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