1 /* 2 * Copyright (c) 1999, 2018, 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/callnode.hpp" 30 #include "opto/castnode.hpp" 31 #include "opto/connode.hpp" 32 #include "opto/castnode.hpp" 33 #include "opto/divnode.hpp" 34 #include "opto/loopnode.hpp" 35 #include "opto/matcher.hpp" 36 #include "opto/mulnode.hpp" 37 #include "opto/movenode.hpp" 38 #include "opto/opaquenode.hpp" 39 #include "opto/rootnode.hpp" 40 #include "opto/subnode.hpp" 41 42 //============================================================================= 43 //------------------------------split_thru_phi--------------------------------- 44 // Split Node 'n' through merge point if there is enough win. 45 Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) { 46 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) { 47 // ConvI2L may have type information on it which is unsafe to push up 48 // so disable this for now 49 return NULL; 50 } 51 52 // Splitting range check CastIIs through a loop induction Phi can 53 // cause new Phis to be created that are left unrelated to the loop 54 // induction Phi and prevent optimizations (vectorization) 55 if (n->Opcode() == Op_CastII && n->as_CastII()->has_range_check() && 56 region->is_CountedLoop() && n->in(1) == region->as_CountedLoop()->phi()) { 57 return NULL; 58 } 59 60 int wins = 0; 61 assert(!n->is_CFG(), ""); 62 assert(region->is_Region(), ""); 63 64 const Type* type = n->bottom_type(); 65 const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr(); 66 Node *phi; 67 if (t_oop != NULL && t_oop->is_known_instance_field()) { 68 int iid = t_oop->instance_id(); 69 int index = C->get_alias_index(t_oop); 70 int offset = t_oop->offset(); 71 phi = new PhiNode(region, type, NULL, iid, index, offset); 72 } else { 73 phi = PhiNode::make_blank(region, n); 74 } 75 uint old_unique = C->unique(); 76 for (uint i = 1; i < region->req(); i++) { 77 Node *x; 78 Node* the_clone = NULL; 79 if (region->in(i) == C->top()) { 80 x = C->top(); // Dead path? Use a dead data op 81 } else { 82 x = n->clone(); // Else clone up the data op 83 the_clone = x; // Remember for possible deletion. 84 // Alter data node to use pre-phi inputs 85 if (n->in(0) == region) 86 x->set_req( 0, region->in(i) ); 87 for (uint j = 1; j < n->req(); j++) { 88 Node *in = n->in(j); 89 if (in->is_Phi() && in->in(0) == region) 90 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone 91 } 92 } 93 // Check for a 'win' on some paths 94 const Type *t = x->Value(&_igvn); 95 96 bool singleton = t->singleton(); 97 98 // A TOP singleton indicates that there are no possible values incoming 99 // along a particular edge. In most cases, this is OK, and the Phi will 100 // be eliminated later in an Ideal call. However, we can't allow this to 101 // happen if the singleton occurs on loop entry, as the elimination of 102 // the PhiNode may cause the resulting node to migrate back to a previous 103 // loop iteration. 104 if (singleton && t == Type::TOP) { 105 // Is_Loop() == false does not confirm the absence of a loop (e.g., an 106 // irreducible loop may not be indicated by an affirmative is_Loop()); 107 // therefore, the only top we can split thru a phi is on a backedge of 108 // a loop. 109 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 110 } 111 112 if (singleton) { 113 wins++; 114 x = ((PhaseGVN&)_igvn).makecon(t); 115 } else { 116 // We now call Identity to try to simplify the cloned node. 117 // Note that some Identity methods call phase->type(this). 118 // Make sure that the type array is big enough for 119 // our new node, even though we may throw the node away. 120 // (Note: This tweaking with igvn only works because x is a new node.) 121 _igvn.set_type(x, t); 122 // If x is a TypeNode, capture any more-precise type permanently into Node 123 // otherwise it will be not updated during igvn->transform since 124 // igvn->type(x) is set to x->Value() already. 125 x->raise_bottom_type(t); 126 Node *y = x->Identity(&_igvn); 127 if (y != x) { 128 wins++; 129 x = y; 130 } else { 131 y = _igvn.hash_find(x); 132 if (y) { 133 wins++; 134 x = y; 135 } else { 136 // Else x is a new node we are keeping 137 // We do not need register_new_node_with_optimizer 138 // because set_type has already been called. 139 _igvn._worklist.push(x); 140 } 141 } 142 } 143 if (x != the_clone && the_clone != NULL) 144 _igvn.remove_dead_node(the_clone); 145 phi->set_req( i, x ); 146 } 147 // Too few wins? 148 if (wins <= policy) { 149 _igvn.remove_dead_node(phi); 150 return NULL; 151 } 152 153 // Record Phi 154 register_new_node( phi, region ); 155 156 for (uint i2 = 1; i2 < phi->req(); i2++) { 157 Node *x = phi->in(i2); 158 // If we commoned up the cloned 'x' with another existing Node, 159 // the existing Node picks up a new use. We need to make the 160 // existing Node occur higher up so it dominates its uses. 161 Node *old_ctrl; 162 IdealLoopTree *old_loop; 163 164 if (x->is_Con()) { 165 // Constant's control is always root. 166 set_ctrl(x, C->root()); 167 continue; 168 } 169 // The occasional new node 170 if (x->_idx >= old_unique) { // Found a new, unplaced node? 171 old_ctrl = NULL; 172 old_loop = NULL; // Not in any prior loop 173 } else { 174 old_ctrl = get_ctrl(x); 175 old_loop = get_loop(old_ctrl); // Get prior loop 176 } 177 // New late point must dominate new use 178 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2)); 179 if (new_ctrl == old_ctrl) // Nothing is changed 180 continue; 181 182 IdealLoopTree *new_loop = get_loop(new_ctrl); 183 184 // Don't move x into a loop if its uses are 185 // outside of loop. Otherwise x will be cloned 186 // for each use outside of this loop. 187 IdealLoopTree *use_loop = get_loop(region); 188 if (!new_loop->is_member(use_loop) && 189 (old_loop == NULL || !new_loop->is_member(old_loop))) { 190 // Take early control, later control will be recalculated 191 // during next iteration of loop optimizations. 192 new_ctrl = get_early_ctrl(x); 193 new_loop = get_loop(new_ctrl); 194 } 195 // Set new location 196 set_ctrl(x, new_ctrl); 197 // If changing loop bodies, see if we need to collect into new body 198 if (old_loop != new_loop) { 199 if (old_loop && !old_loop->_child) 200 old_loop->_body.yank(x); 201 if (!new_loop->_child) 202 new_loop->_body.push(x); // Collect body info 203 } 204 } 205 206 return phi; 207 } 208 209 //------------------------------dominated_by------------------------------------ 210 // Replace the dominated test with an obvious true or false. Place it on the 211 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the 212 // live path up to the dominating control. 213 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff, bool flip, bool exclude_loop_predicate ) { 214 if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); } 215 216 // prevdom is the dominating projection of the dominating test. 217 assert( iff->is_If(), "" ); 218 assert(iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd || iff->Opcode() == Op_RangeCheck, "Check this code when new subtype is added"); 219 int pop = prevdom->Opcode(); 220 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" ); 221 if (flip) { 222 if (pop == Op_IfTrue) 223 pop = Op_IfFalse; 224 else 225 pop = Op_IfTrue; 226 } 227 // 'con' is set to true or false to kill the dominated test. 228 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO); 229 set_ctrl(con, C->root()); // Constant gets a new use 230 // Hack the dominated test 231 _igvn.replace_input_of(iff, 1, con); 232 233 // If I dont have a reachable TRUE and FALSE path following the IfNode then 234 // I can assume this path reaches an infinite loop. In this case it's not 235 // important to optimize the data Nodes - either the whole compilation will 236 // be tossed or this path (and all data Nodes) will go dead. 237 if (iff->outcnt() != 2) return; 238 239 // Make control-dependent data Nodes on the live path (path that will remain 240 // once the dominated IF is removed) become control-dependent on the 241 // dominating projection. 242 Node* dp = iff->as_If()->proj_out_or_null(pop == Op_IfTrue); 243 244 // Loop predicates may have depending checks which should not 245 // be skipped. For example, range check predicate has two checks 246 // for lower and upper bounds. 247 if (dp == NULL) 248 return; 249 250 ProjNode* dp_proj = dp->as_Proj(); 251 ProjNode* unc_proj = iff->as_If()->proj_out(1 - dp_proj->_con)->as_Proj(); 252 if (exclude_loop_predicate && 253 (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) != NULL || 254 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check) != NULL)) { 255 // If this is a range check (IfNode::is_range_check), do not 256 // reorder because Compile::allow_range_check_smearing might have 257 // changed the check. 258 return; // Let IGVN transformation change control dependence. 259 } 260 261 IdealLoopTree *old_loop = get_loop(dp); 262 263 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { 264 Node* cd = dp->fast_out(i); // Control-dependent node 265 if (cd->depends_only_on_test()) { 266 assert(cd->in(0) == dp, ""); 267 _igvn.replace_input_of(cd, 0, prevdom); 268 set_early_ctrl(cd); 269 IdealLoopTree *new_loop = get_loop(get_ctrl(cd)); 270 if (old_loop != new_loop) { 271 if (!old_loop->_child) old_loop->_body.yank(cd); 272 if (!new_loop->_child) new_loop->_body.push(cd); 273 } 274 --i; 275 --imax; 276 } 277 } 278 } 279 280 //------------------------------has_local_phi_input---------------------------- 281 // Return TRUE if 'n' has Phi inputs from its local block and no other 282 // block-local inputs (all non-local-phi inputs come from earlier blocks) 283 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) { 284 Node *n_ctrl = get_ctrl(n); 285 // See if some inputs come from a Phi in this block, or from before 286 // this block. 287 uint i; 288 for( i = 1; i < n->req(); i++ ) { 289 Node *phi = n->in(i); 290 if( phi->is_Phi() && phi->in(0) == n_ctrl ) 291 break; 292 } 293 if( i >= n->req() ) 294 return NULL; // No Phi inputs; nowhere to clone thru 295 296 // Check for inputs created between 'n' and the Phi input. These 297 // must split as well; they have already been given the chance 298 // (courtesy of a post-order visit) and since they did not we must 299 // recover the 'cost' of splitting them by being very profitable 300 // when splitting 'n'. Since this is unlikely we simply give up. 301 for( i = 1; i < n->req(); i++ ) { 302 Node *m = n->in(i); 303 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) { 304 // We allow the special case of AddP's with no local inputs. 305 // This allows us to split-up address expressions. 306 if (m->is_AddP() && 307 get_ctrl(m->in(2)) != n_ctrl && 308 get_ctrl(m->in(3)) != n_ctrl) { 309 // Move the AddP up to dominating point 310 Node* c = find_non_split_ctrl(idom(n_ctrl)); 311 if (c->is_OuterStripMinedLoop()) { 312 c->as_Loop()->verify_strip_mined(1); 313 c = c->in(LoopNode::EntryControl); 314 } 315 set_ctrl_and_loop(m, c); 316 continue; 317 } 318 return NULL; 319 } 320 assert(m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control"); 321 } 322 323 return n_ctrl; 324 } 325 326 //------------------------------remix_address_expressions---------------------- 327 // Rework addressing expressions to get the most loop-invariant stuff 328 // moved out. We'd like to do all associative operators, but it's especially 329 // important (common) to do address expressions. 330 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) { 331 if (!has_ctrl(n)) return NULL; 332 Node *n_ctrl = get_ctrl(n); 333 IdealLoopTree *n_loop = get_loop(n_ctrl); 334 335 // See if 'n' mixes loop-varying and loop-invariant inputs and 336 // itself is loop-varying. 337 338 // Only interested in binary ops (and AddP) 339 if( n->req() < 3 || n->req() > 4 ) return NULL; 340 341 Node *n1_ctrl = get_ctrl(n->in( 1)); 342 Node *n2_ctrl = get_ctrl(n->in( 2)); 343 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3)); 344 IdealLoopTree *n1_loop = get_loop( n1_ctrl ); 345 IdealLoopTree *n2_loop = get_loop( n2_ctrl ); 346 IdealLoopTree *n3_loop = get_loop( n3_ctrl ); 347 348 // Does one of my inputs spin in a tighter loop than self? 349 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) || 350 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) || 351 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) ) 352 return NULL; // Leave well enough alone 353 354 // Is at least one of my inputs loop-invariant? 355 if( n1_loop == n_loop && 356 n2_loop == n_loop && 357 n3_loop == n_loop ) 358 return NULL; // No loop-invariant inputs 359 360 361 int n_op = n->Opcode(); 362 363 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2). 364 if( n_op == Op_LShiftI ) { 365 // Scale is loop invariant 366 Node *scale = n->in(2); 367 Node *scale_ctrl = get_ctrl(scale); 368 IdealLoopTree *scale_loop = get_loop(scale_ctrl ); 369 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) ) 370 return NULL; 371 const TypeInt *scale_t = scale->bottom_type()->isa_int(); 372 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 ) 373 return NULL; // Dont bother with byte/short masking 374 // Add must vary with loop (else shift would be loop-invariant) 375 Node *add = n->in(1); 376 Node *add_ctrl = get_ctrl(add); 377 IdealLoopTree *add_loop = get_loop(add_ctrl); 378 //assert( n_loop == add_loop, "" ); 379 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops 380 381 // Convert I-V into I+ (0-V); same for V-I 382 if( add->Opcode() == Op_SubI && 383 _igvn.type( add->in(1) ) != TypeInt::ZERO ) { 384 Node *zero = _igvn.intcon(0); 385 set_ctrl(zero, C->root()); 386 Node *neg = new SubINode( _igvn.intcon(0), add->in(2) ); 387 register_new_node( neg, get_ctrl(add->in(2) ) ); 388 add = new AddINode( add->in(1), neg ); 389 register_new_node( add, add_ctrl ); 390 } 391 if( add->Opcode() != Op_AddI ) return NULL; 392 // See if one add input is loop invariant 393 Node *add_var = add->in(1); 394 Node *add_var_ctrl = get_ctrl(add_var); 395 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 396 Node *add_invar = add->in(2); 397 Node *add_invar_ctrl = get_ctrl(add_invar); 398 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl ); 399 if( add_var_loop == n_loop ) { 400 } else if( add_invar_loop == n_loop ) { 401 // Swap to find the invariant part 402 add_invar = add_var; 403 add_invar_ctrl = add_var_ctrl; 404 add_invar_loop = add_var_loop; 405 add_var = add->in(2); 406 Node *add_var_ctrl = get_ctrl(add_var); 407 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 408 } else // Else neither input is loop invariant 409 return NULL; 410 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) ) 411 return NULL; // No invariant part of the add? 412 413 // Yes! Reshape address expression! 414 Node *inv_scale = new LShiftINode( add_invar, scale ); 415 Node *inv_scale_ctrl = 416 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ? 417 add_invar_ctrl : scale_ctrl; 418 register_new_node( inv_scale, inv_scale_ctrl ); 419 Node *var_scale = new LShiftINode( add_var, scale ); 420 register_new_node( var_scale, n_ctrl ); 421 Node *var_add = new AddINode( var_scale, inv_scale ); 422 register_new_node( var_add, n_ctrl ); 423 _igvn.replace_node( n, var_add ); 424 return var_add; 425 } 426 427 // Replace (I+V) with (V+I) 428 if( n_op == Op_AddI || 429 n_op == Op_AddL || 430 n_op == Op_AddF || 431 n_op == Op_AddD || 432 n_op == Op_MulI || 433 n_op == Op_MulL || 434 n_op == Op_MulF || 435 n_op == Op_MulD ) { 436 if( n2_loop == n_loop ) { 437 assert( n1_loop != n_loop, "" ); 438 n->swap_edges(1, 2); 439 } 440 } 441 442 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V), 443 // but not if I2 is a constant. 444 if( n_op == Op_AddP ) { 445 if( n2_loop == n_loop && n3_loop != n_loop ) { 446 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) { 447 Node *n22_ctrl = get_ctrl(n->in(2)->in(2)); 448 Node *n23_ctrl = get_ctrl(n->in(2)->in(3)); 449 IdealLoopTree *n22loop = get_loop( n22_ctrl ); 450 IdealLoopTree *n23_loop = get_loop( n23_ctrl ); 451 if( n22loop != n_loop && n22loop->is_member(n_loop) && 452 n23_loop == n_loop ) { 453 Node *add1 = new AddPNode( n->in(1), n->in(2)->in(2), n->in(3) ); 454 // Stuff new AddP in the loop preheader 455 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) ); 456 Node *add2 = new AddPNode( n->in(1), add1, n->in(2)->in(3) ); 457 register_new_node( add2, n_ctrl ); 458 _igvn.replace_node( n, add2 ); 459 return add2; 460 } 461 } 462 } 463 464 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V) 465 if (n2_loop != n_loop && n3_loop == n_loop) { 466 if (n->in(3)->Opcode() == Op_AddX) { 467 Node *V = n->in(3)->in(1); 468 Node *I = n->in(3)->in(2); 469 if (is_member(n_loop,get_ctrl(V))) { 470 } else { 471 Node *tmp = V; V = I; I = tmp; 472 } 473 if (!is_member(n_loop,get_ctrl(I))) { 474 Node *add1 = new AddPNode(n->in(1), n->in(2), I); 475 // Stuff new AddP in the loop preheader 476 register_new_node(add1, n_loop->_head->in(LoopNode::EntryControl)); 477 Node *add2 = new AddPNode(n->in(1), add1, V); 478 register_new_node(add2, n_ctrl); 479 _igvn.replace_node(n, add2); 480 return add2; 481 } 482 } 483 } 484 } 485 486 return NULL; 487 } 488 489 //------------------------------conditional_move------------------------------- 490 // Attempt to replace a Phi with a conditional move. We have some pretty 491 // strict profitability requirements. All Phis at the merge point must 492 // be converted, so we can remove the control flow. We need to limit the 493 // number of c-moves to a small handful. All code that was in the side-arms 494 // of the CFG diamond is now speculatively executed. This code has to be 495 // "cheap enough". We are pretty much limited to CFG diamonds that merge 496 // 1 or 2 items with a total of 1 or 2 ops executed speculatively. 497 Node *PhaseIdealLoop::conditional_move( Node *region ) { 498 499 assert(region->is_Region(), "sanity check"); 500 if (region->req() != 3) return NULL; 501 502 // Check for CFG diamond 503 Node *lp = region->in(1); 504 Node *rp = region->in(2); 505 if (!lp || !rp) return NULL; 506 Node *lp_c = lp->in(0); 507 if (lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If()) return NULL; 508 IfNode *iff = lp_c->as_If(); 509 510 // Check for ops pinned in an arm of the diamond. 511 // Can't remove the control flow in this case 512 if (lp->outcnt() > 1) return NULL; 513 if (rp->outcnt() > 1) return NULL; 514 515 IdealLoopTree* r_loop = get_loop(region); 516 assert(r_loop == get_loop(iff), "sanity"); 517 // Always convert to CMOVE if all results are used only outside this loop. 518 bool used_inside_loop = (r_loop == _ltree_root); 519 520 // Check profitability 521 int cost = 0; 522 int phis = 0; 523 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 524 Node *out = region->fast_out(i); 525 if (!out->is_Phi()) continue; // Ignore other control edges, etc 526 phis++; 527 PhiNode* phi = out->as_Phi(); 528 BasicType bt = phi->type()->basic_type(); 529 switch (bt) { 530 case T_DOUBLE: 531 case T_FLOAT: 532 if (C->use_cmove()) { 533 continue; //TODO: maybe we want to add some cost 534 } 535 cost += Matcher::float_cmove_cost(); // Could be very expensive 536 break; 537 case T_LONG: { 538 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's 539 } 540 case T_INT: // These all CMOV fine 541 case T_ADDRESS: { // (RawPtr) 542 cost++; 543 break; 544 } 545 case T_NARROWOOP: // Fall through 546 case T_OBJECT: { // Base oops are OK, but not derived oops 547 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr(); 548 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a 549 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus 550 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we 551 // have a Phi for the base here that we convert to a CMOVE all is well 552 // and good. But if the base is dead, we'll not make a CMOVE. Later 553 // the allocator will have to produce a base by creating a CMOVE of the 554 // relevant bases. This puts the allocator in the business of 555 // manufacturing expensive instructions, generally a bad plan. 556 // Just Say No to Conditionally-Moved Derived Pointers. 557 if (tp && tp->offset() != 0) 558 return NULL; 559 cost++; 560 break; 561 } 562 default: 563 return NULL; // In particular, can't do memory or I/O 564 } 565 // Add in cost any speculative ops 566 for (uint j = 1; j < region->req(); j++) { 567 Node *proj = region->in(j); 568 Node *inp = phi->in(j); 569 if (get_ctrl(inp) == proj) { // Found local op 570 cost++; 571 // Check for a chain of dependent ops; these will all become 572 // speculative in a CMOV. 573 for (uint k = 1; k < inp->req(); k++) 574 if (get_ctrl(inp->in(k)) == proj) 575 cost += ConditionalMoveLimit; // Too much speculative goo 576 } 577 } 578 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode. 579 // This will likely Split-If, a higher-payoff operation. 580 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) { 581 Node* use = phi->fast_out(k); 582 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr()) 583 cost += ConditionalMoveLimit; 584 // Is there a use inside the loop? 585 // Note: check only basic types since CMoveP is pinned. 586 if (!used_inside_loop && is_java_primitive(bt)) { 587 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use); 588 if (r_loop == u_loop || r_loop->is_member(u_loop)) { 589 used_inside_loop = true; 590 } 591 } 592 } 593 }//for 594 Node* bol = iff->in(1); 595 assert(bol->Opcode() == Op_Bool, ""); 596 int cmp_op = bol->in(1)->Opcode(); 597 // It is expensive to generate flags from a float compare. 598 // Avoid duplicated float compare. 599 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL; 600 601 float infrequent_prob = PROB_UNLIKELY_MAG(3); 602 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop. 603 if (used_inside_loop) { 604 if (cost >= ConditionalMoveLimit) return NULL; // Too much goo 605 606 // BlockLayoutByFrequency optimization moves infrequent branch 607 // from hot path. No point in CMOV'ing in such case (110 is used 608 // instead of 100 to take into account not exactness of float value). 609 if (BlockLayoutByFrequency) { 610 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f); 611 } 612 } 613 // Check for highly predictable branch. No point in CMOV'ing if 614 // we are going to predict accurately all the time. 615 if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) { 616 //keep going 617 } else if (iff->_prob < infrequent_prob || 618 iff->_prob > (1.0f - infrequent_prob)) 619 return NULL; 620 621 // -------------- 622 // Now replace all Phis with CMOV's 623 Node *cmov_ctrl = iff->in(0); 624 uint flip = (lp->Opcode() == Op_IfTrue); 625 Node_List wq; 626 while (1) { 627 PhiNode* phi = NULL; 628 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 629 Node *out = region->fast_out(i); 630 if (out->is_Phi()) { 631 phi = out->as_Phi(); 632 break; 633 } 634 } 635 if (phi == NULL) break; 636 if (PrintOpto && VerifyLoopOptimizations) { tty->print_cr("CMOV"); } 637 // Move speculative ops 638 wq.push(phi); 639 while (wq.size() > 0) { 640 Node *n = wq.pop(); 641 for (uint j = 1; j < n->req(); j++) { 642 Node* m = n->in(j); 643 if (m != NULL && !is_dominator(get_ctrl(m), cmov_ctrl)) { 644 #ifndef PRODUCT 645 if (PrintOpto && VerifyLoopOptimizations) { 646 tty->print(" speculate: "); 647 m->dump(); 648 } 649 #endif 650 set_ctrl(m, cmov_ctrl); 651 wq.push(m); 652 } 653 } 654 } 655 Node *cmov = CMoveNode::make(cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi)); 656 register_new_node( cmov, cmov_ctrl ); 657 _igvn.replace_node( phi, cmov ); 658 #ifndef PRODUCT 659 if (TraceLoopOpts) { 660 tty->print("CMOV "); 661 r_loop->dump_head(); 662 if (Verbose) { 663 bol->in(1)->dump(1); 664 cmov->dump(1); 665 } 666 } 667 if (VerifyLoopOptimizations) verify(); 668 #endif 669 } 670 671 // The useless CFG diamond will fold up later; see the optimization in 672 // RegionNode::Ideal. 673 _igvn._worklist.push(region); 674 675 return iff->in(1); 676 } 677 678 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) { 679 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 680 Node* u = m->fast_out(i); 681 if (u->is_CFG()) { 682 if (u->Opcode() == Op_NeverBranch) { 683 u = ((NeverBranchNode*)u)->proj_out(0); 684 enqueue_cfg_uses(u, wq); 685 } else { 686 wq.push(u); 687 } 688 } 689 } 690 } 691 692 // Try moving a store out of a loop, right before the loop 693 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) { 694 // Store has to be first in the loop body 695 IdealLoopTree *n_loop = get_loop(n_ctrl); 696 if (n->is_Store() && n_loop != _ltree_root && 697 n_loop->is_loop() && n_loop->_head->is_Loop() && 698 n->in(0) != NULL) { 699 Node* address = n->in(MemNode::Address); 700 Node* value = n->in(MemNode::ValueIn); 701 Node* mem = n->in(MemNode::Memory); 702 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 703 IdealLoopTree* value_loop = get_loop(get_ctrl(value)); 704 705 // - address and value must be loop invariant 706 // - memory must be a memory Phi for the loop 707 // - Store must be the only store on this memory slice in the 708 // loop: if there's another store following this one then value 709 // written at iteration i by the second store could be overwritten 710 // at iteration i+n by the first store: it's not safe to move the 711 // first store out of the loop 712 // - nothing must observe the memory Phi: it guarantees no read 713 // before the store, we are also guaranteed the store post 714 // dominates the loop head (ignoring a possible early 715 // exit). Otherwise there would be extra Phi involved between the 716 // loop's Phi and the store. 717 // - there must be no early exit from the loop before the Store 718 // (such an exit most of the time would be an extra use of the 719 // memory Phi but sometimes is a bottom memory Phi that takes the 720 // store as input). 721 722 if (!n_loop->is_member(address_loop) && 723 !n_loop->is_member(value_loop) && 724 mem->is_Phi() && mem->in(0) == n_loop->_head && 725 mem->outcnt() == 1 && 726 mem->in(LoopNode::LoopBackControl) == n) { 727 728 assert(n_loop->_tail != NULL, "need a tail"); 729 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop"); 730 731 // Verify that there's no early exit of the loop before the store. 732 bool ctrl_ok = false; 733 { 734 // Follow control from loop head until n, we exit the loop or 735 // we reach the tail 736 ResourceMark rm; 737 Unique_Node_List wq; 738 wq.push(n_loop->_head); 739 740 for (uint next = 0; next < wq.size(); ++next) { 741 Node *m = wq.at(next); 742 if (m == n->in(0)) { 743 ctrl_ok = true; 744 continue; 745 } 746 assert(!has_ctrl(m), "should be CFG"); 747 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) { 748 ctrl_ok = false; 749 break; 750 } 751 enqueue_cfg_uses(m, wq); 752 if (wq.size() > 10) { 753 ctrl_ok = false; 754 break; 755 } 756 } 757 } 758 if (ctrl_ok) { 759 // move the Store 760 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem); 761 _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl)); 762 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl)); 763 // Disconnect the phi now. An empty phi can confuse other 764 // optimizations in this pass of loop opts. 765 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl)); 766 n_loop->_body.yank(mem); 767 768 set_ctrl_and_loop(n, n->in(0)); 769 770 return n; 771 } 772 } 773 } 774 return NULL; 775 } 776 777 // Try moving a store out of a loop, right after the loop 778 void PhaseIdealLoop::try_move_store_after_loop(Node* n) { 779 if (n->is_Store() && n->in(0) != NULL) { 780 Node *n_ctrl = get_ctrl(n); 781 IdealLoopTree *n_loop = get_loop(n_ctrl); 782 // Store must be in a loop 783 if (n_loop != _ltree_root && !n_loop->_irreducible) { 784 Node* address = n->in(MemNode::Address); 785 Node* value = n->in(MemNode::ValueIn); 786 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 787 // address must be loop invariant 788 if (!n_loop->is_member(address_loop)) { 789 // Store must be last on this memory slice in the loop and 790 // nothing in the loop must observe it 791 Node* phi = NULL; 792 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 793 Node* u = n->fast_out(i); 794 if (has_ctrl(u)) { // control use? 795 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 796 if (!n_loop->is_member(u_loop)) { 797 continue; 798 } 799 if (u->is_Phi() && u->in(0) == n_loop->_head) { 800 assert(_igvn.type(u) == Type::MEMORY, "bad phi"); 801 // multiple phis on the same slice are possible 802 if (phi != NULL) { 803 return; 804 } 805 phi = u; 806 continue; 807 } 808 } 809 return; 810 } 811 if (phi != NULL) { 812 // Nothing in the loop before the store (next iteration) 813 // must observe the stored value 814 bool mem_ok = true; 815 { 816 ResourceMark rm; 817 Unique_Node_List wq; 818 wq.push(phi); 819 for (uint next = 0; next < wq.size() && mem_ok; ++next) { 820 Node *m = wq.at(next); 821 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) { 822 Node* u = m->fast_out(i); 823 if (u->is_Store() || u->is_Phi()) { 824 if (u != n) { 825 wq.push(u); 826 mem_ok = (wq.size() <= 10); 827 } 828 } else { 829 mem_ok = false; 830 break; 831 } 832 } 833 } 834 } 835 if (mem_ok) { 836 // Move the store out of the loop if the LCA of all 837 // users (except for the phi) is outside the loop. 838 Node* hook = new Node(1); 839 _igvn.rehash_node_delayed(phi); 840 int count = phi->replace_edge(n, hook); 841 assert(count > 0, "inconsistent phi"); 842 843 // Compute latest point this store can go 844 Node* lca = get_late_ctrl(n, get_ctrl(n)); 845 if (n_loop->is_member(get_loop(lca))) { 846 // LCA is in the loop - bail out 847 _igvn.replace_node(hook, n); 848 return; 849 } 850 #ifdef ASSERT 851 if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) { 852 assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined"); 853 n_loop->_head->as_Loop()->verify_strip_mined(1); 854 Node* outer = n_loop->_head->as_CountedLoop()->outer_loop(); 855 IdealLoopTree* outer_loop = get_loop(outer); 856 assert(n_loop->_parent == outer_loop, "broken loop tree"); 857 assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state"); 858 } 859 #endif 860 861 // Move store out of the loop 862 _igvn.replace_node(hook, n->in(MemNode::Memory)); 863 _igvn.replace_input_of(n, 0, lca); 864 set_ctrl_and_loop(n, lca); 865 866 // Disconnect the phi now. An empty phi can confuse other 867 // optimizations in this pass of loop opts.. 868 if (phi->in(LoopNode::LoopBackControl) == phi) { 869 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl)); 870 n_loop->_body.yank(phi); 871 } 872 } 873 } 874 } 875 } 876 } 877 } 878 879 //------------------------------split_if_with_blocks_pre----------------------- 880 // Do the real work in a non-recursive function. Data nodes want to be 881 // cloned in the pre-order so they can feed each other nicely. 882 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) { 883 // Cloning these guys is unlikely to win 884 int n_op = n->Opcode(); 885 if( n_op == Op_MergeMem ) return n; 886 if( n->is_Proj() ) return n; 887 // Do not clone-up CmpFXXX variations, as these are always 888 // followed by a CmpI 889 if( n->is_Cmp() ) return n; 890 // Attempt to use a conditional move instead of a phi/branch 891 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) { 892 Node *cmov = conditional_move( n ); 893 if( cmov ) return cmov; 894 } 895 if( n->is_CFG() || n->is_LoadStore() ) 896 return n; 897 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd 898 n_op == Op_Opaque2 ) { 899 if( !C->major_progress() ) // If chance of no more loop opts... 900 _igvn._worklist.push(n); // maybe we'll remove them 901 return n; 902 } 903 904 if( n->is_Con() ) return n; // No cloning for Con nodes 905 906 Node *n_ctrl = get_ctrl(n); 907 if( !n_ctrl ) return n; // Dead node 908 909 Node* res = try_move_store_before_loop(n, n_ctrl); 910 if (res != NULL) { 911 return n; 912 } 913 914 // Attempt to remix address expressions for loop invariants 915 Node *m = remix_address_expressions( n ); 916 if( m ) return m; 917 918 if (n->is_ConstraintCast()) { 919 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this); 920 // ConstraintCastNode::dominating_cast() uses node control input to determine domination. 921 // Node control inputs don't necessarily agree with loop control info (due to 922 // transformations happened in between), thus additional dominance check is needed 923 // to keep loop info valid. 924 if (dom_cast != NULL && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) { 925 _igvn.replace_node(n, dom_cast); 926 return dom_cast; 927 } 928 } 929 930 // Determine if the Node has inputs from some local Phi. 931 // Returns the block to clone thru. 932 Node *n_blk = has_local_phi_input( n ); 933 if( !n_blk ) return n; 934 935 // Do not clone the trip counter through on a CountedLoop 936 // (messes up the canonical shape). 937 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n; 938 939 // Check for having no control input; not pinned. Allow 940 // dominating control. 941 if (n->in(0)) { 942 Node *dom = idom(n_blk); 943 if (dom_lca(n->in(0), dom) != n->in(0)) { 944 return n; 945 } 946 } 947 // Policy: when is it profitable. You must get more wins than 948 // policy before it is considered profitable. Policy is usually 0, 949 // so 1 win is considered profitable. Big merges will require big 950 // cloning, so get a larger policy. 951 int policy = n_blk->req() >> 2; 952 953 // If the loop is a candidate for range check elimination, 954 // delay splitting through it's phi until a later loop optimization 955 if (n_blk->is_CountedLoop()) { 956 IdealLoopTree *lp = get_loop(n_blk); 957 if (lp && lp->_rce_candidate) { 958 return n; 959 } 960 } 961 962 // Use same limit as split_if_with_blocks_post 963 if( C->live_nodes() > 35000 ) return n; // Method too big 964 965 // Split 'n' through the merge point if it is profitable 966 Node *phi = split_thru_phi( n, n_blk, policy ); 967 if (!phi) return n; 968 969 // Found a Phi to split thru! 970 // Replace 'n' with the new phi 971 _igvn.replace_node( n, phi ); 972 // Moved a load around the loop, 'en-registering' something. 973 if (n_blk->is_Loop() && n->is_Load() && 974 !phi->in(LoopNode::LoopBackControl)->is_Load()) 975 C->set_major_progress(); 976 977 return phi; 978 } 979 980 static bool merge_point_too_heavy(Compile* C, Node* region) { 981 // Bail out if the region and its phis have too many users. 982 int weight = 0; 983 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 984 weight += region->fast_out(i)->outcnt(); 985 } 986 int nodes_left = C->max_node_limit() - C->live_nodes(); 987 if (weight * 8 > nodes_left) { 988 if (PrintOpto) { 989 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight); 990 } 991 return true; 992 } else { 993 return false; 994 } 995 } 996 997 static bool merge_point_safe(Node* region) { 998 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode 999 // having a PhiNode input. This sidesteps the dangerous case where the split 1000 // ConvI2LNode may become TOP if the input Value() does not 1001 // overlap the ConvI2L range, leaving a node which may not dominate its 1002 // uses. 1003 // A better fix for this problem can be found in the BugTraq entry, but 1004 // expediency for Mantis demands this hack. 1005 // 6855164: If the merge point has a FastLockNode with a PhiNode input, we stop 1006 // split_if_with_blocks from splitting a block because we could not move around 1007 // the FastLockNode. 1008 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1009 Node* n = region->fast_out(i); 1010 if (n->is_Phi()) { 1011 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1012 Node* m = n->fast_out(j); 1013 if (m->is_FastLock()) 1014 return false; 1015 #ifdef _LP64 1016 if (m->Opcode() == Op_ConvI2L) 1017 return false; 1018 if (m->is_CastII() && m->isa_CastII()->has_range_check()) { 1019 return false; 1020 } 1021 #endif 1022 } 1023 } 1024 } 1025 return true; 1026 } 1027 1028 1029 //------------------------------place_near_use--------------------------------- 1030 // Place some computation next to use but not inside inner loops. 1031 // For inner loop uses move it to the preheader area. 1032 Node *PhaseIdealLoop::place_near_use(Node *useblock) const { 1033 IdealLoopTree *u_loop = get_loop( useblock ); 1034 if (u_loop->_irreducible) { 1035 return useblock; 1036 } 1037 if (u_loop->_child) { 1038 if (useblock == u_loop->_head && u_loop->_head->is_OuterStripMinedLoop()) { 1039 return u_loop->_head->in(LoopNode::EntryControl); 1040 } 1041 return useblock; 1042 } 1043 return u_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl); 1044 } 1045 1046 1047 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) { 1048 if (!n->is_If() || n->is_CountedLoopEnd()) { 1049 return false; 1050 } 1051 if (!n->in(0)->is_Region()) { 1052 return false; 1053 } 1054 Node* region = n->in(0); 1055 Node* dom = idom(region); 1056 if (!dom->is_If() || dom->in(1) != n->in(1)) { 1057 return false; 1058 } 1059 IfNode* dom_if = dom->as_If(); 1060 Node* proj_true = dom_if->proj_out(1); 1061 Node* proj_false = dom_if->proj_out(0); 1062 1063 for (uint i = 1; i < region->req(); i++) { 1064 if (is_dominator(proj_true, region->in(i))) { 1065 continue; 1066 } 1067 if (is_dominator(proj_false, region->in(i))) { 1068 continue; 1069 } 1070 return false; 1071 } 1072 1073 return true; 1074 } 1075 1076 bool PhaseIdealLoop::can_split_if(Node *n_ctrl) { 1077 if (C->live_nodes() > 35000) { 1078 return false; // Method too big 1079 } 1080 1081 // Do not do 'split-if' if irreducible loops are present. 1082 if (_has_irreducible_loops) { 1083 return false; 1084 } 1085 1086 if (merge_point_too_heavy(C, n_ctrl)) { 1087 return false; 1088 } 1089 1090 // Do not do 'split-if' if some paths are dead. First do dead code 1091 // elimination and then see if its still profitable. 1092 for (uint i = 1; i < n_ctrl->req(); i++) { 1093 if (n_ctrl->in(i) == C->top()) { 1094 return false; 1095 } 1096 } 1097 1098 // If trying to do a 'Split-If' at the loop head, it is only 1099 // profitable if the cmp folds up on BOTH paths. Otherwise we 1100 // risk peeling a loop forever. 1101 1102 // CNC - Disabled for now. Requires careful handling of loop 1103 // body selection for the cloned code. Also, make sure we check 1104 // for any input path not being in the same loop as n_ctrl. For 1105 // irreducible loops we cannot check for 'n_ctrl->is_Loop()' 1106 // because the alternative loop entry points won't be converted 1107 // into LoopNodes. 1108 IdealLoopTree *n_loop = get_loop(n_ctrl); 1109 for (uint j = 1; j < n_ctrl->req(); j++) { 1110 if (get_loop(n_ctrl->in(j)) != n_loop) { 1111 return false; 1112 } 1113 } 1114 1115 // Check for safety of the merge point. 1116 if (!merge_point_safe(n_ctrl)) { 1117 return false; 1118 } 1119 1120 return true; 1121 } 1122 1123 bool PhaseIdealLoop::replace_with_dominating_barrier(LoadBarrierNode* lb, bool last_round) { 1124 LoadBarrierNode* lb2 = lb->has_dominating_barrier(this, false, last_round); 1125 if (lb2 != NULL) { 1126 if (lb->in(LoadBarrierNode::Oop) != lb2->in(LoadBarrierNode::Oop)) { 1127 assert(lb->in(LoadBarrierNode::Address) == lb2->in(LoadBarrierNode::Address), ""); 1128 _igvn.replace_input_of(lb, LoadBarrierNode::Similar, lb2->proj_out(LoadBarrierNode::Oop)); 1129 C->set_major_progress(); 1130 } else { 1131 // That transformation may cause the Similar edge on dominated load barriers to be invalid 1132 lb->fix_similar_in_uses(&_igvn); 1133 1134 Node* val = lb->proj_out(LoadBarrierNode::Oop); 1135 assert(lb2->has_true_uses(), ""); 1136 assert(lb2->in(LoadBarrierNode::Oop) == lb->in(LoadBarrierNode::Oop), ""); 1137 1138 lazy_update(lb, lb->in(LoadBarrierNode::Control)); 1139 lazy_replace(lb->proj_out(LoadBarrierNode::Control), lb->in(LoadBarrierNode::Control)); 1140 _igvn.replace_node(val, lb2->proj_out(LoadBarrierNode::Oop)); 1141 1142 return true; 1143 } 1144 } 1145 return false; 1146 } 1147 1148 Node* PhaseIdealLoop::find_dominating_memory(Node* mem, Node* dom, int i) { 1149 assert(dom->is_Region() || i == -1, ""); 1150 Node* m = mem; 1151 while(is_dominator(dom, has_ctrl(m) ? get_ctrl(m) : m->in(0))) { 1152 if (m->is_Mem()) { 1153 assert(m->as_Mem()->adr_type() == TypeRawPtr::BOTTOM, ""); 1154 m = m->in(MemNode::Memory); 1155 } else if (m->is_MergeMem()) { 1156 m = m->as_MergeMem()->memory_at(Compile::AliasIdxRaw); 1157 } else if (m->is_Phi()) { 1158 if (m->in(0) == dom && i != -1) { 1159 m = m->in(i); 1160 break; 1161 } else { 1162 m = m->in(LoopNode::EntryControl); 1163 } 1164 } else if (m->is_Proj()) { 1165 m = m->in(0); 1166 } else if (m->is_SafePoint() || m->is_MemBar()) { 1167 m = m->in(TypeFunc::Memory); 1168 } else { 1169 #ifdef ASSERT 1170 m->dump(); 1171 #endif 1172 ShouldNotReachHere(); 1173 } 1174 } 1175 return m; 1176 } 1177 1178 LoadBarrierNode* PhaseIdealLoop::clone_load_barrier(LoadBarrierNode* lb, Node* ctl, Node* mem, Node* oop_in) { 1179 Node* the_clone = lb->clone(); 1180 the_clone->set_req(LoadBarrierNode::Control, ctl); 1181 the_clone->set_req(LoadBarrierNode::Memory, mem); 1182 if (oop_in != NULL) { 1183 the_clone->set_req(LoadBarrierNode::Oop, oop_in); 1184 } 1185 1186 LoadBarrierNode* new_lb = the_clone->as_LoadBarrier(); 1187 _igvn.register_new_node_with_optimizer(new_lb); 1188 IdealLoopTree *loop = get_loop(new_lb->in(0)); 1189 set_ctrl(new_lb, new_lb->in(0)); 1190 set_loop(new_lb, loop); 1191 set_idom(new_lb, new_lb->in(0), dom_depth(new_lb->in(0))+1); 1192 if (!loop->_child) { 1193 loop->_body.push(new_lb); 1194 } 1195 1196 Node* proj_ctl = new ProjNode(new_lb, LoadBarrierNode::Control); 1197 _igvn.register_new_node_with_optimizer(proj_ctl); 1198 set_ctrl(proj_ctl, proj_ctl->in(0)); 1199 set_loop(proj_ctl, loop); 1200 set_idom(proj_ctl, new_lb, dom_depth(new_lb)+1); 1201 if (!loop->_child) { 1202 loop->_body.push(proj_ctl); 1203 } 1204 1205 Node* proj_oop = new ProjNode(new_lb, LoadBarrierNode::Oop); 1206 register_new_node(proj_oop, new_lb); 1207 1208 if (!new_lb->in(LoadBarrierNode::Similar)->is_top()) { 1209 LoadBarrierNode* similar = new_lb->in(LoadBarrierNode::Similar)->in(0)->as_LoadBarrier(); 1210 if (!is_dominator(similar, ctl)) { 1211 _igvn.replace_input_of(new_lb, LoadBarrierNode::Similar, C->top()); 1212 } 1213 } 1214 1215 return new_lb; 1216 } 1217 1218 void PhaseIdealLoop::replace_barrier(LoadBarrierNode* lb, Node* new_val) { 1219 Node* val = lb->proj_out(LoadBarrierNode::Oop); 1220 _igvn.replace_node(val, new_val); 1221 lazy_update(lb, lb->in(LoadBarrierNode::Control)); 1222 lazy_replace(lb->proj_out(LoadBarrierNode::Control), lb->in(LoadBarrierNode::Control)); 1223 } 1224 1225 bool PhaseIdealLoop::split_barrier_thru_phi(LoadBarrierNode* lb) { 1226 if (lb->in(LoadBarrierNode::Oop)->is_Phi()) { 1227 Node* oop_phi = lb->in(LoadBarrierNode::Oop); 1228 1229 if (oop_phi->req() == 2) { 1230 // Ignore phis with only one input 1231 return false; 1232 } 1233 1234 if (is_dominator(get_ctrl(lb->in(LoadBarrierNode::Address)), oop_phi->in(0)) && get_ctrl(lb->in(LoadBarrierNode::Address)) != oop_phi->in(0) /*&& (get_ctrl(lb->in(LoadBarrierNode::Memory)) != lb->in(0) || lb->in(LoadBarrierNode::Memory)->is_Phi())*/) { 1235 // That transformation may cause the Similar edge on dominated load barriers to be invalid 1236 lb->fix_similar_in_uses(&_igvn); 1237 1238 RegionNode* region = oop_phi->in(0)->as_Region(); 1239 1240 int backedge = LoopNode::LoopBackControl; 1241 if (region->is_Loop() && region->in(backedge)->is_Proj() && region->in(backedge)->in(0)->is_If()) { 1242 Node* c = region->in(backedge)->in(0)->in(0); 1243 assert(c->unique_ctrl_out() == region->in(backedge)->in(0), ""); 1244 Node* oop = lb->in(LoadBarrierNode::Oop)->in(backedge); 1245 Node* oop_c = has_ctrl(oop) ? get_ctrl(oop) : oop; 1246 if (!is_dominator(oop_c, c)) { 1247 return false; 1248 } 1249 } 1250 1251 Node *phi = oop_phi->clone(); 1252 1253 for (uint i = 1; i < region->req(); i++) { 1254 Node* ctrl = region->in(i); 1255 if (ctrl != C->top()) { 1256 assert(!is_dominator(ctrl, region) || region->is_Loop(), ""); 1257 1258 Node* mem = lb->in(LoadBarrierNode::Memory); 1259 Node* m = find_dominating_memory(mem, region, i); 1260 1261 if (region->is_Loop() && i == LoopNode::LoopBackControl && ctrl->is_Proj() && ctrl->in(0)->is_If()) { 1262 ctrl = ctrl->in(0)->in(0); 1263 } 1264 1265 LoadBarrierNode* new_lb = clone_load_barrier(lb, ctrl, m, lb->in(LoadBarrierNode::Oop)->in(i)); 1266 1267 Node* out_ctrl = new_lb->proj_out(LoadBarrierNode::Control); 1268 if (ctrl == region->in(i)) { 1269 _igvn.replace_input_of(region, i, new_lb->proj_out(LoadBarrierNode::Control)); 1270 } else { 1271 Node* iff = region->in(i)->in(0); 1272 Node* out_ctrl = new_lb->proj_out(LoadBarrierNode::Control); 1273 _igvn.replace_input_of(iff, 0, out_ctrl); 1274 set_idom(iff, out_ctrl, dom_depth(out_ctrl)+1); 1275 } 1276 phi->set_req(i, new_lb->proj_out(LoadBarrierNode::Oop)); 1277 } 1278 } 1279 register_new_node(phi, region); 1280 1281 replace_barrier(lb, phi); 1282 1283 if (region->is_Loop()) { 1284 // Load barrier moved to the back edge of the Loop may now 1285 // have a safepoint on the path to the barrier on the Similar 1286 // edge 1287 _igvn.replace_input_of(phi->in(LoopNode::LoopBackControl)->in(0), LoadBarrierNode::Similar, C->top()); 1288 Node* head = region->in(LoopNode::EntryControl); 1289 set_idom(region, head, dom_depth(head)+1); 1290 recompute_dom_depth(); 1291 if (head->is_CountedLoop() && head->as_CountedLoop()->is_main_loop()) { 1292 head->as_CountedLoop()->set_normal_loop(); 1293 } 1294 } 1295 return true; 1296 } 1297 } 1298 return false; 1299 } 1300 1301 bool PhaseIdealLoop::move_out_of_loop(LoadBarrierNode* lb) { 1302 IdealLoopTree *lb_loop = get_loop(lb->in(0)); 1303 if (lb_loop != _ltree_root && !lb_loop->_irreducible) { 1304 Node* oop_ctrl = get_ctrl(lb->in(LoadBarrierNode::Oop)); 1305 IdealLoopTree *oop_loop = get_loop(oop_ctrl); 1306 IdealLoopTree* adr_loop = get_loop(get_ctrl(lb->in(LoadBarrierNode::Address))); 1307 if (!lb_loop->is_member(oop_loop) && !lb_loop->is_member(adr_loop)) { 1308 // That transformation may cause the Similar edge on dominated load barriers to be invalid 1309 lb->fix_similar_in_uses(&_igvn); 1310 1311 Node* head = lb_loop->_head; 1312 assert(head->is_Loop(), ""); 1313 1314 if (is_dominator(head, oop_ctrl)) { 1315 assert(oop_ctrl->Opcode() == Op_CProj && oop_ctrl->in(0)->Opcode() == Op_NeverBranch, ""); 1316 assert(lb_loop->is_member(get_loop(oop_ctrl->in(0)->in(0))), ""); 1317 return false; 1318 } 1319 1320 if (head->is_CountedLoop()) { 1321 CountedLoopNode* cloop = head->as_CountedLoop(); 1322 if (cloop->is_main_loop()) { 1323 cloop->set_normal_loop(); 1324 } 1325 // When we are moving barrier out of a counted loop, 1326 // make sure we move it all the way out of the strip mined outer loop. 1327 if (cloop->is_strip_mined()) { 1328 head = cloop->outer_loop(); 1329 } 1330 } 1331 1332 Node* mem = lb->in(LoadBarrierNode::Memory); 1333 Node* m = find_dominating_memory(mem, head); 1334 1335 LoadBarrierNode* new_lb = clone_load_barrier(lb, head->in(LoopNode::EntryControl), m, NULL); 1336 1337 assert(idom(head) == head->in(LoopNode::EntryControl), ""); 1338 Node* proj_ctl = new_lb->proj_out(LoadBarrierNode::Control); 1339 _igvn.replace_input_of(head, LoopNode::EntryControl, proj_ctl); 1340 set_idom(head, proj_ctl, dom_depth(proj_ctl)+1); 1341 1342 replace_barrier(lb, new_lb->proj_out(LoadBarrierNode::Oop)); 1343 1344 recompute_dom_depth(); 1345 1346 return true; 1347 } 1348 } 1349 return false; 1350 } 1351 1352 bool PhaseIdealLoop::common_barriers(LoadBarrierNode* lb) { 1353 Node* in_val = lb->in(LoadBarrierNode::Oop); 1354 for (DUIterator_Fast imax, i = in_val->fast_outs(imax); i < imax; i++) { 1355 Node* u = in_val->fast_out(i); 1356 if (u != lb && u->is_LoadBarrier() && u->as_LoadBarrier()->has_true_uses()) { 1357 Node* this_ctrl = lb->in(LoadBarrierNode::Control); 1358 Node* other_ctrl = u->in(LoadBarrierNode::Control); 1359 1360 Node* lca = dom_lca(this_ctrl, other_ctrl); 1361 bool ok = true; 1362 1363 Node* proj1 = NULL; 1364 Node* proj2 = NULL; 1365 1366 while (this_ctrl != lca && ok) { 1367 if (this_ctrl->in(0) != NULL && 1368 this_ctrl->in(0)->is_MultiBranch()) { 1369 if (this_ctrl->in(0)->in(0) == lca) { 1370 assert(proj1 == NULL, ""); 1371 assert(this_ctrl->is_Proj(), ""); 1372 proj1 = this_ctrl; 1373 } else if (!(this_ctrl->in(0)->is_If() && this_ctrl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none))) { 1374 ok = false; 1375 } 1376 } 1377 this_ctrl = idom(this_ctrl); 1378 } 1379 while (other_ctrl != lca && ok) { 1380 if (other_ctrl->in(0) != NULL && 1381 other_ctrl->in(0)->is_MultiBranch()) { 1382 if (other_ctrl->in(0)->in(0) == lca) { 1383 assert(other_ctrl->is_Proj(), ""); 1384 assert(proj2 == NULL, ""); 1385 proj2 = other_ctrl; 1386 } else if (!(other_ctrl->in(0)->is_If() && other_ctrl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none))) { 1387 ok = false; 1388 } 1389 } 1390 other_ctrl = idom(other_ctrl); 1391 } 1392 assert(proj1 == NULL || proj2 == NULL || proj1->in(0) == proj2->in(0), ""); 1393 if (ok && proj1 && proj2 && proj1 != proj2 && proj1->in(0)->is_If()) { 1394 // That transformation may cause the Similar edge on dominated load barriers to be invalid 1395 lb->fix_similar_in_uses(&_igvn); 1396 u->as_LoadBarrier()->fix_similar_in_uses(&_igvn); 1397 1398 Node* split = lca->unique_ctrl_out(); 1399 assert(split->in(0) == lca, ""); 1400 1401 Node* mem = lb->in(LoadBarrierNode::Memory); 1402 Node* m = find_dominating_memory(mem, split); 1403 LoadBarrierNode* new_lb = clone_load_barrier(lb, lca, m, NULL); 1404 1405 Node* proj_ctl = new_lb->proj_out(LoadBarrierNode::Control); 1406 _igvn.replace_input_of(split, 0, new_lb->proj_out(LoadBarrierNode::Control)); 1407 set_idom(split, proj_ctl, dom_depth(proj_ctl)+1); 1408 1409 Node* proj_oop = new_lb->proj_out(LoadBarrierNode::Oop); 1410 replace_barrier(lb, proj_oop); 1411 replace_barrier(u->as_LoadBarrier(), proj_oop); 1412 1413 recompute_dom_depth(); 1414 1415 return true; 1416 } 1417 } 1418 } 1419 return false; 1420 } 1421 1422 void PhaseIdealLoop::optimize_load_barrier(LoadBarrierNode* lb, bool last_round) { 1423 if (!C->directive()->OptimizeLoadBarriersOption) { 1424 return; 1425 } 1426 1427 if (lb->has_true_uses()) { 1428 1429 if (replace_with_dominating_barrier(lb, last_round)) { 1430 return; 1431 } 1432 1433 if (split_barrier_thru_phi(lb)) { 1434 return; 1435 } 1436 1437 if (move_out_of_loop(lb)) { 1438 return; 1439 } 1440 1441 if (common_barriers(lb)) { 1442 return; 1443 } 1444 } 1445 } 1446 1447 //------------------------------split_if_with_blocks_post---------------------- 1448 // Do the real work in a non-recursive function. CFG hackery wants to be 1449 // in the post-order, so it can dirty the I-DOM info and not use the dirtied 1450 // info. 1451 void PhaseIdealLoop::split_if_with_blocks_post(Node *n, bool last_round) { 1452 1453 // Cloning Cmp through Phi's involves the split-if transform. 1454 // FastLock is not used by an If 1455 if (n->is_Cmp() && !n->is_FastLock() && !last_round) { 1456 Node *n_ctrl = get_ctrl(n); 1457 // Determine if the Node has inputs from some local Phi. 1458 // Returns the block to clone thru. 1459 Node *n_blk = has_local_phi_input(n); 1460 if (n_blk != n_ctrl) { 1461 return; 1462 } 1463 1464 if (!can_split_if(n_ctrl)) { 1465 return; 1466 } 1467 1468 if (n->outcnt() != 1) { 1469 return; // Multiple bool's from 1 compare? 1470 } 1471 Node *bol = n->unique_out(); 1472 assert(bol->is_Bool(), "expect a bool here"); 1473 if (bol->outcnt() != 1) { 1474 return;// Multiple branches from 1 compare? 1475 } 1476 Node *iff = bol->unique_out(); 1477 1478 // Check some safety conditions 1479 if (iff->is_If()) { // Classic split-if? 1480 if (iff->in(0) != n_ctrl) { 1481 return; // Compare must be in same blk as if 1482 } 1483 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE 1484 // Can't split CMove with different control edge. 1485 if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) { 1486 return; 1487 } 1488 if (get_ctrl(iff->in(2)) == n_ctrl || 1489 get_ctrl(iff->in(3)) == n_ctrl) { 1490 return; // Inputs not yet split-up 1491 } 1492 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) { 1493 return; // Loop-invar test gates loop-varying CMOVE 1494 } 1495 } else { 1496 return; // some other kind of node, such as an Allocate 1497 } 1498 1499 // When is split-if profitable? Every 'win' on means some control flow 1500 // goes dead, so it's almost always a win. 1501 int policy = 0; 1502 // Split compare 'n' through the merge point if it is profitable 1503 Node *phi = split_thru_phi( n, n_ctrl, policy); 1504 if (!phi) { 1505 return; 1506 } 1507 1508 // Found a Phi to split thru! 1509 // Replace 'n' with the new phi 1510 _igvn.replace_node(n, phi); 1511 1512 // Now split the bool up thru the phi 1513 Node *bolphi = split_thru_phi(bol, n_ctrl, -1); 1514 guarantee(bolphi != NULL, "null boolean phi node"); 1515 1516 _igvn.replace_node(bol, bolphi); 1517 assert(iff->in(1) == bolphi, ""); 1518 1519 if (bolphi->Value(&_igvn)->singleton()) { 1520 return; 1521 } 1522 1523 // Conditional-move? Must split up now 1524 if (!iff->is_If()) { 1525 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1); 1526 _igvn.replace_node(iff, cmovphi); 1527 return; 1528 } 1529 1530 // Now split the IF 1531 do_split_if(iff); 1532 return; 1533 } 1534 1535 // Two identical ifs back to back can be merged 1536 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) { 1537 Node *n_ctrl = n->in(0); 1538 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1)); 1539 IfNode* dom_if = idom(n_ctrl)->as_If(); 1540 Node* proj_true = dom_if->proj_out(1); 1541 Node* proj_false = dom_if->proj_out(0); 1542 Node* con_true = _igvn.makecon(TypeInt::ONE); 1543 Node* con_false = _igvn.makecon(TypeInt::ZERO); 1544 1545 for (uint i = 1; i < n_ctrl->req(); i++) { 1546 if (is_dominator(proj_true, n_ctrl->in(i))) { 1547 bolphi->init_req(i, con_true); 1548 } else { 1549 assert(is_dominator(proj_false, n_ctrl->in(i)), "bad if"); 1550 bolphi->init_req(i, con_false); 1551 } 1552 } 1553 register_new_node(bolphi, n_ctrl); 1554 _igvn.replace_input_of(n, 1, bolphi); 1555 1556 // Now split the IF 1557 do_split_if(n); 1558 return; 1559 } 1560 1561 // Check for an IF ready to split; one that has its 1562 // condition codes input coming from a Phi at the block start. 1563 int n_op = n->Opcode(); 1564 1565 // Check for an IF being dominated by another IF same test 1566 if (n_op == Op_If || 1567 n_op == Op_RangeCheck) { 1568 Node *bol = n->in(1); 1569 uint max = bol->outcnt(); 1570 // Check for same test used more than once? 1571 if (max > 1 && bol->is_Bool()) { 1572 // Search up IDOMs to see if this IF is dominated. 1573 Node *cutoff = get_ctrl(bol); 1574 1575 // Now search up IDOMs till cutoff, looking for a dominating test 1576 Node *prevdom = n; 1577 Node *dom = idom(prevdom); 1578 while (dom != cutoff) { 1579 if (dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom) { 1580 // Replace the dominated test with an obvious true or false. 1581 // Place it on the IGVN worklist for later cleanup. 1582 C->set_major_progress(); 1583 dominated_by(prevdom, n, false, true); 1584 #ifndef PRODUCT 1585 if( VerifyLoopOptimizations ) verify(); 1586 #endif 1587 return; 1588 } 1589 prevdom = dom; 1590 dom = idom(prevdom); 1591 } 1592 } 1593 } 1594 1595 // See if a shared loop-varying computation has no loop-varying uses. 1596 // Happens if something is only used for JVM state in uncommon trap exits, 1597 // like various versions of induction variable+offset. Clone the 1598 // computation per usage to allow it to sink out of the loop. 1599 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about) 1600 Node *n_ctrl = get_ctrl(n); 1601 IdealLoopTree *n_loop = get_loop(n_ctrl); 1602 if( n_loop != _ltree_root ) { 1603 DUIterator_Fast imax, i = n->fast_outs(imax); 1604 for (; i < imax; i++) { 1605 Node* u = n->fast_out(i); 1606 if( !has_ctrl(u) ) break; // Found control user 1607 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 1608 if( u_loop == n_loop ) break; // Found loop-varying use 1609 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop 1610 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003 1611 } 1612 bool did_break = (i < imax); // Did we break out of the previous loop? 1613 if (!did_break && n->outcnt() > 1) { // All uses in outer loops! 1614 Node *late_load_ctrl = NULL; 1615 if (n->is_Load()) { 1616 // If n is a load, get and save the result from get_late_ctrl(), 1617 // to be later used in calculating the control for n's clones. 1618 clear_dom_lca_tags(); 1619 late_load_ctrl = get_late_ctrl(n, n_ctrl); 1620 } 1621 // If n is a load, and the late control is the same as the current 1622 // control, then the cloning of n is a pointless exercise, because 1623 // GVN will ensure that we end up where we started. 1624 if (!n->is_Load() || late_load_ctrl != n_ctrl) { 1625 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) { 1626 Node *u = n->last_out(j); // Clone private computation per use 1627 _igvn.rehash_node_delayed(u); 1628 Node *x = n->clone(); // Clone computation 1629 Node *x_ctrl = NULL; 1630 if( u->is_Phi() ) { 1631 // Replace all uses of normal nodes. Replace Phi uses 1632 // individually, so the separate Nodes can sink down 1633 // different paths. 1634 uint k = 1; 1635 while( u->in(k) != n ) k++; 1636 u->set_req( k, x ); 1637 // x goes next to Phi input path 1638 x_ctrl = u->in(0)->in(k); 1639 --j; 1640 } else { // Normal use 1641 // Replace all uses 1642 for( uint k = 0; k < u->req(); k++ ) { 1643 if( u->in(k) == n ) { 1644 u->set_req( k, x ); 1645 --j; 1646 } 1647 } 1648 x_ctrl = get_ctrl(u); 1649 } 1650 1651 // Find control for 'x' next to use but not inside inner loops. 1652 // For inner loop uses get the preheader area. 1653 x_ctrl = place_near_use(x_ctrl); 1654 1655 if (n->is_Load()) { 1656 // For loads, add a control edge to a CFG node outside of the loop 1657 // to force them to not combine and return back inside the loop 1658 // during GVN optimization (4641526). 1659 // 1660 // Because we are setting the actual control input, factor in 1661 // the result from get_late_ctrl() so we respect any 1662 // anti-dependences. (6233005). 1663 x_ctrl = dom_lca(late_load_ctrl, x_ctrl); 1664 1665 // Don't allow the control input to be a CFG splitting node. 1666 // Such nodes should only have ProjNodes as outs, e.g. IfNode 1667 // should only have IfTrueNode and IfFalseNode (4985384). 1668 x_ctrl = find_non_split_ctrl(x_ctrl); 1669 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone"); 1670 1671 x->set_req(0, x_ctrl); 1672 } 1673 register_new_node(x, x_ctrl); 1674 1675 // Some institutional knowledge is needed here: 'x' is 1676 // yanked because if the optimizer runs GVN on it all the 1677 // cloned x's will common up and undo this optimization and 1678 // be forced back in the loop. This is annoying because it 1679 // makes +VerifyOpto report false-positives on progress. I 1680 // tried setting control edges on the x's to force them to 1681 // not combine, but the matching gets worried when it tries 1682 // to fold a StoreP and an AddP together (as part of an 1683 // address expression) and the AddP and StoreP have 1684 // different controls. 1685 if (!x->is_Load() && !x->is_DecodeNarrowPtr()) _igvn._worklist.yank(x); 1686 } 1687 _igvn.remove_dead_node(n); 1688 } 1689 } 1690 } 1691 } 1692 1693 try_move_store_after_loop(n); 1694 1695 // Check for Opaque2's who's loop has disappeared - who's input is in the 1696 // same loop nest as their output. Remove 'em, they are no longer useful. 1697 if( n_op == Op_Opaque2 && 1698 n->in(1) != NULL && 1699 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) { 1700 _igvn.replace_node( n, n->in(1) ); 1701 } 1702 1703 if (n->is_LoadBarrier()) { 1704 optimize_load_barrier(n->as_LoadBarrier(), last_round); 1705 } 1706 } 1707 1708 //------------------------------split_if_with_blocks--------------------------- 1709 // Check for aggressive application of 'split-if' optimization, 1710 // using basic block level info. 1711 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack, bool last_round) { 1712 Node *n = C->root(); 1713 visited.set(n->_idx); // first, mark node as visited 1714 // Do pre-visit work for root 1715 n = split_if_with_blocks_pre( n ); 1716 uint cnt = n->outcnt(); 1717 uint i = 0; 1718 while (true) { 1719 // Visit all children 1720 if (i < cnt) { 1721 Node* use = n->raw_out(i); 1722 ++i; 1723 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) { 1724 // Now do pre-visit work for this use 1725 use = split_if_with_blocks_pre( use ); 1726 nstack.push(n, i); // Save parent and next use's index. 1727 n = use; // Process all children of current use. 1728 cnt = use->outcnt(); 1729 i = 0; 1730 } 1731 } 1732 else { 1733 // All of n's children have been processed, complete post-processing. 1734 if (cnt != 0 && !n->is_Con()) { 1735 assert(has_node(n), "no dead nodes"); 1736 split_if_with_blocks_post(n, last_round); 1737 } 1738 if (nstack.is_empty()) { 1739 // Finished all nodes on stack. 1740 break; 1741 } 1742 // Get saved parent node and next use's index. Visit the rest of uses. 1743 n = nstack.node(); 1744 cnt = n->outcnt(); 1745 i = nstack.index(); 1746 nstack.pop(); 1747 } 1748 } 1749 } 1750 1751 1752 //============================================================================= 1753 // 1754 // C L O N E A L O O P B O D Y 1755 // 1756 1757 //------------------------------clone_iff-------------------------------------- 1758 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1759 // "Nearly" because all Nodes have been cloned from the original in the loop, 1760 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1761 // through the Phi recursively, and return a Bool. 1762 Node* PhaseIdealLoop::clone_iff(PhiNode *phi, IdealLoopTree *loop) { 1763 1764 // Convert this Phi into a Phi merging Bools 1765 uint i; 1766 for (i = 1; i < phi->req(); i++) { 1767 Node *b = phi->in(i); 1768 if (b->is_Phi()) { 1769 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi(), loop)); 1770 } else { 1771 assert(b->is_Bool() || b->Opcode() == Op_Opaque4, ""); 1772 } 1773 } 1774 1775 Node* n = phi->in(1); 1776 Node* sample_opaque = NULL; 1777 Node *sample_bool = NULL; 1778 if (n->Opcode() == Op_Opaque4) { 1779 sample_opaque = n; 1780 sample_bool = n->in(1); 1781 assert(sample_bool->is_Bool(), "wrong type"); 1782 } else { 1783 sample_bool = n; 1784 } 1785 Node *sample_cmp = sample_bool->in(1); 1786 1787 // Make Phis to merge the Cmp's inputs. 1788 PhiNode *phi1 = new PhiNode(phi->in(0), Type::TOP); 1789 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP); 1790 for (i = 1; i < phi->req(); i++) { 1791 Node *n1 = sample_opaque == NULL ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1); 1792 Node *n2 = sample_opaque == NULL ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2); 1793 phi1->set_req(i, n1); 1794 phi2->set_req(i, n2); 1795 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 1796 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 1797 } 1798 // See if these Phis have been made before. 1799 // Register with optimizer 1800 Node *hit1 = _igvn.hash_find_insert(phi1); 1801 if (hit1) { // Hit, toss just made Phi 1802 _igvn.remove_dead_node(phi1); // Remove new phi 1803 assert(hit1->is_Phi(), "" ); 1804 phi1 = (PhiNode*)hit1; // Use existing phi 1805 } else { // Miss 1806 _igvn.register_new_node_with_optimizer(phi1); 1807 } 1808 Node *hit2 = _igvn.hash_find_insert(phi2); 1809 if (hit2) { // Hit, toss just made Phi 1810 _igvn.remove_dead_node(phi2); // Remove new phi 1811 assert(hit2->is_Phi(), "" ); 1812 phi2 = (PhiNode*)hit2; // Use existing phi 1813 } else { // Miss 1814 _igvn.register_new_node_with_optimizer(phi2); 1815 } 1816 // Register Phis with loop/block info 1817 set_ctrl(phi1, phi->in(0)); 1818 set_ctrl(phi2, phi->in(0)); 1819 // Make a new Cmp 1820 Node *cmp = sample_cmp->clone(); 1821 cmp->set_req(1, phi1); 1822 cmp->set_req(2, phi2); 1823 _igvn.register_new_node_with_optimizer(cmp); 1824 set_ctrl(cmp, phi->in(0)); 1825 1826 // Make a new Bool 1827 Node *b = sample_bool->clone(); 1828 b->set_req(1,cmp); 1829 _igvn.register_new_node_with_optimizer(b); 1830 set_ctrl(b, phi->in(0)); 1831 1832 if (sample_opaque != NULL) { 1833 Node* opaque = sample_opaque->clone(); 1834 opaque->set_req(1, b); 1835 _igvn.register_new_node_with_optimizer(opaque); 1836 set_ctrl(opaque, phi->in(0)); 1837 return opaque; 1838 } 1839 1840 assert(b->is_Bool(), ""); 1841 return b; 1842 } 1843 1844 //------------------------------clone_bool------------------------------------- 1845 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1846 // "Nearly" because all Nodes have been cloned from the original in the loop, 1847 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1848 // through the Phi recursively, and return a Bool. 1849 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) { 1850 uint i; 1851 // Convert this Phi into a Phi merging Bools 1852 for( i = 1; i < phi->req(); i++ ) { 1853 Node *b = phi->in(i); 1854 if( b->is_Phi() ) { 1855 _igvn.replace_input_of(phi, i, clone_bool( b->as_Phi(), loop )); 1856 } else { 1857 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" ); 1858 } 1859 } 1860 1861 Node *sample_cmp = phi->in(1); 1862 1863 // Make Phis to merge the Cmp's inputs. 1864 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP ); 1865 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP ); 1866 for( uint j = 1; j < phi->req(); j++ ) { 1867 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP 1868 Node *n1, *n2; 1869 if( cmp_top->is_Cmp() ) { 1870 n1 = cmp_top->in(1); 1871 n2 = cmp_top->in(2); 1872 } else { 1873 n1 = n2 = cmp_top; 1874 } 1875 phi1->set_req( j, n1 ); 1876 phi2->set_req( j, n2 ); 1877 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 1878 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 1879 } 1880 1881 // See if these Phis have been made before. 1882 // Register with optimizer 1883 Node *hit1 = _igvn.hash_find_insert(phi1); 1884 if( hit1 ) { // Hit, toss just made Phi 1885 _igvn.remove_dead_node(phi1); // Remove new phi 1886 assert( hit1->is_Phi(), "" ); 1887 phi1 = (PhiNode*)hit1; // Use existing phi 1888 } else { // Miss 1889 _igvn.register_new_node_with_optimizer(phi1); 1890 } 1891 Node *hit2 = _igvn.hash_find_insert(phi2); 1892 if( hit2 ) { // Hit, toss just made Phi 1893 _igvn.remove_dead_node(phi2); // Remove new phi 1894 assert( hit2->is_Phi(), "" ); 1895 phi2 = (PhiNode*)hit2; // Use existing phi 1896 } else { // Miss 1897 _igvn.register_new_node_with_optimizer(phi2); 1898 } 1899 // Register Phis with loop/block info 1900 set_ctrl(phi1, phi->in(0)); 1901 set_ctrl(phi2, phi->in(0)); 1902 // Make a new Cmp 1903 Node *cmp = sample_cmp->clone(); 1904 cmp->set_req( 1, phi1 ); 1905 cmp->set_req( 2, phi2 ); 1906 _igvn.register_new_node_with_optimizer(cmp); 1907 set_ctrl(cmp, phi->in(0)); 1908 1909 assert( cmp->is_Cmp(), "" ); 1910 return (CmpNode*)cmp; 1911 } 1912 1913 //------------------------------sink_use--------------------------------------- 1914 // If 'use' was in the loop-exit block, it now needs to be sunk 1915 // below the post-loop merge point. 1916 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) { 1917 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) { 1918 set_ctrl(use, post_loop); 1919 for (DUIterator j = use->outs(); use->has_out(j); j++) 1920 sink_use(use->out(j), post_loop); 1921 } 1922 } 1923 1924 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new, 1925 IdealLoopTree* loop, IdealLoopTree* outer_loop, 1926 Node_List*& split_if_set, Node_List*& split_bool_set, 1927 Node_List*& split_cex_set, Node_List& worklist, 1928 uint new_counter, CloneLoopMode mode) { 1929 Node* nnn = old_new[old->_idx]; 1930 // Copy uses to a worklist, so I can munge the def-use info 1931 // with impunity. 1932 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1933 worklist.push(old->fast_out(j)); 1934 1935 while( worklist.size() ) { 1936 Node *use = worklist.pop(); 1937 if (!has_node(use)) continue; // Ignore dead nodes 1938 if (use->in(0) == C->top()) continue; 1939 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1940 // Check for data-use outside of loop - at least one of OLD or USE 1941 // must not be a CFG node. 1942 #ifdef ASSERT 1943 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == NULL) { 1944 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint(); 1945 assert(mode == ControlAroundStripMined && use == sfpt, "missed a node"); 1946 } 1947 #endif 1948 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) { 1949 1950 // If the Data use is an IF, that means we have an IF outside of the 1951 // loop that is switching on a condition that is set inside of the 1952 // loop. Happens if people set a loop-exit flag; then test the flag 1953 // in the loop to break the loop, then test is again outside of the 1954 // loop to determine which way the loop exited. 1955 // Loop predicate If node connects to Bool node through Opaque1 node. 1956 if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use) || use->Opcode() == Op_Opaque4) { 1957 // Since this code is highly unlikely, we lazily build the worklist 1958 // of such Nodes to go split. 1959 if (!split_if_set) { 1960 ResourceArea *area = Thread::current()->resource_area(); 1961 split_if_set = new Node_List(area); 1962 } 1963 split_if_set->push(use); 1964 } 1965 if (use->is_Bool()) { 1966 if (!split_bool_set) { 1967 ResourceArea *area = Thread::current()->resource_area(); 1968 split_bool_set = new Node_List(area); 1969 } 1970 split_bool_set->push(use); 1971 } 1972 if (use->Opcode() == Op_CreateEx) { 1973 if (!split_cex_set) { 1974 ResourceArea *area = Thread::current()->resource_area(); 1975 split_cex_set = new Node_List(area); 1976 } 1977 split_cex_set->push(use); 1978 } 1979 1980 1981 // Get "block" use is in 1982 uint idx = 0; 1983 while( use->in(idx) != old ) idx++; 1984 Node *prev = use->is_CFG() ? use : get_ctrl(use); 1985 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" ); 1986 Node *cfg = prev->_idx >= new_counter 1987 ? prev->in(2) 1988 : idom(prev); 1989 if( use->is_Phi() ) // Phi use is in prior block 1990 cfg = prev->in(idx); // NOT in block of Phi itself 1991 if (cfg->is_top()) { // Use is dead? 1992 _igvn.replace_input_of(use, idx, C->top()); 1993 continue; 1994 } 1995 1996 while(!outer_loop->is_member(get_loop(cfg))) { 1997 prev = cfg; 1998 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg); 1999 } 2000 // If the use occurs after merging several exits from the loop, then 2001 // old value must have dominated all those exits. Since the same old 2002 // value was used on all those exits we did not need a Phi at this 2003 // merge point. NOW we do need a Phi here. Each loop exit value 2004 // is now merged with the peeled body exit; each exit gets its own 2005 // private Phi and those Phis need to be merged here. 2006 Node *phi; 2007 if( prev->is_Region() ) { 2008 if( idx == 0 ) { // Updating control edge? 2009 phi = prev; // Just use existing control 2010 } else { // Else need a new Phi 2011 phi = PhiNode::make( prev, old ); 2012 // Now recursively fix up the new uses of old! 2013 for( uint i = 1; i < prev->req(); i++ ) { 2014 worklist.push(phi); // Onto worklist once for each 'old' input 2015 } 2016 } 2017 } else { 2018 // Get new RegionNode merging old and new loop exits 2019 prev = old_new[prev->_idx]; 2020 assert( prev, "just made this in step 7" ); 2021 if( idx == 0) { // Updating control edge? 2022 phi = prev; // Just use existing control 2023 } else { // Else need a new Phi 2024 // Make a new Phi merging data values properly 2025 phi = PhiNode::make( prev, old ); 2026 phi->set_req( 1, nnn ); 2027 } 2028 } 2029 // If inserting a new Phi, check for prior hits 2030 if( idx != 0 ) { 2031 Node *hit = _igvn.hash_find_insert(phi); 2032 if( hit == NULL ) { 2033 _igvn.register_new_node_with_optimizer(phi); // Register new phi 2034 } else { // or 2035 // Remove the new phi from the graph and use the hit 2036 _igvn.remove_dead_node(phi); 2037 phi = hit; // Use existing phi 2038 } 2039 set_ctrl(phi, prev); 2040 } 2041 // Make 'use' use the Phi instead of the old loop body exit value 2042 _igvn.replace_input_of(use, idx, phi); 2043 if( use->_idx >= new_counter ) { // If updating new phis 2044 // Not needed for correctness, but prevents a weak assert 2045 // in AddPNode from tripping (when we end up with different 2046 // base & derived Phis that will become the same after 2047 // IGVN does CSE). 2048 Node *hit = _igvn.hash_find_insert(use); 2049 if( hit ) // Go ahead and re-hash for hits. 2050 _igvn.replace_node( use, hit ); 2051 } 2052 2053 // If 'use' was in the loop-exit block, it now needs to be sunk 2054 // below the post-loop merge point. 2055 sink_use( use, prev ); 2056 } 2057 } 2058 } 2059 2060 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop, 2061 IdealLoopTree* outer_loop, int dd, Node_List &old_new, 2062 Node_List& extra_data_nodes) { 2063 if (head->is_strip_mined() && mode != IgnoreStripMined) { 2064 CountedLoopNode* cl = head->as_CountedLoop(); 2065 Node* l = cl->outer_loop(); 2066 Node* tail = cl->outer_loop_tail(); 2067 IfNode* le = cl->outer_loop_end(); 2068 Node* sfpt = cl->outer_safepoint(); 2069 CountedLoopEndNode* cle = cl->loopexit(); 2070 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop(); 2071 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null(); 2072 Node* cle_out = cle->proj_out(false); 2073 2074 Node* new_sfpt = NULL; 2075 Node* new_cle_out = cle_out->clone(); 2076 old_new.map(cle_out->_idx, new_cle_out); 2077 if (mode == CloneIncludesStripMined) { 2078 // clone outer loop body 2079 Node* new_l = l->clone(); 2080 Node* new_tail = tail->clone(); 2081 IfNode* new_le = le->clone()->as_If(); 2082 new_sfpt = sfpt->clone(); 2083 2084 set_loop(new_l, outer_loop->_parent); 2085 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd); 2086 set_loop(new_cle_out, outer_loop->_parent); 2087 set_idom(new_cle_out, new_cle, dd); 2088 set_loop(new_sfpt, outer_loop->_parent); 2089 set_idom(new_sfpt, new_cle_out, dd); 2090 set_loop(new_le, outer_loop->_parent); 2091 set_idom(new_le, new_sfpt, dd); 2092 set_loop(new_tail, outer_loop->_parent); 2093 set_idom(new_tail, new_le, dd); 2094 set_idom(new_cl, new_l, dd); 2095 2096 old_new.map(l->_idx, new_l); 2097 old_new.map(tail->_idx, new_tail); 2098 old_new.map(le->_idx, new_le); 2099 old_new.map(sfpt->_idx, new_sfpt); 2100 2101 new_l->set_req(LoopNode::LoopBackControl, new_tail); 2102 new_l->set_req(0, new_l); 2103 new_tail->set_req(0, new_le); 2104 new_le->set_req(0, new_sfpt); 2105 new_sfpt->set_req(0, new_cle_out); 2106 new_cle_out->set_req(0, new_cle); 2107 new_cl->set_req(LoopNode::EntryControl, new_l); 2108 2109 _igvn.register_new_node_with_optimizer(new_l); 2110 _igvn.register_new_node_with_optimizer(new_tail); 2111 _igvn.register_new_node_with_optimizer(new_le); 2112 } else { 2113 Node *newhead = old_new[loop->_head->_idx]; 2114 newhead->as_Loop()->clear_strip_mined(); 2115 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl)); 2116 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 2117 } 2118 // Look at data node that were assigned a control in the outer 2119 // loop: they are kept in the outer loop by the safepoint so start 2120 // from the safepoint node's inputs. 2121 IdealLoopTree* outer_loop = get_loop(l); 2122 Node_Stack stack(2); 2123 stack.push(sfpt, 1); 2124 uint new_counter = C->unique(); 2125 while (stack.size() > 0) { 2126 Node* n = stack.node(); 2127 uint i = stack.index(); 2128 while (i < n->req() && 2129 (n->in(i) == NULL || 2130 !has_ctrl(n->in(i)) || 2131 get_loop(get_ctrl(n->in(i))) != outer_loop || 2132 (old_new[n->in(i)->_idx] != NULL && old_new[n->in(i)->_idx]->_idx >= new_counter))) { 2133 i++; 2134 } 2135 if (i < n->req()) { 2136 stack.set_index(i+1); 2137 stack.push(n->in(i), 0); 2138 } else { 2139 assert(old_new[n->_idx] == NULL || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet"); 2140 Node* m = n == sfpt ? new_sfpt : n->clone(); 2141 if (m != NULL) { 2142 for (uint i = 0; i < n->req(); i++) { 2143 if (m->in(i) != NULL && old_new[m->in(i)->_idx] != NULL) { 2144 m->set_req(i, old_new[m->in(i)->_idx]); 2145 } 2146 } 2147 } else { 2148 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?"); 2149 } 2150 if (n != sfpt) { 2151 extra_data_nodes.push(n); 2152 _igvn.register_new_node_with_optimizer(m); 2153 assert(get_ctrl(n) == cle_out, "what other control?"); 2154 set_ctrl(m, new_cle_out); 2155 old_new.map(n->_idx, m); 2156 } 2157 stack.pop(); 2158 } 2159 } 2160 if (mode == CloneIncludesStripMined) { 2161 _igvn.register_new_node_with_optimizer(new_sfpt); 2162 _igvn.register_new_node_with_optimizer(new_cle_out); 2163 } 2164 } else { 2165 Node *newhead = old_new[loop->_head->_idx]; 2166 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 2167 } 2168 } 2169 2170 //------------------------------clone_loop------------------------------------- 2171 // 2172 // C L O N E A L O O P B O D Y 2173 // 2174 // This is the basic building block of the loop optimizations. It clones an 2175 // entire loop body. It makes an old_new loop body mapping; with this mapping 2176 // you can find the new-loop equivalent to an old-loop node. All new-loop 2177 // nodes are exactly equal to their old-loop counterparts, all edges are the 2178 // same. All exits from the old-loop now have a RegionNode that merges the 2179 // equivalent new-loop path. This is true even for the normal "loop-exit" 2180 // condition. All uses of loop-invariant old-loop values now come from (one 2181 // or more) Phis that merge their new-loop equivalents. 2182 // 2183 // This operation leaves the graph in an illegal state: there are two valid 2184 // control edges coming from the loop pre-header to both loop bodies. I'll 2185 // definitely have to hack the graph after running this transform. 2186 // 2187 // From this building block I will further edit edges to perform loop peeling 2188 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc. 2189 // 2190 // Parameter side_by_size_idom: 2191 // When side_by_size_idom is NULL, the dominator tree is constructed for 2192 // the clone loop to dominate the original. Used in construction of 2193 // pre-main-post loop sequence. 2194 // When nonnull, the clone and original are side-by-side, both are 2195 // dominated by the side_by_side_idom node. Used in construction of 2196 // unswitched loops. 2197 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd, 2198 CloneLoopMode mode, Node* side_by_side_idom) { 2199 2200 LoopNode* head = loop->_head->as_Loop(); 2201 head->verify_strip_mined(1); 2202 2203 if (C->do_vector_loop() && PrintOpto) { 2204 const char* mname = C->method()->name()->as_quoted_ascii(); 2205 if (mname != NULL) { 2206 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname); 2207 } 2208 } 2209 2210 CloneMap& cm = C->clone_map(); 2211 Dict* dict = cm.dict(); 2212 if (C->do_vector_loop()) { 2213 cm.set_clone_idx(cm.max_gen()+1); 2214 #ifndef PRODUCT 2215 if (PrintOpto) { 2216 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx()); 2217 loop->dump_head(); 2218 } 2219 #endif 2220 } 2221 2222 // Step 1: Clone the loop body. Make the old->new mapping. 2223 uint i; 2224 for( i = 0; i < loop->_body.size(); i++ ) { 2225 Node *old = loop->_body.at(i); 2226 Node *nnn = old->clone(); 2227 old_new.map( old->_idx, nnn ); 2228 if (C->do_vector_loop()) { 2229 cm.verify_insert_and_clone(old, nnn, cm.clone_idx()); 2230 } 2231 _igvn.register_new_node_with_optimizer(nnn); 2232 } 2233 2234 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop; 2235 2236 // Step 2: Fix the edges in the new body. If the old input is outside the 2237 // loop use it. If the old input is INside the loop, use the corresponding 2238 // new node instead. 2239 for( i = 0; i < loop->_body.size(); i++ ) { 2240 Node *old = loop->_body.at(i); 2241 Node *nnn = old_new[old->_idx]; 2242 // Fix CFG/Loop controlling the new node 2243 if (has_ctrl(old)) { 2244 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]); 2245 } else { 2246 set_loop(nnn, outer_loop->_parent); 2247 if (old->outcnt() > 0) { 2248 set_idom( nnn, old_new[idom(old)->_idx], dd ); 2249 } 2250 } 2251 // Correct edges to the new node 2252 for( uint j = 0; j < nnn->req(); j++ ) { 2253 Node *n = nnn->in(j); 2254 if( n ) { 2255 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n ); 2256 if( loop->is_member( old_in_loop ) ) 2257 nnn->set_req(j, old_new[n->_idx]); 2258 } 2259 } 2260 _igvn.hash_find_insert(nnn); 2261 } 2262 2263 ResourceArea *area = Thread::current()->resource_area(); 2264 Node_List extra_data_nodes(area); 2265 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes); 2266 2267 // Step 3: Now fix control uses. Loop varying control uses have already 2268 // been fixed up (as part of all input edges in Step 2). Loop invariant 2269 // control uses must be either an IfFalse or an IfTrue. Make a merge 2270 // point to merge the old and new IfFalse/IfTrue nodes; make the use 2271 // refer to this. 2272 Node_List worklist(area); 2273 uint new_counter = C->unique(); 2274 for( i = 0; i < loop->_body.size(); i++ ) { 2275 Node* old = loop->_body.at(i); 2276 if( !old->is_CFG() ) continue; 2277 2278 // Copy uses to a worklist, so I can munge the def-use info 2279 // with impunity. 2280 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 2281 worklist.push(old->fast_out(j)); 2282 2283 while( worklist.size() ) { // Visit all uses 2284 Node *use = worklist.pop(); 2285 if (!has_node(use)) continue; // Ignore dead nodes 2286 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 2287 if( !loop->is_member( use_loop ) && use->is_CFG() ) { 2288 // Both OLD and USE are CFG nodes here. 2289 assert( use->is_Proj(), "" ); 2290 Node* nnn = old_new[old->_idx]; 2291 2292 Node* newuse = NULL; 2293 if (head->is_strip_mined() && mode != IgnoreStripMined) { 2294 CountedLoopNode* cl = head->as_CountedLoop(); 2295 CountedLoopEndNode* cle = cl->loopexit(); 2296 Node* cle_out = cle->proj_out_or_null(false); 2297 if (use == cle_out) { 2298 IfNode* le = cl->outer_loop_end(); 2299 use = le->proj_out(false); 2300 use_loop = get_loop(use); 2301 if (mode == CloneIncludesStripMined) { 2302 nnn = old_new[le->_idx]; 2303 } else { 2304 newuse = old_new[cle_out->_idx]; 2305 } 2306 } 2307 } 2308 if (newuse == NULL) { 2309 newuse = use->clone(); 2310 } 2311 2312 // Clone the loop exit control projection 2313 if (C->do_vector_loop()) { 2314 cm.verify_insert_and_clone(use, newuse, cm.clone_idx()); 2315 } 2316 newuse->set_req(0,nnn); 2317 _igvn.register_new_node_with_optimizer(newuse); 2318 set_loop(newuse, use_loop); 2319 set_idom(newuse, nnn, dom_depth(nnn) + 1 ); 2320 2321 // We need a Region to merge the exit from the peeled body and the 2322 // exit from the old loop body. 2323 RegionNode *r = new RegionNode(3); 2324 // Map the old use to the new merge point 2325 old_new.map( use->_idx, r ); 2326 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use)); 2327 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" ); 2328 2329 // The original user of 'use' uses 'r' instead. 2330 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) { 2331 Node* useuse = use->last_out(l); 2332 _igvn.rehash_node_delayed(useuse); 2333 uint uses_found = 0; 2334 if( useuse->in(0) == use ) { 2335 useuse->set_req(0, r); 2336 uses_found++; 2337 if( useuse->is_CFG() ) { 2338 assert( dom_depth(useuse) > dd_r, "" ); 2339 set_idom(useuse, r, dom_depth(useuse)); 2340 } 2341 } 2342 for( uint k = 1; k < useuse->req(); k++ ) { 2343 if( useuse->in(k) == use ) { 2344 useuse->set_req(k, r); 2345 uses_found++; 2346 if (useuse->is_Loop() && k == LoopNode::EntryControl) { 2347 assert(dom_depth(useuse) > dd_r , ""); 2348 set_idom(useuse, r, dom_depth(useuse)); 2349 } 2350 } 2351 } 2352 l -= uses_found; // we deleted 1 or more copies of this edge 2353 } 2354 2355 // Now finish up 'r' 2356 r->set_req( 1, newuse ); 2357 r->set_req( 2, use ); 2358 _igvn.register_new_node_with_optimizer(r); 2359 set_loop(r, use_loop); 2360 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r); 2361 } // End of if a loop-exit test 2362 } 2363 } 2364 2365 // Step 4: If loop-invariant use is not control, it must be dominated by a 2366 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region 2367 // there if needed. Make a Phi there merging old and new used values. 2368 Node_List *split_if_set = NULL; 2369 Node_List *split_bool_set = NULL; 2370 Node_List *split_cex_set = NULL; 2371 for( i = 0; i < loop->_body.size(); i++ ) { 2372 Node* old = loop->_body.at(i); 2373 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set, 2374 split_bool_set, split_cex_set, worklist, new_counter, 2375 mode); 2376 } 2377 2378 for (i = 0; i < extra_data_nodes.size(); i++) { 2379 Node* old = extra_data_nodes.at(i); 2380 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set, 2381 split_bool_set, split_cex_set, worklist, new_counter, 2382 mode); 2383 } 2384 2385 // Check for IFs that need splitting/cloning. Happens if an IF outside of 2386 // the loop uses a condition set in the loop. The original IF probably 2387 // takes control from one or more OLD Regions (which in turn get from NEW 2388 // Regions). In any case, there will be a set of Phis for each merge point 2389 // from the IF up to where the original BOOL def exists the loop. 2390 if (split_if_set) { 2391 while (split_if_set->size()) { 2392 Node *iff = split_if_set->pop(); 2393 if (iff->in(1)->is_Phi()) { 2394 Node *b = clone_iff(iff->in(1)->as_Phi(), loop); 2395 _igvn.replace_input_of(iff, 1, b); 2396 } 2397 } 2398 } 2399 if (split_bool_set) { 2400 while (split_bool_set->size()) { 2401 Node *b = split_bool_set->pop(); 2402 Node *phi = b->in(1); 2403 assert(phi->is_Phi(), ""); 2404 CmpNode *cmp = clone_bool((PhiNode*)phi, loop); 2405 _igvn.replace_input_of(b, 1, cmp); 2406 } 2407 } 2408 if (split_cex_set) { 2409 while (split_cex_set->size()) { 2410 Node *b = split_cex_set->pop(); 2411 assert(b->in(0)->is_Region(), ""); 2412 assert(b->in(1)->is_Phi(), ""); 2413 assert(b->in(0)->in(0) == b->in(1)->in(0), ""); 2414 split_up(b, b->in(0), NULL); 2415 } 2416 } 2417 2418 } 2419 2420 2421 //---------------------- stride_of_possible_iv ------------------------------------- 2422 // Looks for an iff/bool/comp with one operand of the compare 2423 // being a cycle involving an add and a phi, 2424 // with an optional truncation (left-shift followed by a right-shift) 2425 // of the add. Returns zero if not an iv. 2426 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) { 2427 Node* trunc1 = NULL; 2428 Node* trunc2 = NULL; 2429 const TypeInt* ttype = NULL; 2430 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) { 2431 return 0; 2432 } 2433 BoolNode* bl = iff->in(1)->as_Bool(); 2434 Node* cmp = bl->in(1); 2435 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) { 2436 return 0; 2437 } 2438 // Must have an invariant operand 2439 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) { 2440 return 0; 2441 } 2442 Node* add2 = NULL; 2443 Node* cmp1 = cmp->in(1); 2444 if (cmp1->is_Phi()) { 2445 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) ))) 2446 Node* phi = cmp1; 2447 for (uint i = 1; i < phi->req(); i++) { 2448 Node* in = phi->in(i); 2449 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in, 2450 &trunc1, &trunc2, &ttype); 2451 if (add && add->in(1) == phi) { 2452 add2 = add->in(2); 2453 break; 2454 } 2455 } 2456 } else { 2457 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) ))) 2458 Node* addtrunc = cmp1; 2459 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc, 2460 &trunc1, &trunc2, &ttype); 2461 if (add && add->in(1)->is_Phi()) { 2462 Node* phi = add->in(1); 2463 for (uint i = 1; i < phi->req(); i++) { 2464 if (phi->in(i) == addtrunc) { 2465 add2 = add->in(2); 2466 break; 2467 } 2468 } 2469 } 2470 } 2471 if (add2 != NULL) { 2472 const TypeInt* add2t = _igvn.type(add2)->is_int(); 2473 if (add2t->is_con()) { 2474 return add2t->get_con(); 2475 } 2476 } 2477 return 0; 2478 } 2479 2480 2481 //---------------------- stay_in_loop ------------------------------------- 2482 // Return the (unique) control output node that's in the loop (if it exists.) 2483 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) { 2484 Node* unique = NULL; 2485 if (!n) return NULL; 2486 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2487 Node* use = n->fast_out(i); 2488 if (!has_ctrl(use) && loop->is_member(get_loop(use))) { 2489 if (unique != NULL) { 2490 return NULL; 2491 } 2492 unique = use; 2493 } 2494 } 2495 return unique; 2496 } 2497 2498 //------------------------------ register_node ------------------------------------- 2499 // Utility to register node "n" with PhaseIdealLoop 2500 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) { 2501 _igvn.register_new_node_with_optimizer(n); 2502 loop->_body.push(n); 2503 if (n->is_CFG()) { 2504 set_loop(n, loop); 2505 set_idom(n, pred, ddepth); 2506 } else { 2507 set_ctrl(n, pred); 2508 } 2509 } 2510 2511 //------------------------------ proj_clone ------------------------------------- 2512 // Utility to create an if-projection 2513 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) { 2514 ProjNode* c = p->clone()->as_Proj(); 2515 c->set_req(0, iff); 2516 return c; 2517 } 2518 2519 //------------------------------ short_circuit_if ------------------------------------- 2520 // Force the iff control output to be the live_proj 2521 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) { 2522 guarantee(live_proj != NULL, "null projection"); 2523 int proj_con = live_proj->_con; 2524 assert(proj_con == 0 || proj_con == 1, "false or true projection"); 2525 Node *con = _igvn.intcon(proj_con); 2526 set_ctrl(con, C->root()); 2527 if (iff) { 2528 iff->set_req(1, con); 2529 } 2530 return con; 2531 } 2532 2533 //------------------------------ insert_if_before_proj ------------------------------------- 2534 // Insert a new if before an if projection (* - new node) 2535 // 2536 // before 2537 // if(test) 2538 // / \ 2539 // v v 2540 // other-proj proj (arg) 2541 // 2542 // after 2543 // if(test) 2544 // / \ 2545 // / v 2546 // | * proj-clone 2547 // v | 2548 // other-proj v 2549 // * new_if(relop(cmp[IU](left,right))) 2550 // / \ 2551 // v v 2552 // * new-proj proj 2553 // (returned) 2554 // 2555 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) { 2556 IfNode* iff = proj->in(0)->as_If(); 2557 IdealLoopTree *loop = get_loop(proj); 2558 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2559 int ddepth = dom_depth(proj); 2560 2561 _igvn.rehash_node_delayed(iff); 2562 _igvn.rehash_node_delayed(proj); 2563 2564 proj->set_req(0, NULL); // temporary disconnect 2565 ProjNode* proj2 = proj_clone(proj, iff); 2566 register_node(proj2, loop, iff, ddepth); 2567 2568 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right); 2569 register_node(cmp, loop, proj2, ddepth); 2570 2571 BoolNode* bol = new BoolNode(cmp, relop); 2572 register_node(bol, loop, proj2, ddepth); 2573 2574 int opcode = iff->Opcode(); 2575 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode"); 2576 IfNode* new_if = (opcode == Op_If) ? new IfNode(proj2, bol, iff->_prob, iff->_fcnt): 2577 new RangeCheckNode(proj2, bol, iff->_prob, iff->_fcnt); 2578 register_node(new_if, loop, proj2, ddepth); 2579 2580 proj->set_req(0, new_if); // reattach 2581 set_idom(proj, new_if, ddepth); 2582 2583 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj(); 2584 guarantee(new_exit != NULL, "null exit node"); 2585 register_node(new_exit, get_loop(other_proj), new_if, ddepth); 2586 2587 return new_exit; 2588 } 2589 2590 //------------------------------ insert_region_before_proj ------------------------------------- 2591 // Insert a region before an if projection (* - new node) 2592 // 2593 // before 2594 // if(test) 2595 // / | 2596 // v | 2597 // proj v 2598 // other-proj 2599 // 2600 // after 2601 // if(test) 2602 // / | 2603 // v | 2604 // * proj-clone v 2605 // | other-proj 2606 // v 2607 // * new-region 2608 // | 2609 // v 2610 // * dum_if 2611 // / \ 2612 // v \ 2613 // * dum-proj v 2614 // proj 2615 // 2616 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) { 2617 IfNode* iff = proj->in(0)->as_If(); 2618 IdealLoopTree *loop = get_loop(proj); 2619 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2620 int ddepth = dom_depth(proj); 2621 2622 _igvn.rehash_node_delayed(iff); 2623 _igvn.rehash_node_delayed(proj); 2624 2625 proj->set_req(0, NULL); // temporary disconnect 2626 ProjNode* proj2 = proj_clone(proj, iff); 2627 register_node(proj2, loop, iff, ddepth); 2628 2629 RegionNode* reg = new RegionNode(2); 2630 reg->set_req(1, proj2); 2631 register_node(reg, loop, iff, ddepth); 2632 2633 IfNode* dum_if = new IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt); 2634 register_node(dum_if, loop, reg, ddepth); 2635 2636 proj->set_req(0, dum_if); // reattach 2637 set_idom(proj, dum_if, ddepth); 2638 2639 ProjNode* dum_proj = proj_clone(other_proj, dum_if); 2640 register_node(dum_proj, loop, dum_if, ddepth); 2641 2642 return reg; 2643 } 2644 2645 //------------------------------ insert_cmpi_loop_exit ------------------------------------- 2646 // Clone a signed compare loop exit from an unsigned compare and 2647 // insert it before the unsigned cmp on the stay-in-loop path. 2648 // All new nodes inserted in the dominator tree between the original 2649 // if and it's projections. The original if test is replaced with 2650 // a constant to force the stay-in-loop path. 2651 // 2652 // This is done to make sure that the original if and it's projections 2653 // still dominate the same set of control nodes, that the ctrl() relation 2654 // from data nodes to them is preserved, and that their loop nesting is 2655 // preserved. 2656 // 2657 // before 2658 // if(i <u limit) unsigned compare loop exit 2659 // / | 2660 // v v 2661 // exit-proj stay-in-loop-proj 2662 // 2663 // after 2664 // if(stay-in-loop-const) original if 2665 // / | 2666 // / v 2667 // / if(i < limit) new signed test 2668 // / / | 2669 // / / v 2670 // / / if(i <u limit) new cloned unsigned test 2671 // / / / | 2672 // v v v | 2673 // region | 2674 // | | 2675 // dum-if | 2676 // / | | 2677 // ether | | 2678 // v v 2679 // exit-proj stay-in-loop-proj 2680 // 2681 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) { 2682 const bool Signed = true; 2683 const bool Unsigned = false; 2684 2685 BoolNode* bol = if_cmpu->in(1)->as_Bool(); 2686 if (bol->_test._test != BoolTest::lt) return NULL; 2687 CmpNode* cmpu = bol->in(1)->as_Cmp(); 2688 if (cmpu->Opcode() != Op_CmpU) return NULL; 2689 int stride = stride_of_possible_iv(if_cmpu); 2690 if (stride == 0) return NULL; 2691 2692 Node* lp_proj = stay_in_loop(if_cmpu, loop); 2693 guarantee(lp_proj != NULL, "null loop node"); 2694 2695 ProjNode* lp_continue = lp_proj->as_Proj(); 2696 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj(); 2697 2698 Node* limit = NULL; 2699 if (stride > 0) { 2700 limit = cmpu->in(2); 2701 } else { 2702 limit = _igvn.makecon(TypeInt::ZERO); 2703 set_ctrl(limit, C->root()); 2704 } 2705 // Create a new region on the exit path 2706 RegionNode* reg = insert_region_before_proj(lp_exit); 2707 guarantee(reg != NULL, "null region node"); 2708 2709 // Clone the if-cmpu-true-false using a signed compare 2710 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge; 2711 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue); 2712 reg->add_req(cmpi_exit); 2713 2714 // Clone the if-cmpu-true-false 2715 BoolTest::mask rel_u = bol->_test._test; 2716 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue); 2717 reg->add_req(cmpu_exit); 2718 2719 // Force original if to stay in loop. 2720 short_circuit_if(if_cmpu, lp_continue); 2721 2722 return cmpi_exit->in(0)->as_If(); 2723 } 2724 2725 //------------------------------ remove_cmpi_loop_exit ------------------------------------- 2726 // Remove a previously inserted signed compare loop exit. 2727 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) { 2728 Node* lp_proj = stay_in_loop(if_cmp, loop); 2729 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI && 2730 stay_in_loop(lp_proj, loop)->is_If() && 2731 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu"); 2732 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO); 2733 set_ctrl(con, C->root()); 2734 if_cmp->set_req(1, con); 2735 } 2736 2737 //------------------------------ scheduled_nodelist ------------------------------------- 2738 // Create a post order schedule of nodes that are in the 2739 // "member" set. The list is returned in "sched". 2740 // The first node in "sched" is the loop head, followed by 2741 // nodes which have no inputs in the "member" set, and then 2742 // followed by the nodes that have an immediate input dependence 2743 // on a node in "sched". 2744 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) { 2745 2746 assert(member.test(loop->_head->_idx), "loop head must be in member set"); 2747 Arena *a = Thread::current()->resource_area(); 2748 VectorSet visited(a); 2749 Node_Stack nstack(a, loop->_body.size()); 2750 2751 Node* n = loop->_head; // top of stack is cached in "n" 2752 uint idx = 0; 2753 visited.set(n->_idx); 2754 2755 // Initially push all with no inputs from within member set 2756 for(uint i = 0; i < loop->_body.size(); i++ ) { 2757 Node *elt = loop->_body.at(i); 2758 if (member.test(elt->_idx)) { 2759 bool found = false; 2760 for (uint j = 0; j < elt->req(); j++) { 2761 Node* def = elt->in(j); 2762 if (def && member.test(def->_idx) && def != elt) { 2763 found = true; 2764 break; 2765 } 2766 } 2767 if (!found && elt != loop->_head) { 2768 nstack.push(n, idx); 2769 n = elt; 2770 assert(!visited.test(n->_idx), "not seen yet"); 2771 visited.set(n->_idx); 2772 } 2773 } 2774 } 2775 2776 // traverse out's that are in the member set 2777 while (true) { 2778 if (idx < n->outcnt()) { 2779 Node* use = n->raw_out(idx); 2780 idx++; 2781 if (!visited.test_set(use->_idx)) { 2782 if (member.test(use->_idx)) { 2783 nstack.push(n, idx); 2784 n = use; 2785 idx = 0; 2786 } 2787 } 2788 } else { 2789 // All outputs processed 2790 sched.push(n); 2791 if (nstack.is_empty()) break; 2792 n = nstack.node(); 2793 idx = nstack.index(); 2794 nstack.pop(); 2795 } 2796 } 2797 } 2798 2799 2800 //------------------------------ has_use_in_set ------------------------------------- 2801 // Has a use in the vector set 2802 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) { 2803 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2804 Node* use = n->fast_out(j); 2805 if (vset.test(use->_idx)) { 2806 return true; 2807 } 2808 } 2809 return false; 2810 } 2811 2812 2813 //------------------------------ has_use_internal_to_set ------------------------------------- 2814 // Has use internal to the vector set (ie. not in a phi at the loop head) 2815 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) { 2816 Node* head = loop->_head; 2817 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2818 Node* use = n->fast_out(j); 2819 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) { 2820 return true; 2821 } 2822 } 2823 return false; 2824 } 2825 2826 2827 //------------------------------ clone_for_use_outside_loop ------------------------------------- 2828 // clone "n" for uses that are outside of loop 2829 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) { 2830 int cloned = 0; 2831 assert(worklist.size() == 0, "should be empty"); 2832 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2833 Node* use = n->fast_out(j); 2834 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) { 2835 worklist.push(use); 2836 } 2837 } 2838 while( worklist.size() ) { 2839 Node *use = worklist.pop(); 2840 if (!has_node(use) || use->in(0) == C->top()) continue; 2841 uint j; 2842 for (j = 0; j < use->req(); j++) { 2843 if (use->in(j) == n) break; 2844 } 2845 assert(j < use->req(), "must be there"); 2846 2847 // clone "n" and insert it between the inputs of "n" and the use outside the loop 2848 Node* n_clone = n->clone(); 2849 _igvn.replace_input_of(use, j, n_clone); 2850 cloned++; 2851 Node* use_c; 2852 if (!use->is_Phi()) { 2853 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0); 2854 } else { 2855 // Use in a phi is considered a use in the associated predecessor block 2856 use_c = use->in(0)->in(j); 2857 } 2858 set_ctrl(n_clone, use_c); 2859 assert(!loop->is_member(get_loop(use_c)), "should be outside loop"); 2860 get_loop(use_c)->_body.push(n_clone); 2861 _igvn.register_new_node_with_optimizer(n_clone); 2862 #if !defined(PRODUCT) 2863 if (TracePartialPeeling) { 2864 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx); 2865 } 2866 #endif 2867 } 2868 return cloned; 2869 } 2870 2871 2872 //------------------------------ clone_for_special_use_inside_loop ------------------------------------- 2873 // clone "n" for special uses that are in the not_peeled region. 2874 // If these def-uses occur in separate blocks, the code generator 2875 // marks the method as not compilable. For example, if a "BoolNode" 2876 // is in a different basic block than the "IfNode" that uses it, then 2877 // the compilation is aborted in the code generator. 2878 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 2879 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) { 2880 if (n->is_Phi() || n->is_Load()) { 2881 return; 2882 } 2883 assert(worklist.size() == 0, "should be empty"); 2884 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2885 Node* use = n->fast_out(j); 2886 if ( not_peel.test(use->_idx) && 2887 (use->is_If() || use->is_CMove() || use->is_Bool()) && 2888 use->in(1) == n) { 2889 worklist.push(use); 2890 } 2891 } 2892 if (worklist.size() > 0) { 2893 // clone "n" and insert it between inputs of "n" and the use 2894 Node* n_clone = n->clone(); 2895 loop->_body.push(n_clone); 2896 _igvn.register_new_node_with_optimizer(n_clone); 2897 set_ctrl(n_clone, get_ctrl(n)); 2898 sink_list.push(n_clone); 2899 not_peel <<= n_clone->_idx; // add n_clone to not_peel set. 2900 #if !defined(PRODUCT) 2901 if (TracePartialPeeling) { 2902 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx); 2903 } 2904 #endif 2905 while( worklist.size() ) { 2906 Node *use = worklist.pop(); 2907 _igvn.rehash_node_delayed(use); 2908 for (uint j = 1; j < use->req(); j++) { 2909 if (use->in(j) == n) { 2910 use->set_req(j, n_clone); 2911 } 2912 } 2913 } 2914 } 2915 } 2916 2917 2918 //------------------------------ insert_phi_for_loop ------------------------------------- 2919 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 2920 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) { 2921 Node *phi = PhiNode::make(lp, back_edge_val); 2922 phi->set_req(LoopNode::EntryControl, lp_entry_val); 2923 // Use existing phi if it already exists 2924 Node *hit = _igvn.hash_find_insert(phi); 2925 if( hit == NULL ) { 2926 _igvn.register_new_node_with_optimizer(phi); 2927 set_ctrl(phi, lp); 2928 } else { 2929 // Remove the new phi from the graph and use the hit 2930 _igvn.remove_dead_node(phi); 2931 phi = hit; 2932 } 2933 _igvn.replace_input_of(use, idx, phi); 2934 } 2935 2936 #ifdef ASSERT 2937 //------------------------------ is_valid_loop_partition ------------------------------------- 2938 // Validate the loop partition sets: peel and not_peel 2939 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, 2940 VectorSet& not_peel ) { 2941 uint i; 2942 // Check that peel_list entries are in the peel set 2943 for (i = 0; i < peel_list.size(); i++) { 2944 if (!peel.test(peel_list.at(i)->_idx)) { 2945 return false; 2946 } 2947 } 2948 // Check at loop members are in one of peel set or not_peel set 2949 for (i = 0; i < loop->_body.size(); i++ ) { 2950 Node *def = loop->_body.at(i); 2951 uint di = def->_idx; 2952 // Check that peel set elements are in peel_list 2953 if (peel.test(di)) { 2954 if (not_peel.test(di)) { 2955 return false; 2956 } 2957 // Must be in peel_list also 2958 bool found = false; 2959 for (uint j = 0; j < peel_list.size(); j++) { 2960 if (peel_list.at(j)->_idx == di) { 2961 found = true; 2962 break; 2963 } 2964 } 2965 if (!found) { 2966 return false; 2967 } 2968 } else if (not_peel.test(di)) { 2969 if (peel.test(di)) { 2970 return false; 2971 } 2972 } else { 2973 return false; 2974 } 2975 } 2976 return true; 2977 } 2978 2979 //------------------------------ is_valid_clone_loop_exit_use ------------------------------------- 2980 // Ensure a use outside of loop is of the right form 2981 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) { 2982 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2983 return (use->is_Phi() && 2984 use_c->is_Region() && use_c->req() == 3 && 2985 (use_c->in(exit_idx)->Opcode() == Op_IfTrue || 2986 use_c->in(exit_idx)->Opcode() == Op_IfFalse || 2987 use_c->in(exit_idx)->Opcode() == Op_JumpProj) && 2988 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) ); 2989 } 2990 2991 //------------------------------ is_valid_clone_loop_form ------------------------------------- 2992 // Ensure that all uses outside of loop are of the right form 2993 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 2994 uint orig_exit_idx, uint clone_exit_idx) { 2995 uint len = peel_list.size(); 2996 for (uint i = 0; i < len; i++) { 2997 Node *def = peel_list.at(i); 2998 2999 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 3000 Node *use = def->fast_out(j); 3001 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 3002 if (!loop->is_member(get_loop(use_c))) { 3003 // use is not in the loop, check for correct structure 3004 if (use->in(0) == def) { 3005 // Okay 3006 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) { 3007 return false; 3008 } 3009 } 3010 } 3011 } 3012 return true; 3013 } 3014 #endif 3015 3016 //------------------------------ partial_peel ------------------------------------- 3017 // Partially peel (aka loop rotation) the top portion of a loop (called 3018 // the peel section below) by cloning it and placing one copy just before 3019 // the new loop head and the other copy at the bottom of the new loop. 3020 // 3021 // before after where it came from 3022 // 3023 // stmt1 stmt1 3024 // loop: stmt2 clone 3025 // stmt2 if condA goto exitA clone 3026 // if condA goto exitA new_loop: new 3027 // stmt3 stmt3 clone 3028 // if !condB goto loop if condB goto exitB clone 3029 // exitB: stmt2 orig 3030 // stmt4 if !condA goto new_loop orig 3031 // exitA: goto exitA 3032 // exitB: 3033 // stmt4 3034 // exitA: 3035 // 3036 // Step 1: find the cut point: an exit test on probable 3037 // induction variable. 3038 // Step 2: schedule (with cloning) operations in the peel 3039 // section that can be executed after the cut into 3040 // the section that is not peeled. This may need 3041 // to clone operations into exit blocks. For 3042 // instance, a reference to A[i] in the not-peel 3043 // section and a reference to B[i] in an exit block 3044 // may cause a left-shift of i by 2 to be placed 3045 // in the peel block. This step will clone the left 3046 // shift into the exit block and sink the left shift 3047 // from the peel to the not-peel section. 3048 // Step 3: clone the loop, retarget the control, and insert 3049 // phis for values that are live across the new loop 3050 // head. This is very dependent on the graph structure 3051 // from clone_loop. It creates region nodes for 3052 // exit control and associated phi nodes for values 3053 // flow out of the loop through that exit. The region 3054 // node is dominated by the clone's control projection. 3055 // So the clone's peel section is placed before the 3056 // new loop head, and the clone's not-peel section is 3057 // forms the top part of the new loop. The original 3058 // peel section forms the tail of the new loop. 3059 // Step 4: update the dominator tree and recompute the 3060 // dominator depth. 3061 // 3062 // orig 3063 // 3064 // stmt1 3065 // | 3066 // v 3067 // loop predicate 3068 // | 3069 // v 3070 // loop<----+ 3071 // | | 3072 // stmt2 | 3073 // | | 3074 // v | 3075 // ifA | 3076 // / | | 3077 // v v | 3078 // false true ^ <-- last_peel 3079 // / | | 3080 // / ===|==cut | 3081 // / stmt3 | <-- first_not_peel 3082 // / | | 3083 // | v | 3084 // v ifB | 3085 // exitA: / \ | 3086 // / \ | 3087 // v v | 3088 // false true | 3089 // / \ | 3090 // / ----+ 3091 // | 3092 // v 3093 // exitB: 3094 // stmt4 3095 // 3096 // 3097 // after clone loop 3098 // 3099 // stmt1 3100 // | 3101 // v 3102 // loop predicate 3103 // / \ 3104 // clone / \ orig 3105 // / \ 3106 // / \ 3107 // v v 3108 // +---->loop loop<----+ 3109 // | | | | 3110 // | stmt2 stmt2 | 3111 // | | | | 3112 // | v v | 3113 // | ifA ifA | 3114 // | | \ / | | 3115 // | v v v v | 3116 // ^ true false false true ^ <-- last_peel 3117 // | | ^ \ / | | 3118 // | cut==|== \ \ / ===|==cut | 3119 // | stmt3 \ \ / stmt3 | <-- first_not_peel 3120 // | | dom | | | | 3121 // | v \ 1v v2 v | 3122 // | ifB regionA ifB | 3123 // | / \ | / \ | 3124 // | / \ v / \ | 3125 // | v v exitA: v v | 3126 // | true false false true | 3127 // | / ^ \ / \ | 3128 // +---- \ \ / ----+ 3129 // dom \ / 3130 // \ 1v v2 3131 // regionB 3132 // | 3133 // v 3134 // exitB: 3135 // stmt4 3136 // 3137 // 3138 // after partial peel 3139 // 3140 // stmt1 3141 // | 3142 // v 3143 // loop predicate 3144 // / 3145 // clone / orig 3146 // / TOP 3147 // / \ 3148 // v v 3149 // TOP->loop loop----+ 3150 // | | | 3151 // stmt2 stmt2 | 3152 // | | | 3153 // v v | 3154 // ifA ifA | 3155 // | \ / | | 3156 // v v v v | 3157 // true false false true | <-- last_peel 3158 // | ^ \ / +------|---+ 3159 // +->newloop \ \ / === ==cut | | 3160 // | stmt3 \ \ / TOP | | 3161 // | | dom | | stmt3 | | <-- first_not_peel 3162 // | v \ 1v v2 v | | 3163 // | ifB regionA ifB ^ v 3164 // | / \ | / \ | | 3165 // | / \ v / \ | | 3166 // | v v exitA: v v | | 3167 // | true false false true | | 3168 // | / ^ \ / \ | | 3169 // | | \ \ / v | | 3170 // | | dom \ / TOP | | 3171 // | | \ 1v v2 | | 3172 // ^ v regionB | | 3173 // | | | | | 3174 // | | v ^ v 3175 // | | exitB: | | 3176 // | | stmt4 | | 3177 // | +------------>-----------------+ | 3178 // | | 3179 // +-----------------<---------------------+ 3180 // 3181 // 3182 // final graph 3183 // 3184 // stmt1 3185 // | 3186 // v 3187 // loop predicate 3188 // | 3189 // v 3190 // stmt2 clone 3191 // | 3192 // v 3193 // ........> ifA clone 3194 // : / | 3195 // dom / | 3196 // : v v 3197 // : false true 3198 // : | | 3199 // : | v 3200 // : | newloop<-----+ 3201 // : | | | 3202 // : | stmt3 clone | 3203 // : | | | 3204 // : | v | 3205 // : | ifB | 3206 // : | / \ | 3207 // : | v v | 3208 // : | false true | 3209 // : | | | | 3210 // : | v stmt2 | 3211 // : | exitB: | | 3212 // : | stmt4 v | 3213 // : | ifA orig | 3214 // : | / \ | 3215 // : | / \ | 3216 // : | v v | 3217 // : | false true | 3218 // : | / \ | 3219 // : v v -----+ 3220 // RegionA 3221 // | 3222 // v 3223 // exitA 3224 // 3225 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) { 3226 3227 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only"); 3228 if (!loop->_head->is_Loop()) { 3229 return false; } 3230 3231 LoopNode *head = loop->_head->as_Loop(); 3232 3233 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) { 3234 return false; 3235 } 3236 3237 // Check for complex exit control 3238 for(uint ii = 0; ii < loop->_body.size(); ii++ ) { 3239 Node *n = loop->_body.at(ii); 3240 int opc = n->Opcode(); 3241 if (n->is_Call() || 3242 opc == Op_Catch || 3243 opc == Op_CatchProj || 3244 opc == Op_Jump || 3245 opc == Op_JumpProj) { 3246 #if !defined(PRODUCT) 3247 if (TracePartialPeeling) { 3248 tty->print_cr("\nExit control too complex: lp: %d", head->_idx); 3249 } 3250 #endif 3251 return false; 3252 } 3253 } 3254 3255 int dd = dom_depth(head); 3256 3257 // Step 1: find cut point 3258 3259 // Walk up dominators to loop head looking for first loop exit 3260 // which is executed on every path thru loop. 3261 IfNode *peel_if = NULL; 3262 IfNode *peel_if_cmpu = NULL; 3263 3264 Node *iff = loop->tail(); 3265 while( iff != head ) { 3266 if( iff->is_If() ) { 3267 Node *ctrl = get_ctrl(iff->in(1)); 3268 if (ctrl->is_top()) return false; // Dead test on live IF. 3269 // If loop-varying exit-test, check for induction variable 3270 if( loop->is_member(get_loop(ctrl)) && 3271 loop->is_loop_exit(iff) && 3272 is_possible_iv_test(iff)) { 3273 Node* cmp = iff->in(1)->in(1); 3274 if (cmp->Opcode() == Op_CmpI) { 3275 peel_if = iff->as_If(); 3276 } else { 3277 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU"); 3278 peel_if_cmpu = iff->as_If(); 3279 } 3280 } 3281 } 3282 iff = idom(iff); 3283 } 3284 // Prefer signed compare over unsigned compare. 3285 IfNode* new_peel_if = NULL; 3286 if (peel_if == NULL) { 3287 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) { 3288 return false; // No peel point found 3289 } 3290 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop); 3291 if (new_peel_if == NULL) { 3292 return false; // No peel point found 3293 } 3294 peel_if = new_peel_if; 3295 } 3296 Node* last_peel = stay_in_loop(peel_if, loop); 3297 Node* first_not_peeled = stay_in_loop(last_peel, loop); 3298 if (first_not_peeled == NULL || first_not_peeled == head) { 3299 return false; 3300 } 3301 3302 #if !defined(PRODUCT) 3303 if (TraceLoopOpts) { 3304 tty->print("PartialPeel "); 3305 loop->dump_head(); 3306 } 3307 3308 if (TracePartialPeeling) { 3309 tty->print_cr("before partial peel one iteration"); 3310 Node_List wl; 3311 Node* t = head->in(2); 3312 while (true) { 3313 wl.push(t); 3314 if (t == head) break; 3315 t = idom(t); 3316 } 3317 while (wl.size() > 0) { 3318 Node* tt = wl.pop(); 3319 tt->dump(); 3320 if (tt == last_peel) tty->print_cr("-- cut --"); 3321 } 3322 } 3323 #endif 3324 ResourceArea *area = Thread::current()->resource_area(); 3325 VectorSet peel(area); 3326 VectorSet not_peel(area); 3327 Node_List peel_list(area); 3328 Node_List worklist(area); 3329 Node_List sink_list(area); 3330 3331 // Set of cfg nodes to peel are those that are executable from 3332 // the head through last_peel. 3333 assert(worklist.size() == 0, "should be empty"); 3334 worklist.push(head); 3335 peel.set(head->_idx); 3336 while (worklist.size() > 0) { 3337 Node *n = worklist.pop(); 3338 if (n != last_peel) { 3339 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3340 Node* use = n->fast_out(j); 3341 if (use->is_CFG() && 3342 loop->is_member(get_loop(use)) && 3343 !peel.test_set(use->_idx)) { 3344 worklist.push(use); 3345 } 3346 } 3347 } 3348 } 3349 3350 // Set of non-cfg nodes to peel are those that are control 3351 // dependent on the cfg nodes. 3352 uint i; 3353 for(i = 0; i < loop->_body.size(); i++ ) { 3354 Node *n = loop->_body.at(i); 3355 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n; 3356 if (peel.test(n_c->_idx)) { 3357 peel.set(n->_idx); 3358 } else { 3359 not_peel.set(n->_idx); 3360 } 3361 } 3362 3363 // Step 2: move operations from the peeled section down into the 3364 // not-peeled section 3365 3366 // Get a post order schedule of nodes in the peel region 3367 // Result in right-most operand. 3368 scheduled_nodelist(loop, peel, peel_list ); 3369 3370 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 3371 3372 // For future check for too many new phis 3373 uint old_phi_cnt = 0; 3374 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 3375 Node* use = head->fast_out(j); 3376 if (use->is_Phi()) old_phi_cnt++; 3377 } 3378 3379 #if !defined(PRODUCT) 3380 if (TracePartialPeeling) { 3381 tty->print_cr("\npeeled list"); 3382 } 3383 #endif 3384 3385 // Evacuate nodes in peel region into the not_peeled region if possible 3386 uint new_phi_cnt = 0; 3387 uint cloned_for_outside_use = 0; 3388 for (i = 0; i < peel_list.size();) { 3389 Node* n = peel_list.at(i); 3390 #if !defined(PRODUCT) 3391 if (TracePartialPeeling) n->dump(); 3392 #endif 3393 bool incr = true; 3394 if ( !n->is_CFG() ) { 3395 3396 if ( has_use_in_set(n, not_peel) ) { 3397 3398 // If not used internal to the peeled region, 3399 // move "n" from peeled to not_peeled region. 3400 3401 if ( !has_use_internal_to_set(n, peel, loop) ) { 3402 3403 // if not pinned and not a load (which maybe anti-dependent on a store) 3404 // and not a CMove (Matcher expects only bool->cmove). 3405 if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) { 3406 cloned_for_outside_use += clone_for_use_outside_loop( loop, n, worklist ); 3407 sink_list.push(n); 3408 peel >>= n->_idx; // delete n from peel set. 3409 not_peel <<= n->_idx; // add n to not_peel set. 3410 peel_list.remove(i); 3411 incr = false; 3412 #if !defined(PRODUCT) 3413 if (TracePartialPeeling) { 3414 tty->print_cr("sink to not_peeled region: %d newbb: %d", 3415 n->_idx, get_ctrl(n)->_idx); 3416 } 3417 #endif 3418 } 3419 } else { 3420 // Otherwise check for special def-use cases that span 3421 // the peel/not_peel boundary such as bool->if 3422 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist ); 3423 new_phi_cnt++; 3424 } 3425 } 3426 } 3427 if (incr) i++; 3428 } 3429 3430 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) { 3431 #if !defined(PRODUCT) 3432 if (TracePartialPeeling) { 3433 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c", 3434 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F'); 3435 } 3436 #endif 3437 if (new_peel_if != NULL) { 3438 remove_cmpi_loop_exit(new_peel_if, loop); 3439 } 3440 // Inhibit more partial peeling on this loop 3441 assert(!head->is_partial_peel_loop(), "not partial peeled"); 3442 head->mark_partial_peel_failed(); 3443 if (cloned_for_outside_use > 0) { 3444 // Terminate this round of loop opts because 3445 // the graph outside this loop was changed. 3446 C->set_major_progress(); 3447 return true; 3448 } 3449 return false; 3450 } 3451 3452 // Step 3: clone loop, retarget control, and insert new phis 3453 3454 // Create new loop head for new phis and to hang 3455 // the nodes being moved (sinked) from the peel region. 3456 LoopNode* new_head = new LoopNode(last_peel, last_peel); 3457 new_head->set_unswitch_count(head->unswitch_count()); // Preserve 3458 _igvn.register_new_node_with_optimizer(new_head); 3459 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled"); 3460 _igvn.replace_input_of(first_not_peeled, 0, new_head); 3461 set_loop(new_head, loop); 3462 loop->_body.push(new_head); 3463 not_peel.set(new_head->_idx); 3464 set_idom(new_head, last_peel, dom_depth(first_not_peeled)); 3465 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled)); 3466 3467 while (sink_list.size() > 0) { 3468 Node* n = sink_list.pop(); 3469 set_ctrl(n, new_head); 3470 } 3471 3472 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 3473 3474 clone_loop(loop, old_new, dd, IgnoreStripMined); 3475 3476 const uint clone_exit_idx = 1; 3477 const uint orig_exit_idx = 2; 3478 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop"); 3479 3480 Node* head_clone = old_new[head->_idx]; 3481 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop(); 3482 Node* orig_tail_clone = head_clone->in(2); 3483 3484 // Add phi if "def" node is in peel set and "use" is not 3485 3486 for(i = 0; i < peel_list.size(); i++ ) { 3487 Node *def = peel_list.at(i); 3488 if (!def->is_CFG()) { 3489 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 3490 Node *use = def->fast_out(j); 3491 if (has_node(use) && use->in(0) != C->top() && 3492 (!peel.test(use->_idx) || 3493 (use->is_Phi() && use->in(0) == head)) ) { 3494 worklist.push(use); 3495 } 3496 } 3497 while( worklist.size() ) { 3498 Node *use = worklist.pop(); 3499 for (uint j = 1; j < use->req(); j++) { 3500 Node* n = use->in(j); 3501 if (n == def) { 3502 3503 // "def" is in peel set, "use" is not in peel set 3504 // or "use" is in the entry boundary (a phi) of the peel set 3505 3506 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use; 3507 3508 if ( loop->is_member(get_loop( use_c )) ) { 3509 // use is in loop 3510 if (old_new[use->_idx] != NULL) { // null for dead code 3511 Node* use_clone = old_new[use->_idx]; 3512 _igvn.replace_input_of(use, j, C->top()); 3513 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone ); 3514 } 3515 } else { 3516 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format"); 3517 // use is not in the loop, check if the live range includes the cut 3518 Node* lp_if = use_c->in(orig_exit_idx)->in(0); 3519 if (not_peel.test(lp_if->_idx)) { 3520 assert(j == orig_exit_idx, "use from original loop"); 3521 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone ); 3522 } 3523 } 3524 } 3525 } 3526 } 3527 } 3528 } 3529 3530 // Step 3b: retarget control 3531 3532 // Redirect control to the new loop head if a cloned node in 3533 // the not_peeled region has control that points into the peeled region. 3534 // This necessary because the cloned peeled region will be outside 3535 // the loop. 3536 // from to 3537 // cloned-peeled <---+ 3538 // new_head_clone: | <--+ 3539 // cloned-not_peeled in(0) in(0) 3540 // orig-peeled 3541 3542 for(i = 0; i < loop->_body.size(); i++ ) { 3543 Node *n = loop->_body.at(i); 3544 if (!n->is_CFG() && n->in(0) != NULL && 3545 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) { 3546 Node* n_clone = old_new[n->_idx]; 3547 _igvn.replace_input_of(n_clone, 0, new_head_clone); 3548 } 3549 } 3550 3551 // Backedge of the surviving new_head (the clone) is original last_peel 3552 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel); 3553 3554 // Cut first node in original not_peel set 3555 _igvn.rehash_node_delayed(new_head); // Multiple edge updates: 3556 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of 3557 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls 3558 3559 // Copy head_clone back-branch info to original head 3560 // and remove original head's loop entry and 3561 // clone head's back-branch 3562 _igvn.rehash_node_delayed(head); // Multiple edge updates 3563 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl)); 3564 head->set_req(LoopNode::LoopBackControl, C->top()); 3565 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top()); 3566 3567 // Similarly modify the phis 3568 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) { 3569 Node* use = head->fast_out(k); 3570 if (use->is_Phi() && use->outcnt() > 0) { 3571 Node* use_clone = old_new[use->_idx]; 3572 _igvn.rehash_node_delayed(use); // Multiple edge updates 3573 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl)); 3574 use->set_req(LoopNode::LoopBackControl, C->top()); 3575 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top()); 3576 } 3577 } 3578 3579 // Step 4: update dominator tree and dominator depth 3580 3581 set_idom(head, orig_tail_clone, dd); 3582 recompute_dom_depth(); 3583 3584 // Inhibit more partial peeling on this loop 3585 new_head_clone->set_partial_peel_loop(); 3586 C->set_major_progress(); 3587 loop->record_for_igvn(); 3588 3589 #if !defined(PRODUCT) 3590 if (TracePartialPeeling) { 3591 tty->print_cr("\nafter partial peel one iteration"); 3592 Node_List wl(area); 3593 Node* t = last_peel; 3594 while (true) { 3595 wl.push(t); 3596 if (t == head_clone) break; 3597 t = idom(t); 3598 } 3599 while (wl.size() > 0) { 3600 Node* tt = wl.pop(); 3601 if (tt == head) tty->print_cr("orig head"); 3602 else if (tt == new_head_clone) tty->print_cr("new head"); 3603 else if (tt == head_clone) tty->print_cr("clone head"); 3604 tt->dump(); 3605 } 3606 } 3607 #endif 3608 return true; 3609 } 3610 3611 //------------------------------reorg_offsets---------------------------------- 3612 // Reorganize offset computations to lower register pressure. Mostly 3613 // prevent loop-fallout uses of the pre-incremented trip counter (which are 3614 // then alive with the post-incremented trip counter forcing an extra 3615 // register move) 3616 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) { 3617 // Perform it only for canonical counted loops. 3618 // Loop's shape could be messed up by iteration_split_impl. 3619 if (!loop->_head->is_CountedLoop()) 3620 return; 3621 if (!loop->_head->as_Loop()->is_valid_counted_loop()) 3622 return; 3623 3624 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 3625 CountedLoopEndNode *cle = cl->loopexit(); 3626 Node *exit = cle->proj_out(false); 3627 Node *phi = cl->phi(); 3628 3629 // Check for the special case of folks using the pre-incremented 3630 // trip-counter on the fall-out path (forces the pre-incremented 3631 // and post-incremented trip counter to be live at the same time). 3632 // Fix this by adjusting to use the post-increment trip counter. 3633 3634 bool progress = true; 3635 while (progress) { 3636 progress = false; 3637 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) { 3638 Node* use = phi->fast_out(i); // User of trip-counter 3639 if (!has_ctrl(use)) continue; 3640 Node *u_ctrl = get_ctrl(use); 3641 if (use->is_Phi()) { 3642 u_ctrl = NULL; 3643 for (uint j = 1; j < use->req(); j++) 3644 if (use->in(j) == phi) 3645 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j)); 3646 } 3647 IdealLoopTree *u_loop = get_loop(u_ctrl); 3648 // Look for loop-invariant use 3649 if (u_loop == loop) continue; 3650 if (loop->is_member(u_loop)) continue; 3651 // Check that use is live out the bottom. Assuming the trip-counter 3652 // update is right at the bottom, uses of of the loop middle are ok. 3653 if (dom_lca(exit, u_ctrl) != exit) continue; 3654 // Hit! Refactor use to use the post-incremented tripcounter. 3655 // Compute a post-increment tripcounter. 3656 Node *opaq = new Opaque2Node( C, cle->incr() ); 3657 register_new_node(opaq, exit); 3658 Node *neg_stride = _igvn.intcon(-cle->stride_con()); 3659 set_ctrl(neg_stride, C->root()); 3660 Node *post = new AddINode( opaq, neg_stride); 3661 register_new_node(post, exit); 3662 _igvn.rehash_node_delayed(use); 3663 for (uint j = 1; j < use->req(); j++) { 3664 if (use->in(j) == phi) 3665 use->set_req(j, post); 3666 } 3667 // Since DU info changed, rerun loop 3668 progress = true; 3669 break; 3670 } 3671 } 3672 3673 }