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