1 /* 2 * Copyright (c) 1999, 2010, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "incls/_precompiled.incl" 26 #include "incls/_loopopts.cpp.incl" 27 28 //============================================================================= 29 //------------------------------split_thru_phi--------------------------------- 30 // Split Node 'n' through merge point if there is enough win. 31 Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) { 32 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) { 33 // ConvI2L may have type information on it which is unsafe to push up 34 // so disable this for now 35 return NULL; 36 } 37 int wins = 0; 38 assert( !n->is_CFG(), "" ); 39 assert( region->is_Region(), "" ); 40 41 const Type* type = n->bottom_type(); 42 const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr(); 43 Node *phi; 44 if( t_oop != NULL && t_oop->is_known_instance_field() ) { 45 int iid = t_oop->instance_id(); 46 int index = C->get_alias_index(t_oop); 47 int offset = t_oop->offset(); 48 phi = new (C,region->req()) PhiNode(region, type, NULL, iid, index, offset); 49 } else { 50 phi = PhiNode::make_blank(region, n); 51 } 52 uint old_unique = C->unique(); 53 for( uint i = 1; i < region->req(); i++ ) { 54 Node *x; 55 Node* the_clone = NULL; 56 if( region->in(i) == C->top() ) { 57 x = C->top(); // Dead path? Use a dead data op 58 } else { 59 x = n->clone(); // Else clone up the data op 60 the_clone = x; // Remember for possible deletion. 61 // Alter data node to use pre-phi inputs 62 if( n->in(0) == region ) 63 x->set_req( 0, region->in(i) ); 64 for( uint j = 1; j < n->req(); j++ ) { 65 Node *in = n->in(j); 66 if( in->is_Phi() && in->in(0) == region ) 67 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone 68 } 69 } 70 // Check for a 'win' on some paths 71 const Type *t = x->Value(&_igvn); 72 73 bool singleton = t->singleton(); 74 75 // A TOP singleton indicates that there are no possible values incoming 76 // along a particular edge. In most cases, this is OK, and the Phi will 77 // be eliminated later in an Ideal call. However, we can't allow this to 78 // happen if the singleton occurs on loop entry, as the elimination of 79 // the PhiNode may cause the resulting node to migrate back to a previous 80 // loop iteration. 81 if( singleton && t == Type::TOP ) { 82 // Is_Loop() == false does not confirm the absence of a loop (e.g., an 83 // irreducible loop may not be indicated by an affirmative is_Loop()); 84 // therefore, the only top we can split thru a phi is on a backedge of 85 // a loop. 86 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 87 } 88 89 if( singleton ) { 90 wins++; 91 x = ((PhaseGVN&)_igvn).makecon(t); 92 } else { 93 // We now call Identity to try to simplify the cloned node. 94 // Note that some Identity methods call phase->type(this). 95 // Make sure that the type array is big enough for 96 // our new node, even though we may throw the node away. 97 // (Note: This tweaking with igvn only works because x is a new node.) 98 _igvn.set_type(x, t); 99 // If x is a TypeNode, capture any more-precise type permanently into Node 100 // otherwise it will be not updated during igvn->transform since 101 // igvn->type(x) is set to x->Value() already. 102 x->raise_bottom_type(t); 103 Node *y = x->Identity(&_igvn); 104 if( y != x ) { 105 wins++; 106 x = y; 107 } else { 108 y = _igvn.hash_find(x); 109 if( y ) { 110 wins++; 111 x = y; 112 } else { 113 // Else x is a new node we are keeping 114 // We do not need register_new_node_with_optimizer 115 // because set_type has already been called. 116 _igvn._worklist.push(x); 117 } 118 } 119 } 120 if (x != the_clone && the_clone != NULL) 121 _igvn.remove_dead_node(the_clone); 122 phi->set_req( i, x ); 123 } 124 // Too few wins? 125 if( wins <= policy ) { 126 _igvn.remove_dead_node(phi); 127 return NULL; 128 } 129 130 // Record Phi 131 register_new_node( phi, region ); 132 133 for( uint i2 = 1; i2 < phi->req(); i2++ ) { 134 Node *x = phi->in(i2); 135 // If we commoned up the cloned 'x' with another existing Node, 136 // the existing Node picks up a new use. We need to make the 137 // existing Node occur higher up so it dominates its uses. 138 Node *old_ctrl; 139 IdealLoopTree *old_loop; 140 141 // The occasional new node 142 if( x->_idx >= old_unique ) { // Found a new, unplaced node? 143 old_ctrl = x->is_Con() ? C->root() : NULL; 144 old_loop = NULL; // Not in any prior loop 145 } else { 146 old_ctrl = x->is_Con() ? C->root() : get_ctrl(x); 147 old_loop = get_loop(old_ctrl); // Get prior loop 148 } 149 // New late point must dominate new use 150 Node *new_ctrl = dom_lca( old_ctrl, region->in(i2) ); 151 // Set new location 152 set_ctrl(x, new_ctrl); 153 IdealLoopTree *new_loop = get_loop( new_ctrl ); 154 // If changing loop bodies, see if we need to collect into new body 155 if( old_loop != new_loop ) { 156 if( old_loop && !old_loop->_child ) 157 old_loop->_body.yank(x); 158 if( !new_loop->_child ) 159 new_loop->_body.push(x); // Collect body info 160 } 161 } 162 163 return phi; 164 } 165 166 //------------------------------dominated_by------------------------------------ 167 // Replace the dominated test with an obvious true or false. Place it on the 168 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the 169 // live path up to the dominating control. 170 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff ) { 171 #ifndef PRODUCT 172 if( VerifyLoopOptimizations && PrintOpto ) tty->print_cr("dominating test"); 173 #endif 174 175 176 // prevdom is the dominating projection of the dominating test. 177 assert( iff->is_If(), "" ); 178 assert( iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added"); 179 int pop = prevdom->Opcode(); 180 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" ); 181 // 'con' is set to true or false to kill the dominated test. 182 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO); 183 set_ctrl(con, C->root()); // Constant gets a new use 184 // Hack the dominated test 185 _igvn.hash_delete(iff); 186 iff->set_req(1, con); 187 _igvn._worklist.push(iff); 188 189 // If I dont have a reachable TRUE and FALSE path following the IfNode then 190 // I can assume this path reaches an infinite loop. In this case it's not 191 // important to optimize the data Nodes - either the whole compilation will 192 // be tossed or this path (and all data Nodes) will go dead. 193 if( iff->outcnt() != 2 ) return; 194 195 // Make control-dependent data Nodes on the live path (path that will remain 196 // once the dominated IF is removed) become control-dependent on the 197 // dominating projection. 198 Node* dp = ((IfNode*)iff)->proj_out(pop == Op_IfTrue); 199 IdealLoopTree *old_loop = get_loop(dp); 200 201 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { 202 Node* cd = dp->fast_out(i); // Control-dependent node 203 if( cd->depends_only_on_test() ) { 204 assert( cd->in(0) == dp, "" ); 205 _igvn.hash_delete( cd ); 206 cd->set_req(0, prevdom); 207 set_early_ctrl( cd ); 208 _igvn._worklist.push(cd); 209 IdealLoopTree *new_loop = get_loop(get_ctrl(cd)); 210 if( old_loop != new_loop ) { 211 if( !old_loop->_child ) old_loop->_body.yank(cd); 212 if( !new_loop->_child ) new_loop->_body.push(cd); 213 } 214 --i; 215 --imax; 216 } 217 } 218 } 219 220 //------------------------------has_local_phi_input---------------------------- 221 // Return TRUE if 'n' has Phi inputs from its local block and no other 222 // block-local inputs (all non-local-phi inputs come from earlier blocks) 223 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) { 224 Node *n_ctrl = get_ctrl(n); 225 // See if some inputs come from a Phi in this block, or from before 226 // this block. 227 uint i; 228 for( i = 1; i < n->req(); i++ ) { 229 Node *phi = n->in(i); 230 if( phi->is_Phi() && phi->in(0) == n_ctrl ) 231 break; 232 } 233 if( i >= n->req() ) 234 return NULL; // No Phi inputs; nowhere to clone thru 235 236 // Check for inputs created between 'n' and the Phi input. These 237 // must split as well; they have already been given the chance 238 // (courtesy of a post-order visit) and since they did not we must 239 // recover the 'cost' of splitting them by being very profitable 240 // when splitting 'n'. Since this is unlikely we simply give up. 241 for( i = 1; i < n->req(); i++ ) { 242 Node *m = n->in(i); 243 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) { 244 // We allow the special case of AddP's with no local inputs. 245 // This allows us to split-up address expressions. 246 if (m->is_AddP() && 247 get_ctrl(m->in(2)) != n_ctrl && 248 get_ctrl(m->in(3)) != n_ctrl) { 249 // Move the AddP up to dominating point 250 set_ctrl_and_loop(m, find_non_split_ctrl(idom(n_ctrl))); 251 continue; 252 } 253 return NULL; 254 } 255 } 256 257 return n_ctrl; 258 } 259 260 //------------------------------remix_address_expressions---------------------- 261 // Rework addressing expressions to get the most loop-invariant stuff 262 // moved out. We'd like to do all associative operators, but it's especially 263 // important (common) to do address expressions. 264 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) { 265 if (!has_ctrl(n)) return NULL; 266 Node *n_ctrl = get_ctrl(n); 267 IdealLoopTree *n_loop = get_loop(n_ctrl); 268 269 // See if 'n' mixes loop-varying and loop-invariant inputs and 270 // itself is loop-varying. 271 272 // Only interested in binary ops (and AddP) 273 if( n->req() < 3 || n->req() > 4 ) return NULL; 274 275 Node *n1_ctrl = get_ctrl(n->in( 1)); 276 Node *n2_ctrl = get_ctrl(n->in( 2)); 277 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3)); 278 IdealLoopTree *n1_loop = get_loop( n1_ctrl ); 279 IdealLoopTree *n2_loop = get_loop( n2_ctrl ); 280 IdealLoopTree *n3_loop = get_loop( n3_ctrl ); 281 282 // Does one of my inputs spin in a tighter loop than self? 283 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) || 284 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) || 285 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) ) 286 return NULL; // Leave well enough alone 287 288 // Is at least one of my inputs loop-invariant? 289 if( n1_loop == n_loop && 290 n2_loop == n_loop && 291 n3_loop == n_loop ) 292 return NULL; // No loop-invariant inputs 293 294 295 int n_op = n->Opcode(); 296 297 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2). 298 if( n_op == Op_LShiftI ) { 299 // Scale is loop invariant 300 Node *scale = n->in(2); 301 Node *scale_ctrl = get_ctrl(scale); 302 IdealLoopTree *scale_loop = get_loop(scale_ctrl ); 303 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) ) 304 return NULL; 305 const TypeInt *scale_t = scale->bottom_type()->isa_int(); 306 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 ) 307 return NULL; // Dont bother with byte/short masking 308 // Add must vary with loop (else shift would be loop-invariant) 309 Node *add = n->in(1); 310 Node *add_ctrl = get_ctrl(add); 311 IdealLoopTree *add_loop = get_loop(add_ctrl); 312 //assert( n_loop == add_loop, "" ); 313 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops 314 315 // Convert I-V into I+ (0-V); same for V-I 316 if( add->Opcode() == Op_SubI && 317 _igvn.type( add->in(1) ) != TypeInt::ZERO ) { 318 Node *zero = _igvn.intcon(0); 319 set_ctrl(zero, C->root()); 320 Node *neg = new (C, 3) SubINode( _igvn.intcon(0), add->in(2) ); 321 register_new_node( neg, get_ctrl(add->in(2) ) ); 322 add = new (C, 3) AddINode( add->in(1), neg ); 323 register_new_node( add, add_ctrl ); 324 } 325 if( add->Opcode() != Op_AddI ) return NULL; 326 // See if one add input is loop invariant 327 Node *add_var = add->in(1); 328 Node *add_var_ctrl = get_ctrl(add_var); 329 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 330 Node *add_invar = add->in(2); 331 Node *add_invar_ctrl = get_ctrl(add_invar); 332 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl ); 333 if( add_var_loop == n_loop ) { 334 } else if( add_invar_loop == n_loop ) { 335 // Swap to find the invariant part 336 add_invar = add_var; 337 add_invar_ctrl = add_var_ctrl; 338 add_invar_loop = add_var_loop; 339 add_var = add->in(2); 340 Node *add_var_ctrl = get_ctrl(add_var); 341 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 342 } else // Else neither input is loop invariant 343 return NULL; 344 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) ) 345 return NULL; // No invariant part of the add? 346 347 // Yes! Reshape address expression! 348 Node *inv_scale = new (C, 3) LShiftINode( add_invar, scale ); 349 Node *inv_scale_ctrl = 350 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ? 351 add_invar_ctrl : scale_ctrl; 352 register_new_node( inv_scale, inv_scale_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.replace_node( n, var_add ); 358 return var_add; 359 } 360 361 // Replace (I+V) with (V+I) 362 if( n_op == Op_AddI || 363 n_op == Op_AddL || 364 n_op == Op_AddF || 365 n_op == Op_AddD || 366 n_op == Op_MulI || 367 n_op == Op_MulL || 368 n_op == Op_MulF || 369 n_op == Op_MulD ) { 370 if( n2_loop == n_loop ) { 371 assert( n1_loop != n_loop, "" ); 372 n->swap_edges(1, 2); 373 } 374 } 375 376 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V), 377 // but not if I2 is a constant. 378 if( n_op == Op_AddP ) { 379 if( n2_loop == n_loop && n3_loop != n_loop ) { 380 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) { 381 Node *n22_ctrl = get_ctrl(n->in(2)->in(2)); 382 Node *n23_ctrl = get_ctrl(n->in(2)->in(3)); 383 IdealLoopTree *n22loop = get_loop( n22_ctrl ); 384 IdealLoopTree *n23_loop = get_loop( n23_ctrl ); 385 if( n22loop != n_loop && n22loop->is_member(n_loop) && 386 n23_loop == n_loop ) { 387 Node *add1 = new (C, 4) AddPNode( n->in(1), n->in(2)->in(2), n->in(3) ); 388 // Stuff new AddP in the loop preheader 389 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) ); 390 Node *add2 = new (C, 4) AddPNode( n->in(1), add1, n->in(2)->in(3) ); 391 register_new_node( add2, n_ctrl ); 392 _igvn.replace_node( n, add2 ); 393 return add2; 394 } 395 } 396 } 397 398 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V) 399 if( n2_loop != n_loop && n3_loop == n_loop ) { 400 if( n->in(3)->Opcode() == Op_AddI ) { 401 Node *V = n->in(3)->in(1); 402 Node *I = n->in(3)->in(2); 403 if( is_member(n_loop,get_ctrl(V)) ) { 404 } else { 405 Node *tmp = V; V = I; I = tmp; 406 } 407 if( !is_member(n_loop,get_ctrl(I)) ) { 408 Node *add1 = new (C, 4) AddPNode( n->in(1), n->in(2), I ); 409 // Stuff new AddP in the loop preheader 410 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) ); 411 Node *add2 = new (C, 4) AddPNode( n->in(1), add1, V ); 412 register_new_node( add2, n_ctrl ); 413 _igvn.replace_node( n, add2 ); 414 return add2; 415 } 416 } 417 } 418 } 419 420 return NULL; 421 } 422 423 //------------------------------conditional_move------------------------------- 424 // Attempt to replace a Phi with a conditional move. We have some pretty 425 // strict profitability requirements. All Phis at the merge point must 426 // be converted, so we can remove the control flow. We need to limit the 427 // number of c-moves to a small handful. All code that was in the side-arms 428 // of the CFG diamond is now speculatively executed. This code has to be 429 // "cheap enough". We are pretty much limited to CFG diamonds that merge 430 // 1 or 2 items with a total of 1 or 2 ops executed speculatively. 431 Node *PhaseIdealLoop::conditional_move( Node *region ) { 432 433 assert( region->is_Region(), "sanity check" ); 434 if( region->req() != 3 ) return NULL; 435 436 // Check for CFG diamond 437 Node *lp = region->in(1); 438 Node *rp = region->in(2); 439 if( !lp || !rp ) return NULL; 440 Node *lp_c = lp->in(0); 441 if( lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If() ) return NULL; 442 IfNode *iff = lp_c->as_If(); 443 444 // Check for highly predictable branch. No point in CMOV'ing if 445 // we are going to predict accurately all the time. 446 // %%% This hides patterns produced by utility methods like Math.min. 447 if( iff->_prob < PROB_UNLIKELY_MAG(3) || 448 iff->_prob > PROB_LIKELY_MAG(3) ) 449 return NULL; 450 451 // Check for ops pinned in an arm of the diamond. 452 // Can't remove the control flow in this case 453 if( lp->outcnt() > 1 ) return NULL; 454 if( rp->outcnt() > 1 ) return NULL; 455 456 // Check profitability 457 int cost = 0; 458 int phis = 0; 459 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 460 Node *out = region->fast_out(i); 461 if( !out->is_Phi() ) continue; // Ignore other control edges, etc 462 phis++; 463 PhiNode* phi = out->as_Phi(); 464 switch (phi->type()->basic_type()) { 465 case T_LONG: 466 cost++; // Probably encodes as 2 CMOV's 467 case T_INT: // These all CMOV fine 468 case T_FLOAT: 469 case T_DOUBLE: 470 case T_ADDRESS: // (RawPtr) 471 cost++; 472 break; 473 case T_NARROWOOP: // Fall through 474 case T_OBJECT: { // Base oops are OK, but not derived oops 475 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr(); 476 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a 477 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus 478 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we 479 // have a Phi for the base here that we convert to a CMOVE all is well 480 // and good. But if the base is dead, we'll not make a CMOVE. Later 481 // the allocator will have to produce a base by creating a CMOVE of the 482 // relevant bases. This puts the allocator in the business of 483 // manufacturing expensive instructions, generally a bad plan. 484 // Just Say No to Conditionally-Moved Derived Pointers. 485 if( tp && tp->offset() != 0 ) 486 return NULL; 487 cost++; 488 break; 489 } 490 default: 491 return NULL; // In particular, can't do memory or I/O 492 } 493 // Add in cost any speculative ops 494 for( uint j = 1; j < region->req(); j++ ) { 495 Node *proj = region->in(j); 496 Node *inp = phi->in(j); 497 if (get_ctrl(inp) == proj) { // Found local op 498 cost++; 499 // Check for a chain of dependent ops; these will all become 500 // speculative in a CMOV. 501 for( uint k = 1; k < inp->req(); k++ ) 502 if (get_ctrl(inp->in(k)) == proj) 503 return NULL; // Too much speculative goo 504 } 505 } 506 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode. 507 // This will likely Split-If, a higher-payoff operation. 508 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) { 509 Node* use = phi->fast_out(k); 510 if( use->is_Cmp() || use->is_DecodeN() || use->is_EncodeP() ) 511 return NULL; 512 } 513 } 514 if( cost >= ConditionalMoveLimit ) return NULL; // Too much goo 515 Node* bol = iff->in(1); 516 assert( bol->Opcode() == Op_Bool, "" ); 517 int cmp_op = bol->in(1)->Opcode(); 518 // It is expensive to generate flags from a float compare. 519 // Avoid duplicated float compare. 520 if( phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL; 521 522 // -------------- 523 // Now replace all Phis with CMOV's 524 Node *cmov_ctrl = iff->in(0); 525 uint flip = (lp->Opcode() == Op_IfTrue); 526 while( 1 ) { 527 PhiNode* phi = NULL; 528 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 529 Node *out = region->fast_out(i); 530 if (out->is_Phi()) { 531 phi = out->as_Phi(); 532 break; 533 } 534 } 535 if (phi == NULL) break; 536 #ifndef PRODUCT 537 if( PrintOpto && VerifyLoopOptimizations ) tty->print_cr("CMOV"); 538 #endif 539 // Move speculative ops 540 for( uint j = 1; j < region->req(); j++ ) { 541 Node *proj = region->in(j); 542 Node *inp = phi->in(j); 543 if (get_ctrl(inp) == proj) { // Found local op 544 #ifndef PRODUCT 545 if( PrintOpto && VerifyLoopOptimizations ) { 546 tty->print(" speculate: "); 547 inp->dump(); 548 } 549 #endif 550 set_ctrl(inp, cmov_ctrl); 551 } 552 } 553 Node *cmov = CMoveNode::make( C, cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi) ); 554 register_new_node( cmov, cmov_ctrl ); 555 _igvn.replace_node( phi, cmov ); 556 #ifndef PRODUCT 557 if( VerifyLoopOptimizations ) verify(); 558 #endif 559 } 560 561 // The useless CFG diamond will fold up later; see the optimization in 562 // RegionNode::Ideal. 563 _igvn._worklist.push(region); 564 565 return iff->in(1); 566 } 567 568 //------------------------------split_if_with_blocks_pre----------------------- 569 // Do the real work in a non-recursive function. Data nodes want to be 570 // cloned in the pre-order so they can feed each other nicely. 571 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) { 572 // Cloning these guys is unlikely to win 573 int n_op = n->Opcode(); 574 if( n_op == Op_MergeMem ) return n; 575 if( n->is_Proj() ) return n; 576 // Do not clone-up CmpFXXX variations, as these are always 577 // followed by a CmpI 578 if( n->is_Cmp() ) return n; 579 // Attempt to use a conditional move instead of a phi/branch 580 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) { 581 Node *cmov = conditional_move( n ); 582 if( cmov ) return cmov; 583 } 584 if( n->is_CFG() || n->is_LoadStore() ) 585 return n; 586 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd 587 n_op == Op_Opaque2 ) { 588 if( !C->major_progress() ) // If chance of no more loop opts... 589 _igvn._worklist.push(n); // maybe we'll remove them 590 return n; 591 } 592 593 if( n->is_Con() ) return n; // No cloning for Con nodes 594 595 Node *n_ctrl = get_ctrl(n); 596 if( !n_ctrl ) return n; // Dead node 597 598 // Attempt to remix address expressions for loop invariants 599 Node *m = remix_address_expressions( n ); 600 if( m ) return m; 601 602 // Determine if the Node has inputs from some local Phi. 603 // Returns the block to clone thru. 604 Node *n_blk = has_local_phi_input( n ); 605 if( !n_blk ) return n; 606 // Do not clone the trip counter through on a CountedLoop 607 // (messes up the canonical shape). 608 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n; 609 610 // Check for having no control input; not pinned. Allow 611 // dominating control. 612 if( n->in(0) ) { 613 Node *dom = idom(n_blk); 614 if( dom_lca( n->in(0), dom ) != n->in(0) ) 615 return n; 616 } 617 // Policy: when is it profitable. You must get more wins than 618 // policy before it is considered profitable. Policy is usually 0, 619 // so 1 win is considered profitable. Big merges will require big 620 // cloning, so get a larger policy. 621 int policy = n_blk->req() >> 2; 622 623 // If the loop is a candidate for range check elimination, 624 // delay splitting through it's phi until a later loop optimization 625 if (n_blk->is_CountedLoop()) { 626 IdealLoopTree *lp = get_loop(n_blk); 627 if (lp && lp->_rce_candidate) { 628 return n; 629 } 630 } 631 632 // Use same limit as split_if_with_blocks_post 633 if( C->unique() > 35000 ) return n; // Method too big 634 635 // Split 'n' through the merge point if it is profitable 636 Node *phi = split_thru_phi( n, n_blk, policy ); 637 if( !phi ) return n; 638 639 // Found a Phi to split thru! 640 // Replace 'n' with the new phi 641 _igvn.replace_node( n, phi ); 642 // Moved a load around the loop, 'en-registering' something. 643 if( n_blk->Opcode() == Op_Loop && n->is_Load() && 644 !phi->in(LoopNode::LoopBackControl)->is_Load() ) 645 C->set_major_progress(); 646 647 return phi; 648 } 649 650 static bool merge_point_too_heavy(Compile* C, Node* region) { 651 // Bail out if the region and its phis have too many users. 652 int weight = 0; 653 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 654 weight += region->fast_out(i)->outcnt(); 655 } 656 int nodes_left = MaxNodeLimit - C->unique(); 657 if (weight * 8 > nodes_left) { 658 #ifndef PRODUCT 659 if (PrintOpto) 660 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight); 661 #endif 662 return true; 663 } else { 664 return false; 665 } 666 } 667 668 static bool merge_point_safe(Node* region) { 669 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode 670 // having a PhiNode input. This sidesteps the dangerous case where the split 671 // ConvI2LNode may become TOP if the input Value() does not 672 // overlap the ConvI2L range, leaving a node which may not dominate its 673 // uses. 674 // A better fix for this problem can be found in the BugTraq entry, but 675 // expediency for Mantis demands this hack. 676 // 6855164: If the merge point has a FastLockNode with a PhiNode input, we stop 677 // split_if_with_blocks from splitting a block because we could not move around 678 // the FastLockNode. 679 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 680 Node* n = region->fast_out(i); 681 if (n->is_Phi()) { 682 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 683 Node* m = n->fast_out(j); 684 if (m->is_FastLock()) 685 return false; 686 #ifdef _LP64 687 if (m->Opcode() == Op_ConvI2L) 688 return false; 689 #endif 690 } 691 } 692 } 693 return true; 694 } 695 696 697 //------------------------------place_near_use--------------------------------- 698 // Place some computation next to use but not inside inner loops. 699 // For inner loop uses move it to the preheader area. 700 Node *PhaseIdealLoop::place_near_use( Node *useblock ) const { 701 IdealLoopTree *u_loop = get_loop( useblock ); 702 return (u_loop->_irreducible || u_loop->_child) 703 ? useblock 704 : u_loop->_head->in(LoopNode::EntryControl); 705 } 706 707 708 //------------------------------split_if_with_blocks_post---------------------- 709 // Do the real work in a non-recursive function. CFG hackery wants to be 710 // in the post-order, so it can dirty the I-DOM info and not use the dirtied 711 // info. 712 void PhaseIdealLoop::split_if_with_blocks_post( Node *n ) { 713 714 // Cloning Cmp through Phi's involves the split-if transform. 715 // FastLock is not used by an If 716 if( n->is_Cmp() && !n->is_FastLock() ) { 717 if( C->unique() > 35000 ) return; // Method too big 718 719 // Do not do 'split-if' if irreducible loops are present. 720 if( _has_irreducible_loops ) 721 return; 722 723 Node *n_ctrl = get_ctrl(n); 724 // Determine if the Node has inputs from some local Phi. 725 // Returns the block to clone thru. 726 Node *n_blk = has_local_phi_input( n ); 727 if( n_blk != n_ctrl ) return; 728 729 if( merge_point_too_heavy(C, n_ctrl) ) 730 return; 731 732 if( n->outcnt() != 1 ) return; // Multiple bool's from 1 compare? 733 Node *bol = n->unique_out(); 734 assert( bol->is_Bool(), "expect a bool here" ); 735 if( bol->outcnt() != 1 ) return;// Multiple branches from 1 compare? 736 Node *iff = bol->unique_out(); 737 738 // Check some safety conditions 739 if( iff->is_If() ) { // Classic split-if? 740 if( iff->in(0) != n_ctrl ) return; // Compare must be in same blk as if 741 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE 742 if( get_ctrl(iff->in(2)) == n_ctrl || 743 get_ctrl(iff->in(3)) == n_ctrl ) 744 return; // Inputs not yet split-up 745 if ( get_loop(n_ctrl) != get_loop(get_ctrl(iff)) ) { 746 return; // Loop-invar test gates loop-varying CMOVE 747 } 748 } else { 749 return; // some other kind of node, such as an Allocate 750 } 751 752 // Do not do 'split-if' if some paths are dead. First do dead code 753 // elimination and then see if its still profitable. 754 for( uint i = 1; i < n_ctrl->req(); i++ ) 755 if( n_ctrl->in(i) == C->top() ) 756 return; 757 758 // When is split-if profitable? Every 'win' on means some control flow 759 // goes dead, so it's almost always a win. 760 int policy = 0; 761 // If trying to do a 'Split-If' at the loop head, it is only 762 // profitable if the cmp folds up on BOTH paths. Otherwise we 763 // risk peeling a loop forever. 764 765 // CNC - Disabled for now. Requires careful handling of loop 766 // body selection for the cloned code. Also, make sure we check 767 // for any input path not being in the same loop as n_ctrl. For 768 // irreducible loops we cannot check for 'n_ctrl->is_Loop()' 769 // because the alternative loop entry points won't be converted 770 // into LoopNodes. 771 IdealLoopTree *n_loop = get_loop(n_ctrl); 772 for( uint j = 1; j < n_ctrl->req(); j++ ) 773 if( get_loop(n_ctrl->in(j)) != n_loop ) 774 return; 775 776 // Check for safety of the merge point. 777 if( !merge_point_safe(n_ctrl) ) { 778 return; 779 } 780 781 // Split compare 'n' through the merge point if it is profitable 782 Node *phi = split_thru_phi( n, n_ctrl, policy ); 783 if( !phi ) return; 784 785 // Found a Phi to split thru! 786 // Replace 'n' with the new phi 787 _igvn.replace_node( n, phi ); 788 789 // Now split the bool up thru the phi 790 Node *bolphi = split_thru_phi( bol, n_ctrl, -1 ); 791 _igvn.replace_node( bol, bolphi ); 792 assert( iff->in(1) == bolphi, "" ); 793 if( bolphi->Value(&_igvn)->singleton() ) 794 return; 795 796 // Conditional-move? Must split up now 797 if( !iff->is_If() ) { 798 Node *cmovphi = split_thru_phi( iff, n_ctrl, -1 ); 799 _igvn.replace_node( iff, cmovphi ); 800 return; 801 } 802 803 // Now split the IF 804 do_split_if( iff ); 805 return; 806 } 807 808 // Check for an IF ready to split; one that has its 809 // condition codes input coming from a Phi at the block start. 810 int n_op = n->Opcode(); 811 812 // Check for an IF being dominated by another IF same test 813 if( n_op == Op_If ) { 814 Node *bol = n->in(1); 815 uint max = bol->outcnt(); 816 // Check for same test used more than once? 817 if( n_op == Op_If && max > 1 && bol->is_Bool() ) { 818 // Search up IDOMs to see if this IF is dominated. 819 Node *cutoff = get_ctrl(bol); 820 821 // Now search up IDOMs till cutoff, looking for a dominating test 822 Node *prevdom = n; 823 Node *dom = idom(prevdom); 824 while( dom != cutoff ) { 825 if( dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom ) { 826 // Replace the dominated test with an obvious true or false. 827 // Place it on the IGVN worklist for later cleanup. 828 C->set_major_progress(); 829 dominated_by( prevdom, n ); 830 #ifndef PRODUCT 831 if( VerifyLoopOptimizations ) verify(); 832 #endif 833 return; 834 } 835 prevdom = dom; 836 dom = idom(prevdom); 837 } 838 } 839 } 840 841 // See if a shared loop-varying computation has no loop-varying uses. 842 // Happens if something is only used for JVM state in uncommon trap exits, 843 // like various versions of induction variable+offset. Clone the 844 // computation per usage to allow it to sink out of the loop. 845 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about) 846 Node *n_ctrl = get_ctrl(n); 847 IdealLoopTree *n_loop = get_loop(n_ctrl); 848 if( n_loop != _ltree_root ) { 849 DUIterator_Fast imax, i = n->fast_outs(imax); 850 for (; i < imax; i++) { 851 Node* u = n->fast_out(i); 852 if( !has_ctrl(u) ) break; // Found control user 853 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 854 if( u_loop == n_loop ) break; // Found loop-varying use 855 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop 856 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003 857 } 858 bool did_break = (i < imax); // Did we break out of the previous loop? 859 if (!did_break && n->outcnt() > 1) { // All uses in outer loops! 860 Node *late_load_ctrl; 861 if (n->is_Load()) { 862 // If n is a load, get and save the result from get_late_ctrl(), 863 // to be later used in calculating the control for n's clones. 864 clear_dom_lca_tags(); 865 late_load_ctrl = get_late_ctrl(n, n_ctrl); 866 } 867 // If n is a load, and the late control is the same as the current 868 // control, then the cloning of n is a pointless exercise, because 869 // GVN will ensure that we end up where we started. 870 if (!n->is_Load() || late_load_ctrl != n_ctrl) { 871 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) { 872 Node *u = n->last_out(j); // Clone private computation per use 873 _igvn.hash_delete(u); 874 _igvn._worklist.push(u); 875 Node *x = n->clone(); // Clone computation 876 Node *x_ctrl = NULL; 877 if( u->is_Phi() ) { 878 // Replace all uses of normal nodes. Replace Phi uses 879 // individually, so the separate Nodes can sink down 880 // different paths. 881 uint k = 1; 882 while( u->in(k) != n ) k++; 883 u->set_req( k, x ); 884 // x goes next to Phi input path 885 x_ctrl = u->in(0)->in(k); 886 --j; 887 } else { // Normal use 888 // Replace all uses 889 for( uint k = 0; k < u->req(); k++ ) { 890 if( u->in(k) == n ) { 891 u->set_req( k, x ); 892 --j; 893 } 894 } 895 x_ctrl = get_ctrl(u); 896 } 897 898 // Find control for 'x' next to use but not inside inner loops. 899 // For inner loop uses get the preheader area. 900 x_ctrl = place_near_use(x_ctrl); 901 902 if (n->is_Load()) { 903 // For loads, add a control edge to a CFG node outside of the loop 904 // to force them to not combine and return back inside the loop 905 // during GVN optimization (4641526). 906 // 907 // Because we are setting the actual control input, factor in 908 // the result from get_late_ctrl() so we respect any 909 // anti-dependences. (6233005). 910 x_ctrl = dom_lca(late_load_ctrl, x_ctrl); 911 912 // Don't allow the control input to be a CFG splitting node. 913 // Such nodes should only have ProjNodes as outs, e.g. IfNode 914 // should only have IfTrueNode and IfFalseNode (4985384). 915 x_ctrl = find_non_split_ctrl(x_ctrl); 916 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone"); 917 918 x->set_req(0, x_ctrl); 919 } 920 register_new_node(x, x_ctrl); 921 922 // Some institutional knowledge is needed here: 'x' is 923 // yanked because if the optimizer runs GVN on it all the 924 // cloned x's will common up and undo this optimization and 925 // be forced back in the loop. This is annoying because it 926 // makes +VerifyOpto report false-positives on progress. I 927 // tried setting control edges on the x's to force them to 928 // not combine, but the matching gets worried when it tries 929 // to fold a StoreP and an AddP together (as part of an 930 // address expression) and the AddP and StoreP have 931 // different controls. 932 if( !x->is_Load() && !x->is_DecodeN() ) _igvn._worklist.yank(x); 933 } 934 _igvn.remove_dead_node(n); 935 } 936 } 937 } 938 } 939 940 // Check for Opaque2's who's loop has disappeared - who's input is in the 941 // same loop nest as their output. Remove 'em, they are no longer useful. 942 if( n_op == Op_Opaque2 && 943 n->in(1) != NULL && 944 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) { 945 _igvn.replace_node( n, n->in(1) ); 946 } 947 } 948 949 //------------------------------split_if_with_blocks--------------------------- 950 // Check for aggressive application of 'split-if' optimization, 951 // using basic block level info. 952 void PhaseIdealLoop::split_if_with_blocks( VectorSet &visited, Node_Stack &nstack ) { 953 Node *n = C->root(); 954 visited.set(n->_idx); // first, mark node as visited 955 // Do pre-visit work for root 956 n = split_if_with_blocks_pre( n ); 957 uint cnt = n->outcnt(); 958 uint i = 0; 959 while (true) { 960 // Visit all children 961 if (i < cnt) { 962 Node* use = n->raw_out(i); 963 ++i; 964 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) { 965 // Now do pre-visit work for this use 966 use = split_if_with_blocks_pre( use ); 967 nstack.push(n, i); // Save parent and next use's index. 968 n = use; // Process all children of current use. 969 cnt = use->outcnt(); 970 i = 0; 971 } 972 } 973 else { 974 // All of n's children have been processed, complete post-processing. 975 if (cnt != 0 && !n->is_Con()) { 976 assert(has_node(n), "no dead nodes"); 977 split_if_with_blocks_post( n ); 978 } 979 if (nstack.is_empty()) { 980 // Finished all nodes on stack. 981 break; 982 } 983 // Get saved parent node and next use's index. Visit the rest of uses. 984 n = nstack.node(); 985 cnt = n->outcnt(); 986 i = nstack.index(); 987 nstack.pop(); 988 } 989 } 990 } 991 992 993 //============================================================================= 994 // 995 // C L O N E A L O O P B O D Y 996 // 997 998 //------------------------------clone_iff-------------------------------------- 999 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1000 // "Nearly" because all Nodes have been cloned from the original in the loop, 1001 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1002 // through the Phi recursively, and return a Bool. 1003 BoolNode *PhaseIdealLoop::clone_iff( PhiNode *phi, IdealLoopTree *loop ) { 1004 1005 // Convert this Phi into a Phi merging Bools 1006 uint i; 1007 for( i = 1; i < phi->req(); i++ ) { 1008 Node *b = phi->in(i); 1009 if( b->is_Phi() ) { 1010 _igvn.hash_delete(phi); 1011 _igvn._worklist.push(phi); 1012 phi->set_req(i, clone_iff( b->as_Phi(), loop )); 1013 } else { 1014 assert( b->is_Bool(), "" ); 1015 } 1016 } 1017 1018 Node *sample_bool = phi->in(1); 1019 Node *sample_cmp = sample_bool->in(1); 1020 1021 // Make Phis to merge the Cmp's inputs. 1022 int size = phi->in(0)->req(); 1023 PhiNode *phi1 = new (C, size) PhiNode( phi->in(0), Type::TOP ); 1024 PhiNode *phi2 = new (C, size) PhiNode( phi->in(0), Type::TOP ); 1025 for( i = 1; i < phi->req(); i++ ) { 1026 Node *n1 = phi->in(i)->in(1)->in(1); 1027 Node *n2 = phi->in(i)->in(1)->in(2); 1028 phi1->set_req( i, n1 ); 1029 phi2->set_req( i, n2 ); 1030 phi1->set_type( phi1->type()->meet(n1->bottom_type()) ); 1031 phi2->set_type( phi2->type()->meet(n2->bottom_type()) ); 1032 } 1033 // See if these Phis have been made before. 1034 // Register with optimizer 1035 Node *hit1 = _igvn.hash_find_insert(phi1); 1036 if( hit1 ) { // Hit, toss just made Phi 1037 _igvn.remove_dead_node(phi1); // Remove new phi 1038 assert( hit1->is_Phi(), "" ); 1039 phi1 = (PhiNode*)hit1; // Use existing phi 1040 } else { // Miss 1041 _igvn.register_new_node_with_optimizer(phi1); 1042 } 1043 Node *hit2 = _igvn.hash_find_insert(phi2); 1044 if( hit2 ) { // Hit, toss just made Phi 1045 _igvn.remove_dead_node(phi2); // Remove new phi 1046 assert( hit2->is_Phi(), "" ); 1047 phi2 = (PhiNode*)hit2; // Use existing phi 1048 } else { // Miss 1049 _igvn.register_new_node_with_optimizer(phi2); 1050 } 1051 // Register Phis with loop/block info 1052 set_ctrl(phi1, phi->in(0)); 1053 set_ctrl(phi2, phi->in(0)); 1054 // Make a new Cmp 1055 Node *cmp = sample_cmp->clone(); 1056 cmp->set_req( 1, phi1 ); 1057 cmp->set_req( 2, phi2 ); 1058 _igvn.register_new_node_with_optimizer(cmp); 1059 set_ctrl(cmp, phi->in(0)); 1060 1061 // Make a new Bool 1062 Node *b = sample_bool->clone(); 1063 b->set_req(1,cmp); 1064 _igvn.register_new_node_with_optimizer(b); 1065 set_ctrl(b, phi->in(0)); 1066 1067 assert( b->is_Bool(), "" ); 1068 return (BoolNode*)b; 1069 } 1070 1071 //------------------------------clone_bool------------------------------------- 1072 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1073 // "Nearly" because all Nodes have been cloned from the original in the loop, 1074 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1075 // through the Phi recursively, and return a Bool. 1076 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) { 1077 uint i; 1078 // Convert this Phi into a Phi merging Bools 1079 for( i = 1; i < phi->req(); i++ ) { 1080 Node *b = phi->in(i); 1081 if( b->is_Phi() ) { 1082 _igvn.hash_delete(phi); 1083 _igvn._worklist.push(phi); 1084 phi->set_req(i, clone_bool( b->as_Phi(), loop )); 1085 } else { 1086 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" ); 1087 } 1088 } 1089 1090 Node *sample_cmp = phi->in(1); 1091 1092 // Make Phis to merge the Cmp's inputs. 1093 int size = phi->in(0)->req(); 1094 PhiNode *phi1 = new (C, size) PhiNode( phi->in(0), Type::TOP ); 1095 PhiNode *phi2 = new (C, size) PhiNode( phi->in(0), Type::TOP ); 1096 for( uint j = 1; j < phi->req(); j++ ) { 1097 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP 1098 Node *n1, *n2; 1099 if( cmp_top->is_Cmp() ) { 1100 n1 = cmp_top->in(1); 1101 n2 = cmp_top->in(2); 1102 } else { 1103 n1 = n2 = cmp_top; 1104 } 1105 phi1->set_req( j, n1 ); 1106 phi2->set_req( j, n2 ); 1107 phi1->set_type( phi1->type()->meet(n1->bottom_type()) ); 1108 phi2->set_type( phi2->type()->meet(n2->bottom_type()) ); 1109 } 1110 1111 // See if these Phis have been made before. 1112 // Register with optimizer 1113 Node *hit1 = _igvn.hash_find_insert(phi1); 1114 if( hit1 ) { // Hit, toss just made Phi 1115 _igvn.remove_dead_node(phi1); // Remove new phi 1116 assert( hit1->is_Phi(), "" ); 1117 phi1 = (PhiNode*)hit1; // Use existing phi 1118 } else { // Miss 1119 _igvn.register_new_node_with_optimizer(phi1); 1120 } 1121 Node *hit2 = _igvn.hash_find_insert(phi2); 1122 if( hit2 ) { // Hit, toss just made Phi 1123 _igvn.remove_dead_node(phi2); // Remove new phi 1124 assert( hit2->is_Phi(), "" ); 1125 phi2 = (PhiNode*)hit2; // Use existing phi 1126 } else { // Miss 1127 _igvn.register_new_node_with_optimizer(phi2); 1128 } 1129 // Register Phis with loop/block info 1130 set_ctrl(phi1, phi->in(0)); 1131 set_ctrl(phi2, phi->in(0)); 1132 // Make a new Cmp 1133 Node *cmp = sample_cmp->clone(); 1134 cmp->set_req( 1, phi1 ); 1135 cmp->set_req( 2, phi2 ); 1136 _igvn.register_new_node_with_optimizer(cmp); 1137 set_ctrl(cmp, phi->in(0)); 1138 1139 assert( cmp->is_Cmp(), "" ); 1140 return (CmpNode*)cmp; 1141 } 1142 1143 //------------------------------sink_use--------------------------------------- 1144 // If 'use' was in the loop-exit block, it now needs to be sunk 1145 // below the post-loop merge point. 1146 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) { 1147 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) { 1148 set_ctrl(use, post_loop); 1149 for (DUIterator j = use->outs(); use->has_out(j); j++) 1150 sink_use(use->out(j), post_loop); 1151 } 1152 } 1153 1154 //------------------------------clone_loop------------------------------------- 1155 // 1156 // C L O N E A L O O P B O D Y 1157 // 1158 // This is the basic building block of the loop optimizations. It clones an 1159 // entire loop body. It makes an old_new loop body mapping; with this mapping 1160 // you can find the new-loop equivalent to an old-loop node. All new-loop 1161 // nodes are exactly equal to their old-loop counterparts, all edges are the 1162 // same. All exits from the old-loop now have a RegionNode that merges the 1163 // equivalent new-loop path. This is true even for the normal "loop-exit" 1164 // condition. All uses of loop-invariant old-loop values now come from (one 1165 // or more) Phis that merge their new-loop equivalents. 1166 // 1167 // This operation leaves the graph in an illegal state: there are two valid 1168 // control edges coming from the loop pre-header to both loop bodies. I'll 1169 // definitely have to hack the graph after running this transform. 1170 // 1171 // From this building block I will further edit edges to perform loop peeling 1172 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc. 1173 // 1174 // Parameter side_by_size_idom: 1175 // When side_by_size_idom is NULL, the dominator tree is constructed for 1176 // the clone loop to dominate the original. Used in construction of 1177 // pre-main-post loop sequence. 1178 // When nonnull, the clone and original are side-by-side, both are 1179 // dominated by the side_by_side_idom node. Used in construction of 1180 // unswitched loops. 1181 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd, 1182 Node* side_by_side_idom) { 1183 1184 // Step 1: Clone the loop body. Make the old->new mapping. 1185 uint i; 1186 for( i = 0; i < loop->_body.size(); i++ ) { 1187 Node *old = loop->_body.at(i); 1188 Node *nnn = old->clone(); 1189 old_new.map( old->_idx, nnn ); 1190 _igvn.register_new_node_with_optimizer(nnn); 1191 } 1192 1193 1194 // Step 2: Fix the edges in the new body. If the old input is outside the 1195 // loop use it. If the old input is INside the loop, use the corresponding 1196 // new node instead. 1197 for( i = 0; i < loop->_body.size(); i++ ) { 1198 Node *old = loop->_body.at(i); 1199 Node *nnn = old_new[old->_idx]; 1200 // Fix CFG/Loop controlling the new node 1201 if (has_ctrl(old)) { 1202 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]); 1203 } else { 1204 set_loop(nnn, loop->_parent); 1205 if (old->outcnt() > 0) { 1206 set_idom( nnn, old_new[idom(old)->_idx], dd ); 1207 } 1208 } 1209 // Correct edges to the new node 1210 for( uint j = 0; j < nnn->req(); j++ ) { 1211 Node *n = nnn->in(j); 1212 if( n ) { 1213 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n ); 1214 if( loop->is_member( old_in_loop ) ) 1215 nnn->set_req(j, old_new[n->_idx]); 1216 } 1217 } 1218 _igvn.hash_find_insert(nnn); 1219 } 1220 Node *newhead = old_new[loop->_head->_idx]; 1221 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 1222 1223 1224 // Step 3: Now fix control uses. Loop varying control uses have already 1225 // been fixed up (as part of all input edges in Step 2). Loop invariant 1226 // control uses must be either an IfFalse or an IfTrue. Make a merge 1227 // point to merge the old and new IfFalse/IfTrue nodes; make the use 1228 // refer to this. 1229 ResourceArea *area = Thread::current()->resource_area(); 1230 Node_List worklist(area); 1231 uint new_counter = C->unique(); 1232 for( i = 0; i < loop->_body.size(); i++ ) { 1233 Node* old = loop->_body.at(i); 1234 if( !old->is_CFG() ) continue; 1235 Node* nnn = old_new[old->_idx]; 1236 1237 // Copy uses to a worklist, so I can munge the def-use info 1238 // with impunity. 1239 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1240 worklist.push(old->fast_out(j)); 1241 1242 while( worklist.size() ) { // Visit all uses 1243 Node *use = worklist.pop(); 1244 if (!has_node(use)) continue; // Ignore dead nodes 1245 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1246 if( !loop->is_member( use_loop ) && use->is_CFG() ) { 1247 // Both OLD and USE are CFG nodes here. 1248 assert( use->is_Proj(), "" ); 1249 1250 // Clone the loop exit control projection 1251 Node *newuse = use->clone(); 1252 newuse->set_req(0,nnn); 1253 _igvn.register_new_node_with_optimizer(newuse); 1254 set_loop(newuse, use_loop); 1255 set_idom(newuse, nnn, dom_depth(nnn) + 1 ); 1256 1257 // We need a Region to merge the exit from the peeled body and the 1258 // exit from the old loop body. 1259 RegionNode *r = new (C, 3) RegionNode(3); 1260 // Map the old use to the new merge point 1261 old_new.map( use->_idx, r ); 1262 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use)); 1263 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" ); 1264 1265 // The original user of 'use' uses 'r' instead. 1266 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) { 1267 Node* useuse = use->last_out(l); 1268 _igvn.hash_delete(useuse); 1269 _igvn._worklist.push(useuse); 1270 uint uses_found = 0; 1271 if( useuse->in(0) == use ) { 1272 useuse->set_req(0, r); 1273 uses_found++; 1274 if( useuse->is_CFG() ) { 1275 assert( dom_depth(useuse) > dd_r, "" ); 1276 set_idom(useuse, r, dom_depth(useuse)); 1277 } 1278 } 1279 for( uint k = 1; k < useuse->req(); k++ ) { 1280 if( useuse->in(k) == use ) { 1281 useuse->set_req(k, r); 1282 uses_found++; 1283 } 1284 } 1285 l -= uses_found; // we deleted 1 or more copies of this edge 1286 } 1287 1288 // Now finish up 'r' 1289 r->set_req( 1, newuse ); 1290 r->set_req( 2, use ); 1291 _igvn.register_new_node_with_optimizer(r); 1292 set_loop(r, use_loop); 1293 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r); 1294 } // End of if a loop-exit test 1295 } 1296 } 1297 1298 // Step 4: If loop-invariant use is not control, it must be dominated by a 1299 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region 1300 // there if needed. Make a Phi there merging old and new used values. 1301 Node_List *split_if_set = NULL; 1302 Node_List *split_bool_set = NULL; 1303 Node_List *split_cex_set = NULL; 1304 for( i = 0; i < loop->_body.size(); i++ ) { 1305 Node* old = loop->_body.at(i); 1306 Node* nnn = old_new[old->_idx]; 1307 // Copy uses to a worklist, so I can munge the def-use info 1308 // with impunity. 1309 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1310 worklist.push(old->fast_out(j)); 1311 1312 while( worklist.size() ) { 1313 Node *use = worklist.pop(); 1314 if (!has_node(use)) continue; // Ignore dead nodes 1315 if (use->in(0) == C->top()) continue; 1316 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1317 // Check for data-use outside of loop - at least one of OLD or USE 1318 // must not be a CFG node. 1319 if( !loop->is_member( use_loop ) && (!old->is_CFG() || !use->is_CFG())) { 1320 1321 // If the Data use is an IF, that means we have an IF outside of the 1322 // loop that is switching on a condition that is set inside of the 1323 // loop. Happens if people set a loop-exit flag; then test the flag 1324 // in the loop to break the loop, then test is again outside of the 1325 // loop to determine which way the loop exited. 1326 if( use->is_If() || use->is_CMove() ) { 1327 // Since this code is highly unlikely, we lazily build the worklist 1328 // of such Nodes to go split. 1329 if( !split_if_set ) 1330 split_if_set = new Node_List(area); 1331 split_if_set->push(use); 1332 } 1333 if( use->is_Bool() ) { 1334 if( !split_bool_set ) 1335 split_bool_set = new Node_List(area); 1336 split_bool_set->push(use); 1337 } 1338 if( use->Opcode() == Op_CreateEx ) { 1339 if( !split_cex_set ) 1340 split_cex_set = new Node_List(area); 1341 split_cex_set->push(use); 1342 } 1343 1344 1345 // Get "block" use is in 1346 uint idx = 0; 1347 while( use->in(idx) != old ) idx++; 1348 Node *prev = use->is_CFG() ? use : get_ctrl(use); 1349 assert( !loop->is_member( get_loop( prev ) ), "" ); 1350 Node *cfg = prev->_idx >= new_counter 1351 ? prev->in(2) 1352 : idom(prev); 1353 if( use->is_Phi() ) // Phi use is in prior block 1354 cfg = prev->in(idx); // NOT in block of Phi itself 1355 if (cfg->is_top()) { // Use is dead? 1356 _igvn.hash_delete(use); 1357 _igvn._worklist.push(use); 1358 use->set_req(idx, C->top()); 1359 continue; 1360 } 1361 1362 while( !loop->is_member( get_loop( cfg ) ) ) { 1363 prev = cfg; 1364 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg); 1365 } 1366 // If the use occurs after merging several exits from the loop, then 1367 // old value must have dominated all those exits. Since the same old 1368 // value was used on all those exits we did not need a Phi at this 1369 // merge point. NOW we do need a Phi here. Each loop exit value 1370 // is now merged with the peeled body exit; each exit gets its own 1371 // private Phi and those Phis need to be merged here. 1372 Node *phi; 1373 if( prev->is_Region() ) { 1374 if( idx == 0 ) { // Updating control edge? 1375 phi = prev; // Just use existing control 1376 } else { // Else need a new Phi 1377 phi = PhiNode::make( prev, old ); 1378 // Now recursively fix up the new uses of old! 1379 for( uint i = 1; i < prev->req(); i++ ) { 1380 worklist.push(phi); // Onto worklist once for each 'old' input 1381 } 1382 } 1383 } else { 1384 // Get new RegionNode merging old and new loop exits 1385 prev = old_new[prev->_idx]; 1386 assert( prev, "just made this in step 7" ); 1387 if( idx == 0 ) { // Updating control edge? 1388 phi = prev; // Just use existing control 1389 } else { // Else need a new Phi 1390 // Make a new Phi merging data values properly 1391 phi = PhiNode::make( prev, old ); 1392 phi->set_req( 1, nnn ); 1393 } 1394 } 1395 // If inserting a new Phi, check for prior hits 1396 if( idx != 0 ) { 1397 Node *hit = _igvn.hash_find_insert(phi); 1398 if( hit == NULL ) { 1399 _igvn.register_new_node_with_optimizer(phi); // Register new phi 1400 } else { // or 1401 // Remove the new phi from the graph and use the hit 1402 _igvn.remove_dead_node(phi); 1403 phi = hit; // Use existing phi 1404 } 1405 set_ctrl(phi, prev); 1406 } 1407 // Make 'use' use the Phi instead of the old loop body exit value 1408 _igvn.hash_delete(use); 1409 _igvn._worklist.push(use); 1410 use->set_req(idx, phi); 1411 if( use->_idx >= new_counter ) { // If updating new phis 1412 // Not needed for correctness, but prevents a weak assert 1413 // in AddPNode from tripping (when we end up with different 1414 // base & derived Phis that will become the same after 1415 // IGVN does CSE). 1416 Node *hit = _igvn.hash_find_insert(use); 1417 if( hit ) // Go ahead and re-hash for hits. 1418 _igvn.replace_node( use, hit ); 1419 } 1420 1421 // If 'use' was in the loop-exit block, it now needs to be sunk 1422 // below the post-loop merge point. 1423 sink_use( use, prev ); 1424 } 1425 } 1426 } 1427 1428 // Check for IFs that need splitting/cloning. Happens if an IF outside of 1429 // the loop uses a condition set in the loop. The original IF probably 1430 // takes control from one or more OLD Regions (which in turn get from NEW 1431 // Regions). In any case, there will be a set of Phis for each merge point 1432 // from the IF up to where the original BOOL def exists the loop. 1433 if( split_if_set ) { 1434 while( split_if_set->size() ) { 1435 Node *iff = split_if_set->pop(); 1436 if( iff->in(1)->is_Phi() ) { 1437 BoolNode *b = clone_iff( iff->in(1)->as_Phi(), loop ); 1438 _igvn.hash_delete(iff); 1439 _igvn._worklist.push(iff); 1440 iff->set_req(1, b); 1441 } 1442 } 1443 } 1444 if( split_bool_set ) { 1445 while( split_bool_set->size() ) { 1446 Node *b = split_bool_set->pop(); 1447 Node *phi = b->in(1); 1448 assert( phi->is_Phi(), "" ); 1449 CmpNode *cmp = clone_bool( (PhiNode*)phi, loop ); 1450 _igvn.hash_delete(b); 1451 _igvn._worklist.push(b); 1452 b->set_req(1, cmp); 1453 } 1454 } 1455 if( split_cex_set ) { 1456 while( split_cex_set->size() ) { 1457 Node *b = split_cex_set->pop(); 1458 assert( b->in(0)->is_Region(), "" ); 1459 assert( b->in(1)->is_Phi(), "" ); 1460 assert( b->in(0)->in(0) == b->in(1)->in(0), "" ); 1461 split_up( b, b->in(0), NULL ); 1462 } 1463 } 1464 1465 } 1466 1467 1468 //---------------------- stride_of_possible_iv ------------------------------------- 1469 // Looks for an iff/bool/comp with one operand of the compare 1470 // being a cycle involving an add and a phi, 1471 // with an optional truncation (left-shift followed by a right-shift) 1472 // of the add. Returns zero if not an iv. 1473 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) { 1474 Node* trunc1 = NULL; 1475 Node* trunc2 = NULL; 1476 const TypeInt* ttype = NULL; 1477 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) { 1478 return 0; 1479 } 1480 BoolNode* bl = iff->in(1)->as_Bool(); 1481 Node* cmp = bl->in(1); 1482 if (!cmp || cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU) { 1483 return 0; 1484 } 1485 // Must have an invariant operand 1486 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) { 1487 return 0; 1488 } 1489 Node* add2 = NULL; 1490 Node* cmp1 = cmp->in(1); 1491 if (cmp1->is_Phi()) { 1492 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) ))) 1493 Node* phi = cmp1; 1494 for (uint i = 1; i < phi->req(); i++) { 1495 Node* in = phi->in(i); 1496 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in, 1497 &trunc1, &trunc2, &ttype); 1498 if (add && add->in(1) == phi) { 1499 add2 = add->in(2); 1500 break; 1501 } 1502 } 1503 } else { 1504 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) ))) 1505 Node* addtrunc = cmp1; 1506 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc, 1507 &trunc1, &trunc2, &ttype); 1508 if (add && add->in(1)->is_Phi()) { 1509 Node* phi = add->in(1); 1510 for (uint i = 1; i < phi->req(); i++) { 1511 if (phi->in(i) == addtrunc) { 1512 add2 = add->in(2); 1513 break; 1514 } 1515 } 1516 } 1517 } 1518 if (add2 != NULL) { 1519 const TypeInt* add2t = _igvn.type(add2)->is_int(); 1520 if (add2t->is_con()) { 1521 return add2t->get_con(); 1522 } 1523 } 1524 return 0; 1525 } 1526 1527 1528 //---------------------- stay_in_loop ------------------------------------- 1529 // Return the (unique) control output node that's in the loop (if it exists.) 1530 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) { 1531 Node* unique = NULL; 1532 if (!n) return NULL; 1533 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1534 Node* use = n->fast_out(i); 1535 if (!has_ctrl(use) && loop->is_member(get_loop(use))) { 1536 if (unique != NULL) { 1537 return NULL; 1538 } 1539 unique = use; 1540 } 1541 } 1542 return unique; 1543 } 1544 1545 //------------------------------ register_node ------------------------------------- 1546 // Utility to register node "n" with PhaseIdealLoop 1547 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) { 1548 _igvn.register_new_node_with_optimizer(n); 1549 loop->_body.push(n); 1550 if (n->is_CFG()) { 1551 set_loop(n, loop); 1552 set_idom(n, pred, ddepth); 1553 } else { 1554 set_ctrl(n, pred); 1555 } 1556 } 1557 1558 //------------------------------ proj_clone ------------------------------------- 1559 // Utility to create an if-projection 1560 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) { 1561 ProjNode* c = p->clone()->as_Proj(); 1562 c->set_req(0, iff); 1563 return c; 1564 } 1565 1566 //------------------------------ short_circuit_if ------------------------------------- 1567 // Force the iff control output to be the live_proj 1568 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) { 1569 int proj_con = live_proj->_con; 1570 assert(proj_con == 0 || proj_con == 1, "false or true projection"); 1571 Node *con = _igvn.intcon(proj_con); 1572 set_ctrl(con, C->root()); 1573 if (iff) { 1574 iff->set_req(1, con); 1575 } 1576 return con; 1577 } 1578 1579 //------------------------------ insert_if_before_proj ------------------------------------- 1580 // Insert a new if before an if projection (* - new node) 1581 // 1582 // before 1583 // if(test) 1584 // / \ 1585 // v v 1586 // other-proj proj (arg) 1587 // 1588 // after 1589 // if(test) 1590 // / \ 1591 // / v 1592 // | * proj-clone 1593 // v | 1594 // other-proj v 1595 // * new_if(relop(cmp[IU](left,right))) 1596 // / \ 1597 // v v 1598 // * new-proj proj 1599 // (returned) 1600 // 1601 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) { 1602 IfNode* iff = proj->in(0)->as_If(); 1603 IdealLoopTree *loop = get_loop(proj); 1604 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 1605 int ddepth = dom_depth(proj); 1606 1607 _igvn.hash_delete(iff); 1608 _igvn._worklist.push(iff); 1609 _igvn.hash_delete(proj); 1610 _igvn._worklist.push(proj); 1611 1612 proj->set_req(0, NULL); // temporary disconnect 1613 ProjNode* proj2 = proj_clone(proj, iff); 1614 register_node(proj2, loop, iff, ddepth); 1615 1616 Node* cmp = Signed ? (Node*) new (C,3)CmpINode(left, right) : (Node*) new (C,3)CmpUNode(left, right); 1617 register_node(cmp, loop, proj2, ddepth); 1618 1619 BoolNode* bol = new (C,2)BoolNode(cmp, relop); 1620 register_node(bol, loop, proj2, ddepth); 1621 1622 IfNode* new_if = new (C,2)IfNode(proj2, bol, iff->_prob, iff->_fcnt); 1623 register_node(new_if, loop, proj2, ddepth); 1624 1625 proj->set_req(0, new_if); // reattach 1626 set_idom(proj, new_if, ddepth); 1627 1628 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj(); 1629 register_node(new_exit, get_loop(other_proj), new_if, ddepth); 1630 1631 return new_exit; 1632 } 1633 1634 //------------------------------ insert_region_before_proj ------------------------------------- 1635 // Insert a region before an if projection (* - new node) 1636 // 1637 // before 1638 // if(test) 1639 // / | 1640 // v | 1641 // proj v 1642 // other-proj 1643 // 1644 // after 1645 // if(test) 1646 // / | 1647 // v | 1648 // * proj-clone v 1649 // | other-proj 1650 // v 1651 // * new-region 1652 // | 1653 // v 1654 // * dum_if 1655 // / \ 1656 // v \ 1657 // * dum-proj v 1658 // proj 1659 // 1660 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) { 1661 IfNode* iff = proj->in(0)->as_If(); 1662 IdealLoopTree *loop = get_loop(proj); 1663 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 1664 int ddepth = dom_depth(proj); 1665 1666 _igvn.hash_delete(iff); 1667 _igvn._worklist.push(iff); 1668 _igvn.hash_delete(proj); 1669 _igvn._worklist.push(proj); 1670 1671 proj->set_req(0, NULL); // temporary disconnect 1672 ProjNode* proj2 = proj_clone(proj, iff); 1673 register_node(proj2, loop, iff, ddepth); 1674 1675 RegionNode* reg = new (C,2)RegionNode(2); 1676 reg->set_req(1, proj2); 1677 register_node(reg, loop, iff, ddepth); 1678 1679 IfNode* dum_if = new (C,2)IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt); 1680 register_node(dum_if, loop, reg, ddepth); 1681 1682 proj->set_req(0, dum_if); // reattach 1683 set_idom(proj, dum_if, ddepth); 1684 1685 ProjNode* dum_proj = proj_clone(other_proj, dum_if); 1686 register_node(dum_proj, loop, dum_if, ddepth); 1687 1688 return reg; 1689 } 1690 1691 //------------------------------ insert_cmpi_loop_exit ------------------------------------- 1692 // Clone a signed compare loop exit from an unsigned compare and 1693 // insert it before the unsigned cmp on the stay-in-loop path. 1694 // All new nodes inserted in the dominator tree between the original 1695 // if and it's projections. The original if test is replaced with 1696 // a constant to force the stay-in-loop path. 1697 // 1698 // This is done to make sure that the original if and it's projections 1699 // still dominate the same set of control nodes, that the ctrl() relation 1700 // from data nodes to them is preserved, and that their loop nesting is 1701 // preserved. 1702 // 1703 // before 1704 // if(i <u limit) unsigned compare loop exit 1705 // / | 1706 // v v 1707 // exit-proj stay-in-loop-proj 1708 // 1709 // after 1710 // if(stay-in-loop-const) original if 1711 // / | 1712 // / v 1713 // / if(i < limit) new signed test 1714 // / / | 1715 // / / v 1716 // / / if(i <u limit) new cloned unsigned test 1717 // / / / | 1718 // v v v | 1719 // region | 1720 // | | 1721 // dum-if | 1722 // / | | 1723 // ether | | 1724 // v v 1725 // exit-proj stay-in-loop-proj 1726 // 1727 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) { 1728 const bool Signed = true; 1729 const bool Unsigned = false; 1730 1731 BoolNode* bol = if_cmpu->in(1)->as_Bool(); 1732 if (bol->_test._test != BoolTest::lt) return NULL; 1733 CmpNode* cmpu = bol->in(1)->as_Cmp(); 1734 if (cmpu->Opcode() != Op_CmpU) return NULL; 1735 int stride = stride_of_possible_iv(if_cmpu); 1736 if (stride == 0) return NULL; 1737 1738 ProjNode* lp_continue = stay_in_loop(if_cmpu, loop)->as_Proj(); 1739 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj(); 1740 1741 Node* limit = NULL; 1742 if (stride > 0) { 1743 limit = cmpu->in(2); 1744 } else { 1745 limit = _igvn.makecon(TypeInt::ZERO); 1746 set_ctrl(limit, C->root()); 1747 } 1748 // Create a new region on the exit path 1749 RegionNode* reg = insert_region_before_proj(lp_exit); 1750 1751 // Clone the if-cmpu-true-false using a signed compare 1752 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge; 1753 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue); 1754 reg->add_req(cmpi_exit); 1755 1756 // Clone the if-cmpu-true-false 1757 BoolTest::mask rel_u = bol->_test._test; 1758 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue); 1759 reg->add_req(cmpu_exit); 1760 1761 // Force original if to stay in loop. 1762 short_circuit_if(if_cmpu, lp_continue); 1763 1764 return cmpi_exit->in(0)->as_If(); 1765 } 1766 1767 //------------------------------ remove_cmpi_loop_exit ------------------------------------- 1768 // Remove a previously inserted signed compare loop exit. 1769 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) { 1770 Node* lp_proj = stay_in_loop(if_cmp, loop); 1771 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI && 1772 stay_in_loop(lp_proj, loop)->is_If() && 1773 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu"); 1774 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO); 1775 set_ctrl(con, C->root()); 1776 if_cmp->set_req(1, con); 1777 } 1778 1779 //------------------------------ scheduled_nodelist ------------------------------------- 1780 // Create a post order schedule of nodes that are in the 1781 // "member" set. The list is returned in "sched". 1782 // The first node in "sched" is the loop head, followed by 1783 // nodes which have no inputs in the "member" set, and then 1784 // followed by the nodes that have an immediate input dependence 1785 // on a node in "sched". 1786 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) { 1787 1788 assert(member.test(loop->_head->_idx), "loop head must be in member set"); 1789 Arena *a = Thread::current()->resource_area(); 1790 VectorSet visited(a); 1791 Node_Stack nstack(a, loop->_body.size()); 1792 1793 Node* n = loop->_head; // top of stack is cached in "n" 1794 uint idx = 0; 1795 visited.set(n->_idx); 1796 1797 // Initially push all with no inputs from within member set 1798 for(uint i = 0; i < loop->_body.size(); i++ ) { 1799 Node *elt = loop->_body.at(i); 1800 if (member.test(elt->_idx)) { 1801 bool found = false; 1802 for (uint j = 0; j < elt->req(); j++) { 1803 Node* def = elt->in(j); 1804 if (def && member.test(def->_idx) && def != elt) { 1805 found = true; 1806 break; 1807 } 1808 } 1809 if (!found && elt != loop->_head) { 1810 nstack.push(n, idx); 1811 n = elt; 1812 assert(!visited.test(n->_idx), "not seen yet"); 1813 visited.set(n->_idx); 1814 } 1815 } 1816 } 1817 1818 // traverse out's that are in the member set 1819 while (true) { 1820 if (idx < n->outcnt()) { 1821 Node* use = n->raw_out(idx); 1822 idx++; 1823 if (!visited.test_set(use->_idx)) { 1824 if (member.test(use->_idx)) { 1825 nstack.push(n, idx); 1826 n = use; 1827 idx = 0; 1828 } 1829 } 1830 } else { 1831 // All outputs processed 1832 sched.push(n); 1833 if (nstack.is_empty()) break; 1834 n = nstack.node(); 1835 idx = nstack.index(); 1836 nstack.pop(); 1837 } 1838 } 1839 } 1840 1841 1842 //------------------------------ has_use_in_set ------------------------------------- 1843 // Has a use in the vector set 1844 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) { 1845 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1846 Node* use = n->fast_out(j); 1847 if (vset.test(use->_idx)) { 1848 return true; 1849 } 1850 } 1851 return false; 1852 } 1853 1854 1855 //------------------------------ has_use_internal_to_set ------------------------------------- 1856 // Has use internal to the vector set (ie. not in a phi at the loop head) 1857 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) { 1858 Node* head = loop->_head; 1859 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1860 Node* use = n->fast_out(j); 1861 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) { 1862 return true; 1863 } 1864 } 1865 return false; 1866 } 1867 1868 1869 //------------------------------ clone_for_use_outside_loop ------------------------------------- 1870 // clone "n" for uses that are outside of loop 1871 void PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) { 1872 1873 assert(worklist.size() == 0, "should be empty"); 1874 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1875 Node* use = n->fast_out(j); 1876 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) { 1877 worklist.push(use); 1878 } 1879 } 1880 while( worklist.size() ) { 1881 Node *use = worklist.pop(); 1882 if (!has_node(use) || use->in(0) == C->top()) continue; 1883 uint j; 1884 for (j = 0; j < use->req(); j++) { 1885 if (use->in(j) == n) break; 1886 } 1887 assert(j < use->req(), "must be there"); 1888 1889 // clone "n" and insert it between the inputs of "n" and the use outside the loop 1890 Node* n_clone = n->clone(); 1891 _igvn.hash_delete(use); 1892 use->set_req(j, n_clone); 1893 _igvn._worklist.push(use); 1894 Node* use_c; 1895 if (!use->is_Phi()) { 1896 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0); 1897 } else { 1898 // Use in a phi is considered a use in the associated predecessor block 1899 use_c = use->in(0)->in(j); 1900 } 1901 set_ctrl(n_clone, use_c); 1902 assert(!loop->is_member(get_loop(use_c)), "should be outside loop"); 1903 get_loop(use_c)->_body.push(n_clone); 1904 _igvn.register_new_node_with_optimizer(n_clone); 1905 #if !defined(PRODUCT) 1906 if (TracePartialPeeling) { 1907 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx); 1908 } 1909 #endif 1910 } 1911 } 1912 1913 1914 //------------------------------ clone_for_special_use_inside_loop ------------------------------------- 1915 // clone "n" for special uses that are in the not_peeled region. 1916 // If these def-uses occur in separate blocks, the code generator 1917 // marks the method as not compilable. For example, if a "BoolNode" 1918 // is in a different basic block than the "IfNode" that uses it, then 1919 // the compilation is aborted in the code generator. 1920 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 1921 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) { 1922 if (n->is_Phi() || n->is_Load()) { 1923 return; 1924 } 1925 assert(worklist.size() == 0, "should be empty"); 1926 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1927 Node* use = n->fast_out(j); 1928 if ( not_peel.test(use->_idx) && 1929 (use->is_If() || use->is_CMove() || use->is_Bool()) && 1930 use->in(1) == n) { 1931 worklist.push(use); 1932 } 1933 } 1934 if (worklist.size() > 0) { 1935 // clone "n" and insert it between inputs of "n" and the use 1936 Node* n_clone = n->clone(); 1937 loop->_body.push(n_clone); 1938 _igvn.register_new_node_with_optimizer(n_clone); 1939 set_ctrl(n_clone, get_ctrl(n)); 1940 sink_list.push(n_clone); 1941 not_peel <<= n_clone->_idx; // add n_clone to not_peel set. 1942 #if !defined(PRODUCT) 1943 if (TracePartialPeeling) { 1944 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx); 1945 } 1946 #endif 1947 while( worklist.size() ) { 1948 Node *use = worklist.pop(); 1949 _igvn.hash_delete(use); 1950 _igvn._worklist.push(use); 1951 for (uint j = 1; j < use->req(); j++) { 1952 if (use->in(j) == n) { 1953 use->set_req(j, n_clone); 1954 } 1955 } 1956 } 1957 } 1958 } 1959 1960 1961 //------------------------------ insert_phi_for_loop ------------------------------------- 1962 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 1963 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) { 1964 Node *phi = PhiNode::make(lp, back_edge_val); 1965 phi->set_req(LoopNode::EntryControl, lp_entry_val); 1966 // Use existing phi if it already exists 1967 Node *hit = _igvn.hash_find_insert(phi); 1968 if( hit == NULL ) { 1969 _igvn.register_new_node_with_optimizer(phi); 1970 set_ctrl(phi, lp); 1971 } else { 1972 // Remove the new phi from the graph and use the hit 1973 _igvn.remove_dead_node(phi); 1974 phi = hit; 1975 } 1976 _igvn.hash_delete(use); 1977 _igvn._worklist.push(use); 1978 use->set_req(idx, phi); 1979 } 1980 1981 #ifdef ASSERT 1982 //------------------------------ is_valid_loop_partition ------------------------------------- 1983 // Validate the loop partition sets: peel and not_peel 1984 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, 1985 VectorSet& not_peel ) { 1986 uint i; 1987 // Check that peel_list entries are in the peel set 1988 for (i = 0; i < peel_list.size(); i++) { 1989 if (!peel.test(peel_list.at(i)->_idx)) { 1990 return false; 1991 } 1992 } 1993 // Check at loop members are in one of peel set or not_peel set 1994 for (i = 0; i < loop->_body.size(); i++ ) { 1995 Node *def = loop->_body.at(i); 1996 uint di = def->_idx; 1997 // Check that peel set elements are in peel_list 1998 if (peel.test(di)) { 1999 if (not_peel.test(di)) { 2000 return false; 2001 } 2002 // Must be in peel_list also 2003 bool found = false; 2004 for (uint j = 0; j < peel_list.size(); j++) { 2005 if (peel_list.at(j)->_idx == di) { 2006 found = true; 2007 break; 2008 } 2009 } 2010 if (!found) { 2011 return false; 2012 } 2013 } else if (not_peel.test(di)) { 2014 if (peel.test(di)) { 2015 return false; 2016 } 2017 } else { 2018 return false; 2019 } 2020 } 2021 return true; 2022 } 2023 2024 //------------------------------ is_valid_clone_loop_exit_use ------------------------------------- 2025 // Ensure a use outside of loop is of the right form 2026 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) { 2027 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2028 return (use->is_Phi() && 2029 use_c->is_Region() && use_c->req() == 3 && 2030 (use_c->in(exit_idx)->Opcode() == Op_IfTrue || 2031 use_c->in(exit_idx)->Opcode() == Op_IfFalse || 2032 use_c->in(exit_idx)->Opcode() == Op_JumpProj) && 2033 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) ); 2034 } 2035 2036 //------------------------------ is_valid_clone_loop_form ------------------------------------- 2037 // Ensure that all uses outside of loop are of the right form 2038 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 2039 uint orig_exit_idx, uint clone_exit_idx) { 2040 uint len = peel_list.size(); 2041 for (uint i = 0; i < len; i++) { 2042 Node *def = peel_list.at(i); 2043 2044 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 2045 Node *use = def->fast_out(j); 2046 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2047 if (!loop->is_member(get_loop(use_c))) { 2048 // use is not in the loop, check for correct structure 2049 if (use->in(0) == def) { 2050 // Okay 2051 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) { 2052 return false; 2053 } 2054 } 2055 } 2056 } 2057 return true; 2058 } 2059 #endif 2060 2061 //------------------------------ partial_peel ------------------------------------- 2062 // Partially peel (aka loop rotation) the top portion of a loop (called 2063 // the peel section below) by cloning it and placing one copy just before 2064 // the new loop head and the other copy at the bottom of the new loop. 2065 // 2066 // before after where it came from 2067 // 2068 // stmt1 stmt1 2069 // loop: stmt2 clone 2070 // stmt2 if condA goto exitA clone 2071 // if condA goto exitA new_loop: new 2072 // stmt3 stmt3 clone 2073 // if !condB goto loop if condB goto exitB clone 2074 // exitB: stmt2 orig 2075 // stmt4 if !condA goto new_loop orig 2076 // exitA: goto exitA 2077 // exitB: 2078 // stmt4 2079 // exitA: 2080 // 2081 // Step 1: find the cut point: an exit test on probable 2082 // induction variable. 2083 // Step 2: schedule (with cloning) operations in the peel 2084 // section that can be executed after the cut into 2085 // the section that is not peeled. This may need 2086 // to clone operations into exit blocks. For 2087 // instance, a reference to A[i] in the not-peel 2088 // section and a reference to B[i] in an exit block 2089 // may cause a left-shift of i by 2 to be placed 2090 // in the peel block. This step will clone the left 2091 // shift into the exit block and sink the left shift 2092 // from the peel to the not-peel section. 2093 // Step 3: clone the loop, retarget the control, and insert 2094 // phis for values that are live across the new loop 2095 // head. This is very dependent on the graph structure 2096 // from clone_loop. It creates region nodes for 2097 // exit control and associated phi nodes for values 2098 // flow out of the loop through that exit. The region 2099 // node is dominated by the clone's control projection. 2100 // So the clone's peel section is placed before the 2101 // new loop head, and the clone's not-peel section is 2102 // forms the top part of the new loop. The original 2103 // peel section forms the tail of the new loop. 2104 // Step 4: update the dominator tree and recompute the 2105 // dominator depth. 2106 // 2107 // orig 2108 // 2109 // stmt1 2110 // | 2111 // v 2112 // loop<----+ 2113 // | | 2114 // stmt2 | 2115 // | | 2116 // v | 2117 // ifA | 2118 // / | | 2119 // v v | 2120 // false true ^ <-- last_peel 2121 // / | | 2122 // / ===|==cut | 2123 // / stmt3 | <-- first_not_peel 2124 // / | | 2125 // | v | 2126 // v ifB | 2127 // exitA: / \ | 2128 // / \ | 2129 // v v | 2130 // false true | 2131 // / \ | 2132 // / ----+ 2133 // | 2134 // v 2135 // exitB: 2136 // stmt4 2137 // 2138 // 2139 // after clone loop 2140 // 2141 // stmt1 2142 // / \ 2143 // clone / \ orig 2144 // / \ 2145 // / \ 2146 // v v 2147 // +---->loop loop<----+ 2148 // | | | | 2149 // | stmt2 stmt2 | 2150 // | | | | 2151 // | v v | 2152 // | ifA ifA | 2153 // | | \ / | | 2154 // | v v v v | 2155 // ^ true false false true ^ <-- last_peel 2156 // | | ^ \ / | | 2157 // | cut==|== \ \ / ===|==cut | 2158 // | stmt3 \ \ / stmt3 | <-- first_not_peel 2159 // | | dom | | | | 2160 // | v \ 1v v2 v | 2161 // | ifB regionA ifB | 2162 // | / \ | / \ | 2163 // | / \ v / \ | 2164 // | v v exitA: v v | 2165 // | true false false true | 2166 // | / ^ \ / \ | 2167 // +---- \ \ / ----+ 2168 // dom \ / 2169 // \ 1v v2 2170 // regionB 2171 // | 2172 // v 2173 // exitB: 2174 // stmt4 2175 // 2176 // 2177 // after partial peel 2178 // 2179 // stmt1 2180 // / 2181 // clone / orig 2182 // / TOP 2183 // / \ 2184 // v v 2185 // TOP->region region----+ 2186 // | | | 2187 // stmt2 stmt2 | 2188 // | | | 2189 // v v | 2190 // ifA ifA | 2191 // | \ / | | 2192 // v v v v | 2193 // true false false true | <-- last_peel 2194 // | ^ \ / +------|---+ 2195 // +->newloop \ \ / === ==cut | | 2196 // | stmt3 \ \ / TOP | | 2197 // | | dom | | stmt3 | | <-- first_not_peel 2198 // | v \ 1v v2 v | | 2199 // | ifB regionA ifB ^ v 2200 // | / \ | / \ | | 2201 // | / \ v / \ | | 2202 // | v v exitA: v v | | 2203 // | true false false true | | 2204 // | / ^ \ / \ | | 2205 // | | \ \ / v | | 2206 // | | dom \ / TOP | | 2207 // | | \ 1v v2 | | 2208 // ^ v regionB | | 2209 // | | | | | 2210 // | | v ^ v 2211 // | | exitB: | | 2212 // | | stmt4 | | 2213 // | +------------>-----------------+ | 2214 // | | 2215 // +-----------------<---------------------+ 2216 // 2217 // 2218 // final graph 2219 // 2220 // stmt1 2221 // | 2222 // v 2223 // ........> ifA clone 2224 // : / | 2225 // dom / | 2226 // : v v 2227 // : false true 2228 // : | | 2229 // : | stmt2 clone 2230 // : | | 2231 // : | v 2232 // : | newloop<-----+ 2233 // : | | | 2234 // : | stmt3 clone | 2235 // : | | | 2236 // : | v | 2237 // : | ifB | 2238 // : | / \ | 2239 // : | v v | 2240 // : | false true | 2241 // : | | | | 2242 // : | v stmt2 | 2243 // : | exitB: | | 2244 // : | stmt4 v | 2245 // : | ifA orig | 2246 // : | / \ | 2247 // : | / \ | 2248 // : | v v | 2249 // : | false true | 2250 // : | / \ | 2251 // : v v -----+ 2252 // RegionA 2253 // | 2254 // v 2255 // exitA 2256 // 2257 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) { 2258 2259 if (!loop->_head->is_Loop()) { 2260 return false; } 2261 2262 LoopNode *head = loop->_head->as_Loop(); 2263 2264 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) { 2265 return false; 2266 } 2267 2268 // Check for complex exit control 2269 for(uint ii = 0; ii < loop->_body.size(); ii++ ) { 2270 Node *n = loop->_body.at(ii); 2271 int opc = n->Opcode(); 2272 if (n->is_Call() || 2273 opc == Op_Catch || 2274 opc == Op_CatchProj || 2275 opc == Op_Jump || 2276 opc == Op_JumpProj) { 2277 #if !defined(PRODUCT) 2278 if (TracePartialPeeling) { 2279 tty->print_cr("\nExit control too complex: lp: %d", head->_idx); 2280 } 2281 #endif 2282 return false; 2283 } 2284 } 2285 2286 int dd = dom_depth(head); 2287 2288 // Step 1: find cut point 2289 2290 // Walk up dominators to loop head looking for first loop exit 2291 // which is executed on every path thru loop. 2292 IfNode *peel_if = NULL; 2293 IfNode *peel_if_cmpu = NULL; 2294 2295 Node *iff = loop->tail(); 2296 while( iff != head ) { 2297 if( iff->is_If() ) { 2298 Node *ctrl = get_ctrl(iff->in(1)); 2299 if (ctrl->is_top()) return false; // Dead test on live IF. 2300 // If loop-varying exit-test, check for induction variable 2301 if( loop->is_member(get_loop(ctrl)) && 2302 loop->is_loop_exit(iff) && 2303 is_possible_iv_test(iff)) { 2304 Node* cmp = iff->in(1)->in(1); 2305 if (cmp->Opcode() == Op_CmpI) { 2306 peel_if = iff->as_If(); 2307 } else { 2308 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU"); 2309 peel_if_cmpu = iff->as_If(); 2310 } 2311 } 2312 } 2313 iff = idom(iff); 2314 } 2315 // Prefer signed compare over unsigned compare. 2316 IfNode* new_peel_if = NULL; 2317 if (peel_if == NULL) { 2318 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) { 2319 return false; // No peel point found 2320 } 2321 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop); 2322 if (new_peel_if == NULL) { 2323 return false; // No peel point found 2324 } 2325 peel_if = new_peel_if; 2326 } 2327 Node* last_peel = stay_in_loop(peel_if, loop); 2328 Node* first_not_peeled = stay_in_loop(last_peel, loop); 2329 if (first_not_peeled == NULL || first_not_peeled == head) { 2330 return false; 2331 } 2332 2333 #if !defined(PRODUCT) 2334 if (TracePartialPeeling) { 2335 tty->print_cr("before partial peel one iteration"); 2336 Node_List wl; 2337 Node* t = head->in(2); 2338 while (true) { 2339 wl.push(t); 2340 if (t == head) break; 2341 t = idom(t); 2342 } 2343 while (wl.size() > 0) { 2344 Node* tt = wl.pop(); 2345 tt->dump(); 2346 if (tt == last_peel) tty->print_cr("-- cut --"); 2347 } 2348 } 2349 #endif 2350 ResourceArea *area = Thread::current()->resource_area(); 2351 VectorSet peel(area); 2352 VectorSet not_peel(area); 2353 Node_List peel_list(area); 2354 Node_List worklist(area); 2355 Node_List sink_list(area); 2356 2357 // Set of cfg nodes to peel are those that are executable from 2358 // the head through last_peel. 2359 assert(worklist.size() == 0, "should be empty"); 2360 worklist.push(head); 2361 peel.set(head->_idx); 2362 while (worklist.size() > 0) { 2363 Node *n = worklist.pop(); 2364 if (n != last_peel) { 2365 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2366 Node* use = n->fast_out(j); 2367 if (use->is_CFG() && 2368 loop->is_member(get_loop(use)) && 2369 !peel.test_set(use->_idx)) { 2370 worklist.push(use); 2371 } 2372 } 2373 } 2374 } 2375 2376 // Set of non-cfg nodes to peel are those that are control 2377 // dependent on the cfg nodes. 2378 uint i; 2379 for(i = 0; i < loop->_body.size(); i++ ) { 2380 Node *n = loop->_body.at(i); 2381 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n; 2382 if (peel.test(n_c->_idx)) { 2383 peel.set(n->_idx); 2384 } else { 2385 not_peel.set(n->_idx); 2386 } 2387 } 2388 2389 // Step 2: move operations from the peeled section down into the 2390 // not-peeled section 2391 2392 // Get a post order schedule of nodes in the peel region 2393 // Result in right-most operand. 2394 scheduled_nodelist(loop, peel, peel_list ); 2395 2396 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 2397 2398 // For future check for too many new phis 2399 uint old_phi_cnt = 0; 2400 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 2401 Node* use = head->fast_out(j); 2402 if (use->is_Phi()) old_phi_cnt++; 2403 } 2404 2405 #if !defined(PRODUCT) 2406 if (TracePartialPeeling) { 2407 tty->print_cr("\npeeled list"); 2408 } 2409 #endif 2410 2411 // Evacuate nodes in peel region into the not_peeled region if possible 2412 uint new_phi_cnt = 0; 2413 for (i = 0; i < peel_list.size();) { 2414 Node* n = peel_list.at(i); 2415 #if !defined(PRODUCT) 2416 if (TracePartialPeeling) n->dump(); 2417 #endif 2418 bool incr = true; 2419 if ( !n->is_CFG() ) { 2420 2421 if ( has_use_in_set(n, not_peel) ) { 2422 2423 // If not used internal to the peeled region, 2424 // move "n" from peeled to not_peeled region. 2425 2426 if ( !has_use_internal_to_set(n, peel, loop) ) { 2427 2428 // if not pinned and not a load (which maybe anti-dependent on a store) 2429 // and not a CMove (Matcher expects only bool->cmove). 2430 if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) { 2431 clone_for_use_outside_loop( loop, n, worklist ); 2432 2433 sink_list.push(n); 2434 peel >>= n->_idx; // delete n from peel set. 2435 not_peel <<= n->_idx; // add n to not_peel set. 2436 peel_list.remove(i); 2437 incr = false; 2438 #if !defined(PRODUCT) 2439 if (TracePartialPeeling) { 2440 tty->print_cr("sink to not_peeled region: %d newbb: %d", 2441 n->_idx, get_ctrl(n)->_idx); 2442 } 2443 #endif 2444 } 2445 } else { 2446 // Otherwise check for special def-use cases that span 2447 // the peel/not_peel boundary such as bool->if 2448 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist ); 2449 new_phi_cnt++; 2450 } 2451 } 2452 } 2453 if (incr) i++; 2454 } 2455 2456 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) { 2457 #if !defined(PRODUCT) 2458 if (TracePartialPeeling) { 2459 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c", 2460 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F'); 2461 } 2462 #endif 2463 if (new_peel_if != NULL) { 2464 remove_cmpi_loop_exit(new_peel_if, loop); 2465 } 2466 // Inhibit more partial peeling on this loop 2467 assert(!head->is_partial_peel_loop(), "not partial peeled"); 2468 head->mark_partial_peel_failed(); 2469 return false; 2470 } 2471 2472 // Step 3: clone loop, retarget control, and insert new phis 2473 2474 // Create new loop head for new phis and to hang 2475 // the nodes being moved (sinked) from the peel region. 2476 LoopNode* new_head = new (C, 3) LoopNode(last_peel, last_peel); 2477 _igvn.register_new_node_with_optimizer(new_head); 2478 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled"); 2479 first_not_peeled->set_req(0, new_head); 2480 set_loop(new_head, loop); 2481 loop->_body.push(new_head); 2482 not_peel.set(new_head->_idx); 2483 set_idom(new_head, last_peel, dom_depth(first_not_peeled)); 2484 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled)); 2485 2486 while (sink_list.size() > 0) { 2487 Node* n = sink_list.pop(); 2488 set_ctrl(n, new_head); 2489 } 2490 2491 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 2492 2493 clone_loop( loop, old_new, dd ); 2494 2495 const uint clone_exit_idx = 1; 2496 const uint orig_exit_idx = 2; 2497 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop"); 2498 2499 Node* head_clone = old_new[head->_idx]; 2500 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop(); 2501 Node* orig_tail_clone = head_clone->in(2); 2502 2503 // Add phi if "def" node is in peel set and "use" is not 2504 2505 for(i = 0; i < peel_list.size(); i++ ) { 2506 Node *def = peel_list.at(i); 2507 if (!def->is_CFG()) { 2508 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 2509 Node *use = def->fast_out(j); 2510 if (has_node(use) && use->in(0) != C->top() && 2511 (!peel.test(use->_idx) || 2512 (use->is_Phi() && use->in(0) == head)) ) { 2513 worklist.push(use); 2514 } 2515 } 2516 while( worklist.size() ) { 2517 Node *use = worklist.pop(); 2518 for (uint j = 1; j < use->req(); j++) { 2519 Node* n = use->in(j); 2520 if (n == def) { 2521 2522 // "def" is in peel set, "use" is not in peel set 2523 // or "use" is in the entry boundary (a phi) of the peel set 2524 2525 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use; 2526 2527 if ( loop->is_member(get_loop( use_c )) ) { 2528 // use is in loop 2529 if (old_new[use->_idx] != NULL) { // null for dead code 2530 Node* use_clone = old_new[use->_idx]; 2531 _igvn.hash_delete(use); 2532 use->set_req(j, C->top()); 2533 _igvn._worklist.push(use); 2534 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone ); 2535 } 2536 } else { 2537 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format"); 2538 // use is not in the loop, check if the live range includes the cut 2539 Node* lp_if = use_c->in(orig_exit_idx)->in(0); 2540 if (not_peel.test(lp_if->_idx)) { 2541 assert(j == orig_exit_idx, "use from original loop"); 2542 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone ); 2543 } 2544 } 2545 } 2546 } 2547 } 2548 } 2549 } 2550 2551 // Step 3b: retarget control 2552 2553 // Redirect control to the new loop head if a cloned node in 2554 // the not_peeled region has control that points into the peeled region. 2555 // This necessary because the cloned peeled region will be outside 2556 // the loop. 2557 // from to 2558 // cloned-peeled <---+ 2559 // new_head_clone: | <--+ 2560 // cloned-not_peeled in(0) in(0) 2561 // orig-peeled 2562 2563 for(i = 0; i < loop->_body.size(); i++ ) { 2564 Node *n = loop->_body.at(i); 2565 if (!n->is_CFG() && n->in(0) != NULL && 2566 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) { 2567 Node* n_clone = old_new[n->_idx]; 2568 _igvn.hash_delete(n_clone); 2569 n_clone->set_req(0, new_head_clone); 2570 _igvn._worklist.push(n_clone); 2571 } 2572 } 2573 2574 // Backedge of the surviving new_head (the clone) is original last_peel 2575 _igvn.hash_delete(new_head_clone); 2576 new_head_clone->set_req(LoopNode::LoopBackControl, last_peel); 2577 _igvn._worklist.push(new_head_clone); 2578 2579 // Cut first node in original not_peel set 2580 _igvn.hash_delete(new_head); 2581 new_head->set_req(LoopNode::EntryControl, C->top()); 2582 new_head->set_req(LoopNode::LoopBackControl, C->top()); 2583 _igvn._worklist.push(new_head); 2584 2585 // Copy head_clone back-branch info to original head 2586 // and remove original head's loop entry and 2587 // clone head's back-branch 2588 _igvn.hash_delete(head); 2589 _igvn.hash_delete(head_clone); 2590 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl)); 2591 head->set_req(LoopNode::LoopBackControl, C->top()); 2592 head_clone->set_req(LoopNode::LoopBackControl, C->top()); 2593 _igvn._worklist.push(head); 2594 _igvn._worklist.push(head_clone); 2595 2596 // Similarly modify the phis 2597 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) { 2598 Node* use = head->fast_out(k); 2599 if (use->is_Phi() && use->outcnt() > 0) { 2600 Node* use_clone = old_new[use->_idx]; 2601 _igvn.hash_delete(use); 2602 _igvn.hash_delete(use_clone); 2603 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl)); 2604 use->set_req(LoopNode::LoopBackControl, C->top()); 2605 use_clone->set_req(LoopNode::LoopBackControl, C->top()); 2606 _igvn._worklist.push(use); 2607 _igvn._worklist.push(use_clone); 2608 } 2609 } 2610 2611 // Step 4: update dominator tree and dominator depth 2612 2613 set_idom(head, orig_tail_clone, dd); 2614 recompute_dom_depth(); 2615 2616 // Inhibit more partial peeling on this loop 2617 new_head_clone->set_partial_peel_loop(); 2618 C->set_major_progress(); 2619 2620 #if !defined(PRODUCT) 2621 if (TracePartialPeeling) { 2622 tty->print_cr("\nafter partial peel one iteration"); 2623 Node_List wl(area); 2624 Node* t = last_peel; 2625 while (true) { 2626 wl.push(t); 2627 if (t == head_clone) break; 2628 t = idom(t); 2629 } 2630 while (wl.size() > 0) { 2631 Node* tt = wl.pop(); 2632 if (tt == head) tty->print_cr("orig head"); 2633 else if (tt == new_head_clone) tty->print_cr("new head"); 2634 else if (tt == head_clone) tty->print_cr("clone head"); 2635 tt->dump(); 2636 } 2637 } 2638 #endif 2639 return true; 2640 } 2641 2642 //------------------------------reorg_offsets---------------------------------- 2643 // Reorganize offset computations to lower register pressure. Mostly 2644 // prevent loop-fallout uses of the pre-incremented trip counter (which are 2645 // then alive with the post-incremented trip counter forcing an extra 2646 // register move) 2647 void PhaseIdealLoop::reorg_offsets( IdealLoopTree *loop ) { 2648 2649 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 2650 CountedLoopEndNode *cle = cl->loopexit(); 2651 if( !cle ) return; // The occasional dead loop 2652 // Find loop exit control 2653 Node *exit = cle->proj_out(false); 2654 assert( exit->Opcode() == Op_IfFalse, "" ); 2655 2656 // Check for the special case of folks using the pre-incremented 2657 // trip-counter on the fall-out path (forces the pre-incremented 2658 // and post-incremented trip counter to be live at the same time). 2659 // Fix this by adjusting to use the post-increment trip counter. 2660 Node *phi = cl->phi(); 2661 if( !phi ) return; // Dead infinite loop 2662 2663 // Shape messed up, probably by iteration_split_impl 2664 if (phi->in(LoopNode::LoopBackControl) != cl->incr()) return; 2665 2666 bool progress = true; 2667 while (progress) { 2668 progress = false; 2669 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) { 2670 Node* use = phi->fast_out(i); // User of trip-counter 2671 if (!has_ctrl(use)) continue; 2672 Node *u_ctrl = get_ctrl(use); 2673 if( use->is_Phi() ) { 2674 u_ctrl = NULL; 2675 for( uint j = 1; j < use->req(); j++ ) 2676 if( use->in(j) == phi ) 2677 u_ctrl = dom_lca( u_ctrl, use->in(0)->in(j) ); 2678 } 2679 IdealLoopTree *u_loop = get_loop(u_ctrl); 2680 // Look for loop-invariant use 2681 if( u_loop == loop ) continue; 2682 if( loop->is_member( u_loop ) ) continue; 2683 // Check that use is live out the bottom. Assuming the trip-counter 2684 // update is right at the bottom, uses of of the loop middle are ok. 2685 if( dom_lca( exit, u_ctrl ) != exit ) continue; 2686 // protect against stride not being a constant 2687 if( !cle->stride_is_con() ) continue; 2688 // Hit! Refactor use to use the post-incremented tripcounter. 2689 // Compute a post-increment tripcounter. 2690 Node *opaq = new (C, 2) Opaque2Node( C, cle->incr() ); 2691 register_new_node( opaq, u_ctrl ); 2692 Node *neg_stride = _igvn.intcon(-cle->stride_con()); 2693 set_ctrl(neg_stride, C->root()); 2694 Node *post = new (C, 3) AddINode( opaq, neg_stride); 2695 register_new_node( post, u_ctrl ); 2696 _igvn.hash_delete(use); 2697 _igvn._worklist.push(use); 2698 for( uint j = 1; j < use->req(); j++ ) 2699 if( use->in(j) == phi ) 2700 use->set_req(j, post); 2701 // Since DU info changed, rerun loop 2702 progress = true; 2703 break; 2704 } 2705 } 2706 2707 }