfNode" that uses it, then 1927 // the compilation is aborted in the code generator. 1928 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 1929 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) { 1930 if (n->is_Phi() || n->is_Load()) { 1931 return; 1932 } 1933 assert(worklist.size() == 0, "should be empty"); 1934 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1935 Node* use = n->fast_out(j); 1936 if ( not_peel.test(use->_idx) && 1937 (use->is_If() || use->is_CMove() || use->is_Bool()) && 1938 use->in(1) == n) { 1939 worklist.push(use); 1940 } 1941 } 1942 if (worklist.size() > 0) { 1943 // clone "n" and insert it between inputs of "n" and the use 1944 Node* n_clone = n->clone(); 1945 loop->_body.push(n_clone); 1946 _igvn.register_new_node_with_optimizer(n_clone); 1947 set_ctrl(n_clone, get_ctrl(n)); 1948 sink_list.push(n_clone); 1949 not_peel <<= n_clone->_idx; // add n_clone to not_peel set. 1950 #if !defined(PRODUCT) 1951 if (TracePartialPeeling) { 1952 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx); 1953 } 1954 #endif 1955 while( worklist.size() ) { 1956 Node *use = worklist.pop(); 1957 _igvn.hash_delete(use); 1958 _igvn._worklist.push(use); 1959 for (uint j = 1; j < use->req(); j++) { 1960 if (use->in(j) == n) { 1961 use->set_req(j, n_clone); 1962 } 1963 } 1964 } 1965 } 1966 } 1967 1968 1969 //------------------------------ insert_phi_for_loop ------------------------------------- 1970 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 1971 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) { 1972 Node *phi = PhiNode::make(lp, back_edge_val); 1973 phi->set_req(LoopNode::EntryControl, lp_entry_val); 1974 // Use existing phi if it already exists 1975 Node *hit = _igvn.hash_find_insert(phi); 1976 if( hit == NULL ) { 1977 _igvn.register_new_node_with_optimizer(phi); 1978 set_ctrl(phi, lp); 1979 } else { 1980 // Remove the new phi from the graph and use the hit 1981 _igvn.remove_dead_node(phi); 1982 phi = hit; 1983 } 1984 _igvn.hash_delete(use); 1985 _igvn._worklist.push(use); 1986 use->set_req(idx, phi); 1987 } 1988 1989 #ifdef ASSERT 1990 //------------------------------ is_valid_loop_partition ------------------------------------- 1991 // Validate the loop partition sets: peel and not_peel 1992 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, 1993 VectorSet& not_peel ) { 1994 uint i; 1995 // Check that peel_list entries are in the peel set 1996 for (i = 0; i < peel_list.size(); i++) { 1997 if (!peel.test(peel_list.at(i)->_idx)) { 1998 return false; 1999 } 2000 } 2001 // Check at loop members are in one of peel set or not_peel set 2002 for (i = 0; i < loop->_body.size(); i++ ) { 2003 Node *def = loop->_body.at(i); 2004 uint di = def->_idx; 2005 // Check that peel set elements are in peel_list 2006 if (peel.test(di)) { 2007 if (not_peel.test(di)) { 2008 return false; 2009 } 2010 // Must be in peel_list also 2011 bool found = false; 2012 for (uint j = 0; j < peel_list.size(); j++) { 2013 if (peel_list.at(j)->_idx == di) { 2014 found = true; 2015 break; 2016 } 2017 } 2018 if (!found) { 2019 return false; 2020 } 2021 } else if (not_peel.test(di)) { 2022 if (peel.test(di)) { 2023 return false; 2024 } 2025 } else { 2026 return false; 2027 } 2028 } 2029 return true; 2030 } 2031 2032 //------------------------------ is_valid_clone_loop_exit_use ------------------------------------- 2033 // Ensure a use outside of loop is of the right form 2034 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) { 2035 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2036 return (use->is_Phi() && 2037 use_c->is_Region() && use_c->req() == 3 && 2038 (use_c->in(exit_idx)->Opcode() == Op_IfTrue || 2039 use_c->in(exit_idx)->Opcode() == Op_IfFalse || 2040 use_c->in(exit_idx)->Opcode() == Op_JumpProj) && 2041 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) ); 2042 } 2043 2044 //------------------------------ is_valid_clone_loop_form ------------------------------------- 2045 // Ensure that all uses outside of loop are of the right form 2046 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 2047 uint orig_exit_idx, uint clone_exit_idx) { 2048 uint len = peel_list.size(); 2049 for (uint i = 0; i < len; i++) { 2050 Node *def = peel_list.at(i); 2051 2052 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 2053 Node *use = def->fast_out(j); 2054 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2055 if (!loop->is_member(get_loop(use_c))) { 2056 // use is not in the loop, check for correct structure 2057 if (use->in(0) == def) { 2058 // Okay 2059 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) { 2060 return false; 2061 } 2062 } 2063 } 2064 } 2065 return true; 2066 } 2067 #endif 2068 2069 //------------------------------ partial_peel ------------------------------------- 2070 // Partially peel (aka loop rotation) the top portion of a loop (called 2071 // the peel section below) by cloning it and placing one copy just before 2072 // the new loop head and the other copy at the bottom of the new loop. 2073 // 2074 // before after where it came from 2075 // 2076 // stmt1 stmt1 2077 // loop: stmt2 clone 2078 // stmt2 if condA goto exitA clone 2079 // if condA goto exitA new_loop: new 2080 // stmt3 stmt3 clone 2081 // if !condB goto loop if condB goto exitB clone 2082 // exitB: stmt2 orig 2083 // stmt4 if !condA goto new_loop orig 2084 // exitA: goto exitA 2085 // exitB: 2086 // stmt4 2087 // exitA: 2088 // 2089 // Step 1: find the cut point: an exit test on probable 2090 // induction variable. 2091 // Step 2: schedule (with cloning) operations in the peel 2092 // section that can be executed after the cut into 2093 // the section that is not peeled. This may need 2094 // to clone operations into exit blocks. For 2095 // instance, a reference to A[i] in the not-peel 2096 // section and a reference to B[i] in an exit block 2097 // may cause a left-shift of i by 2 to be placed 2098 // in the peel block. This step will clone the left 2099 // shift into the exit block and sink the left shift 2100 // from the peel to the not-peel section. 2101 // Step 3: clone the loop, retarget the control, and insert 2102 // phis for values that are live across the new loop 2103 // head. This is very dependent on the graph structure 2104 // from clone_loop. It creates region nodes for 2105 // exit control and associated phi nodes for values 2106 // flow out of the loop through that exit. The region 2107 // node is dominated by the clone's control projection. 2108 // So the clone's peel section is placed before the 2109 // new loop head, and the clone's not-peel section is 2110 // forms the top part of the new loop. The original 2111 // peel section forms the tail of the new loop. 2112 // Step 4: update the dominator tree and recompute the 2113 // dominator depth. 2114 // 2115 // orig 2116 // 2117 // stmt1 2118 // | 2119 // v 2120 // loop<----+ 2121 // | | 2122 // stmt2 | 2123 // | | 2124 // v | 2125 // ifA | 2126 // / | | 2127 // v v | 2128 // false true ^ <-- last_peel 2129 // / | | 2130 // / ===|==cut | 2131 // / stmt3 | <-- first_not_peel 2132 // / | | 2133 // | v | 2134 // v ifB | 2135 // exitA: / \ | 2136 // / \ | 2137 // v v | 2138 // false true | 2139 // / \ | 2140 // / ----+ 2141 // | 2142 // v 2143 // exitB: 2144 // stmt4 2145 // 2146 // 2147 // after clone loop 2148 // 2149 // stmt1 2150 // / \ 2151 // clone / \ orig 2152 // / \ 2153 // / \ 2154 // v v 2155 // +---->loop loop<----+ 2156 // | | | | 2157 // | stmt2 stmt2 | 2158 // | | | | 2159 // | v v | 2160 // | ifA ifA | 2161 // | | \ / | | 2162 // | v v v v | 2163 // ^ true false false true ^ <-- last_peel 2164 // | | ^ \ / | | 2165 // | cut==|== \ \ / ===|==cut | 2166 // | stmt3 \ \ / stmt3 | <-- first_not_peel 2167 // | | dom | | | | 2168 // | v \ 1v v2 v | 2169 // | ifB regionA ifB | 2170 // | / \ | / \ | 2171 // | / \ v / \ | 2172 // | v v exitA: v v | 2173 // | true false false true | 2174 // | / ^ \ / \ | 2175 // +---- \ \ / ----+ 2176 // dom \ / 2177 // \ 1v v2 2178 // regionB 2179 // | 2180 // v 2181 // exitB: 2182 // stmt4 2183 // 2184 // 2185 // after partial peel 2186 // 2187 // stmt1 2188 // / 2189 // clone / orig 2190 // / TOP 2191 // / \ 2192 // v v 2193 // TOP->region region----+ 2194 // | | | 2195 // stmt2 stmt2 | 2196 // | | | 2197 // v v | 2198 // ifA ifA | 2199 // | \ / | | 2200 // v v v v | 2201 // true false false true | <-- last_peel 2202 // | ^ \ / +------|---+ 2203 // +->newloop \ \ / === ==cut | | 2204 // | stmt3 \ \ / TOP | | 2205 // | | dom | | stmt3 | | <-- first_not_peel 2206 // | v \ 1v v2 v | | 2207 // | ifB regionA ifB ^ v 2208 // | / \ | / \ | | 2209 // | / \ v / \ | | 2210 // | v v exitA: v v | | 2211 // | true false false true | | 2212 // | / ^ \ / \ | | 2213 // | | \ \ / v | | 2214 // | | dom \ / TOP | | 2215 // | | \ 1v v2 | | 2216 // ^ v regionB | | 2217 // | | | | | 2218 // | | v ^ v 2219 // | | exitB: | | 2220 // | | stmt4 | | 2221 // | +------------>-----------------+ | 2222 // | | 2223 // +-----------------<---------------------+ 2224 // 2225 // 2226 // final graph 2227 // 2228 // stmt1 2229 // | 2230 // v 2231 // ........> ifA clone 2232 // : / | 2233 // dom / | 2234 // : v v 2235 // : false true 2236 // : | | 2237 // : | stmt2 clone 2238 // : | | 2239 // : | v 2240 // : | newloop<-----+ 2241 // : | | | 2242 // : | stmt3 clone | 2243 // : | | | 2244 // : | v | 2245 // : | ifB | 2246 // : | / \ | 2247 // : | v v | 2248 // : | false true | 2249 // : | | | | 2250 // : | v stmt2 | 2251 // : | exitB: | | 2252 // : | stmt4 v | 2253 // : | ifA orig | 2254 // : | / \ | 2255 // : | / \ | 2256 // : | v v | 2257 // : | false true | 2258 // : | / \ | 2259 // : v v -----+ 2260 // RegionA 2261 // | 2262 // v 2263 // exitA 2264 // 2265 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) { 2266 2267 if (!loop->_head->is_Loop()) { 2268 return false; } 2269 2270 LoopNode *head = loop->_head->as_Loop(); 2271 2272 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) { 2273 return false; 2274 } 2275 2276 // Check for complex exit control 2277 for(uint ii = 0; ii < loop->_body.size(); ii++ ) { 2278 Node *n = loop->_body.at(ii); 2279 int opc = n->Opcode(); 2280 if (n->is_Call() || 2281 opc == Op_Catch || 2282 opc == Op_CatchProj || 2283 opc == Op_Jump || 2284 opc == Op_JumpProj) { 2285 #if !defined(PRODUCT) 2286 if (TracePartialPeeling) { 2287 tty->print_cr("\nExit control too complex: lp: %d", head->_idx); 2288 } 2289 #endif 2290 return false; 2291 } 2292 } 2293 2294 int dd = dom_depth(head); 2295 2296 // Step 1: find cut point 2297 2298 // Walk up dominators to loop head looking for first loop exit 2299 // which is executed on every path thru loop. 2300 IfNode *peel_if = NULL; 2301 IfNode *peel_if_cmpu = NULL; 2302 2303 Node *iff = loop->tail(); 2304 while( iff != head ) { 2305 if( iff->is_If() ) { 2306 Node *ctrl = get_ctrl(iff->in(1)); 2307 if (ctrl->is_top()) return false; // Dead test on live IF. 2308 // If loop-varying exit-test, check for induction variable 2309 if( loop->is_member(get_loop(ctrl)) && 2310 loop->is_loop_exit(iff) && 2311 is_possible_iv_test(iff)) { 2312 Node* cmp = iff->in(1)->in(1); 2313 if (cmp->Opcode() == Op_CmpI) { 2314 peel_if = iff->as_If(); 2315 } else { 2316 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU"); 2317 peel_if_cmpu = iff->as_If(); 2318 } 2319 } 2320 } 2321 iff = idom(iff); 2322 } 2323 // Prefer signed compare over unsigned compare. 2324 IfNode* new_peel_if = NULL; 2325 if (peel_if == NULL) { 2326 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) { 2327 return false; // No peel point found 2328 } 2329 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop); 2330 if (new_peel_if == NULL) { 2331 return false; // No peel point found 2332 } 2333 peel_if = new_peel_if; 2334 } 2335 Node* last_peel = stay_in_loop(peel_if, loop); 2336 Node* first_not_peeled = stay_in_loop(last_peel, loop); 2337 if (first_not_peeled == NULL || first_not_peeled == head) { 2338 return false; 2339 } 2340 2341 #if !defined(PRODUCT) 2342 if (TracePartialPeeling) { 2343 tty->print_cr("before partial peel one iteration"); 2344 Node_List wl; 2345 Node* t = head->in(2); 2346 while (true) { 2347 wl.push(t); 2348 if (t == head) break; 2349 t = idom(t); 2350 } 2351 while (wl.size() > 0) { 2352 Node* tt = wl.pop(); 2353 tt->dump(); 2354 if (tt == last_peel) tty->print_cr("-- cut --"); 2355 } 2356 } 2357 #endif 2358 ResourceArea *area = Thread::current()->resource_area(); 2359 VectorSet peel(area); 2360 VectorSet not_peel(area); 2361 Node_List peel_list(area); 2362 Node_List worklist(area); 2363 Node_List sink_list(area); 2364 2365 // Set of cfg nodes to peel are those that are executable from 2366 // the head through last_peel. 2367 assert(worklist.size() == 0, "should be empty"); 2368 worklist.push(head); 2369 peel.set(head->_idx); 2370 while (worklist.size() > 0) { 2371 Node *n = worklist.pop(); 2372 if (n != last_peel) { 2373 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2374 Node* use = n->fast_out(j); 2375 if (use->is_CFG() && 2376 loop->is_member(get_loop(use)) && 2377 !peel.test_set(use->_idx)) { 2378 worklist.push(use); 2379 } 2380 } 2381 } 2382 } 2383 2384 // Set of non-cfg nodes to peel are those that are control 2385 // dependent on the cfg nodes. 2386 uint i; 2387 for(i = 0; i < loop->_body.size(); i++ ) { 2388 Node *n = loop->_body.at(i); 2389 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n; 2390 if (peel.test(n_c->_idx)) { 2391 peel.set(n->_idx); 2392 } else { 2393 not_peel.set(n->_idx); 2394 } 2395 } 2396 2397 // Step 2: move operations from the peeled section down into the 2398 // not-peeled section 2399 2400 // Get a post order schedule of nodes in the peel region 2401 // Result in right-most operand. 2402 scheduled_nodelist(loop, peel, peel_list ); 2403 2404 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 2405 2406 // For future check for too many new phis 2407 uint old_phi_cnt = 0; 2408 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 2409 Node* use = head->fast_out(j); 2410 if (use->is_Phi()) old_phi_cnt++; 2411 } 2412 2413 #if !defined(PRODUCT) 2414 if (TracePartialPeeling) { 2415 tty->print_cr("\npeeled list"); 2416 } 2417 #endif 2418 2419 // Evacuate nodes in peel region into the not_peeled region if possible 2420 uint new_phi_cnt = 0; 2421 for (i = 0; i < peel_list.size();) { 2422 Node* n = peel_list.at(i); 2423 #if !defined(PRODUCT) 2424 if (TracePartialPeeling) n->dump(); 2425 #endif 2426 bool incr = true; 2427 if ( !n->is_CFG() ) { 2428 2429 if ( has_use_in_set(n, not_peel) ) { 2430 2431 // If not used internal to the peeled region, 2432 // move "n" from peeled to not_peeled region. 2433 2434 if ( !has_use_internal_to_set(n, peel, loop) ) { 2435 2436 // if not pinned and not a load (which maybe anti-dependent on a store) 2437 // and not a CMove (Matcher expects only bool->cmove). 2438 if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) { 2439 clone_for_use_outside_loop( loop, n, worklist ); 2440 2441 sink_list.push(n); 2442 peel >>= n->_idx; // delete n from peel set. 2443 not_peel <<= n->_idx; // add n to not_peel set. 2444 peel_list.remove(i); 2445 incr = false; 2446 #if !defined(PRODUCT) 2447 if (TracePartialPeeling) { 2448 tty->print_cr("sink to not_peeled region: %d newbb: %d", 2449 n->_idx, get_ctrl(n)->_idx); 2450 } 2451 #endif 2452 } 2453 } else { 2454 // Otherwise check for special def-use cases that span 2455 // the peel/not_peel boundary such as bool->if 2456 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist ); 2457 new_phi_cnt++; 2458 } 2459 } 2460 } 2461 if (incr) i++; 2462 } 2463 2464 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) { 2465 #if !defined(PRODUCT) 2466 if (TracePartialPeeling) { 2467 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c", 2468 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F'); 2469 } 2470 #endif 2471 if (new_peel_if != NULL) { 2472 remove_cmpi_loop_exit(new_peel_if, loop); 2473 } 2474 // Inhibit more partial peeling on this loop 2475 assert(!head->is_partial_peel_loop(), "not partial peeled"); 2476 head->mark_partial_peel_failed(); 2477 return false; 2478 } 2479 2480 // Step 3: clone loop, retarget control, and insert new phis 2481 2482 // Create new loop head for new phis and to hang 2483 // the nodes being moved (sinked) from the peel region. 2484 LoopNode* new_head = new (C, 3) LoopNode(last_peel, last_peel); 2485 _igvn.register_new_node_with_optimizer(new_head); 2486 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled"); 2487 first_not_peeled->set_req(0, new_head); 2488 set_loop(new_head, loop); 2489 loop->_body.push(new_head); 2490 not_peel.set(new_head->_idx); 2491 set_idom(new_head, last_peel, dom_depth(first_not_peeled)); 2492 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled)); 2493 2494 while (sink_list.size() > 0) { 2495 Node* n = sink_list.pop(); 2496 set_ctrl(n, new_head); 2497 } 2498 2499 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 2500 2501 clone_loop( loop, old_new, dd ); 2502 2503 const uint clone_exit_idx = 1; 2504 const uint orig_exit_idx = 2; 2505 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop"); 2506 2507 Node* head_clone = old_new[head->_idx]; 2508 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop(); 2509 Node* orig_tail_clone = head_clone->in(2); 2510 2511 // Add phi if "def" node is in peel set and "use" is not 2512 2513 for(i = 0; i < peel_list.size(); i++ ) { 2514 Node *def = peel_list.at(i); 2515 if (!def->is_CFG()) { 2516 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 2517 Node *use = def->fast_out(j); 2518 if (has_node(use) && use->in(0) != C->top() && 2519 (!peel.test(use->_idx) || 2520 (use->is_Phi() && use->in(0) == head)) ) { 2521 worklist.push(use); 2522 } 2523 } 2524 while( worklist.size() ) { 2525 Node *use = worklist.pop(); 2526 for (uint j = 1; j < use->req(); j++) { 2527 Node* n = use->in(j); 2528 if (n == def) { 2529 2530 // "def" is in peel set, "use" is not in peel set 2531 // or "use" is in the entry boundary (a phi) of the peel set 2532 2533 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use; 2534 2535 if ( loop->is_member(get_loop( use_c )) ) { 2536 // use is in loop 2537 if (old_new[use->_idx] != NULL) { // null for dead code 2538 Node* use_clone = old_new[use->_idx]; 2539 _igvn.hash_delete(use); 2540 use->set_req(j, C->top()); 2541 _igvn._worklist.push(use); 2542 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone ); 2543 } 2544 } else { 2545 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format"); 2546 // use is not in the loop, check if the live range includes the cut 2547 Node* lp_if = use_c->in(orig_exit_idx)->in(0); 2548 if (not_peel.test(lp_if->_idx)) { 2549 assert(j == orig_exit_idx, "use from original loop"); 2550 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone ); 2551 } 2552 } 2553 } 2554 } 2555 } 2556 } 2557 } 2558 2559 // Step 3b: retarget control 2560 2561 // Redirect control to the new loop head if a cloned node in 2562 // the not_peeled region has control that points into the peeled region. 2563 // This necessary because the cloned peeled region will be outside 2564 // the loop. 2565 // from to 2566 // cloned-peeled <---+ 2567 // new_head_clone: | <--+ 2568 // cloned-not_peeled in(0) in(0) 2569 // orig-peeled 2570 2571 for(i = 0; i < loop->_body.size(); i++ ) { 2572 Node *n = loop->_body.at(i); 2573 if (!n->is_CFG() && n->in(0) != NULL && 2574 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) { 2575 Node* n_clone = old_new[n->_idx]; 2576 _igvn.hash_delete(n_clone); 2577 n_clone->set_req(0, new_head_clone); 2578 _igvn._worklist.push(n_clone); 2579 } 2580 } 2581 2582 // Backedge of the surviving new_head (the clone) is original last_peel 2583 _igvn.hash_delete(new_head_clone); 2584 new_head_clone->set_req(LoopNode::LoopBackControl, last_peel); 2585 _igvn._worklist.push(new_head_clone); 2586 2587 // Cut first node in original not_peel set 2588 _igvn.hash_delete(new_head); 2589 new_head->set_req(LoopNode::EntryControl, C->top()); 2590 new_head->set_req(LoopNode::LoopBackControl, C->top()); 2591 _igvn._worklist.push(new_head); 2592 2593 // Copy head_clone back-branch info to original head 2594 // and remove original head's loop entry and 2595 // clone head's back-branch 2596 _igvn.hash_delete(head); 2597 _igvn.hash_delete(head_clone); 2598 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl)); 2599 head->set_req(LoopNode::LoopBackControl, C->top()); 2600 head_clone->set_req(LoopNode::LoopBackControl, C->top()); 2601 _igvn._worklist.push(head); 2602 _igvn._worklist.push(head_clone); 2603 2604 // Similarly modify the phis 2605 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) { 2606 Node* use = head->fast_out(k); 2607 if (use->is_Phi() && use->outcnt() > 0) { 2608 Node* use_clone = old_new[use->_idx]; 2609 _igvn.hash_delete(use); 2610 _igvn.hash_delete(use_clone); 2611 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl)); 2612 use->set_req(LoopNode::LoopBackControl, C->top()); 2613 use_clone->set_req(LoopNode::LoopBackControl, C->top()); 2614 _igvn._worklist.push(use); 2615 _igvn._worklist.push(use_clone); 2616 } 2617 } 2618 2619 // Step 4: update dominator tree and dominator depth 2620 2621 set_idom(head, orig_tail_clone, dd); 2622 recompute_dom_depth(); 2623 2624 // Inhibit more partial peeling on this loop 2625 new_head_clone->set_partial_peel_loop(); 2626 C->set_major_progress(); 2627 2628 #if !defined(PRODUCT) 2629 if (TracePartialPeeling) { 2630 tty->print_cr("\nafter partial peel one iteration"); 2631 Node_List wl(area); 2632 Node* t = last_peel; 2633 while (true) { 2634 wl.push(t); 2635 if (t == head_clone) break; 2636 t = idom(t); 2637 } 2638 while (wl.size() > 0) { 2639 Node* tt = wl.pop(); 2640 if (tt == head) tty->print_cr("orig head"); 2641 else if (tt == new_head_clone) tty->print_cr("new head"); 2642 else if (tt == head_clone) tty->print_cr("clone head"); 2643 tt->dump(); 2644 } 2645 } 2646 #endif 2647 return true; 2648 } 2649 2650 //------------------------------reorg_offsets---------------------------------- 2651 // Reorganize offset computations to lower register pressure. Mostly 2652 // prevent loop-fallout uses of the pre-incremented trip counter (which are 2653 // then alive with the post-incremented trip counter forcing an extra 2654 // register move) 2655 void PhaseIdealLoop::reorg_offsets( IdealLoopTree *loop ) { 2656 2657 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 2658 CountedLoopEndNode *cle = cl->loopexit(); 2659 if( !cle ) return; // The occasional dead loop 2660 // Find loop exit control 2661 Node *exit = cle->proj_out(false); 2662 assert( exit->Opcode() == Op_IfFalse, "" ); 2663 2664 // Check for the special case of folks using the pre-incremented 2665 // trip-counter on the fall-out path (forces the pre-incremented 2666 // and post-incremented trip counter to be live at the same time). 2667 // Fix this by adjusting to use the post-increment trip counter. 2668 Node *phi = cl->phi(); 2669 if( !phi ) return; // Dead infinite loop 2670 2671 // Shape messed up, probably by iteration_split_impl 2672 if (phi->in(LoopNode::LoopBackControl) != cl->incr()) return; 2673 2674 bool progress = true; 2675 while (progress) { 2676 progress = false; 2677 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) { 2678 Node* use = phi->fast_out(i); // User of trip-counter 2679 if (!has_ctrl(use)) continue; 2680 Node *u_ctrl = get_ctrl(use); 2681 if( use->is_Phi() ) { 2682 u_ctrl = NULL; 2683 for( uint j = 1; j < use->req(); j++ ) 2684 if( use->in(j) == phi ) 2685 u_ctrl = dom_lca( u_ctrl, use->in(0)->in(j) ); 2686 } 2687 IdealLoopTree *u_loop = get_loop(u_ctrl); 2688 // Look for loop-invariant use 2689 if( u_loop == loop ) continue; 2690 if( loop->is_member( u_loop ) ) continue; 2691 // Check that use is live out the bottom. Assuming the trip-counter 2692 // update is right at the bottom, uses of of the loop middle are ok. 2693 if( dom_lca( exit, u_ctrl ) != exit ) continue; 2694 // protect against stride not being a constant 2695 if( !cle->stride_is_con() ) continue; 2696 // Hit! Refactor use to use the post-incremented tripcounter. 2697 // Compute a post-increment tripcounter. 2698 Node *opaq = new (C, 2) Opaque2Node( C, cle->incr() ); 2699 register_new_node( opaq, u_ctrl ); 2700 Node *neg_stride = _igvn.intcon(-cle->stride_con()); 2701 set_ctrl(neg_stride, C->root()); 2702 Node *post = new (C, 3) AddINode( opaq, neg_stride); 2703 register_new_node( post, u_ctrl ); 2704 _igvn.hash_delete(use); 2705 _igvn._worklist.push(use); 2706 for( uint j = 1; j < use->req(); j++ ) 2707 if( use->in(j) == phi ) 2708 use->set_req(j, post); 2709 // Since DU info changed, rerun loop 2710 progress = true; 2711 break; 2712 } 2713 } 2714 2715 } 2716