1 /* 2 * Copyright (c) 1999, 2019, 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 "opto/subtypenode.hpp" 44 #include "utilities/macros.hpp" 45 46 //============================================================================= 47 //------------------------------split_thru_phi--------------------------------- 48 // Split Node 'n' through merge point if there is enough win. 49 Node* PhaseIdealLoop::split_thru_phi(Node* n, Node* region, int policy) { 50 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) { 51 // ConvI2L may have type information on it which is unsafe to push up 52 // so disable this for now 53 return NULL; 54 } 55 56 // Splitting range check CastIIs through a loop induction Phi can 57 // cause new Phis to be created that are left unrelated to the loop 58 // induction Phi and prevent optimizations (vectorization) 59 if (n->Opcode() == Op_CastII && n->as_CastII()->has_range_check() && 60 region->is_CountedLoop() && n->in(1) == region->as_CountedLoop()->phi()) { 61 return NULL; 62 } 63 64 // Bail out if 'n' is a Div or Mod node whose zero check was removed earlier (i.e. control is NULL) and its divisor is an induction variable 65 // phi p of a trip-counted (integer) loop whose inputs could be zero (include zero in their type range). p could have a more precise type 66 // range that does not necessarily include all values of its inputs. Since each of these inputs will be a divisor of the newly cloned nodes 67 // of 'n', we need to bail out of one of these divisors could be zero (zero in its type range). 68 if ((n->Opcode() == Op_DivI || n->Opcode() == Op_ModI) && n->in(0) == NULL 69 && region->is_CountedLoop() && n->in(2) == region->as_CountedLoop()->phi()) { 70 Node* phi = region->as_CountedLoop()->phi(); 71 for (uint i = 1; i < phi->req(); i++) { 72 if (_igvn.type(phi->in(i))->filter_speculative(TypeInt::ZERO) != Type::TOP) { 73 // Zero could be a possible value but we already removed the zero check. Bail out to avoid a possible division by zero at a later point. 74 return NULL; 75 } 76 } 77 } 78 79 int wins = 0; 80 assert(!n->is_CFG(), ""); 81 assert(region->is_Region(), ""); 82 83 const Type* type = n->bottom_type(); 84 const TypeOopPtr* t_oop = _igvn.type(n)->isa_oopptr(); 85 Node* phi; 86 if (t_oop != NULL && t_oop->is_known_instance_field()) { 87 int iid = t_oop->instance_id(); 88 int index = C->get_alias_index(t_oop); 89 int offset = t_oop->offset(); 90 phi = new PhiNode(region, type, NULL, iid, index, offset); 91 } else { 92 phi = PhiNode::make_blank(region, n); 93 } 94 uint old_unique = C->unique(); 95 for (uint i = 1; i < region->req(); i++) { 96 Node* x; 97 Node* the_clone = NULL; 98 if (region->in(i) == C->top()) { 99 x = C->top(); // Dead path? Use a dead data op 100 } else { 101 x = n->clone(); // Else clone up the data op 102 the_clone = x; // Remember for possible deletion. 103 // Alter data node to use pre-phi inputs 104 if (n->in(0) == region) 105 x->set_req( 0, region->in(i) ); 106 for (uint j = 1; j < n->req(); j++) { 107 Node* in = n->in(j); 108 if (in->is_Phi() && in->in(0) == region) 109 x->set_req(j, in->in(i)); // Use pre-Phi input for the clone 110 } 111 } 112 // Check for a 'win' on some paths 113 const Type* t = x->Value(&_igvn); 114 115 bool singleton = t->singleton(); 116 117 // A TOP singleton indicates that there are no possible values incoming 118 // along a particular edge. In most cases, this is OK, and the Phi will 119 // be eliminated later in an Ideal call. However, we can't allow this to 120 // happen if the singleton occurs on loop entry, as the elimination of 121 // the PhiNode may cause the resulting node to migrate back to a previous 122 // loop iteration. 123 if (singleton && t == Type::TOP) { 124 // Is_Loop() == false does not confirm the absence of a loop (e.g., an 125 // irreducible loop may not be indicated by an affirmative is_Loop()); 126 // therefore, the only top we can split thru a phi is on a backedge of 127 // a loop. 128 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 129 } 130 131 if (singleton) { 132 wins++; 133 x = ((PhaseGVN&)_igvn).makecon(t); 134 } else { 135 // We now call Identity to try to simplify the cloned node. 136 // Note that some Identity methods call phase->type(this). 137 // Make sure that the type array is big enough for 138 // our new node, even though we may throw the node away. 139 // (Note: This tweaking with igvn only works because x is a new node.) 140 _igvn.set_type(x, t); 141 // If x is a TypeNode, capture any more-precise type permanently into Node 142 // otherwise it will be not updated during igvn->transform since 143 // igvn->type(x) is set to x->Value() already. 144 x->raise_bottom_type(t); 145 Node* y = x->Identity(&_igvn); 146 if (y != x) { 147 wins++; 148 x = y; 149 } else { 150 y = _igvn.hash_find(x); 151 if (y) { 152 wins++; 153 x = y; 154 } else { 155 // Else x is a new node we are keeping 156 // We do not need register_new_node_with_optimizer 157 // because set_type has already been called. 158 _igvn._worklist.push(x); 159 } 160 } 161 } 162 if (x != the_clone && the_clone != NULL) 163 _igvn.remove_dead_node(the_clone); 164 phi->set_req( i, x ); 165 } 166 // Too few wins? 167 if (wins <= policy) { 168 _igvn.remove_dead_node(phi); 169 return NULL; 170 } 171 172 // Record Phi 173 register_new_node( phi, region ); 174 175 for (uint i2 = 1; i2 < phi->req(); i2++) { 176 Node *x = phi->in(i2); 177 // If we commoned up the cloned 'x' with another existing Node, 178 // the existing Node picks up a new use. We need to make the 179 // existing Node occur higher up so it dominates its uses. 180 Node *old_ctrl; 181 IdealLoopTree *old_loop; 182 183 if (x->is_Con()) { 184 // Constant's control is always root. 185 set_ctrl(x, C->root()); 186 continue; 187 } 188 // The occasional new node 189 if (x->_idx >= old_unique) { // Found a new, unplaced node? 190 old_ctrl = NULL; 191 old_loop = NULL; // Not in any prior loop 192 } else { 193 old_ctrl = get_ctrl(x); 194 old_loop = get_loop(old_ctrl); // Get prior loop 195 } 196 // New late point must dominate new use 197 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2)); 198 if (new_ctrl == old_ctrl) // Nothing is changed 199 continue; 200 201 IdealLoopTree *new_loop = get_loop(new_ctrl); 202 203 // Don't move x into a loop if its uses are 204 // outside of loop. Otherwise x will be cloned 205 // for each use outside of this loop. 206 IdealLoopTree *use_loop = get_loop(region); 207 if (!new_loop->is_member(use_loop) && 208 (old_loop == NULL || !new_loop->is_member(old_loop))) { 209 // Take early control, later control will be recalculated 210 // during next iteration of loop optimizations. 211 new_ctrl = get_early_ctrl(x); 212 new_loop = get_loop(new_ctrl); 213 } 214 // Set new location 215 set_ctrl(x, new_ctrl); 216 // If changing loop bodies, see if we need to collect into new body 217 if (old_loop != new_loop) { 218 if (old_loop && !old_loop->_child) 219 old_loop->_body.yank(x); 220 if (!new_loop->_child) 221 new_loop->_body.push(x); // Collect body info 222 } 223 } 224 225 return phi; 226 } 227 228 //------------------------------dominated_by------------------------------------ 229 // Replace the dominated test with an obvious true or false. Place it on the 230 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the 231 // live path up to the dominating control. 232 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff, bool flip, bool exclude_loop_predicate ) { 233 if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); } 234 235 // prevdom is the dominating projection of the dominating test. 236 assert( iff->is_If(), "" ); 237 assert(iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd || iff->Opcode() == Op_RangeCheck, "Check this code when new subtype is added"); 238 int pop = prevdom->Opcode(); 239 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" ); 240 if (flip) { 241 if (pop == Op_IfTrue) 242 pop = Op_IfFalse; 243 else 244 pop = Op_IfTrue; 245 } 246 // 'con' is set to true or false to kill the dominated test. 247 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO); 248 set_ctrl(con, C->root()); // Constant gets a new use 249 // Hack the dominated test 250 _igvn.replace_input_of(iff, 1, con); 251 252 // If I dont have a reachable TRUE and FALSE path following the IfNode then 253 // I can assume this path reaches an infinite loop. In this case it's not 254 // important to optimize the data Nodes - either the whole compilation will 255 // be tossed or this path (and all data Nodes) will go dead. 256 if (iff->outcnt() != 2) return; 257 258 // Make control-dependent data Nodes on the live path (path that will remain 259 // once the dominated IF is removed) become control-dependent on the 260 // dominating projection. 261 Node* dp = iff->as_If()->proj_out_or_null(pop == Op_IfTrue); 262 263 // Loop predicates may have depending checks which should not 264 // be skipped. For example, range check predicate has two checks 265 // for lower and upper bounds. 266 if (dp == NULL) 267 return; 268 269 ProjNode* dp_proj = dp->as_Proj(); 270 ProjNode* unc_proj = iff->as_If()->proj_out(1 - dp_proj->_con)->as_Proj(); 271 if (exclude_loop_predicate && 272 (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) != NULL || 273 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_profile_predicate) != NULL || 274 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check) != NULL)) { 275 // If this is a range check (IfNode::is_range_check), do not 276 // reorder because Compile::allow_range_check_smearing might have 277 // changed the check. 278 return; // Let IGVN transformation change control dependence. 279 } 280 281 IdealLoopTree *old_loop = get_loop(dp); 282 283 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { 284 Node* cd = dp->fast_out(i); // Control-dependent node 285 if (cd->depends_only_on_test()) { 286 assert(cd->in(0) == dp, ""); 287 _igvn.replace_input_of(cd, 0, prevdom); 288 set_early_ctrl(cd); 289 IdealLoopTree *new_loop = get_loop(get_ctrl(cd)); 290 if (old_loop != new_loop) { 291 if (!old_loop->_child) old_loop->_body.yank(cd); 292 if (!new_loop->_child) new_loop->_body.push(cd); 293 } 294 --i; 295 --imax; 296 } 297 } 298 } 299 300 //------------------------------has_local_phi_input---------------------------- 301 // Return TRUE if 'n' has Phi inputs from its local block and no other 302 // block-local inputs (all non-local-phi inputs come from earlier blocks) 303 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) { 304 Node *n_ctrl = get_ctrl(n); 305 // See if some inputs come from a Phi in this block, or from before 306 // this block. 307 uint i; 308 for( i = 1; i < n->req(); i++ ) { 309 Node *phi = n->in(i); 310 if( phi->is_Phi() && phi->in(0) == n_ctrl ) 311 break; 312 } 313 if( i >= n->req() ) 314 return NULL; // No Phi inputs; nowhere to clone thru 315 316 // Check for inputs created between 'n' and the Phi input. These 317 // must split as well; they have already been given the chance 318 // (courtesy of a post-order visit) and since they did not we must 319 // recover the 'cost' of splitting them by being very profitable 320 // when splitting 'n'. Since this is unlikely we simply give up. 321 for( i = 1; i < n->req(); i++ ) { 322 Node *m = n->in(i); 323 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) { 324 // We allow the special case of AddP's with no local inputs. 325 // This allows us to split-up address expressions. 326 if (m->is_AddP() && 327 get_ctrl(m->in(2)) != n_ctrl && 328 get_ctrl(m->in(3)) != n_ctrl) { 329 // Move the AddP up to dominating point 330 Node* c = find_non_split_ctrl(idom(n_ctrl)); 331 if (c->is_OuterStripMinedLoop()) { 332 c->as_Loop()->verify_strip_mined(1); 333 c = c->in(LoopNode::EntryControl); 334 } 335 set_ctrl_and_loop(m, c); 336 continue; 337 } 338 return NULL; 339 } 340 assert(n->is_Phi() || m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control"); 341 } 342 343 return n_ctrl; 344 } 345 346 //------------------------------remix_address_expressions---------------------- 347 // Rework addressing expressions to get the most loop-invariant stuff 348 // moved out. We'd like to do all associative operators, but it's especially 349 // important (common) to do address expressions. 350 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) { 351 if (!has_ctrl(n)) return NULL; 352 Node *n_ctrl = get_ctrl(n); 353 IdealLoopTree *n_loop = get_loop(n_ctrl); 354 355 // See if 'n' mixes loop-varying and loop-invariant inputs and 356 // itself is loop-varying. 357 358 // Only interested in binary ops (and AddP) 359 if( n->req() < 3 || n->req() > 4 ) return NULL; 360 361 Node *n1_ctrl = get_ctrl(n->in( 1)); 362 Node *n2_ctrl = get_ctrl(n->in( 2)); 363 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3)); 364 IdealLoopTree *n1_loop = get_loop( n1_ctrl ); 365 IdealLoopTree *n2_loop = get_loop( n2_ctrl ); 366 IdealLoopTree *n3_loop = get_loop( n3_ctrl ); 367 368 // Does one of my inputs spin in a tighter loop than self? 369 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) || 370 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) || 371 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) ) 372 return NULL; // Leave well enough alone 373 374 // Is at least one of my inputs loop-invariant? 375 if( n1_loop == n_loop && 376 n2_loop == n_loop && 377 n3_loop == n_loop ) 378 return NULL; // No loop-invariant inputs 379 380 381 int n_op = n->Opcode(); 382 383 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2). 384 if( n_op == Op_LShiftI ) { 385 // Scale is loop invariant 386 Node *scale = n->in(2); 387 Node *scale_ctrl = get_ctrl(scale); 388 IdealLoopTree *scale_loop = get_loop(scale_ctrl ); 389 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) ) 390 return NULL; 391 const TypeInt *scale_t = scale->bottom_type()->isa_int(); 392 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 ) 393 return NULL; // Dont bother with byte/short masking 394 // Add must vary with loop (else shift would be loop-invariant) 395 Node *add = n->in(1); 396 Node *add_ctrl = get_ctrl(add); 397 IdealLoopTree *add_loop = get_loop(add_ctrl); 398 //assert( n_loop == add_loop, "" ); 399 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops 400 401 // Convert I-V into I+ (0-V); same for V-I 402 if( add->Opcode() == Op_SubI && 403 _igvn.type( add->in(1) ) != TypeInt::ZERO ) { 404 Node *zero = _igvn.intcon(0); 405 set_ctrl(zero, C->root()); 406 Node *neg = new SubINode( _igvn.intcon(0), add->in(2) ); 407 register_new_node( neg, get_ctrl(add->in(2) ) ); 408 add = new AddINode( add->in(1), neg ); 409 register_new_node( add, add_ctrl ); 410 } 411 if( add->Opcode() != Op_AddI ) return NULL; 412 // See if one add input is loop invariant 413 Node *add_var = add->in(1); 414 Node *add_var_ctrl = get_ctrl(add_var); 415 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 416 Node *add_invar = add->in(2); 417 Node *add_invar_ctrl = get_ctrl(add_invar); 418 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl ); 419 if( add_var_loop == n_loop ) { 420 } else if( add_invar_loop == n_loop ) { 421 // Swap to find the invariant part 422 add_invar = add_var; 423 add_invar_ctrl = add_var_ctrl; 424 add_invar_loop = add_var_loop; 425 add_var = add->in(2); 426 Node *add_var_ctrl = get_ctrl(add_var); 427 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 428 } else // Else neither input is loop invariant 429 return NULL; 430 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) ) 431 return NULL; // No invariant part of the add? 432 433 // Yes! Reshape address expression! 434 Node *inv_scale = new LShiftINode( add_invar, scale ); 435 Node *inv_scale_ctrl = 436 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ? 437 add_invar_ctrl : scale_ctrl; 438 register_new_node( inv_scale, inv_scale_ctrl ); 439 Node *var_scale = new LShiftINode( add_var, scale ); 440 register_new_node( var_scale, n_ctrl ); 441 Node *var_add = new AddINode( var_scale, inv_scale ); 442 register_new_node( var_add, n_ctrl ); 443 _igvn.replace_node( n, var_add ); 444 return var_add; 445 } 446 447 // Replace (I+V) with (V+I) 448 if( n_op == Op_AddI || 449 n_op == Op_AddL || 450 n_op == Op_AddF || 451 n_op == Op_AddD || 452 n_op == Op_MulI || 453 n_op == Op_MulL || 454 n_op == Op_MulF || 455 n_op == Op_MulD ) { 456 if( n2_loop == n_loop ) { 457 assert( n1_loop != n_loop, "" ); 458 n->swap_edges(1, 2); 459 } 460 } 461 462 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V), 463 // but not if I2 is a constant. 464 if( n_op == Op_AddP ) { 465 if( n2_loop == n_loop && n3_loop != n_loop ) { 466 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) { 467 Node *n22_ctrl = get_ctrl(n->in(2)->in(2)); 468 Node *n23_ctrl = get_ctrl(n->in(2)->in(3)); 469 IdealLoopTree *n22loop = get_loop( n22_ctrl ); 470 IdealLoopTree *n23_loop = get_loop( n23_ctrl ); 471 if( n22loop != n_loop && n22loop->is_member(n_loop) && 472 n23_loop == n_loop ) { 473 Node *add1 = new AddPNode( n->in(1), n->in(2)->in(2), n->in(3) ); 474 // Stuff new AddP in the loop preheader 475 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) ); 476 Node *add2 = new AddPNode( n->in(1), add1, n->in(2)->in(3) ); 477 register_new_node( add2, n_ctrl ); 478 _igvn.replace_node( n, add2 ); 479 return add2; 480 } 481 } 482 } 483 484 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V) 485 if (n2_loop != n_loop && n3_loop == n_loop) { 486 if (n->in(3)->Opcode() == Op_AddX) { 487 Node *V = n->in(3)->in(1); 488 Node *I = n->in(3)->in(2); 489 if (is_member(n_loop,get_ctrl(V))) { 490 } else { 491 Node *tmp = V; V = I; I = tmp; 492 } 493 if (!is_member(n_loop,get_ctrl(I))) { 494 Node *add1 = new AddPNode(n->in(1), n->in(2), I); 495 // Stuff new AddP in the loop preheader 496 register_new_node(add1, n_loop->_head->in(LoopNode::EntryControl)); 497 Node *add2 = new AddPNode(n->in(1), add1, V); 498 register_new_node(add2, n_ctrl); 499 _igvn.replace_node(n, add2); 500 return add2; 501 } 502 } 503 } 504 } 505 506 return NULL; 507 } 508 509 // Optimize ((in1[2*i] * in2[2*i]) + (in1[2*i+1] * in2[2*i+1])) 510 Node *PhaseIdealLoop::convert_add_to_muladd(Node* n) { 511 assert(n->Opcode() == Op_AddI, "sanity"); 512 Node * nn = NULL; 513 Node * in1 = n->in(1); 514 Node * in2 = n->in(2); 515 if (in1->Opcode() == Op_MulI && in2->Opcode() == Op_MulI) { 516 IdealLoopTree* loop_n = get_loop(get_ctrl(n)); 517 if (loop_n->is_counted() && 518 loop_n->_head->as_Loop()->is_valid_counted_loop() && 519 Matcher::match_rule_supported(Op_MulAddVS2VI) && 520 Matcher::match_rule_supported(Op_MulAddS2I)) { 521 Node* mul_in1 = in1->in(1); 522 Node* mul_in2 = in1->in(2); 523 Node* mul_in3 = in2->in(1); 524 Node* mul_in4 = in2->in(2); 525 if (mul_in1->Opcode() == Op_LoadS && 526 mul_in2->Opcode() == Op_LoadS && 527 mul_in3->Opcode() == Op_LoadS && 528 mul_in4->Opcode() == Op_LoadS) { 529 IdealLoopTree* loop1 = get_loop(get_ctrl(mul_in1)); 530 IdealLoopTree* loop2 = get_loop(get_ctrl(mul_in2)); 531 IdealLoopTree* loop3 = get_loop(get_ctrl(mul_in3)); 532 IdealLoopTree* loop4 = get_loop(get_ctrl(mul_in4)); 533 IdealLoopTree* loop5 = get_loop(get_ctrl(in1)); 534 IdealLoopTree* loop6 = get_loop(get_ctrl(in2)); 535 // All nodes should be in the same counted loop. 536 if (loop_n == loop1 && loop_n == loop2 && loop_n == loop3 && 537 loop_n == loop4 && loop_n == loop5 && loop_n == loop6) { 538 Node* adr1 = mul_in1->in(MemNode::Address); 539 Node* adr2 = mul_in2->in(MemNode::Address); 540 Node* adr3 = mul_in3->in(MemNode::Address); 541 Node* adr4 = mul_in4->in(MemNode::Address); 542 if (adr1->is_AddP() && adr2->is_AddP() && adr3->is_AddP() && adr4->is_AddP()) { 543 if ((adr1->in(AddPNode::Base) == adr3->in(AddPNode::Base)) && 544 (adr2->in(AddPNode::Base) == adr4->in(AddPNode::Base))) { 545 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in3, mul_in4); 546 register_new_node(nn, get_ctrl(n)); 547 _igvn.replace_node(n, nn); 548 return nn; 549 } else if ((adr1->in(AddPNode::Base) == adr4->in(AddPNode::Base)) && 550 (adr2->in(AddPNode::Base) == adr3->in(AddPNode::Base))) { 551 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in4, mul_in3); 552 register_new_node(nn, get_ctrl(n)); 553 _igvn.replace_node(n, nn); 554 return nn; 555 } 556 } 557 } 558 } 559 } 560 } 561 return nn; 562 } 563 564 //------------------------------conditional_move------------------------------- 565 // Attempt to replace a Phi with a conditional move. We have some pretty 566 // strict profitability requirements. All Phis at the merge point must 567 // be converted, so we can remove the control flow. We need to limit the 568 // number of c-moves to a small handful. All code that was in the side-arms 569 // of the CFG diamond is now speculatively executed. This code has to be 570 // "cheap enough". We are pretty much limited to CFG diamonds that merge 571 // 1 or 2 items with a total of 1 or 2 ops executed speculatively. 572 Node *PhaseIdealLoop::conditional_move( Node *region ) { 573 574 assert(region->is_Region(), "sanity check"); 575 if (region->req() != 3) return NULL; 576 577 // Check for CFG diamond 578 Node *lp = region->in(1); 579 Node *rp = region->in(2); 580 if (!lp || !rp) return NULL; 581 Node *lp_c = lp->in(0); 582 if (lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If()) return NULL; 583 IfNode *iff = lp_c->as_If(); 584 585 // Check for ops pinned in an arm of the diamond. 586 // Can't remove the control flow in this case 587 if (lp->outcnt() > 1) return NULL; 588 if (rp->outcnt() > 1) return NULL; 589 590 IdealLoopTree* r_loop = get_loop(region); 591 assert(r_loop == get_loop(iff), "sanity"); 592 // Always convert to CMOVE if all results are used only outside this loop. 593 bool used_inside_loop = (r_loop == _ltree_root); 594 595 // Check profitability 596 int cost = 0; 597 int phis = 0; 598 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 599 Node *out = region->fast_out(i); 600 if (!out->is_Phi()) continue; // Ignore other control edges, etc 601 phis++; 602 PhiNode* phi = out->as_Phi(); 603 BasicType bt = phi->type()->basic_type(); 604 switch (bt) { 605 case T_DOUBLE: 606 case T_FLOAT: 607 if (C->use_cmove()) { 608 continue; //TODO: maybe we want to add some cost 609 } 610 cost += Matcher::float_cmove_cost(); // Could be very expensive 611 break; 612 case T_LONG: { 613 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's 614 } 615 case T_INT: // These all CMOV fine 616 case T_ADDRESS: { // (RawPtr) 617 cost++; 618 break; 619 } 620 case T_NARROWOOP: // Fall through 621 case T_OBJECT: { // Base oops are OK, but not derived oops 622 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr(); 623 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a 624 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus 625 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we 626 // have a Phi for the base here that we convert to a CMOVE all is well 627 // and good. But if the base is dead, we'll not make a CMOVE. Later 628 // the allocator will have to produce a base by creating a CMOVE of the 629 // relevant bases. This puts the allocator in the business of 630 // manufacturing expensive instructions, generally a bad plan. 631 // Just Say No to Conditionally-Moved Derived Pointers. 632 if (tp && tp->offset() != 0) 633 return NULL; 634 cost++; 635 break; 636 } 637 default: 638 return NULL; // In particular, can't do memory or I/O 639 } 640 // Add in cost any speculative ops 641 for (uint j = 1; j < region->req(); j++) { 642 Node *proj = region->in(j); 643 Node *inp = phi->in(j); 644 if (get_ctrl(inp) == proj) { // Found local op 645 cost++; 646 // Check for a chain of dependent ops; these will all become 647 // speculative in a CMOV. 648 for (uint k = 1; k < inp->req(); k++) 649 if (get_ctrl(inp->in(k)) == proj) 650 cost += ConditionalMoveLimit; // Too much speculative goo 651 } 652 } 653 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode. 654 // This will likely Split-If, a higher-payoff operation. 655 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) { 656 Node* use = phi->fast_out(k); 657 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr()) 658 cost += ConditionalMoveLimit; 659 // Is there a use inside the loop? 660 // Note: check only basic types since CMoveP is pinned. 661 if (!used_inside_loop && is_java_primitive(bt)) { 662 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use); 663 if (r_loop == u_loop || r_loop->is_member(u_loop)) { 664 used_inside_loop = true; 665 } 666 } 667 } 668 }//for 669 Node* bol = iff->in(1); 670 if (bol->Opcode() == Op_Opaque4) { 671 return NULL; // Ignore loop predicate checks (the Opaque4 ensures they will go away) 672 } 673 assert(bol->Opcode() == Op_Bool, "Unexpected node"); 674 int cmp_op = bol->in(1)->Opcode(); 675 if (cmp_op == Op_SubTypeCheck) { // SubTypeCheck expansion expects an IfNode 676 return NULL; 677 } 678 // It is expensive to generate flags from a float compare. 679 // Avoid duplicated float compare. 680 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL; 681 682 float infrequent_prob = PROB_UNLIKELY_MAG(3); 683 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop. 684 if (used_inside_loop) { 685 if (cost >= ConditionalMoveLimit) return NULL; // Too much goo 686 687 // BlockLayoutByFrequency optimization moves infrequent branch 688 // from hot path. No point in CMOV'ing in such case (110 is used 689 // instead of 100 to take into account not exactness of float value). 690 if (BlockLayoutByFrequency) { 691 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f); 692 } 693 } 694 // Check for highly predictable branch. No point in CMOV'ing if 695 // we are going to predict accurately all the time. 696 if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) { 697 //keep going 698 } else if (iff->_prob < infrequent_prob || 699 iff->_prob > (1.0f - infrequent_prob)) 700 return NULL; 701 702 // -------------- 703 // Now replace all Phis with CMOV's 704 Node *cmov_ctrl = iff->in(0); 705 uint flip = (lp->Opcode() == Op_IfTrue); 706 Node_List wq; 707 while (1) { 708 PhiNode* phi = NULL; 709 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 710 Node *out = region->fast_out(i); 711 if (out->is_Phi()) { 712 phi = out->as_Phi(); 713 break; 714 } 715 } 716 if (phi == NULL) break; 717 if (PrintOpto && VerifyLoopOptimizations) { tty->print_cr("CMOV"); } 718 // Move speculative ops 719 wq.push(phi); 720 while (wq.size() > 0) { 721 Node *n = wq.pop(); 722 for (uint j = 1; j < n->req(); j++) { 723 Node* m = n->in(j); 724 if (m != NULL && !is_dominator(get_ctrl(m), cmov_ctrl)) { 725 #ifndef PRODUCT 726 if (PrintOpto && VerifyLoopOptimizations) { 727 tty->print(" speculate: "); 728 m->dump(); 729 } 730 #endif 731 set_ctrl(m, cmov_ctrl); 732 wq.push(m); 733 } 734 } 735 } 736 Node *cmov = CMoveNode::make(cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi)); 737 register_new_node( cmov, cmov_ctrl ); 738 _igvn.replace_node( phi, cmov ); 739 #ifndef PRODUCT 740 if (TraceLoopOpts) { 741 tty->print("CMOV "); 742 r_loop->dump_head(); 743 if (Verbose) { 744 bol->in(1)->dump(1); 745 cmov->dump(1); 746 } 747 } 748 if (VerifyLoopOptimizations) verify(); 749 #endif 750 } 751 752 // The useless CFG diamond will fold up later; see the optimization in 753 // RegionNode::Ideal. 754 _igvn._worklist.push(region); 755 756 return iff->in(1); 757 } 758 759 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) { 760 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 761 Node* u = m->fast_out(i); 762 if (u->is_CFG()) { 763 if (u->Opcode() == Op_NeverBranch) { 764 u = ((NeverBranchNode*)u)->proj_out(0); 765 enqueue_cfg_uses(u, wq); 766 } else { 767 wq.push(u); 768 } 769 } 770 } 771 } 772 773 // Try moving a store out of a loop, right before the loop 774 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) { 775 // Store has to be first in the loop body 776 IdealLoopTree *n_loop = get_loop(n_ctrl); 777 if (n->is_Store() && n_loop != _ltree_root && 778 n_loop->is_loop() && n_loop->_head->is_Loop() && 779 n->in(0) != NULL) { 780 Node* address = n->in(MemNode::Address); 781 Node* value = n->in(MemNode::ValueIn); 782 Node* mem = n->in(MemNode::Memory); 783 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 784 IdealLoopTree* value_loop = get_loop(get_ctrl(value)); 785 786 // - address and value must be loop invariant 787 // - memory must be a memory Phi for the loop 788 // - Store must be the only store on this memory slice in the 789 // loop: if there's another store following this one then value 790 // written at iteration i by the second store could be overwritten 791 // at iteration i+n by the first store: it's not safe to move the 792 // first store out of the loop 793 // - nothing must observe the memory Phi: it guarantees no read 794 // before the store, we are also guaranteed the store post 795 // dominates the loop head (ignoring a possible early 796 // exit). Otherwise there would be extra Phi involved between the 797 // loop's Phi and the store. 798 // - there must be no early exit from the loop before the Store 799 // (such an exit most of the time would be an extra use of the 800 // memory Phi but sometimes is a bottom memory Phi that takes the 801 // store as input). 802 803 if (!n_loop->is_member(address_loop) && 804 !n_loop->is_member(value_loop) && 805 mem->is_Phi() && mem->in(0) == n_loop->_head && 806 mem->outcnt() == 1 && 807 mem->in(LoopNode::LoopBackControl) == n) { 808 809 assert(n_loop->_tail != NULL, "need a tail"); 810 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop"); 811 812 // Verify that there's no early exit of the loop before the store. 813 bool ctrl_ok = false; 814 { 815 // Follow control from loop head until n, we exit the loop or 816 // we reach the tail 817 ResourceMark rm; 818 Unique_Node_List wq; 819 wq.push(n_loop->_head); 820 821 for (uint next = 0; next < wq.size(); ++next) { 822 Node *m = wq.at(next); 823 if (m == n->in(0)) { 824 ctrl_ok = true; 825 continue; 826 } 827 assert(!has_ctrl(m), "should be CFG"); 828 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) { 829 ctrl_ok = false; 830 break; 831 } 832 enqueue_cfg_uses(m, wq); 833 if (wq.size() > 10) { 834 ctrl_ok = false; 835 break; 836 } 837 } 838 } 839 if (ctrl_ok) { 840 // move the Store 841 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem); 842 _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl)); 843 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl)); 844 // Disconnect the phi now. An empty phi can confuse other 845 // optimizations in this pass of loop opts. 846 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl)); 847 n_loop->_body.yank(mem); 848 849 set_ctrl_and_loop(n, n->in(0)); 850 851 return n; 852 } 853 } 854 } 855 return NULL; 856 } 857 858 // Try moving a store out of a loop, right after the loop 859 void PhaseIdealLoop::try_move_store_after_loop(Node* n) { 860 if (n->is_Store() && n->in(0) != NULL) { 861 Node *n_ctrl = get_ctrl(n); 862 IdealLoopTree *n_loop = get_loop(n_ctrl); 863 // Store must be in a loop 864 if (n_loop != _ltree_root && !n_loop->_irreducible) { 865 Node* address = n->in(MemNode::Address); 866 Node* value = n->in(MemNode::ValueIn); 867 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 868 // address must be loop invariant 869 if (!n_loop->is_member(address_loop)) { 870 // Store must be last on this memory slice in the loop and 871 // nothing in the loop must observe it 872 Node* phi = NULL; 873 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 874 Node* u = n->fast_out(i); 875 if (has_ctrl(u)) { // control use? 876 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 877 if (!n_loop->is_member(u_loop)) { 878 continue; 879 } 880 if (u->is_Phi() && u->in(0) == n_loop->_head) { 881 assert(_igvn.type(u) == Type::MEMORY, "bad phi"); 882 // multiple phis on the same slice are possible 883 if (phi != NULL) { 884 return; 885 } 886 phi = u; 887 continue; 888 } 889 } 890 return; 891 } 892 if (phi != NULL) { 893 // Nothing in the loop before the store (next iteration) 894 // must observe the stored value 895 bool mem_ok = true; 896 { 897 ResourceMark rm; 898 Unique_Node_List wq; 899 wq.push(phi); 900 for (uint next = 0; next < wq.size() && mem_ok; ++next) { 901 Node *m = wq.at(next); 902 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) { 903 Node* u = m->fast_out(i); 904 if (u->is_Store() || u->is_Phi()) { 905 if (u != n) { 906 wq.push(u); 907 mem_ok = (wq.size() <= 10); 908 } 909 } else { 910 mem_ok = false; 911 break; 912 } 913 } 914 } 915 } 916 if (mem_ok) { 917 // Move the store out of the loop if the LCA of all 918 // users (except for the phi) is outside the loop. 919 Node* hook = new Node(1); 920 hook->init_req(0, n_ctrl); // Add an input to prevent hook from being dead 921 _igvn.rehash_node_delayed(phi); 922 int count = phi->replace_edge(n, hook); 923 assert(count > 0, "inconsistent phi"); 924 925 // Compute latest point this store can go 926 Node* lca = get_late_ctrl(n, get_ctrl(n)); 927 if (lca->is_OuterStripMinedLoop()) { 928 lca = lca->in(LoopNode::EntryControl); 929 } 930 if (n_loop->is_member(get_loop(lca))) { 931 // LCA is in the loop - bail out 932 _igvn.replace_node(hook, n); 933 return; 934 } 935 #ifdef ASSERT 936 if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) { 937 assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined"); 938 n_loop->_head->as_Loop()->verify_strip_mined(1); 939 Node* outer = n_loop->_head->as_CountedLoop()->outer_loop(); 940 IdealLoopTree* outer_loop = get_loop(outer); 941 assert(n_loop->_parent == outer_loop, "broken loop tree"); 942 assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state"); 943 } 944 #endif 945 946 // Move store out of the loop 947 _igvn.replace_node(hook, n->in(MemNode::Memory)); 948 _igvn.replace_input_of(n, 0, lca); 949 set_ctrl_and_loop(n, lca); 950 951 // Disconnect the phi now. An empty phi can confuse other 952 // optimizations in this pass of loop opts.. 953 if (phi->in(LoopNode::LoopBackControl) == phi) { 954 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl)); 955 n_loop->_body.yank(phi); 956 } 957 } 958 } 959 } 960 } 961 } 962 } 963 964 //------------------------------split_if_with_blocks_pre----------------------- 965 // Do the real work in a non-recursive function. Data nodes want to be 966 // cloned in the pre-order so they can feed each other nicely. 967 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) { 968 // Cloning these guys is unlikely to win 969 int n_op = n->Opcode(); 970 if (n_op == Op_MergeMem) { 971 return n; 972 } 973 if (n->is_Proj()) { 974 return n; 975 } 976 // Do not clone-up CmpFXXX variations, as these are always 977 // followed by a CmpI 978 if (n->is_Cmp()) { 979 return n; 980 } 981 // Attempt to use a conditional move instead of a phi/branch 982 if (ConditionalMoveLimit > 0 && n_op == Op_Region) { 983 Node *cmov = conditional_move( n ); 984 if (cmov) { 985 return cmov; 986 } 987 } 988 if (n->is_CFG() || n->is_LoadStore()) { 989 return n; 990 } 991 if (n->is_Opaque1() || // Opaque nodes cannot be mod'd 992 n_op == Op_Opaque2) { 993 if (!C->major_progress()) { // If chance of no more loop opts... 994 _igvn._worklist.push(n); // maybe we'll remove them 995 } 996 return n; 997 } 998 999 if (n->is_Con()) { 1000 return n; // No cloning for Con nodes 1001 } 1002 1003 Node *n_ctrl = get_ctrl(n); 1004 if (!n_ctrl) { 1005 return n; // Dead node 1006 } 1007 1008 Node* res = try_move_store_before_loop(n, n_ctrl); 1009 if (res != NULL) { 1010 return n; 1011 } 1012 1013 // Attempt to remix address expressions for loop invariants 1014 Node *m = remix_address_expressions( n ); 1015 if( m ) return m; 1016 1017 if (n_op == Op_AddI) { 1018 Node *nn = convert_add_to_muladd( n ); 1019 if ( nn ) return nn; 1020 } 1021 1022 if (n->is_ConstraintCast()) { 1023 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this); 1024 // ConstraintCastNode::dominating_cast() uses node control input to determine domination. 1025 // Node control inputs don't necessarily agree with loop control info (due to 1026 // transformations happened in between), thus additional dominance check is needed 1027 // to keep loop info valid. 1028 if (dom_cast != NULL && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) { 1029 _igvn.replace_node(n, dom_cast); 1030 return dom_cast; 1031 } 1032 } 1033 1034 // Determine if the Node has inputs from some local Phi. 1035 // Returns the block to clone thru. 1036 Node *n_blk = has_local_phi_input( n ); 1037 if( !n_blk ) return n; 1038 1039 // Do not clone the trip counter through on a CountedLoop 1040 // (messes up the canonical shape). 1041 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n; 1042 1043 // Check for having no control input; not pinned. Allow 1044 // dominating control. 1045 if (n->in(0)) { 1046 Node *dom = idom(n_blk); 1047 if (dom_lca(n->in(0), dom) != n->in(0)) { 1048 return n; 1049 } 1050 } 1051 // Policy: when is it profitable. You must get more wins than 1052 // policy before it is considered profitable. Policy is usually 0, 1053 // so 1 win is considered profitable. Big merges will require big 1054 // cloning, so get a larger policy. 1055 int policy = n_blk->req() >> 2; 1056 1057 // If the loop is a candidate for range check elimination, 1058 // delay splitting through it's phi until a later loop optimization 1059 if (n_blk->is_CountedLoop()) { 1060 IdealLoopTree *lp = get_loop(n_blk); 1061 if (lp && lp->_rce_candidate) { 1062 return n; 1063 } 1064 } 1065 1066 if (must_throttle_split_if()) return n; 1067 1068 // Split 'n' through the merge point if it is profitable 1069 Node *phi = split_thru_phi( n, n_blk, policy ); 1070 if (!phi) return n; 1071 1072 // Found a Phi to split thru! 1073 // Replace 'n' with the new phi 1074 _igvn.replace_node( n, phi ); 1075 // Moved a load around the loop, 'en-registering' something. 1076 if (n_blk->is_Loop() && n->is_Load() && 1077 !phi->in(LoopNode::LoopBackControl)->is_Load()) 1078 C->set_major_progress(); 1079 1080 return phi; 1081 } 1082 1083 static bool merge_point_too_heavy(Compile* C, Node* region) { 1084 // Bail out if the region and its phis have too many users. 1085 int weight = 0; 1086 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1087 weight += region->fast_out(i)->outcnt(); 1088 } 1089 int nodes_left = C->max_node_limit() - C->live_nodes(); 1090 if (weight * 8 > nodes_left) { 1091 if (PrintOpto) { 1092 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight); 1093 } 1094 return true; 1095 } else { 1096 return false; 1097 } 1098 } 1099 1100 static bool merge_point_safe(Node* region) { 1101 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode 1102 // having a PhiNode input. This sidesteps the dangerous case where the split 1103 // ConvI2LNode may become TOP if the input Value() does not 1104 // overlap the ConvI2L range, leaving a node which may not dominate its 1105 // uses. 1106 // A better fix for this problem can be found in the BugTraq entry, but 1107 // expediency for Mantis demands this hack. 1108 #ifdef _LP64 1109 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1110 Node* n = region->fast_out(i); 1111 if (n->is_Phi()) { 1112 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1113 Node* m = n->fast_out(j); 1114 if (m->Opcode() == Op_ConvI2L) 1115 return false; 1116 if (m->is_CastII() && m->isa_CastII()->has_range_check()) { 1117 return false; 1118 } 1119 } 1120 } 1121 } 1122 #endif 1123 return true; 1124 } 1125 1126 1127 //------------------------------place_near_use--------------------------------- 1128 // Place some computation next to use but not inside inner loops. 1129 // For inner loop uses move it to the preheader area. 1130 Node *PhaseIdealLoop::place_near_use(Node *useblock) const { 1131 IdealLoopTree *u_loop = get_loop( useblock ); 1132 if (u_loop->_irreducible) { 1133 return useblock; 1134 } 1135 if (u_loop->_child) { 1136 if (useblock == u_loop->_head && u_loop->_head->is_OuterStripMinedLoop()) { 1137 return u_loop->_head->in(LoopNode::EntryControl); 1138 } 1139 return useblock; 1140 } 1141 return u_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl); 1142 } 1143 1144 1145 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) { 1146 if (!n->is_If() || n->is_CountedLoopEnd()) { 1147 return false; 1148 } 1149 if (!n->in(0)->is_Region()) { 1150 return false; 1151 } 1152 Node* region = n->in(0); 1153 Node* dom = idom(region); 1154 if (!dom->is_If() || dom->in(1) != n->in(1)) { 1155 return false; 1156 } 1157 IfNode* dom_if = dom->as_If(); 1158 Node* proj_true = dom_if->proj_out(1); 1159 Node* proj_false = dom_if->proj_out(0); 1160 1161 for (uint i = 1; i < region->req(); i++) { 1162 if (is_dominator(proj_true, region->in(i))) { 1163 continue; 1164 } 1165 if (is_dominator(proj_false, region->in(i))) { 1166 continue; 1167 } 1168 return false; 1169 } 1170 1171 return true; 1172 } 1173 1174 1175 bool PhaseIdealLoop::can_split_if(Node* n_ctrl) { 1176 if (must_throttle_split_if()) { 1177 return false; 1178 } 1179 1180 // Do not do 'split-if' if irreducible loops are present. 1181 if (_has_irreducible_loops) { 1182 return false; 1183 } 1184 1185 if (merge_point_too_heavy(C, n_ctrl)) { 1186 return false; 1187 } 1188 1189 // Do not do 'split-if' if some paths are dead. First do dead code 1190 // elimination and then see if its still profitable. 1191 for (uint i = 1; i < n_ctrl->req(); i++) { 1192 if (n_ctrl->in(i) == C->top()) { 1193 return false; 1194 } 1195 } 1196 1197 // If trying to do a 'Split-If' at the loop head, it is only 1198 // profitable if the cmp folds up on BOTH paths. Otherwise we 1199 // risk peeling a loop forever. 1200 1201 // CNC - Disabled for now. Requires careful handling of loop 1202 // body selection for the cloned code. Also, make sure we check 1203 // for any input path not being in the same loop as n_ctrl. For 1204 // irreducible loops we cannot check for 'n_ctrl->is_Loop()' 1205 // because the alternative loop entry points won't be converted 1206 // into LoopNodes. 1207 IdealLoopTree *n_loop = get_loop(n_ctrl); 1208 for (uint j = 1; j < n_ctrl->req(); j++) { 1209 if (get_loop(n_ctrl->in(j)) != n_loop) { 1210 return false; 1211 } 1212 } 1213 1214 // Check for safety of the merge point. 1215 if (!merge_point_safe(n_ctrl)) { 1216 return false; 1217 } 1218 1219 return true; 1220 } 1221 1222 // Detect if the node is the inner strip-mined loop 1223 // Return: NULL if it's not the case, or the exit of outer strip-mined loop 1224 static Node* is_inner_of_stripmined_loop(const Node* out) { 1225 Node* out_le = NULL; 1226 1227 if (out->is_CountedLoopEnd()) { 1228 const CountedLoopNode* loop = out->as_CountedLoopEnd()->loopnode(); 1229 1230 if (loop != NULL && loop->is_strip_mined()) { 1231 out_le = loop->in(LoopNode::EntryControl)->as_OuterStripMinedLoop()->outer_loop_exit(); 1232 } 1233 } 1234 1235 return out_le; 1236 } 1237 1238 //------------------------------split_if_with_blocks_post---------------------- 1239 // Do the real work in a non-recursive function. CFG hackery wants to be 1240 // in the post-order, so it can dirty the I-DOM info and not use the dirtied 1241 // info. 1242 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) { 1243 1244 // Cloning Cmp through Phi's involves the split-if transform. 1245 // FastLock is not used by an If 1246 if (n->is_Cmp() && !n->is_FastLock()) { 1247 Node *n_ctrl = get_ctrl(n); 1248 // Determine if the Node has inputs from some local Phi. 1249 // Returns the block to clone thru. 1250 Node *n_blk = has_local_phi_input(n); 1251 if (n_blk != n_ctrl) { 1252 return; 1253 } 1254 1255 if (!can_split_if(n_ctrl)) { 1256 return; 1257 } 1258 1259 if (n->outcnt() != 1) { 1260 return; // Multiple bool's from 1 compare? 1261 } 1262 Node *bol = n->unique_out(); 1263 assert(bol->is_Bool(), "expect a bool here"); 1264 if (bol->outcnt() != 1) { 1265 return;// Multiple branches from 1 compare? 1266 } 1267 Node *iff = bol->unique_out(); 1268 1269 // Check some safety conditions 1270 if (iff->is_If()) { // Classic split-if? 1271 if (iff->in(0) != n_ctrl) { 1272 return; // Compare must be in same blk as if 1273 } 1274 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE 1275 // Can't split CMove with different control edge. 1276 if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) { 1277 return; 1278 } 1279 if (get_ctrl(iff->in(2)) == n_ctrl || 1280 get_ctrl(iff->in(3)) == n_ctrl) { 1281 return; // Inputs not yet split-up 1282 } 1283 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) { 1284 return; // Loop-invar test gates loop-varying CMOVE 1285 } 1286 } else { 1287 return; // some other kind of node, such as an Allocate 1288 } 1289 1290 // When is split-if profitable? Every 'win' on means some control flow 1291 // goes dead, so it's almost always a win. 1292 int policy = 0; 1293 // Split compare 'n' through the merge point if it is profitable 1294 Node *phi = split_thru_phi( n, n_ctrl, policy); 1295 if (!phi) { 1296 return; 1297 } 1298 1299 // Found a Phi to split thru! 1300 // Replace 'n' with the new phi 1301 _igvn.replace_node(n, phi); 1302 1303 // Now split the bool up thru the phi 1304 Node *bolphi = split_thru_phi(bol, n_ctrl, -1); 1305 guarantee(bolphi != NULL, "null boolean phi node"); 1306 1307 _igvn.replace_node(bol, bolphi); 1308 assert(iff->in(1) == bolphi, ""); 1309 1310 if (bolphi->Value(&_igvn)->singleton()) { 1311 return; 1312 } 1313 1314 // Conditional-move? Must split up now 1315 if (!iff->is_If()) { 1316 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1); 1317 _igvn.replace_node(iff, cmovphi); 1318 return; 1319 } 1320 1321 // Now split the IF 1322 do_split_if(iff); 1323 return; 1324 } 1325 1326 // Two identical ifs back to back can be merged 1327 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) { 1328 Node *n_ctrl = n->in(0); 1329 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1)); 1330 IfNode* dom_if = idom(n_ctrl)->as_If(); 1331 Node* proj_true = dom_if->proj_out(1); 1332 Node* proj_false = dom_if->proj_out(0); 1333 Node* con_true = _igvn.makecon(TypeInt::ONE); 1334 Node* con_false = _igvn.makecon(TypeInt::ZERO); 1335 1336 for (uint i = 1; i < n_ctrl->req(); i++) { 1337 if (is_dominator(proj_true, n_ctrl->in(i))) { 1338 bolphi->init_req(i, con_true); 1339 } else { 1340 assert(is_dominator(proj_false, n_ctrl->in(i)), "bad if"); 1341 bolphi->init_req(i, con_false); 1342 } 1343 } 1344 register_new_node(bolphi, n_ctrl); 1345 _igvn.replace_input_of(n, 1, bolphi); 1346 1347 // Now split the IF 1348 do_split_if(n); 1349 return; 1350 } 1351 1352 // Check for an IF ready to split; one that has its 1353 // condition codes input coming from a Phi at the block start. 1354 int n_op = n->Opcode(); 1355 1356 // Check for an IF being dominated by another IF same test 1357 if (n_op == Op_If || 1358 n_op == Op_RangeCheck) { 1359 Node *bol = n->in(1); 1360 uint max = bol->outcnt(); 1361 // Check for same test used more than once? 1362 if (max > 1 && bol->is_Bool()) { 1363 // Search up IDOMs to see if this IF is dominated. 1364 Node *cutoff = get_ctrl(bol); 1365 1366 // Now search up IDOMs till cutoff, looking for a dominating test 1367 Node *prevdom = n; 1368 Node *dom = idom(prevdom); 1369 while (dom != cutoff) { 1370 if (dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom) { 1371 // It's invalid to move control dependent data nodes in the inner 1372 // strip-mined loop, because: 1373 // 1) break validation of LoopNode::verify_strip_mined() 1374 // 2) move code with side-effect in strip-mined loop 1375 // Move to the exit of outer strip-mined loop in that case. 1376 Node* out_le = is_inner_of_stripmined_loop(dom); 1377 if (out_le != NULL) { 1378 prevdom = out_le; 1379 } 1380 // Replace the dominated test with an obvious true or false. 1381 // Place it on the IGVN worklist for later cleanup. 1382 C->set_major_progress(); 1383 dominated_by(prevdom, n, false, true); 1384 #ifndef PRODUCT 1385 if( VerifyLoopOptimizations ) verify(); 1386 #endif 1387 return; 1388 } 1389 prevdom = dom; 1390 dom = idom(prevdom); 1391 } 1392 } 1393 } 1394 1395 // See if a shared loop-varying computation has no loop-varying uses. 1396 // Happens if something is only used for JVM state in uncommon trap exits, 1397 // like various versions of induction variable+offset. Clone the 1398 // computation per usage to allow it to sink out of the loop. 1399 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about) 1400 Node *n_ctrl = get_ctrl(n); 1401 IdealLoopTree *n_loop = get_loop(n_ctrl); 1402 if( n_loop != _ltree_root ) { 1403 DUIterator_Fast imax, i = n->fast_outs(imax); 1404 for (; i < imax; i++) { 1405 Node* u = n->fast_out(i); 1406 if( !has_ctrl(u) ) break; // Found control user 1407 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 1408 if( u_loop == n_loop ) break; // Found loop-varying use 1409 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop 1410 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003 1411 } 1412 bool did_break = (i < imax); // Did we break out of the previous loop? 1413 if (!did_break && n->outcnt() > 1) { // All uses in outer loops! 1414 Node *late_load_ctrl = NULL; 1415 if (n->is_Load()) { 1416 // If n is a load, get and save the result from get_late_ctrl(), 1417 // to be later used in calculating the control for n's clones. 1418 clear_dom_lca_tags(); 1419 late_load_ctrl = get_late_ctrl(n, n_ctrl); 1420 } 1421 // If n is a load, and the late control is the same as the current 1422 // control, then the cloning of n is a pointless exercise, because 1423 // GVN will ensure that we end up where we started. 1424 if (!n->is_Load() || late_load_ctrl != n_ctrl) { 1425 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) { 1426 Node *u = n->last_out(j); // Clone private computation per use 1427 _igvn.rehash_node_delayed(u); 1428 Node *x = n->clone(); // Clone computation 1429 Node *x_ctrl = NULL; 1430 if( u->is_Phi() ) { 1431 // Replace all uses of normal nodes. Replace Phi uses 1432 // individually, so the separate Nodes can sink down 1433 // different paths. 1434 uint k = 1; 1435 while( u->in(k) != n ) k++; 1436 u->set_req( k, x ); 1437 // x goes next to Phi input path 1438 x_ctrl = u->in(0)->in(k); 1439 --j; 1440 } else { // Normal use 1441 // Replace all uses 1442 for( uint k = 0; k < u->req(); k++ ) { 1443 if( u->in(k) == n ) { 1444 u->set_req( k, x ); 1445 --j; 1446 } 1447 } 1448 x_ctrl = get_ctrl(u); 1449 } 1450 1451 // Find control for 'x' next to use but not inside inner loops. 1452 // For inner loop uses get the preheader area. 1453 x_ctrl = place_near_use(x_ctrl); 1454 1455 if (n->is_Load()) { 1456 // For loads, add a control edge to a CFG node outside of the loop 1457 // to force them to not combine and return back inside the loop 1458 // during GVN optimization (4641526). 1459 // 1460 // Because we are setting the actual control input, factor in 1461 // the result from get_late_ctrl() so we respect any 1462 // anti-dependences. (6233005). 1463 x_ctrl = dom_lca(late_load_ctrl, x_ctrl); 1464 1465 // Don't allow the control input to be a CFG splitting node. 1466 // Such nodes should only have ProjNodes as outs, e.g. IfNode 1467 // should only have IfTrueNode and IfFalseNode (4985384). 1468 x_ctrl = find_non_split_ctrl(x_ctrl); 1469 1470 IdealLoopTree* x_loop = get_loop(x_ctrl); 1471 Node* x_head = x_loop->_head; 1472 if (x_head->is_Loop() && (x_head->is_OuterStripMinedLoop() || x_head->as_Loop()->is_strip_mined()) && is_dominator(n_ctrl, x_head)) { 1473 // Anti dependence analysis is sometimes too 1474 // conservative: a store in the outer strip mined loop 1475 // can prevent a load from floating out of the outer 1476 // strip mined loop but the load may not be referenced 1477 // from the safepoint: loop strip mining verification 1478 // code reports a problem in that case. Make sure the 1479 // load is not moved in the outer strip mined loop in 1480 // that case. 1481 x_ctrl = x_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl); 1482 } 1483 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone"); 1484 1485 x->set_req(0, x_ctrl); 1486 } 1487 register_new_node(x, x_ctrl); 1488 1489 // Some institutional knowledge is needed here: 'x' is 1490 // yanked because if the optimizer runs GVN on it all the 1491 // cloned x's will common up and undo this optimization and 1492 // be forced back in the loop. 1493 // I tried setting control edges on the x's to force them to 1494 // not combine, but the matching gets worried when it tries 1495 // to fold a StoreP and an AddP together (as part of an 1496 // address expression) and the AddP and StoreP have 1497 // different controls. 1498 if (!x->is_Load() && !x->is_DecodeNarrowPtr()) _igvn._worklist.yank(x); 1499 } 1500 _igvn.remove_dead_node(n); 1501 } 1502 } 1503 } 1504 } 1505 1506 try_move_store_after_loop(n); 1507 1508 // Check for Opaque2's who's loop has disappeared - who's input is in the 1509 // same loop nest as their output. Remove 'em, they are no longer useful. 1510 if( n_op == Op_Opaque2 && 1511 n->in(1) != NULL && 1512 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) { 1513 _igvn.replace_node( n, n->in(1) ); 1514 } 1515 } 1516 1517 //------------------------------split_if_with_blocks--------------------------- 1518 // Check for aggressive application of 'split-if' optimization, 1519 // using basic block level info. 1520 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack) { 1521 Node* root = C->root(); 1522 visited.set(root->_idx); // first, mark root as visited 1523 // Do pre-visit work for root 1524 Node* n = split_if_with_blocks_pre(root); 1525 uint cnt = n->outcnt(); 1526 uint i = 0; 1527 1528 while (true) { 1529 // Visit all children 1530 if (i < cnt) { 1531 Node* use = n->raw_out(i); 1532 ++i; 1533 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) { 1534 // Now do pre-visit work for this use 1535 use = split_if_with_blocks_pre(use); 1536 nstack.push(n, i); // Save parent and next use's index. 1537 n = use; // Process all children of current use. 1538 cnt = use->outcnt(); 1539 i = 0; 1540 } 1541 } 1542 else { 1543 // All of n's children have been processed, complete post-processing. 1544 if (cnt != 0 && !n->is_Con()) { 1545 assert(has_node(n), "no dead nodes"); 1546 split_if_with_blocks_post(n); 1547 } 1548 if (must_throttle_split_if()) { 1549 nstack.clear(); 1550 } 1551 if (nstack.is_empty()) { 1552 // Finished all nodes on stack. 1553 break; 1554 } 1555 // Get saved parent node and next use's index. Visit the rest of uses. 1556 n = nstack.node(); 1557 cnt = n->outcnt(); 1558 i = nstack.index(); 1559 nstack.pop(); 1560 } 1561 } 1562 } 1563 1564 1565 //============================================================================= 1566 // 1567 // C L O N E A L O O P B O D Y 1568 // 1569 1570 //------------------------------clone_iff-------------------------------------- 1571 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1572 // "Nearly" because all Nodes have been cloned from the original in the loop, 1573 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1574 // through the Phi recursively, and return a Bool. 1575 Node* PhaseIdealLoop::clone_iff(PhiNode *phi, IdealLoopTree *loop) { 1576 1577 // Convert this Phi into a Phi merging Bools 1578 uint i; 1579 for (i = 1; i < phi->req(); i++) { 1580 Node *b = phi->in(i); 1581 if (b->is_Phi()) { 1582 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi(), loop)); 1583 } else { 1584 assert(b->is_Bool() || b->Opcode() == Op_Opaque4, ""); 1585 } 1586 } 1587 1588 Node* n = phi->in(1); 1589 Node* sample_opaque = NULL; 1590 Node *sample_bool = NULL; 1591 if (n->Opcode() == Op_Opaque4) { 1592 sample_opaque = n; 1593 sample_bool = n->in(1); 1594 assert(sample_bool->is_Bool(), "wrong type"); 1595 } else { 1596 sample_bool = n; 1597 } 1598 Node *sample_cmp = sample_bool->in(1); 1599 1600 // Make Phis to merge the Cmp's inputs. 1601 PhiNode *phi1 = new PhiNode(phi->in(0), Type::TOP); 1602 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP); 1603 for (i = 1; i < phi->req(); i++) { 1604 Node *n1 = sample_opaque == NULL ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1); 1605 Node *n2 = sample_opaque == NULL ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2); 1606 phi1->set_req(i, n1); 1607 phi2->set_req(i, n2); 1608 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 1609 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 1610 } 1611 // See if these Phis have been made before. 1612 // Register with optimizer 1613 Node *hit1 = _igvn.hash_find_insert(phi1); 1614 if (hit1) { // Hit, toss just made Phi 1615 _igvn.remove_dead_node(phi1); // Remove new phi 1616 assert(hit1->is_Phi(), "" ); 1617 phi1 = (PhiNode*)hit1; // Use existing phi 1618 } else { // Miss 1619 _igvn.register_new_node_with_optimizer(phi1); 1620 } 1621 Node *hit2 = _igvn.hash_find_insert(phi2); 1622 if (hit2) { // Hit, toss just made Phi 1623 _igvn.remove_dead_node(phi2); // Remove new phi 1624 assert(hit2->is_Phi(), "" ); 1625 phi2 = (PhiNode*)hit2; // Use existing phi 1626 } else { // Miss 1627 _igvn.register_new_node_with_optimizer(phi2); 1628 } 1629 // Register Phis with loop/block info 1630 set_ctrl(phi1, phi->in(0)); 1631 set_ctrl(phi2, phi->in(0)); 1632 // Make a new Cmp 1633 Node *cmp = sample_cmp->clone(); 1634 cmp->set_req(1, phi1); 1635 cmp->set_req(2, phi2); 1636 _igvn.register_new_node_with_optimizer(cmp); 1637 set_ctrl(cmp, phi->in(0)); 1638 1639 // Make a new Bool 1640 Node *b = sample_bool->clone(); 1641 b->set_req(1,cmp); 1642 _igvn.register_new_node_with_optimizer(b); 1643 set_ctrl(b, phi->in(0)); 1644 1645 if (sample_opaque != NULL) { 1646 Node* opaque = sample_opaque->clone(); 1647 opaque->set_req(1, b); 1648 _igvn.register_new_node_with_optimizer(opaque); 1649 set_ctrl(opaque, phi->in(0)); 1650 return opaque; 1651 } 1652 1653 assert(b->is_Bool(), ""); 1654 return b; 1655 } 1656 1657 //------------------------------clone_bool------------------------------------- 1658 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1659 // "Nearly" because all Nodes have been cloned from the original in the loop, 1660 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1661 // through the Phi recursively, and return a Bool. 1662 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) { 1663 uint i; 1664 // Convert this Phi into a Phi merging Bools 1665 for( i = 1; i < phi->req(); i++ ) { 1666 Node *b = phi->in(i); 1667 if( b->is_Phi() ) { 1668 _igvn.replace_input_of(phi, i, clone_bool( b->as_Phi(), loop )); 1669 } else { 1670 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" ); 1671 } 1672 } 1673 1674 Node *sample_cmp = phi->in(1); 1675 1676 // Make Phis to merge the Cmp's inputs. 1677 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP ); 1678 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP ); 1679 for( uint j = 1; j < phi->req(); j++ ) { 1680 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP 1681 Node *n1, *n2; 1682 if( cmp_top->is_Cmp() ) { 1683 n1 = cmp_top->in(1); 1684 n2 = cmp_top->in(2); 1685 } else { 1686 n1 = n2 = cmp_top; 1687 } 1688 phi1->set_req( j, n1 ); 1689 phi2->set_req( j, n2 ); 1690 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 1691 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 1692 } 1693 1694 // See if these Phis have been made before. 1695 // Register with optimizer 1696 Node *hit1 = _igvn.hash_find_insert(phi1); 1697 if( hit1 ) { // Hit, toss just made Phi 1698 _igvn.remove_dead_node(phi1); // Remove new phi 1699 assert( hit1->is_Phi(), "" ); 1700 phi1 = (PhiNode*)hit1; // Use existing phi 1701 } else { // Miss 1702 _igvn.register_new_node_with_optimizer(phi1); 1703 } 1704 Node *hit2 = _igvn.hash_find_insert(phi2); 1705 if( hit2 ) { // Hit, toss just made Phi 1706 _igvn.remove_dead_node(phi2); // Remove new phi 1707 assert( hit2->is_Phi(), "" ); 1708 phi2 = (PhiNode*)hit2; // Use existing phi 1709 } else { // Miss 1710 _igvn.register_new_node_with_optimizer(phi2); 1711 } 1712 // Register Phis with loop/block info 1713 set_ctrl(phi1, phi->in(0)); 1714 set_ctrl(phi2, phi->in(0)); 1715 // Make a new Cmp 1716 Node *cmp = sample_cmp->clone(); 1717 cmp->set_req( 1, phi1 ); 1718 cmp->set_req( 2, phi2 ); 1719 _igvn.register_new_node_with_optimizer(cmp); 1720 set_ctrl(cmp, phi->in(0)); 1721 1722 assert( cmp->is_Cmp(), "" ); 1723 return (CmpNode*)cmp; 1724 } 1725 1726 //------------------------------sink_use--------------------------------------- 1727 // If 'use' was in the loop-exit block, it now needs to be sunk 1728 // below the post-loop merge point. 1729 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) { 1730 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) { 1731 set_ctrl(use, post_loop); 1732 for (DUIterator j = use->outs(); use->has_out(j); j++) 1733 sink_use(use->out(j), post_loop); 1734 } 1735 } 1736 1737 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new, 1738 IdealLoopTree* loop, IdealLoopTree* outer_loop, 1739 Node_List*& split_if_set, Node_List*& split_bool_set, 1740 Node_List*& split_cex_set, Node_List& worklist, 1741 uint new_counter, CloneLoopMode mode) { 1742 Node* nnn = old_new[old->_idx]; 1743 // Copy uses to a worklist, so I can munge the def-use info 1744 // with impunity. 1745 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1746 worklist.push(old->fast_out(j)); 1747 1748 while( worklist.size() ) { 1749 Node *use = worklist.pop(); 1750 if (!has_node(use)) continue; // Ignore dead nodes 1751 if (use->in(0) == C->top()) continue; 1752 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1753 // Check for data-use outside of loop - at least one of OLD or USE 1754 // must not be a CFG node. 1755 #ifdef ASSERT 1756 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == NULL) { 1757 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint(); 1758 assert(mode != IgnoreStripMined, "incorrect cloning mode"); 1759 assert((mode == ControlAroundStripMined && use == sfpt) || !use->is_reachable_from_root(), "missed a node"); 1760 } 1761 #endif 1762 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) { 1763 1764 // If the Data use is an IF, that means we have an IF outside of the 1765 // loop that is switching on a condition that is set inside of the 1766 // loop. Happens if people set a loop-exit flag; then test the flag 1767 // in the loop to break the loop, then test is again outside of the 1768 // loop to determine which way the loop exited. 1769 // Loop predicate If node connects to Bool node through Opaque1 node. 1770 if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use) || use->Opcode() == Op_Opaque4) { 1771 // Since this code is highly unlikely, we lazily build the worklist 1772 // of such Nodes to go split. 1773 if (!split_if_set) { 1774 split_if_set = new Node_List(); 1775 } 1776 split_if_set->push(use); 1777 } 1778 if (use->is_Bool()) { 1779 if (!split_bool_set) { 1780 split_bool_set = new Node_List(); 1781 } 1782 split_bool_set->push(use); 1783 } 1784 if (use->Opcode() == Op_CreateEx) { 1785 if (!split_cex_set) { 1786 split_cex_set = new Node_List(); 1787 } 1788 split_cex_set->push(use); 1789 } 1790 1791 1792 // Get "block" use is in 1793 uint idx = 0; 1794 while( use->in(idx) != old ) idx++; 1795 Node *prev = use->is_CFG() ? use : get_ctrl(use); 1796 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" ); 1797 Node *cfg = prev->_idx >= new_counter 1798 ? prev->in(2) 1799 : idom(prev); 1800 if( use->is_Phi() ) // Phi use is in prior block 1801 cfg = prev->in(idx); // NOT in block of Phi itself 1802 if (cfg->is_top()) { // Use is dead? 1803 _igvn.replace_input_of(use, idx, C->top()); 1804 continue; 1805 } 1806 1807 // If use is referenced through control edge... (idx == 0) 1808 if (mode == IgnoreStripMined && idx == 0) { 1809 LoopNode *head = loop->_head->as_Loop(); 1810 if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) { 1811 // That node is outside the inner loop, leave it outside the 1812 // outer loop as well to not confuse verification code. 1813 assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop"); 1814 _igvn.replace_input_of(use, 0, head->outer_loop_exit()); 1815 continue; 1816 } 1817 } 1818 1819 while(!outer_loop->is_member(get_loop(cfg))) { 1820 prev = cfg; 1821 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg); 1822 } 1823 // If the use occurs after merging several exits from the loop, then 1824 // old value must have dominated all those exits. Since the same old 1825 // value was used on all those exits we did not need a Phi at this 1826 // merge point. NOW we do need a Phi here. Each loop exit value 1827 // is now merged with the peeled body exit; each exit gets its own 1828 // private Phi and those Phis need to be merged here. 1829 Node *phi; 1830 if( prev->is_Region() ) { 1831 if( idx == 0 ) { // Updating control edge? 1832 phi = prev; // Just use existing control 1833 } else { // Else need a new Phi 1834 phi = PhiNode::make( prev, old ); 1835 // Now recursively fix up the new uses of old! 1836 for( uint i = 1; i < prev->req(); i++ ) { 1837 worklist.push(phi); // Onto worklist once for each 'old' input 1838 } 1839 } 1840 } else { 1841 // Get new RegionNode merging old and new loop exits 1842 prev = old_new[prev->_idx]; 1843 assert( prev, "just made this in step 7" ); 1844 if( idx == 0) { // Updating control edge? 1845 phi = prev; // Just use existing control 1846 } else { // Else need a new Phi 1847 // Make a new Phi merging data values properly 1848 phi = PhiNode::make( prev, old ); 1849 phi->set_req( 1, nnn ); 1850 } 1851 } 1852 // If inserting a new Phi, check for prior hits 1853 if( idx != 0 ) { 1854 Node *hit = _igvn.hash_find_insert(phi); 1855 if( hit == NULL ) { 1856 _igvn.register_new_node_with_optimizer(phi); // Register new phi 1857 } else { // or 1858 // Remove the new phi from the graph and use the hit 1859 _igvn.remove_dead_node(phi); 1860 phi = hit; // Use existing phi 1861 } 1862 set_ctrl(phi, prev); 1863 } 1864 // Make 'use' use the Phi instead of the old loop body exit value 1865 _igvn.replace_input_of(use, idx, phi); 1866 if( use->_idx >= new_counter ) { // If updating new phis 1867 // Not needed for correctness, but prevents a weak assert 1868 // in AddPNode from tripping (when we end up with different 1869 // base & derived Phis that will become the same after 1870 // IGVN does CSE). 1871 Node *hit = _igvn.hash_find_insert(use); 1872 if( hit ) // Go ahead and re-hash for hits. 1873 _igvn.replace_node( use, hit ); 1874 } 1875 1876 // If 'use' was in the loop-exit block, it now needs to be sunk 1877 // below the post-loop merge point. 1878 sink_use( use, prev ); 1879 } 1880 } 1881 } 1882 1883 static void clone_outer_loop_helper(Node* n, const IdealLoopTree *loop, const IdealLoopTree* outer_loop, 1884 const Node_List &old_new, Unique_Node_List& wq, PhaseIdealLoop* phase, 1885 bool check_old_new) { 1886 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1887 Node* u = n->fast_out(j); 1888 assert(check_old_new || old_new[u->_idx] == NULL, "shouldn't have been cloned"); 1889 if (!u->is_CFG() && (!check_old_new || old_new[u->_idx] == NULL)) { 1890 Node* c = phase->get_ctrl(u); 1891 IdealLoopTree* u_loop = phase->get_loop(c); 1892 assert(!loop->is_member(u_loop), "can be in outer loop or out of both loops only"); 1893 if (outer_loop->is_member(u_loop)) { 1894 wq.push(u); 1895 } 1896 } 1897 } 1898 } 1899 1900 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop, 1901 IdealLoopTree* outer_loop, int dd, Node_List &old_new, 1902 Node_List& extra_data_nodes) { 1903 if (head->is_strip_mined() && mode != IgnoreStripMined) { 1904 CountedLoopNode* cl = head->as_CountedLoop(); 1905 Node* l = cl->outer_loop(); 1906 Node* tail = cl->outer_loop_tail(); 1907 IfNode* le = cl->outer_loop_end(); 1908 Node* sfpt = cl->outer_safepoint(); 1909 CountedLoopEndNode* cle = cl->loopexit(); 1910 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop(); 1911 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null(); 1912 Node* cle_out = cle->proj_out(false); 1913 1914 Node* new_sfpt = NULL; 1915 Node* new_cle_out = cle_out->clone(); 1916 old_new.map(cle_out->_idx, new_cle_out); 1917 if (mode == CloneIncludesStripMined) { 1918 // clone outer loop body 1919 Node* new_l = l->clone(); 1920 Node* new_tail = tail->clone(); 1921 IfNode* new_le = le->clone()->as_If(); 1922 new_sfpt = sfpt->clone(); 1923 1924 set_loop(new_l, outer_loop->_parent); 1925 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd); 1926 set_loop(new_cle_out, outer_loop->_parent); 1927 set_idom(new_cle_out, new_cle, dd); 1928 set_loop(new_sfpt, outer_loop->_parent); 1929 set_idom(new_sfpt, new_cle_out, dd); 1930 set_loop(new_le, outer_loop->_parent); 1931 set_idom(new_le, new_sfpt, dd); 1932 set_loop(new_tail, outer_loop->_parent); 1933 set_idom(new_tail, new_le, dd); 1934 set_idom(new_cl, new_l, dd); 1935 1936 old_new.map(l->_idx, new_l); 1937 old_new.map(tail->_idx, new_tail); 1938 old_new.map(le->_idx, new_le); 1939 old_new.map(sfpt->_idx, new_sfpt); 1940 1941 new_l->set_req(LoopNode::LoopBackControl, new_tail); 1942 new_l->set_req(0, new_l); 1943 new_tail->set_req(0, new_le); 1944 new_le->set_req(0, new_sfpt); 1945 new_sfpt->set_req(0, new_cle_out); 1946 new_cle_out->set_req(0, new_cle); 1947 new_cl->set_req(LoopNode::EntryControl, new_l); 1948 1949 _igvn.register_new_node_with_optimizer(new_l); 1950 _igvn.register_new_node_with_optimizer(new_tail); 1951 _igvn.register_new_node_with_optimizer(new_le); 1952 } else { 1953 Node *newhead = old_new[loop->_head->_idx]; 1954 newhead->as_Loop()->clear_strip_mined(); 1955 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl)); 1956 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 1957 } 1958 // Look at data node that were assigned a control in the outer 1959 // loop: they are kept in the outer loop by the safepoint so start 1960 // from the safepoint node's inputs. 1961 IdealLoopTree* outer_loop = get_loop(l); 1962 Node_Stack stack(2); 1963 stack.push(sfpt, 1); 1964 uint new_counter = C->unique(); 1965 while (stack.size() > 0) { 1966 Node* n = stack.node(); 1967 uint i = stack.index(); 1968 while (i < n->req() && 1969 (n->in(i) == NULL || 1970 !has_ctrl(n->in(i)) || 1971 get_loop(get_ctrl(n->in(i))) != outer_loop || 1972 (old_new[n->in(i)->_idx] != NULL && old_new[n->in(i)->_idx]->_idx >= new_counter))) { 1973 i++; 1974 } 1975 if (i < n->req()) { 1976 stack.set_index(i+1); 1977 stack.push(n->in(i), 0); 1978 } else { 1979 assert(old_new[n->_idx] == NULL || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet"); 1980 Node* m = n == sfpt ? new_sfpt : n->clone(); 1981 if (m != NULL) { 1982 for (uint i = 0; i < n->req(); i++) { 1983 if (m->in(i) != NULL && old_new[m->in(i)->_idx] != NULL) { 1984 m->set_req(i, old_new[m->in(i)->_idx]); 1985 } 1986 } 1987 } else { 1988 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?"); 1989 } 1990 if (n != sfpt) { 1991 extra_data_nodes.push(n); 1992 _igvn.register_new_node_with_optimizer(m); 1993 assert(get_ctrl(n) == cle_out, "what other control?"); 1994 set_ctrl(m, new_cle_out); 1995 old_new.map(n->_idx, m); 1996 } 1997 stack.pop(); 1998 } 1999 } 2000 if (mode == CloneIncludesStripMined) { 2001 _igvn.register_new_node_with_optimizer(new_sfpt); 2002 _igvn.register_new_node_with_optimizer(new_cle_out); 2003 } 2004 // Some other transformation may have pessimistically assign some 2005 // data nodes to the outer loop. Set their control so they are out 2006 // of the outer loop. 2007 ResourceMark rm; 2008 Unique_Node_List wq; 2009 for (uint i = 0; i < extra_data_nodes.size(); i++) { 2010 Node* old = extra_data_nodes.at(i); 2011 clone_outer_loop_helper(old, loop, outer_loop, old_new, wq, this, true); 2012 } 2013 Node* new_ctrl = cl->outer_loop_exit(); 2014 assert(get_loop(new_ctrl) != outer_loop, "must be out of the loop nest"); 2015 for (uint i = 0; i < wq.size(); i++) { 2016 Node* n = wq.at(i); 2017 set_ctrl(n, new_ctrl); 2018 clone_outer_loop_helper(n, loop, outer_loop, old_new, wq, this, false); 2019 } 2020 } else { 2021 Node *newhead = old_new[loop->_head->_idx]; 2022 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 2023 } 2024 } 2025 2026 //------------------------------clone_loop------------------------------------- 2027 // 2028 // C L O N E A L O O P B O D Y 2029 // 2030 // This is the basic building block of the loop optimizations. It clones an 2031 // entire loop body. It makes an old_new loop body mapping; with this mapping 2032 // you can find the new-loop equivalent to an old-loop node. All new-loop 2033 // nodes are exactly equal to their old-loop counterparts, all edges are the 2034 // same. All exits from the old-loop now have a RegionNode that merges the 2035 // equivalent new-loop path. This is true even for the normal "loop-exit" 2036 // condition. All uses of loop-invariant old-loop values now come from (one 2037 // or more) Phis that merge their new-loop equivalents. 2038 // 2039 // This operation leaves the graph in an illegal state: there are two valid 2040 // control edges coming from the loop pre-header to both loop bodies. I'll 2041 // definitely have to hack the graph after running this transform. 2042 // 2043 // From this building block I will further edit edges to perform loop peeling 2044 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc. 2045 // 2046 // Parameter side_by_size_idom: 2047 // When side_by_size_idom is NULL, the dominator tree is constructed for 2048 // the clone loop to dominate the original. Used in construction of 2049 // pre-main-post loop sequence. 2050 // When nonnull, the clone and original are side-by-side, both are 2051 // dominated by the side_by_side_idom node. Used in construction of 2052 // unswitched loops. 2053 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd, 2054 CloneLoopMode mode, Node* side_by_side_idom) { 2055 2056 LoopNode* head = loop->_head->as_Loop(); 2057 head->verify_strip_mined(1); 2058 2059 if (C->do_vector_loop() && PrintOpto) { 2060 const char* mname = C->method()->name()->as_quoted_ascii(); 2061 if (mname != NULL) { 2062 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname); 2063 } 2064 } 2065 2066 CloneMap& cm = C->clone_map(); 2067 Dict* dict = cm.dict(); 2068 if (C->do_vector_loop()) { 2069 cm.set_clone_idx(cm.max_gen()+1); 2070 #ifndef PRODUCT 2071 if (PrintOpto) { 2072 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx()); 2073 loop->dump_head(); 2074 } 2075 #endif 2076 } 2077 2078 // Step 1: Clone the loop body. Make the old->new mapping. 2079 uint i; 2080 for( i = 0; i < loop->_body.size(); i++ ) { 2081 Node *old = loop->_body.at(i); 2082 Node *nnn = old->clone(); 2083 old_new.map( old->_idx, nnn ); 2084 if (C->do_vector_loop()) { 2085 cm.verify_insert_and_clone(old, nnn, cm.clone_idx()); 2086 } 2087 _igvn.register_new_node_with_optimizer(nnn); 2088 } 2089 2090 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop; 2091 2092 // Step 2: Fix the edges in the new body. If the old input is outside the 2093 // loop use it. If the old input is INside the loop, use the corresponding 2094 // new node instead. 2095 for( i = 0; i < loop->_body.size(); i++ ) { 2096 Node *old = loop->_body.at(i); 2097 Node *nnn = old_new[old->_idx]; 2098 // Fix CFG/Loop controlling the new node 2099 if (has_ctrl(old)) { 2100 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]); 2101 } else { 2102 set_loop(nnn, outer_loop->_parent); 2103 if (old->outcnt() > 0) { 2104 set_idom( nnn, old_new[idom(old)->_idx], dd ); 2105 } 2106 } 2107 // Correct edges to the new node 2108 for( uint j = 0; j < nnn->req(); j++ ) { 2109 Node *n = nnn->in(j); 2110 if( n ) { 2111 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n ); 2112 if( loop->is_member( old_in_loop ) ) 2113 nnn->set_req(j, old_new[n->_idx]); 2114 } 2115 } 2116 _igvn.hash_find_insert(nnn); 2117 } 2118 2119 Node_List extra_data_nodes; // data nodes in the outer strip mined loop 2120 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes); 2121 2122 // Step 3: Now fix control uses. Loop varying control uses have already 2123 // been fixed up (as part of all input edges in Step 2). Loop invariant 2124 // control uses must be either an IfFalse or an IfTrue. Make a merge 2125 // point to merge the old and new IfFalse/IfTrue nodes; make the use 2126 // refer to this. 2127 Node_List worklist; 2128 uint new_counter = C->unique(); 2129 for( i = 0; i < loop->_body.size(); i++ ) { 2130 Node* old = loop->_body.at(i); 2131 if( !old->is_CFG() ) continue; 2132 2133 // Copy uses to a worklist, so I can munge the def-use info 2134 // with impunity. 2135 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 2136 worklist.push(old->fast_out(j)); 2137 2138 while( worklist.size() ) { // Visit all uses 2139 Node *use = worklist.pop(); 2140 if (!has_node(use)) continue; // Ignore dead nodes 2141 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 2142 if( !loop->is_member( use_loop ) && use->is_CFG() ) { 2143 // Both OLD and USE are CFG nodes here. 2144 assert( use->is_Proj(), "" ); 2145 Node* nnn = old_new[old->_idx]; 2146 2147 Node* newuse = NULL; 2148 if (head->is_strip_mined() && mode != IgnoreStripMined) { 2149 CountedLoopNode* cl = head->as_CountedLoop(); 2150 CountedLoopEndNode* cle = cl->loopexit(); 2151 Node* cle_out = cle->proj_out_or_null(false); 2152 if (use == cle_out) { 2153 IfNode* le = cl->outer_loop_end(); 2154 use = le->proj_out(false); 2155 use_loop = get_loop(use); 2156 if (mode == CloneIncludesStripMined) { 2157 nnn = old_new[le->_idx]; 2158 } else { 2159 newuse = old_new[cle_out->_idx]; 2160 } 2161 } 2162 } 2163 if (newuse == NULL) { 2164 newuse = use->clone(); 2165 } 2166 2167 // Clone the loop exit control projection 2168 if (C->do_vector_loop()) { 2169 cm.verify_insert_and_clone(use, newuse, cm.clone_idx()); 2170 } 2171 newuse->set_req(0,nnn); 2172 _igvn.register_new_node_with_optimizer(newuse); 2173 set_loop(newuse, use_loop); 2174 set_idom(newuse, nnn, dom_depth(nnn) + 1 ); 2175 2176 // We need a Region to merge the exit from the peeled body and the 2177 // exit from the old loop body. 2178 RegionNode *r = new RegionNode(3); 2179 // Map the old use to the new merge point 2180 old_new.map( use->_idx, r ); 2181 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use)); 2182 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" ); 2183 2184 // The original user of 'use' uses 'r' instead. 2185 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) { 2186 Node* useuse = use->last_out(l); 2187 _igvn.rehash_node_delayed(useuse); 2188 uint uses_found = 0; 2189 if( useuse->in(0) == use ) { 2190 useuse->set_req(0, r); 2191 uses_found++; 2192 if( useuse->is_CFG() ) { 2193 // This is not a dom_depth > dd_r because when new 2194 // control flow is constructed by a loop opt, a node and 2195 // its dominator can end up at the same dom_depth 2196 assert(dom_depth(useuse) >= dd_r, ""); 2197 set_idom(useuse, r, dom_depth(useuse)); 2198 } 2199 } 2200 for( uint k = 1; k < useuse->req(); k++ ) { 2201 if( useuse->in(k) == use ) { 2202 useuse->set_req(k, r); 2203 uses_found++; 2204 if (useuse->is_Loop() && k == LoopNode::EntryControl) { 2205 // This is not a dom_depth > dd_r because when new 2206 // control flow is constructed by a loop opt, a node 2207 // and its dominator can end up at the same dom_depth 2208 assert(dom_depth(useuse) >= dd_r , ""); 2209 set_idom(useuse, r, dom_depth(useuse)); 2210 } 2211 } 2212 } 2213 l -= uses_found; // we deleted 1 or more copies of this edge 2214 } 2215 2216 // Now finish up 'r' 2217 r->set_req( 1, newuse ); 2218 r->set_req( 2, use ); 2219 _igvn.register_new_node_with_optimizer(r); 2220 set_loop(r, use_loop); 2221 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r); 2222 } // End of if a loop-exit test 2223 } 2224 } 2225 2226 // Step 4: If loop-invariant use is not control, it must be dominated by a 2227 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region 2228 // there if needed. Make a Phi there merging old and new used values. 2229 Node_List *split_if_set = NULL; 2230 Node_List *split_bool_set = NULL; 2231 Node_List *split_cex_set = NULL; 2232 for( i = 0; i < loop->_body.size(); i++ ) { 2233 Node* old = loop->_body.at(i); 2234 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set, 2235 split_bool_set, split_cex_set, worklist, new_counter, 2236 mode); 2237 } 2238 2239 for (i = 0; i < extra_data_nodes.size(); i++) { 2240 Node* old = extra_data_nodes.at(i); 2241 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set, 2242 split_bool_set, split_cex_set, worklist, new_counter, 2243 mode); 2244 } 2245 2246 // Check for IFs that need splitting/cloning. Happens if an IF outside of 2247 // the loop uses a condition set in the loop. The original IF probably 2248 // takes control from one or more OLD Regions (which in turn get from NEW 2249 // Regions). In any case, there will be a set of Phis for each merge point 2250 // from the IF up to where the original BOOL def exists the loop. 2251 if (split_if_set) { 2252 while (split_if_set->size()) { 2253 Node *iff = split_if_set->pop(); 2254 if (iff->in(1)->is_Phi()) { 2255 Node *b = clone_iff(iff->in(1)->as_Phi(), loop); 2256 _igvn.replace_input_of(iff, 1, b); 2257 } 2258 } 2259 } 2260 if (split_bool_set) { 2261 while (split_bool_set->size()) { 2262 Node *b = split_bool_set->pop(); 2263 Node *phi = b->in(1); 2264 assert(phi->is_Phi(), ""); 2265 CmpNode *cmp = clone_bool((PhiNode*)phi, loop); 2266 _igvn.replace_input_of(b, 1, cmp); 2267 } 2268 } 2269 if (split_cex_set) { 2270 while (split_cex_set->size()) { 2271 Node *b = split_cex_set->pop(); 2272 assert(b->in(0)->is_Region(), ""); 2273 assert(b->in(1)->is_Phi(), ""); 2274 assert(b->in(0)->in(0) == b->in(1)->in(0), ""); 2275 split_up(b, b->in(0), NULL); 2276 } 2277 } 2278 2279 } 2280 2281 2282 //---------------------- stride_of_possible_iv ------------------------------------- 2283 // Looks for an iff/bool/comp with one operand of the compare 2284 // being a cycle involving an add and a phi, 2285 // with an optional truncation (left-shift followed by a right-shift) 2286 // of the add. Returns zero if not an iv. 2287 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) { 2288 Node* trunc1 = NULL; 2289 Node* trunc2 = NULL; 2290 const TypeInt* ttype = NULL; 2291 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) { 2292 return 0; 2293 } 2294 BoolNode* bl = iff->in(1)->as_Bool(); 2295 Node* cmp = bl->in(1); 2296 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) { 2297 return 0; 2298 } 2299 // Must have an invariant operand 2300 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) { 2301 return 0; 2302 } 2303 Node* add2 = NULL; 2304 Node* cmp1 = cmp->in(1); 2305 if (cmp1->is_Phi()) { 2306 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) ))) 2307 Node* phi = cmp1; 2308 for (uint i = 1; i < phi->req(); i++) { 2309 Node* in = phi->in(i); 2310 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in, 2311 &trunc1, &trunc2, &ttype); 2312 if (add && add->in(1) == phi) { 2313 add2 = add->in(2); 2314 break; 2315 } 2316 } 2317 } else { 2318 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) ))) 2319 Node* addtrunc = cmp1; 2320 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc, 2321 &trunc1, &trunc2, &ttype); 2322 if (add && add->in(1)->is_Phi()) { 2323 Node* phi = add->in(1); 2324 for (uint i = 1; i < phi->req(); i++) { 2325 if (phi->in(i) == addtrunc) { 2326 add2 = add->in(2); 2327 break; 2328 } 2329 } 2330 } 2331 } 2332 if (add2 != NULL) { 2333 const TypeInt* add2t = _igvn.type(add2)->is_int(); 2334 if (add2t->is_con()) { 2335 return add2t->get_con(); 2336 } 2337 } 2338 return 0; 2339 } 2340 2341 2342 //---------------------- stay_in_loop ------------------------------------- 2343 // Return the (unique) control output node that's in the loop (if it exists.) 2344 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) { 2345 Node* unique = NULL; 2346 if (!n) return NULL; 2347 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2348 Node* use = n->fast_out(i); 2349 if (!has_ctrl(use) && loop->is_member(get_loop(use))) { 2350 if (unique != NULL) { 2351 return NULL; 2352 } 2353 unique = use; 2354 } 2355 } 2356 return unique; 2357 } 2358 2359 //------------------------------ register_node ------------------------------------- 2360 // Utility to register node "n" with PhaseIdealLoop 2361 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) { 2362 _igvn.register_new_node_with_optimizer(n); 2363 loop->_body.push(n); 2364 if (n->is_CFG()) { 2365 set_loop(n, loop); 2366 set_idom(n, pred, ddepth); 2367 } else { 2368 set_ctrl(n, pred); 2369 } 2370 } 2371 2372 //------------------------------ proj_clone ------------------------------------- 2373 // Utility to create an if-projection 2374 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) { 2375 ProjNode* c = p->clone()->as_Proj(); 2376 c->set_req(0, iff); 2377 return c; 2378 } 2379 2380 //------------------------------ short_circuit_if ------------------------------------- 2381 // Force the iff control output to be the live_proj 2382 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) { 2383 guarantee(live_proj != NULL, "null projection"); 2384 int proj_con = live_proj->_con; 2385 assert(proj_con == 0 || proj_con == 1, "false or true projection"); 2386 Node *con = _igvn.intcon(proj_con); 2387 set_ctrl(con, C->root()); 2388 if (iff) { 2389 iff->set_req(1, con); 2390 } 2391 return con; 2392 } 2393 2394 //------------------------------ insert_if_before_proj ------------------------------------- 2395 // Insert a new if before an if projection (* - new node) 2396 // 2397 // before 2398 // if(test) 2399 // / \ 2400 // v v 2401 // other-proj proj (arg) 2402 // 2403 // after 2404 // if(test) 2405 // / \ 2406 // / v 2407 // | * proj-clone 2408 // v | 2409 // other-proj v 2410 // * new_if(relop(cmp[IU](left,right))) 2411 // / \ 2412 // v v 2413 // * new-proj proj 2414 // (returned) 2415 // 2416 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) { 2417 IfNode* iff = proj->in(0)->as_If(); 2418 IdealLoopTree *loop = get_loop(proj); 2419 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2420 int ddepth = dom_depth(proj); 2421 2422 _igvn.rehash_node_delayed(iff); 2423 _igvn.rehash_node_delayed(proj); 2424 2425 proj->set_req(0, NULL); // temporary disconnect 2426 ProjNode* proj2 = proj_clone(proj, iff); 2427 register_node(proj2, loop, iff, ddepth); 2428 2429 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right); 2430 register_node(cmp, loop, proj2, ddepth); 2431 2432 BoolNode* bol = new BoolNode(cmp, relop); 2433 register_node(bol, loop, proj2, ddepth); 2434 2435 int opcode = iff->Opcode(); 2436 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode"); 2437 IfNode* new_if = (opcode == Op_If) ? new IfNode(proj2, bol, iff->_prob, iff->_fcnt): 2438 new RangeCheckNode(proj2, bol, iff->_prob, iff->_fcnt); 2439 register_node(new_if, loop, proj2, ddepth); 2440 2441 proj->set_req(0, new_if); // reattach 2442 set_idom(proj, new_if, ddepth); 2443 2444 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj(); 2445 guarantee(new_exit != NULL, "null exit node"); 2446 register_node(new_exit, get_loop(other_proj), new_if, ddepth); 2447 2448 return new_exit; 2449 } 2450 2451 //------------------------------ insert_region_before_proj ------------------------------------- 2452 // Insert a region before an if projection (* - new node) 2453 // 2454 // before 2455 // if(test) 2456 // / | 2457 // v | 2458 // proj v 2459 // other-proj 2460 // 2461 // after 2462 // if(test) 2463 // / | 2464 // v | 2465 // * proj-clone v 2466 // | other-proj 2467 // v 2468 // * new-region 2469 // | 2470 // v 2471 // * dum_if 2472 // / \ 2473 // v \ 2474 // * dum-proj v 2475 // proj 2476 // 2477 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) { 2478 IfNode* iff = proj->in(0)->as_If(); 2479 IdealLoopTree *loop = get_loop(proj); 2480 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2481 int ddepth = dom_depth(proj); 2482 2483 _igvn.rehash_node_delayed(iff); 2484 _igvn.rehash_node_delayed(proj); 2485 2486 proj->set_req(0, NULL); // temporary disconnect 2487 ProjNode* proj2 = proj_clone(proj, iff); 2488 register_node(proj2, loop, iff, ddepth); 2489 2490 RegionNode* reg = new RegionNode(2); 2491 reg->set_req(1, proj2); 2492 register_node(reg, loop, iff, ddepth); 2493 2494 IfNode* dum_if = new IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt); 2495 register_node(dum_if, loop, reg, ddepth); 2496 2497 proj->set_req(0, dum_if); // reattach 2498 set_idom(proj, dum_if, ddepth); 2499 2500 ProjNode* dum_proj = proj_clone(other_proj, dum_if); 2501 register_node(dum_proj, loop, dum_if, ddepth); 2502 2503 return reg; 2504 } 2505 2506 //------------------------------ insert_cmpi_loop_exit ------------------------------------- 2507 // Clone a signed compare loop exit from an unsigned compare and 2508 // insert it before the unsigned cmp on the stay-in-loop path. 2509 // All new nodes inserted in the dominator tree between the original 2510 // if and it's projections. The original if test is replaced with 2511 // a constant to force the stay-in-loop path. 2512 // 2513 // This is done to make sure that the original if and it's projections 2514 // still dominate the same set of control nodes, that the ctrl() relation 2515 // from data nodes to them is preserved, and that their loop nesting is 2516 // preserved. 2517 // 2518 // before 2519 // if(i <u limit) unsigned compare loop exit 2520 // / | 2521 // v v 2522 // exit-proj stay-in-loop-proj 2523 // 2524 // after 2525 // if(stay-in-loop-const) original if 2526 // / | 2527 // / v 2528 // / if(i < limit) new signed test 2529 // / / | 2530 // / / v 2531 // / / if(i <u limit) new cloned unsigned test 2532 // / / / | 2533 // v v v | 2534 // region | 2535 // | | 2536 // dum-if | 2537 // / | | 2538 // ether | | 2539 // v v 2540 // exit-proj stay-in-loop-proj 2541 // 2542 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) { 2543 const bool Signed = true; 2544 const bool Unsigned = false; 2545 2546 BoolNode* bol = if_cmpu->in(1)->as_Bool(); 2547 if (bol->_test._test != BoolTest::lt) return NULL; 2548 CmpNode* cmpu = bol->in(1)->as_Cmp(); 2549 if (cmpu->Opcode() != Op_CmpU) return NULL; 2550 int stride = stride_of_possible_iv(if_cmpu); 2551 if (stride == 0) return NULL; 2552 2553 Node* lp_proj = stay_in_loop(if_cmpu, loop); 2554 guarantee(lp_proj != NULL, "null loop node"); 2555 2556 ProjNode* lp_continue = lp_proj->as_Proj(); 2557 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj(); 2558 2559 Node* limit = NULL; 2560 if (stride > 0) { 2561 limit = cmpu->in(2); 2562 } else { 2563 limit = _igvn.makecon(TypeInt::ZERO); 2564 set_ctrl(limit, C->root()); 2565 } 2566 // Create a new region on the exit path 2567 RegionNode* reg = insert_region_before_proj(lp_exit); 2568 guarantee(reg != NULL, "null region node"); 2569 2570 // Clone the if-cmpu-true-false using a signed compare 2571 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge; 2572 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue); 2573 reg->add_req(cmpi_exit); 2574 2575 // Clone the if-cmpu-true-false 2576 BoolTest::mask rel_u = bol->_test._test; 2577 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue); 2578 reg->add_req(cmpu_exit); 2579 2580 // Force original if to stay in loop. 2581 short_circuit_if(if_cmpu, lp_continue); 2582 2583 return cmpi_exit->in(0)->as_If(); 2584 } 2585 2586 //------------------------------ remove_cmpi_loop_exit ------------------------------------- 2587 // Remove a previously inserted signed compare loop exit. 2588 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) { 2589 Node* lp_proj = stay_in_loop(if_cmp, loop); 2590 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI && 2591 stay_in_loop(lp_proj, loop)->is_If() && 2592 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu"); 2593 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO); 2594 set_ctrl(con, C->root()); 2595 if_cmp->set_req(1, con); 2596 } 2597 2598 //------------------------------ scheduled_nodelist ------------------------------------- 2599 // Create a post order schedule of nodes that are in the 2600 // "member" set. The list is returned in "sched". 2601 // The first node in "sched" is the loop head, followed by 2602 // nodes which have no inputs in the "member" set, and then 2603 // followed by the nodes that have an immediate input dependence 2604 // on a node in "sched". 2605 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) { 2606 2607 assert(member.test(loop->_head->_idx), "loop head must be in member set"); 2608 VectorSet visited; 2609 Node_Stack nstack(loop->_body.size()); 2610 2611 Node* n = loop->_head; // top of stack is cached in "n" 2612 uint idx = 0; 2613 visited.set(n->_idx); 2614 2615 // Initially push all with no inputs from within member set 2616 for(uint i = 0; i < loop->_body.size(); i++ ) { 2617 Node *elt = loop->_body.at(i); 2618 if (member.test(elt->_idx)) { 2619 bool found = false; 2620 for (uint j = 0; j < elt->req(); j++) { 2621 Node* def = elt->in(j); 2622 if (def && member.test(def->_idx) && def != elt) { 2623 found = true; 2624 break; 2625 } 2626 } 2627 if (!found && elt != loop->_head) { 2628 nstack.push(n, idx); 2629 n = elt; 2630 assert(!visited.test(n->_idx), "not seen yet"); 2631 visited.set(n->_idx); 2632 } 2633 } 2634 } 2635 2636 // traverse out's that are in the member set 2637 while (true) { 2638 if (idx < n->outcnt()) { 2639 Node* use = n->raw_out(idx); 2640 idx++; 2641 if (!visited.test_set(use->_idx)) { 2642 if (member.test(use->_idx)) { 2643 nstack.push(n, idx); 2644 n = use; 2645 idx = 0; 2646 } 2647 } 2648 } else { 2649 // All outputs processed 2650 sched.push(n); 2651 if (nstack.is_empty()) break; 2652 n = nstack.node(); 2653 idx = nstack.index(); 2654 nstack.pop(); 2655 } 2656 } 2657 } 2658 2659 2660 //------------------------------ has_use_in_set ------------------------------------- 2661 // Has a use in the vector set 2662 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) { 2663 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2664 Node* use = n->fast_out(j); 2665 if (vset.test(use->_idx)) { 2666 return true; 2667 } 2668 } 2669 return false; 2670 } 2671 2672 2673 //------------------------------ has_use_internal_to_set ------------------------------------- 2674 // Has use internal to the vector set (ie. not in a phi at the loop head) 2675 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) { 2676 Node* head = loop->_head; 2677 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2678 Node* use = n->fast_out(j); 2679 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) { 2680 return true; 2681 } 2682 } 2683 return false; 2684 } 2685 2686 2687 //------------------------------ clone_for_use_outside_loop ------------------------------------- 2688 // clone "n" for uses that are outside of loop 2689 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) { 2690 int cloned = 0; 2691 assert(worklist.size() == 0, "should be empty"); 2692 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2693 Node* use = n->fast_out(j); 2694 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) { 2695 worklist.push(use); 2696 } 2697 } 2698 while( worklist.size() ) { 2699 Node *use = worklist.pop(); 2700 if (!has_node(use) || use->in(0) == C->top()) continue; 2701 uint j; 2702 for (j = 0; j < use->req(); j++) { 2703 if (use->in(j) == n) break; 2704 } 2705 assert(j < use->req(), "must be there"); 2706 2707 // clone "n" and insert it between the inputs of "n" and the use outside the loop 2708 Node* n_clone = n->clone(); 2709 _igvn.replace_input_of(use, j, n_clone); 2710 cloned++; 2711 Node* use_c; 2712 if (!use->is_Phi()) { 2713 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0); 2714 } else { 2715 // Use in a phi is considered a use in the associated predecessor block 2716 use_c = use->in(0)->in(j); 2717 } 2718 set_ctrl(n_clone, use_c); 2719 assert(!loop->is_member(get_loop(use_c)), "should be outside loop"); 2720 get_loop(use_c)->_body.push(n_clone); 2721 _igvn.register_new_node_with_optimizer(n_clone); 2722 #ifndef PRODUCT 2723 if (TracePartialPeeling) { 2724 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx); 2725 } 2726 #endif 2727 } 2728 return cloned; 2729 } 2730 2731 2732 //------------------------------ clone_for_special_use_inside_loop ------------------------------------- 2733 // clone "n" for special uses that are in the not_peeled region. 2734 // If these def-uses occur in separate blocks, the code generator 2735 // marks the method as not compilable. For example, if a "BoolNode" 2736 // is in a different basic block than the "IfNode" that uses it, then 2737 // the compilation is aborted in the code generator. 2738 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 2739 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) { 2740 if (n->is_Phi() || n->is_Load()) { 2741 return; 2742 } 2743 assert(worklist.size() == 0, "should be empty"); 2744 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2745 Node* use = n->fast_out(j); 2746 if ( not_peel.test(use->_idx) && 2747 (use->is_If() || use->is_CMove() || use->is_Bool()) && 2748 use->in(1) == n) { 2749 worklist.push(use); 2750 } 2751 } 2752 if (worklist.size() > 0) { 2753 // clone "n" and insert it between inputs of "n" and the use 2754 Node* n_clone = n->clone(); 2755 loop->_body.push(n_clone); 2756 _igvn.register_new_node_with_optimizer(n_clone); 2757 set_ctrl(n_clone, get_ctrl(n)); 2758 sink_list.push(n_clone); 2759 not_peel.set(n_clone->_idx); 2760 #ifndef PRODUCT 2761 if (TracePartialPeeling) { 2762 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx); 2763 } 2764 #endif 2765 while( worklist.size() ) { 2766 Node *use = worklist.pop(); 2767 _igvn.rehash_node_delayed(use); 2768 for (uint j = 1; j < use->req(); j++) { 2769 if (use->in(j) == n) { 2770 use->set_req(j, n_clone); 2771 } 2772 } 2773 } 2774 } 2775 } 2776 2777 2778 //------------------------------ insert_phi_for_loop ------------------------------------- 2779 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 2780 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) { 2781 Node *phi = PhiNode::make(lp, back_edge_val); 2782 phi->set_req(LoopNode::EntryControl, lp_entry_val); 2783 // Use existing phi if it already exists 2784 Node *hit = _igvn.hash_find_insert(phi); 2785 if( hit == NULL ) { 2786 _igvn.register_new_node_with_optimizer(phi); 2787 set_ctrl(phi, lp); 2788 } else { 2789 // Remove the new phi from the graph and use the hit 2790 _igvn.remove_dead_node(phi); 2791 phi = hit; 2792 } 2793 _igvn.replace_input_of(use, idx, phi); 2794 } 2795 2796 #ifdef ASSERT 2797 //------------------------------ is_valid_loop_partition ------------------------------------- 2798 // Validate the loop partition sets: peel and not_peel 2799 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, 2800 VectorSet& not_peel ) { 2801 uint i; 2802 // Check that peel_list entries are in the peel set 2803 for (i = 0; i < peel_list.size(); i++) { 2804 if (!peel.test(peel_list.at(i)->_idx)) { 2805 return false; 2806 } 2807 } 2808 // Check at loop members are in one of peel set or not_peel set 2809 for (i = 0; i < loop->_body.size(); i++ ) { 2810 Node *def = loop->_body.at(i); 2811 uint di = def->_idx; 2812 // Check that peel set elements are in peel_list 2813 if (peel.test(di)) { 2814 if (not_peel.test(di)) { 2815 return false; 2816 } 2817 // Must be in peel_list also 2818 bool found = false; 2819 for (uint j = 0; j < peel_list.size(); j++) { 2820 if (peel_list.at(j)->_idx == di) { 2821 found = true; 2822 break; 2823 } 2824 } 2825 if (!found) { 2826 return false; 2827 } 2828 } else if (not_peel.test(di)) { 2829 if (peel.test(di)) { 2830 return false; 2831 } 2832 } else { 2833 return false; 2834 } 2835 } 2836 return true; 2837 } 2838 2839 //------------------------------ is_valid_clone_loop_exit_use ------------------------------------- 2840 // Ensure a use outside of loop is of the right form 2841 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) { 2842 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2843 return (use->is_Phi() && 2844 use_c->is_Region() && use_c->req() == 3 && 2845 (use_c->in(exit_idx)->Opcode() == Op_IfTrue || 2846 use_c->in(exit_idx)->Opcode() == Op_IfFalse || 2847 use_c->in(exit_idx)->Opcode() == Op_JumpProj) && 2848 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) ); 2849 } 2850 2851 //------------------------------ is_valid_clone_loop_form ------------------------------------- 2852 // Ensure that all uses outside of loop are of the right form 2853 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 2854 uint orig_exit_idx, uint clone_exit_idx) { 2855 uint len = peel_list.size(); 2856 for (uint i = 0; i < len; i++) { 2857 Node *def = peel_list.at(i); 2858 2859 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 2860 Node *use = def->fast_out(j); 2861 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2862 if (!loop->is_member(get_loop(use_c))) { 2863 // use is not in the loop, check for correct structure 2864 if (use->in(0) == def) { 2865 // Okay 2866 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) { 2867 return false; 2868 } 2869 } 2870 } 2871 } 2872 return true; 2873 } 2874 #endif 2875 2876 //------------------------------ partial_peel ------------------------------------- 2877 // Partially peel (aka loop rotation) the top portion of a loop (called 2878 // the peel section below) by cloning it and placing one copy just before 2879 // the new loop head and the other copy at the bottom of the new loop. 2880 // 2881 // before after where it came from 2882 // 2883 // stmt1 stmt1 2884 // loop: stmt2 clone 2885 // stmt2 if condA goto exitA clone 2886 // if condA goto exitA new_loop: new 2887 // stmt3 stmt3 clone 2888 // if !condB goto loop if condB goto exitB clone 2889 // exitB: stmt2 orig 2890 // stmt4 if !condA goto new_loop orig 2891 // exitA: goto exitA 2892 // exitB: 2893 // stmt4 2894 // exitA: 2895 // 2896 // Step 1: find the cut point: an exit test on probable 2897 // induction variable. 2898 // Step 2: schedule (with cloning) operations in the peel 2899 // section that can be executed after the cut into 2900 // the section that is not peeled. This may need 2901 // to clone operations into exit blocks. For 2902 // instance, a reference to A[i] in the not-peel 2903 // section and a reference to B[i] in an exit block 2904 // may cause a left-shift of i by 2 to be placed 2905 // in the peel block. This step will clone the left 2906 // shift into the exit block and sink the left shift 2907 // from the peel to the not-peel section. 2908 // Step 3: clone the loop, retarget the control, and insert 2909 // phis for values that are live across the new loop 2910 // head. This is very dependent on the graph structure 2911 // from clone_loop. It creates region nodes for 2912 // exit control and associated phi nodes for values 2913 // flow out of the loop through that exit. The region 2914 // node is dominated by the clone's control projection. 2915 // So the clone's peel section is placed before the 2916 // new loop head, and the clone's not-peel section is 2917 // forms the top part of the new loop. The original 2918 // peel section forms the tail of the new loop. 2919 // Step 4: update the dominator tree and recompute the 2920 // dominator depth. 2921 // 2922 // orig 2923 // 2924 // stmt1 2925 // | 2926 // v 2927 // loop predicate 2928 // | 2929 // v 2930 // loop<----+ 2931 // | | 2932 // stmt2 | 2933 // | | 2934 // v | 2935 // ifA | 2936 // / | | 2937 // v v | 2938 // false true ^ <-- last_peel 2939 // / | | 2940 // / ===|==cut | 2941 // / stmt3 | <-- first_not_peel 2942 // / | | 2943 // | v | 2944 // v ifB | 2945 // exitA: / \ | 2946 // / \ | 2947 // v v | 2948 // false true | 2949 // / \ | 2950 // / ----+ 2951 // | 2952 // v 2953 // exitB: 2954 // stmt4 2955 // 2956 // 2957 // after clone loop 2958 // 2959 // stmt1 2960 // | 2961 // v 2962 // loop predicate 2963 // / \ 2964 // clone / \ orig 2965 // / \ 2966 // / \ 2967 // v v 2968 // +---->loop loop<----+ 2969 // | | | | 2970 // | stmt2 stmt2 | 2971 // | | | | 2972 // | v v | 2973 // | ifA ifA | 2974 // | | \ / | | 2975 // | v v v v | 2976 // ^ true false false true ^ <-- last_peel 2977 // | | ^ \ / | | 2978 // | cut==|== \ \ / ===|==cut | 2979 // | stmt3 \ \ / stmt3 | <-- first_not_peel 2980 // | | dom | | | | 2981 // | v \ 1v v2 v | 2982 // | ifB regionA ifB | 2983 // | / \ | / \ | 2984 // | / \ v / \ | 2985 // | v v exitA: v v | 2986 // | true false false true | 2987 // | / ^ \ / \ | 2988 // +---- \ \ / ----+ 2989 // dom \ / 2990 // \ 1v v2 2991 // regionB 2992 // | 2993 // v 2994 // exitB: 2995 // stmt4 2996 // 2997 // 2998 // after partial peel 2999 // 3000 // stmt1 3001 // | 3002 // v 3003 // loop predicate 3004 // / 3005 // clone / orig 3006 // / TOP 3007 // / \ 3008 // v v 3009 // TOP->loop loop----+ 3010 // | | | 3011 // stmt2 stmt2 | 3012 // | | | 3013 // v v | 3014 // ifA ifA | 3015 // | \ / | | 3016 // v v v v | 3017 // true false false true | <-- last_peel 3018 // | ^ \ / +------|---+ 3019 // +->newloop \ \ / === ==cut | | 3020 // | stmt3 \ \ / TOP | | 3021 // | | dom | | stmt3 | | <-- first_not_peel 3022 // | v \ 1v v2 v | | 3023 // | ifB regionA ifB ^ v 3024 // | / \ | / \ | | 3025 // | / \ v / \ | | 3026 // | v v exitA: v v | | 3027 // | true false false true | | 3028 // | / ^ \ / \ | | 3029 // | | \ \ / v | | 3030 // | | dom \ / TOP | | 3031 // | | \ 1v v2 | | 3032 // ^ v regionB | | 3033 // | | | | | 3034 // | | v ^ v 3035 // | | exitB: | | 3036 // | | stmt4 | | 3037 // | +------------>-----------------+ | 3038 // | | 3039 // +-----------------<---------------------+ 3040 // 3041 // 3042 // final graph 3043 // 3044 // stmt1 3045 // | 3046 // v 3047 // loop predicate 3048 // | 3049 // v 3050 // stmt2 clone 3051 // | 3052 // v 3053 // ........> ifA clone 3054 // : / | 3055 // dom / | 3056 // : v v 3057 // : false true 3058 // : | | 3059 // : | v 3060 // : | newloop<-----+ 3061 // : | | | 3062 // : | stmt3 clone | 3063 // : | | | 3064 // : | v | 3065 // : | ifB | 3066 // : | / \ | 3067 // : | v v | 3068 // : | false true | 3069 // : | | | | 3070 // : | v stmt2 | 3071 // : | exitB: | | 3072 // : | stmt4 v | 3073 // : | ifA orig | 3074 // : | / \ | 3075 // : | / \ | 3076 // : | v v | 3077 // : | false true | 3078 // : | / \ | 3079 // : v v -----+ 3080 // RegionA 3081 // | 3082 // v 3083 // exitA 3084 // 3085 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) { 3086 3087 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only"); 3088 if (!loop->_head->is_Loop()) { 3089 return false; 3090 } 3091 LoopNode *head = loop->_head->as_Loop(); 3092 3093 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) { 3094 return false; 3095 } 3096 3097 // Check for complex exit control 3098 for (uint ii = 0; ii < loop->_body.size(); ii++) { 3099 Node *n = loop->_body.at(ii); 3100 int opc = n->Opcode(); 3101 if (n->is_Call() || 3102 opc == Op_Catch || 3103 opc == Op_CatchProj || 3104 opc == Op_Jump || 3105 opc == Op_JumpProj) { 3106 #ifndef PRODUCT 3107 if (TracePartialPeeling) { 3108 tty->print_cr("\nExit control too complex: lp: %d", head->_idx); 3109 } 3110 #endif 3111 return false; 3112 } 3113 } 3114 3115 int dd = dom_depth(head); 3116 3117 // Step 1: find cut point 3118 3119 // Walk up dominators to loop head looking for first loop exit 3120 // which is executed on every path thru loop. 3121 IfNode *peel_if = NULL; 3122 IfNode *peel_if_cmpu = NULL; 3123 3124 Node *iff = loop->tail(); 3125 while (iff != head) { 3126 if (iff->is_If()) { 3127 Node *ctrl = get_ctrl(iff->in(1)); 3128 if (ctrl->is_top()) return false; // Dead test on live IF. 3129 // If loop-varying exit-test, check for induction variable 3130 if (loop->is_member(get_loop(ctrl)) && 3131 loop->is_loop_exit(iff) && 3132 is_possible_iv_test(iff)) { 3133 Node* cmp = iff->in(1)->in(1); 3134 if (cmp->Opcode() == Op_CmpI) { 3135 peel_if = iff->as_If(); 3136 } else { 3137 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU"); 3138 peel_if_cmpu = iff->as_If(); 3139 } 3140 } 3141 } 3142 iff = idom(iff); 3143 } 3144 3145 // Prefer signed compare over unsigned compare. 3146 IfNode* new_peel_if = NULL; 3147 if (peel_if == NULL) { 3148 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) { 3149 return false; // No peel point found 3150 } 3151 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop); 3152 if (new_peel_if == NULL) { 3153 return false; // No peel point found 3154 } 3155 peel_if = new_peel_if; 3156 } 3157 Node* last_peel = stay_in_loop(peel_if, loop); 3158 Node* first_not_peeled = stay_in_loop(last_peel, loop); 3159 if (first_not_peeled == NULL || first_not_peeled == head) { 3160 return false; 3161 } 3162 3163 #ifndef PRODUCT 3164 if (TraceLoopOpts) { 3165 tty->print("PartialPeel "); 3166 loop->dump_head(); 3167 } 3168 3169 if (TracePartialPeeling) { 3170 tty->print_cr("before partial peel one iteration"); 3171 Node_List wl; 3172 Node* t = head->in(2); 3173 while (true) { 3174 wl.push(t); 3175 if (t == head) break; 3176 t = idom(t); 3177 } 3178 while (wl.size() > 0) { 3179 Node* tt = wl.pop(); 3180 tt->dump(); 3181 if (tt == last_peel) tty->print_cr("-- cut --"); 3182 } 3183 } 3184 #endif 3185 VectorSet peel; 3186 VectorSet not_peel; 3187 Node_List peel_list; 3188 Node_List worklist; 3189 Node_List sink_list; 3190 3191 uint estimate = loop->est_loop_clone_sz(1); 3192 if (exceeding_node_budget(estimate)) { 3193 return false; 3194 } 3195 3196 // Set of cfg nodes to peel are those that are executable from 3197 // the head through last_peel. 3198 assert(worklist.size() == 0, "should be empty"); 3199 worklist.push(head); 3200 peel.set(head->_idx); 3201 while (worklist.size() > 0) { 3202 Node *n = worklist.pop(); 3203 if (n != last_peel) { 3204 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3205 Node* use = n->fast_out(j); 3206 if (use->is_CFG() && 3207 loop->is_member(get_loop(use)) && 3208 !peel.test_set(use->_idx)) { 3209 worklist.push(use); 3210 } 3211 } 3212 } 3213 } 3214 3215 // Set of non-cfg nodes to peel are those that are control 3216 // dependent on the cfg nodes. 3217 for (uint i = 0; i < loop->_body.size(); i++) { 3218 Node *n = loop->_body.at(i); 3219 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n; 3220 if (peel.test(n_c->_idx)) { 3221 peel.set(n->_idx); 3222 } else { 3223 not_peel.set(n->_idx); 3224 } 3225 } 3226 3227 // Step 2: move operations from the peeled section down into the 3228 // not-peeled section 3229 3230 // Get a post order schedule of nodes in the peel region 3231 // Result in right-most operand. 3232 scheduled_nodelist(loop, peel, peel_list); 3233 3234 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 3235 3236 // For future check for too many new phis 3237 uint old_phi_cnt = 0; 3238 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 3239 Node* use = head->fast_out(j); 3240 if (use->is_Phi()) old_phi_cnt++; 3241 } 3242 3243 #ifndef PRODUCT 3244 if (TracePartialPeeling) { 3245 tty->print_cr("\npeeled list"); 3246 } 3247 #endif 3248 3249 // Evacuate nodes in peel region into the not_peeled region if possible 3250 uint new_phi_cnt = 0; 3251 uint cloned_for_outside_use = 0; 3252 for (uint i = 0; i < peel_list.size();) { 3253 Node* n = peel_list.at(i); 3254 #ifndef PRODUCT 3255 if (TracePartialPeeling) n->dump(); 3256 #endif 3257 bool incr = true; 3258 if (!n->is_CFG()) { 3259 if (has_use_in_set(n, not_peel)) { 3260 // If not used internal to the peeled region, 3261 // move "n" from peeled to not_peeled region. 3262 if (!has_use_internal_to_set(n, peel, loop)) { 3263 // if not pinned and not a load (which maybe anti-dependent on a store) 3264 // and not a CMove (Matcher expects only bool->cmove). 3265 if (n->in(0) == NULL && !n->is_Load() && !n->is_CMove()) { 3266 cloned_for_outside_use += clone_for_use_outside_loop(loop, n, worklist); 3267 sink_list.push(n); 3268 peel.remove(n->_idx); 3269 not_peel.set(n->_idx); 3270 peel_list.remove(i); 3271 incr = false; 3272 #ifndef PRODUCT 3273 if (TracePartialPeeling) { 3274 tty->print_cr("sink to not_peeled region: %d newbb: %d", 3275 n->_idx, get_ctrl(n)->_idx); 3276 } 3277 #endif 3278 } 3279 } else { 3280 // Otherwise check for special def-use cases that span 3281 // the peel/not_peel boundary such as bool->if 3282 clone_for_special_use_inside_loop(loop, n, not_peel, sink_list, worklist); 3283 new_phi_cnt++; 3284 } 3285 } 3286 } 3287 if (incr) i++; 3288 } 3289 3290 estimate += cloned_for_outside_use + new_phi_cnt; 3291 bool exceed_node_budget = !may_require_nodes(estimate); 3292 bool exceed_phi_limit = new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta; 3293 3294 if (exceed_node_budget || exceed_phi_limit) { 3295 #ifndef PRODUCT 3296 if (TracePartialPeeling) { 3297 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c", 3298 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F'); 3299 } 3300 #endif 3301 if (new_peel_if != NULL) { 3302 remove_cmpi_loop_exit(new_peel_if, loop); 3303 } 3304 // Inhibit more partial peeling on this loop 3305 assert(!head->is_partial_peel_loop(), "not partial peeled"); 3306 head->mark_partial_peel_failed(); 3307 if (cloned_for_outside_use > 0) { 3308 // Terminate this round of loop opts because 3309 // the graph outside this loop was changed. 3310 C->set_major_progress(); 3311 return true; 3312 } 3313 return false; 3314 } 3315 3316 // Step 3: clone loop, retarget control, and insert new phis 3317 3318 // Create new loop head for new phis and to hang 3319 // the nodes being moved (sinked) from the peel region. 3320 LoopNode* new_head = new LoopNode(last_peel, last_peel); 3321 new_head->set_unswitch_count(head->unswitch_count()); // Preserve 3322 _igvn.register_new_node_with_optimizer(new_head); 3323 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled"); 3324 _igvn.replace_input_of(first_not_peeled, 0, new_head); 3325 set_loop(new_head, loop); 3326 loop->_body.push(new_head); 3327 not_peel.set(new_head->_idx); 3328 set_idom(new_head, last_peel, dom_depth(first_not_peeled)); 3329 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled)); 3330 3331 while (sink_list.size() > 0) { 3332 Node* n = sink_list.pop(); 3333 set_ctrl(n, new_head); 3334 } 3335 3336 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 3337 3338 clone_loop(loop, old_new, dd, IgnoreStripMined); 3339 3340 const uint clone_exit_idx = 1; 3341 const uint orig_exit_idx = 2; 3342 assert(is_valid_clone_loop_form(loop, peel_list, orig_exit_idx, clone_exit_idx), "bad clone loop"); 3343 3344 Node* head_clone = old_new[head->_idx]; 3345 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop(); 3346 Node* orig_tail_clone = head_clone->in(2); 3347 3348 // Add phi if "def" node is in peel set and "use" is not 3349 3350 for (uint i = 0; i < peel_list.size(); i++) { 3351 Node *def = peel_list.at(i); 3352 if (!def->is_CFG()) { 3353 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 3354 Node *use = def->fast_out(j); 3355 if (has_node(use) && use->in(0) != C->top() && 3356 (!peel.test(use->_idx) || 3357 (use->is_Phi() && use->in(0) == head)) ) { 3358 worklist.push(use); 3359 } 3360 } 3361 while( worklist.size() ) { 3362 Node *use = worklist.pop(); 3363 for (uint j = 1; j < use->req(); j++) { 3364 Node* n = use->in(j); 3365 if (n == def) { 3366 3367 // "def" is in peel set, "use" is not in peel set 3368 // or "use" is in the entry boundary (a phi) of the peel set 3369 3370 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use; 3371 3372 if ( loop->is_member(get_loop( use_c )) ) { 3373 // use is in loop 3374 if (old_new[use->_idx] != NULL) { // null for dead code 3375 Node* use_clone = old_new[use->_idx]; 3376 _igvn.replace_input_of(use, j, C->top()); 3377 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone ); 3378 } 3379 } else { 3380 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format"); 3381 // use is not in the loop, check if the live range includes the cut 3382 Node* lp_if = use_c->in(orig_exit_idx)->in(0); 3383 if (not_peel.test(lp_if->_idx)) { 3384 assert(j == orig_exit_idx, "use from original loop"); 3385 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone ); 3386 } 3387 } 3388 } 3389 } 3390 } 3391 } 3392 } 3393 3394 // Step 3b: retarget control 3395 3396 // Redirect control to the new loop head if a cloned node in 3397 // the not_peeled region has control that points into the peeled region. 3398 // This necessary because the cloned peeled region will be outside 3399 // the loop. 3400 // from to 3401 // cloned-peeled <---+ 3402 // new_head_clone: | <--+ 3403 // cloned-not_peeled in(0) in(0) 3404 // orig-peeled 3405 3406 for (uint i = 0; i < loop->_body.size(); i++) { 3407 Node *n = loop->_body.at(i); 3408 if (!n->is_CFG() && n->in(0) != NULL && 3409 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) { 3410 Node* n_clone = old_new[n->_idx]; 3411 _igvn.replace_input_of(n_clone, 0, new_head_clone); 3412 } 3413 } 3414 3415 // Backedge of the surviving new_head (the clone) is original last_peel 3416 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel); 3417 3418 // Cut first node in original not_peel set 3419 _igvn.rehash_node_delayed(new_head); // Multiple edge updates: 3420 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of 3421 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls 3422 3423 // Copy head_clone back-branch info to original head 3424 // and remove original head's loop entry and 3425 // clone head's back-branch 3426 _igvn.rehash_node_delayed(head); // Multiple edge updates 3427 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl)); 3428 head->set_req(LoopNode::LoopBackControl, C->top()); 3429 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top()); 3430 3431 // Similarly modify the phis 3432 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) { 3433 Node* use = head->fast_out(k); 3434 if (use->is_Phi() && use->outcnt() > 0) { 3435 Node* use_clone = old_new[use->_idx]; 3436 _igvn.rehash_node_delayed(use); // Multiple edge updates 3437 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl)); 3438 use->set_req(LoopNode::LoopBackControl, C->top()); 3439 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top()); 3440 } 3441 } 3442 3443 // Step 4: update dominator tree and dominator depth 3444 3445 set_idom(head, orig_tail_clone, dd); 3446 recompute_dom_depth(); 3447 3448 // Inhibit more partial peeling on this loop 3449 new_head_clone->set_partial_peel_loop(); 3450 C->set_major_progress(); 3451 loop->record_for_igvn(); 3452 3453 #ifndef PRODUCT 3454 if (TracePartialPeeling) { 3455 tty->print_cr("\nafter partial peel one iteration"); 3456 Node_List wl; 3457 Node* t = last_peel; 3458 while (true) { 3459 wl.push(t); 3460 if (t == head_clone) break; 3461 t = idom(t); 3462 } 3463 while (wl.size() > 0) { 3464 Node* tt = wl.pop(); 3465 if (tt == head) tty->print_cr("orig head"); 3466 else if (tt == new_head_clone) tty->print_cr("new head"); 3467 else if (tt == head_clone) tty->print_cr("clone head"); 3468 tt->dump(); 3469 } 3470 } 3471 #endif 3472 return true; 3473 } 3474 3475 //------------------------------reorg_offsets---------------------------------- 3476 // Reorganize offset computations to lower register pressure. Mostly 3477 // prevent loop-fallout uses of the pre-incremented trip counter (which are 3478 // then alive with the post-incremented trip counter forcing an extra 3479 // register move) 3480 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) { 3481 // Perform it only for canonical counted loops. 3482 // Loop's shape could be messed up by iteration_split_impl. 3483 if (!loop->_head->is_CountedLoop()) 3484 return; 3485 if (!loop->_head->as_Loop()->is_valid_counted_loop()) 3486 return; 3487 3488 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 3489 CountedLoopEndNode *cle = cl->loopexit(); 3490 Node *exit = cle->proj_out(false); 3491 Node *phi = cl->phi(); 3492 3493 // Check for the special case when using the pre-incremented trip-counter on 3494 // the fall-out path (forces the pre-incremented and post-incremented trip 3495 // counter to be live at the same time). Fix this by adjusting to use the 3496 // post-increment trip counter. 3497 3498 bool progress = true; 3499 while (progress) { 3500 progress = false; 3501 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) { 3502 Node* use = phi->fast_out(i); // User of trip-counter 3503 if (!has_ctrl(use)) continue; 3504 Node *u_ctrl = get_ctrl(use); 3505 if (use->is_Phi()) { 3506 u_ctrl = NULL; 3507 for (uint j = 1; j < use->req(); j++) 3508 if (use->in(j) == phi) 3509 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j)); 3510 } 3511 IdealLoopTree *u_loop = get_loop(u_ctrl); 3512 // Look for loop-invariant use 3513 if (u_loop == loop) continue; 3514 if (loop->is_member(u_loop)) continue; 3515 // Check that use is live out the bottom. Assuming the trip-counter 3516 // update is right at the bottom, uses of of the loop middle are ok. 3517 if (dom_lca(exit, u_ctrl) != exit) continue; 3518 // Hit! Refactor use to use the post-incremented tripcounter. 3519 // Compute a post-increment tripcounter. 3520 Node* c = exit; 3521 if (cl->is_strip_mined()) { 3522 IdealLoopTree* outer_loop = get_loop(cl->outer_loop()); 3523 if (!outer_loop->is_member(u_loop)) { 3524 c = cl->outer_loop_exit(); 3525 } 3526 } 3527 Node *opaq = new Opaque2Node(C, cle->incr()); 3528 register_new_node(opaq, c); 3529 Node *neg_stride = _igvn.intcon(-cle->stride_con()); 3530 set_ctrl(neg_stride, C->root()); 3531 Node *post = new AddINode(opaq, neg_stride); 3532 register_new_node(post, c); 3533 _igvn.rehash_node_delayed(use); 3534 for (uint j = 1; j < use->req(); j++) { 3535 if (use->in(j) == phi) 3536 use->set_req(j, post); 3537 } 3538 // Since DU info changed, rerun loop 3539 progress = true; 3540 break; 3541 } 3542 } 3543 3544 }