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