1 /* 2 * Copyright (c) 1999, 2018, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "memory/allocation.inline.hpp" 27 #include "memory/resourceArea.hpp" 28 #include "opto/addnode.hpp" 29 #include "opto/callnode.hpp" 30 #include "opto/castnode.hpp" 31 #include "opto/connode.hpp" 32 #include "opto/castnode.hpp" 33 #include "opto/divnode.hpp" 34 #include "opto/loopnode.hpp" 35 #include "opto/matcher.hpp" 36 #include "opto/mulnode.hpp" 37 #include "opto/movenode.hpp" 38 #include "opto/opaquenode.hpp" 39 #include "opto/rootnode.hpp" 40 #include "opto/subnode.hpp" 41 42 //============================================================================= 43 //------------------------------split_thru_phi--------------------------------- 44 // Split Node 'n' through merge point if there is enough win. 45 Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) { 46 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) { 47 // ConvI2L may have type information on it which is unsafe to push up 48 // so disable this for now 49 return NULL; 50 } 51 52 // Splitting range check CastIIs through a loop induction Phi can 53 // cause new Phis to be created that are left unrelated to the loop 54 // induction Phi and prevent optimizations (vectorization) 55 if (n->Opcode() == Op_CastII && n->as_CastII()->has_range_check() && 56 region->is_CountedLoop() && n->in(1) == region->as_CountedLoop()->phi()) { 57 return NULL; 58 } 59 60 int wins = 0; 61 assert(!n->is_CFG(), ""); 62 assert(region->is_Region(), ""); 63 64 const Type* type = n->bottom_type(); 65 const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr(); 66 Node *phi; 67 if (t_oop != NULL && t_oop->is_known_instance_field()) { 68 int iid = t_oop->instance_id(); 69 int index = C->get_alias_index(t_oop); 70 int offset = t_oop->offset(); 71 phi = new PhiNode(region, type, NULL, iid, index, offset); 72 } else { 73 phi = PhiNode::make_blank(region, n); 74 } 75 uint old_unique = C->unique(); 76 for (uint i = 1; i < region->req(); i++) { 77 Node *x; 78 Node* the_clone = NULL; 79 if (region->in(i) == C->top()) { 80 x = C->top(); // Dead path? Use a dead data op 81 } else { 82 x = n->clone(); // Else clone up the data op 83 the_clone = x; // Remember for possible deletion. 84 // Alter data node to use pre-phi inputs 85 if (n->in(0) == region) 86 x->set_req( 0, region->in(i) ); 87 for (uint j = 1; j < n->req(); j++) { 88 Node *in = n->in(j); 89 if (in->is_Phi() && in->in(0) == region) 90 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone 91 } 92 } 93 // Check for a 'win' on some paths 94 const Type *t = x->Value(&_igvn); 95 96 bool singleton = t->singleton(); 97 98 // A TOP singleton indicates that there are no possible values incoming 99 // along a particular edge. In most cases, this is OK, and the Phi will 100 // be eliminated later in an Ideal call. However, we can't allow this to 101 // happen if the singleton occurs on loop entry, as the elimination of 102 // the PhiNode may cause the resulting node to migrate back to a previous 103 // loop iteration. 104 if (singleton && t == Type::TOP) { 105 // Is_Loop() == false does not confirm the absence of a loop (e.g., an 106 // irreducible loop may not be indicated by an affirmative is_Loop()); 107 // therefore, the only top we can split thru a phi is on a backedge of 108 // a loop. 109 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 110 } 111 112 if (singleton) { 113 wins++; 114 x = ((PhaseGVN&)_igvn).makecon(t); 115 } else { 116 // We now call Identity to try to simplify the cloned node. 117 // Note that some Identity methods call phase->type(this). 118 // Make sure that the type array is big enough for 119 // our new node, even though we may throw the node away. 120 // (Note: This tweaking with igvn only works because x is a new node.) 121 _igvn.set_type(x, t); 122 // If x is a TypeNode, capture any more-precise type permanently into Node 123 // otherwise it will be not updated during igvn->transform since 124 // igvn->type(x) is set to x->Value() already. 125 x->raise_bottom_type(t); 126 Node *y = x->Identity(&_igvn); 127 if (y != x) { 128 wins++; 129 x = y; 130 } else { 131 y = _igvn.hash_find(x); 132 if (y) { 133 wins++; 134 x = y; 135 } else { 136 // Else x is a new node we are keeping 137 // We do not need register_new_node_with_optimizer 138 // because set_type has already been called. 139 _igvn._worklist.push(x); 140 } 141 } 142 } 143 if (x != the_clone && the_clone != NULL) 144 _igvn.remove_dead_node(the_clone); 145 phi->set_req( i, x ); 146 } 147 // Too few wins? 148 if (wins <= policy) { 149 _igvn.remove_dead_node(phi); 150 return NULL; 151 } 152 153 // Record Phi 154 register_new_node( phi, region ); 155 156 for (uint i2 = 1; i2 < phi->req(); i2++) { 157 Node *x = phi->in(i2); 158 // If we commoned up the cloned 'x' with another existing Node, 159 // the existing Node picks up a new use. We need to make the 160 // existing Node occur higher up so it dominates its uses. 161 Node *old_ctrl; 162 IdealLoopTree *old_loop; 163 164 if (x->is_Con()) { 165 // Constant's control is always root. 166 set_ctrl(x, C->root()); 167 continue; 168 } 169 // The occasional new node 170 if (x->_idx >= old_unique) { // Found a new, unplaced node? 171 old_ctrl = NULL; 172 old_loop = NULL; // Not in any prior loop 173 } else { 174 old_ctrl = get_ctrl(x); 175 old_loop = get_loop(old_ctrl); // Get prior loop 176 } 177 // New late point must dominate new use 178 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2)); 179 if (new_ctrl == old_ctrl) // Nothing is changed 180 continue; 181 182 IdealLoopTree *new_loop = get_loop(new_ctrl); 183 184 // Don't move x into a loop if its uses are 185 // outside of loop. Otherwise x will be cloned 186 // for each use outside of this loop. 187 IdealLoopTree *use_loop = get_loop(region); 188 if (!new_loop->is_member(use_loop) && 189 (old_loop == NULL || !new_loop->is_member(old_loop))) { 190 // Take early control, later control will be recalculated 191 // during next iteration of loop optimizations. 192 new_ctrl = get_early_ctrl(x); 193 new_loop = get_loop(new_ctrl); 194 } 195 // Set new location 196 set_ctrl(x, new_ctrl); 197 // If changing loop bodies, see if we need to collect into new body 198 if (old_loop != new_loop) { 199 if (old_loop && !old_loop->_child) 200 old_loop->_body.yank(x); 201 if (!new_loop->_child) 202 new_loop->_body.push(x); // Collect body info 203 } 204 } 205 206 return phi; 207 } 208 209 //------------------------------dominated_by------------------------------------ 210 // Replace the dominated test with an obvious true or false. Place it on the 211 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the 212 // live path up to the dominating control. 213 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff, bool flip, bool exclude_loop_predicate ) { 214 if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); } 215 216 // prevdom is the dominating projection of the dominating test. 217 assert( iff->is_If(), "" ); 218 assert(iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd || iff->Opcode() == Op_RangeCheck, "Check this code when new subtype is added"); 219 int pop = prevdom->Opcode(); 220 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" ); 221 if (flip) { 222 if (pop == Op_IfTrue) 223 pop = Op_IfFalse; 224 else 225 pop = Op_IfTrue; 226 } 227 // 'con' is set to true or false to kill the dominated test. 228 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO); 229 set_ctrl(con, C->root()); // Constant gets a new use 230 // Hack the dominated test 231 _igvn.replace_input_of(iff, 1, con); 232 233 // If I dont have a reachable TRUE and FALSE path following the IfNode then 234 // I can assume this path reaches an infinite loop. In this case it's not 235 // important to optimize the data Nodes - either the whole compilation will 236 // be tossed or this path (and all data Nodes) will go dead. 237 if (iff->outcnt() != 2) return; 238 239 // Make control-dependent data Nodes on the live path (path that will remain 240 // once the dominated IF is removed) become control-dependent on the 241 // dominating projection. 242 Node* dp = iff->as_If()->proj_out_or_null(pop == Op_IfTrue); 243 244 // Loop predicates may have depending checks which should not 245 // be skipped. For example, range check predicate has two checks 246 // for lower and upper bounds. 247 if (dp == NULL) 248 return; 249 250 ProjNode* dp_proj = dp->as_Proj(); 251 ProjNode* unc_proj = iff->as_If()->proj_out(1 - dp_proj->_con)->as_Proj(); 252 if (exclude_loop_predicate && 253 (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) != NULL || 254 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check) != NULL)) { 255 // If this is a range check (IfNode::is_range_check), do not 256 // reorder because Compile::allow_range_check_smearing might have 257 // changed the check. 258 return; // Let IGVN transformation change control dependence. 259 } 260 261 IdealLoopTree *old_loop = get_loop(dp); 262 263 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { 264 Node* cd = dp->fast_out(i); // Control-dependent node 265 if (cd->depends_only_on_test()) { 266 assert(cd->in(0) == dp, ""); 267 _igvn.replace_input_of(cd, 0, prevdom); 268 set_early_ctrl(cd); 269 IdealLoopTree *new_loop = get_loop(get_ctrl(cd)); 270 if (old_loop != new_loop) { 271 if (!old_loop->_child) old_loop->_body.yank(cd); 272 if (!new_loop->_child) new_loop->_body.push(cd); 273 } 274 --i; 275 --imax; 276 } 277 } 278 } 279 280 //------------------------------has_local_phi_input---------------------------- 281 // Return TRUE if 'n' has Phi inputs from its local block and no other 282 // block-local inputs (all non-local-phi inputs come from earlier blocks) 283 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) { 284 Node *n_ctrl = get_ctrl(n); 285 // See if some inputs come from a Phi in this block, or from before 286 // this block. 287 uint i; 288 for( i = 1; i < n->req(); i++ ) { 289 Node *phi = n->in(i); 290 if( phi->is_Phi() && phi->in(0) == n_ctrl ) 291 break; 292 } 293 if( i >= n->req() ) 294 return NULL; // No Phi inputs; nowhere to clone thru 295 296 // Check for inputs created between 'n' and the Phi input. These 297 // must split as well; they have already been given the chance 298 // (courtesy of a post-order visit) and since they did not we must 299 // recover the 'cost' of splitting them by being very profitable 300 // when splitting 'n'. Since this is unlikely we simply give up. 301 for( i = 1; i < n->req(); i++ ) { 302 Node *m = n->in(i); 303 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) { 304 // We allow the special case of AddP's with no local inputs. 305 // This allows us to split-up address expressions. 306 if (m->is_AddP() && 307 get_ctrl(m->in(2)) != n_ctrl && 308 get_ctrl(m->in(3)) != n_ctrl) { 309 // Move the AddP up to dominating point 310 Node* c = find_non_split_ctrl(idom(n_ctrl)); 311 if (c->is_OuterStripMinedLoop()) { 312 c->as_Loop()->verify_strip_mined(1); 313 c = c->in(LoopNode::EntryControl); 314 } 315 set_ctrl_and_loop(m, c); 316 continue; 317 } 318 return NULL; 319 } 320 assert(m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control"); 321 } 322 323 return n_ctrl; 324 } 325 326 //------------------------------remix_address_expressions---------------------- 327 // Rework addressing expressions to get the most loop-invariant stuff 328 // moved out. We'd like to do all associative operators, but it's especially 329 // important (common) to do address expressions. 330 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) { 331 if (!has_ctrl(n)) return NULL; 332 Node *n_ctrl = get_ctrl(n); 333 IdealLoopTree *n_loop = get_loop(n_ctrl); 334 335 // See if 'n' mixes loop-varying and loop-invariant inputs and 336 // itself is loop-varying. 337 338 // Only interested in binary ops (and AddP) 339 if( n->req() < 3 || n->req() > 4 ) return NULL; 340 341 Node *n1_ctrl = get_ctrl(n->in( 1)); 342 Node *n2_ctrl = get_ctrl(n->in( 2)); 343 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3)); 344 IdealLoopTree *n1_loop = get_loop( n1_ctrl ); 345 IdealLoopTree *n2_loop = get_loop( n2_ctrl ); 346 IdealLoopTree *n3_loop = get_loop( n3_ctrl ); 347 348 // Does one of my inputs spin in a tighter loop than self? 349 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) || 350 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) || 351 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) ) 352 return NULL; // Leave well enough alone 353 354 // Is at least one of my inputs loop-invariant? 355 if( n1_loop == n_loop && 356 n2_loop == n_loop && 357 n3_loop == n_loop ) 358 return NULL; // No loop-invariant inputs 359 360 361 int n_op = n->Opcode(); 362 363 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2). 364 if( n_op == Op_LShiftI ) { 365 // Scale is loop invariant 366 Node *scale = n->in(2); 367 Node *scale_ctrl = get_ctrl(scale); 368 IdealLoopTree *scale_loop = get_loop(scale_ctrl ); 369 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) ) 370 return NULL; 371 const TypeInt *scale_t = scale->bottom_type()->isa_int(); 372 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 ) 373 return NULL; // Dont bother with byte/short masking 374 // Add must vary with loop (else shift would be loop-invariant) 375 Node *add = n->in(1); 376 Node *add_ctrl = get_ctrl(add); 377 IdealLoopTree *add_loop = get_loop(add_ctrl); 378 //assert( n_loop == add_loop, "" ); 379 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops 380 381 // Convert I-V into I+ (0-V); same for V-I 382 if( add->Opcode() == Op_SubI && 383 _igvn.type( add->in(1) ) != TypeInt::ZERO ) { 384 Node *zero = _igvn.intcon(0); 385 set_ctrl(zero, C->root()); 386 Node *neg = new SubINode( _igvn.intcon(0), add->in(2) ); 387 register_new_node( neg, get_ctrl(add->in(2) ) ); 388 add = new AddINode( add->in(1), neg ); 389 register_new_node( add, add_ctrl ); 390 } 391 if( add->Opcode() != Op_AddI ) return NULL; 392 // See if one add input is loop invariant 393 Node *add_var = add->in(1); 394 Node *add_var_ctrl = get_ctrl(add_var); 395 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 396 Node *add_invar = add->in(2); 397 Node *add_invar_ctrl = get_ctrl(add_invar); 398 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl ); 399 if( add_var_loop == n_loop ) { 400 } else if( add_invar_loop == n_loop ) { 401 // Swap to find the invariant part 402 add_invar = add_var; 403 add_invar_ctrl = add_var_ctrl; 404 add_invar_loop = add_var_loop; 405 add_var = add->in(2); 406 Node *add_var_ctrl = get_ctrl(add_var); 407 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 408 } else // Else neither input is loop invariant 409 return NULL; 410 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) ) 411 return NULL; // No invariant part of the add? 412 413 // Yes! Reshape address expression! 414 Node *inv_scale = new LShiftINode( add_invar, scale ); 415 Node *inv_scale_ctrl = 416 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ? 417 add_invar_ctrl : scale_ctrl; 418 register_new_node( inv_scale, inv_scale_ctrl ); 419 Node *var_scale = new LShiftINode( add_var, scale ); 420 register_new_node( var_scale, n_ctrl ); 421 Node *var_add = new AddINode( var_scale, inv_scale ); 422 register_new_node( var_add, n_ctrl ); 423 _igvn.replace_node( n, var_add ); 424 return var_add; 425 } 426 427 // Replace (I+V) with (V+I) 428 if( n_op == Op_AddI || 429 n_op == Op_AddL || 430 n_op == Op_AddF || 431 n_op == Op_AddD || 432 n_op == Op_MulI || 433 n_op == Op_MulL || 434 n_op == Op_MulF || 435 n_op == Op_MulD ) { 436 if( n2_loop == n_loop ) { 437 assert( n1_loop != n_loop, "" ); 438 n->swap_edges(1, 2); 439 } 440 } 441 442 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V), 443 // but not if I2 is a constant. 444 if( n_op == Op_AddP ) { 445 if( n2_loop == n_loop && n3_loop != n_loop ) { 446 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) { 447 Node *n22_ctrl = get_ctrl(n->in(2)->in(2)); 448 Node *n23_ctrl = get_ctrl(n->in(2)->in(3)); 449 IdealLoopTree *n22loop = get_loop( n22_ctrl ); 450 IdealLoopTree *n23_loop = get_loop( n23_ctrl ); 451 if( n22loop != n_loop && n22loop->is_member(n_loop) && 452 n23_loop == n_loop ) { 453 Node *add1 = new AddPNode( n->in(1), n->in(2)->in(2), n->in(3) ); 454 // Stuff new AddP in the loop preheader 455 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) ); 456 Node *add2 = new AddPNode( n->in(1), add1, n->in(2)->in(3) ); 457 register_new_node( add2, n_ctrl ); 458 _igvn.replace_node( n, add2 ); 459 return add2; 460 } 461 } 462 } 463 464 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V) 465 if (n2_loop != n_loop && n3_loop == n_loop) { 466 if (n->in(3)->Opcode() == Op_AddX) { 467 Node *V = n->in(3)->in(1); 468 Node *I = n->in(3)->in(2); 469 if (is_member(n_loop,get_ctrl(V))) { 470 } else { 471 Node *tmp = V; V = I; I = tmp; 472 } 473 if (!is_member(n_loop,get_ctrl(I))) { 474 Node *add1 = new AddPNode(n->in(1), n->in(2), I); 475 // Stuff new AddP in the loop preheader 476 register_new_node(add1, n_loop->_head->in(LoopNode::EntryControl)); 477 Node *add2 = new AddPNode(n->in(1), add1, V); 478 register_new_node(add2, n_ctrl); 479 _igvn.replace_node(n, add2); 480 return add2; 481 } 482 } 483 } 484 } 485 486 return NULL; 487 } 488 489 //------------------------------conditional_move------------------------------- 490 // Attempt to replace a Phi with a conditional move. We have some pretty 491 // strict profitability requirements. All Phis at the merge point must 492 // be converted, so we can remove the control flow. We need to limit the 493 // number of c-moves to a small handful. All code that was in the side-arms 494 // of the CFG diamond is now speculatively executed. This code has to be 495 // "cheap enough". We are pretty much limited to CFG diamonds that merge 496 // 1 or 2 items with a total of 1 or 2 ops executed speculatively. 497 Node *PhaseIdealLoop::conditional_move( Node *region ) { 498 499 assert(region->is_Region(), "sanity check"); 500 if (region->req() != 3) return NULL; 501 502 // Check for CFG diamond 503 Node *lp = region->in(1); 504 Node *rp = region->in(2); 505 if (!lp || !rp) return NULL; 506 Node *lp_c = lp->in(0); 507 if (lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If()) return NULL; 508 IfNode *iff = lp_c->as_If(); 509 510 // Check for ops pinned in an arm of the diamond. 511 // Can't remove the control flow in this case 512 if (lp->outcnt() > 1) return NULL; 513 if (rp->outcnt() > 1) return NULL; 514 515 IdealLoopTree* r_loop = get_loop(region); 516 assert(r_loop == get_loop(iff), "sanity"); 517 // Always convert to CMOVE if all results are used only outside this loop. 518 bool used_inside_loop = (r_loop == _ltree_root); 519 520 // Check profitability 521 int cost = 0; 522 int phis = 0; 523 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 524 Node *out = region->fast_out(i); 525 if (!out->is_Phi()) continue; // Ignore other control edges, etc 526 phis++; 527 PhiNode* phi = out->as_Phi(); 528 BasicType bt = phi->type()->basic_type(); 529 switch (bt) { 530 case T_DOUBLE: 531 case T_FLOAT: 532 if (C->use_cmove()) { 533 continue; //TODO: maybe we want to add some cost 534 } 535 cost += Matcher::float_cmove_cost(); // Could be very expensive 536 break; 537 case T_LONG: { 538 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's 539 } 540 case T_INT: // These all CMOV fine 541 case T_ADDRESS: { // (RawPtr) 542 cost++; 543 break; 544 } 545 case T_NARROWOOP: // Fall through 546 case T_OBJECT: { // Base oops are OK, but not derived oops 547 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr(); 548 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a 549 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus 550 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we 551 // have a Phi for the base here that we convert to a CMOVE all is well 552 // and good. But if the base is dead, we'll not make a CMOVE. Later 553 // the allocator will have to produce a base by creating a CMOVE of the 554 // relevant bases. This puts the allocator in the business of 555 // manufacturing expensive instructions, generally a bad plan. 556 // Just Say No to Conditionally-Moved Derived Pointers. 557 if (tp && tp->offset() != 0) 558 return NULL; 559 cost++; 560 break; 561 } 562 default: 563 return NULL; // In particular, can't do memory or I/O 564 } 565 // Add in cost any speculative ops 566 for (uint j = 1; j < region->req(); j++) { 567 Node *proj = region->in(j); 568 Node *inp = phi->in(j); 569 if (get_ctrl(inp) == proj) { // Found local op 570 cost++; 571 // Check for a chain of dependent ops; these will all become 572 // speculative in a CMOV. 573 for (uint k = 1; k < inp->req(); k++) 574 if (get_ctrl(inp->in(k)) == proj) 575 cost += ConditionalMoveLimit; // Too much speculative goo 576 } 577 } 578 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode. 579 // This will likely Split-If, a higher-payoff operation. 580 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) { 581 Node* use = phi->fast_out(k); 582 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr()) 583 cost += ConditionalMoveLimit; 584 // Is there a use inside the loop? 585 // Note: check only basic types since CMoveP is pinned. 586 if (!used_inside_loop && is_java_primitive(bt)) { 587 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use); 588 if (r_loop == u_loop || r_loop->is_member(u_loop)) { 589 used_inside_loop = true; 590 } 591 } 592 } 593 }//for 594 Node* bol = iff->in(1); 595 assert(bol->Opcode() == Op_Bool, ""); 596 int cmp_op = bol->in(1)->Opcode(); 597 // It is expensive to generate flags from a float compare. 598 // Avoid duplicated float compare. 599 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL; 600 601 float infrequent_prob = PROB_UNLIKELY_MAG(3); 602 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop. 603 if (used_inside_loop) { 604 if (cost >= ConditionalMoveLimit) return NULL; // Too much goo 605 606 // BlockLayoutByFrequency optimization moves infrequent branch 607 // from hot path. No point in CMOV'ing in such case (110 is used 608 // instead of 100 to take into account not exactness of float value). 609 if (BlockLayoutByFrequency) { 610 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f); 611 } 612 } 613 // Check for highly predictable branch. No point in CMOV'ing if 614 // we are going to predict accurately all the time. 615 if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) { 616 //keep going 617 } else if (iff->_prob < infrequent_prob || 618 iff->_prob > (1.0f - infrequent_prob)) 619 return NULL; 620 621 // -------------- 622 // Now replace all Phis with CMOV's 623 Node *cmov_ctrl = iff->in(0); 624 uint flip = (lp->Opcode() == Op_IfTrue); 625 Node_List wq; 626 while (1) { 627 PhiNode* phi = NULL; 628 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 629 Node *out = region->fast_out(i); 630 if (out->is_Phi()) { 631 phi = out->as_Phi(); 632 break; 633 } 634 } 635 if (phi == NULL) break; 636 if (PrintOpto && VerifyLoopOptimizations) { tty->print_cr("CMOV"); } 637 // Move speculative ops 638 wq.push(phi); 639 while (wq.size() > 0) { 640 Node *n = wq.pop(); 641 for (uint j = 1; j < n->req(); j++) { 642 Node* m = n->in(j); 643 if (m != NULL && !is_dominator(get_ctrl(m), cmov_ctrl)) { 644 #ifndef PRODUCT 645 if (PrintOpto && VerifyLoopOptimizations) { 646 tty->print(" speculate: "); 647 m->dump(); 648 } 649 #endif 650 set_ctrl(m, cmov_ctrl); 651 wq.push(m); 652 } 653 } 654 } 655 Node *cmov = CMoveNode::make(cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi)); 656 register_new_node( cmov, cmov_ctrl ); 657 _igvn.replace_node( phi, cmov ); 658 #ifndef PRODUCT 659 if (TraceLoopOpts) { 660 tty->print("CMOV "); 661 r_loop->dump_head(); 662 if (Verbose) { 663 bol->in(1)->dump(1); 664 cmov->dump(1); 665 } 666 } 667 if (VerifyLoopOptimizations) verify(); 668 #endif 669 } 670 671 // The useless CFG diamond will fold up later; see the optimization in 672 // RegionNode::Ideal. 673 _igvn._worklist.push(region); 674 675 return iff->in(1); 676 } 677 678 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) { 679 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 680 Node* u = m->fast_out(i); 681 if (u->is_CFG()) { 682 if (u->Opcode() == Op_NeverBranch) { 683 u = ((NeverBranchNode*)u)->proj_out(0); 684 enqueue_cfg_uses(u, wq); 685 } else { 686 wq.push(u); 687 } 688 } 689 } 690 } 691 692 // Try moving a store out of a loop, right before the loop 693 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) { 694 // Store has to be first in the loop body 695 IdealLoopTree *n_loop = get_loop(n_ctrl); 696 if (n->is_Store() && n_loop != _ltree_root && 697 n_loop->is_loop() && n_loop->_head->is_Loop() && 698 n->in(0) != NULL) { 699 Node* address = n->in(MemNode::Address); 700 Node* value = n->in(MemNode::ValueIn); 701 Node* mem = n->in(MemNode::Memory); 702 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 703 IdealLoopTree* value_loop = get_loop(get_ctrl(value)); 704 705 // - address and value must be loop invariant 706 // - memory must be a memory Phi for the loop 707 // - Store must be the only store on this memory slice in the 708 // loop: if there's another store following this one then value 709 // written at iteration i by the second store could be overwritten 710 // at iteration i+n by the first store: it's not safe to move the 711 // first store out of the loop 712 // - nothing must observe the memory Phi: it guarantees no read 713 // before the store, we are also guaranteed the store post 714 // dominates the loop head (ignoring a possible early 715 // exit). Otherwise there would be extra Phi involved between the 716 // loop's Phi and the store. 717 // - there must be no early exit from the loop before the Store 718 // (such an exit most of the time would be an extra use of the 719 // memory Phi but sometimes is a bottom memory Phi that takes the 720 // store as input). 721 722 if (!n_loop->is_member(address_loop) && 723 !n_loop->is_member(value_loop) && 724 mem->is_Phi() && mem->in(0) == n_loop->_head && 725 mem->outcnt() == 1 && 726 mem->in(LoopNode::LoopBackControl) == n) { 727 728 assert(n_loop->_tail != NULL, "need a tail"); 729 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop"); 730 731 // Verify that there's no early exit of the loop before the store. 732 bool ctrl_ok = false; 733 { 734 // Follow control from loop head until n, we exit the loop or 735 // we reach the tail 736 ResourceMark rm; 737 Unique_Node_List wq; 738 wq.push(n_loop->_head); 739 740 for (uint next = 0; next < wq.size(); ++next) { 741 Node *m = wq.at(next); 742 if (m == n->in(0)) { 743 ctrl_ok = true; 744 continue; 745 } 746 assert(!has_ctrl(m), "should be CFG"); 747 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) { 748 ctrl_ok = false; 749 break; 750 } 751 enqueue_cfg_uses(m, wq); 752 if (wq.size() > 10) { 753 ctrl_ok = false; 754 break; 755 } 756 } 757 } 758 if (ctrl_ok) { 759 // move the Store 760 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem); 761 _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl)); 762 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl)); 763 // Disconnect the phi now. An empty phi can confuse other 764 // optimizations in this pass of loop opts. 765 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl)); 766 n_loop->_body.yank(mem); 767 768 set_ctrl_and_loop(n, n->in(0)); 769 770 return n; 771 } 772 } 773 } 774 return NULL; 775 } 776 777 // Try moving a store out of a loop, right after the loop 778 void PhaseIdealLoop::try_move_store_after_loop(Node* n) { 779 if (n->is_Store() && n->in(0) != NULL) { 780 Node *n_ctrl = get_ctrl(n); 781 IdealLoopTree *n_loop = get_loop(n_ctrl); 782 // Store must be in a loop 783 if (n_loop != _ltree_root && !n_loop->_irreducible) { 784 Node* address = n->in(MemNode::Address); 785 Node* value = n->in(MemNode::ValueIn); 786 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 787 // address must be loop invariant 788 if (!n_loop->is_member(address_loop)) { 789 // Store must be last on this memory slice in the loop and 790 // nothing in the loop must observe it 791 Node* phi = NULL; 792 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 793 Node* u = n->fast_out(i); 794 if (has_ctrl(u)) { // control use? 795 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 796 if (!n_loop->is_member(u_loop)) { 797 continue; 798 } 799 if (u->is_Phi() && u->in(0) == n_loop->_head) { 800 assert(_igvn.type(u) == Type::MEMORY, "bad phi"); 801 // multiple phis on the same slice are possible 802 if (phi != NULL) { 803 return; 804 } 805 phi = u; 806 continue; 807 } 808 } 809 return; 810 } 811 if (phi != NULL) { 812 // Nothing in the loop before the store (next iteration) 813 // must observe the stored value 814 bool mem_ok = true; 815 { 816 ResourceMark rm; 817 Unique_Node_List wq; 818 wq.push(phi); 819 for (uint next = 0; next < wq.size() && mem_ok; ++next) { 820 Node *m = wq.at(next); 821 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) { 822 Node* u = m->fast_out(i); 823 if (u->is_Store() || u->is_Phi()) { 824 if (u != n) { 825 wq.push(u); 826 mem_ok = (wq.size() <= 10); 827 } 828 } else { 829 mem_ok = false; 830 break; 831 } 832 } 833 } 834 } 835 if (mem_ok) { 836 // Move the store out of the loop if the LCA of all 837 // users (except for the phi) is outside the loop. 838 Node* hook = new Node(1); 839 _igvn.rehash_node_delayed(phi); 840 int count = phi->replace_edge(n, hook); 841 assert(count > 0, "inconsistent phi"); 842 843 // Compute latest point this store can go 844 Node* lca = get_late_ctrl(n, get_ctrl(n)); 845 if (n_loop->is_member(get_loop(lca))) { 846 // LCA is in the loop - bail out 847 _igvn.replace_node(hook, n); 848 return; 849 } 850 #ifdef ASSERT 851 if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) { 852 assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined"); 853 n_loop->_head->as_Loop()->verify_strip_mined(1); 854 Node* outer = n_loop->_head->as_CountedLoop()->outer_loop(); 855 IdealLoopTree* outer_loop = get_loop(outer); 856 assert(n_loop->_parent == outer_loop, "broken loop tree"); 857 assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state"); 858 } 859 #endif 860 861 // Move store out of the loop 862 _igvn.replace_node(hook, n->in(MemNode::Memory)); 863 _igvn.replace_input_of(n, 0, lca); 864 set_ctrl_and_loop(n, lca); 865 866 // Disconnect the phi now. An empty phi can confuse other 867 // optimizations in this pass of loop opts.. 868 if (phi->in(LoopNode::LoopBackControl) == phi) { 869 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl)); 870 n_loop->_body.yank(phi); 871 } 872 } 873 } 874 } 875 } 876 } 877 } 878 879 //------------------------------split_if_with_blocks_pre----------------------- 880 // Do the real work in a non-recursive function. Data nodes want to be 881 // cloned in the pre-order so they can feed each other nicely. 882 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) { 883 // Cloning these guys is unlikely to win 884 int n_op = n->Opcode(); 885 if( n_op == Op_MergeMem ) return n; 886 if( n->is_Proj() ) return n; 887 // Do not clone-up CmpFXXX variations, as these are always 888 // followed by a CmpI 889 if( n->is_Cmp() ) return n; 890 // Attempt to use a conditional move instead of a phi/branch 891 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) { 892 Node *cmov = conditional_move( n ); 893 if( cmov ) return cmov; 894 } 895 if( n->is_CFG() || n->is_LoadStore() ) 896 return n; 897 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd 898 n_op == Op_Opaque2 ) { 899 if( !C->major_progress() ) // If chance of no more loop opts... 900 _igvn._worklist.push(n); // maybe we'll remove them 901 return n; 902 } 903 904 if( n->is_Con() ) return n; // No cloning for Con nodes 905 906 Node *n_ctrl = get_ctrl(n); 907 if( !n_ctrl ) return n; // Dead node 908 909 Node* res = try_move_store_before_loop(n, n_ctrl); 910 if (res != NULL) { 911 return n; 912 } 913 914 // Attempt to remix address expressions for loop invariants 915 Node *m = remix_address_expressions( n ); 916 if( m ) return m; 917 918 if (n->is_ConstraintCast()) { 919 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this); 920 // ConstraintCastNode::dominating_cast() uses node control input to determine domination. 921 // Node control inputs don't necessarily agree with loop control info (due to 922 // transformations happened in between), thus additional dominance check is needed 923 // to keep loop info valid. 924 if (dom_cast != NULL && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) { 925 _igvn.replace_node(n, dom_cast); 926 return dom_cast; 927 } 928 } 929 930 // Determine if the Node has inputs from some local Phi. 931 // Returns the block to clone thru. 932 Node *n_blk = has_local_phi_input( n ); 933 if( !n_blk ) return n; 934 935 // Do not clone the trip counter through on a CountedLoop 936 // (messes up the canonical shape). 937 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n; 938 939 // Check for having no control input; not pinned. Allow 940 // dominating control. 941 if (n->in(0)) { 942 Node *dom = idom(n_blk); 943 if (dom_lca(n->in(0), dom) != n->in(0)) { 944 return n; 945 } 946 } 947 // Policy: when is it profitable. You must get more wins than 948 // policy before it is considered profitable. Policy is usually 0, 949 // so 1 win is considered profitable. Big merges will require big 950 // cloning, so get a larger policy. 951 int policy = n_blk->req() >> 2; 952 953 // If the loop is a candidate for range check elimination, 954 // delay splitting through it's phi until a later loop optimization 955 if (n_blk->is_CountedLoop()) { 956 IdealLoopTree *lp = get_loop(n_blk); 957 if (lp && lp->_rce_candidate) { 958 return n; 959 } 960 } 961 962 // Use same limit as split_if_with_blocks_post 963 if( C->live_nodes() > 35000 ) return n; // Method too big 964 965 // Split 'n' through the merge point if it is profitable 966 Node *phi = split_thru_phi( n, n_blk, policy ); 967 if (!phi) return n; 968 969 // Found a Phi to split thru! 970 // Replace 'n' with the new phi 971 _igvn.replace_node( n, phi ); 972 // Moved a load around the loop, 'en-registering' something. 973 if (n_blk->is_Loop() && n->is_Load() && 974 !phi->in(LoopNode::LoopBackControl)->is_Load()) 975 C->set_major_progress(); 976 977 return phi; 978 } 979 980 static bool merge_point_too_heavy(Compile* C, Node* region) { 981 // Bail out if the region and its phis have too many users. 982 int weight = 0; 983 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 984 weight += region->fast_out(i)->outcnt(); 985 } 986 int nodes_left = C->max_node_limit() - C->live_nodes(); 987 if (weight * 8 > nodes_left) { 988 if (PrintOpto) { 989 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight); 990 } 991 return true; 992 } else { 993 return false; 994 } 995 } 996 997 static bool merge_point_safe(Node* region) { 998 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode 999 // having a PhiNode input. This sidesteps the dangerous case where the split 1000 // ConvI2LNode may become TOP if the input Value() does not 1001 // overlap the ConvI2L range, leaving a node which may not dominate its 1002 // uses. 1003 // A better fix for this problem can be found in the BugTraq entry, but 1004 // expediency for Mantis demands this hack. 1005 // 6855164: If the merge point has a FastLockNode with a PhiNode input, we stop 1006 // split_if_with_blocks from splitting a block because we could not move around 1007 // the FastLockNode. 1008 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1009 Node* n = region->fast_out(i); 1010 if (n->is_Phi()) { 1011 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1012 Node* m = n->fast_out(j); 1013 if (m->is_FastLock()) 1014 return false; 1015 #ifdef _LP64 1016 if (m->Opcode() == Op_ConvI2L) 1017 return false; 1018 if (m->is_CastII() && m->isa_CastII()->has_range_check()) { 1019 return false; 1020 } 1021 #endif 1022 } 1023 } 1024 } 1025 return true; 1026 } 1027 1028 1029 //------------------------------place_near_use--------------------------------- 1030 // Place some computation next to use but not inside inner loops. 1031 // For inner loop uses move it to the preheader area. 1032 Node *PhaseIdealLoop::place_near_use(Node *useblock) const { 1033 IdealLoopTree *u_loop = get_loop( useblock ); 1034 if (u_loop->_irreducible) { 1035 return useblock; 1036 } 1037 if (u_loop->_child) { 1038 if (useblock == u_loop->_head && u_loop->_head->is_OuterStripMinedLoop()) { 1039 return u_loop->_head->in(LoopNode::EntryControl); 1040 } 1041 return useblock; 1042 } 1043 return u_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl); 1044 } 1045 1046 1047 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) { 1048 if (!n->is_If() || n->is_CountedLoopEnd()) { 1049 return false; 1050 } 1051 if (!n->in(0)->is_Region()) { 1052 return false; 1053 } 1054 Node* region = n->in(0); 1055 Node* dom = idom(region); 1056 if (!dom->is_If() || dom->in(1) != n->in(1)) { 1057 return false; 1058 } 1059 IfNode* dom_if = dom->as_If(); 1060 Node* proj_true = dom_if->proj_out(1); 1061 Node* proj_false = dom_if->proj_out(0); 1062 1063 for (uint i = 1; i < region->req(); i++) { 1064 if (is_dominator(proj_true, region->in(i))) { 1065 continue; 1066 } 1067 if (is_dominator(proj_false, region->in(i))) { 1068 continue; 1069 } 1070 return false; 1071 } 1072 1073 return true; 1074 } 1075 1076 bool PhaseIdealLoop::can_split_if(Node *n_ctrl) { 1077 if (C->live_nodes() > 35000) { 1078 return false; // Method too big 1079 } 1080 1081 // Do not do 'split-if' if irreducible loops are present. 1082 if (_has_irreducible_loops) { 1083 return false; 1084 } 1085 1086 if (merge_point_too_heavy(C, n_ctrl)) { 1087 return false; 1088 } 1089 1090 // Do not do 'split-if' if some paths are dead. First do dead code 1091 // elimination and then see if its still profitable. 1092 for (uint i = 1; i < n_ctrl->req(); i++) { 1093 if (n_ctrl->in(i) == C->top()) { 1094 return false; 1095 } 1096 } 1097 1098 // If trying to do a 'Split-If' at the loop head, it is only 1099 // profitable if the cmp folds up on BOTH paths. Otherwise we 1100 // risk peeling a loop forever. 1101 1102 // CNC - Disabled for now. Requires careful handling of loop 1103 // body selection for the cloned code. Also, make sure we check 1104 // for any input path not being in the same loop as n_ctrl. For 1105 // irreducible loops we cannot check for 'n_ctrl->is_Loop()' 1106 // because the alternative loop entry points won't be converted 1107 // into LoopNodes. 1108 IdealLoopTree *n_loop = get_loop(n_ctrl); 1109 for (uint j = 1; j < n_ctrl->req(); j++) { 1110 if (get_loop(n_ctrl->in(j)) != n_loop) { 1111 return false; 1112 } 1113 } 1114 1115 // Check for safety of the merge point. 1116 if (!merge_point_safe(n_ctrl)) { 1117 return false; 1118 } 1119 1120 return true; 1121 } 1122 1123 //------------------------------split_if_with_blocks_post---------------------- 1124 // Do the real work in a non-recursive function. CFG hackery wants to be 1125 // in the post-order, so it can dirty the I-DOM info and not use the dirtied 1126 // info. 1127 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) { 1128 1129 // Cloning Cmp through Phi's involves the split-if transform. 1130 // FastLock is not used by an If 1131 if (n->is_Cmp() && !n->is_FastLock()) { 1132 Node *n_ctrl = get_ctrl(n); 1133 // Determine if the Node has inputs from some local Phi. 1134 // Returns the block to clone thru. 1135 Node *n_blk = has_local_phi_input(n); 1136 if (n_blk != n_ctrl) { 1137 return; 1138 } 1139 1140 if (!can_split_if(n_ctrl)) { 1141 return; 1142 } 1143 1144 if (n->outcnt() != 1) { 1145 return; // Multiple bool's from 1 compare? 1146 } 1147 Node *bol = n->unique_out(); 1148 assert(bol->is_Bool(), "expect a bool here"); 1149 if (bol->outcnt() != 1) { 1150 return;// Multiple branches from 1 compare? 1151 } 1152 Node *iff = bol->unique_out(); 1153 1154 // Check some safety conditions 1155 if (iff->is_If()) { // Classic split-if? 1156 if (iff->in(0) != n_ctrl) { 1157 return; // Compare must be in same blk as if 1158 } 1159 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE 1160 // Can't split CMove with different control edge. 1161 if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) { 1162 return; 1163 } 1164 if (get_ctrl(iff->in(2)) == n_ctrl || 1165 get_ctrl(iff->in(3)) == n_ctrl) { 1166 return; // Inputs not yet split-up 1167 } 1168 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) { 1169 return; // Loop-invar test gates loop-varying CMOVE 1170 } 1171 } else { 1172 return; // some other kind of node, such as an Allocate 1173 } 1174 1175 // When is split-if profitable? Every 'win' on means some control flow 1176 // goes dead, so it's almost always a win. 1177 int policy = 0; 1178 // Split compare 'n' through the merge point if it is profitable 1179 Node *phi = split_thru_phi( n, n_ctrl, policy); 1180 if (!phi) { 1181 return; 1182 } 1183 1184 // Found a Phi to split thru! 1185 // Replace 'n' with the new phi 1186 _igvn.replace_node(n, phi); 1187 1188 // Now split the bool up thru the phi 1189 Node *bolphi = split_thru_phi(bol, n_ctrl, -1); 1190 guarantee(bolphi != NULL, "null boolean phi node"); 1191 1192 _igvn.replace_node(bol, bolphi); 1193 assert(iff->in(1) == bolphi, ""); 1194 1195 if (bolphi->Value(&_igvn)->singleton()) { 1196 return; 1197 } 1198 1199 // Conditional-move? Must split up now 1200 if (!iff->is_If()) { 1201 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1); 1202 _igvn.replace_node(iff, cmovphi); 1203 return; 1204 } 1205 1206 // Now split the IF 1207 do_split_if(iff); 1208 return; 1209 } 1210 1211 // Two identical ifs back to back can be merged 1212 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) { 1213 Node *n_ctrl = n->in(0); 1214 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1)); 1215 IfNode* dom_if = idom(n_ctrl)->as_If(); 1216 Node* proj_true = dom_if->proj_out(1); 1217 Node* proj_false = dom_if->proj_out(0); 1218 Node* con_true = _igvn.makecon(TypeInt::ONE); 1219 Node* con_false = _igvn.makecon(TypeInt::ZERO); 1220 1221 for (uint i = 1; i < n_ctrl->req(); i++) { 1222 if (is_dominator(proj_true, n_ctrl->in(i))) { 1223 bolphi->init_req(i, con_true); 1224 } else { 1225 assert(is_dominator(proj_false, n_ctrl->in(i)), "bad if"); 1226 bolphi->init_req(i, con_false); 1227 } 1228 } 1229 register_new_node(bolphi, n_ctrl); 1230 _igvn.replace_input_of(n, 1, bolphi); 1231 1232 // Now split the IF 1233 do_split_if(n); 1234 return; 1235 } 1236 1237 // Check for an IF ready to split; one that has its 1238 // condition codes input coming from a Phi at the block start. 1239 int n_op = n->Opcode(); 1240 1241 // Check for an IF being dominated by another IF same test 1242 if (n_op == Op_If || 1243 n_op == Op_RangeCheck) { 1244 Node *bol = n->in(1); 1245 uint max = bol->outcnt(); 1246 // Check for same test used more than once? 1247 if (max > 1 && bol->is_Bool()) { 1248 // Search up IDOMs to see if this IF is dominated. 1249 Node *cutoff = get_ctrl(bol); 1250 1251 // Now search up IDOMs till cutoff, looking for a dominating test 1252 Node *prevdom = n; 1253 Node *dom = idom(prevdom); 1254 while (dom != cutoff) { 1255 if (dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom) { 1256 // Replace the dominated test with an obvious true or false. 1257 // Place it on the IGVN worklist for later cleanup. 1258 C->set_major_progress(); 1259 dominated_by(prevdom, n, false, true); 1260 #ifndef PRODUCT 1261 if( VerifyLoopOptimizations ) verify(); 1262 #endif 1263 return; 1264 } 1265 prevdom = dom; 1266 dom = idom(prevdom); 1267 } 1268 } 1269 } 1270 1271 // See if a shared loop-varying computation has no loop-varying uses. 1272 // Happens if something is only used for JVM state in uncommon trap exits, 1273 // like various versions of induction variable+offset. Clone the 1274 // computation per usage to allow it to sink out of the loop. 1275 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about) 1276 Node *n_ctrl = get_ctrl(n); 1277 IdealLoopTree *n_loop = get_loop(n_ctrl); 1278 if( n_loop != _ltree_root ) { 1279 DUIterator_Fast imax, i = n->fast_outs(imax); 1280 for (; i < imax; i++) { 1281 Node* u = n->fast_out(i); 1282 if( !has_ctrl(u) ) break; // Found control user 1283 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 1284 if( u_loop == n_loop ) break; // Found loop-varying use 1285 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop 1286 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003 1287 } 1288 bool did_break = (i < imax); // Did we break out of the previous loop? 1289 if (!did_break && n->outcnt() > 1) { // All uses in outer loops! 1290 Node *late_load_ctrl = NULL; 1291 if (n->is_Load()) { 1292 // If n is a load, get and save the result from get_late_ctrl(), 1293 // to be later used in calculating the control for n's clones. 1294 clear_dom_lca_tags(); 1295 late_load_ctrl = get_late_ctrl(n, n_ctrl); 1296 } 1297 // If n is a load, and the late control is the same as the current 1298 // control, then the cloning of n is a pointless exercise, because 1299 // GVN will ensure that we end up where we started. 1300 if (!n->is_Load() || late_load_ctrl != n_ctrl) { 1301 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) { 1302 Node *u = n->last_out(j); // Clone private computation per use 1303 _igvn.rehash_node_delayed(u); 1304 Node *x = n->clone(); // Clone computation 1305 Node *x_ctrl = NULL; 1306 if( u->is_Phi() ) { 1307 // Replace all uses of normal nodes. Replace Phi uses 1308 // individually, so the separate Nodes can sink down 1309 // different paths. 1310 uint k = 1; 1311 while( u->in(k) != n ) k++; 1312 u->set_req( k, x ); 1313 // x goes next to Phi input path 1314 x_ctrl = u->in(0)->in(k); 1315 --j; 1316 } else { // Normal use 1317 // Replace all uses 1318 for( uint k = 0; k < u->req(); k++ ) { 1319 if( u->in(k) == n ) { 1320 u->set_req( k, x ); 1321 --j; 1322 } 1323 } 1324 x_ctrl = get_ctrl(u); 1325 } 1326 1327 // Find control for 'x' next to use but not inside inner loops. 1328 // For inner loop uses get the preheader area. 1329 x_ctrl = place_near_use(x_ctrl); 1330 1331 if (n->is_Load()) { 1332 // For loads, add a control edge to a CFG node outside of the loop 1333 // to force them to not combine and return back inside the loop 1334 // during GVN optimization (4641526). 1335 // 1336 // Because we are setting the actual control input, factor in 1337 // the result from get_late_ctrl() so we respect any 1338 // anti-dependences. (6233005). 1339 x_ctrl = dom_lca(late_load_ctrl, x_ctrl); 1340 1341 // Don't allow the control input to be a CFG splitting node. 1342 // Such nodes should only have ProjNodes as outs, e.g. IfNode 1343 // should only have IfTrueNode and IfFalseNode (4985384). 1344 x_ctrl = find_non_split_ctrl(x_ctrl); 1345 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone"); 1346 1347 x->set_req(0, x_ctrl); 1348 } 1349 register_new_node(x, x_ctrl); 1350 1351 // Some institutional knowledge is needed here: 'x' is 1352 // yanked because if the optimizer runs GVN on it all the 1353 // cloned x's will common up and undo this optimization and 1354 // be forced back in the loop. This is annoying because it 1355 // makes +VerifyOpto report false-positives on progress. I 1356 // tried setting control edges on the x's to force them to 1357 // not combine, but the matching gets worried when it tries 1358 // to fold a StoreP and an AddP together (as part of an 1359 // address expression) and the AddP and StoreP have 1360 // different controls. 1361 if (!x->is_Load() && !x->is_DecodeNarrowPtr()) _igvn._worklist.yank(x); 1362 } 1363 _igvn.remove_dead_node(n); 1364 } 1365 } 1366 } 1367 } 1368 1369 try_move_store_after_loop(n); 1370 1371 // Check for Opaque2's who's loop has disappeared - who's input is in the 1372 // same loop nest as their output. Remove 'em, they are no longer useful. 1373 if( n_op == Op_Opaque2 && 1374 n->in(1) != NULL && 1375 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) { 1376 _igvn.replace_node( n, n->in(1) ); 1377 } 1378 } 1379 1380 //------------------------------split_if_with_blocks--------------------------- 1381 // Check for aggressive application of 'split-if' optimization, 1382 // using basic block level info. 1383 void PhaseIdealLoop::split_if_with_blocks( VectorSet &visited, Node_Stack &nstack ) { 1384 Node *n = C->root(); 1385 visited.set(n->_idx); // first, mark node as visited 1386 // Do pre-visit work for root 1387 n = split_if_with_blocks_pre( n ); 1388 uint cnt = n->outcnt(); 1389 uint i = 0; 1390 while (true) { 1391 // Visit all children 1392 if (i < cnt) { 1393 Node* use = n->raw_out(i); 1394 ++i; 1395 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) { 1396 // Now do pre-visit work for this use 1397 use = split_if_with_blocks_pre( use ); 1398 nstack.push(n, i); // Save parent and next use's index. 1399 n = use; // Process all children of current use. 1400 cnt = use->outcnt(); 1401 i = 0; 1402 } 1403 } 1404 else { 1405 // All of n's children have been processed, complete post-processing. 1406 if (cnt != 0 && !n->is_Con()) { 1407 assert(has_node(n), "no dead nodes"); 1408 split_if_with_blocks_post( n ); 1409 } 1410 if (nstack.is_empty()) { 1411 // Finished all nodes on stack. 1412 break; 1413 } 1414 // Get saved parent node and next use's index. Visit the rest of uses. 1415 n = nstack.node(); 1416 cnt = n->outcnt(); 1417 i = nstack.index(); 1418 nstack.pop(); 1419 } 1420 } 1421 } 1422 1423 1424 //============================================================================= 1425 // 1426 // C L O N E A L O O P B O D Y 1427 // 1428 1429 //------------------------------clone_iff-------------------------------------- 1430 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1431 // "Nearly" because all Nodes have been cloned from the original in the loop, 1432 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1433 // through the Phi recursively, and return a Bool. 1434 Node* PhaseIdealLoop::clone_iff(PhiNode *phi, IdealLoopTree *loop) { 1435 1436 // Convert this Phi into a Phi merging Bools 1437 uint i; 1438 for (i = 1; i < phi->req(); i++) { 1439 Node *b = phi->in(i); 1440 if (b->is_Phi()) { 1441 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi(), loop)); 1442 } else { 1443 assert(b->is_Bool() || b->Opcode() == Op_Opaque4, ""); 1444 } 1445 } 1446 1447 Node* n = phi->in(1); 1448 Node* sample_opaque = NULL; 1449 Node *sample_bool = NULL; 1450 if (n->Opcode() == Op_Opaque4) { 1451 sample_opaque = n; 1452 sample_bool = n->in(1); 1453 assert(sample_bool->is_Bool(), "wrong type"); 1454 } else { 1455 sample_bool = n; 1456 } 1457 Node *sample_cmp = sample_bool->in(1); 1458 1459 // Make Phis to merge the Cmp's inputs. 1460 PhiNode *phi1 = new PhiNode(phi->in(0), Type::TOP); 1461 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP); 1462 for (i = 1; i < phi->req(); i++) { 1463 Node *n1 = sample_opaque == NULL ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1); 1464 Node *n2 = sample_opaque == NULL ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2); 1465 phi1->set_req(i, n1); 1466 phi2->set_req(i, n2); 1467 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 1468 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 1469 } 1470 // See if these Phis have been made before. 1471 // Register with optimizer 1472 Node *hit1 = _igvn.hash_find_insert(phi1); 1473 if (hit1) { // Hit, toss just made Phi 1474 _igvn.remove_dead_node(phi1); // Remove new phi 1475 assert(hit1->is_Phi(), "" ); 1476 phi1 = (PhiNode*)hit1; // Use existing phi 1477 } else { // Miss 1478 _igvn.register_new_node_with_optimizer(phi1); 1479 } 1480 Node *hit2 = _igvn.hash_find_insert(phi2); 1481 if (hit2) { // Hit, toss just made Phi 1482 _igvn.remove_dead_node(phi2); // Remove new phi 1483 assert(hit2->is_Phi(), "" ); 1484 phi2 = (PhiNode*)hit2; // Use existing phi 1485 } else { // Miss 1486 _igvn.register_new_node_with_optimizer(phi2); 1487 } 1488 // Register Phis with loop/block info 1489 set_ctrl(phi1, phi->in(0)); 1490 set_ctrl(phi2, phi->in(0)); 1491 // Make a new Cmp 1492 Node *cmp = sample_cmp->clone(); 1493 cmp->set_req(1, phi1); 1494 cmp->set_req(2, phi2); 1495 _igvn.register_new_node_with_optimizer(cmp); 1496 set_ctrl(cmp, phi->in(0)); 1497 1498 // Make a new Bool 1499 Node *b = sample_bool->clone(); 1500 b->set_req(1,cmp); 1501 _igvn.register_new_node_with_optimizer(b); 1502 set_ctrl(b, phi->in(0)); 1503 1504 if (sample_opaque != NULL) { 1505 Node* opaque = sample_opaque->clone(); 1506 opaque->set_req(1, b); 1507 _igvn.register_new_node_with_optimizer(opaque); 1508 set_ctrl(opaque, phi->in(0)); 1509 return opaque; 1510 } 1511 1512 assert(b->is_Bool(), ""); 1513 return b; 1514 } 1515 1516 //------------------------------clone_bool------------------------------------- 1517 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1518 // "Nearly" because all Nodes have been cloned from the original in the loop, 1519 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1520 // through the Phi recursively, and return a Bool. 1521 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) { 1522 uint i; 1523 // Convert this Phi into a Phi merging Bools 1524 for( i = 1; i < phi->req(); i++ ) { 1525 Node *b = phi->in(i); 1526 if( b->is_Phi() ) { 1527 _igvn.replace_input_of(phi, i, clone_bool( b->as_Phi(), loop )); 1528 } else { 1529 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" ); 1530 } 1531 } 1532 1533 Node *sample_cmp = phi->in(1); 1534 1535 // Make Phis to merge the Cmp's inputs. 1536 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP ); 1537 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP ); 1538 for( uint j = 1; j < phi->req(); j++ ) { 1539 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP 1540 Node *n1, *n2; 1541 if( cmp_top->is_Cmp() ) { 1542 n1 = cmp_top->in(1); 1543 n2 = cmp_top->in(2); 1544 } else { 1545 n1 = n2 = cmp_top; 1546 } 1547 phi1->set_req( j, n1 ); 1548 phi2->set_req( j, n2 ); 1549 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 1550 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 1551 } 1552 1553 // See if these Phis have been made before. 1554 // Register with optimizer 1555 Node *hit1 = _igvn.hash_find_insert(phi1); 1556 if( hit1 ) { // Hit, toss just made Phi 1557 _igvn.remove_dead_node(phi1); // Remove new phi 1558 assert( hit1->is_Phi(), "" ); 1559 phi1 = (PhiNode*)hit1; // Use existing phi 1560 } else { // Miss 1561 _igvn.register_new_node_with_optimizer(phi1); 1562 } 1563 Node *hit2 = _igvn.hash_find_insert(phi2); 1564 if( hit2 ) { // Hit, toss just made Phi 1565 _igvn.remove_dead_node(phi2); // Remove new phi 1566 assert( hit2->is_Phi(), "" ); 1567 phi2 = (PhiNode*)hit2; // Use existing phi 1568 } else { // Miss 1569 _igvn.register_new_node_with_optimizer(phi2); 1570 } 1571 // Register Phis with loop/block info 1572 set_ctrl(phi1, phi->in(0)); 1573 set_ctrl(phi2, phi->in(0)); 1574 // Make a new Cmp 1575 Node *cmp = sample_cmp->clone(); 1576 cmp->set_req( 1, phi1 ); 1577 cmp->set_req( 2, phi2 ); 1578 _igvn.register_new_node_with_optimizer(cmp); 1579 set_ctrl(cmp, phi->in(0)); 1580 1581 assert( cmp->is_Cmp(), "" ); 1582 return (CmpNode*)cmp; 1583 } 1584 1585 //------------------------------sink_use--------------------------------------- 1586 // If 'use' was in the loop-exit block, it now needs to be sunk 1587 // below the post-loop merge point. 1588 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) { 1589 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) { 1590 set_ctrl(use, post_loop); 1591 for (DUIterator j = use->outs(); use->has_out(j); j++) 1592 sink_use(use->out(j), post_loop); 1593 } 1594 } 1595 1596 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new, 1597 IdealLoopTree* loop, IdealLoopTree* outer_loop, 1598 Node_List*& split_if_set, Node_List*& split_bool_set, 1599 Node_List*& split_cex_set, Node_List& worklist, 1600 uint new_counter, CloneLoopMode mode) { 1601 Node* nnn = old_new[old->_idx]; 1602 // Copy uses to a worklist, so I can munge the def-use info 1603 // with impunity. 1604 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1605 worklist.push(old->fast_out(j)); 1606 1607 while( worklist.size() ) { 1608 Node *use = worklist.pop(); 1609 if (!has_node(use)) continue; // Ignore dead nodes 1610 if (use->in(0) == C->top()) continue; 1611 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1612 // Check for data-use outside of loop - at least one of OLD or USE 1613 // must not be a CFG node. 1614 #ifdef ASSERT 1615 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == NULL) { 1616 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint(); 1617 assert(mode == ControlAroundStripMined && use == sfpt, "missed a node"); 1618 } 1619 #endif 1620 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) { 1621 1622 // If the Data use is an IF, that means we have an IF outside of the 1623 // loop that is switching on a condition that is set inside of the 1624 // loop. Happens if people set a loop-exit flag; then test the flag 1625 // in the loop to break the loop, then test is again outside of the 1626 // loop to determine which way the loop exited. 1627 // Loop predicate If node connects to Bool node through Opaque1 node. 1628 if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use) || use->Opcode() == Op_Opaque4) { 1629 // Since this code is highly unlikely, we lazily build the worklist 1630 // of such Nodes to go split. 1631 if (!split_if_set) { 1632 ResourceArea *area = Thread::current()->resource_area(); 1633 split_if_set = new Node_List(area); 1634 } 1635 split_if_set->push(use); 1636 } 1637 if (use->is_Bool()) { 1638 if (!split_bool_set) { 1639 ResourceArea *area = Thread::current()->resource_area(); 1640 split_bool_set = new Node_List(area); 1641 } 1642 split_bool_set->push(use); 1643 } 1644 if (use->Opcode() == Op_CreateEx) { 1645 if (!split_cex_set) { 1646 ResourceArea *area = Thread::current()->resource_area(); 1647 split_cex_set = new Node_List(area); 1648 } 1649 split_cex_set->push(use); 1650 } 1651 1652 1653 // Get "block" use is in 1654 uint idx = 0; 1655 while( use->in(idx) != old ) idx++; 1656 Node *prev = use->is_CFG() ? use : get_ctrl(use); 1657 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" ); 1658 Node *cfg = prev->_idx >= new_counter 1659 ? prev->in(2) 1660 : idom(prev); 1661 if( use->is_Phi() ) // Phi use is in prior block 1662 cfg = prev->in(idx); // NOT in block of Phi itself 1663 if (cfg->is_top()) { // Use is dead? 1664 _igvn.replace_input_of(use, idx, C->top()); 1665 continue; 1666 } 1667 1668 while(!outer_loop->is_member(get_loop(cfg))) { 1669 prev = cfg; 1670 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg); 1671 } 1672 // If the use occurs after merging several exits from the loop, then 1673 // old value must have dominated all those exits. Since the same old 1674 // value was used on all those exits we did not need a Phi at this 1675 // merge point. NOW we do need a Phi here. Each loop exit value 1676 // is now merged with the peeled body exit; each exit gets its own 1677 // private Phi and those Phis need to be merged here. 1678 Node *phi; 1679 if( prev->is_Region() ) { 1680 if( idx == 0 ) { // Updating control edge? 1681 phi = prev; // Just use existing control 1682 } else { // Else need a new Phi 1683 phi = PhiNode::make( prev, old ); 1684 // Now recursively fix up the new uses of old! 1685 for( uint i = 1; i < prev->req(); i++ ) { 1686 worklist.push(phi); // Onto worklist once for each 'old' input 1687 } 1688 } 1689 } else { 1690 // Get new RegionNode merging old and new loop exits 1691 prev = old_new[prev->_idx]; 1692 assert( prev, "just made this in step 7" ); 1693 if( idx == 0) { // Updating control edge? 1694 phi = prev; // Just use existing control 1695 } else { // Else need a new Phi 1696 // Make a new Phi merging data values properly 1697 phi = PhiNode::make( prev, old ); 1698 phi->set_req( 1, nnn ); 1699 } 1700 } 1701 // If inserting a new Phi, check for prior hits 1702 if( idx != 0 ) { 1703 Node *hit = _igvn.hash_find_insert(phi); 1704 if( hit == NULL ) { 1705 _igvn.register_new_node_with_optimizer(phi); // Register new phi 1706 } else { // or 1707 // Remove the new phi from the graph and use the hit 1708 _igvn.remove_dead_node(phi); 1709 phi = hit; // Use existing phi 1710 } 1711 set_ctrl(phi, prev); 1712 } 1713 // Make 'use' use the Phi instead of the old loop body exit value 1714 _igvn.replace_input_of(use, idx, phi); 1715 if( use->_idx >= new_counter ) { // If updating new phis 1716 // Not needed for correctness, but prevents a weak assert 1717 // in AddPNode from tripping (when we end up with different 1718 // base & derived Phis that will become the same after 1719 // IGVN does CSE). 1720 Node *hit = _igvn.hash_find_insert(use); 1721 if( hit ) // Go ahead and re-hash for hits. 1722 _igvn.replace_node( use, hit ); 1723 } 1724 1725 // If 'use' was in the loop-exit block, it now needs to be sunk 1726 // below the post-loop merge point. 1727 sink_use( use, prev ); 1728 } 1729 } 1730 } 1731 1732 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop, 1733 IdealLoopTree* outer_loop, int dd, Node_List &old_new, 1734 Node_List& extra_data_nodes) { 1735 if (head->is_strip_mined() && mode != IgnoreStripMined) { 1736 CountedLoopNode* cl = head->as_CountedLoop(); 1737 Node* l = cl->outer_loop(); 1738 Node* tail = cl->outer_loop_tail(); 1739 IfNode* le = cl->outer_loop_end(); 1740 Node* sfpt = cl->outer_safepoint(); 1741 CountedLoopEndNode* cle = cl->loopexit(); 1742 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop(); 1743 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null(); 1744 Node* cle_out = cle->proj_out(false); 1745 1746 Node* new_sfpt = NULL; 1747 Node* new_cle_out = cle_out->clone(); 1748 old_new.map(cle_out->_idx, new_cle_out); 1749 if (mode == CloneIncludesStripMined) { 1750 // clone outer loop body 1751 Node* new_l = l->clone(); 1752 Node* new_tail = tail->clone(); 1753 IfNode* new_le = le->clone()->as_If(); 1754 new_sfpt = sfpt->clone(); 1755 1756 set_loop(new_l, outer_loop->_parent); 1757 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd); 1758 set_loop(new_cle_out, outer_loop->_parent); 1759 set_idom(new_cle_out, new_cle, dd); 1760 set_loop(new_sfpt, outer_loop->_parent); 1761 set_idom(new_sfpt, new_cle_out, dd); 1762 set_loop(new_le, outer_loop->_parent); 1763 set_idom(new_le, new_sfpt, dd); 1764 set_loop(new_tail, outer_loop->_parent); 1765 set_idom(new_tail, new_le, dd); 1766 set_idom(new_cl, new_l, dd); 1767 1768 old_new.map(l->_idx, new_l); 1769 old_new.map(tail->_idx, new_tail); 1770 old_new.map(le->_idx, new_le); 1771 old_new.map(sfpt->_idx, new_sfpt); 1772 1773 new_l->set_req(LoopNode::LoopBackControl, new_tail); 1774 new_l->set_req(0, new_l); 1775 new_tail->set_req(0, new_le); 1776 new_le->set_req(0, new_sfpt); 1777 new_sfpt->set_req(0, new_cle_out); 1778 new_cle_out->set_req(0, new_cle); 1779 new_cl->set_req(LoopNode::EntryControl, new_l); 1780 1781 _igvn.register_new_node_with_optimizer(new_l); 1782 _igvn.register_new_node_with_optimizer(new_tail); 1783 _igvn.register_new_node_with_optimizer(new_le); 1784 } else { 1785 Node *newhead = old_new[loop->_head->_idx]; 1786 newhead->as_Loop()->clear_strip_mined(); 1787 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl)); 1788 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 1789 } 1790 // Look at data node that were assigned a control in the outer 1791 // loop: they are kept in the outer loop by the safepoint so start 1792 // from the safepoint node's inputs. 1793 IdealLoopTree* outer_loop = get_loop(l); 1794 Node_Stack stack(2); 1795 stack.push(sfpt, 1); 1796 uint new_counter = C->unique(); 1797 while (stack.size() > 0) { 1798 Node* n = stack.node(); 1799 uint i = stack.index(); 1800 while (i < n->req() && 1801 (n->in(i) == NULL || 1802 !has_ctrl(n->in(i)) || 1803 get_loop(get_ctrl(n->in(i))) != outer_loop || 1804 (old_new[n->in(i)->_idx] != NULL && old_new[n->in(i)->_idx]->_idx >= new_counter))) { 1805 i++; 1806 } 1807 if (i < n->req()) { 1808 stack.set_index(i+1); 1809 stack.push(n->in(i), 0); 1810 } else { 1811 assert(old_new[n->_idx] == NULL || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet"); 1812 Node* m = n == sfpt ? new_sfpt : n->clone(); 1813 if (m != NULL) { 1814 for (uint i = 0; i < n->req(); i++) { 1815 if (m->in(i) != NULL && old_new[m->in(i)->_idx] != NULL) { 1816 m->set_req(i, old_new[m->in(i)->_idx]); 1817 } 1818 } 1819 } else { 1820 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?"); 1821 } 1822 if (n != sfpt) { 1823 extra_data_nodes.push(n); 1824 _igvn.register_new_node_with_optimizer(m); 1825 assert(get_ctrl(n) == cle_out, "what other control?"); 1826 set_ctrl(m, new_cle_out); 1827 old_new.map(n->_idx, m); 1828 } 1829 stack.pop(); 1830 } 1831 } 1832 if (mode == CloneIncludesStripMined) { 1833 _igvn.register_new_node_with_optimizer(new_sfpt); 1834 _igvn.register_new_node_with_optimizer(new_cle_out); 1835 } 1836 } else { 1837 Node *newhead = old_new[loop->_head->_idx]; 1838 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 1839 } 1840 } 1841 1842 //------------------------------clone_loop------------------------------------- 1843 // 1844 // C L O N E A L O O P B O D Y 1845 // 1846 // This is the basic building block of the loop optimizations. It clones an 1847 // entire loop body. It makes an old_new loop body mapping; with this mapping 1848 // you can find the new-loop equivalent to an old-loop node. All new-loop 1849 // nodes are exactly equal to their old-loop counterparts, all edges are the 1850 // same. All exits from the old-loop now have a RegionNode that merges the 1851 // equivalent new-loop path. This is true even for the normal "loop-exit" 1852 // condition. All uses of loop-invariant old-loop values now come from (one 1853 // or more) Phis that merge their new-loop equivalents. 1854 // 1855 // This operation leaves the graph in an illegal state: there are two valid 1856 // control edges coming from the loop pre-header to both loop bodies. I'll 1857 // definitely have to hack the graph after running this transform. 1858 // 1859 // From this building block I will further edit edges to perform loop peeling 1860 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc. 1861 // 1862 // Parameter side_by_size_idom: 1863 // When side_by_size_idom is NULL, the dominator tree is constructed for 1864 // the clone loop to dominate the original. Used in construction of 1865 // pre-main-post loop sequence. 1866 // When nonnull, the clone and original are side-by-side, both are 1867 // dominated by the side_by_side_idom node. Used in construction of 1868 // unswitched loops. 1869 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd, 1870 CloneLoopMode mode, Node* side_by_side_idom) { 1871 1872 LoopNode* head = loop->_head->as_Loop(); 1873 head->verify_strip_mined(1); 1874 1875 if (C->do_vector_loop() && PrintOpto) { 1876 const char* mname = C->method()->name()->as_quoted_ascii(); 1877 if (mname != NULL) { 1878 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname); 1879 } 1880 } 1881 1882 CloneMap& cm = C->clone_map(); 1883 Dict* dict = cm.dict(); 1884 if (C->do_vector_loop()) { 1885 cm.set_clone_idx(cm.max_gen()+1); 1886 #ifndef PRODUCT 1887 if (PrintOpto) { 1888 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx()); 1889 loop->dump_head(); 1890 } 1891 #endif 1892 } 1893 1894 // Step 1: Clone the loop body. Make the old->new mapping. 1895 uint i; 1896 for( i = 0; i < loop->_body.size(); i++ ) { 1897 Node *old = loop->_body.at(i); 1898 Node *nnn = old->clone(); 1899 old_new.map( old->_idx, nnn ); 1900 if (C->do_vector_loop()) { 1901 cm.verify_insert_and_clone(old, nnn, cm.clone_idx()); 1902 } 1903 _igvn.register_new_node_with_optimizer(nnn); 1904 } 1905 1906 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop; 1907 1908 // Step 2: Fix the edges in the new body. If the old input is outside the 1909 // loop use it. If the old input is INside the loop, use the corresponding 1910 // new node instead. 1911 for( i = 0; i < loop->_body.size(); i++ ) { 1912 Node *old = loop->_body.at(i); 1913 Node *nnn = old_new[old->_idx]; 1914 // Fix CFG/Loop controlling the new node 1915 if (has_ctrl(old)) { 1916 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]); 1917 } else { 1918 set_loop(nnn, outer_loop->_parent); 1919 if (old->outcnt() > 0) { 1920 set_idom( nnn, old_new[idom(old)->_idx], dd ); 1921 } 1922 } 1923 // Correct edges to the new node 1924 for( uint j = 0; j < nnn->req(); j++ ) { 1925 Node *n = nnn->in(j); 1926 if( n ) { 1927 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n ); 1928 if( loop->is_member( old_in_loop ) ) 1929 nnn->set_req(j, old_new[n->_idx]); 1930 } 1931 } 1932 _igvn.hash_find_insert(nnn); 1933 } 1934 1935 ResourceArea *area = Thread::current()->resource_area(); 1936 Node_List extra_data_nodes(area); 1937 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes); 1938 1939 // Step 3: Now fix control uses. Loop varying control uses have already 1940 // been fixed up (as part of all input edges in Step 2). Loop invariant 1941 // control uses must be either an IfFalse or an IfTrue. Make a merge 1942 // point to merge the old and new IfFalse/IfTrue nodes; make the use 1943 // refer to this. 1944 Node_List worklist(area); 1945 uint new_counter = C->unique(); 1946 for( i = 0; i < loop->_body.size(); i++ ) { 1947 Node* old = loop->_body.at(i); 1948 if( !old->is_CFG() ) continue; 1949 1950 // Copy uses to a worklist, so I can munge the def-use info 1951 // with impunity. 1952 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1953 worklist.push(old->fast_out(j)); 1954 1955 while( worklist.size() ) { // Visit all uses 1956 Node *use = worklist.pop(); 1957 if (!has_node(use)) continue; // Ignore dead nodes 1958 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1959 if( !loop->is_member( use_loop ) && use->is_CFG() ) { 1960 // Both OLD and USE are CFG nodes here. 1961 assert( use->is_Proj(), "" ); 1962 Node* nnn = old_new[old->_idx]; 1963 1964 Node* newuse = NULL; 1965 if (head->is_strip_mined() && mode != IgnoreStripMined) { 1966 CountedLoopNode* cl = head->as_CountedLoop(); 1967 CountedLoopEndNode* cle = cl->loopexit(); 1968 Node* cle_out = cle->proj_out_or_null(false); 1969 if (use == cle_out) { 1970 IfNode* le = cl->outer_loop_end(); 1971 use = le->proj_out(false); 1972 use_loop = get_loop(use); 1973 if (mode == CloneIncludesStripMined) { 1974 nnn = old_new[le->_idx]; 1975 } else { 1976 newuse = old_new[cle_out->_idx]; 1977 } 1978 } 1979 } 1980 if (newuse == NULL) { 1981 newuse = use->clone(); 1982 } 1983 1984 // Clone the loop exit control projection 1985 if (C->do_vector_loop()) { 1986 cm.verify_insert_and_clone(use, newuse, cm.clone_idx()); 1987 } 1988 newuse->set_req(0,nnn); 1989 _igvn.register_new_node_with_optimizer(newuse); 1990 set_loop(newuse, use_loop); 1991 set_idom(newuse, nnn, dom_depth(nnn) + 1 ); 1992 1993 // We need a Region to merge the exit from the peeled body and the 1994 // exit from the old loop body. 1995 RegionNode *r = new RegionNode(3); 1996 // Map the old use to the new merge point 1997 old_new.map( use->_idx, r ); 1998 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use)); 1999 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" ); 2000 2001 // The original user of 'use' uses 'r' instead. 2002 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) { 2003 Node* useuse = use->last_out(l); 2004 _igvn.rehash_node_delayed(useuse); 2005 uint uses_found = 0; 2006 if( useuse->in(0) == use ) { 2007 useuse->set_req(0, r); 2008 uses_found++; 2009 if( useuse->is_CFG() ) { 2010 assert( dom_depth(useuse) > dd_r, "" ); 2011 set_idom(useuse, r, dom_depth(useuse)); 2012 } 2013 } 2014 for( uint k = 1; k < useuse->req(); k++ ) { 2015 if( useuse->in(k) == use ) { 2016 useuse->set_req(k, r); 2017 uses_found++; 2018 if (useuse->is_Loop() && k == LoopNode::EntryControl) { 2019 assert(dom_depth(useuse) > dd_r , ""); 2020 set_idom(useuse, r, dom_depth(useuse)); 2021 } 2022 } 2023 } 2024 l -= uses_found; // we deleted 1 or more copies of this edge 2025 } 2026 2027 // Now finish up 'r' 2028 r->set_req( 1, newuse ); 2029 r->set_req( 2, use ); 2030 _igvn.register_new_node_with_optimizer(r); 2031 set_loop(r, use_loop); 2032 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r); 2033 } // End of if a loop-exit test 2034 } 2035 } 2036 2037 // Step 4: If loop-invariant use is not control, it must be dominated by a 2038 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region 2039 // there if needed. Make a Phi there merging old and new used values. 2040 Node_List *split_if_set = NULL; 2041 Node_List *split_bool_set = NULL; 2042 Node_List *split_cex_set = NULL; 2043 for( i = 0; i < loop->_body.size(); i++ ) { 2044 Node* old = loop->_body.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 for (i = 0; i < extra_data_nodes.size(); i++) { 2051 Node* old = extra_data_nodes.at(i); 2052 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set, 2053 split_bool_set, split_cex_set, worklist, new_counter, 2054 mode); 2055 } 2056 2057 // Check for IFs that need splitting/cloning. Happens if an IF outside of 2058 // the loop uses a condition set in the loop. The original IF probably 2059 // takes control from one or more OLD Regions (which in turn get from NEW 2060 // Regions). In any case, there will be a set of Phis for each merge point 2061 // from the IF up to where the original BOOL def exists the loop. 2062 if (split_if_set) { 2063 while (split_if_set->size()) { 2064 Node *iff = split_if_set->pop(); 2065 if (iff->in(1)->is_Phi()) { 2066 Node *b = clone_iff(iff->in(1)->as_Phi(), loop); 2067 _igvn.replace_input_of(iff, 1, b); 2068 } 2069 } 2070 } 2071 if (split_bool_set) { 2072 while (split_bool_set->size()) { 2073 Node *b = split_bool_set->pop(); 2074 Node *phi = b->in(1); 2075 assert(phi->is_Phi(), ""); 2076 CmpNode *cmp = clone_bool((PhiNode*)phi, loop); 2077 _igvn.replace_input_of(b, 1, cmp); 2078 } 2079 } 2080 if (split_cex_set) { 2081 while (split_cex_set->size()) { 2082 Node *b = split_cex_set->pop(); 2083 assert(b->in(0)->is_Region(), ""); 2084 assert(b->in(1)->is_Phi(), ""); 2085 assert(b->in(0)->in(0) == b->in(1)->in(0), ""); 2086 split_up(b, b->in(0), NULL); 2087 } 2088 } 2089 2090 } 2091 2092 2093 //---------------------- stride_of_possible_iv ------------------------------------- 2094 // Looks for an iff/bool/comp with one operand of the compare 2095 // being a cycle involving an add and a phi, 2096 // with an optional truncation (left-shift followed by a right-shift) 2097 // of the add. Returns zero if not an iv. 2098 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) { 2099 Node* trunc1 = NULL; 2100 Node* trunc2 = NULL; 2101 const TypeInt* ttype = NULL; 2102 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) { 2103 return 0; 2104 } 2105 BoolNode* bl = iff->in(1)->as_Bool(); 2106 Node* cmp = bl->in(1); 2107 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) { 2108 return 0; 2109 } 2110 // Must have an invariant operand 2111 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) { 2112 return 0; 2113 } 2114 Node* add2 = NULL; 2115 Node* cmp1 = cmp->in(1); 2116 if (cmp1->is_Phi()) { 2117 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) ))) 2118 Node* phi = cmp1; 2119 for (uint i = 1; i < phi->req(); i++) { 2120 Node* in = phi->in(i); 2121 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in, 2122 &trunc1, &trunc2, &ttype); 2123 if (add && add->in(1) == phi) { 2124 add2 = add->in(2); 2125 break; 2126 } 2127 } 2128 } else { 2129 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) ))) 2130 Node* addtrunc = cmp1; 2131 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc, 2132 &trunc1, &trunc2, &ttype); 2133 if (add && add->in(1)->is_Phi()) { 2134 Node* phi = add->in(1); 2135 for (uint i = 1; i < phi->req(); i++) { 2136 if (phi->in(i) == addtrunc) { 2137 add2 = add->in(2); 2138 break; 2139 } 2140 } 2141 } 2142 } 2143 if (add2 != NULL) { 2144 const TypeInt* add2t = _igvn.type(add2)->is_int(); 2145 if (add2t->is_con()) { 2146 return add2t->get_con(); 2147 } 2148 } 2149 return 0; 2150 } 2151 2152 2153 //---------------------- stay_in_loop ------------------------------------- 2154 // Return the (unique) control output node that's in the loop (if it exists.) 2155 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) { 2156 Node* unique = NULL; 2157 if (!n) return NULL; 2158 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2159 Node* use = n->fast_out(i); 2160 if (!has_ctrl(use) && loop->is_member(get_loop(use))) { 2161 if (unique != NULL) { 2162 return NULL; 2163 } 2164 unique = use; 2165 } 2166 } 2167 return unique; 2168 } 2169 2170 //------------------------------ register_node ------------------------------------- 2171 // Utility to register node "n" with PhaseIdealLoop 2172 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) { 2173 _igvn.register_new_node_with_optimizer(n); 2174 loop->_body.push(n); 2175 if (n->is_CFG()) { 2176 set_loop(n, loop); 2177 set_idom(n, pred, ddepth); 2178 } else { 2179 set_ctrl(n, pred); 2180 } 2181 } 2182 2183 //------------------------------ proj_clone ------------------------------------- 2184 // Utility to create an if-projection 2185 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) { 2186 ProjNode* c = p->clone()->as_Proj(); 2187 c->set_req(0, iff); 2188 return c; 2189 } 2190 2191 //------------------------------ short_circuit_if ------------------------------------- 2192 // Force the iff control output to be the live_proj 2193 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) { 2194 guarantee(live_proj != NULL, "null projection"); 2195 int proj_con = live_proj->_con; 2196 assert(proj_con == 0 || proj_con == 1, "false or true projection"); 2197 Node *con = _igvn.intcon(proj_con); 2198 set_ctrl(con, C->root()); 2199 if (iff) { 2200 iff->set_req(1, con); 2201 } 2202 return con; 2203 } 2204 2205 //------------------------------ insert_if_before_proj ------------------------------------- 2206 // Insert a new if before an if projection (* - new node) 2207 // 2208 // before 2209 // if(test) 2210 // / \ 2211 // v v 2212 // other-proj proj (arg) 2213 // 2214 // after 2215 // if(test) 2216 // / \ 2217 // / v 2218 // | * proj-clone 2219 // v | 2220 // other-proj v 2221 // * new_if(relop(cmp[IU](left,right))) 2222 // / \ 2223 // v v 2224 // * new-proj proj 2225 // (returned) 2226 // 2227 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) { 2228 IfNode* iff = proj->in(0)->as_If(); 2229 IdealLoopTree *loop = get_loop(proj); 2230 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2231 int ddepth = dom_depth(proj); 2232 2233 _igvn.rehash_node_delayed(iff); 2234 _igvn.rehash_node_delayed(proj); 2235 2236 proj->set_req(0, NULL); // temporary disconnect 2237 ProjNode* proj2 = proj_clone(proj, iff); 2238 register_node(proj2, loop, iff, ddepth); 2239 2240 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right); 2241 register_node(cmp, loop, proj2, ddepth); 2242 2243 BoolNode* bol = new BoolNode(cmp, relop); 2244 register_node(bol, loop, proj2, ddepth); 2245 2246 int opcode = iff->Opcode(); 2247 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode"); 2248 IfNode* new_if = (opcode == Op_If) ? new IfNode(proj2, bol, iff->_prob, iff->_fcnt): 2249 new RangeCheckNode(proj2, bol, iff->_prob, iff->_fcnt); 2250 register_node(new_if, loop, proj2, ddepth); 2251 2252 proj->set_req(0, new_if); // reattach 2253 set_idom(proj, new_if, ddepth); 2254 2255 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj(); 2256 guarantee(new_exit != NULL, "null exit node"); 2257 register_node(new_exit, get_loop(other_proj), new_if, ddepth); 2258 2259 return new_exit; 2260 } 2261 2262 //------------------------------ insert_region_before_proj ------------------------------------- 2263 // Insert a region before an if projection (* - new node) 2264 // 2265 // before 2266 // if(test) 2267 // / | 2268 // v | 2269 // proj v 2270 // other-proj 2271 // 2272 // after 2273 // if(test) 2274 // / | 2275 // v | 2276 // * proj-clone v 2277 // | other-proj 2278 // v 2279 // * new-region 2280 // | 2281 // v 2282 // * dum_if 2283 // / \ 2284 // v \ 2285 // * dum-proj v 2286 // proj 2287 // 2288 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) { 2289 IfNode* iff = proj->in(0)->as_If(); 2290 IdealLoopTree *loop = get_loop(proj); 2291 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2292 int ddepth = dom_depth(proj); 2293 2294 _igvn.rehash_node_delayed(iff); 2295 _igvn.rehash_node_delayed(proj); 2296 2297 proj->set_req(0, NULL); // temporary disconnect 2298 ProjNode* proj2 = proj_clone(proj, iff); 2299 register_node(proj2, loop, iff, ddepth); 2300 2301 RegionNode* reg = new RegionNode(2); 2302 reg->set_req(1, proj2); 2303 register_node(reg, loop, iff, ddepth); 2304 2305 IfNode* dum_if = new IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt); 2306 register_node(dum_if, loop, reg, ddepth); 2307 2308 proj->set_req(0, dum_if); // reattach 2309 set_idom(proj, dum_if, ddepth); 2310 2311 ProjNode* dum_proj = proj_clone(other_proj, dum_if); 2312 register_node(dum_proj, loop, dum_if, ddepth); 2313 2314 return reg; 2315 } 2316 2317 //------------------------------ insert_cmpi_loop_exit ------------------------------------- 2318 // Clone a signed compare loop exit from an unsigned compare and 2319 // insert it before the unsigned cmp on the stay-in-loop path. 2320 // All new nodes inserted in the dominator tree between the original 2321 // if and it's projections. The original if test is replaced with 2322 // a constant to force the stay-in-loop path. 2323 // 2324 // This is done to make sure that the original if and it's projections 2325 // still dominate the same set of control nodes, that the ctrl() relation 2326 // from data nodes to them is preserved, and that their loop nesting is 2327 // preserved. 2328 // 2329 // before 2330 // if(i <u limit) unsigned compare loop exit 2331 // / | 2332 // v v 2333 // exit-proj stay-in-loop-proj 2334 // 2335 // after 2336 // if(stay-in-loop-const) original if 2337 // / | 2338 // / v 2339 // / if(i < limit) new signed test 2340 // / / | 2341 // / / v 2342 // / / if(i <u limit) new cloned unsigned test 2343 // / / / | 2344 // v v v | 2345 // region | 2346 // | | 2347 // dum-if | 2348 // / | | 2349 // ether | | 2350 // v v 2351 // exit-proj stay-in-loop-proj 2352 // 2353 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) { 2354 const bool Signed = true; 2355 const bool Unsigned = false; 2356 2357 BoolNode* bol = if_cmpu->in(1)->as_Bool(); 2358 if (bol->_test._test != BoolTest::lt) return NULL; 2359 CmpNode* cmpu = bol->in(1)->as_Cmp(); 2360 if (cmpu->Opcode() != Op_CmpU) return NULL; 2361 int stride = stride_of_possible_iv(if_cmpu); 2362 if (stride == 0) return NULL; 2363 2364 Node* lp_proj = stay_in_loop(if_cmpu, loop); 2365 guarantee(lp_proj != NULL, "null loop node"); 2366 2367 ProjNode* lp_continue = lp_proj->as_Proj(); 2368 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj(); 2369 2370 Node* limit = NULL; 2371 if (stride > 0) { 2372 limit = cmpu->in(2); 2373 } else { 2374 limit = _igvn.makecon(TypeInt::ZERO); 2375 set_ctrl(limit, C->root()); 2376 } 2377 // Create a new region on the exit path 2378 RegionNode* reg = insert_region_before_proj(lp_exit); 2379 guarantee(reg != NULL, "null region node"); 2380 2381 // Clone the if-cmpu-true-false using a signed compare 2382 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge; 2383 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue); 2384 reg->add_req(cmpi_exit); 2385 2386 // Clone the if-cmpu-true-false 2387 BoolTest::mask rel_u = bol->_test._test; 2388 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue); 2389 reg->add_req(cmpu_exit); 2390 2391 // Force original if to stay in loop. 2392 short_circuit_if(if_cmpu, lp_continue); 2393 2394 return cmpi_exit->in(0)->as_If(); 2395 } 2396 2397 //------------------------------ remove_cmpi_loop_exit ------------------------------------- 2398 // Remove a previously inserted signed compare loop exit. 2399 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) { 2400 Node* lp_proj = stay_in_loop(if_cmp, loop); 2401 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI && 2402 stay_in_loop(lp_proj, loop)->is_If() && 2403 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu"); 2404 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO); 2405 set_ctrl(con, C->root()); 2406 if_cmp->set_req(1, con); 2407 } 2408 2409 //------------------------------ scheduled_nodelist ------------------------------------- 2410 // Create a post order schedule of nodes that are in the 2411 // "member" set. The list is returned in "sched". 2412 // The first node in "sched" is the loop head, followed by 2413 // nodes which have no inputs in the "member" set, and then 2414 // followed by the nodes that have an immediate input dependence 2415 // on a node in "sched". 2416 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) { 2417 2418 assert(member.test(loop->_head->_idx), "loop head must be in member set"); 2419 Arena *a = Thread::current()->resource_area(); 2420 VectorSet visited(a); 2421 Node_Stack nstack(a, loop->_body.size()); 2422 2423 Node* n = loop->_head; // top of stack is cached in "n" 2424 uint idx = 0; 2425 visited.set(n->_idx); 2426 2427 // Initially push all with no inputs from within member set 2428 for(uint i = 0; i < loop->_body.size(); i++ ) { 2429 Node *elt = loop->_body.at(i); 2430 if (member.test(elt->_idx)) { 2431 bool found = false; 2432 for (uint j = 0; j < elt->req(); j++) { 2433 Node* def = elt->in(j); 2434 if (def && member.test(def->_idx) && def != elt) { 2435 found = true; 2436 break; 2437 } 2438 } 2439 if (!found && elt != loop->_head) { 2440 nstack.push(n, idx); 2441 n = elt; 2442 assert(!visited.test(n->_idx), "not seen yet"); 2443 visited.set(n->_idx); 2444 } 2445 } 2446 } 2447 2448 // traverse out's that are in the member set 2449 while (true) { 2450 if (idx < n->outcnt()) { 2451 Node* use = n->raw_out(idx); 2452 idx++; 2453 if (!visited.test_set(use->_idx)) { 2454 if (member.test(use->_idx)) { 2455 nstack.push(n, idx); 2456 n = use; 2457 idx = 0; 2458 } 2459 } 2460 } else { 2461 // All outputs processed 2462 sched.push(n); 2463 if (nstack.is_empty()) break; 2464 n = nstack.node(); 2465 idx = nstack.index(); 2466 nstack.pop(); 2467 } 2468 } 2469 } 2470 2471 2472 //------------------------------ has_use_in_set ------------------------------------- 2473 // Has a use in the vector set 2474 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) { 2475 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2476 Node* use = n->fast_out(j); 2477 if (vset.test(use->_idx)) { 2478 return true; 2479 } 2480 } 2481 return false; 2482 } 2483 2484 2485 //------------------------------ has_use_internal_to_set ------------------------------------- 2486 // Has use internal to the vector set (ie. not in a phi at the loop head) 2487 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) { 2488 Node* head = loop->_head; 2489 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2490 Node* use = n->fast_out(j); 2491 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) { 2492 return true; 2493 } 2494 } 2495 return false; 2496 } 2497 2498 2499 //------------------------------ clone_for_use_outside_loop ------------------------------------- 2500 // clone "n" for uses that are outside of loop 2501 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) { 2502 int cloned = 0; 2503 assert(worklist.size() == 0, "should be empty"); 2504 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2505 Node* use = n->fast_out(j); 2506 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) { 2507 worklist.push(use); 2508 } 2509 } 2510 while( worklist.size() ) { 2511 Node *use = worklist.pop(); 2512 if (!has_node(use) || use->in(0) == C->top()) continue; 2513 uint j; 2514 for (j = 0; j < use->req(); j++) { 2515 if (use->in(j) == n) break; 2516 } 2517 assert(j < use->req(), "must be there"); 2518 2519 // clone "n" and insert it between the inputs of "n" and the use outside the loop 2520 Node* n_clone = n->clone(); 2521 _igvn.replace_input_of(use, j, n_clone); 2522 cloned++; 2523 Node* use_c; 2524 if (!use->is_Phi()) { 2525 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0); 2526 } else { 2527 // Use in a phi is considered a use in the associated predecessor block 2528 use_c = use->in(0)->in(j); 2529 } 2530 set_ctrl(n_clone, use_c); 2531 assert(!loop->is_member(get_loop(use_c)), "should be outside loop"); 2532 get_loop(use_c)->_body.push(n_clone); 2533 _igvn.register_new_node_with_optimizer(n_clone); 2534 #if !defined(PRODUCT) 2535 if (TracePartialPeeling) { 2536 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx); 2537 } 2538 #endif 2539 } 2540 return cloned; 2541 } 2542 2543 2544 //------------------------------ clone_for_special_use_inside_loop ------------------------------------- 2545 // clone "n" for special uses that are in the not_peeled region. 2546 // If these def-uses occur in separate blocks, the code generator 2547 // marks the method as not compilable. For example, if a "BoolNode" 2548 // is in a different basic block than the "IfNode" that uses it, then 2549 // the compilation is aborted in the code generator. 2550 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 2551 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) { 2552 if (n->is_Phi() || n->is_Load()) { 2553 return; 2554 } 2555 assert(worklist.size() == 0, "should be empty"); 2556 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2557 Node* use = n->fast_out(j); 2558 if ( not_peel.test(use->_idx) && 2559 (use->is_If() || use->is_CMove() || use->is_Bool()) && 2560 use->in(1) == n) { 2561 worklist.push(use); 2562 } 2563 } 2564 if (worklist.size() > 0) { 2565 // clone "n" and insert it between inputs of "n" and the use 2566 Node* n_clone = n->clone(); 2567 loop->_body.push(n_clone); 2568 _igvn.register_new_node_with_optimizer(n_clone); 2569 set_ctrl(n_clone, get_ctrl(n)); 2570 sink_list.push(n_clone); 2571 not_peel <<= n_clone->_idx; // add n_clone to not_peel set. 2572 #if !defined(PRODUCT) 2573 if (TracePartialPeeling) { 2574 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx); 2575 } 2576 #endif 2577 while( worklist.size() ) { 2578 Node *use = worklist.pop(); 2579 _igvn.rehash_node_delayed(use); 2580 for (uint j = 1; j < use->req(); j++) { 2581 if (use->in(j) == n) { 2582 use->set_req(j, n_clone); 2583 } 2584 } 2585 } 2586 } 2587 } 2588 2589 2590 //------------------------------ insert_phi_for_loop ------------------------------------- 2591 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 2592 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) { 2593 Node *phi = PhiNode::make(lp, back_edge_val); 2594 phi->set_req(LoopNode::EntryControl, lp_entry_val); 2595 // Use existing phi if it already exists 2596 Node *hit = _igvn.hash_find_insert(phi); 2597 if( hit == NULL ) { 2598 _igvn.register_new_node_with_optimizer(phi); 2599 set_ctrl(phi, lp); 2600 } else { 2601 // Remove the new phi from the graph and use the hit 2602 _igvn.remove_dead_node(phi); 2603 phi = hit; 2604 } 2605 _igvn.replace_input_of(use, idx, phi); 2606 } 2607 2608 #ifdef ASSERT 2609 //------------------------------ is_valid_loop_partition ------------------------------------- 2610 // Validate the loop partition sets: peel and not_peel 2611 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, 2612 VectorSet& not_peel ) { 2613 uint i; 2614 // Check that peel_list entries are in the peel set 2615 for (i = 0; i < peel_list.size(); i++) { 2616 if (!peel.test(peel_list.at(i)->_idx)) { 2617 return false; 2618 } 2619 } 2620 // Check at loop members are in one of peel set or not_peel set 2621 for (i = 0; i < loop->_body.size(); i++ ) { 2622 Node *def = loop->_body.at(i); 2623 uint di = def->_idx; 2624 // Check that peel set elements are in peel_list 2625 if (peel.test(di)) { 2626 if (not_peel.test(di)) { 2627 return false; 2628 } 2629 // Must be in peel_list also 2630 bool found = false; 2631 for (uint j = 0; j < peel_list.size(); j++) { 2632 if (peel_list.at(j)->_idx == di) { 2633 found = true; 2634 break; 2635 } 2636 } 2637 if (!found) { 2638 return false; 2639 } 2640 } else if (not_peel.test(di)) { 2641 if (peel.test(di)) { 2642 return false; 2643 } 2644 } else { 2645 return false; 2646 } 2647 } 2648 return true; 2649 } 2650 2651 //------------------------------ is_valid_clone_loop_exit_use ------------------------------------- 2652 // Ensure a use outside of loop is of the right form 2653 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) { 2654 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2655 return (use->is_Phi() && 2656 use_c->is_Region() && use_c->req() == 3 && 2657 (use_c->in(exit_idx)->Opcode() == Op_IfTrue || 2658 use_c->in(exit_idx)->Opcode() == Op_IfFalse || 2659 use_c->in(exit_idx)->Opcode() == Op_JumpProj) && 2660 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) ); 2661 } 2662 2663 //------------------------------ is_valid_clone_loop_form ------------------------------------- 2664 // Ensure that all uses outside of loop are of the right form 2665 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 2666 uint orig_exit_idx, uint clone_exit_idx) { 2667 uint len = peel_list.size(); 2668 for (uint i = 0; i < len; i++) { 2669 Node *def = peel_list.at(i); 2670 2671 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 2672 Node *use = def->fast_out(j); 2673 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2674 if (!loop->is_member(get_loop(use_c))) { 2675 // use is not in the loop, check for correct structure 2676 if (use->in(0) == def) { 2677 // Okay 2678 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) { 2679 return false; 2680 } 2681 } 2682 } 2683 } 2684 return true; 2685 } 2686 #endif 2687 2688 //------------------------------ partial_peel ------------------------------------- 2689 // Partially peel (aka loop rotation) the top portion of a loop (called 2690 // the peel section below) by cloning it and placing one copy just before 2691 // the new loop head and the other copy at the bottom of the new loop. 2692 // 2693 // before after where it came from 2694 // 2695 // stmt1 stmt1 2696 // loop: stmt2 clone 2697 // stmt2 if condA goto exitA clone 2698 // if condA goto exitA new_loop: new 2699 // stmt3 stmt3 clone 2700 // if !condB goto loop if condB goto exitB clone 2701 // exitB: stmt2 orig 2702 // stmt4 if !condA goto new_loop orig 2703 // exitA: goto exitA 2704 // exitB: 2705 // stmt4 2706 // exitA: 2707 // 2708 // Step 1: find the cut point: an exit test on probable 2709 // induction variable. 2710 // Step 2: schedule (with cloning) operations in the peel 2711 // section that can be executed after the cut into 2712 // the section that is not peeled. This may need 2713 // to clone operations into exit blocks. For 2714 // instance, a reference to A[i] in the not-peel 2715 // section and a reference to B[i] in an exit block 2716 // may cause a left-shift of i by 2 to be placed 2717 // in the peel block. This step will clone the left 2718 // shift into the exit block and sink the left shift 2719 // from the peel to the not-peel section. 2720 // Step 3: clone the loop, retarget the control, and insert 2721 // phis for values that are live across the new loop 2722 // head. This is very dependent on the graph structure 2723 // from clone_loop. It creates region nodes for 2724 // exit control and associated phi nodes for values 2725 // flow out of the loop through that exit. The region 2726 // node is dominated by the clone's control projection. 2727 // So the clone's peel section is placed before the 2728 // new loop head, and the clone's not-peel section is 2729 // forms the top part of the new loop. The original 2730 // peel section forms the tail of the new loop. 2731 // Step 4: update the dominator tree and recompute the 2732 // dominator depth. 2733 // 2734 // orig 2735 // 2736 // stmt1 2737 // | 2738 // v 2739 // loop predicate 2740 // | 2741 // v 2742 // loop<----+ 2743 // | | 2744 // stmt2 | 2745 // | | 2746 // v | 2747 // ifA | 2748 // / | | 2749 // v v | 2750 // false true ^ <-- last_peel 2751 // / | | 2752 // / ===|==cut | 2753 // / stmt3 | <-- first_not_peel 2754 // / | | 2755 // | v | 2756 // v ifB | 2757 // exitA: / \ | 2758 // / \ | 2759 // v v | 2760 // false true | 2761 // / \ | 2762 // / ----+ 2763 // | 2764 // v 2765 // exitB: 2766 // stmt4 2767 // 2768 // 2769 // after clone loop 2770 // 2771 // stmt1 2772 // | 2773 // v 2774 // loop predicate 2775 // / \ 2776 // clone / \ orig 2777 // / \ 2778 // / \ 2779 // v v 2780 // +---->loop loop<----+ 2781 // | | | | 2782 // | stmt2 stmt2 | 2783 // | | | | 2784 // | v v | 2785 // | ifA ifA | 2786 // | | \ / | | 2787 // | v v v v | 2788 // ^ true false false true ^ <-- last_peel 2789 // | | ^ \ / | | 2790 // | cut==|== \ \ / ===|==cut | 2791 // | stmt3 \ \ / stmt3 | <-- first_not_peel 2792 // | | dom | | | | 2793 // | v \ 1v v2 v | 2794 // | ifB regionA ifB | 2795 // | / \ | / \ | 2796 // | / \ v / \ | 2797 // | v v exitA: v v | 2798 // | true false false true | 2799 // | / ^ \ / \ | 2800 // +---- \ \ / ----+ 2801 // dom \ / 2802 // \ 1v v2 2803 // regionB 2804 // | 2805 // v 2806 // exitB: 2807 // stmt4 2808 // 2809 // 2810 // after partial peel 2811 // 2812 // stmt1 2813 // | 2814 // v 2815 // loop predicate 2816 // / 2817 // clone / orig 2818 // / TOP 2819 // / \ 2820 // v v 2821 // TOP->loop loop----+ 2822 // | | | 2823 // stmt2 stmt2 | 2824 // | | | 2825 // v v | 2826 // ifA ifA | 2827 // | \ / | | 2828 // v v v v | 2829 // true false false true | <-- last_peel 2830 // | ^ \ / +------|---+ 2831 // +->newloop \ \ / === ==cut | | 2832 // | stmt3 \ \ / TOP | | 2833 // | | dom | | stmt3 | | <-- first_not_peel 2834 // | v \ 1v v2 v | | 2835 // | ifB regionA ifB ^ v 2836 // | / \ | / \ | | 2837 // | / \ v / \ | | 2838 // | v v exitA: v v | | 2839 // | true false false true | | 2840 // | / ^ \ / \ | | 2841 // | | \ \ / v | | 2842 // | | dom \ / TOP | | 2843 // | | \ 1v v2 | | 2844 // ^ v regionB | | 2845 // | | | | | 2846 // | | v ^ v 2847 // | | exitB: | | 2848 // | | stmt4 | | 2849 // | +------------>-----------------+ | 2850 // | | 2851 // +-----------------<---------------------+ 2852 // 2853 // 2854 // final graph 2855 // 2856 // stmt1 2857 // | 2858 // v 2859 // loop predicate 2860 // | 2861 // v 2862 // stmt2 clone 2863 // | 2864 // v 2865 // ........> ifA clone 2866 // : / | 2867 // dom / | 2868 // : v v 2869 // : false true 2870 // : | | 2871 // : | v 2872 // : | newloop<-----+ 2873 // : | | | 2874 // : | stmt3 clone | 2875 // : | | | 2876 // : | v | 2877 // : | ifB | 2878 // : | / \ | 2879 // : | v v | 2880 // : | false true | 2881 // : | | | | 2882 // : | v stmt2 | 2883 // : | exitB: | | 2884 // : | stmt4 v | 2885 // : | ifA orig | 2886 // : | / \ | 2887 // : | / \ | 2888 // : | v v | 2889 // : | false true | 2890 // : | / \ | 2891 // : v v -----+ 2892 // RegionA 2893 // | 2894 // v 2895 // exitA 2896 // 2897 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) { 2898 2899 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only"); 2900 if (!loop->_head->is_Loop()) { 2901 return false; } 2902 2903 LoopNode *head = loop->_head->as_Loop(); 2904 2905 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) { 2906 return false; 2907 } 2908 2909 // Check for complex exit control 2910 for(uint ii = 0; ii < loop->_body.size(); ii++ ) { 2911 Node *n = loop->_body.at(ii); 2912 int opc = n->Opcode(); 2913 if (n->is_Call() || 2914 opc == Op_Catch || 2915 opc == Op_CatchProj || 2916 opc == Op_Jump || 2917 opc == Op_JumpProj) { 2918 #if !defined(PRODUCT) 2919 if (TracePartialPeeling) { 2920 tty->print_cr("\nExit control too complex: lp: %d", head->_idx); 2921 } 2922 #endif 2923 return false; 2924 } 2925 } 2926 2927 int dd = dom_depth(head); 2928 2929 // Step 1: find cut point 2930 2931 // Walk up dominators to loop head looking for first loop exit 2932 // which is executed on every path thru loop. 2933 IfNode *peel_if = NULL; 2934 IfNode *peel_if_cmpu = NULL; 2935 2936 Node *iff = loop->tail(); 2937 while( iff != head ) { 2938 if( iff->is_If() ) { 2939 Node *ctrl = get_ctrl(iff->in(1)); 2940 if (ctrl->is_top()) return false; // Dead test on live IF. 2941 // If loop-varying exit-test, check for induction variable 2942 if( loop->is_member(get_loop(ctrl)) && 2943 loop->is_loop_exit(iff) && 2944 is_possible_iv_test(iff)) { 2945 Node* cmp = iff->in(1)->in(1); 2946 if (cmp->Opcode() == Op_CmpI) { 2947 peel_if = iff->as_If(); 2948 } else { 2949 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU"); 2950 peel_if_cmpu = iff->as_If(); 2951 } 2952 } 2953 } 2954 iff = idom(iff); 2955 } 2956 // Prefer signed compare over unsigned compare. 2957 IfNode* new_peel_if = NULL; 2958 if (peel_if == NULL) { 2959 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) { 2960 return false; // No peel point found 2961 } 2962 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop); 2963 if (new_peel_if == NULL) { 2964 return false; // No peel point found 2965 } 2966 peel_if = new_peel_if; 2967 } 2968 Node* last_peel = stay_in_loop(peel_if, loop); 2969 Node* first_not_peeled = stay_in_loop(last_peel, loop); 2970 if (first_not_peeled == NULL || first_not_peeled == head) { 2971 return false; 2972 } 2973 2974 #if !defined(PRODUCT) 2975 if (TraceLoopOpts) { 2976 tty->print("PartialPeel "); 2977 loop->dump_head(); 2978 } 2979 2980 if (TracePartialPeeling) { 2981 tty->print_cr("before partial peel one iteration"); 2982 Node_List wl; 2983 Node* t = head->in(2); 2984 while (true) { 2985 wl.push(t); 2986 if (t == head) break; 2987 t = idom(t); 2988 } 2989 while (wl.size() > 0) { 2990 Node* tt = wl.pop(); 2991 tt->dump(); 2992 if (tt == last_peel) tty->print_cr("-- cut --"); 2993 } 2994 } 2995 #endif 2996 ResourceArea *area = Thread::current()->resource_area(); 2997 VectorSet peel(area); 2998 VectorSet not_peel(area); 2999 Node_List peel_list(area); 3000 Node_List worklist(area); 3001 Node_List sink_list(area); 3002 3003 // Set of cfg nodes to peel are those that are executable from 3004 // the head through last_peel. 3005 assert(worklist.size() == 0, "should be empty"); 3006 worklist.push(head); 3007 peel.set(head->_idx); 3008 while (worklist.size() > 0) { 3009 Node *n = worklist.pop(); 3010 if (n != last_peel) { 3011 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3012 Node* use = n->fast_out(j); 3013 if (use->is_CFG() && 3014 loop->is_member(get_loop(use)) && 3015 !peel.test_set(use->_idx)) { 3016 worklist.push(use); 3017 } 3018 } 3019 } 3020 } 3021 3022 // Set of non-cfg nodes to peel are those that are control 3023 // dependent on the cfg nodes. 3024 uint i; 3025 for(i = 0; i < loop->_body.size(); i++ ) { 3026 Node *n = loop->_body.at(i); 3027 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n; 3028 if (peel.test(n_c->_idx)) { 3029 peel.set(n->_idx); 3030 } else { 3031 not_peel.set(n->_idx); 3032 } 3033 } 3034 3035 // Step 2: move operations from the peeled section down into the 3036 // not-peeled section 3037 3038 // Get a post order schedule of nodes in the peel region 3039 // Result in right-most operand. 3040 scheduled_nodelist(loop, peel, peel_list ); 3041 3042 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 3043 3044 // For future check for too many new phis 3045 uint old_phi_cnt = 0; 3046 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 3047 Node* use = head->fast_out(j); 3048 if (use->is_Phi()) old_phi_cnt++; 3049 } 3050 3051 #if !defined(PRODUCT) 3052 if (TracePartialPeeling) { 3053 tty->print_cr("\npeeled list"); 3054 } 3055 #endif 3056 3057 // Evacuate nodes in peel region into the not_peeled region if possible 3058 uint new_phi_cnt = 0; 3059 uint cloned_for_outside_use = 0; 3060 for (i = 0; i < peel_list.size();) { 3061 Node* n = peel_list.at(i); 3062 #if !defined(PRODUCT) 3063 if (TracePartialPeeling) n->dump(); 3064 #endif 3065 bool incr = true; 3066 if ( !n->is_CFG() ) { 3067 3068 if ( has_use_in_set(n, not_peel) ) { 3069 3070 // If not used internal to the peeled region, 3071 // move "n" from peeled to not_peeled region. 3072 3073 if ( !has_use_internal_to_set(n, peel, loop) ) { 3074 3075 // if not pinned and not a load (which maybe anti-dependent on a store) 3076 // and not a CMove (Matcher expects only bool->cmove). 3077 if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) { 3078 cloned_for_outside_use += clone_for_use_outside_loop( loop, n, worklist ); 3079 sink_list.push(n); 3080 peel >>= n->_idx; // delete n from peel set. 3081 not_peel <<= n->_idx; // add n to not_peel set. 3082 peel_list.remove(i); 3083 incr = false; 3084 #if !defined(PRODUCT) 3085 if (TracePartialPeeling) { 3086 tty->print_cr("sink to not_peeled region: %d newbb: %d", 3087 n->_idx, get_ctrl(n)->_idx); 3088 } 3089 #endif 3090 } 3091 } else { 3092 // Otherwise check for special def-use cases that span 3093 // the peel/not_peel boundary such as bool->if 3094 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist ); 3095 new_phi_cnt++; 3096 } 3097 } 3098 } 3099 if (incr) i++; 3100 } 3101 3102 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) { 3103 #if !defined(PRODUCT) 3104 if (TracePartialPeeling) { 3105 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c", 3106 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F'); 3107 } 3108 #endif 3109 if (new_peel_if != NULL) { 3110 remove_cmpi_loop_exit(new_peel_if, loop); 3111 } 3112 // Inhibit more partial peeling on this loop 3113 assert(!head->is_partial_peel_loop(), "not partial peeled"); 3114 head->mark_partial_peel_failed(); 3115 if (cloned_for_outside_use > 0) { 3116 // Terminate this round of loop opts because 3117 // the graph outside this loop was changed. 3118 C->set_major_progress(); 3119 return true; 3120 } 3121 return false; 3122 } 3123 3124 // Step 3: clone loop, retarget control, and insert new phis 3125 3126 // Create new loop head for new phis and to hang 3127 // the nodes being moved (sinked) from the peel region. 3128 LoopNode* new_head = new LoopNode(last_peel, last_peel); 3129 new_head->set_unswitch_count(head->unswitch_count()); // Preserve 3130 _igvn.register_new_node_with_optimizer(new_head); 3131 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled"); 3132 _igvn.replace_input_of(first_not_peeled, 0, new_head); 3133 set_loop(new_head, loop); 3134 loop->_body.push(new_head); 3135 not_peel.set(new_head->_idx); 3136 set_idom(new_head, last_peel, dom_depth(first_not_peeled)); 3137 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled)); 3138 3139 while (sink_list.size() > 0) { 3140 Node* n = sink_list.pop(); 3141 set_ctrl(n, new_head); 3142 } 3143 3144 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 3145 3146 clone_loop(loop, old_new, dd, IgnoreStripMined); 3147 3148 const uint clone_exit_idx = 1; 3149 const uint orig_exit_idx = 2; 3150 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop"); 3151 3152 Node* head_clone = old_new[head->_idx]; 3153 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop(); 3154 Node* orig_tail_clone = head_clone->in(2); 3155 3156 // Add phi if "def" node is in peel set and "use" is not 3157 3158 for(i = 0; i < peel_list.size(); i++ ) { 3159 Node *def = peel_list.at(i); 3160 if (!def->is_CFG()) { 3161 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 3162 Node *use = def->fast_out(j); 3163 if (has_node(use) && use->in(0) != C->top() && 3164 (!peel.test(use->_idx) || 3165 (use->is_Phi() && use->in(0) == head)) ) { 3166 worklist.push(use); 3167 } 3168 } 3169 while( worklist.size() ) { 3170 Node *use = worklist.pop(); 3171 for (uint j = 1; j < use->req(); j++) { 3172 Node* n = use->in(j); 3173 if (n == def) { 3174 3175 // "def" is in peel set, "use" is not in peel set 3176 // or "use" is in the entry boundary (a phi) of the peel set 3177 3178 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use; 3179 3180 if ( loop->is_member(get_loop( use_c )) ) { 3181 // use is in loop 3182 if (old_new[use->_idx] != NULL) { // null for dead code 3183 Node* use_clone = old_new[use->_idx]; 3184 _igvn.replace_input_of(use, j, C->top()); 3185 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone ); 3186 } 3187 } else { 3188 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format"); 3189 // use is not in the loop, check if the live range includes the cut 3190 Node* lp_if = use_c->in(orig_exit_idx)->in(0); 3191 if (not_peel.test(lp_if->_idx)) { 3192 assert(j == orig_exit_idx, "use from original loop"); 3193 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone ); 3194 } 3195 } 3196 } 3197 } 3198 } 3199 } 3200 } 3201 3202 // Step 3b: retarget control 3203 3204 // Redirect control to the new loop head if a cloned node in 3205 // the not_peeled region has control that points into the peeled region. 3206 // This necessary because the cloned peeled region will be outside 3207 // the loop. 3208 // from to 3209 // cloned-peeled <---+ 3210 // new_head_clone: | <--+ 3211 // cloned-not_peeled in(0) in(0) 3212 // orig-peeled 3213 3214 for(i = 0; i < loop->_body.size(); i++ ) { 3215 Node *n = loop->_body.at(i); 3216 if (!n->is_CFG() && n->in(0) != NULL && 3217 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) { 3218 Node* n_clone = old_new[n->_idx]; 3219 _igvn.replace_input_of(n_clone, 0, new_head_clone); 3220 } 3221 } 3222 3223 // Backedge of the surviving new_head (the clone) is original last_peel 3224 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel); 3225 3226 // Cut first node in original not_peel set 3227 _igvn.rehash_node_delayed(new_head); // Multiple edge updates: 3228 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of 3229 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls 3230 3231 // Copy head_clone back-branch info to original head 3232 // and remove original head's loop entry and 3233 // clone head's back-branch 3234 _igvn.rehash_node_delayed(head); // Multiple edge updates 3235 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl)); 3236 head->set_req(LoopNode::LoopBackControl, C->top()); 3237 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top()); 3238 3239 // Similarly modify the phis 3240 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) { 3241 Node* use = head->fast_out(k); 3242 if (use->is_Phi() && use->outcnt() > 0) { 3243 Node* use_clone = old_new[use->_idx]; 3244 _igvn.rehash_node_delayed(use); // Multiple edge updates 3245 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl)); 3246 use->set_req(LoopNode::LoopBackControl, C->top()); 3247 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top()); 3248 } 3249 } 3250 3251 // Step 4: update dominator tree and dominator depth 3252 3253 set_idom(head, orig_tail_clone, dd); 3254 recompute_dom_depth(); 3255 3256 // Inhibit more partial peeling on this loop 3257 new_head_clone->set_partial_peel_loop(); 3258 C->set_major_progress(); 3259 loop->record_for_igvn(); 3260 3261 #if !defined(PRODUCT) 3262 if (TracePartialPeeling) { 3263 tty->print_cr("\nafter partial peel one iteration"); 3264 Node_List wl(area); 3265 Node* t = last_peel; 3266 while (true) { 3267 wl.push(t); 3268 if (t == head_clone) break; 3269 t = idom(t); 3270 } 3271 while (wl.size() > 0) { 3272 Node* tt = wl.pop(); 3273 if (tt == head) tty->print_cr("orig head"); 3274 else if (tt == new_head_clone) tty->print_cr("new head"); 3275 else if (tt == head_clone) tty->print_cr("clone head"); 3276 tt->dump(); 3277 } 3278 } 3279 #endif 3280 return true; 3281 } 3282 3283 //------------------------------reorg_offsets---------------------------------- 3284 // Reorganize offset computations to lower register pressure. Mostly 3285 // prevent loop-fallout uses of the pre-incremented trip counter (which are 3286 // then alive with the post-incremented trip counter forcing an extra 3287 // register move) 3288 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) { 3289 // Perform it only for canonical counted loops. 3290 // Loop's shape could be messed up by iteration_split_impl. 3291 if (!loop->_head->is_CountedLoop()) 3292 return; 3293 if (!loop->_head->as_Loop()->is_valid_counted_loop()) 3294 return; 3295 3296 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 3297 CountedLoopEndNode *cle = cl->loopexit(); 3298 Node *exit = cle->proj_out(false); 3299 Node *phi = cl->phi(); 3300 3301 // Check for the special case of folks using the pre-incremented 3302 // trip-counter on the fall-out path (forces the pre-incremented 3303 // and post-incremented trip counter to be live at the same time). 3304 // Fix this by adjusting to use the post-increment trip counter. 3305 3306 bool progress = true; 3307 while (progress) { 3308 progress = false; 3309 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) { 3310 Node* use = phi->fast_out(i); // User of trip-counter 3311 if (!has_ctrl(use)) continue; 3312 Node *u_ctrl = get_ctrl(use); 3313 if (use->is_Phi()) { 3314 u_ctrl = NULL; 3315 for (uint j = 1; j < use->req(); j++) 3316 if (use->in(j) == phi) 3317 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j)); 3318 } 3319 IdealLoopTree *u_loop = get_loop(u_ctrl); 3320 // Look for loop-invariant use 3321 if (u_loop == loop) continue; 3322 if (loop->is_member(u_loop)) continue; 3323 // Check that use is live out the bottom. Assuming the trip-counter 3324 // update is right at the bottom, uses of of the loop middle are ok. 3325 if (dom_lca(exit, u_ctrl) != exit) continue; 3326 // Hit! Refactor use to use the post-incremented tripcounter. 3327 // Compute a post-increment tripcounter. 3328 Node *opaq = new Opaque2Node( C, cle->incr() ); 3329 register_new_node(opaq, exit); 3330 Node *neg_stride = _igvn.intcon(-cle->stride_con()); 3331 set_ctrl(neg_stride, C->root()); 3332 Node *post = new AddINode( opaq, neg_stride); 3333 register_new_node(post, exit); 3334 _igvn.rehash_node_delayed(use); 3335 for (uint j = 1; j < use->req(); j++) { 3336 if (use->in(j) == phi) 3337 use->set_req(j, post); 3338 } 3339 // Since DU info changed, rerun loop 3340 progress = true; 3341 break; 3342 } 3343 } 3344 3345 }