loopopts.cpp revision 1472:c18cbe5936b8
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. 31Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) { 32 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) { 33 // ConvI2L may have type information on it which is unsafe to push up 34 // so disable this for now 35 return NULL; 36 } 37 int wins = 0; 38 assert( !n->is_CFG(), "" ); 39 assert( region->is_Region(), "" ); 40 41 const Type* type = n->bottom_type(); 42 const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr(); 43 Node *phi; 44 if( t_oop != NULL && t_oop->is_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. 170void 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) 223Node *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. 264Node *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.hash_delete( n ); 358 _igvn.subsume_node( n, var_add ); 359 return var_add; 360 } 361 362 // Replace (I+V) with (V+I) 363 if( n_op == Op_AddI || 364 n_op == Op_AddL || 365 n_op == Op_AddF || 366 n_op == Op_AddD || 367 n_op == Op_MulI || 368 n_op == Op_MulL || 369 n_op == Op_MulF || 370 n_op == Op_MulD ) { 371 if( n2_loop == n_loop ) { 372 assert( n1_loop != n_loop, "" ); 373 n->swap_edges(1, 2); 374 } 375 } 376 377 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V), 378 // but not if I2 is a constant. 379 if( n_op == Op_AddP ) { 380 if( n2_loop == n_loop && n3_loop != n_loop ) { 381 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) { 382 Node *n22_ctrl = get_ctrl(n->in(2)->in(2)); 383 Node *n23_ctrl = get_ctrl(n->in(2)->in(3)); 384 IdealLoopTree *n22loop = get_loop( n22_ctrl ); 385 IdealLoopTree *n23_loop = get_loop( n23_ctrl ); 386 if( n22loop != n_loop && n22loop->is_member(n_loop) && 387 n23_loop == n_loop ) { 388 Node *add1 = new (C, 4) AddPNode( n->in(1), n->in(2)->in(2), n->in(3) ); 389 // Stuff new AddP in the loop preheader 390 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) ); 391 Node *add2 = new (C, 4) AddPNode( n->in(1), add1, n->in(2)->in(3) ); 392 register_new_node( add2, n_ctrl ); 393 _igvn.hash_delete( n ); 394 _igvn.subsume_node( n, add2 ); 395 return add2; 396 } 397 } 398 } 399 400 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V) 401 if( n2_loop != n_loop && n3_loop == n_loop ) { 402 if( n->in(3)->Opcode() == Op_AddI ) { 403 Node *V = n->in(3)->in(1); 404 Node *I = n->in(3)->in(2); 405 if( is_member(n_loop,get_ctrl(V)) ) { 406 } else { 407 Node *tmp = V; V = I; I = tmp; 408 } 409 if( !is_member(n_loop,get_ctrl(I)) ) { 410 Node *add1 = new (C, 4) AddPNode( n->in(1), n->in(2), I ); 411 // Stuff new AddP in the loop preheader 412 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) ); 413 Node *add2 = new (C, 4) AddPNode( n->in(1), add1, V ); 414 register_new_node( add2, n_ctrl ); 415 _igvn.hash_delete( n ); 416 _igvn.subsume_node( n, add2 ); 417 return add2; 418 } 419 } 420 } 421 } 422 423 return NULL; 424} 425 426//------------------------------conditional_move------------------------------- 427// Attempt to replace a Phi with a conditional move. We have some pretty 428// strict profitability requirements. All Phis at the merge point must 429// be converted, so we can remove the control flow. We need to limit the 430// number of c-moves to a small handful. All code that was in the side-arms 431// of the CFG diamond is now speculatively executed. This code has to be 432// "cheap enough". We are pretty much limited to CFG diamonds that merge 433// 1 or 2 items with a total of 1 or 2 ops executed speculatively. 434Node *PhaseIdealLoop::conditional_move( Node *region ) { 435 436 assert( region->is_Region(), "sanity check" ); 437 if( region->req() != 3 ) return NULL; 438 439 // Check for CFG diamond 440 Node *lp = region->in(1); 441 Node *rp = region->in(2); 442 if( !lp || !rp ) return NULL; 443 Node *lp_c = lp->in(0); 444 if( lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If() ) return NULL; 445 IfNode *iff = lp_c->as_If(); 446 447 // Check for highly predictable branch. No point in CMOV'ing if 448 // we are going to predict accurately all the time. 449 // %%% This hides patterns produced by utility methods like Math.min. 450 if( iff->_prob < PROB_UNLIKELY_MAG(3) || 451 iff->_prob > PROB_LIKELY_MAG(3) ) 452 return NULL; 453 454 // Check for ops pinned in an arm of the diamond. 455 // Can't remove the control flow in this case 456 if( lp->outcnt() > 1 ) return NULL; 457 if( rp->outcnt() > 1 ) return NULL; 458 459 // Check profitability 460 int cost = 0; 461 int phis = 0; 462 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 463 Node *out = region->fast_out(i); 464 if( !out->is_Phi() ) continue; // Ignore other control edges, etc 465 phis++; 466 PhiNode* phi = out->as_Phi(); 467 switch (phi->type()->basic_type()) { 468 case T_LONG: 469 cost++; // Probably encodes as 2 CMOV's 470 case T_INT: // These all CMOV fine 471 case T_FLOAT: 472 case T_DOUBLE: 473 case T_ADDRESS: // (RawPtr) 474 cost++; 475 break; 476 case T_NARROWOOP: // Fall through 477 case T_OBJECT: { // Base oops are OK, but not derived oops 478 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr(); 479 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a 480 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus 481 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we 482 // have a Phi for the base here that we convert to a CMOVE all is well 483 // and good. But if the base is dead, we'll not make a CMOVE. Later 484 // the allocator will have to produce a base by creating a CMOVE of the 485 // relevant bases. This puts the allocator in the business of 486 // manufacturing expensive instructions, generally a bad plan. 487 // Just Say No to Conditionally-Moved Derived Pointers. 488 if( tp && tp->offset() != 0 ) 489 return NULL; 490 cost++; 491 break; 492 } 493 default: 494 return NULL; // In particular, can't do memory or I/O 495 } 496 // Add in cost any speculative ops 497 for( uint j = 1; j < region->req(); j++ ) { 498 Node *proj = region->in(j); 499 Node *inp = phi->in(j); 500 if (get_ctrl(inp) == proj) { // Found local op 501 cost++; 502 // Check for a chain of dependent ops; these will all become 503 // speculative in a CMOV. 504 for( uint k = 1; k < inp->req(); k++ ) 505 if (get_ctrl(inp->in(k)) == proj) 506 return NULL; // Too much speculative goo 507 } 508 } 509 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode. 510 // This will likely Split-If, a higher-payoff operation. 511 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) { 512 Node* use = phi->fast_out(k); 513 if( use->is_Cmp() || use->is_DecodeN() || use->is_EncodeP() ) 514 return NULL; 515 } 516 } 517 if( cost >= ConditionalMoveLimit ) return NULL; // Too much goo 518 Node* bol = iff->in(1); 519 assert( bol->Opcode() == Op_Bool, "" ); 520 int cmp_op = bol->in(1)->Opcode(); 521 // It is expensive to generate flags from a float compare. 522 // Avoid duplicated float compare. 523 if( phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL; 524 525 // -------------- 526 // Now replace all Phis with CMOV's 527 Node *cmov_ctrl = iff->in(0); 528 uint flip = (lp->Opcode() == Op_IfTrue); 529 while( 1 ) { 530 PhiNode* phi = NULL; 531 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 532 Node *out = region->fast_out(i); 533 if (out->is_Phi()) { 534 phi = out->as_Phi(); 535 break; 536 } 537 } 538 if (phi == NULL) break; 539#ifndef PRODUCT 540 if( PrintOpto && VerifyLoopOptimizations ) tty->print_cr("CMOV"); 541#endif 542 // Move speculative ops 543 for( uint j = 1; j < region->req(); j++ ) { 544 Node *proj = region->in(j); 545 Node *inp = phi->in(j); 546 if (get_ctrl(inp) == proj) { // Found local op 547#ifndef PRODUCT 548 if( PrintOpto && VerifyLoopOptimizations ) { 549 tty->print(" speculate: "); 550 inp->dump(); 551 } 552#endif 553 set_ctrl(inp, cmov_ctrl); 554 } 555 } 556 Node *cmov = CMoveNode::make( C, cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi) ); 557 register_new_node( cmov, cmov_ctrl ); 558 _igvn.hash_delete(phi); 559 _igvn.subsume_node( phi, cmov ); 560#ifndef PRODUCT 561 if( VerifyLoopOptimizations ) verify(); 562#endif 563 } 564 565 // The useless CFG diamond will fold up later; see the optimization in 566 // RegionNode::Ideal. 567 _igvn._worklist.push(region); 568 569 return iff->in(1); 570} 571 572//------------------------------split_if_with_blocks_pre----------------------- 573// Do the real work in a non-recursive function. Data nodes want to be 574// cloned in the pre-order so they can feed each other nicely. 575Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) { 576 // Cloning these guys is unlikely to win 577 int n_op = n->Opcode(); 578 if( n_op == Op_MergeMem ) return n; 579 if( n->is_Proj() ) return n; 580 // Do not clone-up CmpFXXX variations, as these are always 581 // followed by a CmpI 582 if( n->is_Cmp() ) return n; 583 // Attempt to use a conditional move instead of a phi/branch 584 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) { 585 Node *cmov = conditional_move( n ); 586 if( cmov ) return cmov; 587 } 588 if( n->is_CFG() || n->is_LoadStore() ) 589 return n; 590 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd 591 n_op == Op_Opaque2 ) { 592 if( !C->major_progress() ) // If chance of no more loop opts... 593 _igvn._worklist.push(n); // maybe we'll remove them 594 return n; 595 } 596 597 if( n->is_Con() ) return n; // No cloning for Con nodes 598 599 Node *n_ctrl = get_ctrl(n); 600 if( !n_ctrl ) return n; // Dead node 601 602 // Attempt to remix address expressions for loop invariants 603 Node *m = remix_address_expressions( n ); 604 if( m ) return m; 605 606 // Determine if the Node has inputs from some local Phi. 607 // Returns the block to clone thru. 608 Node *n_blk = has_local_phi_input( n ); 609 if( !n_blk ) return n; 610 // Do not clone the trip counter through on a CountedLoop 611 // (messes up the canonical shape). 612 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n; 613 614 // Check for having no control input; not pinned. Allow 615 // dominating control. 616 if( n->in(0) ) { 617 Node *dom = idom(n_blk); 618 if( dom_lca( n->in(0), dom ) != n->in(0) ) 619 return n; 620 } 621 // Policy: when is it profitable. You must get more wins than 622 // policy before it is considered profitable. Policy is usually 0, 623 // so 1 win is considered profitable. Big merges will require big 624 // cloning, so get a larger policy. 625 int policy = n_blk->req() >> 2; 626 627 // If the loop is a candidate for range check elimination, 628 // delay splitting through it's phi until a later loop optimization 629 if (n_blk->is_CountedLoop()) { 630 IdealLoopTree *lp = get_loop(n_blk); 631 if (lp && lp->_rce_candidate) { 632 return n; 633 } 634 } 635 636 // Use same limit as split_if_with_blocks_post 637 if( C->unique() > 35000 ) return n; // Method too big 638 639 // Split 'n' through the merge point if it is profitable 640 Node *phi = split_thru_phi( n, n_blk, policy ); 641 if( !phi ) return n; 642 643 // Found a Phi to split thru! 644 // Replace 'n' with the new phi 645 _igvn.hash_delete(n); 646 _igvn.subsume_node( n, phi ); 647 // Moved a load around the loop, 'en-registering' something. 648 if( n_blk->Opcode() == Op_Loop && n->is_Load() && 649 !phi->in(LoopNode::LoopBackControl)->is_Load() ) 650 C->set_major_progress(); 651 652 return phi; 653} 654 655static bool merge_point_too_heavy(Compile* C, Node* region) { 656 // Bail out if the region and its phis have too many users. 657 int weight = 0; 658 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 659 weight += region->fast_out(i)->outcnt(); 660 } 661 int nodes_left = MaxNodeLimit - C->unique(); 662 if (weight * 8 > nodes_left) { 663#ifndef PRODUCT 664 if (PrintOpto) 665 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight); 666#endif 667 return true; 668 } else { 669 return false; 670 } 671} 672 673static bool merge_point_safe(Node* region) { 674 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode 675 // having a PhiNode input. This sidesteps the dangerous case where the split 676 // ConvI2LNode may become TOP if the input Value() does not 677 // overlap the ConvI2L range, leaving a node which may not dominate its 678 // uses. 679 // A better fix for this problem can be found in the BugTraq entry, but 680 // expediency for Mantis demands this hack. 681 // 6855164: If the merge point has a FastLockNode with a PhiNode input, we stop 682 // split_if_with_blocks from splitting a block because we could not move around 683 // the FastLockNode. 684 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 685 Node* n = region->fast_out(i); 686 if (n->is_Phi()) { 687 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 688 Node* m = n->fast_out(j); 689 if (m->is_FastLock()) 690 return false; 691#ifdef _LP64 692 if (m->Opcode() == Op_ConvI2L) 693 return false; 694#endif 695 } 696 } 697 } 698 return true; 699} 700 701 702//------------------------------place_near_use--------------------------------- 703// Place some computation next to use but not inside inner loops. 704// For inner loop uses move it to the preheader area. 705Node *PhaseIdealLoop::place_near_use( Node *useblock ) const { 706 IdealLoopTree *u_loop = get_loop( useblock ); 707 return (u_loop->_irreducible || u_loop->_child) 708 ? useblock 709 : u_loop->_head->in(LoopNode::EntryControl); 710} 711 712 713//------------------------------split_if_with_blocks_post---------------------- 714// Do the real work in a non-recursive function. CFG hackery wants to be 715// in the post-order, so it can dirty the I-DOM info and not use the dirtied 716// info. 717void PhaseIdealLoop::split_if_with_blocks_post( Node *n ) { 718 719 // Cloning Cmp through Phi's involves the split-if transform. 720 // FastLock is not used by an If 721 if( n->is_Cmp() && !n->is_FastLock() ) { 722 if( C->unique() > 35000 ) return; // Method too big 723 724 // Do not do 'split-if' if irreducible loops are present. 725 if( _has_irreducible_loops ) 726 return; 727 728 Node *n_ctrl = get_ctrl(n); 729 // Determine if the Node has inputs from some local Phi. 730 // Returns the block to clone thru. 731 Node *n_blk = has_local_phi_input( n ); 732 if( n_blk != n_ctrl ) return; 733 734 if( merge_point_too_heavy(C, n_ctrl) ) 735 return; 736 737 if( n->outcnt() != 1 ) return; // Multiple bool's from 1 compare? 738 Node *bol = n->unique_out(); 739 assert( bol->is_Bool(), "expect a bool here" ); 740 if( bol->outcnt() != 1 ) return;// Multiple branches from 1 compare? 741 Node *iff = bol->unique_out(); 742 743 // Check some safety conditions 744 if( iff->is_If() ) { // Classic split-if? 745 if( iff->in(0) != n_ctrl ) return; // Compare must be in same blk as if 746 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE 747 if( get_ctrl(iff->in(2)) == n_ctrl || 748 get_ctrl(iff->in(3)) == n_ctrl ) 749 return; // Inputs not yet split-up 750 if ( get_loop(n_ctrl) != get_loop(get_ctrl(iff)) ) { 751 return; // Loop-invar test gates loop-varying CMOVE 752 } 753 } else { 754 return; // some other kind of node, such as an Allocate 755 } 756 757 // Do not do 'split-if' if some paths are dead. First do dead code 758 // elimination and then see if its still profitable. 759 for( uint i = 1; i < n_ctrl->req(); i++ ) 760 if( n_ctrl->in(i) == C->top() ) 761 return; 762 763 // When is split-if profitable? Every 'win' on means some control flow 764 // goes dead, so it's almost always a win. 765 int policy = 0; 766 // If trying to do a 'Split-If' at the loop head, it is only 767 // profitable if the cmp folds up on BOTH paths. Otherwise we 768 // risk peeling a loop forever. 769 770 // CNC - Disabled for now. Requires careful handling of loop 771 // body selection for the cloned code. Also, make sure we check 772 // for any input path not being in the same loop as n_ctrl. For 773 // irreducible loops we cannot check for 'n_ctrl->is_Loop()' 774 // because the alternative loop entry points won't be converted 775 // into LoopNodes. 776 IdealLoopTree *n_loop = get_loop(n_ctrl); 777 for( uint j = 1; j < n_ctrl->req(); j++ ) 778 if( get_loop(n_ctrl->in(j)) != n_loop ) 779 return; 780 781 // Check for safety of the merge point. 782 if( !merge_point_safe(n_ctrl) ) { 783 return; 784 } 785 786 // Split compare 'n' through the merge point if it is profitable 787 Node *phi = split_thru_phi( n, n_ctrl, policy ); 788 if( !phi ) return; 789 790 // Found a Phi to split thru! 791 // Replace 'n' with the new phi 792 _igvn.hash_delete(n); 793 _igvn.subsume_node( n, phi ); 794 795 // Now split the bool up thru the phi 796 Node *bolphi = split_thru_phi( bol, n_ctrl, -1 ); 797 _igvn.hash_delete(bol); 798 _igvn.subsume_node( bol, bolphi ); 799 assert( iff->in(1) == bolphi, "" ); 800 if( bolphi->Value(&_igvn)->singleton() ) 801 return; 802 803 // Conditional-move? Must split up now 804 if( !iff->is_If() ) { 805 Node *cmovphi = split_thru_phi( iff, n_ctrl, -1 ); 806 _igvn.hash_delete(iff); 807 _igvn.subsume_node( iff, cmovphi ); 808 return; 809 } 810 811 // Now split the IF 812 do_split_if( iff ); 813 return; 814 } 815 816 // Check for an IF ready to split; one that has its 817 // condition codes input coming from a Phi at the block start. 818 int n_op = n->Opcode(); 819 820 // Check for an IF being dominated by another IF same test 821 if( n_op == Op_If ) { 822 Node *bol = n->in(1); 823 uint max = bol->outcnt(); 824 // Check for same test used more than once? 825 if( n_op == Op_If && max > 1 && bol->is_Bool() ) { 826 // Search up IDOMs to see if this IF is dominated. 827 Node *cutoff = get_ctrl(bol); 828 829 // Now search up IDOMs till cutoff, looking for a dominating test 830 Node *prevdom = n; 831 Node *dom = idom(prevdom); 832 while( dom != cutoff ) { 833 if( dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom ) { 834 // Replace the dominated test with an obvious true or false. 835 // Place it on the IGVN worklist for later cleanup. 836 C->set_major_progress(); 837 dominated_by( prevdom, n ); 838#ifndef PRODUCT 839 if( VerifyLoopOptimizations ) verify(); 840#endif 841 return; 842 } 843 prevdom = dom; 844 dom = idom(prevdom); 845 } 846 } 847 } 848 849 // See if a shared loop-varying computation has no loop-varying uses. 850 // Happens if something is only used for JVM state in uncommon trap exits, 851 // like various versions of induction variable+offset. Clone the 852 // computation per usage to allow it to sink out of the loop. 853 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about) 854 Node *n_ctrl = get_ctrl(n); 855 IdealLoopTree *n_loop = get_loop(n_ctrl); 856 if( n_loop != _ltree_root ) { 857 DUIterator_Fast imax, i = n->fast_outs(imax); 858 for (; i < imax; i++) { 859 Node* u = n->fast_out(i); 860 if( !has_ctrl(u) ) break; // Found control user 861 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 862 if( u_loop == n_loop ) break; // Found loop-varying use 863 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop 864 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003 865 } 866 bool did_break = (i < imax); // Did we break out of the previous loop? 867 if (!did_break && n->outcnt() > 1) { // All uses in outer loops! 868 Node *late_load_ctrl; 869 if (n->is_Load()) { 870 // If n is a load, get and save the result from get_late_ctrl(), 871 // to be later used in calculating the control for n's clones. 872 clear_dom_lca_tags(); 873 late_load_ctrl = get_late_ctrl(n, n_ctrl); 874 } 875 // If n is a load, and the late control is the same as the current 876 // control, then the cloning of n is a pointless exercise, because 877 // GVN will ensure that we end up where we started. 878 if (!n->is_Load() || late_load_ctrl != n_ctrl) { 879 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) { 880 Node *u = n->last_out(j); // Clone private computation per use 881 _igvn.hash_delete(u); 882 _igvn._worklist.push(u); 883 Node *x = n->clone(); // Clone computation 884 Node *x_ctrl = NULL; 885 if( u->is_Phi() ) { 886 // Replace all uses of normal nodes. Replace Phi uses 887 // individually, so the separate Nodes can sink down 888 // different paths. 889 uint k = 1; 890 while( u->in(k) != n ) k++; 891 u->set_req( k, x ); 892 // x goes next to Phi input path 893 x_ctrl = u->in(0)->in(k); 894 --j; 895 } else { // Normal use 896 // Replace all uses 897 for( uint k = 0; k < u->req(); k++ ) { 898 if( u->in(k) == n ) { 899 u->set_req( k, x ); 900 --j; 901 } 902 } 903 x_ctrl = get_ctrl(u); 904 } 905 906 // Find control for 'x' next to use but not inside inner loops. 907 // For inner loop uses get the preheader area. 908 x_ctrl = place_near_use(x_ctrl); 909 910 if (n->is_Load()) { 911 // For loads, add a control edge to a CFG node outside of the loop 912 // to force them to not combine and return back inside the loop 913 // during GVN optimization (4641526). 914 // 915 // Because we are setting the actual control input, factor in 916 // the result from get_late_ctrl() so we respect any 917 // anti-dependences. (6233005). 918 x_ctrl = dom_lca(late_load_ctrl, x_ctrl); 919 920 // Don't allow the control input to be a CFG splitting node. 921 // Such nodes should only have ProjNodes as outs, e.g. IfNode 922 // should only have IfTrueNode and IfFalseNode (4985384). 923 x_ctrl = find_non_split_ctrl(x_ctrl); 924 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone"); 925 926 x->set_req(0, x_ctrl); 927 } 928 register_new_node(x, x_ctrl); 929 930 // Some institutional knowledge is needed here: 'x' is 931 // yanked because if the optimizer runs GVN on it all the 932 // cloned x's will common up and undo this optimization and 933 // be forced back in the loop. This is annoying because it 934 // makes +VerifyOpto report false-positives on progress. I 935 // tried setting control edges on the x's to force them to 936 // not combine, but the matching gets worried when it tries 937 // to fold a StoreP and an AddP together (as part of an 938 // address expression) and the AddP and StoreP have 939 // different controls. 940 if( !x->is_Load() && !x->is_DecodeN() ) _igvn._worklist.yank(x); 941 } 942 _igvn.remove_dead_node(n); 943 } 944 } 945 } 946 } 947 948 // Check for Opaque2's who's loop has disappeared - who's input is in the 949 // same loop nest as their output. Remove 'em, they are no longer useful. 950 if( n_op == Op_Opaque2 && 951 n->in(1) != NULL && 952 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) { 953 _igvn.add_users_to_worklist(n); 954 _igvn.hash_delete(n); 955 _igvn.subsume_node( n, n->in(1) ); 956 } 957} 958 959//------------------------------split_if_with_blocks--------------------------- 960// Check for aggressive application of 'split-if' optimization, 961// using basic block level info. 962void PhaseIdealLoop::split_if_with_blocks( VectorSet &visited, Node_Stack &nstack ) { 963 Node *n = C->root(); 964 visited.set(n->_idx); // first, mark node as visited 965 // Do pre-visit work for root 966 n = split_if_with_blocks_pre( n ); 967 uint cnt = n->outcnt(); 968 uint i = 0; 969 while (true) { 970 // Visit all children 971 if (i < cnt) { 972 Node* use = n->raw_out(i); 973 ++i; 974 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) { 975 // Now do pre-visit work for this use 976 use = split_if_with_blocks_pre( use ); 977 nstack.push(n, i); // Save parent and next use's index. 978 n = use; // Process all children of current use. 979 cnt = use->outcnt(); 980 i = 0; 981 } 982 } 983 else { 984 // All of n's children have been processed, complete post-processing. 985 if (cnt != 0 && !n->is_Con()) { 986 assert(has_node(n), "no dead nodes"); 987 split_if_with_blocks_post( n ); 988 } 989 if (nstack.is_empty()) { 990 // Finished all nodes on stack. 991 break; 992 } 993 // Get saved parent node and next use's index. Visit the rest of uses. 994 n = nstack.node(); 995 cnt = n->outcnt(); 996 i = nstack.index(); 997 nstack.pop(); 998 } 999 } 1000} 1001 1002 1003//============================================================================= 1004// 1005// C L O N E A L O O P B O D Y 1006// 1007 1008//------------------------------clone_iff-------------------------------------- 1009// Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1010// "Nearly" because all Nodes have been cloned from the original in the loop, 1011// but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1012// through the Phi recursively, and return a Bool. 1013BoolNode *PhaseIdealLoop::clone_iff( PhiNode *phi, IdealLoopTree *loop ) { 1014 1015 // Convert this Phi into a Phi merging Bools 1016 uint i; 1017 for( i = 1; i < phi->req(); i++ ) { 1018 Node *b = phi->in(i); 1019 if( b->is_Phi() ) { 1020 _igvn.hash_delete(phi); 1021 _igvn._worklist.push(phi); 1022 phi->set_req(i, clone_iff( b->as_Phi(), loop )); 1023 } else { 1024 assert( b->is_Bool(), "" ); 1025 } 1026 } 1027 1028 Node *sample_bool = phi->in(1); 1029 Node *sample_cmp = sample_bool->in(1); 1030 1031 // Make Phis to merge the Cmp's inputs. 1032 int size = phi->in(0)->req(); 1033 PhiNode *phi1 = new (C, size) PhiNode( phi->in(0), Type::TOP ); 1034 PhiNode *phi2 = new (C, size) PhiNode( phi->in(0), Type::TOP ); 1035 for( i = 1; i < phi->req(); i++ ) { 1036 Node *n1 = phi->in(i)->in(1)->in(1); 1037 Node *n2 = phi->in(i)->in(1)->in(2); 1038 phi1->set_req( i, n1 ); 1039 phi2->set_req( i, n2 ); 1040 phi1->set_type( phi1->type()->meet(n1->bottom_type()) ); 1041 phi2->set_type( phi2->type()->meet(n2->bottom_type()) ); 1042 } 1043 // See if these Phis have been made before. 1044 // Register with optimizer 1045 Node *hit1 = _igvn.hash_find_insert(phi1); 1046 if( hit1 ) { // Hit, toss just made Phi 1047 _igvn.remove_dead_node(phi1); // Remove new phi 1048 assert( hit1->is_Phi(), "" ); 1049 phi1 = (PhiNode*)hit1; // Use existing phi 1050 } else { // Miss 1051 _igvn.register_new_node_with_optimizer(phi1); 1052 } 1053 Node *hit2 = _igvn.hash_find_insert(phi2); 1054 if( hit2 ) { // Hit, toss just made Phi 1055 _igvn.remove_dead_node(phi2); // Remove new phi 1056 assert( hit2->is_Phi(), "" ); 1057 phi2 = (PhiNode*)hit2; // Use existing phi 1058 } else { // Miss 1059 _igvn.register_new_node_with_optimizer(phi2); 1060 } 1061 // Register Phis with loop/block info 1062 set_ctrl(phi1, phi->in(0)); 1063 set_ctrl(phi2, phi->in(0)); 1064 // Make a new Cmp 1065 Node *cmp = sample_cmp->clone(); 1066 cmp->set_req( 1, phi1 ); 1067 cmp->set_req( 2, phi2 ); 1068 _igvn.register_new_node_with_optimizer(cmp); 1069 set_ctrl(cmp, phi->in(0)); 1070 1071 // Make a new Bool 1072 Node *b = sample_bool->clone(); 1073 b->set_req(1,cmp); 1074 _igvn.register_new_node_with_optimizer(b); 1075 set_ctrl(b, phi->in(0)); 1076 1077 assert( b->is_Bool(), "" ); 1078 return (BoolNode*)b; 1079} 1080 1081//------------------------------clone_bool------------------------------------- 1082// Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1083// "Nearly" because all Nodes have been cloned from the original in the loop, 1084// but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1085// through the Phi recursively, and return a Bool. 1086CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) { 1087 uint i; 1088 // Convert this Phi into a Phi merging Bools 1089 for( i = 1; i < phi->req(); i++ ) { 1090 Node *b = phi->in(i); 1091 if( b->is_Phi() ) { 1092 _igvn.hash_delete(phi); 1093 _igvn._worklist.push(phi); 1094 phi->set_req(i, clone_bool( b->as_Phi(), loop )); 1095 } else { 1096 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" ); 1097 } 1098 } 1099 1100 Node *sample_cmp = phi->in(1); 1101 1102 // Make Phis to merge the Cmp's inputs. 1103 int size = phi->in(0)->req(); 1104 PhiNode *phi1 = new (C, size) PhiNode( phi->in(0), Type::TOP ); 1105 PhiNode *phi2 = new (C, size) PhiNode( phi->in(0), Type::TOP ); 1106 for( uint j = 1; j < phi->req(); j++ ) { 1107 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP 1108 Node *n1, *n2; 1109 if( cmp_top->is_Cmp() ) { 1110 n1 = cmp_top->in(1); 1111 n2 = cmp_top->in(2); 1112 } else { 1113 n1 = n2 = cmp_top; 1114 } 1115 phi1->set_req( j, n1 ); 1116 phi2->set_req( j, n2 ); 1117 phi1->set_type( phi1->type()->meet(n1->bottom_type()) ); 1118 phi2->set_type( phi2->type()->meet(n2->bottom_type()) ); 1119 } 1120 1121 // See if these Phis have been made before. 1122 // Register with optimizer 1123 Node *hit1 = _igvn.hash_find_insert(phi1); 1124 if( hit1 ) { // Hit, toss just made Phi 1125 _igvn.remove_dead_node(phi1); // Remove new phi 1126 assert( hit1->is_Phi(), "" ); 1127 phi1 = (PhiNode*)hit1; // Use existing phi 1128 } else { // Miss 1129 _igvn.register_new_node_with_optimizer(phi1); 1130 } 1131 Node *hit2 = _igvn.hash_find_insert(phi2); 1132 if( hit2 ) { // Hit, toss just made Phi 1133 _igvn.remove_dead_node(phi2); // Remove new phi 1134 assert( hit2->is_Phi(), "" ); 1135 phi2 = (PhiNode*)hit2; // Use existing phi 1136 } else { // Miss 1137 _igvn.register_new_node_with_optimizer(phi2); 1138 } 1139 // Register Phis with loop/block info 1140 set_ctrl(phi1, phi->in(0)); 1141 set_ctrl(phi2, phi->in(0)); 1142 // Make a new Cmp 1143 Node *cmp = sample_cmp->clone(); 1144 cmp->set_req( 1, phi1 ); 1145 cmp->set_req( 2, phi2 ); 1146 _igvn.register_new_node_with_optimizer(cmp); 1147 set_ctrl(cmp, phi->in(0)); 1148 1149 assert( cmp->is_Cmp(), "" ); 1150 return (CmpNode*)cmp; 1151} 1152 1153//------------------------------sink_use--------------------------------------- 1154// If 'use' was in the loop-exit block, it now needs to be sunk 1155// below the post-loop merge point. 1156void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) { 1157 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) { 1158 set_ctrl(use, post_loop); 1159 for (DUIterator j = use->outs(); use->has_out(j); j++) 1160 sink_use(use->out(j), post_loop); 1161 } 1162} 1163 1164//------------------------------clone_loop------------------------------------- 1165// 1166// C L O N E A L O O P B O D Y 1167// 1168// This is the basic building block of the loop optimizations. It clones an 1169// entire loop body. It makes an old_new loop body mapping; with this mapping 1170// you can find the new-loop equivalent to an old-loop node. All new-loop 1171// nodes are exactly equal to their old-loop counterparts, all edges are the 1172// same. All exits from the old-loop now have a RegionNode that merges the 1173// equivalent new-loop path. This is true even for the normal "loop-exit" 1174// condition. All uses of loop-invariant old-loop values now come from (one 1175// or more) Phis that merge their new-loop equivalents. 1176// 1177// This operation leaves the graph in an illegal state: there are two valid 1178// control edges coming from the loop pre-header to both loop bodies. I'll 1179// definitely have to hack the graph after running this transform. 1180// 1181// From this building block I will further edit edges to perform loop peeling 1182// or loop unrolling or iteration splitting (Range-Check-Elimination), etc. 1183// 1184// Parameter side_by_size_idom: 1185// When side_by_size_idom is NULL, the dominator tree is constructed for 1186// the clone loop to dominate the original. Used in construction of 1187// pre-main-post loop sequence. 1188// When nonnull, the clone and original are side-by-side, both are 1189// dominated by the side_by_side_idom node. Used in construction of 1190// unswitched loops. 1191void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd, 1192 Node* side_by_side_idom) { 1193 1194 // Step 1: Clone the loop body. Make the old->new mapping. 1195 uint i; 1196 for( i = 0; i < loop->_body.size(); i++ ) { 1197 Node *old = loop->_body.at(i); 1198 Node *nnn = old->clone(); 1199 old_new.map( old->_idx, nnn ); 1200 _igvn.register_new_node_with_optimizer(nnn); 1201 } 1202 1203 1204 // Step 2: Fix the edges in the new body. If the old input is outside the 1205 // loop use it. If the old input is INside the loop, use the corresponding 1206 // new node instead. 1207 for( i = 0; i < loop->_body.size(); i++ ) { 1208 Node *old = loop->_body.at(i); 1209 Node *nnn = old_new[old->_idx]; 1210 // Fix CFG/Loop controlling the new node 1211 if (has_ctrl(old)) { 1212 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]); 1213 } else { 1214 set_loop(nnn, loop->_parent); 1215 if (old->outcnt() > 0) { 1216 set_idom( nnn, old_new[idom(old)->_idx], dd ); 1217 } 1218 } 1219 // Correct edges to the new node 1220 for( uint j = 0; j < nnn->req(); j++ ) { 1221 Node *n = nnn->in(j); 1222 if( n ) { 1223 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n ); 1224 if( loop->is_member( old_in_loop ) ) 1225 nnn->set_req(j, old_new[n->_idx]); 1226 } 1227 } 1228 _igvn.hash_find_insert(nnn); 1229 } 1230 Node *newhead = old_new[loop->_head->_idx]; 1231 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 1232 1233 1234 // Step 3: Now fix control uses. Loop varying control uses have already 1235 // been fixed up (as part of all input edges in Step 2). Loop invariant 1236 // control uses must be either an IfFalse or an IfTrue. Make a merge 1237 // point to merge the old and new IfFalse/IfTrue nodes; make the use 1238 // refer to this. 1239 ResourceArea *area = Thread::current()->resource_area(); 1240 Node_List worklist(area); 1241 uint new_counter = C->unique(); 1242 for( i = 0; i < loop->_body.size(); i++ ) { 1243 Node* old = loop->_body.at(i); 1244 if( !old->is_CFG() ) continue; 1245 Node* nnn = old_new[old->_idx]; 1246 1247 // Copy uses to a worklist, so I can munge the def-use info 1248 // with impunity. 1249 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1250 worklist.push(old->fast_out(j)); 1251 1252 while( worklist.size() ) { // Visit all uses 1253 Node *use = worklist.pop(); 1254 if (!has_node(use)) continue; // Ignore dead nodes 1255 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1256 if( !loop->is_member( use_loop ) && use->is_CFG() ) { 1257 // Both OLD and USE are CFG nodes here. 1258 assert( use->is_Proj(), "" ); 1259 1260 // Clone the loop exit control projection 1261 Node *newuse = use->clone(); 1262 newuse->set_req(0,nnn); 1263 _igvn.register_new_node_with_optimizer(newuse); 1264 set_loop(newuse, use_loop); 1265 set_idom(newuse, nnn, dom_depth(nnn) + 1 ); 1266 1267 // We need a Region to merge the exit from the peeled body and the 1268 // exit from the old loop body. 1269 RegionNode *r = new (C, 3) RegionNode(3); 1270 // Map the old use to the new merge point 1271 old_new.map( use->_idx, r ); 1272 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use)); 1273 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" ); 1274 1275 // The original user of 'use' uses 'r' instead. 1276 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) { 1277 Node* useuse = use->last_out(l); 1278 _igvn.hash_delete(useuse); 1279 _igvn._worklist.push(useuse); 1280 uint uses_found = 0; 1281 if( useuse->in(0) == use ) { 1282 useuse->set_req(0, r); 1283 uses_found++; 1284 if( useuse->is_CFG() ) { 1285 assert( dom_depth(useuse) > dd_r, "" ); 1286 set_idom(useuse, r, dom_depth(useuse)); 1287 } 1288 } 1289 for( uint k = 1; k < useuse->req(); k++ ) { 1290 if( useuse->in(k) == use ) { 1291 useuse->set_req(k, r); 1292 uses_found++; 1293 } 1294 } 1295 l -= uses_found; // we deleted 1 or more copies of this edge 1296 } 1297 1298 // Now finish up 'r' 1299 r->set_req( 1, newuse ); 1300 r->set_req( 2, use ); 1301 _igvn.register_new_node_with_optimizer(r); 1302 set_loop(r, use_loop); 1303 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r); 1304 } // End of if a loop-exit test 1305 } 1306 } 1307 1308 // Step 4: If loop-invariant use is not control, it must be dominated by a 1309 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region 1310 // there if needed. Make a Phi there merging old and new used values. 1311 Node_List *split_if_set = NULL; 1312 Node_List *split_bool_set = NULL; 1313 Node_List *split_cex_set = NULL; 1314 for( i = 0; i < loop->_body.size(); i++ ) { 1315 Node* old = loop->_body.at(i); 1316 Node* nnn = old_new[old->_idx]; 1317 // Copy uses to a worklist, so I can munge the def-use info 1318 // with impunity. 1319 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1320 worklist.push(old->fast_out(j)); 1321 1322 while( worklist.size() ) { 1323 Node *use = worklist.pop(); 1324 if (!has_node(use)) continue; // Ignore dead nodes 1325 if (use->in(0) == C->top()) continue; 1326 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1327 // Check for data-use outside of loop - at least one of OLD or USE 1328 // must not be a CFG node. 1329 if( !loop->is_member( use_loop ) && (!old->is_CFG() || !use->is_CFG())) { 1330 1331 // If the Data use is an IF, that means we have an IF outside of the 1332 // loop that is switching on a condition that is set inside of the 1333 // loop. Happens if people set a loop-exit flag; then test the flag 1334 // in the loop to break the loop, then test is again outside of the 1335 // loop to determine which way the loop exited. 1336 if( use->is_If() || use->is_CMove() ) { 1337 // Since this code is highly unlikely, we lazily build the worklist 1338 // of such Nodes to go split. 1339 if( !split_if_set ) 1340 split_if_set = new Node_List(area); 1341 split_if_set->push(use); 1342 } 1343 if( use->is_Bool() ) { 1344 if( !split_bool_set ) 1345 split_bool_set = new Node_List(area); 1346 split_bool_set->push(use); 1347 } 1348 if( use->Opcode() == Op_CreateEx ) { 1349 if( !split_cex_set ) 1350 split_cex_set = new Node_List(area); 1351 split_cex_set->push(use); 1352 } 1353 1354 1355 // Get "block" use is in 1356 uint idx = 0; 1357 while( use->in(idx) != old ) idx++; 1358 Node *prev = use->is_CFG() ? use : get_ctrl(use); 1359 assert( !loop->is_member( get_loop( prev ) ), "" ); 1360 Node *cfg = prev->_idx >= new_counter 1361 ? prev->in(2) 1362 : idom(prev); 1363 if( use->is_Phi() ) // Phi use is in prior block 1364 cfg = prev->in(idx); // NOT in block of Phi itself 1365 if (cfg->is_top()) { // Use is dead? 1366 _igvn.hash_delete(use); 1367 _igvn._worklist.push(use); 1368 use->set_req(idx, C->top()); 1369 continue; 1370 } 1371 1372 while( !loop->is_member( get_loop( cfg ) ) ) { 1373 prev = cfg; 1374 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg); 1375 } 1376 // If the use occurs after merging several exits from the loop, then 1377 // old value must have dominated all those exits. Since the same old 1378 // value was used on all those exits we did not need a Phi at this 1379 // merge point. NOW we do need a Phi here. Each loop exit value 1380 // is now merged with the peeled body exit; each exit gets its own 1381 // private Phi and those Phis need to be merged here. 1382 Node *phi; 1383 if( prev->is_Region() ) { 1384 if( idx == 0 ) { // Updating control edge? 1385 phi = prev; // Just use existing control 1386 } else { // Else need a new Phi 1387 phi = PhiNode::make( prev, old ); 1388 // Now recursively fix up the new uses of old! 1389 for( uint i = 1; i < prev->req(); i++ ) { 1390 worklist.push(phi); // Onto worklist once for each 'old' input 1391 } 1392 } 1393 } else { 1394 // Get new RegionNode merging old and new loop exits 1395 prev = old_new[prev->_idx]; 1396 assert( prev, "just made this in step 7" ); 1397 if( idx == 0 ) { // Updating control edge? 1398 phi = prev; // Just use existing control 1399 } else { // Else need a new Phi 1400 // Make a new Phi merging data values properly 1401 phi = PhiNode::make( prev, old ); 1402 phi->set_req( 1, nnn ); 1403 } 1404 } 1405 // If inserting a new Phi, check for prior hits 1406 if( idx != 0 ) { 1407 Node *hit = _igvn.hash_find_insert(phi); 1408 if( hit == NULL ) { 1409 _igvn.register_new_node_with_optimizer(phi); // Register new phi 1410 } else { // or 1411 // Remove the new phi from the graph and use the hit 1412 _igvn.remove_dead_node(phi); 1413 phi = hit; // Use existing phi 1414 } 1415 set_ctrl(phi, prev); 1416 } 1417 // Make 'use' use the Phi instead of the old loop body exit value 1418 _igvn.hash_delete(use); 1419 _igvn._worklist.push(use); 1420 use->set_req(idx, phi); 1421 if( use->_idx >= new_counter ) { // If updating new phis 1422 // Not needed for correctness, but prevents a weak assert 1423 // in AddPNode from tripping (when we end up with different 1424 // base & derived Phis that will become the same after 1425 // IGVN does CSE). 1426 Node *hit = _igvn.hash_find_insert(use); 1427 if( hit ) // Go ahead and re-hash for hits. 1428 _igvn.subsume_node( use, hit ); 1429 } 1430 1431 // If 'use' was in the loop-exit block, it now needs to be sunk 1432 // below the post-loop merge point. 1433 sink_use( use, prev ); 1434 } 1435 } 1436 } 1437 1438 // Check for IFs that need splitting/cloning. Happens if an IF outside of 1439 // the loop uses a condition set in the loop. The original IF probably 1440 // takes control from one or more OLD Regions (which in turn get from NEW 1441 // Regions). In any case, there will be a set of Phis for each merge point 1442 // from the IF up to where the original BOOL def exists the loop. 1443 if( split_if_set ) { 1444 while( split_if_set->size() ) { 1445 Node *iff = split_if_set->pop(); 1446 if( iff->in(1)->is_Phi() ) { 1447 BoolNode *b = clone_iff( iff->in(1)->as_Phi(), loop ); 1448 _igvn.hash_delete(iff); 1449 _igvn._worklist.push(iff); 1450 iff->set_req(1, b); 1451 } 1452 } 1453 } 1454 if( split_bool_set ) { 1455 while( split_bool_set->size() ) { 1456 Node *b = split_bool_set->pop(); 1457 Node *phi = b->in(1); 1458 assert( phi->is_Phi(), "" ); 1459 CmpNode *cmp = clone_bool( (PhiNode*)phi, loop ); 1460 _igvn.hash_delete(b); 1461 _igvn._worklist.push(b); 1462 b->set_req(1, cmp); 1463 } 1464 } 1465 if( split_cex_set ) { 1466 while( split_cex_set->size() ) { 1467 Node *b = split_cex_set->pop(); 1468 assert( b->in(0)->is_Region(), "" ); 1469 assert( b->in(1)->is_Phi(), "" ); 1470 assert( b->in(0)->in(0) == b->in(1)->in(0), "" ); 1471 split_up( b, b->in(0), NULL ); 1472 } 1473 } 1474 1475} 1476 1477 1478//---------------------- stride_of_possible_iv ------------------------------------- 1479// Looks for an iff/bool/comp with one operand of the compare 1480// being a cycle involving an add and a phi, 1481// with an optional truncation (left-shift followed by a right-shift) 1482// of the add. Returns zero if not an iv. 1483int PhaseIdealLoop::stride_of_possible_iv(Node* iff) { 1484 Node* trunc1 = NULL; 1485 Node* trunc2 = NULL; 1486 const TypeInt* ttype = NULL; 1487 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) { 1488 return 0; 1489 } 1490 BoolNode* bl = iff->in(1)->as_Bool(); 1491 Node* cmp = bl->in(1); 1492 if (!cmp || cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU) { 1493 return 0; 1494 } 1495 // Must have an invariant operand 1496 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) { 1497 return 0; 1498 } 1499 Node* add2 = NULL; 1500 Node* cmp1 = cmp->in(1); 1501 if (cmp1->is_Phi()) { 1502 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) ))) 1503 Node* phi = cmp1; 1504 for (uint i = 1; i < phi->req(); i++) { 1505 Node* in = phi->in(i); 1506 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in, 1507 &trunc1, &trunc2, &ttype); 1508 if (add && add->in(1) == phi) { 1509 add2 = add->in(2); 1510 break; 1511 } 1512 } 1513 } else { 1514 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) ))) 1515 Node* addtrunc = cmp1; 1516 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc, 1517 &trunc1, &trunc2, &ttype); 1518 if (add && add->in(1)->is_Phi()) { 1519 Node* phi = add->in(1); 1520 for (uint i = 1; i < phi->req(); i++) { 1521 if (phi->in(i) == addtrunc) { 1522 add2 = add->in(2); 1523 break; 1524 } 1525 } 1526 } 1527 } 1528 if (add2 != NULL) { 1529 const TypeInt* add2t = _igvn.type(add2)->is_int(); 1530 if (add2t->is_con()) { 1531 return add2t->get_con(); 1532 } 1533 } 1534 return 0; 1535} 1536 1537 1538//---------------------- stay_in_loop ------------------------------------- 1539// Return the (unique) control output node that's in the loop (if it exists.) 1540Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) { 1541 Node* unique = NULL; 1542 if (!n) return NULL; 1543 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1544 Node* use = n->fast_out(i); 1545 if (!has_ctrl(use) && loop->is_member(get_loop(use))) { 1546 if (unique != NULL) { 1547 return NULL; 1548 } 1549 unique = use; 1550 } 1551 } 1552 return unique; 1553} 1554 1555//------------------------------ register_node ------------------------------------- 1556// Utility to register node "n" with PhaseIdealLoop 1557void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) { 1558 _igvn.register_new_node_with_optimizer(n); 1559 loop->_body.push(n); 1560 if (n->is_CFG()) { 1561 set_loop(n, loop); 1562 set_idom(n, pred, ddepth); 1563 } else { 1564 set_ctrl(n, pred); 1565 } 1566} 1567 1568//------------------------------ proj_clone ------------------------------------- 1569// Utility to create an if-projection 1570ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) { 1571 ProjNode* c = p->clone()->as_Proj(); 1572 c->set_req(0, iff); 1573 return c; 1574} 1575 1576//------------------------------ short_circuit_if ------------------------------------- 1577// Force the iff control output to be the live_proj 1578Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) { 1579 int proj_con = live_proj->_con; 1580 assert(proj_con == 0 || proj_con == 1, "false or true projection"); 1581 Node *con = _igvn.intcon(proj_con); 1582 set_ctrl(con, C->root()); 1583 if (iff) { 1584 iff->set_req(1, con); 1585 } 1586 return con; 1587} 1588 1589//------------------------------ insert_if_before_proj ------------------------------------- 1590// Insert a new if before an if projection (* - new node) 1591// 1592// before 1593// if(test) 1594// / \ 1595// v v 1596// other-proj proj (arg) 1597// 1598// after 1599// if(test) 1600// / \ 1601// / v 1602// | * proj-clone 1603// v | 1604// other-proj v 1605// * new_if(relop(cmp[IU](left,right))) 1606// / \ 1607// v v 1608// * new-proj proj 1609// (returned) 1610// 1611ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) { 1612 IfNode* iff = proj->in(0)->as_If(); 1613 IdealLoopTree *loop = get_loop(proj); 1614 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 1615 int ddepth = dom_depth(proj); 1616 1617 _igvn.hash_delete(iff); 1618 _igvn._worklist.push(iff); 1619 _igvn.hash_delete(proj); 1620 _igvn._worklist.push(proj); 1621 1622 proj->set_req(0, NULL); // temporary disconnect 1623 ProjNode* proj2 = proj_clone(proj, iff); 1624 register_node(proj2, loop, iff, ddepth); 1625 1626 Node* cmp = Signed ? (Node*) new (C,3)CmpINode(left, right) : (Node*) new (C,3)CmpUNode(left, right); 1627 register_node(cmp, loop, proj2, ddepth); 1628 1629 BoolNode* bol = new (C,2)BoolNode(cmp, relop); 1630 register_node(bol, loop, proj2, ddepth); 1631 1632 IfNode* new_if = new (C,2)IfNode(proj2, bol, iff->_prob, iff->_fcnt); 1633 register_node(new_if, loop, proj2, ddepth); 1634 1635 proj->set_req(0, new_if); // reattach 1636 set_idom(proj, new_if, ddepth); 1637 1638 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj(); 1639 register_node(new_exit, get_loop(other_proj), new_if, ddepth); 1640 1641 return new_exit; 1642} 1643 1644//------------------------------ insert_region_before_proj ------------------------------------- 1645// Insert a region before an if projection (* - new node) 1646// 1647// before 1648// if(test) 1649// / | 1650// v | 1651// proj v 1652// other-proj 1653// 1654// after 1655// if(test) 1656// / | 1657// v | 1658// * proj-clone v 1659// | other-proj 1660// v 1661// * new-region 1662// | 1663// v 1664// * dum_if 1665// / \ 1666// v \ 1667// * dum-proj v 1668// proj 1669// 1670RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) { 1671 IfNode* iff = proj->in(0)->as_If(); 1672 IdealLoopTree *loop = get_loop(proj); 1673 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 1674 int ddepth = dom_depth(proj); 1675 1676 _igvn.hash_delete(iff); 1677 _igvn._worklist.push(iff); 1678 _igvn.hash_delete(proj); 1679 _igvn._worklist.push(proj); 1680 1681 proj->set_req(0, NULL); // temporary disconnect 1682 ProjNode* proj2 = proj_clone(proj, iff); 1683 register_node(proj2, loop, iff, ddepth); 1684 1685 RegionNode* reg = new (C,2)RegionNode(2); 1686 reg->set_req(1, proj2); 1687 register_node(reg, loop, iff, ddepth); 1688 1689 IfNode* dum_if = new (C,2)IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt); 1690 register_node(dum_if, loop, reg, ddepth); 1691 1692 proj->set_req(0, dum_if); // reattach 1693 set_idom(proj, dum_if, ddepth); 1694 1695 ProjNode* dum_proj = proj_clone(other_proj, dum_if); 1696 register_node(dum_proj, loop, dum_if, ddepth); 1697 1698 return reg; 1699} 1700 1701//------------------------------ insert_cmpi_loop_exit ------------------------------------- 1702// Clone a signed compare loop exit from an unsigned compare and 1703// insert it before the unsigned cmp on the stay-in-loop path. 1704// All new nodes inserted in the dominator tree between the original 1705// if and it's projections. The original if test is replaced with 1706// a constant to force the stay-in-loop path. 1707// 1708// This is done to make sure that the original if and it's projections 1709// still dominate the same set of control nodes, that the ctrl() relation 1710// from data nodes to them is preserved, and that their loop nesting is 1711// preserved. 1712// 1713// before 1714// if(i <u limit) unsigned compare loop exit 1715// / | 1716// v v 1717// exit-proj stay-in-loop-proj 1718// 1719// after 1720// if(stay-in-loop-const) original if 1721// / | 1722// / v 1723// / if(i < limit) new signed test 1724// / / | 1725// / / v 1726// / / if(i <u limit) new cloned unsigned test 1727// / / / | 1728// v v v | 1729// region | 1730// | | 1731// dum-if | 1732// / | | 1733// ether | | 1734// v v 1735// exit-proj stay-in-loop-proj 1736// 1737IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) { 1738 const bool Signed = true; 1739 const bool Unsigned = false; 1740 1741 BoolNode* bol = if_cmpu->in(1)->as_Bool(); 1742 if (bol->_test._test != BoolTest::lt) return NULL; 1743 CmpNode* cmpu = bol->in(1)->as_Cmp(); 1744 if (cmpu->Opcode() != Op_CmpU) return NULL; 1745 int stride = stride_of_possible_iv(if_cmpu); 1746 if (stride == 0) return NULL; 1747 1748 ProjNode* lp_continue = stay_in_loop(if_cmpu, loop)->as_Proj(); 1749 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj(); 1750 1751 Node* limit = NULL; 1752 if (stride > 0) { 1753 limit = cmpu->in(2); 1754 } else { 1755 limit = _igvn.makecon(TypeInt::ZERO); 1756 set_ctrl(limit, C->root()); 1757 } 1758 // Create a new region on the exit path 1759 RegionNode* reg = insert_region_before_proj(lp_exit); 1760 1761 // Clone the if-cmpu-true-false using a signed compare 1762 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge; 1763 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue); 1764 reg->add_req(cmpi_exit); 1765 1766 // Clone the if-cmpu-true-false 1767 BoolTest::mask rel_u = bol->_test._test; 1768 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue); 1769 reg->add_req(cmpu_exit); 1770 1771 // Force original if to stay in loop. 1772 short_circuit_if(if_cmpu, lp_continue); 1773 1774 return cmpi_exit->in(0)->as_If(); 1775} 1776 1777//------------------------------ remove_cmpi_loop_exit ------------------------------------- 1778// Remove a previously inserted signed compare loop exit. 1779void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) { 1780 Node* lp_proj = stay_in_loop(if_cmp, loop); 1781 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI && 1782 stay_in_loop(lp_proj, loop)->is_If() && 1783 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu"); 1784 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO); 1785 set_ctrl(con, C->root()); 1786 if_cmp->set_req(1, con); 1787} 1788 1789//------------------------------ scheduled_nodelist ------------------------------------- 1790// Create a post order schedule of nodes that are in the 1791// "member" set. The list is returned in "sched". 1792// The first node in "sched" is the loop head, followed by 1793// nodes which have no inputs in the "member" set, and then 1794// followed by the nodes that have an immediate input dependence 1795// on a node in "sched". 1796void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) { 1797 1798 assert(member.test(loop->_head->_idx), "loop head must be in member set"); 1799 Arena *a = Thread::current()->resource_area(); 1800 VectorSet visited(a); 1801 Node_Stack nstack(a, loop->_body.size()); 1802 1803 Node* n = loop->_head; // top of stack is cached in "n" 1804 uint idx = 0; 1805 visited.set(n->_idx); 1806 1807 // Initially push all with no inputs from within member set 1808 for(uint i = 0; i < loop->_body.size(); i++ ) { 1809 Node *elt = loop->_body.at(i); 1810 if (member.test(elt->_idx)) { 1811 bool found = false; 1812 for (uint j = 0; j < elt->req(); j++) { 1813 Node* def = elt->in(j); 1814 if (def && member.test(def->_idx) && def != elt) { 1815 found = true; 1816 break; 1817 } 1818 } 1819 if (!found && elt != loop->_head) { 1820 nstack.push(n, idx); 1821 n = elt; 1822 assert(!visited.test(n->_idx), "not seen yet"); 1823 visited.set(n->_idx); 1824 } 1825 } 1826 } 1827 1828 // traverse out's that are in the member set 1829 while (true) { 1830 if (idx < n->outcnt()) { 1831 Node* use = n->raw_out(idx); 1832 idx++; 1833 if (!visited.test_set(use->_idx)) { 1834 if (member.test(use->_idx)) { 1835 nstack.push(n, idx); 1836 n = use; 1837 idx = 0; 1838 } 1839 } 1840 } else { 1841 // All outputs processed 1842 sched.push(n); 1843 if (nstack.is_empty()) break; 1844 n = nstack.node(); 1845 idx = nstack.index(); 1846 nstack.pop(); 1847 } 1848 } 1849} 1850 1851 1852//------------------------------ has_use_in_set ------------------------------------- 1853// Has a use in the vector set 1854bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) { 1855 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1856 Node* use = n->fast_out(j); 1857 if (vset.test(use->_idx)) { 1858 return true; 1859 } 1860 } 1861 return false; 1862} 1863 1864 1865//------------------------------ has_use_internal_to_set ------------------------------------- 1866// Has use internal to the vector set (ie. not in a phi at the loop head) 1867bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) { 1868 Node* head = loop->_head; 1869 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1870 Node* use = n->fast_out(j); 1871 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) { 1872 return true; 1873 } 1874 } 1875 return false; 1876} 1877 1878 1879//------------------------------ clone_for_use_outside_loop ------------------------------------- 1880// clone "n" for uses that are outside of loop 1881void PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) { 1882 1883 assert(worklist.size() == 0, "should be empty"); 1884 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1885 Node* use = n->fast_out(j); 1886 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) { 1887 worklist.push(use); 1888 } 1889 } 1890 while( worklist.size() ) { 1891 Node *use = worklist.pop(); 1892 if (!has_node(use) || use->in(0) == C->top()) continue; 1893 uint j; 1894 for (j = 0; j < use->req(); j++) { 1895 if (use->in(j) == n) break; 1896 } 1897 assert(j < use->req(), "must be there"); 1898 1899 // clone "n" and insert it between the inputs of "n" and the use outside the loop 1900 Node* n_clone = n->clone(); 1901 _igvn.hash_delete(use); 1902 use->set_req(j, n_clone); 1903 _igvn._worklist.push(use); 1904 Node* use_c; 1905 if (!use->is_Phi()) { 1906 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0); 1907 } else { 1908 // Use in a phi is considered a use in the associated predecessor block 1909 use_c = use->in(0)->in(j); 1910 } 1911 set_ctrl(n_clone, use_c); 1912 assert(!loop->is_member(get_loop(use_c)), "should be outside loop"); 1913 get_loop(use_c)->_body.push(n_clone); 1914 _igvn.register_new_node_with_optimizer(n_clone); 1915#if !defined(PRODUCT) 1916 if (TracePartialPeeling) { 1917 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx); 1918 } 1919#endif 1920 } 1921} 1922 1923 1924//------------------------------ clone_for_special_use_inside_loop ------------------------------------- 1925// clone "n" for special uses that are in the not_peeled region. 1926// If these def-uses occur in separate blocks, the code generator 1927// marks the method as not compilable. For example, if a "BoolNode" 1928// is in a different basic block than the "IfNode" that uses it, then 1929// the compilation is aborted in the code generator. 1930void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 1931 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) { 1932 if (n->is_Phi() || n->is_Load()) { 1933 return; 1934 } 1935 assert(worklist.size() == 0, "should be empty"); 1936 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1937 Node* use = n->fast_out(j); 1938 if ( not_peel.test(use->_idx) && 1939 (use->is_If() || use->is_CMove() || use->is_Bool()) && 1940 use->in(1) == n) { 1941 worklist.push(use); 1942 } 1943 } 1944 if (worklist.size() > 0) { 1945 // clone "n" and insert it between inputs of "n" and the use 1946 Node* n_clone = n->clone(); 1947 loop->_body.push(n_clone); 1948 _igvn.register_new_node_with_optimizer(n_clone); 1949 set_ctrl(n_clone, get_ctrl(n)); 1950 sink_list.push(n_clone); 1951 not_peel <<= n_clone->_idx; // add n_clone to not_peel set. 1952#if !defined(PRODUCT) 1953 if (TracePartialPeeling) { 1954 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx); 1955 } 1956#endif 1957 while( worklist.size() ) { 1958 Node *use = worklist.pop(); 1959 _igvn.hash_delete(use); 1960 _igvn._worklist.push(use); 1961 for (uint j = 1; j < use->req(); j++) { 1962 if (use->in(j) == n) { 1963 use->set_req(j, n_clone); 1964 } 1965 } 1966 } 1967 } 1968} 1969 1970 1971//------------------------------ insert_phi_for_loop ------------------------------------- 1972// Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 1973void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) { 1974 Node *phi = PhiNode::make(lp, back_edge_val); 1975 phi->set_req(LoopNode::EntryControl, lp_entry_val); 1976 // Use existing phi if it already exists 1977 Node *hit = _igvn.hash_find_insert(phi); 1978 if( hit == NULL ) { 1979 _igvn.register_new_node_with_optimizer(phi); 1980 set_ctrl(phi, lp); 1981 } else { 1982 // Remove the new phi from the graph and use the hit 1983 _igvn.remove_dead_node(phi); 1984 phi = hit; 1985 } 1986 _igvn.hash_delete(use); 1987 _igvn._worklist.push(use); 1988 use->set_req(idx, phi); 1989} 1990 1991#ifdef ASSERT 1992//------------------------------ is_valid_loop_partition ------------------------------------- 1993// Validate the loop partition sets: peel and not_peel 1994bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, 1995 VectorSet& not_peel ) { 1996 uint i; 1997 // Check that peel_list entries are in the peel set 1998 for (i = 0; i < peel_list.size(); i++) { 1999 if (!peel.test(peel_list.at(i)->_idx)) { 2000 return false; 2001 } 2002 } 2003 // Check at loop members are in one of peel set or not_peel set 2004 for (i = 0; i < loop->_body.size(); i++ ) { 2005 Node *def = loop->_body.at(i); 2006 uint di = def->_idx; 2007 // Check that peel set elements are in peel_list 2008 if (peel.test(di)) { 2009 if (not_peel.test(di)) { 2010 return false; 2011 } 2012 // Must be in peel_list also 2013 bool found = false; 2014 for (uint j = 0; j < peel_list.size(); j++) { 2015 if (peel_list.at(j)->_idx == di) { 2016 found = true; 2017 break; 2018 } 2019 } 2020 if (!found) { 2021 return false; 2022 } 2023 } else if (not_peel.test(di)) { 2024 if (peel.test(di)) { 2025 return false; 2026 } 2027 } else { 2028 return false; 2029 } 2030 } 2031 return true; 2032} 2033 2034//------------------------------ is_valid_clone_loop_exit_use ------------------------------------- 2035// Ensure a use outside of loop is of the right form 2036bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) { 2037 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2038 return (use->is_Phi() && 2039 use_c->is_Region() && use_c->req() == 3 && 2040 (use_c->in(exit_idx)->Opcode() == Op_IfTrue || 2041 use_c->in(exit_idx)->Opcode() == Op_IfFalse || 2042 use_c->in(exit_idx)->Opcode() == Op_JumpProj) && 2043 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) ); 2044} 2045 2046//------------------------------ is_valid_clone_loop_form ------------------------------------- 2047// Ensure that all uses outside of loop are of the right form 2048bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 2049 uint orig_exit_idx, uint clone_exit_idx) { 2050 uint len = peel_list.size(); 2051 for (uint i = 0; i < len; i++) { 2052 Node *def = peel_list.at(i); 2053 2054 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 2055 Node *use = def->fast_out(j); 2056 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2057 if (!loop->is_member(get_loop(use_c))) { 2058 // use is not in the loop, check for correct structure 2059 if (use->in(0) == def) { 2060 // Okay 2061 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) { 2062 return false; 2063 } 2064 } 2065 } 2066 } 2067 return true; 2068} 2069#endif 2070 2071//------------------------------ partial_peel ------------------------------------- 2072// Partially peel (aka loop rotation) the top portion of a loop (called 2073// the peel section below) by cloning it and placing one copy just before 2074// the new loop head and the other copy at the bottom of the new loop. 2075// 2076// before after where it came from 2077// 2078// stmt1 stmt1 2079// loop: stmt2 clone 2080// stmt2 if condA goto exitA clone 2081// if condA goto exitA new_loop: new 2082// stmt3 stmt3 clone 2083// if !condB goto loop if condB goto exitB clone 2084// exitB: stmt2 orig 2085// stmt4 if !condA goto new_loop orig 2086// exitA: goto exitA 2087// exitB: 2088// stmt4 2089// exitA: 2090// 2091// Step 1: find the cut point: an exit test on probable 2092// induction variable. 2093// Step 2: schedule (with cloning) operations in the peel 2094// section that can be executed after the cut into 2095// the section that is not peeled. This may need 2096// to clone operations into exit blocks. For 2097// instance, a reference to A[i] in the not-peel 2098// section and a reference to B[i] in an exit block 2099// may cause a left-shift of i by 2 to be placed 2100// in the peel block. This step will clone the left 2101// shift into the exit block and sink the left shift 2102// from the peel to the not-peel section. 2103// Step 3: clone the loop, retarget the control, and insert 2104// phis for values that are live across the new loop 2105// head. This is very dependent on the graph structure 2106// from clone_loop. It creates region nodes for 2107// exit control and associated phi nodes for values 2108// flow out of the loop through that exit. The region 2109// node is dominated by the clone's control projection. 2110// So the clone's peel section is placed before the 2111// new loop head, and the clone's not-peel section is 2112// forms the top part of the new loop. The original 2113// peel section forms the tail of the new loop. 2114// Step 4: update the dominator tree and recompute the 2115// dominator depth. 2116// 2117// orig 2118// 2119// stmt1 2120// | 2121// v 2122// loop<----+ 2123// | | 2124// stmt2 | 2125// | | 2126// v | 2127// ifA | 2128// / | | 2129// v v | 2130// false true ^ <-- last_peel 2131// / | | 2132// / ===|==cut | 2133// / stmt3 | <-- first_not_peel 2134// / | | 2135// | v | 2136// v ifB | 2137// exitA: / \ | 2138// / \ | 2139// v v | 2140// false true | 2141// / \ | 2142// / ----+ 2143// | 2144// v 2145// exitB: 2146// stmt4 2147// 2148// 2149// after clone loop 2150// 2151// stmt1 2152// / \ 2153// clone / \ orig 2154// / \ 2155// / \ 2156// v v 2157// +---->loop loop<----+ 2158// | | | | 2159// | stmt2 stmt2 | 2160// | | | | 2161// | v v | 2162// | ifA ifA | 2163// | | \ / | | 2164// | v v v v | 2165// ^ true false false true ^ <-- last_peel 2166// | | ^ \ / | | 2167// | cut==|== \ \ / ===|==cut | 2168// | stmt3 \ \ / stmt3 | <-- first_not_peel 2169// | | dom | | | | 2170// | v \ 1v v2 v | 2171// | ifB regionA ifB | 2172// | / \ | / \ | 2173// | / \ v / \ | 2174// | v v exitA: v v | 2175// | true false false true | 2176// | / ^ \ / \ | 2177// +---- \ \ / ----+ 2178// dom \ / 2179// \ 1v v2 2180// regionB 2181// | 2182// v 2183// exitB: 2184// stmt4 2185// 2186// 2187// after partial peel 2188// 2189// stmt1 2190// / 2191// clone / orig 2192// / TOP 2193// / \ 2194// v v 2195// TOP->region region----+ 2196// | | | 2197// stmt2 stmt2 | 2198// | | | 2199// v v | 2200// ifA ifA | 2201// | \ / | | 2202// v v v v | 2203// true false false true | <-- last_peel 2204// | ^ \ / +------|---+ 2205// +->newloop \ \ / === ==cut | | 2206// | stmt3 \ \ / TOP | | 2207// | | dom | | stmt3 | | <-- first_not_peel 2208// | v \ 1v v2 v | | 2209// | ifB regionA ifB ^ v 2210// | / \ | / \ | | 2211// | / \ v / \ | | 2212// | v v exitA: v v | | 2213// | true false false true | | 2214// | / ^ \ / \ | | 2215// | | \ \ / v | | 2216// | | dom \ / TOP | | 2217// | | \ 1v v2 | | 2218// ^ v regionB | | 2219// | | | | | 2220// | | v ^ v 2221// | | exitB: | | 2222// | | stmt4 | | 2223// | +------------>-----------------+ | 2224// | | 2225// +-----------------<---------------------+ 2226// 2227// 2228// final graph 2229// 2230// stmt1 2231// | 2232// v 2233// ........> ifA clone 2234// : / | 2235// dom / | 2236// : v v 2237// : false true 2238// : | | 2239// : | stmt2 clone 2240// : | | 2241// : | v 2242// : | newloop<-----+ 2243// : | | | 2244// : | stmt3 clone | 2245// : | | | 2246// : | v | 2247// : | ifB | 2248// : | / \ | 2249// : | v v | 2250// : | false true | 2251// : | | | | 2252// : | v stmt2 | 2253// : | exitB: | | 2254// : | stmt4 v | 2255// : | ifA orig | 2256// : | / \ | 2257// : | / \ | 2258// : | v v | 2259// : | false true | 2260// : | / \ | 2261// : v v -----+ 2262// RegionA 2263// | 2264// v 2265// exitA 2266// 2267bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) { 2268 2269 if (!loop->_head->is_Loop()) { 2270 return false; } 2271 2272 LoopNode *head = loop->_head->as_Loop(); 2273 2274 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) { 2275 return false; 2276 } 2277 2278 // Check for complex exit control 2279 for(uint ii = 0; ii < loop->_body.size(); ii++ ) { 2280 Node *n = loop->_body.at(ii); 2281 int opc = n->Opcode(); 2282 if (n->is_Call() || 2283 opc == Op_Catch || 2284 opc == Op_CatchProj || 2285 opc == Op_Jump || 2286 opc == Op_JumpProj) { 2287#if !defined(PRODUCT) 2288 if (TracePartialPeeling) { 2289 tty->print_cr("\nExit control too complex: lp: %d", head->_idx); 2290 } 2291#endif 2292 return false; 2293 } 2294 } 2295 2296 int dd = dom_depth(head); 2297 2298 // Step 1: find cut point 2299 2300 // Walk up dominators to loop head looking for first loop exit 2301 // which is executed on every path thru loop. 2302 IfNode *peel_if = NULL; 2303 IfNode *peel_if_cmpu = NULL; 2304 2305 Node *iff = loop->tail(); 2306 while( iff != head ) { 2307 if( iff->is_If() ) { 2308 Node *ctrl = get_ctrl(iff->in(1)); 2309 if (ctrl->is_top()) return false; // Dead test on live IF. 2310 // If loop-varying exit-test, check for induction variable 2311 if( loop->is_member(get_loop(ctrl)) && 2312 loop->is_loop_exit(iff) && 2313 is_possible_iv_test(iff)) { 2314 Node* cmp = iff->in(1)->in(1); 2315 if (cmp->Opcode() == Op_CmpI) { 2316 peel_if = iff->as_If(); 2317 } else { 2318 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU"); 2319 peel_if_cmpu = iff->as_If(); 2320 } 2321 } 2322 } 2323 iff = idom(iff); 2324 } 2325 // Prefer signed compare over unsigned compare. 2326 IfNode* new_peel_if = NULL; 2327 if (peel_if == NULL) { 2328 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) { 2329 return false; // No peel point found 2330 } 2331 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop); 2332 if (new_peel_if == NULL) { 2333 return false; // No peel point found 2334 } 2335 peel_if = new_peel_if; 2336 } 2337 Node* last_peel = stay_in_loop(peel_if, loop); 2338 Node* first_not_peeled = stay_in_loop(last_peel, loop); 2339 if (first_not_peeled == NULL || first_not_peeled == head) { 2340 return false; 2341 } 2342 2343#if !defined(PRODUCT) 2344 if (TracePartialPeeling) { 2345 tty->print_cr("before partial peel one iteration"); 2346 Node_List wl; 2347 Node* t = head->in(2); 2348 while (true) { 2349 wl.push(t); 2350 if (t == head) break; 2351 t = idom(t); 2352 } 2353 while (wl.size() > 0) { 2354 Node* tt = wl.pop(); 2355 tt->dump(); 2356 if (tt == last_peel) tty->print_cr("-- cut --"); 2357 } 2358 } 2359#endif 2360 ResourceArea *area = Thread::current()->resource_area(); 2361 VectorSet peel(area); 2362 VectorSet not_peel(area); 2363 Node_List peel_list(area); 2364 Node_List worklist(area); 2365 Node_List sink_list(area); 2366 2367 // Set of cfg nodes to peel are those that are executable from 2368 // the head through last_peel. 2369 assert(worklist.size() == 0, "should be empty"); 2370 worklist.push(head); 2371 peel.set(head->_idx); 2372 while (worklist.size() > 0) { 2373 Node *n = worklist.pop(); 2374 if (n != last_peel) { 2375 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2376 Node* use = n->fast_out(j); 2377 if (use->is_CFG() && 2378 loop->is_member(get_loop(use)) && 2379 !peel.test_set(use->_idx)) { 2380 worklist.push(use); 2381 } 2382 } 2383 } 2384 } 2385 2386 // Set of non-cfg nodes to peel are those that are control 2387 // dependent on the cfg nodes. 2388 uint i; 2389 for(i = 0; i < loop->_body.size(); i++ ) { 2390 Node *n = loop->_body.at(i); 2391 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n; 2392 if (peel.test(n_c->_idx)) { 2393 peel.set(n->_idx); 2394 } else { 2395 not_peel.set(n->_idx); 2396 } 2397 } 2398 2399 // Step 2: move operations from the peeled section down into the 2400 // not-peeled section 2401 2402 // Get a post order schedule of nodes in the peel region 2403 // Result in right-most operand. 2404 scheduled_nodelist(loop, peel, peel_list ); 2405 2406 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 2407 2408 // For future check for too many new phis 2409 uint old_phi_cnt = 0; 2410 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 2411 Node* use = head->fast_out(j); 2412 if (use->is_Phi()) old_phi_cnt++; 2413 } 2414 2415#if !defined(PRODUCT) 2416 if (TracePartialPeeling) { 2417 tty->print_cr("\npeeled list"); 2418 } 2419#endif 2420 2421 // Evacuate nodes in peel region into the not_peeled region if possible 2422 uint new_phi_cnt = 0; 2423 for (i = 0; i < peel_list.size();) { 2424 Node* n = peel_list.at(i); 2425#if !defined(PRODUCT) 2426 if (TracePartialPeeling) n->dump(); 2427#endif 2428 bool incr = true; 2429 if ( !n->is_CFG() ) { 2430 2431 if ( has_use_in_set(n, not_peel) ) { 2432 2433 // If not used internal to the peeled region, 2434 // move "n" from peeled to not_peeled region. 2435 2436 if ( !has_use_internal_to_set(n, peel, loop) ) { 2437 2438 // if not pinned and not a load (which maybe anti-dependent on a store) 2439 // and not a CMove (Matcher expects only bool->cmove). 2440 if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) { 2441 clone_for_use_outside_loop( loop, n, worklist ); 2442 2443 sink_list.push(n); 2444 peel >>= n->_idx; // delete n from peel set. 2445 not_peel <<= n->_idx; // add n to not_peel set. 2446 peel_list.remove(i); 2447 incr = false; 2448#if !defined(PRODUCT) 2449 if (TracePartialPeeling) { 2450 tty->print_cr("sink to not_peeled region: %d newbb: %d", 2451 n->_idx, get_ctrl(n)->_idx); 2452 } 2453#endif 2454 } 2455 } else { 2456 // Otherwise check for special def-use cases that span 2457 // the peel/not_peel boundary such as bool->if 2458 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist ); 2459 new_phi_cnt++; 2460 } 2461 } 2462 } 2463 if (incr) i++; 2464 } 2465 2466 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) { 2467#if !defined(PRODUCT) 2468 if (TracePartialPeeling) { 2469 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c", 2470 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F'); 2471 } 2472#endif 2473 if (new_peel_if != NULL) { 2474 remove_cmpi_loop_exit(new_peel_if, loop); 2475 } 2476 // Inhibit more partial peeling on this loop 2477 assert(!head->is_partial_peel_loop(), "not partial peeled"); 2478 head->mark_partial_peel_failed(); 2479 return false; 2480 } 2481 2482 // Step 3: clone loop, retarget control, and insert new phis 2483 2484 // Create new loop head for new phis and to hang 2485 // the nodes being moved (sinked) from the peel region. 2486 LoopNode* new_head = new (C, 3) LoopNode(last_peel, last_peel); 2487 _igvn.register_new_node_with_optimizer(new_head); 2488 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled"); 2489 first_not_peeled->set_req(0, new_head); 2490 set_loop(new_head, loop); 2491 loop->_body.push(new_head); 2492 not_peel.set(new_head->_idx); 2493 set_idom(new_head, last_peel, dom_depth(first_not_peeled)); 2494 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled)); 2495 2496 while (sink_list.size() > 0) { 2497 Node* n = sink_list.pop(); 2498 set_ctrl(n, new_head); 2499 } 2500 2501 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 2502 2503 clone_loop( loop, old_new, dd ); 2504 2505 const uint clone_exit_idx = 1; 2506 const uint orig_exit_idx = 2; 2507 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop"); 2508 2509 Node* head_clone = old_new[head->_idx]; 2510 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop(); 2511 Node* orig_tail_clone = head_clone->in(2); 2512 2513 // Add phi if "def" node is in peel set and "use" is not 2514 2515 for(i = 0; i < peel_list.size(); i++ ) { 2516 Node *def = peel_list.at(i); 2517 if (!def->is_CFG()) { 2518 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 2519 Node *use = def->fast_out(j); 2520 if (has_node(use) && use->in(0) != C->top() && 2521 (!peel.test(use->_idx) || 2522 (use->is_Phi() && use->in(0) == head)) ) { 2523 worklist.push(use); 2524 } 2525 } 2526 while( worklist.size() ) { 2527 Node *use = worklist.pop(); 2528 for (uint j = 1; j < use->req(); j++) { 2529 Node* n = use->in(j); 2530 if (n == def) { 2531 2532 // "def" is in peel set, "use" is not in peel set 2533 // or "use" is in the entry boundary (a phi) of the peel set 2534 2535 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use; 2536 2537 if ( loop->is_member(get_loop( use_c )) ) { 2538 // use is in loop 2539 if (old_new[use->_idx] != NULL) { // null for dead code 2540 Node* use_clone = old_new[use->_idx]; 2541 _igvn.hash_delete(use); 2542 use->set_req(j, C->top()); 2543 _igvn._worklist.push(use); 2544 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone ); 2545 } 2546 } else { 2547 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format"); 2548 // use is not in the loop, check if the live range includes the cut 2549 Node* lp_if = use_c->in(orig_exit_idx)->in(0); 2550 if (not_peel.test(lp_if->_idx)) { 2551 assert(j == orig_exit_idx, "use from original loop"); 2552 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone ); 2553 } 2554 } 2555 } 2556 } 2557 } 2558 } 2559 } 2560 2561 // Step 3b: retarget control 2562 2563 // Redirect control to the new loop head if a cloned node in 2564 // the not_peeled region has control that points into the peeled region. 2565 // This necessary because the cloned peeled region will be outside 2566 // the loop. 2567 // from to 2568 // cloned-peeled <---+ 2569 // new_head_clone: | <--+ 2570 // cloned-not_peeled in(0) in(0) 2571 // orig-peeled 2572 2573 for(i = 0; i < loop->_body.size(); i++ ) { 2574 Node *n = loop->_body.at(i); 2575 if (!n->is_CFG() && n->in(0) != NULL && 2576 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) { 2577 Node* n_clone = old_new[n->_idx]; 2578 _igvn.hash_delete(n_clone); 2579 n_clone->set_req(0, new_head_clone); 2580 _igvn._worklist.push(n_clone); 2581 } 2582 } 2583 2584 // Backedge of the surviving new_head (the clone) is original last_peel 2585 _igvn.hash_delete(new_head_clone); 2586 new_head_clone->set_req(LoopNode::LoopBackControl, last_peel); 2587 _igvn._worklist.push(new_head_clone); 2588 2589 // Cut first node in original not_peel set 2590 _igvn.hash_delete(new_head); 2591 new_head->set_req(LoopNode::EntryControl, C->top()); 2592 new_head->set_req(LoopNode::LoopBackControl, C->top()); 2593 _igvn._worklist.push(new_head); 2594 2595 // Copy head_clone back-branch info to original head 2596 // and remove original head's loop entry and 2597 // clone head's back-branch 2598 _igvn.hash_delete(head); 2599 _igvn.hash_delete(head_clone); 2600 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl)); 2601 head->set_req(LoopNode::LoopBackControl, C->top()); 2602 head_clone->set_req(LoopNode::LoopBackControl, C->top()); 2603 _igvn._worklist.push(head); 2604 _igvn._worklist.push(head_clone); 2605 2606 // Similarly modify the phis 2607 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) { 2608 Node* use = head->fast_out(k); 2609 if (use->is_Phi() && use->outcnt() > 0) { 2610 Node* use_clone = old_new[use->_idx]; 2611 _igvn.hash_delete(use); 2612 _igvn.hash_delete(use_clone); 2613 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl)); 2614 use->set_req(LoopNode::LoopBackControl, C->top()); 2615 use_clone->set_req(LoopNode::LoopBackControl, C->top()); 2616 _igvn._worklist.push(use); 2617 _igvn._worklist.push(use_clone); 2618 } 2619 } 2620 2621 // Step 4: update dominator tree and dominator depth 2622 2623 set_idom(head, orig_tail_clone, dd); 2624 recompute_dom_depth(); 2625 2626 // Inhibit more partial peeling on this loop 2627 new_head_clone->set_partial_peel_loop(); 2628 C->set_major_progress(); 2629 2630#if !defined(PRODUCT) 2631 if (TracePartialPeeling) { 2632 tty->print_cr("\nafter partial peel one iteration"); 2633 Node_List wl(area); 2634 Node* t = last_peel; 2635 while (true) { 2636 wl.push(t); 2637 if (t == head_clone) break; 2638 t = idom(t); 2639 } 2640 while (wl.size() > 0) { 2641 Node* tt = wl.pop(); 2642 if (tt == head) tty->print_cr("orig head"); 2643 else if (tt == new_head_clone) tty->print_cr("new head"); 2644 else if (tt == head_clone) tty->print_cr("clone head"); 2645 tt->dump(); 2646 } 2647 } 2648#endif 2649 return true; 2650} 2651 2652//------------------------------reorg_offsets---------------------------------- 2653// Reorganize offset computations to lower register pressure. Mostly 2654// prevent loop-fallout uses of the pre-incremented trip counter (which are 2655// then alive with the post-incremented trip counter forcing an extra 2656// register move) 2657void PhaseIdealLoop::reorg_offsets( IdealLoopTree *loop ) { 2658 2659 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 2660 CountedLoopEndNode *cle = cl->loopexit(); 2661 if( !cle ) return; // The occasional dead loop 2662 // Find loop exit control 2663 Node *exit = cle->proj_out(false); 2664 assert( exit->Opcode() == Op_IfFalse, "" ); 2665 2666 // Check for the special case of folks using the pre-incremented 2667 // trip-counter on the fall-out path (forces the pre-incremented 2668 // and post-incremented trip counter to be live at the same time). 2669 // Fix this by adjusting to use the post-increment trip counter. 2670 Node *phi = cl->phi(); 2671 if( !phi ) return; // Dead infinite loop 2672 2673 // Shape messed up, probably by iteration_split_impl 2674 if (phi->in(LoopNode::LoopBackControl) != cl->incr()) return; 2675 2676 bool progress = true; 2677 while (progress) { 2678 progress = false; 2679 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) { 2680 Node* use = phi->fast_out(i); // User of trip-counter 2681 if (!has_ctrl(use)) continue; 2682 Node *u_ctrl = get_ctrl(use); 2683 if( use->is_Phi() ) { 2684 u_ctrl = NULL; 2685 for( uint j = 1; j < use->req(); j++ ) 2686 if( use->in(j) == phi ) 2687 u_ctrl = dom_lca( u_ctrl, use->in(0)->in(j) ); 2688 } 2689 IdealLoopTree *u_loop = get_loop(u_ctrl); 2690 // Look for loop-invariant use 2691 if( u_loop == loop ) continue; 2692 if( loop->is_member( u_loop ) ) continue; 2693 // Check that use is live out the bottom. Assuming the trip-counter 2694 // update is right at the bottom, uses of of the loop middle are ok. 2695 if( dom_lca( exit, u_ctrl ) != exit ) continue; 2696 // protect against stride not being a constant 2697 if( !cle->stride_is_con() ) continue; 2698 // Hit! Refactor use to use the post-incremented tripcounter. 2699 // Compute a post-increment tripcounter. 2700 Node *opaq = new (C, 2) Opaque2Node( C, cle->incr() ); 2701 register_new_node( opaq, u_ctrl ); 2702 Node *neg_stride = _igvn.intcon(-cle->stride_con()); 2703 set_ctrl(neg_stride, C->root()); 2704 Node *post = new (C, 3) AddINode( opaq, neg_stride); 2705 register_new_node( post, u_ctrl ); 2706 _igvn.hash_delete(use); 2707 _igvn._worklist.push(use); 2708 for( uint j = 1; j < use->req(); j++ ) 2709 if( use->in(j) == phi ) 2710 use->set_req(j, post); 2711 // Since DU info changed, rerun loop 2712 progress = true; 2713 break; 2714 } 2715 } 2716 2717} 2718