1 /* 2 * Copyright (c) 1999, 2012, 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/matcher.hpp" 32 #include "opto/mulnode.hpp" 33 #include "opto/rootnode.hpp" 34 #include "opto/subnode.hpp" 35 36 //============================================================================= 37 //------------------------------split_thru_phi--------------------------------- 38 // Split Node 'n' through merge point if there is enough win. 39 Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) { 40 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) { 41 // ConvI2L may have type information on it which is unsafe to push up 42 // so disable this for now 43 return NULL; 44 } 45 int wins = 0; 46 assert(!n->is_CFG(), ""); 47 assert(region->is_Region(), ""); 48 49 const Type* type = n->bottom_type(); 50 const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr(); 51 Node *phi; 52 if (t_oop != NULL && t_oop->is_known_instance_field()) { 53 int iid = t_oop->instance_id(); 54 int index = C->get_alias_index(t_oop); 55 int offset = t_oop->offset(); 56 phi = new (C) PhiNode(region, type, NULL, iid, index, offset); 57 } else { 58 phi = PhiNode::make_blank(region, n); 59 } 60 uint old_unique = C->unique(); 61 for (uint i = 1; i < region->req(); i++) { 62 Node *x; 63 Node* the_clone = NULL; 64 if (region->in(i) == C->top()) { 65 x = C->top(); // Dead path? Use a dead data op 66 } else { 67 x = n->clone(); // Else clone up the data op 68 the_clone = x; // Remember for possible deletion. 69 // Alter data node to use pre-phi inputs 70 if (n->in(0) == region) 71 x->set_req( 0, region->in(i) ); 72 for (uint j = 1; j < n->req(); j++) { 73 Node *in = n->in(j); 74 if (in->is_Phi() && in->in(0) == region) 75 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone 76 } 77 } 78 // Check for a 'win' on some paths 79 const Type *t = x->Value(&_igvn); 80 81 bool singleton = t->singleton(); 82 83 // A TOP singleton indicates that there are no possible values incoming 84 // along a particular edge. In most cases, this is OK, and the Phi will 85 // be eliminated later in an Ideal call. However, we can't allow this to 86 // happen if the singleton occurs on loop entry, as the elimination of 87 // the PhiNode may cause the resulting node to migrate back to a previous 88 // loop iteration. 89 if (singleton && t == Type::TOP) { 90 // Is_Loop() == false does not confirm the absence of a loop (e.g., an 91 // irreducible loop may not be indicated by an affirmative is_Loop()); 92 // therefore, the only top we can split thru a phi is on a backedge of 93 // a loop. 94 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 95 } 96 97 if (singleton) { 98 wins++; 99 x = ((PhaseGVN&)_igvn).makecon(t); 100 } else { 101 // We now call Identity to try to simplify the cloned node. 102 // Note that some Identity methods call phase->type(this). 103 // Make sure that the type array is big enough for 104 // our new node, even though we may throw the node away. 105 // (Note: This tweaking with igvn only works because x is a new node.) 106 _igvn.set_type(x, t); 107 // If x is a TypeNode, capture any more-precise type permanently into Node 108 // otherwise it will be not updated during igvn->transform since 109 // igvn->type(x) is set to x->Value() already. 110 x->raise_bottom_type(t); 111 Node *y = x->Identity(&_igvn); 112 if (y != x) { 113 wins++; 114 x = y; 115 } else { 116 y = _igvn.hash_find(x); 117 if (y) { 118 wins++; 119 x = y; 120 } else { 121 // Else x is a new node we are keeping 122 // We do not need register_new_node_with_optimizer 123 // because set_type has already been called. 124 _igvn._worklist.push(x); 125 } 126 } 127 } 128 if (x != the_clone && the_clone != NULL) 129 _igvn.remove_dead_node(the_clone); 130 phi->set_req( i, x ); 131 } 132 // Too few wins? 133 if (wins <= policy) { 134 _igvn.remove_dead_node(phi); 135 return NULL; 136 } 137 138 // Record Phi 139 register_new_node( phi, region ); 140 141 for (uint i2 = 1; i2 < phi->req(); i2++) { 142 Node *x = phi->in(i2); 143 // If we commoned up the cloned 'x' with another existing Node, 144 // the existing Node picks up a new use. We need to make the 145 // existing Node occur higher up so it dominates its uses. 146 Node *old_ctrl; 147 IdealLoopTree *old_loop; 148 149 if (x->is_Con()) { 150 // Constant's control is always root. 151 set_ctrl(x, C->root()); 152 continue; 153 } 154 // The occasional new node 155 if (x->_idx >= old_unique) { // Found a new, unplaced node? 156 old_ctrl = NULL; 157 old_loop = NULL; // Not in any prior loop 158 } else { 159 old_ctrl = get_ctrl(x); 160 old_loop = get_loop(old_ctrl); // Get prior loop 161 } 162 // New late point must dominate new use 163 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2)); 164 if (new_ctrl == old_ctrl) // Nothing is changed 165 continue; 166 167 IdealLoopTree *new_loop = get_loop(new_ctrl); 168 169 // Don't move x into a loop if its uses are 170 // outside of loop. Otherwise x will be cloned 171 // for each use outside of this loop. 172 IdealLoopTree *use_loop = get_loop(region); 173 if (!new_loop->is_member(use_loop) && 174 (old_loop == NULL || !new_loop->is_member(old_loop))) { 175 // Take early control, later control will be recalculated 176 // during next iteration of loop optimizations. 177 new_ctrl = get_early_ctrl(x); 178 new_loop = get_loop(new_ctrl); 179 } 180 // Set new location 181 set_ctrl(x, new_ctrl); 182 // If changing loop bodies, see if we need to collect into new body 183 if (old_loop != new_loop) { 184 if (old_loop && !old_loop->_child) 185 old_loop->_body.yank(x); 186 if (!new_loop->_child) 187 new_loop->_body.push(x); // Collect body info 188 } 189 } 190 191 return phi; 192 } 193 194 //------------------------------dominated_by------------------------------------ 195 // Replace the dominated test with an obvious true or false. Place it on the 196 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the 197 // live path up to the dominating control. 198 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff, bool flip, bool exclude_loop_predicate ) { 199 #ifndef PRODUCT 200 if (VerifyLoopOptimizations && PrintOpto) tty->print_cr("dominating test"); 201 #endif 202 203 204 // prevdom is the dominating projection of the dominating test. 205 assert( iff->is_If(), "" ); 206 assert( iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added"); 207 int pop = prevdom->Opcode(); 208 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" ); 209 if (flip) { 210 if (pop == Op_IfTrue) 211 pop = Op_IfFalse; 212 else 213 pop = Op_IfTrue; 214 } 215 // 'con' is set to true or false to kill the dominated test. 216 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO); 217 set_ctrl(con, C->root()); // Constant gets a new use 218 // Hack the dominated test 219 _igvn.replace_input_of(iff, 1, con); 220 221 // If I dont have a reachable TRUE and FALSE path following the IfNode then 222 // I can assume this path reaches an infinite loop. In this case it's not 223 // important to optimize the data Nodes - either the whole compilation will 224 // be tossed or this path (and all data Nodes) will go dead. 225 if (iff->outcnt() != 2) return; 226 227 // Make control-dependent data Nodes on the live path (path that will remain 228 // once the dominated IF is removed) become control-dependent on the 229 // dominating projection. 230 Node* dp = iff->as_If()->proj_out(pop == Op_IfTrue); 231 232 // Loop predicates may have depending checks which should not 233 // be skipped. For example, range check predicate has two checks 234 // for lower and upper bounds. 235 if (dp == NULL) 236 return; 237 238 ProjNode* dp_proj = dp->as_Proj(); 239 ProjNode* unc_proj = iff->as_If()->proj_out(1 - dp_proj->_con)->as_Proj(); 240 if (exclude_loop_predicate && 241 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate)) 242 return; // Let IGVN transformation change control dependence. 243 244 IdealLoopTree *old_loop = get_loop(dp); 245 246 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { 247 Node* cd = dp->fast_out(i); // Control-dependent node 248 if (cd->depends_only_on_test()) { 249 assert(cd->in(0) == dp, ""); 250 _igvn.replace_input_of(cd, 0, prevdom); 251 set_early_ctrl(cd); 252 IdealLoopTree *new_loop = get_loop(get_ctrl(cd)); 253 if (old_loop != new_loop) { 254 if (!old_loop->_child) old_loop->_body.yank(cd); 255 if (!new_loop->_child) new_loop->_body.push(cd); 256 } 257 --i; 258 --imax; 259 } 260 } 261 } 262 263 //------------------------------has_local_phi_input---------------------------- 264 // Return TRUE if 'n' has Phi inputs from its local block and no other 265 // block-local inputs (all non-local-phi inputs come from earlier blocks) 266 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) { 267 Node *n_ctrl = get_ctrl(n); 268 // See if some inputs come from a Phi in this block, or from before 269 // this block. 270 uint i; 271 for( i = 1; i < n->req(); i++ ) { 272 Node *phi = n->in(i); 273 if( phi->is_Phi() && phi->in(0) == n_ctrl ) 274 break; 275 } 276 if( i >= n->req() ) 277 return NULL; // No Phi inputs; nowhere to clone thru 278 279 // Check for inputs created between 'n' and the Phi input. These 280 // must split as well; they have already been given the chance 281 // (courtesy of a post-order visit) and since they did not we must 282 // recover the 'cost' of splitting them by being very profitable 283 // when splitting 'n'. Since this is unlikely we simply give up. 284 for( i = 1; i < n->req(); i++ ) { 285 Node *m = n->in(i); 286 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) { 287 // We allow the special case of AddP's with no local inputs. 288 // This allows us to split-up address expressions. 289 if (m->is_AddP() && 290 get_ctrl(m->in(2)) != n_ctrl && 291 get_ctrl(m->in(3)) != n_ctrl) { 292 // Move the AddP up to dominating point 293 set_ctrl_and_loop(m, find_non_split_ctrl(idom(n_ctrl))); 294 continue; 295 } 296 return NULL; 297 } 298 } 299 300 return n_ctrl; 301 } 302 303 //------------------------------remix_address_expressions---------------------- 304 // Rework addressing expressions to get the most loop-invariant stuff 305 // moved out. We'd like to do all associative operators, but it's especially 306 // important (common) to do address expressions. 307 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) { 308 if (!has_ctrl(n)) return NULL; 309 Node *n_ctrl = get_ctrl(n); 310 IdealLoopTree *n_loop = get_loop(n_ctrl); 311 312 // See if 'n' mixes loop-varying and loop-invariant inputs and 313 // itself is loop-varying. 314 315 // Only interested in binary ops (and AddP) 316 if( n->req() < 3 || n->req() > 4 ) return NULL; 317 318 Node *n1_ctrl = get_ctrl(n->in( 1)); 319 Node *n2_ctrl = get_ctrl(n->in( 2)); 320 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3)); 321 IdealLoopTree *n1_loop = get_loop( n1_ctrl ); 322 IdealLoopTree *n2_loop = get_loop( n2_ctrl ); 323 IdealLoopTree *n3_loop = get_loop( n3_ctrl ); 324 325 // Does one of my inputs spin in a tighter loop than self? 326 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) || 327 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) || 328 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) ) 329 return NULL; // Leave well enough alone 330 331 // Is at least one of my inputs loop-invariant? 332 if( n1_loop == n_loop && 333 n2_loop == n_loop && 334 n3_loop == n_loop ) 335 return NULL; // No loop-invariant inputs 336 337 338 int n_op = n->Opcode(); 339 340 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2). 341 if( n_op == Op_LShiftI ) { 342 // Scale is loop invariant 343 Node *scale = n->in(2); 344 Node *scale_ctrl = get_ctrl(scale); 345 IdealLoopTree *scale_loop = get_loop(scale_ctrl ); 346 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) ) 347 return NULL; 348 const TypeInt *scale_t = scale->bottom_type()->isa_int(); 349 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 ) 350 return NULL; // Dont bother with byte/short masking 351 // Add must vary with loop (else shift would be loop-invariant) 352 Node *add = n->in(1); 353 Node *add_ctrl = get_ctrl(add); 354 IdealLoopTree *add_loop = get_loop(add_ctrl); 355 //assert( n_loop == add_loop, "" ); 356 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops 357 358 // Convert I-V into I+ (0-V); same for V-I 359 if( add->Opcode() == Op_SubI && 360 _igvn.type( add->in(1) ) != TypeInt::ZERO ) { 361 Node *zero = _igvn.intcon(0); 362 set_ctrl(zero, C->root()); 363 Node *neg = new (C) SubINode( _igvn.intcon(0), add->in(2) ); 364 register_new_node( neg, get_ctrl(add->in(2) ) ); 365 add = new (C) AddINode( add->in(1), neg ); 366 register_new_node( add, add_ctrl ); 367 } 368 if( add->Opcode() != Op_AddI ) return NULL; 369 // See if one add input is loop invariant 370 Node *add_var = add->in(1); 371 Node *add_var_ctrl = get_ctrl(add_var); 372 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 373 Node *add_invar = add->in(2); 374 Node *add_invar_ctrl = get_ctrl(add_invar); 375 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl ); 376 if( add_var_loop == n_loop ) { 377 } else if( add_invar_loop == n_loop ) { 378 // Swap to find the invariant part 379 add_invar = add_var; 380 add_invar_ctrl = add_var_ctrl; 381 add_invar_loop = add_var_loop; 382 add_var = add->in(2); 383 Node *add_var_ctrl = get_ctrl(add_var); 384 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 385 } else // Else neither input is loop invariant 386 return NULL; 387 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) ) 388 return NULL; // No invariant part of the add? 389 390 // Yes! Reshape address expression! 391 Node *inv_scale = new (C) LShiftINode( add_invar, scale ); 392 Node *inv_scale_ctrl = 393 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ? 394 add_invar_ctrl : scale_ctrl; 395 register_new_node( inv_scale, inv_scale_ctrl ); 396 Node *var_scale = new (C) LShiftINode( add_var, scale ); 397 register_new_node( var_scale, n_ctrl ); 398 Node *var_add = new (C) AddINode( var_scale, inv_scale ); 399 register_new_node( var_add, n_ctrl ); 400 _igvn.replace_node( n, var_add ); 401 return var_add; 402 } 403 404 // Replace (I+V) with (V+I) 405 if( n_op == Op_AddI || 406 n_op == Op_AddL || 407 n_op == Op_AddF || 408 n_op == Op_AddD || 409 n_op == Op_MulI || 410 n_op == Op_MulL || 411 n_op == Op_MulF || 412 n_op == Op_MulD ) { 413 if( n2_loop == n_loop ) { 414 assert( n1_loop != n_loop, "" ); 415 n->swap_edges(1, 2); 416 } 417 } 418 419 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V), 420 // but not if I2 is a constant. 421 if( n_op == Op_AddP ) { 422 if( n2_loop == n_loop && n3_loop != n_loop ) { 423 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) { 424 Node *n22_ctrl = get_ctrl(n->in(2)->in(2)); 425 Node *n23_ctrl = get_ctrl(n->in(2)->in(3)); 426 IdealLoopTree *n22loop = get_loop( n22_ctrl ); 427 IdealLoopTree *n23_loop = get_loop( n23_ctrl ); 428 if( n22loop != n_loop && n22loop->is_member(n_loop) && 429 n23_loop == n_loop ) { 430 Node *add1 = new (C) AddPNode( n->in(1), n->in(2)->in(2), n->in(3) ); 431 // Stuff new AddP in the loop preheader 432 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) ); 433 Node *add2 = new (C) AddPNode( n->in(1), add1, n->in(2)->in(3) ); 434 register_new_node( add2, n_ctrl ); 435 _igvn.replace_node( n, add2 ); 436 return add2; 437 } 438 } 439 } 440 441 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V) 442 if( n2_loop != n_loop && n3_loop == n_loop ) { 443 if( n->in(3)->Opcode() == Op_AddI ) { 444 Node *V = n->in(3)->in(1); 445 Node *I = n->in(3)->in(2); 446 if( is_member(n_loop,get_ctrl(V)) ) { 447 } else { 448 Node *tmp = V; V = I; I = tmp; 449 } 450 if( !is_member(n_loop,get_ctrl(I)) ) { 451 Node *add1 = new (C) AddPNode( n->in(1), n->in(2), I ); 452 // Stuff new AddP in the loop preheader 453 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) ); 454 Node *add2 = new (C) AddPNode( n->in(1), add1, V ); 455 register_new_node( add2, n_ctrl ); 456 _igvn.replace_node( n, add2 ); 457 return add2; 458 } 459 } 460 } 461 } 462 463 return NULL; 464 } 465 466 //------------------------------conditional_move------------------------------- 467 // Attempt to replace a Phi with a conditional move. We have some pretty 468 // strict profitability requirements. All Phis at the merge point must 469 // be converted, so we can remove the control flow. We need to limit the 470 // number of c-moves to a small handful. All code that was in the side-arms 471 // of the CFG diamond is now speculatively executed. This code has to be 472 // "cheap enough". We are pretty much limited to CFG diamonds that merge 473 // 1 or 2 items with a total of 1 or 2 ops executed speculatively. 474 Node *PhaseIdealLoop::conditional_move( Node *region ) { 475 476 assert(region->is_Region(), "sanity check"); 477 if (region->req() != 3) return NULL; 478 479 // Check for CFG diamond 480 Node *lp = region->in(1); 481 Node *rp = region->in(2); 482 if (!lp || !rp) return NULL; 483 Node *lp_c = lp->in(0); 484 if (lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If()) return NULL; 485 IfNode *iff = lp_c->as_If(); 486 487 // Check for ops pinned in an arm of the diamond. 488 // Can't remove the control flow in this case 489 if (lp->outcnt() > 1) return NULL; 490 if (rp->outcnt() > 1) return NULL; 491 492 IdealLoopTree* r_loop = get_loop(region); 493 assert(r_loop == get_loop(iff), "sanity"); 494 // Always convert to CMOVE if all results are used only outside this loop. 495 bool used_inside_loop = (r_loop == _ltree_root); 496 497 // Check profitability 498 int cost = 0; 499 int phis = 0; 500 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 501 Node *out = region->fast_out(i); 502 if (!out->is_Phi()) continue; // Ignore other control edges, etc 503 phis++; 504 PhiNode* phi = out->as_Phi(); 505 BasicType bt = phi->type()->basic_type(); 506 switch (bt) { 507 case T_FLOAT: 508 case T_DOUBLE: { 509 cost += Matcher::float_cmove_cost(); // Could be very expensive 510 break; 511 } 512 case T_LONG: { 513 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's 514 } 515 case T_INT: // These all CMOV fine 516 case T_ADDRESS: { // (RawPtr) 517 cost++; 518 break; 519 } 520 case T_NARROWOOP: // Fall through 521 case T_OBJECT: { // Base oops are OK, but not derived oops 522 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr(); 523 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a 524 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus 525 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we 526 // have a Phi for the base here that we convert to a CMOVE all is well 527 // and good. But if the base is dead, we'll not make a CMOVE. Later 528 // the allocator will have to produce a base by creating a CMOVE of the 529 // relevant bases. This puts the allocator in the business of 530 // manufacturing expensive instructions, generally a bad plan. 531 // Just Say No to Conditionally-Moved Derived Pointers. 532 if (tp && tp->offset() != 0) 533 return NULL; 534 cost++; 535 break; 536 } 537 default: 538 return NULL; // In particular, can't do memory or I/O 539 } 540 // Add in cost any speculative ops 541 for (uint j = 1; j < region->req(); j++) { 542 Node *proj = region->in(j); 543 Node *inp = phi->in(j); 544 if (get_ctrl(inp) == proj) { // Found local op 545 cost++; 546 // Check for a chain of dependent ops; these will all become 547 // speculative in a CMOV. 548 for (uint k = 1; k < inp->req(); k++) 549 if (get_ctrl(inp->in(k)) == proj) 550 cost += ConditionalMoveLimit; // Too much speculative goo 551 } 552 } 553 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode. 554 // This will likely Split-If, a higher-payoff operation. 555 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) { 556 Node* use = phi->fast_out(k); 557 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr()) 558 cost += ConditionalMoveLimit; 559 // Is there a use inside the loop? 560 // Note: check only basic types since CMoveP is pinned. 561 if (!used_inside_loop && is_java_primitive(bt)) { 562 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use); 563 if (r_loop == u_loop || r_loop->is_member(u_loop)) { 564 used_inside_loop = true; 565 } 566 } 567 } 568 } 569 Node* bol = iff->in(1); 570 assert(bol->Opcode() == Op_Bool, ""); 571 int cmp_op = bol->in(1)->Opcode(); 572 // It is expensive to generate flags from a float compare. 573 // Avoid duplicated float compare. 574 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL; 575 576 float infrequent_prob = PROB_UNLIKELY_MAG(3); 577 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop. 578 if (used_inside_loop) { 579 if (cost >= ConditionalMoveLimit) return NULL; // Too much goo 580 581 // BlockLayoutByFrequency optimization moves infrequent branch 582 // from hot path. No point in CMOV'ing in such case (110 is used 583 // instead of 100 to take into account not exactness of float value). 584 if (BlockLayoutByFrequency) { 585 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f); 586 } 587 } 588 // Check for highly predictable branch. No point in CMOV'ing if 589 // we are going to predict accurately all the time. 590 if (iff->_prob < infrequent_prob || 591 iff->_prob > (1.0f - infrequent_prob)) 592 return NULL; 593 594 // -------------- 595 // Now replace all Phis with CMOV's 596 Node *cmov_ctrl = iff->in(0); 597 uint flip = (lp->Opcode() == Op_IfTrue); 598 while (1) { 599 PhiNode* phi = NULL; 600 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 601 Node *out = region->fast_out(i); 602 if (out->is_Phi()) { 603 phi = out->as_Phi(); 604 break; 605 } 606 } 607 if (phi == NULL) break; 608 #ifndef PRODUCT 609 if (PrintOpto && VerifyLoopOptimizations) tty->print_cr("CMOV"); 610 #endif 611 // Move speculative ops 612 for (uint j = 1; j < region->req(); j++) { 613 Node *proj = region->in(j); 614 Node *inp = phi->in(j); 615 if (get_ctrl(inp) == proj) { // Found local op 616 #ifndef PRODUCT 617 if (PrintOpto && VerifyLoopOptimizations) { 618 tty->print(" speculate: "); 619 inp->dump(); 620 } 621 #endif 622 set_ctrl(inp, cmov_ctrl); 623 } 624 } 625 Node *cmov = CMoveNode::make( C, cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi) ); 626 register_new_node( cmov, cmov_ctrl ); 627 _igvn.replace_node( phi, cmov ); 628 #ifndef PRODUCT 629 if (TraceLoopOpts) { 630 tty->print("CMOV "); 631 r_loop->dump_head(); 632 if (Verbose) { 633 bol->in(1)->dump(1); 634 cmov->dump(1); 635 } 636 } 637 if (VerifyLoopOptimizations) verify(); 638 #endif 639 } 640 641 // The useless CFG diamond will fold up later; see the optimization in 642 // RegionNode::Ideal. 643 _igvn._worklist.push(region); 644 645 return iff->in(1); 646 } 647 648 //------------------------------split_if_with_blocks_pre----------------------- 649 // Do the real work in a non-recursive function. Data nodes want to be 650 // cloned in the pre-order so they can feed each other nicely. 651 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) { 652 // Cloning these guys is unlikely to win 653 int n_op = n->Opcode(); 654 if( n_op == Op_MergeMem ) return n; 655 if( n->is_Proj() ) return n; 656 // Do not clone-up CmpFXXX variations, as these are always 657 // followed by a CmpI 658 if( n->is_Cmp() ) return n; 659 // Attempt to use a conditional move instead of a phi/branch 660 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) { 661 Node *cmov = conditional_move( n ); 662 if( cmov ) return cmov; 663 } 664 if( n->is_CFG() || n->is_LoadStore() ) 665 return n; 666 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd 667 n_op == Op_Opaque2 ) { 668 if( !C->major_progress() ) // If chance of no more loop opts... 669 _igvn._worklist.push(n); // maybe we'll remove them 670 return n; 671 } 672 673 if( n->is_Con() ) return n; // No cloning for Con nodes 674 675 Node *n_ctrl = get_ctrl(n); 676 if( !n_ctrl ) return n; // Dead node 677 678 // Attempt to remix address expressions for loop invariants 679 Node *m = remix_address_expressions( n ); 680 if( m ) return m; 681 682 // Determine if the Node has inputs from some local Phi. 683 // Returns the block to clone thru. 684 Node *n_blk = has_local_phi_input( n ); 685 if( !n_blk ) return n; 686 // Do not clone the trip counter through on a CountedLoop 687 // (messes up the canonical shape). 688 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n; 689 690 // Check for having no control input; not pinned. Allow 691 // dominating control. 692 if( n->in(0) ) { 693 Node *dom = idom(n_blk); 694 if( dom_lca( n->in(0), dom ) != n->in(0) ) 695 return n; 696 } 697 // Policy: when is it profitable. You must get more wins than 698 // policy before it is considered profitable. Policy is usually 0, 699 // so 1 win is considered profitable. Big merges will require big 700 // cloning, so get a larger policy. 701 int policy = n_blk->req() >> 2; 702 703 // If the loop is a candidate for range check elimination, 704 // delay splitting through it's phi until a later loop optimization 705 if (n_blk->is_CountedLoop()) { 706 IdealLoopTree *lp = get_loop(n_blk); 707 if (lp && lp->_rce_candidate) { 708 return n; 709 } 710 } 711 712 // Use same limit as split_if_with_blocks_post 713 if( C->unique() > 35000 ) return n; // Method too big 714 715 // Split 'n' through the merge point if it is profitable 716 Node *phi = split_thru_phi( n, n_blk, policy ); 717 if (!phi) return n; 718 719 // Found a Phi to split thru! 720 // Replace 'n' with the new phi 721 _igvn.replace_node( n, phi ); 722 // Moved a load around the loop, 'en-registering' something. 723 if (n_blk->is_Loop() && n->is_Load() && 724 !phi->in(LoopNode::LoopBackControl)->is_Load()) 725 C->set_major_progress(); 726 727 return phi; 728 } 729 730 static bool merge_point_too_heavy(Compile* C, Node* region) { 731 // Bail out if the region and its phis have too many users. 732 int weight = 0; 733 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 734 weight += region->fast_out(i)->outcnt(); 735 } 736 int nodes_left = MaxNodeLimit - C->live_nodes(); 737 if (weight * 8 > nodes_left) { 738 #ifndef PRODUCT 739 if (PrintOpto) 740 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight); 741 #endif 742 return true; 743 } else { 744 return false; 745 } 746 } 747 748 static bool merge_point_safe(Node* region) { 749 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode 750 // having a PhiNode input. This sidesteps the dangerous case where the split 751 // ConvI2LNode may become TOP if the input Value() does not 752 // overlap the ConvI2L range, leaving a node which may not dominate its 753 // uses. 754 // A better fix for this problem can be found in the BugTraq entry, but 755 // expediency for Mantis demands this hack. 756 // 6855164: If the merge point has a FastLockNode with a PhiNode input, we stop 757 // split_if_with_blocks from splitting a block because we could not move around 758 // the FastLockNode. 759 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 760 Node* n = region->fast_out(i); 761 if (n->is_Phi()) { 762 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 763 Node* m = n->fast_out(j); 764 if (m->is_FastLock()) 765 return false; 766 #ifdef _LP64 767 if (m->Opcode() == Op_ConvI2L) 768 return false; 769 #endif 770 } 771 } 772 } 773 return true; 774 } 775 776 777 //------------------------------place_near_use--------------------------------- 778 // Place some computation next to use but not inside inner loops. 779 // For inner loop uses move it to the preheader area. 780 Node *PhaseIdealLoop::place_near_use( Node *useblock ) const { 781 IdealLoopTree *u_loop = get_loop( useblock ); 782 return (u_loop->_irreducible || u_loop->_child) 783 ? useblock 784 : u_loop->_head->in(LoopNode::EntryControl); 785 } 786 787 788 //------------------------------split_if_with_blocks_post---------------------- 789 // Do the real work in a non-recursive function. CFG hackery wants to be 790 // in the post-order, so it can dirty the I-DOM info and not use the dirtied 791 // info. 792 void PhaseIdealLoop::split_if_with_blocks_post( Node *n ) { 793 794 // Cloning Cmp through Phi's involves the split-if transform. 795 // FastLock is not used by an If 796 if( n->is_Cmp() && !n->is_FastLock() ) { 797 if( C->unique() > 35000 ) return; // Method too big 798 799 // Do not do 'split-if' if irreducible loops are present. 800 if( _has_irreducible_loops ) 801 return; 802 803 Node *n_ctrl = get_ctrl(n); 804 // Determine if the Node has inputs from some local Phi. 805 // Returns the block to clone thru. 806 Node *n_blk = has_local_phi_input( n ); 807 if( n_blk != n_ctrl ) return; 808 809 if( merge_point_too_heavy(C, n_ctrl) ) 810 return; 811 812 if( n->outcnt() != 1 ) return; // Multiple bool's from 1 compare? 813 Node *bol = n->unique_out(); 814 assert( bol->is_Bool(), "expect a bool here" ); 815 if( bol->outcnt() != 1 ) return;// Multiple branches from 1 compare? 816 Node *iff = bol->unique_out(); 817 818 // Check some safety conditions 819 if( iff->is_If() ) { // Classic split-if? 820 if( iff->in(0) != n_ctrl ) return; // Compare must be in same blk as if 821 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE 822 // Can't split CMove with different control edge. 823 if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) return; 824 if( get_ctrl(iff->in(2)) == n_ctrl || 825 get_ctrl(iff->in(3)) == n_ctrl ) 826 return; // Inputs not yet split-up 827 if ( get_loop(n_ctrl) != get_loop(get_ctrl(iff)) ) { 828 return; // Loop-invar test gates loop-varying CMOVE 829 } 830 } else { 831 return; // some other kind of node, such as an Allocate 832 } 833 834 // Do not do 'split-if' if some paths are dead. First do dead code 835 // elimination and then see if its still profitable. 836 for( uint i = 1; i < n_ctrl->req(); i++ ) 837 if( n_ctrl->in(i) == C->top() ) 838 return; 839 840 // When is split-if profitable? Every 'win' on means some control flow 841 // goes dead, so it's almost always a win. 842 int policy = 0; 843 // If trying to do a 'Split-If' at the loop head, it is only 844 // profitable if the cmp folds up on BOTH paths. Otherwise we 845 // risk peeling a loop forever. 846 847 // CNC - Disabled for now. Requires careful handling of loop 848 // body selection for the cloned code. Also, make sure we check 849 // for any input path not being in the same loop as n_ctrl. For 850 // irreducible loops we cannot check for 'n_ctrl->is_Loop()' 851 // because the alternative loop entry points won't be converted 852 // into LoopNodes. 853 IdealLoopTree *n_loop = get_loop(n_ctrl); 854 for( uint j = 1; j < n_ctrl->req(); j++ ) 855 if( get_loop(n_ctrl->in(j)) != n_loop ) 856 return; 857 858 // Check for safety of the merge point. 859 if( !merge_point_safe(n_ctrl) ) { 860 return; 861 } 862 863 // Split compare 'n' through the merge point if it is profitable 864 Node *phi = split_thru_phi( n, n_ctrl, policy ); 865 if( !phi ) return; 866 867 // Found a Phi to split thru! 868 // Replace 'n' with the new phi 869 _igvn.replace_node( n, phi ); 870 871 // Now split the bool up thru the phi 872 Node *bolphi = split_thru_phi( bol, n_ctrl, -1 ); 873 guarantee(bolphi != NULL, "null boolean phi node"); 874 875 _igvn.replace_node( bol, bolphi ); 876 assert( iff->in(1) == bolphi, "" ); 877 878 if( bolphi->Value(&_igvn)->singleton() ) 879 return; 880 881 // Conditional-move? Must split up now 882 if( !iff->is_If() ) { 883 Node *cmovphi = split_thru_phi( iff, n_ctrl, -1 ); 884 _igvn.replace_node( iff, cmovphi ); 885 return; 886 } 887 888 // Now split the IF 889 do_split_if( iff ); 890 return; 891 } 892 893 // Check for an IF ready to split; one that has its 894 // condition codes input coming from a Phi at the block start. 895 int n_op = n->Opcode(); 896 897 // Check for an IF being dominated by another IF same test 898 if( n_op == Op_If ) { 899 Node *bol = n->in(1); 900 uint max = bol->outcnt(); 901 // Check for same test used more than once? 902 if( n_op == Op_If && max > 1 && bol->is_Bool() ) { 903 // Search up IDOMs to see if this IF is dominated. 904 Node *cutoff = get_ctrl(bol); 905 906 // Now search up IDOMs till cutoff, looking for a dominating test 907 Node *prevdom = n; 908 Node *dom = idom(prevdom); 909 while( dom != cutoff ) { 910 if( dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom ) { 911 // Replace the dominated test with an obvious true or false. 912 // Place it on the IGVN worklist for later cleanup. 913 C->set_major_progress(); 914 dominated_by( prevdom, n, false, true ); 915 #ifndef PRODUCT 916 if( VerifyLoopOptimizations ) verify(); 917 #endif 918 return; 919 } 920 prevdom = dom; 921 dom = idom(prevdom); 922 } 923 } 924 } 925 926 // See if a shared loop-varying computation has no loop-varying uses. 927 // Happens if something is only used for JVM state in uncommon trap exits, 928 // like various versions of induction variable+offset. Clone the 929 // computation per usage to allow it to sink out of the loop. 930 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about) 931 Node *n_ctrl = get_ctrl(n); 932 IdealLoopTree *n_loop = get_loop(n_ctrl); 933 if( n_loop != _ltree_root ) { 934 DUIterator_Fast imax, i = n->fast_outs(imax); 935 for (; i < imax; i++) { 936 Node* u = n->fast_out(i); 937 if( !has_ctrl(u) ) break; // Found control user 938 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 939 if( u_loop == n_loop ) break; // Found loop-varying use 940 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop 941 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003 942 } 943 bool did_break = (i < imax); // Did we break out of the previous loop? 944 if (!did_break && n->outcnt() > 1) { // All uses in outer loops! 945 Node *late_load_ctrl = NULL; 946 if (n->is_Load()) { 947 // If n is a load, get and save the result from get_late_ctrl(), 948 // to be later used in calculating the control for n's clones. 949 clear_dom_lca_tags(); 950 late_load_ctrl = get_late_ctrl(n, n_ctrl); 951 } 952 // If n is a load, and the late control is the same as the current 953 // control, then the cloning of n is a pointless exercise, because 954 // GVN will ensure that we end up where we started. 955 if (!n->is_Load() || late_load_ctrl != n_ctrl) { 956 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) { 957 Node *u = n->last_out(j); // Clone private computation per use 958 _igvn.rehash_node_delayed(u); 959 Node *x = n->clone(); // Clone computation 960 Node *x_ctrl = NULL; 961 if( u->is_Phi() ) { 962 // Replace all uses of normal nodes. Replace Phi uses 963 // individually, so the separate Nodes can sink down 964 // different paths. 965 uint k = 1; 966 while( u->in(k) != n ) k++; 967 u->set_req( k, x ); 968 // x goes next to Phi input path 969 x_ctrl = u->in(0)->in(k); 970 --j; 971 } else { // Normal use 972 // Replace all uses 973 for( uint k = 0; k < u->req(); k++ ) { 974 if( u->in(k) == n ) { 975 u->set_req( k, x ); 976 --j; 977 } 978 } 979 x_ctrl = get_ctrl(u); 980 } 981 982 // Find control for 'x' next to use but not inside inner loops. 983 // For inner loop uses get the preheader area. 984 x_ctrl = place_near_use(x_ctrl); 985 986 if (n->is_Load()) { 987 // For loads, add a control edge to a CFG node outside of the loop 988 // to force them to not combine and return back inside the loop 989 // during GVN optimization (4641526). 990 // 991 // Because we are setting the actual control input, factor in 992 // the result from get_late_ctrl() so we respect any 993 // anti-dependences. (6233005). 994 x_ctrl = dom_lca(late_load_ctrl, x_ctrl); 995 996 // Don't allow the control input to be a CFG splitting node. 997 // Such nodes should only have ProjNodes as outs, e.g. IfNode 998 // should only have IfTrueNode and IfFalseNode (4985384). 999 x_ctrl = find_non_split_ctrl(x_ctrl); 1000 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone"); 1001 1002 x->set_req(0, x_ctrl); 1003 } 1004 register_new_node(x, x_ctrl); 1005 1006 // Some institutional knowledge is needed here: 'x' is 1007 // yanked because if the optimizer runs GVN on it all the 1008 // cloned x's will common up and undo this optimization and 1009 // be forced back in the loop. This is annoying because it 1010 // makes +VerifyOpto report false-positives on progress. I 1011 // tried setting control edges on the x's to force them to 1012 // not combine, but the matching gets worried when it tries 1013 // to fold a StoreP and an AddP together (as part of an 1014 // address expression) and the AddP and StoreP have 1015 // different controls. 1016 if (!x->is_Load() && !x->is_DecodeNarrowPtr()) _igvn._worklist.yank(x); 1017 } 1018 _igvn.remove_dead_node(n); 1019 } 1020 } 1021 } 1022 } 1023 1024 // Check for Opaque2's who's loop has disappeared - who's input is in the 1025 // same loop nest as their output. Remove 'em, they are no longer useful. 1026 if( n_op == Op_Opaque2 && 1027 n->in(1) != NULL && 1028 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) { 1029 _igvn.replace_node( n, n->in(1) ); 1030 } 1031 } 1032 1033 //------------------------------split_if_with_blocks--------------------------- 1034 // Check for aggressive application of 'split-if' optimization, 1035 // using basic block level info. 1036 void PhaseIdealLoop::split_if_with_blocks( VectorSet &visited, Node_Stack &nstack ) { 1037 Node *n = C->root(); 1038 visited.set(n->_idx); // first, mark node as visited 1039 // Do pre-visit work for root 1040 n = split_if_with_blocks_pre( n ); 1041 uint cnt = n->outcnt(); 1042 uint i = 0; 1043 while (true) { 1044 // Visit all children 1045 if (i < cnt) { 1046 Node* use = n->raw_out(i); 1047 ++i; 1048 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) { 1049 // Now do pre-visit work for this use 1050 use = split_if_with_blocks_pre( use ); 1051 nstack.push(n, i); // Save parent and next use's index. 1052 n = use; // Process all children of current use. 1053 cnt = use->outcnt(); 1054 i = 0; 1055 } 1056 } 1057 else { 1058 // All of n's children have been processed, complete post-processing. 1059 if (cnt != 0 && !n->is_Con()) { 1060 assert(has_node(n), "no dead nodes"); 1061 split_if_with_blocks_post( n ); 1062 } 1063 if (nstack.is_empty()) { 1064 // Finished all nodes on stack. 1065 break; 1066 } 1067 // Get saved parent node and next use's index. Visit the rest of uses. 1068 n = nstack.node(); 1069 cnt = n->outcnt(); 1070 i = nstack.index(); 1071 nstack.pop(); 1072 } 1073 } 1074 } 1075 1076 1077 //============================================================================= 1078 // 1079 // C L O N E A L O O P B O D Y 1080 // 1081 1082 //------------------------------clone_iff-------------------------------------- 1083 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1084 // "Nearly" because all Nodes have been cloned from the original in the loop, 1085 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1086 // through the Phi recursively, and return a Bool. 1087 BoolNode *PhaseIdealLoop::clone_iff( PhiNode *phi, IdealLoopTree *loop ) { 1088 1089 // Convert this Phi into a Phi merging Bools 1090 uint i; 1091 for( i = 1; i < phi->req(); i++ ) { 1092 Node *b = phi->in(i); 1093 if( b->is_Phi() ) { 1094 _igvn.replace_input_of(phi, i, clone_iff( b->as_Phi(), loop )); 1095 } else { 1096 assert( b->is_Bool(), "" ); 1097 } 1098 } 1099 1100 Node *sample_bool = phi->in(1); 1101 Node *sample_cmp = sample_bool->in(1); 1102 1103 // Make Phis to merge the Cmp's inputs. 1104 PhiNode *phi1 = new (C) PhiNode( phi->in(0), Type::TOP ); 1105 PhiNode *phi2 = new (C) PhiNode( phi->in(0), Type::TOP ); 1106 for( i = 1; i < phi->req(); i++ ) { 1107 Node *n1 = phi->in(i)->in(1)->in(1); 1108 Node *n2 = phi->in(i)->in(1)->in(2); 1109 phi1->set_req( i, n1 ); 1110 phi2->set_req( i, n2 ); 1111 phi1->set_type( phi1->type()->meet(n1->bottom_type()) ); 1112 phi2->set_type( phi2->type()->meet(n2->bottom_type()) ); 1113 } 1114 // See if these Phis have been made before. 1115 // Register with optimizer 1116 Node *hit1 = _igvn.hash_find_insert(phi1); 1117 if( hit1 ) { // Hit, toss just made Phi 1118 _igvn.remove_dead_node(phi1); // Remove new phi 1119 assert( hit1->is_Phi(), "" ); 1120 phi1 = (PhiNode*)hit1; // Use existing phi 1121 } else { // Miss 1122 _igvn.register_new_node_with_optimizer(phi1); 1123 } 1124 Node *hit2 = _igvn.hash_find_insert(phi2); 1125 if( hit2 ) { // Hit, toss just made Phi 1126 _igvn.remove_dead_node(phi2); // Remove new phi 1127 assert( hit2->is_Phi(), "" ); 1128 phi2 = (PhiNode*)hit2; // Use existing phi 1129 } else { // Miss 1130 _igvn.register_new_node_with_optimizer(phi2); 1131 } 1132 // Register Phis with loop/block info 1133 set_ctrl(phi1, phi->in(0)); 1134 set_ctrl(phi2, phi->in(0)); 1135 // Make a new Cmp 1136 Node *cmp = sample_cmp->clone(); 1137 cmp->set_req( 1, phi1 ); 1138 cmp->set_req( 2, phi2 ); 1139 _igvn.register_new_node_with_optimizer(cmp); 1140 set_ctrl(cmp, phi->in(0)); 1141 1142 // Make a new Bool 1143 Node *b = sample_bool->clone(); 1144 b->set_req(1,cmp); 1145 _igvn.register_new_node_with_optimizer(b); 1146 set_ctrl(b, phi->in(0)); 1147 1148 assert( b->is_Bool(), "" ); 1149 return (BoolNode*)b; 1150 } 1151 1152 //------------------------------clone_bool------------------------------------- 1153 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1154 // "Nearly" because all Nodes have been cloned from the original in the loop, 1155 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1156 // through the Phi recursively, and return a Bool. 1157 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) { 1158 uint i; 1159 // Convert this Phi into a Phi merging Bools 1160 for( i = 1; i < phi->req(); i++ ) { 1161 Node *b = phi->in(i); 1162 if( b->is_Phi() ) { 1163 _igvn.replace_input_of(phi, i, clone_bool( b->as_Phi(), loop )); 1164 } else { 1165 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" ); 1166 } 1167 } 1168 1169 Node *sample_cmp = phi->in(1); 1170 1171 // Make Phis to merge the Cmp's inputs. 1172 PhiNode *phi1 = new (C) PhiNode( phi->in(0), Type::TOP ); 1173 PhiNode *phi2 = new (C) PhiNode( phi->in(0), Type::TOP ); 1174 for( uint j = 1; j < phi->req(); j++ ) { 1175 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP 1176 Node *n1, *n2; 1177 if( cmp_top->is_Cmp() ) { 1178 n1 = cmp_top->in(1); 1179 n2 = cmp_top->in(2); 1180 } else { 1181 n1 = n2 = cmp_top; 1182 } 1183 phi1->set_req( j, n1 ); 1184 phi2->set_req( j, n2 ); 1185 phi1->set_type( phi1->type()->meet(n1->bottom_type()) ); 1186 phi2->set_type( phi2->type()->meet(n2->bottom_type()) ); 1187 } 1188 1189 // See if these Phis have been made before. 1190 // Register with optimizer 1191 Node *hit1 = _igvn.hash_find_insert(phi1); 1192 if( hit1 ) { // Hit, toss just made Phi 1193 _igvn.remove_dead_node(phi1); // Remove new phi 1194 assert( hit1->is_Phi(), "" ); 1195 phi1 = (PhiNode*)hit1; // Use existing phi 1196 } else { // Miss 1197 _igvn.register_new_node_with_optimizer(phi1); 1198 } 1199 Node *hit2 = _igvn.hash_find_insert(phi2); 1200 if( hit2 ) { // Hit, toss just made Phi 1201 _igvn.remove_dead_node(phi2); // Remove new phi 1202 assert( hit2->is_Phi(), "" ); 1203 phi2 = (PhiNode*)hit2; // Use existing phi 1204 } else { // Miss 1205 _igvn.register_new_node_with_optimizer(phi2); 1206 } 1207 // Register Phis with loop/block info 1208 set_ctrl(phi1, phi->in(0)); 1209 set_ctrl(phi2, phi->in(0)); 1210 // Make a new Cmp 1211 Node *cmp = sample_cmp->clone(); 1212 cmp->set_req( 1, phi1 ); 1213 cmp->set_req( 2, phi2 ); 1214 _igvn.register_new_node_with_optimizer(cmp); 1215 set_ctrl(cmp, phi->in(0)); 1216 1217 assert( cmp->is_Cmp(), "" ); 1218 return (CmpNode*)cmp; 1219 } 1220 1221 //------------------------------sink_use--------------------------------------- 1222 // If 'use' was in the loop-exit block, it now needs to be sunk 1223 // below the post-loop merge point. 1224 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) { 1225 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) { 1226 set_ctrl(use, post_loop); 1227 for (DUIterator j = use->outs(); use->has_out(j); j++) 1228 sink_use(use->out(j), post_loop); 1229 } 1230 } 1231 1232 //------------------------------clone_loop------------------------------------- 1233 // 1234 // C L O N E A L O O P B O D Y 1235 // 1236 // This is the basic building block of the loop optimizations. It clones an 1237 // entire loop body. It makes an old_new loop body mapping; with this mapping 1238 // you can find the new-loop equivalent to an old-loop node. All new-loop 1239 // nodes are exactly equal to their old-loop counterparts, all edges are the 1240 // same. All exits from the old-loop now have a RegionNode that merges the 1241 // equivalent new-loop path. This is true even for the normal "loop-exit" 1242 // condition. All uses of loop-invariant old-loop values now come from (one 1243 // or more) Phis that merge their new-loop equivalents. 1244 // 1245 // This operation leaves the graph in an illegal state: there are two valid 1246 // control edges coming from the loop pre-header to both loop bodies. I'll 1247 // definitely have to hack the graph after running this transform. 1248 // 1249 // From this building block I will further edit edges to perform loop peeling 1250 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc. 1251 // 1252 // Parameter side_by_size_idom: 1253 // When side_by_size_idom is NULL, the dominator tree is constructed for 1254 // the clone loop to dominate the original. Used in construction of 1255 // pre-main-post loop sequence. 1256 // When nonnull, the clone and original are side-by-side, both are 1257 // dominated by the side_by_side_idom node. Used in construction of 1258 // unswitched loops. 1259 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd, 1260 Node* side_by_side_idom) { 1261 1262 // Step 1: Clone the loop body. Make the old->new mapping. 1263 uint i; 1264 for( i = 0; i < loop->_body.size(); i++ ) { 1265 Node *old = loop->_body.at(i); 1266 Node *nnn = old->clone(); 1267 old_new.map( old->_idx, nnn ); 1268 _igvn.register_new_node_with_optimizer(nnn); 1269 } 1270 1271 1272 // Step 2: Fix the edges in the new body. If the old input is outside the 1273 // loop use it. If the old input is INside the loop, use the corresponding 1274 // new node instead. 1275 for( i = 0; i < loop->_body.size(); i++ ) { 1276 Node *old = loop->_body.at(i); 1277 Node *nnn = old_new[old->_idx]; 1278 // Fix CFG/Loop controlling the new node 1279 if (has_ctrl(old)) { 1280 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]); 1281 } else { 1282 set_loop(nnn, loop->_parent); 1283 if (old->outcnt() > 0) { 1284 set_idom( nnn, old_new[idom(old)->_idx], dd ); 1285 } 1286 } 1287 // Correct edges to the new node 1288 for( uint j = 0; j < nnn->req(); j++ ) { 1289 Node *n = nnn->in(j); 1290 if( n ) { 1291 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n ); 1292 if( loop->is_member( old_in_loop ) ) 1293 nnn->set_req(j, old_new[n->_idx]); 1294 } 1295 } 1296 _igvn.hash_find_insert(nnn); 1297 } 1298 Node *newhead = old_new[loop->_head->_idx]; 1299 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 1300 1301 1302 // Step 3: Now fix control uses. Loop varying control uses have already 1303 // been fixed up (as part of all input edges in Step 2). Loop invariant 1304 // control uses must be either an IfFalse or an IfTrue. Make a merge 1305 // point to merge the old and new IfFalse/IfTrue nodes; make the use 1306 // refer to this. 1307 ResourceArea *area = Thread::current()->resource_area(); 1308 Node_List worklist(area); 1309 uint new_counter = C->unique(); 1310 for( i = 0; i < loop->_body.size(); i++ ) { 1311 Node* old = loop->_body.at(i); 1312 if( !old->is_CFG() ) continue; 1313 Node* nnn = old_new[old->_idx]; 1314 1315 // Copy uses to a worklist, so I can munge the def-use info 1316 // with impunity. 1317 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1318 worklist.push(old->fast_out(j)); 1319 1320 while( worklist.size() ) { // Visit all uses 1321 Node *use = worklist.pop(); 1322 if (!has_node(use)) continue; // Ignore dead nodes 1323 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1324 if( !loop->is_member( use_loop ) && use->is_CFG() ) { 1325 // Both OLD and USE are CFG nodes here. 1326 assert( use->is_Proj(), "" ); 1327 1328 // Clone the loop exit control projection 1329 Node *newuse = use->clone(); 1330 newuse->set_req(0,nnn); 1331 _igvn.register_new_node_with_optimizer(newuse); 1332 set_loop(newuse, use_loop); 1333 set_idom(newuse, nnn, dom_depth(nnn) + 1 ); 1334 1335 // We need a Region to merge the exit from the peeled body and the 1336 // exit from the old loop body. 1337 RegionNode *r = new (C) RegionNode(3); 1338 // Map the old use to the new merge point 1339 old_new.map( use->_idx, r ); 1340 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use)); 1341 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" ); 1342 1343 // The original user of 'use' uses 'r' instead. 1344 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) { 1345 Node* useuse = use->last_out(l); 1346 _igvn.rehash_node_delayed(useuse); 1347 uint uses_found = 0; 1348 if( useuse->in(0) == use ) { 1349 useuse->set_req(0, r); 1350 uses_found++; 1351 if( useuse->is_CFG() ) { 1352 assert( dom_depth(useuse) > dd_r, "" ); 1353 set_idom(useuse, r, dom_depth(useuse)); 1354 } 1355 } 1356 for( uint k = 1; k < useuse->req(); k++ ) { 1357 if( useuse->in(k) == use ) { 1358 useuse->set_req(k, r); 1359 uses_found++; 1360 } 1361 } 1362 l -= uses_found; // we deleted 1 or more copies of this edge 1363 } 1364 1365 // Now finish up 'r' 1366 r->set_req( 1, newuse ); 1367 r->set_req( 2, use ); 1368 _igvn.register_new_node_with_optimizer(r); 1369 set_loop(r, use_loop); 1370 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r); 1371 } // End of if a loop-exit test 1372 } 1373 } 1374 1375 // Step 4: If loop-invariant use is not control, it must be dominated by a 1376 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region 1377 // there if needed. Make a Phi there merging old and new used values. 1378 Node_List *split_if_set = NULL; 1379 Node_List *split_bool_set = NULL; 1380 Node_List *split_cex_set = NULL; 1381 for( i = 0; i < loop->_body.size(); i++ ) { 1382 Node* old = loop->_body.at(i); 1383 Node* nnn = old_new[old->_idx]; 1384 // Copy uses to a worklist, so I can munge the def-use info 1385 // with impunity. 1386 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1387 worklist.push(old->fast_out(j)); 1388 1389 while( worklist.size() ) { 1390 Node *use = worklist.pop(); 1391 if (!has_node(use)) continue; // Ignore dead nodes 1392 if (use->in(0) == C->top()) continue; 1393 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1394 // Check for data-use outside of loop - at least one of OLD or USE 1395 // must not be a CFG node. 1396 if( !loop->is_member( use_loop ) && (!old->is_CFG() || !use->is_CFG())) { 1397 1398 // If the Data use is an IF, that means we have an IF outside of the 1399 // loop that is switching on a condition that is set inside of the 1400 // loop. Happens if people set a loop-exit flag; then test the flag 1401 // in the loop to break the loop, then test is again outside of the 1402 // loop to determine which way the loop exited. 1403 if( use->is_If() || use->is_CMove() ) { 1404 // Since this code is highly unlikely, we lazily build the worklist 1405 // of such Nodes to go split. 1406 if( !split_if_set ) 1407 split_if_set = new Node_List(area); 1408 split_if_set->push(use); 1409 } 1410 if( use->is_Bool() ) { 1411 if( !split_bool_set ) 1412 split_bool_set = new Node_List(area); 1413 split_bool_set->push(use); 1414 } 1415 if( use->Opcode() == Op_CreateEx ) { 1416 if( !split_cex_set ) 1417 split_cex_set = new Node_List(area); 1418 split_cex_set->push(use); 1419 } 1420 1421 1422 // Get "block" use is in 1423 uint idx = 0; 1424 while( use->in(idx) != old ) idx++; 1425 Node *prev = use->is_CFG() ? use : get_ctrl(use); 1426 assert( !loop->is_member( get_loop( prev ) ), "" ); 1427 Node *cfg = prev->_idx >= new_counter 1428 ? prev->in(2) 1429 : idom(prev); 1430 if( use->is_Phi() ) // Phi use is in prior block 1431 cfg = prev->in(idx); // NOT in block of Phi itself 1432 if (cfg->is_top()) { // Use is dead? 1433 _igvn.replace_input_of(use, idx, C->top()); 1434 continue; 1435 } 1436 1437 while( !loop->is_member( get_loop( cfg ) ) ) { 1438 prev = cfg; 1439 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg); 1440 } 1441 // If the use occurs after merging several exits from the loop, then 1442 // old value must have dominated all those exits. Since the same old 1443 // value was used on all those exits we did not need a Phi at this 1444 // merge point. NOW we do need a Phi here. Each loop exit value 1445 // is now merged with the peeled body exit; each exit gets its own 1446 // private Phi and those Phis need to be merged here. 1447 Node *phi; 1448 if( prev->is_Region() ) { 1449 if( idx == 0 ) { // Updating control edge? 1450 phi = prev; // Just use existing control 1451 } else { // Else need a new Phi 1452 phi = PhiNode::make( prev, old ); 1453 // Now recursively fix up the new uses of old! 1454 for( uint i = 1; i < prev->req(); i++ ) { 1455 worklist.push(phi); // Onto worklist once for each 'old' input 1456 } 1457 } 1458 } else { 1459 // Get new RegionNode merging old and new loop exits 1460 prev = old_new[prev->_idx]; 1461 assert( prev, "just made this in step 7" ); 1462 if( idx == 0 ) { // Updating control edge? 1463 phi = prev; // Just use existing control 1464 } else { // Else need a new Phi 1465 // Make a new Phi merging data values properly 1466 phi = PhiNode::make( prev, old ); 1467 phi->set_req( 1, nnn ); 1468 } 1469 } 1470 // If inserting a new Phi, check for prior hits 1471 if( idx != 0 ) { 1472 Node *hit = _igvn.hash_find_insert(phi); 1473 if( hit == NULL ) { 1474 _igvn.register_new_node_with_optimizer(phi); // Register new phi 1475 } else { // or 1476 // Remove the new phi from the graph and use the hit 1477 _igvn.remove_dead_node(phi); 1478 phi = hit; // Use existing phi 1479 } 1480 set_ctrl(phi, prev); 1481 } 1482 // Make 'use' use the Phi instead of the old loop body exit value 1483 _igvn.replace_input_of(use, idx, phi); 1484 if( use->_idx >= new_counter ) { // If updating new phis 1485 // Not needed for correctness, but prevents a weak assert 1486 // in AddPNode from tripping (when we end up with different 1487 // base & derived Phis that will become the same after 1488 // IGVN does CSE). 1489 Node *hit = _igvn.hash_find_insert(use); 1490 if( hit ) // Go ahead and re-hash for hits. 1491 _igvn.replace_node( use, hit ); 1492 } 1493 1494 // If 'use' was in the loop-exit block, it now needs to be sunk 1495 // below the post-loop merge point. 1496 sink_use( use, prev ); 1497 } 1498 } 1499 } 1500 1501 // Check for IFs that need splitting/cloning. Happens if an IF outside of 1502 // the loop uses a condition set in the loop. The original IF probably 1503 // takes control from one or more OLD Regions (which in turn get from NEW 1504 // Regions). In any case, there will be a set of Phis for each merge point 1505 // from the IF up to where the original BOOL def exists the loop. 1506 if( split_if_set ) { 1507 while( split_if_set->size() ) { 1508 Node *iff = split_if_set->pop(); 1509 if( iff->in(1)->is_Phi() ) { 1510 BoolNode *b = clone_iff( iff->in(1)->as_Phi(), loop ); 1511 _igvn.replace_input_of(iff, 1, b); 1512 } 1513 } 1514 } 1515 if( split_bool_set ) { 1516 while( split_bool_set->size() ) { 1517 Node *b = split_bool_set->pop(); 1518 Node *phi = b->in(1); 1519 assert( phi->is_Phi(), "" ); 1520 CmpNode *cmp = clone_bool( (PhiNode*)phi, loop ); 1521 _igvn.replace_input_of(b, 1, cmp); 1522 } 1523 } 1524 if( split_cex_set ) { 1525 while( split_cex_set->size() ) { 1526 Node *b = split_cex_set->pop(); 1527 assert( b->in(0)->is_Region(), "" ); 1528 assert( b->in(1)->is_Phi(), "" ); 1529 assert( b->in(0)->in(0) == b->in(1)->in(0), "" ); 1530 split_up( b, b->in(0), NULL ); 1531 } 1532 } 1533 1534 } 1535 1536 1537 //---------------------- stride_of_possible_iv ------------------------------------- 1538 // Looks for an iff/bool/comp with one operand of the compare 1539 // being a cycle involving an add and a phi, 1540 // with an optional truncation (left-shift followed by a right-shift) 1541 // of the add. Returns zero if not an iv. 1542 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) { 1543 Node* trunc1 = NULL; 1544 Node* trunc2 = NULL; 1545 const TypeInt* ttype = NULL; 1546 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) { 1547 return 0; 1548 } 1549 BoolNode* bl = iff->in(1)->as_Bool(); 1550 Node* cmp = bl->in(1); 1551 if (!cmp || cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU) { 1552 return 0; 1553 } 1554 // Must have an invariant operand 1555 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) { 1556 return 0; 1557 } 1558 Node* add2 = NULL; 1559 Node* cmp1 = cmp->in(1); 1560 if (cmp1->is_Phi()) { 1561 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) ))) 1562 Node* phi = cmp1; 1563 for (uint i = 1; i < phi->req(); i++) { 1564 Node* in = phi->in(i); 1565 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in, 1566 &trunc1, &trunc2, &ttype); 1567 if (add && add->in(1) == phi) { 1568 add2 = add->in(2); 1569 break; 1570 } 1571 } 1572 } else { 1573 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) ))) 1574 Node* addtrunc = cmp1; 1575 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc, 1576 &trunc1, &trunc2, &ttype); 1577 if (add && add->in(1)->is_Phi()) { 1578 Node* phi = add->in(1); 1579 for (uint i = 1; i < phi->req(); i++) { 1580 if (phi->in(i) == addtrunc) { 1581 add2 = add->in(2); 1582 break; 1583 } 1584 } 1585 } 1586 } 1587 if (add2 != NULL) { 1588 const TypeInt* add2t = _igvn.type(add2)->is_int(); 1589 if (add2t->is_con()) { 1590 return add2t->get_con(); 1591 } 1592 } 1593 return 0; 1594 } 1595 1596 1597 //---------------------- stay_in_loop ------------------------------------- 1598 // Return the (unique) control output node that's in the loop (if it exists.) 1599 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) { 1600 Node* unique = NULL; 1601 if (!n) return NULL; 1602 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1603 Node* use = n->fast_out(i); 1604 if (!has_ctrl(use) && loop->is_member(get_loop(use))) { 1605 if (unique != NULL) { 1606 return NULL; 1607 } 1608 unique = use; 1609 } 1610 } 1611 return unique; 1612 } 1613 1614 //------------------------------ register_node ------------------------------------- 1615 // Utility to register node "n" with PhaseIdealLoop 1616 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) { 1617 _igvn.register_new_node_with_optimizer(n); 1618 loop->_body.push(n); 1619 if (n->is_CFG()) { 1620 set_loop(n, loop); 1621 set_idom(n, pred, ddepth); 1622 } else { 1623 set_ctrl(n, pred); 1624 } 1625 } 1626 1627 //------------------------------ proj_clone ------------------------------------- 1628 // Utility to create an if-projection 1629 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) { 1630 ProjNode* c = p->clone()->as_Proj(); 1631 c->set_req(0, iff); 1632 return c; 1633 } 1634 1635 //------------------------------ short_circuit_if ------------------------------------- 1636 // Force the iff control output to be the live_proj 1637 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) { 1638 guarantee(live_proj != NULL, "null projection"); 1639 int proj_con = live_proj->_con; 1640 assert(proj_con == 0 || proj_con == 1, "false or true projection"); 1641 Node *con = _igvn.intcon(proj_con); 1642 set_ctrl(con, C->root()); 1643 if (iff) { 1644 iff->set_req(1, con); 1645 } 1646 return con; 1647 } 1648 1649 //------------------------------ insert_if_before_proj ------------------------------------- 1650 // Insert a new if before an if projection (* - new node) 1651 // 1652 // before 1653 // if(test) 1654 // / \ 1655 // v v 1656 // other-proj proj (arg) 1657 // 1658 // after 1659 // if(test) 1660 // / \ 1661 // / v 1662 // | * proj-clone 1663 // v | 1664 // other-proj v 1665 // * new_if(relop(cmp[IU](left,right))) 1666 // / \ 1667 // v v 1668 // * new-proj proj 1669 // (returned) 1670 // 1671 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) { 1672 IfNode* iff = proj->in(0)->as_If(); 1673 IdealLoopTree *loop = get_loop(proj); 1674 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 1675 int ddepth = dom_depth(proj); 1676 1677 _igvn.rehash_node_delayed(iff); 1678 _igvn.rehash_node_delayed(proj); 1679 1680 proj->set_req(0, NULL); // temporary disconnect 1681 ProjNode* proj2 = proj_clone(proj, iff); 1682 register_node(proj2, loop, iff, ddepth); 1683 1684 Node* cmp = Signed ? (Node*) new (C)CmpINode(left, right) : (Node*) new (C)CmpUNode(left, right); 1685 register_node(cmp, loop, proj2, ddepth); 1686 1687 BoolNode* bol = new (C)BoolNode(cmp, relop); 1688 register_node(bol, loop, proj2, ddepth); 1689 1690 IfNode* new_if = new (C)IfNode(proj2, bol, iff->_prob, iff->_fcnt); 1691 register_node(new_if, loop, proj2, ddepth); 1692 1693 proj->set_req(0, new_if); // reattach 1694 set_idom(proj, new_if, ddepth); 1695 1696 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj(); 1697 guarantee(new_exit != NULL, "null exit node"); 1698 register_node(new_exit, get_loop(other_proj), new_if, ddepth); 1699 1700 return new_exit; 1701 } 1702 1703 //------------------------------ insert_region_before_proj ------------------------------------- 1704 // Insert a region before an if projection (* - new node) 1705 // 1706 // before 1707 // if(test) 1708 // / | 1709 // v | 1710 // proj v 1711 // other-proj 1712 // 1713 // after 1714 // if(test) 1715 // / | 1716 // v | 1717 // * proj-clone v 1718 // | other-proj 1719 // v 1720 // * new-region 1721 // | 1722 // v 1723 // * dum_if 1724 // / \ 1725 // v \ 1726 // * dum-proj v 1727 // proj 1728 // 1729 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) { 1730 IfNode* iff = proj->in(0)->as_If(); 1731 IdealLoopTree *loop = get_loop(proj); 1732 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 1733 int ddepth = dom_depth(proj); 1734 1735 _igvn.rehash_node_delayed(iff); 1736 _igvn.rehash_node_delayed(proj); 1737 1738 proj->set_req(0, NULL); // temporary disconnect 1739 ProjNode* proj2 = proj_clone(proj, iff); 1740 register_node(proj2, loop, iff, ddepth); 1741 1742 RegionNode* reg = new (C)RegionNode(2); 1743 reg->set_req(1, proj2); 1744 register_node(reg, loop, iff, ddepth); 1745 1746 IfNode* dum_if = new (C)IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt); 1747 register_node(dum_if, loop, reg, ddepth); 1748 1749 proj->set_req(0, dum_if); // reattach 1750 set_idom(proj, dum_if, ddepth); 1751 1752 ProjNode* dum_proj = proj_clone(other_proj, dum_if); 1753 register_node(dum_proj, loop, dum_if, ddepth); 1754 1755 return reg; 1756 } 1757 1758 //------------------------------ insert_cmpi_loop_exit ------------------------------------- 1759 // Clone a signed compare loop exit from an unsigned compare and 1760 // insert it before the unsigned cmp on the stay-in-loop path. 1761 // All new nodes inserted in the dominator tree between the original 1762 // if and it's projections. The original if test is replaced with 1763 // a constant to force the stay-in-loop path. 1764 // 1765 // This is done to make sure that the original if and it's projections 1766 // still dominate the same set of control nodes, that the ctrl() relation 1767 // from data nodes to them is preserved, and that their loop nesting is 1768 // preserved. 1769 // 1770 // before 1771 // if(i <u limit) unsigned compare loop exit 1772 // / | 1773 // v v 1774 // exit-proj stay-in-loop-proj 1775 // 1776 // after 1777 // if(stay-in-loop-const) original if 1778 // / | 1779 // / v 1780 // / if(i < limit) new signed test 1781 // / / | 1782 // / / v 1783 // / / if(i <u limit) new cloned unsigned test 1784 // / / / | 1785 // v v v | 1786 // region | 1787 // | | 1788 // dum-if | 1789 // / | | 1790 // ether | | 1791 // v v 1792 // exit-proj stay-in-loop-proj 1793 // 1794 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) { 1795 const bool Signed = true; 1796 const bool Unsigned = false; 1797 1798 BoolNode* bol = if_cmpu->in(1)->as_Bool(); 1799 if (bol->_test._test != BoolTest::lt) return NULL; 1800 CmpNode* cmpu = bol->in(1)->as_Cmp(); 1801 if (cmpu->Opcode() != Op_CmpU) return NULL; 1802 int stride = stride_of_possible_iv(if_cmpu); 1803 if (stride == 0) return NULL; 1804 1805 Node* lp_proj = stay_in_loop(if_cmpu, loop); 1806 guarantee(lp_proj != NULL, "null loop node"); 1807 1808 ProjNode* lp_continue = lp_proj->as_Proj(); 1809 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj(); 1810 1811 Node* limit = NULL; 1812 if (stride > 0) { 1813 limit = cmpu->in(2); 1814 } else { 1815 limit = _igvn.makecon(TypeInt::ZERO); 1816 set_ctrl(limit, C->root()); 1817 } 1818 // Create a new region on the exit path 1819 RegionNode* reg = insert_region_before_proj(lp_exit); 1820 guarantee(reg != NULL, "null region node"); 1821 1822 // Clone the if-cmpu-true-false using a signed compare 1823 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge; 1824 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue); 1825 reg->add_req(cmpi_exit); 1826 1827 // Clone the if-cmpu-true-false 1828 BoolTest::mask rel_u = bol->_test._test; 1829 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue); 1830 reg->add_req(cmpu_exit); 1831 1832 // Force original if to stay in loop. 1833 short_circuit_if(if_cmpu, lp_continue); 1834 1835 return cmpi_exit->in(0)->as_If(); 1836 } 1837 1838 //------------------------------ remove_cmpi_loop_exit ------------------------------------- 1839 // Remove a previously inserted signed compare loop exit. 1840 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) { 1841 Node* lp_proj = stay_in_loop(if_cmp, loop); 1842 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI && 1843 stay_in_loop(lp_proj, loop)->is_If() && 1844 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu"); 1845 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO); 1846 set_ctrl(con, C->root()); 1847 if_cmp->set_req(1, con); 1848 } 1849 1850 //------------------------------ scheduled_nodelist ------------------------------------- 1851 // Create a post order schedule of nodes that are in the 1852 // "member" set. The list is returned in "sched". 1853 // The first node in "sched" is the loop head, followed by 1854 // nodes which have no inputs in the "member" set, and then 1855 // followed by the nodes that have an immediate input dependence 1856 // on a node in "sched". 1857 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) { 1858 1859 assert(member.test(loop->_head->_idx), "loop head must be in member set"); 1860 Arena *a = Thread::current()->resource_area(); 1861 VectorSet visited(a); 1862 Node_Stack nstack(a, loop->_body.size()); 1863 1864 Node* n = loop->_head; // top of stack is cached in "n" 1865 uint idx = 0; 1866 visited.set(n->_idx); 1867 1868 // Initially push all with no inputs from within member set 1869 for(uint i = 0; i < loop->_body.size(); i++ ) { 1870 Node *elt = loop->_body.at(i); 1871 if (member.test(elt->_idx)) { 1872 bool found = false; 1873 for (uint j = 0; j < elt->req(); j++) { 1874 Node* def = elt->in(j); 1875 if (def && member.test(def->_idx) && def != elt) { 1876 found = true; 1877 break; 1878 } 1879 } 1880 if (!found && elt != loop->_head) { 1881 nstack.push(n, idx); 1882 n = elt; 1883 assert(!visited.test(n->_idx), "not seen yet"); 1884 visited.set(n->_idx); 1885 } 1886 } 1887 } 1888 1889 // traverse out's that are in the member set 1890 while (true) { 1891 if (idx < n->outcnt()) { 1892 Node* use = n->raw_out(idx); 1893 idx++; 1894 if (!visited.test_set(use->_idx)) { 1895 if (member.test(use->_idx)) { 1896 nstack.push(n, idx); 1897 n = use; 1898 idx = 0; 1899 } 1900 } 1901 } else { 1902 // All outputs processed 1903 sched.push(n); 1904 if (nstack.is_empty()) break; 1905 n = nstack.node(); 1906 idx = nstack.index(); 1907 nstack.pop(); 1908 } 1909 } 1910 } 1911 1912 1913 //------------------------------ has_use_in_set ------------------------------------- 1914 // Has a use in the vector set 1915 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) { 1916 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1917 Node* use = n->fast_out(j); 1918 if (vset.test(use->_idx)) { 1919 return true; 1920 } 1921 } 1922 return false; 1923 } 1924 1925 1926 //------------------------------ has_use_internal_to_set ------------------------------------- 1927 // Has use internal to the vector set (ie. not in a phi at the loop head) 1928 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) { 1929 Node* head = loop->_head; 1930 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1931 Node* use = n->fast_out(j); 1932 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) { 1933 return true; 1934 } 1935 } 1936 return false; 1937 } 1938 1939 1940 //------------------------------ clone_for_use_outside_loop ------------------------------------- 1941 // clone "n" for uses that are outside of loop 1942 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) { 1943 int cloned = 0; 1944 assert(worklist.size() == 0, "should be empty"); 1945 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1946 Node* use = n->fast_out(j); 1947 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) { 1948 worklist.push(use); 1949 } 1950 } 1951 while( worklist.size() ) { 1952 Node *use = worklist.pop(); 1953 if (!has_node(use) || use->in(0) == C->top()) continue; 1954 uint j; 1955 for (j = 0; j < use->req(); j++) { 1956 if (use->in(j) == n) break; 1957 } 1958 assert(j < use->req(), "must be there"); 1959 1960 // clone "n" and insert it between the inputs of "n" and the use outside the loop 1961 Node* n_clone = n->clone(); 1962 _igvn.replace_input_of(use, j, n_clone); 1963 cloned++; 1964 Node* use_c; 1965 if (!use->is_Phi()) { 1966 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0); 1967 } else { 1968 // Use in a phi is considered a use in the associated predecessor block 1969 use_c = use->in(0)->in(j); 1970 } 1971 set_ctrl(n_clone, use_c); 1972 assert(!loop->is_member(get_loop(use_c)), "should be outside loop"); 1973 get_loop(use_c)->_body.push(n_clone); 1974 _igvn.register_new_node_with_optimizer(n_clone); 1975 #if !defined(PRODUCT) 1976 if (TracePartialPeeling) { 1977 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx); 1978 } 1979 #endif 1980 } 1981 return cloned; 1982 } 1983 1984 1985 //------------------------------ clone_for_special_use_inside_loop ------------------------------------- 1986 // clone "n" for special uses that are in the not_peeled region. 1987 // If these def-uses occur in separate blocks, the code generator 1988 // marks the method as not compilable. For example, if a "BoolNode" 1989 // is in a different basic block than the "IfNode" that uses it, then 1990 // the compilation is aborted in the code generator. 1991 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 1992 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) { 1993 if (n->is_Phi() || n->is_Load()) { 1994 return; 1995 } 1996 assert(worklist.size() == 0, "should be empty"); 1997 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1998 Node* use = n->fast_out(j); 1999 if ( not_peel.test(use->_idx) && 2000 (use->is_If() || use->is_CMove() || use->is_Bool()) && 2001 use->in(1) == n) { 2002 worklist.push(use); 2003 } 2004 } 2005 if (worklist.size() > 0) { 2006 // clone "n" and insert it between inputs of "n" and the use 2007 Node* n_clone = n->clone(); 2008 loop->_body.push(n_clone); 2009 _igvn.register_new_node_with_optimizer(n_clone); 2010 set_ctrl(n_clone, get_ctrl(n)); 2011 sink_list.push(n_clone); 2012 not_peel <<= n_clone->_idx; // add n_clone to not_peel set. 2013 #if !defined(PRODUCT) 2014 if (TracePartialPeeling) { 2015 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx); 2016 } 2017 #endif 2018 while( worklist.size() ) { 2019 Node *use = worklist.pop(); 2020 _igvn.rehash_node_delayed(use); 2021 for (uint j = 1; j < use->req(); j++) { 2022 if (use->in(j) == n) { 2023 use->set_req(j, n_clone); 2024 } 2025 } 2026 } 2027 } 2028 } 2029 2030 2031 //------------------------------ insert_phi_for_loop ------------------------------------- 2032 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 2033 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) { 2034 Node *phi = PhiNode::make(lp, back_edge_val); 2035 phi->set_req(LoopNode::EntryControl, lp_entry_val); 2036 // Use existing phi if it already exists 2037 Node *hit = _igvn.hash_find_insert(phi); 2038 if( hit == NULL ) { 2039 _igvn.register_new_node_with_optimizer(phi); 2040 set_ctrl(phi, lp); 2041 } else { 2042 // Remove the new phi from the graph and use the hit 2043 _igvn.remove_dead_node(phi); 2044 phi = hit; 2045 } 2046 _igvn.replace_input_of(use, idx, phi); 2047 } 2048 2049 #ifdef ASSERT 2050 //------------------------------ is_valid_loop_partition ------------------------------------- 2051 // Validate the loop partition sets: peel and not_peel 2052 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, 2053 VectorSet& not_peel ) { 2054 uint i; 2055 // Check that peel_list entries are in the peel set 2056 for (i = 0; i < peel_list.size(); i++) { 2057 if (!peel.test(peel_list.at(i)->_idx)) { 2058 return false; 2059 } 2060 } 2061 // Check at loop members are in one of peel set or not_peel set 2062 for (i = 0; i < loop->_body.size(); i++ ) { 2063 Node *def = loop->_body.at(i); 2064 uint di = def->_idx; 2065 // Check that peel set elements are in peel_list 2066 if (peel.test(di)) { 2067 if (not_peel.test(di)) { 2068 return false; 2069 } 2070 // Must be in peel_list also 2071 bool found = false; 2072 for (uint j = 0; j < peel_list.size(); j++) { 2073 if (peel_list.at(j)->_idx == di) { 2074 found = true; 2075 break; 2076 } 2077 } 2078 if (!found) { 2079 return false; 2080 } 2081 } else if (not_peel.test(di)) { 2082 if (peel.test(di)) { 2083 return false; 2084 } 2085 } else { 2086 return false; 2087 } 2088 } 2089 return true; 2090 } 2091 2092 //------------------------------ is_valid_clone_loop_exit_use ------------------------------------- 2093 // Ensure a use outside of loop is of the right form 2094 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) { 2095 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2096 return (use->is_Phi() && 2097 use_c->is_Region() && use_c->req() == 3 && 2098 (use_c->in(exit_idx)->Opcode() == Op_IfTrue || 2099 use_c->in(exit_idx)->Opcode() == Op_IfFalse || 2100 use_c->in(exit_idx)->Opcode() == Op_JumpProj) && 2101 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) ); 2102 } 2103 2104 //------------------------------ is_valid_clone_loop_form ------------------------------------- 2105 // Ensure that all uses outside of loop are of the right form 2106 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 2107 uint orig_exit_idx, uint clone_exit_idx) { 2108 uint len = peel_list.size(); 2109 for (uint i = 0; i < len; i++) { 2110 Node *def = peel_list.at(i); 2111 2112 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 2113 Node *use = def->fast_out(j); 2114 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2115 if (!loop->is_member(get_loop(use_c))) { 2116 // use is not in the loop, check for correct structure 2117 if (use->in(0) == def) { 2118 // Okay 2119 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) { 2120 return false; 2121 } 2122 } 2123 } 2124 } 2125 return true; 2126 } 2127 #endif 2128 2129 //------------------------------ partial_peel ------------------------------------- 2130 // Partially peel (aka loop rotation) the top portion of a loop (called 2131 // the peel section below) by cloning it and placing one copy just before 2132 // the new loop head and the other copy at the bottom of the new loop. 2133 // 2134 // before after where it came from 2135 // 2136 // stmt1 stmt1 2137 // loop: stmt2 clone 2138 // stmt2 if condA goto exitA clone 2139 // if condA goto exitA new_loop: new 2140 // stmt3 stmt3 clone 2141 // if !condB goto loop if condB goto exitB clone 2142 // exitB: stmt2 orig 2143 // stmt4 if !condA goto new_loop orig 2144 // exitA: goto exitA 2145 // exitB: 2146 // stmt4 2147 // exitA: 2148 // 2149 // Step 1: find the cut point: an exit test on probable 2150 // induction variable. 2151 // Step 2: schedule (with cloning) operations in the peel 2152 // section that can be executed after the cut into 2153 // the section that is not peeled. This may need 2154 // to clone operations into exit blocks. For 2155 // instance, a reference to A[i] in the not-peel 2156 // section and a reference to B[i] in an exit block 2157 // may cause a left-shift of i by 2 to be placed 2158 // in the peel block. This step will clone the left 2159 // shift into the exit block and sink the left shift 2160 // from the peel to the not-peel section. 2161 // Step 3: clone the loop, retarget the control, and insert 2162 // phis for values that are live across the new loop 2163 // head. This is very dependent on the graph structure 2164 // from clone_loop. It creates region nodes for 2165 // exit control and associated phi nodes for values 2166 // flow out of the loop through that exit. The region 2167 // node is dominated by the clone's control projection. 2168 // So the clone's peel section is placed before the 2169 // new loop head, and the clone's not-peel section is 2170 // forms the top part of the new loop. The original 2171 // peel section forms the tail of the new loop. 2172 // Step 4: update the dominator tree and recompute the 2173 // dominator depth. 2174 // 2175 // orig 2176 // 2177 // stmt1 2178 // | 2179 // v 2180 // loop predicate 2181 // | 2182 // v 2183 // loop<----+ 2184 // | | 2185 // stmt2 | 2186 // | | 2187 // v | 2188 // ifA | 2189 // / | | 2190 // v v | 2191 // false true ^ <-- last_peel 2192 // / | | 2193 // / ===|==cut | 2194 // / stmt3 | <-- first_not_peel 2195 // / | | 2196 // | v | 2197 // v ifB | 2198 // exitA: / \ | 2199 // / \ | 2200 // v v | 2201 // false true | 2202 // / \ | 2203 // / ----+ 2204 // | 2205 // v 2206 // exitB: 2207 // stmt4 2208 // 2209 // 2210 // after clone loop 2211 // 2212 // stmt1 2213 // | 2214 // v 2215 // loop predicate 2216 // / \ 2217 // clone / \ orig 2218 // / \ 2219 // / \ 2220 // v v 2221 // +---->loop loop<----+ 2222 // | | | | 2223 // | stmt2 stmt2 | 2224 // | | | | 2225 // | v v | 2226 // | ifA ifA | 2227 // | | \ / | | 2228 // | v v v v | 2229 // ^ true false false true ^ <-- last_peel 2230 // | | ^ \ / | | 2231 // | cut==|== \ \ / ===|==cut | 2232 // | stmt3 \ \ / stmt3 | <-- first_not_peel 2233 // | | dom | | | | 2234 // | v \ 1v v2 v | 2235 // | ifB regionA ifB | 2236 // | / \ | / \ | 2237 // | / \ v / \ | 2238 // | v v exitA: v v | 2239 // | true false false true | 2240 // | / ^ \ / \ | 2241 // +---- \ \ / ----+ 2242 // dom \ / 2243 // \ 1v v2 2244 // regionB 2245 // | 2246 // v 2247 // exitB: 2248 // stmt4 2249 // 2250 // 2251 // after partial peel 2252 // 2253 // stmt1 2254 // | 2255 // v 2256 // loop predicate 2257 // / 2258 // clone / orig 2259 // / TOP 2260 // / \ 2261 // v v 2262 // TOP->loop loop----+ 2263 // | | | 2264 // stmt2 stmt2 | 2265 // | | | 2266 // v v | 2267 // ifA ifA | 2268 // | \ / | | 2269 // v v v v | 2270 // true false false true | <-- last_peel 2271 // | ^ \ / +------|---+ 2272 // +->newloop \ \ / === ==cut | | 2273 // | stmt3 \ \ / TOP | | 2274 // | | dom | | stmt3 | | <-- first_not_peel 2275 // | v \ 1v v2 v | | 2276 // | ifB regionA ifB ^ v 2277 // | / \ | / \ | | 2278 // | / \ v / \ | | 2279 // | v v exitA: v v | | 2280 // | true false false true | | 2281 // | / ^ \ / \ | | 2282 // | | \ \ / v | | 2283 // | | dom \ / TOP | | 2284 // | | \ 1v v2 | | 2285 // ^ v regionB | | 2286 // | | | | | 2287 // | | v ^ v 2288 // | | exitB: | | 2289 // | | stmt4 | | 2290 // | +------------>-----------------+ | 2291 // | | 2292 // +-----------------<---------------------+ 2293 // 2294 // 2295 // final graph 2296 // 2297 // stmt1 2298 // | 2299 // v 2300 // loop predicate 2301 // | 2302 // v 2303 // stmt2 clone 2304 // | 2305 // v 2306 // ........> ifA clone 2307 // : / | 2308 // dom / | 2309 // : v v 2310 // : false true 2311 // : | | 2312 // : | v 2313 // : | newloop<-----+ 2314 // : | | | 2315 // : | stmt3 clone | 2316 // : | | | 2317 // : | v | 2318 // : | ifB | 2319 // : | / \ | 2320 // : | v v | 2321 // : | false true | 2322 // : | | | | 2323 // : | v stmt2 | 2324 // : | exitB: | | 2325 // : | stmt4 v | 2326 // : | ifA orig | 2327 // : | / \ | 2328 // : | / \ | 2329 // : | v v | 2330 // : | false true | 2331 // : | / \ | 2332 // : v v -----+ 2333 // RegionA 2334 // | 2335 // v 2336 // exitA 2337 // 2338 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) { 2339 2340 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only"); 2341 if (!loop->_head->is_Loop()) { 2342 return false; } 2343 2344 LoopNode *head = loop->_head->as_Loop(); 2345 2346 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) { 2347 return false; 2348 } 2349 2350 // Check for complex exit control 2351 for(uint ii = 0; ii < loop->_body.size(); ii++ ) { 2352 Node *n = loop->_body.at(ii); 2353 int opc = n->Opcode(); 2354 if (n->is_Call() || 2355 opc == Op_Catch || 2356 opc == Op_CatchProj || 2357 opc == Op_Jump || 2358 opc == Op_JumpProj || 2359 opc == Op_FlagsProj) { 2360 #if !defined(PRODUCT) 2361 if (TracePartialPeeling) { 2362 tty->print_cr("\nExit control too complex: lp: %d", head->_idx); 2363 } 2364 #endif 2365 return false; 2366 } 2367 } 2368 2369 int dd = dom_depth(head); 2370 2371 // Step 1: find cut point 2372 2373 // Walk up dominators to loop head looking for first loop exit 2374 // which is executed on every path thru loop. 2375 IfNode *peel_if = NULL; 2376 IfNode *peel_if_cmpu = NULL; 2377 2378 Node *iff = loop->tail(); 2379 while( iff != head ) { 2380 if( iff->is_If() ) { 2381 Node *ctrl = get_ctrl(iff->in(1)); 2382 if (ctrl->is_top()) return false; // Dead test on live IF. 2383 // If loop-varying exit-test, check for induction variable 2384 if( loop->is_member(get_loop(ctrl)) && 2385 loop->is_loop_exit(iff) && 2386 is_possible_iv_test(iff)) { 2387 Node* cmp = iff->in(1)->in(1); 2388 if (cmp->Opcode() == Op_CmpI) { 2389 peel_if = iff->as_If(); 2390 } else { 2391 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU"); 2392 peel_if_cmpu = iff->as_If(); 2393 } 2394 } 2395 } 2396 iff = idom(iff); 2397 } 2398 // Prefer signed compare over unsigned compare. 2399 IfNode* new_peel_if = NULL; 2400 if (peel_if == NULL) { 2401 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) { 2402 return false; // No peel point found 2403 } 2404 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop); 2405 if (new_peel_if == NULL) { 2406 return false; // No peel point found 2407 } 2408 peel_if = new_peel_if; 2409 } 2410 Node* last_peel = stay_in_loop(peel_if, loop); 2411 Node* first_not_peeled = stay_in_loop(last_peel, loop); 2412 if (first_not_peeled == NULL || first_not_peeled == head) { 2413 return false; 2414 } 2415 2416 #if !defined(PRODUCT) 2417 if (TraceLoopOpts) { 2418 tty->print("PartialPeel "); 2419 loop->dump_head(); 2420 } 2421 2422 if (TracePartialPeeling) { 2423 tty->print_cr("before partial peel one iteration"); 2424 Node_List wl; 2425 Node* t = head->in(2); 2426 while (true) { 2427 wl.push(t); 2428 if (t == head) break; 2429 t = idom(t); 2430 } 2431 while (wl.size() > 0) { 2432 Node* tt = wl.pop(); 2433 tt->dump(); 2434 if (tt == last_peel) tty->print_cr("-- cut --"); 2435 } 2436 } 2437 #endif 2438 ResourceArea *area = Thread::current()->resource_area(); 2439 VectorSet peel(area); 2440 VectorSet not_peel(area); 2441 Node_List peel_list(area); 2442 Node_List worklist(area); 2443 Node_List sink_list(area); 2444 2445 // Set of cfg nodes to peel are those that are executable from 2446 // the head through last_peel. 2447 assert(worklist.size() == 0, "should be empty"); 2448 worklist.push(head); 2449 peel.set(head->_idx); 2450 while (worklist.size() > 0) { 2451 Node *n = worklist.pop(); 2452 if (n != last_peel) { 2453 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2454 Node* use = n->fast_out(j); 2455 if (use->is_CFG() && 2456 loop->is_member(get_loop(use)) && 2457 !peel.test_set(use->_idx)) { 2458 worklist.push(use); 2459 } 2460 } 2461 } 2462 } 2463 2464 // Set of non-cfg nodes to peel are those that are control 2465 // dependent on the cfg nodes. 2466 uint i; 2467 for(i = 0; i < loop->_body.size(); i++ ) { 2468 Node *n = loop->_body.at(i); 2469 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n; 2470 if (peel.test(n_c->_idx)) { 2471 peel.set(n->_idx); 2472 } else { 2473 not_peel.set(n->_idx); 2474 } 2475 } 2476 2477 // Step 2: move operations from the peeled section down into the 2478 // not-peeled section 2479 2480 // Get a post order schedule of nodes in the peel region 2481 // Result in right-most operand. 2482 scheduled_nodelist(loop, peel, peel_list ); 2483 2484 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 2485 2486 // For future check for too many new phis 2487 uint old_phi_cnt = 0; 2488 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 2489 Node* use = head->fast_out(j); 2490 if (use->is_Phi()) old_phi_cnt++; 2491 } 2492 2493 #if !defined(PRODUCT) 2494 if (TracePartialPeeling) { 2495 tty->print_cr("\npeeled list"); 2496 } 2497 #endif 2498 2499 // Evacuate nodes in peel region into the not_peeled region if possible 2500 uint new_phi_cnt = 0; 2501 uint cloned_for_outside_use = 0; 2502 for (i = 0; i < peel_list.size();) { 2503 Node* n = peel_list.at(i); 2504 #if !defined(PRODUCT) 2505 if (TracePartialPeeling) n->dump(); 2506 #endif 2507 bool incr = true; 2508 if ( !n->is_CFG() ) { 2509 2510 if ( has_use_in_set(n, not_peel) ) { 2511 2512 // If not used internal to the peeled region, 2513 // move "n" from peeled to not_peeled region. 2514 2515 if ( !has_use_internal_to_set(n, peel, loop) ) { 2516 2517 // if not pinned and not a load (which maybe anti-dependent on a store) 2518 // and not a CMove (Matcher expects only bool->cmove). 2519 if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) { 2520 cloned_for_outside_use += clone_for_use_outside_loop( loop, n, worklist ); 2521 sink_list.push(n); 2522 peel >>= n->_idx; // delete n from peel set. 2523 not_peel <<= n->_idx; // add n to not_peel set. 2524 peel_list.remove(i); 2525 incr = false; 2526 #if !defined(PRODUCT) 2527 if (TracePartialPeeling) { 2528 tty->print_cr("sink to not_peeled region: %d newbb: %d", 2529 n->_idx, get_ctrl(n)->_idx); 2530 } 2531 #endif 2532 } 2533 } else { 2534 // Otherwise check for special def-use cases that span 2535 // the peel/not_peel boundary such as bool->if 2536 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist ); 2537 new_phi_cnt++; 2538 } 2539 } 2540 } 2541 if (incr) i++; 2542 } 2543 2544 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) { 2545 #if !defined(PRODUCT) 2546 if (TracePartialPeeling) { 2547 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c", 2548 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F'); 2549 } 2550 #endif 2551 if (new_peel_if != NULL) { 2552 remove_cmpi_loop_exit(new_peel_if, loop); 2553 } 2554 // Inhibit more partial peeling on this loop 2555 assert(!head->is_partial_peel_loop(), "not partial peeled"); 2556 head->mark_partial_peel_failed(); 2557 if (cloned_for_outside_use > 0) { 2558 // Terminate this round of loop opts because 2559 // the graph outside this loop was changed. 2560 C->set_major_progress(); 2561 return true; 2562 } 2563 return false; 2564 } 2565 2566 // Step 3: clone loop, retarget control, and insert new phis 2567 2568 // Create new loop head for new phis and to hang 2569 // the nodes being moved (sinked) from the peel region. 2570 LoopNode* new_head = new (C) LoopNode(last_peel, last_peel); 2571 new_head->set_unswitch_count(head->unswitch_count()); // Preserve 2572 _igvn.register_new_node_with_optimizer(new_head); 2573 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled"); 2574 first_not_peeled->set_req(0, new_head); 2575 set_loop(new_head, loop); 2576 loop->_body.push(new_head); 2577 not_peel.set(new_head->_idx); 2578 set_idom(new_head, last_peel, dom_depth(first_not_peeled)); 2579 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled)); 2580 2581 while (sink_list.size() > 0) { 2582 Node* n = sink_list.pop(); 2583 set_ctrl(n, new_head); 2584 } 2585 2586 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 2587 2588 clone_loop( loop, old_new, dd ); 2589 2590 const uint clone_exit_idx = 1; 2591 const uint orig_exit_idx = 2; 2592 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop"); 2593 2594 Node* head_clone = old_new[head->_idx]; 2595 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop(); 2596 Node* orig_tail_clone = head_clone->in(2); 2597 2598 // Add phi if "def" node is in peel set and "use" is not 2599 2600 for(i = 0; i < peel_list.size(); i++ ) { 2601 Node *def = peel_list.at(i); 2602 if (!def->is_CFG()) { 2603 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 2604 Node *use = def->fast_out(j); 2605 if (has_node(use) && use->in(0) != C->top() && 2606 (!peel.test(use->_idx) || 2607 (use->is_Phi() && use->in(0) == head)) ) { 2608 worklist.push(use); 2609 } 2610 } 2611 while( worklist.size() ) { 2612 Node *use = worklist.pop(); 2613 for (uint j = 1; j < use->req(); j++) { 2614 Node* n = use->in(j); 2615 if (n == def) { 2616 2617 // "def" is in peel set, "use" is not in peel set 2618 // or "use" is in the entry boundary (a phi) of the peel set 2619 2620 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use; 2621 2622 if ( loop->is_member(get_loop( use_c )) ) { 2623 // use is in loop 2624 if (old_new[use->_idx] != NULL) { // null for dead code 2625 Node* use_clone = old_new[use->_idx]; 2626 _igvn.replace_input_of(use, j, C->top()); 2627 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone ); 2628 } 2629 } else { 2630 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format"); 2631 // use is not in the loop, check if the live range includes the cut 2632 Node* lp_if = use_c->in(orig_exit_idx)->in(0); 2633 if (not_peel.test(lp_if->_idx)) { 2634 assert(j == orig_exit_idx, "use from original loop"); 2635 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone ); 2636 } 2637 } 2638 } 2639 } 2640 } 2641 } 2642 } 2643 2644 // Step 3b: retarget control 2645 2646 // Redirect control to the new loop head if a cloned node in 2647 // the not_peeled region has control that points into the peeled region. 2648 // This necessary because the cloned peeled region will be outside 2649 // the loop. 2650 // from to 2651 // cloned-peeled <---+ 2652 // new_head_clone: | <--+ 2653 // cloned-not_peeled in(0) in(0) 2654 // orig-peeled 2655 2656 for(i = 0; i < loop->_body.size(); i++ ) { 2657 Node *n = loop->_body.at(i); 2658 if (!n->is_CFG() && n->in(0) != NULL && 2659 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) { 2660 Node* n_clone = old_new[n->_idx]; 2661 _igvn.replace_input_of(n_clone, 0, new_head_clone); 2662 } 2663 } 2664 2665 // Backedge of the surviving new_head (the clone) is original last_peel 2666 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel); 2667 2668 // Cut first node in original not_peel set 2669 _igvn.rehash_node_delayed(new_head); // Multiple edge updates: 2670 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of 2671 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls 2672 2673 // Copy head_clone back-branch info to original head 2674 // and remove original head's loop entry and 2675 // clone head's back-branch 2676 _igvn.rehash_node_delayed(head); // Multiple edge updates 2677 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl)); 2678 head->set_req(LoopNode::LoopBackControl, C->top()); 2679 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top()); 2680 2681 // Similarly modify the phis 2682 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) { 2683 Node* use = head->fast_out(k); 2684 if (use->is_Phi() && use->outcnt() > 0) { 2685 Node* use_clone = old_new[use->_idx]; 2686 _igvn.rehash_node_delayed(use); // Multiple edge updates 2687 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl)); 2688 use->set_req(LoopNode::LoopBackControl, C->top()); 2689 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top()); 2690 } 2691 } 2692 2693 // Step 4: update dominator tree and dominator depth 2694 2695 set_idom(head, orig_tail_clone, dd); 2696 recompute_dom_depth(); 2697 2698 // Inhibit more partial peeling on this loop 2699 new_head_clone->set_partial_peel_loop(); 2700 C->set_major_progress(); 2701 2702 #if !defined(PRODUCT) 2703 if (TracePartialPeeling) { 2704 tty->print_cr("\nafter partial peel one iteration"); 2705 Node_List wl(area); 2706 Node* t = last_peel; 2707 while (true) { 2708 wl.push(t); 2709 if (t == head_clone) break; 2710 t = idom(t); 2711 } 2712 while (wl.size() > 0) { 2713 Node* tt = wl.pop(); 2714 if (tt == head) tty->print_cr("orig head"); 2715 else if (tt == new_head_clone) tty->print_cr("new head"); 2716 else if (tt == head_clone) tty->print_cr("clone head"); 2717 tt->dump(); 2718 } 2719 } 2720 #endif 2721 return true; 2722 } 2723 2724 //------------------------------reorg_offsets---------------------------------- 2725 // Reorganize offset computations to lower register pressure. Mostly 2726 // prevent loop-fallout uses of the pre-incremented trip counter (which are 2727 // then alive with the post-incremented trip counter forcing an extra 2728 // register move) 2729 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) { 2730 // Perform it only for canonical counted loops. 2731 // Loop's shape could be messed up by iteration_split_impl. 2732 if (!loop->_head->is_CountedLoop()) 2733 return; 2734 if (!loop->_head->as_Loop()->is_valid_counted_loop()) 2735 return; 2736 2737 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 2738 CountedLoopEndNode *cle = cl->loopexit(); 2739 Node *exit = cle->proj_out(false); 2740 Node *phi = cl->phi(); 2741 2742 // Check for the special case of folks using the pre-incremented 2743 // trip-counter on the fall-out path (forces the pre-incremented 2744 // and post-incremented trip counter to be live at the same time). 2745 // Fix this by adjusting to use the post-increment trip counter. 2746 2747 bool progress = true; 2748 while (progress) { 2749 progress = false; 2750 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) { 2751 Node* use = phi->fast_out(i); // User of trip-counter 2752 if (!has_ctrl(use)) continue; 2753 Node *u_ctrl = get_ctrl(use); 2754 if (use->is_Phi()) { 2755 u_ctrl = NULL; 2756 for (uint j = 1; j < use->req(); j++) 2757 if (use->in(j) == phi) 2758 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j)); 2759 } 2760 IdealLoopTree *u_loop = get_loop(u_ctrl); 2761 // Look for loop-invariant use 2762 if (u_loop == loop) continue; 2763 if (loop->is_member(u_loop)) continue; 2764 // Check that use is live out the bottom. Assuming the trip-counter 2765 // update is right at the bottom, uses of of the loop middle are ok. 2766 if (dom_lca(exit, u_ctrl) != exit) continue; 2767 // Hit! Refactor use to use the post-incremented tripcounter. 2768 // Compute a post-increment tripcounter. 2769 Node *opaq = new (C) Opaque2Node( C, cle->incr() ); 2770 register_new_node( opaq, u_ctrl ); 2771 Node *neg_stride = _igvn.intcon(-cle->stride_con()); 2772 set_ctrl(neg_stride, C->root()); 2773 Node *post = new (C) AddINode( opaq, neg_stride); 2774 register_new_node( post, u_ctrl ); 2775 _igvn.rehash_node_delayed(use); 2776 for (uint j = 1; j < use->req(); j++) { 2777 if (use->in(j) == phi) 2778 use->set_req(j, post); 2779 } 2780 // Since DU info changed, rerun loop 2781 progress = true; 2782 break; 2783 } 2784 } 2785 2786 }