1 /* 2 * Copyright (c) 1999, 2018, 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 "gc/shared/barrierSet.hpp" 27 #include "gc/shared/c2/barrierSetC2.hpp" 28 #include "memory/allocation.inline.hpp" 29 #include "memory/resourceArea.hpp" 30 #include "opto/addnode.hpp" 31 #include "opto/callnode.hpp" 32 #include "opto/castnode.hpp" 33 #include "opto/connode.hpp" 34 #include "opto/castnode.hpp" 35 #include "opto/divnode.hpp" 36 #include "opto/loopnode.hpp" 37 #include "opto/matcher.hpp" 38 #include "opto/mulnode.hpp" 39 #include "opto/movenode.hpp" 40 #include "opto/opaquenode.hpp" 41 #include "opto/rootnode.hpp" 42 #include "opto/subnode.hpp" 43 #include "utilities/macros.hpp" 44 #if INCLUDE_ZGC 45 #include "gc/z/c2/zBarrierSetC2.hpp" 46 #endif 47 48 //============================================================================= 49 //------------------------------split_thru_phi--------------------------------- 50 // Split Node 'n' through merge point if there is enough win. 51 Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) { 52 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) { 53 // ConvI2L may have type information on it which is unsafe to push up 54 // so disable this for now 55 return NULL; 56 } 57 58 // Splitting range check CastIIs through a loop induction Phi can 59 // cause new Phis to be created that are left unrelated to the loop 60 // induction Phi and prevent optimizations (vectorization) 61 if (n->Opcode() == Op_CastII && n->as_CastII()->has_range_check() && 62 region->is_CountedLoop() && n->in(1) == region->as_CountedLoop()->phi()) { 63 return NULL; 64 } 65 66 int wins = 0; 67 assert(!n->is_CFG(), ""); 68 assert(region->is_Region(), ""); 69 70 const Type* type = n->bottom_type(); 71 const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr(); 72 Node *phi; 73 if (t_oop != NULL && t_oop->is_known_instance_field()) { 74 int iid = t_oop->instance_id(); 75 int index = C->get_alias_index(t_oop); 76 int offset = t_oop->offset(); 77 phi = new PhiNode(region, type, NULL, iid, index, offset); 78 } else { 79 phi = PhiNode::make_blank(region, n); 80 } 81 uint old_unique = C->unique(); 82 for (uint i = 1; i < region->req(); i++) { 83 Node *x; 84 Node* the_clone = NULL; 85 if (region->in(i) == C->top()) { 86 x = C->top(); // Dead path? Use a dead data op 87 } else { 88 x = n->clone(); // Else clone up the data op 89 the_clone = x; // Remember for possible deletion. 90 // Alter data node to use pre-phi inputs 91 if (n->in(0) == region) 92 x->set_req( 0, region->in(i) ); 93 for (uint j = 1; j < n->req(); j++) { 94 Node *in = n->in(j); 95 if (in->is_Phi() && in->in(0) == region) 96 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone 97 } 98 } 99 // Check for a 'win' on some paths 100 const Type *t = x->Value(&_igvn); 101 102 bool singleton = t->singleton(); 103 104 // A TOP singleton indicates that there are no possible values incoming 105 // along a particular edge. In most cases, this is OK, and the Phi will 106 // be eliminated later in an Ideal call. However, we can't allow this to 107 // happen if the singleton occurs on loop entry, as the elimination of 108 // the PhiNode may cause the resulting node to migrate back to a previous 109 // loop iteration. 110 if (singleton && t == Type::TOP) { 111 // Is_Loop() == false does not confirm the absence of a loop (e.g., an 112 // irreducible loop may not be indicated by an affirmative is_Loop()); 113 // therefore, the only top we can split thru a phi is on a backedge of 114 // a loop. 115 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 116 } 117 118 if (singleton) { 119 wins++; 120 x = ((PhaseGVN&)_igvn).makecon(t); 121 } else { 122 // We now call Identity to try to simplify the cloned node. 123 // Note that some Identity methods call phase->type(this). 124 // Make sure that the type array is big enough for 125 // our new node, even though we may throw the node away. 126 // (Note: This tweaking with igvn only works because x is a new node.) 127 _igvn.set_type(x, t); 128 // If x is a TypeNode, capture any more-precise type permanently into Node 129 // otherwise it will be not updated during igvn->transform since 130 // igvn->type(x) is set to x->Value() already. 131 x->raise_bottom_type(t); 132 Node *y = x->Identity(&_igvn); 133 if (y != x) { 134 wins++; 135 x = y; 136 } else { 137 y = _igvn.hash_find(x); 138 if (y) { 139 wins++; 140 x = y; 141 } else { 142 // Else x is a new node we are keeping 143 // We do not need register_new_node_with_optimizer 144 // because set_type has already been called. 145 _igvn._worklist.push(x); 146 } 147 } 148 } 149 if (x != the_clone && the_clone != NULL) 150 _igvn.remove_dead_node(the_clone); 151 phi->set_req( i, x ); 152 } 153 // Too few wins? 154 if (wins <= policy) { 155 _igvn.remove_dead_node(phi); 156 return NULL; 157 } 158 159 // Record Phi 160 register_new_node( phi, region ); 161 162 for (uint i2 = 1; i2 < phi->req(); i2++) { 163 Node *x = phi->in(i2); 164 // If we commoned up the cloned 'x' with another existing Node, 165 // the existing Node picks up a new use. We need to make the 166 // existing Node occur higher up so it dominates its uses. 167 Node *old_ctrl; 168 IdealLoopTree *old_loop; 169 170 if (x->is_Con()) { 171 // Constant's control is always root. 172 set_ctrl(x, C->root()); 173 continue; 174 } 175 // The occasional new node 176 if (x->_idx >= old_unique) { // Found a new, unplaced node? 177 old_ctrl = NULL; 178 old_loop = NULL; // Not in any prior loop 179 } else { 180 old_ctrl = get_ctrl(x); 181 old_loop = get_loop(old_ctrl); // Get prior loop 182 } 183 // New late point must dominate new use 184 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2)); 185 if (new_ctrl == old_ctrl) // Nothing is changed 186 continue; 187 188 IdealLoopTree *new_loop = get_loop(new_ctrl); 189 190 // Don't move x into a loop if its uses are 191 // outside of loop. Otherwise x will be cloned 192 // for each use outside of this loop. 193 IdealLoopTree *use_loop = get_loop(region); 194 if (!new_loop->is_member(use_loop) && 195 (old_loop == NULL || !new_loop->is_member(old_loop))) { 196 // Take early control, later control will be recalculated 197 // during next iteration of loop optimizations. 198 new_ctrl = get_early_ctrl(x); 199 new_loop = get_loop(new_ctrl); 200 } 201 // Set new location 202 set_ctrl(x, new_ctrl); 203 // If changing loop bodies, see if we need to collect into new body 204 if (old_loop != new_loop) { 205 if (old_loop && !old_loop->_child) 206 old_loop->_body.yank(x); 207 if (!new_loop->_child) 208 new_loop->_body.push(x); // Collect body info 209 } 210 } 211 212 return phi; 213 } 214 215 //------------------------------dominated_by------------------------------------ 216 // Replace the dominated test with an obvious true or false. Place it on the 217 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the 218 // live path up to the dominating control. 219 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff, bool flip, bool exclude_loop_predicate ) { 220 if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); } 221 222 // prevdom is the dominating projection of the dominating test. 223 assert( iff->is_If(), "" ); 224 assert(iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd || iff->Opcode() == Op_RangeCheck, "Check this code when new subtype is added"); 225 int pop = prevdom->Opcode(); 226 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" ); 227 if (flip) { 228 if (pop == Op_IfTrue) 229 pop = Op_IfFalse; 230 else 231 pop = Op_IfTrue; 232 } 233 // 'con' is set to true or false to kill the dominated test. 234 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO); 235 set_ctrl(con, C->root()); // Constant gets a new use 236 // Hack the dominated test 237 _igvn.replace_input_of(iff, 1, con); 238 239 // If I dont have a reachable TRUE and FALSE path following the IfNode then 240 // I can assume this path reaches an infinite loop. In this case it's not 241 // important to optimize the data Nodes - either the whole compilation will 242 // be tossed or this path (and all data Nodes) will go dead. 243 if (iff->outcnt() != 2) return; 244 245 // Make control-dependent data Nodes on the live path (path that will remain 246 // once the dominated IF is removed) become control-dependent on the 247 // dominating projection. 248 Node* dp = iff->as_If()->proj_out_or_null(pop == Op_IfTrue); 249 250 // Loop predicates may have depending checks which should not 251 // be skipped. For example, range check predicate has two checks 252 // for lower and upper bounds. 253 if (dp == NULL) 254 return; 255 256 ProjNode* dp_proj = dp->as_Proj(); 257 ProjNode* unc_proj = iff->as_If()->proj_out(1 - dp_proj->_con)->as_Proj(); 258 if (exclude_loop_predicate && 259 (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) != NULL || 260 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check) != NULL)) { 261 // If this is a range check (IfNode::is_range_check), do not 262 // reorder because Compile::allow_range_check_smearing might have 263 // changed the check. 264 return; // Let IGVN transformation change control dependence. 265 } 266 267 IdealLoopTree *old_loop = get_loop(dp); 268 269 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { 270 Node* cd = dp->fast_out(i); // Control-dependent node 271 if (cd->depends_only_on_test()) { 272 assert(cd->in(0) == dp, ""); 273 _igvn.replace_input_of(cd, 0, prevdom); 274 set_early_ctrl(cd); 275 IdealLoopTree *new_loop = get_loop(get_ctrl(cd)); 276 if (old_loop != new_loop) { 277 if (!old_loop->_child) old_loop->_body.yank(cd); 278 if (!new_loop->_child) new_loop->_body.push(cd); 279 } 280 --i; 281 --imax; 282 } 283 } 284 } 285 286 //------------------------------has_local_phi_input---------------------------- 287 // Return TRUE if 'n' has Phi inputs from its local block and no other 288 // block-local inputs (all non-local-phi inputs come from earlier blocks) 289 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) { 290 Node *n_ctrl = get_ctrl(n); 291 // See if some inputs come from a Phi in this block, or from before 292 // this block. 293 uint i; 294 for( i = 1; i < n->req(); i++ ) { 295 Node *phi = n->in(i); 296 if( phi->is_Phi() && phi->in(0) == n_ctrl ) 297 break; 298 } 299 if( i >= n->req() ) 300 return NULL; // No Phi inputs; nowhere to clone thru 301 302 // Check for inputs created between 'n' and the Phi input. These 303 // must split as well; they have already been given the chance 304 // (courtesy of a post-order visit) and since they did not we must 305 // recover the 'cost' of splitting them by being very profitable 306 // when splitting 'n'. Since this is unlikely we simply give up. 307 for( i = 1; i < n->req(); i++ ) { 308 Node *m = n->in(i); 309 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) { 310 // We allow the special case of AddP's with no local inputs. 311 // This allows us to split-up address expressions. 312 if (m->is_AddP() && 313 get_ctrl(m->in(2)) != n_ctrl && 314 get_ctrl(m->in(3)) != n_ctrl) { 315 // Move the AddP up to dominating point 316 Node* c = find_non_split_ctrl(idom(n_ctrl)); 317 if (c->is_OuterStripMinedLoop()) { 318 c->as_Loop()->verify_strip_mined(1); 319 c = c->in(LoopNode::EntryControl); 320 } 321 set_ctrl_and_loop(m, c); 322 continue; 323 } 324 return NULL; 325 } 326 assert(m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control"); 327 } 328 329 return n_ctrl; 330 } 331 332 //------------------------------remix_address_expressions---------------------- 333 // Rework addressing expressions to get the most loop-invariant stuff 334 // moved out. We'd like to do all associative operators, but it's especially 335 // important (common) to do address expressions. 336 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) { 337 if (!has_ctrl(n)) return NULL; 338 Node *n_ctrl = get_ctrl(n); 339 IdealLoopTree *n_loop = get_loop(n_ctrl); 340 341 // See if 'n' mixes loop-varying and loop-invariant inputs and 342 // itself is loop-varying. 343 344 // Only interested in binary ops (and AddP) 345 if( n->req() < 3 || n->req() > 4 ) return NULL; 346 347 Node *n1_ctrl = get_ctrl(n->in( 1)); 348 Node *n2_ctrl = get_ctrl(n->in( 2)); 349 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3)); 350 IdealLoopTree *n1_loop = get_loop( n1_ctrl ); 351 IdealLoopTree *n2_loop = get_loop( n2_ctrl ); 352 IdealLoopTree *n3_loop = get_loop( n3_ctrl ); 353 354 // Does one of my inputs spin in a tighter loop than self? 355 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) || 356 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) || 357 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) ) 358 return NULL; // Leave well enough alone 359 360 // Is at least one of my inputs loop-invariant? 361 if( n1_loop == n_loop && 362 n2_loop == n_loop && 363 n3_loop == n_loop ) 364 return NULL; // No loop-invariant inputs 365 366 367 int n_op = n->Opcode(); 368 369 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2). 370 if( n_op == Op_LShiftI ) { 371 // Scale is loop invariant 372 Node *scale = n->in(2); 373 Node *scale_ctrl = get_ctrl(scale); 374 IdealLoopTree *scale_loop = get_loop(scale_ctrl ); 375 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) ) 376 return NULL; 377 const TypeInt *scale_t = scale->bottom_type()->isa_int(); 378 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 ) 379 return NULL; // Dont bother with byte/short masking 380 // Add must vary with loop (else shift would be loop-invariant) 381 Node *add = n->in(1); 382 Node *add_ctrl = get_ctrl(add); 383 IdealLoopTree *add_loop = get_loop(add_ctrl); 384 //assert( n_loop == add_loop, "" ); 385 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops 386 387 // Convert I-V into I+ (0-V); same for V-I 388 if( add->Opcode() == Op_SubI && 389 _igvn.type( add->in(1) ) != TypeInt::ZERO ) { 390 Node *zero = _igvn.intcon(0); 391 set_ctrl(zero, C->root()); 392 Node *neg = new SubINode( _igvn.intcon(0), add->in(2) ); 393 register_new_node( neg, get_ctrl(add->in(2) ) ); 394 add = new AddINode( add->in(1), neg ); 395 register_new_node( add, add_ctrl ); 396 } 397 if( add->Opcode() != Op_AddI ) return NULL; 398 // See if one add input is loop invariant 399 Node *add_var = add->in(1); 400 Node *add_var_ctrl = get_ctrl(add_var); 401 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 402 Node *add_invar = add->in(2); 403 Node *add_invar_ctrl = get_ctrl(add_invar); 404 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl ); 405 if( add_var_loop == n_loop ) { 406 } else if( add_invar_loop == n_loop ) { 407 // Swap to find the invariant part 408 add_invar = add_var; 409 add_invar_ctrl = add_var_ctrl; 410 add_invar_loop = add_var_loop; 411 add_var = add->in(2); 412 Node *add_var_ctrl = get_ctrl(add_var); 413 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 414 } else // Else neither input is loop invariant 415 return NULL; 416 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) ) 417 return NULL; // No invariant part of the add? 418 419 // Yes! Reshape address expression! 420 Node *inv_scale = new LShiftINode( add_invar, scale ); 421 Node *inv_scale_ctrl = 422 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ? 423 add_invar_ctrl : scale_ctrl; 424 register_new_node( inv_scale, inv_scale_ctrl ); 425 Node *var_scale = new LShiftINode( add_var, scale ); 426 register_new_node( var_scale, n_ctrl ); 427 Node *var_add = new AddINode( var_scale, inv_scale ); 428 register_new_node( var_add, n_ctrl ); 429 _igvn.replace_node( n, var_add ); 430 return var_add; 431 } 432 433 // Replace (I+V) with (V+I) 434 if( n_op == Op_AddI || 435 n_op == Op_AddL || 436 n_op == Op_AddF || 437 n_op == Op_AddD || 438 n_op == Op_MulI || 439 n_op == Op_MulL || 440 n_op == Op_MulF || 441 n_op == Op_MulD ) { 442 if( n2_loop == n_loop ) { 443 assert( n1_loop != n_loop, "" ); 444 n->swap_edges(1, 2); 445 } 446 } 447 448 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V), 449 // but not if I2 is a constant. 450 if( n_op == Op_AddP ) { 451 if( n2_loop == n_loop && n3_loop != n_loop ) { 452 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) { 453 Node *n22_ctrl = get_ctrl(n->in(2)->in(2)); 454 Node *n23_ctrl = get_ctrl(n->in(2)->in(3)); 455 IdealLoopTree *n22loop = get_loop( n22_ctrl ); 456 IdealLoopTree *n23_loop = get_loop( n23_ctrl ); 457 if( n22loop != n_loop && n22loop->is_member(n_loop) && 458 n23_loop == n_loop ) { 459 Node *add1 = new AddPNode( n->in(1), n->in(2)->in(2), n->in(3) ); 460 // Stuff new AddP in the loop preheader 461 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) ); 462 Node *add2 = new AddPNode( n->in(1), add1, n->in(2)->in(3) ); 463 register_new_node( add2, n_ctrl ); 464 _igvn.replace_node( n, add2 ); 465 return add2; 466 } 467 } 468 } 469 470 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V) 471 if (n2_loop != n_loop && n3_loop == n_loop) { 472 if (n->in(3)->Opcode() == Op_AddX) { 473 Node *V = n->in(3)->in(1); 474 Node *I = n->in(3)->in(2); 475 if (is_member(n_loop,get_ctrl(V))) { 476 } else { 477 Node *tmp = V; V = I; I = tmp; 478 } 479 if (!is_member(n_loop,get_ctrl(I))) { 480 Node *add1 = new AddPNode(n->in(1), n->in(2), I); 481 // Stuff new AddP in the loop preheader 482 register_new_node(add1, n_loop->_head->in(LoopNode::EntryControl)); 483 Node *add2 = new AddPNode(n->in(1), add1, V); 484 register_new_node(add2, n_ctrl); 485 _igvn.replace_node(n, add2); 486 return add2; 487 } 488 } 489 } 490 } 491 492 return NULL; 493 } 494 495 //------------------------------conditional_move------------------------------- 496 // Attempt to replace a Phi with a conditional move. We have some pretty 497 // strict profitability requirements. All Phis at the merge point must 498 // be converted, so we can remove the control flow. We need to limit the 499 // number of c-moves to a small handful. All code that was in the side-arms 500 // of the CFG diamond is now speculatively executed. This code has to be 501 // "cheap enough". We are pretty much limited to CFG diamonds that merge 502 // 1 or 2 items with a total of 1 or 2 ops executed speculatively. 503 Node *PhaseIdealLoop::conditional_move( Node *region ) { 504 505 assert(region->is_Region(), "sanity check"); 506 if (region->req() != 3) return NULL; 507 508 // Check for CFG diamond 509 Node *lp = region->in(1); 510 Node *rp = region->in(2); 511 if (!lp || !rp) return NULL; 512 Node *lp_c = lp->in(0); 513 if (lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If()) return NULL; 514 IfNode *iff = lp_c->as_If(); 515 516 // Check for ops pinned in an arm of the diamond. 517 // Can't remove the control flow in this case 518 if (lp->outcnt() > 1) return NULL; 519 if (rp->outcnt() > 1) return NULL; 520 521 IdealLoopTree* r_loop = get_loop(region); 522 assert(r_loop == get_loop(iff), "sanity"); 523 // Always convert to CMOVE if all results are used only outside this loop. 524 bool used_inside_loop = (r_loop == _ltree_root); 525 526 // Check profitability 527 int cost = 0; 528 int phis = 0; 529 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 530 Node *out = region->fast_out(i); 531 if (!out->is_Phi()) continue; // Ignore other control edges, etc 532 phis++; 533 PhiNode* phi = out->as_Phi(); 534 BasicType bt = phi->type()->basic_type(); 535 switch (bt) { 536 case T_DOUBLE: 537 case T_FLOAT: 538 if (C->use_cmove()) { 539 continue; //TODO: maybe we want to add some cost 540 } 541 cost += Matcher::float_cmove_cost(); // Could be very expensive 542 break; 543 case T_LONG: { 544 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's 545 } 546 case T_INT: // These all CMOV fine 547 case T_ADDRESS: { // (RawPtr) 548 cost++; 549 break; 550 } 551 case T_NARROWOOP: // Fall through 552 case T_OBJECT: { // Base oops are OK, but not derived oops 553 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr(); 554 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a 555 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus 556 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we 557 // have a Phi for the base here that we convert to a CMOVE all is well 558 // and good. But if the base is dead, we'll not make a CMOVE. Later 559 // the allocator will have to produce a base by creating a CMOVE of the 560 // relevant bases. This puts the allocator in the business of 561 // manufacturing expensive instructions, generally a bad plan. 562 // Just Say No to Conditionally-Moved Derived Pointers. 563 if (tp && tp->offset() != 0) 564 return NULL; 565 cost++; 566 break; 567 } 568 default: 569 return NULL; // In particular, can't do memory or I/O 570 } 571 // Add in cost any speculative ops 572 for (uint j = 1; j < region->req(); j++) { 573 Node *proj = region->in(j); 574 Node *inp = phi->in(j); 575 if (get_ctrl(inp) == proj) { // Found local op 576 cost++; 577 // Check for a chain of dependent ops; these will all become 578 // speculative in a CMOV. 579 for (uint k = 1; k < inp->req(); k++) 580 if (get_ctrl(inp->in(k)) == proj) 581 cost += ConditionalMoveLimit; // Too much speculative goo 582 } 583 } 584 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode. 585 // This will likely Split-If, a higher-payoff operation. 586 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) { 587 Node* use = phi->fast_out(k); 588 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr()) 589 cost += ConditionalMoveLimit; 590 // Is there a use inside the loop? 591 // Note: check only basic types since CMoveP is pinned. 592 if (!used_inside_loop && is_java_primitive(bt)) { 593 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use); 594 if (r_loop == u_loop || r_loop->is_member(u_loop)) { 595 used_inside_loop = true; 596 } 597 } 598 } 599 }//for 600 Node* bol = iff->in(1); 601 assert(bol->Opcode() == Op_Bool, ""); 602 int cmp_op = bol->in(1)->Opcode(); 603 // It is expensive to generate flags from a float compare. 604 // Avoid duplicated float compare. 605 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL; 606 607 float infrequent_prob = PROB_UNLIKELY_MAG(3); 608 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop. 609 if (used_inside_loop) { 610 if (cost >= ConditionalMoveLimit) return NULL; // Too much goo 611 612 // BlockLayoutByFrequency optimization moves infrequent branch 613 // from hot path. No point in CMOV'ing in such case (110 is used 614 // instead of 100 to take into account not exactness of float value). 615 if (BlockLayoutByFrequency) { 616 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f); 617 } 618 } 619 // Check for highly predictable branch. No point in CMOV'ing if 620 // we are going to predict accurately all the time. 621 if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) { 622 //keep going 623 } else if (iff->_prob < infrequent_prob || 624 iff->_prob > (1.0f - infrequent_prob)) 625 return NULL; 626 627 // -------------- 628 // Now replace all Phis with CMOV's 629 Node *cmov_ctrl = iff->in(0); 630 uint flip = (lp->Opcode() == Op_IfTrue); 631 Node_List wq; 632 while (1) { 633 PhiNode* phi = NULL; 634 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 635 Node *out = region->fast_out(i); 636 if (out->is_Phi()) { 637 phi = out->as_Phi(); 638 break; 639 } 640 } 641 if (phi == NULL) break; 642 if (PrintOpto && VerifyLoopOptimizations) { tty->print_cr("CMOV"); } 643 // Move speculative ops 644 wq.push(phi); 645 while (wq.size() > 0) { 646 Node *n = wq.pop(); 647 for (uint j = 1; j < n->req(); j++) { 648 Node* m = n->in(j); 649 if (m != NULL && !is_dominator(get_ctrl(m), cmov_ctrl)) { 650 #ifndef PRODUCT 651 if (PrintOpto && VerifyLoopOptimizations) { 652 tty->print(" speculate: "); 653 m->dump(); 654 } 655 #endif 656 set_ctrl(m, cmov_ctrl); 657 wq.push(m); 658 } 659 } 660 } 661 Node *cmov = CMoveNode::make(cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi)); 662 register_new_node( cmov, cmov_ctrl ); 663 _igvn.replace_node( phi, cmov ); 664 #ifndef PRODUCT 665 if (TraceLoopOpts) { 666 tty->print("CMOV "); 667 r_loop->dump_head(); 668 if (Verbose) { 669 bol->in(1)->dump(1); 670 cmov->dump(1); 671 } 672 } 673 if (VerifyLoopOptimizations) verify(); 674 #endif 675 } 676 677 // The useless CFG diamond will fold up later; see the optimization in 678 // RegionNode::Ideal. 679 _igvn._worklist.push(region); 680 681 return iff->in(1); 682 } 683 684 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) { 685 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 686 Node* u = m->fast_out(i); 687 if (u->is_CFG()) { 688 if (u->Opcode() == Op_NeverBranch) { 689 u = ((NeverBranchNode*)u)->proj_out(0); 690 enqueue_cfg_uses(u, wq); 691 } else { 692 wq.push(u); 693 } 694 } 695 } 696 } 697 698 // Try moving a store out of a loop, right before the loop 699 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) { 700 // Store has to be first in the loop body 701 IdealLoopTree *n_loop = get_loop(n_ctrl); 702 if (n->is_Store() && n_loop != _ltree_root && 703 n_loop->is_loop() && n_loop->_head->is_Loop() && 704 n->in(0) != NULL) { 705 Node* address = n->in(MemNode::Address); 706 Node* value = n->in(MemNode::ValueIn); 707 Node* mem = n->in(MemNode::Memory); 708 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 709 IdealLoopTree* value_loop = get_loop(get_ctrl(value)); 710 711 // - address and value must be loop invariant 712 // - memory must be a memory Phi for the loop 713 // - Store must be the only store on this memory slice in the 714 // loop: if there's another store following this one then value 715 // written at iteration i by the second store could be overwritten 716 // at iteration i+n by the first store: it's not safe to move the 717 // first store out of the loop 718 // - nothing must observe the memory Phi: it guarantees no read 719 // before the store, we are also guaranteed the store post 720 // dominates the loop head (ignoring a possible early 721 // exit). Otherwise there would be extra Phi involved between the 722 // loop's Phi and the store. 723 // - there must be no early exit from the loop before the Store 724 // (such an exit most of the time would be an extra use of the 725 // memory Phi but sometimes is a bottom memory Phi that takes the 726 // store as input). 727 728 if (!n_loop->is_member(address_loop) && 729 !n_loop->is_member(value_loop) && 730 mem->is_Phi() && mem->in(0) == n_loop->_head && 731 mem->outcnt() == 1 && 732 mem->in(LoopNode::LoopBackControl) == n) { 733 734 assert(n_loop->_tail != NULL, "need a tail"); 735 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop"); 736 737 // Verify that there's no early exit of the loop before the store. 738 bool ctrl_ok = false; 739 { 740 // Follow control from loop head until n, we exit the loop or 741 // we reach the tail 742 ResourceMark rm; 743 Unique_Node_List wq; 744 wq.push(n_loop->_head); 745 746 for (uint next = 0; next < wq.size(); ++next) { 747 Node *m = wq.at(next); 748 if (m == n->in(0)) { 749 ctrl_ok = true; 750 continue; 751 } 752 assert(!has_ctrl(m), "should be CFG"); 753 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) { 754 ctrl_ok = false; 755 break; 756 } 757 enqueue_cfg_uses(m, wq); 758 if (wq.size() > 10) { 759 ctrl_ok = false; 760 break; 761 } 762 } 763 } 764 if (ctrl_ok) { 765 // move the Store 766 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem); 767 _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl)); 768 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl)); 769 // Disconnect the phi now. An empty phi can confuse other 770 // optimizations in this pass of loop opts. 771 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl)); 772 n_loop->_body.yank(mem); 773 774 set_ctrl_and_loop(n, n->in(0)); 775 776 return n; 777 } 778 } 779 } 780 return NULL; 781 } 782 783 // Try moving a store out of a loop, right after the loop 784 void PhaseIdealLoop::try_move_store_after_loop(Node* n) { 785 if (n->is_Store() && n->in(0) != NULL) { 786 Node *n_ctrl = get_ctrl(n); 787 IdealLoopTree *n_loop = get_loop(n_ctrl); 788 // Store must be in a loop 789 if (n_loop != _ltree_root && !n_loop->_irreducible) { 790 Node* address = n->in(MemNode::Address); 791 Node* value = n->in(MemNode::ValueIn); 792 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 793 // address must be loop invariant 794 if (!n_loop->is_member(address_loop)) { 795 // Store must be last on this memory slice in the loop and 796 // nothing in the loop must observe it 797 Node* phi = NULL; 798 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 799 Node* u = n->fast_out(i); 800 if (has_ctrl(u)) { // control use? 801 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 802 if (!n_loop->is_member(u_loop)) { 803 continue; 804 } 805 if (u->is_Phi() && u->in(0) == n_loop->_head) { 806 assert(_igvn.type(u) == Type::MEMORY, "bad phi"); 807 // multiple phis on the same slice are possible 808 if (phi != NULL) { 809 return; 810 } 811 phi = u; 812 continue; 813 } 814 } 815 return; 816 } 817 if (phi != NULL) { 818 // Nothing in the loop before the store (next iteration) 819 // must observe the stored value 820 bool mem_ok = true; 821 { 822 ResourceMark rm; 823 Unique_Node_List wq; 824 wq.push(phi); 825 for (uint next = 0; next < wq.size() && mem_ok; ++next) { 826 Node *m = wq.at(next); 827 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) { 828 Node* u = m->fast_out(i); 829 if (u->is_Store() || u->is_Phi()) { 830 if (u != n) { 831 wq.push(u); 832 mem_ok = (wq.size() <= 10); 833 } 834 } else { 835 mem_ok = false; 836 break; 837 } 838 } 839 } 840 } 841 if (mem_ok) { 842 // Move the store out of the loop if the LCA of all 843 // users (except for the phi) is outside the loop. 844 Node* hook = new Node(1); 845 _igvn.rehash_node_delayed(phi); 846 int count = phi->replace_edge(n, hook); 847 assert(count > 0, "inconsistent phi"); 848 849 // Compute latest point this store can go 850 Node* lca = get_late_ctrl(n, get_ctrl(n)); 851 if (n_loop->is_member(get_loop(lca))) { 852 // LCA is in the loop - bail out 853 _igvn.replace_node(hook, n); 854 return; 855 } 856 #ifdef ASSERT 857 if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) { 858 assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined"); 859 n_loop->_head->as_Loop()->verify_strip_mined(1); 860 Node* outer = n_loop->_head->as_CountedLoop()->outer_loop(); 861 IdealLoopTree* outer_loop = get_loop(outer); 862 assert(n_loop->_parent == outer_loop, "broken loop tree"); 863 assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state"); 864 } 865 #endif 866 867 // Move store out of the loop 868 _igvn.replace_node(hook, n->in(MemNode::Memory)); 869 _igvn.replace_input_of(n, 0, lca); 870 set_ctrl_and_loop(n, lca); 871 872 // Disconnect the phi now. An empty phi can confuse other 873 // optimizations in this pass of loop opts.. 874 if (phi->in(LoopNode::LoopBackControl) == phi) { 875 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl)); 876 n_loop->_body.yank(phi); 877 } 878 } 879 } 880 } 881 } 882 } 883 } 884 885 //------------------------------split_if_with_blocks_pre----------------------- 886 // Do the real work in a non-recursive function. Data nodes want to be 887 // cloned in the pre-order so they can feed each other nicely. 888 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) { 889 // Cloning these guys is unlikely to win 890 int n_op = n->Opcode(); 891 if( n_op == Op_MergeMem ) return n; 892 if( n->is_Proj() ) return n; 893 // Do not clone-up CmpFXXX variations, as these are always 894 // followed by a CmpI 895 if( n->is_Cmp() ) return n; 896 // Attempt to use a conditional move instead of a phi/branch 897 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) { 898 Node *cmov = conditional_move( n ); 899 if( cmov ) return cmov; 900 } 901 if( n->is_CFG() || n->is_LoadStore() ) 902 return n; 903 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd 904 n_op == Op_Opaque2 ) { 905 if( !C->major_progress() ) // If chance of no more loop opts... 906 _igvn._worklist.push(n); // maybe we'll remove them 907 return n; 908 } 909 910 if( n->is_Con() ) return n; // No cloning for Con nodes 911 912 Node *n_ctrl = get_ctrl(n); 913 if( !n_ctrl ) return n; // Dead node 914 915 Node* res = try_move_store_before_loop(n, n_ctrl); 916 if (res != NULL) { 917 return n; 918 } 919 920 // Attempt to remix address expressions for loop invariants 921 Node *m = remix_address_expressions( n ); 922 if( m ) return m; 923 924 if (n->is_ConstraintCast()) { 925 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this); 926 // ConstraintCastNode::dominating_cast() uses node control input to determine domination. 927 // Node control inputs don't necessarily agree with loop control info (due to 928 // transformations happened in between), thus additional dominance check is needed 929 // to keep loop info valid. 930 if (dom_cast != NULL && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) { 931 _igvn.replace_node(n, dom_cast); 932 return dom_cast; 933 } 934 } 935 936 // Determine if the Node has inputs from some local Phi. 937 // Returns the block to clone thru. 938 Node *n_blk = has_local_phi_input( n ); 939 if( !n_blk ) return n; 940 941 // Do not clone the trip counter through on a CountedLoop 942 // (messes up the canonical shape). 943 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n; 944 945 // Check for having no control input; not pinned. Allow 946 // dominating control. 947 if (n->in(0)) { 948 Node *dom = idom(n_blk); 949 if (dom_lca(n->in(0), dom) != n->in(0)) { 950 return n; 951 } 952 } 953 // Policy: when is it profitable. You must get more wins than 954 // policy before it is considered profitable. Policy is usually 0, 955 // so 1 win is considered profitable. Big merges will require big 956 // cloning, so get a larger policy. 957 int policy = n_blk->req() >> 2; 958 959 // If the loop is a candidate for range check elimination, 960 // delay splitting through it's phi until a later loop optimization 961 if (n_blk->is_CountedLoop()) { 962 IdealLoopTree *lp = get_loop(n_blk); 963 if (lp && lp->_rce_candidate) { 964 return n; 965 } 966 } 967 968 // Use same limit as split_if_with_blocks_post 969 if( C->live_nodes() > 35000 ) return n; // Method too big 970 971 // Split 'n' through the merge point if it is profitable 972 Node *phi = split_thru_phi( n, n_blk, policy ); 973 if (!phi) return n; 974 975 // Found a Phi to split thru! 976 // Replace 'n' with the new phi 977 _igvn.replace_node( n, phi ); 978 // Moved a load around the loop, 'en-registering' something. 979 if (n_blk->is_Loop() && n->is_Load() && 980 !phi->in(LoopNode::LoopBackControl)->is_Load()) 981 C->set_major_progress(); 982 983 return phi; 984 } 985 986 static bool merge_point_too_heavy(Compile* C, Node* region) { 987 // Bail out if the region and its phis have too many users. 988 int weight = 0; 989 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 990 weight += region->fast_out(i)->outcnt(); 991 } 992 int nodes_left = C->max_node_limit() - C->live_nodes(); 993 if (weight * 8 > nodes_left) { 994 if (PrintOpto) { 995 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight); 996 } 997 return true; 998 } else { 999 return false; 1000 } 1001 } 1002 1003 static bool merge_point_safe(Node* region) { 1004 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode 1005 // having a PhiNode input. This sidesteps the dangerous case where the split 1006 // ConvI2LNode may become TOP if the input Value() does not 1007 // overlap the ConvI2L range, leaving a node which may not dominate its 1008 // uses. 1009 // A better fix for this problem can be found in the BugTraq entry, but 1010 // expediency for Mantis demands this hack. 1011 // 6855164: If the merge point has a FastLockNode with a PhiNode input, we stop 1012 // split_if_with_blocks from splitting a block because we could not move around 1013 // the FastLockNode. 1014 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1015 Node* n = region->fast_out(i); 1016 if (n->is_Phi()) { 1017 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1018 Node* m = n->fast_out(j); 1019 if (m->is_FastLock()) 1020 return false; 1021 #ifdef _LP64 1022 if (m->Opcode() == Op_ConvI2L) 1023 return false; 1024 if (m->is_CastII() && m->isa_CastII()->has_range_check()) { 1025 return false; 1026 } 1027 #endif 1028 } 1029 } 1030 } 1031 return true; 1032 } 1033 1034 1035 //------------------------------place_near_use--------------------------------- 1036 // Place some computation next to use but not inside inner loops. 1037 // For inner loop uses move it to the preheader area. 1038 Node *PhaseIdealLoop::place_near_use(Node *useblock) const { 1039 IdealLoopTree *u_loop = get_loop( useblock ); 1040 if (u_loop->_irreducible) { 1041 return useblock; 1042 } 1043 if (u_loop->_child) { 1044 if (useblock == u_loop->_head && u_loop->_head->is_OuterStripMinedLoop()) { 1045 return u_loop->_head->in(LoopNode::EntryControl); 1046 } 1047 return useblock; 1048 } 1049 return u_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl); 1050 } 1051 1052 1053 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) { 1054 if (!n->is_If() || n->is_CountedLoopEnd()) { 1055 return false; 1056 } 1057 if (!n->in(0)->is_Region()) { 1058 return false; 1059 } 1060 Node* region = n->in(0); 1061 Node* dom = idom(region); 1062 if (!dom->is_If() || dom->in(1) != n->in(1)) { 1063 return false; 1064 } 1065 IfNode* dom_if = dom->as_If(); 1066 Node* proj_true = dom_if->proj_out(1); 1067 Node* proj_false = dom_if->proj_out(0); 1068 1069 for (uint i = 1; i < region->req(); i++) { 1070 if (is_dominator(proj_true, region->in(i))) { 1071 continue; 1072 } 1073 if (is_dominator(proj_false, region->in(i))) { 1074 continue; 1075 } 1076 return false; 1077 } 1078 1079 return true; 1080 } 1081 1082 bool PhaseIdealLoop::can_split_if(Node *n_ctrl) { 1083 if (C->live_nodes() > 35000) { 1084 return false; // Method too big 1085 } 1086 1087 // Do not do 'split-if' if irreducible loops are present. 1088 if (_has_irreducible_loops) { 1089 return false; 1090 } 1091 1092 if (merge_point_too_heavy(C, n_ctrl)) { 1093 return false; 1094 } 1095 1096 // Do not do 'split-if' if some paths are dead. First do dead code 1097 // elimination and then see if its still profitable. 1098 for (uint i = 1; i < n_ctrl->req(); i++) { 1099 if (n_ctrl->in(i) == C->top()) { 1100 return false; 1101 } 1102 } 1103 1104 // If trying to do a 'Split-If' at the loop head, it is only 1105 // profitable if the cmp folds up on BOTH paths. Otherwise we 1106 // risk peeling a loop forever. 1107 1108 // CNC - Disabled for now. Requires careful handling of loop 1109 // body selection for the cloned code. Also, make sure we check 1110 // for any input path not being in the same loop as n_ctrl. For 1111 // irreducible loops we cannot check for 'n_ctrl->is_Loop()' 1112 // because the alternative loop entry points won't be converted 1113 // into LoopNodes. 1114 IdealLoopTree *n_loop = get_loop(n_ctrl); 1115 for (uint j = 1; j < n_ctrl->req(); j++) { 1116 if (get_loop(n_ctrl->in(j)) != n_loop) { 1117 return false; 1118 } 1119 } 1120 1121 // Check for safety of the merge point. 1122 if (!merge_point_safe(n_ctrl)) { 1123 return false; 1124 } 1125 1126 return true; 1127 } 1128 1129 //------------------------------split_if_with_blocks_post---------------------- 1130 // Do the real work in a non-recursive function. CFG hackery wants to be 1131 // in the post-order, so it can dirty the I-DOM info and not use the dirtied 1132 // info. 1133 void PhaseIdealLoop::split_if_with_blocks_post(Node *n, bool last_round) { 1134 1135 // Cloning Cmp through Phi's involves the split-if transform. 1136 // FastLock is not used by an If 1137 if (n->is_Cmp() && !n->is_FastLock() && !last_round) { 1138 Node *n_ctrl = get_ctrl(n); 1139 // Determine if the Node has inputs from some local Phi. 1140 // Returns the block to clone thru. 1141 Node *n_blk = has_local_phi_input(n); 1142 if (n_blk != n_ctrl) { 1143 return; 1144 } 1145 1146 if (!can_split_if(n_ctrl)) { 1147 return; 1148 } 1149 1150 if (n->outcnt() != 1) { 1151 return; // Multiple bool's from 1 compare? 1152 } 1153 Node *bol = n->unique_out(); 1154 assert(bol->is_Bool(), "expect a bool here"); 1155 if (bol->outcnt() != 1) { 1156 return;// Multiple branches from 1 compare? 1157 } 1158 Node *iff = bol->unique_out(); 1159 1160 // Check some safety conditions 1161 if (iff->is_If()) { // Classic split-if? 1162 if (iff->in(0) != n_ctrl) { 1163 return; // Compare must be in same blk as if 1164 } 1165 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE 1166 // Can't split CMove with different control edge. 1167 if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) { 1168 return; 1169 } 1170 if (get_ctrl(iff->in(2)) == n_ctrl || 1171 get_ctrl(iff->in(3)) == n_ctrl) { 1172 return; // Inputs not yet split-up 1173 } 1174 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) { 1175 return; // Loop-invar test gates loop-varying CMOVE 1176 } 1177 } else { 1178 return; // some other kind of node, such as an Allocate 1179 } 1180 1181 // When is split-if profitable? Every 'win' on means some control flow 1182 // goes dead, so it's almost always a win. 1183 int policy = 0; 1184 // Split compare 'n' through the merge point if it is profitable 1185 Node *phi = split_thru_phi( n, n_ctrl, policy); 1186 if (!phi) { 1187 return; 1188 } 1189 1190 // Found a Phi to split thru! 1191 // Replace 'n' with the new phi 1192 _igvn.replace_node(n, phi); 1193 1194 // Now split the bool up thru the phi 1195 Node *bolphi = split_thru_phi(bol, n_ctrl, -1); 1196 guarantee(bolphi != NULL, "null boolean phi node"); 1197 1198 _igvn.replace_node(bol, bolphi); 1199 assert(iff->in(1) == bolphi, ""); 1200 1201 if (bolphi->Value(&_igvn)->singleton()) { 1202 return; 1203 } 1204 1205 // Conditional-move? Must split up now 1206 if (!iff->is_If()) { 1207 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1); 1208 _igvn.replace_node(iff, cmovphi); 1209 return; 1210 } 1211 1212 // Now split the IF 1213 do_split_if(iff); 1214 return; 1215 } 1216 1217 // Two identical ifs back to back can be merged 1218 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) { 1219 Node *n_ctrl = n->in(0); 1220 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1)); 1221 IfNode* dom_if = idom(n_ctrl)->as_If(); 1222 Node* proj_true = dom_if->proj_out(1); 1223 Node* proj_false = dom_if->proj_out(0); 1224 Node* con_true = _igvn.makecon(TypeInt::ONE); 1225 Node* con_false = _igvn.makecon(TypeInt::ZERO); 1226 1227 for (uint i = 1; i < n_ctrl->req(); i++) { 1228 if (is_dominator(proj_true, n_ctrl->in(i))) { 1229 bolphi->init_req(i, con_true); 1230 } else { 1231 assert(is_dominator(proj_false, n_ctrl->in(i)), "bad if"); 1232 bolphi->init_req(i, con_false); 1233 } 1234 } 1235 register_new_node(bolphi, n_ctrl); 1236 _igvn.replace_input_of(n, 1, bolphi); 1237 1238 // Now split the IF 1239 do_split_if(n); 1240 return; 1241 } 1242 1243 // Check for an IF ready to split; one that has its 1244 // condition codes input coming from a Phi at the block start. 1245 int n_op = n->Opcode(); 1246 1247 // Check for an IF being dominated by another IF same test 1248 if (n_op == Op_If || 1249 n_op == Op_RangeCheck) { 1250 Node *bol = n->in(1); 1251 uint max = bol->outcnt(); 1252 // Check for same test used more than once? 1253 if (max > 1 && bol->is_Bool()) { 1254 // Search up IDOMs to see if this IF is dominated. 1255 Node *cutoff = get_ctrl(bol); 1256 1257 // Now search up IDOMs till cutoff, looking for a dominating test 1258 Node *prevdom = n; 1259 Node *dom = idom(prevdom); 1260 while (dom != cutoff) { 1261 if (dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom) { 1262 // Replace the dominated test with an obvious true or false. 1263 // Place it on the IGVN worklist for later cleanup. 1264 C->set_major_progress(); 1265 dominated_by(prevdom, n, false, true); 1266 #ifndef PRODUCT 1267 if( VerifyLoopOptimizations ) verify(); 1268 #endif 1269 return; 1270 } 1271 prevdom = dom; 1272 dom = idom(prevdom); 1273 } 1274 } 1275 } 1276 1277 // See if a shared loop-varying computation has no loop-varying uses. 1278 // Happens if something is only used for JVM state in uncommon trap exits, 1279 // like various versions of induction variable+offset. Clone the 1280 // computation per usage to allow it to sink out of the loop. 1281 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about) 1282 Node *n_ctrl = get_ctrl(n); 1283 IdealLoopTree *n_loop = get_loop(n_ctrl); 1284 if( n_loop != _ltree_root ) { 1285 DUIterator_Fast imax, i = n->fast_outs(imax); 1286 for (; i < imax; i++) { 1287 Node* u = n->fast_out(i); 1288 if( !has_ctrl(u) ) break; // Found control user 1289 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 1290 if( u_loop == n_loop ) break; // Found loop-varying use 1291 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop 1292 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003 1293 } 1294 bool did_break = (i < imax); // Did we break out of the previous loop? 1295 if (!did_break && n->outcnt() > 1) { // All uses in outer loops! 1296 Node *late_load_ctrl = NULL; 1297 if (n->is_Load()) { 1298 // If n is a load, get and save the result from get_late_ctrl(), 1299 // to be later used in calculating the control for n's clones. 1300 clear_dom_lca_tags(); 1301 late_load_ctrl = get_late_ctrl(n, n_ctrl); 1302 } 1303 // If n is a load, and the late control is the same as the current 1304 // control, then the cloning of n is a pointless exercise, because 1305 // GVN will ensure that we end up where we started. 1306 if (!n->is_Load() || late_load_ctrl != n_ctrl) { 1307 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) { 1308 Node *u = n->last_out(j); // Clone private computation per use 1309 _igvn.rehash_node_delayed(u); 1310 Node *x = n->clone(); // Clone computation 1311 Node *x_ctrl = NULL; 1312 if( u->is_Phi() ) { 1313 // Replace all uses of normal nodes. Replace Phi uses 1314 // individually, so the separate Nodes can sink down 1315 // different paths. 1316 uint k = 1; 1317 while( u->in(k) != n ) k++; 1318 u->set_req( k, x ); 1319 // x goes next to Phi input path 1320 x_ctrl = u->in(0)->in(k); 1321 --j; 1322 } else { // Normal use 1323 // Replace all uses 1324 for( uint k = 0; k < u->req(); k++ ) { 1325 if( u->in(k) == n ) { 1326 u->set_req( k, x ); 1327 --j; 1328 } 1329 } 1330 x_ctrl = get_ctrl(u); 1331 } 1332 1333 // Find control for 'x' next to use but not inside inner loops. 1334 // For inner loop uses get the preheader area. 1335 x_ctrl = place_near_use(x_ctrl); 1336 1337 if (n->is_Load()) { 1338 // For loads, add a control edge to a CFG node outside of the loop 1339 // to force them to not combine and return back inside the loop 1340 // during GVN optimization (4641526). 1341 // 1342 // Because we are setting the actual control input, factor in 1343 // the result from get_late_ctrl() so we respect any 1344 // anti-dependences. (6233005). 1345 x_ctrl = dom_lca(late_load_ctrl, x_ctrl); 1346 1347 // Don't allow the control input to be a CFG splitting node. 1348 // Such nodes should only have ProjNodes as outs, e.g. IfNode 1349 // should only have IfTrueNode and IfFalseNode (4985384). 1350 x_ctrl = find_non_split_ctrl(x_ctrl); 1351 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone"); 1352 1353 x->set_req(0, x_ctrl); 1354 } 1355 register_new_node(x, x_ctrl); 1356 1357 // Some institutional knowledge is needed here: 'x' is 1358 // yanked because if the optimizer runs GVN on it all the 1359 // cloned x's will common up and undo this optimization and 1360 // be forced back in the loop. This is annoying because it 1361 // makes +VerifyOpto report false-positives on progress. I 1362 // tried setting control edges on the x's to force them to 1363 // not combine, but the matching gets worried when it tries 1364 // to fold a StoreP and an AddP together (as part of an 1365 // address expression) and the AddP and StoreP have 1366 // different controls. 1367 if (!x->is_Load() && !x->is_DecodeNarrowPtr()) _igvn._worklist.yank(x); 1368 } 1369 _igvn.remove_dead_node(n); 1370 } 1371 } 1372 } 1373 } 1374 1375 try_move_store_after_loop(n); 1376 1377 // Check for Opaque2's who's loop has disappeared - who's input is in the 1378 // same loop nest as their output. Remove 'em, they are no longer useful. 1379 if( n_op == Op_Opaque2 && 1380 n->in(1) != NULL && 1381 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) { 1382 _igvn.replace_node( n, n->in(1) ); 1383 } 1384 1385 #if INCLUDE_ZGC 1386 if (UseZGC) { 1387 ZBarrierSetC2::loop_optimize_gc_barrier(this, n, last_round); 1388 } 1389 #endif 1390 } 1391 1392 //------------------------------split_if_with_blocks--------------------------- 1393 // Check for aggressive application of 'split-if' optimization, 1394 // using basic block level info. 1395 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack, bool last_round) { 1396 Node *n = C->root(); 1397 visited.set(n->_idx); // first, mark node as visited 1398 // Do pre-visit work for root 1399 n = split_if_with_blocks_pre( n ); 1400 uint cnt = n->outcnt(); 1401 uint i = 0; 1402 while (true) { 1403 // Visit all children 1404 if (i < cnt) { 1405 Node* use = n->raw_out(i); 1406 ++i; 1407 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) { 1408 // Now do pre-visit work for this use 1409 use = split_if_with_blocks_pre( use ); 1410 nstack.push(n, i); // Save parent and next use's index. 1411 n = use; // Process all children of current use. 1412 cnt = use->outcnt(); 1413 i = 0; 1414 } 1415 } 1416 else { 1417 // All of n's children have been processed, complete post-processing. 1418 if (cnt != 0 && !n->is_Con()) { 1419 assert(has_node(n), "no dead nodes"); 1420 split_if_with_blocks_post( n, last_round ); 1421 } 1422 if (nstack.is_empty()) { 1423 // Finished all nodes on stack. 1424 break; 1425 } 1426 // Get saved parent node and next use's index. Visit the rest of uses. 1427 n = nstack.node(); 1428 cnt = n->outcnt(); 1429 i = nstack.index(); 1430 nstack.pop(); 1431 } 1432 } 1433 } 1434 1435 1436 //============================================================================= 1437 // 1438 // C L O N E A L O O P B O D Y 1439 // 1440 1441 //------------------------------clone_iff-------------------------------------- 1442 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1443 // "Nearly" because all Nodes have been cloned from the original in the loop, 1444 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1445 // through the Phi recursively, and return a Bool. 1446 Node* PhaseIdealLoop::clone_iff(PhiNode *phi, IdealLoopTree *loop) { 1447 1448 // Convert this Phi into a Phi merging Bools 1449 uint i; 1450 for (i = 1; i < phi->req(); i++) { 1451 Node *b = phi->in(i); 1452 if (b->is_Phi()) { 1453 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi(), loop)); 1454 } else { 1455 assert(b->is_Bool() || b->Opcode() == Op_Opaque4, ""); 1456 } 1457 } 1458 1459 Node* n = phi->in(1); 1460 Node* sample_opaque = NULL; 1461 Node *sample_bool = NULL; 1462 if (n->Opcode() == Op_Opaque4) { 1463 sample_opaque = n; 1464 sample_bool = n->in(1); 1465 assert(sample_bool->is_Bool(), "wrong type"); 1466 } else { 1467 sample_bool = n; 1468 } 1469 Node *sample_cmp = sample_bool->in(1); 1470 1471 // Make Phis to merge the Cmp's inputs. 1472 PhiNode *phi1 = new PhiNode(phi->in(0), Type::TOP); 1473 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP); 1474 for (i = 1; i < phi->req(); i++) { 1475 Node *n1 = sample_opaque == NULL ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1); 1476 Node *n2 = sample_opaque == NULL ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2); 1477 phi1->set_req(i, n1); 1478 phi2->set_req(i, n2); 1479 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 1480 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 1481 } 1482 // See if these Phis have been made before. 1483 // Register with optimizer 1484 Node *hit1 = _igvn.hash_find_insert(phi1); 1485 if (hit1) { // Hit, toss just made Phi 1486 _igvn.remove_dead_node(phi1); // Remove new phi 1487 assert(hit1->is_Phi(), "" ); 1488 phi1 = (PhiNode*)hit1; // Use existing phi 1489 } else { // Miss 1490 _igvn.register_new_node_with_optimizer(phi1); 1491 } 1492 Node *hit2 = _igvn.hash_find_insert(phi2); 1493 if (hit2) { // Hit, toss just made Phi 1494 _igvn.remove_dead_node(phi2); // Remove new phi 1495 assert(hit2->is_Phi(), "" ); 1496 phi2 = (PhiNode*)hit2; // Use existing phi 1497 } else { // Miss 1498 _igvn.register_new_node_with_optimizer(phi2); 1499 } 1500 // Register Phis with loop/block info 1501 set_ctrl(phi1, phi->in(0)); 1502 set_ctrl(phi2, phi->in(0)); 1503 // Make a new Cmp 1504 Node *cmp = sample_cmp->clone(); 1505 cmp->set_req(1, phi1); 1506 cmp->set_req(2, phi2); 1507 _igvn.register_new_node_with_optimizer(cmp); 1508 set_ctrl(cmp, phi->in(0)); 1509 1510 // Make a new Bool 1511 Node *b = sample_bool->clone(); 1512 b->set_req(1,cmp); 1513 _igvn.register_new_node_with_optimizer(b); 1514 set_ctrl(b, phi->in(0)); 1515 1516 if (sample_opaque != NULL) { 1517 Node* opaque = sample_opaque->clone(); 1518 opaque->set_req(1, b); 1519 _igvn.register_new_node_with_optimizer(opaque); 1520 set_ctrl(opaque, phi->in(0)); 1521 return opaque; 1522 } 1523 1524 assert(b->is_Bool(), ""); 1525 return b; 1526 } 1527 1528 //------------------------------clone_bool------------------------------------- 1529 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1530 // "Nearly" because all Nodes have been cloned from the original in the loop, 1531 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1532 // through the Phi recursively, and return a Bool. 1533 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) { 1534 uint i; 1535 // Convert this Phi into a Phi merging Bools 1536 for( i = 1; i < phi->req(); i++ ) { 1537 Node *b = phi->in(i); 1538 if( b->is_Phi() ) { 1539 _igvn.replace_input_of(phi, i, clone_bool( b->as_Phi(), loop )); 1540 } else { 1541 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" ); 1542 } 1543 } 1544 1545 Node *sample_cmp = phi->in(1); 1546 1547 // Make Phis to merge the Cmp's inputs. 1548 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP ); 1549 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP ); 1550 for( uint j = 1; j < phi->req(); j++ ) { 1551 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP 1552 Node *n1, *n2; 1553 if( cmp_top->is_Cmp() ) { 1554 n1 = cmp_top->in(1); 1555 n2 = cmp_top->in(2); 1556 } else { 1557 n1 = n2 = cmp_top; 1558 } 1559 phi1->set_req( j, n1 ); 1560 phi2->set_req( j, n2 ); 1561 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 1562 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 1563 } 1564 1565 // See if these Phis have been made before. 1566 // Register with optimizer 1567 Node *hit1 = _igvn.hash_find_insert(phi1); 1568 if( hit1 ) { // Hit, toss just made Phi 1569 _igvn.remove_dead_node(phi1); // Remove new phi 1570 assert( hit1->is_Phi(), "" ); 1571 phi1 = (PhiNode*)hit1; // Use existing phi 1572 } else { // Miss 1573 _igvn.register_new_node_with_optimizer(phi1); 1574 } 1575 Node *hit2 = _igvn.hash_find_insert(phi2); 1576 if( hit2 ) { // Hit, toss just made Phi 1577 _igvn.remove_dead_node(phi2); // Remove new phi 1578 assert( hit2->is_Phi(), "" ); 1579 phi2 = (PhiNode*)hit2; // Use existing phi 1580 } else { // Miss 1581 _igvn.register_new_node_with_optimizer(phi2); 1582 } 1583 // Register Phis with loop/block info 1584 set_ctrl(phi1, phi->in(0)); 1585 set_ctrl(phi2, phi->in(0)); 1586 // Make a new Cmp 1587 Node *cmp = sample_cmp->clone(); 1588 cmp->set_req( 1, phi1 ); 1589 cmp->set_req( 2, phi2 ); 1590 _igvn.register_new_node_with_optimizer(cmp); 1591 set_ctrl(cmp, phi->in(0)); 1592 1593 assert( cmp->is_Cmp(), "" ); 1594 return (CmpNode*)cmp; 1595 } 1596 1597 //------------------------------sink_use--------------------------------------- 1598 // If 'use' was in the loop-exit block, it now needs to be sunk 1599 // below the post-loop merge point. 1600 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) { 1601 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) { 1602 set_ctrl(use, post_loop); 1603 for (DUIterator j = use->outs(); use->has_out(j); j++) 1604 sink_use(use->out(j), post_loop); 1605 } 1606 } 1607 1608 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new, 1609 IdealLoopTree* loop, IdealLoopTree* outer_loop, 1610 Node_List*& split_if_set, Node_List*& split_bool_set, 1611 Node_List*& split_cex_set, Node_List& worklist, 1612 uint new_counter, CloneLoopMode mode) { 1613 Node* nnn = old_new[old->_idx]; 1614 // Copy uses to a worklist, so I can munge the def-use info 1615 // with impunity. 1616 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1617 worklist.push(old->fast_out(j)); 1618 1619 while( worklist.size() ) { 1620 Node *use = worklist.pop(); 1621 if (!has_node(use)) continue; // Ignore dead nodes 1622 if (use->in(0) == C->top()) continue; 1623 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1624 // Check for data-use outside of loop - at least one of OLD or USE 1625 // must not be a CFG node. 1626 #ifdef ASSERT 1627 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == NULL) { 1628 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint(); 1629 assert(mode == ControlAroundStripMined && use == sfpt, "missed a node"); 1630 } 1631 #endif 1632 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) { 1633 1634 // If the Data use is an IF, that means we have an IF outside of the 1635 // loop that is switching on a condition that is set inside of the 1636 // loop. Happens if people set a loop-exit flag; then test the flag 1637 // in the loop to break the loop, then test is again outside of the 1638 // loop to determine which way the loop exited. 1639 // Loop predicate If node connects to Bool node through Opaque1 node. 1640 if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use) || use->Opcode() == Op_Opaque4) { 1641 // Since this code is highly unlikely, we lazily build the worklist 1642 // of such Nodes to go split. 1643 if (!split_if_set) { 1644 ResourceArea *area = Thread::current()->resource_area(); 1645 split_if_set = new Node_List(area); 1646 } 1647 split_if_set->push(use); 1648 } 1649 if (use->is_Bool()) { 1650 if (!split_bool_set) { 1651 ResourceArea *area = Thread::current()->resource_area(); 1652 split_bool_set = new Node_List(area); 1653 } 1654 split_bool_set->push(use); 1655 } 1656 if (use->Opcode() == Op_CreateEx) { 1657 if (!split_cex_set) { 1658 ResourceArea *area = Thread::current()->resource_area(); 1659 split_cex_set = new Node_List(area); 1660 } 1661 split_cex_set->push(use); 1662 } 1663 1664 1665 // Get "block" use is in 1666 uint idx = 0; 1667 while( use->in(idx) != old ) idx++; 1668 Node *prev = use->is_CFG() ? use : get_ctrl(use); 1669 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" ); 1670 Node *cfg = prev->_idx >= new_counter 1671 ? prev->in(2) 1672 : idom(prev); 1673 if( use->is_Phi() ) // Phi use is in prior block 1674 cfg = prev->in(idx); // NOT in block of Phi itself 1675 if (cfg->is_top()) { // Use is dead? 1676 _igvn.replace_input_of(use, idx, C->top()); 1677 continue; 1678 } 1679 1680 // If use is referenced through control edge... (idx == 0) 1681 if (mode == IgnoreStripMined && idx == 0) { 1682 LoopNode *head = loop->_head->as_Loop(); 1683 if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) { 1684 // That node is outside the inner loop, leave it outside the 1685 // outer loop as well to not confuse verification code. 1686 assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop"); 1687 _igvn.replace_input_of(use, 0, head->outer_loop_exit()); 1688 continue; 1689 } 1690 } 1691 1692 while(!outer_loop->is_member(get_loop(cfg))) { 1693 prev = cfg; 1694 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg); 1695 } 1696 // If the use occurs after merging several exits from the loop, then 1697 // old value must have dominated all those exits. Since the same old 1698 // value was used on all those exits we did not need a Phi at this 1699 // merge point. NOW we do need a Phi here. Each loop exit value 1700 // is now merged with the peeled body exit; each exit gets its own 1701 // private Phi and those Phis need to be merged here. 1702 Node *phi; 1703 if( prev->is_Region() ) { 1704 if( idx == 0 ) { // Updating control edge? 1705 phi = prev; // Just use existing control 1706 } else { // Else need a new Phi 1707 phi = PhiNode::make( prev, old ); 1708 // Now recursively fix up the new uses of old! 1709 for( uint i = 1; i < prev->req(); i++ ) { 1710 worklist.push(phi); // Onto worklist once for each 'old' input 1711 } 1712 } 1713 } else { 1714 // Get new RegionNode merging old and new loop exits 1715 prev = old_new[prev->_idx]; 1716 assert( prev, "just made this in step 7" ); 1717 if( idx == 0) { // Updating control edge? 1718 phi = prev; // Just use existing control 1719 } else { // Else need a new Phi 1720 // Make a new Phi merging data values properly 1721 phi = PhiNode::make( prev, old ); 1722 phi->set_req( 1, nnn ); 1723 } 1724 } 1725 // If inserting a new Phi, check for prior hits 1726 if( idx != 0 ) { 1727 Node *hit = _igvn.hash_find_insert(phi); 1728 if( hit == NULL ) { 1729 _igvn.register_new_node_with_optimizer(phi); // Register new phi 1730 } else { // or 1731 // Remove the new phi from the graph and use the hit 1732 _igvn.remove_dead_node(phi); 1733 phi = hit; // Use existing phi 1734 } 1735 set_ctrl(phi, prev); 1736 } 1737 // Make 'use' use the Phi instead of the old loop body exit value 1738 _igvn.replace_input_of(use, idx, phi); 1739 if( use->_idx >= new_counter ) { // If updating new phis 1740 // Not needed for correctness, but prevents a weak assert 1741 // in AddPNode from tripping (when we end up with different 1742 // base & derived Phis that will become the same after 1743 // IGVN does CSE). 1744 Node *hit = _igvn.hash_find_insert(use); 1745 if( hit ) // Go ahead and re-hash for hits. 1746 _igvn.replace_node( use, hit ); 1747 } 1748 1749 // If 'use' was in the loop-exit block, it now needs to be sunk 1750 // below the post-loop merge point. 1751 sink_use( use, prev ); 1752 } 1753 } 1754 } 1755 1756 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop, 1757 IdealLoopTree* outer_loop, int dd, Node_List &old_new, 1758 Node_List& extra_data_nodes) { 1759 if (head->is_strip_mined() && mode != IgnoreStripMined) { 1760 CountedLoopNode* cl = head->as_CountedLoop(); 1761 Node* l = cl->outer_loop(); 1762 Node* tail = cl->outer_loop_tail(); 1763 IfNode* le = cl->outer_loop_end(); 1764 Node* sfpt = cl->outer_safepoint(); 1765 CountedLoopEndNode* cle = cl->loopexit(); 1766 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop(); 1767 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null(); 1768 Node* cle_out = cle->proj_out(false); 1769 1770 Node* new_sfpt = NULL; 1771 Node* new_cle_out = cle_out->clone(); 1772 old_new.map(cle_out->_idx, new_cle_out); 1773 if (mode == CloneIncludesStripMined) { 1774 // clone outer loop body 1775 Node* new_l = l->clone(); 1776 Node* new_tail = tail->clone(); 1777 IfNode* new_le = le->clone()->as_If(); 1778 new_sfpt = sfpt->clone(); 1779 1780 set_loop(new_l, outer_loop->_parent); 1781 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd); 1782 set_loop(new_cle_out, outer_loop->_parent); 1783 set_idom(new_cle_out, new_cle, dd); 1784 set_loop(new_sfpt, outer_loop->_parent); 1785 set_idom(new_sfpt, new_cle_out, dd); 1786 set_loop(new_le, outer_loop->_parent); 1787 set_idom(new_le, new_sfpt, dd); 1788 set_loop(new_tail, outer_loop->_parent); 1789 set_idom(new_tail, new_le, dd); 1790 set_idom(new_cl, new_l, dd); 1791 1792 old_new.map(l->_idx, new_l); 1793 old_new.map(tail->_idx, new_tail); 1794 old_new.map(le->_idx, new_le); 1795 old_new.map(sfpt->_idx, new_sfpt); 1796 1797 new_l->set_req(LoopNode::LoopBackControl, new_tail); 1798 new_l->set_req(0, new_l); 1799 new_tail->set_req(0, new_le); 1800 new_le->set_req(0, new_sfpt); 1801 new_sfpt->set_req(0, new_cle_out); 1802 new_cle_out->set_req(0, new_cle); 1803 new_cl->set_req(LoopNode::EntryControl, new_l); 1804 1805 _igvn.register_new_node_with_optimizer(new_l); 1806 _igvn.register_new_node_with_optimizer(new_tail); 1807 _igvn.register_new_node_with_optimizer(new_le); 1808 } else { 1809 Node *newhead = old_new[loop->_head->_idx]; 1810 newhead->as_Loop()->clear_strip_mined(); 1811 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl)); 1812 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 1813 } 1814 // Look at data node that were assigned a control in the outer 1815 // loop: they are kept in the outer loop by the safepoint so start 1816 // from the safepoint node's inputs. 1817 IdealLoopTree* outer_loop = get_loop(l); 1818 Node_Stack stack(2); 1819 stack.push(sfpt, 1); 1820 uint new_counter = C->unique(); 1821 while (stack.size() > 0) { 1822 Node* n = stack.node(); 1823 uint i = stack.index(); 1824 while (i < n->req() && 1825 (n->in(i) == NULL || 1826 !has_ctrl(n->in(i)) || 1827 get_loop(get_ctrl(n->in(i))) != outer_loop || 1828 (old_new[n->in(i)->_idx] != NULL && old_new[n->in(i)->_idx]->_idx >= new_counter))) { 1829 i++; 1830 } 1831 if (i < n->req()) { 1832 stack.set_index(i+1); 1833 stack.push(n->in(i), 0); 1834 } else { 1835 assert(old_new[n->_idx] == NULL || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet"); 1836 Node* m = n == sfpt ? new_sfpt : n->clone(); 1837 if (m != NULL) { 1838 for (uint i = 0; i < n->req(); i++) { 1839 if (m->in(i) != NULL && old_new[m->in(i)->_idx] != NULL) { 1840 m->set_req(i, old_new[m->in(i)->_idx]); 1841 } 1842 } 1843 } else { 1844 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?"); 1845 } 1846 if (n != sfpt) { 1847 extra_data_nodes.push(n); 1848 _igvn.register_new_node_with_optimizer(m); 1849 assert(get_ctrl(n) == cle_out, "what other control?"); 1850 set_ctrl(m, new_cle_out); 1851 old_new.map(n->_idx, m); 1852 } 1853 stack.pop(); 1854 } 1855 } 1856 if (mode == CloneIncludesStripMined) { 1857 _igvn.register_new_node_with_optimizer(new_sfpt); 1858 _igvn.register_new_node_with_optimizer(new_cle_out); 1859 } 1860 } else { 1861 Node *newhead = old_new[loop->_head->_idx]; 1862 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 1863 } 1864 } 1865 1866 //------------------------------clone_loop------------------------------------- 1867 // 1868 // C L O N E A L O O P B O D Y 1869 // 1870 // This is the basic building block of the loop optimizations. It clones an 1871 // entire loop body. It makes an old_new loop body mapping; with this mapping 1872 // you can find the new-loop equivalent to an old-loop node. All new-loop 1873 // nodes are exactly equal to their old-loop counterparts, all edges are the 1874 // same. All exits from the old-loop now have a RegionNode that merges the 1875 // equivalent new-loop path. This is true even for the normal "loop-exit" 1876 // condition. All uses of loop-invariant old-loop values now come from (one 1877 // or more) Phis that merge their new-loop equivalents. 1878 // 1879 // This operation leaves the graph in an illegal state: there are two valid 1880 // control edges coming from the loop pre-header to both loop bodies. I'll 1881 // definitely have to hack the graph after running this transform. 1882 // 1883 // From this building block I will further edit edges to perform loop peeling 1884 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc. 1885 // 1886 // Parameter side_by_size_idom: 1887 // When side_by_size_idom is NULL, the dominator tree is constructed for 1888 // the clone loop to dominate the original. Used in construction of 1889 // pre-main-post loop sequence. 1890 // When nonnull, the clone and original are side-by-side, both are 1891 // dominated by the side_by_side_idom node. Used in construction of 1892 // unswitched loops. 1893 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd, 1894 CloneLoopMode mode, Node* side_by_side_idom) { 1895 1896 LoopNode* head = loop->_head->as_Loop(); 1897 head->verify_strip_mined(1); 1898 1899 if (C->do_vector_loop() && PrintOpto) { 1900 const char* mname = C->method()->name()->as_quoted_ascii(); 1901 if (mname != NULL) { 1902 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname); 1903 } 1904 } 1905 1906 CloneMap& cm = C->clone_map(); 1907 Dict* dict = cm.dict(); 1908 if (C->do_vector_loop()) { 1909 cm.set_clone_idx(cm.max_gen()+1); 1910 #ifndef PRODUCT 1911 if (PrintOpto) { 1912 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx()); 1913 loop->dump_head(); 1914 } 1915 #endif 1916 } 1917 1918 // Step 1: Clone the loop body. Make the old->new mapping. 1919 uint i; 1920 for( i = 0; i < loop->_body.size(); i++ ) { 1921 Node *old = loop->_body.at(i); 1922 Node *nnn = old->clone(); 1923 old_new.map( old->_idx, nnn ); 1924 if (C->do_vector_loop()) { 1925 cm.verify_insert_and_clone(old, nnn, cm.clone_idx()); 1926 } 1927 _igvn.register_new_node_with_optimizer(nnn); 1928 } 1929 1930 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop; 1931 1932 // Step 2: Fix the edges in the new body. If the old input is outside the 1933 // loop use it. If the old input is INside the loop, use the corresponding 1934 // new node instead. 1935 for( i = 0; i < loop->_body.size(); i++ ) { 1936 Node *old = loop->_body.at(i); 1937 Node *nnn = old_new[old->_idx]; 1938 // Fix CFG/Loop controlling the new node 1939 if (has_ctrl(old)) { 1940 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]); 1941 } else { 1942 set_loop(nnn, outer_loop->_parent); 1943 if (old->outcnt() > 0) { 1944 set_idom( nnn, old_new[idom(old)->_idx], dd ); 1945 } 1946 } 1947 // Correct edges to the new node 1948 for( uint j = 0; j < nnn->req(); j++ ) { 1949 Node *n = nnn->in(j); 1950 if( n ) { 1951 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n ); 1952 if( loop->is_member( old_in_loop ) ) 1953 nnn->set_req(j, old_new[n->_idx]); 1954 } 1955 } 1956 _igvn.hash_find_insert(nnn); 1957 } 1958 1959 ResourceArea *area = Thread::current()->resource_area(); 1960 Node_List extra_data_nodes(area); 1961 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes); 1962 1963 // Step 3: Now fix control uses. Loop varying control uses have already 1964 // been fixed up (as part of all input edges in Step 2). Loop invariant 1965 // control uses must be either an IfFalse or an IfTrue. Make a merge 1966 // point to merge the old and new IfFalse/IfTrue nodes; make the use 1967 // refer to this. 1968 Node_List worklist(area); 1969 uint new_counter = C->unique(); 1970 for( i = 0; i < loop->_body.size(); i++ ) { 1971 Node* old = loop->_body.at(i); 1972 if( !old->is_CFG() ) continue; 1973 1974 // Copy uses to a worklist, so I can munge the def-use info 1975 // with impunity. 1976 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1977 worklist.push(old->fast_out(j)); 1978 1979 while( worklist.size() ) { // Visit all uses 1980 Node *use = worklist.pop(); 1981 if (!has_node(use)) continue; // Ignore dead nodes 1982 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1983 if( !loop->is_member( use_loop ) && use->is_CFG() ) { 1984 // Both OLD and USE are CFG nodes here. 1985 assert( use->is_Proj(), "" ); 1986 Node* nnn = old_new[old->_idx]; 1987 1988 Node* newuse = NULL; 1989 if (head->is_strip_mined() && mode != IgnoreStripMined) { 1990 CountedLoopNode* cl = head->as_CountedLoop(); 1991 CountedLoopEndNode* cle = cl->loopexit(); 1992 Node* cle_out = cle->proj_out_or_null(false); 1993 if (use == cle_out) { 1994 IfNode* le = cl->outer_loop_end(); 1995 use = le->proj_out(false); 1996 use_loop = get_loop(use); 1997 if (mode == CloneIncludesStripMined) { 1998 nnn = old_new[le->_idx]; 1999 } else { 2000 newuse = old_new[cle_out->_idx]; 2001 } 2002 } 2003 } 2004 if (newuse == NULL) { 2005 newuse = use->clone(); 2006 } 2007 2008 // Clone the loop exit control projection 2009 if (C->do_vector_loop()) { 2010 cm.verify_insert_and_clone(use, newuse, cm.clone_idx()); 2011 } 2012 newuse->set_req(0,nnn); 2013 _igvn.register_new_node_with_optimizer(newuse); 2014 set_loop(newuse, use_loop); 2015 set_idom(newuse, nnn, dom_depth(nnn) + 1 ); 2016 2017 // We need a Region to merge the exit from the peeled body and the 2018 // exit from the old loop body. 2019 RegionNode *r = new RegionNode(3); 2020 // Map the old use to the new merge point 2021 old_new.map( use->_idx, r ); 2022 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use)); 2023 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" ); 2024 2025 // The original user of 'use' uses 'r' instead. 2026 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) { 2027 Node* useuse = use->last_out(l); 2028 _igvn.rehash_node_delayed(useuse); 2029 uint uses_found = 0; 2030 if( useuse->in(0) == use ) { 2031 useuse->set_req(0, r); 2032 uses_found++; 2033 if( useuse->is_CFG() ) { 2034 assert( dom_depth(useuse) > dd_r, "" ); 2035 set_idom(useuse, r, dom_depth(useuse)); 2036 } 2037 } 2038 for( uint k = 1; k < useuse->req(); k++ ) { 2039 if( useuse->in(k) == use ) { 2040 useuse->set_req(k, r); 2041 uses_found++; 2042 if (useuse->is_Loop() && k == LoopNode::EntryControl) { 2043 assert(dom_depth(useuse) > dd_r , ""); 2044 set_idom(useuse, r, dom_depth(useuse)); 2045 } 2046 } 2047 } 2048 l -= uses_found; // we deleted 1 or more copies of this edge 2049 } 2050 2051 // Now finish up 'r' 2052 r->set_req( 1, newuse ); 2053 r->set_req( 2, use ); 2054 _igvn.register_new_node_with_optimizer(r); 2055 set_loop(r, use_loop); 2056 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r); 2057 } // End of if a loop-exit test 2058 } 2059 } 2060 2061 // Step 4: If loop-invariant use is not control, it must be dominated by a 2062 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region 2063 // there if needed. Make a Phi there merging old and new used values. 2064 Node_List *split_if_set = NULL; 2065 Node_List *split_bool_set = NULL; 2066 Node_List *split_cex_set = NULL; 2067 for( i = 0; i < loop->_body.size(); i++ ) { 2068 Node* old = loop->_body.at(i); 2069 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set, 2070 split_bool_set, split_cex_set, worklist, new_counter, 2071 mode); 2072 } 2073 2074 for (i = 0; i < extra_data_nodes.size(); i++) { 2075 Node* old = extra_data_nodes.at(i); 2076 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set, 2077 split_bool_set, split_cex_set, worklist, new_counter, 2078 mode); 2079 } 2080 2081 // Check for IFs that need splitting/cloning. Happens if an IF outside of 2082 // the loop uses a condition set in the loop. The original IF probably 2083 // takes control from one or more OLD Regions (which in turn get from NEW 2084 // Regions). In any case, there will be a set of Phis for each merge point 2085 // from the IF up to where the original BOOL def exists the loop. 2086 if (split_if_set) { 2087 while (split_if_set->size()) { 2088 Node *iff = split_if_set->pop(); 2089 if (iff->in(1)->is_Phi()) { 2090 Node *b = clone_iff(iff->in(1)->as_Phi(), loop); 2091 _igvn.replace_input_of(iff, 1, b); 2092 } 2093 } 2094 } 2095 if (split_bool_set) { 2096 while (split_bool_set->size()) { 2097 Node *b = split_bool_set->pop(); 2098 Node *phi = b->in(1); 2099 assert(phi->is_Phi(), ""); 2100 CmpNode *cmp = clone_bool((PhiNode*)phi, loop); 2101 _igvn.replace_input_of(b, 1, cmp); 2102 } 2103 } 2104 if (split_cex_set) { 2105 while (split_cex_set->size()) { 2106 Node *b = split_cex_set->pop(); 2107 assert(b->in(0)->is_Region(), ""); 2108 assert(b->in(1)->is_Phi(), ""); 2109 assert(b->in(0)->in(0) == b->in(1)->in(0), ""); 2110 split_up(b, b->in(0), NULL); 2111 } 2112 } 2113 2114 } 2115 2116 2117 //---------------------- stride_of_possible_iv ------------------------------------- 2118 // Looks for an iff/bool/comp with one operand of the compare 2119 // being a cycle involving an add and a phi, 2120 // with an optional truncation (left-shift followed by a right-shift) 2121 // of the add. Returns zero if not an iv. 2122 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) { 2123 Node* trunc1 = NULL; 2124 Node* trunc2 = NULL; 2125 const TypeInt* ttype = NULL; 2126 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) { 2127 return 0; 2128 } 2129 BoolNode* bl = iff->in(1)->as_Bool(); 2130 Node* cmp = bl->in(1); 2131 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) { 2132 return 0; 2133 } 2134 // Must have an invariant operand 2135 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) { 2136 return 0; 2137 } 2138 Node* add2 = NULL; 2139 Node* cmp1 = cmp->in(1); 2140 if (cmp1->is_Phi()) { 2141 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) ))) 2142 Node* phi = cmp1; 2143 for (uint i = 1; i < phi->req(); i++) { 2144 Node* in = phi->in(i); 2145 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in, 2146 &trunc1, &trunc2, &ttype); 2147 if (add && add->in(1) == phi) { 2148 add2 = add->in(2); 2149 break; 2150 } 2151 } 2152 } else { 2153 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) ))) 2154 Node* addtrunc = cmp1; 2155 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc, 2156 &trunc1, &trunc2, &ttype); 2157 if (add && add->in(1)->is_Phi()) { 2158 Node* phi = add->in(1); 2159 for (uint i = 1; i < phi->req(); i++) { 2160 if (phi->in(i) == addtrunc) { 2161 add2 = add->in(2); 2162 break; 2163 } 2164 } 2165 } 2166 } 2167 if (add2 != NULL) { 2168 const TypeInt* add2t = _igvn.type(add2)->is_int(); 2169 if (add2t->is_con()) { 2170 return add2t->get_con(); 2171 } 2172 } 2173 return 0; 2174 } 2175 2176 2177 //---------------------- stay_in_loop ------------------------------------- 2178 // Return the (unique) control output node that's in the loop (if it exists.) 2179 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) { 2180 Node* unique = NULL; 2181 if (!n) return NULL; 2182 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2183 Node* use = n->fast_out(i); 2184 if (!has_ctrl(use) && loop->is_member(get_loop(use))) { 2185 if (unique != NULL) { 2186 return NULL; 2187 } 2188 unique = use; 2189 } 2190 } 2191 return unique; 2192 } 2193 2194 //------------------------------ register_node ------------------------------------- 2195 // Utility to register node "n" with PhaseIdealLoop 2196 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) { 2197 _igvn.register_new_node_with_optimizer(n); 2198 loop->_body.push(n); 2199 if (n->is_CFG()) { 2200 set_loop(n, loop); 2201 set_idom(n, pred, ddepth); 2202 } else { 2203 set_ctrl(n, pred); 2204 } 2205 } 2206 2207 //------------------------------ proj_clone ------------------------------------- 2208 // Utility to create an if-projection 2209 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) { 2210 ProjNode* c = p->clone()->as_Proj(); 2211 c->set_req(0, iff); 2212 return c; 2213 } 2214 2215 //------------------------------ short_circuit_if ------------------------------------- 2216 // Force the iff control output to be the live_proj 2217 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) { 2218 guarantee(live_proj != NULL, "null projection"); 2219 int proj_con = live_proj->_con; 2220 assert(proj_con == 0 || proj_con == 1, "false or true projection"); 2221 Node *con = _igvn.intcon(proj_con); 2222 set_ctrl(con, C->root()); 2223 if (iff) { 2224 iff->set_req(1, con); 2225 } 2226 return con; 2227 } 2228 2229 //------------------------------ insert_if_before_proj ------------------------------------- 2230 // Insert a new if before an if projection (* - new node) 2231 // 2232 // before 2233 // if(test) 2234 // / \ 2235 // v v 2236 // other-proj proj (arg) 2237 // 2238 // after 2239 // if(test) 2240 // / \ 2241 // / v 2242 // | * proj-clone 2243 // v | 2244 // other-proj v 2245 // * new_if(relop(cmp[IU](left,right))) 2246 // / \ 2247 // v v 2248 // * new-proj proj 2249 // (returned) 2250 // 2251 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) { 2252 IfNode* iff = proj->in(0)->as_If(); 2253 IdealLoopTree *loop = get_loop(proj); 2254 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2255 int ddepth = dom_depth(proj); 2256 2257 _igvn.rehash_node_delayed(iff); 2258 _igvn.rehash_node_delayed(proj); 2259 2260 proj->set_req(0, NULL); // temporary disconnect 2261 ProjNode* proj2 = proj_clone(proj, iff); 2262 register_node(proj2, loop, iff, ddepth); 2263 2264 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right); 2265 register_node(cmp, loop, proj2, ddepth); 2266 2267 BoolNode* bol = new BoolNode(cmp, relop); 2268 register_node(bol, loop, proj2, ddepth); 2269 2270 int opcode = iff->Opcode(); 2271 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode"); 2272 IfNode* new_if = (opcode == Op_If) ? new IfNode(proj2, bol, iff->_prob, iff->_fcnt): 2273 new RangeCheckNode(proj2, bol, iff->_prob, iff->_fcnt); 2274 register_node(new_if, loop, proj2, ddepth); 2275 2276 proj->set_req(0, new_if); // reattach 2277 set_idom(proj, new_if, ddepth); 2278 2279 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj(); 2280 guarantee(new_exit != NULL, "null exit node"); 2281 register_node(new_exit, get_loop(other_proj), new_if, ddepth); 2282 2283 return new_exit; 2284 } 2285 2286 //------------------------------ insert_region_before_proj ------------------------------------- 2287 // Insert a region before an if projection (* - new node) 2288 // 2289 // before 2290 // if(test) 2291 // / | 2292 // v | 2293 // proj v 2294 // other-proj 2295 // 2296 // after 2297 // if(test) 2298 // / | 2299 // v | 2300 // * proj-clone v 2301 // | other-proj 2302 // v 2303 // * new-region 2304 // | 2305 // v 2306 // * dum_if 2307 // / \ 2308 // v \ 2309 // * dum-proj v 2310 // proj 2311 // 2312 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) { 2313 IfNode* iff = proj->in(0)->as_If(); 2314 IdealLoopTree *loop = get_loop(proj); 2315 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2316 int ddepth = dom_depth(proj); 2317 2318 _igvn.rehash_node_delayed(iff); 2319 _igvn.rehash_node_delayed(proj); 2320 2321 proj->set_req(0, NULL); // temporary disconnect 2322 ProjNode* proj2 = proj_clone(proj, iff); 2323 register_node(proj2, loop, iff, ddepth); 2324 2325 RegionNode* reg = new RegionNode(2); 2326 reg->set_req(1, proj2); 2327 register_node(reg, loop, iff, ddepth); 2328 2329 IfNode* dum_if = new IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt); 2330 register_node(dum_if, loop, reg, ddepth); 2331 2332 proj->set_req(0, dum_if); // reattach 2333 set_idom(proj, dum_if, ddepth); 2334 2335 ProjNode* dum_proj = proj_clone(other_proj, dum_if); 2336 register_node(dum_proj, loop, dum_if, ddepth); 2337 2338 return reg; 2339 } 2340 2341 //------------------------------ insert_cmpi_loop_exit ------------------------------------- 2342 // Clone a signed compare loop exit from an unsigned compare and 2343 // insert it before the unsigned cmp on the stay-in-loop path. 2344 // All new nodes inserted in the dominator tree between the original 2345 // if and it's projections. The original if test is replaced with 2346 // a constant to force the stay-in-loop path. 2347 // 2348 // This is done to make sure that the original if and it's projections 2349 // still dominate the same set of control nodes, that the ctrl() relation 2350 // from data nodes to them is preserved, and that their loop nesting is 2351 // preserved. 2352 // 2353 // before 2354 // if(i <u limit) unsigned compare loop exit 2355 // / | 2356 // v v 2357 // exit-proj stay-in-loop-proj 2358 // 2359 // after 2360 // if(stay-in-loop-const) original if 2361 // / | 2362 // / v 2363 // / if(i < limit) new signed test 2364 // / / | 2365 // / / v 2366 // / / if(i <u limit) new cloned unsigned test 2367 // / / / | 2368 // v v v | 2369 // region | 2370 // | | 2371 // dum-if | 2372 // / | | 2373 // ether | | 2374 // v v 2375 // exit-proj stay-in-loop-proj 2376 // 2377 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) { 2378 const bool Signed = true; 2379 const bool Unsigned = false; 2380 2381 BoolNode* bol = if_cmpu->in(1)->as_Bool(); 2382 if (bol->_test._test != BoolTest::lt) return NULL; 2383 CmpNode* cmpu = bol->in(1)->as_Cmp(); 2384 if (cmpu->Opcode() != Op_CmpU) return NULL; 2385 int stride = stride_of_possible_iv(if_cmpu); 2386 if (stride == 0) return NULL; 2387 2388 Node* lp_proj = stay_in_loop(if_cmpu, loop); 2389 guarantee(lp_proj != NULL, "null loop node"); 2390 2391 ProjNode* lp_continue = lp_proj->as_Proj(); 2392 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj(); 2393 2394 Node* limit = NULL; 2395 if (stride > 0) { 2396 limit = cmpu->in(2); 2397 } else { 2398 limit = _igvn.makecon(TypeInt::ZERO); 2399 set_ctrl(limit, C->root()); 2400 } 2401 // Create a new region on the exit path 2402 RegionNode* reg = insert_region_before_proj(lp_exit); 2403 guarantee(reg != NULL, "null region node"); 2404 2405 // Clone the if-cmpu-true-false using a signed compare 2406 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge; 2407 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue); 2408 reg->add_req(cmpi_exit); 2409 2410 // Clone the if-cmpu-true-false 2411 BoolTest::mask rel_u = bol->_test._test; 2412 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue); 2413 reg->add_req(cmpu_exit); 2414 2415 // Force original if to stay in loop. 2416 short_circuit_if(if_cmpu, lp_continue); 2417 2418 return cmpi_exit->in(0)->as_If(); 2419 } 2420 2421 //------------------------------ remove_cmpi_loop_exit ------------------------------------- 2422 // Remove a previously inserted signed compare loop exit. 2423 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) { 2424 Node* lp_proj = stay_in_loop(if_cmp, loop); 2425 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI && 2426 stay_in_loop(lp_proj, loop)->is_If() && 2427 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu"); 2428 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO); 2429 set_ctrl(con, C->root()); 2430 if_cmp->set_req(1, con); 2431 } 2432 2433 //------------------------------ scheduled_nodelist ------------------------------------- 2434 // Create a post order schedule of nodes that are in the 2435 // "member" set. The list is returned in "sched". 2436 // The first node in "sched" is the loop head, followed by 2437 // nodes which have no inputs in the "member" set, and then 2438 // followed by the nodes that have an immediate input dependence 2439 // on a node in "sched". 2440 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) { 2441 2442 assert(member.test(loop->_head->_idx), "loop head must be in member set"); 2443 Arena *a = Thread::current()->resource_area(); 2444 VectorSet visited(a); 2445 Node_Stack nstack(a, loop->_body.size()); 2446 2447 Node* n = loop->_head; // top of stack is cached in "n" 2448 uint idx = 0; 2449 visited.set(n->_idx); 2450 2451 // Initially push all with no inputs from within member set 2452 for(uint i = 0; i < loop->_body.size(); i++ ) { 2453 Node *elt = loop->_body.at(i); 2454 if (member.test(elt->_idx)) { 2455 bool found = false; 2456 for (uint j = 0; j < elt->req(); j++) { 2457 Node* def = elt->in(j); 2458 if (def && member.test(def->_idx) && def != elt) { 2459 found = true; 2460 break; 2461 } 2462 } 2463 if (!found && elt != loop->_head) { 2464 nstack.push(n, idx); 2465 n = elt; 2466 assert(!visited.test(n->_idx), "not seen yet"); 2467 visited.set(n->_idx); 2468 } 2469 } 2470 } 2471 2472 // traverse out's that are in the member set 2473 while (true) { 2474 if (idx < n->outcnt()) { 2475 Node* use = n->raw_out(idx); 2476 idx++; 2477 if (!visited.test_set(use->_idx)) { 2478 if (member.test(use->_idx)) { 2479 nstack.push(n, idx); 2480 n = use; 2481 idx = 0; 2482 } 2483 } 2484 } else { 2485 // All outputs processed 2486 sched.push(n); 2487 if (nstack.is_empty()) break; 2488 n = nstack.node(); 2489 idx = nstack.index(); 2490 nstack.pop(); 2491 } 2492 } 2493 } 2494 2495 2496 //------------------------------ has_use_in_set ------------------------------------- 2497 // Has a use in the vector set 2498 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) { 2499 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2500 Node* use = n->fast_out(j); 2501 if (vset.test(use->_idx)) { 2502 return true; 2503 } 2504 } 2505 return false; 2506 } 2507 2508 2509 //------------------------------ has_use_internal_to_set ------------------------------------- 2510 // Has use internal to the vector set (ie. not in a phi at the loop head) 2511 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) { 2512 Node* head = loop->_head; 2513 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2514 Node* use = n->fast_out(j); 2515 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) { 2516 return true; 2517 } 2518 } 2519 return false; 2520 } 2521 2522 2523 //------------------------------ clone_for_use_outside_loop ------------------------------------- 2524 // clone "n" for uses that are outside of loop 2525 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) { 2526 int cloned = 0; 2527 assert(worklist.size() == 0, "should be empty"); 2528 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2529 Node* use = n->fast_out(j); 2530 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) { 2531 worklist.push(use); 2532 } 2533 } 2534 while( worklist.size() ) { 2535 Node *use = worklist.pop(); 2536 if (!has_node(use) || use->in(0) == C->top()) continue; 2537 uint j; 2538 for (j = 0; j < use->req(); j++) { 2539 if (use->in(j) == n) break; 2540 } 2541 assert(j < use->req(), "must be there"); 2542 2543 // clone "n" and insert it between the inputs of "n" and the use outside the loop 2544 Node* n_clone = n->clone(); 2545 _igvn.replace_input_of(use, j, n_clone); 2546 cloned++; 2547 Node* use_c; 2548 if (!use->is_Phi()) { 2549 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0); 2550 } else { 2551 // Use in a phi is considered a use in the associated predecessor block 2552 use_c = use->in(0)->in(j); 2553 } 2554 set_ctrl(n_clone, use_c); 2555 assert(!loop->is_member(get_loop(use_c)), "should be outside loop"); 2556 get_loop(use_c)->_body.push(n_clone); 2557 _igvn.register_new_node_with_optimizer(n_clone); 2558 #if !defined(PRODUCT) 2559 if (TracePartialPeeling) { 2560 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx); 2561 } 2562 #endif 2563 } 2564 return cloned; 2565 } 2566 2567 2568 //------------------------------ clone_for_special_use_inside_loop ------------------------------------- 2569 // clone "n" for special uses that are in the not_peeled region. 2570 // If these def-uses occur in separate blocks, the code generator 2571 // marks the method as not compilable. For example, if a "BoolNode" 2572 // is in a different basic block than the "IfNode" that uses it, then 2573 // the compilation is aborted in the code generator. 2574 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 2575 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) { 2576 if (n->is_Phi() || n->is_Load()) { 2577 return; 2578 } 2579 assert(worklist.size() == 0, "should be empty"); 2580 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2581 Node* use = n->fast_out(j); 2582 if ( not_peel.test(use->_idx) && 2583 (use->is_If() || use->is_CMove() || use->is_Bool()) && 2584 use->in(1) == n) { 2585 worklist.push(use); 2586 } 2587 } 2588 if (worklist.size() > 0) { 2589 // clone "n" and insert it between inputs of "n" and the use 2590 Node* n_clone = n->clone(); 2591 loop->_body.push(n_clone); 2592 _igvn.register_new_node_with_optimizer(n_clone); 2593 set_ctrl(n_clone, get_ctrl(n)); 2594 sink_list.push(n_clone); 2595 not_peel <<= n_clone->_idx; // add n_clone to not_peel set. 2596 #if !defined(PRODUCT) 2597 if (TracePartialPeeling) { 2598 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx); 2599 } 2600 #endif 2601 while( worklist.size() ) { 2602 Node *use = worklist.pop(); 2603 _igvn.rehash_node_delayed(use); 2604 for (uint j = 1; j < use->req(); j++) { 2605 if (use->in(j) == n) { 2606 use->set_req(j, n_clone); 2607 } 2608 } 2609 } 2610 } 2611 } 2612 2613 2614 //------------------------------ insert_phi_for_loop ------------------------------------- 2615 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 2616 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) { 2617 Node *phi = PhiNode::make(lp, back_edge_val); 2618 phi->set_req(LoopNode::EntryControl, lp_entry_val); 2619 // Use existing phi if it already exists 2620 Node *hit = _igvn.hash_find_insert(phi); 2621 if( hit == NULL ) { 2622 _igvn.register_new_node_with_optimizer(phi); 2623 set_ctrl(phi, lp); 2624 } else { 2625 // Remove the new phi from the graph and use the hit 2626 _igvn.remove_dead_node(phi); 2627 phi = hit; 2628 } 2629 _igvn.replace_input_of(use, idx, phi); 2630 } 2631 2632 #ifdef ASSERT 2633 //------------------------------ is_valid_loop_partition ------------------------------------- 2634 // Validate the loop partition sets: peel and not_peel 2635 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, 2636 VectorSet& not_peel ) { 2637 uint i; 2638 // Check that peel_list entries are in the peel set 2639 for (i = 0; i < peel_list.size(); i++) { 2640 if (!peel.test(peel_list.at(i)->_idx)) { 2641 return false; 2642 } 2643 } 2644 // Check at loop members are in one of peel set or not_peel set 2645 for (i = 0; i < loop->_body.size(); i++ ) { 2646 Node *def = loop->_body.at(i); 2647 uint di = def->_idx; 2648 // Check that peel set elements are in peel_list 2649 if (peel.test(di)) { 2650 if (not_peel.test(di)) { 2651 return false; 2652 } 2653 // Must be in peel_list also 2654 bool found = false; 2655 for (uint j = 0; j < peel_list.size(); j++) { 2656 if (peel_list.at(j)->_idx == di) { 2657 found = true; 2658 break; 2659 } 2660 } 2661 if (!found) { 2662 return false; 2663 } 2664 } else if (not_peel.test(di)) { 2665 if (peel.test(di)) { 2666 return false; 2667 } 2668 } else { 2669 return false; 2670 } 2671 } 2672 return true; 2673 } 2674 2675 //------------------------------ is_valid_clone_loop_exit_use ------------------------------------- 2676 // Ensure a use outside of loop is of the right form 2677 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) { 2678 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2679 return (use->is_Phi() && 2680 use_c->is_Region() && use_c->req() == 3 && 2681 (use_c->in(exit_idx)->Opcode() == Op_IfTrue || 2682 use_c->in(exit_idx)->Opcode() == Op_IfFalse || 2683 use_c->in(exit_idx)->Opcode() == Op_JumpProj) && 2684 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) ); 2685 } 2686 2687 //------------------------------ is_valid_clone_loop_form ------------------------------------- 2688 // Ensure that all uses outside of loop are of the right form 2689 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 2690 uint orig_exit_idx, uint clone_exit_idx) { 2691 uint len = peel_list.size(); 2692 for (uint i = 0; i < len; i++) { 2693 Node *def = peel_list.at(i); 2694 2695 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 2696 Node *use = def->fast_out(j); 2697 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2698 if (!loop->is_member(get_loop(use_c))) { 2699 // use is not in the loop, check for correct structure 2700 if (use->in(0) == def) { 2701 // Okay 2702 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) { 2703 return false; 2704 } 2705 } 2706 } 2707 } 2708 return true; 2709 } 2710 #endif 2711 2712 //------------------------------ partial_peel ------------------------------------- 2713 // Partially peel (aka loop rotation) the top portion of a loop (called 2714 // the peel section below) by cloning it and placing one copy just before 2715 // the new loop head and the other copy at the bottom of the new loop. 2716 // 2717 // before after where it came from 2718 // 2719 // stmt1 stmt1 2720 // loop: stmt2 clone 2721 // stmt2 if condA goto exitA clone 2722 // if condA goto exitA new_loop: new 2723 // stmt3 stmt3 clone 2724 // if !condB goto loop if condB goto exitB clone 2725 // exitB: stmt2 orig 2726 // stmt4 if !condA goto new_loop orig 2727 // exitA: goto exitA 2728 // exitB: 2729 // stmt4 2730 // exitA: 2731 // 2732 // Step 1: find the cut point: an exit test on probable 2733 // induction variable. 2734 // Step 2: schedule (with cloning) operations in the peel 2735 // section that can be executed after the cut into 2736 // the section that is not peeled. This may need 2737 // to clone operations into exit blocks. For 2738 // instance, a reference to A[i] in the not-peel 2739 // section and a reference to B[i] in an exit block 2740 // may cause a left-shift of i by 2 to be placed 2741 // in the peel block. This step will clone the left 2742 // shift into the exit block and sink the left shift 2743 // from the peel to the not-peel section. 2744 // Step 3: clone the loop, retarget the control, and insert 2745 // phis for values that are live across the new loop 2746 // head. This is very dependent on the graph structure 2747 // from clone_loop. It creates region nodes for 2748 // exit control and associated phi nodes for values 2749 // flow out of the loop through that exit. The region 2750 // node is dominated by the clone's control projection. 2751 // So the clone's peel section is placed before the 2752 // new loop head, and the clone's not-peel section is 2753 // forms the top part of the new loop. The original 2754 // peel section forms the tail of the new loop. 2755 // Step 4: update the dominator tree and recompute the 2756 // dominator depth. 2757 // 2758 // orig 2759 // 2760 // stmt1 2761 // | 2762 // v 2763 // loop predicate 2764 // | 2765 // v 2766 // loop<----+ 2767 // | | 2768 // stmt2 | 2769 // | | 2770 // v | 2771 // ifA | 2772 // / | | 2773 // v v | 2774 // false true ^ <-- last_peel 2775 // / | | 2776 // / ===|==cut | 2777 // / stmt3 | <-- first_not_peel 2778 // / | | 2779 // | v | 2780 // v ifB | 2781 // exitA: / \ | 2782 // / \ | 2783 // v v | 2784 // false true | 2785 // / \ | 2786 // / ----+ 2787 // | 2788 // v 2789 // exitB: 2790 // stmt4 2791 // 2792 // 2793 // after clone loop 2794 // 2795 // stmt1 2796 // | 2797 // v 2798 // loop predicate 2799 // / \ 2800 // clone / \ orig 2801 // / \ 2802 // / \ 2803 // v v 2804 // +---->loop loop<----+ 2805 // | | | | 2806 // | stmt2 stmt2 | 2807 // | | | | 2808 // | v v | 2809 // | ifA ifA | 2810 // | | \ / | | 2811 // | v v v v | 2812 // ^ true false false true ^ <-- last_peel 2813 // | | ^ \ / | | 2814 // | cut==|== \ \ / ===|==cut | 2815 // | stmt3 \ \ / stmt3 | <-- first_not_peel 2816 // | | dom | | | | 2817 // | v \ 1v v2 v | 2818 // | ifB regionA ifB | 2819 // | / \ | / \ | 2820 // | / \ v / \ | 2821 // | v v exitA: v v | 2822 // | true false false true | 2823 // | / ^ \ / \ | 2824 // +---- \ \ / ----+ 2825 // dom \ / 2826 // \ 1v v2 2827 // regionB 2828 // | 2829 // v 2830 // exitB: 2831 // stmt4 2832 // 2833 // 2834 // after partial peel 2835 // 2836 // stmt1 2837 // | 2838 // v 2839 // loop predicate 2840 // / 2841 // clone / orig 2842 // / TOP 2843 // / \ 2844 // v v 2845 // TOP->loop loop----+ 2846 // | | | 2847 // stmt2 stmt2 | 2848 // | | | 2849 // v v | 2850 // ifA ifA | 2851 // | \ / | | 2852 // v v v v | 2853 // true false false true | <-- last_peel 2854 // | ^ \ / +------|---+ 2855 // +->newloop \ \ / === ==cut | | 2856 // | stmt3 \ \ / TOP | | 2857 // | | dom | | stmt3 | | <-- first_not_peel 2858 // | v \ 1v v2 v | | 2859 // | ifB regionA ifB ^ v 2860 // | / \ | / \ | | 2861 // | / \ v / \ | | 2862 // | v v exitA: v v | | 2863 // | true false false true | | 2864 // | / ^ \ / \ | | 2865 // | | \ \ / v | | 2866 // | | dom \ / TOP | | 2867 // | | \ 1v v2 | | 2868 // ^ v regionB | | 2869 // | | | | | 2870 // | | v ^ v 2871 // | | exitB: | | 2872 // | | stmt4 | | 2873 // | +------------>-----------------+ | 2874 // | | 2875 // +-----------------<---------------------+ 2876 // 2877 // 2878 // final graph 2879 // 2880 // stmt1 2881 // | 2882 // v 2883 // loop predicate 2884 // | 2885 // v 2886 // stmt2 clone 2887 // | 2888 // v 2889 // ........> ifA clone 2890 // : / | 2891 // dom / | 2892 // : v v 2893 // : false true 2894 // : | | 2895 // : | v 2896 // : | newloop<-----+ 2897 // : | | | 2898 // : | stmt3 clone | 2899 // : | | | 2900 // : | v | 2901 // : | ifB | 2902 // : | / \ | 2903 // : | v v | 2904 // : | false true | 2905 // : | | | | 2906 // : | v stmt2 | 2907 // : | exitB: | | 2908 // : | stmt4 v | 2909 // : | ifA orig | 2910 // : | / \ | 2911 // : | / \ | 2912 // : | v v | 2913 // : | false true | 2914 // : | / \ | 2915 // : v v -----+ 2916 // RegionA 2917 // | 2918 // v 2919 // exitA 2920 // 2921 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) { 2922 2923 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only"); 2924 if (!loop->_head->is_Loop()) { 2925 return false; } 2926 2927 LoopNode *head = loop->_head->as_Loop(); 2928 2929 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) { 2930 return false; 2931 } 2932 2933 // Check for complex exit control 2934 for(uint ii = 0; ii < loop->_body.size(); ii++ ) { 2935 Node *n = loop->_body.at(ii); 2936 int opc = n->Opcode(); 2937 if (n->is_Call() || 2938 opc == Op_Catch || 2939 opc == Op_CatchProj || 2940 opc == Op_Jump || 2941 opc == Op_JumpProj) { 2942 #if !defined(PRODUCT) 2943 if (TracePartialPeeling) { 2944 tty->print_cr("\nExit control too complex: lp: %d", head->_idx); 2945 } 2946 #endif 2947 return false; 2948 } 2949 } 2950 2951 int dd = dom_depth(head); 2952 2953 // Step 1: find cut point 2954 2955 // Walk up dominators to loop head looking for first loop exit 2956 // which is executed on every path thru loop. 2957 IfNode *peel_if = NULL; 2958 IfNode *peel_if_cmpu = NULL; 2959 2960 Node *iff = loop->tail(); 2961 while( iff != head ) { 2962 if( iff->is_If() ) { 2963 Node *ctrl = get_ctrl(iff->in(1)); 2964 if (ctrl->is_top()) return false; // Dead test on live IF. 2965 // If loop-varying exit-test, check for induction variable 2966 if( loop->is_member(get_loop(ctrl)) && 2967 loop->is_loop_exit(iff) && 2968 is_possible_iv_test(iff)) { 2969 Node* cmp = iff->in(1)->in(1); 2970 if (cmp->Opcode() == Op_CmpI) { 2971 peel_if = iff->as_If(); 2972 } else { 2973 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU"); 2974 peel_if_cmpu = iff->as_If(); 2975 } 2976 } 2977 } 2978 iff = idom(iff); 2979 } 2980 // Prefer signed compare over unsigned compare. 2981 IfNode* new_peel_if = NULL; 2982 if (peel_if == NULL) { 2983 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) { 2984 return false; // No peel point found 2985 } 2986 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop); 2987 if (new_peel_if == NULL) { 2988 return false; // No peel point found 2989 } 2990 peel_if = new_peel_if; 2991 } 2992 Node* last_peel = stay_in_loop(peel_if, loop); 2993 Node* first_not_peeled = stay_in_loop(last_peel, loop); 2994 if (first_not_peeled == NULL || first_not_peeled == head) { 2995 return false; 2996 } 2997 2998 #if !defined(PRODUCT) 2999 if (TraceLoopOpts) { 3000 tty->print("PartialPeel "); 3001 loop->dump_head(); 3002 } 3003 3004 if (TracePartialPeeling) { 3005 tty->print_cr("before partial peel one iteration"); 3006 Node_List wl; 3007 Node* t = head->in(2); 3008 while (true) { 3009 wl.push(t); 3010 if (t == head) break; 3011 t = idom(t); 3012 } 3013 while (wl.size() > 0) { 3014 Node* tt = wl.pop(); 3015 tt->dump(); 3016 if (tt == last_peel) tty->print_cr("-- cut --"); 3017 } 3018 } 3019 #endif 3020 ResourceArea *area = Thread::current()->resource_area(); 3021 VectorSet peel(area); 3022 VectorSet not_peel(area); 3023 Node_List peel_list(area); 3024 Node_List worklist(area); 3025 Node_List sink_list(area); 3026 3027 // Set of cfg nodes to peel are those that are executable from 3028 // the head through last_peel. 3029 assert(worklist.size() == 0, "should be empty"); 3030 worklist.push(head); 3031 peel.set(head->_idx); 3032 while (worklist.size() > 0) { 3033 Node *n = worklist.pop(); 3034 if (n != last_peel) { 3035 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3036 Node* use = n->fast_out(j); 3037 if (use->is_CFG() && 3038 loop->is_member(get_loop(use)) && 3039 !peel.test_set(use->_idx)) { 3040 worklist.push(use); 3041 } 3042 } 3043 } 3044 } 3045 3046 // Set of non-cfg nodes to peel are those that are control 3047 // dependent on the cfg nodes. 3048 uint i; 3049 for(i = 0; i < loop->_body.size(); i++ ) { 3050 Node *n = loop->_body.at(i); 3051 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n; 3052 if (peel.test(n_c->_idx)) { 3053 peel.set(n->_idx); 3054 } else { 3055 not_peel.set(n->_idx); 3056 } 3057 } 3058 3059 // Step 2: move operations from the peeled section down into the 3060 // not-peeled section 3061 3062 // Get a post order schedule of nodes in the peel region 3063 // Result in right-most operand. 3064 scheduled_nodelist(loop, peel, peel_list ); 3065 3066 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 3067 3068 // For future check for too many new phis 3069 uint old_phi_cnt = 0; 3070 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 3071 Node* use = head->fast_out(j); 3072 if (use->is_Phi()) old_phi_cnt++; 3073 } 3074 3075 #if !defined(PRODUCT) 3076 if (TracePartialPeeling) { 3077 tty->print_cr("\npeeled list"); 3078 } 3079 #endif 3080 3081 // Evacuate nodes in peel region into the not_peeled region if possible 3082 uint new_phi_cnt = 0; 3083 uint cloned_for_outside_use = 0; 3084 for (i = 0; i < peel_list.size();) { 3085 Node* n = peel_list.at(i); 3086 #if !defined(PRODUCT) 3087 if (TracePartialPeeling) n->dump(); 3088 #endif 3089 bool incr = true; 3090 if ( !n->is_CFG() ) { 3091 3092 if ( has_use_in_set(n, not_peel) ) { 3093 3094 // If not used internal to the peeled region, 3095 // move "n" from peeled to not_peeled region. 3096 3097 if ( !has_use_internal_to_set(n, peel, loop) ) { 3098 3099 // if not pinned and not a load (which maybe anti-dependent on a store) 3100 // and not a CMove (Matcher expects only bool->cmove). 3101 if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) { 3102 cloned_for_outside_use += clone_for_use_outside_loop( loop, n, worklist ); 3103 sink_list.push(n); 3104 peel >>= n->_idx; // delete n from peel set. 3105 not_peel <<= n->_idx; // add n to not_peel set. 3106 peel_list.remove(i); 3107 incr = false; 3108 #if !defined(PRODUCT) 3109 if (TracePartialPeeling) { 3110 tty->print_cr("sink to not_peeled region: %d newbb: %d", 3111 n->_idx, get_ctrl(n)->_idx); 3112 } 3113 #endif 3114 } 3115 } else { 3116 // Otherwise check for special def-use cases that span 3117 // the peel/not_peel boundary such as bool->if 3118 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist ); 3119 new_phi_cnt++; 3120 } 3121 } 3122 } 3123 if (incr) i++; 3124 } 3125 3126 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) { 3127 #if !defined(PRODUCT) 3128 if (TracePartialPeeling) { 3129 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c", 3130 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F'); 3131 } 3132 #endif 3133 if (new_peel_if != NULL) { 3134 remove_cmpi_loop_exit(new_peel_if, loop); 3135 } 3136 // Inhibit more partial peeling on this loop 3137 assert(!head->is_partial_peel_loop(), "not partial peeled"); 3138 head->mark_partial_peel_failed(); 3139 if (cloned_for_outside_use > 0) { 3140 // Terminate this round of loop opts because 3141 // the graph outside this loop was changed. 3142 C->set_major_progress(); 3143 return true; 3144 } 3145 return false; 3146 } 3147 3148 // Step 3: clone loop, retarget control, and insert new phis 3149 3150 // Create new loop head for new phis and to hang 3151 // the nodes being moved (sinked) from the peel region. 3152 LoopNode* new_head = new LoopNode(last_peel, last_peel); 3153 new_head->set_unswitch_count(head->unswitch_count()); // Preserve 3154 _igvn.register_new_node_with_optimizer(new_head); 3155 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled"); 3156 _igvn.replace_input_of(first_not_peeled, 0, new_head); 3157 set_loop(new_head, loop); 3158 loop->_body.push(new_head); 3159 not_peel.set(new_head->_idx); 3160 set_idom(new_head, last_peel, dom_depth(first_not_peeled)); 3161 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled)); 3162 3163 while (sink_list.size() > 0) { 3164 Node* n = sink_list.pop(); 3165 set_ctrl(n, new_head); 3166 } 3167 3168 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 3169 3170 clone_loop(loop, old_new, dd, IgnoreStripMined); 3171 3172 const uint clone_exit_idx = 1; 3173 const uint orig_exit_idx = 2; 3174 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop"); 3175 3176 Node* head_clone = old_new[head->_idx]; 3177 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop(); 3178 Node* orig_tail_clone = head_clone->in(2); 3179 3180 // Add phi if "def" node is in peel set and "use" is not 3181 3182 for(i = 0; i < peel_list.size(); i++ ) { 3183 Node *def = peel_list.at(i); 3184 if (!def->is_CFG()) { 3185 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 3186 Node *use = def->fast_out(j); 3187 if (has_node(use) && use->in(0) != C->top() && 3188 (!peel.test(use->_idx) || 3189 (use->is_Phi() && use->in(0) == head)) ) { 3190 worklist.push(use); 3191 } 3192 } 3193 while( worklist.size() ) { 3194 Node *use = worklist.pop(); 3195 for (uint j = 1; j < use->req(); j++) { 3196 Node* n = use->in(j); 3197 if (n == def) { 3198 3199 // "def" is in peel set, "use" is not in peel set 3200 // or "use" is in the entry boundary (a phi) of the peel set 3201 3202 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use; 3203 3204 if ( loop->is_member(get_loop( use_c )) ) { 3205 // use is in loop 3206 if (old_new[use->_idx] != NULL) { // null for dead code 3207 Node* use_clone = old_new[use->_idx]; 3208 _igvn.replace_input_of(use, j, C->top()); 3209 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone ); 3210 } 3211 } else { 3212 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format"); 3213 // use is not in the loop, check if the live range includes the cut 3214 Node* lp_if = use_c->in(orig_exit_idx)->in(0); 3215 if (not_peel.test(lp_if->_idx)) { 3216 assert(j == orig_exit_idx, "use from original loop"); 3217 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone ); 3218 } 3219 } 3220 } 3221 } 3222 } 3223 } 3224 } 3225 3226 // Step 3b: retarget control 3227 3228 // Redirect control to the new loop head if a cloned node in 3229 // the not_peeled region has control that points into the peeled region. 3230 // This necessary because the cloned peeled region will be outside 3231 // the loop. 3232 // from to 3233 // cloned-peeled <---+ 3234 // new_head_clone: | <--+ 3235 // cloned-not_peeled in(0) in(0) 3236 // orig-peeled 3237 3238 for(i = 0; i < loop->_body.size(); i++ ) { 3239 Node *n = loop->_body.at(i); 3240 if (!n->is_CFG() && n->in(0) != NULL && 3241 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) { 3242 Node* n_clone = old_new[n->_idx]; 3243 _igvn.replace_input_of(n_clone, 0, new_head_clone); 3244 } 3245 } 3246 3247 // Backedge of the surviving new_head (the clone) is original last_peel 3248 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel); 3249 3250 // Cut first node in original not_peel set 3251 _igvn.rehash_node_delayed(new_head); // Multiple edge updates: 3252 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of 3253 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls 3254 3255 // Copy head_clone back-branch info to original head 3256 // and remove original head's loop entry and 3257 // clone head's back-branch 3258 _igvn.rehash_node_delayed(head); // Multiple edge updates 3259 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl)); 3260 head->set_req(LoopNode::LoopBackControl, C->top()); 3261 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top()); 3262 3263 // Similarly modify the phis 3264 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) { 3265 Node* use = head->fast_out(k); 3266 if (use->is_Phi() && use->outcnt() > 0) { 3267 Node* use_clone = old_new[use->_idx]; 3268 _igvn.rehash_node_delayed(use); // Multiple edge updates 3269 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl)); 3270 use->set_req(LoopNode::LoopBackControl, C->top()); 3271 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top()); 3272 } 3273 } 3274 3275 // Step 4: update dominator tree and dominator depth 3276 3277 set_idom(head, orig_tail_clone, dd); 3278 recompute_dom_depth(); 3279 3280 // Inhibit more partial peeling on this loop 3281 new_head_clone->set_partial_peel_loop(); 3282 C->set_major_progress(); 3283 loop->record_for_igvn(); 3284 3285 #if !defined(PRODUCT) 3286 if (TracePartialPeeling) { 3287 tty->print_cr("\nafter partial peel one iteration"); 3288 Node_List wl(area); 3289 Node* t = last_peel; 3290 while (true) { 3291 wl.push(t); 3292 if (t == head_clone) break; 3293 t = idom(t); 3294 } 3295 while (wl.size() > 0) { 3296 Node* tt = wl.pop(); 3297 if (tt == head) tty->print_cr("orig head"); 3298 else if (tt == new_head_clone) tty->print_cr("new head"); 3299 else if (tt == head_clone) tty->print_cr("clone head"); 3300 tt->dump(); 3301 } 3302 } 3303 #endif 3304 return true; 3305 } 3306 3307 //------------------------------reorg_offsets---------------------------------- 3308 // Reorganize offset computations to lower register pressure. Mostly 3309 // prevent loop-fallout uses of the pre-incremented trip counter (which are 3310 // then alive with the post-incremented trip counter forcing an extra 3311 // register move) 3312 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) { 3313 // Perform it only for canonical counted loops. 3314 // Loop's shape could be messed up by iteration_split_impl. 3315 if (!loop->_head->is_CountedLoop()) 3316 return; 3317 if (!loop->_head->as_Loop()->is_valid_counted_loop()) 3318 return; 3319 3320 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 3321 CountedLoopEndNode *cle = cl->loopexit(); 3322 Node *exit = cle->proj_out(false); 3323 Node *phi = cl->phi(); 3324 3325 // Check for the special case of folks using the pre-incremented 3326 // trip-counter on the fall-out path (forces the pre-incremented 3327 // and post-incremented trip counter to be live at the same time). 3328 // Fix this by adjusting to use the post-increment trip counter. 3329 3330 bool progress = true; 3331 while (progress) { 3332 progress = false; 3333 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) { 3334 Node* use = phi->fast_out(i); // User of trip-counter 3335 if (!has_ctrl(use)) continue; 3336 Node *u_ctrl = get_ctrl(use); 3337 if (use->is_Phi()) { 3338 u_ctrl = NULL; 3339 for (uint j = 1; j < use->req(); j++) 3340 if (use->in(j) == phi) 3341 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j)); 3342 } 3343 IdealLoopTree *u_loop = get_loop(u_ctrl); 3344 // Look for loop-invariant use 3345 if (u_loop == loop) continue; 3346 if (loop->is_member(u_loop)) continue; 3347 // Check that use is live out the bottom. Assuming the trip-counter 3348 // update is right at the bottom, uses of of the loop middle are ok. 3349 if (dom_lca(exit, u_ctrl) != exit) continue; 3350 // Hit! Refactor use to use the post-incremented tripcounter. 3351 // Compute a post-increment tripcounter. 3352 Node *opaq = new Opaque2Node( C, cle->incr() ); 3353 register_new_node(opaq, exit); 3354 Node *neg_stride = _igvn.intcon(-cle->stride_con()); 3355 set_ctrl(neg_stride, C->root()); 3356 Node *post = new AddINode( opaq, neg_stride); 3357 register_new_node(post, exit); 3358 _igvn.rehash_node_delayed(use); 3359 for (uint j = 1; j < use->req(); j++) { 3360 if (use->in(j) == phi) 3361 use->set_req(j, post); 3362 } 3363 // Since DU info changed, rerun loop 3364 progress = true; 3365 break; 3366 } 3367 } 3368 3369 }