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 "precompiled.hpp" 26 #include "memory/allocation.inline.hpp" 27 #include "opto/addnode.hpp" 28 #include "opto/connode.hpp" 29 #include "opto/divnode.hpp" 30 #include "opto/loopnode.hpp" 31 #include "opto/mulnode.hpp" 32 #include "opto/rootnode.hpp" 33 #include "opto/subnode.hpp" 34 35 //============================================================================= 36 //------------------------------split_thru_phi--------------------------------- 37 // Split Node 'n' through merge point if there is enough win. 38 Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) { 39 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) { 40 // ConvI2L may have type information on it which is unsafe to push up 41 // so disable this for now 42 return NULL; 43 } 44 int wins = 0; 45 assert( !n->is_CFG(), "" ); 46 assert( region->is_Region(), "" ); 47 48 const Type* type = n->bottom_type(); 49 const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr(); 50 Node *phi; 51 if( t_oop != NULL && t_oop->is_known_instance_field() ) { 52 int iid = t_oop->instance_id(); 53 int index = C->get_alias_index(t_oop); 54 int offset = t_oop->offset(); 55 phi = new (C,region->req()) PhiNode(region, type, NULL, iid, index, offset); 56 } else { 57 phi = PhiNode::make_blank(region, n); 58 } 59 uint old_unique = C->unique(); 60 for( uint i = 1; i < region->req(); i++ ) { 61 Node *x; 62 Node* the_clone = NULL; 63 if( region->in(i) == C->top() ) { 64 x = C->top(); // Dead path? Use a dead data op 65 } else { 66 x = n->clone(); // Else clone up the data op 67 the_clone = x; // Remember for possible deletion. 68 // Alter data node to use pre-phi inputs 69 if( n->in(0) == region ) 70 x->set_req( 0, region->in(i) ); 71 for( uint j = 1; j < n->req(); j++ ) { 72 Node *in = n->in(j); 73 if( in->is_Phi() && in->in(0) == region ) 74 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone 75 } 76 } 77 // Check for a 'win' on some paths 78 const Type *t = x->Value(&_igvn); 79 80 bool singleton = t->singleton(); 81 82 // A TOP singleton indicates that there are no possible values incoming 83 // along a particular edge. In most cases, this is OK, and the Phi will 84 // be eliminated later in an Ideal call. However, we can't allow this to 85 // happen if the singleton occurs on loop entry, as the elimination of 86 // the PhiNode may cause the resulting node to migrate back to a previous 87 // loop iteration. 88 if( singleton && t == Type::TOP ) { 89 // Is_Loop() == false does not confirm the absence of a loop (e.g., an 90 // irreducible loop may not be indicated by an affirmative is_Loop()); 91 // therefore, the only top we can split thru a phi is on a backedge of 92 // a loop. 93 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 94 } 95 96 if( singleton ) { 97 wins++; 98 x = ((PhaseGVN&)_igvn).makecon(t); 99 } else { 100 // We now call Identity to try to simplify the cloned node. 101 // Note that some Identity methods call phase->type(this). 102 // Make sure that the type array is big enough for 103 // our new node, even though we may throw the node away. 104 // (Note: This tweaking with igvn only works because x is a new node.) 105 _igvn.set_type(x, t); 106 // If x is a TypeNode, capture any more-precise type permanently into Node 107 // otherwise it will be not updated during igvn->transform since 108 // igvn->type(x) is set to x->Value() already. 109 x->raise_bottom_type(t); 110 Node *y = x->Identity(&_igvn); 111 if( y != x ) { 112 wins++; 113 x = y; 114 } else { 115 y = _igvn.hash_find(x); 116 if( y ) { 117 wins++; 118 x = y; 119 } else { 120 // Else x is a new node we are keeping 121 // We do not need register_new_node_with_optimizer 122 // because set_type has already been called. 123 _igvn._worklist.push(x); 124 } 125 } 126 } 127 if (x != the_clone && the_clone != NULL) 128 _igvn.remove_dead_node(the_clone); 129 phi->set_req( i, x ); 130 } 131 // Too few wins? 132 if( wins <= policy ) { 133 _igvn.remove_dead_node(phi); 134 return NULL; 135 } 136 137 // Record Phi 138 register_new_node( phi, region ); 139 140 for( uint i2 = 1; i2 < phi->req(); i2++ ) { 141 Node *x = phi->in(i2); 142 // If we commoned up the cloned 'x' with another existing Node, 143 // the existing Node picks up a new use. We need to make the 144 // existing Node occur higher up so it dominates its uses. 145 Node *old_ctrl; 146 IdealLoopTree *old_loop; 147 148 // The occasional new node 149 if( x->_idx >= old_unique ) { // Found a new, unplaced node? 150 old_ctrl = x->is_Con() ? C->root() : NULL; 151 old_loop = NULL; // Not in any prior loop 152 } else { 153 old_ctrl = x->is_Con() ? C->root() : get_ctrl(x); 154 old_loop = get_loop(old_ctrl); // Get prior loop 155 } 156 // New late point must dominate new use 157 Node *new_ctrl = dom_lca( old_ctrl, region->in(i2) ); 158 // Set new location 159 set_ctrl(x, new_ctrl); 160 IdealLoopTree *new_loop = get_loop( new_ctrl ); 161 // If changing loop bodies, see if we need to collect into new body 162 if( old_loop != new_loop ) { 163 if( old_loop && !old_loop->_child ) 164 old_loop->_body.yank(x); 165 if( !new_loop->_child ) 166 new_loop->_body.push(x); // Collect body info 167 } 168 } 169 170 return phi; 171 } 172 173 //------------------------------dominated_by------------------------------------ 174 // Replace the dominated test with an obvious true or false. Place it on the 175 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the 176 // live path up to the dominating control. 177 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff ) { 178 #ifndef PRODUCT 179 if( VerifyLoopOptimizations && PrintOpto ) tty->print_cr("dominating test"); 180 #endif 181 182 183 // prevdom is the dominating projection of the dominating test. 184 assert( iff->is_If(), "" ); 185 assert( iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added"); 186 int pop = prevdom->Opcode(); 187 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" ); 188 // 'con' is set to true or false to kill the dominated test. 189 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO); 190 set_ctrl(con, C->root()); // Constant gets a new use 191 // Hack the dominated test 192 _igvn.hash_delete(iff); 193 iff->set_req(1, con); 194 _igvn._worklist.push(iff); 195 196 // If I dont have a reachable TRUE and FALSE path following the IfNode then 197 // I can assume this path reaches an infinite loop. In this case it's not 198 // important to optimize the data Nodes - either the whole compilation will 199 // be tossed or this path (and all data Nodes) will go dead. 200 if( iff->outcnt() != 2 ) return; 201 202 // Make control-dependent data Nodes on the live path (path that will remain 203 // once the dominated IF is removed) become control-dependent on the 204 // dominating projection. 205 Node* dp = ((IfNode*)iff)->proj_out(pop == Op_IfTrue); 206 IdealLoopTree *old_loop = get_loop(dp); 207 208 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { 209 Node* cd = dp->fast_out(i); // Control-dependent node 210 if( cd->depends_only_on_test() ) { 211 assert( cd->in(0) == dp, "" ); 212 _igvn.hash_delete( cd ); 213 cd->set_req(0, prevdom); 214 set_early_ctrl( cd ); 215 _igvn._worklist.push(cd); 216 IdealLoopTree *new_loop = get_loop(get_ctrl(cd)); 217 if( old_loop != new_loop ) { 218 if( !old_loop->_child ) old_loop->_body.yank(cd); 219 if( !new_loop->_child ) new_loop->_body.push(cd); 220 } 221 --i; 222 --imax; 223 } 224 } 225 } 226 227 //------------------------------has_local_phi_input---------------------------- 228 // Return TRUE if 'n' has Phi inputs from its local block and no other 229 // block-local inputs (all non-local-phi inputs come from earlier blocks) 230 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) { 231 Node *n_ctrl = get_ctrl(n); 232 // See if some inputs come from a Phi in this block, or from before 233 // this block. 234 uint i; 235 for( i = 1; i < n->req(); i++ ) { 236 Node *phi = n->in(i); 237 if( phi->is_Phi() && phi->in(0) == n_ctrl ) 238 break; 239 } 240 if( i >= n->req() ) 241 return NULL; // No Phi inputs; nowhere to clone thru 242 243 // Check for inputs created between 'n' and the Phi input. These 244 // must split as well; they have already been given the chance 245 // (courtesy of a post-order visit) and since they did not we must 246 // recover the 'cost' of splitting them by being very profitable 247 // when splitting 'n'. Since this is unlikely we simply give up. 248 for( i = 1; i < n->req(); i++ ) { 249 Node *m = n->in(i); 250 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) { 251 // We allow the special case of AddP's with no local inputs. 252 // This allows us to split-up address expressions. 253 if (m->is_AddP() && 254 get_ctrl(m->in(2)) != n_ctrl && 255 get_ctrl(m->in(3)) != n_ctrl) { 256 // Move the AddP up to dominating point 257 set_ctrl_and_loop(m, find_non_split_ctrl(idom(n_ctrl))); 258 continue; 259 } 260 return NULL; 261 } 262 } 263 264 return n_ctrl; 265 } 266 267 //------------------------------remix_address_expressions---------------------- 268 // Rework addressing expressions to get the most loop-invariant stuff 269 // moved out. We'd like to do all associative operators, but it's especially 270 // important (common) to do address expressions. 271 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) { 272 if (!has_ctrl(n)) return NULL; 273 Node *n_ctrl = get_ctrl(n); 274 IdealLoopTree *n_loop = get_loop(n_ctrl); 275 276 // See if 'n' mixes loop-varying and loop-invariant inputs and 277 // itself is loop-varying. 278 279 // Only interested in binary ops (and AddP) 280 if( n->req() < 3 || n->req() > 4 ) return NULL; 281 282 Node *n1_ctrl = get_ctrl(n->in( 1)); 283 Node *n2_ctrl = get_ctrl(n->in( 2)); 284 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3)); 285 IdealLoopTree *n1_loop = get_loop( n1_ctrl ); 286 IdealLoopTree *n2_loop = get_loop( n2_ctrl ); 287 IdealLoopTree *n3_loop = get_loop( n3_ctrl ); 288 289 // Does one of my inputs spin in a tighter loop than self? 290 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) || 291 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) || 292 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) ) 293 return NULL; // Leave well enough alone 294 295 // Is at least one of my inputs loop-invariant? 296 if( n1_loop == n_loop && 297 n2_loop == n_loop && 298 n3_loop == n_loop ) 299 return NULL; // No loop-invariant inputs 300 301 302 int n_op = n->Opcode(); 303 304 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2). 305 if( n_op == Op_LShiftI ) { 306 // Scale is loop invariant 307 Node *scale = n->in(2); 308 Node *scale_ctrl = get_ctrl(scale); 309 IdealLoopTree *scale_loop = get_loop(scale_ctrl ); 310 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) ) 311 return NULL; 312 const TypeInt *scale_t = scale->bottom_type()->isa_int(); 313 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 ) 314 return NULL; // Dont bother with byte/short masking 315 // Add must vary with loop (else shift would be loop-invariant) 316 Node *add = n->in(1); 317 Node *add_ctrl = get_ctrl(add); 318 IdealLoopTree *add_loop = get_loop(add_ctrl); 319 //assert( n_loop == add_loop, "" ); 320 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops 321 322 // Convert I-V into I+ (0-V); same for V-I 323 if( add->Opcode() == Op_SubI && 324 _igvn.type( add->in(1) ) != TypeInt::ZERO ) { 325 Node *zero = _igvn.intcon(0); 326 set_ctrl(zero, C->root()); 327 Node *neg = new (C, 3) SubINode( _igvn.intcon(0), add->in(2) ); 328 register_new_node( neg, get_ctrl(add->in(2) ) ); 329 add = new (C, 3) AddINode( add->in(1), neg ); 330 register_new_node( add, add_ctrl ); 331 } 332 if( add->Opcode() != Op_AddI ) return NULL; 333 // See if one add input is loop invariant 334 Node *add_var = add->in(1); 335 Node *add_var_ctrl = get_ctrl(add_var); 336 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 337 Node *add_invar = add->in(2); 338 Node *add_invar_ctrl = get_ctrl(add_invar); 339 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl ); 340 if( add_var_loop == n_loop ) { 341 } else if( add_invar_loop == n_loop ) { 342 // Swap to find the invariant part 343 add_invar = add_var; 344 add_invar_ctrl = add_var_ctrl; 345 add_invar_loop = add_var_loop; 346 add_var = add->in(2); 347 Node *add_var_ctrl = get_ctrl(add_var); 348 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 349 } else // Else neither input is loop invariant 350 return NULL; 351 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) ) 352 return NULL; // No invariant part of the add? 353 354 // Yes! Reshape address expression! 355 Node *inv_scale = new (C, 3) LShiftINode( add_invar, scale ); 356 Node *inv_scale_ctrl = 357 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ? 358 add_invar_ctrl : scale_ctrl; 359 register_new_node( inv_scale, inv_scale_ctrl ); 360 Node *var_scale = new (C, 3) LShiftINode( add_var, scale ); 361 register_new_node( var_scale, n_ctrl ); 362 Node *var_add = new (C, 3) AddINode( var_scale, inv_scale ); 363 register_new_node( var_add, n_ctrl ); 364 _igvn.replace_node( n, var_add ); 365 return var_add; 366 } 367 368 // Replace (I+V) with (V+I) 369 if( n_op == Op_AddI || 370 n_op == Op_AddL || 371 n_op == Op_AddF || 372 n_op == Op_AddD || 373 n_op == Op_MulI || 374 n_op == Op_MulL || 375 n_op == Op_MulF || 376 n_op == Op_MulD ) { 377 if( n2_loop == n_loop ) { 378 assert( n1_loop != n_loop, "" ); 379 n->swap_edges(1, 2); 380 } 381 } 382 383 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V), 384 // but not if I2 is a constant. 385 if( n_op == Op_AddP ) { 386 if( n2_loop == n_loop && n3_loop != n_loop ) { 387 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) { 388 Node *n22_ctrl = get_ctrl(n->in(2)->in(2)); 389 Node *n23_ctrl = get_ctrl(n->in(2)->in(3)); 390 IdealLoopTree *n22loop = get_loop( n22_ctrl ); 391 IdealLoopTree *n23_loop = get_loop( n23_ctrl ); 392 if( n22loop != n_loop && n22loop->is_member(n_loop) && 393 n23_loop == n_loop ) { 394 Node *add1 = new (C, 4) AddPNode( n->in(1), n->in(2)->in(2), n->in(3) ); 395 // Stuff new AddP in the loop preheader 396 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) ); 397 Node *add2 = new (C, 4) AddPNode( n->in(1), add1, n->in(2)->in(3) ); 398 register_new_node( add2, n_ctrl ); 399 _igvn.replace_node( n, add2 ); 400 return add2; 401 } 402 } 403 } 404 405 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V) 406 if( n2_loop != n_loop && n3_loop == n_loop ) { 407 if( n->in(3)->Opcode() == Op_AddI ) { 408 Node *V = n->in(3)->in(1); 409 Node *I = n->in(3)->in(2); 410 if( is_member(n_loop,get_ctrl(V)) ) { 411 } else { 412 Node *tmp = V; V = I; I = tmp; 413 } 414 if( !is_member(n_loop,get_ctrl(I)) ) { 415 Node *add1 = new (C, 4) AddPNode( n->in(1), n->in(2), I ); 416 // Stuff new AddP in the loop preheader 417 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) ); 418 Node *add2 = new (C, 4) AddPNode( n->in(1), add1, V ); 419 register_new_node( add2, n_ctrl ); 420 _igvn.replace_node( n, add2 ); 421 return add2; 422 } 423 } 424 } 425 } 426 427 return NULL; 428 } 429 430 //------------------------------conditional_move------------------------------- 431 // Attempt to replace a Phi with a conditional move. We have some pretty 432 // strict profitability requirements. All Phis at the merge point must 433 // be converted, so we can remove the control flow. We need to limit the 434 // number of c-moves to a small handful. All code that was in the side-arms 435 // of the CFG diamond is now speculatively executed. This code has to be 436 // "cheap enough". We are pretty much limited to CFG diamonds that merge 437 // 1 or 2 items with a total of 1 or 2 ops executed speculatively. 438 Node *PhaseIdealLoop::conditional_move( Node *region ) { 439 440 assert( region->is_Region(), "sanity check" ); 441 if( region->req() != 3 ) return NULL; 442 443 // Check for CFG diamond 444 Node *lp = region->in(1); 445 Node *rp = region->in(2); 446 if( !lp || !rp ) return NULL; 447 Node *lp_c = lp->in(0); 448 if( lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If() ) return NULL; 449 IfNode *iff = lp_c->as_If(); 450 451 // Check for highly predictable branch. No point in CMOV'ing if 452 // we are going to predict accurately all the time. 453 // %%% This hides patterns produced by utility methods like Math.min. 454 if( iff->_prob < PROB_UNLIKELY_MAG(3) || 455 iff->_prob > PROB_LIKELY_MAG(3) ) 456 return NULL; 457 458 // Check for ops pinned in an arm of the diamond. 459 // Can't remove the control flow in this case 460 if( lp->outcnt() > 1 ) return NULL; 461 if( rp->outcnt() > 1 ) return NULL; 462 463 // Check profitability 464 int cost = 0; 465 int phis = 0; 466 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 467 Node *out = region->fast_out(i); 468 if( !out->is_Phi() ) continue; // Ignore other control edges, etc 469 phis++; 470 PhiNode* phi = out->as_Phi(); 471 switch (phi->type()->basic_type()) { 472 case T_LONG: 473 cost++; // Probably encodes as 2 CMOV's 474 case T_INT: // These all CMOV fine 475 case T_FLOAT: 476 case T_DOUBLE: 477 case T_ADDRESS: // (RawPtr) 478 cost++; 479 break; 480 case T_NARROWOOP: // Fall through 481 case T_OBJECT: { // Base oops are OK, but not derived oops 482 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr(); 483 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a 484 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus 485 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we 486 // have a Phi for the base here that we convert to a CMOVE all is well 487 // and good. But if the base is dead, we'll not make a CMOVE. Later 488 // the allocator will have to produce a base by creating a CMOVE of the 489 // relevant bases. This puts the allocator in the business of 490 // manufacturing expensive instructions, generally a bad plan. 491 // Just Say No to Conditionally-Moved Derived Pointers. 492 if( tp && tp->offset() != 0 ) 493 return NULL; 494 cost++; 495 break; 496 } 497 default: 498 return NULL; // In particular, can't do memory or I/O 499 } 500 // Add in cost any speculative ops 501 for( uint j = 1; j < region->req(); j++ ) { 502 Node *proj = region->in(j); 503 Node *inp = phi->in(j); 504 if (get_ctrl(inp) == proj) { // Found local op 505 cost++; 506 // Check for a chain of dependent ops; these will all become 507 // speculative in a CMOV. 508 for( uint k = 1; k < inp->req(); k++ ) 509 if (get_ctrl(inp->in(k)) == proj) 510 return NULL; // Too much speculative goo 511 } 512 } 513 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode. 514 // This will likely Split-If, a higher-payoff operation. 515 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) { 516 Node* use = phi->fast_out(k); 517 if( use->is_Cmp() || use->is_DecodeN() || use->is_EncodeP() ) 518 return NULL; 519 } 520 } 521 if( cost >= ConditionalMoveLimit ) return NULL; // Too much goo 522 Node* bol = iff->in(1); 523 assert( bol->Opcode() == Op_Bool, "" ); 524 int cmp_op = bol->in(1)->Opcode(); 525 // It is expensive to generate flags from a float compare. 526 // Avoid duplicated float compare. 527 if( phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL; 528 529 // -------------- 530 // Now replace all Phis with CMOV's 531 Node *cmov_ctrl = iff->in(0); 532 uint flip = (lp->Opcode() == Op_IfTrue); 533 while( 1 ) { 534 PhiNode* phi = NULL; 535 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 536 Node *out = region->fast_out(i); 537 if (out->is_Phi()) { 538 phi = out->as_Phi(); 539 break; 540 } 541 } 542 if (phi == NULL) break; 543 #ifndef PRODUCT 544 if( PrintOpto && VerifyLoopOptimizations ) tty->print_cr("CMOV"); 545 #endif 546 // Move speculative ops 547 for( uint j = 1; j < region->req(); j++ ) { 548 Node *proj = region->in(j); 549 Node *inp = phi->in(j); 550 if (get_ctrl(inp) == proj) { // Found local op 551 #ifndef PRODUCT 552 if( PrintOpto && VerifyLoopOptimizations ) { 553 tty->print(" speculate: "); 554 inp->dump(); 555 } 556 #endif 557 set_ctrl(inp, cmov_ctrl); 558 } 559 } 560 Node *cmov = CMoveNode::make( C, cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi) ); 561 register_new_node( cmov, cmov_ctrl ); 562 _igvn.replace_node( phi, cmov ); 563 #ifndef PRODUCT 564 if( VerifyLoopOptimizations ) verify(); 565 #endif 566 } 567 568 // The useless CFG diamond will fold up later; see the optimization in 569 // RegionNode::Ideal. 570 _igvn._worklist.push(region); 571 572 return iff->in(1); 573 } 574 575 //------------------------------split_if_with_blocks_pre----------------------- 576 // Do the real work in a non-recursive function. Data nodes want to be 577 // cloned in the pre-order so they can feed each other nicely. 578 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) { 579 // Cloning these guys is unlikely to win 580 int n_op = n->Opcode(); 581 if( n_op == Op_MergeMem ) return n; 582 if( n->is_Proj() ) return n; 583 // Do not clone-up CmpFXXX variations, as these are always 584 // followed by a CmpI 585 if( n->is_Cmp() ) return n; 586 // Attempt to use a conditional move instead of a phi/branch 587 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) { 588 Node *cmov = conditional_move( n ); 589 if( cmov ) return cmov; 590 } 591 if( n->is_CFG() || n->is_LoadStore() ) 592 return n; 593 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd 594 n_op == Op_Opaque2 ) { 595 if( !C->major_progress() ) // If chance of no more loop opts... 596 _igvn._worklist.push(n); // maybe we'll remove them 597 return n; 598 } 599 600 if( n->is_Con() ) return n; // No cloning for Con nodes 601 602 Node *n_ctrl = get_ctrl(n); 603 if( !n_ctrl ) return n; // Dead node 604 605 // Attempt to remix address expressions for loop invariants 606 Node *m = remix_address_expressions( n ); 607 if( m ) return m; 608 609 // Determine if the Node has inputs from some local Phi. 610 // Returns the block to clone thru. 611 Node *n_blk = has_local_phi_input( n ); 612 if( !n_blk ) return n; 613 // Do not clone the trip counter through on a CountedLoop 614 // (messes up the canonical shape). 615 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n; 616 617 // Check for having no control input; not pinned. Allow 618 // dominating control. 619 if( n->in(0) ) { 620 Node *dom = idom(n_blk); 621 if( dom_lca( n->in(0), dom ) != n->in(0) ) 622 return n; 623 } 624 // Policy: when is it profitable. You must get more wins than 625 // policy before it is considered profitable. Policy is usually 0, 626 // so 1 win is considered profitable. Big merges will require big 627 // cloning, so get a larger policy. 628 int policy = n_blk->req() >> 2; 629 630 // If the loop is a candidate for range check elimination, 631 // delay splitting through it's phi until a later loop optimization 632 if (n_blk->is_CountedLoop()) { 633 IdealLoopTree *lp = get_loop(n_blk); 634 if (lp && lp->_rce_candidate) { 635 return n; 636 } 637 } 638 639 // Use same limit as split_if_with_blocks_post 640 if( C->unique() > 35000 ) return n; // Method too big 641 642 // Split 'n' through the merge point if it is profitable 643 Node *phi = split_thru_phi( n, n_blk, policy ); 644 if( !phi ) return n; 645 646 // Found a Phi to split thru! 647 // Replace 'n' with the new phi 648 _igvn.replace_node( n, phi ); 649 // Moved a load around the loop, 'en-registering' something. 650 if( n_blk->Opcode() == Op_Loop && n->is_Load() && 651 !phi->in(LoopNode::LoopBackControl)->is_Load() ) 652 C->set_major_progress(); 653 654 return phi; 655 } 656 657 static bool merge_point_too_heavy(Compile* C, Node* region) { 658 // Bail out if the region and its phis have too many users. 659 int weight = 0; 660 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 661 weight += region->fast_out(i)->outcnt(); 662 } 663 int nodes_left = MaxNodeLimit - C->unique(); 664 if (weight * 8 > nodes_left) { 665 #ifndef PRODUCT 666 if (PrintOpto) 667 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight); 668 #endif 669 return true; 670 } else { 671 return false; 672 } 673 } 674 675 static bool merge_point_safe(Node* region) { 676 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode 677 // having a PhiNode input. This sidesteps the dangerous case where the split 678 // ConvI2LNode may become TOP if the input Value() does not 679 // overlap the ConvI2L range, leaving a node which may not dominate its 680 // uses. 681 // A better fix for this problem can be found in the BugTraq entry, but 682 // expediency for Mantis demands this hack. 683 // 6855164: If the merge point has a FastLockNode with a PhiNode input, we stop 684 // split_if_with_blocks from splitting a block because we could not move around 685 // the FastLockNode. 686 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 687 Node* n = region->fast_out(i); 688 if (n->is_Phi()) { 689 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 690 Node* m = n->fast_out(j); 691 if (m->is_FastLock()) 692 return false; 693 #ifdef _LP64 694 if (m->Opcode() == Op_ConvI2L) 695 return false; 696 #endif 697 } 698 } 699 } 700 return true; 701 } 702 703 704 //------------------------------place_near_use--------------------------------- 705 // Place some computation next to use but not inside inner loops. 706 // For inner loop uses move it to the preheader area. 707 Node *PhaseIdealLoop::place_near_use( Node *useblock ) const { 708 IdealLoopTree *u_loop = get_loop( useblock ); 709 return (u_loop->_irreducible || u_loop->_child) 710 ? useblock 711 : u_loop->_head->in(LoopNode::EntryControl); 712 } 713 714 715 //------------------------------split_if_with_blocks_post---------------------- 716 // Do the real work in a non-recursive function. CFG hackery wants to be 717 // in the post-order, so it can dirty the I-DOM info and not use the dirtied 718 // info. 719 void PhaseIdealLoop::split_if_with_blocks_post( Node *n ) { 720 721 // Cloning Cmp through Phi's involves the split-if transform. 722 // FastLock is not used by an If 723 if( n->is_Cmp() && !n->is_FastLock() ) { 724 if( C->unique() > 35000 ) return; // Method too big 725 726 // Do not do 'split-if' if irreducible loops are present. 727 if( _has_irreducible_loops ) 728 return; 729 730 Node *n_ctrl = get_ctrl(n); 731 // Determine if the Node has inputs from some local Phi. 732 // Returns the block to clone thru. 733 Node *n_blk = has_local_phi_input( n ); 734 if( n_blk != n_ctrl ) return; 735 736 if( merge_point_too_heavy(C, n_ctrl) ) 737 return; 738 739 if( n->outcnt() != 1 ) return; // Multiple bool's from 1 compare? 740 Node *bol = n->unique_out(); 741 assert( bol->is_Bool(), "expect a bool here" ); 742 if( bol->outcnt() != 1 ) return;// Multiple branches from 1 compare? 743 Node *iff = bol->unique_out(); 744 745 // Check some safety conditions 746 if( iff->is_If() ) { // Classic split-if? 747 if( iff->in(0) != n_ctrl ) return; // Compare must be in same blk as if 748 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE 749 if( get_ctrl(iff->in(2)) == n_ctrl || 750 get_ctrl(iff->in(3)) == n_ctrl ) 751 return; // Inputs not yet split-up 752 if ( get_loop(n_ctrl) != get_loop(get_ctrl(iff)) ) { 753 return; // Loop-invar test gates loop-varying CMOVE 754 } 755 } else { 756 return; // some other kind of node, such as an Allocate 757 } 758 759 // Do not do 'split-if' if some paths are dead. First do dead code 760 // elimination and then see if its still profitable. 761 for( uint i = 1; i < n_ctrl->req(); i++ ) 762 if( n_ctrl->in(i) == C->top() ) 763 return; 764 765 // When is split-if profitable? Every 'win' on means some control flow 766 // goes dead, so it's almost always a win. 767 int policy = 0; 768 // If trying to do a 'Split-If' at the loop head, it is only 769 // profitable if the cmp folds up on BOTH paths. Otherwise we 770 // risk peeling a loop forever. 771 772 // CNC - Disabled for now. Requires careful handling of loop 773 // body selection for the cloned code. Also, make sure we check 774 // for any input path not being in the same loop as n_ctrl. For 775 // irreducible loops we cannot check for 'n_ctrl->is_Loop()' 776 // because the alternative loop entry points won't be converted 777 // into LoopNodes. 778 IdealLoopTree *n_loop = get_loop(n_ctrl); 779 for( uint j = 1; j < n_ctrl->req(); j++ ) 780 if( get_loop(n_ctrl->in(j)) != n_loop ) 781 return; 782 783 // Check for safety of the merge point. 784 if( !merge_point_safe(n_ctrl) ) { 785 return; 786 } 787 788 // Split compare 'n' through the merge point if it is profitable 789 Node *phi = split_thru_phi( n, n_ctrl, policy ); 790 if( !phi ) return; 791 792 // Found a Phi to split thru! 793 // Replace 'n' with the new phi 794 _igvn.replace_node( n, phi ); 795 796 // Now split the bool up thru the phi 797 Node *bolphi = split_thru_phi( bol, n_ctrl, -1 ); 798 _igvn.replace_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.replace_node( iff, cmovphi ); 807 return; 808 } 809 810 // Now split the IF 811 do_split_if( iff ); 812 return; 813 } 814 815 // Check for an IF ready to split; one that has its 816 // condition codes input coming from a Phi at the block start. 817 int n_op = n->Opcode(); 818 819 // Check for an IF being dominated by another IF same test 820 if( n_op == Op_If ) { 821 Node *bol = n->in(1); 822 uint max = bol->outcnt(); 823 // Check for same test used more than once? 824 if( n_op == Op_If && max > 1 && bol->is_Bool() ) { 825 // Search up IDOMs to see if this IF is dominated. 826 Node *cutoff = get_ctrl(bol); 827 828 // Now search up IDOMs till cutoff, looking for a dominating test 829 Node *prevdom = n; 830 Node *dom = idom(prevdom); 831 while( dom != cutoff ) { 832 if( dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom ) { 833 // Replace the dominated test with an obvious true or false. 834 // Place it on the IGVN worklist for later cleanup. 835 C->set_major_progress(); 836 dominated_by( prevdom, n ); 837 #ifndef PRODUCT 838 if( VerifyLoopOptimizations ) verify(); 839 #endif 840 return; 841 } 842 prevdom = dom; 843 dom = idom(prevdom); 844 } 845 } 846 } 847 848 // See if a shared loop-varying computation has no loop-varying uses. 849 // Happens if something is only used for JVM state in uncommon trap exits, 850 // like various versions of induction variable+offset. Clone the 851 // computation per usage to allow it to sink out of the loop. 852 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about) 853 Node *n_ctrl = get_ctrl(n); 854 IdealLoopTree *n_loop = get_loop(n_ctrl); 855 if( n_loop != _ltree_root ) { 856 DUIterator_Fast imax, i = n->fast_outs(imax); 857 for (; i < imax; i++) { 858 Node* u = n->fast_out(i); 859 if( !has_ctrl(u) ) break; // Found control user 860 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 861 if( u_loop == n_loop ) break; // Found loop-varying use 862 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop 863 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003 864 } 865 bool did_break = (i < imax); // Did we break out of the previous loop? 866 if (!did_break && n->outcnt() > 1) { // All uses in outer loops! 867 Node *late_load_ctrl; 868 if (n->is_Load()) { 869 // If n is a load, get and save the result from get_late_ctrl(), 870 // to be later used in calculating the control for n's clones. 871 clear_dom_lca_tags(); 872 late_load_ctrl = get_late_ctrl(n, n_ctrl); 873 } 874 // If n is a load, and the late control is the same as the current 875 // control, then the cloning of n is a pointless exercise, because 876 // GVN will ensure that we end up where we started. 877 if (!n->is_Load() || late_load_ctrl != n_ctrl) { 878 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) { 879 Node *u = n->last_out(j); // Clone private computation per use 880 _igvn.hash_delete(u); 881 _igvn._worklist.push(u); 882 Node *x = n->clone(); // Clone computation 883 Node *x_ctrl = NULL; 884 if( u->is_Phi() ) { 885 // Replace all uses of normal nodes. Replace Phi uses 886 // individually, so the separate Nodes can sink down 887 // different paths. 888 uint k = 1; 889 while( u->in(k) != n ) k++; 890 u->set_req( k, x ); 891 // x goes next to Phi input path 892 x_ctrl = u->in(0)->in(k); 893 --j; 894 } else { // Normal use 895 // Replace all uses 896 for( uint k = 0; k < u->req(); k++ ) { 897 if( u->in(k) == n ) { 898 u->set_req( k, x ); 899 --j; 900 } 901 } 902 x_ctrl = get_ctrl(u); 903 } 904 905 // Find control for 'x' next to use but not inside inner loops. 906 // For inner loop uses get the preheader area. 907 x_ctrl = place_near_use(x_ctrl); 908 909 if (n->is_Load()) { 910 // For loads, add a control edge to a CFG node outside of the loop 911 // to force them to not combine and return back inside the loop 912 // during GVN optimization (4641526). 913 // 914 // Because we are setting the actual control input, factor in 915 // the result from get_late_ctrl() so we respect any 916 // anti-dependences. (6233005). 917 x_ctrl = dom_lca(late_load_ctrl, x_ctrl); 918 919 // Don't allow the control input to be a CFG splitting node. 920 // Such nodes should only have ProjNodes as outs, e.g. IfNode 921 // should only have IfTrueNode and IfFalseNode (4985384). 922 x_ctrl = find_non_split_ctrl(x_ctrl); 923 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone"); 924 925 x->set_req(0, x_ctrl); 926 } 927 register_new_node(x, x_ctrl); 928 929 // Some institutional knowledge is needed here: 'x' is 930 // yanked because if the optimizer runs GVN on it all the 931 // cloned x's will common up and undo this optimization and 932 // be forced back in the loop. This is annoying because it 933 // makes +VerifyOpto report false-positives on progress. I 934 // tried setting control edges on the x's to force them to 935 // not combine, but the matching gets worried when it tries 936 // to fold a StoreP and an AddP together (as part of an 937 // address expression) and the AddP and StoreP have 938 // different controls. 939 if( !x->is_Load() && !x->is_DecodeN() ) _igvn._worklist.yank(x); 940 } 941 _igvn.remove_dead_node(n); 942 } 943 } 944 } 945 } 946 947 // Check for Opaque2's who's loop has disappeared - who's input is in the 948 // same loop nest as their output. Remove 'em, they are no longer useful. 949 if( n_op == Op_Opaque2 && 950 n->in(1) != NULL && 951 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) { 952 _igvn.replace_node( n, n->in(1) ); 953 } 954 } 955 956 //------------------------------split_if_with_blocks--------------------------- 957 // Check for aggressive application of 'split-if' optimization, 958 // using basic block level info. 959 void PhaseIdealLoop::split_if_with_blocks( VectorSet &visited, Node_Stack &nstack ) { 960 Node *n = C->root(); 961 visited.set(n->_idx); // first, mark node as visited 962 // Do pre-visit work for root 963 n = split_if_with_blocks_pre( n ); 964 uint cnt = n->outcnt(); 965 uint i = 0; 966 while (true) { 967 // Visit all children 968 if (i < cnt) { 969 Node* use = n->raw_out(i); 970 ++i; 971 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) { 972 // Now do pre-visit work for this use 973 use = split_if_with_blocks_pre( use ); 974 nstack.push(n, i); // Save parent and next use's index. 975 n = use; // Process all children of current use. 976 cnt = use->outcnt(); 977 i = 0; 978 } 979 } 980 else { 981 // All of n's children have been processed, complete post-processing. 982 if (cnt != 0 && !n->is_Con()) { 983 assert(has_node(n), "no dead nodes"); 984 split_if_with_blocks_post( n ); 985 } 986 if (nstack.is_empty()) { 987 // Finished all nodes on stack. 988 break; 989 } 990 // Get saved parent node and next use's index. Visit the rest of uses. 991 n = nstack.node(); 992 cnt = n->outcnt(); 993 i = nstack.index(); 994 nstack.pop(); 995 } 996 } 997 } 998 999 1000 //============================================================================= 1001 // 1002 // C L O N E A L O O P B O D Y 1003 // 1004 1005 //------------------------------clone_iff-------------------------------------- 1006 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1007 // "Nearly" because all Nodes have been cloned from the original in the loop, 1008 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1009 // through the Phi recursively, and return a Bool. 1010 BoolNode *PhaseIdealLoop::clone_iff( PhiNode *phi, IdealLoopTree *loop ) { 1011 1012 // Convert this Phi into a Phi merging Bools 1013 uint i; 1014 for( i = 1; i < phi->req(); i++ ) { 1015 Node *b = phi->in(i); 1016 if( b->is_Phi() ) { 1017 _igvn.hash_delete(phi); 1018 _igvn._worklist.push(phi); 1019 phi->set_req(i, clone_iff( b->as_Phi(), loop )); 1020 } else { 1021 assert( b->is_Bool(), "" ); 1022 } 1023 } 1024 1025 Node *sample_bool = phi->in(1); 1026 Node *sample_cmp = sample_bool->in(1); 1027 1028 // Make Phis to merge the Cmp's inputs. 1029 int size = phi->in(0)->req(); 1030 PhiNode *phi1 = new (C, size) PhiNode( phi->in(0), Type::TOP ); 1031 PhiNode *phi2 = new (C, size) PhiNode( phi->in(0), Type::TOP ); 1032 for( i = 1; i < phi->req(); i++ ) { 1033 Node *n1 = phi->in(i)->in(1)->in(1); 1034 Node *n2 = phi->in(i)->in(1)->in(2); 1035 phi1->set_req( i, n1 ); 1036 phi2->set_req( i, n2 ); 1037 phi1->set_type( phi1->type()->meet(n1->bottom_type()) ); 1038 phi2->set_type( phi2->type()->meet(n2->bottom_type()) ); 1039 } 1040 // See if these Phis have been made before. 1041 // Register with optimizer 1042 Node *hit1 = _igvn.hash_find_insert(phi1); 1043 if( hit1 ) { // Hit, toss just made Phi 1044 _igvn.remove_dead_node(phi1); // Remove new phi 1045 assert( hit1->is_Phi(), "" ); 1046 phi1 = (PhiNode*)hit1; // Use existing phi 1047 } else { // Miss 1048 _igvn.register_new_node_with_optimizer(phi1); 1049 } 1050 Node *hit2 = _igvn.hash_find_insert(phi2); 1051 if( hit2 ) { // Hit, toss just made Phi 1052 _igvn.remove_dead_node(phi2); // Remove new phi 1053 assert( hit2->is_Phi(), "" ); 1054 phi2 = (PhiNode*)hit2; // Use existing phi 1055 } else { // Miss 1056 _igvn.register_new_node_with_optimizer(phi2); 1057 } 1058 // Register Phis with loop/block info 1059 set_ctrl(phi1, phi->in(0)); 1060 set_ctrl(phi2, phi->in(0)); 1061 // Make a new Cmp 1062 Node *cmp = sample_cmp->clone(); 1063 cmp->set_req( 1, phi1 ); 1064 cmp->set_req( 2, phi2 ); 1065 _igvn.register_new_node_with_optimizer(cmp); 1066 set_ctrl(cmp, phi->in(0)); 1067 1068 // Make a new Bool 1069 Node *b = sample_bool->clone(); 1070 b->set_req(1,cmp); 1071 _igvn.register_new_node_with_optimizer(b); 1072 set_ctrl(b, phi->in(0)); 1073 1074 assert( b->is_Bool(), "" ); 1075 return (BoolNode*)b; 1076 } 1077 1078 //------------------------------clone_bool------------------------------------- 1079 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1080 // "Nearly" because all Nodes have been cloned from the original in the loop, 1081 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1082 // through the Phi recursively, and return a Bool. 1083 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) { 1084 uint i; 1085 // Convert this Phi into a Phi merging Bools 1086 for( i = 1; i < phi->req(); i++ ) { 1087 Node *b = phi->in(i); 1088 if( b->is_Phi() ) { 1089 _igvn.hash_delete(phi); 1090 _igvn._worklist.push(phi); 1091 phi->set_req(i, clone_bool( b->as_Phi(), loop )); 1092 } else { 1093 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" ); 1094 } 1095 } 1096 1097 Node *sample_cmp = phi->in(1); 1098 1099 // Make Phis to merge the Cmp's inputs. 1100 int size = phi->in(0)->req(); 1101 PhiNode *phi1 = new (C, size) PhiNode( phi->in(0), Type::TOP ); 1102 PhiNode *phi2 = new (C, size) PhiNode( phi->in(0), Type::TOP ); 1103 for( uint j = 1; j < phi->req(); j++ ) { 1104 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP 1105 Node *n1, *n2; 1106 if( cmp_top->is_Cmp() ) { 1107 n1 = cmp_top->in(1); 1108 n2 = cmp_top->in(2); 1109 } else { 1110 n1 = n2 = cmp_top; 1111 } 1112 phi1->set_req( j, n1 ); 1113 phi2->set_req( j, n2 ); 1114 phi1->set_type( phi1->type()->meet(n1->bottom_type()) ); 1115 phi2->set_type( phi2->type()->meet(n2->bottom_type()) ); 1116 } 1117 1118 // See if these Phis have been made before. 1119 // Register with optimizer 1120 Node *hit1 = _igvn.hash_find_insert(phi1); 1121 if( hit1 ) { // Hit, toss just made Phi 1122 _igvn.remove_dead_node(phi1); // Remove new phi 1123 assert( hit1->is_Phi(), "" ); 1124 phi1 = (PhiNode*)hit1; // Use existing phi 1125 } else { // Miss 1126 _igvn.register_new_node_with_optimizer(phi1); 1127 } 1128 Node *hit2 = _igvn.hash_find_insert(phi2); 1129 if( hit2 ) { // Hit, toss just made Phi 1130 _igvn.remove_dead_node(phi2); // Remove new phi 1131 assert( hit2->is_Phi(), "" ); 1132 phi2 = (PhiNode*)hit2; // Use existing phi 1133 } else { // Miss 1134 _igvn.register_new_node_with_optimizer(phi2); 1135 } 1136 // Register Phis with loop/block info 1137 set_ctrl(phi1, phi->in(0)); 1138 set_ctrl(phi2, phi->in(0)); 1139 // Make a new Cmp 1140 Node *cmp = sample_cmp->clone(); 1141 cmp->set_req( 1, phi1 ); 1142 cmp->set_req( 2, phi2 ); 1143 _igvn.register_new_node_with_optimizer(cmp); 1144 set_ctrl(cmp, phi->in(0)); 1145 1146 assert( cmp->is_Cmp(), "" ); 1147 return (CmpNode*)cmp; 1148 } 1149 1150 //------------------------------sink_use--------------------------------------- 1151 // If 'use' was in the loop-exit block, it now needs to be sunk 1152 // below the post-loop merge point. 1153 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) { 1154 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) { 1155 set_ctrl(use, post_loop); 1156 for (DUIterator j = use->outs(); use->has_out(j); j++) 1157 sink_use(use->out(j), post_loop); 1158 } 1159 } 1160 1161 //------------------------------clone_loop------------------------------------- 1162 // 1163 // C L O N E A L O O P B O D Y 1164 // 1165 // This is the basic building block of the loop optimizations. It clones an 1166 // entire loop body. It makes an old_new loop body mapping; with this mapping 1167 // you can find the new-loop equivalent to an old-loop node. All new-loop 1168 // nodes are exactly equal to their old-loop counterparts, all edges are the 1169 // same. All exits from the old-loop now have a RegionNode that merges the 1170 // equivalent new-loop path. This is true even for the normal "loop-exit" 1171 // condition. All uses of loop-invariant old-loop values now come from (one 1172 // or more) Phis that merge their new-loop equivalents. 1173 // 1174 // This operation leaves the graph in an illegal state: there are two valid 1175 // control edges coming from the loop pre-header to both loop bodies. I'll 1176 // definitely have to hack the graph after running this transform. 1177 // 1178 // From this building block I will further edit edges to perform loop peeling 1179 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc. 1180 // 1181 // Parameter side_by_size_idom: 1182 // When side_by_size_idom is NULL, the dominator tree is constructed for 1183 // the clone loop to dominate the original. Used in construction of 1184 // pre-main-post loop sequence. 1185 // When nonnull, the clone and original are side-by-side, both are 1186 // dominated by the side_by_side_idom node. Used in construction of 1187 // unswitched loops. 1188 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd, 1189 Node* side_by_side_idom) { 1190 1191 // Step 1: Clone the loop body. Make the old->new mapping. 1192 uint i; 1193 for( i = 0; i < loop->_body.size(); i++ ) { 1194 Node *old = loop->_body.at(i); 1195 Node *nnn = old->clone(); 1196 old_new.map( old->_idx, nnn ); 1197 _igvn.register_new_node_with_optimizer(nnn); 1198 } 1199 1200 1201 // Step 2: Fix the edges in the new body. If the old input is outside the 1202 // loop use it. If the old input is INside the loop, use the corresponding 1203 // new node instead. 1204 for( i = 0; i < loop->_body.size(); i++ ) { 1205 Node *old = loop->_body.at(i); 1206 Node *nnn = old_new[old->_idx]; 1207 // Fix CFG/Loop controlling the new node 1208 if (has_ctrl(old)) { 1209 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]); 1210 } else { 1211 set_loop(nnn, loop->_parent); 1212 if (old->outcnt() > 0) { 1213 set_idom( nnn, old_new[idom(old)->_idx], dd ); 1214 } 1215 } 1216 // Correct edges to the new node 1217 for( uint j = 0; j < nnn->req(); j++ ) { 1218 Node *n = nnn->in(j); 1219 if( n ) { 1220 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n ); 1221 if( loop->is_member( old_in_loop ) ) 1222 nnn->set_req(j, old_new[n->_idx]); 1223 } 1224 } 1225 _igvn.hash_find_insert(nnn); 1226 } 1227 Node *newhead = old_new[loop->_head->_idx]; 1228 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 1229 1230 1231 // Step 3: Now fix control uses. Loop varying control uses have already 1232 // been fixed up (as part of all input edges in Step 2). Loop invariant 1233 // control uses must be either an IfFalse or an IfTrue. Make a merge 1234 // point to merge the old and new IfFalse/IfTrue nodes; make the use 1235 // refer to this. 1236 ResourceArea *area = Thread::current()->resource_area(); 1237 Node_List worklist(area); 1238 uint new_counter = C->unique(); 1239 for( i = 0; i < loop->_body.size(); i++ ) { 1240 Node* old = loop->_body.at(i); 1241 if( !old->is_CFG() ) continue; 1242 Node* nnn = old_new[old->_idx]; 1243 1244 // Copy uses to a worklist, so I can munge the def-use info 1245 // with impunity. 1246 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1247 worklist.push(old->fast_out(j)); 1248 1249 while( worklist.size() ) { // Visit all uses 1250 Node *use = worklist.pop(); 1251 if (!has_node(use)) continue; // Ignore dead nodes 1252 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1253 if( !loop->is_member( use_loop ) && use->is_CFG() ) { 1254 // Both OLD and USE are CFG nodes here. 1255 assert( use->is_Proj(), "" ); 1256 1257 // Clone the loop exit control projection 1258 Node *newuse = use->clone(); 1259 newuse->set_req(0,nnn); 1260 _igvn.register_new_node_with_optimizer(newuse); 1261 set_loop(newuse, use_loop); 1262 set_idom(newuse, nnn, dom_depth(nnn) + 1 ); 1263 1264 // We need a Region to merge the exit from the peeled body and the 1265 // exit from the old loop body. 1266 RegionNode *r = new (C, 3) RegionNode(3); 1267 // Map the old use to the new merge point 1268 old_new.map( use->_idx, r ); 1269 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use)); 1270 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" ); 1271 1272 // The original user of 'use' uses 'r' instead. 1273 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) { 1274 Node* useuse = use->last_out(l); 1275 _igvn.hash_delete(useuse); 1276 _igvn._worklist.push(useuse); 1277 uint uses_found = 0; 1278 if( useuse->in(0) == use ) { 1279 useuse->set_req(0, r); 1280 uses_found++; 1281 if( useuse->is_CFG() ) { 1282 assert( dom_depth(useuse) > dd_r, "" ); 1283 set_idom(useuse, r, dom_depth(useuse)); 1284 } 1285 } 1286 for( uint k = 1; k < useuse->req(); k++ ) { 1287 if( useuse->in(k) == use ) { 1288 useuse->set_req(k, r); 1289 uses_found++; 1290 } 1291 } 1292 l -= uses_found; // we deleted 1 or more copies of this edge 1293 } 1294 1295 // Now finish up 'r' 1296 r->set_req( 1, newuse ); 1297 r->set_req( 2, use ); 1298 _igvn.register_new_node_with_optimizer(r); 1299 set_loop(r, use_loop); 1300 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r); 1301 } // End of if a loop-exit test 1302 } 1303 } 1304 1305 // Step 4: If loop-invariant use is not control, it must be dominated by a 1306 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region 1307 // there if needed. Make a Phi there merging old and new used values. 1308 Node_List *split_if_set = NULL; 1309 Node_List *split_bool_set = NULL; 1310 Node_List *split_cex_set = NULL; 1311 for( i = 0; i < loop->_body.size(); i++ ) { 1312 Node* old = loop->_body.at(i); 1313 Node* nnn = old_new[old->_idx]; 1314 // Copy uses to a worklist, so I can munge the def-use info 1315 // with impunity. 1316 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1317 worklist.push(old->fast_out(j)); 1318 1319 while( worklist.size() ) { 1320 Node *use = worklist.pop(); 1321 if (!has_node(use)) continue; // Ignore dead nodes 1322 if (use->in(0) == C->top()) continue; 1323 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1324 // Check for data-use outside of loop - at least one of OLD or USE 1325 // must not be a CFG node. 1326 if( !loop->is_member( use_loop ) && (!old->is_CFG() || !use->is_CFG())) { 1327 1328 // If the Data use is an IF, that means we have an IF outside of the 1329 // loop that is switching on a condition that is set inside of the 1330 // loop. Happens if people set a loop-exit flag; then test the flag 1331 // in the loop to break the loop, then test is again outside of the 1332 // loop to determine which way the loop exited. 1333 if( use->is_If() || use->is_CMove() ) { 1334 // Since this code is highly unlikely, we lazily build the worklist 1335 // of such Nodes to go split. 1336 if( !split_if_set ) 1337 split_if_set = new Node_List(area); 1338 split_if_set->push(use); 1339 } 1340 if( use->is_Bool() ) { 1341 if( !split_bool_set ) 1342 split_bool_set = new Node_List(area); 1343 split_bool_set->push(use); 1344 } 1345 if( use->Opcode() == Op_CreateEx ) { 1346 if( !split_cex_set ) 1347 split_cex_set = new Node_List(area); 1348 split_cex_set->push(use); 1349 } 1350 1351 1352 // Get "block" use is in 1353 uint idx = 0; 1354 while( use->in(idx) != old ) idx++; 1355 Node *prev = use->is_CFG() ? use : get_ctrl(use); 1356 assert( !loop->is_member( get_loop( prev ) ), "" ); 1357 Node *cfg = prev->_idx >= new_counter 1358 ? prev->in(2) 1359 : idom(prev); 1360 if( use->is_Phi() ) // Phi use is in prior block 1361 cfg = prev->in(idx); // NOT in block of Phi itself 1362 if (cfg->is_top()) { // Use is dead? 1363 _igvn.hash_delete(use); 1364 _igvn._worklist.push(use); 1365 use->set_req(idx, C->top()); 1366 continue; 1367 } 1368 1369 while( !loop->is_member( get_loop( cfg ) ) ) { 1370 prev = cfg; 1371 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg); 1372 } 1373 // If the use occurs after merging several exits from the loop, then 1374 // old value must have dominated all those exits. Since the same old 1375 // value was used on all those exits we did not need a Phi at this 1376 // merge point. NOW we do need a Phi here. Each loop exit value 1377 // is now merged with the peeled body exit; each exit gets its own 1378 // private Phi and those Phis need to be merged here. 1379 Node *phi; 1380 if( prev->is_Region() ) { 1381 if( idx == 0 ) { // Updating control edge? 1382 phi = prev; // Just use existing control 1383 } else { // Else need a new Phi 1384 phi = PhiNode::make( prev, old ); 1385 // Now recursively fix up the new uses of old! 1386 for( uint i = 1; i < prev->req(); i++ ) { 1387 worklist.push(phi); // Onto worklist once for each 'old' input 1388 } 1389 } 1390 } else { 1391 // Get new RegionNode merging old and new loop exits 1392 prev = old_new[prev->_idx]; 1393 assert( prev, "just made this in step 7" ); 1394 if( idx == 0 ) { // Updating control edge? 1395 phi = prev; // Just use existing control 1396 } else { // Else need a new Phi 1397 // Make a new Phi merging data values properly 1398 phi = PhiNode::make( prev, old ); 1399 phi->set_req( 1, nnn ); 1400 } 1401 } 1402 // If inserting a new Phi, check for prior hits 1403 if( idx != 0 ) { 1404 Node *hit = _igvn.hash_find_insert(phi); 1405 if( hit == NULL ) { 1406 _igvn.register_new_node_with_optimizer(phi); // Register new phi 1407 } else { // or 1408 // Remove the new phi from the graph and use the hit 1409 _igvn.remove_dead_node(phi); 1410 phi = hit; // Use existing phi 1411 } 1412 set_ctrl(phi, prev); 1413 } 1414 // Make 'use' use the Phi instead of the old loop body exit value 1415 _igvn.hash_delete(use); 1416 _igvn._worklist.push(use); 1417 use->set_req(idx, phi); 1418 if( use->_idx >= new_counter ) { // If updating new phis 1419 // Not needed for correctness, but prevents a weak assert 1420 // in AddPNode from tripping (when we end up with different 1421 // base & derived Phis that will become the same after 1422 // IGVN does CSE). 1423 Node *hit = _igvn.hash_find_insert(use); 1424 if( hit ) // Go ahead and re-hash for hits. 1425 _igvn.replace_node( use, hit ); 1426 } 1427 1428 // If 'use' was in the loop-exit block, it now needs to be sunk 1429 // below the post-loop merge point. 1430 sink_use( use, prev ); 1431 } 1432 } 1433 } 1434 1435 // Check for IFs that need splitting/cloning. Happens if an IF outside of 1436 // the loop uses a condition set in the loop. The original IF probably 1437 // takes control from one or more OLD Regions (which in turn get from NEW 1438 // Regions). In any case, there will be a set of Phis for each merge point 1439 // from the IF up to where the original BOOL def exists the loop. 1440 if( split_if_set ) { 1441 while( split_if_set->size() ) { 1442 Node *iff = split_if_set->pop(); 1443 if( iff->in(1)->is_Phi() ) { 1444 BoolNode *b = clone_iff( iff->in(1)->as_Phi(), loop ); 1445 _igvn.hash_delete(iff); 1446 _igvn._worklist.push(iff); 1447 iff->set_req(1, b); 1448 } 1449 } 1450 } 1451 if( split_bool_set ) { 1452 while( split_bool_set->size() ) { 1453 Node *b = split_bool_set->pop(); 1454 Node *phi = b->in(1); 1455 assert( phi->is_Phi(), "" ); 1456 CmpNode *cmp = clone_bool( (PhiNode*)phi, loop ); 1457 _igvn.hash_delete(b); 1458 _igvn._worklist.push(b); 1459 b->set_req(1, cmp); 1460 } 1461 } 1462 if( split_cex_set ) { 1463 while( split_cex_set->size() ) { 1464 Node *b = split_cex_set->pop(); 1465 assert( b->in(0)->is_Region(), "" ); 1466 assert( b->in(1)->is_Phi(), "" ); 1467 assert( b->in(0)->in(0) == b->in(1)->in(0), "" ); 1468 split_up( b, b->in(0), NULL ); 1469 } 1470 } 1471 1472 } 1473 1474 1475 //---------------------- stride_of_possible_iv ------------------------------------- 1476 // Looks for an iff/bool/comp with one operand of the compare 1477 // being a cycle involving an add and a phi, 1478 // with an optional truncation (left-shift followed by a right-shift) 1479 // of the add. Returns zero if not an iv. 1480 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) { 1481 Node* trunc1 = NULL; 1482 Node* trunc2 = NULL; 1483 const TypeInt* ttype = NULL; 1484 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) { 1485 return 0; 1486 } 1487 BoolNode* bl = iff->in(1)->as_Bool(); 1488 Node* cmp = bl->in(1); 1489 if (!cmp || cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU) { 1490 return 0; 1491 } 1492 // Must have an invariant operand 1493 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) { 1494 return 0; 1495 } 1496 Node* add2 = NULL; 1497 Node* cmp1 = cmp->in(1); 1498 if (cmp1->is_Phi()) { 1499 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) ))) 1500 Node* phi = cmp1; 1501 for (uint i = 1; i < phi->req(); i++) { 1502 Node* in = phi->in(i); 1503 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in, 1504 &trunc1, &trunc2, &ttype); 1505 if (add && add->in(1) == phi) { 1506 add2 = add->in(2); 1507 break; 1508 } 1509 } 1510 } else { 1511 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) ))) 1512 Node* addtrunc = cmp1; 1513 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc, 1514 &trunc1, &trunc2, &ttype); 1515 if (add && add->in(1)->is_Phi()) { 1516 Node* phi = add->in(1); 1517 for (uint i = 1; i < phi->req(); i++) { 1518 if (phi->in(i) == addtrunc) { 1519 add2 = add->in(2); 1520 break; 1521 } 1522 } 1523 } 1524 } 1525 if (add2 != NULL) { 1526 const TypeInt* add2t = _igvn.type(add2)->is_int(); 1527 if (add2t->is_con()) { 1528 return add2t->get_con(); 1529 } 1530 } 1531 return 0; 1532 } 1533 1534 1535 //---------------------- stay_in_loop ------------------------------------- 1536 // Return the (unique) control output node that's in the loop (if it exists.) 1537 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) { 1538 Node* unique = NULL; 1539 if (!n) return NULL; 1540 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1541 Node* use = n->fast_out(i); 1542 if (!has_ctrl(use) && loop->is_member(get_loop(use))) { 1543 if (unique != NULL) { 1544 return NULL; 1545 } 1546 unique = use; 1547 } 1548 } 1549 return unique; 1550 } 1551 1552 //------------------------------ register_node ------------------------------------- 1553 // Utility to register node "n" with PhaseIdealLoop 1554 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) { 1555 _igvn.register_new_node_with_optimizer(n); 1556 loop->_body.push(n); 1557 if (n->is_CFG()) { 1558 set_loop(n, loop); 1559 set_idom(n, pred, ddepth); 1560 } else { 1561 set_ctrl(n, pred); 1562 } 1563 } 1564 1565 //------------------------------ proj_clone ------------------------------------- 1566 // Utility to create an if-projection 1567 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) { 1568 ProjNode* c = p->clone()->as_Proj(); 1569 c->set_req(0, iff); 1570 return c; 1571 } 1572 1573 //------------------------------ short_circuit_if ------------------------------------- 1574 // Force the iff control output to be the live_proj 1575 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) { 1576 int proj_con = live_proj->_con; 1577 assert(proj_con == 0 || proj_con == 1, "false or true projection"); 1578 Node *con = _igvn.intcon(proj_con); 1579 set_ctrl(con, C->root()); 1580 if (iff) { 1581 iff->set_req(1, con); 1582 } 1583 return con; 1584 } 1585 1586 //------------------------------ insert_if_before_proj ------------------------------------- 1587 // Insert a new if before an if projection (* - new node) 1588 // 1589 // before 1590 // if(test) 1591 // / \ 1592 // v v 1593 // other-proj proj (arg) 1594 // 1595 // after 1596 // if(test) 1597 // / \ 1598 // / v 1599 // | * proj-clone 1600 // v | 1601 // other-proj v 1602 // * new_if(relop(cmp[IU](left,right))) 1603 // / \ 1604 // v v 1605 // * new-proj proj 1606 // (returned) 1607 // 1608 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) { 1609 IfNode* iff = proj->in(0)->as_If(); 1610 IdealLoopTree *loop = get_loop(proj); 1611 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 1612 int ddepth = dom_depth(proj); 1613 1614 _igvn.hash_delete(iff); 1615 _igvn._worklist.push(iff); 1616 _igvn.hash_delete(proj); 1617 _igvn._worklist.push(proj); 1618 1619 proj->set_req(0, NULL); // temporary disconnect 1620 ProjNode* proj2 = proj_clone(proj, iff); 1621 register_node(proj2, loop, iff, ddepth); 1622 1623 Node* cmp = Signed ? (Node*) new (C,3)CmpINode(left, right) : (Node*) new (C,3)CmpUNode(left, right); 1624 register_node(cmp, loop, proj2, ddepth); 1625 1626 BoolNode* bol = new (C,2)BoolNode(cmp, relop); 1627 register_node(bol, loop, proj2, ddepth); 1628 1629 IfNode* new_if = new (C,2)IfNode(proj2, bol, iff->_prob, iff->_fcnt); 1630 register_node(new_if, loop, proj2, ddepth); 1631 1632 proj->set_req(0, new_if); // reattach 1633 set_idom(proj, new_if, ddepth); 1634 1635 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj(); 1636 register_node(new_exit, get_loop(other_proj), new_if, ddepth); 1637 1638 return new_exit; 1639 } 1640 1641 //------------------------------ insert_region_before_proj ------------------------------------- 1642 // Insert a region before an if projection (* - new node) 1643 // 1644 // before 1645 // if(test) 1646 // / | 1647 // v | 1648 // proj v 1649 // other-proj 1650 // 1651 // after 1652 // if(test) 1653 // / | 1654 // v | 1655 // * proj-clone v 1656 // | other-proj 1657 // v 1658 // * new-region 1659 // | 1660 // v 1661 // * dum_if 1662 // / \ 1663 // v \ 1664 // * dum-proj v 1665 // proj 1666 // 1667 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) { 1668 IfNode* iff = proj->in(0)->as_If(); 1669 IdealLoopTree *loop = get_loop(proj); 1670 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 1671 int ddepth = dom_depth(proj); 1672 1673 _igvn.hash_delete(iff); 1674 _igvn._worklist.push(iff); 1675 _igvn.hash_delete(proj); 1676 _igvn._worklist.push(proj); 1677 1678 proj->set_req(0, NULL); // temporary disconnect 1679 ProjNode* proj2 = proj_clone(proj, iff); 1680 register_node(proj2, loop, iff, ddepth); 1681 1682 RegionNode* reg = new (C,2)RegionNode(2); 1683 reg->set_req(1, proj2); 1684 register_node(reg, loop, iff, ddepth); 1685 1686 IfNode* dum_if = new (C,2)IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt); 1687 register_node(dum_if, loop, reg, ddepth); 1688 1689 proj->set_req(0, dum_if); // reattach 1690 set_idom(proj, dum_if, ddepth); 1691 1692 ProjNode* dum_proj = proj_clone(other_proj, dum_if); 1693 register_node(dum_proj, loop, dum_if, ddepth); 1694 1695 return reg; 1696 } 1697 1698 //------------------------------ insert_cmpi_loop_exit ------------------------------------- 1699 // Clone a signed compare loop exit from an unsigned compare and 1700 // insert it before the unsigned cmp on the stay-in-loop path. 1701 // All new nodes inserted in the dominator tree between the original 1702 // if and it's projections. The original if test is replaced with 1703 // a constant to force the stay-in-loop path. 1704 // 1705 // This is done to make sure that the original if and it's projections 1706 // still dominate the same set of control nodes, that the ctrl() relation 1707 // from data nodes to them is preserved, and that their loop nesting is 1708 // preserved. 1709 // 1710 // before 1711 // if(i <u limit) unsigned compare loop exit 1712 // / | 1713 // v v 1714 // exit-proj stay-in-loop-proj 1715 // 1716 // after 1717 // if(stay-in-loop-const) original if 1718 // / | 1719 // / v 1720 // / if(i < limit) new signed test 1721 // / / | 1722 // / / v 1723 // / / if(i <u limit) new cloned unsigned test 1724 // / / / | 1725 // v v v | 1726 // region | 1727 // | | 1728 // dum-if | 1729 // / | | 1730 // ether | | 1731 // v v 1732 // exit-proj stay-in-loop-proj 1733 // 1734 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) { 1735 const bool Signed = true; 1736 const bool Unsigned = false; 1737 1738 BoolNode* bol = if_cmpu->in(1)->as_Bool(); 1739 if (bol->_test._test != BoolTest::lt) return NULL; 1740 CmpNode* cmpu = bol->in(1)->as_Cmp(); 1741 if (cmpu->Opcode() != Op_CmpU) return NULL; 1742 int stride = stride_of_possible_iv(if_cmpu); 1743 if (stride == 0) return NULL; 1744 1745 ProjNode* lp_continue = stay_in_loop(if_cmpu, loop)->as_Proj(); 1746 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj(); 1747 1748 Node* limit = NULL; 1749 if (stride > 0) { 1750 limit = cmpu->in(2); 1751 } else { 1752 limit = _igvn.makecon(TypeInt::ZERO); 1753 set_ctrl(limit, C->root()); 1754 } 1755 // Create a new region on the exit path 1756 RegionNode* reg = insert_region_before_proj(lp_exit); 1757 1758 // Clone the if-cmpu-true-false using a signed compare 1759 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge; 1760 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue); 1761 reg->add_req(cmpi_exit); 1762 1763 // Clone the if-cmpu-true-false 1764 BoolTest::mask rel_u = bol->_test._test; 1765 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue); 1766 reg->add_req(cmpu_exit); 1767 1768 // Force original if to stay in loop. 1769 short_circuit_if(if_cmpu, lp_continue); 1770 1771 return cmpi_exit->in(0)->as_If(); 1772 } 1773 1774 //------------------------------ remove_cmpi_loop_exit ------------------------------------- 1775 // Remove a previously inserted signed compare loop exit. 1776 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) { 1777 Node* lp_proj = stay_in_loop(if_cmp, loop); 1778 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI && 1779 stay_in_loop(lp_proj, loop)->is_If() && 1780 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu"); 1781 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO); 1782 set_ctrl(con, C->root()); 1783 if_cmp->set_req(1, con); 1784 } 1785 1786 //------------------------------ scheduled_nodelist ------------------------------------- 1787 // Create a post order schedule of nodes that are in the 1788 // "member" set. The list is returned in "sched". 1789 // The first node in "sched" is the loop head, followed by 1790 // nodes which have no inputs in the "member" set, and then 1791 // followed by the nodes that have an immediate input dependence 1792 // on a node in "sched". 1793 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) { 1794 1795 assert(member.test(loop->_head->_idx), "loop head must be in member set"); 1796 Arena *a = Thread::current()->resource_area(); 1797 VectorSet visited(a); 1798 Node_Stack nstack(a, loop->_body.size()); 1799 1800 Node* n = loop->_head; // top of stack is cached in "n" 1801 uint idx = 0; 1802 visited.set(n->_idx); 1803 1804 // Initially push all with no inputs from within member set 1805 for(uint i = 0; i < loop->_body.size(); i++ ) { 1806 Node *elt = loop->_body.at(i); 1807 if (member.test(elt->_idx)) { 1808 bool found = false; 1809 for (uint j = 0; j < elt->req(); j++) { 1810 Node* def = elt->in(j); 1811 if (def && member.test(def->_idx) && def != elt) { 1812 found = true; 1813 break; 1814 } 1815 } 1816 if (!found && elt != loop->_head) { 1817 nstack.push(n, idx); 1818 n = elt; 1819 assert(!visited.test(n->_idx), "not seen yet"); 1820 visited.set(n->_idx); 1821 } 1822 } 1823 } 1824 1825 // traverse out's that are in the member set 1826 while (true) { 1827 if (idx < n->outcnt()) { 1828 Node* use = n->raw_out(idx); 1829 idx++; 1830 if (!visited.test_set(use->_idx)) { 1831 if (member.test(use->_idx)) { 1832 nstack.push(n, idx); 1833 n = use; 1834 idx = 0; 1835 } 1836 } 1837 } else { 1838 // All outputs processed 1839 sched.push(n); 1840 if (nstack.is_empty()) break; 1841 n = nstack.node(); 1842 idx = nstack.index(); 1843 nstack.pop(); 1844 } 1845 } 1846 } 1847 1848 1849 //------------------------------ has_use_in_set ------------------------------------- 1850 // Has a use in the vector set 1851 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) { 1852 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1853 Node* use = n->fast_out(j); 1854 if (vset.test(use->_idx)) { 1855 return true; 1856 } 1857 } 1858 return false; 1859 } 1860 1861 1862 //------------------------------ has_use_internal_to_set ------------------------------------- 1863 // Has use internal to the vector set (ie. not in a phi at the loop head) 1864 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) { 1865 Node* head = loop->_head; 1866 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1867 Node* use = n->fast_out(j); 1868 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) { 1869 return true; 1870 } 1871 } 1872 return false; 1873 } 1874 1875 1876 //------------------------------ clone_for_use_outside_loop ------------------------------------- 1877 // clone "n" for uses that are outside of loop 1878 void PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) { 1879 1880 assert(worklist.size() == 0, "should be empty"); 1881 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1882 Node* use = n->fast_out(j); 1883 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) { 1884 worklist.push(use); 1885 } 1886 } 1887 while( worklist.size() ) { 1888 Node *use = worklist.pop(); 1889 if (!has_node(use) || use->in(0) == C->top()) continue; 1890 uint j; 1891 for (j = 0; j < use->req(); j++) { 1892 if (use->in(j) == n) break; 1893 } 1894 assert(j < use->req(), "must be there"); 1895 1896 // clone "n" and insert it between the inputs of "n" and the use outside the loop 1897 Node* n_clone = n->clone(); 1898 _igvn.hash_delete(use); 1899 use->set_req(j, n_clone); 1900 _igvn._worklist.push(use); 1901 Node* use_c; 1902 if (!use->is_Phi()) { 1903 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0); 1904 } else { 1905 // Use in a phi is considered a use in the associated predecessor block 1906 use_c = use->in(0)->in(j); 1907 } 1908 set_ctrl(n_clone, use_c); 1909 assert(!loop->is_member(get_loop(use_c)), "should be outside loop"); 1910 get_loop(use_c)->_body.push(n_clone); 1911 _igvn.register_new_node_with_optimizer(n_clone); 1912 #if !defined(PRODUCT) 1913 if (TracePartialPeeling) { 1914 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx); 1915 } 1916 #endif 1917 } 1918 } 1919 1920 1921 //------------------------------ clone_for_special_use_inside_loop ------------------------------------- 1922 // clone "n" for special uses that are in the not_peeled region. 1923 // If these def-uses occur in separate blocks, the code generator 1924 // marks the method as not compilable. For example, if a "BoolNode" 1925 // is in a different basic block than the "IfNode" that uses it, then 1926 // the compilation is aborted in the code generator. 1927 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 1928 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) { 1929 if (n->is_Phi() || n->is_Load()) { 1930 return; 1931 } 1932 assert(worklist.size() == 0, "should be empty"); 1933 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1934 Node* use = n->fast_out(j); 1935 if ( not_peel.test(use->_idx) && 1936 (use->is_If() || use->is_CMove() || use->is_Bool()) && 1937 use->in(1) == n) { 1938 worklist.push(use); 1939 } 1940 } 1941 if (worklist.size() > 0) { 1942 // clone "n" and insert it between inputs of "n" and the use 1943 Node* n_clone = n->clone(); 1944 loop->_body.push(n_clone); 1945 _igvn.register_new_node_with_optimizer(n_clone); 1946 set_ctrl(n_clone, get_ctrl(n)); 1947 sink_list.push(n_clone); 1948 not_peel <<= n_clone->_idx; // add n_clone to not_peel set. 1949 #if !defined(PRODUCT) 1950 if (TracePartialPeeling) { 1951 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx); 1952 } 1953 #endif 1954 while( worklist.size() ) { 1955 Node *use = worklist.pop(); 1956 _igvn.hash_delete(use); 1957 _igvn._worklist.push(use); 1958 for (uint j = 1; j < use->req(); j++) { 1959 if (use->in(j) == n) { 1960 use->set_req(j, n_clone); 1961 } 1962 } 1963 } 1964 } 1965 } 1966 1967 1968 //------------------------------ insert_phi_for_loop ------------------------------------- 1969 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 1970 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) { 1971 Node *phi = PhiNode::make(lp, back_edge_val); 1972 phi->set_req(LoopNode::EntryControl, lp_entry_val); 1973 // Use existing phi if it already exists 1974 Node *hit = _igvn.hash_find_insert(phi); 1975 if( hit == NULL ) { 1976 _igvn.register_new_node_with_optimizer(phi); 1977 set_ctrl(phi, lp); 1978 } else { 1979 // Remove the new phi from the graph and use the hit 1980 _igvn.remove_dead_node(phi); 1981 phi = hit; 1982 } 1983 _igvn.hash_delete(use); 1984 _igvn._worklist.push(use); 1985 use->set_req(idx, phi); 1986 } 1987 1988 #ifdef ASSERT 1989 //------------------------------ is_valid_loop_partition ------------------------------------- 1990 // Validate the loop partition sets: peel and not_peel 1991 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, 1992 VectorSet& not_peel ) { 1993 uint i; 1994 // Check that peel_list entries are in the peel set 1995 for (i = 0; i < peel_list.size(); i++) { 1996 if (!peel.test(peel_list.at(i)->_idx)) { 1997 return false; 1998 } 1999 } 2000 // Check at loop members are in one of peel set or not_peel set 2001 for (i = 0; i < loop->_body.size(); i++ ) { 2002 Node *def = loop->_body.at(i); 2003 uint di = def->_idx; 2004 // Check that peel set elements are in peel_list 2005 if (peel.test(di)) { 2006 if (not_peel.test(di)) { 2007 return false; 2008 } 2009 // Must be in peel_list also 2010 bool found = false; 2011 for (uint j = 0; j < peel_list.size(); j++) { 2012 if (peel_list.at(j)->_idx == di) { 2013 found = true; 2014 break; 2015 } 2016 } 2017 if (!found) { 2018 return false; 2019 } 2020 } else if (not_peel.test(di)) { 2021 if (peel.test(di)) { 2022 return false; 2023 } 2024 } else { 2025 return false; 2026 } 2027 } 2028 return true; 2029 } 2030 2031 //------------------------------ is_valid_clone_loop_exit_use ------------------------------------- 2032 // Ensure a use outside of loop is of the right form 2033 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) { 2034 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2035 return (use->is_Phi() && 2036 use_c->is_Region() && use_c->req() == 3 && 2037 (use_c->in(exit_idx)->Opcode() == Op_IfTrue || 2038 use_c->in(exit_idx)->Opcode() == Op_IfFalse || 2039 use_c->in(exit_idx)->Opcode() == Op_JumpProj) && 2040 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) ); 2041 } 2042 2043 //------------------------------ is_valid_clone_loop_form ------------------------------------- 2044 // Ensure that all uses outside of loop are of the right form 2045 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 2046 uint orig_exit_idx, uint clone_exit_idx) { 2047 uint len = peel_list.size(); 2048 for (uint i = 0; i < len; i++) { 2049 Node *def = peel_list.at(i); 2050 2051 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 2052 Node *use = def->fast_out(j); 2053 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2054 if (!loop->is_member(get_loop(use_c))) { 2055 // use is not in the loop, check for correct structure 2056 if (use->in(0) == def) { 2057 // Okay 2058 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) { 2059 return false; 2060 } 2061 } 2062 } 2063 } 2064 return true; 2065 } 2066 #endif 2067 2068 //------------------------------ partial_peel ------------------------------------- 2069 // Partially peel (aka loop rotation) the top portion of a loop (called 2070 // the peel section below) by cloning it and placing one copy just before 2071 // the new loop head and the other copy at the bottom of the new loop. 2072 // 2073 // before after where it came from 2074 // 2075 // stmt1 stmt1 2076 // loop: stmt2 clone 2077 // stmt2 if condA goto exitA clone 2078 // if condA goto exitA new_loop: new 2079 // stmt3 stmt3 clone 2080 // if !condB goto loop if condB goto exitB clone 2081 // exitB: stmt2 orig 2082 // stmt4 if !condA goto new_loop orig 2083 // exitA: goto exitA 2084 // exitB: 2085 // stmt4 2086 // exitA: 2087 // 2088 // Step 1: find the cut point: an exit test on probable 2089 // induction variable. 2090 // Step 2: schedule (with cloning) operations in the peel 2091 // section that can be executed after the cut into 2092 // the section that is not peeled. This may need 2093 // to clone operations into exit blocks. For 2094 // instance, a reference to A[i] in the not-peel 2095 // section and a reference to B[i] in an exit block 2096 // may cause a left-shift of i by 2 to be placed 2097 // in the peel block. This step will clone the left 2098 // shift into the exit block and sink the left shift 2099 // from the peel to the not-peel section. 2100 // Step 3: clone the loop, retarget the control, and insert 2101 // phis for values that are live across the new loop 2102 // head. This is very dependent on the graph structure 2103 // from clone_loop. It creates region nodes for 2104 // exit control and associated phi nodes for values 2105 // flow out of the loop through that exit. The region 2106 // node is dominated by the clone's control projection. 2107 // So the clone's peel section is placed before the 2108 // new loop head, and the clone's not-peel section is 2109 // forms the top part of the new loop. The original 2110 // peel section forms the tail of the new loop. 2111 // Step 4: update the dominator tree and recompute the 2112 // dominator depth. 2113 // 2114 // orig 2115 // 2116 // stmt1 2117 // | 2118 // v 2119 // loop<----+ 2120 // | | 2121 // stmt2 | 2122 // | | 2123 // v | 2124 // ifA | 2125 // / | | 2126 // v v | 2127 // false true ^ <-- last_peel 2128 // / | | 2129 // / ===|==cut | 2130 // / stmt3 | <-- first_not_peel 2131 // / | | 2132 // | v | 2133 // v ifB | 2134 // exitA: / \ | 2135 // / \ | 2136 // v v | 2137 // false true | 2138 // / \ | 2139 // / ----+ 2140 // | 2141 // v 2142 // exitB: 2143 // stmt4 2144 // 2145 // 2146 // after clone loop 2147 // 2148 // stmt1 2149 // / \ 2150 // clone / \ orig 2151 // / \ 2152 // / \ 2153 // v v 2154 // +---->loop loop<----+ 2155 // | | | | 2156 // | stmt2 stmt2 | 2157 // | | | | 2158 // | v v | 2159 // | ifA ifA | 2160 // | | \ / | | 2161 // | v v v v | 2162 // ^ true false false true ^ <-- last_peel 2163 // | | ^ \ / | | 2164 // | cut==|== \ \ / ===|==cut | 2165 // | stmt3 \ \ / stmt3 | <-- first_not_peel 2166 // | | dom | | | | 2167 // | v \ 1v v2 v | 2168 // | ifB regionA ifB | 2169 // | / \ | / \ | 2170 // | / \ v / \ | 2171 // | v v exitA: v v | 2172 // | true false false true | 2173 // | / ^ \ / \ | 2174 // +---- \ \ / ----+ 2175 // dom \ / 2176 // \ 1v v2 2177 // regionB 2178 // | 2179 // v 2180 // exitB: 2181 // stmt4 2182 // 2183 // 2184 // after partial peel 2185 // 2186 // stmt1 2187 // / 2188 // clone / orig 2189 // / TOP 2190 // / \ 2191 // v v 2192 // TOP->region region----+ 2193 // | | | 2194 // stmt2 stmt2 | 2195 // | | | 2196 // v v | 2197 // ifA ifA | 2198 // | \ / | | 2199 // v v v v | 2200 // true false false true | <-- last_peel 2201 // | ^ \ / +------|---+ 2202 // +->newloop \ \ / === ==cut | | 2203 // | stmt3 \ \ / TOP | | 2204 // | | dom | | stmt3 | | <-- first_not_peel 2205 // | v \ 1v v2 v | | 2206 // | ifB regionA ifB ^ v 2207 // | / \ | / \ | | 2208 // | / \ v / \ | | 2209 // | v v exitA: v v | | 2210 // | true false false true | | 2211 // | / ^ \ / \ | | 2212 // | | \ \ / v | | 2213 // | | dom \ / TOP | | 2214 // | | \ 1v v2 | | 2215 // ^ v regionB | | 2216 // | | | | | 2217 // | | v ^ v 2218 // | | exitB: | | 2219 // | | stmt4 | | 2220 // | +------------>-----------------+ | 2221 // | | 2222 // +-----------------<---------------------+ 2223 // 2224 // 2225 // final graph 2226 // 2227 // stmt1 2228 // | 2229 // v 2230 // ........> ifA clone 2231 // : / | 2232 // dom / | 2233 // : v v 2234 // : false true 2235 // : | | 2236 // : | stmt2 clone 2237 // : | | 2238 // : | v 2239 // : | newloop<-----+ 2240 // : | | | 2241 // : | stmt3 clone | 2242 // : | | | 2243 // : | v | 2244 // : | ifB | 2245 // : | / \ | 2246 // : | v v | 2247 // : | false true | 2248 // : | | | | 2249 // : | v stmt2 | 2250 // : | exitB: | | 2251 // : | stmt4 v | 2252 // : | ifA orig | 2253 // : | / \ | 2254 // : | / \ | 2255 // : | v v | 2256 // : | false true | 2257 // : | / \ | 2258 // : v v -----+ 2259 // RegionA 2260 // | 2261 // v 2262 // exitA 2263 // 2264 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) { 2265 2266 if (!loop->_head->is_Loop()) { 2267 return false; } 2268 2269 LoopNode *head = loop->_head->as_Loop(); 2270 2271 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) { 2272 return false; 2273 } 2274 2275 // Check for complex exit control 2276 for(uint ii = 0; ii < loop->_body.size(); ii++ ) { 2277 Node *n = loop->_body.at(ii); 2278 int opc = n->Opcode(); 2279 if (n->is_Call() || 2280 opc == Op_Catch || 2281 opc == Op_CatchProj || 2282 opc == Op_Jump || 2283 opc == Op_JumpProj) { 2284 #if !defined(PRODUCT) 2285 if (TracePartialPeeling) { 2286 tty->print_cr("\nExit control too complex: lp: %d", head->_idx); 2287 } 2288 #endif 2289 return false; 2290 } 2291 } 2292 2293 int dd = dom_depth(head); 2294 2295 // Step 1: find cut point 2296 2297 // Walk up dominators to loop head looking for first loop exit 2298 // which is executed on every path thru loop. 2299 IfNode *peel_if = NULL; 2300 IfNode *peel_if_cmpu = NULL; 2301 2302 Node *iff = loop->tail(); 2303 while( iff != head ) { 2304 if( iff->is_If() ) { 2305 Node *ctrl = get_ctrl(iff->in(1)); 2306 if (ctrl->is_top()) return false; // Dead test on live IF. 2307 // If loop-varying exit-test, check for induction variable 2308 if( loop->is_member(get_loop(ctrl)) && 2309 loop->is_loop_exit(iff) && 2310 is_possible_iv_test(iff)) { 2311 Node* cmp = iff->in(1)->in(1); 2312 if (cmp->Opcode() == Op_CmpI) { 2313 peel_if = iff->as_If(); 2314 } else { 2315 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU"); 2316 peel_if_cmpu = iff->as_If(); 2317 } 2318 } 2319 } 2320 iff = idom(iff); 2321 } 2322 // Prefer signed compare over unsigned compare. 2323 IfNode* new_peel_if = NULL; 2324 if (peel_if == NULL) { 2325 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) { 2326 return false; // No peel point found 2327 } 2328 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop); 2329 if (new_peel_if == NULL) { 2330 return false; // No peel point found 2331 } 2332 peel_if = new_peel_if; 2333 } 2334 Node* last_peel = stay_in_loop(peel_if, loop); 2335 Node* first_not_peeled = stay_in_loop(last_peel, loop); 2336 if (first_not_peeled == NULL || first_not_peeled == head) { 2337 return false; 2338 } 2339 2340 #if !defined(PRODUCT) 2341 if (TracePartialPeeling) { 2342 tty->print_cr("before partial peel one iteration"); 2343 Node_List wl; 2344 Node* t = head->in(2); 2345 while (true) { 2346 wl.push(t); 2347 if (t == head) break; 2348 t = idom(t); 2349 } 2350 while (wl.size() > 0) { 2351 Node* tt = wl.pop(); 2352 tt->dump(); 2353 if (tt == last_peel) tty->print_cr("-- cut --"); 2354 } 2355 } 2356 #endif 2357 ResourceArea *area = Thread::current()->resource_area(); 2358 VectorSet peel(area); 2359 VectorSet not_peel(area); 2360 Node_List peel_list(area); 2361 Node_List worklist(area); 2362 Node_List sink_list(area); 2363 2364 // Set of cfg nodes to peel are those that are executable from 2365 // the head through last_peel. 2366 assert(worklist.size() == 0, "should be empty"); 2367 worklist.push(head); 2368 peel.set(head->_idx); 2369 while (worklist.size() > 0) { 2370 Node *n = worklist.pop(); 2371 if (n != last_peel) { 2372 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2373 Node* use = n->fast_out(j); 2374 if (use->is_CFG() && 2375 loop->is_member(get_loop(use)) && 2376 !peel.test_set(use->_idx)) { 2377 worklist.push(use); 2378 } 2379 } 2380 } 2381 } 2382 2383 // Set of non-cfg nodes to peel are those that are control 2384 // dependent on the cfg nodes. 2385 uint i; 2386 for(i = 0; i < loop->_body.size(); i++ ) { 2387 Node *n = loop->_body.at(i); 2388 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n; 2389 if (peel.test(n_c->_idx)) { 2390 peel.set(n->_idx); 2391 } else { 2392 not_peel.set(n->_idx); 2393 } 2394 } 2395 2396 // Step 2: move operations from the peeled section down into the 2397 // not-peeled section 2398 2399 // Get a post order schedule of nodes in the peel region 2400 // Result in right-most operand. 2401 scheduled_nodelist(loop, peel, peel_list ); 2402 2403 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 2404 2405 // For future check for too many new phis 2406 uint old_phi_cnt = 0; 2407 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 2408 Node* use = head->fast_out(j); 2409 if (use->is_Phi()) old_phi_cnt++; 2410 } 2411 2412 #if !defined(PRODUCT) 2413 if (TracePartialPeeling) { 2414 tty->print_cr("\npeeled list"); 2415 } 2416 #endif 2417 2418 // Evacuate nodes in peel region into the not_peeled region if possible 2419 uint new_phi_cnt = 0; 2420 for (i = 0; i < peel_list.size();) { 2421 Node* n = peel_list.at(i); 2422 #if !defined(PRODUCT) 2423 if (TracePartialPeeling) n->dump(); 2424 #endif 2425 bool incr = true; 2426 if ( !n->is_CFG() ) { 2427 2428 if ( has_use_in_set(n, not_peel) ) { 2429 2430 // If not used internal to the peeled region, 2431 // move "n" from peeled to not_peeled region. 2432 2433 if ( !has_use_internal_to_set(n, peel, loop) ) { 2434 2435 // if not pinned and not a load (which maybe anti-dependent on a store) 2436 // and not a CMove (Matcher expects only bool->cmove). 2437 if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) { 2438 clone_for_use_outside_loop( loop, n, worklist ); 2439 2440 sink_list.push(n); 2441 peel >>= n->_idx; // delete n from peel set. 2442 not_peel <<= n->_idx; // add n to not_peel set. 2443 peel_list.remove(i); 2444 incr = false; 2445 #if !defined(PRODUCT) 2446 if (TracePartialPeeling) { 2447 tty->print_cr("sink to not_peeled region: %d newbb: %d", 2448 n->_idx, get_ctrl(n)->_idx); 2449 } 2450 #endif 2451 } 2452 } else { 2453 // Otherwise check for special def-use cases that span 2454 // the peel/not_peel boundary such as bool->if 2455 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist ); 2456 new_phi_cnt++; 2457 } 2458 } 2459 } 2460 if (incr) i++; 2461 } 2462 2463 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) { 2464 #if !defined(PRODUCT) 2465 if (TracePartialPeeling) { 2466 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c", 2467 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F'); 2468 } 2469 #endif 2470 if (new_peel_if != NULL) { 2471 remove_cmpi_loop_exit(new_peel_if, loop); 2472 } 2473 // Inhibit more partial peeling on this loop 2474 assert(!head->is_partial_peel_loop(), "not partial peeled"); 2475 head->mark_partial_peel_failed(); 2476 return false; 2477 } 2478 2479 // Step 3: clone loop, retarget control, and insert new phis 2480 2481 // Create new loop head for new phis and to hang 2482 // the nodes being moved (sinked) from the peel region. 2483 LoopNode* new_head = new (C, 3) LoopNode(last_peel, last_peel); 2484 _igvn.register_new_node_with_optimizer(new_head); 2485 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled"); 2486 first_not_peeled->set_req(0, new_head); 2487 set_loop(new_head, loop); 2488 loop->_body.push(new_head); 2489 not_peel.set(new_head->_idx); 2490 set_idom(new_head, last_peel, dom_depth(first_not_peeled)); 2491 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled)); 2492 2493 while (sink_list.size() > 0) { 2494 Node* n = sink_list.pop(); 2495 set_ctrl(n, new_head); 2496 } 2497 2498 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 2499 2500 clone_loop( loop, old_new, dd ); 2501 2502 const uint clone_exit_idx = 1; 2503 const uint orig_exit_idx = 2; 2504 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop"); 2505 2506 Node* head_clone = old_new[head->_idx]; 2507 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop(); 2508 Node* orig_tail_clone = head_clone->in(2); 2509 2510 // Add phi if "def" node is in peel set and "use" is not 2511 2512 for(i = 0; i < peel_list.size(); i++ ) { 2513 Node *def = peel_list.at(i); 2514 if (!def->is_CFG()) { 2515 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 2516 Node *use = def->fast_out(j); 2517 if (has_node(use) && use->in(0) != C->top() && 2518 (!peel.test(use->_idx) || 2519 (use->is_Phi() && use->in(0) == head)) ) { 2520 worklist.push(use); 2521 } 2522 } 2523 while( worklist.size() ) { 2524 Node *use = worklist.pop(); 2525 for (uint j = 1; j < use->req(); j++) { 2526 Node* n = use->in(j); 2527 if (n == def) { 2528 2529 // "def" is in peel set, "use" is not in peel set 2530 // or "use" is in the entry boundary (a phi) of the peel set 2531 2532 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use; 2533 2534 if ( loop->is_member(get_loop( use_c )) ) { 2535 // use is in loop 2536 if (old_new[use->_idx] != NULL) { // null for dead code 2537 Node* use_clone = old_new[use->_idx]; 2538 _igvn.hash_delete(use); 2539 use->set_req(j, C->top()); 2540 _igvn._worklist.push(use); 2541 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone ); 2542 } 2543 } else { 2544 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format"); 2545 // use is not in the loop, check if the live range includes the cut 2546 Node* lp_if = use_c->in(orig_exit_idx)->in(0); 2547 if (not_peel.test(lp_if->_idx)) { 2548 assert(j == orig_exit_idx, "use from original loop"); 2549 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone ); 2550 } 2551 } 2552 } 2553 } 2554 } 2555 } 2556 } 2557 2558 // Step 3b: retarget control 2559 2560 // Redirect control to the new loop head if a cloned node in 2561 // the not_peeled region has control that points into the peeled region. 2562 // This necessary because the cloned peeled region will be outside 2563 // the loop. 2564 // from to 2565 // cloned-peeled <---+ 2566 // new_head_clone: | <--+ 2567 // cloned-not_peeled in(0) in(0) 2568 // orig-peeled 2569 2570 for(i = 0; i < loop->_body.size(); i++ ) { 2571 Node *n = loop->_body.at(i); 2572 if (!n->is_CFG() && n->in(0) != NULL && 2573 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) { 2574 Node* n_clone = old_new[n->_idx]; 2575 _igvn.hash_delete(n_clone); 2576 n_clone->set_req(0, new_head_clone); 2577 _igvn._worklist.push(n_clone); 2578 } 2579 } 2580 2581 // Backedge of the surviving new_head (the clone) is original last_peel 2582 _igvn.hash_delete(new_head_clone); 2583 new_head_clone->set_req(LoopNode::LoopBackControl, last_peel); 2584 _igvn._worklist.push(new_head_clone); 2585 2586 // Cut first node in original not_peel set 2587 _igvn.hash_delete(new_head); 2588 new_head->set_req(LoopNode::EntryControl, C->top()); 2589 new_head->set_req(LoopNode::LoopBackControl, C->top()); 2590 _igvn._worklist.push(new_head); 2591 2592 // Copy head_clone back-branch info to original head 2593 // and remove original head's loop entry and 2594 // clone head's back-branch 2595 _igvn.hash_delete(head); 2596 _igvn.hash_delete(head_clone); 2597 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl)); 2598 head->set_req(LoopNode::LoopBackControl, C->top()); 2599 head_clone->set_req(LoopNode::LoopBackControl, C->top()); 2600 _igvn._worklist.push(head); 2601 _igvn._worklist.push(head_clone); 2602 2603 // Similarly modify the phis 2604 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) { 2605 Node* use = head->fast_out(k); 2606 if (use->is_Phi() && use->outcnt() > 0) { 2607 Node* use_clone = old_new[use->_idx]; 2608 _igvn.hash_delete(use); 2609 _igvn.hash_delete(use_clone); 2610 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl)); 2611 use->set_req(LoopNode::LoopBackControl, C->top()); 2612 use_clone->set_req(LoopNode::LoopBackControl, C->top()); 2613 _igvn._worklist.push(use); 2614 _igvn._worklist.push(use_clone); 2615 } 2616 } 2617 2618 // Step 4: update dominator tree and dominator depth 2619 2620 set_idom(head, orig_tail_clone, dd); 2621 recompute_dom_depth(); 2622 2623 // Inhibit more partial peeling on this loop 2624 new_head_clone->set_partial_peel_loop(); 2625 C->set_major_progress(); 2626 2627 #if !defined(PRODUCT) 2628 if (TracePartialPeeling) { 2629 tty->print_cr("\nafter partial peel one iteration"); 2630 Node_List wl(area); 2631 Node* t = last_peel; 2632 while (true) { 2633 wl.push(t); 2634 if (t == head_clone) break; 2635 t = idom(t); 2636 } 2637 while (wl.size() > 0) { 2638 Node* tt = wl.pop(); 2639 if (tt == head) tty->print_cr("orig head"); 2640 else if (tt == new_head_clone) tty->print_cr("new head"); 2641 else if (tt == head_clone) tty->print_cr("clone head"); 2642 tt->dump(); 2643 } 2644 } 2645 #endif 2646 return true; 2647 } 2648 2649 //------------------------------reorg_offsets---------------------------------- 2650 // Reorganize offset computations to lower register pressure. Mostly 2651 // prevent loop-fallout uses of the pre-incremented trip counter (which are 2652 // then alive with the post-incremented trip counter forcing an extra 2653 // register move) 2654 void PhaseIdealLoop::reorg_offsets( IdealLoopTree *loop ) { 2655 2656 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 2657 CountedLoopEndNode *cle = cl->loopexit(); 2658 if( !cle ) return; // The occasional dead loop 2659 // Find loop exit control 2660 Node *exit = cle->proj_out(false); 2661 assert( exit->Opcode() == Op_IfFalse, "" ); 2662 2663 // Check for the special case of folks using the pre-incremented 2664 // trip-counter on the fall-out path (forces the pre-incremented 2665 // and post-incremented trip counter to be live at the same time). 2666 // Fix this by adjusting to use the post-increment trip counter. 2667 Node *phi = cl->phi(); 2668 if( !phi ) return; // Dead infinite loop 2669 2670 // Shape messed up, probably by iteration_split_impl 2671 if (phi->in(LoopNode::LoopBackControl) != cl->incr()) return; 2672 2673 bool progress = true; 2674 while (progress) { 2675 progress = false; 2676 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) { 2677 Node* use = phi->fast_out(i); // User of trip-counter 2678 if (!has_ctrl(use)) continue; 2679 Node *u_ctrl = get_ctrl(use); 2680 if( use->is_Phi() ) { 2681 u_ctrl = NULL; 2682 for( uint j = 1; j < use->req(); j++ ) 2683 if( use->in(j) == phi ) 2684 u_ctrl = dom_lca( u_ctrl, use->in(0)->in(j) ); 2685 } 2686 IdealLoopTree *u_loop = get_loop(u_ctrl); 2687 // Look for loop-invariant use 2688 if( u_loop == loop ) continue; 2689 if( loop->is_member( u_loop ) ) continue; 2690 // Check that use is live out the bottom. Assuming the trip-counter 2691 // update is right at the bottom, uses of of the loop middle are ok. 2692 if( dom_lca( exit, u_ctrl ) != exit ) continue; 2693 // protect against stride not being a constant 2694 if( !cle->stride_is_con() ) continue; 2695 // Hit! Refactor use to use the post-incremented tripcounter. 2696 // Compute a post-increment tripcounter. 2697 Node *opaq = new (C, 2) Opaque2Node( C, cle->incr() ); 2698 register_new_node( opaq, u_ctrl ); 2699 Node *neg_stride = _igvn.intcon(-cle->stride_con()); 2700 set_ctrl(neg_stride, C->root()); 2701 Node *post = new (C, 3) AddINode( opaq, neg_stride); 2702 register_new_node( post, u_ctrl ); 2703 _igvn.hash_delete(use); 2704 _igvn._worklist.push(use); 2705 for( uint j = 1; j < use->req(); j++ ) 2706 if( use->in(j) == phi ) 2707 use->set_req(j, post); 2708 // Since DU info changed, rerun loop 2709 progress = true; 2710 break; 2711 } 2712 } 2713 2714 }