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