1 /* 2 * Copyright (c) 1999, 2016, 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 "memory/resourceArea.hpp" 28 #include "opto/addnode.hpp" 29 #include "opto/castnode.hpp" 30 #include "opto/connode.hpp" 31 #include "opto/castnode.hpp" 32 #include "opto/divnode.hpp" 33 #include "opto/loopnode.hpp" 34 #include "opto/matcher.hpp" 35 #include "opto/mulnode.hpp" 36 #include "opto/movenode.hpp" 37 #include "opto/opaquenode.hpp" 38 #include "opto/rootnode.hpp" 39 #include "opto/subnode.hpp" 40 41 //============================================================================= 42 //------------------------------split_thru_phi--------------------------------- 43 // Split Node 'n' through merge point if there is enough win. 44 Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) { 45 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) { 46 // ConvI2L may have type information on it which is unsafe to push up 47 // so disable this for now 48 return NULL; 49 } 50 51 // Splitting range check CastIIs through a loop induction Phi can 52 // cause new Phis to be created that are left unrelated to the loop 53 // induction Phi and prevent optimizations (vectorization) 54 if (n->Opcode() == Op_CastII && n->as_CastII()->has_range_check() && 55 region->is_CountedLoop() && n->in(1) == region->as_CountedLoop()->phi()) { 56 return NULL; 57 } 58 59 int wins = 0; 60 assert(!n->is_CFG(), ""); 61 assert(region->is_Region(), ""); 62 63 const Type* type = n->bottom_type(); 64 const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr(); 65 Node *phi; 66 if (t_oop != NULL && t_oop->is_known_instance_field()) { 67 int iid = t_oop->instance_id(); 68 int index = C->get_alias_index(t_oop); 69 int offset = t_oop->offset(); 70 phi = new PhiNode(region, type, NULL, iid, index, offset); 71 } else { 72 phi = PhiNode::make_blank(region, n); 73 } 74 uint old_unique = C->unique(); 75 for (uint i = 1; i < region->req(); i++) { 76 Node *x; 77 Node* the_clone = NULL; 78 if (region->in(i) == C->top()) { 79 x = C->top(); // Dead path? Use a dead data op 80 } else { 81 x = n->clone(); // Else clone up the data op 82 the_clone = x; // Remember for possible deletion. 83 // Alter data node to use pre-phi inputs 84 if (n->in(0) == region) 85 x->set_req( 0, region->in(i) ); 86 for (uint j = 1; j < n->req(); j++) { 87 Node *in = n->in(j); 88 if (in->is_Phi() && in->in(0) == region) 89 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone 90 } 91 } 92 // Check for a 'win' on some paths 93 const Type *t = x->Value(&_igvn); 94 95 bool singleton = t->singleton(); 96 97 // A TOP singleton indicates that there are no possible values incoming 98 // along a particular edge. In most cases, this is OK, and the Phi will 99 // be eliminated later in an Ideal call. However, we can't allow this to 100 // happen if the singleton occurs on loop entry, as the elimination of 101 // the PhiNode may cause the resulting node to migrate back to a previous 102 // loop iteration. 103 if (singleton && t == Type::TOP) { 104 // Is_Loop() == false does not confirm the absence of a loop (e.g., an 105 // irreducible loop may not be indicated by an affirmative is_Loop()); 106 // therefore, the only top we can split thru a phi is on a backedge of 107 // a loop. 108 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 109 } 110 111 if (singleton) { 112 wins++; 113 x = ((PhaseGVN&)_igvn).makecon(t); 114 } else { 115 // We now call Identity to try to simplify the cloned node. 116 // Note that some Identity methods call phase->type(this). 117 // Make sure that the type array is big enough for 118 // our new node, even though we may throw the node away. 119 // (Note: This tweaking with igvn only works because x is a new node.) 120 _igvn.set_type(x, t); 121 // If x is a TypeNode, capture any more-precise type permanently into Node 122 // otherwise it will be not updated during igvn->transform since 123 // igvn->type(x) is set to x->Value() already. 124 x->raise_bottom_type(t); 125 Node *y = x->Identity(&_igvn); 126 if (y != x) { 127 wins++; 128 x = y; 129 } else { 130 y = _igvn.hash_find(x); 131 if (y) { 132 wins++; 133 x = y; 134 } else { 135 // Else x is a new node we are keeping 136 // We do not need register_new_node_with_optimizer 137 // because set_type has already been called. 138 _igvn._worklist.push(x); 139 } 140 } 141 } 142 if (x != the_clone && the_clone != NULL) 143 _igvn.remove_dead_node(the_clone); 144 phi->set_req( i, x ); 145 } 146 // Too few wins? 147 if (wins <= policy) { 148 _igvn.remove_dead_node(phi); 149 return NULL; 150 } 151 152 // Record Phi 153 register_new_node( phi, region ); 154 155 for (uint i2 = 1; i2 < phi->req(); i2++) { 156 Node *x = phi->in(i2); 157 // If we commoned up the cloned 'x' with another existing Node, 158 // the existing Node picks up a new use. We need to make the 159 // existing Node occur higher up so it dominates its uses. 160 Node *old_ctrl; 161 IdealLoopTree *old_loop; 162 163 if (x->is_Con()) { 164 // Constant's control is always root. 165 set_ctrl(x, C->root()); 166 continue; 167 } 168 // The occasional new node 169 if (x->_idx >= old_unique) { // Found a new, unplaced node? 170 old_ctrl = NULL; 171 old_loop = NULL; // Not in any prior loop 172 } else { 173 old_ctrl = get_ctrl(x); 174 old_loop = get_loop(old_ctrl); // Get prior loop 175 } 176 // New late point must dominate new use 177 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2)); 178 if (new_ctrl == old_ctrl) // Nothing is changed 179 continue; 180 181 IdealLoopTree *new_loop = get_loop(new_ctrl); 182 183 // Don't move x into a loop if its uses are 184 // outside of loop. Otherwise x will be cloned 185 // for each use outside of this loop. 186 IdealLoopTree *use_loop = get_loop(region); 187 if (!new_loop->is_member(use_loop) && 188 (old_loop == NULL || !new_loop->is_member(old_loop))) { 189 // Take early control, later control will be recalculated 190 // during next iteration of loop optimizations. 191 new_ctrl = get_early_ctrl(x); 192 new_loop = get_loop(new_ctrl); 193 } 194 // Set new location 195 set_ctrl(x, new_ctrl); 196 // If changing loop bodies, see if we need to collect into new body 197 if (old_loop != new_loop) { 198 if (old_loop && !old_loop->_child) 199 old_loop->_body.yank(x); 200 if (!new_loop->_child) 201 new_loop->_body.push(x); // Collect body info 202 } 203 } 204 205 return phi; 206 } 207 208 //------------------------------dominated_by------------------------------------ 209 // Replace the dominated test with an obvious true or false. Place it on the 210 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the 211 // live path up to the dominating control. 212 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff, bool flip, bool exclude_loop_predicate ) { 213 if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); } 214 215 // prevdom is the dominating projection of the dominating test. 216 assert( iff->is_If(), "" ); 217 assert(iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd || iff->Opcode() == Op_RangeCheck, "Check this code when new subtype is added"); 218 int pop = prevdom->Opcode(); 219 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" ); 220 if (flip) { 221 if (pop == Op_IfTrue) 222 pop = Op_IfFalse; 223 else 224 pop = Op_IfTrue; 225 } 226 // 'con' is set to true or false to kill the dominated test. 227 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO); 228 set_ctrl(con, C->root()); // Constant gets a new use 229 // Hack the dominated test 230 _igvn.replace_input_of(iff, 1, con); 231 232 // If I dont have a reachable TRUE and FALSE path following the IfNode then 233 // I can assume this path reaches an infinite loop. In this case it's not 234 // important to optimize the data Nodes - either the whole compilation will 235 // be tossed or this path (and all data Nodes) will go dead. 236 if (iff->outcnt() != 2) return; 237 238 // Make control-dependent data Nodes on the live path (path that will remain 239 // once the dominated IF is removed) become control-dependent on the 240 // dominating projection. 241 Node* dp = iff->as_If()->proj_out(pop == Op_IfTrue); 242 243 // Loop predicates may have depending checks which should not 244 // be skipped. For example, range check predicate has two checks 245 // for lower and upper bounds. 246 if (dp == NULL) 247 return; 248 249 ProjNode* dp_proj = dp->as_Proj(); 250 ProjNode* unc_proj = iff->as_If()->proj_out(1 - dp_proj->_con)->as_Proj(); 251 if (exclude_loop_predicate && 252 (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) != NULL || 253 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check) != NULL)) { 254 // If this is a range check (IfNode::is_range_check), do not 255 // reorder because Compile::allow_range_check_smearing might have 256 // changed the check. 257 return; // Let IGVN transformation change control dependence. 258 } 259 260 IdealLoopTree *old_loop = get_loop(dp); 261 262 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { 263 Node* cd = dp->fast_out(i); // Control-dependent node 264 if (cd->depends_only_on_test()) { 265 assert(cd->in(0) == dp, ""); 266 _igvn.replace_input_of(cd, 0, prevdom); 267 set_early_ctrl(cd); 268 IdealLoopTree *new_loop = get_loop(get_ctrl(cd)); 269 if (old_loop != new_loop) { 270 if (!old_loop->_child) old_loop->_body.yank(cd); 271 if (!new_loop->_child) new_loop->_body.push(cd); 272 } 273 --i; 274 --imax; 275 } 276 } 277 } 278 279 //------------------------------has_local_phi_input---------------------------- 280 // Return TRUE if 'n' has Phi inputs from its local block and no other 281 // block-local inputs (all non-local-phi inputs come from earlier blocks) 282 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) { 283 Node *n_ctrl = get_ctrl(n); 284 // See if some inputs come from a Phi in this block, or from before 285 // this block. 286 uint i; 287 for( i = 1; i < n->req(); i++ ) { 288 Node *phi = n->in(i); 289 if( phi->is_Phi() && phi->in(0) == n_ctrl ) 290 break; 291 } 292 if( i >= n->req() ) 293 return NULL; // No Phi inputs; nowhere to clone thru 294 295 // Check for inputs created between 'n' and the Phi input. These 296 // must split as well; they have already been given the chance 297 // (courtesy of a post-order visit) and since they did not we must 298 // recover the 'cost' of splitting them by being very profitable 299 // when splitting 'n'. Since this is unlikely we simply give up. 300 for( i = 1; i < n->req(); i++ ) { 301 Node *m = n->in(i); 302 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) { 303 // We allow the special case of AddP's with no local inputs. 304 // This allows us to split-up address expressions. 305 if (m->is_AddP() && 306 get_ctrl(m->in(2)) != n_ctrl && 307 get_ctrl(m->in(3)) != n_ctrl) { 308 // Move the AddP up to dominating point 309 set_ctrl_and_loop(m, find_non_split_ctrl(idom(n_ctrl))); 310 continue; 311 } 312 return NULL; 313 } 314 } 315 316 return n_ctrl; 317 } 318 319 //------------------------------remix_address_expressions---------------------- 320 // Rework addressing expressions to get the most loop-invariant stuff 321 // moved out. We'd like to do all associative operators, but it's especially 322 // important (common) to do address expressions. 323 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) { 324 if (!has_ctrl(n)) return NULL; 325 Node *n_ctrl = get_ctrl(n); 326 IdealLoopTree *n_loop = get_loop(n_ctrl); 327 328 // See if 'n' mixes loop-varying and loop-invariant inputs and 329 // itself is loop-varying. 330 331 // Only interested in binary ops (and AddP) 332 if( n->req() < 3 || n->req() > 4 ) return NULL; 333 334 Node *n1_ctrl = get_ctrl(n->in( 1)); 335 Node *n2_ctrl = get_ctrl(n->in( 2)); 336 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3)); 337 IdealLoopTree *n1_loop = get_loop( n1_ctrl ); 338 IdealLoopTree *n2_loop = get_loop( n2_ctrl ); 339 IdealLoopTree *n3_loop = get_loop( n3_ctrl ); 340 341 // Does one of my inputs spin in a tighter loop than self? 342 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) || 343 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) || 344 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) ) 345 return NULL; // Leave well enough alone 346 347 // Is at least one of my inputs loop-invariant? 348 if( n1_loop == n_loop && 349 n2_loop == n_loop && 350 n3_loop == n_loop ) 351 return NULL; // No loop-invariant inputs 352 353 354 int n_op = n->Opcode(); 355 356 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2). 357 if( n_op == Op_LShiftI ) { 358 // Scale is loop invariant 359 Node *scale = n->in(2); 360 Node *scale_ctrl = get_ctrl(scale); 361 IdealLoopTree *scale_loop = get_loop(scale_ctrl ); 362 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) ) 363 return NULL; 364 const TypeInt *scale_t = scale->bottom_type()->isa_int(); 365 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 ) 366 return NULL; // Dont bother with byte/short masking 367 // Add must vary with loop (else shift would be loop-invariant) 368 Node *add = n->in(1); 369 Node *add_ctrl = get_ctrl(add); 370 IdealLoopTree *add_loop = get_loop(add_ctrl); 371 //assert( n_loop == add_loop, "" ); 372 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops 373 374 // Convert I-V into I+ (0-V); same for V-I 375 if( add->Opcode() == Op_SubI && 376 _igvn.type( add->in(1) ) != TypeInt::ZERO ) { 377 Node *zero = _igvn.intcon(0); 378 set_ctrl(zero, C->root()); 379 Node *neg = new SubINode( _igvn.intcon(0), add->in(2) ); 380 register_new_node( neg, get_ctrl(add->in(2) ) ); 381 add = new AddINode( add->in(1), neg ); 382 register_new_node( add, add_ctrl ); 383 } 384 if( add->Opcode() != Op_AddI ) return NULL; 385 // See if one add input is loop invariant 386 Node *add_var = add->in(1); 387 Node *add_var_ctrl = get_ctrl(add_var); 388 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 389 Node *add_invar = add->in(2); 390 Node *add_invar_ctrl = get_ctrl(add_invar); 391 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl ); 392 if( add_var_loop == n_loop ) { 393 } else if( add_invar_loop == n_loop ) { 394 // Swap to find the invariant part 395 add_invar = add_var; 396 add_invar_ctrl = add_var_ctrl; 397 add_invar_loop = add_var_loop; 398 add_var = add->in(2); 399 Node *add_var_ctrl = get_ctrl(add_var); 400 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 401 } else // Else neither input is loop invariant 402 return NULL; 403 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) ) 404 return NULL; // No invariant part of the add? 405 406 // Yes! Reshape address expression! 407 Node *inv_scale = new LShiftINode( add_invar, scale ); 408 Node *inv_scale_ctrl = 409 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ? 410 add_invar_ctrl : scale_ctrl; 411 register_new_node( inv_scale, inv_scale_ctrl ); 412 Node *var_scale = new LShiftINode( add_var, scale ); 413 register_new_node( var_scale, n_ctrl ); 414 Node *var_add = new AddINode( var_scale, inv_scale ); 415 register_new_node( var_add, n_ctrl ); 416 _igvn.replace_node( n, var_add ); 417 return var_add; 418 } 419 420 // Replace (I+V) with (V+I) 421 if( n_op == Op_AddI || 422 n_op == Op_AddL || 423 n_op == Op_AddF || 424 n_op == Op_AddD || 425 n_op == Op_MulI || 426 n_op == Op_MulL || 427 n_op == Op_MulF || 428 n_op == Op_MulD ) { 429 if( n2_loop == n_loop ) { 430 assert( n1_loop != n_loop, "" ); 431 n->swap_edges(1, 2); 432 } 433 } 434 435 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V), 436 // but not if I2 is a constant. 437 if( n_op == Op_AddP ) { 438 if( n2_loop == n_loop && n3_loop != n_loop ) { 439 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) { 440 Node *n22_ctrl = get_ctrl(n->in(2)->in(2)); 441 Node *n23_ctrl = get_ctrl(n->in(2)->in(3)); 442 IdealLoopTree *n22loop = get_loop( n22_ctrl ); 443 IdealLoopTree *n23_loop = get_loop( n23_ctrl ); 444 if( n22loop != n_loop && n22loop->is_member(n_loop) && 445 n23_loop == n_loop ) { 446 Node *add1 = new AddPNode( n->in(1), n->in(2)->in(2), n->in(3) ); 447 // Stuff new AddP in the loop preheader 448 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) ); 449 Node *add2 = new AddPNode( n->in(1), add1, n->in(2)->in(3) ); 450 register_new_node( add2, n_ctrl ); 451 _igvn.replace_node( n, add2 ); 452 return add2; 453 } 454 } 455 } 456 457 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V) 458 if (n2_loop != n_loop && n3_loop == n_loop) { 459 if (n->in(3)->Opcode() == Op_AddX) { 460 Node *V = n->in(3)->in(1); 461 Node *I = n->in(3)->in(2); 462 if (is_member(n_loop,get_ctrl(V))) { 463 } else { 464 Node *tmp = V; V = I; I = tmp; 465 } 466 if (!is_member(n_loop,get_ctrl(I))) { 467 Node *add1 = new AddPNode(n->in(1), n->in(2), I); 468 // Stuff new AddP in the loop preheader 469 register_new_node(add1, n_loop->_head->in(LoopNode::EntryControl)); 470 Node *add2 = new AddPNode(n->in(1), add1, V); 471 register_new_node(add2, n_ctrl); 472 _igvn.replace_node(n, add2); 473 return add2; 474 } 475 } 476 } 477 } 478 479 return NULL; 480 } 481 482 //------------------------------conditional_move------------------------------- 483 // Attempt to replace a Phi with a conditional move. We have some pretty 484 // strict profitability requirements. All Phis at the merge point must 485 // be converted, so we can remove the control flow. We need to limit the 486 // number of c-moves to a small handful. All code that was in the side-arms 487 // of the CFG diamond is now speculatively executed. This code has to be 488 // "cheap enough". We are pretty much limited to CFG diamonds that merge 489 // 1 or 2 items with a total of 1 or 2 ops executed speculatively. 490 Node *PhaseIdealLoop::conditional_move( Node *region ) { 491 492 assert(region->is_Region(), "sanity check"); 493 if (region->req() != 3) return NULL; 494 495 // Check for CFG diamond 496 Node *lp = region->in(1); 497 Node *rp = region->in(2); 498 if (!lp || !rp) return NULL; 499 Node *lp_c = lp->in(0); 500 if (lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If()) return NULL; 501 IfNode *iff = lp_c->as_If(); 502 503 // Check for ops pinned in an arm of the diamond. 504 // Can't remove the control flow in this case 505 if (lp->outcnt() > 1) return NULL; 506 if (rp->outcnt() > 1) return NULL; 507 508 IdealLoopTree* r_loop = get_loop(region); 509 assert(r_loop == get_loop(iff), "sanity"); 510 // Always convert to CMOVE if all results are used only outside this loop. 511 bool used_inside_loop = (r_loop == _ltree_root); 512 513 // Check profitability 514 int cost = 0; 515 int phis = 0; 516 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 517 Node *out = region->fast_out(i); 518 if (!out->is_Phi()) continue; // Ignore other control edges, etc 519 phis++; 520 PhiNode* phi = out->as_Phi(); 521 BasicType bt = phi->type()->basic_type(); 522 switch (bt) { 523 case T_DOUBLE: 524 if (C->use_cmove()) { 525 continue; //TODO: maybe we want to add some cost 526 } 527 case T_FLOAT: { 528 cost += Matcher::float_cmove_cost(); // Could be very expensive 529 break; 530 } 531 case T_LONG: { 532 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's 533 } 534 case T_INT: // These all CMOV fine 535 case T_ADDRESS: { // (RawPtr) 536 cost++; 537 break; 538 } 539 case T_NARROWOOP: // Fall through 540 case T_OBJECT: { // Base oops are OK, but not derived oops 541 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr(); 542 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a 543 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus 544 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we 545 // have a Phi for the base here that we convert to a CMOVE all is well 546 // and good. But if the base is dead, we'll not make a CMOVE. Later 547 // the allocator will have to produce a base by creating a CMOVE of the 548 // relevant bases. This puts the allocator in the business of 549 // manufacturing expensive instructions, generally a bad plan. 550 // Just Say No to Conditionally-Moved Derived Pointers. 551 if (tp && tp->offset() != 0) 552 return NULL; 553 cost++; 554 break; 555 } 556 default: 557 return NULL; // In particular, can't do memory or I/O 558 } 559 // Add in cost any speculative ops 560 for (uint j = 1; j < region->req(); j++) { 561 Node *proj = region->in(j); 562 Node *inp = phi->in(j); 563 if (get_ctrl(inp) == proj) { // Found local op 564 cost++; 565 // Check for a chain of dependent ops; these will all become 566 // speculative in a CMOV. 567 for (uint k = 1; k < inp->req(); k++) 568 if (get_ctrl(inp->in(k)) == proj) 569 cost += ConditionalMoveLimit; // Too much speculative goo 570 } 571 } 572 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode. 573 // This will likely Split-If, a higher-payoff operation. 574 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) { 575 Node* use = phi->fast_out(k); 576 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr()) 577 cost += ConditionalMoveLimit; 578 // Is there a use inside the loop? 579 // Note: check only basic types since CMoveP is pinned. 580 if (!used_inside_loop && is_java_primitive(bt)) { 581 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use); 582 if (r_loop == u_loop || r_loop->is_member(u_loop)) { 583 used_inside_loop = true; 584 } 585 } 586 } 587 }//for 588 Node* bol = iff->in(1); 589 assert(bol->Opcode() == Op_Bool, ""); 590 int cmp_op = bol->in(1)->Opcode(); 591 // It is expensive to generate flags from a float compare. 592 // Avoid duplicated float compare. 593 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL; 594 595 float infrequent_prob = PROB_UNLIKELY_MAG(3); 596 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop. 597 if (used_inside_loop) { 598 if (cost >= ConditionalMoveLimit) return NULL; // Too much goo 599 600 // BlockLayoutByFrequency optimization moves infrequent branch 601 // from hot path. No point in CMOV'ing in such case (110 is used 602 // instead of 100 to take into account not exactness of float value). 603 if (BlockLayoutByFrequency) { 604 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f); 605 } 606 } 607 // Check for highly predictable branch. No point in CMOV'ing if 608 // we are going to predict accurately all the time. 609 if (C->use_cmove() && cmp_op == Op_CmpD) ;//keep going 610 else if (iff->_prob < infrequent_prob || 611 iff->_prob > (1.0f - infrequent_prob)) 612 return NULL; 613 614 // -------------- 615 // Now replace all Phis with CMOV's 616 Node *cmov_ctrl = iff->in(0); 617 uint flip = (lp->Opcode() == Op_IfTrue); 618 while (1) { 619 PhiNode* phi = NULL; 620 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 621 Node *out = region->fast_out(i); 622 if (out->is_Phi()) { 623 phi = out->as_Phi(); 624 break; 625 } 626 } 627 if (phi == NULL) break; 628 if (PrintOpto && VerifyLoopOptimizations) { tty->print_cr("CMOV"); } 629 // Move speculative ops 630 for (uint j = 1; j < region->req(); j++) { 631 Node *proj = region->in(j); 632 Node *inp = phi->in(j); 633 if (get_ctrl(inp) == proj) { // Found local op 634 #ifndef PRODUCT 635 if (PrintOpto && VerifyLoopOptimizations) { 636 tty->print(" speculate: "); 637 inp->dump(); 638 } 639 #endif 640 set_ctrl(inp, cmov_ctrl); 641 } 642 } 643 Node *cmov = CMoveNode::make(cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi)); 644 register_new_node( cmov, cmov_ctrl ); 645 _igvn.replace_node( phi, cmov ); 646 #ifndef PRODUCT 647 if (TraceLoopOpts) { 648 tty->print("CMOV "); 649 r_loop->dump_head(); 650 if (Verbose) { 651 bol->in(1)->dump(1); 652 cmov->dump(1); 653 } 654 } 655 if (VerifyLoopOptimizations) verify(); 656 #endif 657 } 658 659 // The useless CFG diamond will fold up later; see the optimization in 660 // RegionNode::Ideal. 661 _igvn._worklist.push(region); 662 663 return iff->in(1); 664 } 665 666 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) { 667 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 668 Node* u = m->fast_out(i); 669 if (u->is_CFG()) { 670 if (u->Opcode() == Op_NeverBranch) { 671 u = ((NeverBranchNode*)u)->proj_out(0); 672 enqueue_cfg_uses(u, wq); 673 } else { 674 wq.push(u); 675 } 676 } 677 } 678 } 679 680 // Try moving a store out of a loop, right before the loop 681 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) { 682 // Store has to be first in the loop body 683 IdealLoopTree *n_loop = get_loop(n_ctrl); 684 if (n->is_Store() && n_loop != _ltree_root && n_loop->is_loop() && n->in(0) != NULL) { 685 Node* address = n->in(MemNode::Address); 686 Node* value = n->in(MemNode::ValueIn); 687 Node* mem = n->in(MemNode::Memory); 688 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 689 IdealLoopTree* value_loop = get_loop(get_ctrl(value)); 690 691 // - address and value must be loop invariant 692 // - memory must be a memory Phi for the loop 693 // - Store must be the only store on this memory slice in the 694 // loop: if there's another store following this one then value 695 // written at iteration i by the second store could be overwritten 696 // at iteration i+n by the first store: it's not safe to move the 697 // first store out of the loop 698 // - nothing must observe the memory Phi: it guarantees no read 699 // before the store, we are also guaranteed the store post 700 // dominates the loop head (ignoring a possible early 701 // exit). Otherwise there would be extra Phi involved between the 702 // loop's Phi and the store. 703 // - there must be no early exit from the loop before the Store 704 // (such an exit most of the time would be an extra use of the 705 // memory Phi but sometimes is a bottom memory Phi that takes the 706 // store as input). 707 708 if (!n_loop->is_member(address_loop) && 709 !n_loop->is_member(value_loop) && 710 mem->is_Phi() && mem->in(0) == n_loop->_head && 711 mem->outcnt() == 1 && 712 mem->in(LoopNode::LoopBackControl) == n) { 713 714 assert(n_loop->_tail != NULL, "need a tail"); 715 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop"); 716 717 // Verify that there's no early exit of the loop before the store. 718 bool ctrl_ok = false; 719 { 720 // Follow control from loop head until n, we exit the loop or 721 // we reach the tail 722 ResourceMark rm; 723 Unique_Node_List wq; 724 wq.push(n_loop->_head); 725 726 for (uint next = 0; next < wq.size(); ++next) { 727 Node *m = wq.at(next); 728 if (m == n->in(0)) { 729 ctrl_ok = true; 730 continue; 731 } 732 assert(!has_ctrl(m), "should be CFG"); 733 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) { 734 ctrl_ok = false; 735 break; 736 } 737 enqueue_cfg_uses(m, wq); 738 if (wq.size() > 10) { 739 ctrl_ok = false; 740 break; 741 } 742 } 743 } 744 if (ctrl_ok) { 745 // move the Store 746 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem); 747 _igvn.replace_input_of(n, 0, n_loop->_head->in(LoopNode::EntryControl)); 748 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl)); 749 // Disconnect the phi now. An empty phi can confuse other 750 // optimizations in this pass of loop opts. 751 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl)); 752 n_loop->_body.yank(mem); 753 754 IdealLoopTree* new_loop = get_loop(n->in(0)); 755 set_ctrl_and_loop(n, n->in(0)); 756 757 return n; 758 } 759 } 760 } 761 return NULL; 762 } 763 764 // Try moving a store out of a loop, right after the loop 765 void PhaseIdealLoop::try_move_store_after_loop(Node* n) { 766 if (n->is_Store() && n->in(0) != NULL) { 767 Node *n_ctrl = get_ctrl(n); 768 IdealLoopTree *n_loop = get_loop(n_ctrl); 769 // Store must be in a loop 770 if (n_loop != _ltree_root && !n_loop->_irreducible) { 771 Node* address = n->in(MemNode::Address); 772 Node* value = n->in(MemNode::ValueIn); 773 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 774 // address must be loop invariant 775 if (!n_loop->is_member(address_loop)) { 776 // Store must be last on this memory slice in the loop and 777 // nothing in the loop must observe it 778 Node* phi = NULL; 779 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 780 Node* u = n->fast_out(i); 781 if (has_ctrl(u)) { // control use? 782 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 783 if (!n_loop->is_member(u_loop)) { 784 continue; 785 } 786 if (u->is_Phi() && u->in(0) == n_loop->_head) { 787 assert(_igvn.type(u) == Type::MEMORY, "bad phi"); 788 // multiple phis on the same slice are possible 789 if (phi != NULL) { 790 return; 791 } 792 phi = u; 793 continue; 794 } 795 } 796 return; 797 } 798 if (phi != NULL) { 799 // Nothing in the loop before the store (next iteration) 800 // must observe the stored value 801 bool mem_ok = true; 802 { 803 ResourceMark rm; 804 Unique_Node_List wq; 805 wq.push(phi); 806 for (uint next = 0; next < wq.size() && mem_ok; ++next) { 807 Node *m = wq.at(next); 808 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) { 809 Node* u = m->fast_out(i); 810 if (u->is_Store() || u->is_Phi()) { 811 if (u != n) { 812 wq.push(u); 813 mem_ok = (wq.size() <= 10); 814 } 815 } else { 816 mem_ok = false; 817 break; 818 } 819 } 820 } 821 } 822 if (mem_ok) { 823 // Move the Store out of the loop creating clones along 824 // all paths out of the loop that observe the stored value 825 _igvn.rehash_node_delayed(phi); 826 int count = phi->replace_edge(n, n->in(MemNode::Memory)); 827 assert(count > 0, "inconsistent phi"); 828 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 829 Node* u = n->fast_out(i); 830 Node* c = get_ctrl(u); 831 832 if (u->is_Phi()) { 833 c = u->in(0)->in(u->find_edge(n)); 834 } 835 IdealLoopTree *u_loop = get_loop(c); 836 assert (!n_loop->is_member(u_loop), "only the phi should have been a use in the loop"); 837 while(true) { 838 Node* next_c = find_non_split_ctrl(idom(c)); 839 if (n_loop->is_member(get_loop(next_c))) { 840 break; 841 } 842 c = next_c; 843 } 844 845 Node* st = n->clone(); 846 st->set_req(0, c); 847 _igvn.register_new_node_with_optimizer(st); 848 849 set_ctrl(st, c); 850 IdealLoopTree* new_loop = get_loop(c); 851 assert(new_loop != n_loop, "should be moved out of loop"); 852 if (new_loop->_child == NULL) new_loop->_body.push(st); 853 854 _igvn.replace_input_of(u, u->find_edge(n), st); 855 --imax; 856 --i; 857 } 858 859 860 assert(n->outcnt() == 0, "all uses should be gone"); 861 _igvn.replace_input_of(n, MemNode::Memory, C->top()); 862 // Disconnect the phi now. An empty phi can confuse other 863 // optimizations in this pass of loop opts.. 864 if (phi->in(LoopNode::LoopBackControl) == phi) { 865 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl)); 866 n_loop->_body.yank(phi); 867 } 868 } 869 } 870 } 871 } 872 } 873 } 874 875 //------------------------------split_if_with_blocks_pre----------------------- 876 // Do the real work in a non-recursive function. Data nodes want to be 877 // cloned in the pre-order so they can feed each other nicely. 878 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) { 879 // Cloning these guys is unlikely to win 880 int n_op = n->Opcode(); 881 if( n_op == Op_MergeMem ) return n; 882 if( n->is_Proj() ) return n; 883 // Do not clone-up CmpFXXX variations, as these are always 884 // followed by a CmpI 885 if( n->is_Cmp() ) return n; 886 // Attempt to use a conditional move instead of a phi/branch 887 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) { 888 Node *cmov = conditional_move( n ); 889 if( cmov ) return cmov; 890 } 891 if( n->is_CFG() || n->is_LoadStore() ) 892 return n; 893 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd 894 n_op == Op_Opaque2 ) { 895 if( !C->major_progress() ) // If chance of no more loop opts... 896 _igvn._worklist.push(n); // maybe we'll remove them 897 return n; 898 } 899 900 if( n->is_Con() ) return n; // No cloning for Con nodes 901 902 Node *n_ctrl = get_ctrl(n); 903 if( !n_ctrl ) return n; // Dead node 904 905 Node* res = try_move_store_before_loop(n, n_ctrl); 906 if (res != NULL) { 907 return n; 908 } 909 910 // Attempt to remix address expressions for loop invariants 911 Node *m = remix_address_expressions( n ); 912 if( m ) return m; 913 914 if (n->is_ConstraintCast()) { 915 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(this); 916 if (dom_cast != NULL) { 917 _igvn.replace_node(n, dom_cast); 918 return dom_cast; 919 } 920 } 921 922 // Determine if the Node has inputs from some local Phi. 923 // Returns the block to clone thru. 924 Node *n_blk = has_local_phi_input( n ); 925 if( !n_blk ) return n; 926 927 // Do not clone the trip counter through on a CountedLoop 928 // (messes up the canonical shape). 929 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n; 930 931 // Check for having no control input; not pinned. Allow 932 // dominating control. 933 if (n->in(0)) { 934 Node *dom = idom(n_blk); 935 if (dom_lca(n->in(0), dom) != n->in(0)) { 936 return n; 937 } 938 } 939 // Policy: when is it profitable. You must get more wins than 940 // policy before it is considered profitable. Policy is usually 0, 941 // so 1 win is considered profitable. Big merges will require big 942 // cloning, so get a larger policy. 943 int policy = n_blk->req() >> 2; 944 945 // If the loop is a candidate for range check elimination, 946 // delay splitting through it's phi until a later loop optimization 947 if (n_blk->is_CountedLoop()) { 948 IdealLoopTree *lp = get_loop(n_blk); 949 if (lp && lp->_rce_candidate) { 950 return n; 951 } 952 } 953 954 // Use same limit as split_if_with_blocks_post 955 if( C->live_nodes() > 35000 ) return n; // Method too big 956 957 // Split 'n' through the merge point if it is profitable 958 Node *phi = split_thru_phi( n, n_blk, policy ); 959 if (!phi) return n; 960 961 // Found a Phi to split thru! 962 // Replace 'n' with the new phi 963 _igvn.replace_node( n, phi ); 964 // Moved a load around the loop, 'en-registering' something. 965 if (n_blk->is_Loop() && n->is_Load() && 966 !phi->in(LoopNode::LoopBackControl)->is_Load()) 967 C->set_major_progress(); 968 969 return phi; 970 } 971 972 static bool merge_point_too_heavy(Compile* C, Node* region) { 973 // Bail out if the region and its phis have too many users. 974 int weight = 0; 975 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 976 weight += region->fast_out(i)->outcnt(); 977 } 978 int nodes_left = C->max_node_limit() - C->live_nodes(); 979 if (weight * 8 > nodes_left) { 980 if (PrintOpto) { 981 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight); 982 } 983 return true; 984 } else { 985 return false; 986 } 987 } 988 989 static bool merge_point_safe(Node* region) { 990 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode 991 // having a PhiNode input. This sidesteps the dangerous case where the split 992 // ConvI2LNode may become TOP if the input Value() does not 993 // overlap the ConvI2L range, leaving a node which may not dominate its 994 // uses. 995 // A better fix for this problem can be found in the BugTraq entry, but 996 // expediency for Mantis demands this hack. 997 // 6855164: If the merge point has a FastLockNode with a PhiNode input, we stop 998 // split_if_with_blocks from splitting a block because we could not move around 999 // the FastLockNode. 1000 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1001 Node* n = region->fast_out(i); 1002 if (n->is_Phi()) { 1003 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1004 Node* m = n->fast_out(j); 1005 if (m->is_FastLock()) 1006 return false; 1007 #ifdef _LP64 1008 if (m->Opcode() == Op_ConvI2L) 1009 return false; 1010 if (m->is_CastII() && m->isa_CastII()->has_range_check()) { 1011 return false; 1012 } 1013 #endif 1014 } 1015 } 1016 } 1017 return true; 1018 } 1019 1020 1021 //------------------------------place_near_use--------------------------------- 1022 // Place some computation next to use but not inside inner loops. 1023 // For inner loop uses move it to the preheader area. 1024 Node *PhaseIdealLoop::place_near_use( Node *useblock ) const { 1025 IdealLoopTree *u_loop = get_loop( useblock ); 1026 return (u_loop->_irreducible || u_loop->_child) 1027 ? useblock 1028 : u_loop->_head->in(LoopNode::EntryControl); 1029 } 1030 1031 1032 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) { 1033 if (!n->is_If()) { 1034 return false; 1035 } 1036 if (!n->in(0)->is_Region()) { 1037 return false; 1038 } 1039 Node* region = n->in(0); 1040 Node* dom = idom(region); 1041 if (!dom->is_If() || dom->in(1) != n->in(1)) { 1042 return false; 1043 } 1044 IfNode* dom_if = dom->as_If(); 1045 Node* proj_true = dom_if->proj_out(1); 1046 Node* proj_false = dom_if->proj_out(0); 1047 1048 for (uint i = 1; i < region->req(); i++) { 1049 if (is_dominator(proj_true, region->in(i))) { 1050 continue; 1051 } 1052 if (is_dominator(proj_false, region->in(i))) { 1053 continue; 1054 } 1055 return false; 1056 } 1057 1058 return true; 1059 } 1060 1061 bool PhaseIdealLoop::can_split_if(Node *n_ctrl) { 1062 if (C->live_nodes() > 35000) { 1063 return false; // Method too big 1064 } 1065 1066 // Do not do 'split-if' if irreducible loops are present. 1067 if (_has_irreducible_loops) { 1068 return false; 1069 } 1070 1071 if (merge_point_too_heavy(C, n_ctrl)) { 1072 return false; 1073 } 1074 1075 // Do not do 'split-if' if some paths are dead. First do dead code 1076 // elimination and then see if its still profitable. 1077 for (uint i = 1; i < n_ctrl->req(); i++) { 1078 if (n_ctrl->in(i) == C->top()) { 1079 return false; 1080 } 1081 } 1082 1083 // If trying to do a 'Split-If' at the loop head, it is only 1084 // profitable if the cmp folds up on BOTH paths. Otherwise we 1085 // risk peeling a loop forever. 1086 1087 // CNC - Disabled for now. Requires careful handling of loop 1088 // body selection for the cloned code. Also, make sure we check 1089 // for any input path not being in the same loop as n_ctrl. For 1090 // irreducible loops we cannot check for 'n_ctrl->is_Loop()' 1091 // because the alternative loop entry points won't be converted 1092 // into LoopNodes. 1093 IdealLoopTree *n_loop = get_loop(n_ctrl); 1094 for (uint j = 1; j < n_ctrl->req(); j++) { 1095 if (get_loop(n_ctrl->in(j)) != n_loop) { 1096 return false; 1097 } 1098 } 1099 1100 // Check for safety of the merge point. 1101 if (!merge_point_safe(n_ctrl)) { 1102 return false; 1103 } 1104 1105 return true; 1106 } 1107 1108 //------------------------------split_if_with_blocks_post---------------------- 1109 // Do the real work in a non-recursive function. CFG hackery wants to be 1110 // in the post-order, so it can dirty the I-DOM info and not use the dirtied 1111 // info. 1112 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) { 1113 1114 // Cloning Cmp through Phi's involves the split-if transform. 1115 // FastLock is not used by an If 1116 if (n->is_Cmp() && !n->is_FastLock()) { 1117 Node *n_ctrl = get_ctrl(n); 1118 // Determine if the Node has inputs from some local Phi. 1119 // Returns the block to clone thru. 1120 Node *n_blk = has_local_phi_input(n); 1121 if (n_blk != n_ctrl) { 1122 return; 1123 } 1124 1125 if (!can_split_if(n_ctrl)) { 1126 return; 1127 } 1128 1129 if (n->outcnt() != 1) { 1130 return; // Multiple bool's from 1 compare? 1131 } 1132 Node *bol = n->unique_out(); 1133 assert(bol->is_Bool(), "expect a bool here"); 1134 if (bol->outcnt() != 1) { 1135 return;// Multiple branches from 1 compare? 1136 } 1137 Node *iff = bol->unique_out(); 1138 1139 // Check some safety conditions 1140 if (iff->is_If()) { // Classic split-if? 1141 if (iff->in(0) != n_ctrl) { 1142 return; // Compare must be in same blk as if 1143 } 1144 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE 1145 // Can't split CMove with different control edge. 1146 if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) { 1147 return; 1148 } 1149 if (get_ctrl(iff->in(2)) == n_ctrl || 1150 get_ctrl(iff->in(3)) == n_ctrl) { 1151 return; // Inputs not yet split-up 1152 } 1153 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) { 1154 return; // Loop-invar test gates loop-varying CMOVE 1155 } 1156 } else { 1157 return; // some other kind of node, such as an Allocate 1158 } 1159 1160 // When is split-if profitable? Every 'win' on means some control flow 1161 // goes dead, so it's almost always a win. 1162 int policy = 0; 1163 // Split compare 'n' through the merge point if it is profitable 1164 Node *phi = split_thru_phi( n, n_ctrl, policy); 1165 if (!phi) { 1166 return; 1167 } 1168 1169 // Found a Phi to split thru! 1170 // Replace 'n' with the new phi 1171 _igvn.replace_node(n, phi); 1172 1173 // Now split the bool up thru the phi 1174 Node *bolphi = split_thru_phi(bol, n_ctrl, -1); 1175 guarantee(bolphi != NULL, "null boolean phi node"); 1176 1177 _igvn.replace_node(bol, bolphi); 1178 assert(iff->in(1) == bolphi, ""); 1179 1180 if (bolphi->Value(&_igvn)->singleton()) { 1181 return; 1182 } 1183 1184 // Conditional-move? Must split up now 1185 if (!iff->is_If()) { 1186 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1); 1187 _igvn.replace_node(iff, cmovphi); 1188 return; 1189 } 1190 1191 // Now split the IF 1192 do_split_if(iff); 1193 return; 1194 } 1195 1196 // Two identical ifs back to back can be merged 1197 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) { 1198 Node *n_ctrl = n->in(0); 1199 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1)); 1200 IfNode* dom_if = idom(n_ctrl)->as_If(); 1201 Node* proj_true = dom_if->proj_out(1); 1202 Node* proj_false = dom_if->proj_out(0); 1203 Node* con_true = _igvn.makecon(TypeInt::ONE); 1204 Node* con_false = _igvn.makecon(TypeInt::ZERO); 1205 1206 for (uint i = 1; i < n_ctrl->req(); i++) { 1207 if (is_dominator(proj_true, n_ctrl->in(i))) { 1208 bolphi->init_req(i, con_true); 1209 } else { 1210 assert(is_dominator(proj_false, n_ctrl->in(i)), "bad if"); 1211 bolphi->init_req(i, con_false); 1212 } 1213 } 1214 register_new_node(bolphi, n_ctrl); 1215 _igvn.replace_input_of(n, 1, bolphi); 1216 1217 // Now split the IF 1218 do_split_if(n); 1219 return; 1220 } 1221 1222 // Check for an IF ready to split; one that has its 1223 // condition codes input coming from a Phi at the block start. 1224 int n_op = n->Opcode(); 1225 1226 // Check for an IF being dominated by another IF same test 1227 if (n_op == Op_If || 1228 n_op == Op_RangeCheck) { 1229 Node *bol = n->in(1); 1230 uint max = bol->outcnt(); 1231 // Check for same test used more than once? 1232 if (max > 1 && bol->is_Bool()) { 1233 // Search up IDOMs to see if this IF is dominated. 1234 Node *cutoff = get_ctrl(bol); 1235 1236 // Now search up IDOMs till cutoff, looking for a dominating test 1237 Node *prevdom = n; 1238 Node *dom = idom(prevdom); 1239 while (dom != cutoff) { 1240 if (dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom) { 1241 // Replace the dominated test with an obvious true or false. 1242 // Place it on the IGVN worklist for later cleanup. 1243 C->set_major_progress(); 1244 dominated_by(prevdom, n, false, true); 1245 #ifndef PRODUCT 1246 if( VerifyLoopOptimizations ) verify(); 1247 #endif 1248 return; 1249 } 1250 prevdom = dom; 1251 dom = idom(prevdom); 1252 } 1253 } 1254 } 1255 1256 // See if a shared loop-varying computation has no loop-varying uses. 1257 // Happens if something is only used for JVM state in uncommon trap exits, 1258 // like various versions of induction variable+offset. Clone the 1259 // computation per usage to allow it to sink out of the loop. 1260 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about) 1261 Node *n_ctrl = get_ctrl(n); 1262 IdealLoopTree *n_loop = get_loop(n_ctrl); 1263 if( n_loop != _ltree_root ) { 1264 DUIterator_Fast imax, i = n->fast_outs(imax); 1265 for (; i < imax; i++) { 1266 Node* u = n->fast_out(i); 1267 if( !has_ctrl(u) ) break; // Found control user 1268 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 1269 if( u_loop == n_loop ) break; // Found loop-varying use 1270 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop 1271 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003 1272 } 1273 bool did_break = (i < imax); // Did we break out of the previous loop? 1274 if (!did_break && n->outcnt() > 1) { // All uses in outer loops! 1275 Node *late_load_ctrl = NULL; 1276 if (n->is_Load()) { 1277 // If n is a load, get and save the result from get_late_ctrl(), 1278 // to be later used in calculating the control for n's clones. 1279 clear_dom_lca_tags(); 1280 late_load_ctrl = get_late_ctrl(n, n_ctrl); 1281 } 1282 // If n is a load, and the late control is the same as the current 1283 // control, then the cloning of n is a pointless exercise, because 1284 // GVN will ensure that we end up where we started. 1285 if (!n->is_Load() || late_load_ctrl != n_ctrl) { 1286 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) { 1287 Node *u = n->last_out(j); // Clone private computation per use 1288 _igvn.rehash_node_delayed(u); 1289 Node *x = n->clone(); // Clone computation 1290 Node *x_ctrl = NULL; 1291 if( u->is_Phi() ) { 1292 // Replace all uses of normal nodes. Replace Phi uses 1293 // individually, so the separate Nodes can sink down 1294 // different paths. 1295 uint k = 1; 1296 while( u->in(k) != n ) k++; 1297 u->set_req( k, x ); 1298 // x goes next to Phi input path 1299 x_ctrl = u->in(0)->in(k); 1300 --j; 1301 } else { // Normal use 1302 // Replace all uses 1303 for( uint k = 0; k < u->req(); k++ ) { 1304 if( u->in(k) == n ) { 1305 u->set_req( k, x ); 1306 --j; 1307 } 1308 } 1309 x_ctrl = get_ctrl(u); 1310 } 1311 1312 // Find control for 'x' next to use but not inside inner loops. 1313 // For inner loop uses get the preheader area. 1314 x_ctrl = place_near_use(x_ctrl); 1315 1316 if (n->is_Load()) { 1317 // For loads, add a control edge to a CFG node outside of the loop 1318 // to force them to not combine and return back inside the loop 1319 // during GVN optimization (4641526). 1320 // 1321 // Because we are setting the actual control input, factor in 1322 // the result from get_late_ctrl() so we respect any 1323 // anti-dependences. (6233005). 1324 x_ctrl = dom_lca(late_load_ctrl, x_ctrl); 1325 1326 // Don't allow the control input to be a CFG splitting node. 1327 // Such nodes should only have ProjNodes as outs, e.g. IfNode 1328 // should only have IfTrueNode and IfFalseNode (4985384). 1329 x_ctrl = find_non_split_ctrl(x_ctrl); 1330 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone"); 1331 1332 x->set_req(0, x_ctrl); 1333 } 1334 register_new_node(x, x_ctrl); 1335 1336 // Some institutional knowledge is needed here: 'x' is 1337 // yanked because if the optimizer runs GVN on it all the 1338 // cloned x's will common up and undo this optimization and 1339 // be forced back in the loop. This is annoying because it 1340 // makes +VerifyOpto report false-positives on progress. I 1341 // tried setting control edges on the x's to force them to 1342 // not combine, but the matching gets worried when it tries 1343 // to fold a StoreP and an AddP together (as part of an 1344 // address expression) and the AddP and StoreP have 1345 // different controls. 1346 if (!x->is_Load() && !x->is_DecodeNarrowPtr()) _igvn._worklist.yank(x); 1347 } 1348 _igvn.remove_dead_node(n); 1349 } 1350 } 1351 } 1352 } 1353 1354 try_move_store_after_loop(n); 1355 1356 // Check for Opaque2's who's loop has disappeared - who's input is in the 1357 // same loop nest as their output. Remove 'em, they are no longer useful. 1358 if( n_op == Op_Opaque2 && 1359 n->in(1) != NULL && 1360 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) { 1361 _igvn.replace_node( n, n->in(1) ); 1362 } 1363 } 1364 1365 //------------------------------split_if_with_blocks--------------------------- 1366 // Check for aggressive application of 'split-if' optimization, 1367 // using basic block level info. 1368 void PhaseIdealLoop::split_if_with_blocks( VectorSet &visited, Node_Stack &nstack ) { 1369 Node *n = C->root(); 1370 visited.set(n->_idx); // first, mark node as visited 1371 // Do pre-visit work for root 1372 n = split_if_with_blocks_pre( n ); 1373 uint cnt = n->outcnt(); 1374 uint i = 0; 1375 while (true) { 1376 // Visit all children 1377 if (i < cnt) { 1378 Node* use = n->raw_out(i); 1379 ++i; 1380 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) { 1381 // Now do pre-visit work for this use 1382 use = split_if_with_blocks_pre( use ); 1383 nstack.push(n, i); // Save parent and next use's index. 1384 n = use; // Process all children of current use. 1385 cnt = use->outcnt(); 1386 i = 0; 1387 } 1388 } 1389 else { 1390 // All of n's children have been processed, complete post-processing. 1391 if (cnt != 0 && !n->is_Con()) { 1392 assert(has_node(n), "no dead nodes"); 1393 split_if_with_blocks_post( n ); 1394 } 1395 if (nstack.is_empty()) { 1396 // Finished all nodes on stack. 1397 break; 1398 } 1399 // Get saved parent node and next use's index. Visit the rest of uses. 1400 n = nstack.node(); 1401 cnt = n->outcnt(); 1402 i = nstack.index(); 1403 nstack.pop(); 1404 } 1405 } 1406 } 1407 1408 1409 //============================================================================= 1410 // 1411 // C L O N E A L O O P B O D Y 1412 // 1413 1414 //------------------------------clone_iff-------------------------------------- 1415 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1416 // "Nearly" because all Nodes have been cloned from the original in the loop, 1417 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1418 // through the Phi recursively, and return a Bool. 1419 BoolNode *PhaseIdealLoop::clone_iff( PhiNode *phi, IdealLoopTree *loop ) { 1420 1421 // Convert this Phi into a Phi merging Bools 1422 uint i; 1423 for( i = 1; i < phi->req(); i++ ) { 1424 Node *b = phi->in(i); 1425 if( b->is_Phi() ) { 1426 _igvn.replace_input_of(phi, i, clone_iff( b->as_Phi(), loop )); 1427 } else { 1428 assert( b->is_Bool(), "" ); 1429 } 1430 } 1431 1432 Node *sample_bool = phi->in(1); 1433 Node *sample_cmp = sample_bool->in(1); 1434 1435 // Make Phis to merge the Cmp's inputs. 1436 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP ); 1437 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP ); 1438 for( i = 1; i < phi->req(); i++ ) { 1439 Node *n1 = phi->in(i)->in(1)->in(1); 1440 Node *n2 = phi->in(i)->in(1)->in(2); 1441 phi1->set_req( i, n1 ); 1442 phi2->set_req( i, n2 ); 1443 phi1->set_type( phi1->type()->meet_speculative(n1->bottom_type())); 1444 phi2->set_type( phi2->type()->meet_speculative(n2->bottom_type())); 1445 } 1446 // See if these Phis have been made before. 1447 // Register with optimizer 1448 Node *hit1 = _igvn.hash_find_insert(phi1); 1449 if( hit1 ) { // Hit, toss just made Phi 1450 _igvn.remove_dead_node(phi1); // Remove new phi 1451 assert( hit1->is_Phi(), "" ); 1452 phi1 = (PhiNode*)hit1; // Use existing phi 1453 } else { // Miss 1454 _igvn.register_new_node_with_optimizer(phi1); 1455 } 1456 Node *hit2 = _igvn.hash_find_insert(phi2); 1457 if( hit2 ) { // Hit, toss just made Phi 1458 _igvn.remove_dead_node(phi2); // Remove new phi 1459 assert( hit2->is_Phi(), "" ); 1460 phi2 = (PhiNode*)hit2; // Use existing phi 1461 } else { // Miss 1462 _igvn.register_new_node_with_optimizer(phi2); 1463 } 1464 // Register Phis with loop/block info 1465 set_ctrl(phi1, phi->in(0)); 1466 set_ctrl(phi2, phi->in(0)); 1467 // Make a new Cmp 1468 Node *cmp = sample_cmp->clone(); 1469 cmp->set_req( 1, phi1 ); 1470 cmp->set_req( 2, phi2 ); 1471 _igvn.register_new_node_with_optimizer(cmp); 1472 set_ctrl(cmp, phi->in(0)); 1473 1474 // Make a new Bool 1475 Node *b = sample_bool->clone(); 1476 b->set_req(1,cmp); 1477 _igvn.register_new_node_with_optimizer(b); 1478 set_ctrl(b, phi->in(0)); 1479 1480 assert( b->is_Bool(), "" ); 1481 return (BoolNode*)b; 1482 } 1483 1484 //------------------------------clone_bool------------------------------------- 1485 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1486 // "Nearly" because all Nodes have been cloned from the original in the loop, 1487 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1488 // through the Phi recursively, and return a Bool. 1489 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) { 1490 uint i; 1491 // Convert this Phi into a Phi merging Bools 1492 for( i = 1; i < phi->req(); i++ ) { 1493 Node *b = phi->in(i); 1494 if( b->is_Phi() ) { 1495 _igvn.replace_input_of(phi, i, clone_bool( b->as_Phi(), loop )); 1496 } else { 1497 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" ); 1498 } 1499 } 1500 1501 Node *sample_cmp = phi->in(1); 1502 1503 // Make Phis to merge the Cmp's inputs. 1504 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP ); 1505 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP ); 1506 for( uint j = 1; j < phi->req(); j++ ) { 1507 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP 1508 Node *n1, *n2; 1509 if( cmp_top->is_Cmp() ) { 1510 n1 = cmp_top->in(1); 1511 n2 = cmp_top->in(2); 1512 } else { 1513 n1 = n2 = cmp_top; 1514 } 1515 phi1->set_req( j, n1 ); 1516 phi2->set_req( j, n2 ); 1517 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 1518 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 1519 } 1520 1521 // See if these Phis have been made before. 1522 // Register with optimizer 1523 Node *hit1 = _igvn.hash_find_insert(phi1); 1524 if( hit1 ) { // Hit, toss just made Phi 1525 _igvn.remove_dead_node(phi1); // Remove new phi 1526 assert( hit1->is_Phi(), "" ); 1527 phi1 = (PhiNode*)hit1; // Use existing phi 1528 } else { // Miss 1529 _igvn.register_new_node_with_optimizer(phi1); 1530 } 1531 Node *hit2 = _igvn.hash_find_insert(phi2); 1532 if( hit2 ) { // Hit, toss just made Phi 1533 _igvn.remove_dead_node(phi2); // Remove new phi 1534 assert( hit2->is_Phi(), "" ); 1535 phi2 = (PhiNode*)hit2; // Use existing phi 1536 } else { // Miss 1537 _igvn.register_new_node_with_optimizer(phi2); 1538 } 1539 // Register Phis with loop/block info 1540 set_ctrl(phi1, phi->in(0)); 1541 set_ctrl(phi2, phi->in(0)); 1542 // Make a new Cmp 1543 Node *cmp = sample_cmp->clone(); 1544 cmp->set_req( 1, phi1 ); 1545 cmp->set_req( 2, phi2 ); 1546 _igvn.register_new_node_with_optimizer(cmp); 1547 set_ctrl(cmp, phi->in(0)); 1548 1549 assert( cmp->is_Cmp(), "" ); 1550 return (CmpNode*)cmp; 1551 } 1552 1553 //------------------------------sink_use--------------------------------------- 1554 // If 'use' was in the loop-exit block, it now needs to be sunk 1555 // below the post-loop merge point. 1556 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) { 1557 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) { 1558 set_ctrl(use, post_loop); 1559 for (DUIterator j = use->outs(); use->has_out(j); j++) 1560 sink_use(use->out(j), post_loop); 1561 } 1562 } 1563 1564 //------------------------------clone_loop------------------------------------- 1565 // 1566 // C L O N E A L O O P B O D Y 1567 // 1568 // This is the basic building block of the loop optimizations. It clones an 1569 // entire loop body. It makes an old_new loop body mapping; with this mapping 1570 // you can find the new-loop equivalent to an old-loop node. All new-loop 1571 // nodes are exactly equal to their old-loop counterparts, all edges are the 1572 // same. All exits from the old-loop now have a RegionNode that merges the 1573 // equivalent new-loop path. This is true even for the normal "loop-exit" 1574 // condition. All uses of loop-invariant old-loop values now come from (one 1575 // or more) Phis that merge their new-loop equivalents. 1576 // 1577 // This operation leaves the graph in an illegal state: there are two valid 1578 // control edges coming from the loop pre-header to both loop bodies. I'll 1579 // definitely have to hack the graph after running this transform. 1580 // 1581 // From this building block I will further edit edges to perform loop peeling 1582 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc. 1583 // 1584 // Parameter side_by_size_idom: 1585 // When side_by_size_idom is NULL, the dominator tree is constructed for 1586 // the clone loop to dominate the original. Used in construction of 1587 // pre-main-post loop sequence. 1588 // When nonnull, the clone and original are side-by-side, both are 1589 // dominated by the side_by_side_idom node. Used in construction of 1590 // unswitched loops. 1591 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd, 1592 Node* side_by_side_idom) { 1593 1594 if (C->do_vector_loop() && PrintOpto) { 1595 const char* mname = C->method()->name()->as_quoted_ascii(); 1596 if (mname != NULL) { 1597 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname); 1598 } 1599 } 1600 1601 CloneMap& cm = C->clone_map(); 1602 Dict* dict = cm.dict(); 1603 if (C->do_vector_loop()) { 1604 cm.set_clone_idx(cm.max_gen()+1); 1605 #ifndef PRODUCT 1606 if (PrintOpto) { 1607 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx()); 1608 loop->dump_head(); 1609 } 1610 #endif 1611 } 1612 1613 // Step 1: Clone the loop body. Make the old->new mapping. 1614 uint i; 1615 for( i = 0; i < loop->_body.size(); i++ ) { 1616 Node *old = loop->_body.at(i); 1617 Node *nnn = old->clone(); 1618 old_new.map( old->_idx, nnn ); 1619 if (C->do_vector_loop()) { 1620 cm.verify_insert_and_clone(old, nnn, cm.clone_idx()); 1621 } 1622 _igvn.register_new_node_with_optimizer(nnn); 1623 } 1624 1625 1626 // Step 2: Fix the edges in the new body. If the old input is outside the 1627 // loop use it. If the old input is INside the loop, use the corresponding 1628 // new node instead. 1629 for( i = 0; i < loop->_body.size(); i++ ) { 1630 Node *old = loop->_body.at(i); 1631 Node *nnn = old_new[old->_idx]; 1632 // Fix CFG/Loop controlling the new node 1633 if (has_ctrl(old)) { 1634 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]); 1635 } else { 1636 set_loop(nnn, loop->_parent); 1637 if (old->outcnt() > 0) { 1638 set_idom( nnn, old_new[idom(old)->_idx], dd ); 1639 } 1640 } 1641 // Correct edges to the new node 1642 for( uint j = 0; j < nnn->req(); j++ ) { 1643 Node *n = nnn->in(j); 1644 if( n ) { 1645 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n ); 1646 if( loop->is_member( old_in_loop ) ) 1647 nnn->set_req(j, old_new[n->_idx]); 1648 } 1649 } 1650 _igvn.hash_find_insert(nnn); 1651 } 1652 Node *newhead = old_new[loop->_head->_idx]; 1653 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 1654 1655 1656 // Step 3: Now fix control uses. Loop varying control uses have already 1657 // been fixed up (as part of all input edges in Step 2). Loop invariant 1658 // control uses must be either an IfFalse or an IfTrue. Make a merge 1659 // point to merge the old and new IfFalse/IfTrue nodes; make the use 1660 // refer to this. 1661 ResourceArea *area = Thread::current()->resource_area(); 1662 Node_List worklist(area); 1663 uint new_counter = C->unique(); 1664 for( i = 0; i < loop->_body.size(); i++ ) { 1665 Node* old = loop->_body.at(i); 1666 if( !old->is_CFG() ) continue; 1667 Node* nnn = old_new[old->_idx]; 1668 1669 // Copy uses to a worklist, so I can munge the def-use info 1670 // with impunity. 1671 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1672 worklist.push(old->fast_out(j)); 1673 1674 while( worklist.size() ) { // Visit all uses 1675 Node *use = worklist.pop(); 1676 if (!has_node(use)) continue; // Ignore dead nodes 1677 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1678 if( !loop->is_member( use_loop ) && use->is_CFG() ) { 1679 // Both OLD and USE are CFG nodes here. 1680 assert( use->is_Proj(), "" ); 1681 1682 // Clone the loop exit control projection 1683 Node *newuse = use->clone(); 1684 if (C->do_vector_loop()) { 1685 cm.verify_insert_and_clone(use, newuse, cm.clone_idx()); 1686 } 1687 newuse->set_req(0,nnn); 1688 _igvn.register_new_node_with_optimizer(newuse); 1689 set_loop(newuse, use_loop); 1690 set_idom(newuse, nnn, dom_depth(nnn) + 1 ); 1691 1692 // We need a Region to merge the exit from the peeled body and the 1693 // exit from the old loop body. 1694 RegionNode *r = new RegionNode(3); 1695 // Map the old use to the new merge point 1696 old_new.map( use->_idx, r ); 1697 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use)); 1698 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" ); 1699 1700 // The original user of 'use' uses 'r' instead. 1701 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) { 1702 Node* useuse = use->last_out(l); 1703 _igvn.rehash_node_delayed(useuse); 1704 uint uses_found = 0; 1705 if( useuse->in(0) == use ) { 1706 useuse->set_req(0, r); 1707 uses_found++; 1708 if( useuse->is_CFG() ) { 1709 assert( dom_depth(useuse) > dd_r, "" ); 1710 set_idom(useuse, r, dom_depth(useuse)); 1711 } 1712 } 1713 for( uint k = 1; k < useuse->req(); k++ ) { 1714 if( useuse->in(k) == use ) { 1715 useuse->set_req(k, r); 1716 uses_found++; 1717 } 1718 } 1719 l -= uses_found; // we deleted 1 or more copies of this edge 1720 } 1721 1722 // Now finish up 'r' 1723 r->set_req( 1, newuse ); 1724 r->set_req( 2, use ); 1725 _igvn.register_new_node_with_optimizer(r); 1726 set_loop(r, use_loop); 1727 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r); 1728 } // End of if a loop-exit test 1729 } 1730 } 1731 1732 // Step 4: If loop-invariant use is not control, it must be dominated by a 1733 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region 1734 // there if needed. Make a Phi there merging old and new used values. 1735 Node_List *split_if_set = NULL; 1736 Node_List *split_bool_set = NULL; 1737 Node_List *split_cex_set = NULL; 1738 for( i = 0; i < loop->_body.size(); i++ ) { 1739 Node* old = loop->_body.at(i); 1740 Node* nnn = old_new[old->_idx]; 1741 // Copy uses to a worklist, so I can munge the def-use info 1742 // with impunity. 1743 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1744 worklist.push(old->fast_out(j)); 1745 1746 while( worklist.size() ) { 1747 Node *use = worklist.pop(); 1748 if (!has_node(use)) continue; // Ignore dead nodes 1749 if (use->in(0) == C->top()) continue; 1750 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1751 // Check for data-use outside of loop - at least one of OLD or USE 1752 // must not be a CFG node. 1753 if( !loop->is_member( use_loop ) && (!old->is_CFG() || !use->is_CFG())) { 1754 1755 // If the Data use is an IF, that means we have an IF outside of the 1756 // loop that is switching on a condition that is set inside of the 1757 // loop. Happens if people set a loop-exit flag; then test the flag 1758 // in the loop to break the loop, then test is again outside of the 1759 // loop to determine which way the loop exited. 1760 // Loop predicate If node connects to Bool node through Opaque1 node. 1761 if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use)) { 1762 // Since this code is highly unlikely, we lazily build the worklist 1763 // of such Nodes to go split. 1764 if( !split_if_set ) 1765 split_if_set = new Node_List(area); 1766 split_if_set->push(use); 1767 } 1768 if( use->is_Bool() ) { 1769 if( !split_bool_set ) 1770 split_bool_set = new Node_List(area); 1771 split_bool_set->push(use); 1772 } 1773 if( use->Opcode() == Op_CreateEx ) { 1774 if( !split_cex_set ) 1775 split_cex_set = new Node_List(area); 1776 split_cex_set->push(use); 1777 } 1778 1779 1780 // Get "block" use is in 1781 uint idx = 0; 1782 while( use->in(idx) != old ) idx++; 1783 Node *prev = use->is_CFG() ? use : get_ctrl(use); 1784 assert( !loop->is_member( get_loop( prev ) ), "" ); 1785 Node *cfg = prev->_idx >= new_counter 1786 ? prev->in(2) 1787 : idom(prev); 1788 if( use->is_Phi() ) // Phi use is in prior block 1789 cfg = prev->in(idx); // NOT in block of Phi itself 1790 if (cfg->is_top()) { // Use is dead? 1791 _igvn.replace_input_of(use, idx, C->top()); 1792 continue; 1793 } 1794 1795 while( !loop->is_member( get_loop( cfg ) ) ) { 1796 prev = cfg; 1797 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg); 1798 } 1799 // If the use occurs after merging several exits from the loop, then 1800 // old value must have dominated all those exits. Since the same old 1801 // value was used on all those exits we did not need a Phi at this 1802 // merge point. NOW we do need a Phi here. Each loop exit value 1803 // is now merged with the peeled body exit; each exit gets its own 1804 // private Phi and those Phis need to be merged here. 1805 Node *phi; 1806 if( prev->is_Region() ) { 1807 if( idx == 0 ) { // Updating control edge? 1808 phi = prev; // Just use existing control 1809 } else { // Else need a new Phi 1810 phi = PhiNode::make( prev, old ); 1811 // Now recursively fix up the new uses of old! 1812 for( uint i = 1; i < prev->req(); i++ ) { 1813 worklist.push(phi); // Onto worklist once for each 'old' input 1814 } 1815 } 1816 } else { 1817 // Get new RegionNode merging old and new loop exits 1818 prev = old_new[prev->_idx]; 1819 assert( prev, "just made this in step 7" ); 1820 if( idx == 0 ) { // Updating control edge? 1821 phi = prev; // Just use existing control 1822 } else { // Else need a new Phi 1823 // Make a new Phi merging data values properly 1824 phi = PhiNode::make( prev, old ); 1825 phi->set_req( 1, nnn ); 1826 } 1827 } 1828 // If inserting a new Phi, check for prior hits 1829 if( idx != 0 ) { 1830 Node *hit = _igvn.hash_find_insert(phi); 1831 if( hit == NULL ) { 1832 _igvn.register_new_node_with_optimizer(phi); // Register new phi 1833 } else { // or 1834 // Remove the new phi from the graph and use the hit 1835 _igvn.remove_dead_node(phi); 1836 phi = hit; // Use existing phi 1837 } 1838 set_ctrl(phi, prev); 1839 } 1840 // Make 'use' use the Phi instead of the old loop body exit value 1841 _igvn.replace_input_of(use, idx, phi); 1842 if( use->_idx >= new_counter ) { // If updating new phis 1843 // Not needed for correctness, but prevents a weak assert 1844 // in AddPNode from tripping (when we end up with different 1845 // base & derived Phis that will become the same after 1846 // IGVN does CSE). 1847 Node *hit = _igvn.hash_find_insert(use); 1848 if( hit ) // Go ahead and re-hash for hits. 1849 _igvn.replace_node( use, hit ); 1850 } 1851 1852 // If 'use' was in the loop-exit block, it now needs to be sunk 1853 // below the post-loop merge point. 1854 sink_use( use, prev ); 1855 } 1856 } 1857 } 1858 1859 // Check for IFs that need splitting/cloning. Happens if an IF outside of 1860 // the loop uses a condition set in the loop. The original IF probably 1861 // takes control from one or more OLD Regions (which in turn get from NEW 1862 // Regions). In any case, there will be a set of Phis for each merge point 1863 // from the IF up to where the original BOOL def exists the loop. 1864 if( split_if_set ) { 1865 while( split_if_set->size() ) { 1866 Node *iff = split_if_set->pop(); 1867 if( iff->in(1)->is_Phi() ) { 1868 BoolNode *b = clone_iff( iff->in(1)->as_Phi(), loop ); 1869 _igvn.replace_input_of(iff, 1, b); 1870 } 1871 } 1872 } 1873 if( split_bool_set ) { 1874 while( split_bool_set->size() ) { 1875 Node *b = split_bool_set->pop(); 1876 Node *phi = b->in(1); 1877 assert( phi->is_Phi(), "" ); 1878 CmpNode *cmp = clone_bool( (PhiNode*)phi, loop ); 1879 _igvn.replace_input_of(b, 1, cmp); 1880 } 1881 } 1882 if( split_cex_set ) { 1883 while( split_cex_set->size() ) { 1884 Node *b = split_cex_set->pop(); 1885 assert( b->in(0)->is_Region(), "" ); 1886 assert( b->in(1)->is_Phi(), "" ); 1887 assert( b->in(0)->in(0) == b->in(1)->in(0), "" ); 1888 split_up( b, b->in(0), NULL ); 1889 } 1890 } 1891 1892 } 1893 1894 1895 //---------------------- stride_of_possible_iv ------------------------------------- 1896 // Looks for an iff/bool/comp with one operand of the compare 1897 // being a cycle involving an add and a phi, 1898 // with an optional truncation (left-shift followed by a right-shift) 1899 // of the add. Returns zero if not an iv. 1900 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) { 1901 Node* trunc1 = NULL; 1902 Node* trunc2 = NULL; 1903 const TypeInt* ttype = NULL; 1904 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) { 1905 return 0; 1906 } 1907 BoolNode* bl = iff->in(1)->as_Bool(); 1908 Node* cmp = bl->in(1); 1909 if (!cmp || cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU) { 1910 return 0; 1911 } 1912 // Must have an invariant operand 1913 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) { 1914 return 0; 1915 } 1916 Node* add2 = NULL; 1917 Node* cmp1 = cmp->in(1); 1918 if (cmp1->is_Phi()) { 1919 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) ))) 1920 Node* phi = cmp1; 1921 for (uint i = 1; i < phi->req(); i++) { 1922 Node* in = phi->in(i); 1923 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in, 1924 &trunc1, &trunc2, &ttype); 1925 if (add && add->in(1) == phi) { 1926 add2 = add->in(2); 1927 break; 1928 } 1929 } 1930 } else { 1931 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) ))) 1932 Node* addtrunc = cmp1; 1933 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc, 1934 &trunc1, &trunc2, &ttype); 1935 if (add && add->in(1)->is_Phi()) { 1936 Node* phi = add->in(1); 1937 for (uint i = 1; i < phi->req(); i++) { 1938 if (phi->in(i) == addtrunc) { 1939 add2 = add->in(2); 1940 break; 1941 } 1942 } 1943 } 1944 } 1945 if (add2 != NULL) { 1946 const TypeInt* add2t = _igvn.type(add2)->is_int(); 1947 if (add2t->is_con()) { 1948 return add2t->get_con(); 1949 } 1950 } 1951 return 0; 1952 } 1953 1954 1955 //---------------------- stay_in_loop ------------------------------------- 1956 // Return the (unique) control output node that's in the loop (if it exists.) 1957 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) { 1958 Node* unique = NULL; 1959 if (!n) return NULL; 1960 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1961 Node* use = n->fast_out(i); 1962 if (!has_ctrl(use) && loop->is_member(get_loop(use))) { 1963 if (unique != NULL) { 1964 return NULL; 1965 } 1966 unique = use; 1967 } 1968 } 1969 return unique; 1970 } 1971 1972 //------------------------------ register_node ------------------------------------- 1973 // Utility to register node "n" with PhaseIdealLoop 1974 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) { 1975 _igvn.register_new_node_with_optimizer(n); 1976 loop->_body.push(n); 1977 if (n->is_CFG()) { 1978 set_loop(n, loop); 1979 set_idom(n, pred, ddepth); 1980 } else { 1981 set_ctrl(n, pred); 1982 } 1983 } 1984 1985 //------------------------------ proj_clone ------------------------------------- 1986 // Utility to create an if-projection 1987 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) { 1988 ProjNode* c = p->clone()->as_Proj(); 1989 c->set_req(0, iff); 1990 return c; 1991 } 1992 1993 //------------------------------ short_circuit_if ------------------------------------- 1994 // Force the iff control output to be the live_proj 1995 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) { 1996 guarantee(live_proj != NULL, "null projection"); 1997 int proj_con = live_proj->_con; 1998 assert(proj_con == 0 || proj_con == 1, "false or true projection"); 1999 Node *con = _igvn.intcon(proj_con); 2000 set_ctrl(con, C->root()); 2001 if (iff) { 2002 iff->set_req(1, con); 2003 } 2004 return con; 2005 } 2006 2007 //------------------------------ insert_if_before_proj ------------------------------------- 2008 // Insert a new if before an if projection (* - new node) 2009 // 2010 // before 2011 // if(test) 2012 // / \ 2013 // v v 2014 // other-proj proj (arg) 2015 // 2016 // after 2017 // if(test) 2018 // / \ 2019 // / v 2020 // | * proj-clone 2021 // v | 2022 // other-proj v 2023 // * new_if(relop(cmp[IU](left,right))) 2024 // / \ 2025 // v v 2026 // * new-proj proj 2027 // (returned) 2028 // 2029 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) { 2030 IfNode* iff = proj->in(0)->as_If(); 2031 IdealLoopTree *loop = get_loop(proj); 2032 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2033 int ddepth = dom_depth(proj); 2034 2035 _igvn.rehash_node_delayed(iff); 2036 _igvn.rehash_node_delayed(proj); 2037 2038 proj->set_req(0, NULL); // temporary disconnect 2039 ProjNode* proj2 = proj_clone(proj, iff); 2040 register_node(proj2, loop, iff, ddepth); 2041 2042 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right); 2043 register_node(cmp, loop, proj2, ddepth); 2044 2045 BoolNode* bol = new BoolNode(cmp, relop); 2046 register_node(bol, loop, proj2, ddepth); 2047 2048 int opcode = iff->Opcode(); 2049 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode"); 2050 IfNode* new_if = (opcode == Op_If) ? new IfNode(proj2, bol, iff->_prob, iff->_fcnt): 2051 new RangeCheckNode(proj2, bol, iff->_prob, iff->_fcnt); 2052 register_node(new_if, loop, proj2, ddepth); 2053 2054 proj->set_req(0, new_if); // reattach 2055 set_idom(proj, new_if, ddepth); 2056 2057 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj(); 2058 guarantee(new_exit != NULL, "null exit node"); 2059 register_node(new_exit, get_loop(other_proj), new_if, ddepth); 2060 2061 return new_exit; 2062 } 2063 2064 //------------------------------ insert_region_before_proj ------------------------------------- 2065 // Insert a region before an if projection (* - new node) 2066 // 2067 // before 2068 // if(test) 2069 // / | 2070 // v | 2071 // proj v 2072 // other-proj 2073 // 2074 // after 2075 // if(test) 2076 // / | 2077 // v | 2078 // * proj-clone v 2079 // | other-proj 2080 // v 2081 // * new-region 2082 // | 2083 // v 2084 // * dum_if 2085 // / \ 2086 // v \ 2087 // * dum-proj v 2088 // proj 2089 // 2090 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) { 2091 IfNode* iff = proj->in(0)->as_If(); 2092 IdealLoopTree *loop = get_loop(proj); 2093 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2094 int ddepth = dom_depth(proj); 2095 2096 _igvn.rehash_node_delayed(iff); 2097 _igvn.rehash_node_delayed(proj); 2098 2099 proj->set_req(0, NULL); // temporary disconnect 2100 ProjNode* proj2 = proj_clone(proj, iff); 2101 register_node(proj2, loop, iff, ddepth); 2102 2103 RegionNode* reg = new RegionNode(2); 2104 reg->set_req(1, proj2); 2105 register_node(reg, loop, iff, ddepth); 2106 2107 IfNode* dum_if = new IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt); 2108 register_node(dum_if, loop, reg, ddepth); 2109 2110 proj->set_req(0, dum_if); // reattach 2111 set_idom(proj, dum_if, ddepth); 2112 2113 ProjNode* dum_proj = proj_clone(other_proj, dum_if); 2114 register_node(dum_proj, loop, dum_if, ddepth); 2115 2116 return reg; 2117 } 2118 2119 //------------------------------ insert_cmpi_loop_exit ------------------------------------- 2120 // Clone a signed compare loop exit from an unsigned compare and 2121 // insert it before the unsigned cmp on the stay-in-loop path. 2122 // All new nodes inserted in the dominator tree between the original 2123 // if and it's projections. The original if test is replaced with 2124 // a constant to force the stay-in-loop path. 2125 // 2126 // This is done to make sure that the original if and it's projections 2127 // still dominate the same set of control nodes, that the ctrl() relation 2128 // from data nodes to them is preserved, and that their loop nesting is 2129 // preserved. 2130 // 2131 // before 2132 // if(i <u limit) unsigned compare loop exit 2133 // / | 2134 // v v 2135 // exit-proj stay-in-loop-proj 2136 // 2137 // after 2138 // if(stay-in-loop-const) original if 2139 // / | 2140 // / v 2141 // / if(i < limit) new signed test 2142 // / / | 2143 // / / v 2144 // / / if(i <u limit) new cloned unsigned test 2145 // / / / | 2146 // v v v | 2147 // region | 2148 // | | 2149 // dum-if | 2150 // / | | 2151 // ether | | 2152 // v v 2153 // exit-proj stay-in-loop-proj 2154 // 2155 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) { 2156 const bool Signed = true; 2157 const bool Unsigned = false; 2158 2159 BoolNode* bol = if_cmpu->in(1)->as_Bool(); 2160 if (bol->_test._test != BoolTest::lt) return NULL; 2161 CmpNode* cmpu = bol->in(1)->as_Cmp(); 2162 if (cmpu->Opcode() != Op_CmpU) return NULL; 2163 int stride = stride_of_possible_iv(if_cmpu); 2164 if (stride == 0) return NULL; 2165 2166 Node* lp_proj = stay_in_loop(if_cmpu, loop); 2167 guarantee(lp_proj != NULL, "null loop node"); 2168 2169 ProjNode* lp_continue = lp_proj->as_Proj(); 2170 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj(); 2171 2172 Node* limit = NULL; 2173 if (stride > 0) { 2174 limit = cmpu->in(2); 2175 } else { 2176 limit = _igvn.makecon(TypeInt::ZERO); 2177 set_ctrl(limit, C->root()); 2178 } 2179 // Create a new region on the exit path 2180 RegionNode* reg = insert_region_before_proj(lp_exit); 2181 guarantee(reg != NULL, "null region node"); 2182 2183 // Clone the if-cmpu-true-false using a signed compare 2184 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge; 2185 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue); 2186 reg->add_req(cmpi_exit); 2187 2188 // Clone the if-cmpu-true-false 2189 BoolTest::mask rel_u = bol->_test._test; 2190 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue); 2191 reg->add_req(cmpu_exit); 2192 2193 // Force original if to stay in loop. 2194 short_circuit_if(if_cmpu, lp_continue); 2195 2196 return cmpi_exit->in(0)->as_If(); 2197 } 2198 2199 //------------------------------ remove_cmpi_loop_exit ------------------------------------- 2200 // Remove a previously inserted signed compare loop exit. 2201 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) { 2202 Node* lp_proj = stay_in_loop(if_cmp, loop); 2203 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI && 2204 stay_in_loop(lp_proj, loop)->is_If() && 2205 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu"); 2206 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO); 2207 set_ctrl(con, C->root()); 2208 if_cmp->set_req(1, con); 2209 } 2210 2211 //------------------------------ scheduled_nodelist ------------------------------------- 2212 // Create a post order schedule of nodes that are in the 2213 // "member" set. The list is returned in "sched". 2214 // The first node in "sched" is the loop head, followed by 2215 // nodes which have no inputs in the "member" set, and then 2216 // followed by the nodes that have an immediate input dependence 2217 // on a node in "sched". 2218 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) { 2219 2220 assert(member.test(loop->_head->_idx), "loop head must be in member set"); 2221 Arena *a = Thread::current()->resource_area(); 2222 VectorSet visited(a); 2223 Node_Stack nstack(a, loop->_body.size()); 2224 2225 Node* n = loop->_head; // top of stack is cached in "n" 2226 uint idx = 0; 2227 visited.set(n->_idx); 2228 2229 // Initially push all with no inputs from within member set 2230 for(uint i = 0; i < loop->_body.size(); i++ ) { 2231 Node *elt = loop->_body.at(i); 2232 if (member.test(elt->_idx)) { 2233 bool found = false; 2234 for (uint j = 0; j < elt->req(); j++) { 2235 Node* def = elt->in(j); 2236 if (def && member.test(def->_idx) && def != elt) { 2237 found = true; 2238 break; 2239 } 2240 } 2241 if (!found && elt != loop->_head) { 2242 nstack.push(n, idx); 2243 n = elt; 2244 assert(!visited.test(n->_idx), "not seen yet"); 2245 visited.set(n->_idx); 2246 } 2247 } 2248 } 2249 2250 // traverse out's that are in the member set 2251 while (true) { 2252 if (idx < n->outcnt()) { 2253 Node* use = n->raw_out(idx); 2254 idx++; 2255 if (!visited.test_set(use->_idx)) { 2256 if (member.test(use->_idx)) { 2257 nstack.push(n, idx); 2258 n = use; 2259 idx = 0; 2260 } 2261 } 2262 } else { 2263 // All outputs processed 2264 sched.push(n); 2265 if (nstack.is_empty()) break; 2266 n = nstack.node(); 2267 idx = nstack.index(); 2268 nstack.pop(); 2269 } 2270 } 2271 } 2272 2273 2274 //------------------------------ has_use_in_set ------------------------------------- 2275 // Has a use in the vector set 2276 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) { 2277 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2278 Node* use = n->fast_out(j); 2279 if (vset.test(use->_idx)) { 2280 return true; 2281 } 2282 } 2283 return false; 2284 } 2285 2286 2287 //------------------------------ has_use_internal_to_set ------------------------------------- 2288 // Has use internal to the vector set (ie. not in a phi at the loop head) 2289 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) { 2290 Node* head = loop->_head; 2291 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2292 Node* use = n->fast_out(j); 2293 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) { 2294 return true; 2295 } 2296 } 2297 return false; 2298 } 2299 2300 2301 //------------------------------ clone_for_use_outside_loop ------------------------------------- 2302 // clone "n" for uses that are outside of loop 2303 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) { 2304 int cloned = 0; 2305 assert(worklist.size() == 0, "should be empty"); 2306 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2307 Node* use = n->fast_out(j); 2308 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) { 2309 worklist.push(use); 2310 } 2311 } 2312 while( worklist.size() ) { 2313 Node *use = worklist.pop(); 2314 if (!has_node(use) || use->in(0) == C->top()) continue; 2315 uint j; 2316 for (j = 0; j < use->req(); j++) { 2317 if (use->in(j) == n) break; 2318 } 2319 assert(j < use->req(), "must be there"); 2320 2321 // clone "n" and insert it between the inputs of "n" and the use outside the loop 2322 Node* n_clone = n->clone(); 2323 _igvn.replace_input_of(use, j, n_clone); 2324 cloned++; 2325 Node* use_c; 2326 if (!use->is_Phi()) { 2327 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0); 2328 } else { 2329 // Use in a phi is considered a use in the associated predecessor block 2330 use_c = use->in(0)->in(j); 2331 } 2332 set_ctrl(n_clone, use_c); 2333 assert(!loop->is_member(get_loop(use_c)), "should be outside loop"); 2334 get_loop(use_c)->_body.push(n_clone); 2335 _igvn.register_new_node_with_optimizer(n_clone); 2336 #if !defined(PRODUCT) 2337 if (TracePartialPeeling) { 2338 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx); 2339 } 2340 #endif 2341 } 2342 return cloned; 2343 } 2344 2345 2346 //------------------------------ clone_for_special_use_inside_loop ------------------------------------- 2347 // clone "n" for special uses that are in the not_peeled region. 2348 // If these def-uses occur in separate blocks, the code generator 2349 // marks the method as not compilable. For example, if a "BoolNode" 2350 // is in a different basic block than the "IfNode" that uses it, then 2351 // the compilation is aborted in the code generator. 2352 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 2353 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) { 2354 if (n->is_Phi() || n->is_Load()) { 2355 return; 2356 } 2357 assert(worklist.size() == 0, "should be empty"); 2358 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2359 Node* use = n->fast_out(j); 2360 if ( not_peel.test(use->_idx) && 2361 (use->is_If() || use->is_CMove() || use->is_Bool()) && 2362 use->in(1) == n) { 2363 worklist.push(use); 2364 } 2365 } 2366 if (worklist.size() > 0) { 2367 // clone "n" and insert it between inputs of "n" and the use 2368 Node* n_clone = n->clone(); 2369 loop->_body.push(n_clone); 2370 _igvn.register_new_node_with_optimizer(n_clone); 2371 set_ctrl(n_clone, get_ctrl(n)); 2372 sink_list.push(n_clone); 2373 not_peel <<= n_clone->_idx; // add n_clone to not_peel set. 2374 #if !defined(PRODUCT) 2375 if (TracePartialPeeling) { 2376 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx); 2377 } 2378 #endif 2379 while( worklist.size() ) { 2380 Node *use = worklist.pop(); 2381 _igvn.rehash_node_delayed(use); 2382 for (uint j = 1; j < use->req(); j++) { 2383 if (use->in(j) == n) { 2384 use->set_req(j, n_clone); 2385 } 2386 } 2387 } 2388 } 2389 } 2390 2391 2392 //------------------------------ insert_phi_for_loop ------------------------------------- 2393 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 2394 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) { 2395 Node *phi = PhiNode::make(lp, back_edge_val); 2396 phi->set_req(LoopNode::EntryControl, lp_entry_val); 2397 // Use existing phi if it already exists 2398 Node *hit = _igvn.hash_find_insert(phi); 2399 if( hit == NULL ) { 2400 _igvn.register_new_node_with_optimizer(phi); 2401 set_ctrl(phi, lp); 2402 } else { 2403 // Remove the new phi from the graph and use the hit 2404 _igvn.remove_dead_node(phi); 2405 phi = hit; 2406 } 2407 _igvn.replace_input_of(use, idx, phi); 2408 } 2409 2410 #ifdef ASSERT 2411 //------------------------------ is_valid_loop_partition ------------------------------------- 2412 // Validate the loop partition sets: peel and not_peel 2413 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, 2414 VectorSet& not_peel ) { 2415 uint i; 2416 // Check that peel_list entries are in the peel set 2417 for (i = 0; i < peel_list.size(); i++) { 2418 if (!peel.test(peel_list.at(i)->_idx)) { 2419 return false; 2420 } 2421 } 2422 // Check at loop members are in one of peel set or not_peel set 2423 for (i = 0; i < loop->_body.size(); i++ ) { 2424 Node *def = loop->_body.at(i); 2425 uint di = def->_idx; 2426 // Check that peel set elements are in peel_list 2427 if (peel.test(di)) { 2428 if (not_peel.test(di)) { 2429 return false; 2430 } 2431 // Must be in peel_list also 2432 bool found = false; 2433 for (uint j = 0; j < peel_list.size(); j++) { 2434 if (peel_list.at(j)->_idx == di) { 2435 found = true; 2436 break; 2437 } 2438 } 2439 if (!found) { 2440 return false; 2441 } 2442 } else if (not_peel.test(di)) { 2443 if (peel.test(di)) { 2444 return false; 2445 } 2446 } else { 2447 return false; 2448 } 2449 } 2450 return true; 2451 } 2452 2453 //------------------------------ is_valid_clone_loop_exit_use ------------------------------------- 2454 // Ensure a use outside of loop is of the right form 2455 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) { 2456 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2457 return (use->is_Phi() && 2458 use_c->is_Region() && use_c->req() == 3 && 2459 (use_c->in(exit_idx)->Opcode() == Op_IfTrue || 2460 use_c->in(exit_idx)->Opcode() == Op_IfFalse || 2461 use_c->in(exit_idx)->Opcode() == Op_JumpProj) && 2462 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) ); 2463 } 2464 2465 //------------------------------ is_valid_clone_loop_form ------------------------------------- 2466 // Ensure that all uses outside of loop are of the right form 2467 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 2468 uint orig_exit_idx, uint clone_exit_idx) { 2469 uint len = peel_list.size(); 2470 for (uint i = 0; i < len; i++) { 2471 Node *def = peel_list.at(i); 2472 2473 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 2474 Node *use = def->fast_out(j); 2475 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2476 if (!loop->is_member(get_loop(use_c))) { 2477 // use is not in the loop, check for correct structure 2478 if (use->in(0) == def) { 2479 // Okay 2480 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) { 2481 return false; 2482 } 2483 } 2484 } 2485 } 2486 return true; 2487 } 2488 #endif 2489 2490 //------------------------------ partial_peel ------------------------------------- 2491 // Partially peel (aka loop rotation) the top portion of a loop (called 2492 // the peel section below) by cloning it and placing one copy just before 2493 // the new loop head and the other copy at the bottom of the new loop. 2494 // 2495 // before after where it came from 2496 // 2497 // stmt1 stmt1 2498 // loop: stmt2 clone 2499 // stmt2 if condA goto exitA clone 2500 // if condA goto exitA new_loop: new 2501 // stmt3 stmt3 clone 2502 // if !condB goto loop if condB goto exitB clone 2503 // exitB: stmt2 orig 2504 // stmt4 if !condA goto new_loop orig 2505 // exitA: goto exitA 2506 // exitB: 2507 // stmt4 2508 // exitA: 2509 // 2510 // Step 1: find the cut point: an exit test on probable 2511 // induction variable. 2512 // Step 2: schedule (with cloning) operations in the peel 2513 // section that can be executed after the cut into 2514 // the section that is not peeled. This may need 2515 // to clone operations into exit blocks. For 2516 // instance, a reference to A[i] in the not-peel 2517 // section and a reference to B[i] in an exit block 2518 // may cause a left-shift of i by 2 to be placed 2519 // in the peel block. This step will clone the left 2520 // shift into the exit block and sink the left shift 2521 // from the peel to the not-peel section. 2522 // Step 3: clone the loop, retarget the control, and insert 2523 // phis for values that are live across the new loop 2524 // head. This is very dependent on the graph structure 2525 // from clone_loop. It creates region nodes for 2526 // exit control and associated phi nodes for values 2527 // flow out of the loop through that exit. The region 2528 // node is dominated by the clone's control projection. 2529 // So the clone's peel section is placed before the 2530 // new loop head, and the clone's not-peel section is 2531 // forms the top part of the new loop. The original 2532 // peel section forms the tail of the new loop. 2533 // Step 4: update the dominator tree and recompute the 2534 // dominator depth. 2535 // 2536 // orig 2537 // 2538 // stmt1 2539 // | 2540 // v 2541 // loop predicate 2542 // | 2543 // v 2544 // loop<----+ 2545 // | | 2546 // stmt2 | 2547 // | | 2548 // v | 2549 // ifA | 2550 // / | | 2551 // v v | 2552 // false true ^ <-- last_peel 2553 // / | | 2554 // / ===|==cut | 2555 // / stmt3 | <-- first_not_peel 2556 // / | | 2557 // | v | 2558 // v ifB | 2559 // exitA: / \ | 2560 // / \ | 2561 // v v | 2562 // false true | 2563 // / \ | 2564 // / ----+ 2565 // | 2566 // v 2567 // exitB: 2568 // stmt4 2569 // 2570 // 2571 // after clone loop 2572 // 2573 // stmt1 2574 // | 2575 // v 2576 // loop predicate 2577 // / \ 2578 // clone / \ orig 2579 // / \ 2580 // / \ 2581 // v v 2582 // +---->loop loop<----+ 2583 // | | | | 2584 // | stmt2 stmt2 | 2585 // | | | | 2586 // | v v | 2587 // | ifA ifA | 2588 // | | \ / | | 2589 // | v v v v | 2590 // ^ true false false true ^ <-- last_peel 2591 // | | ^ \ / | | 2592 // | cut==|== \ \ / ===|==cut | 2593 // | stmt3 \ \ / stmt3 | <-- first_not_peel 2594 // | | dom | | | | 2595 // | v \ 1v v2 v | 2596 // | ifB regionA ifB | 2597 // | / \ | / \ | 2598 // | / \ v / \ | 2599 // | v v exitA: v v | 2600 // | true false false true | 2601 // | / ^ \ / \ | 2602 // +---- \ \ / ----+ 2603 // dom \ / 2604 // \ 1v v2 2605 // regionB 2606 // | 2607 // v 2608 // exitB: 2609 // stmt4 2610 // 2611 // 2612 // after partial peel 2613 // 2614 // stmt1 2615 // | 2616 // v 2617 // loop predicate 2618 // / 2619 // clone / orig 2620 // / TOP 2621 // / \ 2622 // v v 2623 // TOP->loop loop----+ 2624 // | | | 2625 // stmt2 stmt2 | 2626 // | | | 2627 // v v | 2628 // ifA ifA | 2629 // | \ / | | 2630 // v v v v | 2631 // true false false true | <-- last_peel 2632 // | ^ \ / +------|---+ 2633 // +->newloop \ \ / === ==cut | | 2634 // | stmt3 \ \ / TOP | | 2635 // | | dom | | stmt3 | | <-- first_not_peel 2636 // | v \ 1v v2 v | | 2637 // | ifB regionA ifB ^ v 2638 // | / \ | / \ | | 2639 // | / \ v / \ | | 2640 // | v v exitA: v v | | 2641 // | true false false true | | 2642 // | / ^ \ / \ | | 2643 // | | \ \ / v | | 2644 // | | dom \ / TOP | | 2645 // | | \ 1v v2 | | 2646 // ^ v regionB | | 2647 // | | | | | 2648 // | | v ^ v 2649 // | | exitB: | | 2650 // | | stmt4 | | 2651 // | +------------>-----------------+ | 2652 // | | 2653 // +-----------------<---------------------+ 2654 // 2655 // 2656 // final graph 2657 // 2658 // stmt1 2659 // | 2660 // v 2661 // loop predicate 2662 // | 2663 // v 2664 // stmt2 clone 2665 // | 2666 // v 2667 // ........> ifA clone 2668 // : / | 2669 // dom / | 2670 // : v v 2671 // : false true 2672 // : | | 2673 // : | v 2674 // : | newloop<-----+ 2675 // : | | | 2676 // : | stmt3 clone | 2677 // : | | | 2678 // : | v | 2679 // : | ifB | 2680 // : | / \ | 2681 // : | v v | 2682 // : | false true | 2683 // : | | | | 2684 // : | v stmt2 | 2685 // : | exitB: | | 2686 // : | stmt4 v | 2687 // : | ifA orig | 2688 // : | / \ | 2689 // : | / \ | 2690 // : | v v | 2691 // : | false true | 2692 // : | / \ | 2693 // : v v -----+ 2694 // RegionA 2695 // | 2696 // v 2697 // exitA 2698 // 2699 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) { 2700 2701 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only"); 2702 if (!loop->_head->is_Loop()) { 2703 return false; } 2704 2705 LoopNode *head = loop->_head->as_Loop(); 2706 2707 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) { 2708 return false; 2709 } 2710 2711 // Check for complex exit control 2712 for(uint ii = 0; ii < loop->_body.size(); ii++ ) { 2713 Node *n = loop->_body.at(ii); 2714 int opc = n->Opcode(); 2715 if (n->is_Call() || 2716 opc == Op_Catch || 2717 opc == Op_CatchProj || 2718 opc == Op_Jump || 2719 opc == Op_JumpProj) { 2720 #if !defined(PRODUCT) 2721 if (TracePartialPeeling) { 2722 tty->print_cr("\nExit control too complex: lp: %d", head->_idx); 2723 } 2724 #endif 2725 return false; 2726 } 2727 } 2728 2729 int dd = dom_depth(head); 2730 2731 // Step 1: find cut point 2732 2733 // Walk up dominators to loop head looking for first loop exit 2734 // which is executed on every path thru loop. 2735 IfNode *peel_if = NULL; 2736 IfNode *peel_if_cmpu = NULL; 2737 2738 Node *iff = loop->tail(); 2739 while( iff != head ) { 2740 if( iff->is_If() ) { 2741 Node *ctrl = get_ctrl(iff->in(1)); 2742 if (ctrl->is_top()) return false; // Dead test on live IF. 2743 // If loop-varying exit-test, check for induction variable 2744 if( loop->is_member(get_loop(ctrl)) && 2745 loop->is_loop_exit(iff) && 2746 is_possible_iv_test(iff)) { 2747 Node* cmp = iff->in(1)->in(1); 2748 if (cmp->Opcode() == Op_CmpI) { 2749 peel_if = iff->as_If(); 2750 } else { 2751 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU"); 2752 peel_if_cmpu = iff->as_If(); 2753 } 2754 } 2755 } 2756 iff = idom(iff); 2757 } 2758 // Prefer signed compare over unsigned compare. 2759 IfNode* new_peel_if = NULL; 2760 if (peel_if == NULL) { 2761 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) { 2762 return false; // No peel point found 2763 } 2764 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop); 2765 if (new_peel_if == NULL) { 2766 return false; // No peel point found 2767 } 2768 peel_if = new_peel_if; 2769 } 2770 Node* last_peel = stay_in_loop(peel_if, loop); 2771 Node* first_not_peeled = stay_in_loop(last_peel, loop); 2772 if (first_not_peeled == NULL || first_not_peeled == head) { 2773 return false; 2774 } 2775 2776 #if !defined(PRODUCT) 2777 if (TraceLoopOpts) { 2778 tty->print("PartialPeel "); 2779 loop->dump_head(); 2780 } 2781 2782 if (TracePartialPeeling) { 2783 tty->print_cr("before partial peel one iteration"); 2784 Node_List wl; 2785 Node* t = head->in(2); 2786 while (true) { 2787 wl.push(t); 2788 if (t == head) break; 2789 t = idom(t); 2790 } 2791 while (wl.size() > 0) { 2792 Node* tt = wl.pop(); 2793 tt->dump(); 2794 if (tt == last_peel) tty->print_cr("-- cut --"); 2795 } 2796 } 2797 #endif 2798 ResourceArea *area = Thread::current()->resource_area(); 2799 VectorSet peel(area); 2800 VectorSet not_peel(area); 2801 Node_List peel_list(area); 2802 Node_List worklist(area); 2803 Node_List sink_list(area); 2804 2805 // Set of cfg nodes to peel are those that are executable from 2806 // the head through last_peel. 2807 assert(worklist.size() == 0, "should be empty"); 2808 worklist.push(head); 2809 peel.set(head->_idx); 2810 while (worklist.size() > 0) { 2811 Node *n = worklist.pop(); 2812 if (n != last_peel) { 2813 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2814 Node* use = n->fast_out(j); 2815 if (use->is_CFG() && 2816 loop->is_member(get_loop(use)) && 2817 !peel.test_set(use->_idx)) { 2818 worklist.push(use); 2819 } 2820 } 2821 } 2822 } 2823 2824 // Set of non-cfg nodes to peel are those that are control 2825 // dependent on the cfg nodes. 2826 uint i; 2827 for(i = 0; i < loop->_body.size(); i++ ) { 2828 Node *n = loop->_body.at(i); 2829 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n; 2830 if (peel.test(n_c->_idx)) { 2831 peel.set(n->_idx); 2832 } else { 2833 not_peel.set(n->_idx); 2834 } 2835 } 2836 2837 // Step 2: move operations from the peeled section down into the 2838 // not-peeled section 2839 2840 // Get a post order schedule of nodes in the peel region 2841 // Result in right-most operand. 2842 scheduled_nodelist(loop, peel, peel_list ); 2843 2844 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 2845 2846 // For future check for too many new phis 2847 uint old_phi_cnt = 0; 2848 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 2849 Node* use = head->fast_out(j); 2850 if (use->is_Phi()) old_phi_cnt++; 2851 } 2852 2853 #if !defined(PRODUCT) 2854 if (TracePartialPeeling) { 2855 tty->print_cr("\npeeled list"); 2856 } 2857 #endif 2858 2859 // Evacuate nodes in peel region into the not_peeled region if possible 2860 uint new_phi_cnt = 0; 2861 uint cloned_for_outside_use = 0; 2862 for (i = 0; i < peel_list.size();) { 2863 Node* n = peel_list.at(i); 2864 #if !defined(PRODUCT) 2865 if (TracePartialPeeling) n->dump(); 2866 #endif 2867 bool incr = true; 2868 if ( !n->is_CFG() ) { 2869 2870 if ( has_use_in_set(n, not_peel) ) { 2871 2872 // If not used internal to the peeled region, 2873 // move "n" from peeled to not_peeled region. 2874 2875 if ( !has_use_internal_to_set(n, peel, loop) ) { 2876 2877 // if not pinned and not a load (which maybe anti-dependent on a store) 2878 // and not a CMove (Matcher expects only bool->cmove). 2879 if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) { 2880 cloned_for_outside_use += clone_for_use_outside_loop( loop, n, worklist ); 2881 sink_list.push(n); 2882 peel >>= n->_idx; // delete n from peel set. 2883 not_peel <<= n->_idx; // add n to not_peel set. 2884 peel_list.remove(i); 2885 incr = false; 2886 #if !defined(PRODUCT) 2887 if (TracePartialPeeling) { 2888 tty->print_cr("sink to not_peeled region: %d newbb: %d", 2889 n->_idx, get_ctrl(n)->_idx); 2890 } 2891 #endif 2892 } 2893 } else { 2894 // Otherwise check for special def-use cases that span 2895 // the peel/not_peel boundary such as bool->if 2896 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist ); 2897 new_phi_cnt++; 2898 } 2899 } 2900 } 2901 if (incr) i++; 2902 } 2903 2904 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) { 2905 #if !defined(PRODUCT) 2906 if (TracePartialPeeling) { 2907 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c", 2908 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F'); 2909 } 2910 #endif 2911 if (new_peel_if != NULL) { 2912 remove_cmpi_loop_exit(new_peel_if, loop); 2913 } 2914 // Inhibit more partial peeling on this loop 2915 assert(!head->is_partial_peel_loop(), "not partial peeled"); 2916 head->mark_partial_peel_failed(); 2917 if (cloned_for_outside_use > 0) { 2918 // Terminate this round of loop opts because 2919 // the graph outside this loop was changed. 2920 C->set_major_progress(); 2921 return true; 2922 } 2923 return false; 2924 } 2925 2926 // Step 3: clone loop, retarget control, and insert new phis 2927 2928 // Create new loop head for new phis and to hang 2929 // the nodes being moved (sinked) from the peel region. 2930 LoopNode* new_head = new LoopNode(last_peel, last_peel); 2931 new_head->set_unswitch_count(head->unswitch_count()); // Preserve 2932 _igvn.register_new_node_with_optimizer(new_head); 2933 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled"); 2934 _igvn.replace_input_of(first_not_peeled, 0, new_head); 2935 set_loop(new_head, loop); 2936 loop->_body.push(new_head); 2937 not_peel.set(new_head->_idx); 2938 set_idom(new_head, last_peel, dom_depth(first_not_peeled)); 2939 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled)); 2940 2941 while (sink_list.size() > 0) { 2942 Node* n = sink_list.pop(); 2943 set_ctrl(n, new_head); 2944 } 2945 2946 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 2947 2948 clone_loop( loop, old_new, dd ); 2949 2950 const uint clone_exit_idx = 1; 2951 const uint orig_exit_idx = 2; 2952 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop"); 2953 2954 Node* head_clone = old_new[head->_idx]; 2955 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop(); 2956 Node* orig_tail_clone = head_clone->in(2); 2957 2958 // Add phi if "def" node is in peel set and "use" is not 2959 2960 for(i = 0; i < peel_list.size(); i++ ) { 2961 Node *def = peel_list.at(i); 2962 if (!def->is_CFG()) { 2963 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 2964 Node *use = def->fast_out(j); 2965 if (has_node(use) && use->in(0) != C->top() && 2966 (!peel.test(use->_idx) || 2967 (use->is_Phi() && use->in(0) == head)) ) { 2968 worklist.push(use); 2969 } 2970 } 2971 while( worklist.size() ) { 2972 Node *use = worklist.pop(); 2973 for (uint j = 1; j < use->req(); j++) { 2974 Node* n = use->in(j); 2975 if (n == def) { 2976 2977 // "def" is in peel set, "use" is not in peel set 2978 // or "use" is in the entry boundary (a phi) of the peel set 2979 2980 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use; 2981 2982 if ( loop->is_member(get_loop( use_c )) ) { 2983 // use is in loop 2984 if (old_new[use->_idx] != NULL) { // null for dead code 2985 Node* use_clone = old_new[use->_idx]; 2986 _igvn.replace_input_of(use, j, C->top()); 2987 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone ); 2988 } 2989 } else { 2990 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format"); 2991 // use is not in the loop, check if the live range includes the cut 2992 Node* lp_if = use_c->in(orig_exit_idx)->in(0); 2993 if (not_peel.test(lp_if->_idx)) { 2994 assert(j == orig_exit_idx, "use from original loop"); 2995 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone ); 2996 } 2997 } 2998 } 2999 } 3000 } 3001 } 3002 } 3003 3004 // Step 3b: retarget control 3005 3006 // Redirect control to the new loop head if a cloned node in 3007 // the not_peeled region has control that points into the peeled region. 3008 // This necessary because the cloned peeled region will be outside 3009 // the loop. 3010 // from to 3011 // cloned-peeled <---+ 3012 // new_head_clone: | <--+ 3013 // cloned-not_peeled in(0) in(0) 3014 // orig-peeled 3015 3016 for(i = 0; i < loop->_body.size(); i++ ) { 3017 Node *n = loop->_body.at(i); 3018 if (!n->is_CFG() && n->in(0) != NULL && 3019 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) { 3020 Node* n_clone = old_new[n->_idx]; 3021 _igvn.replace_input_of(n_clone, 0, new_head_clone); 3022 } 3023 } 3024 3025 // Backedge of the surviving new_head (the clone) is original last_peel 3026 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel); 3027 3028 // Cut first node in original not_peel set 3029 _igvn.rehash_node_delayed(new_head); // Multiple edge updates: 3030 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of 3031 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls 3032 3033 // Copy head_clone back-branch info to original head 3034 // and remove original head's loop entry and 3035 // clone head's back-branch 3036 _igvn.rehash_node_delayed(head); // Multiple edge updates 3037 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl)); 3038 head->set_req(LoopNode::LoopBackControl, C->top()); 3039 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top()); 3040 3041 // Similarly modify the phis 3042 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) { 3043 Node* use = head->fast_out(k); 3044 if (use->is_Phi() && use->outcnt() > 0) { 3045 Node* use_clone = old_new[use->_idx]; 3046 _igvn.rehash_node_delayed(use); // Multiple edge updates 3047 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl)); 3048 use->set_req(LoopNode::LoopBackControl, C->top()); 3049 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top()); 3050 } 3051 } 3052 3053 // Step 4: update dominator tree and dominator depth 3054 3055 set_idom(head, orig_tail_clone, dd); 3056 recompute_dom_depth(); 3057 3058 // Inhibit more partial peeling on this loop 3059 new_head_clone->set_partial_peel_loop(); 3060 C->set_major_progress(); 3061 loop->record_for_igvn(); 3062 3063 #if !defined(PRODUCT) 3064 if (TracePartialPeeling) { 3065 tty->print_cr("\nafter partial peel one iteration"); 3066 Node_List wl(area); 3067 Node* t = last_peel; 3068 while (true) { 3069 wl.push(t); 3070 if (t == head_clone) break; 3071 t = idom(t); 3072 } 3073 while (wl.size() > 0) { 3074 Node* tt = wl.pop(); 3075 if (tt == head) tty->print_cr("orig head"); 3076 else if (tt == new_head_clone) tty->print_cr("new head"); 3077 else if (tt == head_clone) tty->print_cr("clone head"); 3078 tt->dump(); 3079 } 3080 } 3081 #endif 3082 return true; 3083 } 3084 3085 //------------------------------reorg_offsets---------------------------------- 3086 // Reorganize offset computations to lower register pressure. Mostly 3087 // prevent loop-fallout uses of the pre-incremented trip counter (which are 3088 // then alive with the post-incremented trip counter forcing an extra 3089 // register move) 3090 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) { 3091 // Perform it only for canonical counted loops. 3092 // Loop's shape could be messed up by iteration_split_impl. 3093 if (!loop->_head->is_CountedLoop()) 3094 return; 3095 if (!loop->_head->as_Loop()->is_valid_counted_loop()) 3096 return; 3097 3098 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 3099 CountedLoopEndNode *cle = cl->loopexit(); 3100 Node *exit = cle->proj_out(false); 3101 Node *phi = cl->phi(); 3102 3103 // Check for the special case of folks using the pre-incremented 3104 // trip-counter on the fall-out path (forces the pre-incremented 3105 // and post-incremented trip counter to be live at the same time). 3106 // Fix this by adjusting to use the post-increment trip counter. 3107 3108 bool progress = true; 3109 while (progress) { 3110 progress = false; 3111 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) { 3112 Node* use = phi->fast_out(i); // User of trip-counter 3113 if (!has_ctrl(use)) continue; 3114 Node *u_ctrl = get_ctrl(use); 3115 if (use->is_Phi()) { 3116 u_ctrl = NULL; 3117 for (uint j = 1; j < use->req(); j++) 3118 if (use->in(j) == phi) 3119 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j)); 3120 } 3121 IdealLoopTree *u_loop = get_loop(u_ctrl); 3122 // Look for loop-invariant use 3123 if (u_loop == loop) continue; 3124 if (loop->is_member(u_loop)) continue; 3125 // Check that use is live out the bottom. Assuming the trip-counter 3126 // update is right at the bottom, uses of of the loop middle are ok. 3127 if (dom_lca(exit, u_ctrl) != exit) continue; 3128 // Hit! Refactor use to use the post-incremented tripcounter. 3129 // Compute a post-increment tripcounter. 3130 Node *opaq = new Opaque2Node( C, cle->incr() ); 3131 register_new_node(opaq, exit); 3132 Node *neg_stride = _igvn.intcon(-cle->stride_con()); 3133 set_ctrl(neg_stride, C->root()); 3134 Node *post = new AddINode( opaq, neg_stride); 3135 register_new_node(post, exit); 3136 _igvn.rehash_node_delayed(use); 3137 for (uint j = 1; j < use->req(); j++) { 3138 if (use->in(j) == phi) 3139 use->set_req(j, post); 3140 } 3141 // Since DU info changed, rerun loop 3142 progress = true; 3143 break; 3144 } 3145 } 3146 3147 }