1 /* 2 * Copyright (c) 1999, 2017, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "memory/allocation.inline.hpp" 27 #include "memory/resourceArea.hpp" 28 #include "opto/addnode.hpp" 29 #include "opto/castnode.hpp" 30 #include "opto/connode.hpp" 31 #include "opto/castnode.hpp" 32 #include "opto/divnode.hpp" 33 #include "opto/loopnode.hpp" 34 #include "opto/matcher.hpp" 35 #include "opto/mulnode.hpp" 36 #include "opto/movenode.hpp" 37 #include "opto/opaquenode.hpp" 38 #include "opto/rootnode.hpp" 39 #include "opto/subnode.hpp" 40 41 //============================================================================= 42 //------------------------------split_thru_phi--------------------------------- 43 // Split Node 'n' through merge point if there is enough win. 44 Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) { 45 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) { 46 // ConvI2L may have type information on it which is unsafe to push up 47 // so disable this for now 48 return NULL; 49 } 50 51 // Splitting range check CastIIs through a loop induction Phi can 52 // cause new Phis to be created that are left unrelated to the loop 53 // induction Phi and prevent optimizations (vectorization) 54 if (n->Opcode() == Op_CastII && n->as_CastII()->has_range_check() && 55 region->is_CountedLoop() && n->in(1) == region->as_CountedLoop()->phi()) { 56 return NULL; 57 } 58 59 int wins = 0; 60 assert(!n->is_CFG(), ""); 61 assert(region->is_Region(), ""); 62 63 const Type* type = n->bottom_type(); 64 const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr(); 65 Node *phi; 66 if (t_oop != NULL && t_oop->is_known_instance_field()) { 67 int iid = t_oop->instance_id(); 68 int index = C->get_alias_index(t_oop); 69 int offset = t_oop->offset(); 70 phi = new PhiNode(region, type, NULL, iid, index, offset); 71 } else { 72 phi = PhiNode::make_blank(region, n); 73 } 74 uint old_unique = C->unique(); 75 for (uint i = 1; i < region->req(); i++) { 76 Node *x; 77 Node* the_clone = NULL; 78 if (region->in(i) == C->top()) { 79 x = C->top(); // Dead path? Use a dead data op 80 } else { 81 x = n->clone(); // Else clone up the data op 82 the_clone = x; // Remember for possible deletion. 83 // Alter data node to use pre-phi inputs 84 if (n->in(0) == region) 85 x->set_req( 0, region->in(i) ); 86 for (uint j = 1; j < n->req(); j++) { 87 Node *in = n->in(j); 88 if (in->is_Phi() && in->in(0) == region) 89 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone 90 } 91 } 92 // Check for a 'win' on some paths 93 const Type *t = x->Value(&_igvn); 94 95 bool singleton = t->singleton(); 96 97 // A TOP singleton indicates that there are no possible values incoming 98 // along a particular edge. In most cases, this is OK, and the Phi will 99 // be eliminated later in an Ideal call. However, we can't allow this to 100 // happen if the singleton occurs on loop entry, as the elimination of 101 // the PhiNode may cause the resulting node to migrate back to a previous 102 // loop iteration. 103 if (singleton && t == Type::TOP) { 104 // Is_Loop() == false does not confirm the absence of a loop (e.g., an 105 // irreducible loop may not be indicated by an affirmative is_Loop()); 106 // therefore, the only top we can split thru a phi is on a backedge of 107 // a loop. 108 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 109 } 110 111 if (singleton) { 112 wins++; 113 x = ((PhaseGVN&)_igvn).makecon(t); 114 } else { 115 // We now call Identity to try to simplify the cloned node. 116 // Note that some Identity methods call phase->type(this). 117 // Make sure that the type array is big enough for 118 // our new node, even though we may throw the node away. 119 // (Note: This tweaking with igvn only works because x is a new node.) 120 _igvn.set_type(x, t); 121 // If x is a TypeNode, capture any more-precise type permanently into Node 122 // otherwise it will be not updated during igvn->transform since 123 // igvn->type(x) is set to x->Value() already. 124 x->raise_bottom_type(t); 125 Node *y = x->Identity(&_igvn); 126 if (y != x) { 127 wins++; 128 x = y; 129 } else { 130 y = _igvn.hash_find(x); 131 if (y) { 132 wins++; 133 x = y; 134 } else { 135 // Else x is a new node we are keeping 136 // We do not need register_new_node_with_optimizer 137 // because set_type has already been called. 138 _igvn._worklist.push(x); 139 } 140 } 141 } 142 if (x != the_clone && the_clone != NULL) 143 _igvn.remove_dead_node(the_clone); 144 phi->set_req( i, x ); 145 } 146 // Too few wins? 147 if (wins <= policy) { 148 _igvn.remove_dead_node(phi); 149 return NULL; 150 } 151 152 // Record Phi 153 register_new_node( phi, region ); 154 155 for (uint i2 = 1; i2 < phi->req(); i2++) { 156 Node *x = phi->in(i2); 157 // If we commoned up the cloned 'x' with another existing Node, 158 // the existing Node picks up a new use. We need to make the 159 // existing Node occur higher up so it dominates its uses. 160 Node *old_ctrl; 161 IdealLoopTree *old_loop; 162 163 if (x->is_Con()) { 164 // Constant's control is always root. 165 set_ctrl(x, C->root()); 166 continue; 167 } 168 // The occasional new node 169 if (x->_idx >= old_unique) { // Found a new, unplaced node? 170 old_ctrl = NULL; 171 old_loop = NULL; // Not in any prior loop 172 } else { 173 old_ctrl = get_ctrl(x); 174 old_loop = get_loop(old_ctrl); // Get prior loop 175 } 176 // New late point must dominate new use 177 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2)); 178 if (new_ctrl == old_ctrl) // Nothing is changed 179 continue; 180 181 IdealLoopTree *new_loop = get_loop(new_ctrl); 182 183 // Don't move x into a loop if its uses are 184 // outside of loop. Otherwise x will be cloned 185 // for each use outside of this loop. 186 IdealLoopTree *use_loop = get_loop(region); 187 if (!new_loop->is_member(use_loop) && 188 (old_loop == NULL || !new_loop->is_member(old_loop))) { 189 // Take early control, later control will be recalculated 190 // during next iteration of loop optimizations. 191 new_ctrl = get_early_ctrl(x); 192 new_loop = get_loop(new_ctrl); 193 } 194 // Set new location 195 set_ctrl(x, new_ctrl); 196 // If changing loop bodies, see if we need to collect into new body 197 if (old_loop != new_loop) { 198 if (old_loop && !old_loop->_child) 199 old_loop->_body.yank(x); 200 if (!new_loop->_child) 201 new_loop->_body.push(x); // Collect body info 202 } 203 } 204 205 return phi; 206 } 207 208 //------------------------------dominated_by------------------------------------ 209 // Replace the dominated test with an obvious true or false. Place it on the 210 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the 211 // live path up to the dominating control. 212 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff, bool flip, bool exclude_loop_predicate ) { 213 if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); } 214 215 // prevdom is the dominating projection of the dominating test. 216 assert( iff->is_If(), "" ); 217 assert(iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd || iff->Opcode() == Op_RangeCheck, "Check this code when new subtype is added"); 218 int pop = prevdom->Opcode(); 219 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" ); 220 if (flip) { 221 if (pop == Op_IfTrue) 222 pop = Op_IfFalse; 223 else 224 pop = Op_IfTrue; 225 } 226 // 'con' is set to true or false to kill the dominated test. 227 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO); 228 set_ctrl(con, C->root()); // Constant gets a new use 229 // Hack the dominated test 230 _igvn.replace_input_of(iff, 1, con); 231 232 // If I dont have a reachable TRUE and FALSE path following the IfNode then 233 // I can assume this path reaches an infinite loop. In this case it's not 234 // important to optimize the data Nodes - either the whole compilation will 235 // be tossed or this path (and all data Nodes) will go dead. 236 if (iff->outcnt() != 2) return; 237 238 // Make control-dependent data Nodes on the live path (path that will remain 239 // once the dominated IF is removed) become control-dependent on the 240 // dominating projection. 241 Node* dp = iff->as_If()->proj_out(pop == Op_IfTrue); 242 243 // Loop predicates may have depending checks which should not 244 // be skipped. For example, range check predicate has two checks 245 // for lower and upper bounds. 246 if (dp == NULL) 247 return; 248 249 ProjNode* dp_proj = dp->as_Proj(); 250 ProjNode* unc_proj = iff->as_If()->proj_out(1 - dp_proj->_con)->as_Proj(); 251 if (exclude_loop_predicate && 252 (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) != NULL || 253 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check) != NULL)) { 254 // If this is a range check (IfNode::is_range_check), do not 255 // reorder because Compile::allow_range_check_smearing might have 256 // changed the check. 257 return; // Let IGVN transformation change control dependence. 258 } 259 260 IdealLoopTree *old_loop = get_loop(dp); 261 262 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { 263 Node* cd = dp->fast_out(i); // Control-dependent node 264 if (cd->depends_only_on_test()) { 265 assert(cd->in(0) == dp, ""); 266 _igvn.replace_input_of(cd, 0, prevdom); 267 set_early_ctrl(cd); 268 IdealLoopTree *new_loop = get_loop(get_ctrl(cd)); 269 if (old_loop != new_loop) { 270 if (!old_loop->_child) old_loop->_body.yank(cd); 271 if (!new_loop->_child) new_loop->_body.push(cd); 272 } 273 --i; 274 --imax; 275 } 276 } 277 } 278 279 //------------------------------has_local_phi_input---------------------------- 280 // Return TRUE if 'n' has Phi inputs from its local block and no other 281 // block-local inputs (all non-local-phi inputs come from earlier blocks) 282 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) { 283 Node *n_ctrl = get_ctrl(n); 284 // See if some inputs come from a Phi in this block, or from before 285 // this block. 286 uint i; 287 for( i = 1; i < n->req(); i++ ) { 288 Node *phi = n->in(i); 289 if( phi->is_Phi() && phi->in(0) == n_ctrl ) 290 break; 291 } 292 if( i >= n->req() ) 293 return NULL; // No Phi inputs; nowhere to clone thru 294 295 // Check for inputs created between 'n' and the Phi input. These 296 // must split as well; they have already been given the chance 297 // (courtesy of a post-order visit) and since they did not we must 298 // recover the 'cost' of splitting them by being very profitable 299 // when splitting 'n'. Since this is unlikely we simply give up. 300 for( i = 1; i < n->req(); i++ ) { 301 Node *m = n->in(i); 302 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) { 303 // We allow the special case of AddP's with no local inputs. 304 // This allows us to split-up address expressions. 305 if (m->is_AddP() && 306 get_ctrl(m->in(2)) != n_ctrl && 307 get_ctrl(m->in(3)) != n_ctrl) { 308 // Move the AddP up to dominating point 309 Node* c = find_non_split_ctrl(idom(n_ctrl)); 310 if (c->Opcode() == Op_Loop && c->as_Loop()->is_strip_mined()) { 311 c->as_Loop()->verify_strip_mined(1); 312 c = c->in(LoopNode::EntryControl); 313 } 314 set_ctrl_and_loop(m, c); 315 continue; 316 } 317 return NULL; 318 } 319 } 320 321 return n_ctrl; 322 } 323 324 //------------------------------remix_address_expressions---------------------- 325 // Rework addressing expressions to get the most loop-invariant stuff 326 // moved out. We'd like to do all associative operators, but it's especially 327 // important (common) to do address expressions. 328 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) { 329 if (!has_ctrl(n)) return NULL; 330 Node *n_ctrl = get_ctrl(n); 331 IdealLoopTree *n_loop = get_loop(n_ctrl); 332 333 // See if 'n' mixes loop-varying and loop-invariant inputs and 334 // itself is loop-varying. 335 336 // Only interested in binary ops (and AddP) 337 if( n->req() < 3 || n->req() > 4 ) return NULL; 338 339 Node *n1_ctrl = get_ctrl(n->in( 1)); 340 Node *n2_ctrl = get_ctrl(n->in( 2)); 341 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3)); 342 IdealLoopTree *n1_loop = get_loop( n1_ctrl ); 343 IdealLoopTree *n2_loop = get_loop( n2_ctrl ); 344 IdealLoopTree *n3_loop = get_loop( n3_ctrl ); 345 346 // Does one of my inputs spin in a tighter loop than self? 347 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) || 348 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) || 349 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) ) 350 return NULL; // Leave well enough alone 351 352 // Is at least one of my inputs loop-invariant? 353 if( n1_loop == n_loop && 354 n2_loop == n_loop && 355 n3_loop == n_loop ) 356 return NULL; // No loop-invariant inputs 357 358 359 int n_op = n->Opcode(); 360 361 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2). 362 if( n_op == Op_LShiftI ) { 363 // Scale is loop invariant 364 Node *scale = n->in(2); 365 Node *scale_ctrl = get_ctrl(scale); 366 IdealLoopTree *scale_loop = get_loop(scale_ctrl ); 367 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) ) 368 return NULL; 369 const TypeInt *scale_t = scale->bottom_type()->isa_int(); 370 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 ) 371 return NULL; // Dont bother with byte/short masking 372 // Add must vary with loop (else shift would be loop-invariant) 373 Node *add = n->in(1); 374 Node *add_ctrl = get_ctrl(add); 375 IdealLoopTree *add_loop = get_loop(add_ctrl); 376 //assert( n_loop == add_loop, "" ); 377 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops 378 379 // Convert I-V into I+ (0-V); same for V-I 380 if( add->Opcode() == Op_SubI && 381 _igvn.type( add->in(1) ) != TypeInt::ZERO ) { 382 Node *zero = _igvn.intcon(0); 383 set_ctrl(zero, C->root()); 384 Node *neg = new SubINode( _igvn.intcon(0), add->in(2) ); 385 register_new_node( neg, get_ctrl(add->in(2) ) ); 386 add = new AddINode( add->in(1), neg ); 387 register_new_node( add, add_ctrl ); 388 } 389 if( add->Opcode() != Op_AddI ) return NULL; 390 // See if one add input is loop invariant 391 Node *add_var = add->in(1); 392 Node *add_var_ctrl = get_ctrl(add_var); 393 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 394 Node *add_invar = add->in(2); 395 Node *add_invar_ctrl = get_ctrl(add_invar); 396 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl ); 397 if( add_var_loop == n_loop ) { 398 } else if( add_invar_loop == n_loop ) { 399 // Swap to find the invariant part 400 add_invar = add_var; 401 add_invar_ctrl = add_var_ctrl; 402 add_invar_loop = add_var_loop; 403 add_var = add->in(2); 404 Node *add_var_ctrl = get_ctrl(add_var); 405 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 406 } else // Else neither input is loop invariant 407 return NULL; 408 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) ) 409 return NULL; // No invariant part of the add? 410 411 // Yes! Reshape address expression! 412 Node *inv_scale = new LShiftINode( add_invar, scale ); 413 Node *inv_scale_ctrl = 414 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ? 415 add_invar_ctrl : scale_ctrl; 416 register_new_node( inv_scale, inv_scale_ctrl ); 417 Node *var_scale = new LShiftINode( add_var, scale ); 418 register_new_node( var_scale, n_ctrl ); 419 Node *var_add = new AddINode( var_scale, inv_scale ); 420 register_new_node( var_add, n_ctrl ); 421 _igvn.replace_node( n, var_add ); 422 return var_add; 423 } 424 425 // Replace (I+V) with (V+I) 426 if( n_op == Op_AddI || 427 n_op == Op_AddL || 428 n_op == Op_AddF || 429 n_op == Op_AddD || 430 n_op == Op_MulI || 431 n_op == Op_MulL || 432 n_op == Op_MulF || 433 n_op == Op_MulD ) { 434 if( n2_loop == n_loop ) { 435 assert( n1_loop != n_loop, "" ); 436 n->swap_edges(1, 2); 437 } 438 } 439 440 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V), 441 // but not if I2 is a constant. 442 if( n_op == Op_AddP ) { 443 if( n2_loop == n_loop && n3_loop != n_loop ) { 444 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) { 445 Node *n22_ctrl = get_ctrl(n->in(2)->in(2)); 446 Node *n23_ctrl = get_ctrl(n->in(2)->in(3)); 447 IdealLoopTree *n22loop = get_loop( n22_ctrl ); 448 IdealLoopTree *n23_loop = get_loop( n23_ctrl ); 449 if( n22loop != n_loop && n22loop->is_member(n_loop) && 450 n23_loop == n_loop ) { 451 Node *add1 = new AddPNode( n->in(1), n->in(2)->in(2), n->in(3) ); 452 // Stuff new AddP in the loop preheader 453 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) ); 454 Node *add2 = new AddPNode( n->in(1), add1, n->in(2)->in(3) ); 455 register_new_node( add2, n_ctrl ); 456 _igvn.replace_node( n, add2 ); 457 return add2; 458 } 459 } 460 } 461 462 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V) 463 if (n2_loop != n_loop && n3_loop == n_loop) { 464 if (n->in(3)->Opcode() == Op_AddX) { 465 Node *V = n->in(3)->in(1); 466 Node *I = n->in(3)->in(2); 467 if (is_member(n_loop,get_ctrl(V))) { 468 } else { 469 Node *tmp = V; V = I; I = tmp; 470 } 471 if (!is_member(n_loop,get_ctrl(I))) { 472 Node *add1 = new AddPNode(n->in(1), n->in(2), I); 473 // Stuff new AddP in the loop preheader 474 register_new_node(add1, n_loop->_head->in(LoopNode::EntryControl)); 475 Node *add2 = new AddPNode(n->in(1), add1, V); 476 register_new_node(add2, n_ctrl); 477 _igvn.replace_node(n, add2); 478 return add2; 479 } 480 } 481 } 482 } 483 484 return NULL; 485 } 486 487 //------------------------------conditional_move------------------------------- 488 // Attempt to replace a Phi with a conditional move. We have some pretty 489 // strict profitability requirements. All Phis at the merge point must 490 // be converted, so we can remove the control flow. We need to limit the 491 // number of c-moves to a small handful. All code that was in the side-arms 492 // of the CFG diamond is now speculatively executed. This code has to be 493 // "cheap enough". We are pretty much limited to CFG diamonds that merge 494 // 1 or 2 items with a total of 1 or 2 ops executed speculatively. 495 Node *PhaseIdealLoop::conditional_move( Node *region ) { 496 497 assert(region->is_Region(), "sanity check"); 498 if (region->req() != 3) return NULL; 499 500 // Check for CFG diamond 501 Node *lp = region->in(1); 502 Node *rp = region->in(2); 503 if (!lp || !rp) return NULL; 504 Node *lp_c = lp->in(0); 505 if (lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If()) return NULL; 506 IfNode *iff = lp_c->as_If(); 507 508 // Check for ops pinned in an arm of the diamond. 509 // Can't remove the control flow in this case 510 if (lp->outcnt() > 1) return NULL; 511 if (rp->outcnt() > 1) return NULL; 512 513 IdealLoopTree* r_loop = get_loop(region); 514 assert(r_loop == get_loop(iff), "sanity"); 515 // Always convert to CMOVE if all results are used only outside this loop. 516 bool used_inside_loop = (r_loop == _ltree_root); 517 518 // Check profitability 519 int cost = 0; 520 int phis = 0; 521 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 522 Node *out = region->fast_out(i); 523 if (!out->is_Phi()) continue; // Ignore other control edges, etc 524 phis++; 525 PhiNode* phi = out->as_Phi(); 526 BasicType bt = phi->type()->basic_type(); 527 switch (bt) { 528 case T_DOUBLE: 529 if (C->use_cmove()) { 530 continue; //TODO: maybe we want to add some cost 531 } 532 case T_FLOAT: { 533 cost += Matcher::float_cmove_cost(); // Could be very expensive 534 break; 535 } 536 case T_LONG: { 537 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's 538 } 539 case T_INT: // These all CMOV fine 540 case T_ADDRESS: { // (RawPtr) 541 cost++; 542 break; 543 } 544 case T_NARROWOOP: // Fall through 545 case T_OBJECT: { // Base oops are OK, but not derived oops 546 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr(); 547 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a 548 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus 549 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we 550 // have a Phi for the base here that we convert to a CMOVE all is well 551 // and good. But if the base is dead, we'll not make a CMOVE. Later 552 // the allocator will have to produce a base by creating a CMOVE of the 553 // relevant bases. This puts the allocator in the business of 554 // manufacturing expensive instructions, generally a bad plan. 555 // Just Say No to Conditionally-Moved Derived Pointers. 556 if (tp && tp->offset() != 0) 557 return NULL; 558 cost++; 559 break; 560 } 561 default: 562 return NULL; // In particular, can't do memory or I/O 563 } 564 // Add in cost any speculative ops 565 for (uint j = 1; j < region->req(); j++) { 566 Node *proj = region->in(j); 567 Node *inp = phi->in(j); 568 if (get_ctrl(inp) == proj) { // Found local op 569 cost++; 570 // Check for a chain of dependent ops; these will all become 571 // speculative in a CMOV. 572 for (uint k = 1; k < inp->req(); k++) 573 if (get_ctrl(inp->in(k)) == proj) 574 cost += ConditionalMoveLimit; // Too much speculative goo 575 } 576 } 577 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode. 578 // This will likely Split-If, a higher-payoff operation. 579 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) { 580 Node* use = phi->fast_out(k); 581 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr()) 582 cost += ConditionalMoveLimit; 583 // Is there a use inside the loop? 584 // Note: check only basic types since CMoveP is pinned. 585 if (!used_inside_loop && is_java_primitive(bt)) { 586 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use); 587 if (r_loop == u_loop || r_loop->is_member(u_loop)) { 588 used_inside_loop = true; 589 } 590 } 591 } 592 }//for 593 Node* bol = iff->in(1); 594 assert(bol->Opcode() == Op_Bool, ""); 595 int cmp_op = bol->in(1)->Opcode(); 596 // It is expensive to generate flags from a float compare. 597 // Avoid duplicated float compare. 598 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL; 599 600 float infrequent_prob = PROB_UNLIKELY_MAG(3); 601 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop. 602 if (used_inside_loop) { 603 if (cost >= ConditionalMoveLimit) return NULL; // Too much goo 604 605 // BlockLayoutByFrequency optimization moves infrequent branch 606 // from hot path. No point in CMOV'ing in such case (110 is used 607 // instead of 100 to take into account not exactness of float value). 608 if (BlockLayoutByFrequency) { 609 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f); 610 } 611 } 612 // Check for highly predictable branch. No point in CMOV'ing if 613 // we are going to predict accurately all the time. 614 if (C->use_cmove() && cmp_op == Op_CmpD) ;//keep going 615 else if (iff->_prob < infrequent_prob || 616 iff->_prob > (1.0f - infrequent_prob)) 617 return NULL; 618 619 // -------------- 620 // Now replace all Phis with CMOV's 621 Node *cmov_ctrl = iff->in(0); 622 uint flip = (lp->Opcode() == Op_IfTrue); 623 while (1) { 624 PhiNode* phi = NULL; 625 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 626 Node *out = region->fast_out(i); 627 if (out->is_Phi()) { 628 phi = out->as_Phi(); 629 break; 630 } 631 } 632 if (phi == NULL) break; 633 if (PrintOpto && VerifyLoopOptimizations) { tty->print_cr("CMOV"); } 634 // Move speculative ops 635 for (uint j = 1; j < region->req(); j++) { 636 Node *proj = region->in(j); 637 Node *inp = phi->in(j); 638 if (get_ctrl(inp) == proj) { // Found local op 639 #ifndef PRODUCT 640 if (PrintOpto && VerifyLoopOptimizations) { 641 tty->print(" speculate: "); 642 inp->dump(); 643 } 644 #endif 645 set_ctrl(inp, cmov_ctrl); 646 } 647 } 648 Node *cmov = CMoveNode::make(cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi)); 649 register_new_node( cmov, cmov_ctrl ); 650 _igvn.replace_node( phi, cmov ); 651 #ifndef PRODUCT 652 if (TraceLoopOpts) { 653 tty->print("CMOV "); 654 r_loop->dump_head(); 655 if (Verbose) { 656 bol->in(1)->dump(1); 657 cmov->dump(1); 658 } 659 } 660 if (VerifyLoopOptimizations) verify(); 661 #endif 662 } 663 664 // The useless CFG diamond will fold up later; see the optimization in 665 // RegionNode::Ideal. 666 _igvn._worklist.push(region); 667 668 return iff->in(1); 669 } 670 671 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) { 672 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 673 Node* u = m->fast_out(i); 674 if (u->is_CFG()) { 675 if (u->Opcode() == Op_NeverBranch) { 676 u = ((NeverBranchNode*)u)->proj_out(0); 677 enqueue_cfg_uses(u, wq); 678 } else { 679 wq.push(u); 680 } 681 } 682 } 683 } 684 685 // Try moving a store out of a loop, right before the loop 686 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) { 687 // Store has to be first in the loop body 688 IdealLoopTree *n_loop = get_loop(n_ctrl); 689 if (n->is_Store() && n_loop != _ltree_root && n_loop->is_loop() && n->in(0) != NULL) { 690 Node* address = n->in(MemNode::Address); 691 Node* value = n->in(MemNode::ValueIn); 692 Node* mem = n->in(MemNode::Memory); 693 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 694 IdealLoopTree* value_loop = get_loop(get_ctrl(value)); 695 696 // - address and value must be loop invariant 697 // - memory must be a memory Phi for the loop 698 // - Store must be the only store on this memory slice in the 699 // loop: if there's another store following this one then value 700 // written at iteration i by the second store could be overwritten 701 // at iteration i+n by the first store: it's not safe to move the 702 // first store out of the loop 703 // - nothing must observe the memory Phi: it guarantees no read 704 // before the store, we are also guaranteed the store post 705 // dominates the loop head (ignoring a possible early 706 // exit). Otherwise there would be extra Phi involved between the 707 // loop's Phi and the store. 708 // - there must be no early exit from the loop before the Store 709 // (such an exit most of the time would be an extra use of the 710 // memory Phi but sometimes is a bottom memory Phi that takes the 711 // store as input). 712 713 if (!n_loop->is_member(address_loop) && 714 !n_loop->is_member(value_loop) && 715 mem->is_Phi() && mem->in(0) == n_loop->_head && 716 mem->outcnt() == 1 && 717 mem->in(LoopNode::LoopBackControl) == n) { 718 719 assert(n_loop->_tail != NULL, "need a tail"); 720 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop"); 721 722 // Verify that there's no early exit of the loop before the store. 723 bool ctrl_ok = false; 724 { 725 // Follow control from loop head until n, we exit the loop or 726 // we reach the tail 727 ResourceMark rm; 728 Unique_Node_List wq; 729 wq.push(n_loop->_head); 730 731 for (uint next = 0; next < wq.size(); ++next) { 732 Node *m = wq.at(next); 733 if (m == n->in(0)) { 734 ctrl_ok = true; 735 continue; 736 } 737 assert(!has_ctrl(m), "should be CFG"); 738 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) { 739 ctrl_ok = false; 740 break; 741 } 742 enqueue_cfg_uses(m, wq); 743 if (wq.size() > 10) { 744 ctrl_ok = false; 745 break; 746 } 747 } 748 } 749 if (ctrl_ok) { 750 // move the Store 751 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem); 752 _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl)); 753 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl)); 754 // Disconnect the phi now. An empty phi can confuse other 755 // optimizations in this pass of loop opts. 756 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl)); 757 n_loop->_body.yank(mem); 758 759 set_ctrl_and_loop(n, n->in(0)); 760 761 return n; 762 } 763 } 764 } 765 return NULL; 766 } 767 768 // Try moving a store out of a loop, right after the loop 769 void PhaseIdealLoop::try_move_store_after_loop(Node* n) { 770 if (n->is_Store() && n->in(0) != NULL) { 771 Node *n_ctrl = get_ctrl(n); 772 IdealLoopTree *n_loop = get_loop(n_ctrl); 773 // Store must be in a loop 774 if (n_loop != _ltree_root && !n_loop->_irreducible) { 775 Node* address = n->in(MemNode::Address); 776 Node* value = n->in(MemNode::ValueIn); 777 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 778 // address must be loop invariant 779 if (!n_loop->is_member(address_loop)) { 780 // Store must be last on this memory slice in the loop and 781 // nothing in the loop must observe it 782 Node* phi = NULL; 783 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 784 Node* u = n->fast_out(i); 785 if (has_ctrl(u)) { // control use? 786 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 787 if (!n_loop->is_member(u_loop)) { 788 continue; 789 } 790 if (u->is_Phi() && u->in(0) == n_loop->_head) { 791 assert(_igvn.type(u) == Type::MEMORY, "bad phi"); 792 // multiple phis on the same slice are possible 793 if (phi != NULL) { 794 return; 795 } 796 phi = u; 797 continue; 798 } 799 } 800 return; 801 } 802 if (phi != NULL) { 803 // Nothing in the loop before the store (next iteration) 804 // must observe the stored value 805 bool mem_ok = true; 806 { 807 ResourceMark rm; 808 Unique_Node_List wq; 809 wq.push(phi); 810 for (uint next = 0; next < wq.size() && mem_ok; ++next) { 811 Node *m = wq.at(next); 812 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) { 813 Node* u = m->fast_out(i); 814 if (u->is_Store() || u->is_Phi()) { 815 if (u != n) { 816 wq.push(u); 817 mem_ok = (wq.size() <= 10); 818 } 819 } else { 820 mem_ok = false; 821 break; 822 } 823 } 824 } 825 } 826 if (mem_ok) { 827 // Move the Store out of the loop creating clones along 828 // all paths out of the loop that observe the stored value 829 _igvn.rehash_node_delayed(phi); 830 IdealLoopTree* outer_loop = n_loop; 831 if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) { 832 assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined"); 833 n_loop->_head->as_Loop()->verify_strip_mined(1); 834 Node* outer = n_loop->_head->as_CountedLoop()->outer_loop(); 835 outer_loop = get_loop(outer); 836 assert(n_loop->_parent == outer_loop, "broken loop tree"); 837 } 838 int count = phi->replace_edge(n, n->in(MemNode::Memory)); 839 assert(count > 0, "inconsistent phi"); 840 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 841 Node* u = n->fast_out(i); 842 Node* c = get_ctrl(u); 843 844 if (u->is_Phi()) { 845 c = u->in(0)->in(u->find_edge(n)); 846 } 847 IdealLoopTree *u_loop = get_loop(c); 848 assert (!n_loop->is_member(u_loop), "only the phi should have been a use in the loop"); 849 while(true) { 850 Node* next_c = find_non_split_ctrl(idom(c)); 851 if (outer_loop->is_member(get_loop(next_c))) { 852 break; 853 } 854 c = next_c; 855 } 856 857 Node* st = n->clone(); 858 st->set_req(0, c); 859 _igvn.register_new_node_with_optimizer(st); 860 861 set_ctrl(st, c); 862 IdealLoopTree* new_loop = get_loop(c); 863 assert(new_loop != n_loop, "should be moved out of loop"); 864 if (new_loop->_child == NULL) new_loop->_body.push(st); 865 866 _igvn.replace_input_of(u, u->find_edge(n), st); 867 --imax; 868 --i; 869 } 870 871 872 assert(n->outcnt() == 0, "all uses should be gone"); 873 _igvn.replace_input_of(n, MemNode::Memory, C->top()); 874 // Disconnect the phi now. An empty phi can confuse other 875 // optimizations in this pass of loop opts.. 876 if (phi->in(LoopNode::LoopBackControl) == phi) { 877 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl)); 878 n_loop->_body.yank(phi); 879 } 880 } 881 } 882 } 883 } 884 } 885 } 886 887 //------------------------------split_if_with_blocks_pre----------------------- 888 // Do the real work in a non-recursive function. Data nodes want to be 889 // cloned in the pre-order so they can feed each other nicely. 890 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) { 891 // Cloning these guys is unlikely to win 892 int n_op = n->Opcode(); 893 if( n_op == Op_MergeMem ) return n; 894 if( n->is_Proj() ) return n; 895 // Do not clone-up CmpFXXX variations, as these are always 896 // followed by a CmpI 897 if( n->is_Cmp() ) return n; 898 // Attempt to use a conditional move instead of a phi/branch 899 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) { 900 Node *cmov = conditional_move( n ); 901 if( cmov ) return cmov; 902 } 903 if( n->is_CFG() || n->is_LoadStore() ) 904 return n; 905 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd 906 n_op == Op_Opaque2 || 907 n_op == Op_Opaque5) { 908 if( !C->major_progress() ) // If chance of no more loop opts... 909 _igvn._worklist.push(n); // maybe we'll remove them 910 return n; 911 } 912 913 if( n->is_Con() ) return n; // No cloning for Con nodes 914 915 Node *n_ctrl = get_ctrl(n); 916 if( !n_ctrl ) return n; // Dead node 917 918 Node* res = try_move_store_before_loop(n, n_ctrl); 919 if (res != NULL) { 920 return n; 921 } 922 923 // Attempt to remix address expressions for loop invariants 924 Node *m = remix_address_expressions( n ); 925 if( m ) return m; 926 927 if (n->is_ConstraintCast()) { 928 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this); 929 // ConstraintCastNode::dominating_cast() uses node control input to determine domination. 930 // Node control inputs don't necessarily agree with loop control info (due to 931 // transformations happened in between), thus additional dominance check is needed 932 // to keep loop info valid. 933 if (dom_cast != NULL && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) { 934 _igvn.replace_node(n, dom_cast); 935 return dom_cast; 936 } 937 } 938 939 // Determine if the Node has inputs from some local Phi. 940 // Returns the block to clone thru. 941 Node *n_blk = has_local_phi_input( n ); 942 if( !n_blk ) return n; 943 944 // Do not clone the trip counter through on a CountedLoop 945 // (messes up the canonical shape). 946 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n; 947 948 // Check for having no control input; not pinned. Allow 949 // dominating control. 950 if (n->in(0)) { 951 Node *dom = idom(n_blk); 952 if (dom_lca(n->in(0), dom) != n->in(0)) { 953 return n; 954 } 955 } 956 // Policy: when is it profitable. You must get more wins than 957 // policy before it is considered profitable. Policy is usually 0, 958 // so 1 win is considered profitable. Big merges will require big 959 // cloning, so get a larger policy. 960 int policy = n_blk->req() >> 2; 961 962 // If the loop is a candidate for range check elimination, 963 // delay splitting through it's phi until a later loop optimization 964 if (n_blk->is_CountedLoop()) { 965 IdealLoopTree *lp = get_loop(n_blk); 966 if (lp && lp->_rce_candidate) { 967 return n; 968 } 969 } 970 971 // Use same limit as split_if_with_blocks_post 972 if( C->live_nodes() > 35000 ) return n; // Method too big 973 974 // Split 'n' through the merge point if it is profitable 975 Node *phi = split_thru_phi( n, n_blk, policy ); 976 if (!phi) return n; 977 978 // Found a Phi to split thru! 979 // Replace 'n' with the new phi 980 _igvn.replace_node( n, phi ); 981 // Moved a load around the loop, 'en-registering' something. 982 if (n_blk->is_Loop() && n->is_Load() && 983 !phi->in(LoopNode::LoopBackControl)->is_Load()) 984 C->set_major_progress(); 985 986 return phi; 987 } 988 989 static bool merge_point_too_heavy(Compile* C, Node* region) { 990 // Bail out if the region and its phis have too many users. 991 int weight = 0; 992 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 993 weight += region->fast_out(i)->outcnt(); 994 } 995 int nodes_left = C->max_node_limit() - C->live_nodes(); 996 if (weight * 8 > nodes_left) { 997 if (PrintOpto) { 998 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight); 999 } 1000 return true; 1001 } else { 1002 return false; 1003 } 1004 } 1005 1006 static bool merge_point_safe(Node* region) { 1007 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode 1008 // having a PhiNode input. This sidesteps the dangerous case where the split 1009 // ConvI2LNode may become TOP if the input Value() does not 1010 // overlap the ConvI2L range, leaving a node which may not dominate its 1011 // uses. 1012 // A better fix for this problem can be found in the BugTraq entry, but 1013 // expediency for Mantis demands this hack. 1014 // 6855164: If the merge point has a FastLockNode with a PhiNode input, we stop 1015 // split_if_with_blocks from splitting a block because we could not move around 1016 // the FastLockNode. 1017 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1018 Node* n = region->fast_out(i); 1019 if (n->is_Phi()) { 1020 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1021 Node* m = n->fast_out(j); 1022 if (m->is_FastLock()) 1023 return false; 1024 #ifdef _LP64 1025 if (m->Opcode() == Op_ConvI2L) 1026 return false; 1027 if (m->is_CastII() && m->isa_CastII()->has_range_check()) { 1028 return false; 1029 } 1030 #endif 1031 } 1032 } 1033 } 1034 return true; 1035 } 1036 1037 1038 //------------------------------place_near_use--------------------------------- 1039 // Place some computation next to use but not inside inner loops. 1040 // For inner loop uses move it to the preheader area. 1041 Node *PhaseIdealLoop::place_near_use( Node *useblock ) const { 1042 IdealLoopTree *u_loop = get_loop( useblock ); 1043 return (u_loop->_irreducible || u_loop->_child) 1044 ? useblock 1045 : u_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl); 1046 } 1047 1048 1049 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) { 1050 if (!n->is_If()) { 1051 return false; 1052 } 1053 if (!n->in(0)->is_Region()) { 1054 return false; 1055 } 1056 Node* region = n->in(0); 1057 Node* dom = idom(region); 1058 if (!dom->is_If() || dom->in(1) != n->in(1)) { 1059 return false; 1060 } 1061 IfNode* dom_if = dom->as_If(); 1062 Node* proj_true = dom_if->proj_out(1); 1063 Node* proj_false = dom_if->proj_out(0); 1064 1065 for (uint i = 1; i < region->req(); i++) { 1066 if (is_dominator(proj_true, region->in(i))) { 1067 continue; 1068 } 1069 if (is_dominator(proj_false, region->in(i))) { 1070 continue; 1071 } 1072 return false; 1073 } 1074 1075 return true; 1076 } 1077 1078 bool PhaseIdealLoop::can_split_if(Node *n_ctrl) { 1079 if (C->live_nodes() > 35000) { 1080 return false; // Method too big 1081 } 1082 1083 // Do not do 'split-if' if irreducible loops are present. 1084 if (_has_irreducible_loops) { 1085 return false; 1086 } 1087 1088 if (merge_point_too_heavy(C, n_ctrl)) { 1089 return false; 1090 } 1091 1092 // Do not do 'split-if' if some paths are dead. First do dead code 1093 // elimination and then see if its still profitable. 1094 for (uint i = 1; i < n_ctrl->req(); i++) { 1095 if (n_ctrl->in(i) == C->top()) { 1096 return false; 1097 } 1098 } 1099 1100 // If trying to do a 'Split-If' at the loop head, it is only 1101 // profitable if the cmp folds up on BOTH paths. Otherwise we 1102 // risk peeling a loop forever. 1103 1104 // CNC - Disabled for now. Requires careful handling of loop 1105 // body selection for the cloned code. Also, make sure we check 1106 // for any input path not being in the same loop as n_ctrl. For 1107 // irreducible loops we cannot check for 'n_ctrl->is_Loop()' 1108 // because the alternative loop entry points won't be converted 1109 // into LoopNodes. 1110 IdealLoopTree *n_loop = get_loop(n_ctrl); 1111 for (uint j = 1; j < n_ctrl->req(); j++) { 1112 if (get_loop(n_ctrl->in(j)) != n_loop) { 1113 return false; 1114 } 1115 } 1116 1117 // Check for safety of the merge point. 1118 if (!merge_point_safe(n_ctrl)) { 1119 return false; 1120 } 1121 1122 return true; 1123 } 1124 1125 //------------------------------split_if_with_blocks_post---------------------- 1126 // Do the real work in a non-recursive function. CFG hackery wants to be 1127 // in the post-order, so it can dirty the I-DOM info and not use the dirtied 1128 // info. 1129 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) { 1130 1131 // Cloning Cmp through Phi's involves the split-if transform. 1132 // FastLock is not used by an If 1133 if (n->is_Cmp() && !n->is_FastLock()) { 1134 Node *n_ctrl = get_ctrl(n); 1135 // Determine if the Node has inputs from some local Phi. 1136 // Returns the block to clone thru. 1137 Node *n_blk = has_local_phi_input(n); 1138 if (n_blk != n_ctrl) { 1139 return; 1140 } 1141 1142 if (!can_split_if(n_ctrl)) { 1143 return; 1144 } 1145 1146 if (n->outcnt() != 1) { 1147 return; // Multiple bool's from 1 compare? 1148 } 1149 Node *bol = n->unique_out(); 1150 assert(bol->is_Bool(), "expect a bool here"); 1151 if (bol->outcnt() != 1) { 1152 return;// Multiple branches from 1 compare? 1153 } 1154 Node *iff = bol->unique_out(); 1155 1156 // Check some safety conditions 1157 if (iff->is_If()) { // Classic split-if? 1158 if (iff->in(0) != n_ctrl) { 1159 return; // Compare must be in same blk as if 1160 } 1161 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE 1162 // Can't split CMove with different control edge. 1163 if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) { 1164 return; 1165 } 1166 if (get_ctrl(iff->in(2)) == n_ctrl || 1167 get_ctrl(iff->in(3)) == n_ctrl) { 1168 return; // Inputs not yet split-up 1169 } 1170 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) { 1171 return; // Loop-invar test gates loop-varying CMOVE 1172 } 1173 } else { 1174 return; // some other kind of node, such as an Allocate 1175 } 1176 1177 // When is split-if profitable? Every 'win' on means some control flow 1178 // goes dead, so it's almost always a win. 1179 int policy = 0; 1180 // Split compare 'n' through the merge point if it is profitable 1181 Node *phi = split_thru_phi( n, n_ctrl, policy); 1182 if (!phi) { 1183 return; 1184 } 1185 1186 // Found a Phi to split thru! 1187 // Replace 'n' with the new phi 1188 _igvn.replace_node(n, phi); 1189 1190 // Now split the bool up thru the phi 1191 Node *bolphi = split_thru_phi(bol, n_ctrl, -1); 1192 guarantee(bolphi != NULL, "null boolean phi node"); 1193 1194 _igvn.replace_node(bol, bolphi); 1195 assert(iff->in(1) == bolphi, ""); 1196 1197 if (bolphi->Value(&_igvn)->singleton()) { 1198 return; 1199 } 1200 1201 // Conditional-move? Must split up now 1202 if (!iff->is_If()) { 1203 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1); 1204 _igvn.replace_node(iff, cmovphi); 1205 return; 1206 } 1207 1208 // Now split the IF 1209 do_split_if(iff); 1210 return; 1211 } 1212 1213 // Two identical ifs back to back can be merged 1214 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) { 1215 Node *n_ctrl = n->in(0); 1216 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1)); 1217 IfNode* dom_if = idom(n_ctrl)->as_If(); 1218 Node* proj_true = dom_if->proj_out(1); 1219 Node* proj_false = dom_if->proj_out(0); 1220 Node* con_true = _igvn.makecon(TypeInt::ONE); 1221 Node* con_false = _igvn.makecon(TypeInt::ZERO); 1222 1223 for (uint i = 1; i < n_ctrl->req(); i++) { 1224 if (is_dominator(proj_true, n_ctrl->in(i))) { 1225 bolphi->init_req(i, con_true); 1226 } else { 1227 assert(is_dominator(proj_false, n_ctrl->in(i)), "bad if"); 1228 bolphi->init_req(i, con_false); 1229 } 1230 } 1231 register_new_node(bolphi, n_ctrl); 1232 _igvn.replace_input_of(n, 1, bolphi); 1233 1234 // Now split the IF 1235 do_split_if(n); 1236 return; 1237 } 1238 1239 // Check for an IF ready to split; one that has its 1240 // condition codes input coming from a Phi at the block start. 1241 int n_op = n->Opcode(); 1242 1243 // Check for an IF being dominated by another IF same test 1244 if (n_op == Op_If || 1245 n_op == Op_RangeCheck) { 1246 Node *bol = n->in(1); 1247 uint max = bol->outcnt(); 1248 // Check for same test used more than once? 1249 if (max > 1 && bol->is_Bool()) { 1250 // Search up IDOMs to see if this IF is dominated. 1251 Node *cutoff = get_ctrl(bol); 1252 1253 // Now search up IDOMs till cutoff, looking for a dominating test 1254 Node *prevdom = n; 1255 Node *dom = idom(prevdom); 1256 while (dom != cutoff) { 1257 if (dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom) { 1258 // Replace the dominated test with an obvious true or false. 1259 // Place it on the IGVN worklist for later cleanup. 1260 C->set_major_progress(); 1261 dominated_by(prevdom, n, false, true); 1262 #ifndef PRODUCT 1263 if( VerifyLoopOptimizations ) verify(); 1264 #endif 1265 return; 1266 } 1267 prevdom = dom; 1268 dom = idom(prevdom); 1269 } 1270 } 1271 } 1272 1273 // See if a shared loop-varying computation has no loop-varying uses. 1274 // Happens if something is only used for JVM state in uncommon trap exits, 1275 // like various versions of induction variable+offset. Clone the 1276 // computation per usage to allow it to sink out of the loop. 1277 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about) 1278 Node *n_ctrl = get_ctrl(n); 1279 IdealLoopTree *n_loop = get_loop(n_ctrl); 1280 if( n_loop != _ltree_root ) { 1281 DUIterator_Fast imax, i = n->fast_outs(imax); 1282 for (; i < imax; i++) { 1283 Node* u = n->fast_out(i); 1284 if( !has_ctrl(u) ) break; // Found control user 1285 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 1286 if( u_loop == n_loop ) break; // Found loop-varying use 1287 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop 1288 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003 1289 } 1290 bool did_break = (i < imax); // Did we break out of the previous loop? 1291 if (!did_break && n->outcnt() > 1) { // All uses in outer loops! 1292 Node *late_load_ctrl = NULL; 1293 if (n->is_Load()) { 1294 // If n is a load, get and save the result from get_late_ctrl(), 1295 // to be later used in calculating the control for n's clones. 1296 clear_dom_lca_tags(); 1297 late_load_ctrl = get_late_ctrl(n, n_ctrl); 1298 } 1299 // If n is a load, and the late control is the same as the current 1300 // control, then the cloning of n is a pointless exercise, because 1301 // GVN will ensure that we end up where we started. 1302 if (!n->is_Load() || late_load_ctrl != n_ctrl) { 1303 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) { 1304 Node *u = n->last_out(j); // Clone private computation per use 1305 _igvn.rehash_node_delayed(u); 1306 Node *x = n->clone(); // Clone computation 1307 Node *x_ctrl = NULL; 1308 if( u->is_Phi() ) { 1309 // Replace all uses of normal nodes. Replace Phi uses 1310 // individually, so the separate Nodes can sink down 1311 // different paths. 1312 uint k = 1; 1313 while( u->in(k) != n ) k++; 1314 u->set_req( k, x ); 1315 // x goes next to Phi input path 1316 x_ctrl = u->in(0)->in(k); 1317 --j; 1318 } else { // Normal use 1319 // Replace all uses 1320 for( uint k = 0; k < u->req(); k++ ) { 1321 if( u->in(k) == n ) { 1322 u->set_req( k, x ); 1323 --j; 1324 } 1325 } 1326 x_ctrl = get_ctrl(u); 1327 } 1328 1329 // Find control for 'x' next to use but not inside inner loops. 1330 // For inner loop uses get the preheader area. 1331 x_ctrl = place_near_use(x_ctrl); 1332 1333 if (n->is_Load()) { 1334 // For loads, add a control edge to a CFG node outside of the loop 1335 // to force them to not combine and return back inside the loop 1336 // during GVN optimization (4641526). 1337 // 1338 // Because we are setting the actual control input, factor in 1339 // the result from get_late_ctrl() so we respect any 1340 // anti-dependences. (6233005). 1341 x_ctrl = dom_lca(late_load_ctrl, x_ctrl); 1342 1343 // Don't allow the control input to be a CFG splitting node. 1344 // Such nodes should only have ProjNodes as outs, e.g. IfNode 1345 // should only have IfTrueNode and IfFalseNode (4985384). 1346 x_ctrl = find_non_split_ctrl(x_ctrl); 1347 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone"); 1348 1349 x->set_req(0, x_ctrl); 1350 } 1351 register_new_node(x, x_ctrl); 1352 1353 // Some institutional knowledge is needed here: 'x' is 1354 // yanked because if the optimizer runs GVN on it all the 1355 // cloned x's will common up and undo this optimization and 1356 // be forced back in the loop. This is annoying because it 1357 // makes +VerifyOpto report false-positives on progress. I 1358 // tried setting control edges on the x's to force them to 1359 // not combine, but the matching gets worried when it tries 1360 // to fold a StoreP and an AddP together (as part of an 1361 // address expression) and the AddP and StoreP have 1362 // different controls. 1363 if (!x->is_Load() && !x->is_DecodeNarrowPtr()) _igvn._worklist.yank(x); 1364 } 1365 _igvn.remove_dead_node(n); 1366 } 1367 } 1368 } 1369 } 1370 1371 try_move_store_after_loop(n); 1372 1373 // Check for Opaque2's who's loop has disappeared - who's input is in the 1374 // same loop nest as their output. Remove 'em, they are no longer useful. 1375 if( n_op == Op_Opaque2 && 1376 n->in(1) != NULL && 1377 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) { 1378 _igvn.replace_node( n, n->in(1) ); 1379 } 1380 } 1381 1382 //------------------------------split_if_with_blocks--------------------------- 1383 // Check for aggressive application of 'split-if' optimization, 1384 // using basic block level info. 1385 void PhaseIdealLoop::split_if_with_blocks( VectorSet &visited, Node_Stack &nstack ) { 1386 Node *n = C->root(); 1387 visited.set(n->_idx); // first, mark node as visited 1388 // Do pre-visit work for root 1389 n = split_if_with_blocks_pre( n ); 1390 uint cnt = n->outcnt(); 1391 uint i = 0; 1392 while (true) { 1393 // Visit all children 1394 if (i < cnt) { 1395 Node* use = n->raw_out(i); 1396 ++i; 1397 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) { 1398 // Now do pre-visit work for this use 1399 use = split_if_with_blocks_pre( use ); 1400 nstack.push(n, i); // Save parent and next use's index. 1401 n = use; // Process all children of current use. 1402 cnt = use->outcnt(); 1403 i = 0; 1404 } 1405 } 1406 else { 1407 // All of n's children have been processed, complete post-processing. 1408 if (cnt != 0 && !n->is_Con()) { 1409 assert(has_node(n), "no dead nodes"); 1410 split_if_with_blocks_post( n ); 1411 } 1412 if (nstack.is_empty()) { 1413 // Finished all nodes on stack. 1414 break; 1415 } 1416 // Get saved parent node and next use's index. Visit the rest of uses. 1417 n = nstack.node(); 1418 cnt = n->outcnt(); 1419 i = nstack.index(); 1420 nstack.pop(); 1421 } 1422 } 1423 } 1424 1425 1426 //============================================================================= 1427 // 1428 // C L O N E A L O O P B O D Y 1429 // 1430 1431 //------------------------------clone_iff-------------------------------------- 1432 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1433 // "Nearly" because all Nodes have been cloned from the original in the loop, 1434 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1435 // through the Phi recursively, and return a Bool. 1436 BoolNode *PhaseIdealLoop::clone_iff( PhiNode *phi, IdealLoopTree *loop ) { 1437 1438 // Convert this Phi into a Phi merging Bools 1439 uint i; 1440 for( i = 1; i < phi->req(); i++ ) { 1441 Node *b = phi->in(i); 1442 if( b->is_Phi() ) { 1443 _igvn.replace_input_of(phi, i, clone_iff( b->as_Phi(), loop )); 1444 } else { 1445 assert( b->is_Bool(), "" ); 1446 } 1447 } 1448 1449 Node *sample_bool = phi->in(1); 1450 Node *sample_cmp = sample_bool->in(1); 1451 1452 // Make Phis to merge the Cmp's inputs. 1453 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP ); 1454 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP ); 1455 for( i = 1; i < phi->req(); i++ ) { 1456 Node *n1 = phi->in(i)->in(1)->in(1); 1457 Node *n2 = phi->in(i)->in(1)->in(2); 1458 phi1->set_req( i, n1 ); 1459 phi2->set_req( i, n2 ); 1460 phi1->set_type( phi1->type()->meet_speculative(n1->bottom_type())); 1461 phi2->set_type( phi2->type()->meet_speculative(n2->bottom_type())); 1462 } 1463 // See if these Phis have been made before. 1464 // Register with optimizer 1465 Node *hit1 = _igvn.hash_find_insert(phi1); 1466 if( hit1 ) { // Hit, toss just made Phi 1467 _igvn.remove_dead_node(phi1); // Remove new phi 1468 assert( hit1->is_Phi(), "" ); 1469 phi1 = (PhiNode*)hit1; // Use existing phi 1470 } else { // Miss 1471 _igvn.register_new_node_with_optimizer(phi1); 1472 } 1473 Node *hit2 = _igvn.hash_find_insert(phi2); 1474 if( hit2 ) { // Hit, toss just made Phi 1475 _igvn.remove_dead_node(phi2); // Remove new phi 1476 assert( hit2->is_Phi(), "" ); 1477 phi2 = (PhiNode*)hit2; // Use existing phi 1478 } else { // Miss 1479 _igvn.register_new_node_with_optimizer(phi2); 1480 } 1481 // Register Phis with loop/block info 1482 set_ctrl(phi1, phi->in(0)); 1483 set_ctrl(phi2, phi->in(0)); 1484 // Make a new Cmp 1485 Node *cmp = sample_cmp->clone(); 1486 cmp->set_req( 1, phi1 ); 1487 cmp->set_req( 2, phi2 ); 1488 _igvn.register_new_node_with_optimizer(cmp); 1489 set_ctrl(cmp, phi->in(0)); 1490 1491 // Make a new Bool 1492 Node *b = sample_bool->clone(); 1493 b->set_req(1,cmp); 1494 _igvn.register_new_node_with_optimizer(b); 1495 set_ctrl(b, phi->in(0)); 1496 1497 assert( b->is_Bool(), "" ); 1498 return (BoolNode*)b; 1499 } 1500 1501 //------------------------------clone_bool------------------------------------- 1502 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1503 // "Nearly" because all Nodes have been cloned from the original in the loop, 1504 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1505 // through the Phi recursively, and return a Bool. 1506 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) { 1507 uint i; 1508 // Convert this Phi into a Phi merging Bools 1509 for( i = 1; i < phi->req(); i++ ) { 1510 Node *b = phi->in(i); 1511 if( b->is_Phi() ) { 1512 _igvn.replace_input_of(phi, i, clone_bool( b->as_Phi(), loop )); 1513 } else { 1514 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" ); 1515 } 1516 } 1517 1518 Node *sample_cmp = phi->in(1); 1519 1520 // Make Phis to merge the Cmp's inputs. 1521 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP ); 1522 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP ); 1523 for( uint j = 1; j < phi->req(); j++ ) { 1524 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP 1525 Node *n1, *n2; 1526 if( cmp_top->is_Cmp() ) { 1527 n1 = cmp_top->in(1); 1528 n2 = cmp_top->in(2); 1529 } else { 1530 n1 = n2 = cmp_top; 1531 } 1532 phi1->set_req( j, n1 ); 1533 phi2->set_req( j, n2 ); 1534 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 1535 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 1536 } 1537 1538 // See if these Phis have been made before. 1539 // Register with optimizer 1540 Node *hit1 = _igvn.hash_find_insert(phi1); 1541 if( hit1 ) { // Hit, toss just made Phi 1542 _igvn.remove_dead_node(phi1); // Remove new phi 1543 assert( hit1->is_Phi(), "" ); 1544 phi1 = (PhiNode*)hit1; // Use existing phi 1545 } else { // Miss 1546 _igvn.register_new_node_with_optimizer(phi1); 1547 } 1548 Node *hit2 = _igvn.hash_find_insert(phi2); 1549 if( hit2 ) { // Hit, toss just made Phi 1550 _igvn.remove_dead_node(phi2); // Remove new phi 1551 assert( hit2->is_Phi(), "" ); 1552 phi2 = (PhiNode*)hit2; // Use existing phi 1553 } else { // Miss 1554 _igvn.register_new_node_with_optimizer(phi2); 1555 } 1556 // Register Phis with loop/block info 1557 set_ctrl(phi1, phi->in(0)); 1558 set_ctrl(phi2, phi->in(0)); 1559 // Make a new Cmp 1560 Node *cmp = sample_cmp->clone(); 1561 cmp->set_req( 1, phi1 ); 1562 cmp->set_req( 2, phi2 ); 1563 _igvn.register_new_node_with_optimizer(cmp); 1564 set_ctrl(cmp, phi->in(0)); 1565 1566 assert( cmp->is_Cmp(), "" ); 1567 return (CmpNode*)cmp; 1568 } 1569 1570 //------------------------------sink_use--------------------------------------- 1571 // If 'use' was in the loop-exit block, it now needs to be sunk 1572 // below the post-loop merge point. 1573 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) { 1574 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) { 1575 set_ctrl(use, post_loop); 1576 for (DUIterator j = use->outs(); use->has_out(j); j++) 1577 sink_use(use->out(j), post_loop); 1578 } 1579 } 1580 1581 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new, 1582 IdealLoopTree* loop, IdealLoopTree* outer_loop, 1583 Node_List*& split_if_set, Node_List*& split_bool_set, 1584 Node_List*& split_cex_set, Node_List& worklist, 1585 uint new_counter, CloneLoopMode mode) { 1586 Node* nnn = old_new[old->_idx]; 1587 // Copy uses to a worklist, so I can munge the def-use info 1588 // with impunity. 1589 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1590 worklist.push(old->fast_out(j)); 1591 1592 while( worklist.size() ) { 1593 Node *use = worklist.pop(); 1594 if (!has_node(use)) continue; // Ignore dead nodes 1595 if (use->in(0) == C->top()) continue; 1596 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1597 // Check for data-use outside of loop - at least one of OLD or USE 1598 // must not be a CFG node. 1599 #ifdef ASSERT 1600 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == NULL) { 1601 Node* cmp = loop->_head->as_CountedLoop()->outer_cmp(); 1602 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint(); 1603 assert(mode == ControlAroundStripMined && (use == cmp || use == sfpt), "missed a node"); 1604 } 1605 #endif 1606 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) { 1607 1608 // If the Data use is an IF, that means we have an IF outside of the 1609 // loop that is switching on a condition that is set inside of the 1610 // loop. Happens if people set a loop-exit flag; then test the flag 1611 // in the loop to break the loop, then test is again outside of the 1612 // loop to determine which way the loop exited. 1613 // Loop predicate If node connects to Bool node through Opaque1 node. 1614 if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use)) { 1615 // Since this code is highly unlikely, we lazily build the worklist 1616 // of such Nodes to go split. 1617 if (!split_if_set) { 1618 ResourceArea *area = Thread::current()->resource_area(); 1619 split_if_set = new Node_List(area); 1620 } 1621 split_if_set->push(use); 1622 } 1623 if (use->is_Bool()) { 1624 if (!split_bool_set) { 1625 ResourceArea *area = Thread::current()->resource_area(); 1626 split_bool_set = new Node_List(area); 1627 } 1628 split_bool_set->push(use); 1629 } 1630 if (use->Opcode() == Op_CreateEx) { 1631 if (!split_cex_set) { 1632 ResourceArea *area = Thread::current()->resource_area(); 1633 split_cex_set = new Node_List(area); 1634 } 1635 split_cex_set->push(use); 1636 } 1637 1638 1639 // Get "block" use is in 1640 uint idx = 0; 1641 while( use->in(idx) != old ) idx++; 1642 Node *prev = use->is_CFG() ? use : get_ctrl(use); 1643 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" ); 1644 Node *cfg = prev->_idx >= new_counter 1645 ? prev->in(2) 1646 : idom(prev); 1647 if( use->is_Phi() ) // Phi use is in prior block 1648 cfg = prev->in(idx); // NOT in block of Phi itself 1649 if (cfg->is_top()) { // Use is dead? 1650 _igvn.replace_input_of(use, idx, C->top()); 1651 continue; 1652 } 1653 1654 while(!outer_loop->is_member(get_loop(cfg))) { 1655 prev = cfg; 1656 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg); 1657 } 1658 // If the use occurs after merging several exits from the loop, then 1659 // old value must have dominated all those exits. Since the same old 1660 // value was used on all those exits we did not need a Phi at this 1661 // merge point. NOW we do need a Phi here. Each loop exit value 1662 // is now merged with the peeled body exit; each exit gets its own 1663 // private Phi and those Phis need to be merged here. 1664 Node *phi; 1665 if( prev->is_Region() ) { 1666 if( idx == 0 ) { // Updating control edge? 1667 phi = prev; // Just use existing control 1668 } else { // Else need a new Phi 1669 phi = PhiNode::make( prev, old ); 1670 // Now recursively fix up the new uses of old! 1671 for( uint i = 1; i < prev->req(); i++ ) { 1672 worklist.push(phi); // Onto worklist once for each 'old' input 1673 } 1674 } 1675 } else { 1676 // Get new RegionNode merging old and new loop exits 1677 prev = old_new[prev->_idx]; 1678 assert( prev, "just made this in step 7" ); 1679 if( idx == 0) { // Updating control edge? 1680 phi = prev; // Just use existing control 1681 } else { // Else need a new Phi 1682 // Make a new Phi merging data values properly 1683 phi = PhiNode::make( prev, old ); 1684 phi->set_req( 1, nnn ); 1685 } 1686 } 1687 // If inserting a new Phi, check for prior hits 1688 if( idx != 0 ) { 1689 Node *hit = _igvn.hash_find_insert(phi); 1690 if( hit == NULL ) { 1691 _igvn.register_new_node_with_optimizer(phi); // Register new phi 1692 } else { // or 1693 // Remove the new phi from the graph and use the hit 1694 _igvn.remove_dead_node(phi); 1695 phi = hit; // Use existing phi 1696 } 1697 set_ctrl(phi, prev); 1698 } 1699 // Make 'use' use the Phi instead of the old loop body exit value 1700 _igvn.replace_input_of(use, idx, phi); 1701 if( use->_idx >= new_counter ) { // If updating new phis 1702 // Not needed for correctness, but prevents a weak assert 1703 // in AddPNode from tripping (when we end up with different 1704 // base & derived Phis that will become the same after 1705 // IGVN does CSE). 1706 Node *hit = _igvn.hash_find_insert(use); 1707 if( hit ) // Go ahead and re-hash for hits. 1708 _igvn.replace_node( use, hit ); 1709 } 1710 1711 // If 'use' was in the loop-exit block, it now needs to be sunk 1712 // below the post-loop merge point. 1713 sink_use( use, prev ); 1714 } 1715 } 1716 } 1717 1718 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop, 1719 IdealLoopTree* outer_loop, int dd, Node_List &old_new, 1720 Node_List& extra_data_nodes) { 1721 if (head->is_strip_mined() && mode != IgnoreStripMined) { 1722 CountedLoopNode* cl = head->as_CountedLoop(); 1723 Node* l = cl->outer_loop(); 1724 Node* tail = cl->outer_loop_tail(); 1725 IfNode* le = cl->outer_loop_end(); 1726 Node* sfpt = cl->outer_safepoint(); 1727 Node* bol = cl->outer_bol(); 1728 Node* cmp = cl->outer_cmp(); 1729 Node* opaq = cl->outer_opaq(); 1730 CountedLoopEndNode* cle = cl->loopexit(); 1731 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop(); 1732 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit(); 1733 Node* cle_out = cle->proj_out(false); 1734 1735 Node* new_sfpt = NULL; 1736 Node* new_cle_out = cle_out->clone(); 1737 old_new.map(cle_out->_idx, new_cle_out); 1738 if (mode == CloneIncludesStripMined) { 1739 // clone outer loop body 1740 Node* new_l = l->clone(); 1741 Node* new_tail = tail->clone(); 1742 IfNode* new_le = le->clone()->as_If(); 1743 new_sfpt = sfpt->clone(); 1744 Node* new_bol = bol->clone(); 1745 Node* new_cmp = cmp->clone(); 1746 Node* new_opaq = opaq->clone(); 1747 1748 set_loop(new_l, outer_loop->_parent); 1749 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd); 1750 set_loop(new_cle_out, outer_loop->_parent); 1751 set_idom(new_cle_out, new_cle, dd); 1752 set_loop(new_sfpt, outer_loop->_parent); 1753 set_idom(new_sfpt, new_cle_out, dd); 1754 set_loop(new_le, outer_loop->_parent); 1755 set_idom(new_le, new_sfpt, dd); 1756 set_loop(new_tail, outer_loop->_parent); 1757 set_idom(new_tail, new_le, dd); 1758 set_ctrl(new_bol, new_sfpt); 1759 set_ctrl(new_cmp, new_sfpt); 1760 set_ctrl(new_cmp, new_sfpt); 1761 set_idom(new_cl, new_l, dd); 1762 1763 old_new.map(l->_idx, new_l); 1764 old_new.map(tail->_idx, new_tail); 1765 old_new.map(le->_idx, new_le); 1766 old_new.map(sfpt->_idx, new_sfpt); 1767 old_new.map(bol->_idx, new_bol); 1768 old_new.map(cmp->_idx, new_cmp); 1769 old_new.map(opaq->_idx, new_opaq); 1770 1771 new_l->set_req(LoopNode::LoopBackControl, new_tail); 1772 new_l->set_req(0, new_l); 1773 new_tail->set_req(0, new_le); 1774 new_le->set_req(0, new_sfpt); 1775 new_sfpt->set_req(0, new_cle_out); 1776 new_le->set_req(1, new_bol); 1777 new_bol->set_req(1, new_cmp); 1778 new_cmp->set_req(2, new_opaq); 1779 new_cmp->set_req(1, new_cle->incr()); 1780 new_cle_out->set_req(0, new_cle); 1781 new_cl->set_req(LoopNode::EntryControl, new_l); 1782 new_opaq->set_req(0, new_sfpt); 1783 1784 _igvn.register_new_node_with_optimizer(new_l); 1785 _igvn.register_new_node_with_optimizer(new_tail); 1786 _igvn.register_new_node_with_optimizer(new_le); 1787 _igvn.register_new_node_with_optimizer(new_bol); 1788 _igvn.register_new_node_with_optimizer(new_cmp); 1789 _igvn.register_new_node_with_optimizer(new_opaq); 1790 } else { 1791 Node *newhead = old_new[loop->_head->_idx]; 1792 newhead->as_Loop()->clear_strip_mined(); 1793 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl)); 1794 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 1795 } 1796 // Look at data node that were assigned a control in the outer 1797 // loop: they are kept in the outer loop by the safepoint so start 1798 // from the safepoint node's inputs. 1799 IdealLoopTree* outer_loop = get_loop(l); 1800 Node_Stack stack(2); 1801 stack.push(sfpt, 1); 1802 uint new_counter = C->unique(); 1803 while (stack.size() > 0) { 1804 Node* n = stack.node(); 1805 uint i = stack.index(); 1806 while (i < n->req() && 1807 (n->in(i) == NULL || 1808 !has_ctrl(n->in(i)) || 1809 get_loop(get_ctrl(n->in(i))) != outer_loop || 1810 (old_new[n->in(i)->_idx] != NULL && old_new[n->in(i)->_idx]->_idx >= new_counter))) { 1811 i++; 1812 } 1813 if (i < n->req()) { 1814 stack.set_index(i+1); 1815 stack.push(n->in(i), 0); 1816 } else { 1817 assert(old_new[n->_idx] == NULL || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet"); 1818 Node* m = n == sfpt ? new_sfpt : n->clone(); 1819 if (m != NULL) { 1820 for (uint i = 0; i < n->req(); i++) { 1821 if (m->in(i) != NULL && old_new[m->in(i)->_idx] != NULL) { 1822 m->set_req(i, old_new[m->in(i)->_idx]); 1823 } 1824 } 1825 } else { 1826 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?"); 1827 } 1828 if (n != sfpt) { 1829 extra_data_nodes.push(n); 1830 _igvn.register_new_node_with_optimizer(m); 1831 assert(get_ctrl(n) == cle_out, "what other control?"); 1832 set_ctrl(m, new_cle_out); 1833 old_new.map(n->_idx, m); 1834 } 1835 stack.pop(); 1836 } 1837 } 1838 if (mode == CloneIncludesStripMined) { 1839 _igvn.register_new_node_with_optimizer(new_sfpt); 1840 _igvn.register_new_node_with_optimizer(new_cle_out); 1841 } 1842 } else { 1843 Node *newhead = old_new[loop->_head->_idx]; 1844 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 1845 } 1846 } 1847 1848 //------------------------------clone_loop------------------------------------- 1849 // 1850 // C L O N E A L O O P B O D Y 1851 // 1852 // This is the basic building block of the loop optimizations. It clones an 1853 // entire loop body. It makes an old_new loop body mapping; with this mapping 1854 // you can find the new-loop equivalent to an old-loop node. All new-loop 1855 // nodes are exactly equal to their old-loop counterparts, all edges are the 1856 // same. All exits from the old-loop now have a RegionNode that merges the 1857 // equivalent new-loop path. This is true even for the normal "loop-exit" 1858 // condition. All uses of loop-invariant old-loop values now come from (one 1859 // or more) Phis that merge their new-loop equivalents. 1860 // 1861 // This operation leaves the graph in an illegal state: there are two valid 1862 // control edges coming from the loop pre-header to both loop bodies. I'll 1863 // definitely have to hack the graph after running this transform. 1864 // 1865 // From this building block I will further edit edges to perform loop peeling 1866 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc. 1867 // 1868 // Parameter side_by_size_idom: 1869 // When side_by_size_idom is NULL, the dominator tree is constructed for 1870 // the clone loop to dominate the original. Used in construction of 1871 // pre-main-post loop sequence. 1872 // When nonnull, the clone and original are side-by-side, both are 1873 // dominated by the side_by_side_idom node. Used in construction of 1874 // unswitched loops. 1875 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd, 1876 CloneLoopMode mode, Node* side_by_side_idom) { 1877 1878 LoopNode* head = loop->_head->as_Loop(); 1879 head->verify_strip_mined(1); 1880 1881 if (C->do_vector_loop() && PrintOpto) { 1882 const char* mname = C->method()->name()->as_quoted_ascii(); 1883 if (mname != NULL) { 1884 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname); 1885 } 1886 } 1887 1888 CloneMap& cm = C->clone_map(); 1889 Dict* dict = cm.dict(); 1890 if (C->do_vector_loop()) { 1891 cm.set_clone_idx(cm.max_gen()+1); 1892 #ifndef PRODUCT 1893 if (PrintOpto) { 1894 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx()); 1895 loop->dump_head(); 1896 } 1897 #endif 1898 } 1899 1900 // Step 1: Clone the loop body. Make the old->new mapping. 1901 uint i; 1902 for( i = 0; i < loop->_body.size(); i++ ) { 1903 Node *old = loop->_body.at(i); 1904 Node *nnn = old->clone(); 1905 old_new.map( old->_idx, nnn ); 1906 if (C->do_vector_loop()) { 1907 cm.verify_insert_and_clone(old, nnn, cm.clone_idx()); 1908 } 1909 _igvn.register_new_node_with_optimizer(nnn); 1910 } 1911 1912 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop; 1913 1914 // Step 2: Fix the edges in the new body. If the old input is outside the 1915 // loop use it. If the old input is INside the loop, use the corresponding 1916 // new node instead. 1917 for( i = 0; i < loop->_body.size(); i++ ) { 1918 Node *old = loop->_body.at(i); 1919 Node *nnn = old_new[old->_idx]; 1920 // Fix CFG/Loop controlling the new node 1921 if (has_ctrl(old)) { 1922 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]); 1923 } else { 1924 set_loop(nnn, outer_loop->_parent); 1925 if (old->outcnt() > 0) { 1926 set_idom( nnn, old_new[idom(old)->_idx], dd ); 1927 } 1928 } 1929 // Correct edges to the new node 1930 for( uint j = 0; j < nnn->req(); j++ ) { 1931 Node *n = nnn->in(j); 1932 if( n ) { 1933 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n ); 1934 if( loop->is_member( old_in_loop ) ) 1935 nnn->set_req(j, old_new[n->_idx]); 1936 } 1937 } 1938 _igvn.hash_find_insert(nnn); 1939 } 1940 1941 ResourceArea *area = Thread::current()->resource_area(); 1942 Node_List extra_data_nodes(area); 1943 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes); 1944 1945 // Step 3: Now fix control uses. Loop varying control uses have already 1946 // been fixed up (as part of all input edges in Step 2). Loop invariant 1947 // control uses must be either an IfFalse or an IfTrue. Make a merge 1948 // point to merge the old and new IfFalse/IfTrue nodes; make the use 1949 // refer to this. 1950 Node_List worklist(area); 1951 uint new_counter = C->unique(); 1952 for( i = 0; i < loop->_body.size(); i++ ) { 1953 Node* old = loop->_body.at(i); 1954 if( !old->is_CFG() ) continue; 1955 1956 // Copy uses to a worklist, so I can munge the def-use info 1957 // with impunity. 1958 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1959 worklist.push(old->fast_out(j)); 1960 1961 while( worklist.size() ) { // Visit all uses 1962 Node *use = worklist.pop(); 1963 if (!has_node(use)) continue; // Ignore dead nodes 1964 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1965 if( !loop->is_member( use_loop ) && use->is_CFG() ) { 1966 // Both OLD and USE are CFG nodes here. 1967 assert( use->is_Proj(), "" ); 1968 Node* nnn = old_new[old->_idx]; 1969 1970 Node* newuse = NULL; 1971 if (head->is_strip_mined() && mode != IgnoreStripMined) { 1972 CountedLoopNode* cl = head->as_CountedLoop(); 1973 CountedLoopEndNode* cle = cl->loopexit(); 1974 Node* cle_out = cle->proj_out(false); 1975 if (use == cle_out) { 1976 IfNode* le = cl->outer_loop_end(); 1977 use = le->proj_out(false); 1978 use_loop = get_loop(use); 1979 if (mode == CloneIncludesStripMined) { 1980 nnn = old_new[le->_idx]; 1981 } else { 1982 newuse = old_new[cle_out->_idx]; 1983 } 1984 } 1985 } 1986 if (newuse == NULL) { 1987 newuse = use->clone(); 1988 } 1989 1990 // Clone the loop exit control projection 1991 if (C->do_vector_loop()) { 1992 cm.verify_insert_and_clone(use, newuse, cm.clone_idx()); 1993 } 1994 newuse->set_req(0,nnn); 1995 _igvn.register_new_node_with_optimizer(newuse); 1996 set_loop(newuse, use_loop); 1997 set_idom(newuse, nnn, dom_depth(nnn) + 1 ); 1998 1999 // We need a Region to merge the exit from the peeled body and the 2000 // exit from the old loop body. 2001 RegionNode *r = new RegionNode(3); 2002 // Map the old use to the new merge point 2003 old_new.map( use->_idx, r ); 2004 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use)); 2005 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" ); 2006 2007 // The original user of 'use' uses 'r' instead. 2008 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) { 2009 Node* useuse = use->last_out(l); 2010 _igvn.rehash_node_delayed(useuse); 2011 uint uses_found = 0; 2012 if( useuse->in(0) == use ) { 2013 useuse->set_req(0, r); 2014 uses_found++; 2015 if( useuse->is_CFG() ) { 2016 assert( dom_depth(useuse) > dd_r, "" ); 2017 set_idom(useuse, r, dom_depth(useuse)); 2018 } 2019 } 2020 for( uint k = 1; k < useuse->req(); k++ ) { 2021 if( useuse->in(k) == use ) { 2022 useuse->set_req(k, r); 2023 uses_found++; 2024 if (useuse->is_Loop() && k == LoopNode::EntryControl) { 2025 assert(dom_depth(useuse) > dd_r , ""); 2026 set_idom(useuse, r, dom_depth(useuse)); 2027 } 2028 } 2029 } 2030 l -= uses_found; // we deleted 1 or more copies of this edge 2031 } 2032 2033 // Now finish up 'r' 2034 r->set_req( 1, newuse ); 2035 r->set_req( 2, use ); 2036 _igvn.register_new_node_with_optimizer(r); 2037 set_loop(r, use_loop); 2038 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r); 2039 } // End of if a loop-exit test 2040 } 2041 } 2042 2043 // Step 4: If loop-invariant use is not control, it must be dominated by a 2044 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region 2045 // there if needed. Make a Phi there merging old and new used values. 2046 Node_List *split_if_set = NULL; 2047 Node_List *split_bool_set = NULL; 2048 Node_List *split_cex_set = NULL; 2049 for( i = 0; i < loop->_body.size(); i++ ) { 2050 Node* old = loop->_body.at(i); 2051 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set, 2052 split_bool_set, split_cex_set, worklist, new_counter, 2053 mode); 2054 } 2055 2056 for (i = 0; i < extra_data_nodes.size(); i++) { 2057 Node* old = extra_data_nodes.at(i); 2058 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set, 2059 split_bool_set, split_cex_set, worklist, new_counter, 2060 mode); 2061 } 2062 2063 // Check for IFs that need splitting/cloning. Happens if an IF outside of 2064 // the loop uses a condition set in the loop. The original IF probably 2065 // takes control from one or more OLD Regions (which in turn get from NEW 2066 // Regions). In any case, there will be a set of Phis for each merge point 2067 // from the IF up to where the original BOOL def exists the loop. 2068 if( split_if_set ) { 2069 while( split_if_set->size() ) { 2070 Node *iff = split_if_set->pop(); 2071 if( iff->in(1)->is_Phi() ) { 2072 BoolNode *b = clone_iff( iff->in(1)->as_Phi(), loop ); 2073 _igvn.replace_input_of(iff, 1, b); 2074 } 2075 } 2076 } 2077 if( split_bool_set ) { 2078 while( split_bool_set->size() ) { 2079 Node *b = split_bool_set->pop(); 2080 Node *phi = b->in(1); 2081 assert( phi->is_Phi(), "" ); 2082 CmpNode *cmp = clone_bool( (PhiNode*)phi, loop ); 2083 _igvn.replace_input_of(b, 1, cmp); 2084 } 2085 } 2086 if( split_cex_set ) { 2087 while( split_cex_set->size() ) { 2088 Node *b = split_cex_set->pop(); 2089 assert( b->in(0)->is_Region(), "" ); 2090 assert( b->in(1)->is_Phi(), "" ); 2091 assert( b->in(0)->in(0) == b->in(1)->in(0), "" ); 2092 split_up( b, b->in(0), NULL ); 2093 } 2094 } 2095 2096 } 2097 2098 2099 //---------------------- stride_of_possible_iv ------------------------------------- 2100 // Looks for an iff/bool/comp with one operand of the compare 2101 // being a cycle involving an add and a phi, 2102 // with an optional truncation (left-shift followed by a right-shift) 2103 // of the add. Returns zero if not an iv. 2104 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) { 2105 Node* trunc1 = NULL; 2106 Node* trunc2 = NULL; 2107 const TypeInt* ttype = NULL; 2108 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) { 2109 return 0; 2110 } 2111 BoolNode* bl = iff->in(1)->as_Bool(); 2112 Node* cmp = bl->in(1); 2113 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) { 2114 return 0; 2115 } 2116 // Must have an invariant operand 2117 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) { 2118 return 0; 2119 } 2120 Node* add2 = NULL; 2121 Node* cmp1 = cmp->in(1); 2122 if (cmp1->is_Phi()) { 2123 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) ))) 2124 Node* phi = cmp1; 2125 for (uint i = 1; i < phi->req(); i++) { 2126 Node* in = phi->in(i); 2127 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in, 2128 &trunc1, &trunc2, &ttype); 2129 if (add && add->in(1) == phi) { 2130 add2 = add->in(2); 2131 break; 2132 } 2133 } 2134 } else { 2135 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) ))) 2136 Node* addtrunc = cmp1; 2137 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc, 2138 &trunc1, &trunc2, &ttype); 2139 if (add && add->in(1)->is_Phi()) { 2140 Node* phi = add->in(1); 2141 for (uint i = 1; i < phi->req(); i++) { 2142 if (phi->in(i) == addtrunc) { 2143 add2 = add->in(2); 2144 break; 2145 } 2146 } 2147 } 2148 } 2149 if (add2 != NULL) { 2150 const TypeInt* add2t = _igvn.type(add2)->is_int(); 2151 if (add2t->is_con()) { 2152 return add2t->get_con(); 2153 } 2154 } 2155 return 0; 2156 } 2157 2158 2159 //---------------------- stay_in_loop ------------------------------------- 2160 // Return the (unique) control output node that's in the loop (if it exists.) 2161 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) { 2162 Node* unique = NULL; 2163 if (!n) return NULL; 2164 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2165 Node* use = n->fast_out(i); 2166 if (!has_ctrl(use) && loop->is_member(get_loop(use))) { 2167 if (unique != NULL) { 2168 return NULL; 2169 } 2170 unique = use; 2171 } 2172 } 2173 return unique; 2174 } 2175 2176 //------------------------------ register_node ------------------------------------- 2177 // Utility to register node "n" with PhaseIdealLoop 2178 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) { 2179 _igvn.register_new_node_with_optimizer(n); 2180 loop->_body.push(n); 2181 if (n->is_CFG()) { 2182 set_loop(n, loop); 2183 set_idom(n, pred, ddepth); 2184 } else { 2185 set_ctrl(n, pred); 2186 } 2187 } 2188 2189 //------------------------------ proj_clone ------------------------------------- 2190 // Utility to create an if-projection 2191 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) { 2192 ProjNode* c = p->clone()->as_Proj(); 2193 c->set_req(0, iff); 2194 return c; 2195 } 2196 2197 //------------------------------ short_circuit_if ------------------------------------- 2198 // Force the iff control output to be the live_proj 2199 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) { 2200 guarantee(live_proj != NULL, "null projection"); 2201 int proj_con = live_proj->_con; 2202 assert(proj_con == 0 || proj_con == 1, "false or true projection"); 2203 Node *con = _igvn.intcon(proj_con); 2204 set_ctrl(con, C->root()); 2205 if (iff) { 2206 iff->set_req(1, con); 2207 } 2208 return con; 2209 } 2210 2211 //------------------------------ insert_if_before_proj ------------------------------------- 2212 // Insert a new if before an if projection (* - new node) 2213 // 2214 // before 2215 // if(test) 2216 // / \ 2217 // v v 2218 // other-proj proj (arg) 2219 // 2220 // after 2221 // if(test) 2222 // / \ 2223 // / v 2224 // | * proj-clone 2225 // v | 2226 // other-proj v 2227 // * new_if(relop(cmp[IU](left,right))) 2228 // / \ 2229 // v v 2230 // * new-proj proj 2231 // (returned) 2232 // 2233 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) { 2234 IfNode* iff = proj->in(0)->as_If(); 2235 IdealLoopTree *loop = get_loop(proj); 2236 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2237 int ddepth = dom_depth(proj); 2238 2239 _igvn.rehash_node_delayed(iff); 2240 _igvn.rehash_node_delayed(proj); 2241 2242 proj->set_req(0, NULL); // temporary disconnect 2243 ProjNode* proj2 = proj_clone(proj, iff); 2244 register_node(proj2, loop, iff, ddepth); 2245 2246 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right); 2247 register_node(cmp, loop, proj2, ddepth); 2248 2249 BoolNode* bol = new BoolNode(cmp, relop); 2250 register_node(bol, loop, proj2, ddepth); 2251 2252 int opcode = iff->Opcode(); 2253 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode"); 2254 IfNode* new_if = (opcode == Op_If) ? new IfNode(proj2, bol, iff->_prob, iff->_fcnt): 2255 new RangeCheckNode(proj2, bol, iff->_prob, iff->_fcnt); 2256 register_node(new_if, loop, proj2, ddepth); 2257 2258 proj->set_req(0, new_if); // reattach 2259 set_idom(proj, new_if, ddepth); 2260 2261 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj(); 2262 guarantee(new_exit != NULL, "null exit node"); 2263 register_node(new_exit, get_loop(other_proj), new_if, ddepth); 2264 2265 return new_exit; 2266 } 2267 2268 //------------------------------ insert_region_before_proj ------------------------------------- 2269 // Insert a region before an if projection (* - new node) 2270 // 2271 // before 2272 // if(test) 2273 // / | 2274 // v | 2275 // proj v 2276 // other-proj 2277 // 2278 // after 2279 // if(test) 2280 // / | 2281 // v | 2282 // * proj-clone v 2283 // | other-proj 2284 // v 2285 // * new-region 2286 // | 2287 // v 2288 // * dum_if 2289 // / \ 2290 // v \ 2291 // * dum-proj v 2292 // proj 2293 // 2294 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) { 2295 IfNode* iff = proj->in(0)->as_If(); 2296 IdealLoopTree *loop = get_loop(proj); 2297 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2298 int ddepth = dom_depth(proj); 2299 2300 _igvn.rehash_node_delayed(iff); 2301 _igvn.rehash_node_delayed(proj); 2302 2303 proj->set_req(0, NULL); // temporary disconnect 2304 ProjNode* proj2 = proj_clone(proj, iff); 2305 register_node(proj2, loop, iff, ddepth); 2306 2307 RegionNode* reg = new RegionNode(2); 2308 reg->set_req(1, proj2); 2309 register_node(reg, loop, iff, ddepth); 2310 2311 IfNode* dum_if = new IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt); 2312 register_node(dum_if, loop, reg, ddepth); 2313 2314 proj->set_req(0, dum_if); // reattach 2315 set_idom(proj, dum_if, ddepth); 2316 2317 ProjNode* dum_proj = proj_clone(other_proj, dum_if); 2318 register_node(dum_proj, loop, dum_if, ddepth); 2319 2320 return reg; 2321 } 2322 2323 //------------------------------ insert_cmpi_loop_exit ------------------------------------- 2324 // Clone a signed compare loop exit from an unsigned compare and 2325 // insert it before the unsigned cmp on the stay-in-loop path. 2326 // All new nodes inserted in the dominator tree between the original 2327 // if and it's projections. The original if test is replaced with 2328 // a constant to force the stay-in-loop path. 2329 // 2330 // This is done to make sure that the original if and it's projections 2331 // still dominate the same set of control nodes, that the ctrl() relation 2332 // from data nodes to them is preserved, and that their loop nesting is 2333 // preserved. 2334 // 2335 // before 2336 // if(i <u limit) unsigned compare loop exit 2337 // / | 2338 // v v 2339 // exit-proj stay-in-loop-proj 2340 // 2341 // after 2342 // if(stay-in-loop-const) original if 2343 // / | 2344 // / v 2345 // / if(i < limit) new signed test 2346 // / / | 2347 // / / v 2348 // / / if(i <u limit) new cloned unsigned test 2349 // / / / | 2350 // v v v | 2351 // region | 2352 // | | 2353 // dum-if | 2354 // / | | 2355 // ether | | 2356 // v v 2357 // exit-proj stay-in-loop-proj 2358 // 2359 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) { 2360 const bool Signed = true; 2361 const bool Unsigned = false; 2362 2363 BoolNode* bol = if_cmpu->in(1)->as_Bool(); 2364 if (bol->_test._test != BoolTest::lt) return NULL; 2365 CmpNode* cmpu = bol->in(1)->as_Cmp(); 2366 if (cmpu->Opcode() != Op_CmpU) return NULL; 2367 int stride = stride_of_possible_iv(if_cmpu); 2368 if (stride == 0) return NULL; 2369 2370 Node* lp_proj = stay_in_loop(if_cmpu, loop); 2371 guarantee(lp_proj != NULL, "null loop node"); 2372 2373 ProjNode* lp_continue = lp_proj->as_Proj(); 2374 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj(); 2375 2376 Node* limit = NULL; 2377 if (stride > 0) { 2378 limit = cmpu->in(2); 2379 } else { 2380 limit = _igvn.makecon(TypeInt::ZERO); 2381 set_ctrl(limit, C->root()); 2382 } 2383 // Create a new region on the exit path 2384 RegionNode* reg = insert_region_before_proj(lp_exit); 2385 guarantee(reg != NULL, "null region node"); 2386 2387 // Clone the if-cmpu-true-false using a signed compare 2388 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge; 2389 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue); 2390 reg->add_req(cmpi_exit); 2391 2392 // Clone the if-cmpu-true-false 2393 BoolTest::mask rel_u = bol->_test._test; 2394 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue); 2395 reg->add_req(cmpu_exit); 2396 2397 // Force original if to stay in loop. 2398 short_circuit_if(if_cmpu, lp_continue); 2399 2400 return cmpi_exit->in(0)->as_If(); 2401 } 2402 2403 //------------------------------ remove_cmpi_loop_exit ------------------------------------- 2404 // Remove a previously inserted signed compare loop exit. 2405 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) { 2406 Node* lp_proj = stay_in_loop(if_cmp, loop); 2407 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI && 2408 stay_in_loop(lp_proj, loop)->is_If() && 2409 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu"); 2410 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO); 2411 set_ctrl(con, C->root()); 2412 if_cmp->set_req(1, con); 2413 } 2414 2415 //------------------------------ scheduled_nodelist ------------------------------------- 2416 // Create a post order schedule of nodes that are in the 2417 // "member" set. The list is returned in "sched". 2418 // The first node in "sched" is the loop head, followed by 2419 // nodes which have no inputs in the "member" set, and then 2420 // followed by the nodes that have an immediate input dependence 2421 // on a node in "sched". 2422 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) { 2423 2424 assert(member.test(loop->_head->_idx), "loop head must be in member set"); 2425 Arena *a = Thread::current()->resource_area(); 2426 VectorSet visited(a); 2427 Node_Stack nstack(a, loop->_body.size()); 2428 2429 Node* n = loop->_head; // top of stack is cached in "n" 2430 uint idx = 0; 2431 visited.set(n->_idx); 2432 2433 // Initially push all with no inputs from within member set 2434 for(uint i = 0; i < loop->_body.size(); i++ ) { 2435 Node *elt = loop->_body.at(i); 2436 if (member.test(elt->_idx)) { 2437 bool found = false; 2438 for (uint j = 0; j < elt->req(); j++) { 2439 Node* def = elt->in(j); 2440 if (def && member.test(def->_idx) && def != elt) { 2441 found = true; 2442 break; 2443 } 2444 } 2445 if (!found && elt != loop->_head) { 2446 nstack.push(n, idx); 2447 n = elt; 2448 assert(!visited.test(n->_idx), "not seen yet"); 2449 visited.set(n->_idx); 2450 } 2451 } 2452 } 2453 2454 // traverse out's that are in the member set 2455 while (true) { 2456 if (idx < n->outcnt()) { 2457 Node* use = n->raw_out(idx); 2458 idx++; 2459 if (!visited.test_set(use->_idx)) { 2460 if (member.test(use->_idx)) { 2461 nstack.push(n, idx); 2462 n = use; 2463 idx = 0; 2464 } 2465 } 2466 } else { 2467 // All outputs processed 2468 sched.push(n); 2469 if (nstack.is_empty()) break; 2470 n = nstack.node(); 2471 idx = nstack.index(); 2472 nstack.pop(); 2473 } 2474 } 2475 } 2476 2477 2478 //------------------------------ has_use_in_set ------------------------------------- 2479 // Has a use in the vector set 2480 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) { 2481 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2482 Node* use = n->fast_out(j); 2483 if (vset.test(use->_idx)) { 2484 return true; 2485 } 2486 } 2487 return false; 2488 } 2489 2490 2491 //------------------------------ has_use_internal_to_set ------------------------------------- 2492 // Has use internal to the vector set (ie. not in a phi at the loop head) 2493 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) { 2494 Node* head = loop->_head; 2495 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2496 Node* use = n->fast_out(j); 2497 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) { 2498 return true; 2499 } 2500 } 2501 return false; 2502 } 2503 2504 2505 //------------------------------ clone_for_use_outside_loop ------------------------------------- 2506 // clone "n" for uses that are outside of loop 2507 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) { 2508 int cloned = 0; 2509 assert(worklist.size() == 0, "should be empty"); 2510 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2511 Node* use = n->fast_out(j); 2512 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) { 2513 worklist.push(use); 2514 } 2515 } 2516 while( worklist.size() ) { 2517 Node *use = worklist.pop(); 2518 if (!has_node(use) || use->in(0) == C->top()) continue; 2519 uint j; 2520 for (j = 0; j < use->req(); j++) { 2521 if (use->in(j) == n) break; 2522 } 2523 assert(j < use->req(), "must be there"); 2524 2525 // clone "n" and insert it between the inputs of "n" and the use outside the loop 2526 Node* n_clone = n->clone(); 2527 _igvn.replace_input_of(use, j, n_clone); 2528 cloned++; 2529 Node* use_c; 2530 if (!use->is_Phi()) { 2531 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0); 2532 } else { 2533 // Use in a phi is considered a use in the associated predecessor block 2534 use_c = use->in(0)->in(j); 2535 } 2536 set_ctrl(n_clone, use_c); 2537 assert(!loop->is_member(get_loop(use_c)), "should be outside loop"); 2538 get_loop(use_c)->_body.push(n_clone); 2539 _igvn.register_new_node_with_optimizer(n_clone); 2540 #if !defined(PRODUCT) 2541 if (TracePartialPeeling) { 2542 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx); 2543 } 2544 #endif 2545 } 2546 return cloned; 2547 } 2548 2549 2550 //------------------------------ clone_for_special_use_inside_loop ------------------------------------- 2551 // clone "n" for special uses that are in the not_peeled region. 2552 // If these def-uses occur in separate blocks, the code generator 2553 // marks the method as not compilable. For example, if a "BoolNode" 2554 // is in a different basic block than the "IfNode" that uses it, then 2555 // the compilation is aborted in the code generator. 2556 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 2557 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) { 2558 if (n->is_Phi() || n->is_Load()) { 2559 return; 2560 } 2561 assert(worklist.size() == 0, "should be empty"); 2562 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2563 Node* use = n->fast_out(j); 2564 if ( not_peel.test(use->_idx) && 2565 (use->is_If() || use->is_CMove() || use->is_Bool()) && 2566 use->in(1) == n) { 2567 worklist.push(use); 2568 } 2569 } 2570 if (worklist.size() > 0) { 2571 // clone "n" and insert it between inputs of "n" and the use 2572 Node* n_clone = n->clone(); 2573 loop->_body.push(n_clone); 2574 _igvn.register_new_node_with_optimizer(n_clone); 2575 set_ctrl(n_clone, get_ctrl(n)); 2576 sink_list.push(n_clone); 2577 not_peel <<= n_clone->_idx; // add n_clone to not_peel set. 2578 #if !defined(PRODUCT) 2579 if (TracePartialPeeling) { 2580 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx); 2581 } 2582 #endif 2583 while( worklist.size() ) { 2584 Node *use = worklist.pop(); 2585 _igvn.rehash_node_delayed(use); 2586 for (uint j = 1; j < use->req(); j++) { 2587 if (use->in(j) == n) { 2588 use->set_req(j, n_clone); 2589 } 2590 } 2591 } 2592 } 2593 } 2594 2595 2596 //------------------------------ insert_phi_for_loop ------------------------------------- 2597 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 2598 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) { 2599 Node *phi = PhiNode::make(lp, back_edge_val); 2600 phi->set_req(LoopNode::EntryControl, lp_entry_val); 2601 // Use existing phi if it already exists 2602 Node *hit = _igvn.hash_find_insert(phi); 2603 if( hit == NULL ) { 2604 _igvn.register_new_node_with_optimizer(phi); 2605 set_ctrl(phi, lp); 2606 } else { 2607 // Remove the new phi from the graph and use the hit 2608 _igvn.remove_dead_node(phi); 2609 phi = hit; 2610 } 2611 _igvn.replace_input_of(use, idx, phi); 2612 } 2613 2614 #ifdef ASSERT 2615 //------------------------------ is_valid_loop_partition ------------------------------------- 2616 // Validate the loop partition sets: peel and not_peel 2617 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, 2618 VectorSet& not_peel ) { 2619 uint i; 2620 // Check that peel_list entries are in the peel set 2621 for (i = 0; i < peel_list.size(); i++) { 2622 if (!peel.test(peel_list.at(i)->_idx)) { 2623 return false; 2624 } 2625 } 2626 // Check at loop members are in one of peel set or not_peel set 2627 for (i = 0; i < loop->_body.size(); i++ ) { 2628 Node *def = loop->_body.at(i); 2629 uint di = def->_idx; 2630 // Check that peel set elements are in peel_list 2631 if (peel.test(di)) { 2632 if (not_peel.test(di)) { 2633 return false; 2634 } 2635 // Must be in peel_list also 2636 bool found = false; 2637 for (uint j = 0; j < peel_list.size(); j++) { 2638 if (peel_list.at(j)->_idx == di) { 2639 found = true; 2640 break; 2641 } 2642 } 2643 if (!found) { 2644 return false; 2645 } 2646 } else if (not_peel.test(di)) { 2647 if (peel.test(di)) { 2648 return false; 2649 } 2650 } else { 2651 return false; 2652 } 2653 } 2654 return true; 2655 } 2656 2657 //------------------------------ is_valid_clone_loop_exit_use ------------------------------------- 2658 // Ensure a use outside of loop is of the right form 2659 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) { 2660 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2661 return (use->is_Phi() && 2662 use_c->is_Region() && use_c->req() == 3 && 2663 (use_c->in(exit_idx)->Opcode() == Op_IfTrue || 2664 use_c->in(exit_idx)->Opcode() == Op_IfFalse || 2665 use_c->in(exit_idx)->Opcode() == Op_JumpProj) && 2666 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) ); 2667 } 2668 2669 //------------------------------ is_valid_clone_loop_form ------------------------------------- 2670 // Ensure that all uses outside of loop are of the right form 2671 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 2672 uint orig_exit_idx, uint clone_exit_idx) { 2673 uint len = peel_list.size(); 2674 for (uint i = 0; i < len; i++) { 2675 Node *def = peel_list.at(i); 2676 2677 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 2678 Node *use = def->fast_out(j); 2679 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2680 if (!loop->is_member(get_loop(use_c))) { 2681 // use is not in the loop, check for correct structure 2682 if (use->in(0) == def) { 2683 // Okay 2684 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) { 2685 return false; 2686 } 2687 } 2688 } 2689 } 2690 return true; 2691 } 2692 #endif 2693 2694 //------------------------------ partial_peel ------------------------------------- 2695 // Partially peel (aka loop rotation) the top portion of a loop (called 2696 // the peel section below) by cloning it and placing one copy just before 2697 // the new loop head and the other copy at the bottom of the new loop. 2698 // 2699 // before after where it came from 2700 // 2701 // stmt1 stmt1 2702 // loop: stmt2 clone 2703 // stmt2 if condA goto exitA clone 2704 // if condA goto exitA new_loop: new 2705 // stmt3 stmt3 clone 2706 // if !condB goto loop if condB goto exitB clone 2707 // exitB: stmt2 orig 2708 // stmt4 if !condA goto new_loop orig 2709 // exitA: goto exitA 2710 // exitB: 2711 // stmt4 2712 // exitA: 2713 // 2714 // Step 1: find the cut point: an exit test on probable 2715 // induction variable. 2716 // Step 2: schedule (with cloning) operations in the peel 2717 // section that can be executed after the cut into 2718 // the section that is not peeled. This may need 2719 // to clone operations into exit blocks. For 2720 // instance, a reference to A[i] in the not-peel 2721 // section and a reference to B[i] in an exit block 2722 // may cause a left-shift of i by 2 to be placed 2723 // in the peel block. This step will clone the left 2724 // shift into the exit block and sink the left shift 2725 // from the peel to the not-peel section. 2726 // Step 3: clone the loop, retarget the control, and insert 2727 // phis for values that are live across the new loop 2728 // head. This is very dependent on the graph structure 2729 // from clone_loop. It creates region nodes for 2730 // exit control and associated phi nodes for values 2731 // flow out of the loop through that exit. The region 2732 // node is dominated by the clone's control projection. 2733 // So the clone's peel section is placed before the 2734 // new loop head, and the clone's not-peel section is 2735 // forms the top part of the new loop. The original 2736 // peel section forms the tail of the new loop. 2737 // Step 4: update the dominator tree and recompute the 2738 // dominator depth. 2739 // 2740 // orig 2741 // 2742 // stmt1 2743 // | 2744 // v 2745 // loop predicate 2746 // | 2747 // v 2748 // loop<----+ 2749 // | | 2750 // stmt2 | 2751 // | | 2752 // v | 2753 // ifA | 2754 // / | | 2755 // v v | 2756 // false true ^ <-- last_peel 2757 // / | | 2758 // / ===|==cut | 2759 // / stmt3 | <-- first_not_peel 2760 // / | | 2761 // | v | 2762 // v ifB | 2763 // exitA: / \ | 2764 // / \ | 2765 // v v | 2766 // false true | 2767 // / \ | 2768 // / ----+ 2769 // | 2770 // v 2771 // exitB: 2772 // stmt4 2773 // 2774 // 2775 // after clone loop 2776 // 2777 // stmt1 2778 // | 2779 // v 2780 // loop predicate 2781 // / \ 2782 // clone / \ orig 2783 // / \ 2784 // / \ 2785 // v v 2786 // +---->loop loop<----+ 2787 // | | | | 2788 // | stmt2 stmt2 | 2789 // | | | | 2790 // | v v | 2791 // | ifA ifA | 2792 // | | \ / | | 2793 // | v v v v | 2794 // ^ true false false true ^ <-- last_peel 2795 // | | ^ \ / | | 2796 // | cut==|== \ \ / ===|==cut | 2797 // | stmt3 \ \ / stmt3 | <-- first_not_peel 2798 // | | dom | | | | 2799 // | v \ 1v v2 v | 2800 // | ifB regionA ifB | 2801 // | / \ | / \ | 2802 // | / \ v / \ | 2803 // | v v exitA: v v | 2804 // | true false false true | 2805 // | / ^ \ / \ | 2806 // +---- \ \ / ----+ 2807 // dom \ / 2808 // \ 1v v2 2809 // regionB 2810 // | 2811 // v 2812 // exitB: 2813 // stmt4 2814 // 2815 // 2816 // after partial peel 2817 // 2818 // stmt1 2819 // | 2820 // v 2821 // loop predicate 2822 // / 2823 // clone / orig 2824 // / TOP 2825 // / \ 2826 // v v 2827 // TOP->loop loop----+ 2828 // | | | 2829 // stmt2 stmt2 | 2830 // | | | 2831 // v v | 2832 // ifA ifA | 2833 // | \ / | | 2834 // v v v v | 2835 // true false false true | <-- last_peel 2836 // | ^ \ / +------|---+ 2837 // +->newloop \ \ / === ==cut | | 2838 // | stmt3 \ \ / TOP | | 2839 // | | dom | | stmt3 | | <-- first_not_peel 2840 // | v \ 1v v2 v | | 2841 // | ifB regionA ifB ^ v 2842 // | / \ | / \ | | 2843 // | / \ v / \ | | 2844 // | v v exitA: v v | | 2845 // | true false false true | | 2846 // | / ^ \ / \ | | 2847 // | | \ \ / v | | 2848 // | | dom \ / TOP | | 2849 // | | \ 1v v2 | | 2850 // ^ v regionB | | 2851 // | | | | | 2852 // | | v ^ v 2853 // | | exitB: | | 2854 // | | stmt4 | | 2855 // | +------------>-----------------+ | 2856 // | | 2857 // +-----------------<---------------------+ 2858 // 2859 // 2860 // final graph 2861 // 2862 // stmt1 2863 // | 2864 // v 2865 // loop predicate 2866 // | 2867 // v 2868 // stmt2 clone 2869 // | 2870 // v 2871 // ........> ifA clone 2872 // : / | 2873 // dom / | 2874 // : v v 2875 // : false true 2876 // : | | 2877 // : | v 2878 // : | newloop<-----+ 2879 // : | | | 2880 // : | stmt3 clone | 2881 // : | | | 2882 // : | v | 2883 // : | ifB | 2884 // : | / \ | 2885 // : | v v | 2886 // : | false true | 2887 // : | | | | 2888 // : | v stmt2 | 2889 // : | exitB: | | 2890 // : | stmt4 v | 2891 // : | ifA orig | 2892 // : | / \ | 2893 // : | / \ | 2894 // : | v v | 2895 // : | false true | 2896 // : | / \ | 2897 // : v v -----+ 2898 // RegionA 2899 // | 2900 // v 2901 // exitA 2902 // 2903 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) { 2904 2905 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only"); 2906 if (!loop->_head->is_Loop()) { 2907 return false; } 2908 2909 LoopNode *head = loop->_head->as_Loop(); 2910 2911 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) { 2912 return false; 2913 } 2914 2915 // Check for complex exit control 2916 for(uint ii = 0; ii < loop->_body.size(); ii++ ) { 2917 Node *n = loop->_body.at(ii); 2918 int opc = n->Opcode(); 2919 if (n->is_Call() || 2920 opc == Op_Catch || 2921 opc == Op_CatchProj || 2922 opc == Op_Jump || 2923 opc == Op_JumpProj) { 2924 #if !defined(PRODUCT) 2925 if (TracePartialPeeling) { 2926 tty->print_cr("\nExit control too complex: lp: %d", head->_idx); 2927 } 2928 #endif 2929 return false; 2930 } 2931 } 2932 2933 int dd = dom_depth(head); 2934 2935 // Step 1: find cut point 2936 2937 // Walk up dominators to loop head looking for first loop exit 2938 // which is executed on every path thru loop. 2939 IfNode *peel_if = NULL; 2940 IfNode *peel_if_cmpu = NULL; 2941 2942 Node *iff = loop->tail(); 2943 while( iff != head ) { 2944 if( iff->is_If() ) { 2945 Node *ctrl = get_ctrl(iff->in(1)); 2946 if (ctrl->is_top()) return false; // Dead test on live IF. 2947 // If loop-varying exit-test, check for induction variable 2948 if( loop->is_member(get_loop(ctrl)) && 2949 loop->is_loop_exit(iff) && 2950 is_possible_iv_test(iff)) { 2951 Node* cmp = iff->in(1)->in(1); 2952 if (cmp->Opcode() == Op_CmpI) { 2953 peel_if = iff->as_If(); 2954 } else { 2955 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU"); 2956 peel_if_cmpu = iff->as_If(); 2957 } 2958 } 2959 } 2960 iff = idom(iff); 2961 } 2962 // Prefer signed compare over unsigned compare. 2963 IfNode* new_peel_if = NULL; 2964 if (peel_if == NULL) { 2965 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) { 2966 return false; // No peel point found 2967 } 2968 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop); 2969 if (new_peel_if == NULL) { 2970 return false; // No peel point found 2971 } 2972 peel_if = new_peel_if; 2973 } 2974 Node* last_peel = stay_in_loop(peel_if, loop); 2975 Node* first_not_peeled = stay_in_loop(last_peel, loop); 2976 if (first_not_peeled == NULL || first_not_peeled == head) { 2977 return false; 2978 } 2979 2980 #if !defined(PRODUCT) 2981 if (TraceLoopOpts) { 2982 tty->print("PartialPeel "); 2983 loop->dump_head(); 2984 } 2985 2986 if (TracePartialPeeling) { 2987 tty->print_cr("before partial peel one iteration"); 2988 Node_List wl; 2989 Node* t = head->in(2); 2990 while (true) { 2991 wl.push(t); 2992 if (t == head) break; 2993 t = idom(t); 2994 } 2995 while (wl.size() > 0) { 2996 Node* tt = wl.pop(); 2997 tt->dump(); 2998 if (tt == last_peel) tty->print_cr("-- cut --"); 2999 } 3000 } 3001 #endif 3002 ResourceArea *area = Thread::current()->resource_area(); 3003 VectorSet peel(area); 3004 VectorSet not_peel(area); 3005 Node_List peel_list(area); 3006 Node_List worklist(area); 3007 Node_List sink_list(area); 3008 3009 // Set of cfg nodes to peel are those that are executable from 3010 // the head through last_peel. 3011 assert(worklist.size() == 0, "should be empty"); 3012 worklist.push(head); 3013 peel.set(head->_idx); 3014 while (worklist.size() > 0) { 3015 Node *n = worklist.pop(); 3016 if (n != last_peel) { 3017 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3018 Node* use = n->fast_out(j); 3019 if (use->is_CFG() && 3020 loop->is_member(get_loop(use)) && 3021 !peel.test_set(use->_idx)) { 3022 worklist.push(use); 3023 } 3024 } 3025 } 3026 } 3027 3028 // Set of non-cfg nodes to peel are those that are control 3029 // dependent on the cfg nodes. 3030 uint i; 3031 for(i = 0; i < loop->_body.size(); i++ ) { 3032 Node *n = loop->_body.at(i); 3033 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n; 3034 if (peel.test(n_c->_idx)) { 3035 peel.set(n->_idx); 3036 } else { 3037 not_peel.set(n->_idx); 3038 } 3039 } 3040 3041 // Step 2: move operations from the peeled section down into the 3042 // not-peeled section 3043 3044 // Get a post order schedule of nodes in the peel region 3045 // Result in right-most operand. 3046 scheduled_nodelist(loop, peel, peel_list ); 3047 3048 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 3049 3050 // For future check for too many new phis 3051 uint old_phi_cnt = 0; 3052 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 3053 Node* use = head->fast_out(j); 3054 if (use->is_Phi()) old_phi_cnt++; 3055 } 3056 3057 #if !defined(PRODUCT) 3058 if (TracePartialPeeling) { 3059 tty->print_cr("\npeeled list"); 3060 } 3061 #endif 3062 3063 // Evacuate nodes in peel region into the not_peeled region if possible 3064 uint new_phi_cnt = 0; 3065 uint cloned_for_outside_use = 0; 3066 for (i = 0; i < peel_list.size();) { 3067 Node* n = peel_list.at(i); 3068 #if !defined(PRODUCT) 3069 if (TracePartialPeeling) n->dump(); 3070 #endif 3071 bool incr = true; 3072 if ( !n->is_CFG() ) { 3073 3074 if ( has_use_in_set(n, not_peel) ) { 3075 3076 // If not used internal to the peeled region, 3077 // move "n" from peeled to not_peeled region. 3078 3079 if ( !has_use_internal_to_set(n, peel, loop) ) { 3080 3081 // if not pinned and not a load (which maybe anti-dependent on a store) 3082 // and not a CMove (Matcher expects only bool->cmove). 3083 if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) { 3084 cloned_for_outside_use += clone_for_use_outside_loop( loop, n, worklist ); 3085 sink_list.push(n); 3086 peel >>= n->_idx; // delete n from peel set. 3087 not_peel <<= n->_idx; // add n to not_peel set. 3088 peel_list.remove(i); 3089 incr = false; 3090 #if !defined(PRODUCT) 3091 if (TracePartialPeeling) { 3092 tty->print_cr("sink to not_peeled region: %d newbb: %d", 3093 n->_idx, get_ctrl(n)->_idx); 3094 } 3095 #endif 3096 } 3097 } else { 3098 // Otherwise check for special def-use cases that span 3099 // the peel/not_peel boundary such as bool->if 3100 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist ); 3101 new_phi_cnt++; 3102 } 3103 } 3104 } 3105 if (incr) i++; 3106 } 3107 3108 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) { 3109 #if !defined(PRODUCT) 3110 if (TracePartialPeeling) { 3111 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c", 3112 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F'); 3113 } 3114 #endif 3115 if (new_peel_if != NULL) { 3116 remove_cmpi_loop_exit(new_peel_if, loop); 3117 } 3118 // Inhibit more partial peeling on this loop 3119 assert(!head->is_partial_peel_loop(), "not partial peeled"); 3120 head->mark_partial_peel_failed(); 3121 if (cloned_for_outside_use > 0) { 3122 // Terminate this round of loop opts because 3123 // the graph outside this loop was changed. 3124 C->set_major_progress(); 3125 return true; 3126 } 3127 return false; 3128 } 3129 3130 // Step 3: clone loop, retarget control, and insert new phis 3131 3132 // Create new loop head for new phis and to hang 3133 // the nodes being moved (sinked) from the peel region. 3134 LoopNode* new_head = new LoopNode(last_peel, last_peel); 3135 new_head->set_unswitch_count(head->unswitch_count()); // Preserve 3136 _igvn.register_new_node_with_optimizer(new_head); 3137 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled"); 3138 _igvn.replace_input_of(first_not_peeled, 0, new_head); 3139 set_loop(new_head, loop); 3140 loop->_body.push(new_head); 3141 not_peel.set(new_head->_idx); 3142 set_idom(new_head, last_peel, dom_depth(first_not_peeled)); 3143 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled)); 3144 3145 while (sink_list.size() > 0) { 3146 Node* n = sink_list.pop(); 3147 set_ctrl(n, new_head); 3148 } 3149 3150 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 3151 3152 clone_loop(loop, old_new, dd, IgnoreStripMined); 3153 3154 const uint clone_exit_idx = 1; 3155 const uint orig_exit_idx = 2; 3156 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop"); 3157 3158 Node* head_clone = old_new[head->_idx]; 3159 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop(); 3160 Node* orig_tail_clone = head_clone->in(2); 3161 3162 // Add phi if "def" node is in peel set and "use" is not 3163 3164 for(i = 0; i < peel_list.size(); i++ ) { 3165 Node *def = peel_list.at(i); 3166 if (!def->is_CFG()) { 3167 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 3168 Node *use = def->fast_out(j); 3169 if (has_node(use) && use->in(0) != C->top() && 3170 (!peel.test(use->_idx) || 3171 (use->is_Phi() && use->in(0) == head)) ) { 3172 worklist.push(use); 3173 } 3174 } 3175 while( worklist.size() ) { 3176 Node *use = worklist.pop(); 3177 for (uint j = 1; j < use->req(); j++) { 3178 Node* n = use->in(j); 3179 if (n == def) { 3180 3181 // "def" is in peel set, "use" is not in peel set 3182 // or "use" is in the entry boundary (a phi) of the peel set 3183 3184 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use; 3185 3186 if ( loop->is_member(get_loop( use_c )) ) { 3187 // use is in loop 3188 if (old_new[use->_idx] != NULL) { // null for dead code 3189 Node* use_clone = old_new[use->_idx]; 3190 _igvn.replace_input_of(use, j, C->top()); 3191 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone ); 3192 } 3193 } else { 3194 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format"); 3195 // use is not in the loop, check if the live range includes the cut 3196 Node* lp_if = use_c->in(orig_exit_idx)->in(0); 3197 if (not_peel.test(lp_if->_idx)) { 3198 assert(j == orig_exit_idx, "use from original loop"); 3199 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone ); 3200 } 3201 } 3202 } 3203 } 3204 } 3205 } 3206 } 3207 3208 // Step 3b: retarget control 3209 3210 // Redirect control to the new loop head if a cloned node in 3211 // the not_peeled region has control that points into the peeled region. 3212 // This necessary because the cloned peeled region will be outside 3213 // the loop. 3214 // from to 3215 // cloned-peeled <---+ 3216 // new_head_clone: | <--+ 3217 // cloned-not_peeled in(0) in(0) 3218 // orig-peeled 3219 3220 for(i = 0; i < loop->_body.size(); i++ ) { 3221 Node *n = loop->_body.at(i); 3222 if (!n->is_CFG() && n->in(0) != NULL && 3223 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) { 3224 Node* n_clone = old_new[n->_idx]; 3225 _igvn.replace_input_of(n_clone, 0, new_head_clone); 3226 } 3227 } 3228 3229 // Backedge of the surviving new_head (the clone) is original last_peel 3230 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel); 3231 3232 // Cut first node in original not_peel set 3233 _igvn.rehash_node_delayed(new_head); // Multiple edge updates: 3234 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of 3235 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls 3236 3237 // Copy head_clone back-branch info to original head 3238 // and remove original head's loop entry and 3239 // clone head's back-branch 3240 _igvn.rehash_node_delayed(head); // Multiple edge updates 3241 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl)); 3242 head->set_req(LoopNode::LoopBackControl, C->top()); 3243 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top()); 3244 3245 // Similarly modify the phis 3246 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) { 3247 Node* use = head->fast_out(k); 3248 if (use->is_Phi() && use->outcnt() > 0) { 3249 Node* use_clone = old_new[use->_idx]; 3250 _igvn.rehash_node_delayed(use); // Multiple edge updates 3251 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl)); 3252 use->set_req(LoopNode::LoopBackControl, C->top()); 3253 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top()); 3254 } 3255 } 3256 3257 // Step 4: update dominator tree and dominator depth 3258 3259 set_idom(head, orig_tail_clone, dd); 3260 recompute_dom_depth(); 3261 3262 // Inhibit more partial peeling on this loop 3263 new_head_clone->set_partial_peel_loop(); 3264 C->set_major_progress(); 3265 loop->record_for_igvn(); 3266 3267 #if !defined(PRODUCT) 3268 if (TracePartialPeeling) { 3269 tty->print_cr("\nafter partial peel one iteration"); 3270 Node_List wl(area); 3271 Node* t = last_peel; 3272 while (true) { 3273 wl.push(t); 3274 if (t == head_clone) break; 3275 t = idom(t); 3276 } 3277 while (wl.size() > 0) { 3278 Node* tt = wl.pop(); 3279 if (tt == head) tty->print_cr("orig head"); 3280 else if (tt == new_head_clone) tty->print_cr("new head"); 3281 else if (tt == head_clone) tty->print_cr("clone head"); 3282 tt->dump(); 3283 } 3284 } 3285 #endif 3286 return true; 3287 } 3288 3289 //------------------------------reorg_offsets---------------------------------- 3290 // Reorganize offset computations to lower register pressure. Mostly 3291 // prevent loop-fallout uses of the pre-incremented trip counter (which are 3292 // then alive with the post-incremented trip counter forcing an extra 3293 // register move) 3294 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) { 3295 // Perform it only for canonical counted loops. 3296 // Loop's shape could be messed up by iteration_split_impl. 3297 if (!loop->_head->is_CountedLoop()) 3298 return; 3299 if (!loop->_head->as_Loop()->is_valid_counted_loop()) 3300 return; 3301 3302 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 3303 CountedLoopEndNode *cle = cl->loopexit(); 3304 Node *exit = cle->proj_out(false); 3305 Node *phi = cl->phi(); 3306 3307 // Check for the special case of folks using the pre-incremented 3308 // trip-counter on the fall-out path (forces the pre-incremented 3309 // and post-incremented trip counter to be live at the same time). 3310 // Fix this by adjusting to use the post-increment trip counter. 3311 3312 bool progress = true; 3313 while (progress) { 3314 progress = false; 3315 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) { 3316 Node* use = phi->fast_out(i); // User of trip-counter 3317 if (!has_ctrl(use)) continue; 3318 Node *u_ctrl = get_ctrl(use); 3319 if (use->is_Phi()) { 3320 u_ctrl = NULL; 3321 for (uint j = 1; j < use->req(); j++) 3322 if (use->in(j) == phi) 3323 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j)); 3324 } 3325 IdealLoopTree *u_loop = get_loop(u_ctrl); 3326 // Look for loop-invariant use 3327 if (u_loop == loop) continue; 3328 if (loop->is_member(u_loop)) continue; 3329 // Check that use is live out the bottom. Assuming the trip-counter 3330 // update is right at the bottom, uses of of the loop middle are ok. 3331 if (dom_lca(exit, u_ctrl) != exit) continue; 3332 // Hit! Refactor use to use the post-incremented tripcounter. 3333 // Compute a post-increment tripcounter. 3334 Node *opaq = new Opaque2Node( C, cle->incr() ); 3335 register_new_node(opaq, exit); 3336 Node *neg_stride = _igvn.intcon(-cle->stride_con()); 3337 set_ctrl(neg_stride, C->root()); 3338 Node *post = new AddINode( opaq, neg_stride); 3339 register_new_node(post, exit); 3340 _igvn.rehash_node_delayed(use); 3341 for (uint j = 1; j < use->req(); j++) { 3342 if (use->in(j) == phi) 3343 use->set_req(j, post); 3344 } 3345 // Since DU info changed, rerun loop 3346 progress = true; 3347 break; 3348 } 3349 } 3350 3351 }