1 /* 2 * Copyright (c) 1999, 2018, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "memory/allocation.inline.hpp" 27 #include "memory/resourceArea.hpp" 28 #include "opto/addnode.hpp" 29 #include "opto/callnode.hpp" 30 #include "opto/castnode.hpp" 31 #include "opto/connode.hpp" 32 #include "opto/castnode.hpp" 33 #include "opto/divnode.hpp" 34 #include "opto/loopnode.hpp" 35 #include "opto/matcher.hpp" 36 #include "opto/mulnode.hpp" 37 #include "opto/movenode.hpp" 38 #include "opto/opaquenode.hpp" 39 #include "opto/rootnode.hpp" 40 #include "opto/subnode.hpp" 41 42 //============================================================================= 43 //------------------------------split_thru_phi--------------------------------- 44 // Split Node 'n' through merge point if there is enough win. 45 Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) { 46 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) { 47 // ConvI2L may have type information on it which is unsafe to push up 48 // so disable this for now 49 return NULL; 50 } 51 52 // Splitting range check CastIIs through a loop induction Phi can 53 // cause new Phis to be created that are left unrelated to the loop 54 // induction Phi and prevent optimizations (vectorization) 55 if (n->Opcode() == Op_CastII && n->as_CastII()->has_range_check() && 56 region->is_CountedLoop() && n->in(1) == region->as_CountedLoop()->phi()) { 57 return NULL; 58 } 59 60 int wins = 0; 61 assert(!n->is_CFG(), ""); 62 assert(region->is_Region(), ""); 63 64 const Type* type = n->bottom_type(); 65 const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr(); 66 Node *phi; 67 if (t_oop != NULL && t_oop->is_known_instance_field()) { 68 int iid = t_oop->instance_id(); 69 int index = C->get_alias_index(t_oop); 70 int offset = t_oop->offset(); 71 phi = new PhiNode(region, type, NULL, iid, index, offset); 72 } else { 73 phi = PhiNode::make_blank(region, n); 74 } 75 uint old_unique = C->unique(); 76 for (uint i = 1; i < region->req(); i++) { 77 Node *x; 78 Node* the_clone = NULL; 79 if (region->in(i) == C->top()) { 80 x = C->top(); // Dead path? Use a dead data op 81 } else { 82 x = n->clone(); // Else clone up the data op 83 the_clone = x; // Remember for possible deletion. 84 // Alter data node to use pre-phi inputs 85 if (n->in(0) == region) 86 x->set_req( 0, region->in(i) ); 87 for (uint j = 1; j < n->req(); j++) { 88 Node *in = n->in(j); 89 if (in->is_Phi() && in->in(0) == region) 90 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone 91 } 92 } 93 // Check for a 'win' on some paths 94 const Type *t = x->Value(&_igvn); 95 96 bool singleton = t->singleton(); 97 98 // A TOP singleton indicates that there are no possible values incoming 99 // along a particular edge. In most cases, this is OK, and the Phi will 100 // be eliminated later in an Ideal call. However, we can't allow this to 101 // happen if the singleton occurs on loop entry, as the elimination of 102 // the PhiNode may cause the resulting node to migrate back to a previous 103 // loop iteration. 104 if (singleton && t == Type::TOP) { 105 // Is_Loop() == false does not confirm the absence of a loop (e.g., an 106 // irreducible loop may not be indicated by an affirmative is_Loop()); 107 // therefore, the only top we can split thru a phi is on a backedge of 108 // a loop. 109 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 110 } 111 112 if (singleton) { 113 wins++; 114 x = ((PhaseGVN&)_igvn).makecon(t); 115 } else { 116 // We now call Identity to try to simplify the cloned node. 117 // Note that some Identity methods call phase->type(this). 118 // Make sure that the type array is big enough for 119 // our new node, even though we may throw the node away. 120 // (Note: This tweaking with igvn only works because x is a new node.) 121 _igvn.set_type(x, t); 122 // If x is a TypeNode, capture any more-precise type permanently into Node 123 // otherwise it will be not updated during igvn->transform since 124 // igvn->type(x) is set to x->Value() already. 125 x->raise_bottom_type(t); 126 Node *y = x->Identity(&_igvn); 127 if (y != x) { 128 wins++; 129 x = y; 130 } else { 131 y = _igvn.hash_find(x); 132 if (y) { 133 wins++; 134 x = y; 135 } else { 136 // Else x is a new node we are keeping 137 // We do not need register_new_node_with_optimizer 138 // because set_type has already been called. 139 _igvn._worklist.push(x); 140 } 141 } 142 } 143 if (x != the_clone && the_clone != NULL) 144 _igvn.remove_dead_node(the_clone); 145 phi->set_req( i, x ); 146 } 147 // Too few wins? 148 if (wins <= policy) { 149 _igvn.remove_dead_node(phi); 150 return NULL; 151 } 152 153 // Record Phi 154 register_new_node( phi, region ); 155 156 for (uint i2 = 1; i2 < phi->req(); i2++) { 157 Node *x = phi->in(i2); 158 // If we commoned up the cloned 'x' with another existing Node, 159 // the existing Node picks up a new use. We need to make the 160 // existing Node occur higher up so it dominates its uses. 161 Node *old_ctrl; 162 IdealLoopTree *old_loop; 163 164 if (x->is_Con()) { 165 // Constant's control is always root. 166 set_ctrl(x, C->root()); 167 continue; 168 } 169 // The occasional new node 170 if (x->_idx >= old_unique) { // Found a new, unplaced node? 171 old_ctrl = NULL; 172 old_loop = NULL; // Not in any prior loop 173 } else { 174 old_ctrl = get_ctrl(x); 175 old_loop = get_loop(old_ctrl); // Get prior loop 176 } 177 // New late point must dominate new use 178 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2)); 179 if (new_ctrl == old_ctrl) // Nothing is changed 180 continue; 181 182 IdealLoopTree *new_loop = get_loop(new_ctrl); 183 184 // Don't move x into a loop if its uses are 185 // outside of loop. Otherwise x will be cloned 186 // for each use outside of this loop. 187 IdealLoopTree *use_loop = get_loop(region); 188 if (!new_loop->is_member(use_loop) && 189 (old_loop == NULL || !new_loop->is_member(old_loop))) { 190 // Take early control, later control will be recalculated 191 // during next iteration of loop optimizations. 192 new_ctrl = get_early_ctrl(x); 193 new_loop = get_loop(new_ctrl); 194 } 195 // Set new location 196 set_ctrl(x, new_ctrl); 197 // If changing loop bodies, see if we need to collect into new body 198 if (old_loop != new_loop) { 199 if (old_loop && !old_loop->_child) 200 old_loop->_body.yank(x); 201 if (!new_loop->_child) 202 new_loop->_body.push(x); // Collect body info 203 } 204 } 205 206 return phi; 207 } 208 209 //------------------------------dominated_by------------------------------------ 210 // Replace the dominated test with an obvious true or false. Place it on the 211 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the 212 // live path up to the dominating control. 213 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff, bool flip, bool exclude_loop_predicate ) { 214 if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); } 215 216 // prevdom is the dominating projection of the dominating test. 217 assert( iff->is_If(), "" ); 218 assert(iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd || iff->Opcode() == Op_RangeCheck, "Check this code when new subtype is added"); 219 int pop = prevdom->Opcode(); 220 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" ); 221 if (flip) { 222 if (pop == Op_IfTrue) 223 pop = Op_IfFalse; 224 else 225 pop = Op_IfTrue; 226 } 227 // 'con' is set to true or false to kill the dominated test. 228 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO); 229 set_ctrl(con, C->root()); // Constant gets a new use 230 // Hack the dominated test 231 _igvn.replace_input_of(iff, 1, con); 232 233 // If I dont have a reachable TRUE and FALSE path following the IfNode then 234 // I can assume this path reaches an infinite loop. In this case it's not 235 // important to optimize the data Nodes - either the whole compilation will 236 // be tossed or this path (and all data Nodes) will go dead. 237 if (iff->outcnt() != 2) return; 238 239 // Make control-dependent data Nodes on the live path (path that will remain 240 // once the dominated IF is removed) become control-dependent on the 241 // dominating projection. 242 Node* dp = iff->as_If()->proj_out_or_null(pop == Op_IfTrue); 243 244 // Loop predicates may have depending checks which should not 245 // be skipped. For example, range check predicate has two checks 246 // for lower and upper bounds. 247 if (dp == NULL) 248 return; 249 250 ProjNode* dp_proj = dp->as_Proj(); 251 ProjNode* unc_proj = iff->as_If()->proj_out(1 - dp_proj->_con)->as_Proj(); 252 if (exclude_loop_predicate && 253 (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) != NULL || 254 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check) != NULL)) { 255 // If this is a range check (IfNode::is_range_check), do not 256 // reorder because Compile::allow_range_check_smearing might have 257 // changed the check. 258 return; // Let IGVN transformation change control dependence. 259 } 260 261 IdealLoopTree *old_loop = get_loop(dp); 262 263 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { 264 Node* cd = dp->fast_out(i); // Control-dependent node 265 if (cd->depends_only_on_test()) { 266 assert(cd->in(0) == dp, ""); 267 _igvn.replace_input_of(cd, 0, prevdom); 268 set_early_ctrl(cd); 269 IdealLoopTree *new_loop = get_loop(get_ctrl(cd)); 270 if (old_loop != new_loop) { 271 if (!old_loop->_child) old_loop->_body.yank(cd); 272 if (!new_loop->_child) new_loop->_body.push(cd); 273 } 274 --i; 275 --imax; 276 } 277 } 278 } 279 280 //------------------------------has_local_phi_input---------------------------- 281 // Return TRUE if 'n' has Phi inputs from its local block and no other 282 // block-local inputs (all non-local-phi inputs come from earlier blocks) 283 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) { 284 Node *n_ctrl = get_ctrl(n); 285 // See if some inputs come from a Phi in this block, or from before 286 // this block. 287 uint i; 288 for( i = 1; i < n->req(); i++ ) { 289 Node *phi = n->in(i); 290 if( phi->is_Phi() && phi->in(0) == n_ctrl ) 291 break; 292 } 293 if( i >= n->req() ) 294 return NULL; // No Phi inputs; nowhere to clone thru 295 296 // Check for inputs created between 'n' and the Phi input. These 297 // must split as well; they have already been given the chance 298 // (courtesy of a post-order visit) and since they did not we must 299 // recover the 'cost' of splitting them by being very profitable 300 // when splitting 'n'. Since this is unlikely we simply give up. 301 for( i = 1; i < n->req(); i++ ) { 302 Node *m = n->in(i); 303 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) { 304 // We allow the special case of AddP's with no local inputs. 305 // This allows us to split-up address expressions. 306 if (m->is_AddP() && 307 get_ctrl(m->in(2)) != n_ctrl && 308 get_ctrl(m->in(3)) != n_ctrl) { 309 // Move the AddP up to dominating point 310 Node* c = find_non_split_ctrl(idom(n_ctrl)); 311 if (c->is_OuterStripMinedLoop()) { 312 c->as_Loop()->verify_strip_mined(1); 313 c = c->in(LoopNode::EntryControl); 314 } 315 set_ctrl_and_loop(m, c); 316 continue; 317 } 318 return NULL; 319 } 320 assert(m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control"); 321 } 322 323 return n_ctrl; 324 } 325 326 //------------------------------remix_address_expressions---------------------- 327 // Rework addressing expressions to get the most loop-invariant stuff 328 // moved out. We'd like to do all associative operators, but it's especially 329 // important (common) to do address expressions. 330 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) { 331 if (!has_ctrl(n)) return NULL; 332 Node *n_ctrl = get_ctrl(n); 333 IdealLoopTree *n_loop = get_loop(n_ctrl); 334 335 // See if 'n' mixes loop-varying and loop-invariant inputs and 336 // itself is loop-varying. 337 338 // Only interested in binary ops (and AddP) 339 if( n->req() < 3 || n->req() > 4 ) return NULL; 340 341 Node *n1_ctrl = get_ctrl(n->in( 1)); 342 Node *n2_ctrl = get_ctrl(n->in( 2)); 343 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3)); 344 IdealLoopTree *n1_loop = get_loop( n1_ctrl ); 345 IdealLoopTree *n2_loop = get_loop( n2_ctrl ); 346 IdealLoopTree *n3_loop = get_loop( n3_ctrl ); 347 348 // Does one of my inputs spin in a tighter loop than self? 349 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) || 350 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) || 351 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) ) 352 return NULL; // Leave well enough alone 353 354 // Is at least one of my inputs loop-invariant? 355 if( n1_loop == n_loop && 356 n2_loop == n_loop && 357 n3_loop == n_loop ) 358 return NULL; // No loop-invariant inputs 359 360 361 int n_op = n->Opcode(); 362 363 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2). 364 if( n_op == Op_LShiftI ) { 365 // Scale is loop invariant 366 Node *scale = n->in(2); 367 Node *scale_ctrl = get_ctrl(scale); 368 IdealLoopTree *scale_loop = get_loop(scale_ctrl ); 369 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) ) 370 return NULL; 371 const TypeInt *scale_t = scale->bottom_type()->isa_int(); 372 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 ) 373 return NULL; // Dont bother with byte/short masking 374 // Add must vary with loop (else shift would be loop-invariant) 375 Node *add = n->in(1); 376 Node *add_ctrl = get_ctrl(add); 377 IdealLoopTree *add_loop = get_loop(add_ctrl); 378 //assert( n_loop == add_loop, "" ); 379 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops 380 381 // Convert I-V into I+ (0-V); same for V-I 382 if( add->Opcode() == Op_SubI && 383 _igvn.type( add->in(1) ) != TypeInt::ZERO ) { 384 Node *zero = _igvn.intcon(0); 385 set_ctrl(zero, C->root()); 386 Node *neg = new SubINode( _igvn.intcon(0), add->in(2) ); 387 register_new_node( neg, get_ctrl(add->in(2) ) ); 388 add = new AddINode( add->in(1), neg ); 389 register_new_node( add, add_ctrl ); 390 } 391 if( add->Opcode() != Op_AddI ) return NULL; 392 // See if one add input is loop invariant 393 Node *add_var = add->in(1); 394 Node *add_var_ctrl = get_ctrl(add_var); 395 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 396 Node *add_invar = add->in(2); 397 Node *add_invar_ctrl = get_ctrl(add_invar); 398 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl ); 399 if( add_var_loop == n_loop ) { 400 } else if( add_invar_loop == n_loop ) { 401 // Swap to find the invariant part 402 add_invar = add_var; 403 add_invar_ctrl = add_var_ctrl; 404 add_invar_loop = add_var_loop; 405 add_var = add->in(2); 406 Node *add_var_ctrl = get_ctrl(add_var); 407 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 408 } else // Else neither input is loop invariant 409 return NULL; 410 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) ) 411 return NULL; // No invariant part of the add? 412 413 // Yes! Reshape address expression! 414 Node *inv_scale = new LShiftINode( add_invar, scale ); 415 Node *inv_scale_ctrl = 416 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ? 417 add_invar_ctrl : scale_ctrl; 418 register_new_node( inv_scale, inv_scale_ctrl ); 419 Node *var_scale = new LShiftINode( add_var, scale ); 420 register_new_node( var_scale, n_ctrl ); 421 Node *var_add = new AddINode( var_scale, inv_scale ); 422 register_new_node( var_add, n_ctrl ); 423 _igvn.replace_node( n, var_add ); 424 return var_add; 425 } 426 427 // Replace (I+V) with (V+I) 428 if( n_op == Op_AddI || 429 n_op == Op_AddL || 430 n_op == Op_AddF || 431 n_op == Op_AddD || 432 n_op == Op_MulI || 433 n_op == Op_MulL || 434 n_op == Op_MulF || 435 n_op == Op_MulD ) { 436 if( n2_loop == n_loop ) { 437 assert( n1_loop != n_loop, "" ); 438 n->swap_edges(1, 2); 439 } 440 } 441 442 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V), 443 // but not if I2 is a constant. 444 if( n_op == Op_AddP ) { 445 if( n2_loop == n_loop && n3_loop != n_loop ) { 446 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) { 447 Node *n22_ctrl = get_ctrl(n->in(2)->in(2)); 448 Node *n23_ctrl = get_ctrl(n->in(2)->in(3)); 449 IdealLoopTree *n22loop = get_loop( n22_ctrl ); 450 IdealLoopTree *n23_loop = get_loop( n23_ctrl ); 451 if( n22loop != n_loop && n22loop->is_member(n_loop) && 452 n23_loop == n_loop ) { 453 Node *add1 = new AddPNode( n->in(1), n->in(2)->in(2), n->in(3) ); 454 // Stuff new AddP in the loop preheader 455 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) ); 456 Node *add2 = new AddPNode( n->in(1), add1, n->in(2)->in(3) ); 457 register_new_node( add2, n_ctrl ); 458 _igvn.replace_node( n, add2 ); 459 return add2; 460 } 461 } 462 } 463 464 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V) 465 if (n2_loop != n_loop && n3_loop == n_loop) { 466 if (n->in(3)->Opcode() == Op_AddX) { 467 Node *V = n->in(3)->in(1); 468 Node *I = n->in(3)->in(2); 469 if (is_member(n_loop,get_ctrl(V))) { 470 } else { 471 Node *tmp = V; V = I; I = tmp; 472 } 473 if (!is_member(n_loop,get_ctrl(I))) { 474 Node *add1 = new AddPNode(n->in(1), n->in(2), I); 475 // Stuff new AddP in the loop preheader 476 register_new_node(add1, n_loop->_head->in(LoopNode::EntryControl)); 477 Node *add2 = new AddPNode(n->in(1), add1, V); 478 register_new_node(add2, n_ctrl); 479 _igvn.replace_node(n, add2); 480 return add2; 481 } 482 } 483 } 484 } 485 486 return NULL; 487 } 488 489 //------------------------------conditional_move------------------------------- 490 // Attempt to replace a Phi with a conditional move. We have some pretty 491 // strict profitability requirements. All Phis at the merge point must 492 // be converted, so we can remove the control flow. We need to limit the 493 // number of c-moves to a small handful. All code that was in the side-arms 494 // of the CFG diamond is now speculatively executed. This code has to be 495 // "cheap enough". We are pretty much limited to CFG diamonds that merge 496 // 1 or 2 items with a total of 1 or 2 ops executed speculatively. 497 Node *PhaseIdealLoop::conditional_move( Node *region ) { 498 499 assert(region->is_Region(), "sanity check"); 500 if (region->req() != 3) return NULL; 501 502 // Check for CFG diamond 503 Node *lp = region->in(1); 504 Node *rp = region->in(2); 505 if (!lp || !rp) return NULL; 506 Node *lp_c = lp->in(0); 507 if (lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If()) return NULL; 508 IfNode *iff = lp_c->as_If(); 509 510 // Check for ops pinned in an arm of the diamond. 511 // Can't remove the control flow in this case 512 if (lp->outcnt() > 1) return NULL; 513 if (rp->outcnt() > 1) return NULL; 514 515 IdealLoopTree* r_loop = get_loop(region); 516 assert(r_loop == get_loop(iff), "sanity"); 517 // Always convert to CMOVE if all results are used only outside this loop. 518 bool used_inside_loop = (r_loop == _ltree_root); 519 520 // Check profitability 521 int cost = 0; 522 int phis = 0; 523 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 524 Node *out = region->fast_out(i); 525 if (!out->is_Phi()) continue; // Ignore other control edges, etc 526 phis++; 527 PhiNode* phi = out->as_Phi(); 528 BasicType bt = phi->type()->basic_type(); 529 switch (bt) { 530 case T_DOUBLE: 531 case T_FLOAT: 532 if (C->use_cmove()) { 533 continue; //TODO: maybe we want to add some cost 534 } 535 cost += Matcher::float_cmove_cost(); // Could be very expensive 536 break; 537 case T_LONG: { 538 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's 539 } 540 case T_INT: // These all CMOV fine 541 case T_ADDRESS: { // (RawPtr) 542 cost++; 543 break; 544 } 545 case T_NARROWOOP: // Fall through 546 case T_OBJECT: { // Base oops are OK, but not derived oops 547 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr(); 548 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a 549 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus 550 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we 551 // have a Phi for the base here that we convert to a CMOVE all is well 552 // and good. But if the base is dead, we'll not make a CMOVE. Later 553 // the allocator will have to produce a base by creating a CMOVE of the 554 // relevant bases. This puts the allocator in the business of 555 // manufacturing expensive instructions, generally a bad plan. 556 // Just Say No to Conditionally-Moved Derived Pointers. 557 if (tp && tp->offset() != 0) 558 return NULL; 559 cost++; 560 break; 561 } 562 default: 563 return NULL; // In particular, can't do memory or I/O 564 } 565 // Add in cost any speculative ops 566 for (uint j = 1; j < region->req(); j++) { 567 Node *proj = region->in(j); 568 Node *inp = phi->in(j); 569 if (get_ctrl(inp) == proj) { // Found local op 570 cost++; 571 // Check for a chain of dependent ops; these will all become 572 // speculative in a CMOV. 573 for (uint k = 1; k < inp->req(); k++) 574 if (get_ctrl(inp->in(k)) == proj) 575 cost += ConditionalMoveLimit; // Too much speculative goo 576 } 577 } 578 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode. 579 // This will likely Split-If, a higher-payoff operation. 580 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) { 581 Node* use = phi->fast_out(k); 582 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr()) 583 cost += ConditionalMoveLimit; 584 // Is there a use inside the loop? 585 // Note: check only basic types since CMoveP is pinned. 586 if (!used_inside_loop && is_java_primitive(bt)) { 587 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use); 588 if (r_loop == u_loop || r_loop->is_member(u_loop)) { 589 used_inside_loop = true; 590 } 591 } 592 } 593 }//for 594 Node* bol = iff->in(1); 595 assert(bol->Opcode() == Op_Bool, ""); 596 int cmp_op = bol->in(1)->Opcode(); 597 // It is expensive to generate flags from a float compare. 598 // Avoid duplicated float compare. 599 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL; 600 601 float infrequent_prob = PROB_UNLIKELY_MAG(3); 602 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop. 603 if (used_inside_loop) { 604 if (cost >= ConditionalMoveLimit) return NULL; // Too much goo 605 606 // BlockLayoutByFrequency optimization moves infrequent branch 607 // from hot path. No point in CMOV'ing in such case (110 is used 608 // instead of 100 to take into account not exactness of float value). 609 if (BlockLayoutByFrequency) { 610 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f); 611 } 612 } 613 // Check for highly predictable branch. No point in CMOV'ing if 614 // we are going to predict accurately all the time. 615 if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) { 616 //keep going 617 } else if (iff->_prob < infrequent_prob || 618 iff->_prob > (1.0f - infrequent_prob)) 619 return NULL; 620 621 // -------------- 622 // Now replace all Phis with CMOV's 623 Node *cmov_ctrl = iff->in(0); 624 uint flip = (lp->Opcode() == Op_IfTrue); 625 Node_List wq; 626 while (1) { 627 PhiNode* phi = NULL; 628 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 629 Node *out = region->fast_out(i); 630 if (out->is_Phi()) { 631 phi = out->as_Phi(); 632 break; 633 } 634 } 635 if (phi == NULL) break; 636 if (PrintOpto && VerifyLoopOptimizations) { tty->print_cr("CMOV"); } 637 // Move speculative ops 638 wq.push(phi); 639 while (wq.size() > 0) { 640 Node *n = wq.pop(); 641 for (uint j = 1; j < n->req(); j++) { 642 Node* m = n->in(j); 643 if (m != NULL && !is_dominator(get_ctrl(m), cmov_ctrl)) { 644 #ifndef PRODUCT 645 if (PrintOpto && VerifyLoopOptimizations) { 646 tty->print(" speculate: "); 647 m->dump(); 648 } 649 #endif 650 set_ctrl(m, cmov_ctrl); 651 wq.push(m); 652 } 653 } 654 } 655 Node *cmov = CMoveNode::make(cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi)); 656 register_new_node( cmov, cmov_ctrl ); 657 _igvn.replace_node( phi, cmov ); 658 #ifndef PRODUCT 659 if (TraceLoopOpts) { 660 tty->print("CMOV "); 661 r_loop->dump_head(); 662 if (Verbose) { 663 bol->in(1)->dump(1); 664 cmov->dump(1); 665 } 666 } 667 if (VerifyLoopOptimizations) verify(); 668 #endif 669 } 670 671 // The useless CFG diamond will fold up later; see the optimization in 672 // RegionNode::Ideal. 673 _igvn._worklist.push(region); 674 675 return iff->in(1); 676 } 677 678 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) { 679 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 680 Node* u = m->fast_out(i); 681 if (u->is_CFG()) { 682 if (u->Opcode() == Op_NeverBranch) { 683 u = ((NeverBranchNode*)u)->proj_out(0); 684 enqueue_cfg_uses(u, wq); 685 } else { 686 wq.push(u); 687 } 688 } 689 } 690 } 691 692 // Try moving a store out of a loop, right before the loop 693 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) { 694 // Store has to be first in the loop body 695 IdealLoopTree *n_loop = get_loop(n_ctrl); 696 if (n->is_Store() && n_loop != _ltree_root && 697 n_loop->is_loop() && n_loop->_head->is_Loop() && 698 n->in(0) != NULL) { 699 Node* address = n->in(MemNode::Address); 700 Node* value = n->in(MemNode::ValueIn); 701 Node* mem = n->in(MemNode::Memory); 702 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 703 IdealLoopTree* value_loop = get_loop(get_ctrl(value)); 704 705 // - address and value must be loop invariant 706 // - memory must be a memory Phi for the loop 707 // - Store must be the only store on this memory slice in the 708 // loop: if there's another store following this one then value 709 // written at iteration i by the second store could be overwritten 710 // at iteration i+n by the first store: it's not safe to move the 711 // first store out of the loop 712 // - nothing must observe the memory Phi: it guarantees no read 713 // before the store, we are also guaranteed the store post 714 // dominates the loop head (ignoring a possible early 715 // exit). Otherwise there would be extra Phi involved between the 716 // loop's Phi and the store. 717 // - there must be no early exit from the loop before the Store 718 // (such an exit most of the time would be an extra use of the 719 // memory Phi but sometimes is a bottom memory Phi that takes the 720 // store as input). 721 722 if (!n_loop->is_member(address_loop) && 723 !n_loop->is_member(value_loop) && 724 mem->is_Phi() && mem->in(0) == n_loop->_head && 725 mem->outcnt() == 1 && 726 mem->in(LoopNode::LoopBackControl) == n) { 727 728 assert(n_loop->_tail != NULL, "need a tail"); 729 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop"); 730 731 // Verify that there's no early exit of the loop before the store. 732 bool ctrl_ok = false; 733 { 734 // Follow control from loop head until n, we exit the loop or 735 // we reach the tail 736 ResourceMark rm; 737 Unique_Node_List wq; 738 wq.push(n_loop->_head); 739 740 for (uint next = 0; next < wq.size(); ++next) { 741 Node *m = wq.at(next); 742 if (m == n->in(0)) { 743 ctrl_ok = true; 744 continue; 745 } 746 assert(!has_ctrl(m), "should be CFG"); 747 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) { 748 ctrl_ok = false; 749 break; 750 } 751 enqueue_cfg_uses(m, wq); 752 if (wq.size() > 10) { 753 ctrl_ok = false; 754 break; 755 } 756 } 757 } 758 if (ctrl_ok) { 759 // move the Store 760 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem); 761 _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl)); 762 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl)); 763 // Disconnect the phi now. An empty phi can confuse other 764 // optimizations in this pass of loop opts. 765 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl)); 766 n_loop->_body.yank(mem); 767 768 set_ctrl_and_loop(n, n->in(0)); 769 770 return n; 771 } 772 } 773 } 774 return NULL; 775 } 776 777 // Try moving a store out of a loop, right after the loop 778 void PhaseIdealLoop::try_move_store_after_loop(Node* n) { 779 if (n->is_Store() && n->in(0) != NULL) { 780 Node *n_ctrl = get_ctrl(n); 781 IdealLoopTree *n_loop = get_loop(n_ctrl); 782 // Store must be in a loop 783 if (n_loop != _ltree_root && !n_loop->_irreducible) { 784 Node* address = n->in(MemNode::Address); 785 Node* value = n->in(MemNode::ValueIn); 786 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 787 // address must be loop invariant 788 if (!n_loop->is_member(address_loop)) { 789 // Store must be last on this memory slice in the loop and 790 // nothing in the loop must observe it 791 Node* phi = NULL; 792 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 793 Node* u = n->fast_out(i); 794 if (has_ctrl(u)) { // control use? 795 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 796 if (!n_loop->is_member(u_loop)) { 797 continue; 798 } 799 if (u->is_Phi() && u->in(0) == n_loop->_head) { 800 assert(_igvn.type(u) == Type::MEMORY, "bad phi"); 801 // multiple phis on the same slice are possible 802 if (phi != NULL) { 803 return; 804 } 805 phi = u; 806 continue; 807 } 808 } 809 return; 810 } 811 if (phi != NULL) { 812 // Nothing in the loop before the store (next iteration) 813 // must observe the stored value 814 bool mem_ok = true; 815 { 816 ResourceMark rm; 817 Unique_Node_List wq; 818 wq.push(phi); 819 for (uint next = 0; next < wq.size() && mem_ok; ++next) { 820 Node *m = wq.at(next); 821 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) { 822 Node* u = m->fast_out(i); 823 if (u->is_Store() || u->is_Phi()) { 824 if (u != n) { 825 wq.push(u); 826 mem_ok = (wq.size() <= 10); 827 } 828 } else { 829 mem_ok = false; 830 break; 831 } 832 } 833 } 834 } 835 if (mem_ok) { 836 // Move the store out of the loop if the LCA of all 837 // users (except for the phi) is outside the loop. 838 Node* hook = new Node(1); 839 _igvn.rehash_node_delayed(phi); 840 int count = phi->replace_edge(n, hook); 841 assert(count > 0, "inconsistent phi"); 842 843 // Compute latest point this store can go 844 Node* lca = get_late_ctrl(n, get_ctrl(n)); 845 if (n_loop->is_member(get_loop(lca))) { 846 // LCA is in the loop - bail out 847 _igvn.replace_node(hook, n); 848 return; 849 } 850 #ifdef ASSERT 851 if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) { 852 assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined"); 853 n_loop->_head->as_Loop()->verify_strip_mined(1); 854 Node* outer = n_loop->_head->as_CountedLoop()->outer_loop(); 855 IdealLoopTree* outer_loop = get_loop(outer); 856 assert(n_loop->_parent == outer_loop, "broken loop tree"); 857 assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state"); 858 } 859 #endif 860 861 // Move store out of the loop 862 _igvn.replace_node(hook, n->in(MemNode::Memory)); 863 _igvn.replace_input_of(n, 0, lca); 864 set_ctrl_and_loop(n, lca); 865 866 // Disconnect the phi now. An empty phi can confuse other 867 // optimizations in this pass of loop opts.. 868 if (phi->in(LoopNode::LoopBackControl) == phi) { 869 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl)); 870 n_loop->_body.yank(phi); 871 } 872 } 873 } 874 } 875 } 876 } 877 } 878 879 //------------------------------split_if_with_blocks_pre----------------------- 880 // Do the real work in a non-recursive function. Data nodes want to be 881 // cloned in the pre-order so they can feed each other nicely. 882 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) { 883 // Cloning these guys is unlikely to win 884 int n_op = n->Opcode(); 885 if( n_op == Op_MergeMem ) return n; 886 if( n->is_Proj() ) return n; 887 // Do not clone-up CmpFXXX variations, as these are always 888 // followed by a CmpI 889 if( n->is_Cmp() ) return n; 890 // Attempt to use a conditional move instead of a phi/branch 891 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) { 892 Node *cmov = conditional_move( n ); 893 if( cmov ) return cmov; 894 } 895 if( n->is_CFG() || n->is_LoadStore() ) 896 return n; 897 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd 898 n_op == Op_Opaque2 ) { 899 if( !C->major_progress() ) // If chance of no more loop opts... 900 _igvn._worklist.push(n); // maybe we'll remove them 901 return n; 902 } 903 904 if( n->is_Con() ) return n; // No cloning for Con nodes 905 906 Node *n_ctrl = get_ctrl(n); 907 if( !n_ctrl ) return n; // Dead node 908 909 Node* res = try_move_store_before_loop(n, n_ctrl); 910 if (res != NULL) { 911 return n; 912 } 913 914 // Attempt to remix address expressions for loop invariants 915 Node *m = remix_address_expressions( n ); 916 if( m ) return m; 917 918 if (n->is_ConstraintCast()) { 919 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this); 920 // ConstraintCastNode::dominating_cast() uses node control input to determine domination. 921 // Node control inputs don't necessarily agree with loop control info (due to 922 // transformations happened in between), thus additional dominance check is needed 923 // to keep loop info valid. 924 if (dom_cast != NULL && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) { 925 _igvn.replace_node(n, dom_cast); 926 return dom_cast; 927 } 928 } 929 930 // Determine if the Node has inputs from some local Phi. 931 // Returns the block to clone thru. 932 Node *n_blk = has_local_phi_input( n ); 933 if( !n_blk ) return n; 934 935 // Do not clone the trip counter through on a CountedLoop 936 // (messes up the canonical shape). 937 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n; 938 939 // Check for having no control input; not pinned. Allow 940 // dominating control. 941 if (n->in(0)) { 942 Node *dom = idom(n_blk); 943 if (dom_lca(n->in(0), dom) != n->in(0)) { 944 return n; 945 } 946 } 947 // Policy: when is it profitable. You must get more wins than 948 // policy before it is considered profitable. Policy is usually 0, 949 // so 1 win is considered profitable. Big merges will require big 950 // cloning, so get a larger policy. 951 int policy = n_blk->req() >> 2; 952 953 // If the loop is a candidate for range check elimination, 954 // delay splitting through it's phi until a later loop optimization 955 if (n_blk->is_CountedLoop()) { 956 IdealLoopTree *lp = get_loop(n_blk); 957 if (lp && lp->_rce_candidate) { 958 return n; 959 } 960 } 961 962 // Use same limit as split_if_with_blocks_post 963 if( C->live_nodes() > 35000 ) return n; // Method too big 964 965 // Split 'n' through the merge point if it is profitable 966 Node *phi = split_thru_phi( n, n_blk, policy ); 967 if (!phi) return n; 968 969 // Found a Phi to split thru! 970 // Replace 'n' with the new phi 971 _igvn.replace_node( n, phi ); 972 // Moved a load around the loop, 'en-registering' something. 973 if (n_blk->is_Loop() && n->is_Load() && 974 !phi->in(LoopNode::LoopBackControl)->is_Load()) 975 C->set_major_progress(); 976 977 return phi; 978 } 979 980 static bool merge_point_too_heavy(Compile* C, Node* region) { 981 // Bail out if the region and its phis have too many users. 982 int weight = 0; 983 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 984 weight += region->fast_out(i)->outcnt(); 985 } 986 int nodes_left = C->max_node_limit() - C->live_nodes(); 987 if (weight * 8 > nodes_left) { 988 if (PrintOpto) { 989 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight); 990 } 991 return true; 992 } else { 993 return false; 994 } 995 } 996 997 static bool merge_point_safe(Node* region) { 998 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode 999 // having a PhiNode input. This sidesteps the dangerous case where the split 1000 // ConvI2LNode may become TOP if the input Value() does not 1001 // overlap the ConvI2L range, leaving a node which may not dominate its 1002 // uses. 1003 // A better fix for this problem can be found in the BugTraq entry, but 1004 // expediency for Mantis demands this hack. 1005 // 6855164: If the merge point has a FastLockNode with a PhiNode input, we stop 1006 // split_if_with_blocks from splitting a block because we could not move around 1007 // the FastLockNode. 1008 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1009 Node* n = region->fast_out(i); 1010 if (n->is_Phi()) { 1011 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1012 Node* m = n->fast_out(j); 1013 if (m->is_FastLock()) 1014 return false; 1015 #ifdef _LP64 1016 if (m->Opcode() == Op_ConvI2L) 1017 return false; 1018 if (m->is_CastII() && m->isa_CastII()->has_range_check()) { 1019 return false; 1020 } 1021 #endif 1022 } 1023 } 1024 } 1025 return true; 1026 } 1027 1028 1029 //------------------------------place_near_use--------------------------------- 1030 // Place some computation next to use but not inside inner loops. 1031 // For inner loop uses move it to the preheader area. 1032 Node *PhaseIdealLoop::place_near_use( Node *useblock ) const { 1033 IdealLoopTree *u_loop = get_loop( useblock ); 1034 return (u_loop->_irreducible || u_loop->_child) 1035 ? useblock 1036 : u_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl); 1037 } 1038 1039 1040 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) { 1041 if (!n->is_If() || n->is_CountedLoopEnd()) { 1042 return false; 1043 } 1044 if (!n->in(0)->is_Region()) { 1045 return false; 1046 } 1047 Node* region = n->in(0); 1048 Node* dom = idom(region); 1049 if (!dom->is_If() || dom->in(1) != n->in(1)) { 1050 return false; 1051 } 1052 IfNode* dom_if = dom->as_If(); 1053 Node* proj_true = dom_if->proj_out(1); 1054 Node* proj_false = dom_if->proj_out(0); 1055 1056 for (uint i = 1; i < region->req(); i++) { 1057 if (is_dominator(proj_true, region->in(i))) { 1058 continue; 1059 } 1060 if (is_dominator(proj_false, region->in(i))) { 1061 continue; 1062 } 1063 return false; 1064 } 1065 1066 return true; 1067 } 1068 1069 bool PhaseIdealLoop::can_split_if(Node *n_ctrl) { 1070 if (C->live_nodes() > 35000) { 1071 return false; // Method too big 1072 } 1073 1074 // Do not do 'split-if' if irreducible loops are present. 1075 if (_has_irreducible_loops) { 1076 return false; 1077 } 1078 1079 if (merge_point_too_heavy(C, n_ctrl)) { 1080 return false; 1081 } 1082 1083 // Do not do 'split-if' if some paths are dead. First do dead code 1084 // elimination and then see if its still profitable. 1085 for (uint i = 1; i < n_ctrl->req(); i++) { 1086 if (n_ctrl->in(i) == C->top()) { 1087 return false; 1088 } 1089 } 1090 1091 // If trying to do a 'Split-If' at the loop head, it is only 1092 // profitable if the cmp folds up on BOTH paths. Otherwise we 1093 // risk peeling a loop forever. 1094 1095 // CNC - Disabled for now. Requires careful handling of loop 1096 // body selection for the cloned code. Also, make sure we check 1097 // for any input path not being in the same loop as n_ctrl. For 1098 // irreducible loops we cannot check for 'n_ctrl->is_Loop()' 1099 // because the alternative loop entry points won't be converted 1100 // into LoopNodes. 1101 IdealLoopTree *n_loop = get_loop(n_ctrl); 1102 for (uint j = 1; j < n_ctrl->req(); j++) { 1103 if (get_loop(n_ctrl->in(j)) != n_loop) { 1104 return false; 1105 } 1106 } 1107 1108 // Check for safety of the merge point. 1109 if (!merge_point_safe(n_ctrl)) { 1110 return false; 1111 } 1112 1113 return true; 1114 } 1115 1116 //------------------------------split_if_with_blocks_post---------------------- 1117 // Do the real work in a non-recursive function. CFG hackery wants to be 1118 // in the post-order, so it can dirty the I-DOM info and not use the dirtied 1119 // info. 1120 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) { 1121 1122 // Cloning Cmp through Phi's involves the split-if transform. 1123 // FastLock is not used by an If 1124 if (n->is_Cmp() && !n->is_FastLock()) { 1125 Node *n_ctrl = get_ctrl(n); 1126 // Determine if the Node has inputs from some local Phi. 1127 // Returns the block to clone thru. 1128 Node *n_blk = has_local_phi_input(n); 1129 if (n_blk != n_ctrl) { 1130 return; 1131 } 1132 1133 if (!can_split_if(n_ctrl)) { 1134 return; 1135 } 1136 1137 if (n->outcnt() != 1) { 1138 return; // Multiple bool's from 1 compare? 1139 } 1140 Node *bol = n->unique_out(); 1141 assert(bol->is_Bool(), "expect a bool here"); 1142 if (bol->outcnt() != 1) { 1143 return;// Multiple branches from 1 compare? 1144 } 1145 Node *iff = bol->unique_out(); 1146 1147 // Check some safety conditions 1148 if (iff->is_If()) { // Classic split-if? 1149 if (iff->in(0) != n_ctrl) { 1150 return; // Compare must be in same blk as if 1151 } 1152 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE 1153 // Can't split CMove with different control edge. 1154 if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) { 1155 return; 1156 } 1157 if (get_ctrl(iff->in(2)) == n_ctrl || 1158 get_ctrl(iff->in(3)) == n_ctrl) { 1159 return; // Inputs not yet split-up 1160 } 1161 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) { 1162 return; // Loop-invar test gates loop-varying CMOVE 1163 } 1164 } else { 1165 return; // some other kind of node, such as an Allocate 1166 } 1167 1168 // When is split-if profitable? Every 'win' on means some control flow 1169 // goes dead, so it's almost always a win. 1170 int policy = 0; 1171 // Split compare 'n' through the merge point if it is profitable 1172 Node *phi = split_thru_phi( n, n_ctrl, policy); 1173 if (!phi) { 1174 return; 1175 } 1176 1177 // Found a Phi to split thru! 1178 // Replace 'n' with the new phi 1179 _igvn.replace_node(n, phi); 1180 1181 // Now split the bool up thru the phi 1182 Node *bolphi = split_thru_phi(bol, n_ctrl, -1); 1183 guarantee(bolphi != NULL, "null boolean phi node"); 1184 1185 _igvn.replace_node(bol, bolphi); 1186 assert(iff->in(1) == bolphi, ""); 1187 1188 if (bolphi->Value(&_igvn)->singleton()) { 1189 return; 1190 } 1191 1192 // Conditional-move? Must split up now 1193 if (!iff->is_If()) { 1194 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1); 1195 _igvn.replace_node(iff, cmovphi); 1196 return; 1197 } 1198 1199 // Now split the IF 1200 do_split_if(iff); 1201 return; 1202 } 1203 1204 // Two identical ifs back to back can be merged 1205 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) { 1206 Node *n_ctrl = n->in(0); 1207 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1)); 1208 IfNode* dom_if = idom(n_ctrl)->as_If(); 1209 Node* proj_true = dom_if->proj_out(1); 1210 Node* proj_false = dom_if->proj_out(0); 1211 Node* con_true = _igvn.makecon(TypeInt::ONE); 1212 Node* con_false = _igvn.makecon(TypeInt::ZERO); 1213 1214 for (uint i = 1; i < n_ctrl->req(); i++) { 1215 if (is_dominator(proj_true, n_ctrl->in(i))) { 1216 bolphi->init_req(i, con_true); 1217 } else { 1218 assert(is_dominator(proj_false, n_ctrl->in(i)), "bad if"); 1219 bolphi->init_req(i, con_false); 1220 } 1221 } 1222 register_new_node(bolphi, n_ctrl); 1223 _igvn.replace_input_of(n, 1, bolphi); 1224 1225 // Now split the IF 1226 do_split_if(n); 1227 return; 1228 } 1229 1230 // Check for an IF ready to split; one that has its 1231 // condition codes input coming from a Phi at the block start. 1232 int n_op = n->Opcode(); 1233 1234 // Check for an IF being dominated by another IF same test 1235 if (n_op == Op_If || 1236 n_op == Op_RangeCheck) { 1237 Node *bol = n->in(1); 1238 uint max = bol->outcnt(); 1239 // Check for same test used more than once? 1240 if (max > 1 && bol->is_Bool()) { 1241 // Search up IDOMs to see if this IF is dominated. 1242 Node *cutoff = get_ctrl(bol); 1243 1244 // Now search up IDOMs till cutoff, looking for a dominating test 1245 Node *prevdom = n; 1246 Node *dom = idom(prevdom); 1247 while (dom != cutoff) { 1248 if (dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom) { 1249 // Replace the dominated test with an obvious true or false. 1250 // Place it on the IGVN worklist for later cleanup. 1251 C->set_major_progress(); 1252 dominated_by(prevdom, n, false, true); 1253 #ifndef PRODUCT 1254 if( VerifyLoopOptimizations ) verify(); 1255 #endif 1256 return; 1257 } 1258 prevdom = dom; 1259 dom = idom(prevdom); 1260 } 1261 } 1262 } 1263 1264 // See if a shared loop-varying computation has no loop-varying uses. 1265 // Happens if something is only used for JVM state in uncommon trap exits, 1266 // like various versions of induction variable+offset. Clone the 1267 // computation per usage to allow it to sink out of the loop. 1268 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about) 1269 Node *n_ctrl = get_ctrl(n); 1270 IdealLoopTree *n_loop = get_loop(n_ctrl); 1271 if( n_loop != _ltree_root ) { 1272 DUIterator_Fast imax, i = n->fast_outs(imax); 1273 for (; i < imax; i++) { 1274 Node* u = n->fast_out(i); 1275 if( !has_ctrl(u) ) break; // Found control user 1276 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 1277 if( u_loop == n_loop ) break; // Found loop-varying use 1278 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop 1279 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003 1280 } 1281 bool did_break = (i < imax); // Did we break out of the previous loop? 1282 if (!did_break && n->outcnt() > 1) { // All uses in outer loops! 1283 Node *late_load_ctrl = NULL; 1284 if (n->is_Load()) { 1285 // If n is a load, get and save the result from get_late_ctrl(), 1286 // to be later used in calculating the control for n's clones. 1287 clear_dom_lca_tags(); 1288 late_load_ctrl = get_late_ctrl(n, n_ctrl); 1289 } 1290 // If n is a load, and the late control is the same as the current 1291 // control, then the cloning of n is a pointless exercise, because 1292 // GVN will ensure that we end up where we started. 1293 if (!n->is_Load() || late_load_ctrl != n_ctrl) { 1294 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) { 1295 Node *u = n->last_out(j); // Clone private computation per use 1296 _igvn.rehash_node_delayed(u); 1297 Node *x = n->clone(); // Clone computation 1298 Node *x_ctrl = NULL; 1299 if( u->is_Phi() ) { 1300 // Replace all uses of normal nodes. Replace Phi uses 1301 // individually, so the separate Nodes can sink down 1302 // different paths. 1303 uint k = 1; 1304 while( u->in(k) != n ) k++; 1305 u->set_req( k, x ); 1306 // x goes next to Phi input path 1307 x_ctrl = u->in(0)->in(k); 1308 --j; 1309 } else { // Normal use 1310 // Replace all uses 1311 for( uint k = 0; k < u->req(); k++ ) { 1312 if( u->in(k) == n ) { 1313 u->set_req( k, x ); 1314 --j; 1315 } 1316 } 1317 x_ctrl = get_ctrl(u); 1318 } 1319 1320 // Find control for 'x' next to use but not inside inner loops. 1321 // For inner loop uses get the preheader area. 1322 x_ctrl = place_near_use(x_ctrl); 1323 1324 if (n->is_Load()) { 1325 // For loads, add a control edge to a CFG node outside of the loop 1326 // to force them to not combine and return back inside the loop 1327 // during GVN optimization (4641526). 1328 // 1329 // Because we are setting the actual control input, factor in 1330 // the result from get_late_ctrl() so we respect any 1331 // anti-dependences. (6233005). 1332 x_ctrl = dom_lca(late_load_ctrl, x_ctrl); 1333 1334 // Don't allow the control input to be a CFG splitting node. 1335 // Such nodes should only have ProjNodes as outs, e.g. IfNode 1336 // should only have IfTrueNode and IfFalseNode (4985384). 1337 x_ctrl = find_non_split_ctrl(x_ctrl); 1338 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone"); 1339 1340 x->set_req(0, x_ctrl); 1341 } 1342 register_new_node(x, x_ctrl); 1343 1344 // Some institutional knowledge is needed here: 'x' is 1345 // yanked because if the optimizer runs GVN on it all the 1346 // cloned x's will common up and undo this optimization and 1347 // be forced back in the loop. This is annoying because it 1348 // makes +VerifyOpto report false-positives on progress. I 1349 // tried setting control edges on the x's to force them to 1350 // not combine, but the matching gets worried when it tries 1351 // to fold a StoreP and an AddP together (as part of an 1352 // address expression) and the AddP and StoreP have 1353 // different controls. 1354 if (!x->is_Load() && !x->is_DecodeNarrowPtr()) _igvn._worklist.yank(x); 1355 } 1356 _igvn.remove_dead_node(n); 1357 } 1358 } 1359 } 1360 } 1361 1362 try_move_store_after_loop(n); 1363 1364 // Check for Opaque2's who's loop has disappeared - who's input is in the 1365 // same loop nest as their output. Remove 'em, they are no longer useful. 1366 if( n_op == Op_Opaque2 && 1367 n->in(1) != NULL && 1368 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) { 1369 _igvn.replace_node( n, n->in(1) ); 1370 } 1371 } 1372 1373 //------------------------------split_if_with_blocks--------------------------- 1374 // Check for aggressive application of 'split-if' optimization, 1375 // using basic block level info. 1376 void PhaseIdealLoop::split_if_with_blocks( VectorSet &visited, Node_Stack &nstack ) { 1377 Node *n = C->root(); 1378 visited.set(n->_idx); // first, mark node as visited 1379 // Do pre-visit work for root 1380 n = split_if_with_blocks_pre( n ); 1381 uint cnt = n->outcnt(); 1382 uint i = 0; 1383 while (true) { 1384 // Visit all children 1385 if (i < cnt) { 1386 Node* use = n->raw_out(i); 1387 ++i; 1388 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) { 1389 // Now do pre-visit work for this use 1390 use = split_if_with_blocks_pre( use ); 1391 nstack.push(n, i); // Save parent and next use's index. 1392 n = use; // Process all children of current use. 1393 cnt = use->outcnt(); 1394 i = 0; 1395 } 1396 } 1397 else { 1398 // All of n's children have been processed, complete post-processing. 1399 if (cnt != 0 && !n->is_Con()) { 1400 assert(has_node(n), "no dead nodes"); 1401 split_if_with_blocks_post( n ); 1402 } 1403 if (nstack.is_empty()) { 1404 // Finished all nodes on stack. 1405 break; 1406 } 1407 // Get saved parent node and next use's index. Visit the rest of uses. 1408 n = nstack.node(); 1409 cnt = n->outcnt(); 1410 i = nstack.index(); 1411 nstack.pop(); 1412 } 1413 } 1414 } 1415 1416 1417 //============================================================================= 1418 // 1419 // C L O N E A L O O P B O D Y 1420 // 1421 1422 //------------------------------clone_iff-------------------------------------- 1423 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1424 // "Nearly" because all Nodes have been cloned from the original in the loop, 1425 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1426 // through the Phi recursively, and return a Bool. 1427 Node* PhaseIdealLoop::clone_iff(PhiNode *phi, IdealLoopTree *loop) { 1428 1429 // Convert this Phi into a Phi merging Bools 1430 uint i; 1431 for (i = 1; i < phi->req(); i++) { 1432 Node *b = phi->in(i); 1433 if (b->is_Phi()) { 1434 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi(), loop)); 1435 } else { 1436 assert(b->is_Bool() || b->Opcode() == Op_Opaque4, ""); 1437 } 1438 } 1439 1440 Node* n = phi->in(1); 1441 Node* sample_opaque = NULL; 1442 Node *sample_bool = NULL; 1443 if (n->Opcode() == Op_Opaque4) { 1444 sample_opaque = n; 1445 sample_bool = n->in(1); 1446 assert(sample_bool->is_Bool(), "wrong type"); 1447 } else { 1448 sample_bool = n; 1449 } 1450 Node *sample_cmp = sample_bool->in(1); 1451 1452 // Make Phis to merge the Cmp's inputs. 1453 PhiNode *phi1 = new PhiNode(phi->in(0), Type::TOP); 1454 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP); 1455 for (i = 1; i < phi->req(); i++) { 1456 Node *n1 = sample_opaque == NULL ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1); 1457 Node *n2 = sample_opaque == NULL ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2); 1458 phi1->set_req(i, n1); 1459 phi2->set_req(i, n2); 1460 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 1461 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 1462 } 1463 // See if these Phis have been made before. 1464 // Register with optimizer 1465 Node *hit1 = _igvn.hash_find_insert(phi1); 1466 if (hit1) { // Hit, toss just made Phi 1467 _igvn.remove_dead_node(phi1); // Remove new phi 1468 assert(hit1->is_Phi(), "" ); 1469 phi1 = (PhiNode*)hit1; // Use existing phi 1470 } else { // Miss 1471 _igvn.register_new_node_with_optimizer(phi1); 1472 } 1473 Node *hit2 = _igvn.hash_find_insert(phi2); 1474 if (hit2) { // Hit, toss just made Phi 1475 _igvn.remove_dead_node(phi2); // Remove new phi 1476 assert(hit2->is_Phi(), "" ); 1477 phi2 = (PhiNode*)hit2; // Use existing phi 1478 } else { // Miss 1479 _igvn.register_new_node_with_optimizer(phi2); 1480 } 1481 // Register Phis with loop/block info 1482 set_ctrl(phi1, phi->in(0)); 1483 set_ctrl(phi2, phi->in(0)); 1484 // Make a new Cmp 1485 Node *cmp = sample_cmp->clone(); 1486 cmp->set_req(1, phi1); 1487 cmp->set_req(2, phi2); 1488 _igvn.register_new_node_with_optimizer(cmp); 1489 set_ctrl(cmp, phi->in(0)); 1490 1491 // Make a new Bool 1492 Node *b = sample_bool->clone(); 1493 b->set_req(1,cmp); 1494 _igvn.register_new_node_with_optimizer(b); 1495 set_ctrl(b, phi->in(0)); 1496 1497 if (sample_opaque != NULL) { 1498 Node* opaque = sample_opaque->clone(); 1499 opaque->set_req(1, b); 1500 _igvn.register_new_node_with_optimizer(opaque); 1501 set_ctrl(opaque, phi->in(0)); 1502 return opaque; 1503 } 1504 1505 assert(b->is_Bool(), ""); 1506 return b; 1507 } 1508 1509 //------------------------------clone_bool------------------------------------- 1510 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1511 // "Nearly" because all Nodes have been cloned from the original in the loop, 1512 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1513 // through the Phi recursively, and return a Bool. 1514 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) { 1515 uint i; 1516 // Convert this Phi into a Phi merging Bools 1517 for( i = 1; i < phi->req(); i++ ) { 1518 Node *b = phi->in(i); 1519 if( b->is_Phi() ) { 1520 _igvn.replace_input_of(phi, i, clone_bool( b->as_Phi(), loop )); 1521 } else { 1522 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" ); 1523 } 1524 } 1525 1526 Node *sample_cmp = phi->in(1); 1527 1528 // Make Phis to merge the Cmp's inputs. 1529 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP ); 1530 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP ); 1531 for( uint j = 1; j < phi->req(); j++ ) { 1532 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP 1533 Node *n1, *n2; 1534 if( cmp_top->is_Cmp() ) { 1535 n1 = cmp_top->in(1); 1536 n2 = cmp_top->in(2); 1537 } else { 1538 n1 = n2 = cmp_top; 1539 } 1540 phi1->set_req( j, n1 ); 1541 phi2->set_req( j, n2 ); 1542 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 1543 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 1544 } 1545 1546 // See if these Phis have been made before. 1547 // Register with optimizer 1548 Node *hit1 = _igvn.hash_find_insert(phi1); 1549 if( hit1 ) { // Hit, toss just made Phi 1550 _igvn.remove_dead_node(phi1); // Remove new phi 1551 assert( hit1->is_Phi(), "" ); 1552 phi1 = (PhiNode*)hit1; // Use existing phi 1553 } else { // Miss 1554 _igvn.register_new_node_with_optimizer(phi1); 1555 } 1556 Node *hit2 = _igvn.hash_find_insert(phi2); 1557 if( hit2 ) { // Hit, toss just made Phi 1558 _igvn.remove_dead_node(phi2); // Remove new phi 1559 assert( hit2->is_Phi(), "" ); 1560 phi2 = (PhiNode*)hit2; // Use existing phi 1561 } else { // Miss 1562 _igvn.register_new_node_with_optimizer(phi2); 1563 } 1564 // Register Phis with loop/block info 1565 set_ctrl(phi1, phi->in(0)); 1566 set_ctrl(phi2, phi->in(0)); 1567 // Make a new Cmp 1568 Node *cmp = sample_cmp->clone(); 1569 cmp->set_req( 1, phi1 ); 1570 cmp->set_req( 2, phi2 ); 1571 _igvn.register_new_node_with_optimizer(cmp); 1572 set_ctrl(cmp, phi->in(0)); 1573 1574 assert( cmp->is_Cmp(), "" ); 1575 return (CmpNode*)cmp; 1576 } 1577 1578 //------------------------------sink_use--------------------------------------- 1579 // If 'use' was in the loop-exit block, it now needs to be sunk 1580 // below the post-loop merge point. 1581 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) { 1582 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) { 1583 set_ctrl(use, post_loop); 1584 for (DUIterator j = use->outs(); use->has_out(j); j++) 1585 sink_use(use->out(j), post_loop); 1586 } 1587 } 1588 1589 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new, 1590 IdealLoopTree* loop, IdealLoopTree* outer_loop, 1591 Node_List*& split_if_set, Node_List*& split_bool_set, 1592 Node_List*& split_cex_set, Node_List& worklist, 1593 uint new_counter, CloneLoopMode mode) { 1594 Node* nnn = old_new[old->_idx]; 1595 // Copy uses to a worklist, so I can munge the def-use info 1596 // with impunity. 1597 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1598 worklist.push(old->fast_out(j)); 1599 1600 while( worklist.size() ) { 1601 Node *use = worklist.pop(); 1602 if (!has_node(use)) continue; // Ignore dead nodes 1603 if (use->in(0) == C->top()) continue; 1604 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1605 // Check for data-use outside of loop - at least one of OLD or USE 1606 // must not be a CFG node. 1607 #ifdef ASSERT 1608 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == NULL) { 1609 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint(); 1610 assert(mode == ControlAroundStripMined && use == sfpt, "missed a node"); 1611 } 1612 #endif 1613 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) { 1614 1615 // If the Data use is an IF, that means we have an IF outside of the 1616 // loop that is switching on a condition that is set inside of the 1617 // loop. Happens if people set a loop-exit flag; then test the flag 1618 // in the loop to break the loop, then test is again outside of the 1619 // loop to determine which way the loop exited. 1620 // Loop predicate If node connects to Bool node through Opaque1 node. 1621 if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use) || use->Opcode() == Op_Opaque4) { 1622 // Since this code is highly unlikely, we lazily build the worklist 1623 // of such Nodes to go split. 1624 if (!split_if_set) { 1625 ResourceArea *area = Thread::current()->resource_area(); 1626 split_if_set = new Node_List(area); 1627 } 1628 split_if_set->push(use); 1629 } 1630 if (use->is_Bool()) { 1631 if (!split_bool_set) { 1632 ResourceArea *area = Thread::current()->resource_area(); 1633 split_bool_set = new Node_List(area); 1634 } 1635 split_bool_set->push(use); 1636 } 1637 if (use->Opcode() == Op_CreateEx) { 1638 if (!split_cex_set) { 1639 ResourceArea *area = Thread::current()->resource_area(); 1640 split_cex_set = new Node_List(area); 1641 } 1642 split_cex_set->push(use); 1643 } 1644 1645 1646 // Get "block" use is in 1647 uint idx = 0; 1648 while( use->in(idx) != old ) idx++; 1649 Node *prev = use->is_CFG() ? use : get_ctrl(use); 1650 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" ); 1651 Node *cfg = prev->_idx >= new_counter 1652 ? prev->in(2) 1653 : idom(prev); 1654 if( use->is_Phi() ) // Phi use is in prior block 1655 cfg = prev->in(idx); // NOT in block of Phi itself 1656 if (cfg->is_top()) { // Use is dead? 1657 _igvn.replace_input_of(use, idx, C->top()); 1658 continue; 1659 } 1660 1661 while(!outer_loop->is_member(get_loop(cfg))) { 1662 prev = cfg; 1663 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg); 1664 } 1665 // If the use occurs after merging several exits from the loop, then 1666 // old value must have dominated all those exits. Since the same old 1667 // value was used on all those exits we did not need a Phi at this 1668 // merge point. NOW we do need a Phi here. Each loop exit value 1669 // is now merged with the peeled body exit; each exit gets its own 1670 // private Phi and those Phis need to be merged here. 1671 Node *phi; 1672 if( prev->is_Region() ) { 1673 if( idx == 0 ) { // Updating control edge? 1674 phi = prev; // Just use existing control 1675 } else { // Else need a new Phi 1676 phi = PhiNode::make( prev, old ); 1677 // Now recursively fix up the new uses of old! 1678 for( uint i = 1; i < prev->req(); i++ ) { 1679 worklist.push(phi); // Onto worklist once for each 'old' input 1680 } 1681 } 1682 } else { 1683 // Get new RegionNode merging old and new loop exits 1684 prev = old_new[prev->_idx]; 1685 assert( prev, "just made this in step 7" ); 1686 if( idx == 0) { // Updating control edge? 1687 phi = prev; // Just use existing control 1688 } else { // Else need a new Phi 1689 // Make a new Phi merging data values properly 1690 phi = PhiNode::make( prev, old ); 1691 phi->set_req( 1, nnn ); 1692 } 1693 } 1694 // If inserting a new Phi, check for prior hits 1695 if( idx != 0 ) { 1696 Node *hit = _igvn.hash_find_insert(phi); 1697 if( hit == NULL ) { 1698 _igvn.register_new_node_with_optimizer(phi); // Register new phi 1699 } else { // or 1700 // Remove the new phi from the graph and use the hit 1701 _igvn.remove_dead_node(phi); 1702 phi = hit; // Use existing phi 1703 } 1704 set_ctrl(phi, prev); 1705 } 1706 // Make 'use' use the Phi instead of the old loop body exit value 1707 _igvn.replace_input_of(use, idx, phi); 1708 if( use->_idx >= new_counter ) { // If updating new phis 1709 // Not needed for correctness, but prevents a weak assert 1710 // in AddPNode from tripping (when we end up with different 1711 // base & derived Phis that will become the same after 1712 // IGVN does CSE). 1713 Node *hit = _igvn.hash_find_insert(use); 1714 if( hit ) // Go ahead and re-hash for hits. 1715 _igvn.replace_node( use, hit ); 1716 } 1717 1718 // If 'use' was in the loop-exit block, it now needs to be sunk 1719 // below the post-loop merge point. 1720 sink_use( use, prev ); 1721 } 1722 } 1723 } 1724 1725 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop, 1726 IdealLoopTree* outer_loop, int dd, Node_List &old_new, 1727 Node_List& extra_data_nodes) { 1728 if (head->is_strip_mined() && mode != IgnoreStripMined) { 1729 CountedLoopNode* cl = head->as_CountedLoop(); 1730 Node* l = cl->outer_loop(); 1731 Node* tail = cl->outer_loop_tail(); 1732 IfNode* le = cl->outer_loop_end(); 1733 Node* sfpt = cl->outer_safepoint(); 1734 CountedLoopEndNode* cle = cl->loopexit(); 1735 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop(); 1736 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null(); 1737 Node* cle_out = cle->proj_out(false); 1738 1739 Node* new_sfpt = NULL; 1740 Node* new_cle_out = cle_out->clone(); 1741 old_new.map(cle_out->_idx, new_cle_out); 1742 if (mode == CloneIncludesStripMined) { 1743 // clone outer loop body 1744 Node* new_l = l->clone(); 1745 Node* new_tail = tail->clone(); 1746 IfNode* new_le = le->clone()->as_If(); 1747 new_sfpt = sfpt->clone(); 1748 1749 set_loop(new_l, outer_loop->_parent); 1750 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd); 1751 set_loop(new_cle_out, outer_loop->_parent); 1752 set_idom(new_cle_out, new_cle, dd); 1753 set_loop(new_sfpt, outer_loop->_parent); 1754 set_idom(new_sfpt, new_cle_out, dd); 1755 set_loop(new_le, outer_loop->_parent); 1756 set_idom(new_le, new_sfpt, dd); 1757 set_loop(new_tail, outer_loop->_parent); 1758 set_idom(new_tail, new_le, dd); 1759 set_idom(new_cl, new_l, dd); 1760 1761 old_new.map(l->_idx, new_l); 1762 old_new.map(tail->_idx, new_tail); 1763 old_new.map(le->_idx, new_le); 1764 old_new.map(sfpt->_idx, new_sfpt); 1765 1766 new_l->set_req(LoopNode::LoopBackControl, new_tail); 1767 new_l->set_req(0, new_l); 1768 new_tail->set_req(0, new_le); 1769 new_le->set_req(0, new_sfpt); 1770 new_sfpt->set_req(0, new_cle_out); 1771 new_cle_out->set_req(0, new_cle); 1772 new_cl->set_req(LoopNode::EntryControl, new_l); 1773 1774 _igvn.register_new_node_with_optimizer(new_l); 1775 _igvn.register_new_node_with_optimizer(new_tail); 1776 _igvn.register_new_node_with_optimizer(new_le); 1777 } else { 1778 Node *newhead = old_new[loop->_head->_idx]; 1779 newhead->as_Loop()->clear_strip_mined(); 1780 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl)); 1781 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 1782 } 1783 // Look at data node that were assigned a control in the outer 1784 // loop: they are kept in the outer loop by the safepoint so start 1785 // from the safepoint node's inputs. 1786 IdealLoopTree* outer_loop = get_loop(l); 1787 Node_Stack stack(2); 1788 stack.push(sfpt, 1); 1789 uint new_counter = C->unique(); 1790 while (stack.size() > 0) { 1791 Node* n = stack.node(); 1792 uint i = stack.index(); 1793 while (i < n->req() && 1794 (n->in(i) == NULL || 1795 !has_ctrl(n->in(i)) || 1796 get_loop(get_ctrl(n->in(i))) != outer_loop || 1797 (old_new[n->in(i)->_idx] != NULL && old_new[n->in(i)->_idx]->_idx >= new_counter))) { 1798 i++; 1799 } 1800 if (i < n->req()) { 1801 stack.set_index(i+1); 1802 stack.push(n->in(i), 0); 1803 } else { 1804 assert(old_new[n->_idx] == NULL || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet"); 1805 Node* m = n == sfpt ? new_sfpt : n->clone(); 1806 if (m != NULL) { 1807 for (uint i = 0; i < n->req(); i++) { 1808 if (m->in(i) != NULL && old_new[m->in(i)->_idx] != NULL) { 1809 m->set_req(i, old_new[m->in(i)->_idx]); 1810 } 1811 } 1812 } else { 1813 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?"); 1814 } 1815 if (n != sfpt) { 1816 extra_data_nodes.push(n); 1817 _igvn.register_new_node_with_optimizer(m); 1818 assert(get_ctrl(n) == cle_out, "what other control?"); 1819 set_ctrl(m, new_cle_out); 1820 old_new.map(n->_idx, m); 1821 } 1822 stack.pop(); 1823 } 1824 } 1825 if (mode == CloneIncludesStripMined) { 1826 _igvn.register_new_node_with_optimizer(new_sfpt); 1827 _igvn.register_new_node_with_optimizer(new_cle_out); 1828 } 1829 } else { 1830 Node *newhead = old_new[loop->_head->_idx]; 1831 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 1832 } 1833 } 1834 1835 //------------------------------clone_loop------------------------------------- 1836 // 1837 // C L O N E A L O O P B O D Y 1838 // 1839 // This is the basic building block of the loop optimizations. It clones an 1840 // entire loop body. It makes an old_new loop body mapping; with this mapping 1841 // you can find the new-loop equivalent to an old-loop node. All new-loop 1842 // nodes are exactly equal to their old-loop counterparts, all edges are the 1843 // same. All exits from the old-loop now have a RegionNode that merges the 1844 // equivalent new-loop path. This is true even for the normal "loop-exit" 1845 // condition. All uses of loop-invariant old-loop values now come from (one 1846 // or more) Phis that merge their new-loop equivalents. 1847 // 1848 // This operation leaves the graph in an illegal state: there are two valid 1849 // control edges coming from the loop pre-header to both loop bodies. I'll 1850 // definitely have to hack the graph after running this transform. 1851 // 1852 // From this building block I will further edit edges to perform loop peeling 1853 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc. 1854 // 1855 // Parameter side_by_size_idom: 1856 // When side_by_size_idom is NULL, the dominator tree is constructed for 1857 // the clone loop to dominate the original. Used in construction of 1858 // pre-main-post loop sequence. 1859 // When nonnull, the clone and original are side-by-side, both are 1860 // dominated by the side_by_side_idom node. Used in construction of 1861 // unswitched loops. 1862 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd, 1863 CloneLoopMode mode, Node* side_by_side_idom) { 1864 1865 LoopNode* head = loop->_head->as_Loop(); 1866 head->verify_strip_mined(1); 1867 1868 if (C->do_vector_loop() && PrintOpto) { 1869 const char* mname = C->method()->name()->as_quoted_ascii(); 1870 if (mname != NULL) { 1871 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname); 1872 } 1873 } 1874 1875 CloneMap& cm = C->clone_map(); 1876 Dict* dict = cm.dict(); 1877 if (C->do_vector_loop()) { 1878 cm.set_clone_idx(cm.max_gen()+1); 1879 #ifndef PRODUCT 1880 if (PrintOpto) { 1881 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx()); 1882 loop->dump_head(); 1883 } 1884 #endif 1885 } 1886 1887 // Step 1: Clone the loop body. Make the old->new mapping. 1888 uint i; 1889 for( i = 0; i < loop->_body.size(); i++ ) { 1890 Node *old = loop->_body.at(i); 1891 Node *nnn = old->clone(); 1892 old_new.map( old->_idx, nnn ); 1893 if (C->do_vector_loop()) { 1894 cm.verify_insert_and_clone(old, nnn, cm.clone_idx()); 1895 } 1896 _igvn.register_new_node_with_optimizer(nnn); 1897 } 1898 1899 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop; 1900 1901 // Step 2: Fix the edges in the new body. If the old input is outside the 1902 // loop use it. If the old input is INside the loop, use the corresponding 1903 // new node instead. 1904 for( i = 0; i < loop->_body.size(); i++ ) { 1905 Node *old = loop->_body.at(i); 1906 Node *nnn = old_new[old->_idx]; 1907 // Fix CFG/Loop controlling the new node 1908 if (has_ctrl(old)) { 1909 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]); 1910 } else { 1911 set_loop(nnn, outer_loop->_parent); 1912 if (old->outcnt() > 0) { 1913 set_idom( nnn, old_new[idom(old)->_idx], dd ); 1914 } 1915 } 1916 // Correct edges to the new node 1917 for( uint j = 0; j < nnn->req(); j++ ) { 1918 Node *n = nnn->in(j); 1919 if( n ) { 1920 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n ); 1921 if( loop->is_member( old_in_loop ) ) 1922 nnn->set_req(j, old_new[n->_idx]); 1923 } 1924 } 1925 _igvn.hash_find_insert(nnn); 1926 } 1927 1928 ResourceArea *area = Thread::current()->resource_area(); 1929 Node_List extra_data_nodes(area); 1930 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes); 1931 1932 // Step 3: Now fix control uses. Loop varying control uses have already 1933 // been fixed up (as part of all input edges in Step 2). Loop invariant 1934 // control uses must be either an IfFalse or an IfTrue. Make a merge 1935 // point to merge the old and new IfFalse/IfTrue nodes; make the use 1936 // refer to this. 1937 Node_List worklist(area); 1938 uint new_counter = C->unique(); 1939 for( i = 0; i < loop->_body.size(); i++ ) { 1940 Node* old = loop->_body.at(i); 1941 if( !old->is_CFG() ) continue; 1942 1943 // Copy uses to a worklist, so I can munge the def-use info 1944 // with impunity. 1945 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1946 worklist.push(old->fast_out(j)); 1947 1948 while( worklist.size() ) { // Visit all uses 1949 Node *use = worklist.pop(); 1950 if (!has_node(use)) continue; // Ignore dead nodes 1951 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1952 if( !loop->is_member( use_loop ) && use->is_CFG() ) { 1953 // Both OLD and USE are CFG nodes here. 1954 assert( use->is_Proj(), "" ); 1955 Node* nnn = old_new[old->_idx]; 1956 1957 Node* newuse = NULL; 1958 if (head->is_strip_mined() && mode != IgnoreStripMined) { 1959 CountedLoopNode* cl = head->as_CountedLoop(); 1960 CountedLoopEndNode* cle = cl->loopexit(); 1961 Node* cle_out = cle->proj_out_or_null(false); 1962 if (use == cle_out) { 1963 IfNode* le = cl->outer_loop_end(); 1964 use = le->proj_out(false); 1965 use_loop = get_loop(use); 1966 if (mode == CloneIncludesStripMined) { 1967 nnn = old_new[le->_idx]; 1968 } else { 1969 newuse = old_new[cle_out->_idx]; 1970 } 1971 } 1972 } 1973 if (newuse == NULL) { 1974 newuse = use->clone(); 1975 } 1976 1977 // Clone the loop exit control projection 1978 if (C->do_vector_loop()) { 1979 cm.verify_insert_and_clone(use, newuse, cm.clone_idx()); 1980 } 1981 newuse->set_req(0,nnn); 1982 _igvn.register_new_node_with_optimizer(newuse); 1983 set_loop(newuse, use_loop); 1984 set_idom(newuse, nnn, dom_depth(nnn) + 1 ); 1985 1986 // We need a Region to merge the exit from the peeled body and the 1987 // exit from the old loop body. 1988 RegionNode *r = new RegionNode(3); 1989 // Map the old use to the new merge point 1990 old_new.map( use->_idx, r ); 1991 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use)); 1992 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" ); 1993 1994 // The original user of 'use' uses 'r' instead. 1995 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) { 1996 Node* useuse = use->last_out(l); 1997 _igvn.rehash_node_delayed(useuse); 1998 uint uses_found = 0; 1999 if( useuse->in(0) == use ) { 2000 useuse->set_req(0, r); 2001 uses_found++; 2002 if( useuse->is_CFG() ) { 2003 assert( dom_depth(useuse) > dd_r, "" ); 2004 set_idom(useuse, r, dom_depth(useuse)); 2005 } 2006 } 2007 for( uint k = 1; k < useuse->req(); k++ ) { 2008 if( useuse->in(k) == use ) { 2009 useuse->set_req(k, r); 2010 uses_found++; 2011 if (useuse->is_Loop() && k == LoopNode::EntryControl) { 2012 assert(dom_depth(useuse) > dd_r , ""); 2013 set_idom(useuse, r, dom_depth(useuse)); 2014 } 2015 } 2016 } 2017 l -= uses_found; // we deleted 1 or more copies of this edge 2018 } 2019 2020 // Now finish up 'r' 2021 r->set_req( 1, newuse ); 2022 r->set_req( 2, use ); 2023 _igvn.register_new_node_with_optimizer(r); 2024 set_loop(r, use_loop); 2025 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r); 2026 } // End of if a loop-exit test 2027 } 2028 } 2029 2030 // Step 4: If loop-invariant use is not control, it must be dominated by a 2031 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region 2032 // there if needed. Make a Phi there merging old and new used values. 2033 Node_List *split_if_set = NULL; 2034 Node_List *split_bool_set = NULL; 2035 Node_List *split_cex_set = NULL; 2036 for( i = 0; i < loop->_body.size(); i++ ) { 2037 Node* old = loop->_body.at(i); 2038 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set, 2039 split_bool_set, split_cex_set, worklist, new_counter, 2040 mode); 2041 } 2042 2043 for (i = 0; i < extra_data_nodes.size(); i++) { 2044 Node* old = extra_data_nodes.at(i); 2045 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set, 2046 split_bool_set, split_cex_set, worklist, new_counter, 2047 mode); 2048 } 2049 2050 // Check for IFs that need splitting/cloning. Happens if an IF outside of 2051 // the loop uses a condition set in the loop. The original IF probably 2052 // takes control from one or more OLD Regions (which in turn get from NEW 2053 // Regions). In any case, there will be a set of Phis for each merge point 2054 // from the IF up to where the original BOOL def exists the loop. 2055 if (split_if_set) { 2056 while (split_if_set->size()) { 2057 Node *iff = split_if_set->pop(); 2058 if (iff->in(1)->is_Phi()) { 2059 Node *b = clone_iff(iff->in(1)->as_Phi(), loop); 2060 _igvn.replace_input_of(iff, 1, b); 2061 } 2062 } 2063 } 2064 if (split_bool_set) { 2065 while (split_bool_set->size()) { 2066 Node *b = split_bool_set->pop(); 2067 Node *phi = b->in(1); 2068 assert(phi->is_Phi(), ""); 2069 CmpNode *cmp = clone_bool((PhiNode*)phi, loop); 2070 _igvn.replace_input_of(b, 1, cmp); 2071 } 2072 } 2073 if (split_cex_set) { 2074 while (split_cex_set->size()) { 2075 Node *b = split_cex_set->pop(); 2076 assert(b->in(0)->is_Region(), ""); 2077 assert(b->in(1)->is_Phi(), ""); 2078 assert(b->in(0)->in(0) == b->in(1)->in(0), ""); 2079 split_up(b, b->in(0), NULL); 2080 } 2081 } 2082 2083 } 2084 2085 2086 //---------------------- stride_of_possible_iv ------------------------------------- 2087 // Looks for an iff/bool/comp with one operand of the compare 2088 // being a cycle involving an add and a phi, 2089 // with an optional truncation (left-shift followed by a right-shift) 2090 // of the add. Returns zero if not an iv. 2091 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) { 2092 Node* trunc1 = NULL; 2093 Node* trunc2 = NULL; 2094 const TypeInt* ttype = NULL; 2095 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) { 2096 return 0; 2097 } 2098 BoolNode* bl = iff->in(1)->as_Bool(); 2099 Node* cmp = bl->in(1); 2100 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) { 2101 return 0; 2102 } 2103 // Must have an invariant operand 2104 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) { 2105 return 0; 2106 } 2107 Node* add2 = NULL; 2108 Node* cmp1 = cmp->in(1); 2109 if (cmp1->is_Phi()) { 2110 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) ))) 2111 Node* phi = cmp1; 2112 for (uint i = 1; i < phi->req(); i++) { 2113 Node* in = phi->in(i); 2114 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in, 2115 &trunc1, &trunc2, &ttype); 2116 if (add && add->in(1) == phi) { 2117 add2 = add->in(2); 2118 break; 2119 } 2120 } 2121 } else { 2122 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) ))) 2123 Node* addtrunc = cmp1; 2124 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc, 2125 &trunc1, &trunc2, &ttype); 2126 if (add && add->in(1)->is_Phi()) { 2127 Node* phi = add->in(1); 2128 for (uint i = 1; i < phi->req(); i++) { 2129 if (phi->in(i) == addtrunc) { 2130 add2 = add->in(2); 2131 break; 2132 } 2133 } 2134 } 2135 } 2136 if (add2 != NULL) { 2137 const TypeInt* add2t = _igvn.type(add2)->is_int(); 2138 if (add2t->is_con()) { 2139 return add2t->get_con(); 2140 } 2141 } 2142 return 0; 2143 } 2144 2145 2146 //---------------------- stay_in_loop ------------------------------------- 2147 // Return the (unique) control output node that's in the loop (if it exists.) 2148 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) { 2149 Node* unique = NULL; 2150 if (!n) return NULL; 2151 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2152 Node* use = n->fast_out(i); 2153 if (!has_ctrl(use) && loop->is_member(get_loop(use))) { 2154 if (unique != NULL) { 2155 return NULL; 2156 } 2157 unique = use; 2158 } 2159 } 2160 return unique; 2161 } 2162 2163 //------------------------------ register_node ------------------------------------- 2164 // Utility to register node "n" with PhaseIdealLoop 2165 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) { 2166 _igvn.register_new_node_with_optimizer(n); 2167 loop->_body.push(n); 2168 if (n->is_CFG()) { 2169 set_loop(n, loop); 2170 set_idom(n, pred, ddepth); 2171 } else { 2172 set_ctrl(n, pred); 2173 } 2174 } 2175 2176 //------------------------------ proj_clone ------------------------------------- 2177 // Utility to create an if-projection 2178 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) { 2179 ProjNode* c = p->clone()->as_Proj(); 2180 c->set_req(0, iff); 2181 return c; 2182 } 2183 2184 //------------------------------ short_circuit_if ------------------------------------- 2185 // Force the iff control output to be the live_proj 2186 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) { 2187 guarantee(live_proj != NULL, "null projection"); 2188 int proj_con = live_proj->_con; 2189 assert(proj_con == 0 || proj_con == 1, "false or true projection"); 2190 Node *con = _igvn.intcon(proj_con); 2191 set_ctrl(con, C->root()); 2192 if (iff) { 2193 iff->set_req(1, con); 2194 } 2195 return con; 2196 } 2197 2198 //------------------------------ insert_if_before_proj ------------------------------------- 2199 // Insert a new if before an if projection (* - new node) 2200 // 2201 // before 2202 // if(test) 2203 // / \ 2204 // v v 2205 // other-proj proj (arg) 2206 // 2207 // after 2208 // if(test) 2209 // / \ 2210 // / v 2211 // | * proj-clone 2212 // v | 2213 // other-proj v 2214 // * new_if(relop(cmp[IU](left,right))) 2215 // / \ 2216 // v v 2217 // * new-proj proj 2218 // (returned) 2219 // 2220 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) { 2221 IfNode* iff = proj->in(0)->as_If(); 2222 IdealLoopTree *loop = get_loop(proj); 2223 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2224 int ddepth = dom_depth(proj); 2225 2226 _igvn.rehash_node_delayed(iff); 2227 _igvn.rehash_node_delayed(proj); 2228 2229 proj->set_req(0, NULL); // temporary disconnect 2230 ProjNode* proj2 = proj_clone(proj, iff); 2231 register_node(proj2, loop, iff, ddepth); 2232 2233 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right); 2234 register_node(cmp, loop, proj2, ddepth); 2235 2236 BoolNode* bol = new BoolNode(cmp, relop); 2237 register_node(bol, loop, proj2, ddepth); 2238 2239 int opcode = iff->Opcode(); 2240 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode"); 2241 IfNode* new_if = (opcode == Op_If) ? new IfNode(proj2, bol, iff->_prob, iff->_fcnt): 2242 new RangeCheckNode(proj2, bol, iff->_prob, iff->_fcnt); 2243 register_node(new_if, loop, proj2, ddepth); 2244 2245 proj->set_req(0, new_if); // reattach 2246 set_idom(proj, new_if, ddepth); 2247 2248 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj(); 2249 guarantee(new_exit != NULL, "null exit node"); 2250 register_node(new_exit, get_loop(other_proj), new_if, ddepth); 2251 2252 return new_exit; 2253 } 2254 2255 //------------------------------ insert_region_before_proj ------------------------------------- 2256 // Insert a region before an if projection (* - new node) 2257 // 2258 // before 2259 // if(test) 2260 // / | 2261 // v | 2262 // proj v 2263 // other-proj 2264 // 2265 // after 2266 // if(test) 2267 // / | 2268 // v | 2269 // * proj-clone v 2270 // | other-proj 2271 // v 2272 // * new-region 2273 // | 2274 // v 2275 // * dum_if 2276 // / \ 2277 // v \ 2278 // * dum-proj v 2279 // proj 2280 // 2281 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) { 2282 IfNode* iff = proj->in(0)->as_If(); 2283 IdealLoopTree *loop = get_loop(proj); 2284 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2285 int ddepth = dom_depth(proj); 2286 2287 _igvn.rehash_node_delayed(iff); 2288 _igvn.rehash_node_delayed(proj); 2289 2290 proj->set_req(0, NULL); // temporary disconnect 2291 ProjNode* proj2 = proj_clone(proj, iff); 2292 register_node(proj2, loop, iff, ddepth); 2293 2294 RegionNode* reg = new RegionNode(2); 2295 reg->set_req(1, proj2); 2296 register_node(reg, loop, iff, ddepth); 2297 2298 IfNode* dum_if = new IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt); 2299 register_node(dum_if, loop, reg, ddepth); 2300 2301 proj->set_req(0, dum_if); // reattach 2302 set_idom(proj, dum_if, ddepth); 2303 2304 ProjNode* dum_proj = proj_clone(other_proj, dum_if); 2305 register_node(dum_proj, loop, dum_if, ddepth); 2306 2307 return reg; 2308 } 2309 2310 //------------------------------ insert_cmpi_loop_exit ------------------------------------- 2311 // Clone a signed compare loop exit from an unsigned compare and 2312 // insert it before the unsigned cmp on the stay-in-loop path. 2313 // All new nodes inserted in the dominator tree between the original 2314 // if and it's projections. The original if test is replaced with 2315 // a constant to force the stay-in-loop path. 2316 // 2317 // This is done to make sure that the original if and it's projections 2318 // still dominate the same set of control nodes, that the ctrl() relation 2319 // from data nodes to them is preserved, and that their loop nesting is 2320 // preserved. 2321 // 2322 // before 2323 // if(i <u limit) unsigned compare loop exit 2324 // / | 2325 // v v 2326 // exit-proj stay-in-loop-proj 2327 // 2328 // after 2329 // if(stay-in-loop-const) original if 2330 // / | 2331 // / v 2332 // / if(i < limit) new signed test 2333 // / / | 2334 // / / v 2335 // / / if(i <u limit) new cloned unsigned test 2336 // / / / | 2337 // v v v | 2338 // region | 2339 // | | 2340 // dum-if | 2341 // / | | 2342 // ether | | 2343 // v v 2344 // exit-proj stay-in-loop-proj 2345 // 2346 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) { 2347 const bool Signed = true; 2348 const bool Unsigned = false; 2349 2350 BoolNode* bol = if_cmpu->in(1)->as_Bool(); 2351 if (bol->_test._test != BoolTest::lt) return NULL; 2352 CmpNode* cmpu = bol->in(1)->as_Cmp(); 2353 if (cmpu->Opcode() != Op_CmpU) return NULL; 2354 int stride = stride_of_possible_iv(if_cmpu); 2355 if (stride == 0) return NULL; 2356 2357 Node* lp_proj = stay_in_loop(if_cmpu, loop); 2358 guarantee(lp_proj != NULL, "null loop node"); 2359 2360 ProjNode* lp_continue = lp_proj->as_Proj(); 2361 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj(); 2362 2363 Node* limit = NULL; 2364 if (stride > 0) { 2365 limit = cmpu->in(2); 2366 } else { 2367 limit = _igvn.makecon(TypeInt::ZERO); 2368 set_ctrl(limit, C->root()); 2369 } 2370 // Create a new region on the exit path 2371 RegionNode* reg = insert_region_before_proj(lp_exit); 2372 guarantee(reg != NULL, "null region node"); 2373 2374 // Clone the if-cmpu-true-false using a signed compare 2375 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge; 2376 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue); 2377 reg->add_req(cmpi_exit); 2378 2379 // Clone the if-cmpu-true-false 2380 BoolTest::mask rel_u = bol->_test._test; 2381 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue); 2382 reg->add_req(cmpu_exit); 2383 2384 // Force original if to stay in loop. 2385 short_circuit_if(if_cmpu, lp_continue); 2386 2387 return cmpi_exit->in(0)->as_If(); 2388 } 2389 2390 //------------------------------ remove_cmpi_loop_exit ------------------------------------- 2391 // Remove a previously inserted signed compare loop exit. 2392 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) { 2393 Node* lp_proj = stay_in_loop(if_cmp, loop); 2394 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI && 2395 stay_in_loop(lp_proj, loop)->is_If() && 2396 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu"); 2397 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO); 2398 set_ctrl(con, C->root()); 2399 if_cmp->set_req(1, con); 2400 } 2401 2402 //------------------------------ scheduled_nodelist ------------------------------------- 2403 // Create a post order schedule of nodes that are in the 2404 // "member" set. The list is returned in "sched". 2405 // The first node in "sched" is the loop head, followed by 2406 // nodes which have no inputs in the "member" set, and then 2407 // followed by the nodes that have an immediate input dependence 2408 // on a node in "sched". 2409 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) { 2410 2411 assert(member.test(loop->_head->_idx), "loop head must be in member set"); 2412 Arena *a = Thread::current()->resource_area(); 2413 VectorSet visited(a); 2414 Node_Stack nstack(a, loop->_body.size()); 2415 2416 Node* n = loop->_head; // top of stack is cached in "n" 2417 uint idx = 0; 2418 visited.set(n->_idx); 2419 2420 // Initially push all with no inputs from within member set 2421 for(uint i = 0; i < loop->_body.size(); i++ ) { 2422 Node *elt = loop->_body.at(i); 2423 if (member.test(elt->_idx)) { 2424 bool found = false; 2425 for (uint j = 0; j < elt->req(); j++) { 2426 Node* def = elt->in(j); 2427 if (def && member.test(def->_idx) && def != elt) { 2428 found = true; 2429 break; 2430 } 2431 } 2432 if (!found && elt != loop->_head) { 2433 nstack.push(n, idx); 2434 n = elt; 2435 assert(!visited.test(n->_idx), "not seen yet"); 2436 visited.set(n->_idx); 2437 } 2438 } 2439 } 2440 2441 // traverse out's that are in the member set 2442 while (true) { 2443 if (idx < n->outcnt()) { 2444 Node* use = n->raw_out(idx); 2445 idx++; 2446 if (!visited.test_set(use->_idx)) { 2447 if (member.test(use->_idx)) { 2448 nstack.push(n, idx); 2449 n = use; 2450 idx = 0; 2451 } 2452 } 2453 } else { 2454 // All outputs processed 2455 sched.push(n); 2456 if (nstack.is_empty()) break; 2457 n = nstack.node(); 2458 idx = nstack.index(); 2459 nstack.pop(); 2460 } 2461 } 2462 } 2463 2464 2465 //------------------------------ has_use_in_set ------------------------------------- 2466 // Has a use in the vector set 2467 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) { 2468 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2469 Node* use = n->fast_out(j); 2470 if (vset.test(use->_idx)) { 2471 return true; 2472 } 2473 } 2474 return false; 2475 } 2476 2477 2478 //------------------------------ has_use_internal_to_set ------------------------------------- 2479 // Has use internal to the vector set (ie. not in a phi at the loop head) 2480 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) { 2481 Node* head = loop->_head; 2482 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2483 Node* use = n->fast_out(j); 2484 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) { 2485 return true; 2486 } 2487 } 2488 return false; 2489 } 2490 2491 2492 //------------------------------ clone_for_use_outside_loop ------------------------------------- 2493 // clone "n" for uses that are outside of loop 2494 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) { 2495 int cloned = 0; 2496 assert(worklist.size() == 0, "should be empty"); 2497 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2498 Node* use = n->fast_out(j); 2499 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) { 2500 worklist.push(use); 2501 } 2502 } 2503 while( worklist.size() ) { 2504 Node *use = worklist.pop(); 2505 if (!has_node(use) || use->in(0) == C->top()) continue; 2506 uint j; 2507 for (j = 0; j < use->req(); j++) { 2508 if (use->in(j) == n) break; 2509 } 2510 assert(j < use->req(), "must be there"); 2511 2512 // clone "n" and insert it between the inputs of "n" and the use outside the loop 2513 Node* n_clone = n->clone(); 2514 _igvn.replace_input_of(use, j, n_clone); 2515 cloned++; 2516 Node* use_c; 2517 if (!use->is_Phi()) { 2518 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0); 2519 } else { 2520 // Use in a phi is considered a use in the associated predecessor block 2521 use_c = use->in(0)->in(j); 2522 } 2523 set_ctrl(n_clone, use_c); 2524 assert(!loop->is_member(get_loop(use_c)), "should be outside loop"); 2525 get_loop(use_c)->_body.push(n_clone); 2526 _igvn.register_new_node_with_optimizer(n_clone); 2527 #if !defined(PRODUCT) 2528 if (TracePartialPeeling) { 2529 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx); 2530 } 2531 #endif 2532 } 2533 return cloned; 2534 } 2535 2536 2537 //------------------------------ clone_for_special_use_inside_loop ------------------------------------- 2538 // clone "n" for special uses that are in the not_peeled region. 2539 // If these def-uses occur in separate blocks, the code generator 2540 // marks the method as not compilable. For example, if a "BoolNode" 2541 // is in a different basic block than the "IfNode" that uses it, then 2542 // the compilation is aborted in the code generator. 2543 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 2544 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) { 2545 if (n->is_Phi() || n->is_Load()) { 2546 return; 2547 } 2548 assert(worklist.size() == 0, "should be empty"); 2549 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2550 Node* use = n->fast_out(j); 2551 if ( not_peel.test(use->_idx) && 2552 (use->is_If() || use->is_CMove() || use->is_Bool()) && 2553 use->in(1) == n) { 2554 worklist.push(use); 2555 } 2556 } 2557 if (worklist.size() > 0) { 2558 // clone "n" and insert it between inputs of "n" and the use 2559 Node* n_clone = n->clone(); 2560 loop->_body.push(n_clone); 2561 _igvn.register_new_node_with_optimizer(n_clone); 2562 set_ctrl(n_clone, get_ctrl(n)); 2563 sink_list.push(n_clone); 2564 not_peel <<= n_clone->_idx; // add n_clone to not_peel set. 2565 #if !defined(PRODUCT) 2566 if (TracePartialPeeling) { 2567 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx); 2568 } 2569 #endif 2570 while( worklist.size() ) { 2571 Node *use = worklist.pop(); 2572 _igvn.rehash_node_delayed(use); 2573 for (uint j = 1; j < use->req(); j++) { 2574 if (use->in(j) == n) { 2575 use->set_req(j, n_clone); 2576 } 2577 } 2578 } 2579 } 2580 } 2581 2582 2583 //------------------------------ insert_phi_for_loop ------------------------------------- 2584 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 2585 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) { 2586 Node *phi = PhiNode::make(lp, back_edge_val); 2587 phi->set_req(LoopNode::EntryControl, lp_entry_val); 2588 // Use existing phi if it already exists 2589 Node *hit = _igvn.hash_find_insert(phi); 2590 if( hit == NULL ) { 2591 _igvn.register_new_node_with_optimizer(phi); 2592 set_ctrl(phi, lp); 2593 } else { 2594 // Remove the new phi from the graph and use the hit 2595 _igvn.remove_dead_node(phi); 2596 phi = hit; 2597 } 2598 _igvn.replace_input_of(use, idx, phi); 2599 } 2600 2601 #ifdef ASSERT 2602 //------------------------------ is_valid_loop_partition ------------------------------------- 2603 // Validate the loop partition sets: peel and not_peel 2604 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, 2605 VectorSet& not_peel ) { 2606 uint i; 2607 // Check that peel_list entries are in the peel set 2608 for (i = 0; i < peel_list.size(); i++) { 2609 if (!peel.test(peel_list.at(i)->_idx)) { 2610 return false; 2611 } 2612 } 2613 // Check at loop members are in one of peel set or not_peel set 2614 for (i = 0; i < loop->_body.size(); i++ ) { 2615 Node *def = loop->_body.at(i); 2616 uint di = def->_idx; 2617 // Check that peel set elements are in peel_list 2618 if (peel.test(di)) { 2619 if (not_peel.test(di)) { 2620 return false; 2621 } 2622 // Must be in peel_list also 2623 bool found = false; 2624 for (uint j = 0; j < peel_list.size(); j++) { 2625 if (peel_list.at(j)->_idx == di) { 2626 found = true; 2627 break; 2628 } 2629 } 2630 if (!found) { 2631 return false; 2632 } 2633 } else if (not_peel.test(di)) { 2634 if (peel.test(di)) { 2635 return false; 2636 } 2637 } else { 2638 return false; 2639 } 2640 } 2641 return true; 2642 } 2643 2644 //------------------------------ is_valid_clone_loop_exit_use ------------------------------------- 2645 // Ensure a use outside of loop is of the right form 2646 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) { 2647 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2648 return (use->is_Phi() && 2649 use_c->is_Region() && use_c->req() == 3 && 2650 (use_c->in(exit_idx)->Opcode() == Op_IfTrue || 2651 use_c->in(exit_idx)->Opcode() == Op_IfFalse || 2652 use_c->in(exit_idx)->Opcode() == Op_JumpProj) && 2653 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) ); 2654 } 2655 2656 //------------------------------ is_valid_clone_loop_form ------------------------------------- 2657 // Ensure that all uses outside of loop are of the right form 2658 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 2659 uint orig_exit_idx, uint clone_exit_idx) { 2660 uint len = peel_list.size(); 2661 for (uint i = 0; i < len; i++) { 2662 Node *def = peel_list.at(i); 2663 2664 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 2665 Node *use = def->fast_out(j); 2666 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2667 if (!loop->is_member(get_loop(use_c))) { 2668 // use is not in the loop, check for correct structure 2669 if (use->in(0) == def) { 2670 // Okay 2671 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) { 2672 return false; 2673 } 2674 } 2675 } 2676 } 2677 return true; 2678 } 2679 #endif 2680 2681 //------------------------------ partial_peel ------------------------------------- 2682 // Partially peel (aka loop rotation) the top portion of a loop (called 2683 // the peel section below) by cloning it and placing one copy just before 2684 // the new loop head and the other copy at the bottom of the new loop. 2685 // 2686 // before after where it came from 2687 // 2688 // stmt1 stmt1 2689 // loop: stmt2 clone 2690 // stmt2 if condA goto exitA clone 2691 // if condA goto exitA new_loop: new 2692 // stmt3 stmt3 clone 2693 // if !condB goto loop if condB goto exitB clone 2694 // exitB: stmt2 orig 2695 // stmt4 if !condA goto new_loop orig 2696 // exitA: goto exitA 2697 // exitB: 2698 // stmt4 2699 // exitA: 2700 // 2701 // Step 1: find the cut point: an exit test on probable 2702 // induction variable. 2703 // Step 2: schedule (with cloning) operations in the peel 2704 // section that can be executed after the cut into 2705 // the section that is not peeled. This may need 2706 // to clone operations into exit blocks. For 2707 // instance, a reference to A[i] in the not-peel 2708 // section and a reference to B[i] in an exit block 2709 // may cause a left-shift of i by 2 to be placed 2710 // in the peel block. This step will clone the left 2711 // shift into the exit block and sink the left shift 2712 // from the peel to the not-peel section. 2713 // Step 3: clone the loop, retarget the control, and insert 2714 // phis for values that are live across the new loop 2715 // head. This is very dependent on the graph structure 2716 // from clone_loop. It creates region nodes for 2717 // exit control and associated phi nodes for values 2718 // flow out of the loop through that exit. The region 2719 // node is dominated by the clone's control projection. 2720 // So the clone's peel section is placed before the 2721 // new loop head, and the clone's not-peel section is 2722 // forms the top part of the new loop. The original 2723 // peel section forms the tail of the new loop. 2724 // Step 4: update the dominator tree and recompute the 2725 // dominator depth. 2726 // 2727 // orig 2728 // 2729 // stmt1 2730 // | 2731 // v 2732 // loop predicate 2733 // | 2734 // v 2735 // loop<----+ 2736 // | | 2737 // stmt2 | 2738 // | | 2739 // v | 2740 // ifA | 2741 // / | | 2742 // v v | 2743 // false true ^ <-- last_peel 2744 // / | | 2745 // / ===|==cut | 2746 // / stmt3 | <-- first_not_peel 2747 // / | | 2748 // | v | 2749 // v ifB | 2750 // exitA: / \ | 2751 // / \ | 2752 // v v | 2753 // false true | 2754 // / \ | 2755 // / ----+ 2756 // | 2757 // v 2758 // exitB: 2759 // stmt4 2760 // 2761 // 2762 // after clone loop 2763 // 2764 // stmt1 2765 // | 2766 // v 2767 // loop predicate 2768 // / \ 2769 // clone / \ orig 2770 // / \ 2771 // / \ 2772 // v v 2773 // +---->loop loop<----+ 2774 // | | | | 2775 // | stmt2 stmt2 | 2776 // | | | | 2777 // | v v | 2778 // | ifA ifA | 2779 // | | \ / | | 2780 // | v v v v | 2781 // ^ true false false true ^ <-- last_peel 2782 // | | ^ \ / | | 2783 // | cut==|== \ \ / ===|==cut | 2784 // | stmt3 \ \ / stmt3 | <-- first_not_peel 2785 // | | dom | | | | 2786 // | v \ 1v v2 v | 2787 // | ifB regionA ifB | 2788 // | / \ | / \ | 2789 // | / \ v / \ | 2790 // | v v exitA: v v | 2791 // | true false false true | 2792 // | / ^ \ / \ | 2793 // +---- \ \ / ----+ 2794 // dom \ / 2795 // \ 1v v2 2796 // regionB 2797 // | 2798 // v 2799 // exitB: 2800 // stmt4 2801 // 2802 // 2803 // after partial peel 2804 // 2805 // stmt1 2806 // | 2807 // v 2808 // loop predicate 2809 // / 2810 // clone / orig 2811 // / TOP 2812 // / \ 2813 // v v 2814 // TOP->loop loop----+ 2815 // | | | 2816 // stmt2 stmt2 | 2817 // | | | 2818 // v v | 2819 // ifA ifA | 2820 // | \ / | | 2821 // v v v v | 2822 // true false false true | <-- last_peel 2823 // | ^ \ / +------|---+ 2824 // +->newloop \ \ / === ==cut | | 2825 // | stmt3 \ \ / TOP | | 2826 // | | dom | | stmt3 | | <-- first_not_peel 2827 // | v \ 1v v2 v | | 2828 // | ifB regionA ifB ^ v 2829 // | / \ | / \ | | 2830 // | / \ v / \ | | 2831 // | v v exitA: v v | | 2832 // | true false false true | | 2833 // | / ^ \ / \ | | 2834 // | | \ \ / v | | 2835 // | | dom \ / TOP | | 2836 // | | \ 1v v2 | | 2837 // ^ v regionB | | 2838 // | | | | | 2839 // | | v ^ v 2840 // | | exitB: | | 2841 // | | stmt4 | | 2842 // | +------------>-----------------+ | 2843 // | | 2844 // +-----------------<---------------------+ 2845 // 2846 // 2847 // final graph 2848 // 2849 // stmt1 2850 // | 2851 // v 2852 // loop predicate 2853 // | 2854 // v 2855 // stmt2 clone 2856 // | 2857 // v 2858 // ........> ifA clone 2859 // : / | 2860 // dom / | 2861 // : v v 2862 // : false true 2863 // : | | 2864 // : | v 2865 // : | newloop<-----+ 2866 // : | | | 2867 // : | stmt3 clone | 2868 // : | | | 2869 // : | v | 2870 // : | ifB | 2871 // : | / \ | 2872 // : | v v | 2873 // : | false true | 2874 // : | | | | 2875 // : | v stmt2 | 2876 // : | exitB: | | 2877 // : | stmt4 v | 2878 // : | ifA orig | 2879 // : | / \ | 2880 // : | / \ | 2881 // : | v v | 2882 // : | false true | 2883 // : | / \ | 2884 // : v v -----+ 2885 // RegionA 2886 // | 2887 // v 2888 // exitA 2889 // 2890 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) { 2891 2892 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only"); 2893 if (!loop->_head->is_Loop()) { 2894 return false; } 2895 2896 LoopNode *head = loop->_head->as_Loop(); 2897 2898 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) { 2899 return false; 2900 } 2901 2902 // Check for complex exit control 2903 for(uint ii = 0; ii < loop->_body.size(); ii++ ) { 2904 Node *n = loop->_body.at(ii); 2905 int opc = n->Opcode(); 2906 if (n->is_Call() || 2907 opc == Op_Catch || 2908 opc == Op_CatchProj || 2909 opc == Op_Jump || 2910 opc == Op_JumpProj) { 2911 #if !defined(PRODUCT) 2912 if (TracePartialPeeling) { 2913 tty->print_cr("\nExit control too complex: lp: %d", head->_idx); 2914 } 2915 #endif 2916 return false; 2917 } 2918 } 2919 2920 int dd = dom_depth(head); 2921 2922 // Step 1: find cut point 2923 2924 // Walk up dominators to loop head looking for first loop exit 2925 // which is executed on every path thru loop. 2926 IfNode *peel_if = NULL; 2927 IfNode *peel_if_cmpu = NULL; 2928 2929 Node *iff = loop->tail(); 2930 while( iff != head ) { 2931 if( iff->is_If() ) { 2932 Node *ctrl = get_ctrl(iff->in(1)); 2933 if (ctrl->is_top()) return false; // Dead test on live IF. 2934 // If loop-varying exit-test, check for induction variable 2935 if( loop->is_member(get_loop(ctrl)) && 2936 loop->is_loop_exit(iff) && 2937 is_possible_iv_test(iff)) { 2938 Node* cmp = iff->in(1)->in(1); 2939 if (cmp->Opcode() == Op_CmpI) { 2940 peel_if = iff->as_If(); 2941 } else { 2942 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU"); 2943 peel_if_cmpu = iff->as_If(); 2944 } 2945 } 2946 } 2947 iff = idom(iff); 2948 } 2949 // Prefer signed compare over unsigned compare. 2950 IfNode* new_peel_if = NULL; 2951 if (peel_if == NULL) { 2952 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) { 2953 return false; // No peel point found 2954 } 2955 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop); 2956 if (new_peel_if == NULL) { 2957 return false; // No peel point found 2958 } 2959 peel_if = new_peel_if; 2960 } 2961 Node* last_peel = stay_in_loop(peel_if, loop); 2962 Node* first_not_peeled = stay_in_loop(last_peel, loop); 2963 if (first_not_peeled == NULL || first_not_peeled == head) { 2964 return false; 2965 } 2966 2967 #if !defined(PRODUCT) 2968 if (TraceLoopOpts) { 2969 tty->print("PartialPeel "); 2970 loop->dump_head(); 2971 } 2972 2973 if (TracePartialPeeling) { 2974 tty->print_cr("before partial peel one iteration"); 2975 Node_List wl; 2976 Node* t = head->in(2); 2977 while (true) { 2978 wl.push(t); 2979 if (t == head) break; 2980 t = idom(t); 2981 } 2982 while (wl.size() > 0) { 2983 Node* tt = wl.pop(); 2984 tt->dump(); 2985 if (tt == last_peel) tty->print_cr("-- cut --"); 2986 } 2987 } 2988 #endif 2989 ResourceArea *area = Thread::current()->resource_area(); 2990 VectorSet peel(area); 2991 VectorSet not_peel(area); 2992 Node_List peel_list(area); 2993 Node_List worklist(area); 2994 Node_List sink_list(area); 2995 2996 // Set of cfg nodes to peel are those that are executable from 2997 // the head through last_peel. 2998 assert(worklist.size() == 0, "should be empty"); 2999 worklist.push(head); 3000 peel.set(head->_idx); 3001 while (worklist.size() > 0) { 3002 Node *n = worklist.pop(); 3003 if (n != last_peel) { 3004 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3005 Node* use = n->fast_out(j); 3006 if (use->is_CFG() && 3007 loop->is_member(get_loop(use)) && 3008 !peel.test_set(use->_idx)) { 3009 worklist.push(use); 3010 } 3011 } 3012 } 3013 } 3014 3015 // Set of non-cfg nodes to peel are those that are control 3016 // dependent on the cfg nodes. 3017 uint i; 3018 for(i = 0; i < loop->_body.size(); i++ ) { 3019 Node *n = loop->_body.at(i); 3020 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n; 3021 if (peel.test(n_c->_idx)) { 3022 peel.set(n->_idx); 3023 } else { 3024 not_peel.set(n->_idx); 3025 } 3026 } 3027 3028 // Step 2: move operations from the peeled section down into the 3029 // not-peeled section 3030 3031 // Get a post order schedule of nodes in the peel region 3032 // Result in right-most operand. 3033 scheduled_nodelist(loop, peel, peel_list ); 3034 3035 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 3036 3037 // For future check for too many new phis 3038 uint old_phi_cnt = 0; 3039 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 3040 Node* use = head->fast_out(j); 3041 if (use->is_Phi()) old_phi_cnt++; 3042 } 3043 3044 #if !defined(PRODUCT) 3045 if (TracePartialPeeling) { 3046 tty->print_cr("\npeeled list"); 3047 } 3048 #endif 3049 3050 // Evacuate nodes in peel region into the not_peeled region if possible 3051 uint new_phi_cnt = 0; 3052 uint cloned_for_outside_use = 0; 3053 for (i = 0; i < peel_list.size();) { 3054 Node* n = peel_list.at(i); 3055 #if !defined(PRODUCT) 3056 if (TracePartialPeeling) n->dump(); 3057 #endif 3058 bool incr = true; 3059 if ( !n->is_CFG() ) { 3060 3061 if ( has_use_in_set(n, not_peel) ) { 3062 3063 // If not used internal to the peeled region, 3064 // move "n" from peeled to not_peeled region. 3065 3066 if ( !has_use_internal_to_set(n, peel, loop) ) { 3067 3068 // if not pinned and not a load (which maybe anti-dependent on a store) 3069 // and not a CMove (Matcher expects only bool->cmove). 3070 if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) { 3071 cloned_for_outside_use += clone_for_use_outside_loop( loop, n, worklist ); 3072 sink_list.push(n); 3073 peel >>= n->_idx; // delete n from peel set. 3074 not_peel <<= n->_idx; // add n to not_peel set. 3075 peel_list.remove(i); 3076 incr = false; 3077 #if !defined(PRODUCT) 3078 if (TracePartialPeeling) { 3079 tty->print_cr("sink to not_peeled region: %d newbb: %d", 3080 n->_idx, get_ctrl(n)->_idx); 3081 } 3082 #endif 3083 } 3084 } else { 3085 // Otherwise check for special def-use cases that span 3086 // the peel/not_peel boundary such as bool->if 3087 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist ); 3088 new_phi_cnt++; 3089 } 3090 } 3091 } 3092 if (incr) i++; 3093 } 3094 3095 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) { 3096 #if !defined(PRODUCT) 3097 if (TracePartialPeeling) { 3098 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c", 3099 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F'); 3100 } 3101 #endif 3102 if (new_peel_if != NULL) { 3103 remove_cmpi_loop_exit(new_peel_if, loop); 3104 } 3105 // Inhibit more partial peeling on this loop 3106 assert(!head->is_partial_peel_loop(), "not partial peeled"); 3107 head->mark_partial_peel_failed(); 3108 if (cloned_for_outside_use > 0) { 3109 // Terminate this round of loop opts because 3110 // the graph outside this loop was changed. 3111 C->set_major_progress(); 3112 return true; 3113 } 3114 return false; 3115 } 3116 3117 // Step 3: clone loop, retarget control, and insert new phis 3118 3119 // Create new loop head for new phis and to hang 3120 // the nodes being moved (sinked) from the peel region. 3121 LoopNode* new_head = new LoopNode(last_peel, last_peel); 3122 new_head->set_unswitch_count(head->unswitch_count()); // Preserve 3123 _igvn.register_new_node_with_optimizer(new_head); 3124 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled"); 3125 _igvn.replace_input_of(first_not_peeled, 0, new_head); 3126 set_loop(new_head, loop); 3127 loop->_body.push(new_head); 3128 not_peel.set(new_head->_idx); 3129 set_idom(new_head, last_peel, dom_depth(first_not_peeled)); 3130 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled)); 3131 3132 while (sink_list.size() > 0) { 3133 Node* n = sink_list.pop(); 3134 set_ctrl(n, new_head); 3135 } 3136 3137 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 3138 3139 clone_loop(loop, old_new, dd, IgnoreStripMined); 3140 3141 const uint clone_exit_idx = 1; 3142 const uint orig_exit_idx = 2; 3143 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop"); 3144 3145 Node* head_clone = old_new[head->_idx]; 3146 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop(); 3147 Node* orig_tail_clone = head_clone->in(2); 3148 3149 // Add phi if "def" node is in peel set and "use" is not 3150 3151 for(i = 0; i < peel_list.size(); i++ ) { 3152 Node *def = peel_list.at(i); 3153 if (!def->is_CFG()) { 3154 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 3155 Node *use = def->fast_out(j); 3156 if (has_node(use) && use->in(0) != C->top() && 3157 (!peel.test(use->_idx) || 3158 (use->is_Phi() && use->in(0) == head)) ) { 3159 worklist.push(use); 3160 } 3161 } 3162 while( worklist.size() ) { 3163 Node *use = worklist.pop(); 3164 for (uint j = 1; j < use->req(); j++) { 3165 Node* n = use->in(j); 3166 if (n == def) { 3167 3168 // "def" is in peel set, "use" is not in peel set 3169 // or "use" is in the entry boundary (a phi) of the peel set 3170 3171 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use; 3172 3173 if ( loop->is_member(get_loop( use_c )) ) { 3174 // use is in loop 3175 if (old_new[use->_idx] != NULL) { // null for dead code 3176 Node* use_clone = old_new[use->_idx]; 3177 _igvn.replace_input_of(use, j, C->top()); 3178 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone ); 3179 } 3180 } else { 3181 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format"); 3182 // use is not in the loop, check if the live range includes the cut 3183 Node* lp_if = use_c->in(orig_exit_idx)->in(0); 3184 if (not_peel.test(lp_if->_idx)) { 3185 assert(j == orig_exit_idx, "use from original loop"); 3186 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone ); 3187 } 3188 } 3189 } 3190 } 3191 } 3192 } 3193 } 3194 3195 // Step 3b: retarget control 3196 3197 // Redirect control to the new loop head if a cloned node in 3198 // the not_peeled region has control that points into the peeled region. 3199 // This necessary because the cloned peeled region will be outside 3200 // the loop. 3201 // from to 3202 // cloned-peeled <---+ 3203 // new_head_clone: | <--+ 3204 // cloned-not_peeled in(0) in(0) 3205 // orig-peeled 3206 3207 for(i = 0; i < loop->_body.size(); i++ ) { 3208 Node *n = loop->_body.at(i); 3209 if (!n->is_CFG() && n->in(0) != NULL && 3210 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) { 3211 Node* n_clone = old_new[n->_idx]; 3212 _igvn.replace_input_of(n_clone, 0, new_head_clone); 3213 } 3214 } 3215 3216 // Backedge of the surviving new_head (the clone) is original last_peel 3217 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel); 3218 3219 // Cut first node in original not_peel set 3220 _igvn.rehash_node_delayed(new_head); // Multiple edge updates: 3221 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of 3222 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls 3223 3224 // Copy head_clone back-branch info to original head 3225 // and remove original head's loop entry and 3226 // clone head's back-branch 3227 _igvn.rehash_node_delayed(head); // Multiple edge updates 3228 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl)); 3229 head->set_req(LoopNode::LoopBackControl, C->top()); 3230 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top()); 3231 3232 // Similarly modify the phis 3233 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) { 3234 Node* use = head->fast_out(k); 3235 if (use->is_Phi() && use->outcnt() > 0) { 3236 Node* use_clone = old_new[use->_idx]; 3237 _igvn.rehash_node_delayed(use); // Multiple edge updates 3238 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl)); 3239 use->set_req(LoopNode::LoopBackControl, C->top()); 3240 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top()); 3241 } 3242 } 3243 3244 // Step 4: update dominator tree and dominator depth 3245 3246 set_idom(head, orig_tail_clone, dd); 3247 recompute_dom_depth(); 3248 3249 // Inhibit more partial peeling on this loop 3250 new_head_clone->set_partial_peel_loop(); 3251 C->set_major_progress(); 3252 loop->record_for_igvn(); 3253 3254 #if !defined(PRODUCT) 3255 if (TracePartialPeeling) { 3256 tty->print_cr("\nafter partial peel one iteration"); 3257 Node_List wl(area); 3258 Node* t = last_peel; 3259 while (true) { 3260 wl.push(t); 3261 if (t == head_clone) break; 3262 t = idom(t); 3263 } 3264 while (wl.size() > 0) { 3265 Node* tt = wl.pop(); 3266 if (tt == head) tty->print_cr("orig head"); 3267 else if (tt == new_head_clone) tty->print_cr("new head"); 3268 else if (tt == head_clone) tty->print_cr("clone head"); 3269 tt->dump(); 3270 } 3271 } 3272 #endif 3273 return true; 3274 } 3275 3276 //------------------------------reorg_offsets---------------------------------- 3277 // Reorganize offset computations to lower register pressure. Mostly 3278 // prevent loop-fallout uses of the pre-incremented trip counter (which are 3279 // then alive with the post-incremented trip counter forcing an extra 3280 // register move) 3281 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) { 3282 // Perform it only for canonical counted loops. 3283 // Loop's shape could be messed up by iteration_split_impl. 3284 if (!loop->_head->is_CountedLoop()) 3285 return; 3286 if (!loop->_head->as_Loop()->is_valid_counted_loop()) 3287 return; 3288 3289 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 3290 CountedLoopEndNode *cle = cl->loopexit(); 3291 Node *exit = cle->proj_out(false); 3292 Node *phi = cl->phi(); 3293 3294 // Check for the special case of folks using the pre-incremented 3295 // trip-counter on the fall-out path (forces the pre-incremented 3296 // and post-incremented trip counter to be live at the same time). 3297 // Fix this by adjusting to use the post-increment trip counter. 3298 3299 bool progress = true; 3300 while (progress) { 3301 progress = false; 3302 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) { 3303 Node* use = phi->fast_out(i); // User of trip-counter 3304 if (!has_ctrl(use)) continue; 3305 Node *u_ctrl = get_ctrl(use); 3306 if (use->is_Phi()) { 3307 u_ctrl = NULL; 3308 for (uint j = 1; j < use->req(); j++) 3309 if (use->in(j) == phi) 3310 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j)); 3311 } 3312 IdealLoopTree *u_loop = get_loop(u_ctrl); 3313 // Look for loop-invariant use 3314 if (u_loop == loop) continue; 3315 if (loop->is_member(u_loop)) continue; 3316 // Check that use is live out the bottom. Assuming the trip-counter 3317 // update is right at the bottom, uses of of the loop middle are ok. 3318 if (dom_lca(exit, u_ctrl) != exit) continue; 3319 // Hit! Refactor use to use the post-incremented tripcounter. 3320 // Compute a post-increment tripcounter. 3321 Node *opaq = new Opaque2Node( C, cle->incr() ); 3322 register_new_node(opaq, exit); 3323 Node *neg_stride = _igvn.intcon(-cle->stride_con()); 3324 set_ctrl(neg_stride, C->root()); 3325 Node *post = new AddINode( opaq, neg_stride); 3326 register_new_node(post, exit); 3327 _igvn.rehash_node_delayed(use); 3328 for (uint j = 1; j < use->req(); j++) { 3329 if (use->in(j) == phi) 3330 use->set_req(j, post); 3331 } 3332 // Since DU info changed, rerun loop 3333 progress = true; 3334 break; 3335 } 3336 } 3337 3338 }