1 /* 2 * Copyright (c) 1999, 2011, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "memory/allocation.inline.hpp" 27 #include "opto/addnode.hpp" 28 #include "opto/connode.hpp" 29 #include "opto/divnode.hpp" 30 #include "opto/loopnode.hpp" 31 #include "opto/matcher.hpp" 32 #include "opto/mulnode.hpp" 33 #include "opto/rootnode.hpp" 34 #include "opto/subnode.hpp" 35 36 //============================================================================= 37 //------------------------------split_thru_phi--------------------------------- 38 // Split Node 'n' through merge point if there is enough win. 39 Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) { 40 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) { 41 // ConvI2L may have type information on it which is unsafe to push up 42 // so disable this for now 43 return NULL; 44 } 45 int wins = 0; 46 assert(!n->is_CFG(), ""); 47 assert(region->is_Region(), ""); 48 49 const Type* type = n->bottom_type(); 50 const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr(); 51 Node *phi; 52 if (t_oop != NULL && t_oop->is_known_instance_field()) { 53 int iid = t_oop->instance_id(); 54 int index = C->get_alias_index(t_oop); 55 int offset = t_oop->offset(); 56 phi = new (C,region->req()) PhiNode(region, type, NULL, iid, index, offset); 57 } else { 58 phi = PhiNode::make_blank(region, n); 59 } 60 uint old_unique = C->unique(); 61 for (uint i = 1; i < region->req(); i++) { 62 Node *x; 63 Node* the_clone = NULL; 64 if (region->in(i) == C->top()) { 65 x = C->top(); // Dead path? Use a dead data op 66 } else { 67 x = n->clone(); // Else clone up the data op 68 the_clone = x; // Remember for possible deletion. 69 // Alter data node to use pre-phi inputs 70 if (n->in(0) == region) 71 x->set_req( 0, region->in(i) ); 72 for (uint j = 1; j < n->req(); j++) { 73 Node *in = n->in(j); 74 if (in->is_Phi() && in->in(0) == region) 75 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone 76 } 77 } 78 // Check for a 'win' on some paths 79 const Type *t = x->Value(&_igvn); 80 81 bool singleton = t->singleton(); 82 83 // A TOP singleton indicates that there are no possible values incoming 84 // along a particular edge. In most cases, this is OK, and the Phi will 85 // be eliminated later in an Ideal call. However, we can't allow this to 86 // happen if the singleton occurs on loop entry, as the elimination of 87 // the PhiNode may cause the resulting node to migrate back to a previous 88 // loop iteration. 89 if (singleton && t == Type::TOP) { 90 // Is_Loop() == false does not confirm the absence of a loop (e.g., an 91 // irreducible loop may not be indicated by an affirmative is_Loop()); 92 // therefore, the only top we can split thru a phi is on a backedge of 93 // a loop. 94 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 95 } 96 97 if (singleton) { 98 wins++; 99 x = ((PhaseGVN&)_igvn).makecon(t); 100 } else { 101 // We now call Identity to try to simplify the cloned node. 102 // Note that some Identity methods call phase->type(this). 103 // Make sure that the type array is big enough for 104 // our new node, even though we may throw the node away. 105 // (Note: This tweaking with igvn only works because x is a new node.) 106 _igvn.set_type(x, t); 107 // If x is a TypeNode, capture any more-precise type permanently into Node 108 // otherwise it will be not updated during igvn->transform since 109 // igvn->type(x) is set to x->Value() already. 110 x->raise_bottom_type(t); 111 Node *y = x->Identity(&_igvn); 112 if (y != x) { 113 wins++; 114 x = y; 115 } else { 116 y = _igvn.hash_find(x); 117 if (y) { 118 wins++; 119 x = y; 120 } else { 121 // Else x is a new node we are keeping 122 // We do not need register_new_node_with_optimizer 123 // because set_type has already been called. 124 _igvn._worklist.push(x); 125 } 126 } 127 } 128 if (x != the_clone && the_clone != NULL) 129 _igvn.remove_dead_node(the_clone); 130 phi->set_req( i, x ); 131 } 132 // Too few wins? 133 if (wins <= policy) { 134 _igvn.remove_dead_node(phi); 135 return NULL; 136 } 137 138 // Record Phi 139 register_new_node( phi, region ); 140 141 for (uint i2 = 1; i2 < phi->req(); i2++) { 142 Node *x = phi->in(i2); 143 // If we commoned up the cloned 'x' with another existing Node, 144 // the existing Node picks up a new use. We need to make the 145 // existing Node occur higher up so it dominates its uses. 146 Node *old_ctrl; 147 IdealLoopTree *old_loop; 148 149 if (x->is_Con()) { 150 // Constant's control is always root. 151 set_ctrl(x, C->root()); 152 continue; 153 } 154 // The occasional new node 155 if (x->_idx >= old_unique) { // Found a new, unplaced node? 156 old_ctrl = NULL; 157 old_loop = NULL; // Not in any prior loop 158 } else { 159 old_ctrl = get_ctrl(x); 160 old_loop = get_loop(old_ctrl); // Get prior loop 161 } 162 // New late point must dominate new use 163 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2)); 164 if (new_ctrl == old_ctrl) // Nothing is changed 165 continue; 166 167 IdealLoopTree *new_loop = get_loop(new_ctrl); 168 169 // Don't move x into a loop if its uses are 170 // outside of loop. Otherwise x will be cloned 171 // for each use outside of this loop. 172 IdealLoopTree *use_loop = get_loop(region); 173 if (!new_loop->is_member(use_loop) && 174 (old_loop == NULL || !new_loop->is_member(old_loop))) { 175 // Take early control, later control will be recalculated 176 // during next iteration of loop optimizations. 177 new_ctrl = get_early_ctrl(x); 178 new_loop = get_loop(new_ctrl); 179 } 180 // Set new location 181 set_ctrl(x, new_ctrl); 182 // If changing loop bodies, see if we need to collect into new body 183 if (old_loop != new_loop) { 184 if (old_loop && !old_loop->_child) 185 old_loop->_body.yank(x); 186 if (!new_loop->_child) 187 new_loop->_body.push(x); // Collect body info 188 } 189 } 190 191 return phi; 192 } 193 194 //------------------------------dominated_by------------------------------------ 195 // Replace the dominated test with an obvious true or false. Place it on the 196 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the 197 // live path up to the dominating control. 198 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff, bool flip, bool exclude_loop_predicate ) { 199 #ifndef PRODUCT 200 if (VerifyLoopOptimizations && PrintOpto) tty->print_cr("dominating test"); 201 #endif 202 203 204 // prevdom is the dominating projection of the dominating test. 205 assert( iff->is_If(), "" ); 206 assert( iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added"); 207 int pop = prevdom->Opcode(); 208 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" ); 209 if (flip) { 210 if (pop == Op_IfTrue) 211 pop = Op_IfFalse; 212 else 213 pop = Op_IfTrue; 214 } 215 // 'con' is set to true or false to kill the dominated test. 216 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO); 217 set_ctrl(con, C->root()); // Constant gets a new use 218 // Hack the dominated test 219 _igvn.hash_delete(iff); 220 iff->set_req(1, con); 221 _igvn._worklist.push(iff); 222 223 // If I dont have a reachable TRUE and FALSE path following the IfNode then 224 // I can assume this path reaches an infinite loop. In this case it's not 225 // important to optimize the data Nodes - either the whole compilation will 226 // be tossed or this path (and all data Nodes) will go dead. 227 if (iff->outcnt() != 2) return; 228 229 // Make control-dependent data Nodes on the live path (path that will remain 230 // once the dominated IF is removed) become control-dependent on the 231 // dominating projection. 232 Node* dp = iff->as_If()->proj_out(pop == Op_IfTrue); 233 234 // Loop predicates may have depending checks which should not 235 // be skipped. For example, range check predicate has two checks 236 // for lower and upper bounds. 237 ProjNode* unc_proj = iff->as_If()->proj_out(1 - dp->as_Proj()->_con)->as_Proj(); 238 if (exclude_loop_predicate && 239 is_uncommon_trap_proj(unc_proj, Deoptimization::Reason_predicate)) 240 return; // Let IGVN transformation change control dependence. 241 242 IdealLoopTree *old_loop = get_loop(dp); 243 244 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { 245 Node* cd = dp->fast_out(i); // Control-dependent node 246 if (cd->depends_only_on_test()) { 247 assert(cd->in(0) == dp, ""); 248 _igvn.hash_delete(cd); 249 cd->set_req(0, prevdom); 250 set_early_ctrl(cd); 251 _igvn._worklist.push(cd); 252 IdealLoopTree *new_loop = get_loop(get_ctrl(cd)); 253 if (old_loop != new_loop) { 254 if (!old_loop->_child) old_loop->_body.yank(cd); 255 if (!new_loop->_child) new_loop->_body.push(cd); 256 } 257 --i; 258 --imax; 259 } 260 } 261 } 262 263 //------------------------------has_local_phi_input---------------------------- 264 // Return TRUE if 'n' has Phi inputs from its local block and no other 265 // block-local inputs (all non-local-phi inputs come from earlier blocks) 266 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) { 267 Node *n_ctrl = get_ctrl(n); 268 // See if some inputs come from a Phi in this block, or from before 269 // this block. 270 uint i; 271 for( i = 1; i < n->req(); i++ ) { 272 Node *phi = n->in(i); 273 if( phi->is_Phi() && phi->in(0) == n_ctrl ) 274 break; 275 } 276 if( i >= n->req() ) 277 return NULL; // No Phi inputs; nowhere to clone thru 278 279 // Check for inputs created between 'n' and the Phi input. These 280 // must split as well; they have already been given the chance 281 // (courtesy of a post-order visit) and since they did not we must 282 // recover the 'cost' of splitting them by being very profitable 283 // when splitting 'n'. Since this is unlikely we simply give up. 284 for( i = 1; i < n->req(); i++ ) { 285 Node *m = n->in(i); 286 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) { 287 // We allow the special case of AddP's with no local inputs. 288 // This allows us to split-up address expressions. 289 if (m->is_AddP() && 290 get_ctrl(m->in(2)) != n_ctrl && 291 get_ctrl(m->in(3)) != n_ctrl) { 292 // Move the AddP up to dominating point 293 set_ctrl_and_loop(m, find_non_split_ctrl(idom(n_ctrl))); 294 continue; 295 } 296 return NULL; 297 } 298 } 299 300 return n_ctrl; 301 } 302 303 //------------------------------remix_address_expressions---------------------- 304 // Rework addressing expressions to get the most loop-invariant stuff 305 // moved out. We'd like to do all associative operators, but it's especially 306 // important (common) to do address expressions. 307 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) { 308 if (!has_ctrl(n)) return NULL; 309 Node *n_ctrl = get_ctrl(n); 310 IdealLoopTree *n_loop = get_loop(n_ctrl); 311 312 // See if 'n' mixes loop-varying and loop-invariant inputs and 313 // itself is loop-varying. 314 315 // Only interested in binary ops (and AddP) 316 if( n->req() < 3 || n->req() > 4 ) return NULL; 317 318 Node *n1_ctrl = get_ctrl(n->in( 1)); 319 Node *n2_ctrl = get_ctrl(n->in( 2)); 320 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3)); 321 IdealLoopTree *n1_loop = get_loop( n1_ctrl ); 322 IdealLoopTree *n2_loop = get_loop( n2_ctrl ); 323 IdealLoopTree *n3_loop = get_loop( n3_ctrl ); 324 325 // Does one of my inputs spin in a tighter loop than self? 326 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) || 327 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) || 328 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) ) 329 return NULL; // Leave well enough alone 330 331 // Is at least one of my inputs loop-invariant? 332 if( n1_loop == n_loop && 333 n2_loop == n_loop && 334 n3_loop == n_loop ) 335 return NULL; // No loop-invariant inputs 336 337 338 int n_op = n->Opcode(); 339 340 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2). 341 if( n_op == Op_LShiftI ) { 342 // Scale is loop invariant 343 Node *scale = n->in(2); 344 Node *scale_ctrl = get_ctrl(scale); 345 IdealLoopTree *scale_loop = get_loop(scale_ctrl ); 346 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) ) 347 return NULL; 348 const TypeInt *scale_t = scale->bottom_type()->isa_int(); 349 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 ) 350 return NULL; // Dont bother with byte/short masking 351 // Add must vary with loop (else shift would be loop-invariant) 352 Node *add = n->in(1); 353 Node *add_ctrl = get_ctrl(add); 354 IdealLoopTree *add_loop = get_loop(add_ctrl); 355 //assert( n_loop == add_loop, "" ); 356 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops 357 358 // Convert I-V into I+ (0-V); same for V-I 359 if( add->Opcode() == Op_SubI && 360 _igvn.type( add->in(1) ) != TypeInt::ZERO ) { 361 Node *zero = _igvn.intcon(0); 362 set_ctrl(zero, C->root()); 363 Node *neg = new (C, 3) SubINode( _igvn.intcon(0), add->in(2) ); 364 register_new_node( neg, get_ctrl(add->in(2) ) ); 365 add = new (C, 3) AddINode( add->in(1), neg ); 366 register_new_node( add, add_ctrl ); 367 } 368 if( add->Opcode() != Op_AddI ) return NULL; 369 // See if one add input is loop invariant 370 Node *add_var = add->in(1); 371 Node *add_var_ctrl = get_ctrl(add_var); 372 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 373 Node *add_invar = add->in(2); 374 Node *add_invar_ctrl = get_ctrl(add_invar); 375 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl ); 376 if( add_var_loop == n_loop ) { 377 } else if( add_invar_loop == n_loop ) { 378 // Swap to find the invariant part 379 add_invar = add_var; 380 add_invar_ctrl = add_var_ctrl; 381 add_invar_loop = add_var_loop; 382 add_var = add->in(2); 383 Node *add_var_ctrl = get_ctrl(add_var); 384 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 385 } else // Else neither input is loop invariant 386 return NULL; 387 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) ) 388 return NULL; // No invariant part of the add? 389 390 // Yes! Reshape address expression! 391 Node *inv_scale = new (C, 3) LShiftINode( add_invar, scale ); 392 Node *inv_scale_ctrl = 393 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ? 394 add_invar_ctrl : scale_ctrl; 395 register_new_node( inv_scale, inv_scale_ctrl ); 396 Node *var_scale = new (C, 3) LShiftINode( add_var, scale ); 397 register_new_node( var_scale, n_ctrl ); 398 Node *var_add = new (C, 3) AddINode( var_scale, inv_scale ); 399 register_new_node( var_add, n_ctrl ); 400 _igvn.replace_node( n, var_add ); 401 return var_add; 402 } 403 404 // Replace (I+V) with (V+I) 405 if( n_op == Op_AddI || 406 n_op == Op_AddL || 407 n_op == Op_AddF || 408 n_op == Op_AddD || 409 n_op == Op_MulI || 410 n_op == Op_MulL || 411 n_op == Op_MulF || 412 n_op == Op_MulD ) { 413 if( n2_loop == n_loop ) { 414 assert( n1_loop != n_loop, "" ); 415 n->swap_edges(1, 2); 416 } 417 } 418 419 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V), 420 // but not if I2 is a constant. 421 if( n_op == Op_AddP ) { 422 if( n2_loop == n_loop && n3_loop != n_loop ) { 423 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) { 424 Node *n22_ctrl = get_ctrl(n->in(2)->in(2)); 425 Node *n23_ctrl = get_ctrl(n->in(2)->in(3)); 426 IdealLoopTree *n22loop = get_loop( n22_ctrl ); 427 IdealLoopTree *n23_loop = get_loop( n23_ctrl ); 428 if( n22loop != n_loop && n22loop->is_member(n_loop) && 429 n23_loop == n_loop ) { 430 Node *add1 = new (C, 4) AddPNode( n->in(1), n->in(2)->in(2), n->in(3) ); 431 // Stuff new AddP in the loop preheader 432 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) ); 433 Node *add2 = new (C, 4) AddPNode( n->in(1), add1, n->in(2)->in(3) ); 434 register_new_node( add2, n_ctrl ); 435 _igvn.replace_node( n, add2 ); 436 return add2; 437 } 438 } 439 } 440 441 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V) 442 if( n2_loop != n_loop && n3_loop == n_loop ) { 443 if( n->in(3)->Opcode() == Op_AddI ) { 444 Node *V = n->in(3)->in(1); 445 Node *I = n->in(3)->in(2); 446 if( is_member(n_loop,get_ctrl(V)) ) { 447 } else { 448 Node *tmp = V; V = I; I = tmp; 449 } 450 if( !is_member(n_loop,get_ctrl(I)) ) { 451 Node *add1 = new (C, 4) AddPNode( n->in(1), n->in(2), I ); 452 // Stuff new AddP in the loop preheader 453 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) ); 454 Node *add2 = new (C, 4) AddPNode( n->in(1), add1, V ); 455 register_new_node( add2, n_ctrl ); 456 _igvn.replace_node( n, add2 ); 457 return add2; 458 } 459 } 460 } 461 } 462 463 return NULL; 464 } 465 466 //------------------------------conditional_move------------------------------- 467 // Attempt to replace a Phi with a conditional move. We have some pretty 468 // strict profitability requirements. All Phis at the merge point must 469 // be converted, so we can remove the control flow. We need to limit the 470 // number of c-moves to a small handful. All code that was in the side-arms 471 // of the CFG diamond is now speculatively executed. This code has to be 472 // "cheap enough". We are pretty much limited to CFG diamonds that merge 473 // 1 or 2 items with a total of 1 or 2 ops executed speculatively. 474 Node *PhaseIdealLoop::conditional_move( Node *region ) { 475 476 assert(region->is_Region(), "sanity check"); 477 if (region->req() != 3) return NULL; 478 479 // Check for CFG diamond 480 Node *lp = region->in(1); 481 Node *rp = region->in(2); 482 if (!lp || !rp) return NULL; 483 Node *lp_c = lp->in(0); 484 if (lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If()) return NULL; 485 IfNode *iff = lp_c->as_If(); 486 487 // Check for ops pinned in an arm of the diamond. 488 // Can't remove the control flow in this case 489 if (lp->outcnt() > 1) return NULL; 490 if (rp->outcnt() > 1) return NULL; 491 492 IdealLoopTree* r_loop = get_loop(region); 493 assert(r_loop == get_loop(iff), "sanity"); 494 // Always convert to CMOVE if all results are used only outside this loop. 495 bool used_inside_loop = (r_loop == _ltree_root); 496 497 // Check profitability 498 int cost = 0; 499 int phis = 0; 500 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 501 Node *out = region->fast_out(i); 502 if (!out->is_Phi()) continue; // Ignore other control edges, etc 503 phis++; 504 PhiNode* phi = out->as_Phi(); 505 BasicType bt = phi->type()->basic_type(); 506 switch (bt) { 507 case T_FLOAT: 508 case T_DOUBLE: { 509 cost += Matcher::float_cmove_cost(); // Could be very expensive 510 break; 511 } 512 case T_LONG: { 513 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's 514 } 515 case T_INT: // These all CMOV fine 516 case T_ADDRESS: { // (RawPtr) 517 cost++; 518 break; 519 } 520 case T_NARROWOOP: // Fall through 521 case T_OBJECT: { // Base oops are OK, but not derived oops 522 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr(); 523 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a 524 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus 525 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we 526 // have a Phi for the base here that we convert to a CMOVE all is well 527 // and good. But if the base is dead, we'll not make a CMOVE. Later 528 // the allocator will have to produce a base by creating a CMOVE of the 529 // relevant bases. This puts the allocator in the business of 530 // manufacturing expensive instructions, generally a bad plan. 531 // Just Say No to Conditionally-Moved Derived Pointers. 532 if (tp && tp->offset() != 0) 533 return NULL; 534 cost++; 535 break; 536 } 537 default: 538 return NULL; // In particular, can't do memory or I/O 539 } 540 // Add in cost any speculative ops 541 for (uint j = 1; j < region->req(); j++) { 542 Node *proj = region->in(j); 543 Node *inp = phi->in(j); 544 if (get_ctrl(inp) == proj) { // Found local op 545 cost++; 546 // Check for a chain of dependent ops; these will all become 547 // speculative in a CMOV. 548 for (uint k = 1; k < inp->req(); k++) 549 if (get_ctrl(inp->in(k)) == proj) 550 cost += ConditionalMoveLimit; // Too much speculative goo 551 } 552 } 553 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode. 554 // This will likely Split-If, a higher-payoff operation. 555 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) { 556 Node* use = phi->fast_out(k); 557 if (use->is_Cmp() || use->is_DecodeN() || use->is_EncodeP()) 558 cost += ConditionalMoveLimit; 559 // Is there a use inside the loop? 560 // Note: check only basic types since CMoveP is pinned. 561 if (!used_inside_loop && is_java_primitive(bt)) { 562 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use); 563 if (r_loop == u_loop || r_loop->is_member(u_loop)) { 564 used_inside_loop = true; 565 } 566 } 567 } 568 } 569 Node* bol = iff->in(1); 570 assert(bol->Opcode() == Op_Bool, ""); 571 int cmp_op = bol->in(1)->Opcode(); 572 // It is expensive to generate flags from a float compare. 573 // Avoid duplicated float compare. 574 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL; 575 576 float infrequent_prob = PROB_UNLIKELY_MAG(3); 577 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop. 578 if (used_inside_loop) { 579 if (cost >= ConditionalMoveLimit) return NULL; // Too much goo 580 581 // BlockLayoutByFrequency optimization moves infrequent branch 582 // from hot path. No point in CMOV'ing in such case (110 is used 583 // instead of 100 to take into account not exactness of float value). 584 if (BlockLayoutByFrequency) { 585 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f); 586 } 587 } 588 // Check for highly predictable branch. No point in CMOV'ing if 589 // we are going to predict accurately all the time. 590 if (iff->_prob < infrequent_prob || 591 iff->_prob > (1.0f - infrequent_prob)) 592 return NULL; 593 594 // -------------- 595 // Now replace all Phis with CMOV's 596 Node *cmov_ctrl = iff->in(0); 597 uint flip = (lp->Opcode() == Op_IfTrue); 598 while (1) { 599 PhiNode* phi = NULL; 600 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 601 Node *out = region->fast_out(i); 602 if (out->is_Phi()) { 603 phi = out->as_Phi(); 604 break; 605 } 606 } 607 if (phi == NULL) break; 608 #ifndef PRODUCT 609 if (PrintOpto && VerifyLoopOptimizations) tty->print_cr("CMOV"); 610 #endif 611 // Move speculative ops 612 for (uint j = 1; j < region->req(); j++) { 613 Node *proj = region->in(j); 614 Node *inp = phi->in(j); 615 if (get_ctrl(inp) == proj) { // Found local op 616 #ifndef PRODUCT 617 if (PrintOpto && VerifyLoopOptimizations) { 618 tty->print(" speculate: "); 619 inp->dump(); 620 } 621 #endif 622 set_ctrl(inp, cmov_ctrl); 623 } 624 } 625 Node *cmov = CMoveNode::make( C, cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi) ); 626 register_new_node( cmov, cmov_ctrl ); 627 _igvn.replace_node( phi, cmov ); 628 #ifndef PRODUCT 629 if (TraceLoopOpts) { 630 tty->print("CMOV "); 631 r_loop->dump_head(); 632 if (Verbose) 633 bol->in(1)->dump(1); 634 cmov->dump(1); 635 } 636 if (VerifyLoopOptimizations) verify(); 637 #endif 638 } 639 640 // The useless CFG diamond will fold up later; see the optimization in 641 // RegionNode::Ideal. 642 _igvn._worklist.push(region); 643 644 return iff->in(1); 645 } 646 647 //------------------------------split_if_with_blocks_pre----------------------- 648 // Do the real work in a non-recursive function. Data nodes want to be 649 // cloned in the pre-order so they can feed each other nicely. 650 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) { 651 // Cloning these guys is unlikely to win 652 int n_op = n->Opcode(); 653 if( n_op == Op_MergeMem ) return n; 654 if( n->is_Proj() ) return n; 655 // Do not clone-up CmpFXXX variations, as these are always 656 // followed by a CmpI 657 if( n->is_Cmp() ) return n; 658 // Attempt to use a conditional move instead of a phi/branch 659 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) { 660 Node *cmov = conditional_move( n ); 661 if( cmov ) return cmov; 662 } 663 if( n->is_CFG() || n->is_LoadStore() ) 664 return n; 665 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd 666 n_op == Op_Opaque2 ) { 667 if( !C->major_progress() ) // If chance of no more loop opts... 668 _igvn._worklist.push(n); // maybe we'll remove them 669 return n; 670 } 671 672 if( n->is_Con() ) return n; // No cloning for Con nodes 673 674 Node *n_ctrl = get_ctrl(n); 675 if( !n_ctrl ) return n; // Dead node 676 677 // Attempt to remix address expressions for loop invariants 678 Node *m = remix_address_expressions( n ); 679 if( m ) return m; 680 681 // Determine if the Node has inputs from some local Phi. 682 // Returns the block to clone thru. 683 Node *n_blk = has_local_phi_input( n ); 684 if( !n_blk ) return n; 685 // Do not clone the trip counter through on a CountedLoop 686 // (messes up the canonical shape). 687 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n; 688 689 // Check for having no control input; not pinned. Allow 690 // dominating control. 691 if( n->in(0) ) { 692 Node *dom = idom(n_blk); 693 if( dom_lca( n->in(0), dom ) != n->in(0) ) 694 return n; 695 } 696 // Policy: when is it profitable. You must get more wins than 697 // policy before it is considered profitable. Policy is usually 0, 698 // so 1 win is considered profitable. Big merges will require big 699 // cloning, so get a larger policy. 700 int policy = n_blk->req() >> 2; 701 702 // If the loop is a candidate for range check elimination, 703 // delay splitting through it's phi until a later loop optimization 704 if (n_blk->is_CountedLoop()) { 705 IdealLoopTree *lp = get_loop(n_blk); 706 if (lp && lp->_rce_candidate) { 707 return n; 708 } 709 } 710 711 // Use same limit as split_if_with_blocks_post 712 if( C->unique() > 35000 ) return n; // Method too big 713 714 // Split 'n' through the merge point if it is profitable 715 Node *phi = split_thru_phi( n, n_blk, policy ); 716 if (!phi) return n; 717 718 // Found a Phi to split thru! 719 // Replace 'n' with the new phi 720 _igvn.replace_node( n, phi ); 721 // Moved a load around the loop, 'en-registering' something. 722 if (n_blk->is_Loop() && n->is_Load() && 723 !phi->in(LoopNode::LoopBackControl)->is_Load()) 724 C->set_major_progress(); 725 726 return phi; 727 } 728 729 static bool merge_point_too_heavy(Compile* C, Node* region) { 730 // Bail out if the region and its phis have too many users. 731 int weight = 0; 732 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 733 weight += region->fast_out(i)->outcnt(); 734 } 735 int nodes_left = MaxNodeLimit - C->unique(); 736 if (weight * 8 > nodes_left) { 737 #ifndef PRODUCT 738 if (PrintOpto) 739 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight); 740 #endif 741 return true; 742 } else { 743 return false; 744 } 745 } 746 747 static bool merge_point_safe(Node* region) { 748 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode 749 // having a PhiNode input. This sidesteps the dangerous case where the split 750 // ConvI2LNode may become TOP if the input Value() does not 751 // overlap the ConvI2L range, leaving a node which may not dominate its 752 // uses. 753 // A better fix for this problem can be found in the BugTraq entry, but 754 // expediency for Mantis demands this hack. 755 // 6855164: If the merge point has a FastLockNode with a PhiNode input, we stop 756 // split_if_with_blocks from splitting a block because we could not move around 757 // the FastLockNode. 758 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 759 Node* n = region->fast_out(i); 760 if (n->is_Phi()) { 761 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 762 Node* m = n->fast_out(j); 763 if (m->is_FastLock()) 764 return false; 765 #ifdef _LP64 766 if (m->Opcode() == Op_ConvI2L) 767 return false; 768 #endif 769 } 770 } 771 } 772 return true; 773 } 774 775 776 //------------------------------place_near_use--------------------------------- 777 // Place some computation next to use but not inside inner loops. 778 // For inner loop uses move it to the preheader area. 779 Node *PhaseIdealLoop::place_near_use( Node *useblock ) const { 780 IdealLoopTree *u_loop = get_loop( useblock ); 781 return (u_loop->_irreducible || u_loop->_child) 782 ? useblock 783 : u_loop->_head->in(LoopNode::EntryControl); 784 } 785 786 787 //------------------------------split_if_with_blocks_post---------------------- 788 // Do the real work in a non-recursive function. CFG hackery wants to be 789 // in the post-order, so it can dirty the I-DOM info and not use the dirtied 790 // info. 791 void PhaseIdealLoop::split_if_with_blocks_post( Node *n ) { 792 793 // Cloning Cmp through Phi's involves the split-if transform. 794 // FastLock is not used by an If 795 if( n->is_Cmp() && !n->is_FastLock() ) { 796 if( C->unique() > 35000 ) return; // Method too big 797 798 // Do not do 'split-if' if irreducible loops are present. 799 if( _has_irreducible_loops ) 800 return; 801 802 Node *n_ctrl = get_ctrl(n); 803 // Determine if the Node has inputs from some local Phi. 804 // Returns the block to clone thru. 805 Node *n_blk = has_local_phi_input( n ); 806 if( n_blk != n_ctrl ) return; 807 808 if( merge_point_too_heavy(C, n_ctrl) ) 809 return; 810 811 if( n->outcnt() != 1 ) return; // Multiple bool's from 1 compare? 812 Node *bol = n->unique_out(); 813 assert( bol->is_Bool(), "expect a bool here" ); 814 if( bol->outcnt() != 1 ) return;// Multiple branches from 1 compare? 815 Node *iff = bol->unique_out(); 816 817 // Check some safety conditions 818 if( iff->is_If() ) { // Classic split-if? 819 if( iff->in(0) != n_ctrl ) return; // Compare must be in same blk as if 820 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE 821 if( get_ctrl(iff->in(2)) == n_ctrl || 822 get_ctrl(iff->in(3)) == n_ctrl ) 823 return; // Inputs not yet split-up 824 if ( get_loop(n_ctrl) != get_loop(get_ctrl(iff)) ) { 825 return; // Loop-invar test gates loop-varying CMOVE 826 } 827 } else { 828 return; // some other kind of node, such as an Allocate 829 } 830 831 // Do not do 'split-if' if some paths are dead. First do dead code 832 // elimination and then see if its still profitable. 833 for( uint i = 1; i < n_ctrl->req(); i++ ) 834 if( n_ctrl->in(i) == C->top() ) 835 return; 836 837 // When is split-if profitable? Every 'win' on means some control flow 838 // goes dead, so it's almost always a win. 839 int policy = 0; 840 // If trying to do a 'Split-If' at the loop head, it is only 841 // profitable if the cmp folds up on BOTH paths. Otherwise we 842 // risk peeling a loop forever. 843 844 // CNC - Disabled for now. Requires careful handling of loop 845 // body selection for the cloned code. Also, make sure we check 846 // for any input path not being in the same loop as n_ctrl. For 847 // irreducible loops we cannot check for 'n_ctrl->is_Loop()' 848 // because the alternative loop entry points won't be converted 849 // into LoopNodes. 850 IdealLoopTree *n_loop = get_loop(n_ctrl); 851 for( uint j = 1; j < n_ctrl->req(); j++ ) 852 if( get_loop(n_ctrl->in(j)) != n_loop ) 853 return; 854 855 // Check for safety of the merge point. 856 if( !merge_point_safe(n_ctrl) ) { 857 return; 858 } 859 860 // Split compare 'n' through the merge point if it is profitable 861 Node *phi = split_thru_phi( n, n_ctrl, policy ); 862 if( !phi ) return; 863 864 // Found a Phi to split thru! 865 // Replace 'n' with the new phi 866 _igvn.replace_node( n, phi ); 867 868 // Now split the bool up thru the phi 869 Node *bolphi = split_thru_phi( bol, n_ctrl, -1 ); 870 _igvn.replace_node( bol, bolphi ); 871 assert( iff->in(1) == bolphi, "" ); 872 if( bolphi->Value(&_igvn)->singleton() ) 873 return; 874 875 // Conditional-move? Must split up now 876 if( !iff->is_If() ) { 877 Node *cmovphi = split_thru_phi( iff, n_ctrl, -1 ); 878 _igvn.replace_node( iff, cmovphi ); 879 return; 880 } 881 882 // Now split the IF 883 do_split_if( iff ); 884 return; 885 } 886 887 // Check for an IF ready to split; one that has its 888 // condition codes input coming from a Phi at the block start. 889 int n_op = n->Opcode(); 890 891 // Check for an IF being dominated by another IF same test 892 if( n_op == Op_If ) { 893 Node *bol = n->in(1); 894 uint max = bol->outcnt(); 895 // Check for same test used more than once? 896 if( n_op == Op_If && max > 1 && bol->is_Bool() ) { 897 // Search up IDOMs to see if this IF is dominated. 898 Node *cutoff = get_ctrl(bol); 899 900 // Now search up IDOMs till cutoff, looking for a dominating test 901 Node *prevdom = n; 902 Node *dom = idom(prevdom); 903 while( dom != cutoff ) { 904 if( dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom ) { 905 // Replace the dominated test with an obvious true or false. 906 // Place it on the IGVN worklist for later cleanup. 907 C->set_major_progress(); 908 dominated_by( prevdom, n, false, true ); 909 #ifndef PRODUCT 910 if( VerifyLoopOptimizations ) verify(); 911 #endif 912 return; 913 } 914 prevdom = dom; 915 dom = idom(prevdom); 916 } 917 } 918 } 919 920 // See if a shared loop-varying computation has no loop-varying uses. 921 // Happens if something is only used for JVM state in uncommon trap exits, 922 // like various versions of induction variable+offset. Clone the 923 // computation per usage to allow it to sink out of the loop. 924 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about) 925 Node *n_ctrl = get_ctrl(n); 926 IdealLoopTree *n_loop = get_loop(n_ctrl); 927 if( n_loop != _ltree_root ) { 928 DUIterator_Fast imax, i = n->fast_outs(imax); 929 for (; i < imax; i++) { 930 Node* u = n->fast_out(i); 931 if( !has_ctrl(u) ) break; // Found control user 932 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 933 if( u_loop == n_loop ) break; // Found loop-varying use 934 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop 935 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003 936 } 937 bool did_break = (i < imax); // Did we break out of the previous loop? 938 if (!did_break && n->outcnt() > 1) { // All uses in outer loops! 939 Node *late_load_ctrl; 940 if (n->is_Load()) { 941 // If n is a load, get and save the result from get_late_ctrl(), 942 // to be later used in calculating the control for n's clones. 943 clear_dom_lca_tags(); 944 late_load_ctrl = get_late_ctrl(n, n_ctrl); 945 } 946 // If n is a load, and the late control is the same as the current 947 // control, then the cloning of n is a pointless exercise, because 948 // GVN will ensure that we end up where we started. 949 if (!n->is_Load() || late_load_ctrl != n_ctrl) { 950 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) { 951 Node *u = n->last_out(j); // Clone private computation per use 952 _igvn.hash_delete(u); 953 _igvn._worklist.push(u); 954 Node *x = n->clone(); // Clone computation 955 Node *x_ctrl = NULL; 956 if( u->is_Phi() ) { 957 // Replace all uses of normal nodes. Replace Phi uses 958 // individually, so the separate Nodes can sink down 959 // different paths. 960 uint k = 1; 961 while( u->in(k) != n ) k++; 962 u->set_req( k, x ); 963 // x goes next to Phi input path 964 x_ctrl = u->in(0)->in(k); 965 --j; 966 } else { // Normal use 967 // Replace all uses 968 for( uint k = 0; k < u->req(); k++ ) { 969 if( u->in(k) == n ) { 970 u->set_req( k, x ); 971 --j; 972 } 973 } 974 x_ctrl = get_ctrl(u); 975 } 976 977 // Find control for 'x' next to use but not inside inner loops. 978 // For inner loop uses get the preheader area. 979 x_ctrl = place_near_use(x_ctrl); 980 981 if (n->is_Load()) { 982 // For loads, add a control edge to a CFG node outside of the loop 983 // to force them to not combine and return back inside the loop 984 // during GVN optimization (4641526). 985 // 986 // Because we are setting the actual control input, factor in 987 // the result from get_late_ctrl() so we respect any 988 // anti-dependences. (6233005). 989 x_ctrl = dom_lca(late_load_ctrl, x_ctrl); 990 991 // Don't allow the control input to be a CFG splitting node. 992 // Such nodes should only have ProjNodes as outs, e.g. IfNode 993 // should only have IfTrueNode and IfFalseNode (4985384). 994 x_ctrl = find_non_split_ctrl(x_ctrl); 995 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone"); 996 997 x->set_req(0, x_ctrl); 998 } 999 register_new_node(x, x_ctrl); 1000 1001 // Some institutional knowledge is needed here: 'x' is 1002 // yanked because if the optimizer runs GVN on it all the 1003 // cloned x's will common up and undo this optimization and 1004 // be forced back in the loop. This is annoying because it 1005 // makes +VerifyOpto report false-positives on progress. I 1006 // tried setting control edges on the x's to force them to 1007 // not combine, but the matching gets worried when it tries 1008 // to fold a StoreP and an AddP together (as part of an 1009 // address expression) and the AddP and StoreP have 1010 // different controls. 1011 if( !x->is_Load() && !x->is_DecodeN() ) _igvn._worklist.yank(x); 1012 } 1013 _igvn.remove_dead_node(n); 1014 } 1015 } 1016 } 1017 } 1018 1019 // Check for Opaque2's who's loop has disappeared - who's input is in the 1020 // same loop nest as their output. Remove 'em, they are no longer useful. 1021 if( n_op == Op_Opaque2 && 1022 n->in(1) != NULL && 1023 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) { 1024 _igvn.replace_node( n, n->in(1) ); 1025 } 1026 } 1027 1028 //------------------------------split_if_with_blocks--------------------------- 1029 // Check for aggressive application of 'split-if' optimization, 1030 // using basic block level info. 1031 void PhaseIdealLoop::split_if_with_blocks( VectorSet &visited, Node_Stack &nstack ) { 1032 Node *n = C->root(); 1033 visited.set(n->_idx); // first, mark node as visited 1034 // Do pre-visit work for root 1035 n = split_if_with_blocks_pre( n ); 1036 uint cnt = n->outcnt(); 1037 uint i = 0; 1038 while (true) { 1039 // Visit all children 1040 if (i < cnt) { 1041 Node* use = n->raw_out(i); 1042 ++i; 1043 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) { 1044 // Now do pre-visit work for this use 1045 use = split_if_with_blocks_pre( use ); 1046 nstack.push(n, i); // Save parent and next use's index. 1047 n = use; // Process all children of current use. 1048 cnt = use->outcnt(); 1049 i = 0; 1050 } 1051 } 1052 else { 1053 // All of n's children have been processed, complete post-processing. 1054 if (cnt != 0 && !n->is_Con()) { 1055 assert(has_node(n), "no dead nodes"); 1056 split_if_with_blocks_post( n ); 1057 } 1058 if (nstack.is_empty()) { 1059 // Finished all nodes on stack. 1060 break; 1061 } 1062 // Get saved parent node and next use's index. Visit the rest of uses. 1063 n = nstack.node(); 1064 cnt = n->outcnt(); 1065 i = nstack.index(); 1066 nstack.pop(); 1067 } 1068 } 1069 } 1070 1071 1072 //============================================================================= 1073 // 1074 // C L O N E A L O O P B O D Y 1075 // 1076 1077 //------------------------------clone_iff-------------------------------------- 1078 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1079 // "Nearly" because all Nodes have been cloned from the original in the loop, 1080 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1081 // through the Phi recursively, and return a Bool. 1082 BoolNode *PhaseIdealLoop::clone_iff( PhiNode *phi, IdealLoopTree *loop ) { 1083 1084 // Convert this Phi into a Phi merging Bools 1085 uint i; 1086 for( i = 1; i < phi->req(); i++ ) { 1087 Node *b = phi->in(i); 1088 if( b->is_Phi() ) { 1089 _igvn.hash_delete(phi); 1090 _igvn._worklist.push(phi); 1091 phi->set_req(i, clone_iff( b->as_Phi(), loop )); 1092 } else { 1093 assert( b->is_Bool(), "" ); 1094 } 1095 } 1096 1097 Node *sample_bool = phi->in(1); 1098 Node *sample_cmp = sample_bool->in(1); 1099 1100 // Make Phis to merge the Cmp's inputs. 1101 int size = phi->in(0)->req(); 1102 PhiNode *phi1 = new (C, size) PhiNode( phi->in(0), Type::TOP ); 1103 PhiNode *phi2 = new (C, size) PhiNode( phi->in(0), Type::TOP ); 1104 for( i = 1; i < phi->req(); i++ ) { 1105 Node *n1 = phi->in(i)->in(1)->in(1); 1106 Node *n2 = phi->in(i)->in(1)->in(2); 1107 phi1->set_req( i, n1 ); 1108 phi2->set_req( i, n2 ); 1109 phi1->set_type( phi1->type()->meet(n1->bottom_type()) ); 1110 phi2->set_type( phi2->type()->meet(n2->bottom_type()) ); 1111 } 1112 // See if these Phis have been made before. 1113 // Register with optimizer 1114 Node *hit1 = _igvn.hash_find_insert(phi1); 1115 if( hit1 ) { // Hit, toss just made Phi 1116 _igvn.remove_dead_node(phi1); // Remove new phi 1117 assert( hit1->is_Phi(), "" ); 1118 phi1 = (PhiNode*)hit1; // Use existing phi 1119 } else { // Miss 1120 _igvn.register_new_node_with_optimizer(phi1); 1121 } 1122 Node *hit2 = _igvn.hash_find_insert(phi2); 1123 if( hit2 ) { // Hit, toss just made Phi 1124 _igvn.remove_dead_node(phi2); // Remove new phi 1125 assert( hit2->is_Phi(), "" ); 1126 phi2 = (PhiNode*)hit2; // Use existing phi 1127 } else { // Miss 1128 _igvn.register_new_node_with_optimizer(phi2); 1129 } 1130 // Register Phis with loop/block info 1131 set_ctrl(phi1, phi->in(0)); 1132 set_ctrl(phi2, phi->in(0)); 1133 // Make a new Cmp 1134 Node *cmp = sample_cmp->clone(); 1135 cmp->set_req( 1, phi1 ); 1136 cmp->set_req( 2, phi2 ); 1137 _igvn.register_new_node_with_optimizer(cmp); 1138 set_ctrl(cmp, phi->in(0)); 1139 1140 // Make a new Bool 1141 Node *b = sample_bool->clone(); 1142 b->set_req(1,cmp); 1143 _igvn.register_new_node_with_optimizer(b); 1144 set_ctrl(b, phi->in(0)); 1145 1146 assert( b->is_Bool(), "" ); 1147 return (BoolNode*)b; 1148 } 1149 1150 //------------------------------clone_bool------------------------------------- 1151 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1152 // "Nearly" because all Nodes have been cloned from the original in the loop, 1153 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1154 // through the Phi recursively, and return a Bool. 1155 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) { 1156 uint i; 1157 // Convert this Phi into a Phi merging Bools 1158 for( i = 1; i < phi->req(); i++ ) { 1159 Node *b = phi->in(i); 1160 if( b->is_Phi() ) { 1161 _igvn.hash_delete(phi); 1162 _igvn._worklist.push(phi); 1163 phi->set_req(i, clone_bool( b->as_Phi(), loop )); 1164 } else { 1165 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" ); 1166 } 1167 } 1168 1169 Node *sample_cmp = phi->in(1); 1170 1171 // Make Phis to merge the Cmp's inputs. 1172 int size = phi->in(0)->req(); 1173 PhiNode *phi1 = new (C, size) PhiNode( phi->in(0), Type::TOP ); 1174 PhiNode *phi2 = new (C, size) PhiNode( phi->in(0), Type::TOP ); 1175 for( uint j = 1; j < phi->req(); j++ ) { 1176 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP 1177 Node *n1, *n2; 1178 if( cmp_top->is_Cmp() ) { 1179 n1 = cmp_top->in(1); 1180 n2 = cmp_top->in(2); 1181 } else { 1182 n1 = n2 = cmp_top; 1183 } 1184 phi1->set_req( j, n1 ); 1185 phi2->set_req( j, n2 ); 1186 phi1->set_type( phi1->type()->meet(n1->bottom_type()) ); 1187 phi2->set_type( phi2->type()->meet(n2->bottom_type()) ); 1188 } 1189 1190 // See if these Phis have been made before. 1191 // Register with optimizer 1192 Node *hit1 = _igvn.hash_find_insert(phi1); 1193 if( hit1 ) { // Hit, toss just made Phi 1194 _igvn.remove_dead_node(phi1); // Remove new phi 1195 assert( hit1->is_Phi(), "" ); 1196 phi1 = (PhiNode*)hit1; // Use existing phi 1197 } else { // Miss 1198 _igvn.register_new_node_with_optimizer(phi1); 1199 } 1200 Node *hit2 = _igvn.hash_find_insert(phi2); 1201 if( hit2 ) { // Hit, toss just made Phi 1202 _igvn.remove_dead_node(phi2); // Remove new phi 1203 assert( hit2->is_Phi(), "" ); 1204 phi2 = (PhiNode*)hit2; // Use existing phi 1205 } else { // Miss 1206 _igvn.register_new_node_with_optimizer(phi2); 1207 } 1208 // Register Phis with loop/block info 1209 set_ctrl(phi1, phi->in(0)); 1210 set_ctrl(phi2, phi->in(0)); 1211 // Make a new Cmp 1212 Node *cmp = sample_cmp->clone(); 1213 cmp->set_req( 1, phi1 ); 1214 cmp->set_req( 2, phi2 ); 1215 _igvn.register_new_node_with_optimizer(cmp); 1216 set_ctrl(cmp, phi->in(0)); 1217 1218 assert( cmp->is_Cmp(), "" ); 1219 return (CmpNode*)cmp; 1220 } 1221 1222 //------------------------------sink_use--------------------------------------- 1223 // If 'use' was in the loop-exit block, it now needs to be sunk 1224 // below the post-loop merge point. 1225 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) { 1226 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) { 1227 set_ctrl(use, post_loop); 1228 for (DUIterator j = use->outs(); use->has_out(j); j++) 1229 sink_use(use->out(j), post_loop); 1230 } 1231 } 1232 1233 //------------------------------clone_loop------------------------------------- 1234 // 1235 // C L O N E A L O O P B O D Y 1236 // 1237 // This is the basic building block of the loop optimizations. It clones an 1238 // entire loop body. It makes an old_new loop body mapping; with this mapping 1239 // you can find the new-loop equivalent to an old-loop node. All new-loop 1240 // nodes are exactly equal to their old-loop counterparts, all edges are the 1241 // same. All exits from the old-loop now have a RegionNode that merges the 1242 // equivalent new-loop path. This is true even for the normal "loop-exit" 1243 // condition. All uses of loop-invariant old-loop values now come from (one 1244 // or more) Phis that merge their new-loop equivalents. 1245 // 1246 // This operation leaves the graph in an illegal state: there are two valid 1247 // control edges coming from the loop pre-header to both loop bodies. I'll 1248 // definitely have to hack the graph after running this transform. 1249 // 1250 // From this building block I will further edit edges to perform loop peeling 1251 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc. 1252 // 1253 // Parameter side_by_size_idom: 1254 // When side_by_size_idom is NULL, the dominator tree is constructed for 1255 // the clone loop to dominate the original. Used in construction of 1256 // pre-main-post loop sequence. 1257 // When nonnull, the clone and original are side-by-side, both are 1258 // dominated by the side_by_side_idom node. Used in construction of 1259 // unswitched loops. 1260 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd, 1261 Node* side_by_side_idom) { 1262 1263 // Step 1: Clone the loop body. Make the old->new mapping. 1264 uint i; 1265 for( i = 0; i < loop->_body.size(); i++ ) { 1266 Node *old = loop->_body.at(i); 1267 Node *nnn = old->clone(); 1268 old_new.map( old->_idx, nnn ); 1269 _igvn.register_new_node_with_optimizer(nnn); 1270 } 1271 1272 1273 // Step 2: Fix the edges in the new body. If the old input is outside the 1274 // loop use it. If the old input is INside the loop, use the corresponding 1275 // new node instead. 1276 for( i = 0; i < loop->_body.size(); i++ ) { 1277 Node *old = loop->_body.at(i); 1278 Node *nnn = old_new[old->_idx]; 1279 // Fix CFG/Loop controlling the new node 1280 if (has_ctrl(old)) { 1281 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]); 1282 } else { 1283 set_loop(nnn, loop->_parent); 1284 if (old->outcnt() > 0) { 1285 set_idom( nnn, old_new[idom(old)->_idx], dd ); 1286 } 1287 } 1288 // Correct edges to the new node 1289 for( uint j = 0; j < nnn->req(); j++ ) { 1290 Node *n = nnn->in(j); 1291 if( n ) { 1292 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n ); 1293 if( loop->is_member( old_in_loop ) ) 1294 nnn->set_req(j, old_new[n->_idx]); 1295 } 1296 } 1297 _igvn.hash_find_insert(nnn); 1298 } 1299 Node *newhead = old_new[loop->_head->_idx]; 1300 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 1301 1302 1303 // Step 3: Now fix control uses. Loop varying control uses have already 1304 // been fixed up (as part of all input edges in Step 2). Loop invariant 1305 // control uses must be either an IfFalse or an IfTrue. Make a merge 1306 // point to merge the old and new IfFalse/IfTrue nodes; make the use 1307 // refer to this. 1308 ResourceArea *area = Thread::current()->resource_area(); 1309 Node_List worklist(area); 1310 uint new_counter = C->unique(); 1311 for( i = 0; i < loop->_body.size(); i++ ) { 1312 Node* old = loop->_body.at(i); 1313 if( !old->is_CFG() ) continue; 1314 Node* nnn = old_new[old->_idx]; 1315 1316 // Copy uses to a worklist, so I can munge the def-use info 1317 // with impunity. 1318 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1319 worklist.push(old->fast_out(j)); 1320 1321 while( worklist.size() ) { // Visit all uses 1322 Node *use = worklist.pop(); 1323 if (!has_node(use)) continue; // Ignore dead nodes 1324 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1325 if( !loop->is_member( use_loop ) && use->is_CFG() ) { 1326 // Both OLD and USE are CFG nodes here. 1327 assert( use->is_Proj(), "" ); 1328 1329 // Clone the loop exit control projection 1330 Node *newuse = use->clone(); 1331 newuse->set_req(0,nnn); 1332 _igvn.register_new_node_with_optimizer(newuse); 1333 set_loop(newuse, use_loop); 1334 set_idom(newuse, nnn, dom_depth(nnn) + 1 ); 1335 1336 // We need a Region to merge the exit from the peeled body and the 1337 // exit from the old loop body. 1338 RegionNode *r = new (C, 3) RegionNode(3); 1339 // Map the old use to the new merge point 1340 old_new.map( use->_idx, r ); 1341 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use)); 1342 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" ); 1343 1344 // The original user of 'use' uses 'r' instead. 1345 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) { 1346 Node* useuse = use->last_out(l); 1347 _igvn.hash_delete(useuse); 1348 _igvn._worklist.push(useuse); 1349 uint uses_found = 0; 1350 if( useuse->in(0) == use ) { 1351 useuse->set_req(0, r); 1352 uses_found++; 1353 if( useuse->is_CFG() ) { 1354 assert( dom_depth(useuse) > dd_r, "" ); 1355 set_idom(useuse, r, dom_depth(useuse)); 1356 } 1357 } 1358 for( uint k = 1; k < useuse->req(); k++ ) { 1359 if( useuse->in(k) == use ) { 1360 useuse->set_req(k, r); 1361 uses_found++; 1362 } 1363 } 1364 l -= uses_found; // we deleted 1 or more copies of this edge 1365 } 1366 1367 // Now finish up 'r' 1368 r->set_req( 1, newuse ); 1369 r->set_req( 2, use ); 1370 _igvn.register_new_node_with_optimizer(r); 1371 set_loop(r, use_loop); 1372 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r); 1373 } // End of if a loop-exit test 1374 } 1375 } 1376 1377 // Step 4: If loop-invariant use is not control, it must be dominated by a 1378 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region 1379 // there if needed. Make a Phi there merging old and new used values. 1380 Node_List *split_if_set = NULL; 1381 Node_List *split_bool_set = NULL; 1382 Node_List *split_cex_set = NULL; 1383 for( i = 0; i < loop->_body.size(); i++ ) { 1384 Node* old = loop->_body.at(i); 1385 Node* nnn = old_new[old->_idx]; 1386 // Copy uses to a worklist, so I can munge the def-use info 1387 // with impunity. 1388 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1389 worklist.push(old->fast_out(j)); 1390 1391 while( worklist.size() ) { 1392 Node *use = worklist.pop(); 1393 if (!has_node(use)) continue; // Ignore dead nodes 1394 if (use->in(0) == C->top()) continue; 1395 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1396 // Check for data-use outside of loop - at least one of OLD or USE 1397 // must not be a CFG node. 1398 if( !loop->is_member( use_loop ) && (!old->is_CFG() || !use->is_CFG())) { 1399 1400 // If the Data use is an IF, that means we have an IF outside of the 1401 // loop that is switching on a condition that is set inside of the 1402 // loop. Happens if people set a loop-exit flag; then test the flag 1403 // in the loop to break the loop, then test is again outside of the 1404 // loop to determine which way the loop exited. 1405 if( use->is_If() || use->is_CMove() ) { 1406 // Since this code is highly unlikely, we lazily build the worklist 1407 // of such Nodes to go split. 1408 if( !split_if_set ) 1409 split_if_set = new Node_List(area); 1410 split_if_set->push(use); 1411 } 1412 if( use->is_Bool() ) { 1413 if( !split_bool_set ) 1414 split_bool_set = new Node_List(area); 1415 split_bool_set->push(use); 1416 } 1417 if( use->Opcode() == Op_CreateEx ) { 1418 if( !split_cex_set ) 1419 split_cex_set = new Node_List(area); 1420 split_cex_set->push(use); 1421 } 1422 1423 1424 // Get "block" use is in 1425 uint idx = 0; 1426 while( use->in(idx) != old ) idx++; 1427 Node *prev = use->is_CFG() ? use : get_ctrl(use); 1428 assert( !loop->is_member( get_loop( prev ) ), "" ); 1429 Node *cfg = prev->_idx >= new_counter 1430 ? prev->in(2) 1431 : idom(prev); 1432 if( use->is_Phi() ) // Phi use is in prior block 1433 cfg = prev->in(idx); // NOT in block of Phi itself 1434 if (cfg->is_top()) { // Use is dead? 1435 _igvn.hash_delete(use); 1436 _igvn._worklist.push(use); 1437 use->set_req(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.hash_delete(use); 1488 _igvn._worklist.push(use); 1489 use->set_req(idx, phi); 1490 if( use->_idx >= new_counter ) { // If updating new phis 1491 // Not needed for correctness, but prevents a weak assert 1492 // in AddPNode from tripping (when we end up with different 1493 // base & derived Phis that will become the same after 1494 // IGVN does CSE). 1495 Node *hit = _igvn.hash_find_insert(use); 1496 if( hit ) // Go ahead and re-hash for hits. 1497 _igvn.replace_node( use, hit ); 1498 } 1499 1500 // If 'use' was in the loop-exit block, it now needs to be sunk 1501 // below the post-loop merge point. 1502 sink_use( use, prev ); 1503 } 1504 } 1505 } 1506 1507 // Check for IFs that need splitting/cloning. Happens if an IF outside of 1508 // the loop uses a condition set in the loop. The original IF probably 1509 // takes control from one or more OLD Regions (which in turn get from NEW 1510 // Regions). In any case, there will be a set of Phis for each merge point 1511 // from the IF up to where the original BOOL def exists the loop. 1512 if( split_if_set ) { 1513 while( split_if_set->size() ) { 1514 Node *iff = split_if_set->pop(); 1515 if( iff->in(1)->is_Phi() ) { 1516 BoolNode *b = clone_iff( iff->in(1)->as_Phi(), loop ); 1517 _igvn.hash_delete(iff); 1518 _igvn._worklist.push(iff); 1519 iff->set_req(1, b); 1520 } 1521 } 1522 } 1523 if( split_bool_set ) { 1524 while( split_bool_set->size() ) { 1525 Node *b = split_bool_set->pop(); 1526 Node *phi = b->in(1); 1527 assert( phi->is_Phi(), "" ); 1528 CmpNode *cmp = clone_bool( (PhiNode*)phi, loop ); 1529 _igvn.hash_delete(b); 1530 _igvn._worklist.push(b); 1531 b->set_req(1, cmp); 1532 } 1533 } 1534 if( split_cex_set ) { 1535 while( split_cex_set->size() ) { 1536 Node *b = split_cex_set->pop(); 1537 assert( b->in(0)->is_Region(), "" ); 1538 assert( b->in(1)->is_Phi(), "" ); 1539 assert( b->in(0)->in(0) == b->in(1)->in(0), "" ); 1540 split_up( b, b->in(0), NULL ); 1541 } 1542 } 1543 1544 } 1545 1546 1547 //---------------------- stride_of_possible_iv ------------------------------------- 1548 // Looks for an iff/bool/comp with one operand of the compare 1549 // being a cycle involving an add and a phi, 1550 // with an optional truncation (left-shift followed by a right-shift) 1551 // of the add. Returns zero if not an iv. 1552 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) { 1553 Node* trunc1 = NULL; 1554 Node* trunc2 = NULL; 1555 const TypeInt* ttype = NULL; 1556 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) { 1557 return 0; 1558 } 1559 BoolNode* bl = iff->in(1)->as_Bool(); 1560 Node* cmp = bl->in(1); 1561 if (!cmp || cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU) { 1562 return 0; 1563 } 1564 // Must have an invariant operand 1565 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) { 1566 return 0; 1567 } 1568 Node* add2 = NULL; 1569 Node* cmp1 = cmp->in(1); 1570 if (cmp1->is_Phi()) { 1571 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) ))) 1572 Node* phi = cmp1; 1573 for (uint i = 1; i < phi->req(); i++) { 1574 Node* in = phi->in(i); 1575 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in, 1576 &trunc1, &trunc2, &ttype); 1577 if (add && add->in(1) == phi) { 1578 add2 = add->in(2); 1579 break; 1580 } 1581 } 1582 } else { 1583 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) ))) 1584 Node* addtrunc = cmp1; 1585 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc, 1586 &trunc1, &trunc2, &ttype); 1587 if (add && add->in(1)->is_Phi()) { 1588 Node* phi = add->in(1); 1589 for (uint i = 1; i < phi->req(); i++) { 1590 if (phi->in(i) == addtrunc) { 1591 add2 = add->in(2); 1592 break; 1593 } 1594 } 1595 } 1596 } 1597 if (add2 != NULL) { 1598 const TypeInt* add2t = _igvn.type(add2)->is_int(); 1599 if (add2t->is_con()) { 1600 return add2t->get_con(); 1601 } 1602 } 1603 return 0; 1604 } 1605 1606 1607 //---------------------- stay_in_loop ------------------------------------- 1608 // Return the (unique) control output node that's in the loop (if it exists.) 1609 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) { 1610 Node* unique = NULL; 1611 if (!n) return NULL; 1612 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1613 Node* use = n->fast_out(i); 1614 if (!has_ctrl(use) && loop->is_member(get_loop(use))) { 1615 if (unique != NULL) { 1616 return NULL; 1617 } 1618 unique = use; 1619 } 1620 } 1621 return unique; 1622 } 1623 1624 //------------------------------ register_node ------------------------------------- 1625 // Utility to register node "n" with PhaseIdealLoop 1626 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) { 1627 _igvn.register_new_node_with_optimizer(n); 1628 loop->_body.push(n); 1629 if (n->is_CFG()) { 1630 set_loop(n, loop); 1631 set_idom(n, pred, ddepth); 1632 } else { 1633 set_ctrl(n, pred); 1634 } 1635 } 1636 1637 //------------------------------ proj_clone ------------------------------------- 1638 // Utility to create an if-projection 1639 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) { 1640 ProjNode* c = p->clone()->as_Proj(); 1641 c->set_req(0, iff); 1642 return c; 1643 } 1644 1645 //------------------------------ short_circuit_if ------------------------------------- 1646 // Force the iff control output to be the live_proj 1647 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) { 1648 int proj_con = live_proj->_con; 1649 assert(proj_con == 0 || proj_con == 1, "false or true projection"); 1650 Node *con = _igvn.intcon(proj_con); 1651 set_ctrl(con, C->root()); 1652 if (iff) { 1653 iff->set_req(1, con); 1654 } 1655 return con; 1656 } 1657 1658 //------------------------------ insert_if_before_proj ------------------------------------- 1659 // Insert a new if before an if projection (* - new node) 1660 // 1661 // before 1662 // if(test) 1663 // / \ 1664 // v v 1665 // other-proj proj (arg) 1666 // 1667 // after 1668 // if(test) 1669 // / \ 1670 // / v 1671 // | * proj-clone 1672 // v | 1673 // other-proj v 1674 // * new_if(relop(cmp[IU](left,right))) 1675 // / \ 1676 // v v 1677 // * new-proj proj 1678 // (returned) 1679 // 1680 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) { 1681 IfNode* iff = proj->in(0)->as_If(); 1682 IdealLoopTree *loop = get_loop(proj); 1683 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 1684 int ddepth = dom_depth(proj); 1685 1686 _igvn.hash_delete(iff); 1687 _igvn._worklist.push(iff); 1688 _igvn.hash_delete(proj); 1689 _igvn._worklist.push(proj); 1690 1691 proj->set_req(0, NULL); // temporary disconnect 1692 ProjNode* proj2 = proj_clone(proj, iff); 1693 register_node(proj2, loop, iff, ddepth); 1694 1695 Node* cmp = Signed ? (Node*) new (C,3)CmpINode(left, right) : (Node*) new (C,3)CmpUNode(left, right); 1696 register_node(cmp, loop, proj2, ddepth); 1697 1698 BoolNode* bol = new (C,2)BoolNode(cmp, relop); 1699 register_node(bol, loop, proj2, ddepth); 1700 1701 IfNode* new_if = new (C,2)IfNode(proj2, bol, iff->_prob, iff->_fcnt); 1702 register_node(new_if, loop, proj2, ddepth); 1703 1704 proj->set_req(0, new_if); // reattach 1705 set_idom(proj, new_if, ddepth); 1706 1707 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj(); 1708 register_node(new_exit, get_loop(other_proj), new_if, ddepth); 1709 1710 return new_exit; 1711 } 1712 1713 //------------------------------ insert_region_before_proj ------------------------------------- 1714 // Insert a region before an if projection (* - new node) 1715 // 1716 // before 1717 // if(test) 1718 // / | 1719 // v | 1720 // proj v 1721 // other-proj 1722 // 1723 // after 1724 // if(test) 1725 // / | 1726 // v | 1727 // * proj-clone v 1728 // | other-proj 1729 // v 1730 // * new-region 1731 // | 1732 // v 1733 // * dum_if 1734 // / \ 1735 // v \ 1736 // * dum-proj v 1737 // proj 1738 // 1739 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) { 1740 IfNode* iff = proj->in(0)->as_If(); 1741 IdealLoopTree *loop = get_loop(proj); 1742 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 1743 int ddepth = dom_depth(proj); 1744 1745 _igvn.hash_delete(iff); 1746 _igvn._worklist.push(iff); 1747 _igvn.hash_delete(proj); 1748 _igvn._worklist.push(proj); 1749 1750 proj->set_req(0, NULL); // temporary disconnect 1751 ProjNode* proj2 = proj_clone(proj, iff); 1752 register_node(proj2, loop, iff, ddepth); 1753 1754 RegionNode* reg = new (C,2)RegionNode(2); 1755 reg->set_req(1, proj2); 1756 register_node(reg, loop, iff, ddepth); 1757 1758 IfNode* dum_if = new (C,2)IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt); 1759 register_node(dum_if, loop, reg, ddepth); 1760 1761 proj->set_req(0, dum_if); // reattach 1762 set_idom(proj, dum_if, ddepth); 1763 1764 ProjNode* dum_proj = proj_clone(other_proj, dum_if); 1765 register_node(dum_proj, loop, dum_if, ddepth); 1766 1767 return reg; 1768 } 1769 1770 //------------------------------ insert_cmpi_loop_exit ------------------------------------- 1771 // Clone a signed compare loop exit from an unsigned compare and 1772 // insert it before the unsigned cmp on the stay-in-loop path. 1773 // All new nodes inserted in the dominator tree between the original 1774 // if and it's projections. The original if test is replaced with 1775 // a constant to force the stay-in-loop path. 1776 // 1777 // This is done to make sure that the original if and it's projections 1778 // still dominate the same set of control nodes, that the ctrl() relation 1779 // from data nodes to them is preserved, and that their loop nesting is 1780 // preserved. 1781 // 1782 // before 1783 // if(i <u limit) unsigned compare loop exit 1784 // / | 1785 // v v 1786 // exit-proj stay-in-loop-proj 1787 // 1788 // after 1789 // if(stay-in-loop-const) original if 1790 // / | 1791 // / v 1792 // / if(i < limit) new signed test 1793 // / / | 1794 // / / v 1795 // / / if(i <u limit) new cloned unsigned test 1796 // / / / | 1797 // v v v | 1798 // region | 1799 // | | 1800 // dum-if | 1801 // / | | 1802 // ether | | 1803 // v v 1804 // exit-proj stay-in-loop-proj 1805 // 1806 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) { 1807 const bool Signed = true; 1808 const bool Unsigned = false; 1809 1810 BoolNode* bol = if_cmpu->in(1)->as_Bool(); 1811 if (bol->_test._test != BoolTest::lt) return NULL; 1812 CmpNode* cmpu = bol->in(1)->as_Cmp(); 1813 if (cmpu->Opcode() != Op_CmpU) return NULL; 1814 int stride = stride_of_possible_iv(if_cmpu); 1815 if (stride == 0) return NULL; 1816 1817 ProjNode* lp_continue = stay_in_loop(if_cmpu, loop)->as_Proj(); 1818 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj(); 1819 1820 Node* limit = NULL; 1821 if (stride > 0) { 1822 limit = cmpu->in(2); 1823 } else { 1824 limit = _igvn.makecon(TypeInt::ZERO); 1825 set_ctrl(limit, C->root()); 1826 } 1827 // Create a new region on the exit path 1828 RegionNode* reg = insert_region_before_proj(lp_exit); 1829 1830 // Clone the if-cmpu-true-false using a signed compare 1831 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge; 1832 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue); 1833 reg->add_req(cmpi_exit); 1834 1835 // Clone the if-cmpu-true-false 1836 BoolTest::mask rel_u = bol->_test._test; 1837 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue); 1838 reg->add_req(cmpu_exit); 1839 1840 // Force original if to stay in loop. 1841 short_circuit_if(if_cmpu, lp_continue); 1842 1843 return cmpi_exit->in(0)->as_If(); 1844 } 1845 1846 //------------------------------ remove_cmpi_loop_exit ------------------------------------- 1847 // Remove a previously inserted signed compare loop exit. 1848 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) { 1849 Node* lp_proj = stay_in_loop(if_cmp, loop); 1850 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI && 1851 stay_in_loop(lp_proj, loop)->is_If() && 1852 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu"); 1853 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO); 1854 set_ctrl(con, C->root()); 1855 if_cmp->set_req(1, con); 1856 } 1857 1858 //------------------------------ scheduled_nodelist ------------------------------------- 1859 // Create a post order schedule of nodes that are in the 1860 // "member" set. The list is returned in "sched". 1861 // The first node in "sched" is the loop head, followed by 1862 // nodes which have no inputs in the "member" set, and then 1863 // followed by the nodes that have an immediate input dependence 1864 // on a node in "sched". 1865 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) { 1866 1867 assert(member.test(loop->_head->_idx), "loop head must be in member set"); 1868 Arena *a = Thread::current()->resource_area(); 1869 VectorSet visited(a); 1870 Node_Stack nstack(a, loop->_body.size()); 1871 1872 Node* n = loop->_head; // top of stack is cached in "n" 1873 uint idx = 0; 1874 visited.set(n->_idx); 1875 1876 // Initially push all with no inputs from within member set 1877 for(uint i = 0; i < loop->_body.size(); i++ ) { 1878 Node *elt = loop->_body.at(i); 1879 if (member.test(elt->_idx)) { 1880 bool found = false; 1881 for (uint j = 0; j < elt->req(); j++) { 1882 Node* def = elt->in(j); 1883 if (def && member.test(def->_idx) && def != elt) { 1884 found = true; 1885 break; 1886 } 1887 } 1888 if (!found && elt != loop->_head) { 1889 nstack.push(n, idx); 1890 n = elt; 1891 assert(!visited.test(n->_idx), "not seen yet"); 1892 visited.set(n->_idx); 1893 } 1894 } 1895 } 1896 1897 // traverse out's that are in the member set 1898 while (true) { 1899 if (idx < n->outcnt()) { 1900 Node* use = n->raw_out(idx); 1901 idx++; 1902 if (!visited.test_set(use->_idx)) { 1903 if (member.test(use->_idx)) { 1904 nstack.push(n, idx); 1905 n = use; 1906 idx = 0; 1907 } 1908 } 1909 } else { 1910 // All outputs processed 1911 sched.push(n); 1912 if (nstack.is_empty()) break; 1913 n = nstack.node(); 1914 idx = nstack.index(); 1915 nstack.pop(); 1916 } 1917 } 1918 } 1919 1920 1921 //------------------------------ has_use_in_set ------------------------------------- 1922 // Has a use in the vector set 1923 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) { 1924 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1925 Node* use = n->fast_out(j); 1926 if (vset.test(use->_idx)) { 1927 return true; 1928 } 1929 } 1930 return false; 1931 } 1932 1933 1934 //------------------------------ has_use_internal_to_set ------------------------------------- 1935 // Has use internal to the vector set (ie. not in a phi at the loop head) 1936 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) { 1937 Node* head = loop->_head; 1938 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1939 Node* use = n->fast_out(j); 1940 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) { 1941 return true; 1942 } 1943 } 1944 return false; 1945 } 1946 1947 1948 //------------------------------ clone_for_use_outside_loop ------------------------------------- 1949 // clone "n" for uses that are outside of loop 1950 void PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) { 1951 1952 assert(worklist.size() == 0, "should be empty"); 1953 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1954 Node* use = n->fast_out(j); 1955 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) { 1956 worklist.push(use); 1957 } 1958 } 1959 while( worklist.size() ) { 1960 Node *use = worklist.pop(); 1961 if (!has_node(use) || use->in(0) == C->top()) continue; 1962 uint j; 1963 for (j = 0; j < use->req(); j++) { 1964 if (use->in(j) == n) break; 1965 } 1966 assert(j < use->req(), "must be there"); 1967 1968 // clone "n" and insert it between the inputs of "n" and the use outside the loop 1969 Node* n_clone = n->clone(); 1970 _igvn.hash_delete(use); 1971 use->set_req(j, n_clone); 1972 _igvn._worklist.push(use); 1973 Node* use_c; 1974 if (!use->is_Phi()) { 1975 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0); 1976 } else { 1977 // Use in a phi is considered a use in the associated predecessor block 1978 use_c = use->in(0)->in(j); 1979 } 1980 set_ctrl(n_clone, use_c); 1981 assert(!loop->is_member(get_loop(use_c)), "should be outside loop"); 1982 get_loop(use_c)->_body.push(n_clone); 1983 _igvn.register_new_node_with_optimizer(n_clone); 1984 #if !defined(PRODUCT) 1985 if (TracePartialPeeling) { 1986 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx); 1987 } 1988 #endif 1989 } 1990 } 1991 1992 1993 //------------------------------ clone_for_special_use_inside_loop ------------------------------------- 1994 // clone "n" for special uses that are in the not_peeled region. 1995 // If these def-uses occur in separate blocks, the code generator 1996 // marks the method as not compilable. For example, if a "BoolNode" 1997 // is in a different basic block than the "IfNode" that uses it, then 1998 // the compilation is aborted in the code generator. 1999 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 2000 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) { 2001 if (n->is_Phi() || n->is_Load()) { 2002 return; 2003 } 2004 assert(worklist.size() == 0, "should be empty"); 2005 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2006 Node* use = n->fast_out(j); 2007 if ( not_peel.test(use->_idx) && 2008 (use->is_If() || use->is_CMove() || use->is_Bool()) && 2009 use->in(1) == n) { 2010 worklist.push(use); 2011 } 2012 } 2013 if (worklist.size() > 0) { 2014 // clone "n" and insert it between inputs of "n" and the use 2015 Node* n_clone = n->clone(); 2016 loop->_body.push(n_clone); 2017 _igvn.register_new_node_with_optimizer(n_clone); 2018 set_ctrl(n_clone, get_ctrl(n)); 2019 sink_list.push(n_clone); 2020 not_peel <<= n_clone->_idx; // add n_clone to not_peel set. 2021 #if !defined(PRODUCT) 2022 if (TracePartialPeeling) { 2023 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx); 2024 } 2025 #endif 2026 while( worklist.size() ) { 2027 Node *use = worklist.pop(); 2028 _igvn.hash_delete(use); 2029 _igvn._worklist.push(use); 2030 for (uint j = 1; j < use->req(); j++) { 2031 if (use->in(j) == n) { 2032 use->set_req(j, n_clone); 2033 } 2034 } 2035 } 2036 } 2037 } 2038 2039 2040 //------------------------------ insert_phi_for_loop ------------------------------------- 2041 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 2042 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) { 2043 Node *phi = PhiNode::make(lp, back_edge_val); 2044 phi->set_req(LoopNode::EntryControl, lp_entry_val); 2045 // Use existing phi if it already exists 2046 Node *hit = _igvn.hash_find_insert(phi); 2047 if( hit == NULL ) { 2048 _igvn.register_new_node_with_optimizer(phi); 2049 set_ctrl(phi, lp); 2050 } else { 2051 // Remove the new phi from the graph and use the hit 2052 _igvn.remove_dead_node(phi); 2053 phi = hit; 2054 } 2055 _igvn.hash_delete(use); 2056 _igvn._worklist.push(use); 2057 use->set_req(idx, phi); 2058 } 2059 2060 #ifdef ASSERT 2061 //------------------------------ is_valid_loop_partition ------------------------------------- 2062 // Validate the loop partition sets: peel and not_peel 2063 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, 2064 VectorSet& not_peel ) { 2065 uint i; 2066 // Check that peel_list entries are in the peel set 2067 for (i = 0; i < peel_list.size(); i++) { 2068 if (!peel.test(peel_list.at(i)->_idx)) { 2069 return false; 2070 } 2071 } 2072 // Check at loop members are in one of peel set or not_peel set 2073 for (i = 0; i < loop->_body.size(); i++ ) { 2074 Node *def = loop->_body.at(i); 2075 uint di = def->_idx; 2076 // Check that peel set elements are in peel_list 2077 if (peel.test(di)) { 2078 if (not_peel.test(di)) { 2079 return false; 2080 } 2081 // Must be in peel_list also 2082 bool found = false; 2083 for (uint j = 0; j < peel_list.size(); j++) { 2084 if (peel_list.at(j)->_idx == di) { 2085 found = true; 2086 break; 2087 } 2088 } 2089 if (!found) { 2090 return false; 2091 } 2092 } else if (not_peel.test(di)) { 2093 if (peel.test(di)) { 2094 return false; 2095 } 2096 } else { 2097 return false; 2098 } 2099 } 2100 return true; 2101 } 2102 2103 //------------------------------ is_valid_clone_loop_exit_use ------------------------------------- 2104 // Ensure a use outside of loop is of the right form 2105 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) { 2106 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2107 return (use->is_Phi() && 2108 use_c->is_Region() && use_c->req() == 3 && 2109 (use_c->in(exit_idx)->Opcode() == Op_IfTrue || 2110 use_c->in(exit_idx)->Opcode() == Op_IfFalse || 2111 use_c->in(exit_idx)->Opcode() == Op_JumpProj) && 2112 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) ); 2113 } 2114 2115 //------------------------------ is_valid_clone_loop_form ------------------------------------- 2116 // Ensure that all uses outside of loop are of the right form 2117 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 2118 uint orig_exit_idx, uint clone_exit_idx) { 2119 uint len = peel_list.size(); 2120 for (uint i = 0; i < len; i++) { 2121 Node *def = peel_list.at(i); 2122 2123 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 2124 Node *use = def->fast_out(j); 2125 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2126 if (!loop->is_member(get_loop(use_c))) { 2127 // use is not in the loop, check for correct structure 2128 if (use->in(0) == def) { 2129 // Okay 2130 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) { 2131 return false; 2132 } 2133 } 2134 } 2135 } 2136 return true; 2137 } 2138 #endif 2139 2140 //------------------------------ partial_peel ------------------------------------- 2141 // Partially peel (aka loop rotation) the top portion of a loop (called 2142 // the peel section below) by cloning it and placing one copy just before 2143 // the new loop head and the other copy at the bottom of the new loop. 2144 // 2145 // before after where it came from 2146 // 2147 // stmt1 stmt1 2148 // loop: stmt2 clone 2149 // stmt2 if condA goto exitA clone 2150 // if condA goto exitA new_loop: new 2151 // stmt3 stmt3 clone 2152 // if !condB goto loop if condB goto exitB clone 2153 // exitB: stmt2 orig 2154 // stmt4 if !condA goto new_loop orig 2155 // exitA: goto exitA 2156 // exitB: 2157 // stmt4 2158 // exitA: 2159 // 2160 // Step 1: find the cut point: an exit test on probable 2161 // induction variable. 2162 // Step 2: schedule (with cloning) operations in the peel 2163 // section that can be executed after the cut into 2164 // the section that is not peeled. This may need 2165 // to clone operations into exit blocks. For 2166 // instance, a reference to A[i] in the not-peel 2167 // section and a reference to B[i] in an exit block 2168 // may cause a left-shift of i by 2 to be placed 2169 // in the peel block. This step will clone the left 2170 // shift into the exit block and sink the left shift 2171 // from the peel to the not-peel section. 2172 // Step 3: clone the loop, retarget the control, and insert 2173 // phis for values that are live across the new loop 2174 // head. This is very dependent on the graph structure 2175 // from clone_loop. It creates region nodes for 2176 // exit control and associated phi nodes for values 2177 // flow out of the loop through that exit. The region 2178 // node is dominated by the clone's control projection. 2179 // So the clone's peel section is placed before the 2180 // new loop head, and the clone's not-peel section is 2181 // forms the top part of the new loop. The original 2182 // peel section forms the tail of the new loop. 2183 // Step 4: update the dominator tree and recompute the 2184 // dominator depth. 2185 // 2186 // orig 2187 // 2188 // stmt1 2189 // | 2190 // v 2191 // loop predicate 2192 // | 2193 // v 2194 // loop<----+ 2195 // | | 2196 // stmt2 | 2197 // | | 2198 // v | 2199 // ifA | 2200 // / | | 2201 // v v | 2202 // false true ^ <-- last_peel 2203 // / | | 2204 // / ===|==cut | 2205 // / stmt3 | <-- first_not_peel 2206 // / | | 2207 // | v | 2208 // v ifB | 2209 // exitA: / \ | 2210 // / \ | 2211 // v v | 2212 // false true | 2213 // / \ | 2214 // / ----+ 2215 // | 2216 // v 2217 // exitB: 2218 // stmt4 2219 // 2220 // 2221 // after clone loop 2222 // 2223 // stmt1 2224 // | 2225 // v 2226 // loop predicate 2227 // / \ 2228 // clone / \ orig 2229 // / \ 2230 // / \ 2231 // v v 2232 // +---->loop loop<----+ 2233 // | | | | 2234 // | stmt2 stmt2 | 2235 // | | | | 2236 // | v v | 2237 // | ifA ifA | 2238 // | | \ / | | 2239 // | v v v v | 2240 // ^ true false false true ^ <-- last_peel 2241 // | | ^ \ / | | 2242 // | cut==|== \ \ / ===|==cut | 2243 // | stmt3 \ \ / stmt3 | <-- first_not_peel 2244 // | | dom | | | | 2245 // | v \ 1v v2 v | 2246 // | ifB regionA ifB | 2247 // | / \ | / \ | 2248 // | / \ v / \ | 2249 // | v v exitA: v v | 2250 // | true false false true | 2251 // | / ^ \ / \ | 2252 // +---- \ \ / ----+ 2253 // dom \ / 2254 // \ 1v v2 2255 // regionB 2256 // | 2257 // v 2258 // exitB: 2259 // stmt4 2260 // 2261 // 2262 // after partial peel 2263 // 2264 // stmt1 2265 // | 2266 // v 2267 // loop predicate 2268 // / 2269 // clone / orig 2270 // / TOP 2271 // / \ 2272 // v v 2273 // TOP->loop loop----+ 2274 // | | | 2275 // stmt2 stmt2 | 2276 // | | | 2277 // v v | 2278 // ifA ifA | 2279 // | \ / | | 2280 // v v v v | 2281 // true false false true | <-- last_peel 2282 // | ^ \ / +------|---+ 2283 // +->newloop \ \ / === ==cut | | 2284 // | stmt3 \ \ / TOP | | 2285 // | | dom | | stmt3 | | <-- first_not_peel 2286 // | v \ 1v v2 v | | 2287 // | ifB regionA ifB ^ v 2288 // | / \ | / \ | | 2289 // | / \ v / \ | | 2290 // | v v exitA: v v | | 2291 // | true false false true | | 2292 // | / ^ \ / \ | | 2293 // | | \ \ / v | | 2294 // | | dom \ / TOP | | 2295 // | | \ 1v v2 | | 2296 // ^ v regionB | | 2297 // | | | | | 2298 // | | v ^ v 2299 // | | exitB: | | 2300 // | | stmt4 | | 2301 // | +------------>-----------------+ | 2302 // | | 2303 // +-----------------<---------------------+ 2304 // 2305 // 2306 // final graph 2307 // 2308 // stmt1 2309 // | 2310 // v 2311 // loop predicate 2312 // | 2313 // v 2314 // stmt2 clone 2315 // | 2316 // v 2317 // ........> ifA clone 2318 // : / | 2319 // dom / | 2320 // : v v 2321 // : false true 2322 // : | | 2323 // : | v 2324 // : | newloop<-----+ 2325 // : | | | 2326 // : | stmt3 clone | 2327 // : | | | 2328 // : | v | 2329 // : | ifB | 2330 // : | / \ | 2331 // : | v v | 2332 // : | false true | 2333 // : | | | | 2334 // : | v stmt2 | 2335 // : | exitB: | | 2336 // : | stmt4 v | 2337 // : | ifA orig | 2338 // : | / \ | 2339 // : | / \ | 2340 // : | v v | 2341 // : | false true | 2342 // : | / \ | 2343 // : v v -----+ 2344 // RegionA 2345 // | 2346 // v 2347 // exitA 2348 // 2349 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) { 2350 2351 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only"); 2352 if (!loop->_head->is_Loop()) { 2353 return false; } 2354 2355 LoopNode *head = loop->_head->as_Loop(); 2356 2357 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) { 2358 return false; 2359 } 2360 2361 // Check for complex exit control 2362 for(uint ii = 0; ii < loop->_body.size(); ii++ ) { 2363 Node *n = loop->_body.at(ii); 2364 int opc = n->Opcode(); 2365 if (n->is_Call() || 2366 opc == Op_Catch || 2367 opc == Op_CatchProj || 2368 opc == Op_Jump || 2369 opc == Op_JumpProj) { 2370 #if !defined(PRODUCT) 2371 if (TracePartialPeeling) { 2372 tty->print_cr("\nExit control too complex: lp: %d", head->_idx); 2373 } 2374 #endif 2375 return false; 2376 } 2377 } 2378 2379 int dd = dom_depth(head); 2380 2381 // Step 1: find cut point 2382 2383 // Walk up dominators to loop head looking for first loop exit 2384 // which is executed on every path thru loop. 2385 IfNode *peel_if = NULL; 2386 IfNode *peel_if_cmpu = NULL; 2387 2388 Node *iff = loop->tail(); 2389 while( iff != head ) { 2390 if( iff->is_If() ) { 2391 Node *ctrl = get_ctrl(iff->in(1)); 2392 if (ctrl->is_top()) return false; // Dead test on live IF. 2393 // If loop-varying exit-test, check for induction variable 2394 if( loop->is_member(get_loop(ctrl)) && 2395 loop->is_loop_exit(iff) && 2396 is_possible_iv_test(iff)) { 2397 Node* cmp = iff->in(1)->in(1); 2398 if (cmp->Opcode() == Op_CmpI) { 2399 peel_if = iff->as_If(); 2400 } else { 2401 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU"); 2402 peel_if_cmpu = iff->as_If(); 2403 } 2404 } 2405 } 2406 iff = idom(iff); 2407 } 2408 // Prefer signed compare over unsigned compare. 2409 IfNode* new_peel_if = NULL; 2410 if (peel_if == NULL) { 2411 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) { 2412 return false; // No peel point found 2413 } 2414 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop); 2415 if (new_peel_if == NULL) { 2416 return false; // No peel point found 2417 } 2418 peel_if = new_peel_if; 2419 } 2420 Node* last_peel = stay_in_loop(peel_if, loop); 2421 Node* first_not_peeled = stay_in_loop(last_peel, loop); 2422 if (first_not_peeled == NULL || first_not_peeled == head) { 2423 return false; 2424 } 2425 2426 #if !defined(PRODUCT) 2427 if (TraceLoopOpts) { 2428 tty->print("PartialPeel "); 2429 loop->dump_head(); 2430 } 2431 2432 if (TracePartialPeeling) { 2433 tty->print_cr("before partial peel one iteration"); 2434 Node_List wl; 2435 Node* t = head->in(2); 2436 while (true) { 2437 wl.push(t); 2438 if (t == head) break; 2439 t = idom(t); 2440 } 2441 while (wl.size() > 0) { 2442 Node* tt = wl.pop(); 2443 tt->dump(); 2444 if (tt == last_peel) tty->print_cr("-- cut --"); 2445 } 2446 } 2447 #endif 2448 ResourceArea *area = Thread::current()->resource_area(); 2449 VectorSet peel(area); 2450 VectorSet not_peel(area); 2451 Node_List peel_list(area); 2452 Node_List worklist(area); 2453 Node_List sink_list(area); 2454 2455 // Set of cfg nodes to peel are those that are executable from 2456 // the head through last_peel. 2457 assert(worklist.size() == 0, "should be empty"); 2458 worklist.push(head); 2459 peel.set(head->_idx); 2460 while (worklist.size() > 0) { 2461 Node *n = worklist.pop(); 2462 if (n != last_peel) { 2463 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2464 Node* use = n->fast_out(j); 2465 if (use->is_CFG() && 2466 loop->is_member(get_loop(use)) && 2467 !peel.test_set(use->_idx)) { 2468 worklist.push(use); 2469 } 2470 } 2471 } 2472 } 2473 2474 // Set of non-cfg nodes to peel are those that are control 2475 // dependent on the cfg nodes. 2476 uint i; 2477 for(i = 0; i < loop->_body.size(); i++ ) { 2478 Node *n = loop->_body.at(i); 2479 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n; 2480 if (peel.test(n_c->_idx)) { 2481 peel.set(n->_idx); 2482 } else { 2483 not_peel.set(n->_idx); 2484 } 2485 } 2486 2487 // Step 2: move operations from the peeled section down into the 2488 // not-peeled section 2489 2490 // Get a post order schedule of nodes in the peel region 2491 // Result in right-most operand. 2492 scheduled_nodelist(loop, peel, peel_list ); 2493 2494 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 2495 2496 // For future check for too many new phis 2497 uint old_phi_cnt = 0; 2498 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 2499 Node* use = head->fast_out(j); 2500 if (use->is_Phi()) old_phi_cnt++; 2501 } 2502 2503 #if !defined(PRODUCT) 2504 if (TracePartialPeeling) { 2505 tty->print_cr("\npeeled list"); 2506 } 2507 #endif 2508 2509 // Evacuate nodes in peel region into the not_peeled region if possible 2510 uint new_phi_cnt = 0; 2511 for (i = 0; i < peel_list.size();) { 2512 Node* n = peel_list.at(i); 2513 #if !defined(PRODUCT) 2514 if (TracePartialPeeling) n->dump(); 2515 #endif 2516 bool incr = true; 2517 if ( !n->is_CFG() ) { 2518 2519 if ( has_use_in_set(n, not_peel) ) { 2520 2521 // If not used internal to the peeled region, 2522 // move "n" from peeled to not_peeled region. 2523 2524 if ( !has_use_internal_to_set(n, peel, loop) ) { 2525 2526 // if not pinned and not a load (which maybe anti-dependent on a store) 2527 // and not a CMove (Matcher expects only bool->cmove). 2528 if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) { 2529 clone_for_use_outside_loop( loop, n, worklist ); 2530 2531 sink_list.push(n); 2532 peel >>= n->_idx; // delete n from peel set. 2533 not_peel <<= n->_idx; // add n to not_peel set. 2534 peel_list.remove(i); 2535 incr = false; 2536 #if !defined(PRODUCT) 2537 if (TracePartialPeeling) { 2538 tty->print_cr("sink to not_peeled region: %d newbb: %d", 2539 n->_idx, get_ctrl(n)->_idx); 2540 } 2541 #endif 2542 } 2543 } else { 2544 // Otherwise check for special def-use cases that span 2545 // the peel/not_peel boundary such as bool->if 2546 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist ); 2547 new_phi_cnt++; 2548 } 2549 } 2550 } 2551 if (incr) i++; 2552 } 2553 2554 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) { 2555 #if !defined(PRODUCT) 2556 if (TracePartialPeeling) { 2557 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c", 2558 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F'); 2559 } 2560 #endif 2561 if (new_peel_if != NULL) { 2562 remove_cmpi_loop_exit(new_peel_if, loop); 2563 } 2564 // Inhibit more partial peeling on this loop 2565 assert(!head->is_partial_peel_loop(), "not partial peeled"); 2566 head->mark_partial_peel_failed(); 2567 return false; 2568 } 2569 2570 // Step 3: clone loop, retarget control, and insert new phis 2571 2572 // Create new loop head for new phis and to hang 2573 // the nodes being moved (sinked) from the peel region. 2574 LoopNode* new_head = new (C, 3) LoopNode(last_peel, last_peel); 2575 new_head->set_unswitch_count(head->unswitch_count()); // Preserve 2576 _igvn.register_new_node_with_optimizer(new_head); 2577 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled"); 2578 first_not_peeled->set_req(0, new_head); 2579 set_loop(new_head, loop); 2580 loop->_body.push(new_head); 2581 not_peel.set(new_head->_idx); 2582 set_idom(new_head, last_peel, dom_depth(first_not_peeled)); 2583 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled)); 2584 2585 while (sink_list.size() > 0) { 2586 Node* n = sink_list.pop(); 2587 set_ctrl(n, new_head); 2588 } 2589 2590 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 2591 2592 clone_loop( loop, old_new, dd ); 2593 2594 const uint clone_exit_idx = 1; 2595 const uint orig_exit_idx = 2; 2596 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop"); 2597 2598 Node* head_clone = old_new[head->_idx]; 2599 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop(); 2600 Node* orig_tail_clone = head_clone->in(2); 2601 2602 // Add phi if "def" node is in peel set and "use" is not 2603 2604 for(i = 0; i < peel_list.size(); i++ ) { 2605 Node *def = peel_list.at(i); 2606 if (!def->is_CFG()) { 2607 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 2608 Node *use = def->fast_out(j); 2609 if (has_node(use) && use->in(0) != C->top() && 2610 (!peel.test(use->_idx) || 2611 (use->is_Phi() && use->in(0) == head)) ) { 2612 worklist.push(use); 2613 } 2614 } 2615 while( worklist.size() ) { 2616 Node *use = worklist.pop(); 2617 for (uint j = 1; j < use->req(); j++) { 2618 Node* n = use->in(j); 2619 if (n == def) { 2620 2621 // "def" is in peel set, "use" is not in peel set 2622 // or "use" is in the entry boundary (a phi) of the peel set 2623 2624 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use; 2625 2626 if ( loop->is_member(get_loop( use_c )) ) { 2627 // use is in loop 2628 if (old_new[use->_idx] != NULL) { // null for dead code 2629 Node* use_clone = old_new[use->_idx]; 2630 _igvn.hash_delete(use); 2631 use->set_req(j, C->top()); 2632 _igvn._worklist.push(use); 2633 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone ); 2634 } 2635 } else { 2636 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format"); 2637 // use is not in the loop, check if the live range includes the cut 2638 Node* lp_if = use_c->in(orig_exit_idx)->in(0); 2639 if (not_peel.test(lp_if->_idx)) { 2640 assert(j == orig_exit_idx, "use from original loop"); 2641 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone ); 2642 } 2643 } 2644 } 2645 } 2646 } 2647 } 2648 } 2649 2650 // Step 3b: retarget control 2651 2652 // Redirect control to the new loop head if a cloned node in 2653 // the not_peeled region has control that points into the peeled region. 2654 // This necessary because the cloned peeled region will be outside 2655 // the loop. 2656 // from to 2657 // cloned-peeled <---+ 2658 // new_head_clone: | <--+ 2659 // cloned-not_peeled in(0) in(0) 2660 // orig-peeled 2661 2662 for(i = 0; i < loop->_body.size(); i++ ) { 2663 Node *n = loop->_body.at(i); 2664 if (!n->is_CFG() && n->in(0) != NULL && 2665 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) { 2666 Node* n_clone = old_new[n->_idx]; 2667 _igvn.hash_delete(n_clone); 2668 n_clone->set_req(0, new_head_clone); 2669 _igvn._worklist.push(n_clone); 2670 } 2671 } 2672 2673 // Backedge of the surviving new_head (the clone) is original last_peel 2674 _igvn.hash_delete(new_head_clone); 2675 new_head_clone->set_req(LoopNode::LoopBackControl, last_peel); 2676 _igvn._worklist.push(new_head_clone); 2677 2678 // Cut first node in original not_peel set 2679 _igvn.hash_delete(new_head); 2680 new_head->set_req(LoopNode::EntryControl, C->top()); 2681 new_head->set_req(LoopNode::LoopBackControl, C->top()); 2682 _igvn._worklist.push(new_head); 2683 2684 // Copy head_clone back-branch info to original head 2685 // and remove original head's loop entry and 2686 // clone head's back-branch 2687 _igvn.hash_delete(head); 2688 _igvn.hash_delete(head_clone); 2689 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl)); 2690 head->set_req(LoopNode::LoopBackControl, C->top()); 2691 head_clone->set_req(LoopNode::LoopBackControl, C->top()); 2692 _igvn._worklist.push(head); 2693 _igvn._worklist.push(head_clone); 2694 2695 // Similarly modify the phis 2696 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) { 2697 Node* use = head->fast_out(k); 2698 if (use->is_Phi() && use->outcnt() > 0) { 2699 Node* use_clone = old_new[use->_idx]; 2700 _igvn.hash_delete(use); 2701 _igvn.hash_delete(use_clone); 2702 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl)); 2703 use->set_req(LoopNode::LoopBackControl, C->top()); 2704 use_clone->set_req(LoopNode::LoopBackControl, C->top()); 2705 _igvn._worklist.push(use); 2706 _igvn._worklist.push(use_clone); 2707 } 2708 } 2709 2710 // Step 4: update dominator tree and dominator depth 2711 2712 set_idom(head, orig_tail_clone, dd); 2713 recompute_dom_depth(); 2714 2715 // Inhibit more partial peeling on this loop 2716 new_head_clone->set_partial_peel_loop(); 2717 C->set_major_progress(); 2718 2719 #if !defined(PRODUCT) 2720 if (TracePartialPeeling) { 2721 tty->print_cr("\nafter partial peel one iteration"); 2722 Node_List wl(area); 2723 Node* t = last_peel; 2724 while (true) { 2725 wl.push(t); 2726 if (t == head_clone) break; 2727 t = idom(t); 2728 } 2729 while (wl.size() > 0) { 2730 Node* tt = wl.pop(); 2731 if (tt == head) tty->print_cr("orig head"); 2732 else if (tt == new_head_clone) tty->print_cr("new head"); 2733 else if (tt == head_clone) tty->print_cr("clone head"); 2734 tt->dump(); 2735 } 2736 } 2737 #endif 2738 return true; 2739 } 2740 2741 //------------------------------reorg_offsets---------------------------------- 2742 // Reorganize offset computations to lower register pressure. Mostly 2743 // prevent loop-fallout uses of the pre-incremented trip counter (which are 2744 // then alive with the post-incremented trip counter forcing an extra 2745 // register move) 2746 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) { 2747 // Perform it only for canonical counted loops. 2748 // Loop's shape could be messed up by iteration_split_impl. 2749 if (!loop->_head->is_CountedLoop()) 2750 return; 2751 if (!loop->_head->as_Loop()->is_valid_counted_loop()) 2752 return; 2753 2754 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 2755 CountedLoopEndNode *cle = cl->loopexit(); 2756 Node *exit = cle->proj_out(false); 2757 Node *phi = cl->phi(); 2758 2759 // Check for the special case of folks using the pre-incremented 2760 // trip-counter on the fall-out path (forces the pre-incremented 2761 // and post-incremented trip counter to be live at the same time). 2762 // Fix this by adjusting to use the post-increment trip counter. 2763 2764 bool progress = true; 2765 while (progress) { 2766 progress = false; 2767 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) { 2768 Node* use = phi->fast_out(i); // User of trip-counter 2769 if (!has_ctrl(use)) continue; 2770 Node *u_ctrl = get_ctrl(use); 2771 if (use->is_Phi()) { 2772 u_ctrl = NULL; 2773 for (uint j = 1; j < use->req(); j++) 2774 if (use->in(j) == phi) 2775 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j)); 2776 } 2777 IdealLoopTree *u_loop = get_loop(u_ctrl); 2778 // Look for loop-invariant use 2779 if (u_loop == loop) continue; 2780 if (loop->is_member(u_loop)) continue; 2781 // Check that use is live out the bottom. Assuming the trip-counter 2782 // update is right at the bottom, uses of of the loop middle are ok. 2783 if (dom_lca(exit, u_ctrl) != exit) continue; 2784 // Hit! Refactor use to use the post-incremented tripcounter. 2785 // Compute a post-increment tripcounter. 2786 Node *opaq = new (C, 2) Opaque2Node( C, cle->incr() ); 2787 register_new_node( opaq, u_ctrl ); 2788 Node *neg_stride = _igvn.intcon(-cle->stride_con()); 2789 set_ctrl(neg_stride, C->root()); 2790 Node *post = new (C, 3) AddINode( opaq, neg_stride); 2791 register_new_node( post, u_ctrl ); 2792 _igvn.hash_delete(use); 2793 _igvn._worklist.push(use); 2794 for (uint j = 1; j < use->req(); j++) { 2795 if (use->in(j) == phi) 2796 use->set_req(j, post); 2797 } 2798 // Since DU info changed, rerun loop 2799 progress = true; 2800 break; 2801 } 2802 } 2803 2804 }