1 /* 2 * Copyright (c) 1999, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "memory/allocation.inline.hpp" 27 #include "opto/addnode.hpp" 28 #include "opto/connode.hpp" 29 #include "opto/divnode.hpp" 30 #include "opto/loopnode.hpp" 31 #include "opto/matcher.hpp" 32 #include "opto/mulnode.hpp" 33 #include "opto/movenode.hpp" 34 #include "opto/opaquenode.hpp" 35 #include "opto/rootnode.hpp" 36 #include "opto/shenandoahSupport.hpp" 37 #include "opto/subnode.hpp" 38 39 //============================================================================= 40 //------------------------------split_thru_phi--------------------------------- 41 // Split Node 'n' through merge point if there is enough win. 42 Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) { 43 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) { 44 // ConvI2L may have type information on it which is unsafe to push up 45 // so disable this for now 46 return NULL; 47 } 48 49 int wins = 0; 50 assert(!n->is_CFG(), ""); 51 assert(region->is_Region(), ""); 52 53 const Type* type = n->bottom_type(); 54 const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr(); 55 Node *phi; 56 if (t_oop != NULL && t_oop->is_known_instance_field()) { 57 int iid = t_oop->instance_id(); 58 int index = C->get_alias_index(t_oop); 59 int offset = t_oop->offset(); 60 phi = new PhiNode(region, type, NULL, iid, index, offset); 61 } else { 62 phi = PhiNode::make_blank(region, n); 63 } 64 uint old_unique = C->unique(); 65 for (uint i = 1; i < region->req(); i++) { 66 Node *x; 67 Node* the_clone = NULL; 68 if (region->in(i) == C->top()) { 69 x = C->top(); // Dead path? Use a dead data op 70 } else { 71 #ifdef ASSERT 72 if (n->is_ShenandoahBarrier()) { n->as_ShenandoahBarrier()->check_invariants(); } 73 #endif 74 x = n->clone(); // Else clone up the data op 75 the_clone = x; // Remember for possible deletion. 76 // Alter data node to use pre-phi inputs 77 if (n->in(0) == region) 78 x->set_req( 0, region->in(i) ); 79 for (uint j = 1; j < n->req(); j++) { 80 Node *in = n->in(j); 81 if (in->is_Phi() && in->in(0) == region) 82 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone 83 } 84 } 85 // Check for a 'win' on some paths 86 const Type *t = x->Value(&_igvn); 87 88 bool singleton = t->singleton(); 89 90 // A TOP singleton indicates that there are no possible values incoming 91 // along a particular edge. In most cases, this is OK, and the Phi will 92 // be eliminated later in an Ideal call. However, we can't allow this to 93 // happen if the singleton occurs on loop entry, as the elimination of 94 // the PhiNode may cause the resulting node to migrate back to a previous 95 // loop iteration. 96 if (singleton && t == Type::TOP) { 97 // Is_Loop() == false does not confirm the absence of a loop (e.g., an 98 // irreducible loop may not be indicated by an affirmative is_Loop()); 99 // therefore, the only top we can split thru a phi is on a backedge of 100 // a loop. 101 singleton &= region->is_Loop() && (i != LoopNode::EntryControl); 102 } 103 104 if (singleton) { 105 wins++; 106 x = ((PhaseGVN&)_igvn).makecon(t); 107 } else { 108 // We now call Identity to try to simplify the cloned node. 109 // Note that some Identity methods call phase->type(this). 110 // Make sure that the type array is big enough for 111 // our new node, even though we may throw the node away. 112 // (Note: This tweaking with igvn only works because x is a new node.) 113 _igvn.set_type(x, t); 114 // If x is a TypeNode, capture any more-precise type permanently into Node 115 // otherwise it will be not updated during igvn->transform since 116 // igvn->type(x) is set to x->Value() already. 117 x->raise_bottom_type(t); 118 if (x->Opcode() != Op_ShenandoahWriteBarrier) { 119 Node *y = x->Identity(&_igvn); 120 if (y != x) { 121 wins++; 122 x = y; 123 } else { 124 y = _igvn.hash_find(x); 125 if (y) { 126 wins++; 127 x = y; 128 } else { 129 // Else x is a new node we are keeping 130 // We do not need register_new_node_with_optimizer 131 // because set_type has already been called. 132 _igvn._worklist.push(x); 133 } 134 } 135 } else { 136 // wins++; 137 _igvn._worklist.push(x); 138 } 139 } 140 if (x != the_clone && the_clone != NULL) 141 _igvn.remove_dead_node(the_clone); 142 phi->set_req( i, x ); 143 } 144 // Too few wins? 145 if (wins <= policy) { 146 _igvn.remove_dead_node(phi); 147 return NULL; 148 } 149 150 // Record Phi 151 register_new_node( phi, region ); 152 153 for (uint i2 = 1; i2 < phi->req(); i2++) { 154 Node *x = phi->in(i2); 155 // If we commoned up the cloned 'x' with another existing Node, 156 // the existing Node picks up a new use. We need to make the 157 // existing Node occur higher up so it dominates its uses. 158 Node *old_ctrl; 159 IdealLoopTree *old_loop; 160 161 if (x->is_Con()) { 162 // Constant's control is always root. 163 set_ctrl(x, C->root()); 164 continue; 165 } 166 // The occasional new node 167 if (x->_idx >= old_unique) { // Found a new, unplaced node? 168 old_ctrl = NULL; 169 old_loop = NULL; // Not in any prior loop 170 } else { 171 old_ctrl = get_ctrl(x); 172 old_loop = get_loop(old_ctrl); // Get prior loop 173 } 174 // New late point must dominate new use 175 Node *new_ctrl = dom_lca(old_ctrl, region->in(i2)); 176 if (new_ctrl == old_ctrl) // Nothing is changed 177 continue; 178 179 IdealLoopTree *new_loop = get_loop(new_ctrl); 180 181 // Don't move x into a loop if its uses are 182 // outside of loop. Otherwise x will be cloned 183 // for each use outside of this loop. 184 IdealLoopTree *use_loop = get_loop(region); 185 if (!new_loop->is_member(use_loop) && 186 (old_loop == NULL || !new_loop->is_member(old_loop))) { 187 // Take early control, later control will be recalculated 188 // during next iteration of loop optimizations. 189 new_ctrl = get_early_ctrl(x); 190 new_loop = get_loop(new_ctrl); 191 } 192 // Set new location 193 set_ctrl(x, new_ctrl); 194 // If changing loop bodies, see if we need to collect into new body 195 if (old_loop != new_loop) { 196 if (old_loop && !old_loop->_child) 197 old_loop->_body.yank(x); 198 if (!new_loop->_child) 199 new_loop->_body.push(x); // Collect body info 200 } 201 } 202 203 return phi; 204 } 205 206 void PhaseIdealLoop::split_mem_thru_phi(Node* n, Node* r, Node* phi) { 207 if (n->Opcode() == Op_ShenandoahWriteBarrier) { 208 #ifdef ASSERT 209 n->as_ShenandoahBarrier()->check_invariants(); 210 #endif 211 if (n->has_out_with(Op_ShenandoahWBMemProj)) { 212 Node* old_mem_phi = n->in(ShenandoahBarrierNode::Memory); 213 assert(r->is_Region(), "need region to control phi"); 214 assert(phi->is_Phi(), "expect phi"); 215 Node* memphi = PhiNode::make(r, old_mem_phi, Type::MEMORY, C->alias_type(n->adr_type())->adr_type()); 216 for (uint i = 1; i < r->req(); i++) { 217 Node* wb = phi->in(i); 218 if (wb->Opcode() == Op_ShenandoahWriteBarrier) { 219 assert(! wb->has_out_with(Op_ShenandoahWBMemProj), "new clone does not have mem proj"); 220 Node* new_proj = new ShenandoahWBMemProjNode(wb); 221 register_new_node(new_proj, r->in(i)); 222 memphi->set_req(i, new_proj); 223 } else { 224 if (old_mem_phi->is_Phi() && old_mem_phi->in(0) == r) { 225 memphi->set_req(i, old_mem_phi->in(i)); 226 } 227 } 228 } 229 register_new_node(memphi, r); 230 Node* old_mem_out = n->find_out_with(Op_ShenandoahWBMemProj); 231 assert(old_mem_out != NULL, "expect memory projection"); 232 _igvn.replace_node(old_mem_out, memphi); 233 } 234 assert(! n->has_out_with(Op_ShenandoahWBMemProj), "no more memory outs"); 235 } 236 } 237 238 //------------------------------dominated_by------------------------------------ 239 // Replace the dominated test with an obvious true or false. Place it on the 240 // IGVN worklist for later cleanup. Move control-dependent data Nodes on the 241 // live path up to the dominating control. 242 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff, bool flip, bool exclude_loop_predicate ) { 243 #ifndef PRODUCT 244 if (VerifyLoopOptimizations && PrintOpto) tty->print_cr("dominating test"); 245 #endif 246 247 248 // prevdom is the dominating projection of the dominating test. 249 assert( iff->is_If(), "" ); 250 assert( iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added"); 251 int pop = prevdom->Opcode(); 252 assert( pop == Op_IfFalse || pop == Op_IfTrue, "" ); 253 if (flip) { 254 if (pop == Op_IfTrue) 255 pop = Op_IfFalse; 256 else 257 pop = Op_IfTrue; 258 } 259 // 'con' is set to true or false to kill the dominated test. 260 Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO); 261 set_ctrl(con, C->root()); // Constant gets a new use 262 // Hack the dominated test 263 _igvn.replace_input_of(iff, 1, con); 264 265 // If I dont have a reachable TRUE and FALSE path following the IfNode then 266 // I can assume this path reaches an infinite loop. In this case it's not 267 // important to optimize the data Nodes - either the whole compilation will 268 // be tossed or this path (and all data Nodes) will go dead. 269 if (iff->outcnt() != 2) return; 270 271 // Make control-dependent data Nodes on the live path (path that will remain 272 // once the dominated IF is removed) become control-dependent on the 273 // dominating projection. 274 Node* dp = iff->as_If()->proj_out(pop == Op_IfTrue); 275 276 // Loop predicates may have depending checks which should not 277 // be skipped. For example, range check predicate has two checks 278 // for lower and upper bounds. 279 if (dp == NULL) 280 return; 281 282 ProjNode* dp_proj = dp->as_Proj(); 283 ProjNode* unc_proj = iff->as_If()->proj_out(1 - dp_proj->_con)->as_Proj(); 284 if (exclude_loop_predicate && 285 (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) != NULL || 286 unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check) != NULL)) { 287 // If this is a range check (IfNode::is_range_check), do not 288 // reorder because Compile::allow_range_check_smearing might have 289 // changed the check. 290 return; // Let IGVN transformation change control dependence. 291 } 292 293 IdealLoopTree *old_loop = get_loop(dp); 294 295 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { 296 Node* cd = dp->fast_out(i); // Control-dependent node 297 if (cd->depends_only_on_test()) { 298 assert(cd->in(0) == dp, ""); 299 _igvn.replace_input_of(cd, 0, prevdom); 300 set_early_ctrl(cd); 301 IdealLoopTree *new_loop = get_loop(get_ctrl(cd)); 302 if (old_loop != new_loop) { 303 if (!old_loop->_child) old_loop->_body.yank(cd); 304 if (!new_loop->_child) new_loop->_body.push(cd); 305 } 306 --i; 307 --imax; 308 } 309 } 310 } 311 312 //------------------------------has_local_phi_input---------------------------- 313 // Return TRUE if 'n' has Phi inputs from its local block and no other 314 // block-local inputs (all non-local-phi inputs come from earlier blocks) 315 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) { 316 Node *n_ctrl = get_ctrl(n); 317 // See if some inputs come from a Phi in this block, or from before 318 // this block. 319 uint i; 320 for( i = 1; i < n->req(); i++ ) { 321 Node *phi = n->in(i); 322 if( phi->is_Phi() && phi->in(0) == n_ctrl ) 323 break; 324 } 325 if( i >= n->req() ) 326 return NULL; // No Phi inputs; nowhere to clone thru 327 328 // Check for inputs created between 'n' and the Phi input. These 329 // must split as well; they have already been given the chance 330 // (courtesy of a post-order visit) and since they did not we must 331 // recover the 'cost' of splitting them by being very profitable 332 // when splitting 'n'. Since this is unlikely we simply give up. 333 for( i = 1; i < n->req(); i++ ) { 334 Node *m = n->in(i); 335 if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) { 336 // We allow the special case of AddP's with no local inputs. 337 // This allows us to split-up address expressions. 338 if (m->is_AddP() && 339 get_ctrl(m->in(2)) != n_ctrl && 340 get_ctrl(m->in(3)) != n_ctrl) { 341 // Move the AddP up to dominating point 342 set_ctrl_and_loop(m, find_non_split_ctrl(idom(n_ctrl))); 343 continue; 344 } 345 return NULL; 346 } 347 } 348 349 return n_ctrl; 350 } 351 352 //------------------------------remix_address_expressions---------------------- 353 // Rework addressing expressions to get the most loop-invariant stuff 354 // moved out. We'd like to do all associative operators, but it's especially 355 // important (common) to do address expressions. 356 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) { 357 if (!has_ctrl(n)) return NULL; 358 Node *n_ctrl = get_ctrl(n); 359 IdealLoopTree *n_loop = get_loop(n_ctrl); 360 361 // See if 'n' mixes loop-varying and loop-invariant inputs and 362 // itself is loop-varying. 363 364 // Only interested in binary ops (and AddP) 365 if( n->req() < 3 || n->req() > 4 ) return NULL; 366 367 Node *n1_ctrl = get_ctrl(n->in( 1)); 368 Node *n2_ctrl = get_ctrl(n->in( 2)); 369 Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3)); 370 IdealLoopTree *n1_loop = get_loop( n1_ctrl ); 371 IdealLoopTree *n2_loop = get_loop( n2_ctrl ); 372 IdealLoopTree *n3_loop = get_loop( n3_ctrl ); 373 374 // Does one of my inputs spin in a tighter loop than self? 375 if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) || 376 (n_loop->is_member( n2_loop ) && n_loop != n2_loop) || 377 (n_loop->is_member( n3_loop ) && n_loop != n3_loop) ) 378 return NULL; // Leave well enough alone 379 380 // Is at least one of my inputs loop-invariant? 381 if( n1_loop == n_loop && 382 n2_loop == n_loop && 383 n3_loop == n_loop ) 384 return NULL; // No loop-invariant inputs 385 386 387 int n_op = n->Opcode(); 388 389 // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2). 390 if( n_op == Op_LShiftI ) { 391 // Scale is loop invariant 392 Node *scale = n->in(2); 393 Node *scale_ctrl = get_ctrl(scale); 394 IdealLoopTree *scale_loop = get_loop(scale_ctrl ); 395 if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) ) 396 return NULL; 397 const TypeInt *scale_t = scale->bottom_type()->isa_int(); 398 if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 ) 399 return NULL; // Dont bother with byte/short masking 400 // Add must vary with loop (else shift would be loop-invariant) 401 Node *add = n->in(1); 402 Node *add_ctrl = get_ctrl(add); 403 IdealLoopTree *add_loop = get_loop(add_ctrl); 404 //assert( n_loop == add_loop, "" ); 405 if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops 406 407 // Convert I-V into I+ (0-V); same for V-I 408 if( add->Opcode() == Op_SubI && 409 _igvn.type( add->in(1) ) != TypeInt::ZERO ) { 410 Node *zero = _igvn.intcon(0); 411 set_ctrl(zero, C->root()); 412 Node *neg = new SubINode( _igvn.intcon(0), add->in(2) ); 413 register_new_node( neg, get_ctrl(add->in(2) ) ); 414 add = new AddINode( add->in(1), neg ); 415 register_new_node( add, add_ctrl ); 416 } 417 if( add->Opcode() != Op_AddI ) return NULL; 418 // See if one add input is loop invariant 419 Node *add_var = add->in(1); 420 Node *add_var_ctrl = get_ctrl(add_var); 421 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 422 Node *add_invar = add->in(2); 423 Node *add_invar_ctrl = get_ctrl(add_invar); 424 IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl ); 425 if( add_var_loop == n_loop ) { 426 } else if( add_invar_loop == n_loop ) { 427 // Swap to find the invariant part 428 add_invar = add_var; 429 add_invar_ctrl = add_var_ctrl; 430 add_invar_loop = add_var_loop; 431 add_var = add->in(2); 432 Node *add_var_ctrl = get_ctrl(add_var); 433 IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); 434 } else // Else neither input is loop invariant 435 return NULL; 436 if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) ) 437 return NULL; // No invariant part of the add? 438 439 // Yes! Reshape address expression! 440 Node *inv_scale = new LShiftINode( add_invar, scale ); 441 Node *inv_scale_ctrl = 442 dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ? 443 add_invar_ctrl : scale_ctrl; 444 register_new_node( inv_scale, inv_scale_ctrl ); 445 Node *var_scale = new LShiftINode( add_var, scale ); 446 register_new_node( var_scale, n_ctrl ); 447 Node *var_add = new AddINode( var_scale, inv_scale ); 448 register_new_node( var_add, n_ctrl ); 449 _igvn.replace_node( n, var_add ); 450 return var_add; 451 } 452 453 // Replace (I+V) with (V+I) 454 if( n_op == Op_AddI || 455 n_op == Op_AddL || 456 n_op == Op_AddF || 457 n_op == Op_AddD || 458 n_op == Op_MulI || 459 n_op == Op_MulL || 460 n_op == Op_MulF || 461 n_op == Op_MulD ) { 462 if( n2_loop == n_loop ) { 463 assert( n1_loop != n_loop, "" ); 464 n->swap_edges(1, 2); 465 } 466 } 467 468 // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V), 469 // but not if I2 is a constant. 470 if( n_op == Op_AddP ) { 471 if( n2_loop == n_loop && n3_loop != n_loop ) { 472 if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) { 473 Node *n22_ctrl = get_ctrl(n->in(2)->in(2)); 474 Node *n23_ctrl = get_ctrl(n->in(2)->in(3)); 475 IdealLoopTree *n22loop = get_loop( n22_ctrl ); 476 IdealLoopTree *n23_loop = get_loop( n23_ctrl ); 477 if( n22loop != n_loop && n22loop->is_member(n_loop) && 478 n23_loop == n_loop ) { 479 Node *add1 = new AddPNode( n->in(1), n->in(2)->in(2), n->in(3) ); 480 // Stuff new AddP in the loop preheader 481 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) ); 482 Node *add2 = new AddPNode( n->in(1), add1, n->in(2)->in(3) ); 483 register_new_node( add2, n_ctrl ); 484 _igvn.replace_node( n, add2 ); 485 return add2; 486 } 487 } 488 } 489 490 // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V) 491 if( n2_loop != n_loop && n3_loop == n_loop ) { 492 if( n->in(3)->Opcode() == Op_AddI ) { 493 Node *V = n->in(3)->in(1); 494 Node *I = n->in(3)->in(2); 495 if( is_member(n_loop,get_ctrl(V)) ) { 496 } else { 497 Node *tmp = V; V = I; I = tmp; 498 } 499 if( !is_member(n_loop,get_ctrl(I)) ) { 500 Node *add1 = new AddPNode( n->in(1), n->in(2), I ); 501 // Stuff new AddP in the loop preheader 502 register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) ); 503 Node *add2 = new AddPNode( n->in(1), add1, V ); 504 register_new_node( add2, n_ctrl ); 505 _igvn.replace_node( n, add2 ); 506 return add2; 507 } 508 } 509 } 510 } 511 512 return NULL; 513 } 514 515 //------------------------------conditional_move------------------------------- 516 // Attempt to replace a Phi with a conditional move. We have some pretty 517 // strict profitability requirements. All Phis at the merge point must 518 // be converted, so we can remove the control flow. We need to limit the 519 // number of c-moves to a small handful. All code that was in the side-arms 520 // of the CFG diamond is now speculatively executed. This code has to be 521 // "cheap enough". We are pretty much limited to CFG diamonds that merge 522 // 1 or 2 items with a total of 1 or 2 ops executed speculatively. 523 Node *PhaseIdealLoop::conditional_move( Node *region ) { 524 525 assert(region->is_Region(), "sanity check"); 526 if (region->req() != 3) return NULL; 527 528 // Check for CFG diamond 529 Node *lp = region->in(1); 530 Node *rp = region->in(2); 531 if (!lp || !rp) return NULL; 532 Node *lp_c = lp->in(0); 533 if (lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If()) return NULL; 534 IfNode *iff = lp_c->as_If(); 535 536 // Check for ops pinned in an arm of the diamond. 537 // Can't remove the control flow in this case 538 if (lp->outcnt() > 1) return NULL; 539 if (rp->outcnt() > 1) return NULL; 540 541 IdealLoopTree* r_loop = get_loop(region); 542 assert(r_loop == get_loop(iff), "sanity"); 543 // Always convert to CMOVE if all results are used only outside this loop. 544 bool used_inside_loop = (r_loop == _ltree_root); 545 546 // Check profitability 547 int cost = 0; 548 int phis = 0; 549 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 550 Node *out = region->fast_out(i); 551 if (!out->is_Phi()) continue; // Ignore other control edges, etc 552 phis++; 553 PhiNode* phi = out->as_Phi(); 554 BasicType bt = phi->type()->basic_type(); 555 switch (bt) { 556 case T_FLOAT: 557 case T_DOUBLE: { 558 cost += Matcher::float_cmove_cost(); // Could be very expensive 559 break; 560 } 561 case T_LONG: { 562 cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's 563 } 564 case T_INT: // These all CMOV fine 565 case T_ADDRESS: { // (RawPtr) 566 cost++; 567 break; 568 } 569 case T_NARROWOOP: // Fall through 570 case T_OBJECT: { // Base oops are OK, but not derived oops 571 const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr(); 572 // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a 573 // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus 574 // CMOVE'ing a derived pointer requires we also CMOVE the base. If we 575 // have a Phi for the base here that we convert to a CMOVE all is well 576 // and good. But if the base is dead, we'll not make a CMOVE. Later 577 // the allocator will have to produce a base by creating a CMOVE of the 578 // relevant bases. This puts the allocator in the business of 579 // manufacturing expensive instructions, generally a bad plan. 580 // Just Say No to Conditionally-Moved Derived Pointers. 581 if (tp && tp->offset() != 0) 582 return NULL; 583 cost++; 584 break; 585 } 586 default: 587 return NULL; // In particular, can't do memory or I/O 588 } 589 // Add in cost any speculative ops 590 for (uint j = 1; j < region->req(); j++) { 591 Node *proj = region->in(j); 592 Node *inp = phi->in(j); 593 if (get_ctrl(inp) == proj) { // Found local op 594 cost++; 595 // Check for a chain of dependent ops; these will all become 596 // speculative in a CMOV. 597 for (uint k = 1; k < inp->req(); k++) 598 if (get_ctrl(inp->in(k)) == proj) 599 cost += ConditionalMoveLimit; // Too much speculative goo 600 } 601 } 602 // See if the Phi is used by a Cmp or Narrow oop Decode/Encode. 603 // This will likely Split-If, a higher-payoff operation. 604 for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) { 605 Node* use = phi->fast_out(k); 606 if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr()) 607 cost += ConditionalMoveLimit; 608 // Is there a use inside the loop? 609 // Note: check only basic types since CMoveP is pinned. 610 if (!used_inside_loop && is_java_primitive(bt)) { 611 IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use); 612 if (r_loop == u_loop || r_loop->is_member(u_loop)) { 613 used_inside_loop = true; 614 } 615 } 616 } 617 } 618 Node* bol = iff->in(1); 619 assert(bol->Opcode() == Op_Bool, ""); 620 int cmp_op = bol->in(1)->Opcode(); 621 // It is expensive to generate flags from a float compare. 622 // Avoid duplicated float compare. 623 if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL; 624 625 float infrequent_prob = PROB_UNLIKELY_MAG(3); 626 // Ignore cost and blocks frequency if CMOVE can be moved outside the loop. 627 if (used_inside_loop) { 628 if (cost >= ConditionalMoveLimit) return NULL; // Too much goo 629 630 // BlockLayoutByFrequency optimization moves infrequent branch 631 // from hot path. No point in CMOV'ing in such case (110 is used 632 // instead of 100 to take into account not exactness of float value). 633 if (BlockLayoutByFrequency) { 634 infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f); 635 } 636 } 637 // Check for highly predictable branch. No point in CMOV'ing if 638 // we are going to predict accurately all the time. 639 if (iff->_prob < infrequent_prob || 640 iff->_prob > (1.0f - infrequent_prob)) 641 return NULL; 642 643 // -------------- 644 // Now replace all Phis with CMOV's 645 Node *cmov_ctrl = iff->in(0); 646 uint flip = (lp->Opcode() == Op_IfTrue); 647 while (1) { 648 PhiNode* phi = NULL; 649 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 650 Node *out = region->fast_out(i); 651 if (out->is_Phi()) { 652 phi = out->as_Phi(); 653 break; 654 } 655 } 656 if (phi == NULL) break; 657 #ifndef PRODUCT 658 if (PrintOpto && VerifyLoopOptimizations) tty->print_cr("CMOV"); 659 #endif 660 // Move speculative ops 661 for (uint j = 1; j < region->req(); j++) { 662 Node *proj = region->in(j); 663 Node *inp = phi->in(j); 664 if (get_ctrl(inp) == proj) { // Found local op 665 #ifndef PRODUCT 666 if (PrintOpto && VerifyLoopOptimizations) { 667 tty->print(" speculate: "); 668 inp->dump(); 669 } 670 #endif 671 set_ctrl(inp, cmov_ctrl); 672 } 673 } 674 Node *cmov = CMoveNode::make(cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi)); 675 register_new_node( cmov, cmov_ctrl ); 676 _igvn.replace_node( phi, cmov ); 677 #ifndef PRODUCT 678 if (TraceLoopOpts) { 679 tty->print("CMOV "); 680 r_loop->dump_head(); 681 if (Verbose) { 682 bol->in(1)->dump(1); 683 cmov->dump(1); 684 } 685 } 686 if (VerifyLoopOptimizations) verify(); 687 #endif 688 } 689 690 // The useless CFG diamond will fold up later; see the optimization in 691 // RegionNode::Ideal. 692 _igvn._worklist.push(region); 693 694 return iff->in(1); 695 } 696 697 #ifdef ASSERT 698 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) { 699 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 700 Node* u = m->fast_out(i); 701 if (u->is_CFG()) { 702 if (u->Opcode() == Op_NeverBranch) { 703 u = ((NeverBranchNode*)u)->proj_out(0); 704 enqueue_cfg_uses(u, wq); 705 } else { 706 wq.push(u); 707 } 708 } 709 } 710 } 711 #endif 712 713 // Try moving a store out of a loop, right before the loop 714 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) { 715 // Store has to be first in the loop body 716 IdealLoopTree *n_loop = get_loop(n_ctrl); 717 if (n->is_Store() && n_loop != _ltree_root && n_loop->is_loop() && n->in(0) != NULL) { 718 Node* address = n->in(MemNode::Address); 719 Node* value = n->in(MemNode::ValueIn); 720 Node* mem = n->in(MemNode::Memory); 721 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 722 IdealLoopTree* value_loop = get_loop(get_ctrl(value)); 723 724 // - address and value must be loop invariant 725 // - memory must be a memory Phi for the loop 726 // - Store must be the only store on this memory slice in the 727 // loop: if there's another store following this one then value 728 // written at iteration i by the second store could be overwritten 729 // at iteration i+n by the first store: it's not safe to move the 730 // first store out of the loop 731 // - nothing must observe the Phi memory: it guarantees no read 732 // before the store and no early exit out of the loop 733 // With those conditions, we are also guaranteed the store post 734 // dominates the loop head. Otherwise there would be extra Phi 735 // involved between the loop's Phi and the store. 736 737 if (!n_loop->is_member(address_loop) && 738 !n_loop->is_member(value_loop) && 739 mem->is_Phi() && mem->in(0) == n_loop->_head && 740 mem->outcnt() == 1 && 741 mem->in(LoopNode::LoopBackControl) == n) { 742 743 #ifdef ASSERT 744 // Verify that store's control does post dominate loop entry and 745 // that there's no early exit of the loop before the store. 746 bool ctrl_ok = false; 747 { 748 // Follow control from loop head until n, we exit the loop or 749 // we reach the tail 750 ResourceMark rm; 751 Unique_Node_List wq; 752 wq.push(n_loop->_head); 753 assert(n_loop->_tail != NULL, "need a tail"); 754 for (uint next = 0; next < wq.size(); ++next) { 755 Node *m = wq.at(next); 756 if (m == n->in(0)) { 757 ctrl_ok = true; 758 continue; 759 } 760 assert(!has_ctrl(m), "should be CFG"); 761 if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) { 762 ctrl_ok = false; 763 break; 764 } 765 enqueue_cfg_uses(m, wq); 766 } 767 } 768 assert(ctrl_ok, "bad control"); 769 #endif 770 771 // move the Store 772 _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem); 773 _igvn.replace_input_of(n, 0, n_loop->_head->in(LoopNode::EntryControl)); 774 _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl)); 775 // Disconnect the phi now. An empty phi can confuse other 776 // optimizations in this pass of loop opts. 777 _igvn.replace_node(mem, mem->in(LoopNode::EntryControl)); 778 n_loop->_body.yank(mem); 779 780 IdealLoopTree* new_loop = get_loop(n->in(0)); 781 set_ctrl_and_loop(n, n->in(0)); 782 783 return n; 784 } 785 } 786 return NULL; 787 } 788 789 // Try moving a store out of a loop, right after the loop 790 void PhaseIdealLoop::try_move_store_after_loop(Node* n) { 791 if (n->is_Store() && n->in(0) != NULL) { 792 Node *n_ctrl = get_ctrl(n); 793 IdealLoopTree *n_loop = get_loop(n_ctrl); 794 // Store must be in a loop 795 if (n_loop != _ltree_root && !n_loop->_irreducible) { 796 Node* address = n->in(MemNode::Address); 797 Node* value = n->in(MemNode::ValueIn); 798 IdealLoopTree* address_loop = get_loop(get_ctrl(address)); 799 // address must be loop invariant 800 if (!n_loop->is_member(address_loop)) { 801 // Store must be last on this memory slice in the loop and 802 // nothing in the loop must observe it 803 Node* phi = NULL; 804 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 805 Node* u = n->fast_out(i); 806 if (has_ctrl(u)) { // control use? 807 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 808 if (!n_loop->is_member(u_loop)) { 809 continue; 810 } 811 if (u->is_Phi() && u->in(0) == n_loop->_head) { 812 assert(_igvn.type(u) == Type::MEMORY, "bad phi"); 813 assert(phi == NULL, "already found"); 814 phi = u; 815 continue; 816 } 817 } 818 phi = NULL; 819 break; 820 } 821 if (phi != NULL) { 822 // Nothing in the loop before the store (next iteration) 823 // must observe the stored value 824 bool mem_ok = true; 825 { 826 ResourceMark rm; 827 Unique_Node_List wq; 828 wq.push(phi); 829 for (uint next = 0; next < wq.size() && mem_ok; ++next) { 830 Node *m = wq.at(next); 831 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) { 832 Node* u = m->fast_out(i); 833 if (u->is_Store() || u->is_Phi()) { 834 if (u != n) { 835 wq.push(u); 836 mem_ok = (wq.size() <= 10); 837 } 838 } else { 839 mem_ok = false; 840 break; 841 } 842 } 843 } 844 } 845 if (mem_ok) { 846 // Move the Store out of the loop creating clones along 847 // all paths out of the loop that observe the stored value 848 _igvn.rehash_node_delayed(phi); 849 int count = phi->replace_edge(n, n->in(MemNode::Memory)); 850 assert(count > 0, "inconsistent phi"); 851 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 852 Node* u = n->fast_out(i); 853 Node* c = get_ctrl(u); 854 855 if (u->is_Phi()) { 856 c = u->in(0)->in(u->find_edge(n)); 857 } 858 IdealLoopTree *u_loop = get_loop(c); 859 assert (!n_loop->is_member(u_loop), "only the phi should have been a use in the loop"); 860 while(true) { 861 Node* next_c = find_non_split_ctrl(idom(c)); 862 if (n_loop->is_member(get_loop(next_c))) { 863 break; 864 } 865 c = next_c; 866 } 867 868 Node* st = n->clone(); 869 st->set_req(0, c); 870 _igvn.register_new_node_with_optimizer(st); 871 872 set_ctrl(st, c); 873 IdealLoopTree* new_loop = get_loop(c); 874 assert(new_loop != n_loop, "should be moved out of loop"); 875 if (new_loop->_child == NULL) new_loop->_body.push(st); 876 877 _igvn.replace_input_of(u, u->find_edge(n), st); 878 --imax; 879 --i; 880 } 881 882 883 assert(n->outcnt() == 0, "all uses should be gone"); 884 _igvn.replace_input_of(n, MemNode::Memory, C->top()); 885 // Disconnect the phi now. An empty phi can confuse other 886 // optimizations in this pass of loop opts.. 887 if (phi->in(LoopNode::LoopBackControl) == phi) { 888 _igvn.replace_node(phi, phi->in(LoopNode::EntryControl)); 889 n_loop->_body.yank(phi); 890 } 891 } 892 } 893 } 894 } 895 } 896 } 897 898 //------------------------------split_if_with_blocks_pre----------------------- 899 // Do the real work in a non-recursive function. Data nodes want to be 900 // cloned in the pre-order so they can feed each other nicely. 901 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) { 902 // Cloning these guys is unlikely to win 903 int n_op = n->Opcode(); 904 if( n_op == Op_MergeMem ) return n; 905 if( n->is_Proj() ) return n; 906 // Do not clone-up CmpFXXX variations, as these are always 907 // followed by a CmpI 908 if( n->is_Cmp() ) return n; 909 // Attempt to use a conditional move instead of a phi/branch 910 if( ConditionalMoveLimit > 0 && n_op == Op_Region ) { 911 Node *cmov = conditional_move( n ); 912 if( cmov ) return cmov; 913 } 914 if( n->is_CFG() || n->is_LoadStore() ) 915 return n; 916 if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd 917 n_op == Op_Opaque2 ) { 918 if( !C->major_progress() ) // If chance of no more loop opts... 919 _igvn._worklist.push(n); // maybe we'll remove them 920 return n; 921 } 922 923 if( n->is_Con() ) return n; // No cloning for Con nodes 924 925 Node *n_ctrl = get_ctrl(n); 926 if( !n_ctrl ) return n; // Dead node 927 928 Node* res = try_move_store_before_loop(n, n_ctrl); 929 if (res != NULL) { 930 return n; 931 } 932 933 // Attempt to remix address expressions for loop invariants 934 Node *m = remix_address_expressions( n ); 935 if( m ) return m; 936 937 // Determine if the Node has inputs from some local Phi. 938 // Returns the block to clone thru. 939 Node *n_blk = has_local_phi_input( n ); 940 if( !n_blk ) return n; 941 942 // Do not clone the trip counter through on a CountedLoop 943 // (messes up the canonical shape). 944 if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n; 945 946 // Check for having no control input; not pinned. Allow 947 // dominating control. 948 if (n->in(0)) { 949 Node *dom = idom(n_blk); 950 if (dom_lca(n->in(0), dom) != n->in(0)) { 951 return n; 952 } 953 } 954 // Policy: when is it profitable. You must get more wins than 955 // policy before it is considered profitable. Policy is usually 0, 956 // so 1 win is considered profitable. Big merges will require big 957 // cloning, so get a larger policy. 958 int policy = n_blk->req() >> 2; 959 960 // If the loop is a candidate for range check elimination, 961 // delay splitting through it's phi until a later loop optimization 962 if (n_blk->is_CountedLoop()) { 963 IdealLoopTree *lp = get_loop(n_blk); 964 if (lp && lp->_rce_candidate) { 965 return n; 966 } 967 } 968 969 // Use same limit as split_if_with_blocks_post 970 if( C->live_nodes() > 35000 ) return n; // Method too big 971 972 // Split 'n' through the merge point if it is profitable 973 Node *phi = split_thru_phi( n, n_blk, policy ); 974 if (!phi) return n; 975 976 // Found a Phi to split thru! 977 // Replace 'n' with the new phi 978 split_mem_thru_phi(n, n_blk, phi); 979 _igvn.replace_node( n, phi ); 980 // Moved a load around the loop, 'en-registering' something. 981 if (n_blk->is_Loop() && n->is_Load() && 982 !phi->in(LoopNode::LoopBackControl)->is_Load()) 983 C->set_major_progress(); 984 985 // Moved a barrier around the loop, 'en-registering' something. 986 if (n_blk->is_Loop() && n->is_ShenandoahBarrier() && 987 !phi->in(LoopNode::LoopBackControl)->is_ShenandoahBarrier()) 988 C->set_major_progress(); 989 990 return phi; 991 } 992 993 static bool merge_point_too_heavy(Compile* C, Node* region) { 994 // Bail out if the region and its phis have too many users. 995 int weight = 0; 996 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 997 weight += region->fast_out(i)->outcnt(); 998 } 999 int nodes_left = C->max_node_limit() - C->live_nodes(); 1000 if (weight * 8 > nodes_left) { 1001 #ifndef PRODUCT 1002 if (PrintOpto) 1003 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight); 1004 #endif 1005 return true; 1006 } else { 1007 return false; 1008 } 1009 } 1010 1011 static bool merge_point_safe(Node* region) { 1012 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode 1013 // having a PhiNode input. This sidesteps the dangerous case where the split 1014 // ConvI2LNode may become TOP if the input Value() does not 1015 // overlap the ConvI2L range, leaving a node which may not dominate its 1016 // uses. 1017 // A better fix for this problem can be found in the BugTraq entry, but 1018 // expediency for Mantis demands this hack. 1019 // 6855164: If the merge point has a FastLockNode with a PhiNode input, we stop 1020 // split_if_with_blocks from splitting a block because we could not move around 1021 // the FastLockNode. 1022 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1023 Node* n = region->fast_out(i); 1024 if (n->is_Phi()) { 1025 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 1026 Node* m = n->fast_out(j); 1027 if (m->is_FastLock()) 1028 return false; 1029 #ifdef _LP64 1030 if (m->Opcode() == Op_ConvI2L) 1031 return false; 1032 #endif 1033 } 1034 } 1035 } 1036 return true; 1037 } 1038 1039 1040 //------------------------------place_near_use--------------------------------- 1041 // Place some computation next to use but not inside inner loops. 1042 // For inner loop uses move it to the preheader area. 1043 Node *PhaseIdealLoop::place_near_use( Node *useblock ) const { 1044 IdealLoopTree *u_loop = get_loop( useblock ); 1045 return (u_loop->_irreducible || u_loop->_child) 1046 ? useblock 1047 : u_loop->_head->in(LoopNode::EntryControl); 1048 } 1049 1050 1051 //------------------------------split_if_with_blocks_post---------------------- 1052 // Do the real work in a non-recursive function. CFG hackery wants to be 1053 // in the post-order, so it can dirty the I-DOM info and not use the dirtied 1054 // info. 1055 void PhaseIdealLoop::split_if_with_blocks_post( Node *n ) { 1056 1057 // Cloning Cmp through Phi's involves the split-if transform. 1058 // FastLock is not used by an If 1059 if( n->is_Cmp() && !n->is_FastLock() ) { 1060 if( C->live_nodes() > 35000 ) return; // Method too big 1061 1062 // Do not do 'split-if' if irreducible loops are present. 1063 if( _has_irreducible_loops ) 1064 return; 1065 1066 Node *n_ctrl = get_ctrl(n); 1067 // Determine if the Node has inputs from some local Phi. 1068 // Returns the block to clone thru. 1069 Node *n_blk = has_local_phi_input( n ); 1070 if( n_blk != n_ctrl ) return; 1071 1072 if( merge_point_too_heavy(C, n_ctrl) ) 1073 return; 1074 1075 if( n->outcnt() != 1 ) return; // Multiple bool's from 1 compare? 1076 Node *bol = n->unique_out(); 1077 assert( bol->is_Bool(), "expect a bool here" ); 1078 if( bol->outcnt() != 1 ) return;// Multiple branches from 1 compare? 1079 Node *iff = bol->unique_out(); 1080 1081 // Check some safety conditions 1082 if( iff->is_If() ) { // Classic split-if? 1083 if( iff->in(0) != n_ctrl ) return; // Compare must be in same blk as if 1084 } else if (iff->is_CMove()) { // Trying to split-up a CMOVE 1085 // Can't split CMove with different control edge. 1086 if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) return; 1087 if( get_ctrl(iff->in(2)) == n_ctrl || 1088 get_ctrl(iff->in(3)) == n_ctrl ) 1089 return; // Inputs not yet split-up 1090 if ( get_loop(n_ctrl) != get_loop(get_ctrl(iff)) ) { 1091 return; // Loop-invar test gates loop-varying CMOVE 1092 } 1093 } else { 1094 return; // some other kind of node, such as an Allocate 1095 } 1096 1097 // Do not do 'split-if' if some paths are dead. First do dead code 1098 // elimination and then see if its still profitable. 1099 for( uint i = 1; i < n_ctrl->req(); i++ ) 1100 if( n_ctrl->in(i) == C->top() ) 1101 return; 1102 1103 // When is split-if profitable? Every 'win' on means some control flow 1104 // goes dead, so it's almost always a win. 1105 int policy = 0; 1106 // If trying to do a 'Split-If' at the loop head, it is only 1107 // profitable if the cmp folds up on BOTH paths. Otherwise we 1108 // risk peeling a loop forever. 1109 1110 // CNC - Disabled for now. Requires careful handling of loop 1111 // body selection for the cloned code. Also, make sure we check 1112 // for any input path not being in the same loop as n_ctrl. For 1113 // irreducible loops we cannot check for 'n_ctrl->is_Loop()' 1114 // because the alternative loop entry points won't be converted 1115 // into LoopNodes. 1116 IdealLoopTree *n_loop = get_loop(n_ctrl); 1117 for( uint j = 1; j < n_ctrl->req(); j++ ) 1118 if( get_loop(n_ctrl->in(j)) != n_loop ) 1119 return; 1120 1121 // Check for safety of the merge point. 1122 if( !merge_point_safe(n_ctrl) ) { 1123 return; 1124 } 1125 1126 // Split compare 'n' through the merge point if it is profitable 1127 Node *phi = split_thru_phi( n, n_ctrl, policy ); 1128 if( !phi ) return; 1129 1130 // Found a Phi to split thru! 1131 // Replace 'n' with the new phi 1132 assert(n->Opcode() != Op_ShenandoahWriteBarrier, "not with write barriers yet"); 1133 _igvn.replace_node( n, phi ); 1134 1135 // Now split the bool up thru the phi 1136 assert(bol->Opcode() != Op_ShenandoahWriteBarrier, "not with write barriers yet"); 1137 Node *bolphi = split_thru_phi( bol, n_ctrl, -1 ); 1138 guarantee(bolphi != NULL, "null boolean phi node"); 1139 1140 _igvn.replace_node( bol, bolphi ); 1141 assert( iff->in(1) == bolphi, "" ); 1142 1143 if( bolphi->Value(&_igvn)->singleton() ) 1144 return; 1145 1146 // Conditional-move? Must split up now 1147 if( !iff->is_If() ) { 1148 assert(iff->Opcode() != Op_ShenandoahWriteBarrier, "not with write barriers yet"); 1149 Node *cmovphi = split_thru_phi( iff, n_ctrl, -1 ); 1150 _igvn.replace_node( iff, cmovphi ); 1151 return; 1152 } 1153 1154 // Now split the IF 1155 do_split_if( iff ); 1156 return; 1157 } 1158 1159 // Check for an IF ready to split; one that has its 1160 // condition codes input coming from a Phi at the block start. 1161 int n_op = n->Opcode(); 1162 1163 // Check for an IF being dominated by another IF same test 1164 if (n_op == Op_If) { 1165 Node *bol = n->in(1); 1166 uint max = bol->outcnt(); 1167 // Check for same test used more than once? 1168 if (max > 1 && bol->is_Bool()) { 1169 // Search up IDOMs to see if this IF is dominated. 1170 Node *cutoff = get_ctrl(bol); 1171 1172 // Now search up IDOMs till cutoff, looking for a dominating test 1173 Node *prevdom = n; 1174 Node *dom = idom(prevdom); 1175 while (dom != cutoff) { 1176 if (dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom) { 1177 // Replace the dominated test with an obvious true or false. 1178 // Place it on the IGVN worklist for later cleanup. 1179 C->set_major_progress(); 1180 dominated_by(prevdom, n, false, true); 1181 #ifndef PRODUCT 1182 if( VerifyLoopOptimizations ) verify(); 1183 #endif 1184 return; 1185 } 1186 prevdom = dom; 1187 dom = idom(prevdom); 1188 } 1189 } 1190 } 1191 1192 // See if a shared loop-varying computation has no loop-varying uses. 1193 // Happens if something is only used for JVM state in uncommon trap exits, 1194 // like various versions of induction variable+offset. Clone the 1195 // computation per usage to allow it to sink out of the loop. 1196 if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about) 1197 Node *n_ctrl = get_ctrl(n); 1198 IdealLoopTree *n_loop = get_loop(n_ctrl); 1199 if( n_loop != _ltree_root ) { 1200 DUIterator_Fast imax, i = n->fast_outs(imax); 1201 for (; i < imax; i++) { 1202 Node* u = n->fast_out(i); 1203 if( !has_ctrl(u) ) break; // Found control user 1204 IdealLoopTree *u_loop = get_loop(get_ctrl(u)); 1205 if( u_loop == n_loop ) break; // Found loop-varying use 1206 if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop 1207 if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003 1208 } 1209 bool did_break = (i < imax); // Did we break out of the previous loop? 1210 if (!did_break && n->outcnt() > 1) { // All uses in outer loops! 1211 Node *late_load_ctrl = NULL; 1212 if (n->is_Load()) { 1213 // If n is a load, get and save the result from get_late_ctrl(), 1214 // to be later used in calculating the control for n's clones. 1215 clear_dom_lca_tags(); 1216 late_load_ctrl = get_late_ctrl(n, n_ctrl); 1217 } 1218 // If n is a load, and the late control is the same as the current 1219 // control, then the cloning of n is a pointless exercise, because 1220 // GVN will ensure that we end up where we started. 1221 if (!n->is_Load() || late_load_ctrl != n_ctrl) { 1222 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) { 1223 Node *u = n->last_out(j); // Clone private computation per use 1224 _igvn.rehash_node_delayed(u); 1225 Node *x = n->clone(); // Clone computation 1226 Node *x_ctrl = NULL; 1227 if( u->is_Phi() ) { 1228 // Replace all uses of normal nodes. Replace Phi uses 1229 // individually, so the separate Nodes can sink down 1230 // different paths. 1231 uint k = 1; 1232 while( u->in(k) != n ) k++; 1233 u->set_req( k, x ); 1234 // x goes next to Phi input path 1235 x_ctrl = u->in(0)->in(k); 1236 --j; 1237 } else { // Normal use 1238 // Replace all uses 1239 for( uint k = 0; k < u->req(); k++ ) { 1240 if( u->in(k) == n ) { 1241 u->set_req( k, x ); 1242 --j; 1243 } 1244 } 1245 x_ctrl = get_ctrl(u); 1246 } 1247 1248 // Find control for 'x' next to use but not inside inner loops. 1249 // For inner loop uses get the preheader area. 1250 x_ctrl = place_near_use(x_ctrl); 1251 1252 if (n->is_Load()) { 1253 // For loads, add a control edge to a CFG node outside of the loop 1254 // to force them to not combine and return back inside the loop 1255 // during GVN optimization (4641526). 1256 // 1257 // Because we are setting the actual control input, factor in 1258 // the result from get_late_ctrl() so we respect any 1259 // anti-dependences. (6233005). 1260 x_ctrl = dom_lca(late_load_ctrl, x_ctrl); 1261 1262 // Don't allow the control input to be a CFG splitting node. 1263 // Such nodes should only have ProjNodes as outs, e.g. IfNode 1264 // should only have IfTrueNode and IfFalseNode (4985384). 1265 x_ctrl = find_non_split_ctrl(x_ctrl); 1266 assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone"); 1267 1268 x->set_req(0, x_ctrl); 1269 } 1270 register_new_node(x, x_ctrl); 1271 1272 // Some institutional knowledge is needed here: 'x' is 1273 // yanked because if the optimizer runs GVN on it all the 1274 // cloned x's will common up and undo this optimization and 1275 // be forced back in the loop. This is annoying because it 1276 // makes +VerifyOpto report false-positives on progress. I 1277 // tried setting control edges on the x's to force them to 1278 // not combine, but the matching gets worried when it tries 1279 // to fold a StoreP and an AddP together (as part of an 1280 // address expression) and the AddP and StoreP have 1281 // different controls. 1282 if (!x->is_Load() && !x->is_DecodeNarrowPtr() && !x->is_ShenandoahBarrier()) _igvn._worklist.yank(x); 1283 } 1284 _igvn.remove_dead_node(n); 1285 } 1286 } 1287 } 1288 } 1289 1290 try_move_store_after_loop(n); 1291 1292 // Check for Opaque2's who's loop has disappeared - who's input is in the 1293 // same loop nest as their output. Remove 'em, they are no longer useful. 1294 if( n_op == Op_Opaque2 && 1295 n->in(1) != NULL && 1296 get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) { 1297 _igvn.replace_node( n, n->in(1) ); 1298 } 1299 } 1300 1301 //------------------------------split_if_with_blocks--------------------------- 1302 // Check for aggressive application of 'split-if' optimization, 1303 // using basic block level info. 1304 void PhaseIdealLoop::split_if_with_blocks( VectorSet &visited, Node_Stack &nstack ) { 1305 Node *n = C->root(); 1306 visited.set(n->_idx); // first, mark node as visited 1307 // Do pre-visit work for root 1308 n = split_if_with_blocks_pre( n ); 1309 uint cnt = n->outcnt(); 1310 uint i = 0; 1311 while (true) { 1312 // Visit all children 1313 if (i < cnt) { 1314 Node* use = n->raw_out(i); 1315 ++i; 1316 if (use->outcnt() != 0 && !visited.test_set(use->_idx)) { 1317 // Now do pre-visit work for this use 1318 use = split_if_with_blocks_pre( use ); 1319 nstack.push(n, i); // Save parent and next use's index. 1320 n = use; // Process all children of current use. 1321 cnt = use->outcnt(); 1322 i = 0; 1323 } 1324 } 1325 else { 1326 // All of n's children have been processed, complete post-processing. 1327 if (cnt != 0 && !n->is_Con()) { 1328 assert(has_node(n), "no dead nodes"); 1329 split_if_with_blocks_post( n ); 1330 } 1331 if (nstack.is_empty()) { 1332 // Finished all nodes on stack. 1333 break; 1334 } 1335 // Get saved parent node and next use's index. Visit the rest of uses. 1336 n = nstack.node(); 1337 cnt = n->outcnt(); 1338 i = nstack.index(); 1339 nstack.pop(); 1340 } 1341 } 1342 } 1343 1344 1345 //============================================================================= 1346 // 1347 // C L O N E A L O O P B O D Y 1348 // 1349 1350 //------------------------------clone_iff-------------------------------------- 1351 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1352 // "Nearly" because all Nodes have been cloned from the original in the loop, 1353 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1354 // through the Phi recursively, and return a Bool. 1355 BoolNode *PhaseIdealLoop::clone_iff( PhiNode *phi, IdealLoopTree *loop ) { 1356 1357 // Convert this Phi into a Phi merging Bools 1358 uint i; 1359 for( i = 1; i < phi->req(); i++ ) { 1360 Node *b = phi->in(i); 1361 if( b->is_Phi() ) { 1362 _igvn.replace_input_of(phi, i, clone_iff( b->as_Phi(), loop )); 1363 } else { 1364 assert( b->is_Bool(), "" ); 1365 } 1366 } 1367 1368 Node *sample_bool = phi->in(1); 1369 Node *sample_cmp = sample_bool->in(1); 1370 1371 // Make Phis to merge the Cmp's inputs. 1372 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP ); 1373 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP ); 1374 for( i = 1; i < phi->req(); i++ ) { 1375 Node *n1 = phi->in(i)->in(1)->in(1); 1376 Node *n2 = phi->in(i)->in(1)->in(2); 1377 phi1->set_req( i, n1 ); 1378 phi2->set_req( i, n2 ); 1379 phi1->set_type( phi1->type()->meet_speculative(n1->bottom_type())); 1380 phi2->set_type( phi2->type()->meet_speculative(n2->bottom_type())); 1381 } 1382 // See if these Phis have been made before. 1383 // Register with optimizer 1384 Node *hit1 = _igvn.hash_find_insert(phi1); 1385 if( hit1 ) { // Hit, toss just made Phi 1386 _igvn.remove_dead_node(phi1); // Remove new phi 1387 assert( hit1->is_Phi(), "" ); 1388 phi1 = (PhiNode*)hit1; // Use existing phi 1389 } else { // Miss 1390 _igvn.register_new_node_with_optimizer(phi1); 1391 } 1392 Node *hit2 = _igvn.hash_find_insert(phi2); 1393 if( hit2 ) { // Hit, toss just made Phi 1394 _igvn.remove_dead_node(phi2); // Remove new phi 1395 assert( hit2->is_Phi(), "" ); 1396 phi2 = (PhiNode*)hit2; // Use existing phi 1397 } else { // Miss 1398 _igvn.register_new_node_with_optimizer(phi2); 1399 } 1400 // Register Phis with loop/block info 1401 set_ctrl(phi1, phi->in(0)); 1402 set_ctrl(phi2, phi->in(0)); 1403 // Make a new Cmp 1404 Node *cmp = sample_cmp->clone(); 1405 cmp->set_req( 1, phi1 ); 1406 cmp->set_req( 2, phi2 ); 1407 _igvn.register_new_node_with_optimizer(cmp); 1408 set_ctrl(cmp, phi->in(0)); 1409 1410 // Make a new Bool 1411 Node *b = sample_bool->clone(); 1412 b->set_req(1,cmp); 1413 _igvn.register_new_node_with_optimizer(b); 1414 set_ctrl(b, phi->in(0)); 1415 1416 assert( b->is_Bool(), "" ); 1417 return (BoolNode*)b; 1418 } 1419 1420 //------------------------------clone_bool------------------------------------- 1421 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1422 // "Nearly" because all Nodes have been cloned from the original in the loop, 1423 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1424 // through the Phi recursively, and return a Bool. 1425 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) { 1426 uint i; 1427 // Convert this Phi into a Phi merging Bools 1428 for( i = 1; i < phi->req(); i++ ) { 1429 Node *b = phi->in(i); 1430 if( b->is_Phi() ) { 1431 _igvn.replace_input_of(phi, i, clone_bool( b->as_Phi(), loop )); 1432 } else { 1433 assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" ); 1434 } 1435 } 1436 1437 Node *sample_cmp = phi->in(1); 1438 1439 // Make Phis to merge the Cmp's inputs. 1440 PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP ); 1441 PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP ); 1442 for( uint j = 1; j < phi->req(); j++ ) { 1443 Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP 1444 Node *n1, *n2; 1445 if( cmp_top->is_Cmp() ) { 1446 n1 = cmp_top->in(1); 1447 n2 = cmp_top->in(2); 1448 } else { 1449 n1 = n2 = cmp_top; 1450 } 1451 phi1->set_req( j, n1 ); 1452 phi2->set_req( j, n2 ); 1453 phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); 1454 phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); 1455 } 1456 1457 // See if these Phis have been made before. 1458 // Register with optimizer 1459 Node *hit1 = _igvn.hash_find_insert(phi1); 1460 if( hit1 ) { // Hit, toss just made Phi 1461 _igvn.remove_dead_node(phi1); // Remove new phi 1462 assert( hit1->is_Phi(), "" ); 1463 phi1 = (PhiNode*)hit1; // Use existing phi 1464 } else { // Miss 1465 _igvn.register_new_node_with_optimizer(phi1); 1466 } 1467 Node *hit2 = _igvn.hash_find_insert(phi2); 1468 if( hit2 ) { // Hit, toss just made Phi 1469 _igvn.remove_dead_node(phi2); // Remove new phi 1470 assert( hit2->is_Phi(), "" ); 1471 phi2 = (PhiNode*)hit2; // Use existing phi 1472 } else { // Miss 1473 _igvn.register_new_node_with_optimizer(phi2); 1474 } 1475 // Register Phis with loop/block info 1476 set_ctrl(phi1, phi->in(0)); 1477 set_ctrl(phi2, phi->in(0)); 1478 // Make a new Cmp 1479 Node *cmp = sample_cmp->clone(); 1480 cmp->set_req( 1, phi1 ); 1481 cmp->set_req( 2, phi2 ); 1482 _igvn.register_new_node_with_optimizer(cmp); 1483 set_ctrl(cmp, phi->in(0)); 1484 1485 assert( cmp->is_Cmp(), "" ); 1486 return (CmpNode*)cmp; 1487 } 1488 1489 //------------------------------sink_use--------------------------------------- 1490 // If 'use' was in the loop-exit block, it now needs to be sunk 1491 // below the post-loop merge point. 1492 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) { 1493 if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) { 1494 set_ctrl(use, post_loop); 1495 for (DUIterator j = use->outs(); use->has_out(j); j++) 1496 sink_use(use->out(j), post_loop); 1497 } 1498 } 1499 1500 //------------------------------clone_loop------------------------------------- 1501 // 1502 // C L O N E A L O O P B O D Y 1503 // 1504 // This is the basic building block of the loop optimizations. It clones an 1505 // entire loop body. It makes an old_new loop body mapping; with this mapping 1506 // you can find the new-loop equivalent to an old-loop node. All new-loop 1507 // nodes are exactly equal to their old-loop counterparts, all edges are the 1508 // same. All exits from the old-loop now have a RegionNode that merges the 1509 // equivalent new-loop path. This is true even for the normal "loop-exit" 1510 // condition. All uses of loop-invariant old-loop values now come from (one 1511 // or more) Phis that merge their new-loop equivalents. 1512 // 1513 // This operation leaves the graph in an illegal state: there are two valid 1514 // control edges coming from the loop pre-header to both loop bodies. I'll 1515 // definitely have to hack the graph after running this transform. 1516 // 1517 // From this building block I will further edit edges to perform loop peeling 1518 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc. 1519 // 1520 // Parameter side_by_size_idom: 1521 // When side_by_size_idom is NULL, the dominator tree is constructed for 1522 // the clone loop to dominate the original. Used in construction of 1523 // pre-main-post loop sequence. 1524 // When nonnull, the clone and original are side-by-side, both are 1525 // dominated by the side_by_side_idom node. Used in construction of 1526 // unswitched loops. 1527 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd, 1528 Node* side_by_side_idom) { 1529 1530 #ifndef PRODUCT 1531 if (C->do_vector_loop() && PrintOpto) { 1532 const char* mname = C->method()->name()->as_quoted_ascii(); 1533 if (mname != NULL) { 1534 tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname); 1535 } 1536 } 1537 #endif 1538 1539 CloneMap& cm = C->clone_map(); 1540 Dict* dict = cm.dict(); 1541 if (C->do_vector_loop()) { 1542 cm.set_clone_idx(cm.max_gen()+1); 1543 #ifndef PRODUCT 1544 if (PrintOpto) { 1545 tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx()); 1546 loop->dump_head(); 1547 } 1548 #endif 1549 } 1550 1551 // Step 1: Clone the loop body. Make the old->new mapping. 1552 uint i; 1553 for( i = 0; i < loop->_body.size(); i++ ) { 1554 Node *old = loop->_body.at(i); 1555 #ifdef ASSERT 1556 if (old->Opcode() == Op_ShenandoahWriteBarrier) { 1557 Node* memproj = old->find_out_with(Op_ShenandoahWBMemProj); 1558 assert(memproj == NULL || (loop->is_member(get_loop(has_ctrl(memproj) ? get_ctrl(memproj) : memproj))), "WB's mem-proj must be in loop too"); 1559 } 1560 if (old->Opcode() == Op_ShenandoahWBMemProj) { 1561 Node* wb = old->in(0); 1562 assert(loop->is_member(get_loop(has_ctrl(wb) ? get_ctrl(wb) : wb)), "WB must be in loop too"); 1563 } 1564 #endif 1565 Node *nnn = old->clone(); 1566 old_new.map( old->_idx, nnn ); 1567 if (C->do_vector_loop()) { 1568 cm.verify_insert_and_clone(old, nnn, cm.clone_idx()); 1569 } 1570 _igvn.register_new_node_with_optimizer(nnn); 1571 } 1572 1573 1574 // Step 2: Fix the edges in the new body. If the old input is outside the 1575 // loop use it. If the old input is INside the loop, use the corresponding 1576 // new node instead. 1577 for( i = 0; i < loop->_body.size(); i++ ) { 1578 Node *old = loop->_body.at(i); 1579 Node *nnn = old_new[old->_idx]; 1580 // Fix CFG/Loop controlling the new node 1581 if (has_ctrl(old)) { 1582 set_ctrl(nnn, old_new[get_ctrl(old)->_idx]); 1583 } else { 1584 set_loop(nnn, loop->_parent); 1585 if (old->outcnt() > 0) { 1586 set_idom( nnn, old_new[idom(old)->_idx], dd ); 1587 } 1588 } 1589 // Correct edges to the new node 1590 for( uint j = 0; j < nnn->req(); j++ ) { 1591 Node *n = nnn->in(j); 1592 if( n ) { 1593 IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n ); 1594 if( loop->is_member( old_in_loop ) ) 1595 nnn->set_req(j, old_new[n->_idx]); 1596 } 1597 } 1598 _igvn.hash_find_insert(nnn); 1599 } 1600 Node *newhead = old_new[loop->_head->_idx]; 1601 set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); 1602 1603 1604 // Step 3: Now fix control uses. Loop varying control uses have already 1605 // been fixed up (as part of all input edges in Step 2). Loop invariant 1606 // control uses must be either an IfFalse or an IfTrue. Make a merge 1607 // point to merge the old and new IfFalse/IfTrue nodes; make the use 1608 // refer to this. 1609 ResourceArea *area = Thread::current()->resource_area(); 1610 Node_List worklist(area); 1611 uint new_counter = C->unique(); 1612 for( i = 0; i < loop->_body.size(); i++ ) { 1613 Node* old = loop->_body.at(i); 1614 if( !old->is_CFG() ) continue; 1615 Node* nnn = old_new[old->_idx]; 1616 1617 // Copy uses to a worklist, so I can munge the def-use info 1618 // with impunity. 1619 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1620 worklist.push(old->fast_out(j)); 1621 1622 while( worklist.size() ) { // Visit all uses 1623 Node *use = worklist.pop(); 1624 if (!has_node(use)) continue; // Ignore dead nodes 1625 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1626 if( !loop->is_member( use_loop ) && use->is_CFG() ) { 1627 // Both OLD and USE are CFG nodes here. 1628 assert( use->is_Proj(), "" ); 1629 1630 // Clone the loop exit control projection 1631 Node *newuse = use->clone(); 1632 if (C->do_vector_loop()) { 1633 cm.verify_insert_and_clone(use, newuse, cm.clone_idx()); 1634 } 1635 newuse->set_req(0,nnn); 1636 _igvn.register_new_node_with_optimizer(newuse); 1637 set_loop(newuse, use_loop); 1638 set_idom(newuse, nnn, dom_depth(nnn) + 1 ); 1639 1640 // We need a Region to merge the exit from the peeled body and the 1641 // exit from the old loop body. 1642 RegionNode *r = new RegionNode(3); 1643 // Map the old use to the new merge point 1644 old_new.map( use->_idx, r ); 1645 uint dd_r = MIN2(dom_depth(newuse),dom_depth(use)); 1646 assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" ); 1647 1648 // The original user of 'use' uses 'r' instead. 1649 for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) { 1650 Node* useuse = use->last_out(l); 1651 _igvn.rehash_node_delayed(useuse); 1652 uint uses_found = 0; 1653 if( useuse->in(0) == use ) { 1654 useuse->set_req(0, r); 1655 uses_found++; 1656 if( useuse->is_CFG() ) { 1657 assert( dom_depth(useuse) > dd_r, "" ); 1658 set_idom(useuse, r, dom_depth(useuse)); 1659 } 1660 } 1661 for( uint k = 1; k < useuse->req(); k++ ) { 1662 if( useuse->in(k) == use ) { 1663 useuse->set_req(k, r); 1664 uses_found++; 1665 } 1666 } 1667 l -= uses_found; // we deleted 1 or more copies of this edge 1668 } 1669 1670 // Now finish up 'r' 1671 r->set_req( 1, newuse ); 1672 r->set_req( 2, use ); 1673 _igvn.register_new_node_with_optimizer(r); 1674 set_loop(r, use_loop); 1675 set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r); 1676 } // End of if a loop-exit test 1677 } 1678 } 1679 1680 // Step 4: If loop-invariant use is not control, it must be dominated by a 1681 // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region 1682 // there if needed. Make a Phi there merging old and new used values. 1683 Node_List *split_if_set = NULL; 1684 Node_List *split_bool_set = NULL; 1685 Node_List *split_cex_set = NULL; 1686 for( i = 0; i < loop->_body.size(); i++ ) { 1687 Node* old = loop->_body.at(i); 1688 Node* nnn = old_new[old->_idx]; 1689 // Copy uses to a worklist, so I can munge the def-use info 1690 // with impunity. 1691 for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) 1692 worklist.push(old->fast_out(j)); 1693 1694 while( worklist.size() ) { 1695 Node *use = worklist.pop(); 1696 if (!has_node(use)) continue; // Ignore dead nodes 1697 if (use->in(0) == C->top()) continue; 1698 IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); 1699 // Check for data-use outside of loop - at least one of OLD or USE 1700 // must not be a CFG node. 1701 if( !loop->is_member( use_loop ) && (!old->is_CFG() || !use->is_CFG())) { 1702 1703 // If the Data use is an IF, that means we have an IF outside of the 1704 // loop that is switching on a condition that is set inside of the 1705 // loop. Happens if people set a loop-exit flag; then test the flag 1706 // in the loop to break the loop, then test is again outside of the 1707 // loop to determine which way the loop exited. 1708 // Loop predicate If node connects to Bool node through Opaque1 node. 1709 if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use)) { 1710 // Since this code is highly unlikely, we lazily build the worklist 1711 // of such Nodes to go split. 1712 if( !split_if_set ) 1713 split_if_set = new Node_List(area); 1714 split_if_set->push(use); 1715 } 1716 if( use->is_Bool() ) { 1717 if( !split_bool_set ) 1718 split_bool_set = new Node_List(area); 1719 split_bool_set->push(use); 1720 } 1721 if( use->Opcode() == Op_CreateEx ) { 1722 if( !split_cex_set ) 1723 split_cex_set = new Node_List(area); 1724 split_cex_set->push(use); 1725 } 1726 1727 1728 // Get "block" use is in 1729 uint idx = 0; 1730 while( use->in(idx) != old ) idx++; 1731 Node *prev = use->is_CFG() ? use : get_ctrl(use); 1732 assert( !loop->is_member( get_loop( prev ) ), "" ); 1733 Node *cfg = prev->_idx >= new_counter 1734 ? prev->in(2) 1735 : idom(prev); 1736 if( use->is_Phi() ) // Phi use is in prior block 1737 cfg = prev->in(idx); // NOT in block of Phi itself 1738 if (cfg->is_top()) { // Use is dead? 1739 _igvn.replace_input_of(use, idx, C->top()); 1740 continue; 1741 } 1742 1743 while( !loop->is_member( get_loop( cfg ) ) ) { 1744 prev = cfg; 1745 cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg); 1746 } 1747 // If the use occurs after merging several exits from the loop, then 1748 // old value must have dominated all those exits. Since the same old 1749 // value was used on all those exits we did not need a Phi at this 1750 // merge point. NOW we do need a Phi here. Each loop exit value 1751 // is now merged with the peeled body exit; each exit gets its own 1752 // private Phi and those Phis need to be merged here. 1753 Node *phi; 1754 if( prev->is_Region() ) { 1755 if( idx == 0 && use->Opcode() != Op_ShenandoahWBMemProj) { // Updating control edge? 1756 phi = prev; // Just use existing control 1757 } else { // Else need a new Phi 1758 phi = PhiNode::make( prev, old ); 1759 // Now recursively fix up the new uses of old! 1760 uint first = use->Opcode() != Op_ShenandoahWBMemProj ? 1 : 0; 1761 for( uint i = first; i < prev->req(); i++ ) { 1762 worklist.push(phi); // Onto worklist once for each 'old' input 1763 } 1764 } 1765 } else { 1766 // Get new RegionNode merging old and new loop exits 1767 prev = old_new[prev->_idx]; 1768 assert( prev, "just made this in step 7" ); 1769 if( idx == 0 && use->Opcode() != Op_ShenandoahWBMemProj) { // Updating control edge? 1770 phi = prev; // Just use existing control 1771 } else { // Else need a new Phi 1772 // Make a new Phi merging data values properly 1773 phi = PhiNode::make( prev, old ); 1774 phi->set_req( 1, nnn ); 1775 } 1776 } 1777 // If inserting a new Phi, check for prior hits 1778 if( idx != 0 ) { 1779 Node *hit = _igvn.hash_find_insert(phi); 1780 if( hit == NULL ) { 1781 _igvn.register_new_node_with_optimizer(phi); // Register new phi 1782 } else { // or 1783 // Remove the new phi from the graph and use the hit 1784 _igvn.remove_dead_node(phi); 1785 phi = hit; // Use existing phi 1786 } 1787 set_ctrl(phi, prev); 1788 } 1789 // Make 'use' use the Phi instead of the old loop body exit value 1790 _igvn.replace_input_of(use, idx, phi); 1791 if( use->_idx >= new_counter ) { // If updating new phis 1792 // Not needed for correctness, but prevents a weak assert 1793 // in AddPNode from tripping (when we end up with different 1794 // base & derived Phis that will become the same after 1795 // IGVN does CSE). 1796 Node *hit = _igvn.hash_find_insert(use); 1797 if( hit ) // Go ahead and re-hash for hits. 1798 _igvn.replace_node( use, hit ); 1799 } 1800 1801 // If 'use' was in the loop-exit block, it now needs to be sunk 1802 // below the post-loop merge point. 1803 sink_use( use, prev ); 1804 } 1805 } 1806 } 1807 1808 // Check for IFs that need splitting/cloning. Happens if an IF outside of 1809 // the loop uses a condition set in the loop. The original IF probably 1810 // takes control from one or more OLD Regions (which in turn get from NEW 1811 // Regions). In any case, there will be a set of Phis for each merge point 1812 // from the IF up to where the original BOOL def exists the loop. 1813 if( split_if_set ) { 1814 while( split_if_set->size() ) { 1815 Node *iff = split_if_set->pop(); 1816 if( iff->in(1)->is_Phi() ) { 1817 BoolNode *b = clone_iff( iff->in(1)->as_Phi(), loop ); 1818 _igvn.replace_input_of(iff, 1, b); 1819 } 1820 } 1821 } 1822 if( split_bool_set ) { 1823 while( split_bool_set->size() ) { 1824 Node *b = split_bool_set->pop(); 1825 Node *phi = b->in(1); 1826 assert( phi->is_Phi(), "" ); 1827 CmpNode *cmp = clone_bool( (PhiNode*)phi, loop ); 1828 _igvn.replace_input_of(b, 1, cmp); 1829 } 1830 } 1831 if( split_cex_set ) { 1832 while( split_cex_set->size() ) { 1833 Node *b = split_cex_set->pop(); 1834 assert( b->in(0)->is_Region(), "" ); 1835 assert( b->in(1)->is_Phi(), "" ); 1836 assert( b->in(0)->in(0) == b->in(1)->in(0), "" ); 1837 split_up( b, b->in(0), NULL ); 1838 } 1839 } 1840 1841 } 1842 1843 1844 //---------------------- stride_of_possible_iv ------------------------------------- 1845 // Looks for an iff/bool/comp with one operand of the compare 1846 // being a cycle involving an add and a phi, 1847 // with an optional truncation (left-shift followed by a right-shift) 1848 // of the add. Returns zero if not an iv. 1849 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) { 1850 Node* trunc1 = NULL; 1851 Node* trunc2 = NULL; 1852 const TypeInt* ttype = NULL; 1853 if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) { 1854 return 0; 1855 } 1856 BoolNode* bl = iff->in(1)->as_Bool(); 1857 Node* cmp = bl->in(1); 1858 if (!cmp || cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU) { 1859 return 0; 1860 } 1861 // Must have an invariant operand 1862 if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) { 1863 return 0; 1864 } 1865 Node* add2 = NULL; 1866 Node* cmp1 = cmp->in(1); 1867 if (cmp1->is_Phi()) { 1868 // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) ))) 1869 Node* phi = cmp1; 1870 for (uint i = 1; i < phi->req(); i++) { 1871 Node* in = phi->in(i); 1872 Node* add = CountedLoopNode::match_incr_with_optional_truncation(in, 1873 &trunc1, &trunc2, &ttype); 1874 if (add && add->in(1) == phi) { 1875 add2 = add->in(2); 1876 break; 1877 } 1878 } 1879 } else { 1880 // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) ))) 1881 Node* addtrunc = cmp1; 1882 Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc, 1883 &trunc1, &trunc2, &ttype); 1884 if (add && add->in(1)->is_Phi()) { 1885 Node* phi = add->in(1); 1886 for (uint i = 1; i < phi->req(); i++) { 1887 if (phi->in(i) == addtrunc) { 1888 add2 = add->in(2); 1889 break; 1890 } 1891 } 1892 } 1893 } 1894 if (add2 != NULL) { 1895 const TypeInt* add2t = _igvn.type(add2)->is_int(); 1896 if (add2t->is_con()) { 1897 return add2t->get_con(); 1898 } 1899 } 1900 return 0; 1901 } 1902 1903 1904 //---------------------- stay_in_loop ------------------------------------- 1905 // Return the (unique) control output node that's in the loop (if it exists.) 1906 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) { 1907 Node* unique = NULL; 1908 if (!n) return NULL; 1909 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1910 Node* use = n->fast_out(i); 1911 if (!has_ctrl(use) && loop->is_member(get_loop(use))) { 1912 if (unique != NULL) { 1913 return NULL; 1914 } 1915 unique = use; 1916 } 1917 } 1918 return unique; 1919 } 1920 1921 //------------------------------ register_node ------------------------------------- 1922 // Utility to register node "n" with PhaseIdealLoop 1923 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) { 1924 _igvn.register_new_node_with_optimizer(n); 1925 loop->_body.push(n); 1926 if (n->is_CFG()) { 1927 set_loop(n, loop); 1928 set_idom(n, pred, ddepth); 1929 } else { 1930 set_ctrl(n, pred); 1931 } 1932 } 1933 1934 //------------------------------ proj_clone ------------------------------------- 1935 // Utility to create an if-projection 1936 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) { 1937 ProjNode* c = p->clone()->as_Proj(); 1938 c->set_req(0, iff); 1939 return c; 1940 } 1941 1942 //------------------------------ short_circuit_if ------------------------------------- 1943 // Force the iff control output to be the live_proj 1944 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) { 1945 guarantee(live_proj != NULL, "null projection"); 1946 int proj_con = live_proj->_con; 1947 assert(proj_con == 0 || proj_con == 1, "false or true projection"); 1948 Node *con = _igvn.intcon(proj_con); 1949 set_ctrl(con, C->root()); 1950 if (iff) { 1951 iff->set_req(1, con); 1952 } 1953 return con; 1954 } 1955 1956 //------------------------------ insert_if_before_proj ------------------------------------- 1957 // Insert a new if before an if projection (* - new node) 1958 // 1959 // before 1960 // if(test) 1961 // / \ 1962 // v v 1963 // other-proj proj (arg) 1964 // 1965 // after 1966 // if(test) 1967 // / \ 1968 // / v 1969 // | * proj-clone 1970 // v | 1971 // other-proj v 1972 // * new_if(relop(cmp[IU](left,right))) 1973 // / \ 1974 // v v 1975 // * new-proj proj 1976 // (returned) 1977 // 1978 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) { 1979 IfNode* iff = proj->in(0)->as_If(); 1980 IdealLoopTree *loop = get_loop(proj); 1981 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 1982 int ddepth = dom_depth(proj); 1983 1984 _igvn.rehash_node_delayed(iff); 1985 _igvn.rehash_node_delayed(proj); 1986 1987 proj->set_req(0, NULL); // temporary disconnect 1988 ProjNode* proj2 = proj_clone(proj, iff); 1989 register_node(proj2, loop, iff, ddepth); 1990 1991 Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right); 1992 register_node(cmp, loop, proj2, ddepth); 1993 1994 BoolNode* bol = new BoolNode(cmp, relop); 1995 register_node(bol, loop, proj2, ddepth); 1996 1997 IfNode* new_if = new IfNode(proj2, bol, iff->_prob, iff->_fcnt); 1998 register_node(new_if, loop, proj2, ddepth); 1999 2000 proj->set_req(0, new_if); // reattach 2001 set_idom(proj, new_if, ddepth); 2002 2003 ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj(); 2004 guarantee(new_exit != NULL, "null exit node"); 2005 register_node(new_exit, get_loop(other_proj), new_if, ddepth); 2006 2007 return new_exit; 2008 } 2009 2010 //------------------------------ insert_region_before_proj ------------------------------------- 2011 // Insert a region before an if projection (* - new node) 2012 // 2013 // before 2014 // if(test) 2015 // / | 2016 // v | 2017 // proj v 2018 // other-proj 2019 // 2020 // after 2021 // if(test) 2022 // / | 2023 // v | 2024 // * proj-clone v 2025 // | other-proj 2026 // v 2027 // * new-region 2028 // | 2029 // v 2030 // * dum_if 2031 // / \ 2032 // v \ 2033 // * dum-proj v 2034 // proj 2035 // 2036 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) { 2037 IfNode* iff = proj->in(0)->as_If(); 2038 IdealLoopTree *loop = get_loop(proj); 2039 ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); 2040 int ddepth = dom_depth(proj); 2041 2042 _igvn.rehash_node_delayed(iff); 2043 _igvn.rehash_node_delayed(proj); 2044 2045 proj->set_req(0, NULL); // temporary disconnect 2046 ProjNode* proj2 = proj_clone(proj, iff); 2047 register_node(proj2, loop, iff, ddepth); 2048 2049 RegionNode* reg = new RegionNode(2); 2050 reg->set_req(1, proj2); 2051 register_node(reg, loop, iff, ddepth); 2052 2053 IfNode* dum_if = new IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt); 2054 register_node(dum_if, loop, reg, ddepth); 2055 2056 proj->set_req(0, dum_if); // reattach 2057 set_idom(proj, dum_if, ddepth); 2058 2059 ProjNode* dum_proj = proj_clone(other_proj, dum_if); 2060 register_node(dum_proj, loop, dum_if, ddepth); 2061 2062 return reg; 2063 } 2064 2065 //------------------------------ insert_cmpi_loop_exit ------------------------------------- 2066 // Clone a signed compare loop exit from an unsigned compare and 2067 // insert it before the unsigned cmp on the stay-in-loop path. 2068 // All new nodes inserted in the dominator tree between the original 2069 // if and it's projections. The original if test is replaced with 2070 // a constant to force the stay-in-loop path. 2071 // 2072 // This is done to make sure that the original if and it's projections 2073 // still dominate the same set of control nodes, that the ctrl() relation 2074 // from data nodes to them is preserved, and that their loop nesting is 2075 // preserved. 2076 // 2077 // before 2078 // if(i <u limit) unsigned compare loop exit 2079 // / | 2080 // v v 2081 // exit-proj stay-in-loop-proj 2082 // 2083 // after 2084 // if(stay-in-loop-const) original if 2085 // / | 2086 // / v 2087 // / if(i < limit) new signed test 2088 // / / | 2089 // / / v 2090 // / / if(i <u limit) new cloned unsigned test 2091 // / / / | 2092 // v v v | 2093 // region | 2094 // | | 2095 // dum-if | 2096 // / | | 2097 // ether | | 2098 // v v 2099 // exit-proj stay-in-loop-proj 2100 // 2101 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) { 2102 const bool Signed = true; 2103 const bool Unsigned = false; 2104 2105 BoolNode* bol = if_cmpu->in(1)->as_Bool(); 2106 if (bol->_test._test != BoolTest::lt) return NULL; 2107 CmpNode* cmpu = bol->in(1)->as_Cmp(); 2108 if (cmpu->Opcode() != Op_CmpU) return NULL; 2109 int stride = stride_of_possible_iv(if_cmpu); 2110 if (stride == 0) return NULL; 2111 2112 Node* lp_proj = stay_in_loop(if_cmpu, loop); 2113 guarantee(lp_proj != NULL, "null loop node"); 2114 2115 ProjNode* lp_continue = lp_proj->as_Proj(); 2116 ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj(); 2117 2118 Node* limit = NULL; 2119 if (stride > 0) { 2120 limit = cmpu->in(2); 2121 } else { 2122 limit = _igvn.makecon(TypeInt::ZERO); 2123 set_ctrl(limit, C->root()); 2124 } 2125 // Create a new region on the exit path 2126 RegionNode* reg = insert_region_before_proj(lp_exit); 2127 guarantee(reg != NULL, "null region node"); 2128 2129 // Clone the if-cmpu-true-false using a signed compare 2130 BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge; 2131 ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue); 2132 reg->add_req(cmpi_exit); 2133 2134 // Clone the if-cmpu-true-false 2135 BoolTest::mask rel_u = bol->_test._test; 2136 ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue); 2137 reg->add_req(cmpu_exit); 2138 2139 // Force original if to stay in loop. 2140 short_circuit_if(if_cmpu, lp_continue); 2141 2142 return cmpi_exit->in(0)->as_If(); 2143 } 2144 2145 //------------------------------ remove_cmpi_loop_exit ------------------------------------- 2146 // Remove a previously inserted signed compare loop exit. 2147 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) { 2148 Node* lp_proj = stay_in_loop(if_cmp, loop); 2149 assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI && 2150 stay_in_loop(lp_proj, loop)->is_If() && 2151 stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu"); 2152 Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO); 2153 set_ctrl(con, C->root()); 2154 if_cmp->set_req(1, con); 2155 } 2156 2157 //------------------------------ scheduled_nodelist ------------------------------------- 2158 // Create a post order schedule of nodes that are in the 2159 // "member" set. The list is returned in "sched". 2160 // The first node in "sched" is the loop head, followed by 2161 // nodes which have no inputs in the "member" set, and then 2162 // followed by the nodes that have an immediate input dependence 2163 // on a node in "sched". 2164 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) { 2165 2166 assert(member.test(loop->_head->_idx), "loop head must be in member set"); 2167 Arena *a = Thread::current()->resource_area(); 2168 VectorSet visited(a); 2169 Node_Stack nstack(a, loop->_body.size()); 2170 2171 Node* n = loop->_head; // top of stack is cached in "n" 2172 uint idx = 0; 2173 visited.set(n->_idx); 2174 2175 // Initially push all with no inputs from within member set 2176 for(uint i = 0; i < loop->_body.size(); i++ ) { 2177 Node *elt = loop->_body.at(i); 2178 if (member.test(elt->_idx)) { 2179 bool found = false; 2180 for (uint j = 0; j < elt->req(); j++) { 2181 Node* def = elt->in(j); 2182 if (def && member.test(def->_idx) && def != elt) { 2183 found = true; 2184 break; 2185 } 2186 } 2187 if (!found && elt != loop->_head) { 2188 nstack.push(n, idx); 2189 n = elt; 2190 assert(!visited.test(n->_idx), "not seen yet"); 2191 visited.set(n->_idx); 2192 } 2193 } 2194 } 2195 2196 // traverse out's that are in the member set 2197 while (true) { 2198 if (idx < n->outcnt()) { 2199 Node* use = n->raw_out(idx); 2200 idx++; 2201 if (!visited.test_set(use->_idx)) { 2202 if (member.test(use->_idx)) { 2203 nstack.push(n, idx); 2204 n = use; 2205 idx = 0; 2206 } 2207 } 2208 } else { 2209 // All outputs processed 2210 sched.push(n); 2211 if (nstack.is_empty()) break; 2212 n = nstack.node(); 2213 idx = nstack.index(); 2214 nstack.pop(); 2215 } 2216 } 2217 } 2218 2219 2220 //------------------------------ has_use_in_set ------------------------------------- 2221 // Has a use in the vector set 2222 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) { 2223 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2224 Node* use = n->fast_out(j); 2225 if (vset.test(use->_idx)) { 2226 return true; 2227 } 2228 } 2229 return false; 2230 } 2231 2232 2233 //------------------------------ has_use_internal_to_set ------------------------------------- 2234 // Has use internal to the vector set (ie. not in a phi at the loop head) 2235 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) { 2236 Node* head = loop->_head; 2237 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2238 Node* use = n->fast_out(j); 2239 if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) { 2240 return true; 2241 } 2242 } 2243 return false; 2244 } 2245 2246 2247 //------------------------------ clone_for_use_outside_loop ------------------------------------- 2248 // clone "n" for uses that are outside of loop 2249 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) { 2250 int cloned = 0; 2251 assert(worklist.size() == 0, "should be empty"); 2252 assert(n->Opcode() != Op_ShenandoahWriteBarrier, "not with write barriers yet"); 2253 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2254 Node* use = n->fast_out(j); 2255 if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) { 2256 worklist.push(use); 2257 } 2258 } 2259 while( worklist.size() ) { 2260 Node *use = worklist.pop(); 2261 if (!has_node(use) || use->in(0) == C->top()) continue; 2262 uint j; 2263 for (j = 0; j < use->req(); j++) { 2264 if (use->in(j) == n) break; 2265 } 2266 assert(j < use->req(), "must be there"); 2267 2268 // clone "n" and insert it between the inputs of "n" and the use outside the loop 2269 assert(n->Opcode() != Op_ShenandoahWriteBarrier, "not with write barriers"); 2270 Node* n_clone = n->clone(); 2271 _igvn.replace_input_of(use, j, n_clone); 2272 cloned++; 2273 Node* use_c; 2274 if (!use->is_Phi()) { 2275 use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0); 2276 } else { 2277 // Use in a phi is considered a use in the associated predecessor block 2278 use_c = use->in(0)->in(j); 2279 } 2280 set_ctrl(n_clone, use_c); 2281 assert(!loop->is_member(get_loop(use_c)), "should be outside loop"); 2282 get_loop(use_c)->_body.push(n_clone); 2283 _igvn.register_new_node_with_optimizer(n_clone); 2284 #if !defined(PRODUCT) 2285 if (TracePartialPeeling) { 2286 tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx); 2287 } 2288 #endif 2289 } 2290 return cloned; 2291 } 2292 2293 2294 //------------------------------ clone_for_special_use_inside_loop ------------------------------------- 2295 // clone "n" for special uses that are in the not_peeled region. 2296 // If these def-uses occur in separate blocks, the code generator 2297 // marks the method as not compilable. For example, if a "BoolNode" 2298 // is in a different basic block than the "IfNode" that uses it, then 2299 // the compilation is aborted in the code generator. 2300 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 2301 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) { 2302 if (n->is_Phi() || n->is_Load()) { 2303 return; 2304 } 2305 assert(worklist.size() == 0, "should be empty"); 2306 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2307 Node* use = n->fast_out(j); 2308 if ( not_peel.test(use->_idx) && 2309 (use->is_If() || use->is_CMove() || use->is_Bool()) && 2310 use->in(1) == n) { 2311 worklist.push(use); 2312 } 2313 } 2314 if (worklist.size() > 0) { 2315 // clone "n" and insert it between inputs of "n" and the use 2316 assert(n->Opcode() != Op_ShenandoahWriteBarrier, "not with write barriers"); 2317 Node* n_clone = n->clone(); 2318 loop->_body.push(n_clone); 2319 _igvn.register_new_node_with_optimizer(n_clone); 2320 set_ctrl(n_clone, get_ctrl(n)); 2321 sink_list.push(n_clone); 2322 not_peel <<= n_clone->_idx; // add n_clone to not_peel set. 2323 #if !defined(PRODUCT) 2324 if (TracePartialPeeling) { 2325 tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx); 2326 } 2327 #endif 2328 while( worklist.size() ) { 2329 Node *use = worklist.pop(); 2330 _igvn.rehash_node_delayed(use); 2331 for (uint j = 1; j < use->req(); j++) { 2332 if (use->in(j) == n) { 2333 use->set_req(j, n_clone); 2334 } 2335 } 2336 } 2337 } 2338 } 2339 2340 2341 //------------------------------ insert_phi_for_loop ------------------------------------- 2342 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 2343 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) { 2344 Node *phi = PhiNode::make(lp, back_edge_val); 2345 phi->set_req(LoopNode::EntryControl, lp_entry_val); 2346 // Use existing phi if it already exists 2347 Node *hit = _igvn.hash_find_insert(phi); 2348 if( hit == NULL ) { 2349 _igvn.register_new_node_with_optimizer(phi); 2350 set_ctrl(phi, lp); 2351 } else { 2352 // Remove the new phi from the graph and use the hit 2353 _igvn.remove_dead_node(phi); 2354 phi = hit; 2355 } 2356 _igvn.replace_input_of(use, idx, phi); 2357 } 2358 2359 #ifdef ASSERT 2360 //------------------------------ is_valid_loop_partition ------------------------------------- 2361 // Validate the loop partition sets: peel and not_peel 2362 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, 2363 VectorSet& not_peel ) { 2364 uint i; 2365 // Check that peel_list entries are in the peel set 2366 for (i = 0; i < peel_list.size(); i++) { 2367 if (!peel.test(peel_list.at(i)->_idx)) { 2368 return false; 2369 } 2370 } 2371 // Check at loop members are in one of peel set or not_peel set 2372 for (i = 0; i < loop->_body.size(); i++ ) { 2373 Node *def = loop->_body.at(i); 2374 uint di = def->_idx; 2375 // Check that peel set elements are in peel_list 2376 if (peel.test(di)) { 2377 if (not_peel.test(di)) { 2378 return false; 2379 } 2380 // Must be in peel_list also 2381 bool found = false; 2382 for (uint j = 0; j < peel_list.size(); j++) { 2383 if (peel_list.at(j)->_idx == di) { 2384 found = true; 2385 break; 2386 } 2387 } 2388 if (!found) { 2389 return false; 2390 } 2391 } else if (not_peel.test(di)) { 2392 if (peel.test(di)) { 2393 return false; 2394 } 2395 } else { 2396 return false; 2397 } 2398 } 2399 return true; 2400 } 2401 2402 //------------------------------ is_valid_clone_loop_exit_use ------------------------------------- 2403 // Ensure a use outside of loop is of the right form 2404 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) { 2405 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2406 return (use->is_Phi() && 2407 use_c->is_Region() && use_c->req() == 3 && 2408 (use_c->in(exit_idx)->Opcode() == Op_IfTrue || 2409 use_c->in(exit_idx)->Opcode() == Op_IfFalse || 2410 use_c->in(exit_idx)->Opcode() == Op_JumpProj) && 2411 loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) ); 2412 } 2413 2414 //------------------------------ is_valid_clone_loop_form ------------------------------------- 2415 // Ensure that all uses outside of loop are of the right form 2416 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 2417 uint orig_exit_idx, uint clone_exit_idx) { 2418 uint len = peel_list.size(); 2419 for (uint i = 0; i < len; i++) { 2420 Node *def = peel_list.at(i); 2421 2422 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 2423 Node *use = def->fast_out(j); 2424 Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; 2425 if (!loop->is_member(get_loop(use_c))) { 2426 // use is not in the loop, check for correct structure 2427 if (use->in(0) == def) { 2428 // Okay 2429 } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) { 2430 return false; 2431 } 2432 } 2433 } 2434 } 2435 return true; 2436 } 2437 #endif 2438 2439 //------------------------------ partial_peel ------------------------------------- 2440 // Partially peel (aka loop rotation) the top portion of a loop (called 2441 // the peel section below) by cloning it and placing one copy just before 2442 // the new loop head and the other copy at the bottom of the new loop. 2443 // 2444 // before after where it came from 2445 // 2446 // stmt1 stmt1 2447 // loop: stmt2 clone 2448 // stmt2 if condA goto exitA clone 2449 // if condA goto exitA new_loop: new 2450 // stmt3 stmt3 clone 2451 // if !condB goto loop if condB goto exitB clone 2452 // exitB: stmt2 orig 2453 // stmt4 if !condA goto new_loop orig 2454 // exitA: goto exitA 2455 // exitB: 2456 // stmt4 2457 // exitA: 2458 // 2459 // Step 1: find the cut point: an exit test on probable 2460 // induction variable. 2461 // Step 2: schedule (with cloning) operations in the peel 2462 // section that can be executed after the cut into 2463 // the section that is not peeled. This may need 2464 // to clone operations into exit blocks. For 2465 // instance, a reference to A[i] in the not-peel 2466 // section and a reference to B[i] in an exit block 2467 // may cause a left-shift of i by 2 to be placed 2468 // in the peel block. This step will clone the left 2469 // shift into the exit block and sink the left shift 2470 // from the peel to the not-peel section. 2471 // Step 3: clone the loop, retarget the control, and insert 2472 // phis for values that are live across the new loop 2473 // head. This is very dependent on the graph structure 2474 // from clone_loop. It creates region nodes for 2475 // exit control and associated phi nodes for values 2476 // flow out of the loop through that exit. The region 2477 // node is dominated by the clone's control projection. 2478 // So the clone's peel section is placed before the 2479 // new loop head, and the clone's not-peel section is 2480 // forms the top part of the new loop. The original 2481 // peel section forms the tail of the new loop. 2482 // Step 4: update the dominator tree and recompute the 2483 // dominator depth. 2484 // 2485 // orig 2486 // 2487 // stmt1 2488 // | 2489 // v 2490 // loop predicate 2491 // | 2492 // v 2493 // loop<----+ 2494 // | | 2495 // stmt2 | 2496 // | | 2497 // v | 2498 // ifA | 2499 // / | | 2500 // v v | 2501 // false true ^ <-- last_peel 2502 // / | | 2503 // / ===|==cut | 2504 // / stmt3 | <-- first_not_peel 2505 // / | | 2506 // | v | 2507 // v ifB | 2508 // exitA: / \ | 2509 // / \ | 2510 // v v | 2511 // false true | 2512 // / \ | 2513 // / ----+ 2514 // | 2515 // v 2516 // exitB: 2517 // stmt4 2518 // 2519 // 2520 // after clone loop 2521 // 2522 // stmt1 2523 // | 2524 // v 2525 // loop predicate 2526 // / \ 2527 // clone / \ orig 2528 // / \ 2529 // / \ 2530 // v v 2531 // +---->loop loop<----+ 2532 // | | | | 2533 // | stmt2 stmt2 | 2534 // | | | | 2535 // | v v | 2536 // | ifA ifA | 2537 // | | \ / | | 2538 // | v v v v | 2539 // ^ true false false true ^ <-- last_peel 2540 // | | ^ \ / | | 2541 // | cut==|== \ \ / ===|==cut | 2542 // | stmt3 \ \ / stmt3 | <-- first_not_peel 2543 // | | dom | | | | 2544 // | v \ 1v v2 v | 2545 // | ifB regionA ifB | 2546 // | / \ | / \ | 2547 // | / \ v / \ | 2548 // | v v exitA: v v | 2549 // | true false false true | 2550 // | / ^ \ / \ | 2551 // +---- \ \ / ----+ 2552 // dom \ / 2553 // \ 1v v2 2554 // regionB 2555 // | 2556 // v 2557 // exitB: 2558 // stmt4 2559 // 2560 // 2561 // after partial peel 2562 // 2563 // stmt1 2564 // | 2565 // v 2566 // loop predicate 2567 // / 2568 // clone / orig 2569 // / TOP 2570 // / \ 2571 // v v 2572 // TOP->loop loop----+ 2573 // | | | 2574 // stmt2 stmt2 | 2575 // | | | 2576 // v v | 2577 // ifA ifA | 2578 // | \ / | | 2579 // v v v v | 2580 // true false false true | <-- last_peel 2581 // | ^ \ / +------|---+ 2582 // +->newloop \ \ / === ==cut | | 2583 // | stmt3 \ \ / TOP | | 2584 // | | dom | | stmt3 | | <-- first_not_peel 2585 // | v \ 1v v2 v | | 2586 // | ifB regionA ifB ^ v 2587 // | / \ | / \ | | 2588 // | / \ v / \ | | 2589 // | v v exitA: v v | | 2590 // | true false false true | | 2591 // | / ^ \ / \ | | 2592 // | | \ \ / v | | 2593 // | | dom \ / TOP | | 2594 // | | \ 1v v2 | | 2595 // ^ v regionB | | 2596 // | | | | | 2597 // | | v ^ v 2598 // | | exitB: | | 2599 // | | stmt4 | | 2600 // | +------------>-----------------+ | 2601 // | | 2602 // +-----------------<---------------------+ 2603 // 2604 // 2605 // final graph 2606 // 2607 // stmt1 2608 // | 2609 // v 2610 // loop predicate 2611 // | 2612 // v 2613 // stmt2 clone 2614 // | 2615 // v 2616 // ........> ifA clone 2617 // : / | 2618 // dom / | 2619 // : v v 2620 // : false true 2621 // : | | 2622 // : | v 2623 // : | newloop<-----+ 2624 // : | | | 2625 // : | stmt3 clone | 2626 // : | | | 2627 // : | v | 2628 // : | ifB | 2629 // : | / \ | 2630 // : | v v | 2631 // : | false true | 2632 // : | | | | 2633 // : | v stmt2 | 2634 // : | exitB: | | 2635 // : | stmt4 v | 2636 // : | ifA orig | 2637 // : | / \ | 2638 // : | / \ | 2639 // : | v v | 2640 // : | false true | 2641 // : | / \ | 2642 // : v v -----+ 2643 // RegionA 2644 // | 2645 // v 2646 // exitA 2647 // 2648 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) { 2649 2650 assert(!loop->_head->is_CountedLoop(), "Non-counted loop only"); 2651 if (!loop->_head->is_Loop()) { 2652 return false; } 2653 2654 LoopNode *head = loop->_head->as_Loop(); 2655 2656 if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) { 2657 return false; 2658 } 2659 2660 // Check for complex exit control 2661 for(uint ii = 0; ii < loop->_body.size(); ii++ ) { 2662 Node *n = loop->_body.at(ii); 2663 int opc = n->Opcode(); 2664 if (n->is_Call() || 2665 opc == Op_Catch || 2666 opc == Op_CatchProj || 2667 opc == Op_Jump || 2668 opc == Op_JumpProj) { 2669 #if !defined(PRODUCT) 2670 if (TracePartialPeeling) { 2671 tty->print_cr("\nExit control too complex: lp: %d", head->_idx); 2672 } 2673 #endif 2674 return false; 2675 } 2676 } 2677 2678 int dd = dom_depth(head); 2679 2680 // Step 1: find cut point 2681 2682 // Walk up dominators to loop head looking for first loop exit 2683 // which is executed on every path thru loop. 2684 IfNode *peel_if = NULL; 2685 IfNode *peel_if_cmpu = NULL; 2686 2687 Node *iff = loop->tail(); 2688 while( iff != head ) { 2689 if( iff->is_If() ) { 2690 Node *ctrl = get_ctrl(iff->in(1)); 2691 if (ctrl->is_top()) return false; // Dead test on live IF. 2692 // If loop-varying exit-test, check for induction variable 2693 if( loop->is_member(get_loop(ctrl)) && 2694 loop->is_loop_exit(iff) && 2695 is_possible_iv_test(iff)) { 2696 Node* cmp = iff->in(1)->in(1); 2697 if (cmp->Opcode() == Op_CmpI) { 2698 peel_if = iff->as_If(); 2699 } else { 2700 assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU"); 2701 peel_if_cmpu = iff->as_If(); 2702 } 2703 } 2704 } 2705 iff = idom(iff); 2706 } 2707 // Prefer signed compare over unsigned compare. 2708 IfNode* new_peel_if = NULL; 2709 if (peel_if == NULL) { 2710 if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) { 2711 return false; // No peel point found 2712 } 2713 new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop); 2714 if (new_peel_if == NULL) { 2715 return false; // No peel point found 2716 } 2717 peel_if = new_peel_if; 2718 } 2719 Node* last_peel = stay_in_loop(peel_if, loop); 2720 Node* first_not_peeled = stay_in_loop(last_peel, loop); 2721 if (first_not_peeled == NULL || first_not_peeled == head) { 2722 return false; 2723 } 2724 2725 #if !defined(PRODUCT) 2726 if (TraceLoopOpts) { 2727 tty->print("PartialPeel "); 2728 loop->dump_head(); 2729 } 2730 2731 if (TracePartialPeeling) { 2732 tty->print_cr("before partial peel one iteration"); 2733 Node_List wl; 2734 Node* t = head->in(2); 2735 while (true) { 2736 wl.push(t); 2737 if (t == head) break; 2738 t = idom(t); 2739 } 2740 while (wl.size() > 0) { 2741 Node* tt = wl.pop(); 2742 tt->dump(); 2743 if (tt == last_peel) tty->print_cr("-- cut --"); 2744 } 2745 } 2746 #endif 2747 ResourceArea *area = Thread::current()->resource_area(); 2748 VectorSet peel(area); 2749 VectorSet not_peel(area); 2750 Node_List peel_list(area); 2751 Node_List worklist(area); 2752 Node_List sink_list(area); 2753 2754 // Set of cfg nodes to peel are those that are executable from 2755 // the head through last_peel. 2756 assert(worklist.size() == 0, "should be empty"); 2757 worklist.push(head); 2758 peel.set(head->_idx); 2759 while (worklist.size() > 0) { 2760 Node *n = worklist.pop(); 2761 if (n != last_peel) { 2762 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2763 Node* use = n->fast_out(j); 2764 if (use->is_CFG() && 2765 loop->is_member(get_loop(use)) && 2766 !peel.test_set(use->_idx)) { 2767 worklist.push(use); 2768 } 2769 } 2770 } 2771 } 2772 2773 // Set of non-cfg nodes to peel are those that are control 2774 // dependent on the cfg nodes. 2775 uint i; 2776 for(i = 0; i < loop->_body.size(); i++ ) { 2777 Node *n = loop->_body.at(i); 2778 Node *n_c = has_ctrl(n) ? get_ctrl(n) : n; 2779 if (peel.test(n_c->_idx)) { 2780 peel.set(n->_idx); 2781 } else { 2782 not_peel.set(n->_idx); 2783 } 2784 } 2785 2786 // Step 2: move operations from the peeled section down into the 2787 // not-peeled section 2788 2789 // Get a post order schedule of nodes in the peel region 2790 // Result in right-most operand. 2791 scheduled_nodelist(loop, peel, peel_list ); 2792 2793 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 2794 2795 // For future check for too many new phis 2796 uint old_phi_cnt = 0; 2797 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 2798 Node* use = head->fast_out(j); 2799 if (use->is_Phi()) old_phi_cnt++; 2800 } 2801 2802 #if !defined(PRODUCT) 2803 if (TracePartialPeeling) { 2804 tty->print_cr("\npeeled list"); 2805 } 2806 #endif 2807 2808 // Evacuate nodes in peel region into the not_peeled region if possible 2809 uint new_phi_cnt = 0; 2810 uint cloned_for_outside_use = 0; 2811 for (i = 0; i < peel_list.size();) { 2812 Node* n = peel_list.at(i); 2813 #if !defined(PRODUCT) 2814 if (TracePartialPeeling) n->dump(); 2815 #endif 2816 bool incr = true; 2817 if ( !n->is_CFG() ) { 2818 2819 if ( has_use_in_set(n, not_peel) ) { 2820 2821 // If not used internal to the peeled region, 2822 // move "n" from peeled to not_peeled region. 2823 2824 if ( !has_use_internal_to_set(n, peel, loop) ) { 2825 2826 // if not pinned and not a load (which maybe anti-dependent on a store) 2827 // and not a CMove (Matcher expects only bool->cmove). 2828 if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) { 2829 cloned_for_outside_use += clone_for_use_outside_loop( loop, n, worklist ); 2830 sink_list.push(n); 2831 peel >>= n->_idx; // delete n from peel set. 2832 not_peel <<= n->_idx; // add n to not_peel set. 2833 peel_list.remove(i); 2834 incr = false; 2835 #if !defined(PRODUCT) 2836 if (TracePartialPeeling) { 2837 tty->print_cr("sink to not_peeled region: %d newbb: %d", 2838 n->_idx, get_ctrl(n)->_idx); 2839 } 2840 #endif 2841 } 2842 } else { 2843 // Otherwise check for special def-use cases that span 2844 // the peel/not_peel boundary such as bool->if 2845 clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist ); 2846 new_phi_cnt++; 2847 } 2848 } 2849 } 2850 if (incr) i++; 2851 } 2852 2853 if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) { 2854 #if !defined(PRODUCT) 2855 if (TracePartialPeeling) { 2856 tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c", 2857 new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F'); 2858 } 2859 #endif 2860 if (new_peel_if != NULL) { 2861 remove_cmpi_loop_exit(new_peel_if, loop); 2862 } 2863 // Inhibit more partial peeling on this loop 2864 assert(!head->is_partial_peel_loop(), "not partial peeled"); 2865 head->mark_partial_peel_failed(); 2866 if (cloned_for_outside_use > 0) { 2867 // Terminate this round of loop opts because 2868 // the graph outside this loop was changed. 2869 C->set_major_progress(); 2870 return true; 2871 } 2872 return false; 2873 } 2874 2875 // Step 3: clone loop, retarget control, and insert new phis 2876 2877 // Create new loop head for new phis and to hang 2878 // the nodes being moved (sinked) from the peel region. 2879 LoopNode* new_head = new LoopNode(last_peel, last_peel); 2880 new_head->set_unswitch_count(head->unswitch_count()); // Preserve 2881 _igvn.register_new_node_with_optimizer(new_head); 2882 assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled"); 2883 _igvn.replace_input_of(first_not_peeled, 0, new_head); 2884 set_loop(new_head, loop); 2885 loop->_body.push(new_head); 2886 not_peel.set(new_head->_idx); 2887 set_idom(new_head, last_peel, dom_depth(first_not_peeled)); 2888 set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled)); 2889 2890 while (sink_list.size() > 0) { 2891 Node* n = sink_list.pop(); 2892 set_ctrl(n, new_head); 2893 } 2894 2895 assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); 2896 2897 clone_loop( loop, old_new, dd ); 2898 2899 const uint clone_exit_idx = 1; 2900 const uint orig_exit_idx = 2; 2901 assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop"); 2902 2903 Node* head_clone = old_new[head->_idx]; 2904 LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop(); 2905 Node* orig_tail_clone = head_clone->in(2); 2906 2907 // Add phi if "def" node is in peel set and "use" is not 2908 2909 for(i = 0; i < peel_list.size(); i++ ) { 2910 Node *def = peel_list.at(i); 2911 if (!def->is_CFG()) { 2912 for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { 2913 Node *use = def->fast_out(j); 2914 if (has_node(use) && use->in(0) != C->top() && 2915 (!peel.test(use->_idx) || 2916 (use->is_Phi() && use->in(0) == head)) ) { 2917 worklist.push(use); 2918 } 2919 } 2920 while( worklist.size() ) { 2921 Node *use = worklist.pop(); 2922 for (uint j = 1; j < use->req(); j++) { 2923 Node* n = use->in(j); 2924 if (n == def) { 2925 2926 // "def" is in peel set, "use" is not in peel set 2927 // or "use" is in the entry boundary (a phi) of the peel set 2928 2929 Node* use_c = has_ctrl(use) ? get_ctrl(use) : use; 2930 2931 if ( loop->is_member(get_loop( use_c )) ) { 2932 // use is in loop 2933 if (old_new[use->_idx] != NULL) { // null for dead code 2934 Node* use_clone = old_new[use->_idx]; 2935 _igvn.replace_input_of(use, j, C->top()); 2936 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone ); 2937 } 2938 } else { 2939 assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format"); 2940 // use is not in the loop, check if the live range includes the cut 2941 Node* lp_if = use_c->in(orig_exit_idx)->in(0); 2942 if (not_peel.test(lp_if->_idx)) { 2943 assert(j == orig_exit_idx, "use from original loop"); 2944 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone ); 2945 } 2946 } 2947 } 2948 } 2949 } 2950 } 2951 } 2952 2953 // Step 3b: retarget control 2954 2955 // Redirect control to the new loop head if a cloned node in 2956 // the not_peeled region has control that points into the peeled region. 2957 // This necessary because the cloned peeled region will be outside 2958 // the loop. 2959 // from to 2960 // cloned-peeled <---+ 2961 // new_head_clone: | <--+ 2962 // cloned-not_peeled in(0) in(0) 2963 // orig-peeled 2964 2965 for(i = 0; i < loop->_body.size(); i++ ) { 2966 Node *n = loop->_body.at(i); 2967 if (!n->is_CFG() && n->in(0) != NULL && 2968 not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) { 2969 Node* n_clone = old_new[n->_idx]; 2970 _igvn.replace_input_of(n_clone, 0, new_head_clone); 2971 } 2972 } 2973 2974 // Backedge of the surviving new_head (the clone) is original last_peel 2975 _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel); 2976 2977 // Cut first node in original not_peel set 2978 _igvn.rehash_node_delayed(new_head); // Multiple edge updates: 2979 new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of 2980 new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls 2981 2982 // Copy head_clone back-branch info to original head 2983 // and remove original head's loop entry and 2984 // clone head's back-branch 2985 _igvn.rehash_node_delayed(head); // Multiple edge updates 2986 head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl)); 2987 head->set_req(LoopNode::LoopBackControl, C->top()); 2988 _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top()); 2989 2990 // Similarly modify the phis 2991 for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) { 2992 Node* use = head->fast_out(k); 2993 if (use->is_Phi() && use->outcnt() > 0) { 2994 Node* use_clone = old_new[use->_idx]; 2995 _igvn.rehash_node_delayed(use); // Multiple edge updates 2996 use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl)); 2997 use->set_req(LoopNode::LoopBackControl, C->top()); 2998 _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top()); 2999 } 3000 } 3001 3002 // Step 4: update dominator tree and dominator depth 3003 3004 set_idom(head, orig_tail_clone, dd); 3005 recompute_dom_depth(); 3006 3007 // Inhibit more partial peeling on this loop 3008 new_head_clone->set_partial_peel_loop(); 3009 C->set_major_progress(); 3010 loop->record_for_igvn(); 3011 3012 #if !defined(PRODUCT) 3013 if (TracePartialPeeling) { 3014 tty->print_cr("\nafter partial peel one iteration"); 3015 Node_List wl(area); 3016 Node* t = last_peel; 3017 while (true) { 3018 wl.push(t); 3019 if (t == head_clone) break; 3020 t = idom(t); 3021 } 3022 while (wl.size() > 0) { 3023 Node* tt = wl.pop(); 3024 if (tt == head) tty->print_cr("orig head"); 3025 else if (tt == new_head_clone) tty->print_cr("new head"); 3026 else if (tt == head_clone) tty->print_cr("clone head"); 3027 tt->dump(); 3028 } 3029 } 3030 #endif 3031 return true; 3032 } 3033 3034 //------------------------------reorg_offsets---------------------------------- 3035 // Reorganize offset computations to lower register pressure. Mostly 3036 // prevent loop-fallout uses of the pre-incremented trip counter (which are 3037 // then alive with the post-incremented trip counter forcing an extra 3038 // register move) 3039 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) { 3040 // Perform it only for canonical counted loops. 3041 // Loop's shape could be messed up by iteration_split_impl. 3042 if (!loop->_head->is_CountedLoop()) 3043 return; 3044 if (!loop->_head->as_Loop()->is_valid_counted_loop()) 3045 return; 3046 3047 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 3048 CountedLoopEndNode *cle = cl->loopexit(); 3049 Node *exit = cle->proj_out(false); 3050 Node *phi = cl->phi(); 3051 3052 // Check for the special case of folks using the pre-incremented 3053 // trip-counter on the fall-out path (forces the pre-incremented 3054 // and post-incremented trip counter to be live at the same time). 3055 // Fix this by adjusting to use the post-increment trip counter. 3056 3057 bool progress = true; 3058 while (progress) { 3059 progress = false; 3060 for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) { 3061 Node* use = phi->fast_out(i); // User of trip-counter 3062 if (!has_ctrl(use)) continue; 3063 Node *u_ctrl = get_ctrl(use); 3064 if (use->is_Phi()) { 3065 u_ctrl = NULL; 3066 for (uint j = 1; j < use->req(); j++) 3067 if (use->in(j) == phi) 3068 u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j)); 3069 } 3070 IdealLoopTree *u_loop = get_loop(u_ctrl); 3071 // Look for loop-invariant use 3072 if (u_loop == loop) continue; 3073 if (loop->is_member(u_loop)) continue; 3074 // Check that use is live out the bottom. Assuming the trip-counter 3075 // update is right at the bottom, uses of of the loop middle are ok. 3076 if (dom_lca(exit, u_ctrl) != exit) continue; 3077 // Hit! Refactor use to use the post-incremented tripcounter. 3078 // Compute a post-increment tripcounter. 3079 Node *opaq = new Opaque2Node( C, cle->incr() ); 3080 register_new_node(opaq, exit); 3081 Node *neg_stride = _igvn.intcon(-cle->stride_con()); 3082 set_ctrl(neg_stride, C->root()); 3083 Node *post = new AddINode( opaq, neg_stride); 3084 register_new_node(post, exit); 3085 _igvn.rehash_node_delayed(use); 3086 for (uint j = 1; j < use->req(); j++) { 3087 if (use->in(j) == phi) 3088 use->set_req(j, post); 3089 } 3090 // Since DU info changed, rerun loop 3091 progress = true; 3092 break; 3093 } 3094 } 3095 3096 }