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 return (u_loop->_irreducible || u_loop->_child) 1035 ? useblock 1036 : u_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl); 1037 } 1038 1039 1040 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) { 1041 if (!n->is_If() || n->is_CountedLoopEnd()) { 1042 return false; 1043 } 1044 if (!n->in(0)->is_Region()) { 1045 return false; 1046 } 1047 Node* region = n->in(0); 1048 Node* dom = idom(region); 1049 if (!dom->is_If() || dom->in(1) != n->in(1)) { 1050 return false; 1051 } 1052 IfNode* dom_if = dom->as_If(); 1053 Node* proj_true = dom_if->proj_out(1); 1054 Node* proj_false = dom_if->proj_out(0); 1055 1056 for (uint i = 1; i < region->req(); i++) { 1057 if (is_dominator(proj_true, region->in(i))) { 1058 continue; 1059 } 1060 if (is_dominator(proj_false, region->in(i))) { 1061 continue; 1062 } 1063 return false; 1064 } 1065 1066 return true; 1067 } 1068 1069 bool PhaseIdealLoop::can_split_if(Node *n_ctrl) { 1070 if (C->live_nodes() > 35000) { 1071 return false; // Method too big 1072 } 1073 1074 // Do not do 'split-if' if irreducible loops are present. 1075 if (_has_irreducible_loops) { 1076 return false; 1077 } 1078 1079 if (merge_point_too_heavy(C, n_ctrl)) { 1080 return false; 1081 } 1082 1083 // Do not do 'split-if' if some paths are dead. First do dead code 1084 // elimination and then see if its still profitable. 1085 for (uint i = 1; i < n_ctrl->req(); i++) { 1086 if (n_ctrl->in(i) == C->top()) { 1087 return false; 1088 } 1089 } 1090 1091 // If trying to do a 'Split-If' at the loop head, it is only 1092 // profitable if the cmp folds up on BOTH paths. Otherwise we 1093 // risk peeling a loop forever. 1094 1095 // CNC - Disabled for now. Requires careful handling of loop 1096 // body selection for the cloned code. Also, make sure we check 1097 // for any input path not being in the same loop as n_ctrl. For 1098 // irreducible loops we cannot check for 'n_ctrl->is_Loop()' 1099 // because the alternative loop entry points won't be converted 1100 // into LoopNodes. 1101 IdealLoopTree *n_loop = get_loop(n_ctrl); 1102 for (uint j = 1; j < n_ctrl->req(); j++) { 1103 if (get_loop(n_ctrl->in(j)) != n_loop) { 1104 return false; 1105 } 1106 } 1107 1108 // Check for safety of the merge point. 1109 if (!merge_point_safe(n_ctrl)) { 1110 return false; 1111 } 1112 1113 return true; 1114 } 1115 1116 bool PhaseIdealLoop::replace_with_dominating_barrier(LoadBarrierNode* lb, bool last_round) { 1117 LoadBarrierNode* lb2 = lb->has_dominating_barrier(this, false, last_round); 1118 if (lb2 != NULL) { 1119 if (lb->in(LoadBarrierNode::Oop) != lb2->in(LoadBarrierNode::Oop)) { 1120 assert(lb->in(LoadBarrierNode::Address) == lb2->in(LoadBarrierNode::Address), ""); 1121 _igvn.replace_input_of(lb, LoadBarrierNode::Similar, lb2->proj_out(LoadBarrierNode::Oop)); 1122 C->set_major_progress(); 1123 } else { 1124 // That transformation may cause the Similar edge on dominated load barriers to be invalid 1125 lb->fix_similar_in_uses(&_igvn); 1126 1127 Node* val = lb->proj_out(LoadBarrierNode::Oop); 1128 assert(lb2->has_true_uses(), ""); 1129 assert(lb2->in(LoadBarrierNode::Oop) == lb->in(LoadBarrierNode::Oop), ""); 1130 1131 lazy_update(lb, lb->in(LoadBarrierNode::Control)); 1132 lazy_replace(lb->proj_out(LoadBarrierNode::Control), lb->in(LoadBarrierNode::Control)); 1133 _igvn.replace_node(val, lb2->proj_out(LoadBarrierNode::Oop)); 1134 1135 return true; 1136 } 1137 } 1138 return false; 1139 } 1140 1141 Node* PhaseIdealLoop::find_dominating_memory(Node* mem, Node* dom, int i) { 1142 assert(dom->is_Region() || i == -1, ""); 1143 Node* m = mem; 1144 while(is_dominator(dom, has_ctrl(m) ? get_ctrl(m) : m->in(0))) { 1145 if (m->is_Mem()) { 1146 assert(m->as_Mem()->adr_type() == TypeRawPtr::BOTTOM, ""); 1147 m = m->in(MemNode::Memory); 1148 } else if (m->is_MergeMem()) { 1149 m = m->as_MergeMem()->memory_at(Compile::AliasIdxRaw); 1150 } else if (m->is_Phi()) { 1151 if (m->in(0) == dom && i != -1) { 1152 m = m->in(i); 1153 break; 1154 } else { 1155 m = m->in(LoopNode::EntryControl); 1156 } 1157 } else if (m->is_Proj()) { 1158 m = m->in(0); 1159 } else if (m->is_SafePoint() || m->is_MemBar()) { 1160 m = m->in(TypeFunc::Memory); 1161 } else { 1162 #ifdef ASSERT 1163 m->dump(); 1164 #endif 1165 ShouldNotReachHere(); 1166 } 1167 } 1168 return m; 1169 } 1170 1171 LoadBarrierNode* PhaseIdealLoop::clone_load_barrier(LoadBarrierNode* lb, Node* ctl, Node* mem, Node* oop_in) { 1172 Node* the_clone = lb->clone(); 1173 the_clone->set_req(LoadBarrierNode::Control, ctl); 1174 the_clone->set_req(LoadBarrierNode::Memory, mem); 1175 if (oop_in != NULL) { 1176 the_clone->set_req(LoadBarrierNode::Oop, oop_in); 1177 } 1178 1179 LoadBarrierNode* new_lb = the_clone->as_LoadBarrier(); 1180 _igvn.register_new_node_with_optimizer(new_lb); 1181 IdealLoopTree *loop = get_loop(new_lb->in(0)); 1182 set_ctrl(new_lb, new_lb->in(0)); 1183 set_loop(new_lb, loop); 1184 set_idom(new_lb, new_lb->in(0), dom_depth(new_lb->in(0))+1); 1185 if (!loop->_child) { 1186 loop->_body.push(new_lb); 1187 } 1188 1189 Node* proj_ctl = new ProjNode(new_lb, LoadBarrierNode::Control); 1190 _igvn.register_new_node_with_optimizer(proj_ctl); 1191 set_ctrl(proj_ctl, proj_ctl->in(0)); 1192 set_loop(proj_ctl, loop); 1193 set_idom(proj_ctl, new_lb, dom_depth(new_lb)+1); 1194 if (!loop->_child) { 1195 loop->_body.push(proj_ctl); 1196 } 1197 1198 Node* proj_oop = new ProjNode(new_lb, LoadBarrierNode::Oop); 1199 register_new_node(proj_oop, new_lb); 1200 1201 if (!new_lb->in(LoadBarrierNode::Similar)->is_top()) { 1202 LoadBarrierNode* similar = new_lb->in(LoadBarrierNode::Similar)->in(0)->as_LoadBarrier(); 1203 if (!is_dominator(similar, ctl)) { 1204 _igvn.replace_input_of(new_lb, LoadBarrierNode::Similar, C->top()); 1205 } 1206 } 1207 1208 return new_lb; 1209 } 1210 1211 void PhaseIdealLoop::replace_barrier(LoadBarrierNode* lb, Node* new_val) { 1212 Node* val = lb->proj_out(LoadBarrierNode::Oop); 1213 _igvn.replace_node(val, new_val); 1214 lazy_update(lb, lb->in(LoadBarrierNode::Control)); 1215 lazy_replace(lb->proj_out(LoadBarrierNode::Control), lb->in(LoadBarrierNode::Control)); 1216 } 1217 1218 bool PhaseIdealLoop::split_barrier_thru_phi(LoadBarrierNode* lb) { 1219 if (lb->in(LoadBarrierNode::Oop)->is_Phi()) { 1220 Node* oop_phi = lb->in(LoadBarrierNode::Oop); 1221 1222 if (oop_phi->req() == 2) { 1223 // Ignore phis with only one input 1224 return false; 1225 } 1226 1227 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())*/) { 1228 // That transformation may cause the Similar edge on dominated load barriers to be invalid 1229 lb->fix_similar_in_uses(&_igvn); 1230 1231 RegionNode* region = oop_phi->in(0)->as_Region(); 1232 1233 int backedge = LoopNode::LoopBackControl; 1234 if (region->is_Loop() && region->in(backedge)->is_Proj() && region->in(backedge)->in(0)->is_If()) { 1235 Node* c = region->in(backedge)->in(0)->in(0); 1236 assert(c->unique_ctrl_out() == region->in(backedge)->in(0), ""); 1237 Node* oop = lb->in(LoadBarrierNode::Oop)->in(backedge); 1238 Node* oop_c = has_ctrl(oop) ? get_ctrl(oop) : oop; 1239 if (!is_dominator(oop_c, c)) { 1240 return false; 1241 } 1242 } 1243 1244 Node *phi = oop_phi->clone(); 1245 1246 for (uint i = 1; i < region->req(); i++) { 1247 Node* ctrl = region->in(i); 1248 if (ctrl != C->top()) { 1249 assert(!is_dominator(ctrl, region) || region->is_Loop(), ""); 1250 1251 Node* mem = lb->in(LoadBarrierNode::Memory); 1252 Node* m = find_dominating_memory(mem, region, i); 1253 1254 if (region->is_Loop() && i == LoopNode::LoopBackControl && ctrl->is_Proj() && ctrl->in(0)->is_If()) { 1255 ctrl = ctrl->in(0)->in(0); 1256 } 1257 1258 LoadBarrierNode* new_lb = clone_load_barrier(lb, ctrl, m, lb->in(LoadBarrierNode::Oop)->in(i)); 1259 1260 Node* out_ctrl = new_lb->proj_out(LoadBarrierNode::Control); 1261 if (ctrl == region->in(i)) { 1262 _igvn.replace_input_of(region, i, new_lb->proj_out(LoadBarrierNode::Control)); 1263 } else { 1264 Node* iff = region->in(i)->in(0); 1265 Node* out_ctrl = new_lb->proj_out(LoadBarrierNode::Control); 1266 _igvn.replace_input_of(iff, 0, out_ctrl); 1267 set_idom(iff, out_ctrl, dom_depth(out_ctrl)+1); 1268 } 1269 phi->set_req(i, new_lb->proj_out(LoadBarrierNode::Oop)); 1270 } 1271 } 1272 register_new_node(phi, region); 1273 1274 replace_barrier(lb, phi); 1275 1276 if (region->is_Loop()) { 1277 // Load barrier moved to the back edge of the Loop may now 1278 // have a safepoint on the path to the barrier on the Similar 1279 // edge 1280 _igvn.replace_input_of(phi->in(LoopNode::LoopBackControl)->in(0), LoadBarrierNode::Similar, C->top()); 1281 Node* head = region->in(LoopNode::EntryControl); 1282 set_idom(region, head, dom_depth(head)+1); 1283 recompute_dom_depth(); 1284 if (head->is_CountedLoop() && head->as_CountedLoop()->is_main_loop()) { 1285 head->as_CountedLoop()->set_normal_loop(); 1286 } 1287 } 1288 return true; 1289 } 1290 } 1291 return false; 1292 } 1293 1294 bool PhaseIdealLoop::move_out_of_loop(LoadBarrierNode* lb) { 1295 IdealLoopTree *lb_loop = get_loop(lb->in(0)); 1296 if (lb_loop != _ltree_root && !lb_loop->_irreducible) { 1297 Node* oop_ctrl = get_ctrl(lb->in(LoadBarrierNode::Oop)); 1298 IdealLoopTree *oop_loop = get_loop(oop_ctrl); 1299 IdealLoopTree* adr_loop = get_loop(get_ctrl(lb->in(LoadBarrierNode::Address))); 1300 if (!lb_loop->is_member(oop_loop) && !lb_loop->is_member(adr_loop)) { 1301 // That transformation may cause the Similar edge on dominated load barriers to be invalid 1302 lb->fix_similar_in_uses(&_igvn); 1303 1304 Node* head = lb_loop->_head; 1305 assert(head->is_Loop(), ""); 1306 1307 if (is_dominator(head, oop_ctrl)) { 1308 assert(oop_ctrl->Opcode() == Op_CProj && oop_ctrl->in(0)->Opcode() == Op_NeverBranch, ""); 1309 assert(lb_loop->is_member(get_loop(oop_ctrl->in(0)->in(0))), ""); 1310 return false; 1311 } 1312 1313 if (head->is_CountedLoop() && head->as_CountedLoop()->is_main_loop()) { 1314 head->as_CountedLoop()->set_normal_loop(); 1315 } 1316 1317 Node* mem = lb->in(LoadBarrierNode::Memory); 1318 Node* m = find_dominating_memory(mem, head); 1319 1320 LoadBarrierNode* new_lb = clone_load_barrier(lb, head->in(LoopNode::EntryControl), m, NULL); 1321 1322 assert(idom(head) == head->in(LoopNode::EntryControl), ""); 1323 Node* proj_ctl = new_lb->proj_out(LoadBarrierNode::Control); 1324 _igvn.replace_input_of(head, LoopNode::EntryControl, proj_ctl); 1325 set_idom(head, proj_ctl, dom_depth(proj_ctl)+1); 1326 1327 replace_barrier(lb, new_lb->proj_out(LoadBarrierNode::Oop)); 1328 1329 recompute_dom_depth(); 1330 1331 return true; 1332 } 1333 } 1334 return false; 1335 } 1336 1337 bool PhaseIdealLoop::common_barriers(LoadBarrierNode* lb) { 1338 Node* in_val = lb->in(LoadBarrierNode::Oop); 1339 for (DUIterator_Fast imax, i = in_val->fast_outs(imax); i < imax; i++) { 1340 Node* u = in_val->fast_out(i); 1341 if (u != lb && u->is_LoadBarrier() && u->as_LoadBarrier()->has_true_uses()) { 1342 Node* this_ctrl = lb->in(LoadBarrierNode::Control); 1343 Node* other_ctrl = u->in(LoadBarrierNode::Control); 1344 1345 Node* lca = dom_lca(this_ctrl, other_ctrl); 1346 bool ok = true; 1347 1348 Node* proj1 = NULL; 1349 Node* proj2 = NULL; 1350 1351 while (this_ctrl != lca && ok) { 1352 if (this_ctrl->in(0) != NULL && 1353 this_ctrl->in(0)->is_MultiBranch()) { 1354 if (this_ctrl->in(0)->in(0) == lca) { 1355 assert(proj1 == NULL, ""); 1356 assert(this_ctrl->is_Proj(), ""); 1357 proj1 = this_ctrl; 1358 } else if (!(this_ctrl->in(0)->is_If() && this_ctrl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none))) { 1359 ok = false; 1360 } 1361 } 1362 this_ctrl = idom(this_ctrl); 1363 } 1364 while (other_ctrl != lca && ok) { 1365 if (other_ctrl->in(0) != NULL && 1366 other_ctrl->in(0)->is_MultiBranch()) { 1367 if (other_ctrl->in(0)->in(0) == lca) { 1368 assert(other_ctrl->is_Proj(), ""); 1369 assert(proj2 == NULL, ""); 1370 proj2 = other_ctrl; 1371 } else if (!(other_ctrl->in(0)->is_If() && other_ctrl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none))) { 1372 ok = false; 1373 } 1374 } 1375 other_ctrl = idom(other_ctrl); 1376 } 1377 assert(proj1 == NULL || proj2 == NULL || proj1->in(0) == proj2->in(0), ""); 1378 if (ok && proj1 && proj2 && proj1 != proj2 && proj1->in(0)->is_If()) { 1379 // That transformation may cause the Similar edge on dominated load barriers to be invalid 1380 lb->fix_similar_in_uses(&_igvn); 1381 u->as_LoadBarrier()->fix_similar_in_uses(&_igvn); 1382 1383 Node* split = lca->unique_ctrl_out(); 1384 assert(split->in(0) == lca, ""); 1385 1386 Node* mem = lb->in(LoadBarrierNode::Memory); 1387 Node* m = find_dominating_memory(mem, split); 1388 LoadBarrierNode* new_lb = clone_load_barrier(lb, lca, m, NULL); 1389 1390 Node* proj_ctl = new_lb->proj_out(LoadBarrierNode::Control); 1391 _igvn.replace_input_of(split, 0, new_lb->proj_out(LoadBarrierNode::Control)); 1392 set_idom(split, proj_ctl, dom_depth(proj_ctl)+1); 1393 1394 Node* proj_oop = new_lb->proj_out(LoadBarrierNode::Oop); 1395 replace_barrier(lb, proj_oop); 1396 replace_barrier(u->as_LoadBarrier(), proj_oop); 1397 1398 recompute_dom_depth(); 1399 1400 return true; 1401 } 1402 } 1403 } 1404 return false; 1405 } 1406 1407 void PhaseIdealLoop::optimize_load_barrier(LoadBarrierNode* lb, bool last_round) { 1408 if (!C->directive()->OptimizeLoadBarriersOption) { 1409 return; 1410 } 1411 1412 if (lb->has_true_uses()) { 1413 1414 if (replace_with_dominating_barrier(lb, last_round)) { 1415 return; 1416 } 1417 1418 if (split_barrier_thru_phi(lb)) { 1419 return; 1420 } 1421 1422 if (move_out_of_loop(lb)) { 1423 return; 1424 } 1425 1426 if (common_barriers(lb)) { 1427 return; 1428 } 1429 } 1430 } 1431 1432 //------------------------------split_if_with_blocks_post---------------------- 1433 // Do the real work in a non-recursive function. CFG hackery wants to be 1434 // in the post-order, so it can dirty the I-DOM info and not use the dirtied 1435 // info. 1436 void PhaseIdealLoop::split_if_with_blocks_post(Node *n, bool last_round) { 1437 1438 // Cloning Cmp through Phi's involves the split-if transform. 1439 // FastLock is not used by an If 1440 if (n->is_Cmp() && !n->is_FastLock() && !last_round) { 1441 Node *n_ctrl = get_ctrl(n); 1442 // Determine if the Node has inputs from some local Phi. 1443 // Returns the block to clone thru. 1444 Node *n_blk = has_local_phi_input(n); 1445 if (n_blk != n_ctrl) { 1446 return; 1447 } 1448 1449 if (!can_split_if(n_ctrl)) { 1450 return; 1451 } 1452 1453 if (n->outcnt() != 1) { 1454 return; // Multiple bool's from 1 compare? 1455 } 1456 Node *bol = n->unique_out(); 1457 assert(bol->is_Bool(), "expect a bool here"); 1458 if (bol->outcnt() != 1) { 1459 return;// Multiple branches from 1 compare? 1460 } 1461 Node *iff = bol->unique_out(); 1462 1463 // Check some safety conditions 1464 if (iff->is_If()) { // Classic split-if? 1465 if (iff->in(0) != n_ctrl) { 1466 return; // Compare must be in same blk as if 1467 } 1468 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE 1469 // Can't split CMove with different control edge. 1470 if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) { 1471 return; 1472 } 1473 if (get_ctrl(iff->in(2)) == n_ctrl || 1474 get_ctrl(iff->in(3)) == n_ctrl) { 1475 return; // Inputs not yet split-up 1476 } 1477 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) { 1478 return; // Loop-invar test gates loop-varying CMOVE 1479 } 1480 } else { 1481 return; // some other kind of node, such as an Allocate 1482 } 1483 1484 // When is split-if profitable? Every 'win' on means some control flow 1485 // goes dead, so it's almost always a win. 1486 int policy = 0; 1487 // Split compare 'n' through the merge point if it is profitable 1488 Node *phi = split_thru_phi( n, n_ctrl, policy); 1489 if (!phi) { 1490 return; 1491 } 1492 1493 // Found a Phi to split thru! 1494 // Replace 'n' with the new phi 1495 _igvn.replace_node(n, phi); 1496 1497 // Now split the bool up thru the phi 1498 Node *bolphi = split_thru_phi(bol, n_ctrl, -1); 1499 guarantee(bolphi != NULL, "null boolean phi node"); 1500 1501 _igvn.replace_node(bol, bolphi); 1502 assert(iff->in(1) == bolphi, ""); 1503 1504 if (bolphi->Value(&_igvn)->singleton()) { 1505 return; 1506 } 1507 1508 // Conditional-move? Must split up now 1509 if (!iff->is_If()) { 1510 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1); 1511 _igvn.replace_node(iff, cmovphi); 1512 return; 1513 } 1514 1515 // Now split the IF 1516 do_split_if(iff); 1517 return; 1518 } 1519 1520 // Two identical ifs back to back can be merged 1521 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) { 1522 Node *n_ctrl = n->in(0); 1523 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1)); 1524 IfNode* dom_if = idom(n_ctrl)->as_If(); 1525 Node* proj_true = dom_if->proj_out(1); 1526 Node* proj_false = dom_if->proj_out(0); 1527 Node* con_true = _igvn.makecon(TypeInt::ONE); 1528 Node* con_false = _igvn.makecon(TypeInt::ZERO); 1529 1530 for (uint i = 1; i < n_ctrl->req(); i++) { 1531 if (is_dominator(proj_true, n_ctrl->in(i))) { 1532 bolphi->init_req(i, con_true); 1533 } else { 1534 assert(is_dominator(proj_false, n_ctrl->in(i)), "bad if"); 1535 bolphi->init_req(i, con_false); 1536 } 1537 } 1538 register_new_node(bolphi, n_ctrl); 1539 _igvn.replace_input_of(n, 1, bolphi); 1540 1541 // Now split the IF 1542 do_split_if(n); 1543 return; 1544 } 1545 1546 // Check for an IF ready to split; one that has its 1547 // condition codes input coming from a Phi at the block start. 1548 int n_op = n->Opcode(); 1549 1550 // Check for an IF being dominated by another IF same test 1551 if (n_op == Op_If || 1552 n_op == Op_RangeCheck) { 1553 Node *bol = n->in(1); 1554 uint max = bol->outcnt(); 1555 // Check for same test used more than once? 1556 if (max > 1 && bol->is_Bool()) { 1557 // Search up IDOMs to see if this IF is dominated. 1558 Node *cutoff = get_ctrl(bol); 1559 1560 // Now search up IDOMs till cutoff, looking for a dominating test 1561 Node *prevdom = n; 1562 Node *dom = idom(prevdom); 1563 while (dom != cutoff) { 1564 if (dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom) { 1565 // Replace the dominated test with an obvious true or false. 1566 // Place it on the IGVN worklist for later cleanup. 1567 C->set_major_progress(); 1568 dominated_by(prevdom, n, false, true); 1569 #ifndef PRODUCT 1570 if( VerifyLoopOptimizations ) verify(); 1571 #endif 1572 return; 1573 } 1574 prevdom = dom; 1575 dom = idom(prevdom); 1576 } 1577 } 1578 } 1579 1580 // See if a shared loop-varying computation has no loop-varying uses. 1581 // Happens if something is only used for JVM state in uncommon trap exits, 1582 // like various versions of induction variable+offset. Clone the 1583 // computation per usage to allow it to sink out of the loop. 1584 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about) 1585 Node *n_ctrl = get_ctrl(n); 1586 IdealLoopTree *n_loop = get_loop(n_ctrl); 1587 if( n_loop != _ltree_root ) { 1588 DUIterator_Fast imax, i = n->fast_outs(imax); 1589 for (; i < imax; i++) { 1590 Node* u = n->fast_out(i); 1591 if( !has_ctrl(u) ) break; // Found control user 1592 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 1593 if( u_loop == n_loop ) break; // Found loop-varying use 1594 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop 1595 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003 1596 } 1597 bool did_break = (i < imax); // Did we break out of the previous loop? 1598 if (!did_break && n->outcnt() > 1) { // All uses in outer loops! 1599 Node *late_load_ctrl = NULL; 1600 if (n->is_Load()) { 1601 // If n is a load, get and save the result from get_late_ctrl(), 1602 // to be later used in calculating the control for n's clones. 1603 clear_dom_lca_tags(); 1604 late_load_ctrl = get_late_ctrl(n, n_ctrl); 1605 } 1606 // If n is a load, and the late control is the same as the current 1607 // control, then the cloning of n is a pointless exercise, because 1608 // GVN will ensure that we end up where we started. 1609 if (!n->is_Load() || late_load_ctrl != n_ctrl) { 1610 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) { 1611 Node *u = n->last_out(j); // Clone private computation per use 1612 _igvn.rehash_node_delayed(u); 1613 Node *x = n->clone(); // Clone computation 1614 Node *x_ctrl = NULL; 1615 if( u->is_Phi() ) { 1616 // Replace all uses of normal nodes. Replace Phi uses 1617 // individually, so the separate Nodes can sink down 1618 // different paths. 1619 uint k = 1; 1620 while( u->in(k) != n ) k++; 1621 u->set_req( k, x ); 1622 // x goes next to Phi input path 1623 x_ctrl = u->in(0)->in(k); 1624 --j; 1625 } else { // Normal use 1626 // Replace all uses 1627 for( uint k = 0; k < u->req(); k++ ) { 1628 if( u->in(k) == n ) { 1629 u->set_req( k, x ); 1630 --j; 1631 } 1632 } 1633 x_ctrl = get_ctrl(u); 1634 } 1635 1636 // Find control for 'x' next to use but not inside inner loops. 1637 // For inner loop uses get the preheader area. 1638 x_ctrl = place_near_use(x_ctrl); 1639 1640 if (n->is_Load()) { 1641 // For loads, add a control edge to a CFG node outside of the loop 1642 // to force them to not combine and return back inside the loop 1643 // during GVN optimization (4641526). 1644 // 1645 // Because we are setting the actual control input, factor in 1646 // the result from get_late_ctrl() so we respect any 1647 // anti-dependences. (6233005). 1648 x_ctrl = dom_lca(late_load_ctrl, x_ctrl); 1649 1650 // Don't allow the control input to be a CFG splitting node. 1651 // Such nodes should only have ProjNodes as outs, e.g. IfNode 1652 // should only have IfTrueNode and IfFalseNode (4985384). 1653 x_ctrl = find_non_split_ctrl(x_ctrl); 1654 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone"); 1655 1656 x->set_req(0, x_ctrl); 1657 } 1658 register_new_node(x, x_ctrl); 1659 1660 // Some institutional knowledge is needed here: 'x' is 1661 // yanked because if the optimizer runs GVN on it all the 1662 // cloned x's will common up and undo this optimization and 1663 // be forced back in the loop. This is annoying because it 1664 // makes +VerifyOpto report false-positives on progress. I 1665 // tried setting control edges on the x's to force them to 1666 // not combine, but the matching gets worried when it tries 1667 // to fold a StoreP and an AddP together (as part of an 1668 // address expression) and the AddP and StoreP have 1669 // different controls. 1670 if (!x->is_Load() && !x->is_DecodeNarrowPtr()) _igvn._worklist.yank(x); 1671 } 1672 _igvn.remove_dead_node(n); 1673 } 1674 } 1675 } 1676 } 1677 1678 try_move_store_after_loop(n); 1679 1680 // Check for Opaque2's who's loop has disappeared - who's input is in the 1681 // same loop nest as their output. Remove 'em, they are no longer useful. 1682 if( n_op == Op_Opaque2 && 1683 n->in(1) != NULL && 1684 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) { 1685 _igvn.replace_node( n, n->in(1) ); 1686 } 1687 1688 if (n->is_LoadBarrier()) { 1689 optimize_load_barrier(n->as_LoadBarrier(), last_round); 1690 } 1691 } 1692 1693 //------------------------------split_if_with_blocks--------------------------- 1694 // Check for aggressive application of 'split-if' optimization, 1695 // using basic block level info. 1696 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack, bool last_round) { 1697 Node *n = C->root(); 1698 visited.set(n->_idx); // first, mark node as visited 1699 // Do pre-visit work for root 1700 n = split_if_with_blocks_pre( n ); 1701 uint cnt = n->outcnt(); 1702 uint i = 0; 1703 while (true) { 1704 // Visit all children 1705 if (i < cnt) { 1706 Node* use = n->raw_out(i); 1707 ++i; 1708 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) { 1709 // Now do pre-visit work for this use 1710 use = split_if_with_blocks_pre( use ); 1711 nstack.push(n, i); // Save parent and next use's index. 1712 n = use; // Process all children of current use. 1713 cnt = use->outcnt(); 1714 i = 0; 1715 } 1716 } 1717 else { 1718 // All of n's children have been processed, complete post-processing. 1719 if (cnt != 0 && !n->is_Con()) { 1720 assert(has_node(n), "no dead nodes"); 1721 split_if_with_blocks_post(n, last_round); 1722 } 1723 if (nstack.is_empty()) { 1724 // Finished all nodes on stack. 1725 break; 1726 } 1727 // Get saved parent node and next use's index. Visit the rest of uses. 1728 n = nstack.node(); 1729 cnt = n->outcnt(); 1730 i = nstack.index(); 1731 nstack.pop(); 1732 } 1733 } 1734 } 1735 1736 1737 //============================================================================= 1738 // 1739 // C L O N E A L O O P B O D Y 1740 // 1741 1742 //------------------------------clone_iff-------------------------------------- 1743 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1744 // "Nearly" because all Nodes have been cloned from the original in the loop, 1745 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1746 // through the Phi recursively, and return a Bool. 1747 Node* PhaseIdealLoop::clone_iff(PhiNode *phi, IdealLoopTree *loop) { 1748 1749 // Convert this Phi into a Phi merging Bools 1750 uint i; 1751 for (i = 1; i < phi->req(); i++) { 1752 Node *b = phi->in(i); 1753 if (b->is_Phi()) { 1754 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi(), loop)); 1755 } else { 1756 assert(b->is_Bool() || b->Opcode() == Op_Opaque4, ""); 1757 } 1758 } 1759 1760 Node* n = phi->in(1); 1761 Node* sample_opaque = NULL; 1762 Node *sample_bool = NULL; 1763 if (n->Opcode() == Op_Opaque4) { 1764 sample_opaque = n; 1765 sample_bool = n->in(1); 1766 assert(sample_bool->is_Bool(), "wrong type"); 1767 } else { 1768 sample_bool = n; 1769 } 1770 Node *sample_cmp = sample_bool->in(1); 1771 1772 // Make Phis to merge the Cmp's inputs. 1773 PhiNode *phi1 = new PhiNode(phi->in(0), Type::TOP); 1774 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP); 1775 for (i = 1; i < phi->req(); i++) { 1776 Node *n1 = sample_opaque == NULL ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1); 1777 Node *n2 = sample_opaque == NULL ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2); 1778 phi1->set_req(i, n1); 1779 phi2->set_req(i, n2); 1780 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 1781 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 1782 } 1783 // See if these Phis have been made before. 1784 // Register with optimizer 1785 Node *hit1 = _igvn.hash_find_insert(phi1); 1786 if (hit1) { // Hit, toss just made Phi 1787 _igvn.remove_dead_node(phi1); // Remove new phi 1788 assert(hit1->is_Phi(), "" ); 1789 phi1 = (PhiNode*)hit1; // Use existing phi 1790 } else { // Miss 1791 _igvn.register_new_node_with_optimizer(phi1); 1792 } 1793 Node *hit2 = _igvn.hash_find_insert(phi2); 1794 if (hit2) { // Hit, toss just made Phi 1795 _igvn.remove_dead_node(phi2); // Remove new phi 1796 assert(hit2->is_Phi(), "" ); 1797 phi2 = (PhiNode*)hit2; // Use existing phi 1798 } else { // Miss 1799 _igvn.register_new_node_with_optimizer(phi2); 1800 } 1801 // Register Phis with loop/block info 1802 set_ctrl(phi1, phi->in(0)); 1803 set_ctrl(phi2, phi->in(0)); 1804 // Make a new Cmp 1805 Node *cmp = sample_cmp->clone(); 1806 cmp->set_req(1, phi1); 1807 cmp->set_req(2, phi2); 1808 _igvn.register_new_node_with_optimizer(cmp); 1809 set_ctrl(cmp, phi->in(0)); 1810 1811 // Make a new Bool 1812 Node *b = sample_bool->clone(); 1813 b->set_req(1,cmp); 1814 _igvn.register_new_node_with_optimizer(b); 1815 set_ctrl(b, phi->in(0)); 1816 1817 if (sample_opaque != NULL) { 1818 Node* opaque = sample_opaque->clone(); 1819 opaque->set_req(1, b); 1820 _igvn.register_new_node_with_optimizer(opaque); 1821 set_ctrl(opaque, phi->in(0)); 1822 return opaque; 1823 } 1824 1825 assert(b->is_Bool(), ""); 1826 return b; 1827 } 1828 1829 //------------------------------clone_bool------------------------------------- 1830 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1831 // "Nearly" because all Nodes have been cloned from the original in the loop, 1832 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1833 // through the Phi recursively, and return a Bool. 1834 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) { 1835 uint i; 1836 // Convert this Phi into a Phi merging Bools 1837 for( i = 1; i < phi->req(); i++ ) { 1838 Node *b = phi->in(i); 1839 if( b->is_Phi() ) { 1840 _igvn.replace_input_of(phi, i, clone_bool( b->as_Phi(), loop )); 1841 } else { 1842 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" ); 1843 } 1844 } 1845 1846 Node *sample_cmp = phi->in(1); 1847 1848 // Make Phis to merge the Cmp's inputs. 1849 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP ); 1850 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP ); 1851 for( uint j = 1; j < phi->req(); j++ ) { 1852 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP 1853 Node *n1, *n2; 1854 if( cmp_top->is_Cmp() ) { 1855 n1 = cmp_top->in(1); 1856 n2 = cmp_top->in(2); 1857 } else { 1858 n1 = n2 = cmp_top; 1859 } 1860 phi1->set_req( j, n1 ); 1861 phi2->set_req( j, n2 ); 1862 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 1863 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 1864 } 1865 1866 // See if these Phis have been made before. 1867 // Register with optimizer 1868 Node *hit1 = _igvn.hash_find_insert(phi1); 1869 if( hit1 ) { // Hit, toss just made Phi 1870 _igvn.remove_dead_node(phi1); // Remove new phi 1871 assert( hit1->is_Phi(), "" ); 1872 phi1 = (PhiNode*)hit1; // Use existing phi 1873 } else { // Miss 1874 _igvn.register_new_node_with_optimizer(phi1); 1875 } 1876 Node *hit2 = _igvn.hash_find_insert(phi2); 1877 if( hit2 ) { // Hit, toss just made Phi 1878 _igvn.remove_dead_node(phi2); // Remove new phi 1879 assert( hit2->is_Phi(), "" ); 1880 phi2 = (PhiNode*)hit2; // Use existing phi 1881 } else { // Miss 1882 _igvn.register_new_node_with_optimizer(phi2); 1883 } 1884 // Register Phis with loop/block info 1885 set_ctrl(phi1, phi->in(0)); 1886 set_ctrl(phi2, phi->in(0)); 1887 // Make a new Cmp 1888 Node *cmp = sample_cmp->clone(); 1889 cmp->set_req( 1, phi1 ); 1890 cmp->set_req( 2, phi2 ); 1891 _igvn.register_new_node_with_optimizer(cmp); 1892 set_ctrl(cmp, phi->in(0)); 1893 1894 assert( cmp->is_Cmp(), "" ); 1895 return (CmpNode*)cmp; 1896 } 1897 1898 //------------------------------sink_use--------------------------------------- 1899 // If 'use' was in the loop-exit block, it now needs to be sunk 1900 // below the post-loop merge point. 1901 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) { 1902 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) { 1903 set_ctrl(use, post_loop); 1904 for (DUIterator j = use->outs(); use->has_out(j); j++) 1905 sink_use(use->out(j), post_loop); 1906 } 1907 } 1908 1909 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new, 1910 IdealLoopTree* loop, IdealLoopTree* outer_loop, 1911 Node_List*& split_if_set, Node_List*& split_bool_set, 1912 Node_List*& split_cex_set, Node_List& worklist, 1913 uint new_counter, CloneLoopMode mode) { 1914 Node* nnn = old_new[old->_idx]; 1915 // Copy uses to a worklist, so I can munge the def-use info 1916 // with impunity. 1917 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1918 worklist.push(old->fast_out(j)); 1919 1920 while( worklist.size() ) { 1921 Node *use = worklist.pop(); 1922 if (!has_node(use)) continue; // Ignore dead nodes 1923 if (use->in(0) == C->top()) continue; 1924 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1925 // Check for data-use outside of loop - at least one of OLD or USE 1926 // must not be a CFG node. 1927 #ifdef ASSERT 1928 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == NULL) { 1929 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint(); 1930 assert(mode == ControlAroundStripMined && use == sfpt, "missed a node"); 1931 } 1932 #endif 1933 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) { 1934 1935 // If the Data use is an IF, that means we have an IF outside of the 1936 // loop that is switching on a condition that is set inside of the 1937 // loop. Happens if people set a loop-exit flag; then test the flag 1938 // in the loop to break the loop, then test is again outside of the 1939 // loop to determine which way the loop exited. 1940 // Loop predicate If node connects to Bool node through Opaque1 node. 1941 if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use) || use->Opcode() == Op_Opaque4) { 1942 // Since this code is highly unlikely, we lazily build the worklist 1943 // of such Nodes to go split. 1944 if (!split_if_set) { 1945 ResourceArea *area = Thread::current()->resource_area(); 1946 split_if_set = new Node_List(area); 1947 } 1948 split_if_set->push(use); 1949 } 1950 if (use->is_Bool()) { 1951 if (!split_bool_set) { 1952 ResourceArea *area = Thread::current()->resource_area(); 1953 split_bool_set = new Node_List(area); 1954 } 1955 split_bool_set->push(use); 1956 } 1957 if (use->Opcode() == Op_CreateEx) { 1958 if (!split_cex_set) { 1959 ResourceArea *area = Thread::current()->resource_area(); 1960 split_cex_set = new Node_List(area); 1961 } 1962 split_cex_set->push(use); 1963 } 1964 1965 1966 // Get "block" use is in 1967 uint idx = 0; 1968 while( use->in(idx) != old ) idx++; 1969 Node *prev = use->is_CFG() ? use : get_ctrl(use); 1970 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" ); 1971 Node *cfg = prev->_idx >= new_counter 1972 ? prev->in(2) 1973 : idom(prev); 1974 if( use->is_Phi() ) // Phi use is in prior block 1975 cfg = prev->in(idx); // NOT in block of Phi itself 1976 if (cfg->is_top()) { // Use is dead? 1977 _igvn.replace_input_of(use, idx, C->top()); 1978 continue; 1979 } 1980 1981 while(!outer_loop->is_member(get_loop(cfg))) { 1982 prev = cfg; 1983 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg); 1984 } 1985 // If the use occurs after merging several exits from the loop, then 1986 // old value must have dominated all those exits. Since the same old 1987 // value was used on all those exits we did not need a Phi at this 1988 // merge point. NOW we do need a Phi here. Each loop exit value 1989 // is now merged with the peeled body exit; each exit gets its own 1990 // private Phi and those Phis need to be merged here. 1991 Node *phi; 1992 if( prev->is_Region() ) { 1993 if( idx == 0 ) { // Updating control edge? 1994 phi = prev; // Just use existing control 1995 } else { // Else need a new Phi 1996 phi = PhiNode::make( prev, old ); 1997 // Now recursively fix up the new uses of old! 1998 for( uint i = 1; i < prev->req(); i++ ) { 1999 worklist.push(phi); // Onto worklist once for each 'old' input 2000 } 2001 } 2002 } else { 2003 // Get new RegionNode merging old and new loop exits 2004 prev = old_new[prev->_idx]; 2005 assert( prev, "just made this in step 7" ); 2006 if( idx == 0) { // Updating control edge? 2007 phi = prev; // Just use existing control 2008 } else { // Else need a new Phi 2009 // Make a new Phi merging data values properly 2010 phi = PhiNode::make( prev, old ); 2011 phi->set_req( 1, nnn ); 2012 } 2013 } 2014 // If inserting a new Phi, check for prior hits 2015 if( idx != 0 ) { 2016 Node *hit = _igvn.hash_find_insert(phi); 2017 if( hit == NULL ) { 2018 _igvn.register_new_node_with_optimizer(phi); // Register new phi 2019 } else { // or 2020 // Remove the new phi from the graph and use the hit 2021 _igvn.remove_dead_node(phi); 2022 phi = hit; // Use existing phi 2023 } 2024 set_ctrl(phi, prev); 2025 } 2026 // Make 'use' use the Phi instead of the old loop body exit value 2027 _igvn.replace_input_of(use, idx, phi); 2028 if( use->_idx >= new_counter ) { // If updating new phis 2029 // Not needed for correctness, but prevents a weak assert 2030 // in AddPNode from tripping (when we end up with different 2031 // base & derived Phis that will become the same after 2032 // IGVN does CSE). 2033 Node *hit = _igvn.hash_find_insert(use); 2034 if( hit ) // Go ahead and re-hash for hits. 2035 _igvn.replace_node( use, hit ); 2036 } 2037 2038 // If 'use' was in the loop-exit block, it now needs to be sunk 2039 // below the post-loop merge point. 2040 sink_use( use, prev ); 2041 } 2042 } 2043 } 2044 2045 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop, 2046 IdealLoopTree* outer_loop, int dd, Node_List &old_new, 2047 Node_List& extra_data_nodes) { 2048 if (head->is_strip_mined() && mode != IgnoreStripMined) { 2049 CountedLoopNode* cl = head->as_CountedLoop(); 2050 Node* l = cl->outer_loop(); 2051 Node* tail = cl->outer_loop_tail(); 2052 IfNode* le = cl->outer_loop_end(); 2053 Node* sfpt = cl->outer_safepoint(); 2054 CountedLoopEndNode* cle = cl->loopexit(); 2055 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop(); 2056 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null(); 2057 Node* cle_out = cle->proj_out(false); 2058 2059 Node* new_sfpt = NULL; 2060 Node* new_cle_out = cle_out->clone(); 2061 old_new.map(cle_out->_idx, new_cle_out); 2062 if (mode == CloneIncludesStripMined) { 2063 // clone outer loop body 2064 Node* new_l = l->clone(); 2065 Node* new_tail = tail->clone(); 2066 IfNode* new_le = le->clone()->as_If(); 2067 new_sfpt = sfpt->clone(); 2068 2069 set_loop(new_l, outer_loop->_parent); 2070 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd); 2071 set_loop(new_cle_out, outer_loop->_parent); 2072 set_idom(new_cle_out, new_cle, dd); 2073 set_loop(new_sfpt, outer_loop->_parent); 2074 set_idom(new_sfpt, new_cle_out, dd); 2075 set_loop(new_le, outer_loop->_parent); 2076 set_idom(new_le, new_sfpt, dd); 2077 set_loop(new_tail, outer_loop->_parent); 2078 set_idom(new_tail, new_le, dd); 2079 set_idom(new_cl, new_l, dd); 2080 2081 old_new.map(l->_idx, new_l); 2082 old_new.map(tail->_idx, new_tail); 2083 old_new.map(le->_idx, new_le); 2084 old_new.map(sfpt->_idx, new_sfpt); 2085 2086 new_l->set_req(LoopNode::LoopBackControl, new_tail); 2087 new_l->set_req(0, new_l); 2088 new_tail->set_req(0, new_le); 2089 new_le->set_req(0, new_sfpt); 2090 new_sfpt->set_req(0, new_cle_out); 2091 new_cle_out->set_req(0, new_cle); 2092 new_cl->set_req(LoopNode::EntryControl, new_l); 2093 2094 _igvn.register_new_node_with_optimizer(new_l); 2095 _igvn.register_new_node_with_optimizer(new_tail); 2096 _igvn.register_new_node_with_optimizer(new_le); 2097 } else { 2098 Node *newhead = old_new[loop->_head->_idx]; 2099 newhead->as_Loop()->clear_strip_mined(); 2100 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl)); 2101 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 2102 } 2103 // Look at data node that were assigned a control in the outer 2104 // loop: they are kept in the outer loop by the safepoint so start 2105 // from the safepoint node's inputs. 2106 IdealLoopTree* outer_loop = get_loop(l); 2107 Node_Stack stack(2); 2108 stack.push(sfpt, 1); 2109 uint new_counter = C->unique(); 2110 while (stack.size() > 0) { 2111 Node* n = stack.node(); 2112 uint i = stack.index(); 2113 while (i < n->req() && 2114 (n->in(i) == NULL || 2115 !has_ctrl(n->in(i)) || 2116 get_loop(get_ctrl(n->in(i))) != outer_loop || 2117 (old_new[n->in(i)->_idx] != NULL && old_new[n->in(i)->_idx]->_idx >= new_counter))) { 2118 i++; 2119 } 2120 if (i < n->req()) { 2121 stack.set_index(i+1); 2122 stack.push(n->in(i), 0); 2123 } else { 2124 assert(old_new[n->_idx] == NULL || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet"); 2125 Node* m = n == sfpt ? new_sfpt : n->clone(); 2126 if (m != NULL) { 2127 for (uint i = 0; i < n->req(); i++) { 2128 if (m->in(i) != NULL && old_new[m->in(i)->_idx] != NULL) { 2129 m->set_req(i, old_new[m->in(i)->_idx]); 2130 } 2131 } 2132 } else { 2133 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?"); 2134 } 2135 if (n != sfpt) { 2136 extra_data_nodes.push(n); 2137 _igvn.register_new_node_with_optimizer(m); 2138 assert(get_ctrl(n) == cle_out, "what other control?"); 2139 set_ctrl(m, new_cle_out); 2140 old_new.map(n->_idx, m); 2141 } 2142 stack.pop(); 2143 } 2144 } 2145 if (mode == CloneIncludesStripMined) { 2146 _igvn.register_new_node_with_optimizer(new_sfpt); 2147 _igvn.register_new_node_with_optimizer(new_cle_out); 2148 } 2149 } else { 2150 Node *newhead = old_new[loop->_head->_idx]; 2151 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 2152 } 2153 } 2154 2155 //------------------------------clone_loop------------------------------------- 2156 // 2157 // C L O N E A L O O P B O D Y 2158 // 2159 // This is the basic building block of the loop optimizations. It clones an 2160 // entire loop body. It makes an old_new loop body mapping; with this mapping 2161 // you can find the new-loop equivalent to an old-loop node. All new-loop 2162 // nodes are exactly equal to their old-loop counterparts, all edges are the 2163 // same. All exits from the old-loop now have a RegionNode that merges the 2164 // equivalent new-loop path. This is true even for the normal "loop-exit" 2165 // condition. All uses of loop-invariant old-loop values now come from (one 2166 // or more) Phis that merge their new-loop equivalents. 2167 // 2168 // This operation leaves the graph in an illegal state: there are two valid 2169 // control edges coming from the loop pre-header to both loop bodies. I'll 2170 // definitely have to hack the graph after running this transform. 2171 // 2172 // From this building block I will further edit edges to perform loop peeling 2173 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc. 2174 // 2175 // Parameter side_by_size_idom: 2176 // When side_by_size_idom is NULL, the dominator tree is constructed for 2177 // the clone loop to dominate the original. Used in construction of 2178 // pre-main-post loop sequence. 2179 // When nonnull, the clone and original are side-by-side, both are 2180 // dominated by the side_by_side_idom node. Used in construction of 2181 // unswitched loops. 2182 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd, 2183 CloneLoopMode mode, Node* side_by_side_idom) { 2184 2185 LoopNode* head = loop->_head->as_Loop(); 2186 head->verify_strip_mined(1); 2187 2188 if (C->do_vector_loop() && PrintOpto) { 2189 const char* mname = C->method()->name()->as_quoted_ascii(); 2190 if (mname != NULL) { 2191 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname); 2192 } 2193 } 2194 2195 CloneMap& cm = C->clone_map(); 2196 Dict* dict = cm.dict(); 2197 if (C->do_vector_loop()) { 2198 cm.set_clone_idx(cm.max_gen()+1); 2199 #ifndef PRODUCT 2200 if (PrintOpto) { 2201 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx()); 2202 loop->dump_head(); 2203 } 2204 #endif 2205 } 2206 2207 // Step 1: Clone the loop body. Make the old->new mapping. 2208 uint i; 2209 for( i = 0; i < loop->_body.size(); i++ ) { 2210 Node *old = loop->_body.at(i); 2211 Node *nnn = old->clone(); 2212 old_new.map( old->_idx, nnn ); 2213 if (C->do_vector_loop()) { 2214 cm.verify_insert_and_clone(old, nnn, cm.clone_idx()); 2215 } 2216 _igvn.register_new_node_with_optimizer(nnn); 2217 } 2218 2219 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop; 2220 2221 // Step 2: Fix the edges in the new body. If the old input is outside the 2222 // loop use it. If the old input is INside the loop, use the corresponding 2223 // new node instead. 2224 for( i = 0; i < loop->_body.size(); i++ ) { 2225 Node *old = loop->_body.at(i); 2226 Node *nnn = old_new[old->_idx]; 2227 // Fix CFG/Loop controlling the new node 2228 if (has_ctrl(old)) { 2229 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]); 2230 } else { 2231 set_loop(nnn, outer_loop->_parent); 2232 if (old->outcnt() > 0) { 2233 set_idom( nnn, old_new[idom(old)->_idx], dd ); 2234 } 2235 } 2236 // Correct edges to the new node 2237 for( uint j = 0; j < nnn->req(); j++ ) { 2238 Node *n = nnn->in(j); 2239 if( n ) { 2240 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n ); 2241 if( loop->is_member( old_in_loop ) ) 2242 nnn->set_req(j, old_new[n->_idx]); 2243 } 2244 } 2245 _igvn.hash_find_insert(nnn); 2246 } 2247 2248 ResourceArea *area = Thread::current()->resource_area(); 2249 Node_List extra_data_nodes(area); 2250 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes); 2251 2252 // Step 3: Now fix control uses. Loop varying control uses have already 2253 // been fixed up (as part of all input edges in Step 2). Loop invariant 2254 // control uses must be either an IfFalse or an IfTrue. Make a merge 2255 // point to merge the old and new IfFalse/IfTrue nodes; make the use 2256 // refer to this. 2257 Node_List worklist(area); 2258 uint new_counter = C->unique(); 2259 for( i = 0; i < loop->_body.size(); i++ ) { 2260 Node* old = loop->_body.at(i); 2261 if( !old->is_CFG() ) continue; 2262 2263 // Copy uses to a worklist, so I can munge the def-use info 2264 // with impunity. 2265 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 2266 worklist.push(old->fast_out(j)); 2267 2268 while( worklist.size() ) { // Visit all uses 2269 Node *use = worklist.pop(); 2270 if (!has_node(use)) continue; // Ignore dead nodes 2271 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 2272 if( !loop->is_member( use_loop ) && use->is_CFG() ) { 2273 // Both OLD and USE are CFG nodes here. 2274 assert( use->is_Proj(), "" ); 2275 Node* nnn = old_new[old->_idx]; 2276 2277 Node* newuse = NULL; 2278 if (head->is_strip_mined() && mode != IgnoreStripMined) { 2279 CountedLoopNode* cl = head->as_CountedLoop(); 2280 CountedLoopEndNode* cle = cl->loopexit(); 2281 Node* cle_out = cle->proj_out_or_null(false); 2282 if (use == cle_out) { 2283 IfNode* le = cl->outer_loop_end(); 2284 use = le->proj_out(false); 2285 use_loop = get_loop(use); 2286 if (mode == CloneIncludesStripMined) { 2287 nnn = old_new[le->_idx]; 2288 } else { 2289 newuse = old_new[cle_out->_idx]; 2290 } 2291 } 2292 } 2293 if (newuse == NULL) { 2294 newuse = use->clone(); 2295 } 2296 2297 // Clone the loop exit control projection 2298 if (C->do_vector_loop()) { 2299 cm.verify_insert_and_clone(use, newuse, cm.clone_idx()); 2300 } 2301 newuse->set_req(0,nnn); 2302 _igvn.register_new_node_with_optimizer(newuse); 2303 set_loop(newuse, use_loop); 2304 set_idom(newuse, nnn, dom_depth(nnn) + 1 ); 2305 2306 // We need a Region to merge the exit from the peeled body and the 2307 // exit from the old loop body. 2308 RegionNode *r = new RegionNode(3); 2309 // Map the old use to the new merge point 2310 old_new.map( use->_idx, r ); 2311 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use)); 2312 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" ); 2313 2314 // The original user of 'use' uses 'r' instead. 2315 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) { 2316 Node* useuse = use->last_out(l); 2317 _igvn.rehash_node_delayed(useuse); 2318 uint uses_found = 0; 2319 if( useuse->in(0) == use ) { 2320 useuse->set_req(0, r); 2321 uses_found++; 2322 if( useuse->is_CFG() ) { 2323 assert( dom_depth(useuse) > dd_r, "" ); 2324 set_idom(useuse, r, dom_depth(useuse)); 2325 } 2326 } 2327 for( uint k = 1; k < useuse->req(); k++ ) { 2328 if( useuse->in(k) == use ) { 2329 useuse->set_req(k, r); 2330 uses_found++; 2331 if (useuse->is_Loop() && k == LoopNode::EntryControl) { 2332 assert(dom_depth(useuse) > dd_r , ""); 2333 set_idom(useuse, r, dom_depth(useuse)); 2334 } 2335 } 2336 } 2337 l -= uses_found; // we deleted 1 or more copies of this edge 2338 } 2339 2340 // Now finish up 'r' 2341 r->set_req( 1, newuse ); 2342 r->set_req( 2, use ); 2343 _igvn.register_new_node_with_optimizer(r); 2344 set_loop(r, use_loop); 2345 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r); 2346 } // End of if a loop-exit test 2347 } 2348 } 2349 2350 // Step 4: If loop-invariant use is not control, it must be dominated by a 2351 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region 2352 // there if needed. Make a Phi there merging old and new used values. 2353 Node_List *split_if_set = NULL; 2354 Node_List *split_bool_set = NULL; 2355 Node_List *split_cex_set = NULL; 2356 for( i = 0; i < loop->_body.size(); i++ ) { 2357 Node* old = loop->_body.at(i); 2358 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set, 2359 split_bool_set, split_cex_set, worklist, new_counter, 2360 mode); 2361 } 2362 2363 for (i = 0; i < extra_data_nodes.size(); i++) { 2364 Node* old = extra_data_nodes.at(i); 2365 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set, 2366 split_bool_set, split_cex_set, worklist, new_counter, 2367 mode); 2368 } 2369 2370 // Check for IFs that need splitting/cloning. Happens if an IF outside of 2371 // the loop uses a condition set in the loop. The original IF probably 2372 // takes control from one or more OLD Regions (which in turn get from NEW 2373 // Regions). In any case, there will be a set of Phis for each merge point 2374 // from the IF up to where the original BOOL def exists the loop. 2375 if (split_if_set) { 2376 while (split_if_set->size()) { 2377 Node *iff = split_if_set->pop(); 2378 if (iff->in(1)->is_Phi()) { 2379 Node *b = clone_iff(iff->in(1)->as_Phi(), loop); 2380 _igvn.replace_input_of(iff, 1, b); 2381 } 2382 } 2383 } 2384 if (split_bool_set) { 2385 while (split_bool_set->size()) { 2386 Node *b = split_bool_set->pop(); 2387 Node *phi = b->in(1); 2388 assert(phi->is_Phi(), ""); 2389 CmpNode *cmp = clone_bool((PhiNode*)phi, loop); 2390 _igvn.replace_input_of(b, 1, cmp); 2391 } 2392 } 2393 if (split_cex_set) { 2394 while (split_cex_set->size()) { 2395 Node *b = split_cex_set->pop(); 2396 assert(b->in(0)->is_Region(), ""); 2397 assert(b->in(1)->is_Phi(), ""); 2398 assert(b->in(0)->in(0) == b->in(1)->in(0), ""); 2399 split_up(b, b->in(0), NULL); 2400 } 2401 } 2402 2403 } 2404 2405 2406 //---------------------- stride_of_possible_iv ------------------------------------- 2407 // Looks for an iff/bool/comp with one operand of the compare 2408 // being a cycle involving an add and a phi, 2409 // with an optional truncation (left-shift followed by a right-shift) 2410 // of the add. Returns zero if not an iv. 2411 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) { 2412 Node* trunc1 = NULL; 2413 Node* trunc2 = NULL; 2414 const TypeInt* ttype = NULL; 2415 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) { 2416 return 0; 2417 } 2418 BoolNode* bl = iff->in(1)->as_Bool(); 2419 Node* cmp = bl->in(1); 2420 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) { 2421 return 0; 2422 } 2423 // Must have an invariant operand 2424 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) { 2425 return 0; 2426 } 2427 Node* add2 = NULL; 2428 Node* cmp1 = cmp->in(1); 2429 if (cmp1->is_Phi()) { 2430 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) ))) 2431 Node* phi = cmp1; 2432 for (uint i = 1; i < phi->req(); i++) { 2433 Node* in = phi->in(i); 2434 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in, 2435 &trunc1, &trunc2, &ttype); 2436 if (add && add->in(1) == phi) { 2437 add2 = add->in(2); 2438 break; 2439 } 2440 } 2441 } else { 2442 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) ))) 2443 Node* addtrunc = cmp1; 2444 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc, 2445 &trunc1, &trunc2, &ttype); 2446 if (add && add->in(1)->is_Phi()) { 2447 Node* phi = add->in(1); 2448 for (uint i = 1; i < phi->req(); i++) { 2449 if (phi->in(i) == addtrunc) { 2450 add2 = add->in(2); 2451 break; 2452 } 2453 } 2454 } 2455 } 2456 if (add2 != NULL) { 2457 const TypeInt* add2t = _igvn.type(add2)->is_int(); 2458 if (add2t->is_con()) { 2459 return add2t->get_con(); 2460 } 2461 } 2462 return 0; 2463 } 2464 2465 2466 //---------------------- stay_in_loop ------------------------------------- 2467 // Return the (unique) control output node that's in the loop (if it exists.) 2468 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) { 2469 Node* unique = NULL; 2470 if (!n) return NULL; 2471 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2472 Node* use = n->fast_out(i); 2473 if (!has_ctrl(use) && loop->is_member(get_loop(use))) { 2474 if (unique != NULL) { 2475 return NULL; 2476 } 2477 unique = use; 2478 } 2479 } 2480 return unique; 2481 } 2482 2483 //------------------------------ register_node ------------------------------------- 2484 // Utility to register node "n" with PhaseIdealLoop 2485 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) { 2486 _igvn.register_new_node_with_optimizer(n); 2487 loop->_body.push(n); 2488 if (n->is_CFG()) { 2489 set_loop(n, loop); 2490 set_idom(n, pred, ddepth); 2491 } else { 2492 set_ctrl(n, pred); 2493 } 2494 } 2495 2496 //------------------------------ proj_clone ------------------------------------- 2497 // Utility to create an if-projection 2498 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) { 2499 ProjNode* c = p->clone()->as_Proj(); 2500 c->set_req(0, iff); 2501 return c; 2502 } 2503 2504 //------------------------------ short_circuit_if ------------------------------------- 2505 // Force the iff control output to be the live_proj 2506 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) { 2507 guarantee(live_proj != NULL, "null projection"); 2508 int proj_con = live_proj->_con; 2509 assert(proj_con == 0 || proj_con == 1, "false or true projection"); 2510 Node *con = _igvn.intcon(proj_con); 2511 set_ctrl(con, C->root()); 2512 if (iff) { 2513 iff->set_req(1, con); 2514 } 2515 return con; 2516 } 2517 2518 //------------------------------ insert_if_before_proj ------------------------------------- 2519 // Insert a new if before an if projection (* - new node) 2520 // 2521 // before 2522 // if(test) 2523 // / \ 2524 // v v 2525 // other-proj proj (arg) 2526 // 2527 // after 2528 // if(test) 2529 // / \ 2530 // / v 2531 // | * proj-clone 2532 // v | 2533 // other-proj v 2534 // * new_if(relop(cmp[IU](left,right))) 2535 // / \ 2536 // v v 2537 // * new-proj proj 2538 // (returned) 2539 // 2540 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) { 2541 IfNode* iff = proj->in(0)->as_If(); 2542 IdealLoopTree *loop = get_loop(proj); 2543 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2544 int ddepth = dom_depth(proj); 2545 2546 _igvn.rehash_node_delayed(iff); 2547 _igvn.rehash_node_delayed(proj); 2548 2549 proj->set_req(0, NULL); // temporary disconnect 2550 ProjNode* proj2 = proj_clone(proj, iff); 2551 register_node(proj2, loop, iff, ddepth); 2552 2553 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right); 2554 register_node(cmp, loop, proj2, ddepth); 2555 2556 BoolNode* bol = new BoolNode(cmp, relop); 2557 register_node(bol, loop, proj2, ddepth); 2558 2559 int opcode = iff->Opcode(); 2560 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode"); 2561 IfNode* new_if = (opcode == Op_If) ? new IfNode(proj2, bol, iff->_prob, iff->_fcnt): 2562 new RangeCheckNode(proj2, bol, iff->_prob, iff->_fcnt); 2563 register_node(new_if, loop, proj2, ddepth); 2564 2565 proj->set_req(0, new_if); // reattach 2566 set_idom(proj, new_if, ddepth); 2567 2568 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj(); 2569 guarantee(new_exit != NULL, "null exit node"); 2570 register_node(new_exit, get_loop(other_proj), new_if, ddepth); 2571 2572 return new_exit; 2573 } 2574 2575 //------------------------------ insert_region_before_proj ------------------------------------- 2576 // Insert a region before an if projection (* - new node) 2577 // 2578 // before 2579 // if(test) 2580 // / | 2581 // v | 2582 // proj v 2583 // other-proj 2584 // 2585 // after 2586 // if(test) 2587 // / | 2588 // v | 2589 // * proj-clone v 2590 // | other-proj 2591 // v 2592 // * new-region 2593 // | 2594 // v 2595 // * dum_if 2596 // / \ 2597 // v \ 2598 // * dum-proj v 2599 // proj 2600 // 2601 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) { 2602 IfNode* iff = proj->in(0)->as_If(); 2603 IdealLoopTree *loop = get_loop(proj); 2604 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2605 int ddepth = dom_depth(proj); 2606 2607 _igvn.rehash_node_delayed(iff); 2608 _igvn.rehash_node_delayed(proj); 2609 2610 proj->set_req(0, NULL); // temporary disconnect 2611 ProjNode* proj2 = proj_clone(proj, iff); 2612 register_node(proj2, loop, iff, ddepth); 2613 2614 RegionNode* reg = new RegionNode(2); 2615 reg->set_req(1, proj2); 2616 register_node(reg, loop, iff, ddepth); 2617 2618 IfNode* dum_if = new IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt); 2619 register_node(dum_if, loop, reg, ddepth); 2620 2621 proj->set_req(0, dum_if); // reattach 2622 set_idom(proj, dum_if, ddepth); 2623 2624 ProjNode* dum_proj = proj_clone(other_proj, dum_if); 2625 register_node(dum_proj, loop, dum_if, ddepth); 2626 2627 return reg; 2628 } 2629 2630 //------------------------------ insert_cmpi_loop_exit ------------------------------------- 2631 // Clone a signed compare loop exit from an unsigned compare and 2632 // insert it before the unsigned cmp on the stay-in-loop path. 2633 // All new nodes inserted in the dominator tree between the original 2634 // if and it's projections. The original if test is replaced with 2635 // a constant to force the stay-in-loop path. 2636 // 2637 // This is done to make sure that the original if and it's projections 2638 // still dominate the same set of control nodes, that the ctrl() relation 2639 // from data nodes to them is preserved, and that their loop nesting is 2640 // preserved. 2641 // 2642 // before 2643 // if(i <u limit) unsigned compare loop exit 2644 // / | 2645 // v v 2646 // exit-proj stay-in-loop-proj 2647 // 2648 // after 2649 // if(stay-in-loop-const) original if 2650 // / | 2651 // / v 2652 // / if(i < limit) new signed test 2653 // / / | 2654 // / / v 2655 // / / if(i <u limit) new cloned unsigned test 2656 // / / / | 2657 // v v v | 2658 // region | 2659 // | | 2660 // dum-if | 2661 // / | | 2662 // ether | | 2663 // v v 2664 // exit-proj stay-in-loop-proj 2665 // 2666 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) { 2667 const bool Signed = true; 2668 const bool Unsigned = false; 2669 2670 BoolNode* bol = if_cmpu->in(1)->as_Bool(); 2671 if (bol->_test._test != BoolTest::lt) return NULL; 2672 CmpNode* cmpu = bol->in(1)->as_Cmp(); 2673 if (cmpu->Opcode() != Op_CmpU) return NULL; 2674 int stride = stride_of_possible_iv(if_cmpu); 2675 if (stride == 0) return NULL; 2676 2677 Node* lp_proj = stay_in_loop(if_cmpu, loop); 2678 guarantee(lp_proj != NULL, "null loop node"); 2679 2680 ProjNode* lp_continue = lp_proj->as_Proj(); 2681 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj(); 2682 2683 Node* limit = NULL; 2684 if (stride > 0) { 2685 limit = cmpu->in(2); 2686 } else { 2687 limit = _igvn.makecon(TypeInt::ZERO); 2688 set_ctrl(limit, C->root()); 2689 } 2690 // Create a new region on the exit path 2691 RegionNode* reg = insert_region_before_proj(lp_exit); 2692 guarantee(reg != NULL, "null region node"); 2693 2694 // Clone the if-cmpu-true-false using a signed compare 2695 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge; 2696 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue); 2697 reg->add_req(cmpi_exit); 2698 2699 // Clone the if-cmpu-true-false 2700 BoolTest::mask rel_u = bol->_test._test; 2701 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue); 2702 reg->add_req(cmpu_exit); 2703 2704 // Force original if to stay in loop. 2705 short_circuit_if(if_cmpu, lp_continue); 2706 2707 return cmpi_exit->in(0)->as_If(); 2708 } 2709 2710 //------------------------------ remove_cmpi_loop_exit ------------------------------------- 2711 // Remove a previously inserted signed compare loop exit. 2712 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) { 2713 Node* lp_proj = stay_in_loop(if_cmp, loop); 2714 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI && 2715 stay_in_loop(lp_proj, loop)->is_If() && 2716 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu"); 2717 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO); 2718 set_ctrl(con, C->root()); 2719 if_cmp->set_req(1, con); 2720 } 2721 2722 //------------------------------ scheduled_nodelist ------------------------------------- 2723 // Create a post order schedule of nodes that are in the 2724 // "member" set. The list is returned in "sched". 2725 // The first node in "sched" is the loop head, followed by 2726 // nodes which have no inputs in the "member" set, and then 2727 // followed by the nodes that have an immediate input dependence 2728 // on a node in "sched". 2729 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) { 2730 2731 assert(member.test(loop->_head->_idx), "loop head must be in member set"); 2732 Arena *a = Thread::current()->resource_area(); 2733 VectorSet visited(a); 2734 Node_Stack nstack(a, loop->_body.size()); 2735 2736 Node* n = loop->_head; // top of stack is cached in "n" 2737 uint idx = 0; 2738 visited.set(n->_idx); 2739 2740 // Initially push all with no inputs from within member set 2741 for(uint i = 0; i < loop->_body.size(); i++ ) { 2742 Node *elt = loop->_body.at(i); 2743 if (member.test(elt->_idx)) { 2744 bool found = false; 2745 for (uint j = 0; j < elt->req(); j++) { 2746 Node* def = elt->in(j); 2747 if (def && member.test(def->_idx) && def != elt) { 2748 found = true; 2749 break; 2750 } 2751 } 2752 if (!found && elt != loop->_head) { 2753 nstack.push(n, idx); 2754 n = elt; 2755 assert(!visited.test(n->_idx), "not seen yet"); 2756 visited.set(n->_idx); 2757 } 2758 } 2759 } 2760 2761 // traverse out's that are in the member set 2762 while (true) { 2763 if (idx < n->outcnt()) { 2764 Node* use = n->raw_out(idx); 2765 idx++; 2766 if (!visited.test_set(use->_idx)) { 2767 if (member.test(use->_idx)) { 2768 nstack.push(n, idx); 2769 n = use; 2770 idx = 0; 2771 } 2772 } 2773 } else { 2774 // All outputs processed 2775 sched.push(n); 2776 if (nstack.is_empty()) break; 2777 n = nstack.node(); 2778 idx = nstack.index(); 2779 nstack.pop(); 2780 } 2781 } 2782 } 2783 2784 2785 //------------------------------ has_use_in_set ------------------------------------- 2786 // Has a use in the vector set 2787 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) { 2788 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2789 Node* use = n->fast_out(j); 2790 if (vset.test(use->_idx)) { 2791 return true; 2792 } 2793 } 2794 return false; 2795 } 2796 2797 2798 //------------------------------ has_use_internal_to_set ------------------------------------- 2799 // Has use internal to the vector set (ie. not in a phi at the loop head) 2800 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) { 2801 Node* head = loop->_head; 2802 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2803 Node* use = n->fast_out(j); 2804 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) { 2805 return true; 2806 } 2807 } 2808 return false; 2809 } 2810 2811 2812 //------------------------------ clone_for_use_outside_loop ------------------------------------- 2813 // clone "n" for uses that are outside of loop 2814 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) { 2815 int cloned = 0; 2816 assert(worklist.size() == 0, "should be empty"); 2817 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2818 Node* use = n->fast_out(j); 2819 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) { 2820 worklist.push(use); 2821 } 2822 } 2823 while( worklist.size() ) { 2824 Node *use = worklist.pop(); 2825 if (!has_node(use) || use->in(0) == C->top()) continue; 2826 uint j; 2827 for (j = 0; j < use->req(); j++) { 2828 if (use->in(j) == n) break; 2829 } 2830 assert(j < use->req(), "must be there"); 2831 2832 // clone "n" and insert it between the inputs of "n" and the use outside the loop 2833 Node* n_clone = n->clone(); 2834 _igvn.replace_input_of(use, j, n_clone); 2835 cloned++; 2836 Node* use_c; 2837 if (!use->is_Phi()) { 2838 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0); 2839 } else { 2840 // Use in a phi is considered a use in the associated predecessor block 2841 use_c = use->in(0)->in(j); 2842 } 2843 set_ctrl(n_clone, use_c); 2844 assert(!loop->is_member(get_loop(use_c)), "should be outside loop"); 2845 get_loop(use_c)->_body.push(n_clone); 2846 _igvn.register_new_node_with_optimizer(n_clone); 2847 #if !defined(PRODUCT) 2848 if (TracePartialPeeling) { 2849 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx); 2850 } 2851 #endif 2852 } 2853 return cloned; 2854 } 2855 2856 2857 //------------------------------ clone_for_special_use_inside_loop ------------------------------------- 2858 // clone "n" for special uses that are in the not_peeled region. 2859 // If these def-uses occur in separate blocks, the code generator 2860 // marks the method as not compilable. For example, if a "BoolNode" 2861 // is in a different basic block than the "IfNode" that uses it, then 2862 // the compilation is aborted in the code generator. 2863 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 2864 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) { 2865 if (n->is_Phi() || n->is_Load()) { 2866 return; 2867 } 2868 assert(worklist.size() == 0, "should be empty"); 2869 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2870 Node* use = n->fast_out(j); 2871 if ( not_peel.test(use->_idx) && 2872 (use->is_If() || use->is_CMove() || use->is_Bool()) && 2873 use->in(1) == n) { 2874 worklist.push(use); 2875 } 2876 } 2877 if (worklist.size() > 0) { 2878 // clone "n" and insert it between inputs of "n" and the use 2879 Node* n_clone = n->clone(); 2880 loop->_body.push(n_clone); 2881 _igvn.register_new_node_with_optimizer(n_clone); 2882 set_ctrl(n_clone, get_ctrl(n)); 2883 sink_list.push(n_clone); 2884 not_peel <<= n_clone->_idx; // add n_clone to not_peel set. 2885 #if !defined(PRODUCT) 2886 if (TracePartialPeeling) { 2887 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx); 2888 } 2889 #endif 2890 while( worklist.size() ) { 2891 Node *use = worklist.pop(); 2892 _igvn.rehash_node_delayed(use); 2893 for (uint j = 1; j < use->req(); j++) { 2894 if (use->in(j) == n) { 2895 use->set_req(j, n_clone); 2896 } 2897 } 2898 } 2899 } 2900 } 2901 2902 2903 //------------------------------ insert_phi_for_loop ------------------------------------- 2904 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 2905 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) { 2906 Node *phi = PhiNode::make(lp, back_edge_val); 2907 phi->set_req(LoopNode::EntryControl, lp_entry_val); 2908 // Use existing phi if it already exists 2909 Node *hit = _igvn.hash_find_insert(phi); 2910 if( hit == NULL ) { 2911 _igvn.register_new_node_with_optimizer(phi); 2912 set_ctrl(phi, lp); 2913 } else { 2914 // Remove the new phi from the graph and use the hit 2915 _igvn.remove_dead_node(phi); 2916 phi = hit; 2917 } 2918 _igvn.replace_input_of(use, idx, phi); 2919 } 2920 2921 #ifdef ASSERT 2922 //------------------------------ is_valid_loop_partition ------------------------------------- 2923 // Validate the loop partition sets: peel and not_peel 2924 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, 2925 VectorSet& not_peel ) { 2926 uint i; 2927 // Check that peel_list entries are in the peel set 2928 for (i = 0; i < peel_list.size(); i++) { 2929 if (!peel.test(peel_list.at(i)->_idx)) { 2930 return false; 2931 } 2932 } 2933 // Check at loop members are in one of peel set or not_peel set 2934 for (i = 0; i < loop->_body.size(); i++ ) { 2935 Node *def = loop->_body.at(i); 2936 uint di = def->_idx; 2937 // Check that peel set elements are in peel_list 2938 if (peel.test(di)) { 2939 if (not_peel.test(di)) { 2940 return false; 2941 } 2942 // Must be in peel_list also 2943 bool found = false; 2944 for (uint j = 0; j < peel_list.size(); j++) { 2945 if (peel_list.at(j)->_idx == di) { 2946 found = true; 2947 break; 2948 } 2949 } 2950 if (!found) { 2951 return false; 2952 } 2953 } else if (not_peel.test(di)) { 2954 if (peel.test(di)) { 2955 return false; 2956 } 2957 } else { 2958 return false; 2959 } 2960 } 2961 return true; 2962 } 2963 2964 //------------------------------ is_valid_clone_loop_exit_use ------------------------------------- 2965 // Ensure a use outside of loop is of the right form 2966 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) { 2967 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2968 return (use->is_Phi() && 2969 use_c->is_Region() && use_c->req() == 3 && 2970 (use_c->in(exit_idx)->Opcode() == Op_IfTrue || 2971 use_c->in(exit_idx)->Opcode() == Op_IfFalse || 2972 use_c->in(exit_idx)->Opcode() == Op_JumpProj) && 2973 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) ); 2974 } 2975 2976 //------------------------------ is_valid_clone_loop_form ------------------------------------- 2977 // Ensure that all uses outside of loop are of the right form 2978 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 2979 uint orig_exit_idx, uint clone_exit_idx) { 2980 uint len = peel_list.size(); 2981 for (uint i = 0; i < len; i++) { 2982 Node *def = peel_list.at(i); 2983 2984 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 2985 Node *use = def->fast_out(j); 2986 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2987 if (!loop->is_member(get_loop(use_c))) { 2988 // use is not in the loop, check for correct structure 2989 if (use->in(0) == def) { 2990 // Okay 2991 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) { 2992 return false; 2993 } 2994 } 2995 } 2996 } 2997 return true; 2998 } 2999 #endif 3000 3001 //------------------------------ partial_peel ------------------------------------- 3002 // Partially peel (aka loop rotation) the top portion of a loop (called 3003 // the peel section below) by cloning it and placing one copy just before 3004 // the new loop head and the other copy at the bottom of the new loop. 3005 // 3006 // before after where it came from 3007 // 3008 // stmt1 stmt1 3009 // loop: stmt2 clone 3010 // stmt2 if condA goto exitA clone 3011 // if condA goto exitA new_loop: new 3012 // stmt3 stmt3 clone 3013 // if !condB goto loop if condB goto exitB clone 3014 // exitB: stmt2 orig 3015 // stmt4 if !condA goto new_loop orig 3016 // exitA: goto exitA 3017 // exitB: 3018 // stmt4 3019 // exitA: 3020 // 3021 // Step 1: find the cut point: an exit test on probable 3022 // induction variable. 3023 // Step 2: schedule (with cloning) operations in the peel 3024 // section that can be executed after the cut into 3025 // the section that is not peeled. This may need 3026 // to clone operations into exit blocks. For 3027 // instance, a reference to A[i] in the not-peel 3028 // section and a reference to B[i] in an exit block 3029 // may cause a left-shift of i by 2 to be placed 3030 // in the peel block. This step will clone the left 3031 // shift into the exit block and sink the left shift 3032 // from the peel to the not-peel section. 3033 // Step 3: clone the loop, retarget the control, and insert 3034 // phis for values that are live across the new loop 3035 // head. This is very dependent on the graph structure 3036 // from clone_loop. It creates region nodes for 3037 // exit control and associated phi nodes for values 3038 // flow out of the loop through that exit. The region 3039 // node is dominated by the clone's control projection. 3040 // So the clone's peel section is placed before the 3041 // new loop head, and the clone's not-peel section is 3042 // forms the top part of the new loop. The original 3043 // peel section forms the tail of the new loop. 3044 // Step 4: update the dominator tree and recompute the 3045 // dominator depth. 3046 // 3047 // orig 3048 // 3049 // stmt1 3050 // | 3051 // v 3052 // loop predicate 3053 // | 3054 // v 3055 // loop<----+ 3056 // | | 3057 // stmt2 | 3058 // | | 3059 // v | 3060 // ifA | 3061 // / | | 3062 // v v | 3063 // false true ^ <-- last_peel 3064 // / | | 3065 // / ===|==cut | 3066 // / stmt3 | <-- first_not_peel 3067 // / | | 3068 // | v | 3069 // v ifB | 3070 // exitA: / \ | 3071 // / \ | 3072 // v v | 3073 // false true | 3074 // / \ | 3075 // / ----+ 3076 // | 3077 // v 3078 // exitB: 3079 // stmt4 3080 // 3081 // 3082 // after clone loop 3083 // 3084 // stmt1 3085 // | 3086 // v 3087 // loop predicate 3088 // / \ 3089 // clone / \ orig 3090 // / \ 3091 // / \ 3092 // v v 3093 // +---->loop loop<----+ 3094 // | | | | 3095 // | stmt2 stmt2 | 3096 // | | | | 3097 // | v v | 3098 // | ifA ifA | 3099 // | | \ / | | 3100 // | v v v v | 3101 // ^ true false false true ^ <-- last_peel 3102 // | | ^ \ / | | 3103 // | cut==|== \ \ / ===|==cut | 3104 // | stmt3 \ \ / stmt3 | <-- first_not_peel 3105 // | | dom | | | | 3106 // | v \ 1v v2 v | 3107 // | ifB regionA ifB | 3108 // | / \ | / \ | 3109 // | / \ v / \ | 3110 // | v v exitA: v v | 3111 // | true false false true | 3112 // | / ^ \ / \ | 3113 // +---- \ \ / ----+ 3114 // dom \ / 3115 // \ 1v v2 3116 // regionB 3117 // | 3118 // v 3119 // exitB: 3120 // stmt4 3121 // 3122 // 3123 // after partial peel 3124 // 3125 // stmt1 3126 // | 3127 // v 3128 // loop predicate 3129 // / 3130 // clone / orig 3131 // / TOP 3132 // / \ 3133 // v v 3134 // TOP->loop loop----+ 3135 // | | | 3136 // stmt2 stmt2 | 3137 // | | | 3138 // v v | 3139 // ifA ifA | 3140 // | \ / | | 3141 // v v v v | 3142 // true false false true | <-- last_peel 3143 // | ^ \ / +------|---+ 3144 // +->newloop \ \ / === ==cut | | 3145 // | stmt3 \ \ / TOP | | 3146 // | | dom | | stmt3 | | <-- first_not_peel 3147 // | v \ 1v v2 v | | 3148 // | ifB regionA ifB ^ v 3149 // | / \ | / \ | | 3150 // | / \ v / \ | | 3151 // | v v exitA: v v | | 3152 // | true false false true | | 3153 // | / ^ \ / \ | | 3154 // | | \ \ / v | | 3155 // | | dom \ / TOP | | 3156 // | | \ 1v v2 | | 3157 // ^ v regionB | | 3158 // | | | | | 3159 // | | v ^ v 3160 // | | exitB: | | 3161 // | | stmt4 | | 3162 // | +------------>-----------------+ | 3163 // | | 3164 // +-----------------<---------------------+ 3165 // 3166 // 3167 // final graph 3168 // 3169 // stmt1 3170 // | 3171 // v 3172 // loop predicate 3173 // | 3174 // v 3175 // stmt2 clone 3176 // | 3177 // v 3178 // ........> ifA clone 3179 // : / | 3180 // dom / | 3181 // : v v 3182 // : false true 3183 // : | | 3184 // : | v 3185 // : | newloop<-----+ 3186 // : | | | 3187 // : | stmt3 clone | 3188 // : | | | 3189 // : | v | 3190 // : | ifB | 3191 // : | / \ | 3192 // : | v v | 3193 // : | false true | 3194 // : | | | | 3195 // : | v stmt2 | 3196 // : | exitB: | | 3197 // : | stmt4 v | 3198 // : | ifA orig | 3199 // : | / \ | 3200 // : | / \ | 3201 // : | v v | 3202 // : | false true | 3203 // : | / \ | 3204 // : v v -----+ 3205 // RegionA 3206 // | 3207 // v 3208 // exitA 3209 // 3210 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) { 3211 3212 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only"); 3213 if (!loop->_head->is_Loop()) { 3214 return false; } 3215 3216 LoopNode *head = loop->_head->as_Loop(); 3217 3218 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) { 3219 return false; 3220 } 3221 3222 // Check for complex exit control 3223 for(uint ii = 0; ii < loop->_body.size(); ii++ ) { 3224 Node *n = loop->_body.at(ii); 3225 int opc = n->Opcode(); 3226 if (n->is_Call() || 3227 opc == Op_Catch || 3228 opc == Op_CatchProj || 3229 opc == Op_Jump || 3230 opc == Op_JumpProj) { 3231 #if !defined(PRODUCT) 3232 if (TracePartialPeeling) { 3233 tty->print_cr("\nExit control too complex: lp: %d", head->_idx); 3234 } 3235 #endif 3236 return false; 3237 } 3238 } 3239 3240 int dd = dom_depth(head); 3241 3242 // Step 1: find cut point 3243 3244 // Walk up dominators to loop head looking for first loop exit 3245 // which is executed on every path thru loop. 3246 IfNode *peel_if = NULL; 3247 IfNode *peel_if_cmpu = NULL; 3248 3249 Node *iff = loop->tail(); 3250 while( iff != head ) { 3251 if( iff->is_If() ) { 3252 Node *ctrl = get_ctrl(iff->in(1)); 3253 if (ctrl->is_top()) return false; // Dead test on live IF. 3254 // If loop-varying exit-test, check for induction variable 3255 if( loop->is_member(get_loop(ctrl)) && 3256 loop->is_loop_exit(iff) && 3257 is_possible_iv_test(iff)) { 3258 Node* cmp = iff->in(1)->in(1); 3259 if (cmp->Opcode() == Op_CmpI) { 3260 peel_if = iff->as_If(); 3261 } else { 3262 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU"); 3263 peel_if_cmpu = iff->as_If(); 3264 } 3265 } 3266 } 3267 iff = idom(iff); 3268 } 3269 // Prefer signed compare over unsigned compare. 3270 IfNode* new_peel_if = NULL; 3271 if (peel_if == NULL) { 3272 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) { 3273 return false; // No peel point found 3274 } 3275 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop); 3276 if (new_peel_if == NULL) { 3277 return false; // No peel point found 3278 } 3279 peel_if = new_peel_if; 3280 } 3281 Node* last_peel = stay_in_loop(peel_if, loop); 3282 Node* first_not_peeled = stay_in_loop(last_peel, loop); 3283 if (first_not_peeled == NULL || first_not_peeled == head) { 3284 return false; 3285 } 3286 3287 #if !defined(PRODUCT) 3288 if (TraceLoopOpts) { 3289 tty->print("PartialPeel "); 3290 loop->dump_head(); 3291 } 3292 3293 if (TracePartialPeeling) { 3294 tty->print_cr("before partial peel one iteration"); 3295 Node_List wl; 3296 Node* t = head->in(2); 3297 while (true) { 3298 wl.push(t); 3299 if (t == head) break; 3300 t = idom(t); 3301 } 3302 while (wl.size() > 0) { 3303 Node* tt = wl.pop(); 3304 tt->dump(); 3305 if (tt == last_peel) tty->print_cr("-- cut --"); 3306 } 3307 } 3308 #endif 3309 ResourceArea *area = Thread::current()->resource_area(); 3310 VectorSet peel(area); 3311 VectorSet not_peel(area); 3312 Node_List peel_list(area); 3313 Node_List worklist(area); 3314 Node_List sink_list(area); 3315 3316 // Set of cfg nodes to peel are those that are executable from 3317 // the head through last_peel. 3318 assert(worklist.size() == 0, "should be empty"); 3319 worklist.push(head); 3320 peel.set(head->_idx); 3321 while (worklist.size() > 0) { 3322 Node *n = worklist.pop(); 3323 if (n != last_peel) { 3324 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3325 Node* use = n->fast_out(j); 3326 if (use->is_CFG() && 3327 loop->is_member(get_loop(use)) && 3328 !peel.test_set(use->_idx)) { 3329 worklist.push(use); 3330 } 3331 } 3332 } 3333 } 3334 3335 // Set of non-cfg nodes to peel are those that are control 3336 // dependent on the cfg nodes. 3337 uint i; 3338 for(i = 0; i < loop->_body.size(); i++ ) { 3339 Node *n = loop->_body.at(i); 3340 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n; 3341 if (peel.test(n_c->_idx)) { 3342 peel.set(n->_idx); 3343 } else { 3344 not_peel.set(n->_idx); 3345 } 3346 } 3347 3348 // Step 2: move operations from the peeled section down into the 3349 // not-peeled section 3350 3351 // Get a post order schedule of nodes in the peel region 3352 // Result in right-most operand. 3353 scheduled_nodelist(loop, peel, peel_list ); 3354 3355 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 3356 3357 // For future check for too many new phis 3358 uint old_phi_cnt = 0; 3359 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 3360 Node* use = head->fast_out(j); 3361 if (use->is_Phi()) old_phi_cnt++; 3362 } 3363 3364 #if !defined(PRODUCT) 3365 if (TracePartialPeeling) { 3366 tty->print_cr("\npeeled list"); 3367 } 3368 #endif 3369 3370 // Evacuate nodes in peel region into the not_peeled region if possible 3371 uint new_phi_cnt = 0; 3372 uint cloned_for_outside_use = 0; 3373 for (i = 0; i < peel_list.size();) { 3374 Node* n = peel_list.at(i); 3375 #if !defined(PRODUCT) 3376 if (TracePartialPeeling) n->dump(); 3377 #endif 3378 bool incr = true; 3379 if ( !n->is_CFG() ) { 3380 3381 if ( has_use_in_set(n, not_peel) ) { 3382 3383 // If not used internal to the peeled region, 3384 // move "n" from peeled to not_peeled region. 3385 3386 if ( !has_use_internal_to_set(n, peel, loop) ) { 3387 3388 // if not pinned and not a load (which maybe anti-dependent on a store) 3389 // and not a CMove (Matcher expects only bool->cmove). 3390 if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) { 3391 cloned_for_outside_use += clone_for_use_outside_loop( loop, n, worklist ); 3392 sink_list.push(n); 3393 peel >>= n->_idx; // delete n from peel set. 3394 not_peel <<= n->_idx; // add n to not_peel set. 3395 peel_list.remove(i); 3396 incr = false; 3397 #if !defined(PRODUCT) 3398 if (TracePartialPeeling) { 3399 tty->print_cr("sink to not_peeled region: %d newbb: %d", 3400 n->_idx, get_ctrl(n)->_idx); 3401 } 3402 #endif 3403 } 3404 } else { 3405 // Otherwise check for special def-use cases that span 3406 // the peel/not_peel boundary such as bool->if 3407 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist ); 3408 new_phi_cnt++; 3409 } 3410 } 3411 } 3412 if (incr) i++; 3413 } 3414 3415 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) { 3416 #if !defined(PRODUCT) 3417 if (TracePartialPeeling) { 3418 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c", 3419 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F'); 3420 } 3421 #endif 3422 if (new_peel_if != NULL) { 3423 remove_cmpi_loop_exit(new_peel_if, loop); 3424 } 3425 // Inhibit more partial peeling on this loop 3426 assert(!head->is_partial_peel_loop(), "not partial peeled"); 3427 head->mark_partial_peel_failed(); 3428 if (cloned_for_outside_use > 0) { 3429 // Terminate this round of loop opts because 3430 // the graph outside this loop was changed. 3431 C->set_major_progress(); 3432 return true; 3433 } 3434 return false; 3435 } 3436 3437 // Step 3: clone loop, retarget control, and insert new phis 3438 3439 // Create new loop head for new phis and to hang 3440 // the nodes being moved (sinked) from the peel region. 3441 LoopNode* new_head = new LoopNode(last_peel, last_peel); 3442 new_head->set_unswitch_count(head->unswitch_count()); // Preserve 3443 _igvn.register_new_node_with_optimizer(new_head); 3444 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled"); 3445 _igvn.replace_input_of(first_not_peeled, 0, new_head); 3446 set_loop(new_head, loop); 3447 loop->_body.push(new_head); 3448 not_peel.set(new_head->_idx); 3449 set_idom(new_head, last_peel, dom_depth(first_not_peeled)); 3450 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled)); 3451 3452 while (sink_list.size() > 0) { 3453 Node* n = sink_list.pop(); 3454 set_ctrl(n, new_head); 3455 } 3456 3457 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 3458 3459 clone_loop(loop, old_new, dd, IgnoreStripMined); 3460 3461 const uint clone_exit_idx = 1; 3462 const uint orig_exit_idx = 2; 3463 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop"); 3464 3465 Node* head_clone = old_new[head->_idx]; 3466 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop(); 3467 Node* orig_tail_clone = head_clone->in(2); 3468 3469 // Add phi if "def" node is in peel set and "use" is not 3470 3471 for(i = 0; i < peel_list.size(); i++ ) { 3472 Node *def = peel_list.at(i); 3473 if (!def->is_CFG()) { 3474 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 3475 Node *use = def->fast_out(j); 3476 if (has_node(use) && use->in(0) != C->top() && 3477 (!peel.test(use->_idx) || 3478 (use->is_Phi() && use->in(0) == head)) ) { 3479 worklist.push(use); 3480 } 3481 } 3482 while( worklist.size() ) { 3483 Node *use = worklist.pop(); 3484 for (uint j = 1; j < use->req(); j++) { 3485 Node* n = use->in(j); 3486 if (n == def) { 3487 3488 // "def" is in peel set, "use" is not in peel set 3489 // or "use" is in the entry boundary (a phi) of the peel set 3490 3491 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use; 3492 3493 if ( loop->is_member(get_loop( use_c )) ) { 3494 // use is in loop 3495 if (old_new[use->_idx] != NULL) { // null for dead code 3496 Node* use_clone = old_new[use->_idx]; 3497 _igvn.replace_input_of(use, j, C->top()); 3498 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone ); 3499 } 3500 } else { 3501 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format"); 3502 // use is not in the loop, check if the live range includes the cut 3503 Node* lp_if = use_c->in(orig_exit_idx)->in(0); 3504 if (not_peel.test(lp_if->_idx)) { 3505 assert(j == orig_exit_idx, "use from original loop"); 3506 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone ); 3507 } 3508 } 3509 } 3510 } 3511 } 3512 } 3513 } 3514 3515 // Step 3b: retarget control 3516 3517 // Redirect control to the new loop head if a cloned node in 3518 // the not_peeled region has control that points into the peeled region. 3519 // This necessary because the cloned peeled region will be outside 3520 // the loop. 3521 // from to 3522 // cloned-peeled <---+ 3523 // new_head_clone: | <--+ 3524 // cloned-not_peeled in(0) in(0) 3525 // orig-peeled 3526 3527 for(i = 0; i < loop->_body.size(); i++ ) { 3528 Node *n = loop->_body.at(i); 3529 if (!n->is_CFG() && n->in(0) != NULL && 3530 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) { 3531 Node* n_clone = old_new[n->_idx]; 3532 _igvn.replace_input_of(n_clone, 0, new_head_clone); 3533 } 3534 } 3535 3536 // Backedge of the surviving new_head (the clone) is original last_peel 3537 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel); 3538 3539 // Cut first node in original not_peel set 3540 _igvn.rehash_node_delayed(new_head); // Multiple edge updates: 3541 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of 3542 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls 3543 3544 // Copy head_clone back-branch info to original head 3545 // and remove original head's loop entry and 3546 // clone head's back-branch 3547 _igvn.rehash_node_delayed(head); // Multiple edge updates 3548 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl)); 3549 head->set_req(LoopNode::LoopBackControl, C->top()); 3550 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top()); 3551 3552 // Similarly modify the phis 3553 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) { 3554 Node* use = head->fast_out(k); 3555 if (use->is_Phi() && use->outcnt() > 0) { 3556 Node* use_clone = old_new[use->_idx]; 3557 _igvn.rehash_node_delayed(use); // Multiple edge updates 3558 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl)); 3559 use->set_req(LoopNode::LoopBackControl, C->top()); 3560 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top()); 3561 } 3562 } 3563 3564 // Step 4: update dominator tree and dominator depth 3565 3566 set_idom(head, orig_tail_clone, dd); 3567 recompute_dom_depth(); 3568 3569 // Inhibit more partial peeling on this loop 3570 new_head_clone->set_partial_peel_loop(); 3571 C->set_major_progress(); 3572 loop->record_for_igvn(); 3573 3574 #if !defined(PRODUCT) 3575 if (TracePartialPeeling) { 3576 tty->print_cr("\nafter partial peel one iteration"); 3577 Node_List wl(area); 3578 Node* t = last_peel; 3579 while (true) { 3580 wl.push(t); 3581 if (t == head_clone) break; 3582 t = idom(t); 3583 } 3584 while (wl.size() > 0) { 3585 Node* tt = wl.pop(); 3586 if (tt == head) tty->print_cr("orig head"); 3587 else if (tt == new_head_clone) tty->print_cr("new head"); 3588 else if (tt == head_clone) tty->print_cr("clone head"); 3589 tt->dump(); 3590 } 3591 } 3592 #endif 3593 return true; 3594 } 3595 3596 //------------------------------reorg_offsets---------------------------------- 3597 // Reorganize offset computations to lower register pressure. Mostly 3598 // prevent loop-fallout uses of the pre-incremented trip counter (which are 3599 // then alive with the post-incremented trip counter forcing an extra 3600 // register move) 3601 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) { 3602 // Perform it only for canonical counted loops. 3603 // Loop's shape could be messed up by iteration_split_impl. 3604 if (!loop->_head->is_CountedLoop()) 3605 return; 3606 if (!loop->_head->as_Loop()->is_valid_counted_loop()) 3607 return; 3608 3609 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 3610 CountedLoopEndNode *cle = cl->loopexit(); 3611 Node *exit = cle->proj_out(false); 3612 Node *phi = cl->phi(); 3613 3614 // Check for the special case of folks using the pre-incremented 3615 // trip-counter on the fall-out path (forces the pre-incremented 3616 // and post-incremented trip counter to be live at the same time). 3617 // Fix this by adjusting to use the post-increment trip counter. 3618 3619 bool progress = true; 3620 while (progress) { 3621 progress = false; 3622 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) { 3623 Node* use = phi->fast_out(i); // User of trip-counter 3624 if (!has_ctrl(use)) continue; 3625 Node *u_ctrl = get_ctrl(use); 3626 if (use->is_Phi()) { 3627 u_ctrl = NULL; 3628 for (uint j = 1; j < use->req(); j++) 3629 if (use->in(j) == phi) 3630 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j)); 3631 } 3632 IdealLoopTree *u_loop = get_loop(u_ctrl); 3633 // Look for loop-invariant use 3634 if (u_loop == loop) continue; 3635 if (loop->is_member(u_loop)) continue; 3636 // Check that use is live out the bottom. Assuming the trip-counter 3637 // update is right at the bottom, uses of of the loop middle are ok. 3638 if (dom_lca(exit, u_ctrl) != exit) continue; 3639 // Hit! Refactor use to use the post-incremented tripcounter. 3640 // Compute a post-increment tripcounter. 3641 Node *opaq = new Opaque2Node( C, cle->incr() ); 3642 register_new_node(opaq, exit); 3643 Node *neg_stride = _igvn.intcon(-cle->stride_con()); 3644 set_ctrl(neg_stride, C->root()); 3645 Node *post = new AddINode( opaq, neg_stride); 3646 register_new_node(post, exit); 3647 _igvn.rehash_node_delayed(use); 3648 for (uint j = 1; j < use->req(); j++) { 3649 if (use->in(j) == phi) 3650 use->set_req(j, post); 3651 } 3652 // Since DU info changed, rerun loop 3653 progress = true; 3654 break; 3655 } 3656 } 3657 3658 }