1 /* 2 * Copyright (c) 1999, 2018, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "gc/shared/barrierSet.hpp" 27 #include "gc/shared/c2/barrierSetC2.hpp" 28 #include "memory/allocation.inline.hpp" 29 #include "memory/resourceArea.hpp" 30 #include "opto/addnode.hpp" 31 #include "opto/callnode.hpp" 32 #include "opto/castnode.hpp" 33 #include "opto/connode.hpp" 34 #include "opto/castnode.hpp" 35 #include "opto/divnode.hpp" 36 #include "opto/loopnode.hpp" 37 #include "opto/matcher.hpp" 38 #include "opto/mulnode.hpp" 39 #include "opto/movenode.hpp" 40 #include "opto/opaquenode.hpp" 41 #include "opto/rootnode.hpp" 42 #include "opto/subnode.hpp" 43 44 //============================================================================= 45 //------------------------------split_thru_phi--------------------------------- 46 // Split Node 'n' through merge point if there is enough win. 47 Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) { 48 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) { 49 // ConvI2L may have type information on it which is unsafe to push up 50 // so disable this for now 51 return NULL; 52 } 53 54 // Splitting range check CastIIs through a loop induction Phi can 55 // cause new Phis to be created that are left unrelated to the loop 56 // induction Phi and prevent optimizations (vectorization) 57 if (n->Opcode() == Op_CastII && n->as_CastII()->has_range_check() && 58 region->is_CountedLoop() && n->in(1) == region->as_CountedLoop()->phi()) { 59 return NULL; 60 } 61 62 int wins = 0; 63 assert(!n->is_CFG(), ""); 64 assert(region->is_Region(), ""); 65 66 const Type* type = n->bottom_type(); 67 const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr(); 68 Node *phi; 69 if (t_oop != NULL && t_oop->is_known_instance_field()) { 70 int iid = t_oop->instance_id(); 71 int index = C->get_alias_index(t_oop); 72 int offset = t_oop->offset(); 73 phi = new PhiNode(region, type, NULL, iid, index, offset); 74 } else { 75 phi = PhiNode::make_blank(region, n); 76 } 77 uint old_unique = C->unique(); 78 for (uint i = 1; i < region->req(); i++) { 79 Node *x; 80 Node* the_clone = NULL; 81 if (region->in(i) == C->top()) { 82 x = C->top(); // Dead path? Use a dead data op 83 } else { 84 x = n->clone(); // Else clone up the data op 85 the_clone = x; // Remember for possible deletion. 86 // Alter data node to use pre-phi inputs 87 if (n->in(0) == region) 88 x->set_req( 0, region->in(i) ); 89 for (uint j = 1; j < n->req(); j++) { 90 Node *in = n->in(j); 91 if (in->is_Phi() && in->in(0) == region) 92 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone 93 } 94 } 95 // Check for a 'win' on some paths 96 const Type *t = x->Value(&_igvn); 97 98 bool singleton = t->singleton(); 99 100 // A TOP singleton indicates that there are no possible values incoming 101 // along a particular edge. In most cases, this is OK, and the Phi will 102 // be eliminated later in an Ideal call. However, we can't allow this to 103 // happen if the singleton occurs on loop entry, as the elimination of 104 // the PhiNode may cause the resulting node to migrate back to a previous 105 // loop iteration. 106 if (singleton && t == Type::TOP) { 107 // Is_Loop() == false does not confirm the absence of a loop (e.g., an 108 // irreducible loop may not be indicated by an affirmative is_Loop()); 109 // therefore, the only top we can split thru a phi is on a backedge of 110 // a loop. 111 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 112 } 113 114 if (singleton) { 115 wins++; 116 x = ((PhaseGVN&)_igvn).makecon(t); 117 } else { 118 // We now call Identity to try to simplify the cloned node. 119 // Note that some Identity methods call phase->type(this). 120 // Make sure that the type array is big enough for 121 // our new node, even though we may throw the node away. 122 // (Note: This tweaking with igvn only works because x is a new node.) 123 _igvn.set_type(x, t); 124 // If x is a TypeNode, capture any more-precise type permanently into Node 125 // otherwise it will be not updated during igvn->transform since 126 // igvn->type(x) is set to x->Value() already. 127 x->raise_bottom_type(t); 128 Node *y = x->Identity(&_igvn); 129 if (y != x) { 130 wins++; 131 x = y; 132 } else { 133 y = _igvn.hash_find(x); 134 if (y) { 135 wins++; 136 x = y; 137 } else { 138 // Else x is a new node we are keeping 139 // We do not need register_new_node_with_optimizer 140 // because set_type has already been called. 141 _igvn._worklist.push(x); 142 } 143 } 144 } 145 if (x != the_clone && the_clone != NULL) 146 _igvn.remove_dead_node(the_clone); 147 phi->set_req( i, x ); 148 } 149 // Too few wins? 150 if (wins <= policy) { 151 _igvn.remove_dead_node(phi); 152 return NULL; 153 } 154 155 // Record Phi 156 register_new_node( phi, region ); 157 158 for (uint i2 = 1; i2 < phi->req(); i2++) { 159 Node *x = phi->in(i2); 160 // If we commoned up the cloned 'x' with another existing Node, 161 // the existing Node picks up a new use. We need to make the 162 // existing Node occur higher up so it dominates its uses. 163 Node *old_ctrl; 164 IdealLoopTree *old_loop; 165 166 if (x->is_Con()) { 167 // Constant's control is always root. 168 set_ctrl(x, C->root()); 169 continue; 170 } 171 // The occasional new node 172 if (x->_idx >= old_unique) { // Found a new, unplaced node? 173 old_ctrl = NULL; 174 old_loop = NULL; // Not in any prior loop 175 } else { 176 old_ctrl = get_ctrl(x); 177 old_loop = get_loop(old_ctrl); // Get prior loop 178 } 179 // New late point must dominate new use 180 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2)); 181 if (new_ctrl == old_ctrl) // Nothing is changed 182 continue; 183 184 IdealLoopTree *new_loop = get_loop(new_ctrl); 185 186 // Don't move x into a loop if its uses are 187 // outside of loop. Otherwise x will be cloned 188 // for each use outside of this loop. 189 IdealLoopTree *use_loop = get_loop(region); 190 if (!new_loop->is_member(use_loop) && 191 (old_loop == NULL || !new_loop->is_member(old_loop))) { 192 // Take early control, later control will be recalculated 193 // during next iteration of loop optimizations. 194 new_ctrl = get_early_ctrl(x); 195 new_loop = get_loop(new_ctrl); 196 } 197 // Set new location 198 set_ctrl(x, new_ctrl); 199 // If changing loop bodies, see if we need to collect into new body 200 if (old_loop != new_loop) { 201 if (old_loop && !old_loop->_child) 202 old_loop->_body.yank(x); 203 if (!new_loop->_child) 204 new_loop->_body.push(x); // Collect body info 205 } 206 } 207 208 return phi; 209 } 210 211 //------------------------------dominated_by------------------------------------ 212 // Replace the dominated test with an obvious true or false. Place it on the 213 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the 214 // live path up to the dominating control. 215 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff, bool flip, bool exclude_loop_predicate ) { 216 if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); } 217 218 // prevdom is the dominating projection of the dominating test. 219 assert( iff->is_If(), "" ); 220 assert(iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd || iff->Opcode() == Op_RangeCheck, "Check this code when new subtype is added"); 221 int pop = prevdom->Opcode(); 222 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" ); 223 if (flip) { 224 if (pop == Op_IfTrue) 225 pop = Op_IfFalse; 226 else 227 pop = Op_IfTrue; 228 } 229 // 'con' is set to true or false to kill the dominated test. 230 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO); 231 set_ctrl(con, C->root()); // Constant gets a new use 232 // Hack the dominated test 233 _igvn.replace_input_of(iff, 1, con); 234 235 // If I dont have a reachable TRUE and FALSE path following the IfNode then 236 // I can assume this path reaches an infinite loop. In this case it's not 237 // important to optimize the data Nodes - either the whole compilation will 238 // be tossed or this path (and all data Nodes) will go dead. 239 if (iff->outcnt() != 2) return; 240 241 // Make control-dependent data Nodes on the live path (path that will remain 242 // once the dominated IF is removed) become control-dependent on the 243 // dominating projection. 244 Node* dp = iff->as_If()->proj_out_or_null(pop == Op_IfTrue); 245 246 // Loop predicates may have depending checks which should not 247 // be skipped. For example, range check predicate has two checks 248 // for lower and upper bounds. 249 if (dp == NULL) 250 return; 251 252 ProjNode* dp_proj = dp->as_Proj(); 253 ProjNode* unc_proj = iff->as_If()->proj_out(1 - dp_proj->_con)->as_Proj(); 254 if (exclude_loop_predicate && 255 (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) != NULL || 256 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check) != NULL)) { 257 // If this is a range check (IfNode::is_range_check), do not 258 // reorder because Compile::allow_range_check_smearing might have 259 // changed the check. 260 return; // Let IGVN transformation change control dependence. 261 } 262 263 IdealLoopTree *old_loop = get_loop(dp); 264 265 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { 266 Node* cd = dp->fast_out(i); // Control-dependent node 267 if (cd->depends_only_on_test()) { 268 assert(cd->in(0) == dp, ""); 269 _igvn.replace_input_of(cd, 0, prevdom); 270 set_early_ctrl(cd); 271 IdealLoopTree *new_loop = get_loop(get_ctrl(cd)); 272 if (old_loop != new_loop) { 273 if (!old_loop->_child) old_loop->_body.yank(cd); 274 if (!new_loop->_child) new_loop->_body.push(cd); 275 } 276 --i; 277 --imax; 278 } 279 } 280 } 281 282 //------------------------------has_local_phi_input---------------------------- 283 // Return TRUE if 'n' has Phi inputs from its local block and no other 284 // block-local inputs (all non-local-phi inputs come from earlier blocks) 285 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) { 286 Node *n_ctrl = get_ctrl(n); 287 // See if some inputs come from a Phi in this block, or from before 288 // this block. 289 uint i; 290 for( i = 1; i < n->req(); i++ ) { 291 Node *phi = n->in(i); 292 if( phi->is_Phi() && phi->in(0) == n_ctrl ) 293 break; 294 } 295 if( i >= n->req() ) 296 return NULL; // No Phi inputs; nowhere to clone thru 297 298 // Check for inputs created between 'n' and the Phi input. These 299 // must split as well; they have already been given the chance 300 // (courtesy of a post-order visit) and since they did not we must 301 // recover the 'cost' of splitting them by being very profitable 302 // when splitting 'n'. Since this is unlikely we simply give up. 303 for( i = 1; i < n->req(); i++ ) { 304 Node *m = n->in(i); 305 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) { 306 // We allow the special case of AddP's with no local inputs. 307 // This allows us to split-up address expressions. 308 if (m->is_AddP() && 309 get_ctrl(m->in(2)) != n_ctrl && 310 get_ctrl(m->in(3)) != n_ctrl) { 311 // Move the AddP up to dominating point 312 Node* c = find_non_split_ctrl(idom(n_ctrl)); 313 if (c->is_OuterStripMinedLoop()) { 314 c->as_Loop()->verify_strip_mined(1); 315 c = c->in(LoopNode::EntryControl); 316 } 317 set_ctrl_and_loop(m, c); 318 continue; 319 } 320 return NULL; 321 } 322 assert(m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control"); 323 } 324 325 return n_ctrl; 326 } 327 328 //------------------------------remix_address_expressions---------------------- 329 // Rework addressing expressions to get the most loop-invariant stuff 330 // moved out. We'd like to do all associative operators, but it's especially 331 // important (common) to do address expressions. 332 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) { 333 if (!has_ctrl(n)) return NULL; 334 Node *n_ctrl = get_ctrl(n); 335 IdealLoopTree *n_loop = get_loop(n_ctrl); 336 337 // See if 'n' mixes loop-varying and loop-invariant inputs and 338 // itself is loop-varying. 339 340 // Only interested in binary ops (and AddP) 341 if( n->req() < 3 || n->req() > 4 ) return NULL; 342 343 Node *n1_ctrl = get_ctrl(n->in( 1)); 344 Node *n2_ctrl = get_ctrl(n->in( 2)); 345 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3)); 346 IdealLoopTree *n1_loop = get_loop( n1_ctrl ); 347 IdealLoopTree *n2_loop = get_loop( n2_ctrl ); 348 IdealLoopTree *n3_loop = get_loop( n3_ctrl ); 349 350 // Does one of my inputs spin in a tighter loop than self? 351 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) || 352 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) || 353 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) ) 354 return NULL; // Leave well enough alone 355 356 // Is at least one of my inputs loop-invariant? 357 if( n1_loop == n_loop && 358 n2_loop == n_loop && 359 n3_loop == n_loop ) 360 return NULL; // No loop-invariant inputs 361 362 363 int n_op = n->Opcode(); 364 365 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2). 366 if( n_op == Op_LShiftI ) { 367 // Scale is loop invariant 368 Node *scale = n->in(2); 369 Node *scale_ctrl = get_ctrl(scale); 370 IdealLoopTree *scale_loop = get_loop(scale_ctrl ); 371 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) ) 372 return NULL; 373 const TypeInt *scale_t = scale->bottom_type()->isa_int(); 374 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 ) 375 return NULL; // Dont bother with byte/short masking 376 // Add must vary with loop (else shift would be loop-invariant) 377 Node *add = n->in(1); 378 Node *add_ctrl = get_ctrl(add); 379 IdealLoopTree *add_loop = get_loop(add_ctrl); 380 //assert( n_loop == add_loop, "" ); 381 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops 382 383 // Convert I-V into I+ (0-V); same for V-I 384 if( add->Opcode() == Op_SubI && 385 _igvn.type( add->in(1) ) != TypeInt::ZERO ) { 386 Node *zero = _igvn.intcon(0); 387 set_ctrl(zero, C->root()); 388 Node *neg = new SubINode( _igvn.intcon(0), add->in(2) ); 389 register_new_node( neg, get_ctrl(add->in(2) ) ); 390 add = new AddINode( add->in(1), neg ); 391 register_new_node( add, add_ctrl ); 392 } 393 if( add->Opcode() != Op_AddI ) return NULL; 394 // See if one add input is loop invariant 395 Node *add_var = add->in(1); 396 Node *add_var_ctrl = get_ctrl(add_var); 397 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 398 Node *add_invar = add->in(2); 399 Node *add_invar_ctrl = get_ctrl(add_invar); 400 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl ); 401 if( add_var_loop == n_loop ) { 402 } else if( add_invar_loop == n_loop ) { 403 // Swap to find the invariant part 404 add_invar = add_var; 405 add_invar_ctrl = add_var_ctrl; 406 add_invar_loop = add_var_loop; 407 add_var = add->in(2); 408 Node *add_var_ctrl = get_ctrl(add_var); 409 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 410 } else // Else neither input is loop invariant 411 return NULL; 412 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) ) 413 return NULL; // No invariant part of the add? 414 415 // Yes! Reshape address expression! 416 Node *inv_scale = new LShiftINode( add_invar, scale ); 417 Node *inv_scale_ctrl = 418 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ? 419 add_invar_ctrl : scale_ctrl; 420 register_new_node( inv_scale, inv_scale_ctrl ); 421 Node *var_scale = new LShiftINode( add_var, scale ); 422 register_new_node( var_scale, n_ctrl ); 423 Node *var_add = new AddINode( var_scale, inv_scale ); 424 register_new_node( var_add, n_ctrl ); 425 _igvn.replace_node( n, var_add ); 426 return var_add; 427 } 428 429 // Replace (I+V) with (V+I) 430 if( n_op == Op_AddI || 431 n_op == Op_AddL || 432 n_op == Op_AddF || 433 n_op == Op_AddD || 434 n_op == Op_MulI || 435 n_op == Op_MulL || 436 n_op == Op_MulF || 437 n_op == Op_MulD ) { 438 if( n2_loop == n_loop ) { 439 assert( n1_loop != n_loop, "" ); 440 n->swap_edges(1, 2); 441 } 442 } 443 444 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V), 445 // but not if I2 is a constant. 446 if( n_op == Op_AddP ) { 447 if( n2_loop == n_loop && n3_loop != n_loop ) { 448 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) { 449 Node *n22_ctrl = get_ctrl(n->in(2)->in(2)); 450 Node *n23_ctrl = get_ctrl(n->in(2)->in(3)); 451 IdealLoopTree *n22loop = get_loop( n22_ctrl ); 452 IdealLoopTree *n23_loop = get_loop( n23_ctrl ); 453 if( n22loop != n_loop && n22loop->is_member(n_loop) && 454 n23_loop == n_loop ) { 455 Node *add1 = new AddPNode( n->in(1), n->in(2)->in(2), n->in(3) ); 456 // Stuff new AddP in the loop preheader 457 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) ); 458 Node *add2 = new AddPNode( n->in(1), add1, n->in(2)->in(3) ); 459 register_new_node( add2, n_ctrl ); 460 _igvn.replace_node( n, add2 ); 461 return add2; 462 } 463 } 464 } 465 466 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V) 467 if (n2_loop != n_loop && n3_loop == n_loop) { 468 if (n->in(3)->Opcode() == Op_AddX) { 469 Node *V = n->in(3)->in(1); 470 Node *I = n->in(3)->in(2); 471 if (is_member(n_loop,get_ctrl(V))) { 472 } else { 473 Node *tmp = V; V = I; I = tmp; 474 } 475 if (!is_member(n_loop,get_ctrl(I))) { 476 Node *add1 = new AddPNode(n->in(1), n->in(2), I); 477 // Stuff new AddP in the loop preheader 478 register_new_node(add1, n_loop->_head->in(LoopNode::EntryControl)); 479 Node *add2 = new AddPNode(n->in(1), add1, V); 480 register_new_node(add2, n_ctrl); 481 _igvn.replace_node(n, add2); 482 return add2; 483 } 484 } 485 } 486 } 487 488 return NULL; 489 } 490 491 //------------------------------conditional_move------------------------------- 492 // Attempt to replace a Phi with a conditional move. We have some pretty 493 // strict profitability requirements. All Phis at the merge point must 494 // be converted, so we can remove the control flow. We need to limit the 495 // number of c-moves to a small handful. All code that was in the side-arms 496 // of the CFG diamond is now speculatively executed. This code has to be 497 // "cheap enough". We are pretty much limited to CFG diamonds that merge 498 // 1 or 2 items with a total of 1 or 2 ops executed speculatively. 499 Node *PhaseIdealLoop::conditional_move( Node *region ) { 500 501 assert(region->is_Region(), "sanity check"); 502 if (region->req() != 3) return NULL; 503 504 // Check for CFG diamond 505 Node *lp = region->in(1); 506 Node *rp = region->in(2); 507 if (!lp || !rp) return NULL; 508 Node *lp_c = lp->in(0); 509 if (lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If()) return NULL; 510 IfNode *iff = lp_c->as_If(); 511 512 // Check for ops pinned in an arm of the diamond. 513 // Can't remove the control flow in this case 514 if (lp->outcnt() > 1) return NULL; 515 if (rp->outcnt() > 1) return NULL; 516 517 IdealLoopTree* r_loop = get_loop(region); 518 assert(r_loop == get_loop(iff), "sanity"); 519 // Always convert to CMOVE if all results are used only outside this loop. 520 bool used_inside_loop = (r_loop == _ltree_root); 521 522 // Check profitability 523 int cost = 0; 524 int phis = 0; 525 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 526 Node *out = region->fast_out(i); 527 if (!out->is_Phi()) continue; // Ignore other control edges, etc 528 phis++; 529 PhiNode* phi = out->as_Phi(); 530 BasicType bt = phi->type()->basic_type(); 531 switch (bt) { 532 case T_DOUBLE: 533 case T_FLOAT: 534 if (C->use_cmove()) { 535 continue; //TODO: maybe we want to add some cost 536 } 537 cost += Matcher::float_cmove_cost(); // Could be very expensive 538 break; 539 case T_LONG: { 540 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's 541 } 542 case T_INT: // These all CMOV fine 543 case T_ADDRESS: { // (RawPtr) 544 cost++; 545 break; 546 } 547 case T_NARROWOOP: // Fall through 548 case T_OBJECT: { // Base oops are OK, but not derived oops 549 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr(); 550 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a 551 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus 552 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we 553 // have a Phi for the base here that we convert to a CMOVE all is well 554 // and good. But if the base is dead, we'll not make a CMOVE. Later 555 // the allocator will have to produce a base by creating a CMOVE of the 556 // relevant bases. This puts the allocator in the business of 557 // manufacturing expensive instructions, generally a bad plan. 558 // Just Say No to Conditionally-Moved Derived Pointers. 559 if (tp && tp->offset() != 0) 560 return NULL; 561 cost++; 562 break; 563 } 564 default: 565 return NULL; // In particular, can't do memory or I/O 566 } 567 // Add in cost any speculative ops 568 for (uint j = 1; j < region->req(); j++) { 569 Node *proj = region->in(j); 570 Node *inp = phi->in(j); 571 if (get_ctrl(inp) == proj) { // Found local op 572 cost++; 573 // Check for a chain of dependent ops; these will all become 574 // speculative in a CMOV. 575 for (uint k = 1; k < inp->req(); k++) 576 if (get_ctrl(inp->in(k)) == proj) 577 cost += ConditionalMoveLimit; // Too much speculative goo 578 } 579 } 580 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode. 581 // This will likely Split-If, a higher-payoff operation. 582 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) { 583 Node* use = phi->fast_out(k); 584 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr()) 585 cost += ConditionalMoveLimit; 586 // Is there a use inside the loop? 587 // Note: check only basic types since CMoveP is pinned. 588 if (!used_inside_loop && is_java_primitive(bt)) { 589 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use); 590 if (r_loop == u_loop || r_loop->is_member(u_loop)) { 591 used_inside_loop = true; 592 } 593 } 594 } 595 }//for 596 Node* bol = iff->in(1); 597 assert(bol->Opcode() == Op_Bool, ""); 598 int cmp_op = bol->in(1)->Opcode(); 599 // It is expensive to generate flags from a float compare. 600 // Avoid duplicated float compare. 601 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL; 602 603 float infrequent_prob = PROB_UNLIKELY_MAG(3); 604 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop. 605 if (used_inside_loop) { 606 if (cost >= ConditionalMoveLimit) return NULL; // Too much goo 607 608 // BlockLayoutByFrequency optimization moves infrequent branch 609 // from hot path. No point in CMOV'ing in such case (110 is used 610 // instead of 100 to take into account not exactness of float value). 611 if (BlockLayoutByFrequency) { 612 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f); 613 } 614 } 615 // Check for highly predictable branch. No point in CMOV'ing if 616 // we are going to predict accurately all the time. 617 if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) { 618 //keep going 619 } else if (iff->_prob < infrequent_prob || 620 iff->_prob > (1.0f - infrequent_prob)) 621 return NULL; 622 623 // -------------- 624 // Now replace all Phis with CMOV's 625 Node *cmov_ctrl = iff->in(0); 626 uint flip = (lp->Opcode() == Op_IfTrue); 627 Node_List wq; 628 while (1) { 629 PhiNode* phi = NULL; 630 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 631 Node *out = region->fast_out(i); 632 if (out->is_Phi()) { 633 phi = out->as_Phi(); 634 break; 635 } 636 } 637 if (phi == NULL) break; 638 if (PrintOpto && VerifyLoopOptimizations) { tty->print_cr("CMOV"); } 639 // Move speculative ops 640 wq.push(phi); 641 while (wq.size() > 0) { 642 Node *n = wq.pop(); 643 for (uint j = 1; j < n->req(); j++) { 644 Node* m = n->in(j); 645 if (m != NULL && !is_dominator(get_ctrl(m), cmov_ctrl)) { 646 #ifndef PRODUCT 647 if (PrintOpto && VerifyLoopOptimizations) { 648 tty->print(" speculate: "); 649 m->dump(); 650 } 651 #endif 652 set_ctrl(m, cmov_ctrl); 653 wq.push(m); 654 } 655 } 656 } 657 Node *cmov = CMoveNode::make(cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi)); 658 register_new_node( cmov, cmov_ctrl ); 659 _igvn.replace_node( phi, cmov ); 660 #ifndef PRODUCT 661 if (TraceLoopOpts) { 662 tty->print("CMOV "); 663 r_loop->dump_head(); 664 if (Verbose) { 665 bol->in(1)->dump(1); 666 cmov->dump(1); 667 } 668 } 669 if (VerifyLoopOptimizations) verify(); 670 #endif 671 } 672 673 // The useless CFG diamond will fold up later; see the optimization in 674 // RegionNode::Ideal. 675 _igvn._worklist.push(region); 676 677 return iff->in(1); 678 } 679 680 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) { 681 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 682 Node* u = m->fast_out(i); 683 if (u->is_CFG()) { 684 if (u->Opcode() == Op_NeverBranch) { 685 u = ((NeverBranchNode*)u)->proj_out(0); 686 enqueue_cfg_uses(u, wq); 687 } else { 688 wq.push(u); 689 } 690 } 691 } 692 } 693 694 // Try moving a store out of a loop, right before the loop 695 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) { 696 // Store has to be first in the loop body 697 IdealLoopTree *n_loop = get_loop(n_ctrl); 698 if (n->is_Store() && n_loop != _ltree_root && 699 n_loop->is_loop() && n_loop->_head->is_Loop() && 700 n->in(0) != NULL) { 701 Node* address = n->in(MemNode::Address); 702 Node* value = n->in(MemNode::ValueIn); 703 Node* mem = n->in(MemNode::Memory); 704 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 705 IdealLoopTree* value_loop = get_loop(get_ctrl(value)); 706 707 // - address and value must be loop invariant 708 // - memory must be a memory Phi for the loop 709 // - Store must be the only store on this memory slice in the 710 // loop: if there's another store following this one then value 711 // written at iteration i by the second store could be overwritten 712 // at iteration i+n by the first store: it's not safe to move the 713 // first store out of the loop 714 // - nothing must observe the memory Phi: it guarantees no read 715 // before the store, we are also guaranteed the store post 716 // dominates the loop head (ignoring a possible early 717 // exit). Otherwise there would be extra Phi involved between the 718 // loop's Phi and the store. 719 // - there must be no early exit from the loop before the Store 720 // (such an exit most of the time would be an extra use of the 721 // memory Phi but sometimes is a bottom memory Phi that takes the 722 // store as input). 723 724 if (!n_loop->is_member(address_loop) && 725 !n_loop->is_member(value_loop) && 726 mem->is_Phi() && mem->in(0) == n_loop->_head && 727 mem->outcnt() == 1 && 728 mem->in(LoopNode::LoopBackControl) == n) { 729 730 assert(n_loop->_tail != NULL, "need a tail"); 731 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop"); 732 733 // Verify that there's no early exit of the loop before the store. 734 bool ctrl_ok = false; 735 { 736 // Follow control from loop head until n, we exit the loop or 737 // we reach the tail 738 ResourceMark rm; 739 Unique_Node_List wq; 740 wq.push(n_loop->_head); 741 742 for (uint next = 0; next < wq.size(); ++next) { 743 Node *m = wq.at(next); 744 if (m == n->in(0)) { 745 ctrl_ok = true; 746 continue; 747 } 748 assert(!has_ctrl(m), "should be CFG"); 749 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) { 750 ctrl_ok = false; 751 break; 752 } 753 enqueue_cfg_uses(m, wq); 754 if (wq.size() > 10) { 755 ctrl_ok = false; 756 break; 757 } 758 } 759 } 760 if (ctrl_ok) { 761 // move the Store 762 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem); 763 _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl)); 764 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl)); 765 // Disconnect the phi now. An empty phi can confuse other 766 // optimizations in this pass of loop opts. 767 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl)); 768 n_loop->_body.yank(mem); 769 770 set_ctrl_and_loop(n, n->in(0)); 771 772 return n; 773 } 774 } 775 } 776 return NULL; 777 } 778 779 // Try moving a store out of a loop, right after the loop 780 void PhaseIdealLoop::try_move_store_after_loop(Node* n) { 781 if (n->is_Store() && n->in(0) != NULL) { 782 Node *n_ctrl = get_ctrl(n); 783 IdealLoopTree *n_loop = get_loop(n_ctrl); 784 // Store must be in a loop 785 if (n_loop != _ltree_root && !n_loop->_irreducible) { 786 Node* address = n->in(MemNode::Address); 787 Node* value = n->in(MemNode::ValueIn); 788 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 789 // address must be loop invariant 790 if (!n_loop->is_member(address_loop)) { 791 // Store must be last on this memory slice in the loop and 792 // nothing in the loop must observe it 793 Node* phi = NULL; 794 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 795 Node* u = n->fast_out(i); 796 if (has_ctrl(u)) { // control use? 797 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 798 if (!n_loop->is_member(u_loop)) { 799 continue; 800 } 801 if (u->is_Phi() && u->in(0) == n_loop->_head) { 802 assert(_igvn.type(u) == Type::MEMORY, "bad phi"); 803 // multiple phis on the same slice are possible 804 if (phi != NULL) { 805 return; 806 } 807 phi = u; 808 continue; 809 } 810 } 811 return; 812 } 813 if (phi != NULL) { 814 // Nothing in the loop before the store (next iteration) 815 // must observe the stored value 816 bool mem_ok = true; 817 { 818 ResourceMark rm; 819 Unique_Node_List wq; 820 wq.push(phi); 821 for (uint next = 0; next < wq.size() && mem_ok; ++next) { 822 Node *m = wq.at(next); 823 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) { 824 Node* u = m->fast_out(i); 825 if (u->is_Store() || u->is_Phi()) { 826 if (u != n) { 827 wq.push(u); 828 mem_ok = (wq.size() <= 10); 829 } 830 } else { 831 mem_ok = false; 832 break; 833 } 834 } 835 } 836 } 837 if (mem_ok) { 838 // Move the store out of the loop if the LCA of all 839 // users (except for the phi) is outside the loop. 840 Node* hook = new Node(1); 841 _igvn.rehash_node_delayed(phi); 842 int count = phi->replace_edge(n, hook); 843 assert(count > 0, "inconsistent phi"); 844 845 // Compute latest point this store can go 846 Node* lca = get_late_ctrl(n, get_ctrl(n)); 847 if (n_loop->is_member(get_loop(lca))) { 848 // LCA is in the loop - bail out 849 _igvn.replace_node(hook, n); 850 return; 851 } 852 #ifdef ASSERT 853 if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) { 854 assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined"); 855 n_loop->_head->as_Loop()->verify_strip_mined(1); 856 Node* outer = n_loop->_head->as_CountedLoop()->outer_loop(); 857 IdealLoopTree* outer_loop = get_loop(outer); 858 assert(n_loop->_parent == outer_loop, "broken loop tree"); 859 assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state"); 860 } 861 #endif 862 863 // Move store out of the loop 864 _igvn.replace_node(hook, n->in(MemNode::Memory)); 865 _igvn.replace_input_of(n, 0, lca); 866 set_ctrl_and_loop(n, lca); 867 868 // Disconnect the phi now. An empty phi can confuse other 869 // optimizations in this pass of loop opts.. 870 if (phi->in(LoopNode::LoopBackControl) == phi) { 871 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl)); 872 n_loop->_body.yank(phi); 873 } 874 } 875 } 876 } 877 } 878 } 879 } 880 881 //------------------------------split_if_with_blocks_pre----------------------- 882 // Do the real work in a non-recursive function. Data nodes want to be 883 // cloned in the pre-order so they can feed each other nicely. 884 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) { 885 // Cloning these guys is unlikely to win 886 int n_op = n->Opcode(); 887 if( n_op == Op_MergeMem ) return n; 888 if( n->is_Proj() ) return n; 889 // Do not clone-up CmpFXXX variations, as these are always 890 // followed by a CmpI 891 if( n->is_Cmp() ) return n; 892 // Attempt to use a conditional move instead of a phi/branch 893 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) { 894 Node *cmov = conditional_move( n ); 895 if( cmov ) return cmov; 896 } 897 if( n->is_CFG() || n->is_LoadStore() ) 898 return n; 899 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd 900 n_op == Op_Opaque2 ) { 901 if( !C->major_progress() ) // If chance of no more loop opts... 902 _igvn._worklist.push(n); // maybe we'll remove them 903 return n; 904 } 905 906 if( n->is_Con() ) return n; // No cloning for Con nodes 907 908 Node *n_ctrl = get_ctrl(n); 909 if( !n_ctrl ) return n; // Dead node 910 911 Node* res = try_move_store_before_loop(n, n_ctrl); 912 if (res != NULL) { 913 return n; 914 } 915 916 // Attempt to remix address expressions for loop invariants 917 Node *m = remix_address_expressions( n ); 918 if( m ) return m; 919 920 if (n->is_ConstraintCast()) { 921 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this); 922 // ConstraintCastNode::dominating_cast() uses node control input to determine domination. 923 // Node control inputs don't necessarily agree with loop control info (due to 924 // transformations happened in between), thus additional dominance check is needed 925 // to keep loop info valid. 926 if (dom_cast != NULL && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) { 927 _igvn.replace_node(n, dom_cast); 928 return dom_cast; 929 } 930 } 931 932 // Determine if the Node has inputs from some local Phi. 933 // Returns the block to clone thru. 934 Node *n_blk = has_local_phi_input( n ); 935 if( !n_blk ) return n; 936 937 // Do not clone the trip counter through on a CountedLoop 938 // (messes up the canonical shape). 939 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n; 940 941 // Check for having no control input; not pinned. Allow 942 // dominating control. 943 if (n->in(0)) { 944 Node *dom = idom(n_blk); 945 if (dom_lca(n->in(0), dom) != n->in(0)) { 946 return n; 947 } 948 } 949 // Policy: when is it profitable. You must get more wins than 950 // policy before it is considered profitable. Policy is usually 0, 951 // so 1 win is considered profitable. Big merges will require big 952 // cloning, so get a larger policy. 953 int policy = n_blk->req() >> 2; 954 955 // If the loop is a candidate for range check elimination, 956 // delay splitting through it's phi until a later loop optimization 957 if (n_blk->is_CountedLoop()) { 958 IdealLoopTree *lp = get_loop(n_blk); 959 if (lp && lp->_rce_candidate) { 960 return n; 961 } 962 } 963 964 // Use same limit as split_if_with_blocks_post 965 if( C->live_nodes() > 35000 ) return n; // Method too big 966 967 // Split 'n' through the merge point if it is profitable 968 Node *phi = split_thru_phi( n, n_blk, policy ); 969 if (!phi) return n; 970 971 // Found a Phi to split thru! 972 // Replace 'n' with the new phi 973 _igvn.replace_node( n, phi ); 974 // Moved a load around the loop, 'en-registering' something. 975 if (n_blk->is_Loop() && n->is_Load() && 976 !phi->in(LoopNode::LoopBackControl)->is_Load()) 977 C->set_major_progress(); 978 979 return phi; 980 } 981 982 static bool merge_point_too_heavy(Compile* C, Node* region) { 983 // Bail out if the region and its phis have too many users. 984 int weight = 0; 985 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 986 weight += region->fast_out(i)->outcnt(); 987 } 988 int nodes_left = C->max_node_limit() - C->live_nodes(); 989 if (weight * 8 > nodes_left) { 990 if (PrintOpto) { 991 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight); 992 } 993 return true; 994 } else { 995 return false; 996 } 997 } 998 999 static bool merge_point_safe(Node* region) { 1000 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode 1001 // having a PhiNode input. This sidesteps the dangerous case where the split 1002 // ConvI2LNode may become TOP if the input Value() does not 1003 // overlap the ConvI2L range, leaving a node which may not dominate its 1004 // uses. 1005 // A better fix for this problem can be found in the BugTraq entry, but 1006 // expediency for Mantis demands this hack. 1007 // 6855164: If the merge point has a FastLockNode with a PhiNode input, we stop 1008 // split_if_with_blocks from splitting a block because we could not move around 1009 // the FastLockNode. 1010 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1011 Node* n = region->fast_out(i); 1012 if (n->is_Phi()) { 1013 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1014 Node* m = n->fast_out(j); 1015 if (m->is_FastLock()) 1016 return false; 1017 #ifdef _LP64 1018 if (m->Opcode() == Op_ConvI2L) 1019 return false; 1020 if (m->is_CastII() && m->isa_CastII()->has_range_check()) { 1021 return false; 1022 } 1023 #endif 1024 } 1025 } 1026 } 1027 return true; 1028 } 1029 1030 1031 //------------------------------place_near_use--------------------------------- 1032 // Place some computation next to use but not inside inner loops. 1033 // For inner loop uses move it to the preheader area. 1034 Node *PhaseIdealLoop::place_near_use(Node *useblock) const { 1035 IdealLoopTree *u_loop = get_loop( useblock ); 1036 if (u_loop->_irreducible) { 1037 return useblock; 1038 } 1039 if (u_loop->_child) { 1040 if (useblock == u_loop->_head && u_loop->_head->is_OuterStripMinedLoop()) { 1041 return u_loop->_head->in(LoopNode::EntryControl); 1042 } 1043 return useblock; 1044 } 1045 return 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() || n->is_CountedLoopEnd()) { 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 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1382 bs->loop_optimize_gc_barrier(n); 1383 } 1384 1385 //------------------------------split_if_with_blocks--------------------------- 1386 // Check for aggressive application of 'split-if' optimization, 1387 // using basic block level info. 1388 void PhaseIdealLoop::split_if_with_blocks( VectorSet &visited, Node_Stack &nstack ) { 1389 Node *n = C->root(); 1390 visited.set(n->_idx); // first, mark node as visited 1391 // Do pre-visit work for root 1392 n = split_if_with_blocks_pre( n ); 1393 uint cnt = n->outcnt(); 1394 uint i = 0; 1395 while (true) { 1396 // Visit all children 1397 if (i < cnt) { 1398 Node* use = n->raw_out(i); 1399 ++i; 1400 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) { 1401 // Now do pre-visit work for this use 1402 use = split_if_with_blocks_pre( use ); 1403 nstack.push(n, i); // Save parent and next use's index. 1404 n = use; // Process all children of current use. 1405 cnt = use->outcnt(); 1406 i = 0; 1407 } 1408 } 1409 else { 1410 // All of n's children have been processed, complete post-processing. 1411 if (cnt != 0 && !n->is_Con()) { 1412 assert(has_node(n), "no dead nodes"); 1413 split_if_with_blocks_post( n ); 1414 } 1415 if (nstack.is_empty()) { 1416 // Finished all nodes on stack. 1417 break; 1418 } 1419 // Get saved parent node and next use's index. Visit the rest of uses. 1420 n = nstack.node(); 1421 cnt = n->outcnt(); 1422 i = nstack.index(); 1423 nstack.pop(); 1424 } 1425 } 1426 } 1427 1428 1429 //============================================================================= 1430 // 1431 // C L O N E A L O O P B O D Y 1432 // 1433 1434 //------------------------------clone_iff-------------------------------------- 1435 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1436 // "Nearly" because all Nodes have been cloned from the original in the loop, 1437 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1438 // through the Phi recursively, and return a Bool. 1439 Node* PhaseIdealLoop::clone_iff(PhiNode *phi, IdealLoopTree *loop) { 1440 1441 // Convert this Phi into a Phi merging Bools 1442 uint i; 1443 for (i = 1; i < phi->req(); i++) { 1444 Node *b = phi->in(i); 1445 if (b->is_Phi()) { 1446 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi(), loop)); 1447 } else { 1448 assert(b->is_Bool() || b->Opcode() == Op_Opaque4, ""); 1449 } 1450 } 1451 1452 Node* n = phi->in(1); 1453 Node* sample_opaque = NULL; 1454 Node *sample_bool = NULL; 1455 if (n->Opcode() == Op_Opaque4) { 1456 sample_opaque = n; 1457 sample_bool = n->in(1); 1458 assert(sample_bool->is_Bool(), "wrong type"); 1459 } else { 1460 sample_bool = n; 1461 } 1462 Node *sample_cmp = sample_bool->in(1); 1463 1464 // Make Phis to merge the Cmp's inputs. 1465 PhiNode *phi1 = new PhiNode(phi->in(0), Type::TOP); 1466 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP); 1467 for (i = 1; i < phi->req(); i++) { 1468 Node *n1 = sample_opaque == NULL ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1); 1469 Node *n2 = sample_opaque == NULL ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2); 1470 phi1->set_req(i, n1); 1471 phi2->set_req(i, n2); 1472 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 1473 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 1474 } 1475 // See if these Phis have been made before. 1476 // Register with optimizer 1477 Node *hit1 = _igvn.hash_find_insert(phi1); 1478 if (hit1) { // Hit, toss just made Phi 1479 _igvn.remove_dead_node(phi1); // Remove new phi 1480 assert(hit1->is_Phi(), "" ); 1481 phi1 = (PhiNode*)hit1; // Use existing phi 1482 } else { // Miss 1483 _igvn.register_new_node_with_optimizer(phi1); 1484 } 1485 Node *hit2 = _igvn.hash_find_insert(phi2); 1486 if (hit2) { // Hit, toss just made Phi 1487 _igvn.remove_dead_node(phi2); // Remove new phi 1488 assert(hit2->is_Phi(), "" ); 1489 phi2 = (PhiNode*)hit2; // Use existing phi 1490 } else { // Miss 1491 _igvn.register_new_node_with_optimizer(phi2); 1492 } 1493 // Register Phis with loop/block info 1494 set_ctrl(phi1, phi->in(0)); 1495 set_ctrl(phi2, phi->in(0)); 1496 // Make a new Cmp 1497 Node *cmp = sample_cmp->clone(); 1498 cmp->set_req(1, phi1); 1499 cmp->set_req(2, phi2); 1500 _igvn.register_new_node_with_optimizer(cmp); 1501 set_ctrl(cmp, phi->in(0)); 1502 1503 // Make a new Bool 1504 Node *b = sample_bool->clone(); 1505 b->set_req(1,cmp); 1506 _igvn.register_new_node_with_optimizer(b); 1507 set_ctrl(b, phi->in(0)); 1508 1509 if (sample_opaque != NULL) { 1510 Node* opaque = sample_opaque->clone(); 1511 opaque->set_req(1, b); 1512 _igvn.register_new_node_with_optimizer(opaque); 1513 set_ctrl(opaque, phi->in(0)); 1514 return opaque; 1515 } 1516 1517 assert(b->is_Bool(), ""); 1518 return b; 1519 } 1520 1521 //------------------------------clone_bool------------------------------------- 1522 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1523 // "Nearly" because all Nodes have been cloned from the original in the loop, 1524 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1525 // through the Phi recursively, and return a Bool. 1526 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) { 1527 uint i; 1528 // Convert this Phi into a Phi merging Bools 1529 for( i = 1; i < phi->req(); i++ ) { 1530 Node *b = phi->in(i); 1531 if( b->is_Phi() ) { 1532 _igvn.replace_input_of(phi, i, clone_bool( b->as_Phi(), loop )); 1533 } else { 1534 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" ); 1535 } 1536 } 1537 1538 Node *sample_cmp = phi->in(1); 1539 1540 // Make Phis to merge the Cmp's inputs. 1541 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP ); 1542 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP ); 1543 for( uint j = 1; j < phi->req(); j++ ) { 1544 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP 1545 Node *n1, *n2; 1546 if( cmp_top->is_Cmp() ) { 1547 n1 = cmp_top->in(1); 1548 n2 = cmp_top->in(2); 1549 } else { 1550 n1 = n2 = cmp_top; 1551 } 1552 phi1->set_req( j, n1 ); 1553 phi2->set_req( j, n2 ); 1554 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 1555 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 1556 } 1557 1558 // See if these Phis have been made before. 1559 // Register with optimizer 1560 Node *hit1 = _igvn.hash_find_insert(phi1); 1561 if( hit1 ) { // Hit, toss just made Phi 1562 _igvn.remove_dead_node(phi1); // Remove new phi 1563 assert( hit1->is_Phi(), "" ); 1564 phi1 = (PhiNode*)hit1; // Use existing phi 1565 } else { // Miss 1566 _igvn.register_new_node_with_optimizer(phi1); 1567 } 1568 Node *hit2 = _igvn.hash_find_insert(phi2); 1569 if( hit2 ) { // Hit, toss just made Phi 1570 _igvn.remove_dead_node(phi2); // Remove new phi 1571 assert( hit2->is_Phi(), "" ); 1572 phi2 = (PhiNode*)hit2; // Use existing phi 1573 } else { // Miss 1574 _igvn.register_new_node_with_optimizer(phi2); 1575 } 1576 // Register Phis with loop/block info 1577 set_ctrl(phi1, phi->in(0)); 1578 set_ctrl(phi2, phi->in(0)); 1579 // Make a new Cmp 1580 Node *cmp = sample_cmp->clone(); 1581 cmp->set_req( 1, phi1 ); 1582 cmp->set_req( 2, phi2 ); 1583 _igvn.register_new_node_with_optimizer(cmp); 1584 set_ctrl(cmp, phi->in(0)); 1585 1586 assert( cmp->is_Cmp(), "" ); 1587 return (CmpNode*)cmp; 1588 } 1589 1590 //------------------------------sink_use--------------------------------------- 1591 // If 'use' was in the loop-exit block, it now needs to be sunk 1592 // below the post-loop merge point. 1593 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) { 1594 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) { 1595 set_ctrl(use, post_loop); 1596 for (DUIterator j = use->outs(); use->has_out(j); j++) 1597 sink_use(use->out(j), post_loop); 1598 } 1599 } 1600 1601 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new, 1602 IdealLoopTree* loop, IdealLoopTree* outer_loop, 1603 Node_List*& split_if_set, Node_List*& split_bool_set, 1604 Node_List*& split_cex_set, Node_List& worklist, 1605 uint new_counter, CloneLoopMode mode) { 1606 Node* nnn = old_new[old->_idx]; 1607 // Copy uses to a worklist, so I can munge the def-use info 1608 // with impunity. 1609 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1610 worklist.push(old->fast_out(j)); 1611 1612 while( worklist.size() ) { 1613 Node *use = worklist.pop(); 1614 if (!has_node(use)) continue; // Ignore dead nodes 1615 if (use->in(0) == C->top()) continue; 1616 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1617 // Check for data-use outside of loop - at least one of OLD or USE 1618 // must not be a CFG node. 1619 #ifdef ASSERT 1620 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == NULL) { 1621 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint(); 1622 assert(mode == ControlAroundStripMined && use == sfpt, "missed a node"); 1623 } 1624 #endif 1625 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) { 1626 1627 // If the Data use is an IF, that means we have an IF outside of the 1628 // loop that is switching on a condition that is set inside of the 1629 // loop. Happens if people set a loop-exit flag; then test the flag 1630 // in the loop to break the loop, then test is again outside of the 1631 // loop to determine which way the loop exited. 1632 // Loop predicate If node connects to Bool node through Opaque1 node. 1633 if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use) || use->Opcode() == Op_Opaque4) { 1634 // Since this code is highly unlikely, we lazily build the worklist 1635 // of such Nodes to go split. 1636 if (!split_if_set) { 1637 ResourceArea *area = Thread::current()->resource_area(); 1638 split_if_set = new Node_List(area); 1639 } 1640 split_if_set->push(use); 1641 } 1642 if (use->is_Bool()) { 1643 if (!split_bool_set) { 1644 ResourceArea *area = Thread::current()->resource_area(); 1645 split_bool_set = new Node_List(area); 1646 } 1647 split_bool_set->push(use); 1648 } 1649 if (use->Opcode() == Op_CreateEx) { 1650 if (!split_cex_set) { 1651 ResourceArea *area = Thread::current()->resource_area(); 1652 split_cex_set = new Node_List(area); 1653 } 1654 split_cex_set->push(use); 1655 } 1656 1657 1658 // Get "block" use is in 1659 uint idx = 0; 1660 while( use->in(idx) != old ) idx++; 1661 Node *prev = use->is_CFG() ? use : get_ctrl(use); 1662 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" ); 1663 Node *cfg = prev->_idx >= new_counter 1664 ? prev->in(2) 1665 : idom(prev); 1666 if( use->is_Phi() ) // Phi use is in prior block 1667 cfg = prev->in(idx); // NOT in block of Phi itself 1668 if (cfg->is_top()) { // Use is dead? 1669 _igvn.replace_input_of(use, idx, C->top()); 1670 continue; 1671 } 1672 1673 while(!outer_loop->is_member(get_loop(cfg))) { 1674 prev = cfg; 1675 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg); 1676 } 1677 // If the use occurs after merging several exits from the loop, then 1678 // old value must have dominated all those exits. Since the same old 1679 // value was used on all those exits we did not need a Phi at this 1680 // merge point. NOW we do need a Phi here. Each loop exit value 1681 // is now merged with the peeled body exit; each exit gets its own 1682 // private Phi and those Phis need to be merged here. 1683 Node *phi; 1684 if( prev->is_Region() ) { 1685 if( idx == 0 ) { // Updating control edge? 1686 phi = prev; // Just use existing control 1687 } else { // Else need a new Phi 1688 phi = PhiNode::make( prev, old ); 1689 // Now recursively fix up the new uses of old! 1690 for( uint i = 1; i < prev->req(); i++ ) { 1691 worklist.push(phi); // Onto worklist once for each 'old' input 1692 } 1693 } 1694 } else { 1695 // Get new RegionNode merging old and new loop exits 1696 prev = old_new[prev->_idx]; 1697 assert( prev, "just made this in step 7" ); 1698 if( idx == 0) { // Updating control edge? 1699 phi = prev; // Just use existing control 1700 } else { // Else need a new Phi 1701 // Make a new Phi merging data values properly 1702 phi = PhiNode::make( prev, old ); 1703 phi->set_req( 1, nnn ); 1704 } 1705 } 1706 // If inserting a new Phi, check for prior hits 1707 if( idx != 0 ) { 1708 Node *hit = _igvn.hash_find_insert(phi); 1709 if( hit == NULL ) { 1710 _igvn.register_new_node_with_optimizer(phi); // Register new phi 1711 } else { // or 1712 // Remove the new phi from the graph and use the hit 1713 _igvn.remove_dead_node(phi); 1714 phi = hit; // Use existing phi 1715 } 1716 set_ctrl(phi, prev); 1717 } 1718 // Make 'use' use the Phi instead of the old loop body exit value 1719 _igvn.replace_input_of(use, idx, phi); 1720 if( use->_idx >= new_counter ) { // If updating new phis 1721 // Not needed for correctness, but prevents a weak assert 1722 // in AddPNode from tripping (when we end up with different 1723 // base & derived Phis that will become the same after 1724 // IGVN does CSE). 1725 Node *hit = _igvn.hash_find_insert(use); 1726 if( hit ) // Go ahead and re-hash for hits. 1727 _igvn.replace_node( use, hit ); 1728 } 1729 1730 // If 'use' was in the loop-exit block, it now needs to be sunk 1731 // below the post-loop merge point. 1732 sink_use( use, prev ); 1733 } 1734 } 1735 } 1736 1737 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop, 1738 IdealLoopTree* outer_loop, int dd, Node_List &old_new, 1739 Node_List& extra_data_nodes) { 1740 if (head->is_strip_mined() && mode != IgnoreStripMined) { 1741 CountedLoopNode* cl = head->as_CountedLoop(); 1742 Node* l = cl->outer_loop(); 1743 Node* tail = cl->outer_loop_tail(); 1744 IfNode* le = cl->outer_loop_end(); 1745 Node* sfpt = cl->outer_safepoint(); 1746 CountedLoopEndNode* cle = cl->loopexit(); 1747 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop(); 1748 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null(); 1749 Node* cle_out = cle->proj_out(false); 1750 1751 Node* new_sfpt = NULL; 1752 Node* new_cle_out = cle_out->clone(); 1753 old_new.map(cle_out->_idx, new_cle_out); 1754 if (mode == CloneIncludesStripMined) { 1755 // clone outer loop body 1756 Node* new_l = l->clone(); 1757 Node* new_tail = tail->clone(); 1758 IfNode* new_le = le->clone()->as_If(); 1759 new_sfpt = sfpt->clone(); 1760 1761 set_loop(new_l, outer_loop->_parent); 1762 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd); 1763 set_loop(new_cle_out, outer_loop->_parent); 1764 set_idom(new_cle_out, new_cle, dd); 1765 set_loop(new_sfpt, outer_loop->_parent); 1766 set_idom(new_sfpt, new_cle_out, dd); 1767 set_loop(new_le, outer_loop->_parent); 1768 set_idom(new_le, new_sfpt, dd); 1769 set_loop(new_tail, outer_loop->_parent); 1770 set_idom(new_tail, new_le, dd); 1771 set_idom(new_cl, new_l, dd); 1772 1773 old_new.map(l->_idx, new_l); 1774 old_new.map(tail->_idx, new_tail); 1775 old_new.map(le->_idx, new_le); 1776 old_new.map(sfpt->_idx, new_sfpt); 1777 1778 new_l->set_req(LoopNode::LoopBackControl, new_tail); 1779 new_l->set_req(0, new_l); 1780 new_tail->set_req(0, new_le); 1781 new_le->set_req(0, new_sfpt); 1782 new_sfpt->set_req(0, new_cle_out); 1783 new_cle_out->set_req(0, new_cle); 1784 new_cl->set_req(LoopNode::EntryControl, new_l); 1785 1786 _igvn.register_new_node_with_optimizer(new_l); 1787 _igvn.register_new_node_with_optimizer(new_tail); 1788 _igvn.register_new_node_with_optimizer(new_le); 1789 } else { 1790 Node *newhead = old_new[loop->_head->_idx]; 1791 newhead->as_Loop()->clear_strip_mined(); 1792 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl)); 1793 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 1794 } 1795 // Look at data node that were assigned a control in the outer 1796 // loop: they are kept in the outer loop by the safepoint so start 1797 // from the safepoint node's inputs. 1798 IdealLoopTree* outer_loop = get_loop(l); 1799 Node_Stack stack(2); 1800 stack.push(sfpt, 1); 1801 uint new_counter = C->unique(); 1802 while (stack.size() > 0) { 1803 Node* n = stack.node(); 1804 uint i = stack.index(); 1805 while (i < n->req() && 1806 (n->in(i) == NULL || 1807 !has_ctrl(n->in(i)) || 1808 get_loop(get_ctrl(n->in(i))) != outer_loop || 1809 (old_new[n->in(i)->_idx] != NULL && old_new[n->in(i)->_idx]->_idx >= new_counter))) { 1810 i++; 1811 } 1812 if (i < n->req()) { 1813 stack.set_index(i+1); 1814 stack.push(n->in(i), 0); 1815 } else { 1816 assert(old_new[n->_idx] == NULL || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet"); 1817 Node* m = n == sfpt ? new_sfpt : n->clone(); 1818 if (m != NULL) { 1819 for (uint i = 0; i < n->req(); i++) { 1820 if (m->in(i) != NULL && old_new[m->in(i)->_idx] != NULL) { 1821 m->set_req(i, old_new[m->in(i)->_idx]); 1822 } 1823 } 1824 } else { 1825 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?"); 1826 } 1827 if (n != sfpt) { 1828 extra_data_nodes.push(n); 1829 _igvn.register_new_node_with_optimizer(m); 1830 assert(get_ctrl(n) == cle_out, "what other control?"); 1831 set_ctrl(m, new_cle_out); 1832 old_new.map(n->_idx, m); 1833 } 1834 stack.pop(); 1835 } 1836 } 1837 if (mode == CloneIncludesStripMined) { 1838 _igvn.register_new_node_with_optimizer(new_sfpt); 1839 _igvn.register_new_node_with_optimizer(new_cle_out); 1840 } 1841 } else { 1842 Node *newhead = old_new[loop->_head->_idx]; 1843 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 1844 } 1845 } 1846 1847 //------------------------------clone_loop------------------------------------- 1848 // 1849 // C L O N E A L O O P B O D Y 1850 // 1851 // This is the basic building block of the loop optimizations. It clones an 1852 // entire loop body. It makes an old_new loop body mapping; with this mapping 1853 // you can find the new-loop equivalent to an old-loop node. All new-loop 1854 // nodes are exactly equal to their old-loop counterparts, all edges are the 1855 // same. All exits from the old-loop now have a RegionNode that merges the 1856 // equivalent new-loop path. This is true even for the normal "loop-exit" 1857 // condition. All uses of loop-invariant old-loop values now come from (one 1858 // or more) Phis that merge their new-loop equivalents. 1859 // 1860 // This operation leaves the graph in an illegal state: there are two valid 1861 // control edges coming from the loop pre-header to both loop bodies. I'll 1862 // definitely have to hack the graph after running this transform. 1863 // 1864 // From this building block I will further edit edges to perform loop peeling 1865 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc. 1866 // 1867 // Parameter side_by_size_idom: 1868 // When side_by_size_idom is NULL, the dominator tree is constructed for 1869 // the clone loop to dominate the original. Used in construction of 1870 // pre-main-post loop sequence. 1871 // When nonnull, the clone and original are side-by-side, both are 1872 // dominated by the side_by_side_idom node. Used in construction of 1873 // unswitched loops. 1874 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd, 1875 CloneLoopMode mode, Node* side_by_side_idom) { 1876 1877 LoopNode* head = loop->_head->as_Loop(); 1878 head->verify_strip_mined(1); 1879 1880 if (C->do_vector_loop() && PrintOpto) { 1881 const char* mname = C->method()->name()->as_quoted_ascii(); 1882 if (mname != NULL) { 1883 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname); 1884 } 1885 } 1886 1887 CloneMap& cm = C->clone_map(); 1888 Dict* dict = cm.dict(); 1889 if (C->do_vector_loop()) { 1890 cm.set_clone_idx(cm.max_gen()+1); 1891 #ifndef PRODUCT 1892 if (PrintOpto) { 1893 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx()); 1894 loop->dump_head(); 1895 } 1896 #endif 1897 } 1898 1899 // Step 1: Clone the loop body. Make the old->new mapping. 1900 uint i; 1901 for( i = 0; i < loop->_body.size(); i++ ) { 1902 Node *old = loop->_body.at(i); 1903 Node *nnn = old->clone(); 1904 old_new.map( old->_idx, nnn ); 1905 if (C->do_vector_loop()) { 1906 cm.verify_insert_and_clone(old, nnn, cm.clone_idx()); 1907 } 1908 _igvn.register_new_node_with_optimizer(nnn); 1909 } 1910 1911 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop; 1912 1913 // Step 2: Fix the edges in the new body. If the old input is outside the 1914 // loop use it. If the old input is INside the loop, use the corresponding 1915 // new node instead. 1916 for( i = 0; i < loop->_body.size(); i++ ) { 1917 Node *old = loop->_body.at(i); 1918 Node *nnn = old_new[old->_idx]; 1919 // Fix CFG/Loop controlling the new node 1920 if (has_ctrl(old)) { 1921 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]); 1922 } else { 1923 set_loop(nnn, outer_loop->_parent); 1924 if (old->outcnt() > 0) { 1925 set_idom( nnn, old_new[idom(old)->_idx], dd ); 1926 } 1927 } 1928 // Correct edges to the new node 1929 for( uint j = 0; j < nnn->req(); j++ ) { 1930 Node *n = nnn->in(j); 1931 if( n ) { 1932 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n ); 1933 if( loop->is_member( old_in_loop ) ) 1934 nnn->set_req(j, old_new[n->_idx]); 1935 } 1936 } 1937 _igvn.hash_find_insert(nnn); 1938 } 1939 1940 ResourceArea *area = Thread::current()->resource_area(); 1941 Node_List extra_data_nodes(area); 1942 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes); 1943 1944 // Step 3: Now fix control uses. Loop varying control uses have already 1945 // been fixed up (as part of all input edges in Step 2). Loop invariant 1946 // control uses must be either an IfFalse or an IfTrue. Make a merge 1947 // point to merge the old and new IfFalse/IfTrue nodes; make the use 1948 // refer to this. 1949 Node_List worklist(area); 1950 uint new_counter = C->unique(); 1951 for( i = 0; i < loop->_body.size(); i++ ) { 1952 Node* old = loop->_body.at(i); 1953 if( !old->is_CFG() ) continue; 1954 1955 // Copy uses to a worklist, so I can munge the def-use info 1956 // with impunity. 1957 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1958 worklist.push(old->fast_out(j)); 1959 1960 while( worklist.size() ) { // Visit all uses 1961 Node *use = worklist.pop(); 1962 if (!has_node(use)) continue; // Ignore dead nodes 1963 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1964 if( !loop->is_member( use_loop ) && use->is_CFG() ) { 1965 // Both OLD and USE are CFG nodes here. 1966 assert( use->is_Proj(), "" ); 1967 Node* nnn = old_new[old->_idx]; 1968 1969 Node* newuse = NULL; 1970 if (head->is_strip_mined() && mode != IgnoreStripMined) { 1971 CountedLoopNode* cl = head->as_CountedLoop(); 1972 CountedLoopEndNode* cle = cl->loopexit(); 1973 Node* cle_out = cle->proj_out_or_null(false); 1974 if (use == cle_out) { 1975 IfNode* le = cl->outer_loop_end(); 1976 use = le->proj_out(false); 1977 use_loop = get_loop(use); 1978 if (mode == CloneIncludesStripMined) { 1979 nnn = old_new[le->_idx]; 1980 } else { 1981 newuse = old_new[cle_out->_idx]; 1982 } 1983 } 1984 } 1985 if (newuse == NULL) { 1986 newuse = use->clone(); 1987 } 1988 1989 // Clone the loop exit control projection 1990 if (C->do_vector_loop()) { 1991 cm.verify_insert_and_clone(use, newuse, cm.clone_idx()); 1992 } 1993 newuse->set_req(0,nnn); 1994 _igvn.register_new_node_with_optimizer(newuse); 1995 set_loop(newuse, use_loop); 1996 set_idom(newuse, nnn, dom_depth(nnn) + 1 ); 1997 1998 // We need a Region to merge the exit from the peeled body and the 1999 // exit from the old loop body. 2000 RegionNode *r = new RegionNode(3); 2001 // Map the old use to the new merge point 2002 old_new.map( use->_idx, r ); 2003 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use)); 2004 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" ); 2005 2006 // The original user of 'use' uses 'r' instead. 2007 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) { 2008 Node* useuse = use->last_out(l); 2009 _igvn.rehash_node_delayed(useuse); 2010 uint uses_found = 0; 2011 if( useuse->in(0) == use ) { 2012 useuse->set_req(0, r); 2013 uses_found++; 2014 if( useuse->is_CFG() ) { 2015 assert( dom_depth(useuse) > dd_r, "" ); 2016 set_idom(useuse, r, dom_depth(useuse)); 2017 } 2018 } 2019 for( uint k = 1; k < useuse->req(); k++ ) { 2020 if( useuse->in(k) == use ) { 2021 useuse->set_req(k, r); 2022 uses_found++; 2023 if (useuse->is_Loop() && k == LoopNode::EntryControl) { 2024 assert(dom_depth(useuse) > dd_r , ""); 2025 set_idom(useuse, r, dom_depth(useuse)); 2026 } 2027 } 2028 } 2029 l -= uses_found; // we deleted 1 or more copies of this edge 2030 } 2031 2032 // Now finish up 'r' 2033 r->set_req( 1, newuse ); 2034 r->set_req( 2, use ); 2035 _igvn.register_new_node_with_optimizer(r); 2036 set_loop(r, use_loop); 2037 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r); 2038 } // End of if a loop-exit test 2039 } 2040 } 2041 2042 // Step 4: If loop-invariant use is not control, it must be dominated by a 2043 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region 2044 // there if needed. Make a Phi there merging old and new used values. 2045 Node_List *split_if_set = NULL; 2046 Node_List *split_bool_set = NULL; 2047 Node_List *split_cex_set = NULL; 2048 for( i = 0; i < loop->_body.size(); i++ ) { 2049 Node* old = loop->_body.at(i); 2050 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set, 2051 split_bool_set, split_cex_set, worklist, new_counter, 2052 mode); 2053 } 2054 2055 for (i = 0; i < extra_data_nodes.size(); i++) { 2056 Node* old = extra_data_nodes.at(i); 2057 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set, 2058 split_bool_set, split_cex_set, worklist, new_counter, 2059 mode); 2060 } 2061 2062 // Check for IFs that need splitting/cloning. Happens if an IF outside of 2063 // the loop uses a condition set in the loop. The original IF probably 2064 // takes control from one or more OLD Regions (which in turn get from NEW 2065 // Regions). In any case, there will be a set of Phis for each merge point 2066 // from the IF up to where the original BOOL def exists the loop. 2067 if (split_if_set) { 2068 while (split_if_set->size()) { 2069 Node *iff = split_if_set->pop(); 2070 if (iff->in(1)->is_Phi()) { 2071 Node *b = clone_iff(iff->in(1)->as_Phi(), loop); 2072 _igvn.replace_input_of(iff, 1, b); 2073 } 2074 } 2075 } 2076 if (split_bool_set) { 2077 while (split_bool_set->size()) { 2078 Node *b = split_bool_set->pop(); 2079 Node *phi = b->in(1); 2080 assert(phi->is_Phi(), ""); 2081 CmpNode *cmp = clone_bool((PhiNode*)phi, loop); 2082 _igvn.replace_input_of(b, 1, cmp); 2083 } 2084 } 2085 if (split_cex_set) { 2086 while (split_cex_set->size()) { 2087 Node *b = split_cex_set->pop(); 2088 assert(b->in(0)->is_Region(), ""); 2089 assert(b->in(1)->is_Phi(), ""); 2090 assert(b->in(0)->in(0) == b->in(1)->in(0), ""); 2091 split_up(b, b->in(0), NULL); 2092 } 2093 } 2094 2095 } 2096 2097 2098 //---------------------- stride_of_possible_iv ------------------------------------- 2099 // Looks for an iff/bool/comp with one operand of the compare 2100 // being a cycle involving an add and a phi, 2101 // with an optional truncation (left-shift followed by a right-shift) 2102 // of the add. Returns zero if not an iv. 2103 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) { 2104 Node* trunc1 = NULL; 2105 Node* trunc2 = NULL; 2106 const TypeInt* ttype = NULL; 2107 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) { 2108 return 0; 2109 } 2110 BoolNode* bl = iff->in(1)->as_Bool(); 2111 Node* cmp = bl->in(1); 2112 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) { 2113 return 0; 2114 } 2115 // Must have an invariant operand 2116 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) { 2117 return 0; 2118 } 2119 Node* add2 = NULL; 2120 Node* cmp1 = cmp->in(1); 2121 if (cmp1->is_Phi()) { 2122 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) ))) 2123 Node* phi = cmp1; 2124 for (uint i = 1; i < phi->req(); i++) { 2125 Node* in = phi->in(i); 2126 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in, 2127 &trunc1, &trunc2, &ttype); 2128 if (add && add->in(1) == phi) { 2129 add2 = add->in(2); 2130 break; 2131 } 2132 } 2133 } else { 2134 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) ))) 2135 Node* addtrunc = cmp1; 2136 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc, 2137 &trunc1, &trunc2, &ttype); 2138 if (add && add->in(1)->is_Phi()) { 2139 Node* phi = add->in(1); 2140 for (uint i = 1; i < phi->req(); i++) { 2141 if (phi->in(i) == addtrunc) { 2142 add2 = add->in(2); 2143 break; 2144 } 2145 } 2146 } 2147 } 2148 if (add2 != NULL) { 2149 const TypeInt* add2t = _igvn.type(add2)->is_int(); 2150 if (add2t->is_con()) { 2151 return add2t->get_con(); 2152 } 2153 } 2154 return 0; 2155 } 2156 2157 2158 //---------------------- stay_in_loop ------------------------------------- 2159 // Return the (unique) control output node that's in the loop (if it exists.) 2160 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) { 2161 Node* unique = NULL; 2162 if (!n) return NULL; 2163 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2164 Node* use = n->fast_out(i); 2165 if (!has_ctrl(use) && loop->is_member(get_loop(use))) { 2166 if (unique != NULL) { 2167 return NULL; 2168 } 2169 unique = use; 2170 } 2171 } 2172 return unique; 2173 } 2174 2175 //------------------------------ register_node ------------------------------------- 2176 // Utility to register node "n" with PhaseIdealLoop 2177 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) { 2178 _igvn.register_new_node_with_optimizer(n); 2179 loop->_body.push(n); 2180 if (n->is_CFG()) { 2181 set_loop(n, loop); 2182 set_idom(n, pred, ddepth); 2183 } else { 2184 set_ctrl(n, pred); 2185 } 2186 } 2187 2188 //------------------------------ proj_clone ------------------------------------- 2189 // Utility to create an if-projection 2190 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) { 2191 ProjNode* c = p->clone()->as_Proj(); 2192 c->set_req(0, iff); 2193 return c; 2194 } 2195 2196 //------------------------------ short_circuit_if ------------------------------------- 2197 // Force the iff control output to be the live_proj 2198 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) { 2199 guarantee(live_proj != NULL, "null projection"); 2200 int proj_con = live_proj->_con; 2201 assert(proj_con == 0 || proj_con == 1, "false or true projection"); 2202 Node *con = _igvn.intcon(proj_con); 2203 set_ctrl(con, C->root()); 2204 if (iff) { 2205 iff->set_req(1, con); 2206 } 2207 return con; 2208 } 2209 2210 //------------------------------ insert_if_before_proj ------------------------------------- 2211 // Insert a new if before an if projection (* - new node) 2212 // 2213 // before 2214 // if(test) 2215 // / \ 2216 // v v 2217 // other-proj proj (arg) 2218 // 2219 // after 2220 // if(test) 2221 // / \ 2222 // / v 2223 // | * proj-clone 2224 // v | 2225 // other-proj v 2226 // * new_if(relop(cmp[IU](left,right))) 2227 // / \ 2228 // v v 2229 // * new-proj proj 2230 // (returned) 2231 // 2232 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) { 2233 IfNode* iff = proj->in(0)->as_If(); 2234 IdealLoopTree *loop = get_loop(proj); 2235 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2236 int ddepth = dom_depth(proj); 2237 2238 _igvn.rehash_node_delayed(iff); 2239 _igvn.rehash_node_delayed(proj); 2240 2241 proj->set_req(0, NULL); // temporary disconnect 2242 ProjNode* proj2 = proj_clone(proj, iff); 2243 register_node(proj2, loop, iff, ddepth); 2244 2245 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right); 2246 register_node(cmp, loop, proj2, ddepth); 2247 2248 BoolNode* bol = new BoolNode(cmp, relop); 2249 register_node(bol, loop, proj2, ddepth); 2250 2251 int opcode = iff->Opcode(); 2252 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode"); 2253 IfNode* new_if = (opcode == Op_If) ? new IfNode(proj2, bol, iff->_prob, iff->_fcnt): 2254 new RangeCheckNode(proj2, bol, iff->_prob, iff->_fcnt); 2255 register_node(new_if, loop, proj2, ddepth); 2256 2257 proj->set_req(0, new_if); // reattach 2258 set_idom(proj, new_if, ddepth); 2259 2260 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj(); 2261 guarantee(new_exit != NULL, "null exit node"); 2262 register_node(new_exit, get_loop(other_proj), new_if, ddepth); 2263 2264 return new_exit; 2265 } 2266 2267 //------------------------------ insert_region_before_proj ------------------------------------- 2268 // Insert a region before an if projection (* - new node) 2269 // 2270 // before 2271 // if(test) 2272 // / | 2273 // v | 2274 // proj v 2275 // other-proj 2276 // 2277 // after 2278 // if(test) 2279 // / | 2280 // v | 2281 // * proj-clone v 2282 // | other-proj 2283 // v 2284 // * new-region 2285 // | 2286 // v 2287 // * dum_if 2288 // / \ 2289 // v \ 2290 // * dum-proj v 2291 // proj 2292 // 2293 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) { 2294 IfNode* iff = proj->in(0)->as_If(); 2295 IdealLoopTree *loop = get_loop(proj); 2296 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2297 int ddepth = dom_depth(proj); 2298 2299 _igvn.rehash_node_delayed(iff); 2300 _igvn.rehash_node_delayed(proj); 2301 2302 proj->set_req(0, NULL); // temporary disconnect 2303 ProjNode* proj2 = proj_clone(proj, iff); 2304 register_node(proj2, loop, iff, ddepth); 2305 2306 RegionNode* reg = new RegionNode(2); 2307 reg->set_req(1, proj2); 2308 register_node(reg, loop, iff, ddepth); 2309 2310 IfNode* dum_if = new IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt); 2311 register_node(dum_if, loop, reg, ddepth); 2312 2313 proj->set_req(0, dum_if); // reattach 2314 set_idom(proj, dum_if, ddepth); 2315 2316 ProjNode* dum_proj = proj_clone(other_proj, dum_if); 2317 register_node(dum_proj, loop, dum_if, ddepth); 2318 2319 return reg; 2320 } 2321 2322 //------------------------------ insert_cmpi_loop_exit ------------------------------------- 2323 // Clone a signed compare loop exit from an unsigned compare and 2324 // insert it before the unsigned cmp on the stay-in-loop path. 2325 // All new nodes inserted in the dominator tree between the original 2326 // if and it's projections. The original if test is replaced with 2327 // a constant to force the stay-in-loop path. 2328 // 2329 // This is done to make sure that the original if and it's projections 2330 // still dominate the same set of control nodes, that the ctrl() relation 2331 // from data nodes to them is preserved, and that their loop nesting is 2332 // preserved. 2333 // 2334 // before 2335 // if(i <u limit) unsigned compare loop exit 2336 // / | 2337 // v v 2338 // exit-proj stay-in-loop-proj 2339 // 2340 // after 2341 // if(stay-in-loop-const) original if 2342 // / | 2343 // / v 2344 // / if(i < limit) new signed test 2345 // / / | 2346 // / / v 2347 // / / if(i <u limit) new cloned unsigned test 2348 // / / / | 2349 // v v v | 2350 // region | 2351 // | | 2352 // dum-if | 2353 // / | | 2354 // ether | | 2355 // v v 2356 // exit-proj stay-in-loop-proj 2357 // 2358 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) { 2359 const bool Signed = true; 2360 const bool Unsigned = false; 2361 2362 BoolNode* bol = if_cmpu->in(1)->as_Bool(); 2363 if (bol->_test._test != BoolTest::lt) return NULL; 2364 CmpNode* cmpu = bol->in(1)->as_Cmp(); 2365 if (cmpu->Opcode() != Op_CmpU) return NULL; 2366 int stride = stride_of_possible_iv(if_cmpu); 2367 if (stride == 0) return NULL; 2368 2369 Node* lp_proj = stay_in_loop(if_cmpu, loop); 2370 guarantee(lp_proj != NULL, "null loop node"); 2371 2372 ProjNode* lp_continue = lp_proj->as_Proj(); 2373 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj(); 2374 2375 Node* limit = NULL; 2376 if (stride > 0) { 2377 limit = cmpu->in(2); 2378 } else { 2379 limit = _igvn.makecon(TypeInt::ZERO); 2380 set_ctrl(limit, C->root()); 2381 } 2382 // Create a new region on the exit path 2383 RegionNode* reg = insert_region_before_proj(lp_exit); 2384 guarantee(reg != NULL, "null region node"); 2385 2386 // Clone the if-cmpu-true-false using a signed compare 2387 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge; 2388 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue); 2389 reg->add_req(cmpi_exit); 2390 2391 // Clone the if-cmpu-true-false 2392 BoolTest::mask rel_u = bol->_test._test; 2393 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue); 2394 reg->add_req(cmpu_exit); 2395 2396 // Force original if to stay in loop. 2397 short_circuit_if(if_cmpu, lp_continue); 2398 2399 return cmpi_exit->in(0)->as_If(); 2400 } 2401 2402 //------------------------------ remove_cmpi_loop_exit ------------------------------------- 2403 // Remove a previously inserted signed compare loop exit. 2404 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) { 2405 Node* lp_proj = stay_in_loop(if_cmp, loop); 2406 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI && 2407 stay_in_loop(lp_proj, loop)->is_If() && 2408 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu"); 2409 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO); 2410 set_ctrl(con, C->root()); 2411 if_cmp->set_req(1, con); 2412 } 2413 2414 //------------------------------ scheduled_nodelist ------------------------------------- 2415 // Create a post order schedule of nodes that are in the 2416 // "member" set. The list is returned in "sched". 2417 // The first node in "sched" is the loop head, followed by 2418 // nodes which have no inputs in the "member" set, and then 2419 // followed by the nodes that have an immediate input dependence 2420 // on a node in "sched". 2421 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) { 2422 2423 assert(member.test(loop->_head->_idx), "loop head must be in member set"); 2424 Arena *a = Thread::current()->resource_area(); 2425 VectorSet visited(a); 2426 Node_Stack nstack(a, loop->_body.size()); 2427 2428 Node* n = loop->_head; // top of stack is cached in "n" 2429 uint idx = 0; 2430 visited.set(n->_idx); 2431 2432 // Initially push all with no inputs from within member set 2433 for(uint i = 0; i < loop->_body.size(); i++ ) { 2434 Node *elt = loop->_body.at(i); 2435 if (member.test(elt->_idx)) { 2436 bool found = false; 2437 for (uint j = 0; j < elt->req(); j++) { 2438 Node* def = elt->in(j); 2439 if (def && member.test(def->_idx) && def != elt) { 2440 found = true; 2441 break; 2442 } 2443 } 2444 if (!found && elt != loop->_head) { 2445 nstack.push(n, idx); 2446 n = elt; 2447 assert(!visited.test(n->_idx), "not seen yet"); 2448 visited.set(n->_idx); 2449 } 2450 } 2451 } 2452 2453 // traverse out's that are in the member set 2454 while (true) { 2455 if (idx < n->outcnt()) { 2456 Node* use = n->raw_out(idx); 2457 idx++; 2458 if (!visited.test_set(use->_idx)) { 2459 if (member.test(use->_idx)) { 2460 nstack.push(n, idx); 2461 n = use; 2462 idx = 0; 2463 } 2464 } 2465 } else { 2466 // All outputs processed 2467 sched.push(n); 2468 if (nstack.is_empty()) break; 2469 n = nstack.node(); 2470 idx = nstack.index(); 2471 nstack.pop(); 2472 } 2473 } 2474 } 2475 2476 2477 //------------------------------ has_use_in_set ------------------------------------- 2478 // Has a use in the vector set 2479 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) { 2480 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2481 Node* use = n->fast_out(j); 2482 if (vset.test(use->_idx)) { 2483 return true; 2484 } 2485 } 2486 return false; 2487 } 2488 2489 2490 //------------------------------ has_use_internal_to_set ------------------------------------- 2491 // Has use internal to the vector set (ie. not in a phi at the loop head) 2492 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) { 2493 Node* head = loop->_head; 2494 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2495 Node* use = n->fast_out(j); 2496 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) { 2497 return true; 2498 } 2499 } 2500 return false; 2501 } 2502 2503 2504 //------------------------------ clone_for_use_outside_loop ------------------------------------- 2505 // clone "n" for uses that are outside of loop 2506 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) { 2507 int cloned = 0; 2508 assert(worklist.size() == 0, "should be empty"); 2509 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2510 Node* use = n->fast_out(j); 2511 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) { 2512 worklist.push(use); 2513 } 2514 } 2515 while( worklist.size() ) { 2516 Node *use = worklist.pop(); 2517 if (!has_node(use) || use->in(0) == C->top()) continue; 2518 uint j; 2519 for (j = 0; j < use->req(); j++) { 2520 if (use->in(j) == n) break; 2521 } 2522 assert(j < use->req(), "must be there"); 2523 2524 // clone "n" and insert it between the inputs of "n" and the use outside the loop 2525 Node* n_clone = n->clone(); 2526 _igvn.replace_input_of(use, j, n_clone); 2527 cloned++; 2528 Node* use_c; 2529 if (!use->is_Phi()) { 2530 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0); 2531 } else { 2532 // Use in a phi is considered a use in the associated predecessor block 2533 use_c = use->in(0)->in(j); 2534 } 2535 set_ctrl(n_clone, use_c); 2536 assert(!loop->is_member(get_loop(use_c)), "should be outside loop"); 2537 get_loop(use_c)->_body.push(n_clone); 2538 _igvn.register_new_node_with_optimizer(n_clone); 2539 #if !defined(PRODUCT) 2540 if (TracePartialPeeling) { 2541 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx); 2542 } 2543 #endif 2544 } 2545 return cloned; 2546 } 2547 2548 2549 //------------------------------ clone_for_special_use_inside_loop ------------------------------------- 2550 // clone "n" for special uses that are in the not_peeled region. 2551 // If these def-uses occur in separate blocks, the code generator 2552 // marks the method as not compilable. For example, if a "BoolNode" 2553 // is in a different basic block than the "IfNode" that uses it, then 2554 // the compilation is aborted in the code generator. 2555 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 2556 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) { 2557 if (n->is_Phi() || n->is_Load()) { 2558 return; 2559 } 2560 assert(worklist.size() == 0, "should be empty"); 2561 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2562 Node* use = n->fast_out(j); 2563 if ( not_peel.test(use->_idx) && 2564 (use->is_If() || use->is_CMove() || use->is_Bool()) && 2565 use->in(1) == n) { 2566 worklist.push(use); 2567 } 2568 } 2569 if (worklist.size() > 0) { 2570 // clone "n" and insert it between inputs of "n" and the use 2571 Node* n_clone = n->clone(); 2572 loop->_body.push(n_clone); 2573 _igvn.register_new_node_with_optimizer(n_clone); 2574 set_ctrl(n_clone, get_ctrl(n)); 2575 sink_list.push(n_clone); 2576 not_peel <<= n_clone->_idx; // add n_clone to not_peel set. 2577 #if !defined(PRODUCT) 2578 if (TracePartialPeeling) { 2579 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx); 2580 } 2581 #endif 2582 while( worklist.size() ) { 2583 Node *use = worklist.pop(); 2584 _igvn.rehash_node_delayed(use); 2585 for (uint j = 1; j < use->req(); j++) { 2586 if (use->in(j) == n) { 2587 use->set_req(j, n_clone); 2588 } 2589 } 2590 } 2591 } 2592 } 2593 2594 2595 //------------------------------ insert_phi_for_loop ------------------------------------- 2596 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 2597 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) { 2598 Node *phi = PhiNode::make(lp, back_edge_val); 2599 phi->set_req(LoopNode::EntryControl, lp_entry_val); 2600 // Use existing phi if it already exists 2601 Node *hit = _igvn.hash_find_insert(phi); 2602 if( hit == NULL ) { 2603 _igvn.register_new_node_with_optimizer(phi); 2604 set_ctrl(phi, lp); 2605 } else { 2606 // Remove the new phi from the graph and use the hit 2607 _igvn.remove_dead_node(phi); 2608 phi = hit; 2609 } 2610 _igvn.replace_input_of(use, idx, phi); 2611 } 2612 2613 #ifdef ASSERT 2614 //------------------------------ is_valid_loop_partition ------------------------------------- 2615 // Validate the loop partition sets: peel and not_peel 2616 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, 2617 VectorSet& not_peel ) { 2618 uint i; 2619 // Check that peel_list entries are in the peel set 2620 for (i = 0; i < peel_list.size(); i++) { 2621 if (!peel.test(peel_list.at(i)->_idx)) { 2622 return false; 2623 } 2624 } 2625 // Check at loop members are in one of peel set or not_peel set 2626 for (i = 0; i < loop->_body.size(); i++ ) { 2627 Node *def = loop->_body.at(i); 2628 uint di = def->_idx; 2629 // Check that peel set elements are in peel_list 2630 if (peel.test(di)) { 2631 if (not_peel.test(di)) { 2632 return false; 2633 } 2634 // Must be in peel_list also 2635 bool found = false; 2636 for (uint j = 0; j < peel_list.size(); j++) { 2637 if (peel_list.at(j)->_idx == di) { 2638 found = true; 2639 break; 2640 } 2641 } 2642 if (!found) { 2643 return false; 2644 } 2645 } else if (not_peel.test(di)) { 2646 if (peel.test(di)) { 2647 return false; 2648 } 2649 } else { 2650 return false; 2651 } 2652 } 2653 return true; 2654 } 2655 2656 //------------------------------ is_valid_clone_loop_exit_use ------------------------------------- 2657 // Ensure a use outside of loop is of the right form 2658 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) { 2659 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2660 return (use->is_Phi() && 2661 use_c->is_Region() && use_c->req() == 3 && 2662 (use_c->in(exit_idx)->Opcode() == Op_IfTrue || 2663 use_c->in(exit_idx)->Opcode() == Op_IfFalse || 2664 use_c->in(exit_idx)->Opcode() == Op_JumpProj) && 2665 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) ); 2666 } 2667 2668 //------------------------------ is_valid_clone_loop_form ------------------------------------- 2669 // Ensure that all uses outside of loop are of the right form 2670 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 2671 uint orig_exit_idx, uint clone_exit_idx) { 2672 uint len = peel_list.size(); 2673 for (uint i = 0; i < len; i++) { 2674 Node *def = peel_list.at(i); 2675 2676 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 2677 Node *use = def->fast_out(j); 2678 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2679 if (!loop->is_member(get_loop(use_c))) { 2680 // use is not in the loop, check for correct structure 2681 if (use->in(0) == def) { 2682 // Okay 2683 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) { 2684 return false; 2685 } 2686 } 2687 } 2688 } 2689 return true; 2690 } 2691 #endif 2692 2693 //------------------------------ partial_peel ------------------------------------- 2694 // Partially peel (aka loop rotation) the top portion of a loop (called 2695 // the peel section below) by cloning it and placing one copy just before 2696 // the new loop head and the other copy at the bottom of the new loop. 2697 // 2698 // before after where it came from 2699 // 2700 // stmt1 stmt1 2701 // loop: stmt2 clone 2702 // stmt2 if condA goto exitA clone 2703 // if condA goto exitA new_loop: new 2704 // stmt3 stmt3 clone 2705 // if !condB goto loop if condB goto exitB clone 2706 // exitB: stmt2 orig 2707 // stmt4 if !condA goto new_loop orig 2708 // exitA: goto exitA 2709 // exitB: 2710 // stmt4 2711 // exitA: 2712 // 2713 // Step 1: find the cut point: an exit test on probable 2714 // induction variable. 2715 // Step 2: schedule (with cloning) operations in the peel 2716 // section that can be executed after the cut into 2717 // the section that is not peeled. This may need 2718 // to clone operations into exit blocks. For 2719 // instance, a reference to A[i] in the not-peel 2720 // section and a reference to B[i] in an exit block 2721 // may cause a left-shift of i by 2 to be placed 2722 // in the peel block. This step will clone the left 2723 // shift into the exit block and sink the left shift 2724 // from the peel to the not-peel section. 2725 // Step 3: clone the loop, retarget the control, and insert 2726 // phis for values that are live across the new loop 2727 // head. This is very dependent on the graph structure 2728 // from clone_loop. It creates region nodes for 2729 // exit control and associated phi nodes for values 2730 // flow out of the loop through that exit. The region 2731 // node is dominated by the clone's control projection. 2732 // So the clone's peel section is placed before the 2733 // new loop head, and the clone's not-peel section is 2734 // forms the top part of the new loop. The original 2735 // peel section forms the tail of the new loop. 2736 // Step 4: update the dominator tree and recompute the 2737 // dominator depth. 2738 // 2739 // orig 2740 // 2741 // stmt1 2742 // | 2743 // v 2744 // loop predicate 2745 // | 2746 // v 2747 // loop<----+ 2748 // | | 2749 // stmt2 | 2750 // | | 2751 // v | 2752 // ifA | 2753 // / | | 2754 // v v | 2755 // false true ^ <-- last_peel 2756 // / | | 2757 // / ===|==cut | 2758 // / stmt3 | <-- first_not_peel 2759 // / | | 2760 // | v | 2761 // v ifB | 2762 // exitA: / \ | 2763 // / \ | 2764 // v v | 2765 // false true | 2766 // / \ | 2767 // / ----+ 2768 // | 2769 // v 2770 // exitB: 2771 // stmt4 2772 // 2773 // 2774 // after clone loop 2775 // 2776 // stmt1 2777 // | 2778 // v 2779 // loop predicate 2780 // / \ 2781 // clone / \ orig 2782 // / \ 2783 // / \ 2784 // v v 2785 // +---->loop loop<----+ 2786 // | | | | 2787 // | stmt2 stmt2 | 2788 // | | | | 2789 // | v v | 2790 // | ifA ifA | 2791 // | | \ / | | 2792 // | v v v v | 2793 // ^ true false false true ^ <-- last_peel 2794 // | | ^ \ / | | 2795 // | cut==|== \ \ / ===|==cut | 2796 // | stmt3 \ \ / stmt3 | <-- first_not_peel 2797 // | | dom | | | | 2798 // | v \ 1v v2 v | 2799 // | ifB regionA ifB | 2800 // | / \ | / \ | 2801 // | / \ v / \ | 2802 // | v v exitA: v v | 2803 // | true false false true | 2804 // | / ^ \ / \ | 2805 // +---- \ \ / ----+ 2806 // dom \ / 2807 // \ 1v v2 2808 // regionB 2809 // | 2810 // v 2811 // exitB: 2812 // stmt4 2813 // 2814 // 2815 // after partial peel 2816 // 2817 // stmt1 2818 // | 2819 // v 2820 // loop predicate 2821 // / 2822 // clone / orig 2823 // / TOP 2824 // / \ 2825 // v v 2826 // TOP->loop loop----+ 2827 // | | | 2828 // stmt2 stmt2 | 2829 // | | | 2830 // v v | 2831 // ifA ifA | 2832 // | \ / | | 2833 // v v v v | 2834 // true false false true | <-- last_peel 2835 // | ^ \ / +------|---+ 2836 // +->newloop \ \ / === ==cut | | 2837 // | stmt3 \ \ / TOP | | 2838 // | | dom | | stmt3 | | <-- first_not_peel 2839 // | v \ 1v v2 v | | 2840 // | ifB regionA ifB ^ v 2841 // | / \ | / \ | | 2842 // | / \ v / \ | | 2843 // | v v exitA: v v | | 2844 // | true false false true | | 2845 // | / ^ \ / \ | | 2846 // | | \ \ / v | | 2847 // | | dom \ / TOP | | 2848 // | | \ 1v v2 | | 2849 // ^ v regionB | | 2850 // | | | | | 2851 // | | v ^ v 2852 // | | exitB: | | 2853 // | | stmt4 | | 2854 // | +------------>-----------------+ | 2855 // | | 2856 // +-----------------<---------------------+ 2857 // 2858 // 2859 // final graph 2860 // 2861 // stmt1 2862 // | 2863 // v 2864 // loop predicate 2865 // | 2866 // v 2867 // stmt2 clone 2868 // | 2869 // v 2870 // ........> ifA clone 2871 // : / | 2872 // dom / | 2873 // : v v 2874 // : false true 2875 // : | | 2876 // : | v 2877 // : | newloop<-----+ 2878 // : | | | 2879 // : | stmt3 clone | 2880 // : | | | 2881 // : | v | 2882 // : | ifB | 2883 // : | / \ | 2884 // : | v v | 2885 // : | false true | 2886 // : | | | | 2887 // : | v stmt2 | 2888 // : | exitB: | | 2889 // : | stmt4 v | 2890 // : | ifA orig | 2891 // : | / \ | 2892 // : | / \ | 2893 // : | v v | 2894 // : | false true | 2895 // : | / \ | 2896 // : v v -----+ 2897 // RegionA 2898 // | 2899 // v 2900 // exitA 2901 // 2902 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) { 2903 2904 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only"); 2905 if (!loop->_head->is_Loop()) { 2906 return false; } 2907 2908 LoopNode *head = loop->_head->as_Loop(); 2909 2910 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) { 2911 return false; 2912 } 2913 2914 // Check for complex exit control 2915 for(uint ii = 0; ii < loop->_body.size(); ii++ ) { 2916 Node *n = loop->_body.at(ii); 2917 int opc = n->Opcode(); 2918 if (n->is_Call() || 2919 opc == Op_Catch || 2920 opc == Op_CatchProj || 2921 opc == Op_Jump || 2922 opc == Op_JumpProj) { 2923 #if !defined(PRODUCT) 2924 if (TracePartialPeeling) { 2925 tty->print_cr("\nExit control too complex: lp: %d", head->_idx); 2926 } 2927 #endif 2928 return false; 2929 } 2930 } 2931 2932 int dd = dom_depth(head); 2933 2934 // Step 1: find cut point 2935 2936 // Walk up dominators to loop head looking for first loop exit 2937 // which is executed on every path thru loop. 2938 IfNode *peel_if = NULL; 2939 IfNode *peel_if_cmpu = NULL; 2940 2941 Node *iff = loop->tail(); 2942 while( iff != head ) { 2943 if( iff->is_If() ) { 2944 Node *ctrl = get_ctrl(iff->in(1)); 2945 if (ctrl->is_top()) return false; // Dead test on live IF. 2946 // If loop-varying exit-test, check for induction variable 2947 if( loop->is_member(get_loop(ctrl)) && 2948 loop->is_loop_exit(iff) && 2949 is_possible_iv_test(iff)) { 2950 Node* cmp = iff->in(1)->in(1); 2951 if (cmp->Opcode() == Op_CmpI) { 2952 peel_if = iff->as_If(); 2953 } else { 2954 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU"); 2955 peel_if_cmpu = iff->as_If(); 2956 } 2957 } 2958 } 2959 iff = idom(iff); 2960 } 2961 // Prefer signed compare over unsigned compare. 2962 IfNode* new_peel_if = NULL; 2963 if (peel_if == NULL) { 2964 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) { 2965 return false; // No peel point found 2966 } 2967 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop); 2968 if (new_peel_if == NULL) { 2969 return false; // No peel point found 2970 } 2971 peel_if = new_peel_if; 2972 } 2973 Node* last_peel = stay_in_loop(peel_if, loop); 2974 Node* first_not_peeled = stay_in_loop(last_peel, loop); 2975 if (first_not_peeled == NULL || first_not_peeled == head) { 2976 return false; 2977 } 2978 2979 #if !defined(PRODUCT) 2980 if (TraceLoopOpts) { 2981 tty->print("PartialPeel "); 2982 loop->dump_head(); 2983 } 2984 2985 if (TracePartialPeeling) { 2986 tty->print_cr("before partial peel one iteration"); 2987 Node_List wl; 2988 Node* t = head->in(2); 2989 while (true) { 2990 wl.push(t); 2991 if (t == head) break; 2992 t = idom(t); 2993 } 2994 while (wl.size() > 0) { 2995 Node* tt = wl.pop(); 2996 tt->dump(); 2997 if (tt == last_peel) tty->print_cr("-- cut --"); 2998 } 2999 } 3000 #endif 3001 ResourceArea *area = Thread::current()->resource_area(); 3002 VectorSet peel(area); 3003 VectorSet not_peel(area); 3004 Node_List peel_list(area); 3005 Node_List worklist(area); 3006 Node_List sink_list(area); 3007 3008 // Set of cfg nodes to peel are those that are executable from 3009 // the head through last_peel. 3010 assert(worklist.size() == 0, "should be empty"); 3011 worklist.push(head); 3012 peel.set(head->_idx); 3013 while (worklist.size() > 0) { 3014 Node *n = worklist.pop(); 3015 if (n != last_peel) { 3016 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3017 Node* use = n->fast_out(j); 3018 if (use->is_CFG() && 3019 loop->is_member(get_loop(use)) && 3020 !peel.test_set(use->_idx)) { 3021 worklist.push(use); 3022 } 3023 } 3024 } 3025 } 3026 3027 // Set of non-cfg nodes to peel are those that are control 3028 // dependent on the cfg nodes. 3029 uint i; 3030 for(i = 0; i < loop->_body.size(); i++ ) { 3031 Node *n = loop->_body.at(i); 3032 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n; 3033 if (peel.test(n_c->_idx)) { 3034 peel.set(n->_idx); 3035 } else { 3036 not_peel.set(n->_idx); 3037 } 3038 } 3039 3040 // Step 2: move operations from the peeled section down into the 3041 // not-peeled section 3042 3043 // Get a post order schedule of nodes in the peel region 3044 // Result in right-most operand. 3045 scheduled_nodelist(loop, peel, peel_list ); 3046 3047 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 3048 3049 // For future check for too many new phis 3050 uint old_phi_cnt = 0; 3051 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 3052 Node* use = head->fast_out(j); 3053 if (use->is_Phi()) old_phi_cnt++; 3054 } 3055 3056 #if !defined(PRODUCT) 3057 if (TracePartialPeeling) { 3058 tty->print_cr("\npeeled list"); 3059 } 3060 #endif 3061 3062 // Evacuate nodes in peel region into the not_peeled region if possible 3063 uint new_phi_cnt = 0; 3064 uint cloned_for_outside_use = 0; 3065 for (i = 0; i < peel_list.size();) { 3066 Node* n = peel_list.at(i); 3067 #if !defined(PRODUCT) 3068 if (TracePartialPeeling) n->dump(); 3069 #endif 3070 bool incr = true; 3071 if ( !n->is_CFG() ) { 3072 3073 if ( has_use_in_set(n, not_peel) ) { 3074 3075 // If not used internal to the peeled region, 3076 // move "n" from peeled to not_peeled region. 3077 3078 if ( !has_use_internal_to_set(n, peel, loop) ) { 3079 3080 // if not pinned and not a load (which maybe anti-dependent on a store) 3081 // and not a CMove (Matcher expects only bool->cmove). 3082 if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) { 3083 cloned_for_outside_use += clone_for_use_outside_loop( loop, n, worklist ); 3084 sink_list.push(n); 3085 peel >>= n->_idx; // delete n from peel set. 3086 not_peel <<= n->_idx; // add n to not_peel set. 3087 peel_list.remove(i); 3088 incr = false; 3089 #if !defined(PRODUCT) 3090 if (TracePartialPeeling) { 3091 tty->print_cr("sink to not_peeled region: %d newbb: %d", 3092 n->_idx, get_ctrl(n)->_idx); 3093 } 3094 #endif 3095 } 3096 } else { 3097 // Otherwise check for special def-use cases that span 3098 // the peel/not_peel boundary such as bool->if 3099 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist ); 3100 new_phi_cnt++; 3101 } 3102 } 3103 } 3104 if (incr) i++; 3105 } 3106 3107 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) { 3108 #if !defined(PRODUCT) 3109 if (TracePartialPeeling) { 3110 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c", 3111 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F'); 3112 } 3113 #endif 3114 if (new_peel_if != NULL) { 3115 remove_cmpi_loop_exit(new_peel_if, loop); 3116 } 3117 // Inhibit more partial peeling on this loop 3118 assert(!head->is_partial_peel_loop(), "not partial peeled"); 3119 head->mark_partial_peel_failed(); 3120 if (cloned_for_outside_use > 0) { 3121 // Terminate this round of loop opts because 3122 // the graph outside this loop was changed. 3123 C->set_major_progress(); 3124 return true; 3125 } 3126 return false; 3127 } 3128 3129 // Step 3: clone loop, retarget control, and insert new phis 3130 3131 // Create new loop head for new phis and to hang 3132 // the nodes being moved (sinked) from the peel region. 3133 LoopNode* new_head = new LoopNode(last_peel, last_peel); 3134 new_head->set_unswitch_count(head->unswitch_count()); // Preserve 3135 _igvn.register_new_node_with_optimizer(new_head); 3136 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled"); 3137 _igvn.replace_input_of(first_not_peeled, 0, new_head); 3138 set_loop(new_head, loop); 3139 loop->_body.push(new_head); 3140 not_peel.set(new_head->_idx); 3141 set_idom(new_head, last_peel, dom_depth(first_not_peeled)); 3142 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled)); 3143 3144 while (sink_list.size() > 0) { 3145 Node* n = sink_list.pop(); 3146 set_ctrl(n, new_head); 3147 } 3148 3149 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 3150 3151 clone_loop(loop, old_new, dd, IgnoreStripMined); 3152 3153 const uint clone_exit_idx = 1; 3154 const uint orig_exit_idx = 2; 3155 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop"); 3156 3157 Node* head_clone = old_new[head->_idx]; 3158 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop(); 3159 Node* orig_tail_clone = head_clone->in(2); 3160 3161 // Add phi if "def" node is in peel set and "use" is not 3162 3163 for(i = 0; i < peel_list.size(); i++ ) { 3164 Node *def = peel_list.at(i); 3165 if (!def->is_CFG()) { 3166 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 3167 Node *use = def->fast_out(j); 3168 if (has_node(use) && use->in(0) != C->top() && 3169 (!peel.test(use->_idx) || 3170 (use->is_Phi() && use->in(0) == head)) ) { 3171 worklist.push(use); 3172 } 3173 } 3174 while( worklist.size() ) { 3175 Node *use = worklist.pop(); 3176 for (uint j = 1; j < use->req(); j++) { 3177 Node* n = use->in(j); 3178 if (n == def) { 3179 3180 // "def" is in peel set, "use" is not in peel set 3181 // or "use" is in the entry boundary (a phi) of the peel set 3182 3183 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use; 3184 3185 if ( loop->is_member(get_loop( use_c )) ) { 3186 // use is in loop 3187 if (old_new[use->_idx] != NULL) { // null for dead code 3188 Node* use_clone = old_new[use->_idx]; 3189 _igvn.replace_input_of(use, j, C->top()); 3190 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone ); 3191 } 3192 } else { 3193 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format"); 3194 // use is not in the loop, check if the live range includes the cut 3195 Node* lp_if = use_c->in(orig_exit_idx)->in(0); 3196 if (not_peel.test(lp_if->_idx)) { 3197 assert(j == orig_exit_idx, "use from original loop"); 3198 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone ); 3199 } 3200 } 3201 } 3202 } 3203 } 3204 } 3205 } 3206 3207 // Step 3b: retarget control 3208 3209 // Redirect control to the new loop head if a cloned node in 3210 // the not_peeled region has control that points into the peeled region. 3211 // This necessary because the cloned peeled region will be outside 3212 // the loop. 3213 // from to 3214 // cloned-peeled <---+ 3215 // new_head_clone: | <--+ 3216 // cloned-not_peeled in(0) in(0) 3217 // orig-peeled 3218 3219 for(i = 0; i < loop->_body.size(); i++ ) { 3220 Node *n = loop->_body.at(i); 3221 if (!n->is_CFG() && n->in(0) != NULL && 3222 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) { 3223 Node* n_clone = old_new[n->_idx]; 3224 _igvn.replace_input_of(n_clone, 0, new_head_clone); 3225 } 3226 } 3227 3228 // Backedge of the surviving new_head (the clone) is original last_peel 3229 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel); 3230 3231 // Cut first node in original not_peel set 3232 _igvn.rehash_node_delayed(new_head); // Multiple edge updates: 3233 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of 3234 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls 3235 3236 // Copy head_clone back-branch info to original head 3237 // and remove original head's loop entry and 3238 // clone head's back-branch 3239 _igvn.rehash_node_delayed(head); // Multiple edge updates 3240 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl)); 3241 head->set_req(LoopNode::LoopBackControl, C->top()); 3242 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top()); 3243 3244 // Similarly modify the phis 3245 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) { 3246 Node* use = head->fast_out(k); 3247 if (use->is_Phi() && use->outcnt() > 0) { 3248 Node* use_clone = old_new[use->_idx]; 3249 _igvn.rehash_node_delayed(use); // Multiple edge updates 3250 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl)); 3251 use->set_req(LoopNode::LoopBackControl, C->top()); 3252 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top()); 3253 } 3254 } 3255 3256 // Step 4: update dominator tree and dominator depth 3257 3258 set_idom(head, orig_tail_clone, dd); 3259 recompute_dom_depth(); 3260 3261 // Inhibit more partial peeling on this loop 3262 new_head_clone->set_partial_peel_loop(); 3263 C->set_major_progress(); 3264 loop->record_for_igvn(); 3265 3266 #if !defined(PRODUCT) 3267 if (TracePartialPeeling) { 3268 tty->print_cr("\nafter partial peel one iteration"); 3269 Node_List wl(area); 3270 Node* t = last_peel; 3271 while (true) { 3272 wl.push(t); 3273 if (t == head_clone) break; 3274 t = idom(t); 3275 } 3276 while (wl.size() > 0) { 3277 Node* tt = wl.pop(); 3278 if (tt == head) tty->print_cr("orig head"); 3279 else if (tt == new_head_clone) tty->print_cr("new head"); 3280 else if (tt == head_clone) tty->print_cr("clone head"); 3281 tt->dump(); 3282 } 3283 } 3284 #endif 3285 return true; 3286 } 3287 3288 //------------------------------reorg_offsets---------------------------------- 3289 // Reorganize offset computations to lower register pressure. Mostly 3290 // prevent loop-fallout uses of the pre-incremented trip counter (which are 3291 // then alive with the post-incremented trip counter forcing an extra 3292 // register move) 3293 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) { 3294 // Perform it only for canonical counted loops. 3295 // Loop's shape could be messed up by iteration_split_impl. 3296 if (!loop->_head->is_CountedLoop()) 3297 return; 3298 if (!loop->_head->as_Loop()->is_valid_counted_loop()) 3299 return; 3300 3301 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 3302 CountedLoopEndNode *cle = cl->loopexit(); 3303 Node *exit = cle->proj_out(false); 3304 Node *phi = cl->phi(); 3305 3306 // Check for the special case of folks using the pre-incremented 3307 // trip-counter on the fall-out path (forces the pre-incremented 3308 // and post-incremented trip counter to be live at the same time). 3309 // Fix this by adjusting to use the post-increment trip counter. 3310 3311 bool progress = true; 3312 while (progress) { 3313 progress = false; 3314 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) { 3315 Node* use = phi->fast_out(i); // User of trip-counter 3316 if (!has_ctrl(use)) continue; 3317 Node *u_ctrl = get_ctrl(use); 3318 if (use->is_Phi()) { 3319 u_ctrl = NULL; 3320 for (uint j = 1; j < use->req(); j++) 3321 if (use->in(j) == phi) 3322 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j)); 3323 } 3324 IdealLoopTree *u_loop = get_loop(u_ctrl); 3325 // Look for loop-invariant use 3326 if (u_loop == loop) continue; 3327 if (loop->is_member(u_loop)) continue; 3328 // Check that use is live out the bottom. Assuming the trip-counter 3329 // update is right at the bottom, uses of of the loop middle are ok. 3330 if (dom_lca(exit, u_ctrl) != exit) continue; 3331 // Hit! Refactor use to use the post-incremented tripcounter. 3332 // Compute a post-increment tripcounter. 3333 Node *opaq = new Opaque2Node( C, cle->incr() ); 3334 register_new_node(opaq, exit); 3335 Node *neg_stride = _igvn.intcon(-cle->stride_con()); 3336 set_ctrl(neg_stride, C->root()); 3337 Node *post = new AddINode( opaq, neg_stride); 3338 register_new_node(post, exit); 3339 _igvn.rehash_node_delayed(use); 3340 for (uint j = 1; j < use->req(); j++) { 3341 if (use->in(j) == phi) 3342 use->set_req(j, post); 3343 } 3344 // Since DU info changed, rerun loop 3345 progress = true; 3346 break; 3347 } 3348 } 3349 3350 }