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