1 /* 2 * Copyright (c) 1999, 2018, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "gc/shared/barrierSet.hpp" 27 #include "gc/shared/c2/barrierSetC2.hpp" 28 #include "memory/allocation.inline.hpp" 29 #include "memory/resourceArea.hpp" 30 #include "opto/addnode.hpp" 31 #include "opto/callnode.hpp" 32 #include "opto/castnode.hpp" 33 #include "opto/connode.hpp" 34 #include "opto/castnode.hpp" 35 #include "opto/divnode.hpp" 36 #include "opto/loopnode.hpp" 37 #include "opto/matcher.hpp" 38 #include "opto/mulnode.hpp" 39 #include "opto/movenode.hpp" 40 #include "opto/opaquenode.hpp" 41 #include "opto/rootnode.hpp" 42 #include "opto/subnode.hpp" 43 #include "utilities/macros.hpp" 44 #if INCLUDE_ZGC 45 #include "gc/z/c2/zBarrierSetC2.hpp" 46 #endif 47 48 //============================================================================= 49 //------------------------------split_thru_phi--------------------------------- 50 // Split Node 'n' through merge point if there is enough win. 51 Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) { 52 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) { 53 // ConvI2L may have type information on it which is unsafe to push up 54 // so disable this for now 55 return NULL; 56 } 57 58 // Splitting range check CastIIs through a loop induction Phi can 59 // cause new Phis to be created that are left unrelated to the loop 60 // induction Phi and prevent optimizations (vectorization) 61 if (n->Opcode() == Op_CastII && n->as_CastII()->has_range_check() && 62 region->is_CountedLoop() && n->in(1) == region->as_CountedLoop()->phi()) { 63 return NULL; 64 } 65 66 int wins = 0; 67 assert(!n->is_CFG(), ""); 68 assert(region->is_Region(), ""); 69 70 const Type* type = n->bottom_type(); 71 const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr(); 72 Node *phi; 73 if (t_oop != NULL && t_oop->is_known_instance_field()) { 74 int iid = t_oop->instance_id(); 75 int index = C->get_alias_index(t_oop); 76 int offset = t_oop->offset(); 77 phi = new PhiNode(region, type, NULL, iid, index, offset); 78 } else { 79 phi = PhiNode::make_blank(region, n); 80 } 81 uint old_unique = C->unique(); 82 for (uint i = 1; i < region->req(); i++) { 83 Node *x; 84 Node* the_clone = NULL; 85 if (region->in(i) == C->top()) { 86 x = C->top(); // Dead path? Use a dead data op 87 } else { 88 x = n->clone(); // Else clone up the data op 89 the_clone = x; // Remember for possible deletion. 90 // Alter data node to use pre-phi inputs 91 if (n->in(0) == region) 92 x->set_req( 0, region->in(i) ); 93 for (uint j = 1; j < n->req(); j++) { 94 Node *in = n->in(j); 95 if (in->is_Phi() && in->in(0) == region) 96 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone 97 } 98 } 99 // Check for a 'win' on some paths 100 const Type *t = x->Value(&_igvn); 101 102 bool singleton = t->singleton(); 103 104 // A TOP singleton indicates that there are no possible values incoming 105 // along a particular edge. In most cases, this is OK, and the Phi will 106 // be eliminated later in an Ideal call. However, we can't allow this to 107 // happen if the singleton occurs on loop entry, as the elimination of 108 // the PhiNode may cause the resulting node to migrate back to a previous 109 // loop iteration. 110 if (singleton && t == Type::TOP) { 111 // Is_Loop() == false does not confirm the absence of a loop (e.g., an 112 // irreducible loop may not be indicated by an affirmative is_Loop()); 113 // therefore, the only top we can split thru a phi is on a backedge of 114 // a loop. 115 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 116 } 117 118 if (singleton) { 119 wins++; 120 x = ((PhaseGVN&)_igvn).makecon(t); 121 } else { 122 // We now call Identity to try to simplify the cloned node. 123 // Note that some Identity methods call phase->type(this). 124 // Make sure that the type array is big enough for 125 // our new node, even though we may throw the node away. 126 // (Note: This tweaking with igvn only works because x is a new node.) 127 _igvn.set_type(x, t); 128 // If x is a TypeNode, capture any more-precise type permanently into Node 129 // otherwise it will be not updated during igvn->transform since 130 // igvn->type(x) is set to x->Value() already. 131 x->raise_bottom_type(t); 132 Node *y = _igvn.apply_identity(x); 133 if (y != x) { 134 wins++; 135 x = y; 136 } else { 137 y = _igvn.hash_find(x); 138 if (y) { 139 wins++; 140 x = y; 141 } else { 142 // Else x is a new node we are keeping 143 // We do not need register_new_node_with_optimizer 144 // because set_type has already been called. 145 _igvn._worklist.push(x); 146 } 147 } 148 } 149 if (x != the_clone && the_clone != NULL) 150 _igvn.remove_dead_node(the_clone); 151 phi->set_req( i, x ); 152 } 153 // Too few wins? 154 if (wins <= policy) { 155 _igvn.remove_dead_node(phi); 156 return NULL; 157 } 158 159 // Record Phi 160 register_new_node( phi, region ); 161 162 for (uint i2 = 1; i2 < phi->req(); i2++) { 163 Node *x = phi->in(i2); 164 // If we commoned up the cloned 'x' with another existing Node, 165 // the existing Node picks up a new use. We need to make the 166 // existing Node occur higher up so it dominates its uses. 167 Node *old_ctrl; 168 IdealLoopTree *old_loop; 169 170 if (x->is_Con()) { 171 // Constant's control is always root. 172 set_ctrl(x, C->root()); 173 continue; 174 } 175 // The occasional new node 176 if (x->_idx >= old_unique) { // Found a new, unplaced node? 177 old_ctrl = NULL; 178 old_loop = NULL; // Not in any prior loop 179 } else { 180 old_ctrl = get_ctrl(x); 181 old_loop = get_loop(old_ctrl); // Get prior loop 182 } 183 // New late point must dominate new use 184 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2)); 185 if (new_ctrl == old_ctrl) // Nothing is changed 186 continue; 187 188 IdealLoopTree *new_loop = get_loop(new_ctrl); 189 190 // Don't move x into a loop if its uses are 191 // outside of loop. Otherwise x will be cloned 192 // for each use outside of this loop. 193 IdealLoopTree *use_loop = get_loop(region); 194 if (!new_loop->is_member(use_loop) && 195 (old_loop == NULL || !new_loop->is_member(old_loop))) { 196 // Take early control, later control will be recalculated 197 // during next iteration of loop optimizations. 198 new_ctrl = get_early_ctrl(x); 199 new_loop = get_loop(new_ctrl); 200 } 201 // Set new location 202 set_ctrl(x, new_ctrl); 203 // If changing loop bodies, see if we need to collect into new body 204 if (old_loop != new_loop) { 205 if (old_loop && !old_loop->_child) 206 old_loop->_body.yank(x); 207 if (!new_loop->_child) 208 new_loop->_body.push(x); // Collect body info 209 } 210 } 211 212 return phi; 213 } 214 215 //------------------------------dominated_by------------------------------------ 216 // Replace the dominated test with an obvious true or false. Place it on the 217 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the 218 // live path up to the dominating control. 219 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff, bool flip, bool exclude_loop_predicate ) { 220 if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); } 221 222 // prevdom is the dominating projection of the dominating test. 223 assert( iff->is_If(), "" ); 224 assert(iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd || iff->Opcode() == Op_RangeCheck, "Check this code when new subtype is added"); 225 int pop = prevdom->Opcode(); 226 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" ); 227 if (flip) { 228 if (pop == Op_IfTrue) 229 pop = Op_IfFalse; 230 else 231 pop = Op_IfTrue; 232 } 233 // 'con' is set to true or false to kill the dominated test. 234 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO); 235 set_ctrl(con, C->root()); // Constant gets a new use 236 // Hack the dominated test 237 _igvn.replace_input_of(iff, 1, con); 238 239 // If I dont have a reachable TRUE and FALSE path following the IfNode then 240 // I can assume this path reaches an infinite loop. In this case it's not 241 // important to optimize the data Nodes - either the whole compilation will 242 // be tossed or this path (and all data Nodes) will go dead. 243 if (iff->outcnt() != 2) return; 244 245 // Make control-dependent data Nodes on the live path (path that will remain 246 // once the dominated IF is removed) become control-dependent on the 247 // dominating projection. 248 Node* dp = iff->as_If()->proj_out_or_null(pop == Op_IfTrue); 249 250 // Loop predicates may have depending checks which should not 251 // be skipped. For example, range check predicate has two checks 252 // for lower and upper bounds. 253 if (dp == NULL) 254 return; 255 256 ProjNode* dp_proj = dp->as_Proj(); 257 ProjNode* unc_proj = iff->as_If()->proj_out(1 - dp_proj->_con)->as_Proj(); 258 if (exclude_loop_predicate && 259 (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) != NULL || 260 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_profile_predicate) != NULL || 261 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check) != NULL)) { 262 // If this is a range check (IfNode::is_range_check), do not 263 // reorder because Compile::allow_range_check_smearing might have 264 // changed the check. 265 return; // Let IGVN transformation change control dependence. 266 } 267 268 IdealLoopTree *old_loop = get_loop(dp); 269 270 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { 271 Node* cd = dp->fast_out(i); // Control-dependent node 272 if (cd->depends_only_on_test()) { 273 assert(cd->in(0) == dp, ""); 274 _igvn.replace_input_of(cd, 0, prevdom); 275 set_early_ctrl(cd); 276 IdealLoopTree *new_loop = get_loop(get_ctrl(cd)); 277 if (old_loop != new_loop) { 278 if (!old_loop->_child) old_loop->_body.yank(cd); 279 if (!new_loop->_child) new_loop->_body.push(cd); 280 } 281 --i; 282 --imax; 283 } 284 } 285 } 286 287 //------------------------------has_local_phi_input---------------------------- 288 // Return TRUE if 'n' has Phi inputs from its local block and no other 289 // block-local inputs (all non-local-phi inputs come from earlier blocks) 290 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) { 291 Node *n_ctrl = get_ctrl(n); 292 // See if some inputs come from a Phi in this block, or from before 293 // this block. 294 uint i; 295 for( i = 1; i < n->req(); i++ ) { 296 Node *phi = n->in(i); 297 if( phi->is_Phi() && phi->in(0) == n_ctrl ) 298 break; 299 } 300 if( i >= n->req() ) 301 return NULL; // No Phi inputs; nowhere to clone thru 302 303 // Check for inputs created between 'n' and the Phi input. These 304 // must split as well; they have already been given the chance 305 // (courtesy of a post-order visit) and since they did not we must 306 // recover the 'cost' of splitting them by being very profitable 307 // when splitting 'n'. Since this is unlikely we simply give up. 308 for( i = 1; i < n->req(); i++ ) { 309 Node *m = n->in(i); 310 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) { 311 // We allow the special case of AddP's with no local inputs. 312 // This allows us to split-up address expressions. 313 if (m->is_AddP() && 314 get_ctrl(m->in(2)) != n_ctrl && 315 get_ctrl(m->in(3)) != n_ctrl) { 316 // Move the AddP up to dominating point 317 Node* c = find_non_split_ctrl(idom(n_ctrl)); 318 if (c->is_OuterStripMinedLoop()) { 319 c->as_Loop()->verify_strip_mined(1); 320 c = c->in(LoopNode::EntryControl); 321 } 322 set_ctrl_and_loop(m, c); 323 continue; 324 } 325 return NULL; 326 } 327 assert(m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control"); 328 } 329 330 return n_ctrl; 331 } 332 333 //------------------------------remix_address_expressions---------------------- 334 // Rework addressing expressions to get the most loop-invariant stuff 335 // moved out. We'd like to do all associative operators, but it's especially 336 // important (common) to do address expressions. 337 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) { 338 if (!has_ctrl(n)) return NULL; 339 Node *n_ctrl = get_ctrl(n); 340 IdealLoopTree *n_loop = get_loop(n_ctrl); 341 342 // See if 'n' mixes loop-varying and loop-invariant inputs and 343 // itself is loop-varying. 344 345 // Only interested in binary ops (and AddP) 346 if( n->req() < 3 || n->req() > 4 ) return NULL; 347 348 Node *n1_ctrl = get_ctrl(n->in( 1)); 349 Node *n2_ctrl = get_ctrl(n->in( 2)); 350 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3)); 351 IdealLoopTree *n1_loop = get_loop( n1_ctrl ); 352 IdealLoopTree *n2_loop = get_loop( n2_ctrl ); 353 IdealLoopTree *n3_loop = get_loop( n3_ctrl ); 354 355 // Does one of my inputs spin in a tighter loop than self? 356 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) || 357 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) || 358 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) ) 359 return NULL; // Leave well enough alone 360 361 // Is at least one of my inputs loop-invariant? 362 if( n1_loop == n_loop && 363 n2_loop == n_loop && 364 n3_loop == n_loop ) 365 return NULL; // No loop-invariant inputs 366 367 368 int n_op = n->Opcode(); 369 370 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2). 371 if( n_op == Op_LShiftI ) { 372 // Scale is loop invariant 373 Node *scale = n->in(2); 374 Node *scale_ctrl = get_ctrl(scale); 375 IdealLoopTree *scale_loop = get_loop(scale_ctrl ); 376 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) ) 377 return NULL; 378 const TypeInt *scale_t = scale->bottom_type()->isa_int(); 379 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 ) 380 return NULL; // Dont bother with byte/short masking 381 // Add must vary with loop (else shift would be loop-invariant) 382 Node *add = n->in(1); 383 Node *add_ctrl = get_ctrl(add); 384 IdealLoopTree *add_loop = get_loop(add_ctrl); 385 //assert( n_loop == add_loop, "" ); 386 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops 387 388 // Convert I-V into I+ (0-V); same for V-I 389 if( add->Opcode() == Op_SubI && 390 _igvn.type( add->in(1) ) != TypeInt::ZERO ) { 391 Node *zero = _igvn.intcon(0); 392 set_ctrl(zero, C->root()); 393 Node *neg = new SubINode( _igvn.intcon(0), add->in(2) ); 394 register_new_node( neg, get_ctrl(add->in(2) ) ); 395 add = new AddINode( add->in(1), neg ); 396 register_new_node( add, add_ctrl ); 397 } 398 if( add->Opcode() != Op_AddI ) return NULL; 399 // See if one add input is loop invariant 400 Node *add_var = add->in(1); 401 Node *add_var_ctrl = get_ctrl(add_var); 402 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 403 Node *add_invar = add->in(2); 404 Node *add_invar_ctrl = get_ctrl(add_invar); 405 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl ); 406 if( add_var_loop == n_loop ) { 407 } else if( add_invar_loop == n_loop ) { 408 // Swap to find the invariant part 409 add_invar = add_var; 410 add_invar_ctrl = add_var_ctrl; 411 add_invar_loop = add_var_loop; 412 add_var = add->in(2); 413 Node *add_var_ctrl = get_ctrl(add_var); 414 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 415 } else // Else neither input is loop invariant 416 return NULL; 417 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) ) 418 return NULL; // No invariant part of the add? 419 420 // Yes! Reshape address expression! 421 Node *inv_scale = new LShiftINode( add_invar, scale ); 422 Node *inv_scale_ctrl = 423 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ? 424 add_invar_ctrl : scale_ctrl; 425 register_new_node( inv_scale, inv_scale_ctrl ); 426 Node *var_scale = new LShiftINode( add_var, scale ); 427 register_new_node( var_scale, n_ctrl ); 428 Node *var_add = new AddINode( var_scale, inv_scale ); 429 register_new_node( var_add, n_ctrl ); 430 _igvn.replace_node( n, var_add ); 431 return var_add; 432 } 433 434 // Replace (I+V) with (V+I) 435 if( n_op == Op_AddI || 436 n_op == Op_AddL || 437 n_op == Op_AddF || 438 n_op == Op_AddD || 439 n_op == Op_MulI || 440 n_op == Op_MulL || 441 n_op == Op_MulF || 442 n_op == Op_MulD ) { 443 if( n2_loop == n_loop ) { 444 assert( n1_loop != n_loop, "" ); 445 n->swap_edges(1, 2); 446 } 447 } 448 449 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V), 450 // but not if I2 is a constant. 451 if( n_op == Op_AddP ) { 452 if( n2_loop == n_loop && n3_loop != n_loop ) { 453 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) { 454 Node *n22_ctrl = get_ctrl(n->in(2)->in(2)); 455 Node *n23_ctrl = get_ctrl(n->in(2)->in(3)); 456 IdealLoopTree *n22loop = get_loop( n22_ctrl ); 457 IdealLoopTree *n23_loop = get_loop( n23_ctrl ); 458 if( n22loop != n_loop && n22loop->is_member(n_loop) && 459 n23_loop == n_loop ) { 460 Node *add1 = new AddPNode( n->in(1), n->in(2)->in(2), n->in(3) ); 461 // Stuff new AddP in the loop preheader 462 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) ); 463 Node *add2 = new AddPNode( n->in(1), add1, n->in(2)->in(3) ); 464 register_new_node( add2, n_ctrl ); 465 _igvn.replace_node( n, add2 ); 466 return add2; 467 } 468 } 469 } 470 471 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V) 472 if (n2_loop != n_loop && n3_loop == n_loop) { 473 if (n->in(3)->Opcode() == Op_AddX) { 474 Node *V = n->in(3)->in(1); 475 Node *I = n->in(3)->in(2); 476 if (is_member(n_loop,get_ctrl(V))) { 477 } else { 478 Node *tmp = V; V = I; I = tmp; 479 } 480 if (!is_member(n_loop,get_ctrl(I))) { 481 Node *add1 = new AddPNode(n->in(1), n->in(2), I); 482 // Stuff new AddP in the loop preheader 483 register_new_node(add1, n_loop->_head->in(LoopNode::EntryControl)); 484 Node *add2 = new AddPNode(n->in(1), add1, V); 485 register_new_node(add2, n_ctrl); 486 _igvn.replace_node(n, add2); 487 return add2; 488 } 489 } 490 } 491 } 492 493 return NULL; 494 } 495 496 // Optimize ((in1[2*i] * in2[2*i]) + (in1[2*i+1] * in2[2*i+1])) 497 Node *PhaseIdealLoop::convert_add_to_muladd(Node* n) { 498 assert(n->Opcode() == Op_AddI, "sanity"); 499 Node * nn = NULL; 500 Node * in1 = n->in(1); 501 Node * in2 = n->in(2); 502 if (in1->Opcode() == Op_MulI && in2->Opcode() == Op_MulI) { 503 IdealLoopTree* loop_n = get_loop(get_ctrl(n)); 504 if (loop_n->_head->as_Loop()->is_valid_counted_loop() && 505 Matcher::match_rule_supported(Op_MulAddS2I) && 506 Matcher::match_rule_supported(Op_MulAddVS2VI)) { 507 Node* mul_in1 = in1->in(1); 508 Node* mul_in2 = in1->in(2); 509 Node* mul_in3 = in2->in(1); 510 Node* mul_in4 = in2->in(2); 511 if (mul_in1->Opcode() == Op_LoadS && 512 mul_in2->Opcode() == Op_LoadS && 513 mul_in3->Opcode() == Op_LoadS && 514 mul_in4->Opcode() == Op_LoadS) { 515 IdealLoopTree* loop1 = get_loop(get_ctrl(mul_in1)); 516 IdealLoopTree* loop2 = get_loop(get_ctrl(mul_in2)); 517 IdealLoopTree* loop3 = get_loop(get_ctrl(mul_in3)); 518 IdealLoopTree* loop4 = get_loop(get_ctrl(mul_in4)); 519 IdealLoopTree* loop5 = get_loop(get_ctrl(in1)); 520 IdealLoopTree* loop6 = get_loop(get_ctrl(in2)); 521 // All nodes should be in the same counted loop. 522 if (loop_n == loop1 && loop_n == loop2 && loop_n == loop3 && 523 loop_n == loop4 && loop_n == loop5 && loop_n == loop6) { 524 Node* adr1 = mul_in1->in(MemNode::Address); 525 Node* adr2 = mul_in2->in(MemNode::Address); 526 Node* adr3 = mul_in3->in(MemNode::Address); 527 Node* adr4 = mul_in4->in(MemNode::Address); 528 if (adr1->is_AddP() && adr2->is_AddP() && adr3->is_AddP() && adr4->is_AddP()) { 529 if ((adr1->in(AddPNode::Base) == adr3->in(AddPNode::Base)) && 530 (adr2->in(AddPNode::Base) == adr4->in(AddPNode::Base))) { 531 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in3, mul_in4); 532 register_new_node(nn, get_ctrl(n)); 533 _igvn.replace_node(n, nn); 534 return nn; 535 } else if ((adr1->in(AddPNode::Base) == adr4->in(AddPNode::Base)) && 536 (adr2->in(AddPNode::Base) == adr3->in(AddPNode::Base))) { 537 nn = new MulAddS2INode(mul_in1, mul_in2, mul_in4, mul_in3); 538 register_new_node(nn, get_ctrl(n)); 539 _igvn.replace_node(n, nn); 540 return nn; 541 } 542 } 543 } 544 } 545 } 546 } 547 return nn; 548 } 549 550 //------------------------------conditional_move------------------------------- 551 // Attempt to replace a Phi with a conditional move. We have some pretty 552 // strict profitability requirements. All Phis at the merge point must 553 // be converted, so we can remove the control flow. We need to limit the 554 // number of c-moves to a small handful. All code that was in the side-arms 555 // of the CFG diamond is now speculatively executed. This code has to be 556 // "cheap enough". We are pretty much limited to CFG diamonds that merge 557 // 1 or 2 items with a total of 1 or 2 ops executed speculatively. 558 Node *PhaseIdealLoop::conditional_move( Node *region ) { 559 560 assert(region->is_Region(), "sanity check"); 561 if (region->req() != 3) return NULL; 562 563 // Check for CFG diamond 564 Node *lp = region->in(1); 565 Node *rp = region->in(2); 566 if (!lp || !rp) return NULL; 567 Node *lp_c = lp->in(0); 568 if (lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If()) return NULL; 569 IfNode *iff = lp_c->as_If(); 570 571 // Check for ops pinned in an arm of the diamond. 572 // Can't remove the control flow in this case 573 if (lp->outcnt() > 1) return NULL; 574 if (rp->outcnt() > 1) return NULL; 575 576 IdealLoopTree* r_loop = get_loop(region); 577 assert(r_loop == get_loop(iff), "sanity"); 578 // Always convert to CMOVE if all results are used only outside this loop. 579 bool used_inside_loop = (r_loop == _ltree_root); 580 581 // Check profitability 582 int cost = 0; 583 int phis = 0; 584 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 585 Node *out = region->fast_out(i); 586 if (!out->is_Phi()) continue; // Ignore other control edges, etc 587 phis++; 588 PhiNode* phi = out->as_Phi(); 589 BasicType bt = phi->type()->basic_type(); 590 switch (bt) { 591 case T_DOUBLE: 592 case T_FLOAT: 593 if (C->use_cmove()) { 594 continue; //TODO: maybe we want to add some cost 595 } 596 cost += Matcher::float_cmove_cost(); // Could be very expensive 597 break; 598 case T_LONG: { 599 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's 600 } 601 case T_INT: // These all CMOV fine 602 case T_ADDRESS: { // (RawPtr) 603 cost++; 604 break; 605 } 606 case T_NARROWOOP: // Fall through 607 case T_OBJECT: { // Base oops are OK, but not derived oops 608 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr(); 609 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a 610 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus 611 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we 612 // have a Phi for the base here that we convert to a CMOVE all is well 613 // and good. But if the base is dead, we'll not make a CMOVE. Later 614 // the allocator will have to produce a base by creating a CMOVE of the 615 // relevant bases. This puts the allocator in the business of 616 // manufacturing expensive instructions, generally a bad plan. 617 // Just Say No to Conditionally-Moved Derived Pointers. 618 if (tp && tp->offset() != 0) 619 return NULL; 620 cost++; 621 break; 622 } 623 default: 624 return NULL; // In particular, can't do memory or I/O 625 } 626 // Add in cost any speculative ops 627 for (uint j = 1; j < region->req(); j++) { 628 Node *proj = region->in(j); 629 Node *inp = phi->in(j); 630 if (get_ctrl(inp) == proj) { // Found local op 631 cost++; 632 // Check for a chain of dependent ops; these will all become 633 // speculative in a CMOV. 634 for (uint k = 1; k < inp->req(); k++) 635 if (get_ctrl(inp->in(k)) == proj) 636 cost += ConditionalMoveLimit; // Too much speculative goo 637 } 638 } 639 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode. 640 // This will likely Split-If, a higher-payoff operation. 641 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) { 642 Node* use = phi->fast_out(k); 643 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr()) 644 cost += ConditionalMoveLimit; 645 // Is there a use inside the loop? 646 // Note: check only basic types since CMoveP is pinned. 647 if (!used_inside_loop && is_java_primitive(bt)) { 648 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use); 649 if (r_loop == u_loop || r_loop->is_member(u_loop)) { 650 used_inside_loop = true; 651 } 652 } 653 } 654 }//for 655 Node* bol = iff->in(1); 656 assert(bol->Opcode() == Op_Bool, ""); 657 int cmp_op = bol->in(1)->Opcode(); 658 // It is expensive to generate flags from a float compare. 659 // Avoid duplicated float compare. 660 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL; 661 662 float infrequent_prob = PROB_UNLIKELY_MAG(3); 663 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop. 664 if (used_inside_loop) { 665 if (cost >= ConditionalMoveLimit) return NULL; // Too much goo 666 667 // BlockLayoutByFrequency optimization moves infrequent branch 668 // from hot path. No point in CMOV'ing in such case (110 is used 669 // instead of 100 to take into account not exactness of float value). 670 if (BlockLayoutByFrequency) { 671 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f); 672 } 673 } 674 // Check for highly predictable branch. No point in CMOV'ing if 675 // we are going to predict accurately all the time. 676 if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) { 677 //keep going 678 } else if (iff->_prob < infrequent_prob || 679 iff->_prob > (1.0f - infrequent_prob)) 680 return NULL; 681 682 // -------------- 683 // Now replace all Phis with CMOV's 684 Node *cmov_ctrl = iff->in(0); 685 uint flip = (lp->Opcode() == Op_IfTrue); 686 Node_List wq; 687 while (1) { 688 PhiNode* phi = NULL; 689 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 690 Node *out = region->fast_out(i); 691 if (out->is_Phi()) { 692 phi = out->as_Phi(); 693 break; 694 } 695 } 696 if (phi == NULL) break; 697 if (PrintOpto && VerifyLoopOptimizations) { tty->print_cr("CMOV"); } 698 // Move speculative ops 699 wq.push(phi); 700 while (wq.size() > 0) { 701 Node *n = wq.pop(); 702 for (uint j = 1; j < n->req(); j++) { 703 Node* m = n->in(j); 704 if (m != NULL && !is_dominator(get_ctrl(m), cmov_ctrl)) { 705 #ifndef PRODUCT 706 if (PrintOpto && VerifyLoopOptimizations) { 707 tty->print(" speculate: "); 708 m->dump(); 709 } 710 #endif 711 set_ctrl(m, cmov_ctrl); 712 wq.push(m); 713 } 714 } 715 } 716 Node *cmov = CMoveNode::make(cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi)); 717 register_new_node( cmov, cmov_ctrl ); 718 _igvn.replace_node( phi, cmov ); 719 #ifndef PRODUCT 720 if (TraceLoopOpts) { 721 tty->print("CMOV "); 722 r_loop->dump_head(); 723 if (Verbose) { 724 bol->in(1)->dump(1); 725 cmov->dump(1); 726 } 727 } 728 if (VerifyLoopOptimizations) verify(); 729 #endif 730 } 731 732 // The useless CFG diamond will fold up later; see the optimization in 733 // RegionNode::Ideal. 734 _igvn._worklist.push(region); 735 736 return iff->in(1); 737 } 738 739 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) { 740 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 741 Node* u = m->fast_out(i); 742 if (u->is_CFG()) { 743 if (u->Opcode() == Op_NeverBranch) { 744 u = ((NeverBranchNode*)u)->proj_out(0); 745 enqueue_cfg_uses(u, wq); 746 } else { 747 wq.push(u); 748 } 749 } 750 } 751 } 752 753 // Try moving a store out of a loop, right before the loop 754 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) { 755 // Store has to be first in the loop body 756 IdealLoopTree *n_loop = get_loop(n_ctrl); 757 if (n->is_Store() && n_loop != _ltree_root && 758 n_loop->is_loop() && n_loop->_head->is_Loop() && 759 n->in(0) != NULL) { 760 Node* address = n->in(MemNode::Address); 761 Node* value = n->in(MemNode::ValueIn); 762 Node* mem = n->in(MemNode::Memory); 763 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 764 IdealLoopTree* value_loop = get_loop(get_ctrl(value)); 765 766 // - address and value must be loop invariant 767 // - memory must be a memory Phi for the loop 768 // - Store must be the only store on this memory slice in the 769 // loop: if there's another store following this one then value 770 // written at iteration i by the second store could be overwritten 771 // at iteration i+n by the first store: it's not safe to move the 772 // first store out of the loop 773 // - nothing must observe the memory Phi: it guarantees no read 774 // before the store, we are also guaranteed the store post 775 // dominates the loop head (ignoring a possible early 776 // exit). Otherwise there would be extra Phi involved between the 777 // loop's Phi and the store. 778 // - there must be no early exit from the loop before the Store 779 // (such an exit most of the time would be an extra use of the 780 // memory Phi but sometimes is a bottom memory Phi that takes the 781 // store as input). 782 783 if (!n_loop->is_member(address_loop) && 784 !n_loop->is_member(value_loop) && 785 mem->is_Phi() && mem->in(0) == n_loop->_head && 786 mem->outcnt() == 1 && 787 mem->in(LoopNode::LoopBackControl) == n) { 788 789 assert(n_loop->_tail != NULL, "need a tail"); 790 assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop"); 791 792 // Verify that there's no early exit of the loop before the store. 793 bool ctrl_ok = false; 794 { 795 // Follow control from loop head until n, we exit the loop or 796 // we reach the tail 797 ResourceMark rm; 798 Unique_Node_List wq; 799 wq.push(n_loop->_head); 800 801 for (uint next = 0; next < wq.size(); ++next) { 802 Node *m = wq.at(next); 803 if (m == n->in(0)) { 804 ctrl_ok = true; 805 continue; 806 } 807 assert(!has_ctrl(m), "should be CFG"); 808 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) { 809 ctrl_ok = false; 810 break; 811 } 812 enqueue_cfg_uses(m, wq); 813 if (wq.size() > 10) { 814 ctrl_ok = false; 815 break; 816 } 817 } 818 } 819 if (ctrl_ok) { 820 // move the Store 821 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem); 822 _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl)); 823 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl)); 824 // Disconnect the phi now. An empty phi can confuse other 825 // optimizations in this pass of loop opts. 826 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl)); 827 n_loop->_body.yank(mem); 828 829 set_ctrl_and_loop(n, n->in(0)); 830 831 return n; 832 } 833 } 834 } 835 return NULL; 836 } 837 838 // Try moving a store out of a loop, right after the loop 839 void PhaseIdealLoop::try_move_store_after_loop(Node* n) { 840 if (n->is_Store() && n->in(0) != NULL) { 841 Node *n_ctrl = get_ctrl(n); 842 IdealLoopTree *n_loop = get_loop(n_ctrl); 843 // Store must be in a loop 844 if (n_loop != _ltree_root && !n_loop->_irreducible) { 845 Node* address = n->in(MemNode::Address); 846 Node* value = n->in(MemNode::ValueIn); 847 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 848 // address must be loop invariant 849 if (!n_loop->is_member(address_loop)) { 850 // Store must be last on this memory slice in the loop and 851 // nothing in the loop must observe it 852 Node* phi = NULL; 853 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 854 Node* u = n->fast_out(i); 855 if (has_ctrl(u)) { // control use? 856 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 857 if (!n_loop->is_member(u_loop)) { 858 continue; 859 } 860 if (u->is_Phi() && u->in(0) == n_loop->_head) { 861 assert(_igvn.type(u) == Type::MEMORY, "bad phi"); 862 // multiple phis on the same slice are possible 863 if (phi != NULL) { 864 return; 865 } 866 phi = u; 867 continue; 868 } 869 } 870 return; 871 } 872 if (phi != NULL) { 873 // Nothing in the loop before the store (next iteration) 874 // must observe the stored value 875 bool mem_ok = true; 876 { 877 ResourceMark rm; 878 Unique_Node_List wq; 879 wq.push(phi); 880 for (uint next = 0; next < wq.size() && mem_ok; ++next) { 881 Node *m = wq.at(next); 882 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) { 883 Node* u = m->fast_out(i); 884 if (u->is_Store() || u->is_Phi()) { 885 if (u != n) { 886 wq.push(u); 887 mem_ok = (wq.size() <= 10); 888 } 889 } else { 890 mem_ok = false; 891 break; 892 } 893 } 894 } 895 } 896 if (mem_ok) { 897 // Move the store out of the loop if the LCA of all 898 // users (except for the phi) is outside the loop. 899 Node* hook = new Node(1); 900 _igvn.rehash_node_delayed(phi); 901 int count = phi->replace_edge(n, hook); 902 assert(count > 0, "inconsistent phi"); 903 904 // Compute latest point this store can go 905 Node* lca = get_late_ctrl(n, get_ctrl(n)); 906 if (n_loop->is_member(get_loop(lca))) { 907 // LCA is in the loop - bail out 908 _igvn.replace_node(hook, n); 909 return; 910 } 911 #ifdef ASSERT 912 if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) { 913 assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined"); 914 n_loop->_head->as_Loop()->verify_strip_mined(1); 915 Node* outer = n_loop->_head->as_CountedLoop()->outer_loop(); 916 IdealLoopTree* outer_loop = get_loop(outer); 917 assert(n_loop->_parent == outer_loop, "broken loop tree"); 918 assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state"); 919 } 920 #endif 921 922 // Move store out of the loop 923 _igvn.replace_node(hook, n->in(MemNode::Memory)); 924 _igvn.replace_input_of(n, 0, lca); 925 set_ctrl_and_loop(n, lca); 926 927 // Disconnect the phi now. An empty phi can confuse other 928 // optimizations in this pass of loop opts.. 929 if (phi->in(LoopNode::LoopBackControl) == phi) { 930 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl)); 931 n_loop->_body.yank(phi); 932 } 933 } 934 } 935 } 936 } 937 } 938 } 939 940 //------------------------------split_if_with_blocks_pre----------------------- 941 // Do the real work in a non-recursive function. Data nodes want to be 942 // cloned in the pre-order so they can feed each other nicely. 943 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) { 944 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 945 Node* bs_res = bs->split_if_pre(this, n); 946 if (bs_res != NULL) { 947 return bs_res; 948 } 949 // Cloning these guys is unlikely to win 950 int n_op = n->Opcode(); 951 if( n_op == Op_MergeMem ) return n; 952 if( n->is_Proj() ) return n; 953 // Do not clone-up CmpFXXX variations, as these are always 954 // followed by a CmpI 955 if( n->is_Cmp() ) return n; 956 // Attempt to use a conditional move instead of a phi/branch 957 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) { 958 Node *cmov = conditional_move( n ); 959 if( cmov ) return cmov; 960 } 961 if( n->is_CFG() || n->is_LoadStore() ) 962 return n; 963 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd 964 n_op == Op_Opaque2 ) { 965 if( !C->major_progress() ) // If chance of no more loop opts... 966 _igvn._worklist.push(n); // maybe we'll remove them 967 return n; 968 } 969 970 if( n->is_Con() ) return n; // No cloning for Con nodes 971 972 Node *n_ctrl = get_ctrl(n); 973 if( !n_ctrl ) return n; // Dead node 974 975 Node* res = try_move_store_before_loop(n, n_ctrl); 976 if (res != NULL) { 977 return n; 978 } 979 980 // Attempt to remix address expressions for loop invariants 981 Node *m = remix_address_expressions( n ); 982 if( m ) return m; 983 984 if (n_op == Op_AddI) { 985 Node *nn = convert_add_to_muladd( n ); 986 if ( nn ) return nn; 987 } 988 989 if (n->is_ConstraintCast()) { 990 Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this); 991 // ConstraintCastNode::dominating_cast() uses node control input to determine domination. 992 // Node control inputs don't necessarily agree with loop control info (due to 993 // transformations happened in between), thus additional dominance check is needed 994 // to keep loop info valid. 995 if (dom_cast != NULL && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) { 996 _igvn.replace_node(n, dom_cast); 997 return dom_cast; 998 } 999 } 1000 1001 // Determine if the Node has inputs from some local Phi. 1002 // Returns the block to clone thru. 1003 Node *n_blk = has_local_phi_input( n ); 1004 if( !n_blk ) return n; 1005 1006 // Do not clone the trip counter through on a CountedLoop 1007 // (messes up the canonical shape). 1008 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n; 1009 1010 // Check for having no control input; not pinned. Allow 1011 // dominating control. 1012 if (n->in(0)) { 1013 Node *dom = idom(n_blk); 1014 if (dom_lca(n->in(0), dom) != n->in(0)) { 1015 return n; 1016 } 1017 } 1018 // Policy: when is it profitable. You must get more wins than 1019 // policy before it is considered profitable. Policy is usually 0, 1020 // so 1 win is considered profitable. Big merges will require big 1021 // cloning, so get a larger policy. 1022 int policy = n_blk->req() >> 2; 1023 1024 // If the loop is a candidate for range check elimination, 1025 // delay splitting through it's phi until a later loop optimization 1026 if (n_blk->is_CountedLoop()) { 1027 IdealLoopTree *lp = get_loop(n_blk); 1028 if (lp && lp->_rce_candidate) { 1029 return n; 1030 } 1031 } 1032 1033 if (must_throttle_split_if()) return n; 1034 1035 // Split 'n' through the merge point if it is profitable 1036 Node *phi = split_thru_phi( n, n_blk, policy ); 1037 if (!phi) return n; 1038 1039 // Found a Phi to split thru! 1040 // Replace 'n' with the new phi 1041 _igvn.replace_node( n, phi ); 1042 // Moved a load around the loop, 'en-registering' something. 1043 if (n_blk->is_Loop() && n->is_Load() && 1044 !phi->in(LoopNode::LoopBackControl)->is_Load()) 1045 C->set_major_progress(); 1046 1047 return phi; 1048 } 1049 1050 static bool merge_point_too_heavy(Compile* C, Node* region) { 1051 // Bail out if the region and its phis have too many users. 1052 int weight = 0; 1053 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1054 weight += region->fast_out(i)->outcnt(); 1055 } 1056 int nodes_left = C->max_node_limit() - C->live_nodes(); 1057 if (weight * 8 > nodes_left) { 1058 if (PrintOpto) { 1059 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight); 1060 } 1061 return true; 1062 } else { 1063 return false; 1064 } 1065 } 1066 1067 static bool merge_point_safe(Node* region) { 1068 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode 1069 // having a PhiNode input. This sidesteps the dangerous case where the split 1070 // ConvI2LNode may become TOP if the input Value() does not 1071 // overlap the ConvI2L range, leaving a node which may not dominate its 1072 // uses. 1073 // A better fix for this problem can be found in the BugTraq entry, but 1074 // expediency for Mantis demands this hack. 1075 // 6855164: If the merge point has a FastLockNode with a PhiNode input, we stop 1076 // split_if_with_blocks from splitting a block because we could not move around 1077 // the FastLockNode. 1078 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1079 Node* n = region->fast_out(i); 1080 if (n->is_Phi()) { 1081 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1082 Node* m = n->fast_out(j); 1083 if (m->is_FastLock()) 1084 return false; 1085 #if INCLUDE_SHENANDOAHGC 1086 if (m->is_ShenandoahBarrier() && m->has_out_with(Op_FastLock)) { 1087 return false; 1088 } 1089 #endif 1090 #ifdef _LP64 1091 if (m->Opcode() == Op_ConvI2L) 1092 return false; 1093 if (m->is_CastII() && m->isa_CastII()->has_range_check()) { 1094 return false; 1095 } 1096 #endif 1097 } 1098 } 1099 } 1100 return true; 1101 } 1102 1103 1104 //------------------------------place_near_use--------------------------------- 1105 // Place some computation next to use but not inside inner loops. 1106 // For inner loop uses move it to the preheader area. 1107 Node *PhaseIdealLoop::place_near_use(Node *useblock) const { 1108 IdealLoopTree *u_loop = get_loop( useblock ); 1109 if (u_loop->_irreducible) { 1110 return useblock; 1111 } 1112 if (u_loop->_child) { 1113 if (useblock == u_loop->_head && u_loop->_head->is_OuterStripMinedLoop()) { 1114 return u_loop->_head->in(LoopNode::EntryControl); 1115 } 1116 return useblock; 1117 } 1118 return u_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl); 1119 } 1120 1121 1122 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) { 1123 if (!n->is_If() || n->is_CountedLoopEnd()) { 1124 return false; 1125 } 1126 if (!n->in(0)->is_Region()) { 1127 return false; 1128 } 1129 Node* region = n->in(0); 1130 Node* dom = idom(region); 1131 if (!dom->is_If() || dom->in(1) != n->in(1)) { 1132 return false; 1133 } 1134 IfNode* dom_if = dom->as_If(); 1135 Node* proj_true = dom_if->proj_out(1); 1136 Node* proj_false = dom_if->proj_out(0); 1137 1138 for (uint i = 1; i < region->req(); i++) { 1139 if (is_dominator(proj_true, region->in(i))) { 1140 continue; 1141 } 1142 if (is_dominator(proj_false, region->in(i))) { 1143 continue; 1144 } 1145 return false; 1146 } 1147 1148 return true; 1149 } 1150 1151 1152 bool PhaseIdealLoop::can_split_if(Node* n_ctrl) { 1153 if (must_throttle_split_if()) { 1154 return false; 1155 } 1156 1157 // Do not do 'split-if' if irreducible loops are present. 1158 if (_has_irreducible_loops) { 1159 return false; 1160 } 1161 1162 if (merge_point_too_heavy(C, n_ctrl)) { 1163 return false; 1164 } 1165 1166 // Do not do 'split-if' if some paths are dead. First do dead code 1167 // elimination and then see if its still profitable. 1168 for (uint i = 1; i < n_ctrl->req(); i++) { 1169 if (n_ctrl->in(i) == C->top()) { 1170 return false; 1171 } 1172 } 1173 1174 // If trying to do a 'Split-If' at the loop head, it is only 1175 // profitable if the cmp folds up on BOTH paths. Otherwise we 1176 // risk peeling a loop forever. 1177 1178 // CNC - Disabled for now. Requires careful handling of loop 1179 // body selection for the cloned code. Also, make sure we check 1180 // for any input path not being in the same loop as n_ctrl. For 1181 // irreducible loops we cannot check for 'n_ctrl->is_Loop()' 1182 // because the alternative loop entry points won't be converted 1183 // into LoopNodes. 1184 IdealLoopTree *n_loop = get_loop(n_ctrl); 1185 for (uint j = 1; j < n_ctrl->req(); j++) { 1186 if (get_loop(n_ctrl->in(j)) != n_loop) { 1187 return false; 1188 } 1189 } 1190 1191 // Check for safety of the merge point. 1192 if (!merge_point_safe(n_ctrl)) { 1193 return false; 1194 } 1195 1196 return true; 1197 } 1198 1199 //------------------------------split_if_with_blocks_post---------------------- 1200 // Do the real work in a non-recursive function. CFG hackery wants to be 1201 // in the post-order, so it can dirty the I-DOM info and not use the dirtied 1202 // info. 1203 void PhaseIdealLoop::split_if_with_blocks_post(Node *n, bool last_round) { 1204 1205 // Cloning Cmp through Phi's involves the split-if transform. 1206 // FastLock is not used by an If 1207 if (n->is_Cmp() && !n->is_FastLock() && !last_round) { 1208 Node *n_ctrl = get_ctrl(n); 1209 // Determine if the Node has inputs from some local Phi. 1210 // Returns the block to clone thru. 1211 Node *n_blk = has_local_phi_input(n); 1212 if (n_blk != n_ctrl) { 1213 return; 1214 } 1215 1216 if (!can_split_if(n_ctrl)) { 1217 return; 1218 } 1219 1220 if (n->outcnt() != 1) { 1221 return; // Multiple bool's from 1 compare? 1222 } 1223 Node *bol = n->unique_out(); 1224 assert(bol->is_Bool(), "expect a bool here"); 1225 if (bol->outcnt() != 1) { 1226 return;// Multiple branches from 1 compare? 1227 } 1228 Node *iff = bol->unique_out(); 1229 1230 // Check some safety conditions 1231 if (iff->is_If()) { // Classic split-if? 1232 if (iff->in(0) != n_ctrl) { 1233 return; // Compare must be in same blk as if 1234 } 1235 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE 1236 // Can't split CMove with different control edge. 1237 if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) { 1238 return; 1239 } 1240 if (get_ctrl(iff->in(2)) == n_ctrl || 1241 get_ctrl(iff->in(3)) == n_ctrl) { 1242 return; // Inputs not yet split-up 1243 } 1244 if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) { 1245 return; // Loop-invar test gates loop-varying CMOVE 1246 } 1247 } else { 1248 return; // some other kind of node, such as an Allocate 1249 } 1250 1251 // When is split-if profitable? Every 'win' on means some control flow 1252 // goes dead, so it's almost always a win. 1253 int policy = 0; 1254 // Split compare 'n' through the merge point if it is profitable 1255 Node *phi = split_thru_phi( n, n_ctrl, policy); 1256 if (!phi) { 1257 return; 1258 } 1259 1260 // Found a Phi to split thru! 1261 // Replace 'n' with the new phi 1262 _igvn.replace_node(n, phi); 1263 1264 // Now split the bool up thru the phi 1265 Node *bolphi = split_thru_phi(bol, n_ctrl, -1); 1266 guarantee(bolphi != NULL, "null boolean phi node"); 1267 1268 _igvn.replace_node(bol, bolphi); 1269 assert(iff->in(1) == bolphi, ""); 1270 1271 if (bolphi->Value(&_igvn)->singleton()) { 1272 return; 1273 } 1274 1275 // Conditional-move? Must split up now 1276 if (!iff->is_If()) { 1277 Node *cmovphi = split_thru_phi(iff, n_ctrl, -1); 1278 _igvn.replace_node(iff, cmovphi); 1279 return; 1280 } 1281 1282 // Now split the IF 1283 do_split_if(iff); 1284 return; 1285 } 1286 1287 // Two identical ifs back to back can be merged 1288 if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) { 1289 Node *n_ctrl = n->in(0); 1290 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1)); 1291 IfNode* dom_if = idom(n_ctrl)->as_If(); 1292 Node* proj_true = dom_if->proj_out(1); 1293 Node* proj_false = dom_if->proj_out(0); 1294 Node* con_true = _igvn.makecon(TypeInt::ONE); 1295 Node* con_false = _igvn.makecon(TypeInt::ZERO); 1296 1297 for (uint i = 1; i < n_ctrl->req(); i++) { 1298 if (is_dominator(proj_true, n_ctrl->in(i))) { 1299 bolphi->init_req(i, con_true); 1300 } else { 1301 assert(is_dominator(proj_false, n_ctrl->in(i)), "bad if"); 1302 bolphi->init_req(i, con_false); 1303 } 1304 } 1305 register_new_node(bolphi, n_ctrl); 1306 _igvn.replace_input_of(n, 1, bolphi); 1307 1308 // Now split the IF 1309 do_split_if(n); 1310 return; 1311 } 1312 1313 // Check for an IF ready to split; one that has its 1314 // condition codes input coming from a Phi at the block start. 1315 int n_op = n->Opcode(); 1316 1317 // Check for an IF being dominated by another IF same test 1318 if (n_op == Op_If || 1319 n_op == Op_RangeCheck) { 1320 Node *bol = n->in(1); 1321 uint max = bol->outcnt(); 1322 // Check for same test used more than once? 1323 if (max > 1 && bol->is_Bool()) { 1324 // Search up IDOMs to see if this IF is dominated. 1325 Node *cutoff = get_ctrl(bol); 1326 1327 // Now search up IDOMs till cutoff, looking for a dominating test 1328 Node *prevdom = n; 1329 Node *dom = idom(prevdom); 1330 while (dom != cutoff) { 1331 if (dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom) { 1332 // Replace the dominated test with an obvious true or false. 1333 // Place it on the IGVN worklist for later cleanup. 1334 C->set_major_progress(); 1335 dominated_by(prevdom, n, false, true); 1336 #ifndef PRODUCT 1337 if( VerifyLoopOptimizations ) verify(); 1338 #endif 1339 return; 1340 } 1341 prevdom = dom; 1342 dom = idom(prevdom); 1343 } 1344 } 1345 } 1346 1347 // See if a shared loop-varying computation has no loop-varying uses. 1348 // Happens if something is only used for JVM state in uncommon trap exits, 1349 // like various versions of induction variable+offset. Clone the 1350 // computation per usage to allow it to sink out of the loop. 1351 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about) 1352 Node *n_ctrl = get_ctrl(n); 1353 IdealLoopTree *n_loop = get_loop(n_ctrl); 1354 if( n_loop != _ltree_root ) { 1355 DUIterator_Fast imax, i = n->fast_outs(imax); 1356 for (; i < imax; i++) { 1357 Node* u = n->fast_out(i); 1358 if( !has_ctrl(u) ) break; // Found control user 1359 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 1360 if( u_loop == n_loop ) break; // Found loop-varying use 1361 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop 1362 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003 1363 } 1364 bool did_break = (i < imax); // Did we break out of the previous loop? 1365 if (!did_break && n->outcnt() > 1) { // All uses in outer loops! 1366 Node *late_load_ctrl = NULL; 1367 if (n->is_Load()) { 1368 // If n is a load, get and save the result from get_late_ctrl(), 1369 // to be later used in calculating the control for n's clones. 1370 clear_dom_lca_tags(); 1371 late_load_ctrl = get_late_ctrl(n, n_ctrl); 1372 } 1373 // If n is a load, and the late control is the same as the current 1374 // control, then the cloning of n is a pointless exercise, because 1375 // GVN will ensure that we end up where we started. 1376 if (!n->is_Load() || late_load_ctrl != n_ctrl) { 1377 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1378 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) { 1379 Node *u = n->last_out(j); // Clone private computation per use 1380 _igvn.rehash_node_delayed(u); 1381 Node *x = n->clone(); // Clone computation 1382 Node *x_ctrl = NULL; 1383 if( u->is_Phi() ) { 1384 // Replace all uses of normal nodes. Replace Phi uses 1385 // individually, so the separate Nodes can sink down 1386 // different paths. 1387 uint k = 1; 1388 while( u->in(k) != n ) k++; 1389 u->set_req( k, x ); 1390 // x goes next to Phi input path 1391 x_ctrl = u->in(0)->in(k); 1392 --j; 1393 } else { // Normal use 1394 // Replace all uses 1395 for( uint k = 0; k < u->req(); k++ ) { 1396 if( u->in(k) == n ) { 1397 u->set_req( k, x ); 1398 --j; 1399 } 1400 } 1401 x_ctrl = get_ctrl(u); 1402 } 1403 1404 // Find control for 'x' next to use but not inside inner loops. 1405 // For inner loop uses get the preheader area. 1406 x_ctrl = place_near_use(x_ctrl); 1407 1408 if (bs->sink_node(this, n, x, x_ctrl, n_ctrl)) { 1409 continue; 1410 } 1411 1412 if (n->is_Load()) { 1413 // For loads, add a control edge to a CFG node outside of the loop 1414 // to force them to not combine and return back inside the loop 1415 // during GVN optimization (4641526). 1416 // 1417 // Because we are setting the actual control input, factor in 1418 // the result from get_late_ctrl() so we respect any 1419 // anti-dependences. (6233005). 1420 x_ctrl = dom_lca(late_load_ctrl, x_ctrl); 1421 1422 // Don't allow the control input to be a CFG splitting node. 1423 // Such nodes should only have ProjNodes as outs, e.g. IfNode 1424 // should only have IfTrueNode and IfFalseNode (4985384). 1425 x_ctrl = find_non_split_ctrl(x_ctrl); 1426 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone"); 1427 1428 x->set_req(0, x_ctrl); 1429 } 1430 register_new_node(x, x_ctrl); 1431 1432 // Some institutional knowledge is needed here: 'x' is 1433 // yanked because if the optimizer runs GVN on it all the 1434 // cloned x's will common up and undo this optimization and 1435 // be forced back in the loop. This is annoying because it 1436 // makes +VerifyOpto report false-positives on progress. I 1437 // tried setting control edges on the x's to force them to 1438 // not combine, but the matching gets worried when it tries 1439 // to fold a StoreP and an AddP together (as part of an 1440 // address expression) and the AddP and StoreP have 1441 // different controls. 1442 if (!x->is_Load() && !x->is_DecodeNarrowPtr()) _igvn._worklist.yank(x); 1443 } 1444 _igvn.remove_dead_node(n); 1445 } 1446 } 1447 } 1448 } 1449 1450 try_move_store_after_loop(n); 1451 1452 // Check for Opaque2's who's loop has disappeared - who's input is in the 1453 // same loop nest as their output. Remove 'em, they are no longer useful. 1454 if( n_op == Op_Opaque2 && 1455 n->in(1) != NULL && 1456 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) { 1457 _igvn.replace_node( n, n->in(1) ); 1458 } 1459 1460 #if INCLUDE_ZGC 1461 if (UseZGC) { 1462 ZBarrierSetC2::loop_optimize_gc_barrier(this, n, last_round); 1463 } 1464 #endif 1465 } 1466 1467 //------------------------------split_if_with_blocks--------------------------- 1468 // Check for aggressive application of 'split-if' optimization, 1469 // using basic block level info. 1470 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack, bool last_round) { 1471 Node* root = C->root(); 1472 visited.set(root->_idx); // first, mark root as visited 1473 // Do pre-visit work for root 1474 Node* n = split_if_with_blocks_pre(root); 1475 uint cnt = n->outcnt(); 1476 uint i = 0; 1477 1478 while (true) { 1479 // Visit all children 1480 if (i < cnt) { 1481 Node* use = n->raw_out(i); 1482 ++i; 1483 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) { 1484 // Now do pre-visit work for this use 1485 use = split_if_with_blocks_pre(use); 1486 nstack.push(n, i); // Save parent and next use's index. 1487 n = use; // Process all children of current use. 1488 cnt = use->outcnt(); 1489 i = 0; 1490 } 1491 } 1492 else { 1493 // All of n's children have been processed, complete post-processing. 1494 if (cnt != 0 && !n->is_Con()) { 1495 assert(has_node(n), "no dead nodes"); 1496 split_if_with_blocks_post(n, last_round); 1497 } 1498 if (must_throttle_split_if()) { 1499 nstack.clear(); 1500 } 1501 if (nstack.is_empty()) { 1502 // Finished all nodes on stack. 1503 break; 1504 } 1505 // Get saved parent node and next use's index. Visit the rest of uses. 1506 n = nstack.node(); 1507 cnt = n->outcnt(); 1508 i = nstack.index(); 1509 nstack.pop(); 1510 } 1511 } 1512 } 1513 1514 1515 //============================================================================= 1516 // 1517 // C L O N E A L O O P B O D Y 1518 // 1519 1520 //------------------------------clone_iff-------------------------------------- 1521 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1522 // "Nearly" because all Nodes have been cloned from the original in the loop, 1523 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1524 // through the Phi recursively, and return a Bool. 1525 Node* PhaseIdealLoop::clone_iff(PhiNode *phi, IdealLoopTree *loop) { 1526 1527 // Convert this Phi into a Phi merging Bools 1528 uint i; 1529 for (i = 1; i < phi->req(); i++) { 1530 Node *b = phi->in(i); 1531 if (b->is_Phi()) { 1532 _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi(), loop)); 1533 } else { 1534 assert(b->is_Bool() || b->Opcode() == Op_Opaque4, ""); 1535 } 1536 } 1537 1538 Node* n = phi->in(1); 1539 Node* sample_opaque = NULL; 1540 Node *sample_bool = NULL; 1541 if (n->Opcode() == Op_Opaque4) { 1542 sample_opaque = n; 1543 sample_bool = n->in(1); 1544 assert(sample_bool->is_Bool(), "wrong type"); 1545 } else { 1546 sample_bool = n; 1547 } 1548 Node *sample_cmp = sample_bool->in(1); 1549 1550 // Make Phis to merge the Cmp's inputs. 1551 PhiNode *phi1 = new PhiNode(phi->in(0), Type::TOP); 1552 PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP); 1553 for (i = 1; i < phi->req(); i++) { 1554 Node *n1 = sample_opaque == NULL ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1); 1555 Node *n2 = sample_opaque == NULL ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2); 1556 phi1->set_req(i, n1); 1557 phi2->set_req(i, n2); 1558 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 1559 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 1560 } 1561 // See if these Phis have been made before. 1562 // Register with optimizer 1563 Node *hit1 = _igvn.hash_find_insert(phi1); 1564 if (hit1) { // Hit, toss just made Phi 1565 _igvn.remove_dead_node(phi1); // Remove new phi 1566 assert(hit1->is_Phi(), "" ); 1567 phi1 = (PhiNode*)hit1; // Use existing phi 1568 } else { // Miss 1569 _igvn.register_new_node_with_optimizer(phi1); 1570 } 1571 Node *hit2 = _igvn.hash_find_insert(phi2); 1572 if (hit2) { // Hit, toss just made Phi 1573 _igvn.remove_dead_node(phi2); // Remove new phi 1574 assert(hit2->is_Phi(), "" ); 1575 phi2 = (PhiNode*)hit2; // Use existing phi 1576 } else { // Miss 1577 _igvn.register_new_node_with_optimizer(phi2); 1578 } 1579 // Register Phis with loop/block info 1580 set_ctrl(phi1, phi->in(0)); 1581 set_ctrl(phi2, phi->in(0)); 1582 // Make a new Cmp 1583 Node *cmp = sample_cmp->clone(); 1584 cmp->set_req(1, phi1); 1585 cmp->set_req(2, phi2); 1586 _igvn.register_new_node_with_optimizer(cmp); 1587 set_ctrl(cmp, phi->in(0)); 1588 1589 // Make a new Bool 1590 Node *b = sample_bool->clone(); 1591 b->set_req(1,cmp); 1592 _igvn.register_new_node_with_optimizer(b); 1593 set_ctrl(b, phi->in(0)); 1594 1595 if (sample_opaque != NULL) { 1596 Node* opaque = sample_opaque->clone(); 1597 opaque->set_req(1, b); 1598 _igvn.register_new_node_with_optimizer(opaque); 1599 set_ctrl(opaque, phi->in(0)); 1600 return opaque; 1601 } 1602 1603 assert(b->is_Bool(), ""); 1604 return b; 1605 } 1606 1607 //------------------------------clone_bool------------------------------------- 1608 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1609 // "Nearly" because all Nodes have been cloned from the original in the loop, 1610 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1611 // through the Phi recursively, and return a Bool. 1612 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) { 1613 uint i; 1614 // Convert this Phi into a Phi merging Bools 1615 for( i = 1; i < phi->req(); i++ ) { 1616 Node *b = phi->in(i); 1617 if( b->is_Phi() ) { 1618 _igvn.replace_input_of(phi, i, clone_bool( b->as_Phi(), loop )); 1619 } else { 1620 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" ); 1621 } 1622 } 1623 1624 Node *sample_cmp = phi->in(1); 1625 1626 // Make Phis to merge the Cmp's inputs. 1627 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP ); 1628 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP ); 1629 for( uint j = 1; j < phi->req(); j++ ) { 1630 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP 1631 Node *n1, *n2; 1632 if( cmp_top->is_Cmp() ) { 1633 n1 = cmp_top->in(1); 1634 n2 = cmp_top->in(2); 1635 } else { 1636 n1 = n2 = cmp_top; 1637 } 1638 phi1->set_req( j, n1 ); 1639 phi2->set_req( j, n2 ); 1640 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 1641 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 1642 } 1643 1644 // See if these Phis have been made before. 1645 // Register with optimizer 1646 Node *hit1 = _igvn.hash_find_insert(phi1); 1647 if( hit1 ) { // Hit, toss just made Phi 1648 _igvn.remove_dead_node(phi1); // Remove new phi 1649 assert( hit1->is_Phi(), "" ); 1650 phi1 = (PhiNode*)hit1; // Use existing phi 1651 } else { // Miss 1652 _igvn.register_new_node_with_optimizer(phi1); 1653 } 1654 Node *hit2 = _igvn.hash_find_insert(phi2); 1655 if( hit2 ) { // Hit, toss just made Phi 1656 _igvn.remove_dead_node(phi2); // Remove new phi 1657 assert( hit2->is_Phi(), "" ); 1658 phi2 = (PhiNode*)hit2; // Use existing phi 1659 } else { // Miss 1660 _igvn.register_new_node_with_optimizer(phi2); 1661 } 1662 // Register Phis with loop/block info 1663 set_ctrl(phi1, phi->in(0)); 1664 set_ctrl(phi2, phi->in(0)); 1665 // Make a new Cmp 1666 Node *cmp = sample_cmp->clone(); 1667 cmp->set_req( 1, phi1 ); 1668 cmp->set_req( 2, phi2 ); 1669 _igvn.register_new_node_with_optimizer(cmp); 1670 set_ctrl(cmp, phi->in(0)); 1671 1672 assert( cmp->is_Cmp(), "" ); 1673 return (CmpNode*)cmp; 1674 } 1675 1676 //------------------------------sink_use--------------------------------------- 1677 // If 'use' was in the loop-exit block, it now needs to be sunk 1678 // below the post-loop merge point. 1679 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) { 1680 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) { 1681 set_ctrl(use, post_loop); 1682 for (DUIterator j = use->outs(); use->has_out(j); j++) 1683 sink_use(use->out(j), post_loop); 1684 } 1685 } 1686 1687 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new, 1688 IdealLoopTree* loop, IdealLoopTree* outer_loop, 1689 Node_List*& split_if_set, Node_List*& split_bool_set, 1690 Node_List*& split_cex_set, Node_List& worklist, 1691 uint new_counter, CloneLoopMode mode) { 1692 Node* nnn = old_new[old->_idx]; 1693 // Copy uses to a worklist, so I can munge the def-use info 1694 // with impunity. 1695 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1696 worklist.push(old->fast_out(j)); 1697 1698 while( worklist.size() ) { 1699 Node *use = worklist.pop(); 1700 if (!has_node(use)) continue; // Ignore dead nodes 1701 if (use->in(0) == C->top()) continue; 1702 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1703 // Check for data-use outside of loop - at least one of OLD or USE 1704 // must not be a CFG node. 1705 #ifdef ASSERT 1706 if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == NULL) { 1707 Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint(); 1708 assert(mode == ControlAroundStripMined && use == sfpt, "missed a node"); 1709 } 1710 #endif 1711 if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) { 1712 1713 // If the Data use is an IF, that means we have an IF outside of the 1714 // loop that is switching on a condition that is set inside of the 1715 // loop. Happens if people set a loop-exit flag; then test the flag 1716 // in the loop to break the loop, then test is again outside of the 1717 // loop to determine which way the loop exited. 1718 // Loop predicate If node connects to Bool node through Opaque1 node. 1719 if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use) || use->Opcode() == Op_Opaque4) { 1720 // Since this code is highly unlikely, we lazily build the worklist 1721 // of such Nodes to go split. 1722 if (!split_if_set) { 1723 ResourceArea *area = Thread::current()->resource_area(); 1724 split_if_set = new Node_List(area); 1725 } 1726 split_if_set->push(use); 1727 } 1728 if (use->is_Bool()) { 1729 if (!split_bool_set) { 1730 ResourceArea *area = Thread::current()->resource_area(); 1731 split_bool_set = new Node_List(area); 1732 } 1733 split_bool_set->push(use); 1734 } 1735 if (use->Opcode() == Op_CreateEx) { 1736 if (!split_cex_set) { 1737 ResourceArea *area = Thread::current()->resource_area(); 1738 split_cex_set = new Node_List(area); 1739 } 1740 split_cex_set->push(use); 1741 } 1742 1743 1744 // Get "block" use is in 1745 uint idx = 0; 1746 while( use->in(idx) != old ) idx++; 1747 Node *prev = use->is_CFG() ? use : get_ctrl(use); 1748 assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" ); 1749 Node *cfg = prev->_idx >= new_counter 1750 ? prev->in(2) 1751 : idom(prev); 1752 if( use->is_Phi() ) // Phi use is in prior block 1753 cfg = prev->in(idx); // NOT in block of Phi itself 1754 if (cfg->is_top()) { // Use is dead? 1755 _igvn.replace_input_of(use, idx, C->top()); 1756 continue; 1757 } 1758 1759 // If use is referenced through control edge... (idx == 0) 1760 if (mode == IgnoreStripMined && idx == 0) { 1761 LoopNode *head = loop->_head->as_Loop(); 1762 if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) { 1763 // That node is outside the inner loop, leave it outside the 1764 // outer loop as well to not confuse verification code. 1765 assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop"); 1766 _igvn.replace_input_of(use, 0, head->outer_loop_exit()); 1767 continue; 1768 } 1769 } 1770 1771 while(!outer_loop->is_member(get_loop(cfg))) { 1772 prev = cfg; 1773 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg); 1774 } 1775 // If the use occurs after merging several exits from the loop, then 1776 // old value must have dominated all those exits. Since the same old 1777 // value was used on all those exits we did not need a Phi at this 1778 // merge point. NOW we do need a Phi here. Each loop exit value 1779 // is now merged with the peeled body exit; each exit gets its own 1780 // private Phi and those Phis need to be merged here. 1781 Node *phi; 1782 if( prev->is_Region() ) { 1783 if( idx == 0 ) { // Updating control edge? 1784 phi = prev; // Just use existing control 1785 } else { // Else need a new Phi 1786 phi = PhiNode::make( prev, old ); 1787 // Now recursively fix up the new uses of old! 1788 for( uint i = 1; i < prev->req(); i++ ) { 1789 worklist.push(phi); // Onto worklist once for each 'old' input 1790 } 1791 } 1792 } else { 1793 // Get new RegionNode merging old and new loop exits 1794 prev = old_new[prev->_idx]; 1795 assert( prev, "just made this in step 7" ); 1796 if( idx == 0) { // Updating control edge? 1797 phi = prev; // Just use existing control 1798 } else { // Else need a new Phi 1799 // Make a new Phi merging data values properly 1800 phi = PhiNode::make( prev, old ); 1801 phi->set_req( 1, nnn ); 1802 } 1803 } 1804 // If inserting a new Phi, check for prior hits 1805 if( idx != 0 ) { 1806 Node *hit = _igvn.hash_find_insert(phi); 1807 if( hit == NULL ) { 1808 _igvn.register_new_node_with_optimizer(phi); // Register new phi 1809 } else { // or 1810 // Remove the new phi from the graph and use the hit 1811 _igvn.remove_dead_node(phi); 1812 phi = hit; // Use existing phi 1813 } 1814 set_ctrl(phi, prev); 1815 } 1816 // Make 'use' use the Phi instead of the old loop body exit value 1817 _igvn.replace_input_of(use, idx, phi); 1818 if( use->_idx >= new_counter ) { // If updating new phis 1819 // Not needed for correctness, but prevents a weak assert 1820 // in AddPNode from tripping (when we end up with different 1821 // base & derived Phis that will become the same after 1822 // IGVN does CSE). 1823 Node *hit = _igvn.hash_find_insert(use); 1824 if( hit ) // Go ahead and re-hash for hits. 1825 _igvn.replace_node( use, hit ); 1826 } 1827 1828 // If 'use' was in the loop-exit block, it now needs to be sunk 1829 // below the post-loop merge point. 1830 sink_use( use, prev ); 1831 } 1832 } 1833 } 1834 1835 static void clone_outer_loop_helper(Node* n, const IdealLoopTree *loop, const IdealLoopTree* outer_loop, 1836 const Node_List &old_new, Unique_Node_List& wq, PhaseIdealLoop* phase, 1837 bool check_old_new) { 1838 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1839 Node* u = n->fast_out(j); 1840 assert(check_old_new || old_new[u->_idx] == NULL, "shouldn't have been cloned"); 1841 if (!u->is_CFG() && (!check_old_new || old_new[u->_idx] == NULL)) { 1842 Node* c = phase->get_ctrl(u); 1843 IdealLoopTree* u_loop = phase->get_loop(c); 1844 assert(!loop->is_member(u_loop), "can be in outer loop or out of both loops only"); 1845 if (outer_loop->is_member(u_loop)) { 1846 wq.push(u); 1847 } 1848 } 1849 } 1850 } 1851 1852 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop, 1853 IdealLoopTree* outer_loop, int dd, Node_List &old_new, 1854 Node_List& extra_data_nodes) { 1855 if (head->is_strip_mined() && mode != IgnoreStripMined) { 1856 CountedLoopNode* cl = head->as_CountedLoop(); 1857 Node* l = cl->outer_loop(); 1858 Node* tail = cl->outer_loop_tail(); 1859 IfNode* le = cl->outer_loop_end(); 1860 Node* sfpt = cl->outer_safepoint(); 1861 CountedLoopEndNode* cle = cl->loopexit(); 1862 CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop(); 1863 CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null(); 1864 Node* cle_out = cle->proj_out(false); 1865 1866 Node* new_sfpt = NULL; 1867 Node* new_cle_out = cle_out->clone(); 1868 old_new.map(cle_out->_idx, new_cle_out); 1869 if (mode == CloneIncludesStripMined) { 1870 // clone outer loop body 1871 Node* new_l = l->clone(); 1872 Node* new_tail = tail->clone(); 1873 IfNode* new_le = le->clone()->as_If(); 1874 new_sfpt = sfpt->clone(); 1875 1876 set_loop(new_l, outer_loop->_parent); 1877 set_idom(new_l, new_l->in(LoopNode::EntryControl), dd); 1878 set_loop(new_cle_out, outer_loop->_parent); 1879 set_idom(new_cle_out, new_cle, dd); 1880 set_loop(new_sfpt, outer_loop->_parent); 1881 set_idom(new_sfpt, new_cle_out, dd); 1882 set_loop(new_le, outer_loop->_parent); 1883 set_idom(new_le, new_sfpt, dd); 1884 set_loop(new_tail, outer_loop->_parent); 1885 set_idom(new_tail, new_le, dd); 1886 set_idom(new_cl, new_l, dd); 1887 1888 old_new.map(l->_idx, new_l); 1889 old_new.map(tail->_idx, new_tail); 1890 old_new.map(le->_idx, new_le); 1891 old_new.map(sfpt->_idx, new_sfpt); 1892 1893 new_l->set_req(LoopNode::LoopBackControl, new_tail); 1894 new_l->set_req(0, new_l); 1895 new_tail->set_req(0, new_le); 1896 new_le->set_req(0, new_sfpt); 1897 new_sfpt->set_req(0, new_cle_out); 1898 new_cle_out->set_req(0, new_cle); 1899 new_cl->set_req(LoopNode::EntryControl, new_l); 1900 1901 _igvn.register_new_node_with_optimizer(new_l); 1902 _igvn.register_new_node_with_optimizer(new_tail); 1903 _igvn.register_new_node_with_optimizer(new_le); 1904 } else { 1905 Node *newhead = old_new[loop->_head->_idx]; 1906 newhead->as_Loop()->clear_strip_mined(); 1907 _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl)); 1908 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 1909 } 1910 // Look at data node that were assigned a control in the outer 1911 // loop: they are kept in the outer loop by the safepoint so start 1912 // from the safepoint node's inputs. 1913 IdealLoopTree* outer_loop = get_loop(l); 1914 Node_Stack stack(2); 1915 stack.push(sfpt, 1); 1916 uint new_counter = C->unique(); 1917 while (stack.size() > 0) { 1918 Node* n = stack.node(); 1919 uint i = stack.index(); 1920 while (i < n->req() && 1921 (n->in(i) == NULL || 1922 !has_ctrl(n->in(i)) || 1923 get_loop(get_ctrl(n->in(i))) != outer_loop || 1924 (old_new[n->in(i)->_idx] != NULL && old_new[n->in(i)->_idx]->_idx >= new_counter))) { 1925 i++; 1926 } 1927 if (i < n->req()) { 1928 stack.set_index(i+1); 1929 stack.push(n->in(i), 0); 1930 } else { 1931 assert(old_new[n->_idx] == NULL || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet"); 1932 Node* m = n == sfpt ? new_sfpt : n->clone(); 1933 if (m != NULL) { 1934 for (uint i = 0; i < n->req(); i++) { 1935 if (m->in(i) != NULL && old_new[m->in(i)->_idx] != NULL) { 1936 m->set_req(i, old_new[m->in(i)->_idx]); 1937 } 1938 } 1939 } else { 1940 assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?"); 1941 } 1942 if (n != sfpt) { 1943 extra_data_nodes.push(n); 1944 _igvn.register_new_node_with_optimizer(m); 1945 assert(get_ctrl(n) == cle_out, "what other control?"); 1946 set_ctrl(m, new_cle_out); 1947 old_new.map(n->_idx, m); 1948 } 1949 stack.pop(); 1950 } 1951 } 1952 if (mode == CloneIncludesStripMined) { 1953 _igvn.register_new_node_with_optimizer(new_sfpt); 1954 _igvn.register_new_node_with_optimizer(new_cle_out); 1955 } 1956 // Some other transformation may have pessimistically assign some 1957 // data nodes to the outer loop. Set their control so they are out 1958 // of the outer loop. 1959 ResourceMark rm; 1960 Unique_Node_List wq; 1961 for (uint i = 0; i < extra_data_nodes.size(); i++) { 1962 Node* old = extra_data_nodes.at(i); 1963 clone_outer_loop_helper(old, loop, outer_loop, old_new, wq, this, true); 1964 } 1965 Node* new_ctrl = cl->outer_loop_exit(); 1966 assert(get_loop(new_ctrl) != outer_loop, "must be out of the loop nest"); 1967 for (uint i = 0; i < wq.size(); i++) { 1968 Node* n = wq.at(i); 1969 set_ctrl(n, new_ctrl); 1970 clone_outer_loop_helper(n, loop, outer_loop, old_new, wq, this, false); 1971 } 1972 } else { 1973 Node *newhead = old_new[loop->_head->_idx]; 1974 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 1975 } 1976 } 1977 1978 //------------------------------clone_loop------------------------------------- 1979 // 1980 // C L O N E A L O O P B O D Y 1981 // 1982 // This is the basic building block of the loop optimizations. It clones an 1983 // entire loop body. It makes an old_new loop body mapping; with this mapping 1984 // you can find the new-loop equivalent to an old-loop node. All new-loop 1985 // nodes are exactly equal to their old-loop counterparts, all edges are the 1986 // same. All exits from the old-loop now have a RegionNode that merges the 1987 // equivalent new-loop path. This is true even for the normal "loop-exit" 1988 // condition. All uses of loop-invariant old-loop values now come from (one 1989 // or more) Phis that merge their new-loop equivalents. 1990 // 1991 // This operation leaves the graph in an illegal state: there are two valid 1992 // control edges coming from the loop pre-header to both loop bodies. I'll 1993 // definitely have to hack the graph after running this transform. 1994 // 1995 // From this building block I will further edit edges to perform loop peeling 1996 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc. 1997 // 1998 // Parameter side_by_size_idom: 1999 // When side_by_size_idom is NULL, the dominator tree is constructed for 2000 // the clone loop to dominate the original. Used in construction of 2001 // pre-main-post loop sequence. 2002 // When nonnull, the clone and original are side-by-side, both are 2003 // dominated by the side_by_side_idom node. Used in construction of 2004 // unswitched loops. 2005 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd, 2006 CloneLoopMode mode, Node* side_by_side_idom) { 2007 2008 LoopNode* head = loop->_head->as_Loop(); 2009 head->verify_strip_mined(1); 2010 2011 if (C->do_vector_loop() && PrintOpto) { 2012 const char* mname = C->method()->name()->as_quoted_ascii(); 2013 if (mname != NULL) { 2014 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname); 2015 } 2016 } 2017 2018 CloneMap& cm = C->clone_map(); 2019 Dict* dict = cm.dict(); 2020 if (C->do_vector_loop()) { 2021 cm.set_clone_idx(cm.max_gen()+1); 2022 #ifndef PRODUCT 2023 if (PrintOpto) { 2024 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx()); 2025 loop->dump_head(); 2026 } 2027 #endif 2028 } 2029 2030 // Step 1: Clone the loop body. Make the old->new mapping. 2031 uint i; 2032 for( i = 0; i < loop->_body.size(); i++ ) { 2033 Node *old = loop->_body.at(i); 2034 Node *nnn = old->clone(); 2035 old_new.map( old->_idx, nnn ); 2036 if (C->do_vector_loop()) { 2037 cm.verify_insert_and_clone(old, nnn, cm.clone_idx()); 2038 } 2039 _igvn.register_new_node_with_optimizer(nnn); 2040 } 2041 2042 IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop; 2043 2044 // Step 2: Fix the edges in the new body. If the old input is outside the 2045 // loop use it. If the old input is INside the loop, use the corresponding 2046 // new node instead. 2047 for( i = 0; i < loop->_body.size(); i++ ) { 2048 Node *old = loop->_body.at(i); 2049 Node *nnn = old_new[old->_idx]; 2050 // Fix CFG/Loop controlling the new node 2051 if (has_ctrl(old)) { 2052 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]); 2053 } else { 2054 set_loop(nnn, outer_loop->_parent); 2055 if (old->outcnt() > 0) { 2056 set_idom( nnn, old_new[idom(old)->_idx], dd ); 2057 } 2058 } 2059 // Correct edges to the new node 2060 for( uint j = 0; j < nnn->req(); j++ ) { 2061 Node *n = nnn->in(j); 2062 if( n ) { 2063 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n ); 2064 if( loop->is_member( old_in_loop ) ) 2065 nnn->set_req(j, old_new[n->_idx]); 2066 } 2067 } 2068 _igvn.hash_find_insert(nnn); 2069 } 2070 2071 ResourceArea *area = Thread::current()->resource_area(); 2072 Node_List extra_data_nodes(area); // data nodes in the outer strip mined loop 2073 clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes); 2074 2075 // Step 3: Now fix control uses. Loop varying control uses have already 2076 // been fixed up (as part of all input edges in Step 2). Loop invariant 2077 // control uses must be either an IfFalse or an IfTrue. Make a merge 2078 // point to merge the old and new IfFalse/IfTrue nodes; make the use 2079 // refer to this. 2080 Node_List worklist(area); 2081 uint new_counter = C->unique(); 2082 for( i = 0; i < loop->_body.size(); i++ ) { 2083 Node* old = loop->_body.at(i); 2084 if( !old->is_CFG() ) continue; 2085 2086 // Copy uses to a worklist, so I can munge the def-use info 2087 // with impunity. 2088 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 2089 worklist.push(old->fast_out(j)); 2090 2091 while( worklist.size() ) { // Visit all uses 2092 Node *use = worklist.pop(); 2093 if (!has_node(use)) continue; // Ignore dead nodes 2094 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 2095 if( !loop->is_member( use_loop ) && use->is_CFG() ) { 2096 // Both OLD and USE are CFG nodes here. 2097 assert( use->is_Proj(), "" ); 2098 Node* nnn = old_new[old->_idx]; 2099 2100 Node* newuse = NULL; 2101 if (head->is_strip_mined() && mode != IgnoreStripMined) { 2102 CountedLoopNode* cl = head->as_CountedLoop(); 2103 CountedLoopEndNode* cle = cl->loopexit(); 2104 Node* cle_out = cle->proj_out_or_null(false); 2105 if (use == cle_out) { 2106 IfNode* le = cl->outer_loop_end(); 2107 use = le->proj_out(false); 2108 use_loop = get_loop(use); 2109 if (mode == CloneIncludesStripMined) { 2110 nnn = old_new[le->_idx]; 2111 } else { 2112 newuse = old_new[cle_out->_idx]; 2113 } 2114 } 2115 } 2116 if (newuse == NULL) { 2117 newuse = use->clone(); 2118 } 2119 2120 // Clone the loop exit control projection 2121 if (C->do_vector_loop()) { 2122 cm.verify_insert_and_clone(use, newuse, cm.clone_idx()); 2123 } 2124 newuse->set_req(0,nnn); 2125 _igvn.register_new_node_with_optimizer(newuse); 2126 set_loop(newuse, use_loop); 2127 set_idom(newuse, nnn, dom_depth(nnn) + 1 ); 2128 2129 // We need a Region to merge the exit from the peeled body and the 2130 // exit from the old loop body. 2131 RegionNode *r = new RegionNode(3); 2132 // Map the old use to the new merge point 2133 old_new.map( use->_idx, r ); 2134 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use)); 2135 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" ); 2136 2137 // The original user of 'use' uses 'r' instead. 2138 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) { 2139 Node* useuse = use->last_out(l); 2140 _igvn.rehash_node_delayed(useuse); 2141 uint uses_found = 0; 2142 if( useuse->in(0) == use ) { 2143 useuse->set_req(0, r); 2144 uses_found++; 2145 if( useuse->is_CFG() ) { 2146 assert( dom_depth(useuse) > dd_r, "" ); 2147 set_idom(useuse, r, dom_depth(useuse)); 2148 } 2149 } 2150 for( uint k = 1; k < useuse->req(); k++ ) { 2151 if( useuse->in(k) == use ) { 2152 useuse->set_req(k, r); 2153 uses_found++; 2154 if (useuse->is_Loop() && k == LoopNode::EntryControl) { 2155 assert(dom_depth(useuse) > dd_r , ""); 2156 set_idom(useuse, r, dom_depth(useuse)); 2157 } 2158 } 2159 } 2160 l -= uses_found; // we deleted 1 or more copies of this edge 2161 } 2162 2163 // Now finish up 'r' 2164 r->set_req( 1, newuse ); 2165 r->set_req( 2, use ); 2166 _igvn.register_new_node_with_optimizer(r); 2167 set_loop(r, use_loop); 2168 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r); 2169 } // End of if a loop-exit test 2170 } 2171 } 2172 2173 // Step 4: If loop-invariant use is not control, it must be dominated by a 2174 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region 2175 // there if needed. Make a Phi there merging old and new used values. 2176 Node_List *split_if_set = NULL; 2177 Node_List *split_bool_set = NULL; 2178 Node_List *split_cex_set = NULL; 2179 for( i = 0; i < loop->_body.size(); i++ ) { 2180 Node* old = loop->_body.at(i); 2181 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set, 2182 split_bool_set, split_cex_set, worklist, new_counter, 2183 mode); 2184 } 2185 2186 for (i = 0; i < extra_data_nodes.size(); i++) { 2187 Node* old = extra_data_nodes.at(i); 2188 clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set, 2189 split_bool_set, split_cex_set, worklist, new_counter, 2190 mode); 2191 } 2192 2193 // Check for IFs that need splitting/cloning. Happens if an IF outside of 2194 // the loop uses a condition set in the loop. The original IF probably 2195 // takes control from one or more OLD Regions (which in turn get from NEW 2196 // Regions). In any case, there will be a set of Phis for each merge point 2197 // from the IF up to where the original BOOL def exists the loop. 2198 if (split_if_set) { 2199 while (split_if_set->size()) { 2200 Node *iff = split_if_set->pop(); 2201 if (iff->in(1)->is_Phi()) { 2202 Node *b = clone_iff(iff->in(1)->as_Phi(), loop); 2203 _igvn.replace_input_of(iff, 1, b); 2204 } 2205 } 2206 } 2207 if (split_bool_set) { 2208 while (split_bool_set->size()) { 2209 Node *b = split_bool_set->pop(); 2210 Node *phi = b->in(1); 2211 assert(phi->is_Phi(), ""); 2212 CmpNode *cmp = clone_bool((PhiNode*)phi, loop); 2213 _igvn.replace_input_of(b, 1, cmp); 2214 } 2215 } 2216 if (split_cex_set) { 2217 while (split_cex_set->size()) { 2218 Node *b = split_cex_set->pop(); 2219 assert(b->in(0)->is_Region(), ""); 2220 assert(b->in(1)->is_Phi(), ""); 2221 assert(b->in(0)->in(0) == b->in(1)->in(0), ""); 2222 split_up(b, b->in(0), NULL); 2223 } 2224 } 2225 2226 } 2227 2228 2229 //---------------------- stride_of_possible_iv ------------------------------------- 2230 // Looks for an iff/bool/comp with one operand of the compare 2231 // being a cycle involving an add and a phi, 2232 // with an optional truncation (left-shift followed by a right-shift) 2233 // of the add. Returns zero if not an iv. 2234 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) { 2235 Node* trunc1 = NULL; 2236 Node* trunc2 = NULL; 2237 const TypeInt* ttype = NULL; 2238 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) { 2239 return 0; 2240 } 2241 BoolNode* bl = iff->in(1)->as_Bool(); 2242 Node* cmp = bl->in(1); 2243 if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) { 2244 return 0; 2245 } 2246 // Must have an invariant operand 2247 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) { 2248 return 0; 2249 } 2250 Node* add2 = NULL; 2251 Node* cmp1 = cmp->in(1); 2252 if (cmp1->is_Phi()) { 2253 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) ))) 2254 Node* phi = cmp1; 2255 for (uint i = 1; i < phi->req(); i++) { 2256 Node* in = phi->in(i); 2257 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in, 2258 &trunc1, &trunc2, &ttype); 2259 if (add && add->in(1) == phi) { 2260 add2 = add->in(2); 2261 break; 2262 } 2263 } 2264 } else { 2265 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) ))) 2266 Node* addtrunc = cmp1; 2267 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc, 2268 &trunc1, &trunc2, &ttype); 2269 if (add && add->in(1)->is_Phi()) { 2270 Node* phi = add->in(1); 2271 for (uint i = 1; i < phi->req(); i++) { 2272 if (phi->in(i) == addtrunc) { 2273 add2 = add->in(2); 2274 break; 2275 } 2276 } 2277 } 2278 } 2279 if (add2 != NULL) { 2280 const TypeInt* add2t = _igvn.type(add2)->is_int(); 2281 if (add2t->is_con()) { 2282 return add2t->get_con(); 2283 } 2284 } 2285 return 0; 2286 } 2287 2288 2289 //---------------------- stay_in_loop ------------------------------------- 2290 // Return the (unique) control output node that's in the loop (if it exists.) 2291 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) { 2292 Node* unique = NULL; 2293 if (!n) return NULL; 2294 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2295 Node* use = n->fast_out(i); 2296 if (!has_ctrl(use) && loop->is_member(get_loop(use))) { 2297 if (unique != NULL) { 2298 return NULL; 2299 } 2300 unique = use; 2301 } 2302 } 2303 return unique; 2304 } 2305 2306 //------------------------------ register_node ------------------------------------- 2307 // Utility to register node "n" with PhaseIdealLoop 2308 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) { 2309 _igvn.register_new_node_with_optimizer(n); 2310 loop->_body.push(n); 2311 if (n->is_CFG()) { 2312 set_loop(n, loop); 2313 set_idom(n, pred, ddepth); 2314 } else { 2315 set_ctrl(n, pred); 2316 } 2317 } 2318 2319 //------------------------------ proj_clone ------------------------------------- 2320 // Utility to create an if-projection 2321 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) { 2322 ProjNode* c = p->clone()->as_Proj(); 2323 c->set_req(0, iff); 2324 return c; 2325 } 2326 2327 //------------------------------ short_circuit_if ------------------------------------- 2328 // Force the iff control output to be the live_proj 2329 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) { 2330 guarantee(live_proj != NULL, "null projection"); 2331 int proj_con = live_proj->_con; 2332 assert(proj_con == 0 || proj_con == 1, "false or true projection"); 2333 Node *con = _igvn.intcon(proj_con); 2334 set_ctrl(con, C->root()); 2335 if (iff) { 2336 iff->set_req(1, con); 2337 } 2338 return con; 2339 } 2340 2341 //------------------------------ insert_if_before_proj ------------------------------------- 2342 // Insert a new if before an if projection (* - new node) 2343 // 2344 // before 2345 // if(test) 2346 // / \ 2347 // v v 2348 // other-proj proj (arg) 2349 // 2350 // after 2351 // if(test) 2352 // / \ 2353 // / v 2354 // | * proj-clone 2355 // v | 2356 // other-proj v 2357 // * new_if(relop(cmp[IU](left,right))) 2358 // / \ 2359 // v v 2360 // * new-proj proj 2361 // (returned) 2362 // 2363 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) { 2364 IfNode* iff = proj->in(0)->as_If(); 2365 IdealLoopTree *loop = get_loop(proj); 2366 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2367 int ddepth = dom_depth(proj); 2368 2369 _igvn.rehash_node_delayed(iff); 2370 _igvn.rehash_node_delayed(proj); 2371 2372 proj->set_req(0, NULL); // temporary disconnect 2373 ProjNode* proj2 = proj_clone(proj, iff); 2374 register_node(proj2, loop, iff, ddepth); 2375 2376 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right); 2377 register_node(cmp, loop, proj2, ddepth); 2378 2379 BoolNode* bol = new BoolNode(cmp, relop); 2380 register_node(bol, loop, proj2, ddepth); 2381 2382 int opcode = iff->Opcode(); 2383 assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode"); 2384 IfNode* new_if = (opcode == Op_If) ? new IfNode(proj2, bol, iff->_prob, iff->_fcnt): 2385 new RangeCheckNode(proj2, bol, iff->_prob, iff->_fcnt); 2386 register_node(new_if, loop, proj2, ddepth); 2387 2388 proj->set_req(0, new_if); // reattach 2389 set_idom(proj, new_if, ddepth); 2390 2391 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj(); 2392 guarantee(new_exit != NULL, "null exit node"); 2393 register_node(new_exit, get_loop(other_proj), new_if, ddepth); 2394 2395 return new_exit; 2396 } 2397 2398 //------------------------------ insert_region_before_proj ------------------------------------- 2399 // Insert a region before an if projection (* - new node) 2400 // 2401 // before 2402 // if(test) 2403 // / | 2404 // v | 2405 // proj v 2406 // other-proj 2407 // 2408 // after 2409 // if(test) 2410 // / | 2411 // v | 2412 // * proj-clone v 2413 // | other-proj 2414 // v 2415 // * new-region 2416 // | 2417 // v 2418 // * dum_if 2419 // / \ 2420 // v \ 2421 // * dum-proj v 2422 // proj 2423 // 2424 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) { 2425 IfNode* iff = proj->in(0)->as_If(); 2426 IdealLoopTree *loop = get_loop(proj); 2427 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2428 int ddepth = dom_depth(proj); 2429 2430 _igvn.rehash_node_delayed(iff); 2431 _igvn.rehash_node_delayed(proj); 2432 2433 proj->set_req(0, NULL); // temporary disconnect 2434 ProjNode* proj2 = proj_clone(proj, iff); 2435 register_node(proj2, loop, iff, ddepth); 2436 2437 RegionNode* reg = new RegionNode(2); 2438 reg->set_req(1, proj2); 2439 register_node(reg, loop, iff, ddepth); 2440 2441 IfNode* dum_if = new IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt); 2442 register_node(dum_if, loop, reg, ddepth); 2443 2444 proj->set_req(0, dum_if); // reattach 2445 set_idom(proj, dum_if, ddepth); 2446 2447 ProjNode* dum_proj = proj_clone(other_proj, dum_if); 2448 register_node(dum_proj, loop, dum_if, ddepth); 2449 2450 return reg; 2451 } 2452 2453 //------------------------------ insert_cmpi_loop_exit ------------------------------------- 2454 // Clone a signed compare loop exit from an unsigned compare and 2455 // insert it before the unsigned cmp on the stay-in-loop path. 2456 // All new nodes inserted in the dominator tree between the original 2457 // if and it's projections. The original if test is replaced with 2458 // a constant to force the stay-in-loop path. 2459 // 2460 // This is done to make sure that the original if and it's projections 2461 // still dominate the same set of control nodes, that the ctrl() relation 2462 // from data nodes to them is preserved, and that their loop nesting is 2463 // preserved. 2464 // 2465 // before 2466 // if(i <u limit) unsigned compare loop exit 2467 // / | 2468 // v v 2469 // exit-proj stay-in-loop-proj 2470 // 2471 // after 2472 // if(stay-in-loop-const) original if 2473 // / | 2474 // / v 2475 // / if(i < limit) new signed test 2476 // / / | 2477 // / / v 2478 // / / if(i <u limit) new cloned unsigned test 2479 // / / / | 2480 // v v v | 2481 // region | 2482 // | | 2483 // dum-if | 2484 // / | | 2485 // ether | | 2486 // v v 2487 // exit-proj stay-in-loop-proj 2488 // 2489 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) { 2490 const bool Signed = true; 2491 const bool Unsigned = false; 2492 2493 BoolNode* bol = if_cmpu->in(1)->as_Bool(); 2494 if (bol->_test._test != BoolTest::lt) return NULL; 2495 CmpNode* cmpu = bol->in(1)->as_Cmp(); 2496 if (cmpu->Opcode() != Op_CmpU) return NULL; 2497 int stride = stride_of_possible_iv(if_cmpu); 2498 if (stride == 0) return NULL; 2499 2500 Node* lp_proj = stay_in_loop(if_cmpu, loop); 2501 guarantee(lp_proj != NULL, "null loop node"); 2502 2503 ProjNode* lp_continue = lp_proj->as_Proj(); 2504 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj(); 2505 2506 Node* limit = NULL; 2507 if (stride > 0) { 2508 limit = cmpu->in(2); 2509 } else { 2510 limit = _igvn.makecon(TypeInt::ZERO); 2511 set_ctrl(limit, C->root()); 2512 } 2513 // Create a new region on the exit path 2514 RegionNode* reg = insert_region_before_proj(lp_exit); 2515 guarantee(reg != NULL, "null region node"); 2516 2517 // Clone the if-cmpu-true-false using a signed compare 2518 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge; 2519 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue); 2520 reg->add_req(cmpi_exit); 2521 2522 // Clone the if-cmpu-true-false 2523 BoolTest::mask rel_u = bol->_test._test; 2524 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue); 2525 reg->add_req(cmpu_exit); 2526 2527 // Force original if to stay in loop. 2528 short_circuit_if(if_cmpu, lp_continue); 2529 2530 return cmpi_exit->in(0)->as_If(); 2531 } 2532 2533 //------------------------------ remove_cmpi_loop_exit ------------------------------------- 2534 // Remove a previously inserted signed compare loop exit. 2535 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) { 2536 Node* lp_proj = stay_in_loop(if_cmp, loop); 2537 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI && 2538 stay_in_loop(lp_proj, loop)->is_If() && 2539 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu"); 2540 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO); 2541 set_ctrl(con, C->root()); 2542 if_cmp->set_req(1, con); 2543 } 2544 2545 //------------------------------ scheduled_nodelist ------------------------------------- 2546 // Create a post order schedule of nodes that are in the 2547 // "member" set. The list is returned in "sched". 2548 // The first node in "sched" is the loop head, followed by 2549 // nodes which have no inputs in the "member" set, and then 2550 // followed by the nodes that have an immediate input dependence 2551 // on a node in "sched". 2552 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) { 2553 2554 assert(member.test(loop->_head->_idx), "loop head must be in member set"); 2555 Arena *a = Thread::current()->resource_area(); 2556 VectorSet visited(a); 2557 Node_Stack nstack(a, loop->_body.size()); 2558 2559 Node* n = loop->_head; // top of stack is cached in "n" 2560 uint idx = 0; 2561 visited.set(n->_idx); 2562 2563 // Initially push all with no inputs from within member set 2564 for(uint i = 0; i < loop->_body.size(); i++ ) { 2565 Node *elt = loop->_body.at(i); 2566 if (member.test(elt->_idx)) { 2567 bool found = false; 2568 for (uint j = 0; j < elt->req(); j++) { 2569 Node* def = elt->in(j); 2570 if (def && member.test(def->_idx) && def != elt) { 2571 found = true; 2572 break; 2573 } 2574 } 2575 if (!found && elt != loop->_head) { 2576 nstack.push(n, idx); 2577 n = elt; 2578 assert(!visited.test(n->_idx), "not seen yet"); 2579 visited.set(n->_idx); 2580 } 2581 } 2582 } 2583 2584 // traverse out's that are in the member set 2585 while (true) { 2586 if (idx < n->outcnt()) { 2587 Node* use = n->raw_out(idx); 2588 idx++; 2589 if (!visited.test_set(use->_idx)) { 2590 if (member.test(use->_idx)) { 2591 nstack.push(n, idx); 2592 n = use; 2593 idx = 0; 2594 } 2595 } 2596 } else { 2597 // All outputs processed 2598 sched.push(n); 2599 if (nstack.is_empty()) break; 2600 n = nstack.node(); 2601 idx = nstack.index(); 2602 nstack.pop(); 2603 } 2604 } 2605 } 2606 2607 2608 //------------------------------ has_use_in_set ------------------------------------- 2609 // Has a use in the vector set 2610 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) { 2611 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2612 Node* use = n->fast_out(j); 2613 if (vset.test(use->_idx)) { 2614 return true; 2615 } 2616 } 2617 return false; 2618 } 2619 2620 2621 //------------------------------ has_use_internal_to_set ------------------------------------- 2622 // Has use internal to the vector set (ie. not in a phi at the loop head) 2623 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) { 2624 Node* head = loop->_head; 2625 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2626 Node* use = n->fast_out(j); 2627 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) { 2628 return true; 2629 } 2630 } 2631 return false; 2632 } 2633 2634 2635 //------------------------------ clone_for_use_outside_loop ------------------------------------- 2636 // clone "n" for uses that are outside of loop 2637 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) { 2638 int cloned = 0; 2639 assert(worklist.size() == 0, "should be empty"); 2640 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2641 Node* use = n->fast_out(j); 2642 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) { 2643 worklist.push(use); 2644 } 2645 } 2646 while( worklist.size() ) { 2647 Node *use = worklist.pop(); 2648 if (!has_node(use) || use->in(0) == C->top()) continue; 2649 uint j; 2650 for (j = 0; j < use->req(); j++) { 2651 if (use->in(j) == n) break; 2652 } 2653 assert(j < use->req(), "must be there"); 2654 2655 // clone "n" and insert it between the inputs of "n" and the use outside the loop 2656 Node* n_clone = n->clone(); 2657 _igvn.replace_input_of(use, j, n_clone); 2658 cloned++; 2659 Node* use_c; 2660 if (!use->is_Phi()) { 2661 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0); 2662 } else { 2663 // Use in a phi is considered a use in the associated predecessor block 2664 use_c = use->in(0)->in(j); 2665 } 2666 set_ctrl(n_clone, use_c); 2667 assert(!loop->is_member(get_loop(use_c)), "should be outside loop"); 2668 get_loop(use_c)->_body.push(n_clone); 2669 _igvn.register_new_node_with_optimizer(n_clone); 2670 #if !defined(PRODUCT) 2671 if (TracePartialPeeling) { 2672 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx); 2673 } 2674 #endif 2675 } 2676 return cloned; 2677 } 2678 2679 2680 //------------------------------ clone_for_special_use_inside_loop ------------------------------------- 2681 // clone "n" for special uses that are in the not_peeled region. 2682 // If these def-uses occur in separate blocks, the code generator 2683 // marks the method as not compilable. For example, if a "BoolNode" 2684 // is in a different basic block than the "IfNode" that uses it, then 2685 // the compilation is aborted in the code generator. 2686 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 2687 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) { 2688 if (n->is_Phi() || n->is_Load()) { 2689 return; 2690 } 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 ( not_peel.test(use->_idx) && 2695 (use->is_If() || use->is_CMove() || use->is_Bool()) && 2696 use->in(1) == n) { 2697 worklist.push(use); 2698 } 2699 } 2700 if (worklist.size() > 0) { 2701 // clone "n" and insert it between inputs of "n" and the use 2702 Node* n_clone = n->clone(); 2703 loop->_body.push(n_clone); 2704 _igvn.register_new_node_with_optimizer(n_clone); 2705 set_ctrl(n_clone, get_ctrl(n)); 2706 sink_list.push(n_clone); 2707 not_peel <<= n_clone->_idx; // add n_clone to not_peel set. 2708 #if !defined(PRODUCT) 2709 if (TracePartialPeeling) { 2710 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx); 2711 } 2712 #endif 2713 while( worklist.size() ) { 2714 Node *use = worklist.pop(); 2715 _igvn.rehash_node_delayed(use); 2716 for (uint j = 1; j < use->req(); j++) { 2717 if (use->in(j) == n) { 2718 use->set_req(j, n_clone); 2719 } 2720 } 2721 } 2722 } 2723 } 2724 2725 2726 //------------------------------ insert_phi_for_loop ------------------------------------- 2727 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 2728 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) { 2729 Node *phi = PhiNode::make(lp, back_edge_val); 2730 phi->set_req(LoopNode::EntryControl, lp_entry_val); 2731 // Use existing phi if it already exists 2732 Node *hit = _igvn.hash_find_insert(phi); 2733 if( hit == NULL ) { 2734 _igvn.register_new_node_with_optimizer(phi); 2735 set_ctrl(phi, lp); 2736 } else { 2737 // Remove the new phi from the graph and use the hit 2738 _igvn.remove_dead_node(phi); 2739 phi = hit; 2740 } 2741 _igvn.replace_input_of(use, idx, phi); 2742 } 2743 2744 #ifdef ASSERT 2745 //------------------------------ is_valid_loop_partition ------------------------------------- 2746 // Validate the loop partition sets: peel and not_peel 2747 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, 2748 VectorSet& not_peel ) { 2749 uint i; 2750 // Check that peel_list entries are in the peel set 2751 for (i = 0; i < peel_list.size(); i++) { 2752 if (!peel.test(peel_list.at(i)->_idx)) { 2753 return false; 2754 } 2755 } 2756 // Check at loop members are in one of peel set or not_peel set 2757 for (i = 0; i < loop->_body.size(); i++ ) { 2758 Node *def = loop->_body.at(i); 2759 uint di = def->_idx; 2760 // Check that peel set elements are in peel_list 2761 if (peel.test(di)) { 2762 if (not_peel.test(di)) { 2763 return false; 2764 } 2765 // Must be in peel_list also 2766 bool found = false; 2767 for (uint j = 0; j < peel_list.size(); j++) { 2768 if (peel_list.at(j)->_idx == di) { 2769 found = true; 2770 break; 2771 } 2772 } 2773 if (!found) { 2774 return false; 2775 } 2776 } else if (not_peel.test(di)) { 2777 if (peel.test(di)) { 2778 return false; 2779 } 2780 } else { 2781 return false; 2782 } 2783 } 2784 return true; 2785 } 2786 2787 //------------------------------ is_valid_clone_loop_exit_use ------------------------------------- 2788 // Ensure a use outside of loop is of the right form 2789 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) { 2790 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2791 return (use->is_Phi() && 2792 use_c->is_Region() && use_c->req() == 3 && 2793 (use_c->in(exit_idx)->Opcode() == Op_IfTrue || 2794 use_c->in(exit_idx)->Opcode() == Op_IfFalse || 2795 use_c->in(exit_idx)->Opcode() == Op_JumpProj) && 2796 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) ); 2797 } 2798 2799 //------------------------------ is_valid_clone_loop_form ------------------------------------- 2800 // Ensure that all uses outside of loop are of the right form 2801 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 2802 uint orig_exit_idx, uint clone_exit_idx) { 2803 uint len = peel_list.size(); 2804 for (uint i = 0; i < len; i++) { 2805 Node *def = peel_list.at(i); 2806 2807 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 2808 Node *use = def->fast_out(j); 2809 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2810 if (!loop->is_member(get_loop(use_c))) { 2811 // use is not in the loop, check for correct structure 2812 if (use->in(0) == def) { 2813 // Okay 2814 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) { 2815 return false; 2816 } 2817 } 2818 } 2819 } 2820 return true; 2821 } 2822 #endif 2823 2824 //------------------------------ partial_peel ------------------------------------- 2825 // Partially peel (aka loop rotation) the top portion of a loop (called 2826 // the peel section below) by cloning it and placing one copy just before 2827 // the new loop head and the other copy at the bottom of the new loop. 2828 // 2829 // before after where it came from 2830 // 2831 // stmt1 stmt1 2832 // loop: stmt2 clone 2833 // stmt2 if condA goto exitA clone 2834 // if condA goto exitA new_loop: new 2835 // stmt3 stmt3 clone 2836 // if !condB goto loop if condB goto exitB clone 2837 // exitB: stmt2 orig 2838 // stmt4 if !condA goto new_loop orig 2839 // exitA: goto exitA 2840 // exitB: 2841 // stmt4 2842 // exitA: 2843 // 2844 // Step 1: find the cut point: an exit test on probable 2845 // induction variable. 2846 // Step 2: schedule (with cloning) operations in the peel 2847 // section that can be executed after the cut into 2848 // the section that is not peeled. This may need 2849 // to clone operations into exit blocks. For 2850 // instance, a reference to A[i] in the not-peel 2851 // section and a reference to B[i] in an exit block 2852 // may cause a left-shift of i by 2 to be placed 2853 // in the peel block. This step will clone the left 2854 // shift into the exit block and sink the left shift 2855 // from the peel to the not-peel section. 2856 // Step 3: clone the loop, retarget the control, and insert 2857 // phis for values that are live across the new loop 2858 // head. This is very dependent on the graph structure 2859 // from clone_loop. It creates region nodes for 2860 // exit control and associated phi nodes for values 2861 // flow out of the loop through that exit. The region 2862 // node is dominated by the clone's control projection. 2863 // So the clone's peel section is placed before the 2864 // new loop head, and the clone's not-peel section is 2865 // forms the top part of the new loop. The original 2866 // peel section forms the tail of the new loop. 2867 // Step 4: update the dominator tree and recompute the 2868 // dominator depth. 2869 // 2870 // orig 2871 // 2872 // stmt1 2873 // | 2874 // v 2875 // loop predicate 2876 // | 2877 // v 2878 // loop<----+ 2879 // | | 2880 // stmt2 | 2881 // | | 2882 // v | 2883 // ifA | 2884 // / | | 2885 // v v | 2886 // false true ^ <-- last_peel 2887 // / | | 2888 // / ===|==cut | 2889 // / stmt3 | <-- first_not_peel 2890 // / | | 2891 // | v | 2892 // v ifB | 2893 // exitA: / \ | 2894 // / \ | 2895 // v v | 2896 // false true | 2897 // / \ | 2898 // / ----+ 2899 // | 2900 // v 2901 // exitB: 2902 // stmt4 2903 // 2904 // 2905 // after clone loop 2906 // 2907 // stmt1 2908 // | 2909 // v 2910 // loop predicate 2911 // / \ 2912 // clone / \ orig 2913 // / \ 2914 // / \ 2915 // v v 2916 // +---->loop loop<----+ 2917 // | | | | 2918 // | stmt2 stmt2 | 2919 // | | | | 2920 // | v v | 2921 // | ifA ifA | 2922 // | | \ / | | 2923 // | v v v v | 2924 // ^ true false false true ^ <-- last_peel 2925 // | | ^ \ / | | 2926 // | cut==|== \ \ / ===|==cut | 2927 // | stmt3 \ \ / stmt3 | <-- first_not_peel 2928 // | | dom | | | | 2929 // | v \ 1v v2 v | 2930 // | ifB regionA ifB | 2931 // | / \ | / \ | 2932 // | / \ v / \ | 2933 // | v v exitA: v v | 2934 // | true false false true | 2935 // | / ^ \ / \ | 2936 // +---- \ \ / ----+ 2937 // dom \ / 2938 // \ 1v v2 2939 // regionB 2940 // | 2941 // v 2942 // exitB: 2943 // stmt4 2944 // 2945 // 2946 // after partial peel 2947 // 2948 // stmt1 2949 // | 2950 // v 2951 // loop predicate 2952 // / 2953 // clone / orig 2954 // / TOP 2955 // / \ 2956 // v v 2957 // TOP->loop loop----+ 2958 // | | | 2959 // stmt2 stmt2 | 2960 // | | | 2961 // v v | 2962 // ifA ifA | 2963 // | \ / | | 2964 // v v v v | 2965 // true false false true | <-- last_peel 2966 // | ^ \ / +------|---+ 2967 // +->newloop \ \ / === ==cut | | 2968 // | stmt3 \ \ / TOP | | 2969 // | | dom | | stmt3 | | <-- first_not_peel 2970 // | v \ 1v v2 v | | 2971 // | ifB regionA ifB ^ v 2972 // | / \ | / \ | | 2973 // | / \ v / \ | | 2974 // | v v exitA: v v | | 2975 // | true false false true | | 2976 // | / ^ \ / \ | | 2977 // | | \ \ / v | | 2978 // | | dom \ / TOP | | 2979 // | | \ 1v v2 | | 2980 // ^ v regionB | | 2981 // | | | | | 2982 // | | v ^ v 2983 // | | exitB: | | 2984 // | | stmt4 | | 2985 // | +------------>-----------------+ | 2986 // | | 2987 // +-----------------<---------------------+ 2988 // 2989 // 2990 // final graph 2991 // 2992 // stmt1 2993 // | 2994 // v 2995 // loop predicate 2996 // | 2997 // v 2998 // stmt2 clone 2999 // | 3000 // v 3001 // ........> ifA clone 3002 // : / | 3003 // dom / | 3004 // : v v 3005 // : false true 3006 // : | | 3007 // : | v 3008 // : | newloop<-----+ 3009 // : | | | 3010 // : | stmt3 clone | 3011 // : | | | 3012 // : | v | 3013 // : | ifB | 3014 // : | / \ | 3015 // : | v v | 3016 // : | false true | 3017 // : | | | | 3018 // : | v stmt2 | 3019 // : | exitB: | | 3020 // : | stmt4 v | 3021 // : | ifA orig | 3022 // : | / \ | 3023 // : | / \ | 3024 // : | v v | 3025 // : | false true | 3026 // : | / \ | 3027 // : v v -----+ 3028 // RegionA 3029 // | 3030 // v 3031 // exitA 3032 // 3033 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) { 3034 3035 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only"); 3036 if (!loop->_head->is_Loop()) { 3037 return false; } 3038 3039 LoopNode *head = loop->_head->as_Loop(); 3040 3041 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) { 3042 return false; 3043 } 3044 3045 // Check for complex exit control 3046 for(uint ii = 0; ii < loop->_body.size(); ii++ ) { 3047 Node *n = loop->_body.at(ii); 3048 int opc = n->Opcode(); 3049 if (n->is_Call() || 3050 opc == Op_Catch || 3051 opc == Op_CatchProj || 3052 opc == Op_Jump || 3053 opc == Op_JumpProj) { 3054 #if !defined(PRODUCT) 3055 if (TracePartialPeeling) { 3056 tty->print_cr("\nExit control too complex: lp: %d", head->_idx); 3057 } 3058 #endif 3059 return false; 3060 } 3061 } 3062 3063 int dd = dom_depth(head); 3064 3065 // Step 1: find cut point 3066 3067 // Walk up dominators to loop head looking for first loop exit 3068 // which is executed on every path thru loop. 3069 IfNode *peel_if = NULL; 3070 IfNode *peel_if_cmpu = NULL; 3071 3072 Node *iff = loop->tail(); 3073 while( iff != head ) { 3074 if( iff->is_If() ) { 3075 Node *ctrl = get_ctrl(iff->in(1)); 3076 if (ctrl->is_top()) return false; // Dead test on live IF. 3077 // If loop-varying exit-test, check for induction variable 3078 if( loop->is_member(get_loop(ctrl)) && 3079 loop->is_loop_exit(iff) && 3080 is_possible_iv_test(iff)) { 3081 Node* cmp = iff->in(1)->in(1); 3082 if (cmp->Opcode() == Op_CmpI) { 3083 peel_if = iff->as_If(); 3084 } else { 3085 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU"); 3086 peel_if_cmpu = iff->as_If(); 3087 } 3088 } 3089 } 3090 iff = idom(iff); 3091 } 3092 // Prefer signed compare over unsigned compare. 3093 IfNode* new_peel_if = NULL; 3094 if (peel_if == NULL) { 3095 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) { 3096 return false; // No peel point found 3097 } 3098 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop); 3099 if (new_peel_if == NULL) { 3100 return false; // No peel point found 3101 } 3102 peel_if = new_peel_if; 3103 } 3104 Node* last_peel = stay_in_loop(peel_if, loop); 3105 Node* first_not_peeled = stay_in_loop(last_peel, loop); 3106 if (first_not_peeled == NULL || first_not_peeled == head) { 3107 return false; 3108 } 3109 3110 #if !defined(PRODUCT) 3111 if (TraceLoopOpts) { 3112 tty->print("PartialPeel "); 3113 loop->dump_head(); 3114 } 3115 3116 if (TracePartialPeeling) { 3117 tty->print_cr("before partial peel one iteration"); 3118 Node_List wl; 3119 Node* t = head->in(2); 3120 while (true) { 3121 wl.push(t); 3122 if (t == head) break; 3123 t = idom(t); 3124 } 3125 while (wl.size() > 0) { 3126 Node* tt = wl.pop(); 3127 tt->dump(); 3128 if (tt == last_peel) tty->print_cr("-- cut --"); 3129 } 3130 } 3131 #endif 3132 ResourceArea *area = Thread::current()->resource_area(); 3133 VectorSet peel(area); 3134 VectorSet not_peel(area); 3135 Node_List peel_list(area); 3136 Node_List worklist(area); 3137 Node_List sink_list(area); 3138 3139 // Set of cfg nodes to peel are those that are executable from 3140 // the head through last_peel. 3141 assert(worklist.size() == 0, "should be empty"); 3142 worklist.push(head); 3143 peel.set(head->_idx); 3144 while (worklist.size() > 0) { 3145 Node *n = worklist.pop(); 3146 if (n != last_peel) { 3147 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3148 Node* use = n->fast_out(j); 3149 if (use->is_CFG() && 3150 loop->is_member(get_loop(use)) && 3151 !peel.test_set(use->_idx)) { 3152 worklist.push(use); 3153 } 3154 } 3155 } 3156 } 3157 3158 // Set of non-cfg nodes to peel are those that are control 3159 // dependent on the cfg nodes. 3160 uint i; 3161 for(i = 0; i < loop->_body.size(); i++ ) { 3162 Node *n = loop->_body.at(i); 3163 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n; 3164 if (peel.test(n_c->_idx)) { 3165 peel.set(n->_idx); 3166 } else { 3167 not_peel.set(n->_idx); 3168 } 3169 } 3170 3171 // Step 2: move operations from the peeled section down into the 3172 // not-peeled section 3173 3174 // Get a post order schedule of nodes in the peel region 3175 // Result in right-most operand. 3176 scheduled_nodelist(loop, peel, peel_list ); 3177 3178 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 3179 3180 // For future check for too many new phis 3181 uint old_phi_cnt = 0; 3182 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 3183 Node* use = head->fast_out(j); 3184 if (use->is_Phi()) old_phi_cnt++; 3185 } 3186 3187 #if !defined(PRODUCT) 3188 if (TracePartialPeeling) { 3189 tty->print_cr("\npeeled list"); 3190 } 3191 #endif 3192 3193 // Evacuate nodes in peel region into the not_peeled region if possible 3194 uint new_phi_cnt = 0; 3195 uint cloned_for_outside_use = 0; 3196 for (i = 0; i < peel_list.size();) { 3197 Node* n = peel_list.at(i); 3198 #if !defined(PRODUCT) 3199 if (TracePartialPeeling) n->dump(); 3200 #endif 3201 bool incr = true; 3202 if ( !n->is_CFG() ) { 3203 3204 if ( has_use_in_set(n, not_peel) ) { 3205 3206 // If not used internal to the peeled region, 3207 // move "n" from peeled to not_peeled region. 3208 3209 if ( !has_use_internal_to_set(n, peel, loop) ) { 3210 3211 // if not pinned and not a load (which maybe anti-dependent on a store) 3212 // and not a CMove (Matcher expects only bool->cmove). 3213 if (n->in(0) == NULL && !n->is_Load() && !n->is_CMove() && n->Opcode() != Op_ShenandoahWBMemProj) { 3214 cloned_for_outside_use += clone_for_use_outside_loop( loop, n, worklist ); 3215 sink_list.push(n); 3216 peel >>= n->_idx; // delete n from peel set. 3217 not_peel <<= n->_idx; // add n to not_peel set. 3218 peel_list.remove(i); 3219 incr = false; 3220 #if !defined(PRODUCT) 3221 if (TracePartialPeeling) { 3222 tty->print_cr("sink to not_peeled region: %d newbb: %d", 3223 n->_idx, get_ctrl(n)->_idx); 3224 } 3225 #endif 3226 } 3227 } else { 3228 // Otherwise check for special def-use cases that span 3229 // the peel/not_peel boundary such as bool->if 3230 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist ); 3231 new_phi_cnt++; 3232 } 3233 } 3234 } 3235 if (incr) i++; 3236 } 3237 3238 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) { 3239 #if !defined(PRODUCT) 3240 if (TracePartialPeeling) { 3241 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c", 3242 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F'); 3243 } 3244 #endif 3245 if (new_peel_if != NULL) { 3246 remove_cmpi_loop_exit(new_peel_if, loop); 3247 } 3248 // Inhibit more partial peeling on this loop 3249 assert(!head->is_partial_peel_loop(), "not partial peeled"); 3250 head->mark_partial_peel_failed(); 3251 if (cloned_for_outside_use > 0) { 3252 // Terminate this round of loop opts because 3253 // the graph outside this loop was changed. 3254 C->set_major_progress(); 3255 return true; 3256 } 3257 return false; 3258 } 3259 3260 // Step 3: clone loop, retarget control, and insert new phis 3261 3262 // Create new loop head for new phis and to hang 3263 // the nodes being moved (sinked) from the peel region. 3264 LoopNode* new_head = new LoopNode(last_peel, last_peel); 3265 new_head->set_unswitch_count(head->unswitch_count()); // Preserve 3266 _igvn.register_new_node_with_optimizer(new_head); 3267 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled"); 3268 _igvn.replace_input_of(first_not_peeled, 0, new_head); 3269 set_loop(new_head, loop); 3270 loop->_body.push(new_head); 3271 not_peel.set(new_head->_idx); 3272 set_idom(new_head, last_peel, dom_depth(first_not_peeled)); 3273 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled)); 3274 3275 while (sink_list.size() > 0) { 3276 Node* n = sink_list.pop(); 3277 set_ctrl(n, new_head); 3278 } 3279 3280 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 3281 3282 clone_loop(loop, old_new, dd, IgnoreStripMined); 3283 3284 const uint clone_exit_idx = 1; 3285 const uint orig_exit_idx = 2; 3286 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop"); 3287 3288 Node* head_clone = old_new[head->_idx]; 3289 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop(); 3290 Node* orig_tail_clone = head_clone->in(2); 3291 3292 // Add phi if "def" node is in peel set and "use" is not 3293 3294 for(i = 0; i < peel_list.size(); i++ ) { 3295 Node *def = peel_list.at(i); 3296 if (!def->is_CFG()) { 3297 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 3298 Node *use = def->fast_out(j); 3299 if (has_node(use) && use->in(0) != C->top() && 3300 (!peel.test(use->_idx) || 3301 (use->is_Phi() && use->in(0) == head)) ) { 3302 worklist.push(use); 3303 } 3304 } 3305 while( worklist.size() ) { 3306 Node *use = worklist.pop(); 3307 for (uint j = 1; j < use->req(); j++) { 3308 Node* n = use->in(j); 3309 if (n == def) { 3310 3311 // "def" is in peel set, "use" is not in peel set 3312 // or "use" is in the entry boundary (a phi) of the peel set 3313 3314 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use; 3315 3316 if ( loop->is_member(get_loop( use_c )) ) { 3317 // use is in loop 3318 if (old_new[use->_idx] != NULL) { // null for dead code 3319 Node* use_clone = old_new[use->_idx]; 3320 _igvn.replace_input_of(use, j, C->top()); 3321 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone ); 3322 } 3323 } else { 3324 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format"); 3325 // use is not in the loop, check if the live range includes the cut 3326 Node* lp_if = use_c->in(orig_exit_idx)->in(0); 3327 if (not_peel.test(lp_if->_idx)) { 3328 assert(j == orig_exit_idx, "use from original loop"); 3329 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone ); 3330 } 3331 } 3332 } 3333 } 3334 } 3335 } 3336 } 3337 3338 // Step 3b: retarget control 3339 3340 // Redirect control to the new loop head if a cloned node in 3341 // the not_peeled region has control that points into the peeled region. 3342 // This necessary because the cloned peeled region will be outside 3343 // the loop. 3344 // from to 3345 // cloned-peeled <---+ 3346 // new_head_clone: | <--+ 3347 // cloned-not_peeled in(0) in(0) 3348 // orig-peeled 3349 3350 for(i = 0; i < loop->_body.size(); i++ ) { 3351 Node *n = loop->_body.at(i); 3352 if (!n->is_CFG() && n->in(0) != NULL && 3353 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) { 3354 Node* n_clone = old_new[n->_idx]; 3355 _igvn.replace_input_of(n_clone, 0, new_head_clone); 3356 } 3357 } 3358 3359 // Backedge of the surviving new_head (the clone) is original last_peel 3360 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel); 3361 3362 // Cut first node in original not_peel set 3363 _igvn.rehash_node_delayed(new_head); // Multiple edge updates: 3364 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of 3365 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls 3366 3367 // Copy head_clone back-branch info to original head 3368 // and remove original head's loop entry and 3369 // clone head's back-branch 3370 _igvn.rehash_node_delayed(head); // Multiple edge updates 3371 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl)); 3372 head->set_req(LoopNode::LoopBackControl, C->top()); 3373 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top()); 3374 3375 // Similarly modify the phis 3376 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) { 3377 Node* use = head->fast_out(k); 3378 if (use->is_Phi() && use->outcnt() > 0) { 3379 Node* use_clone = old_new[use->_idx]; 3380 _igvn.rehash_node_delayed(use); // Multiple edge updates 3381 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl)); 3382 use->set_req(LoopNode::LoopBackControl, C->top()); 3383 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top()); 3384 } 3385 } 3386 3387 // Step 4: update dominator tree and dominator depth 3388 3389 set_idom(head, orig_tail_clone, dd); 3390 recompute_dom_depth(); 3391 3392 // Inhibit more partial peeling on this loop 3393 new_head_clone->set_partial_peel_loop(); 3394 C->set_major_progress(); 3395 loop->record_for_igvn(); 3396 3397 #if !defined(PRODUCT) 3398 if (TracePartialPeeling) { 3399 tty->print_cr("\nafter partial peel one iteration"); 3400 Node_List wl(area); 3401 Node* t = last_peel; 3402 while (true) { 3403 wl.push(t); 3404 if (t == head_clone) break; 3405 t = idom(t); 3406 } 3407 while (wl.size() > 0) { 3408 Node* tt = wl.pop(); 3409 if (tt == head) tty->print_cr("orig head"); 3410 else if (tt == new_head_clone) tty->print_cr("new head"); 3411 else if (tt == head_clone) tty->print_cr("clone head"); 3412 tt->dump(); 3413 } 3414 } 3415 #endif 3416 return true; 3417 } 3418 3419 //------------------------------reorg_offsets---------------------------------- 3420 // Reorganize offset computations to lower register pressure. Mostly 3421 // prevent loop-fallout uses of the pre-incremented trip counter (which are 3422 // then alive with the post-incremented trip counter forcing an extra 3423 // register move) 3424 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) { 3425 // Perform it only for canonical counted loops. 3426 // Loop's shape could be messed up by iteration_split_impl. 3427 if (!loop->_head->is_CountedLoop()) 3428 return; 3429 if (!loop->_head->as_Loop()->is_valid_counted_loop()) 3430 return; 3431 3432 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 3433 CountedLoopEndNode *cle = cl->loopexit(); 3434 Node *exit = cle->proj_out(false); 3435 Node *phi = cl->phi(); 3436 3437 // Check for the special case of folks using the pre-incremented 3438 // trip-counter on the fall-out path (forces the pre-incremented 3439 // and post-incremented trip counter to be live at the same time). 3440 // Fix this by adjusting to use the post-increment trip counter. 3441 3442 bool progress = true; 3443 while (progress) { 3444 progress = false; 3445 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) { 3446 Node* use = phi->fast_out(i); // User of trip-counter 3447 if (!has_ctrl(use)) continue; 3448 Node *u_ctrl = get_ctrl(use); 3449 if (use->is_Phi()) { 3450 u_ctrl = NULL; 3451 for (uint j = 1; j < use->req(); j++) 3452 if (use->in(j) == phi) 3453 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j)); 3454 } 3455 IdealLoopTree *u_loop = get_loop(u_ctrl); 3456 // Look for loop-invariant use 3457 if (u_loop == loop) continue; 3458 if (loop->is_member(u_loop)) continue; 3459 // Check that use is live out the bottom. Assuming the trip-counter 3460 // update is right at the bottom, uses of of the loop middle are ok. 3461 if (dom_lca(exit, u_ctrl) != exit) continue; 3462 // Hit! Refactor use to use the post-incremented tripcounter. 3463 // Compute a post-increment tripcounter. 3464 Node* c = exit; 3465 if (cl->is_strip_mined()) { 3466 IdealLoopTree* outer_loop = get_loop(cl->outer_loop()); 3467 if (!outer_loop->is_member(u_loop)) { 3468 c = cl->outer_loop_exit(); 3469 } 3470 } 3471 Node *opaq = new Opaque2Node(C, cle->incr()); 3472 register_new_node(opaq, c); 3473 Node *neg_stride = _igvn.intcon(-cle->stride_con()); 3474 set_ctrl(neg_stride, C->root()); 3475 Node *post = new AddINode(opaq, neg_stride); 3476 register_new_node(post, c); 3477 _igvn.rehash_node_delayed(use); 3478 for (uint j = 1; j < use->req(); j++) { 3479 if (use->in(j) == phi) 3480 use->set_req(j, post); 3481 } 3482 // Since DU info changed, rerun loop 3483 progress = true; 3484 break; 3485 } 3486 } 3487 3488 }