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