Thu, 13 Mar 2008 16:31:32 -0700
6673473: (Escape Analysis) Add the instance's field information to PhiNode
Summary: Avoid an infinite generation of instance's field values Phi nodes.
Reviewed-by: never
duke@435 | 1 | /* |
duke@435 | 2 | * Copyright 1999-2006 Sun Microsystems, Inc. All Rights Reserved. |
duke@435 | 3 | * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
duke@435 | 4 | * |
duke@435 | 5 | * This code is free software; you can redistribute it and/or modify it |
duke@435 | 6 | * under the terms of the GNU General Public License version 2 only, as |
duke@435 | 7 | * published by the Free Software Foundation. |
duke@435 | 8 | * |
duke@435 | 9 | * This code is distributed in the hope that it will be useful, but WITHOUT |
duke@435 | 10 | * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or |
duke@435 | 11 | * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
duke@435 | 12 | * version 2 for more details (a copy is included in the LICENSE file that |
duke@435 | 13 | * accompanied this code). |
duke@435 | 14 | * |
duke@435 | 15 | * You should have received a copy of the GNU General Public License version |
duke@435 | 16 | * 2 along with this work; if not, write to the Free Software Foundation, |
duke@435 | 17 | * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. |
duke@435 | 18 | * |
duke@435 | 19 | * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, |
duke@435 | 20 | * CA 95054 USA or visit www.sun.com if you need additional information or |
duke@435 | 21 | * have any questions. |
duke@435 | 22 | * |
duke@435 | 23 | */ |
duke@435 | 24 | |
duke@435 | 25 | #include "incls/_precompiled.incl" |
duke@435 | 26 | #include "incls/_loopopts.cpp.incl" |
duke@435 | 27 | |
duke@435 | 28 | //============================================================================= |
duke@435 | 29 | //------------------------------split_thru_phi--------------------------------- |
duke@435 | 30 | // Split Node 'n' through merge point if there is enough win. |
duke@435 | 31 | Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) { |
duke@435 | 32 | int wins = 0; |
duke@435 | 33 | assert( !n->is_CFG(), "" ); |
duke@435 | 34 | assert( region->is_Region(), "" ); |
kvn@499 | 35 | |
kvn@499 | 36 | const Type* type = n->bottom_type(); |
kvn@499 | 37 | const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr(); |
kvn@499 | 38 | Node *phi; |
kvn@499 | 39 | if( t_oop != NULL && t_oop->is_instance_field() ) { |
kvn@499 | 40 | int iid = t_oop->instance_id(); |
kvn@499 | 41 | int index = C->get_alias_index(t_oop); |
kvn@499 | 42 | int offset = t_oop->offset(); |
kvn@499 | 43 | phi = new (C,region->req()) PhiNode(region, type, NULL, iid, index, offset); |
kvn@499 | 44 | } else { |
kvn@499 | 45 | phi = new (C,region->req()) PhiNode(region, type); |
kvn@499 | 46 | } |
duke@435 | 47 | uint old_unique = C->unique(); |
duke@435 | 48 | for( uint i = 1; i < region->req(); i++ ) { |
duke@435 | 49 | Node *x; |
duke@435 | 50 | Node* the_clone = NULL; |
duke@435 | 51 | if( region->in(i) == C->top() ) { |
duke@435 | 52 | x = C->top(); // Dead path? Use a dead data op |
duke@435 | 53 | } else { |
duke@435 | 54 | x = n->clone(); // Else clone up the data op |
duke@435 | 55 | the_clone = x; // Remember for possible deletion. |
duke@435 | 56 | // Alter data node to use pre-phi inputs |
duke@435 | 57 | if( n->in(0) == region ) |
duke@435 | 58 | x->set_req( 0, region->in(i) ); |
duke@435 | 59 | for( uint j = 1; j < n->req(); j++ ) { |
duke@435 | 60 | Node *in = n->in(j); |
duke@435 | 61 | if( in->is_Phi() && in->in(0) == region ) |
duke@435 | 62 | x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone |
duke@435 | 63 | } |
duke@435 | 64 | } |
duke@435 | 65 | // Check for a 'win' on some paths |
duke@435 | 66 | const Type *t = x->Value(&_igvn); |
duke@435 | 67 | |
duke@435 | 68 | bool singleton = t->singleton(); |
duke@435 | 69 | |
duke@435 | 70 | // A TOP singleton indicates that there are no possible values incoming |
duke@435 | 71 | // along a particular edge. In most cases, this is OK, and the Phi will |
duke@435 | 72 | // be eliminated later in an Ideal call. However, we can't allow this to |
duke@435 | 73 | // happen if the singleton occurs on loop entry, as the elimination of |
duke@435 | 74 | // the PhiNode may cause the resulting node to migrate back to a previous |
duke@435 | 75 | // loop iteration. |
duke@435 | 76 | if( singleton && t == Type::TOP ) { |
duke@435 | 77 | // Is_Loop() == false does not confirm the absence of a loop (e.g., an |
duke@435 | 78 | // irreducible loop may not be indicated by an affirmative is_Loop()); |
duke@435 | 79 | // therefore, the only top we can split thru a phi is on a backedge of |
duke@435 | 80 | // a loop. |
duke@435 | 81 | singleton &= region->is_Loop() && (i != LoopNode::EntryControl); |
duke@435 | 82 | } |
duke@435 | 83 | |
duke@435 | 84 | if( singleton ) { |
duke@435 | 85 | wins++; |
duke@435 | 86 | x = ((PhaseGVN&)_igvn).makecon(t); |
duke@435 | 87 | } else { |
duke@435 | 88 | // We now call Identity to try to simplify the cloned node. |
duke@435 | 89 | // Note that some Identity methods call phase->type(this). |
duke@435 | 90 | // Make sure that the type array is big enough for |
duke@435 | 91 | // our new node, even though we may throw the node away. |
duke@435 | 92 | // (Note: This tweaking with igvn only works because x is a new node.) |
duke@435 | 93 | _igvn.set_type(x, t); |
duke@435 | 94 | Node *y = x->Identity(&_igvn); |
duke@435 | 95 | if( y != x ) { |
duke@435 | 96 | wins++; |
duke@435 | 97 | x = y; |
duke@435 | 98 | } else { |
duke@435 | 99 | y = _igvn.hash_find(x); |
duke@435 | 100 | if( y ) { |
duke@435 | 101 | wins++; |
duke@435 | 102 | x = y; |
duke@435 | 103 | } else { |
duke@435 | 104 | // Else x is a new node we are keeping |
duke@435 | 105 | // We do not need register_new_node_with_optimizer |
duke@435 | 106 | // because set_type has already been called. |
duke@435 | 107 | _igvn._worklist.push(x); |
duke@435 | 108 | } |
duke@435 | 109 | } |
duke@435 | 110 | } |
duke@435 | 111 | if (x != the_clone && the_clone != NULL) |
duke@435 | 112 | _igvn.remove_dead_node(the_clone); |
duke@435 | 113 | phi->set_req( i, x ); |
duke@435 | 114 | } |
duke@435 | 115 | // Too few wins? |
duke@435 | 116 | if( wins <= policy ) { |
duke@435 | 117 | _igvn.remove_dead_node(phi); |
duke@435 | 118 | return NULL; |
duke@435 | 119 | } |
duke@435 | 120 | |
duke@435 | 121 | // Record Phi |
duke@435 | 122 | register_new_node( phi, region ); |
duke@435 | 123 | |
duke@435 | 124 | for( uint i2 = 1; i2 < phi->req(); i2++ ) { |
duke@435 | 125 | Node *x = phi->in(i2); |
duke@435 | 126 | // If we commoned up the cloned 'x' with another existing Node, |
duke@435 | 127 | // the existing Node picks up a new use. We need to make the |
duke@435 | 128 | // existing Node occur higher up so it dominates its uses. |
duke@435 | 129 | Node *old_ctrl; |
duke@435 | 130 | IdealLoopTree *old_loop; |
duke@435 | 131 | |
duke@435 | 132 | // The occasional new node |
duke@435 | 133 | if( x->_idx >= old_unique ) { // Found a new, unplaced node? |
duke@435 | 134 | old_ctrl = x->is_Con() ? C->root() : NULL; |
duke@435 | 135 | old_loop = NULL; // Not in any prior loop |
duke@435 | 136 | } else { |
duke@435 | 137 | old_ctrl = x->is_Con() ? C->root() : get_ctrl(x); |
duke@435 | 138 | old_loop = get_loop(old_ctrl); // Get prior loop |
duke@435 | 139 | } |
duke@435 | 140 | // New late point must dominate new use |
duke@435 | 141 | Node *new_ctrl = dom_lca( old_ctrl, region->in(i2) ); |
duke@435 | 142 | // Set new location |
duke@435 | 143 | set_ctrl(x, new_ctrl); |
duke@435 | 144 | IdealLoopTree *new_loop = get_loop( new_ctrl ); |
duke@435 | 145 | // If changing loop bodies, see if we need to collect into new body |
duke@435 | 146 | if( old_loop != new_loop ) { |
duke@435 | 147 | if( old_loop && !old_loop->_child ) |
duke@435 | 148 | old_loop->_body.yank(x); |
duke@435 | 149 | if( !new_loop->_child ) |
duke@435 | 150 | new_loop->_body.push(x); // Collect body info |
duke@435 | 151 | } |
duke@435 | 152 | } |
duke@435 | 153 | |
duke@435 | 154 | return phi; |
duke@435 | 155 | } |
duke@435 | 156 | |
duke@435 | 157 | //------------------------------dominated_by------------------------------------ |
duke@435 | 158 | // Replace the dominated test with an obvious true or false. Place it on the |
duke@435 | 159 | // IGVN worklist for later cleanup. Move control-dependent data Nodes on the |
duke@435 | 160 | // live path up to the dominating control. |
duke@435 | 161 | void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff ) { |
duke@435 | 162 | #ifndef PRODUCT |
duke@435 | 163 | if( VerifyLoopOptimizations && PrintOpto ) tty->print_cr("dominating test"); |
duke@435 | 164 | #endif |
duke@435 | 165 | |
duke@435 | 166 | |
duke@435 | 167 | // prevdom is the dominating projection of the dominating test. |
duke@435 | 168 | assert( iff->is_If(), "" ); |
duke@435 | 169 | assert( iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added"); |
duke@435 | 170 | int pop = prevdom->Opcode(); |
duke@435 | 171 | assert( pop == Op_IfFalse || pop == Op_IfTrue, "" ); |
duke@435 | 172 | // 'con' is set to true or false to kill the dominated test. |
duke@435 | 173 | Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO); |
duke@435 | 174 | set_ctrl(con, C->root()); // Constant gets a new use |
duke@435 | 175 | // Hack the dominated test |
duke@435 | 176 | _igvn.hash_delete(iff); |
duke@435 | 177 | iff->set_req(1, con); |
duke@435 | 178 | _igvn._worklist.push(iff); |
duke@435 | 179 | |
duke@435 | 180 | // If I dont have a reachable TRUE and FALSE path following the IfNode then |
duke@435 | 181 | // I can assume this path reaches an infinite loop. In this case it's not |
duke@435 | 182 | // important to optimize the data Nodes - either the whole compilation will |
duke@435 | 183 | // be tossed or this path (and all data Nodes) will go dead. |
duke@435 | 184 | if( iff->outcnt() != 2 ) return; |
duke@435 | 185 | |
duke@435 | 186 | // Make control-dependent data Nodes on the live path (path that will remain |
duke@435 | 187 | // once the dominated IF is removed) become control-dependent on the |
duke@435 | 188 | // dominating projection. |
duke@435 | 189 | Node* dp = ((IfNode*)iff)->proj_out(pop == Op_IfTrue); |
duke@435 | 190 | IdealLoopTree *old_loop = get_loop(dp); |
duke@435 | 191 | |
duke@435 | 192 | for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { |
duke@435 | 193 | Node* cd = dp->fast_out(i); // Control-dependent node |
duke@435 | 194 | if( cd->depends_only_on_test() ) { |
duke@435 | 195 | assert( cd->in(0) == dp, "" ); |
duke@435 | 196 | _igvn.hash_delete( cd ); |
duke@435 | 197 | cd->set_req(0, prevdom); |
duke@435 | 198 | set_early_ctrl( cd ); |
duke@435 | 199 | _igvn._worklist.push(cd); |
duke@435 | 200 | IdealLoopTree *new_loop = get_loop(get_ctrl(cd)); |
duke@435 | 201 | if( old_loop != new_loop ) { |
duke@435 | 202 | if( !old_loop->_child ) old_loop->_body.yank(cd); |
duke@435 | 203 | if( !new_loop->_child ) new_loop->_body.push(cd); |
duke@435 | 204 | } |
duke@435 | 205 | --i; |
duke@435 | 206 | --imax; |
duke@435 | 207 | } |
duke@435 | 208 | } |
duke@435 | 209 | } |
duke@435 | 210 | |
duke@435 | 211 | //------------------------------has_local_phi_input---------------------------- |
duke@435 | 212 | // Return TRUE if 'n' has Phi inputs from its local block and no other |
duke@435 | 213 | // block-local inputs (all non-local-phi inputs come from earlier blocks) |
duke@435 | 214 | Node *PhaseIdealLoop::has_local_phi_input( Node *n ) { |
duke@435 | 215 | Node *n_ctrl = get_ctrl(n); |
duke@435 | 216 | // See if some inputs come from a Phi in this block, or from before |
duke@435 | 217 | // this block. |
duke@435 | 218 | uint i; |
duke@435 | 219 | for( i = 1; i < n->req(); i++ ) { |
duke@435 | 220 | Node *phi = n->in(i); |
duke@435 | 221 | if( phi->is_Phi() && phi->in(0) == n_ctrl ) |
duke@435 | 222 | break; |
duke@435 | 223 | } |
duke@435 | 224 | if( i >= n->req() ) |
duke@435 | 225 | return NULL; // No Phi inputs; nowhere to clone thru |
duke@435 | 226 | |
duke@435 | 227 | // Check for inputs created between 'n' and the Phi input. These |
duke@435 | 228 | // must split as well; they have already been given the chance |
duke@435 | 229 | // (courtesy of a post-order visit) and since they did not we must |
duke@435 | 230 | // recover the 'cost' of splitting them by being very profitable |
duke@435 | 231 | // when splitting 'n'. Since this is unlikely we simply give up. |
duke@435 | 232 | for( i = 1; i < n->req(); i++ ) { |
duke@435 | 233 | Node *m = n->in(i); |
duke@435 | 234 | if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) { |
duke@435 | 235 | // We allow the special case of AddP's with no local inputs. |
duke@435 | 236 | // This allows us to split-up address expressions. |
duke@435 | 237 | if (m->is_AddP() && |
duke@435 | 238 | get_ctrl(m->in(2)) != n_ctrl && |
duke@435 | 239 | get_ctrl(m->in(3)) != n_ctrl) { |
duke@435 | 240 | // Move the AddP up to dominating point |
duke@435 | 241 | set_ctrl_and_loop(m, find_non_split_ctrl(idom(n_ctrl))); |
duke@435 | 242 | continue; |
duke@435 | 243 | } |
duke@435 | 244 | return NULL; |
duke@435 | 245 | } |
duke@435 | 246 | } |
duke@435 | 247 | |
duke@435 | 248 | return n_ctrl; |
duke@435 | 249 | } |
duke@435 | 250 | |
duke@435 | 251 | //------------------------------remix_address_expressions---------------------- |
duke@435 | 252 | // Rework addressing expressions to get the most loop-invariant stuff |
duke@435 | 253 | // moved out. We'd like to do all associative operators, but it's especially |
duke@435 | 254 | // important (common) to do address expressions. |
duke@435 | 255 | Node *PhaseIdealLoop::remix_address_expressions( Node *n ) { |
duke@435 | 256 | if (!has_ctrl(n)) return NULL; |
duke@435 | 257 | Node *n_ctrl = get_ctrl(n); |
duke@435 | 258 | IdealLoopTree *n_loop = get_loop(n_ctrl); |
duke@435 | 259 | |
duke@435 | 260 | // See if 'n' mixes loop-varying and loop-invariant inputs and |
duke@435 | 261 | // itself is loop-varying. |
duke@435 | 262 | |
duke@435 | 263 | // Only interested in binary ops (and AddP) |
duke@435 | 264 | if( n->req() < 3 || n->req() > 4 ) return NULL; |
duke@435 | 265 | |
duke@435 | 266 | Node *n1_ctrl = get_ctrl(n->in( 1)); |
duke@435 | 267 | Node *n2_ctrl = get_ctrl(n->in( 2)); |
duke@435 | 268 | Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3)); |
duke@435 | 269 | IdealLoopTree *n1_loop = get_loop( n1_ctrl ); |
duke@435 | 270 | IdealLoopTree *n2_loop = get_loop( n2_ctrl ); |
duke@435 | 271 | IdealLoopTree *n3_loop = get_loop( n3_ctrl ); |
duke@435 | 272 | |
duke@435 | 273 | // Does one of my inputs spin in a tighter loop than self? |
duke@435 | 274 | if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) || |
duke@435 | 275 | (n_loop->is_member( n2_loop ) && n_loop != n2_loop) || |
duke@435 | 276 | (n_loop->is_member( n3_loop ) && n_loop != n3_loop) ) |
duke@435 | 277 | return NULL; // Leave well enough alone |
duke@435 | 278 | |
duke@435 | 279 | // Is at least one of my inputs loop-invariant? |
duke@435 | 280 | if( n1_loop == n_loop && |
duke@435 | 281 | n2_loop == n_loop && |
duke@435 | 282 | n3_loop == n_loop ) |
duke@435 | 283 | return NULL; // No loop-invariant inputs |
duke@435 | 284 | |
duke@435 | 285 | |
duke@435 | 286 | int n_op = n->Opcode(); |
duke@435 | 287 | |
duke@435 | 288 | // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2). |
duke@435 | 289 | if( n_op == Op_LShiftI ) { |
duke@435 | 290 | // Scale is loop invariant |
duke@435 | 291 | Node *scale = n->in(2); |
duke@435 | 292 | Node *scale_ctrl = get_ctrl(scale); |
duke@435 | 293 | IdealLoopTree *scale_loop = get_loop(scale_ctrl ); |
duke@435 | 294 | if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) ) |
duke@435 | 295 | return NULL; |
duke@435 | 296 | const TypeInt *scale_t = scale->bottom_type()->isa_int(); |
duke@435 | 297 | if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 ) |
duke@435 | 298 | return NULL; // Dont bother with byte/short masking |
duke@435 | 299 | // Add must vary with loop (else shift would be loop-invariant) |
duke@435 | 300 | Node *add = n->in(1); |
duke@435 | 301 | Node *add_ctrl = get_ctrl(add); |
duke@435 | 302 | IdealLoopTree *add_loop = get_loop(add_ctrl); |
duke@435 | 303 | //assert( n_loop == add_loop, "" ); |
duke@435 | 304 | if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops |
duke@435 | 305 | |
duke@435 | 306 | // Convert I-V into I+ (0-V); same for V-I |
duke@435 | 307 | if( add->Opcode() == Op_SubI && |
duke@435 | 308 | _igvn.type( add->in(1) ) != TypeInt::ZERO ) { |
duke@435 | 309 | Node *zero = _igvn.intcon(0); |
duke@435 | 310 | set_ctrl(zero, C->root()); |
duke@435 | 311 | Node *neg = new (C, 3) SubINode( _igvn.intcon(0), add->in(2) ); |
duke@435 | 312 | register_new_node( neg, get_ctrl(add->in(2) ) ); |
duke@435 | 313 | add = new (C, 3) AddINode( add->in(1), neg ); |
duke@435 | 314 | register_new_node( add, add_ctrl ); |
duke@435 | 315 | } |
duke@435 | 316 | if( add->Opcode() != Op_AddI ) return NULL; |
duke@435 | 317 | // See if one add input is loop invariant |
duke@435 | 318 | Node *add_var = add->in(1); |
duke@435 | 319 | Node *add_var_ctrl = get_ctrl(add_var); |
duke@435 | 320 | IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); |
duke@435 | 321 | Node *add_invar = add->in(2); |
duke@435 | 322 | Node *add_invar_ctrl = get_ctrl(add_invar); |
duke@435 | 323 | IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl ); |
duke@435 | 324 | if( add_var_loop == n_loop ) { |
duke@435 | 325 | } else if( add_invar_loop == n_loop ) { |
duke@435 | 326 | // Swap to find the invariant part |
duke@435 | 327 | add_invar = add_var; |
duke@435 | 328 | add_invar_ctrl = add_var_ctrl; |
duke@435 | 329 | add_invar_loop = add_var_loop; |
duke@435 | 330 | add_var = add->in(2); |
duke@435 | 331 | Node *add_var_ctrl = get_ctrl(add_var); |
duke@435 | 332 | IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); |
duke@435 | 333 | } else // Else neither input is loop invariant |
duke@435 | 334 | return NULL; |
duke@435 | 335 | if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) ) |
duke@435 | 336 | return NULL; // No invariant part of the add? |
duke@435 | 337 | |
duke@435 | 338 | // Yes! Reshape address expression! |
duke@435 | 339 | Node *inv_scale = new (C, 3) LShiftINode( add_invar, scale ); |
duke@435 | 340 | register_new_node( inv_scale, add_invar_ctrl ); |
duke@435 | 341 | Node *var_scale = new (C, 3) LShiftINode( add_var, scale ); |
duke@435 | 342 | register_new_node( var_scale, n_ctrl ); |
duke@435 | 343 | Node *var_add = new (C, 3) AddINode( var_scale, inv_scale ); |
duke@435 | 344 | register_new_node( var_add, n_ctrl ); |
duke@435 | 345 | _igvn.hash_delete( n ); |
duke@435 | 346 | _igvn.subsume_node( n, var_add ); |
duke@435 | 347 | return var_add; |
duke@435 | 348 | } |
duke@435 | 349 | |
duke@435 | 350 | // Replace (I+V) with (V+I) |
duke@435 | 351 | if( n_op == Op_AddI || |
duke@435 | 352 | n_op == Op_AddL || |
duke@435 | 353 | n_op == Op_AddF || |
duke@435 | 354 | n_op == Op_AddD || |
duke@435 | 355 | n_op == Op_MulI || |
duke@435 | 356 | n_op == Op_MulL || |
duke@435 | 357 | n_op == Op_MulF || |
duke@435 | 358 | n_op == Op_MulD ) { |
duke@435 | 359 | if( n2_loop == n_loop ) { |
duke@435 | 360 | assert( n1_loop != n_loop, "" ); |
duke@435 | 361 | n->swap_edges(1, 2); |
duke@435 | 362 | } |
duke@435 | 363 | } |
duke@435 | 364 | |
duke@435 | 365 | // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V), |
duke@435 | 366 | // but not if I2 is a constant. |
duke@435 | 367 | if( n_op == Op_AddP ) { |
duke@435 | 368 | if( n2_loop == n_loop && n3_loop != n_loop ) { |
duke@435 | 369 | if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) { |
duke@435 | 370 | Node *n22_ctrl = get_ctrl(n->in(2)->in(2)); |
duke@435 | 371 | Node *n23_ctrl = get_ctrl(n->in(2)->in(3)); |
duke@435 | 372 | IdealLoopTree *n22loop = get_loop( n22_ctrl ); |
duke@435 | 373 | IdealLoopTree *n23_loop = get_loop( n23_ctrl ); |
duke@435 | 374 | if( n22loop != n_loop && n22loop->is_member(n_loop) && |
duke@435 | 375 | n23_loop == n_loop ) { |
duke@435 | 376 | Node *add1 = new (C, 4) AddPNode( n->in(1), n->in(2)->in(2), n->in(3) ); |
duke@435 | 377 | // Stuff new AddP in the loop preheader |
duke@435 | 378 | register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) ); |
duke@435 | 379 | Node *add2 = new (C, 4) AddPNode( n->in(1), add1, n->in(2)->in(3) ); |
duke@435 | 380 | register_new_node( add2, n_ctrl ); |
duke@435 | 381 | _igvn.hash_delete( n ); |
duke@435 | 382 | _igvn.subsume_node( n, add2 ); |
duke@435 | 383 | return add2; |
duke@435 | 384 | } |
duke@435 | 385 | } |
duke@435 | 386 | } |
duke@435 | 387 | |
duke@435 | 388 | // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V) |
duke@435 | 389 | if( n2_loop != n_loop && n3_loop == n_loop ) { |
duke@435 | 390 | if( n->in(3)->Opcode() == Op_AddI ) { |
duke@435 | 391 | Node *V = n->in(3)->in(1); |
duke@435 | 392 | Node *I = n->in(3)->in(2); |
duke@435 | 393 | if( is_member(n_loop,get_ctrl(V)) ) { |
duke@435 | 394 | } else { |
duke@435 | 395 | Node *tmp = V; V = I; I = tmp; |
duke@435 | 396 | } |
duke@435 | 397 | if( !is_member(n_loop,get_ctrl(I)) ) { |
duke@435 | 398 | Node *add1 = new (C, 4) AddPNode( n->in(1), n->in(2), I ); |
duke@435 | 399 | // Stuff new AddP in the loop preheader |
duke@435 | 400 | register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) ); |
duke@435 | 401 | Node *add2 = new (C, 4) AddPNode( n->in(1), add1, V ); |
duke@435 | 402 | register_new_node( add2, n_ctrl ); |
duke@435 | 403 | _igvn.hash_delete( n ); |
duke@435 | 404 | _igvn.subsume_node( n, add2 ); |
duke@435 | 405 | return add2; |
duke@435 | 406 | } |
duke@435 | 407 | } |
duke@435 | 408 | } |
duke@435 | 409 | } |
duke@435 | 410 | |
duke@435 | 411 | return NULL; |
duke@435 | 412 | } |
duke@435 | 413 | |
duke@435 | 414 | //------------------------------conditional_move------------------------------- |
duke@435 | 415 | // Attempt to replace a Phi with a conditional move. We have some pretty |
duke@435 | 416 | // strict profitability requirements. All Phis at the merge point must |
duke@435 | 417 | // be converted, so we can remove the control flow. We need to limit the |
duke@435 | 418 | // number of c-moves to a small handful. All code that was in the side-arms |
duke@435 | 419 | // of the CFG diamond is now speculatively executed. This code has to be |
duke@435 | 420 | // "cheap enough". We are pretty much limited to CFG diamonds that merge |
duke@435 | 421 | // 1 or 2 items with a total of 1 or 2 ops executed speculatively. |
duke@435 | 422 | Node *PhaseIdealLoop::conditional_move( Node *region ) { |
duke@435 | 423 | |
duke@435 | 424 | assert( region->is_Region(), "sanity check" ); |
duke@435 | 425 | if( region->req() != 3 ) return NULL; |
duke@435 | 426 | |
duke@435 | 427 | // Check for CFG diamond |
duke@435 | 428 | Node *lp = region->in(1); |
duke@435 | 429 | Node *rp = region->in(2); |
duke@435 | 430 | if( !lp || !rp ) return NULL; |
duke@435 | 431 | Node *lp_c = lp->in(0); |
duke@435 | 432 | if( lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If() ) return NULL; |
duke@435 | 433 | IfNode *iff = lp_c->as_If(); |
duke@435 | 434 | |
duke@435 | 435 | // Check for highly predictable branch. No point in CMOV'ing if |
duke@435 | 436 | // we are going to predict accurately all the time. |
duke@435 | 437 | // %%% This hides patterns produced by utility methods like Math.min. |
duke@435 | 438 | if( iff->_prob < PROB_UNLIKELY_MAG(3) || |
duke@435 | 439 | iff->_prob > PROB_LIKELY_MAG(3) ) |
duke@435 | 440 | return NULL; |
duke@435 | 441 | |
duke@435 | 442 | // Check for ops pinned in an arm of the diamond. |
duke@435 | 443 | // Can't remove the control flow in this case |
duke@435 | 444 | if( lp->outcnt() > 1 ) return NULL; |
duke@435 | 445 | if( rp->outcnt() > 1 ) return NULL; |
duke@435 | 446 | |
duke@435 | 447 | // Check profitability |
duke@435 | 448 | int cost = 0; |
kvn@470 | 449 | int phis = 0; |
duke@435 | 450 | for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { |
duke@435 | 451 | Node *out = region->fast_out(i); |
duke@435 | 452 | if( !out->is_Phi() ) continue; // Ignore other control edges, etc |
kvn@470 | 453 | phis++; |
duke@435 | 454 | PhiNode* phi = out->as_Phi(); |
duke@435 | 455 | switch (phi->type()->basic_type()) { |
duke@435 | 456 | case T_LONG: |
duke@435 | 457 | cost++; // Probably encodes as 2 CMOV's |
duke@435 | 458 | case T_INT: // These all CMOV fine |
duke@435 | 459 | case T_FLOAT: |
duke@435 | 460 | case T_DOUBLE: |
duke@435 | 461 | case T_ADDRESS: // (RawPtr) |
duke@435 | 462 | cost++; |
duke@435 | 463 | break; |
duke@435 | 464 | case T_OBJECT: { // Base oops are OK, but not derived oops |
duke@435 | 465 | const TypeOopPtr *tp = phi->type()->isa_oopptr(); |
duke@435 | 466 | // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a |
duke@435 | 467 | // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus |
duke@435 | 468 | // CMOVE'ing a derived pointer requires we also CMOVE the base. If we |
duke@435 | 469 | // have a Phi for the base here that we convert to a CMOVE all is well |
duke@435 | 470 | // and good. But if the base is dead, we'll not make a CMOVE. Later |
duke@435 | 471 | // the allocator will have to produce a base by creating a CMOVE of the |
duke@435 | 472 | // relevant bases. This puts the allocator in the business of |
duke@435 | 473 | // manufacturing expensive instructions, generally a bad plan. |
duke@435 | 474 | // Just Say No to Conditionally-Moved Derived Pointers. |
duke@435 | 475 | if( tp && tp->offset() != 0 ) |
duke@435 | 476 | return NULL; |
duke@435 | 477 | cost++; |
duke@435 | 478 | break; |
duke@435 | 479 | } |
duke@435 | 480 | default: |
duke@435 | 481 | return NULL; // In particular, can't do memory or I/O |
duke@435 | 482 | } |
duke@435 | 483 | // Add in cost any speculative ops |
duke@435 | 484 | for( uint j = 1; j < region->req(); j++ ) { |
duke@435 | 485 | Node *proj = region->in(j); |
duke@435 | 486 | Node *inp = phi->in(j); |
duke@435 | 487 | if (get_ctrl(inp) == proj) { // Found local op |
duke@435 | 488 | cost++; |
duke@435 | 489 | // Check for a chain of dependent ops; these will all become |
duke@435 | 490 | // speculative in a CMOV. |
duke@435 | 491 | for( uint k = 1; k < inp->req(); k++ ) |
duke@435 | 492 | if (get_ctrl(inp->in(k)) == proj) |
duke@435 | 493 | return NULL; // Too much speculative goo |
duke@435 | 494 | } |
duke@435 | 495 | } |
duke@435 | 496 | // See if the Phi is used by a Cmp. This will likely Split-If, a |
duke@435 | 497 | // higher-payoff operation. |
duke@435 | 498 | for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) { |
duke@435 | 499 | Node* use = phi->fast_out(k); |
duke@435 | 500 | if( use->is_Cmp() ) |
duke@435 | 501 | return NULL; |
duke@435 | 502 | } |
duke@435 | 503 | } |
duke@435 | 504 | if( cost >= ConditionalMoveLimit ) return NULL; // Too much goo |
kvn@470 | 505 | Node* bol = iff->in(1); |
kvn@470 | 506 | assert( bol->Opcode() == Op_Bool, "" ); |
kvn@470 | 507 | int cmp_op = bol->in(1)->Opcode(); |
kvn@470 | 508 | // It is expensive to generate flags from a float compare. |
kvn@470 | 509 | // Avoid duplicated float compare. |
kvn@470 | 510 | if( phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL; |
duke@435 | 511 | |
duke@435 | 512 | // -------------- |
duke@435 | 513 | // Now replace all Phis with CMOV's |
duke@435 | 514 | Node *cmov_ctrl = iff->in(0); |
duke@435 | 515 | uint flip = (lp->Opcode() == Op_IfTrue); |
duke@435 | 516 | while( 1 ) { |
duke@435 | 517 | PhiNode* phi = NULL; |
duke@435 | 518 | for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { |
duke@435 | 519 | Node *out = region->fast_out(i); |
duke@435 | 520 | if (out->is_Phi()) { |
duke@435 | 521 | phi = out->as_Phi(); |
duke@435 | 522 | break; |
duke@435 | 523 | } |
duke@435 | 524 | } |
duke@435 | 525 | if (phi == NULL) break; |
duke@435 | 526 | #ifndef PRODUCT |
duke@435 | 527 | if( PrintOpto && VerifyLoopOptimizations ) tty->print_cr("CMOV"); |
duke@435 | 528 | #endif |
duke@435 | 529 | // Move speculative ops |
duke@435 | 530 | for( uint j = 1; j < region->req(); j++ ) { |
duke@435 | 531 | Node *proj = region->in(j); |
duke@435 | 532 | Node *inp = phi->in(j); |
duke@435 | 533 | if (get_ctrl(inp) == proj) { // Found local op |
duke@435 | 534 | #ifndef PRODUCT |
duke@435 | 535 | if( PrintOpto && VerifyLoopOptimizations ) { |
duke@435 | 536 | tty->print(" speculate: "); |
duke@435 | 537 | inp->dump(); |
duke@435 | 538 | } |
duke@435 | 539 | #endif |
duke@435 | 540 | set_ctrl(inp, cmov_ctrl); |
duke@435 | 541 | } |
duke@435 | 542 | } |
duke@435 | 543 | Node *cmov = CMoveNode::make( C, cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi) ); |
duke@435 | 544 | register_new_node( cmov, cmov_ctrl ); |
duke@435 | 545 | _igvn.hash_delete(phi); |
duke@435 | 546 | _igvn.subsume_node( phi, cmov ); |
duke@435 | 547 | #ifndef PRODUCT |
duke@435 | 548 | if( VerifyLoopOptimizations ) verify(); |
duke@435 | 549 | #endif |
duke@435 | 550 | } |
duke@435 | 551 | |
duke@435 | 552 | // The useless CFG diamond will fold up later; see the optimization in |
duke@435 | 553 | // RegionNode::Ideal. |
duke@435 | 554 | _igvn._worklist.push(region); |
duke@435 | 555 | |
duke@435 | 556 | return iff->in(1); |
duke@435 | 557 | } |
duke@435 | 558 | |
duke@435 | 559 | //------------------------------split_if_with_blocks_pre----------------------- |
duke@435 | 560 | // Do the real work in a non-recursive function. Data nodes want to be |
duke@435 | 561 | // cloned in the pre-order so they can feed each other nicely. |
duke@435 | 562 | Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) { |
duke@435 | 563 | // Cloning these guys is unlikely to win |
duke@435 | 564 | int n_op = n->Opcode(); |
duke@435 | 565 | if( n_op == Op_MergeMem ) return n; |
duke@435 | 566 | if( n->is_Proj() ) return n; |
duke@435 | 567 | // Do not clone-up CmpFXXX variations, as these are always |
duke@435 | 568 | // followed by a CmpI |
duke@435 | 569 | if( n->is_Cmp() ) return n; |
duke@435 | 570 | // Attempt to use a conditional move instead of a phi/branch |
duke@435 | 571 | if( ConditionalMoveLimit > 0 && n_op == Op_Region ) { |
duke@435 | 572 | Node *cmov = conditional_move( n ); |
duke@435 | 573 | if( cmov ) return cmov; |
duke@435 | 574 | } |
duke@435 | 575 | if( n->is_CFG() || n_op == Op_StorePConditional || n_op == Op_StoreLConditional || n_op == Op_CompareAndSwapI || n_op == Op_CompareAndSwapL ||n_op == Op_CompareAndSwapP) return n; |
duke@435 | 576 | if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd |
duke@435 | 577 | n_op == Op_Opaque2 ) { |
duke@435 | 578 | if( !C->major_progress() ) // If chance of no more loop opts... |
duke@435 | 579 | _igvn._worklist.push(n); // maybe we'll remove them |
duke@435 | 580 | return n; |
duke@435 | 581 | } |
duke@435 | 582 | |
duke@435 | 583 | if( n->is_Con() ) return n; // No cloning for Con nodes |
duke@435 | 584 | |
duke@435 | 585 | Node *n_ctrl = get_ctrl(n); |
duke@435 | 586 | if( !n_ctrl ) return n; // Dead node |
duke@435 | 587 | |
duke@435 | 588 | // Attempt to remix address expressions for loop invariants |
duke@435 | 589 | Node *m = remix_address_expressions( n ); |
duke@435 | 590 | if( m ) return m; |
duke@435 | 591 | |
duke@435 | 592 | // Determine if the Node has inputs from some local Phi. |
duke@435 | 593 | // Returns the block to clone thru. |
duke@435 | 594 | Node *n_blk = has_local_phi_input( n ); |
duke@435 | 595 | if( !n_blk ) return n; |
duke@435 | 596 | // Do not clone the trip counter through on a CountedLoop |
duke@435 | 597 | // (messes up the canonical shape). |
duke@435 | 598 | if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n; |
duke@435 | 599 | |
duke@435 | 600 | // Check for having no control input; not pinned. Allow |
duke@435 | 601 | // dominating control. |
duke@435 | 602 | if( n->in(0) ) { |
duke@435 | 603 | Node *dom = idom(n_blk); |
duke@435 | 604 | if( dom_lca( n->in(0), dom ) != n->in(0) ) |
duke@435 | 605 | return n; |
duke@435 | 606 | } |
duke@435 | 607 | // Policy: when is it profitable. You must get more wins than |
duke@435 | 608 | // policy before it is considered profitable. Policy is usually 0, |
duke@435 | 609 | // so 1 win is considered profitable. Big merges will require big |
duke@435 | 610 | // cloning, so get a larger policy. |
duke@435 | 611 | int policy = n_blk->req() >> 2; |
duke@435 | 612 | |
duke@435 | 613 | // If the loop is a candidate for range check elimination, |
duke@435 | 614 | // delay splitting through it's phi until a later loop optimization |
duke@435 | 615 | if (n_blk->is_CountedLoop()) { |
duke@435 | 616 | IdealLoopTree *lp = get_loop(n_blk); |
duke@435 | 617 | if (lp && lp->_rce_candidate) { |
duke@435 | 618 | return n; |
duke@435 | 619 | } |
duke@435 | 620 | } |
duke@435 | 621 | |
duke@435 | 622 | // Use same limit as split_if_with_blocks_post |
duke@435 | 623 | if( C->unique() > 35000 ) return n; // Method too big |
duke@435 | 624 | |
duke@435 | 625 | // Split 'n' through the merge point if it is profitable |
duke@435 | 626 | Node *phi = split_thru_phi( n, n_blk, policy ); |
duke@435 | 627 | if( !phi ) return n; |
duke@435 | 628 | |
duke@435 | 629 | // Found a Phi to split thru! |
duke@435 | 630 | // Replace 'n' with the new phi |
duke@435 | 631 | _igvn.hash_delete(n); |
duke@435 | 632 | _igvn.subsume_node( n, phi ); |
duke@435 | 633 | // Moved a load around the loop, 'en-registering' something. |
duke@435 | 634 | if( n_blk->Opcode() == Op_Loop && n->is_Load() && |
duke@435 | 635 | !phi->in(LoopNode::LoopBackControl)->is_Load() ) |
duke@435 | 636 | C->set_major_progress(); |
duke@435 | 637 | |
duke@435 | 638 | return phi; |
duke@435 | 639 | } |
duke@435 | 640 | |
duke@435 | 641 | static bool merge_point_too_heavy(Compile* C, Node* region) { |
duke@435 | 642 | // Bail out if the region and its phis have too many users. |
duke@435 | 643 | int weight = 0; |
duke@435 | 644 | for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { |
duke@435 | 645 | weight += region->fast_out(i)->outcnt(); |
duke@435 | 646 | } |
duke@435 | 647 | int nodes_left = MaxNodeLimit - C->unique(); |
duke@435 | 648 | if (weight * 8 > nodes_left) { |
duke@435 | 649 | #ifndef PRODUCT |
duke@435 | 650 | if (PrintOpto) |
duke@435 | 651 | tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight); |
duke@435 | 652 | #endif |
duke@435 | 653 | return true; |
duke@435 | 654 | } else { |
duke@435 | 655 | return false; |
duke@435 | 656 | } |
duke@435 | 657 | } |
duke@435 | 658 | |
duke@435 | 659 | #ifdef _LP64 |
duke@435 | 660 | static bool merge_point_safe(Node* region) { |
duke@435 | 661 | // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode |
duke@435 | 662 | // having a PhiNode input. This sidesteps the dangerous case where the split |
duke@435 | 663 | // ConvI2LNode may become TOP if the input Value() does not |
duke@435 | 664 | // overlap the ConvI2L range, leaving a node which may not dominate its |
duke@435 | 665 | // uses. |
duke@435 | 666 | // A better fix for this problem can be found in the BugTraq entry, but |
duke@435 | 667 | // expediency for Mantis demands this hack. |
duke@435 | 668 | for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { |
duke@435 | 669 | Node* n = region->fast_out(i); |
duke@435 | 670 | if (n->is_Phi()) { |
duke@435 | 671 | for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { |
duke@435 | 672 | Node* m = n->fast_out(j); |
duke@435 | 673 | if (m->Opcode() == Op_ConvI2L) { |
duke@435 | 674 | return false; |
duke@435 | 675 | } |
duke@435 | 676 | } |
duke@435 | 677 | } |
duke@435 | 678 | } |
duke@435 | 679 | return true; |
duke@435 | 680 | } |
duke@435 | 681 | #endif |
duke@435 | 682 | |
duke@435 | 683 | |
duke@435 | 684 | //------------------------------place_near_use--------------------------------- |
duke@435 | 685 | // Place some computation next to use but not inside inner loops. |
duke@435 | 686 | // For inner loop uses move it to the preheader area. |
duke@435 | 687 | Node *PhaseIdealLoop::place_near_use( Node *useblock ) const { |
duke@435 | 688 | IdealLoopTree *u_loop = get_loop( useblock ); |
duke@435 | 689 | return (u_loop->_irreducible || u_loop->_child) |
duke@435 | 690 | ? useblock |
duke@435 | 691 | : u_loop->_head->in(LoopNode::EntryControl); |
duke@435 | 692 | } |
duke@435 | 693 | |
duke@435 | 694 | |
duke@435 | 695 | //------------------------------split_if_with_blocks_post---------------------- |
duke@435 | 696 | // Do the real work in a non-recursive function. CFG hackery wants to be |
duke@435 | 697 | // in the post-order, so it can dirty the I-DOM info and not use the dirtied |
duke@435 | 698 | // info. |
duke@435 | 699 | void PhaseIdealLoop::split_if_with_blocks_post( Node *n ) { |
duke@435 | 700 | |
duke@435 | 701 | // Cloning Cmp through Phi's involves the split-if transform. |
duke@435 | 702 | // FastLock is not used by an If |
duke@435 | 703 | if( n->is_Cmp() && !n->is_FastLock() ) { |
duke@435 | 704 | if( C->unique() > 35000 ) return; // Method too big |
duke@435 | 705 | |
duke@435 | 706 | // Do not do 'split-if' if irreducible loops are present. |
duke@435 | 707 | if( _has_irreducible_loops ) |
duke@435 | 708 | return; |
duke@435 | 709 | |
duke@435 | 710 | Node *n_ctrl = get_ctrl(n); |
duke@435 | 711 | // Determine if the Node has inputs from some local Phi. |
duke@435 | 712 | // Returns the block to clone thru. |
duke@435 | 713 | Node *n_blk = has_local_phi_input( n ); |
duke@435 | 714 | if( n_blk != n_ctrl ) return; |
duke@435 | 715 | |
duke@435 | 716 | if( merge_point_too_heavy(C, n_ctrl) ) |
duke@435 | 717 | return; |
duke@435 | 718 | |
duke@435 | 719 | if( n->outcnt() != 1 ) return; // Multiple bool's from 1 compare? |
duke@435 | 720 | Node *bol = n->unique_out(); |
duke@435 | 721 | assert( bol->is_Bool(), "expect a bool here" ); |
duke@435 | 722 | if( bol->outcnt() != 1 ) return;// Multiple branches from 1 compare? |
duke@435 | 723 | Node *iff = bol->unique_out(); |
duke@435 | 724 | |
duke@435 | 725 | // Check some safety conditions |
duke@435 | 726 | if( iff->is_If() ) { // Classic split-if? |
duke@435 | 727 | if( iff->in(0) != n_ctrl ) return; // Compare must be in same blk as if |
duke@435 | 728 | } else if (iff->is_CMove()) { // Trying to split-up a CMOVE |
duke@435 | 729 | if( get_ctrl(iff->in(2)) == n_ctrl || |
duke@435 | 730 | get_ctrl(iff->in(3)) == n_ctrl ) |
duke@435 | 731 | return; // Inputs not yet split-up |
duke@435 | 732 | if ( get_loop(n_ctrl) != get_loop(get_ctrl(iff)) ) { |
duke@435 | 733 | return; // Loop-invar test gates loop-varying CMOVE |
duke@435 | 734 | } |
duke@435 | 735 | } else { |
duke@435 | 736 | return; // some other kind of node, such as an Allocate |
duke@435 | 737 | } |
duke@435 | 738 | |
duke@435 | 739 | // Do not do 'split-if' if some paths are dead. First do dead code |
duke@435 | 740 | // elimination and then see if its still profitable. |
duke@435 | 741 | for( uint i = 1; i < n_ctrl->req(); i++ ) |
duke@435 | 742 | if( n_ctrl->in(i) == C->top() ) |
duke@435 | 743 | return; |
duke@435 | 744 | |
duke@435 | 745 | // When is split-if profitable? Every 'win' on means some control flow |
duke@435 | 746 | // goes dead, so it's almost always a win. |
duke@435 | 747 | int policy = 0; |
duke@435 | 748 | // If trying to do a 'Split-If' at the loop head, it is only |
duke@435 | 749 | // profitable if the cmp folds up on BOTH paths. Otherwise we |
duke@435 | 750 | // risk peeling a loop forever. |
duke@435 | 751 | |
duke@435 | 752 | // CNC - Disabled for now. Requires careful handling of loop |
duke@435 | 753 | // body selection for the cloned code. Also, make sure we check |
duke@435 | 754 | // for any input path not being in the same loop as n_ctrl. For |
duke@435 | 755 | // irreducible loops we cannot check for 'n_ctrl->is_Loop()' |
duke@435 | 756 | // because the alternative loop entry points won't be converted |
duke@435 | 757 | // into LoopNodes. |
duke@435 | 758 | IdealLoopTree *n_loop = get_loop(n_ctrl); |
duke@435 | 759 | for( uint j = 1; j < n_ctrl->req(); j++ ) |
duke@435 | 760 | if( get_loop(n_ctrl->in(j)) != n_loop ) |
duke@435 | 761 | return; |
duke@435 | 762 | |
duke@435 | 763 | #ifdef _LP64 |
duke@435 | 764 | // Check for safety of the merge point. |
duke@435 | 765 | if( !merge_point_safe(n_ctrl) ) { |
duke@435 | 766 | return; |
duke@435 | 767 | } |
duke@435 | 768 | #endif |
duke@435 | 769 | |
duke@435 | 770 | // Split compare 'n' through the merge point if it is profitable |
duke@435 | 771 | Node *phi = split_thru_phi( n, n_ctrl, policy ); |
duke@435 | 772 | if( !phi ) return; |
duke@435 | 773 | |
duke@435 | 774 | // Found a Phi to split thru! |
duke@435 | 775 | // Replace 'n' with the new phi |
duke@435 | 776 | _igvn.hash_delete(n); |
duke@435 | 777 | _igvn.subsume_node( n, phi ); |
duke@435 | 778 | |
duke@435 | 779 | // Now split the bool up thru the phi |
duke@435 | 780 | Node *bolphi = split_thru_phi( bol, n_ctrl, -1 ); |
duke@435 | 781 | _igvn.hash_delete(bol); |
duke@435 | 782 | _igvn.subsume_node( bol, bolphi ); |
duke@435 | 783 | assert( iff->in(1) == bolphi, "" ); |
duke@435 | 784 | if( bolphi->Value(&_igvn)->singleton() ) |
duke@435 | 785 | return; |
duke@435 | 786 | |
duke@435 | 787 | // Conditional-move? Must split up now |
duke@435 | 788 | if( !iff->is_If() ) { |
duke@435 | 789 | Node *cmovphi = split_thru_phi( iff, n_ctrl, -1 ); |
duke@435 | 790 | _igvn.hash_delete(iff); |
duke@435 | 791 | _igvn.subsume_node( iff, cmovphi ); |
duke@435 | 792 | return; |
duke@435 | 793 | } |
duke@435 | 794 | |
duke@435 | 795 | // Now split the IF |
duke@435 | 796 | do_split_if( iff ); |
duke@435 | 797 | return; |
duke@435 | 798 | } |
duke@435 | 799 | |
duke@435 | 800 | // Check for an IF ready to split; one that has its |
duke@435 | 801 | // condition codes input coming from a Phi at the block start. |
duke@435 | 802 | int n_op = n->Opcode(); |
duke@435 | 803 | |
duke@435 | 804 | // Check for an IF being dominated by another IF same test |
duke@435 | 805 | if( n_op == Op_If ) { |
duke@435 | 806 | Node *bol = n->in(1); |
duke@435 | 807 | uint max = bol->outcnt(); |
duke@435 | 808 | // Check for same test used more than once? |
duke@435 | 809 | if( n_op == Op_If && max > 1 && bol->is_Bool() ) { |
duke@435 | 810 | // Search up IDOMs to see if this IF is dominated. |
duke@435 | 811 | Node *cutoff = get_ctrl(bol); |
duke@435 | 812 | |
duke@435 | 813 | // Now search up IDOMs till cutoff, looking for a dominating test |
duke@435 | 814 | Node *prevdom = n; |
duke@435 | 815 | Node *dom = idom(prevdom); |
duke@435 | 816 | while( dom != cutoff ) { |
duke@435 | 817 | if( dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom ) { |
duke@435 | 818 | // Replace the dominated test with an obvious true or false. |
duke@435 | 819 | // Place it on the IGVN worklist for later cleanup. |
duke@435 | 820 | C->set_major_progress(); |
duke@435 | 821 | dominated_by( prevdom, n ); |
duke@435 | 822 | #ifndef PRODUCT |
duke@435 | 823 | if( VerifyLoopOptimizations ) verify(); |
duke@435 | 824 | #endif |
duke@435 | 825 | return; |
duke@435 | 826 | } |
duke@435 | 827 | prevdom = dom; |
duke@435 | 828 | dom = idom(prevdom); |
duke@435 | 829 | } |
duke@435 | 830 | } |
duke@435 | 831 | } |
duke@435 | 832 | |
duke@435 | 833 | // See if a shared loop-varying computation has no loop-varying uses. |
duke@435 | 834 | // Happens if something is only used for JVM state in uncommon trap exits, |
duke@435 | 835 | // like various versions of induction variable+offset. Clone the |
duke@435 | 836 | // computation per usage to allow it to sink out of the loop. |
duke@435 | 837 | if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about) |
duke@435 | 838 | Node *n_ctrl = get_ctrl(n); |
duke@435 | 839 | IdealLoopTree *n_loop = get_loop(n_ctrl); |
duke@435 | 840 | if( n_loop != _ltree_root ) { |
duke@435 | 841 | DUIterator_Fast imax, i = n->fast_outs(imax); |
duke@435 | 842 | for (; i < imax; i++) { |
duke@435 | 843 | Node* u = n->fast_out(i); |
duke@435 | 844 | if( !has_ctrl(u) ) break; // Found control user |
duke@435 | 845 | IdealLoopTree *u_loop = get_loop(get_ctrl(u)); |
duke@435 | 846 | if( u_loop == n_loop ) break; // Found loop-varying use |
duke@435 | 847 | if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop |
duke@435 | 848 | if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003 |
duke@435 | 849 | } |
duke@435 | 850 | bool did_break = (i < imax); // Did we break out of the previous loop? |
duke@435 | 851 | if (!did_break && n->outcnt() > 1) { // All uses in outer loops! |
duke@435 | 852 | Node *late_load_ctrl; |
duke@435 | 853 | if (n->is_Load()) { |
duke@435 | 854 | // If n is a load, get and save the result from get_late_ctrl(), |
duke@435 | 855 | // to be later used in calculating the control for n's clones. |
duke@435 | 856 | clear_dom_lca_tags(); |
duke@435 | 857 | late_load_ctrl = get_late_ctrl(n, n_ctrl); |
duke@435 | 858 | } |
duke@435 | 859 | // If n is a load, and the late control is the same as the current |
duke@435 | 860 | // control, then the cloning of n is a pointless exercise, because |
duke@435 | 861 | // GVN will ensure that we end up where we started. |
duke@435 | 862 | if (!n->is_Load() || late_load_ctrl != n_ctrl) { |
duke@435 | 863 | for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) { |
duke@435 | 864 | Node *u = n->last_out(j); // Clone private computation per use |
duke@435 | 865 | _igvn.hash_delete(u); |
duke@435 | 866 | _igvn._worklist.push(u); |
duke@435 | 867 | Node *x = n->clone(); // Clone computation |
duke@435 | 868 | Node *x_ctrl = NULL; |
duke@435 | 869 | if( u->is_Phi() ) { |
duke@435 | 870 | // Replace all uses of normal nodes. Replace Phi uses |
duke@435 | 871 | // individually, so the seperate Nodes can sink down |
duke@435 | 872 | // different paths. |
duke@435 | 873 | uint k = 1; |
duke@435 | 874 | while( u->in(k) != n ) k++; |
duke@435 | 875 | u->set_req( k, x ); |
duke@435 | 876 | // x goes next to Phi input path |
duke@435 | 877 | x_ctrl = u->in(0)->in(k); |
duke@435 | 878 | --j; |
duke@435 | 879 | } else { // Normal use |
duke@435 | 880 | // Replace all uses |
duke@435 | 881 | for( uint k = 0; k < u->req(); k++ ) { |
duke@435 | 882 | if( u->in(k) == n ) { |
duke@435 | 883 | u->set_req( k, x ); |
duke@435 | 884 | --j; |
duke@435 | 885 | } |
duke@435 | 886 | } |
duke@435 | 887 | x_ctrl = get_ctrl(u); |
duke@435 | 888 | } |
duke@435 | 889 | |
duke@435 | 890 | // Find control for 'x' next to use but not inside inner loops. |
duke@435 | 891 | // For inner loop uses get the preheader area. |
duke@435 | 892 | x_ctrl = place_near_use(x_ctrl); |
duke@435 | 893 | |
duke@435 | 894 | if (n->is_Load()) { |
duke@435 | 895 | // For loads, add a control edge to a CFG node outside of the loop |
duke@435 | 896 | // to force them to not combine and return back inside the loop |
duke@435 | 897 | // during GVN optimization (4641526). |
duke@435 | 898 | // |
duke@435 | 899 | // Because we are setting the actual control input, factor in |
duke@435 | 900 | // the result from get_late_ctrl() so we respect any |
duke@435 | 901 | // anti-dependences. (6233005). |
duke@435 | 902 | x_ctrl = dom_lca(late_load_ctrl, x_ctrl); |
duke@435 | 903 | |
duke@435 | 904 | // Don't allow the control input to be a CFG splitting node. |
duke@435 | 905 | // Such nodes should only have ProjNodes as outs, e.g. IfNode |
duke@435 | 906 | // should only have IfTrueNode and IfFalseNode (4985384). |
duke@435 | 907 | x_ctrl = find_non_split_ctrl(x_ctrl); |
duke@435 | 908 | assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone"); |
duke@435 | 909 | |
duke@435 | 910 | x->set_req(0, x_ctrl); |
duke@435 | 911 | } |
duke@435 | 912 | register_new_node(x, x_ctrl); |
duke@435 | 913 | |
duke@435 | 914 | // Some institutional knowledge is needed here: 'x' is |
duke@435 | 915 | // yanked because if the optimizer runs GVN on it all the |
duke@435 | 916 | // cloned x's will common up and undo this optimization and |
duke@435 | 917 | // be forced back in the loop. This is annoying because it |
duke@435 | 918 | // makes +VerifyOpto report false-positives on progress. I |
duke@435 | 919 | // tried setting control edges on the x's to force them to |
duke@435 | 920 | // not combine, but the matching gets worried when it tries |
duke@435 | 921 | // to fold a StoreP and an AddP together (as part of an |
duke@435 | 922 | // address expression) and the AddP and StoreP have |
duke@435 | 923 | // different controls. |
duke@435 | 924 | if( !x->is_Load() ) _igvn._worklist.yank(x); |
duke@435 | 925 | } |
duke@435 | 926 | _igvn.remove_dead_node(n); |
duke@435 | 927 | } |
duke@435 | 928 | } |
duke@435 | 929 | } |
duke@435 | 930 | } |
duke@435 | 931 | |
duke@435 | 932 | // Check for Opaque2's who's loop has disappeared - who's input is in the |
duke@435 | 933 | // same loop nest as their output. Remove 'em, they are no longer useful. |
duke@435 | 934 | if( n_op == Op_Opaque2 && |
duke@435 | 935 | n->in(1) != NULL && |
duke@435 | 936 | get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) { |
duke@435 | 937 | _igvn.add_users_to_worklist(n); |
duke@435 | 938 | _igvn.hash_delete(n); |
duke@435 | 939 | _igvn.subsume_node( n, n->in(1) ); |
duke@435 | 940 | } |
duke@435 | 941 | } |
duke@435 | 942 | |
duke@435 | 943 | //------------------------------split_if_with_blocks--------------------------- |
duke@435 | 944 | // Check for aggressive application of 'split-if' optimization, |
duke@435 | 945 | // using basic block level info. |
duke@435 | 946 | void PhaseIdealLoop::split_if_with_blocks( VectorSet &visited, Node_Stack &nstack ) { |
duke@435 | 947 | Node *n = C->root(); |
duke@435 | 948 | visited.set(n->_idx); // first, mark node as visited |
duke@435 | 949 | // Do pre-visit work for root |
duke@435 | 950 | n = split_if_with_blocks_pre( n ); |
duke@435 | 951 | uint cnt = n->outcnt(); |
duke@435 | 952 | uint i = 0; |
duke@435 | 953 | while (true) { |
duke@435 | 954 | // Visit all children |
duke@435 | 955 | if (i < cnt) { |
duke@435 | 956 | Node* use = n->raw_out(i); |
duke@435 | 957 | ++i; |
duke@435 | 958 | if (use->outcnt() != 0 && !visited.test_set(use->_idx)) { |
duke@435 | 959 | // Now do pre-visit work for this use |
duke@435 | 960 | use = split_if_with_blocks_pre( use ); |
duke@435 | 961 | nstack.push(n, i); // Save parent and next use's index. |
duke@435 | 962 | n = use; // Process all children of current use. |
duke@435 | 963 | cnt = use->outcnt(); |
duke@435 | 964 | i = 0; |
duke@435 | 965 | } |
duke@435 | 966 | } |
duke@435 | 967 | else { |
duke@435 | 968 | // All of n's children have been processed, complete post-processing. |
duke@435 | 969 | if (cnt != 0 && !n->is_Con()) { |
duke@435 | 970 | assert(has_node(n), "no dead nodes"); |
duke@435 | 971 | split_if_with_blocks_post( n ); |
duke@435 | 972 | } |
duke@435 | 973 | if (nstack.is_empty()) { |
duke@435 | 974 | // Finished all nodes on stack. |
duke@435 | 975 | break; |
duke@435 | 976 | } |
duke@435 | 977 | // Get saved parent node and next use's index. Visit the rest of uses. |
duke@435 | 978 | n = nstack.node(); |
duke@435 | 979 | cnt = n->outcnt(); |
duke@435 | 980 | i = nstack.index(); |
duke@435 | 981 | nstack.pop(); |
duke@435 | 982 | } |
duke@435 | 983 | } |
duke@435 | 984 | } |
duke@435 | 985 | |
duke@435 | 986 | |
duke@435 | 987 | //============================================================================= |
duke@435 | 988 | // |
duke@435 | 989 | // C L O N E A L O O P B O D Y |
duke@435 | 990 | // |
duke@435 | 991 | |
duke@435 | 992 | //------------------------------clone_iff-------------------------------------- |
duke@435 | 993 | // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. |
duke@435 | 994 | // "Nearly" because all Nodes have been cloned from the original in the loop, |
duke@435 | 995 | // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs |
duke@435 | 996 | // through the Phi recursively, and return a Bool. |
duke@435 | 997 | BoolNode *PhaseIdealLoop::clone_iff( PhiNode *phi, IdealLoopTree *loop ) { |
duke@435 | 998 | |
duke@435 | 999 | // Convert this Phi into a Phi merging Bools |
duke@435 | 1000 | uint i; |
duke@435 | 1001 | for( i = 1; i < phi->req(); i++ ) { |
duke@435 | 1002 | Node *b = phi->in(i); |
duke@435 | 1003 | if( b->is_Phi() ) { |
duke@435 | 1004 | _igvn.hash_delete(phi); |
duke@435 | 1005 | _igvn._worklist.push(phi); |
duke@435 | 1006 | phi->set_req(i, clone_iff( b->as_Phi(), loop )); |
duke@435 | 1007 | } else { |
duke@435 | 1008 | assert( b->is_Bool(), "" ); |
duke@435 | 1009 | } |
duke@435 | 1010 | } |
duke@435 | 1011 | |
duke@435 | 1012 | Node *sample_bool = phi->in(1); |
duke@435 | 1013 | Node *sample_cmp = sample_bool->in(1); |
duke@435 | 1014 | |
duke@435 | 1015 | // Make Phis to merge the Cmp's inputs. |
duke@435 | 1016 | int size = phi->in(0)->req(); |
duke@435 | 1017 | PhiNode *phi1 = new (C, size) PhiNode( phi->in(0), Type::TOP ); |
duke@435 | 1018 | PhiNode *phi2 = new (C, size) PhiNode( phi->in(0), Type::TOP ); |
duke@435 | 1019 | for( i = 1; i < phi->req(); i++ ) { |
duke@435 | 1020 | Node *n1 = phi->in(i)->in(1)->in(1); |
duke@435 | 1021 | Node *n2 = phi->in(i)->in(1)->in(2); |
duke@435 | 1022 | phi1->set_req( i, n1 ); |
duke@435 | 1023 | phi2->set_req( i, n2 ); |
duke@435 | 1024 | phi1->set_type( phi1->type()->meet(n1->bottom_type()) ); |
duke@435 | 1025 | phi2->set_type( phi2->type()->meet(n2->bottom_type()) ); |
duke@435 | 1026 | } |
duke@435 | 1027 | // See if these Phis have been made before. |
duke@435 | 1028 | // Register with optimizer |
duke@435 | 1029 | Node *hit1 = _igvn.hash_find_insert(phi1); |
duke@435 | 1030 | if( hit1 ) { // Hit, toss just made Phi |
duke@435 | 1031 | _igvn.remove_dead_node(phi1); // Remove new phi |
duke@435 | 1032 | assert( hit1->is_Phi(), "" ); |
duke@435 | 1033 | phi1 = (PhiNode*)hit1; // Use existing phi |
duke@435 | 1034 | } else { // Miss |
duke@435 | 1035 | _igvn.register_new_node_with_optimizer(phi1); |
duke@435 | 1036 | } |
duke@435 | 1037 | Node *hit2 = _igvn.hash_find_insert(phi2); |
duke@435 | 1038 | if( hit2 ) { // Hit, toss just made Phi |
duke@435 | 1039 | _igvn.remove_dead_node(phi2); // Remove new phi |
duke@435 | 1040 | assert( hit2->is_Phi(), "" ); |
duke@435 | 1041 | phi2 = (PhiNode*)hit2; // Use existing phi |
duke@435 | 1042 | } else { // Miss |
duke@435 | 1043 | _igvn.register_new_node_with_optimizer(phi2); |
duke@435 | 1044 | } |
duke@435 | 1045 | // Register Phis with loop/block info |
duke@435 | 1046 | set_ctrl(phi1, phi->in(0)); |
duke@435 | 1047 | set_ctrl(phi2, phi->in(0)); |
duke@435 | 1048 | // Make a new Cmp |
duke@435 | 1049 | Node *cmp = sample_cmp->clone(); |
duke@435 | 1050 | cmp->set_req( 1, phi1 ); |
duke@435 | 1051 | cmp->set_req( 2, phi2 ); |
duke@435 | 1052 | _igvn.register_new_node_with_optimizer(cmp); |
duke@435 | 1053 | set_ctrl(cmp, phi->in(0)); |
duke@435 | 1054 | |
duke@435 | 1055 | // Make a new Bool |
duke@435 | 1056 | Node *b = sample_bool->clone(); |
duke@435 | 1057 | b->set_req(1,cmp); |
duke@435 | 1058 | _igvn.register_new_node_with_optimizer(b); |
duke@435 | 1059 | set_ctrl(b, phi->in(0)); |
duke@435 | 1060 | |
duke@435 | 1061 | assert( b->is_Bool(), "" ); |
duke@435 | 1062 | return (BoolNode*)b; |
duke@435 | 1063 | } |
duke@435 | 1064 | |
duke@435 | 1065 | //------------------------------clone_bool------------------------------------- |
duke@435 | 1066 | // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. |
duke@435 | 1067 | // "Nearly" because all Nodes have been cloned from the original in the loop, |
duke@435 | 1068 | // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs |
duke@435 | 1069 | // through the Phi recursively, and return a Bool. |
duke@435 | 1070 | CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) { |
duke@435 | 1071 | uint i; |
duke@435 | 1072 | // Convert this Phi into a Phi merging Bools |
duke@435 | 1073 | for( i = 1; i < phi->req(); i++ ) { |
duke@435 | 1074 | Node *b = phi->in(i); |
duke@435 | 1075 | if( b->is_Phi() ) { |
duke@435 | 1076 | _igvn.hash_delete(phi); |
duke@435 | 1077 | _igvn._worklist.push(phi); |
duke@435 | 1078 | phi->set_req(i, clone_bool( b->as_Phi(), loop )); |
duke@435 | 1079 | } else { |
duke@435 | 1080 | assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" ); |
duke@435 | 1081 | } |
duke@435 | 1082 | } |
duke@435 | 1083 | |
duke@435 | 1084 | Node *sample_cmp = phi->in(1); |
duke@435 | 1085 | |
duke@435 | 1086 | // Make Phis to merge the Cmp's inputs. |
duke@435 | 1087 | int size = phi->in(0)->req(); |
duke@435 | 1088 | PhiNode *phi1 = new (C, size) PhiNode( phi->in(0), Type::TOP ); |
duke@435 | 1089 | PhiNode *phi2 = new (C, size) PhiNode( phi->in(0), Type::TOP ); |
duke@435 | 1090 | for( uint j = 1; j < phi->req(); j++ ) { |
duke@435 | 1091 | Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP |
duke@435 | 1092 | Node *n1, *n2; |
duke@435 | 1093 | if( cmp_top->is_Cmp() ) { |
duke@435 | 1094 | n1 = cmp_top->in(1); |
duke@435 | 1095 | n2 = cmp_top->in(2); |
duke@435 | 1096 | } else { |
duke@435 | 1097 | n1 = n2 = cmp_top; |
duke@435 | 1098 | } |
duke@435 | 1099 | phi1->set_req( j, n1 ); |
duke@435 | 1100 | phi2->set_req( j, n2 ); |
duke@435 | 1101 | phi1->set_type( phi1->type()->meet(n1->bottom_type()) ); |
duke@435 | 1102 | phi2->set_type( phi2->type()->meet(n2->bottom_type()) ); |
duke@435 | 1103 | } |
duke@435 | 1104 | |
duke@435 | 1105 | // See if these Phis have been made before. |
duke@435 | 1106 | // Register with optimizer |
duke@435 | 1107 | Node *hit1 = _igvn.hash_find_insert(phi1); |
duke@435 | 1108 | if( hit1 ) { // Hit, toss just made Phi |
duke@435 | 1109 | _igvn.remove_dead_node(phi1); // Remove new phi |
duke@435 | 1110 | assert( hit1->is_Phi(), "" ); |
duke@435 | 1111 | phi1 = (PhiNode*)hit1; // Use existing phi |
duke@435 | 1112 | } else { // Miss |
duke@435 | 1113 | _igvn.register_new_node_with_optimizer(phi1); |
duke@435 | 1114 | } |
duke@435 | 1115 | Node *hit2 = _igvn.hash_find_insert(phi2); |
duke@435 | 1116 | if( hit2 ) { // Hit, toss just made Phi |
duke@435 | 1117 | _igvn.remove_dead_node(phi2); // Remove new phi |
duke@435 | 1118 | assert( hit2->is_Phi(), "" ); |
duke@435 | 1119 | phi2 = (PhiNode*)hit2; // Use existing phi |
duke@435 | 1120 | } else { // Miss |
duke@435 | 1121 | _igvn.register_new_node_with_optimizer(phi2); |
duke@435 | 1122 | } |
duke@435 | 1123 | // Register Phis with loop/block info |
duke@435 | 1124 | set_ctrl(phi1, phi->in(0)); |
duke@435 | 1125 | set_ctrl(phi2, phi->in(0)); |
duke@435 | 1126 | // Make a new Cmp |
duke@435 | 1127 | Node *cmp = sample_cmp->clone(); |
duke@435 | 1128 | cmp->set_req( 1, phi1 ); |
duke@435 | 1129 | cmp->set_req( 2, phi2 ); |
duke@435 | 1130 | _igvn.register_new_node_with_optimizer(cmp); |
duke@435 | 1131 | set_ctrl(cmp, phi->in(0)); |
duke@435 | 1132 | |
duke@435 | 1133 | assert( cmp->is_Cmp(), "" ); |
duke@435 | 1134 | return (CmpNode*)cmp; |
duke@435 | 1135 | } |
duke@435 | 1136 | |
duke@435 | 1137 | //------------------------------sink_use--------------------------------------- |
duke@435 | 1138 | // If 'use' was in the loop-exit block, it now needs to be sunk |
duke@435 | 1139 | // below the post-loop merge point. |
duke@435 | 1140 | void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) { |
duke@435 | 1141 | if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) { |
duke@435 | 1142 | set_ctrl(use, post_loop); |
duke@435 | 1143 | for (DUIterator j = use->outs(); use->has_out(j); j++) |
duke@435 | 1144 | sink_use(use->out(j), post_loop); |
duke@435 | 1145 | } |
duke@435 | 1146 | } |
duke@435 | 1147 | |
duke@435 | 1148 | //------------------------------clone_loop------------------------------------- |
duke@435 | 1149 | // |
duke@435 | 1150 | // C L O N E A L O O P B O D Y |
duke@435 | 1151 | // |
duke@435 | 1152 | // This is the basic building block of the loop optimizations. It clones an |
duke@435 | 1153 | // entire loop body. It makes an old_new loop body mapping; with this mapping |
duke@435 | 1154 | // you can find the new-loop equivalent to an old-loop node. All new-loop |
duke@435 | 1155 | // nodes are exactly equal to their old-loop counterparts, all edges are the |
duke@435 | 1156 | // same. All exits from the old-loop now have a RegionNode that merges the |
duke@435 | 1157 | // equivalent new-loop path. This is true even for the normal "loop-exit" |
duke@435 | 1158 | // condition. All uses of loop-invariant old-loop values now come from (one |
duke@435 | 1159 | // or more) Phis that merge their new-loop equivalents. |
duke@435 | 1160 | // |
duke@435 | 1161 | // This operation leaves the graph in an illegal state: there are two valid |
duke@435 | 1162 | // control edges coming from the loop pre-header to both loop bodies. I'll |
duke@435 | 1163 | // definitely have to hack the graph after running this transform. |
duke@435 | 1164 | // |
duke@435 | 1165 | // From this building block I will further edit edges to perform loop peeling |
duke@435 | 1166 | // or loop unrolling or iteration splitting (Range-Check-Elimination), etc. |
duke@435 | 1167 | // |
duke@435 | 1168 | // Parameter side_by_size_idom: |
duke@435 | 1169 | // When side_by_size_idom is NULL, the dominator tree is constructed for |
duke@435 | 1170 | // the clone loop to dominate the original. Used in construction of |
duke@435 | 1171 | // pre-main-post loop sequence. |
duke@435 | 1172 | // When nonnull, the clone and original are side-by-side, both are |
duke@435 | 1173 | // dominated by the side_by_side_idom node. Used in construction of |
duke@435 | 1174 | // unswitched loops. |
duke@435 | 1175 | void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd, |
duke@435 | 1176 | Node* side_by_side_idom) { |
duke@435 | 1177 | |
duke@435 | 1178 | // Step 1: Clone the loop body. Make the old->new mapping. |
duke@435 | 1179 | uint i; |
duke@435 | 1180 | for( i = 0; i < loop->_body.size(); i++ ) { |
duke@435 | 1181 | Node *old = loop->_body.at(i); |
duke@435 | 1182 | Node *nnn = old->clone(); |
duke@435 | 1183 | old_new.map( old->_idx, nnn ); |
duke@435 | 1184 | _igvn.register_new_node_with_optimizer(nnn); |
duke@435 | 1185 | } |
duke@435 | 1186 | |
duke@435 | 1187 | |
duke@435 | 1188 | // Step 2: Fix the edges in the new body. If the old input is outside the |
duke@435 | 1189 | // loop use it. If the old input is INside the loop, use the corresponding |
duke@435 | 1190 | // new node instead. |
duke@435 | 1191 | for( i = 0; i < loop->_body.size(); i++ ) { |
duke@435 | 1192 | Node *old = loop->_body.at(i); |
duke@435 | 1193 | Node *nnn = old_new[old->_idx]; |
duke@435 | 1194 | // Fix CFG/Loop controlling the new node |
duke@435 | 1195 | if (has_ctrl(old)) { |
duke@435 | 1196 | set_ctrl(nnn, old_new[get_ctrl(old)->_idx]); |
duke@435 | 1197 | } else { |
duke@435 | 1198 | set_loop(nnn, loop->_parent); |
duke@435 | 1199 | if (old->outcnt() > 0) { |
duke@435 | 1200 | set_idom( nnn, old_new[idom(old)->_idx], dd ); |
duke@435 | 1201 | } |
duke@435 | 1202 | } |
duke@435 | 1203 | // Correct edges to the new node |
duke@435 | 1204 | for( uint j = 0; j < nnn->req(); j++ ) { |
duke@435 | 1205 | Node *n = nnn->in(j); |
duke@435 | 1206 | if( n ) { |
duke@435 | 1207 | IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n ); |
duke@435 | 1208 | if( loop->is_member( old_in_loop ) ) |
duke@435 | 1209 | nnn->set_req(j, old_new[n->_idx]); |
duke@435 | 1210 | } |
duke@435 | 1211 | } |
duke@435 | 1212 | _igvn.hash_find_insert(nnn); |
duke@435 | 1213 | } |
duke@435 | 1214 | Node *newhead = old_new[loop->_head->_idx]; |
duke@435 | 1215 | set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); |
duke@435 | 1216 | |
duke@435 | 1217 | |
duke@435 | 1218 | // Step 3: Now fix control uses. Loop varying control uses have already |
duke@435 | 1219 | // been fixed up (as part of all input edges in Step 2). Loop invariant |
duke@435 | 1220 | // control uses must be either an IfFalse or an IfTrue. Make a merge |
duke@435 | 1221 | // point to merge the old and new IfFalse/IfTrue nodes; make the use |
duke@435 | 1222 | // refer to this. |
duke@435 | 1223 | ResourceArea *area = Thread::current()->resource_area(); |
duke@435 | 1224 | Node_List worklist(area); |
duke@435 | 1225 | uint new_counter = C->unique(); |
duke@435 | 1226 | for( i = 0; i < loop->_body.size(); i++ ) { |
duke@435 | 1227 | Node* old = loop->_body.at(i); |
duke@435 | 1228 | if( !old->is_CFG() ) continue; |
duke@435 | 1229 | Node* nnn = old_new[old->_idx]; |
duke@435 | 1230 | |
duke@435 | 1231 | // Copy uses to a worklist, so I can munge the def-use info |
duke@435 | 1232 | // with impunity. |
duke@435 | 1233 | for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) |
duke@435 | 1234 | worklist.push(old->fast_out(j)); |
duke@435 | 1235 | |
duke@435 | 1236 | while( worklist.size() ) { // Visit all uses |
duke@435 | 1237 | Node *use = worklist.pop(); |
duke@435 | 1238 | if (!has_node(use)) continue; // Ignore dead nodes |
duke@435 | 1239 | IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); |
duke@435 | 1240 | if( !loop->is_member( use_loop ) && use->is_CFG() ) { |
duke@435 | 1241 | // Both OLD and USE are CFG nodes here. |
duke@435 | 1242 | assert( use->is_Proj(), "" ); |
duke@435 | 1243 | |
duke@435 | 1244 | // Clone the loop exit control projection |
duke@435 | 1245 | Node *newuse = use->clone(); |
duke@435 | 1246 | newuse->set_req(0,nnn); |
duke@435 | 1247 | _igvn.register_new_node_with_optimizer(newuse); |
duke@435 | 1248 | set_loop(newuse, use_loop); |
duke@435 | 1249 | set_idom(newuse, nnn, dom_depth(nnn) + 1 ); |
duke@435 | 1250 | |
duke@435 | 1251 | // We need a Region to merge the exit from the peeled body and the |
duke@435 | 1252 | // exit from the old loop body. |
duke@435 | 1253 | RegionNode *r = new (C, 3) RegionNode(3); |
duke@435 | 1254 | // Map the old use to the new merge point |
duke@435 | 1255 | old_new.map( use->_idx, r ); |
duke@435 | 1256 | uint dd_r = MIN2(dom_depth(newuse),dom_depth(use)); |
duke@435 | 1257 | assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" ); |
duke@435 | 1258 | |
duke@435 | 1259 | // The original user of 'use' uses 'r' instead. |
duke@435 | 1260 | for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) { |
duke@435 | 1261 | Node* useuse = use->last_out(l); |
duke@435 | 1262 | _igvn.hash_delete(useuse); |
duke@435 | 1263 | _igvn._worklist.push(useuse); |
duke@435 | 1264 | uint uses_found = 0; |
duke@435 | 1265 | if( useuse->in(0) == use ) { |
duke@435 | 1266 | useuse->set_req(0, r); |
duke@435 | 1267 | uses_found++; |
duke@435 | 1268 | if( useuse->is_CFG() ) { |
duke@435 | 1269 | assert( dom_depth(useuse) > dd_r, "" ); |
duke@435 | 1270 | set_idom(useuse, r, dom_depth(useuse)); |
duke@435 | 1271 | } |
duke@435 | 1272 | } |
duke@435 | 1273 | for( uint k = 1; k < useuse->req(); k++ ) { |
duke@435 | 1274 | if( useuse->in(k) == use ) { |
duke@435 | 1275 | useuse->set_req(k, r); |
duke@435 | 1276 | uses_found++; |
duke@435 | 1277 | } |
duke@435 | 1278 | } |
duke@435 | 1279 | l -= uses_found; // we deleted 1 or more copies of this edge |
duke@435 | 1280 | } |
duke@435 | 1281 | |
duke@435 | 1282 | // Now finish up 'r' |
duke@435 | 1283 | r->set_req( 1, newuse ); |
duke@435 | 1284 | r->set_req( 2, use ); |
duke@435 | 1285 | _igvn.register_new_node_with_optimizer(r); |
duke@435 | 1286 | set_loop(r, use_loop); |
duke@435 | 1287 | set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r); |
duke@435 | 1288 | } // End of if a loop-exit test |
duke@435 | 1289 | } |
duke@435 | 1290 | } |
duke@435 | 1291 | |
duke@435 | 1292 | // Step 4: If loop-invariant use is not control, it must be dominated by a |
duke@435 | 1293 | // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region |
duke@435 | 1294 | // there if needed. Make a Phi there merging old and new used values. |
duke@435 | 1295 | Node_List *split_if_set = NULL; |
duke@435 | 1296 | Node_List *split_bool_set = NULL; |
duke@435 | 1297 | Node_List *split_cex_set = NULL; |
duke@435 | 1298 | for( i = 0; i < loop->_body.size(); i++ ) { |
duke@435 | 1299 | Node* old = loop->_body.at(i); |
duke@435 | 1300 | Node* nnn = old_new[old->_idx]; |
duke@435 | 1301 | // Copy uses to a worklist, so I can munge the def-use info |
duke@435 | 1302 | // with impunity. |
duke@435 | 1303 | for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) |
duke@435 | 1304 | worklist.push(old->fast_out(j)); |
duke@435 | 1305 | |
duke@435 | 1306 | while( worklist.size() ) { |
duke@435 | 1307 | Node *use = worklist.pop(); |
duke@435 | 1308 | if (!has_node(use)) continue; // Ignore dead nodes |
duke@435 | 1309 | if (use->in(0) == C->top()) continue; |
duke@435 | 1310 | IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); |
duke@435 | 1311 | // Check for data-use outside of loop - at least one of OLD or USE |
duke@435 | 1312 | // must not be a CFG node. |
duke@435 | 1313 | if( !loop->is_member( use_loop ) && (!old->is_CFG() || !use->is_CFG())) { |
duke@435 | 1314 | |
duke@435 | 1315 | // If the Data use is an IF, that means we have an IF outside of the |
duke@435 | 1316 | // loop that is switching on a condition that is set inside of the |
duke@435 | 1317 | // loop. Happens if people set a loop-exit flag; then test the flag |
duke@435 | 1318 | // in the loop to break the loop, then test is again outside of the |
duke@435 | 1319 | // loop to determine which way the loop exited. |
duke@435 | 1320 | if( use->is_If() || use->is_CMove() ) { |
duke@435 | 1321 | // Since this code is highly unlikely, we lazily build the worklist |
duke@435 | 1322 | // of such Nodes to go split. |
duke@435 | 1323 | if( !split_if_set ) |
duke@435 | 1324 | split_if_set = new Node_List(area); |
duke@435 | 1325 | split_if_set->push(use); |
duke@435 | 1326 | } |
duke@435 | 1327 | if( use->is_Bool() ) { |
duke@435 | 1328 | if( !split_bool_set ) |
duke@435 | 1329 | split_bool_set = new Node_List(area); |
duke@435 | 1330 | split_bool_set->push(use); |
duke@435 | 1331 | } |
duke@435 | 1332 | if( use->Opcode() == Op_CreateEx ) { |
duke@435 | 1333 | if( !split_cex_set ) |
duke@435 | 1334 | split_cex_set = new Node_List(area); |
duke@435 | 1335 | split_cex_set->push(use); |
duke@435 | 1336 | } |
duke@435 | 1337 | |
duke@435 | 1338 | |
duke@435 | 1339 | // Get "block" use is in |
duke@435 | 1340 | uint idx = 0; |
duke@435 | 1341 | while( use->in(idx) != old ) idx++; |
duke@435 | 1342 | Node *prev = use->is_CFG() ? use : get_ctrl(use); |
duke@435 | 1343 | assert( !loop->is_member( get_loop( prev ) ), "" ); |
duke@435 | 1344 | Node *cfg = prev->_idx >= new_counter |
duke@435 | 1345 | ? prev->in(2) |
duke@435 | 1346 | : idom(prev); |
duke@435 | 1347 | if( use->is_Phi() ) // Phi use is in prior block |
duke@435 | 1348 | cfg = prev->in(idx); // NOT in block of Phi itself |
duke@435 | 1349 | if (cfg->is_top()) { // Use is dead? |
duke@435 | 1350 | _igvn.hash_delete(use); |
duke@435 | 1351 | _igvn._worklist.push(use); |
duke@435 | 1352 | use->set_req(idx, C->top()); |
duke@435 | 1353 | continue; |
duke@435 | 1354 | } |
duke@435 | 1355 | |
duke@435 | 1356 | while( !loop->is_member( get_loop( cfg ) ) ) { |
duke@435 | 1357 | prev = cfg; |
duke@435 | 1358 | cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg); |
duke@435 | 1359 | } |
duke@435 | 1360 | // If the use occurs after merging several exits from the loop, then |
duke@435 | 1361 | // old value must have dominated all those exits. Since the same old |
duke@435 | 1362 | // value was used on all those exits we did not need a Phi at this |
duke@435 | 1363 | // merge point. NOW we do need a Phi here. Each loop exit value |
duke@435 | 1364 | // is now merged with the peeled body exit; each exit gets its own |
duke@435 | 1365 | // private Phi and those Phis need to be merged here. |
duke@435 | 1366 | Node *phi; |
duke@435 | 1367 | if( prev->is_Region() ) { |
duke@435 | 1368 | if( idx == 0 ) { // Updating control edge? |
duke@435 | 1369 | phi = prev; // Just use existing control |
duke@435 | 1370 | } else { // Else need a new Phi |
duke@435 | 1371 | phi = PhiNode::make( prev, old ); |
duke@435 | 1372 | // Now recursively fix up the new uses of old! |
duke@435 | 1373 | for( uint i = 1; i < prev->req(); i++ ) { |
duke@435 | 1374 | worklist.push(phi); // Onto worklist once for each 'old' input |
duke@435 | 1375 | } |
duke@435 | 1376 | } |
duke@435 | 1377 | } else { |
duke@435 | 1378 | // Get new RegionNode merging old and new loop exits |
duke@435 | 1379 | prev = old_new[prev->_idx]; |
duke@435 | 1380 | assert( prev, "just made this in step 7" ); |
duke@435 | 1381 | if( idx == 0 ) { // Updating control edge? |
duke@435 | 1382 | phi = prev; // Just use existing control |
duke@435 | 1383 | } else { // Else need a new Phi |
duke@435 | 1384 | // Make a new Phi merging data values properly |
duke@435 | 1385 | phi = PhiNode::make( prev, old ); |
duke@435 | 1386 | phi->set_req( 1, nnn ); |
duke@435 | 1387 | } |
duke@435 | 1388 | } |
duke@435 | 1389 | // If inserting a new Phi, check for prior hits |
duke@435 | 1390 | if( idx != 0 ) { |
duke@435 | 1391 | Node *hit = _igvn.hash_find_insert(phi); |
duke@435 | 1392 | if( hit == NULL ) { |
duke@435 | 1393 | _igvn.register_new_node_with_optimizer(phi); // Register new phi |
duke@435 | 1394 | } else { // or |
duke@435 | 1395 | // Remove the new phi from the graph and use the hit |
duke@435 | 1396 | _igvn.remove_dead_node(phi); |
duke@435 | 1397 | phi = hit; // Use existing phi |
duke@435 | 1398 | } |
duke@435 | 1399 | set_ctrl(phi, prev); |
duke@435 | 1400 | } |
duke@435 | 1401 | // Make 'use' use the Phi instead of the old loop body exit value |
duke@435 | 1402 | _igvn.hash_delete(use); |
duke@435 | 1403 | _igvn._worklist.push(use); |
duke@435 | 1404 | use->set_req(idx, phi); |
duke@435 | 1405 | if( use->_idx >= new_counter ) { // If updating new phis |
duke@435 | 1406 | // Not needed for correctness, but prevents a weak assert |
duke@435 | 1407 | // in AddPNode from tripping (when we end up with different |
duke@435 | 1408 | // base & derived Phis that will become the same after |
duke@435 | 1409 | // IGVN does CSE). |
duke@435 | 1410 | Node *hit = _igvn.hash_find_insert(use); |
duke@435 | 1411 | if( hit ) // Go ahead and re-hash for hits. |
duke@435 | 1412 | _igvn.subsume_node( use, hit ); |
duke@435 | 1413 | } |
duke@435 | 1414 | |
duke@435 | 1415 | // If 'use' was in the loop-exit block, it now needs to be sunk |
duke@435 | 1416 | // below the post-loop merge point. |
duke@435 | 1417 | sink_use( use, prev ); |
duke@435 | 1418 | } |
duke@435 | 1419 | } |
duke@435 | 1420 | } |
duke@435 | 1421 | |
duke@435 | 1422 | // Check for IFs that need splitting/cloning. Happens if an IF outside of |
duke@435 | 1423 | // the loop uses a condition set in the loop. The original IF probably |
duke@435 | 1424 | // takes control from one or more OLD Regions (which in turn get from NEW |
duke@435 | 1425 | // Regions). In any case, there will be a set of Phis for each merge point |
duke@435 | 1426 | // from the IF up to where the original BOOL def exists the loop. |
duke@435 | 1427 | if( split_if_set ) { |
duke@435 | 1428 | while( split_if_set->size() ) { |
duke@435 | 1429 | Node *iff = split_if_set->pop(); |
duke@435 | 1430 | if( iff->in(1)->is_Phi() ) { |
duke@435 | 1431 | BoolNode *b = clone_iff( iff->in(1)->as_Phi(), loop ); |
duke@435 | 1432 | _igvn.hash_delete(iff); |
duke@435 | 1433 | _igvn._worklist.push(iff); |
duke@435 | 1434 | iff->set_req(1, b); |
duke@435 | 1435 | } |
duke@435 | 1436 | } |
duke@435 | 1437 | } |
duke@435 | 1438 | if( split_bool_set ) { |
duke@435 | 1439 | while( split_bool_set->size() ) { |
duke@435 | 1440 | Node *b = split_bool_set->pop(); |
duke@435 | 1441 | Node *phi = b->in(1); |
duke@435 | 1442 | assert( phi->is_Phi(), "" ); |
duke@435 | 1443 | CmpNode *cmp = clone_bool( (PhiNode*)phi, loop ); |
duke@435 | 1444 | _igvn.hash_delete(b); |
duke@435 | 1445 | _igvn._worklist.push(b); |
duke@435 | 1446 | b->set_req(1, cmp); |
duke@435 | 1447 | } |
duke@435 | 1448 | } |
duke@435 | 1449 | if( split_cex_set ) { |
duke@435 | 1450 | while( split_cex_set->size() ) { |
duke@435 | 1451 | Node *b = split_cex_set->pop(); |
duke@435 | 1452 | assert( b->in(0)->is_Region(), "" ); |
duke@435 | 1453 | assert( b->in(1)->is_Phi(), "" ); |
duke@435 | 1454 | assert( b->in(0)->in(0) == b->in(1)->in(0), "" ); |
duke@435 | 1455 | split_up( b, b->in(0), NULL ); |
duke@435 | 1456 | } |
duke@435 | 1457 | } |
duke@435 | 1458 | |
duke@435 | 1459 | } |
duke@435 | 1460 | |
duke@435 | 1461 | |
duke@435 | 1462 | //---------------------- stride_of_possible_iv ------------------------------------- |
duke@435 | 1463 | // Looks for an iff/bool/comp with one operand of the compare |
duke@435 | 1464 | // being a cycle involving an add and a phi, |
duke@435 | 1465 | // with an optional truncation (left-shift followed by a right-shift) |
duke@435 | 1466 | // of the add. Returns zero if not an iv. |
duke@435 | 1467 | int PhaseIdealLoop::stride_of_possible_iv(Node* iff) { |
duke@435 | 1468 | Node* trunc1 = NULL; |
duke@435 | 1469 | Node* trunc2 = NULL; |
duke@435 | 1470 | const TypeInt* ttype = NULL; |
duke@435 | 1471 | if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) { |
duke@435 | 1472 | return 0; |
duke@435 | 1473 | } |
duke@435 | 1474 | BoolNode* bl = iff->in(1)->as_Bool(); |
duke@435 | 1475 | Node* cmp = bl->in(1); |
duke@435 | 1476 | if (!cmp || cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU) { |
duke@435 | 1477 | return 0; |
duke@435 | 1478 | } |
duke@435 | 1479 | // Must have an invariant operand |
duke@435 | 1480 | if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) { |
duke@435 | 1481 | return 0; |
duke@435 | 1482 | } |
duke@435 | 1483 | Node* add2 = NULL; |
duke@435 | 1484 | Node* cmp1 = cmp->in(1); |
duke@435 | 1485 | if (cmp1->is_Phi()) { |
duke@435 | 1486 | // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) ))) |
duke@435 | 1487 | Node* phi = cmp1; |
duke@435 | 1488 | for (uint i = 1; i < phi->req(); i++) { |
duke@435 | 1489 | Node* in = phi->in(i); |
duke@435 | 1490 | Node* add = CountedLoopNode::match_incr_with_optional_truncation(in, |
duke@435 | 1491 | &trunc1, &trunc2, &ttype); |
duke@435 | 1492 | if (add && add->in(1) == phi) { |
duke@435 | 1493 | add2 = add->in(2); |
duke@435 | 1494 | break; |
duke@435 | 1495 | } |
duke@435 | 1496 | } |
duke@435 | 1497 | } else { |
duke@435 | 1498 | // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) ))) |
duke@435 | 1499 | Node* addtrunc = cmp1; |
duke@435 | 1500 | Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc, |
duke@435 | 1501 | &trunc1, &trunc2, &ttype); |
duke@435 | 1502 | if (add && add->in(1)->is_Phi()) { |
duke@435 | 1503 | Node* phi = add->in(1); |
duke@435 | 1504 | for (uint i = 1; i < phi->req(); i++) { |
duke@435 | 1505 | if (phi->in(i) == addtrunc) { |
duke@435 | 1506 | add2 = add->in(2); |
duke@435 | 1507 | break; |
duke@435 | 1508 | } |
duke@435 | 1509 | } |
duke@435 | 1510 | } |
duke@435 | 1511 | } |
duke@435 | 1512 | if (add2 != NULL) { |
duke@435 | 1513 | const TypeInt* add2t = _igvn.type(add2)->is_int(); |
duke@435 | 1514 | if (add2t->is_con()) { |
duke@435 | 1515 | return add2t->get_con(); |
duke@435 | 1516 | } |
duke@435 | 1517 | } |
duke@435 | 1518 | return 0; |
duke@435 | 1519 | } |
duke@435 | 1520 | |
duke@435 | 1521 | |
duke@435 | 1522 | //---------------------- stay_in_loop ------------------------------------- |
duke@435 | 1523 | // Return the (unique) control output node that's in the loop (if it exists.) |
duke@435 | 1524 | Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) { |
duke@435 | 1525 | Node* unique = NULL; |
duke@435 | 1526 | if (!n) return NULL; |
duke@435 | 1527 | for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { |
duke@435 | 1528 | Node* use = n->fast_out(i); |
duke@435 | 1529 | if (!has_ctrl(use) && loop->is_member(get_loop(use))) { |
duke@435 | 1530 | if (unique != NULL) { |
duke@435 | 1531 | return NULL; |
duke@435 | 1532 | } |
duke@435 | 1533 | unique = use; |
duke@435 | 1534 | } |
duke@435 | 1535 | } |
duke@435 | 1536 | return unique; |
duke@435 | 1537 | } |
duke@435 | 1538 | |
duke@435 | 1539 | //------------------------------ register_node ------------------------------------- |
duke@435 | 1540 | // Utility to register node "n" with PhaseIdealLoop |
duke@435 | 1541 | void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) { |
duke@435 | 1542 | _igvn.register_new_node_with_optimizer(n); |
duke@435 | 1543 | loop->_body.push(n); |
duke@435 | 1544 | if (n->is_CFG()) { |
duke@435 | 1545 | set_loop(n, loop); |
duke@435 | 1546 | set_idom(n, pred, ddepth); |
duke@435 | 1547 | } else { |
duke@435 | 1548 | set_ctrl(n, pred); |
duke@435 | 1549 | } |
duke@435 | 1550 | } |
duke@435 | 1551 | |
duke@435 | 1552 | //------------------------------ proj_clone ------------------------------------- |
duke@435 | 1553 | // Utility to create an if-projection |
duke@435 | 1554 | ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) { |
duke@435 | 1555 | ProjNode* c = p->clone()->as_Proj(); |
duke@435 | 1556 | c->set_req(0, iff); |
duke@435 | 1557 | return c; |
duke@435 | 1558 | } |
duke@435 | 1559 | |
duke@435 | 1560 | //------------------------------ short_circuit_if ------------------------------------- |
duke@435 | 1561 | // Force the iff control output to be the live_proj |
duke@435 | 1562 | Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) { |
duke@435 | 1563 | int proj_con = live_proj->_con; |
duke@435 | 1564 | assert(proj_con == 0 || proj_con == 1, "false or true projection"); |
duke@435 | 1565 | Node *con = _igvn.intcon(proj_con); |
duke@435 | 1566 | set_ctrl(con, C->root()); |
duke@435 | 1567 | if (iff) { |
duke@435 | 1568 | iff->set_req(1, con); |
duke@435 | 1569 | } |
duke@435 | 1570 | return con; |
duke@435 | 1571 | } |
duke@435 | 1572 | |
duke@435 | 1573 | //------------------------------ insert_if_before_proj ------------------------------------- |
duke@435 | 1574 | // Insert a new if before an if projection (* - new node) |
duke@435 | 1575 | // |
duke@435 | 1576 | // before |
duke@435 | 1577 | // if(test) |
duke@435 | 1578 | // / \ |
duke@435 | 1579 | // v v |
duke@435 | 1580 | // other-proj proj (arg) |
duke@435 | 1581 | // |
duke@435 | 1582 | // after |
duke@435 | 1583 | // if(test) |
duke@435 | 1584 | // / \ |
duke@435 | 1585 | // / v |
duke@435 | 1586 | // | * proj-clone |
duke@435 | 1587 | // v | |
duke@435 | 1588 | // other-proj v |
duke@435 | 1589 | // * new_if(relop(cmp[IU](left,right))) |
duke@435 | 1590 | // / \ |
duke@435 | 1591 | // v v |
duke@435 | 1592 | // * new-proj proj |
duke@435 | 1593 | // (returned) |
duke@435 | 1594 | // |
duke@435 | 1595 | ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) { |
duke@435 | 1596 | IfNode* iff = proj->in(0)->as_If(); |
duke@435 | 1597 | IdealLoopTree *loop = get_loop(proj); |
duke@435 | 1598 | ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); |
duke@435 | 1599 | int ddepth = dom_depth(proj); |
duke@435 | 1600 | |
duke@435 | 1601 | _igvn.hash_delete(iff); |
duke@435 | 1602 | _igvn._worklist.push(iff); |
duke@435 | 1603 | _igvn.hash_delete(proj); |
duke@435 | 1604 | _igvn._worklist.push(proj); |
duke@435 | 1605 | |
duke@435 | 1606 | proj->set_req(0, NULL); // temporary disconnect |
duke@435 | 1607 | ProjNode* proj2 = proj_clone(proj, iff); |
duke@435 | 1608 | register_node(proj2, loop, iff, ddepth); |
duke@435 | 1609 | |
duke@435 | 1610 | Node* cmp = Signed ? (Node*) new (C,3)CmpINode(left, right) : (Node*) new (C,3)CmpUNode(left, right); |
duke@435 | 1611 | register_node(cmp, loop, proj2, ddepth); |
duke@435 | 1612 | |
duke@435 | 1613 | BoolNode* bol = new (C,2)BoolNode(cmp, relop); |
duke@435 | 1614 | register_node(bol, loop, proj2, ddepth); |
duke@435 | 1615 | |
duke@435 | 1616 | IfNode* new_if = new (C,2)IfNode(proj2, bol, iff->_prob, iff->_fcnt); |
duke@435 | 1617 | register_node(new_if, loop, proj2, ddepth); |
duke@435 | 1618 | |
duke@435 | 1619 | proj->set_req(0, new_if); // reattach |
duke@435 | 1620 | set_idom(proj, new_if, ddepth); |
duke@435 | 1621 | |
duke@435 | 1622 | ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj(); |
duke@435 | 1623 | register_node(new_exit, get_loop(other_proj), new_if, ddepth); |
duke@435 | 1624 | |
duke@435 | 1625 | return new_exit; |
duke@435 | 1626 | } |
duke@435 | 1627 | |
duke@435 | 1628 | //------------------------------ insert_region_before_proj ------------------------------------- |
duke@435 | 1629 | // Insert a region before an if projection (* - new node) |
duke@435 | 1630 | // |
duke@435 | 1631 | // before |
duke@435 | 1632 | // if(test) |
duke@435 | 1633 | // / | |
duke@435 | 1634 | // v | |
duke@435 | 1635 | // proj v |
duke@435 | 1636 | // other-proj |
duke@435 | 1637 | // |
duke@435 | 1638 | // after |
duke@435 | 1639 | // if(test) |
duke@435 | 1640 | // / | |
duke@435 | 1641 | // v | |
duke@435 | 1642 | // * proj-clone v |
duke@435 | 1643 | // | other-proj |
duke@435 | 1644 | // v |
duke@435 | 1645 | // * new-region |
duke@435 | 1646 | // | |
duke@435 | 1647 | // v |
duke@435 | 1648 | // * dum_if |
duke@435 | 1649 | // / \ |
duke@435 | 1650 | // v \ |
duke@435 | 1651 | // * dum-proj v |
duke@435 | 1652 | // proj |
duke@435 | 1653 | // |
duke@435 | 1654 | RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) { |
duke@435 | 1655 | IfNode* iff = proj->in(0)->as_If(); |
duke@435 | 1656 | IdealLoopTree *loop = get_loop(proj); |
duke@435 | 1657 | ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); |
duke@435 | 1658 | int ddepth = dom_depth(proj); |
duke@435 | 1659 | |
duke@435 | 1660 | _igvn.hash_delete(iff); |
duke@435 | 1661 | _igvn._worklist.push(iff); |
duke@435 | 1662 | _igvn.hash_delete(proj); |
duke@435 | 1663 | _igvn._worklist.push(proj); |
duke@435 | 1664 | |
duke@435 | 1665 | proj->set_req(0, NULL); // temporary disconnect |
duke@435 | 1666 | ProjNode* proj2 = proj_clone(proj, iff); |
duke@435 | 1667 | register_node(proj2, loop, iff, ddepth); |
duke@435 | 1668 | |
duke@435 | 1669 | RegionNode* reg = new (C,2)RegionNode(2); |
duke@435 | 1670 | reg->set_req(1, proj2); |
duke@435 | 1671 | register_node(reg, loop, iff, ddepth); |
duke@435 | 1672 | |
duke@435 | 1673 | IfNode* dum_if = new (C,2)IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt); |
duke@435 | 1674 | register_node(dum_if, loop, reg, ddepth); |
duke@435 | 1675 | |
duke@435 | 1676 | proj->set_req(0, dum_if); // reattach |
duke@435 | 1677 | set_idom(proj, dum_if, ddepth); |
duke@435 | 1678 | |
duke@435 | 1679 | ProjNode* dum_proj = proj_clone(other_proj, dum_if); |
duke@435 | 1680 | register_node(dum_proj, loop, dum_if, ddepth); |
duke@435 | 1681 | |
duke@435 | 1682 | return reg; |
duke@435 | 1683 | } |
duke@435 | 1684 | |
duke@435 | 1685 | //------------------------------ insert_cmpi_loop_exit ------------------------------------- |
duke@435 | 1686 | // Clone a signed compare loop exit from an unsigned compare and |
duke@435 | 1687 | // insert it before the unsigned cmp on the stay-in-loop path. |
duke@435 | 1688 | // All new nodes inserted in the dominator tree between the original |
duke@435 | 1689 | // if and it's projections. The original if test is replaced with |
duke@435 | 1690 | // a constant to force the stay-in-loop path. |
duke@435 | 1691 | // |
duke@435 | 1692 | // This is done to make sure that the original if and it's projections |
duke@435 | 1693 | // still dominate the same set of control nodes, that the ctrl() relation |
duke@435 | 1694 | // from data nodes to them is preserved, and that their loop nesting is |
duke@435 | 1695 | // preserved. |
duke@435 | 1696 | // |
duke@435 | 1697 | // before |
duke@435 | 1698 | // if(i <u limit) unsigned compare loop exit |
duke@435 | 1699 | // / | |
duke@435 | 1700 | // v v |
duke@435 | 1701 | // exit-proj stay-in-loop-proj |
duke@435 | 1702 | // |
duke@435 | 1703 | // after |
duke@435 | 1704 | // if(stay-in-loop-const) original if |
duke@435 | 1705 | // / | |
duke@435 | 1706 | // / v |
duke@435 | 1707 | // / if(i < limit) new signed test |
duke@435 | 1708 | // / / | |
duke@435 | 1709 | // / / v |
duke@435 | 1710 | // / / if(i <u limit) new cloned unsigned test |
duke@435 | 1711 | // / / / | |
duke@435 | 1712 | // v v v | |
duke@435 | 1713 | // region | |
duke@435 | 1714 | // | | |
duke@435 | 1715 | // dum-if | |
duke@435 | 1716 | // / | | |
duke@435 | 1717 | // ether | | |
duke@435 | 1718 | // v v |
duke@435 | 1719 | // exit-proj stay-in-loop-proj |
duke@435 | 1720 | // |
duke@435 | 1721 | IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) { |
duke@435 | 1722 | const bool Signed = true; |
duke@435 | 1723 | const bool Unsigned = false; |
duke@435 | 1724 | |
duke@435 | 1725 | BoolNode* bol = if_cmpu->in(1)->as_Bool(); |
duke@435 | 1726 | if (bol->_test._test != BoolTest::lt) return NULL; |
duke@435 | 1727 | CmpNode* cmpu = bol->in(1)->as_Cmp(); |
duke@435 | 1728 | if (cmpu->Opcode() != Op_CmpU) return NULL; |
duke@435 | 1729 | int stride = stride_of_possible_iv(if_cmpu); |
duke@435 | 1730 | if (stride == 0) return NULL; |
duke@435 | 1731 | |
duke@435 | 1732 | ProjNode* lp_continue = stay_in_loop(if_cmpu, loop)->as_Proj(); |
duke@435 | 1733 | ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj(); |
duke@435 | 1734 | |
duke@435 | 1735 | Node* limit = NULL; |
duke@435 | 1736 | if (stride > 0) { |
duke@435 | 1737 | limit = cmpu->in(2); |
duke@435 | 1738 | } else { |
duke@435 | 1739 | limit = _igvn.makecon(TypeInt::ZERO); |
duke@435 | 1740 | set_ctrl(limit, C->root()); |
duke@435 | 1741 | } |
duke@435 | 1742 | // Create a new region on the exit path |
duke@435 | 1743 | RegionNode* reg = insert_region_before_proj(lp_exit); |
duke@435 | 1744 | |
duke@435 | 1745 | // Clone the if-cmpu-true-false using a signed compare |
duke@435 | 1746 | BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge; |
duke@435 | 1747 | ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue); |
duke@435 | 1748 | reg->add_req(cmpi_exit); |
duke@435 | 1749 | |
duke@435 | 1750 | // Clone the if-cmpu-true-false |
duke@435 | 1751 | BoolTest::mask rel_u = bol->_test._test; |
duke@435 | 1752 | ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue); |
duke@435 | 1753 | reg->add_req(cmpu_exit); |
duke@435 | 1754 | |
duke@435 | 1755 | // Force original if to stay in loop. |
duke@435 | 1756 | short_circuit_if(if_cmpu, lp_continue); |
duke@435 | 1757 | |
duke@435 | 1758 | return cmpi_exit->in(0)->as_If(); |
duke@435 | 1759 | } |
duke@435 | 1760 | |
duke@435 | 1761 | //------------------------------ remove_cmpi_loop_exit ------------------------------------- |
duke@435 | 1762 | // Remove a previously inserted signed compare loop exit. |
duke@435 | 1763 | void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) { |
duke@435 | 1764 | Node* lp_proj = stay_in_loop(if_cmp, loop); |
duke@435 | 1765 | assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI && |
duke@435 | 1766 | stay_in_loop(lp_proj, loop)->is_If() && |
duke@435 | 1767 | stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu"); |
duke@435 | 1768 | Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO); |
duke@435 | 1769 | set_ctrl(con, C->root()); |
duke@435 | 1770 | if_cmp->set_req(1, con); |
duke@435 | 1771 | } |
duke@435 | 1772 | |
duke@435 | 1773 | //------------------------------ scheduled_nodelist ------------------------------------- |
duke@435 | 1774 | // Create a post order schedule of nodes that are in the |
duke@435 | 1775 | // "member" set. The list is returned in "sched". |
duke@435 | 1776 | // The first node in "sched" is the loop head, followed by |
duke@435 | 1777 | // nodes which have no inputs in the "member" set, and then |
duke@435 | 1778 | // followed by the nodes that have an immediate input dependence |
duke@435 | 1779 | // on a node in "sched". |
duke@435 | 1780 | void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) { |
duke@435 | 1781 | |
duke@435 | 1782 | assert(member.test(loop->_head->_idx), "loop head must be in member set"); |
duke@435 | 1783 | Arena *a = Thread::current()->resource_area(); |
duke@435 | 1784 | VectorSet visited(a); |
duke@435 | 1785 | Node_Stack nstack(a, loop->_body.size()); |
duke@435 | 1786 | |
duke@435 | 1787 | Node* n = loop->_head; // top of stack is cached in "n" |
duke@435 | 1788 | uint idx = 0; |
duke@435 | 1789 | visited.set(n->_idx); |
duke@435 | 1790 | |
duke@435 | 1791 | // Initially push all with no inputs from within member set |
duke@435 | 1792 | for(uint i = 0; i < loop->_body.size(); i++ ) { |
duke@435 | 1793 | Node *elt = loop->_body.at(i); |
duke@435 | 1794 | if (member.test(elt->_idx)) { |
duke@435 | 1795 | bool found = false; |
duke@435 | 1796 | for (uint j = 0; j < elt->req(); j++) { |
duke@435 | 1797 | Node* def = elt->in(j); |
duke@435 | 1798 | if (def && member.test(def->_idx) && def != elt) { |
duke@435 | 1799 | found = true; |
duke@435 | 1800 | break; |
duke@435 | 1801 | } |
duke@435 | 1802 | } |
duke@435 | 1803 | if (!found && elt != loop->_head) { |
duke@435 | 1804 | nstack.push(n, idx); |
duke@435 | 1805 | n = elt; |
duke@435 | 1806 | assert(!visited.test(n->_idx), "not seen yet"); |
duke@435 | 1807 | visited.set(n->_idx); |
duke@435 | 1808 | } |
duke@435 | 1809 | } |
duke@435 | 1810 | } |
duke@435 | 1811 | |
duke@435 | 1812 | // traverse out's that are in the member set |
duke@435 | 1813 | while (true) { |
duke@435 | 1814 | if (idx < n->outcnt()) { |
duke@435 | 1815 | Node* use = n->raw_out(idx); |
duke@435 | 1816 | idx++; |
duke@435 | 1817 | if (!visited.test_set(use->_idx)) { |
duke@435 | 1818 | if (member.test(use->_idx)) { |
duke@435 | 1819 | nstack.push(n, idx); |
duke@435 | 1820 | n = use; |
duke@435 | 1821 | idx = 0; |
duke@435 | 1822 | } |
duke@435 | 1823 | } |
duke@435 | 1824 | } else { |
duke@435 | 1825 | // All outputs processed |
duke@435 | 1826 | sched.push(n); |
duke@435 | 1827 | if (nstack.is_empty()) break; |
duke@435 | 1828 | n = nstack.node(); |
duke@435 | 1829 | idx = nstack.index(); |
duke@435 | 1830 | nstack.pop(); |
duke@435 | 1831 | } |
duke@435 | 1832 | } |
duke@435 | 1833 | } |
duke@435 | 1834 | |
duke@435 | 1835 | |
duke@435 | 1836 | //------------------------------ has_use_in_set ------------------------------------- |
duke@435 | 1837 | // Has a use in the vector set |
duke@435 | 1838 | bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) { |
duke@435 | 1839 | for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { |
duke@435 | 1840 | Node* use = n->fast_out(j); |
duke@435 | 1841 | if (vset.test(use->_idx)) { |
duke@435 | 1842 | return true; |
duke@435 | 1843 | } |
duke@435 | 1844 | } |
duke@435 | 1845 | return false; |
duke@435 | 1846 | } |
duke@435 | 1847 | |
duke@435 | 1848 | |
duke@435 | 1849 | //------------------------------ has_use_internal_to_set ------------------------------------- |
duke@435 | 1850 | // Has use internal to the vector set (ie. not in a phi at the loop head) |
duke@435 | 1851 | bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) { |
duke@435 | 1852 | Node* head = loop->_head; |
duke@435 | 1853 | for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { |
duke@435 | 1854 | Node* use = n->fast_out(j); |
duke@435 | 1855 | if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) { |
duke@435 | 1856 | return true; |
duke@435 | 1857 | } |
duke@435 | 1858 | } |
duke@435 | 1859 | return false; |
duke@435 | 1860 | } |
duke@435 | 1861 | |
duke@435 | 1862 | |
duke@435 | 1863 | //------------------------------ clone_for_use_outside_loop ------------------------------------- |
duke@435 | 1864 | // clone "n" for uses that are outside of loop |
duke@435 | 1865 | void PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) { |
duke@435 | 1866 | |
duke@435 | 1867 | assert(worklist.size() == 0, "should be empty"); |
duke@435 | 1868 | for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { |
duke@435 | 1869 | Node* use = n->fast_out(j); |
duke@435 | 1870 | if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) { |
duke@435 | 1871 | worklist.push(use); |
duke@435 | 1872 | } |
duke@435 | 1873 | } |
duke@435 | 1874 | while( worklist.size() ) { |
duke@435 | 1875 | Node *use = worklist.pop(); |
duke@435 | 1876 | if (!has_node(use) || use->in(0) == C->top()) continue; |
duke@435 | 1877 | uint j; |
duke@435 | 1878 | for (j = 0; j < use->req(); j++) { |
duke@435 | 1879 | if (use->in(j) == n) break; |
duke@435 | 1880 | } |
duke@435 | 1881 | assert(j < use->req(), "must be there"); |
duke@435 | 1882 | |
duke@435 | 1883 | // clone "n" and insert it between the inputs of "n" and the use outside the loop |
duke@435 | 1884 | Node* n_clone = n->clone(); |
duke@435 | 1885 | _igvn.hash_delete(use); |
duke@435 | 1886 | use->set_req(j, n_clone); |
duke@435 | 1887 | _igvn._worklist.push(use); |
duke@435 | 1888 | if (!use->is_Phi()) { |
duke@435 | 1889 | Node* use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0); |
duke@435 | 1890 | set_ctrl(n_clone, use_c); |
duke@435 | 1891 | assert(!loop->is_member(get_loop(use_c)), "should be outside loop"); |
duke@435 | 1892 | get_loop(use_c)->_body.push(n_clone); |
duke@435 | 1893 | } else { |
duke@435 | 1894 | // Use in a phi is considered a use in the associated predecessor block |
duke@435 | 1895 | Node *prevbb = use->in(0)->in(j); |
duke@435 | 1896 | set_ctrl(n_clone, prevbb); |
duke@435 | 1897 | assert(!loop->is_member(get_loop(prevbb)), "should be outside loop"); |
duke@435 | 1898 | get_loop(prevbb)->_body.push(n_clone); |
duke@435 | 1899 | } |
duke@435 | 1900 | _igvn.register_new_node_with_optimizer(n_clone); |
duke@435 | 1901 | #if !defined(PRODUCT) |
duke@435 | 1902 | if (TracePartialPeeling) { |
duke@435 | 1903 | tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx); |
duke@435 | 1904 | } |
duke@435 | 1905 | #endif |
duke@435 | 1906 | } |
duke@435 | 1907 | } |
duke@435 | 1908 | |
duke@435 | 1909 | |
duke@435 | 1910 | //------------------------------ clone_for_special_use_inside_loop ------------------------------------- |
duke@435 | 1911 | // clone "n" for special uses that are in the not_peeled region. |
duke@435 | 1912 | // If these def-uses occur in separate blocks, the code generator |
duke@435 | 1913 | // marks the method as not compilable. For example, if a "BoolNode" |
duke@435 | 1914 | // is in a different basic block than the "IfNode" that uses it, then |
duke@435 | 1915 | // the compilation is aborted in the code generator. |
duke@435 | 1916 | void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, |
duke@435 | 1917 | VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) { |
duke@435 | 1918 | if (n->is_Phi() || n->is_Load()) { |
duke@435 | 1919 | return; |
duke@435 | 1920 | } |
duke@435 | 1921 | assert(worklist.size() == 0, "should be empty"); |
duke@435 | 1922 | for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { |
duke@435 | 1923 | Node* use = n->fast_out(j); |
duke@435 | 1924 | if ( not_peel.test(use->_idx) && |
duke@435 | 1925 | (use->is_If() || use->is_CMove() || use->is_Bool()) && |
duke@435 | 1926 | use->in(1) == n) { |
duke@435 | 1927 | worklist.push(use); |
duke@435 | 1928 | } |
duke@435 | 1929 | } |
duke@435 | 1930 | if (worklist.size() > 0) { |
duke@435 | 1931 | // clone "n" and insert it between inputs of "n" and the use |
duke@435 | 1932 | Node* n_clone = n->clone(); |
duke@435 | 1933 | loop->_body.push(n_clone); |
duke@435 | 1934 | _igvn.register_new_node_with_optimizer(n_clone); |
duke@435 | 1935 | set_ctrl(n_clone, get_ctrl(n)); |
duke@435 | 1936 | sink_list.push(n_clone); |
duke@435 | 1937 | not_peel <<= n_clone->_idx; // add n_clone to not_peel set. |
duke@435 | 1938 | #if !defined(PRODUCT) |
duke@435 | 1939 | if (TracePartialPeeling) { |
duke@435 | 1940 | tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx); |
duke@435 | 1941 | } |
duke@435 | 1942 | #endif |
duke@435 | 1943 | while( worklist.size() ) { |
duke@435 | 1944 | Node *use = worklist.pop(); |
duke@435 | 1945 | _igvn.hash_delete(use); |
duke@435 | 1946 | _igvn._worklist.push(use); |
duke@435 | 1947 | for (uint j = 1; j < use->req(); j++) { |
duke@435 | 1948 | if (use->in(j) == n) { |
duke@435 | 1949 | use->set_req(j, n_clone); |
duke@435 | 1950 | } |
duke@435 | 1951 | } |
duke@435 | 1952 | } |
duke@435 | 1953 | } |
duke@435 | 1954 | } |
duke@435 | 1955 | |
duke@435 | 1956 | |
duke@435 | 1957 | //------------------------------ insert_phi_for_loop ------------------------------------- |
duke@435 | 1958 | // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist |
duke@435 | 1959 | void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) { |
duke@435 | 1960 | Node *phi = PhiNode::make(lp, back_edge_val); |
duke@435 | 1961 | phi->set_req(LoopNode::EntryControl, lp_entry_val); |
duke@435 | 1962 | // Use existing phi if it already exists |
duke@435 | 1963 | Node *hit = _igvn.hash_find_insert(phi); |
duke@435 | 1964 | if( hit == NULL ) { |
duke@435 | 1965 | _igvn.register_new_node_with_optimizer(phi); |
duke@435 | 1966 | set_ctrl(phi, lp); |
duke@435 | 1967 | } else { |
duke@435 | 1968 | // Remove the new phi from the graph and use the hit |
duke@435 | 1969 | _igvn.remove_dead_node(phi); |
duke@435 | 1970 | phi = hit; |
duke@435 | 1971 | } |
duke@435 | 1972 | _igvn.hash_delete(use); |
duke@435 | 1973 | _igvn._worklist.push(use); |
duke@435 | 1974 | use->set_req(idx, phi); |
duke@435 | 1975 | } |
duke@435 | 1976 | |
duke@435 | 1977 | #ifdef ASSERT |
duke@435 | 1978 | //------------------------------ is_valid_loop_partition ------------------------------------- |
duke@435 | 1979 | // Validate the loop partition sets: peel and not_peel |
duke@435 | 1980 | bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, |
duke@435 | 1981 | VectorSet& not_peel ) { |
duke@435 | 1982 | uint i; |
duke@435 | 1983 | // Check that peel_list entries are in the peel set |
duke@435 | 1984 | for (i = 0; i < peel_list.size(); i++) { |
duke@435 | 1985 | if (!peel.test(peel_list.at(i)->_idx)) { |
duke@435 | 1986 | return false; |
duke@435 | 1987 | } |
duke@435 | 1988 | } |
duke@435 | 1989 | // Check at loop members are in one of peel set or not_peel set |
duke@435 | 1990 | for (i = 0; i < loop->_body.size(); i++ ) { |
duke@435 | 1991 | Node *def = loop->_body.at(i); |
duke@435 | 1992 | uint di = def->_idx; |
duke@435 | 1993 | // Check that peel set elements are in peel_list |
duke@435 | 1994 | if (peel.test(di)) { |
duke@435 | 1995 | if (not_peel.test(di)) { |
duke@435 | 1996 | return false; |
duke@435 | 1997 | } |
duke@435 | 1998 | // Must be in peel_list also |
duke@435 | 1999 | bool found = false; |
duke@435 | 2000 | for (uint j = 0; j < peel_list.size(); j++) { |
duke@435 | 2001 | if (peel_list.at(j)->_idx == di) { |
duke@435 | 2002 | found = true; |
duke@435 | 2003 | break; |
duke@435 | 2004 | } |
duke@435 | 2005 | } |
duke@435 | 2006 | if (!found) { |
duke@435 | 2007 | return false; |
duke@435 | 2008 | } |
duke@435 | 2009 | } else if (not_peel.test(di)) { |
duke@435 | 2010 | if (peel.test(di)) { |
duke@435 | 2011 | return false; |
duke@435 | 2012 | } |
duke@435 | 2013 | } else { |
duke@435 | 2014 | return false; |
duke@435 | 2015 | } |
duke@435 | 2016 | } |
duke@435 | 2017 | return true; |
duke@435 | 2018 | } |
duke@435 | 2019 | |
duke@435 | 2020 | //------------------------------ is_valid_clone_loop_exit_use ------------------------------------- |
duke@435 | 2021 | // Ensure a use outside of loop is of the right form |
duke@435 | 2022 | bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) { |
duke@435 | 2023 | Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; |
duke@435 | 2024 | return (use->is_Phi() && |
duke@435 | 2025 | use_c->is_Region() && use_c->req() == 3 && |
duke@435 | 2026 | (use_c->in(exit_idx)->Opcode() == Op_IfTrue || |
duke@435 | 2027 | use_c->in(exit_idx)->Opcode() == Op_IfFalse || |
duke@435 | 2028 | use_c->in(exit_idx)->Opcode() == Op_JumpProj) && |
duke@435 | 2029 | loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) ); |
duke@435 | 2030 | } |
duke@435 | 2031 | |
duke@435 | 2032 | //------------------------------ is_valid_clone_loop_form ------------------------------------- |
duke@435 | 2033 | // Ensure that all uses outside of loop are of the right form |
duke@435 | 2034 | bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, |
duke@435 | 2035 | uint orig_exit_idx, uint clone_exit_idx) { |
duke@435 | 2036 | uint len = peel_list.size(); |
duke@435 | 2037 | for (uint i = 0; i < len; i++) { |
duke@435 | 2038 | Node *def = peel_list.at(i); |
duke@435 | 2039 | |
duke@435 | 2040 | for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { |
duke@435 | 2041 | Node *use = def->fast_out(j); |
duke@435 | 2042 | Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; |
duke@435 | 2043 | if (!loop->is_member(get_loop(use_c))) { |
duke@435 | 2044 | // use is not in the loop, check for correct structure |
duke@435 | 2045 | if (use->in(0) == def) { |
duke@435 | 2046 | // Okay |
duke@435 | 2047 | } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) { |
duke@435 | 2048 | return false; |
duke@435 | 2049 | } |
duke@435 | 2050 | } |
duke@435 | 2051 | } |
duke@435 | 2052 | } |
duke@435 | 2053 | return true; |
duke@435 | 2054 | } |
duke@435 | 2055 | #endif |
duke@435 | 2056 | |
duke@435 | 2057 | //------------------------------ partial_peel ------------------------------------- |
duke@435 | 2058 | // Partially peel (aka loop rotation) the top portion of a loop (called |
duke@435 | 2059 | // the peel section below) by cloning it and placing one copy just before |
duke@435 | 2060 | // the new loop head and the other copy at the bottom of the new loop. |
duke@435 | 2061 | // |
duke@435 | 2062 | // before after where it came from |
duke@435 | 2063 | // |
duke@435 | 2064 | // stmt1 stmt1 |
duke@435 | 2065 | // loop: stmt2 clone |
duke@435 | 2066 | // stmt2 if condA goto exitA clone |
duke@435 | 2067 | // if condA goto exitA new_loop: new |
duke@435 | 2068 | // stmt3 stmt3 clone |
duke@435 | 2069 | // if !condB goto loop if condB goto exitB clone |
duke@435 | 2070 | // exitB: stmt2 orig |
duke@435 | 2071 | // stmt4 if !condA goto new_loop orig |
duke@435 | 2072 | // exitA: goto exitA |
duke@435 | 2073 | // exitB: |
duke@435 | 2074 | // stmt4 |
duke@435 | 2075 | // exitA: |
duke@435 | 2076 | // |
duke@435 | 2077 | // Step 1: find the cut point: an exit test on probable |
duke@435 | 2078 | // induction variable. |
duke@435 | 2079 | // Step 2: schedule (with cloning) operations in the peel |
duke@435 | 2080 | // section that can be executed after the cut into |
duke@435 | 2081 | // the section that is not peeled. This may need |
duke@435 | 2082 | // to clone operations into exit blocks. For |
duke@435 | 2083 | // instance, a reference to A[i] in the not-peel |
duke@435 | 2084 | // section and a reference to B[i] in an exit block |
duke@435 | 2085 | // may cause a left-shift of i by 2 to be placed |
duke@435 | 2086 | // in the peel block. This step will clone the left |
duke@435 | 2087 | // shift into the exit block and sink the left shift |
duke@435 | 2088 | // from the peel to the not-peel section. |
duke@435 | 2089 | // Step 3: clone the loop, retarget the control, and insert |
duke@435 | 2090 | // phis for values that are live across the new loop |
duke@435 | 2091 | // head. This is very dependent on the graph structure |
duke@435 | 2092 | // from clone_loop. It creates region nodes for |
duke@435 | 2093 | // exit control and associated phi nodes for values |
duke@435 | 2094 | // flow out of the loop through that exit. The region |
duke@435 | 2095 | // node is dominated by the clone's control projection. |
duke@435 | 2096 | // So the clone's peel section is placed before the |
duke@435 | 2097 | // new loop head, and the clone's not-peel section is |
duke@435 | 2098 | // forms the top part of the new loop. The original |
duke@435 | 2099 | // peel section forms the tail of the new loop. |
duke@435 | 2100 | // Step 4: update the dominator tree and recompute the |
duke@435 | 2101 | // dominator depth. |
duke@435 | 2102 | // |
duke@435 | 2103 | // orig |
duke@435 | 2104 | // |
duke@435 | 2105 | // stmt1 |
duke@435 | 2106 | // | |
duke@435 | 2107 | // v |
duke@435 | 2108 | // loop<----+ |
duke@435 | 2109 | // | | |
duke@435 | 2110 | // stmt2 | |
duke@435 | 2111 | // | | |
duke@435 | 2112 | // v | |
duke@435 | 2113 | // ifA | |
duke@435 | 2114 | // / | | |
duke@435 | 2115 | // v v | |
duke@435 | 2116 | // false true ^ <-- last_peel |
duke@435 | 2117 | // / | | |
duke@435 | 2118 | // / ===|==cut | |
duke@435 | 2119 | // / stmt3 | <-- first_not_peel |
duke@435 | 2120 | // / | | |
duke@435 | 2121 | // | v | |
duke@435 | 2122 | // v ifB | |
duke@435 | 2123 | // exitA: / \ | |
duke@435 | 2124 | // / \ | |
duke@435 | 2125 | // v v | |
duke@435 | 2126 | // false true | |
duke@435 | 2127 | // / \ | |
duke@435 | 2128 | // / ----+ |
duke@435 | 2129 | // | |
duke@435 | 2130 | // v |
duke@435 | 2131 | // exitB: |
duke@435 | 2132 | // stmt4 |
duke@435 | 2133 | // |
duke@435 | 2134 | // |
duke@435 | 2135 | // after clone loop |
duke@435 | 2136 | // |
duke@435 | 2137 | // stmt1 |
duke@435 | 2138 | // / \ |
duke@435 | 2139 | // clone / \ orig |
duke@435 | 2140 | // / \ |
duke@435 | 2141 | // / \ |
duke@435 | 2142 | // v v |
duke@435 | 2143 | // +---->loop loop<----+ |
duke@435 | 2144 | // | | | | |
duke@435 | 2145 | // | stmt2 stmt2 | |
duke@435 | 2146 | // | | | | |
duke@435 | 2147 | // | v v | |
duke@435 | 2148 | // | ifA ifA | |
duke@435 | 2149 | // | | \ / | | |
duke@435 | 2150 | // | v v v v | |
duke@435 | 2151 | // ^ true false false true ^ <-- last_peel |
duke@435 | 2152 | // | | ^ \ / | | |
duke@435 | 2153 | // | cut==|== \ \ / ===|==cut | |
duke@435 | 2154 | // | stmt3 \ \ / stmt3 | <-- first_not_peel |
duke@435 | 2155 | // | | dom | | | | |
duke@435 | 2156 | // | v \ 1v v2 v | |
duke@435 | 2157 | // | ifB regionA ifB | |
duke@435 | 2158 | // | / \ | / \ | |
duke@435 | 2159 | // | / \ v / \ | |
duke@435 | 2160 | // | v v exitA: v v | |
duke@435 | 2161 | // | true false false true | |
duke@435 | 2162 | // | / ^ \ / \ | |
duke@435 | 2163 | // +---- \ \ / ----+ |
duke@435 | 2164 | // dom \ / |
duke@435 | 2165 | // \ 1v v2 |
duke@435 | 2166 | // regionB |
duke@435 | 2167 | // | |
duke@435 | 2168 | // v |
duke@435 | 2169 | // exitB: |
duke@435 | 2170 | // stmt4 |
duke@435 | 2171 | // |
duke@435 | 2172 | // |
duke@435 | 2173 | // after partial peel |
duke@435 | 2174 | // |
duke@435 | 2175 | // stmt1 |
duke@435 | 2176 | // / |
duke@435 | 2177 | // clone / orig |
duke@435 | 2178 | // / TOP |
duke@435 | 2179 | // / \ |
duke@435 | 2180 | // v v |
duke@435 | 2181 | // TOP->region region----+ |
duke@435 | 2182 | // | | | |
duke@435 | 2183 | // stmt2 stmt2 | |
duke@435 | 2184 | // | | | |
duke@435 | 2185 | // v v | |
duke@435 | 2186 | // ifA ifA | |
duke@435 | 2187 | // | \ / | | |
duke@435 | 2188 | // v v v v | |
duke@435 | 2189 | // true false false true | <-- last_peel |
duke@435 | 2190 | // | ^ \ / +------|---+ |
duke@435 | 2191 | // +->newloop \ \ / === ==cut | | |
duke@435 | 2192 | // | stmt3 \ \ / TOP | | |
duke@435 | 2193 | // | | dom | | stmt3 | | <-- first_not_peel |
duke@435 | 2194 | // | v \ 1v v2 v | | |
duke@435 | 2195 | // | ifB regionA ifB ^ v |
duke@435 | 2196 | // | / \ | / \ | | |
duke@435 | 2197 | // | / \ v / \ | | |
duke@435 | 2198 | // | v v exitA: v v | | |
duke@435 | 2199 | // | true false false true | | |
duke@435 | 2200 | // | / ^ \ / \ | | |
duke@435 | 2201 | // | | \ \ / v | | |
duke@435 | 2202 | // | | dom \ / TOP | | |
duke@435 | 2203 | // | | \ 1v v2 | | |
duke@435 | 2204 | // ^ v regionB | | |
duke@435 | 2205 | // | | | | | |
duke@435 | 2206 | // | | v ^ v |
duke@435 | 2207 | // | | exitB: | | |
duke@435 | 2208 | // | | stmt4 | | |
duke@435 | 2209 | // | +------------>-----------------+ | |
duke@435 | 2210 | // | | |
duke@435 | 2211 | // +-----------------<---------------------+ |
duke@435 | 2212 | // |
duke@435 | 2213 | // |
duke@435 | 2214 | // final graph |
duke@435 | 2215 | // |
duke@435 | 2216 | // stmt1 |
duke@435 | 2217 | // | |
duke@435 | 2218 | // v |
duke@435 | 2219 | // ........> ifA clone |
duke@435 | 2220 | // : / | |
duke@435 | 2221 | // dom / | |
duke@435 | 2222 | // : v v |
duke@435 | 2223 | // : false true |
duke@435 | 2224 | // : | | |
duke@435 | 2225 | // : | stmt2 clone |
duke@435 | 2226 | // : | | |
duke@435 | 2227 | // : | v |
duke@435 | 2228 | // : | newloop<-----+ |
duke@435 | 2229 | // : | | | |
duke@435 | 2230 | // : | stmt3 clone | |
duke@435 | 2231 | // : | | | |
duke@435 | 2232 | // : | v | |
duke@435 | 2233 | // : | ifB | |
duke@435 | 2234 | // : | / \ | |
duke@435 | 2235 | // : | v v | |
duke@435 | 2236 | // : | false true | |
duke@435 | 2237 | // : | | | | |
duke@435 | 2238 | // : | v stmt2 | |
duke@435 | 2239 | // : | exitB: | | |
duke@435 | 2240 | // : | stmt4 v | |
duke@435 | 2241 | // : | ifA orig | |
duke@435 | 2242 | // : | / \ | |
duke@435 | 2243 | // : | / \ | |
duke@435 | 2244 | // : | v v | |
duke@435 | 2245 | // : | false true | |
duke@435 | 2246 | // : | / \ | |
duke@435 | 2247 | // : v v -----+ |
duke@435 | 2248 | // RegionA |
duke@435 | 2249 | // | |
duke@435 | 2250 | // v |
duke@435 | 2251 | // exitA |
duke@435 | 2252 | // |
duke@435 | 2253 | bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) { |
duke@435 | 2254 | |
duke@435 | 2255 | LoopNode *head = loop->_head->as_Loop(); |
duke@435 | 2256 | |
duke@435 | 2257 | if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) { |
duke@435 | 2258 | return false; |
duke@435 | 2259 | } |
duke@435 | 2260 | |
duke@435 | 2261 | // Check for complex exit control |
duke@435 | 2262 | for(uint ii = 0; ii < loop->_body.size(); ii++ ) { |
duke@435 | 2263 | Node *n = loop->_body.at(ii); |
duke@435 | 2264 | int opc = n->Opcode(); |
duke@435 | 2265 | if (n->is_Call() || |
duke@435 | 2266 | opc == Op_Catch || |
duke@435 | 2267 | opc == Op_CatchProj || |
duke@435 | 2268 | opc == Op_Jump || |
duke@435 | 2269 | opc == Op_JumpProj) { |
duke@435 | 2270 | #if !defined(PRODUCT) |
duke@435 | 2271 | if (TracePartialPeeling) { |
duke@435 | 2272 | tty->print_cr("\nExit control too complex: lp: %d", head->_idx); |
duke@435 | 2273 | } |
duke@435 | 2274 | #endif |
duke@435 | 2275 | return false; |
duke@435 | 2276 | } |
duke@435 | 2277 | } |
duke@435 | 2278 | |
duke@435 | 2279 | int dd = dom_depth(head); |
duke@435 | 2280 | |
duke@435 | 2281 | // Step 1: find cut point |
duke@435 | 2282 | |
duke@435 | 2283 | // Walk up dominators to loop head looking for first loop exit |
duke@435 | 2284 | // which is executed on every path thru loop. |
duke@435 | 2285 | IfNode *peel_if = NULL; |
duke@435 | 2286 | IfNode *peel_if_cmpu = NULL; |
duke@435 | 2287 | |
duke@435 | 2288 | Node *iff = loop->tail(); |
duke@435 | 2289 | while( iff != head ) { |
duke@435 | 2290 | if( iff->is_If() ) { |
duke@435 | 2291 | Node *ctrl = get_ctrl(iff->in(1)); |
duke@435 | 2292 | if (ctrl->is_top()) return false; // Dead test on live IF. |
duke@435 | 2293 | // If loop-varying exit-test, check for induction variable |
duke@435 | 2294 | if( loop->is_member(get_loop(ctrl)) && |
duke@435 | 2295 | loop->is_loop_exit(iff) && |
duke@435 | 2296 | is_possible_iv_test(iff)) { |
duke@435 | 2297 | Node* cmp = iff->in(1)->in(1); |
duke@435 | 2298 | if (cmp->Opcode() == Op_CmpI) { |
duke@435 | 2299 | peel_if = iff->as_If(); |
duke@435 | 2300 | } else { |
duke@435 | 2301 | assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU"); |
duke@435 | 2302 | peel_if_cmpu = iff->as_If(); |
duke@435 | 2303 | } |
duke@435 | 2304 | } |
duke@435 | 2305 | } |
duke@435 | 2306 | iff = idom(iff); |
duke@435 | 2307 | } |
duke@435 | 2308 | // Prefer signed compare over unsigned compare. |
duke@435 | 2309 | IfNode* new_peel_if = NULL; |
duke@435 | 2310 | if (peel_if == NULL) { |
duke@435 | 2311 | if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) { |
duke@435 | 2312 | return false; // No peel point found |
duke@435 | 2313 | } |
duke@435 | 2314 | new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop); |
duke@435 | 2315 | if (new_peel_if == NULL) { |
duke@435 | 2316 | return false; // No peel point found |
duke@435 | 2317 | } |
duke@435 | 2318 | peel_if = new_peel_if; |
duke@435 | 2319 | } |
duke@435 | 2320 | Node* last_peel = stay_in_loop(peel_if, loop); |
duke@435 | 2321 | Node* first_not_peeled = stay_in_loop(last_peel, loop); |
duke@435 | 2322 | if (first_not_peeled == NULL || first_not_peeled == head) { |
duke@435 | 2323 | return false; |
duke@435 | 2324 | } |
duke@435 | 2325 | |
duke@435 | 2326 | #if !defined(PRODUCT) |
duke@435 | 2327 | if (TracePartialPeeling) { |
duke@435 | 2328 | tty->print_cr("before partial peel one iteration"); |
duke@435 | 2329 | Node_List wl; |
duke@435 | 2330 | Node* t = head->in(2); |
duke@435 | 2331 | while (true) { |
duke@435 | 2332 | wl.push(t); |
duke@435 | 2333 | if (t == head) break; |
duke@435 | 2334 | t = idom(t); |
duke@435 | 2335 | } |
duke@435 | 2336 | while (wl.size() > 0) { |
duke@435 | 2337 | Node* tt = wl.pop(); |
duke@435 | 2338 | tt->dump(); |
duke@435 | 2339 | if (tt == last_peel) tty->print_cr("-- cut --"); |
duke@435 | 2340 | } |
duke@435 | 2341 | } |
duke@435 | 2342 | #endif |
duke@435 | 2343 | ResourceArea *area = Thread::current()->resource_area(); |
duke@435 | 2344 | VectorSet peel(area); |
duke@435 | 2345 | VectorSet not_peel(area); |
duke@435 | 2346 | Node_List peel_list(area); |
duke@435 | 2347 | Node_List worklist(area); |
duke@435 | 2348 | Node_List sink_list(area); |
duke@435 | 2349 | |
duke@435 | 2350 | // Set of cfg nodes to peel are those that are executable from |
duke@435 | 2351 | // the head through last_peel. |
duke@435 | 2352 | assert(worklist.size() == 0, "should be empty"); |
duke@435 | 2353 | worklist.push(head); |
duke@435 | 2354 | peel.set(head->_idx); |
duke@435 | 2355 | while (worklist.size() > 0) { |
duke@435 | 2356 | Node *n = worklist.pop(); |
duke@435 | 2357 | if (n != last_peel) { |
duke@435 | 2358 | for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { |
duke@435 | 2359 | Node* use = n->fast_out(j); |
duke@435 | 2360 | if (use->is_CFG() && |
duke@435 | 2361 | loop->is_member(get_loop(use)) && |
duke@435 | 2362 | !peel.test_set(use->_idx)) { |
duke@435 | 2363 | worklist.push(use); |
duke@435 | 2364 | } |
duke@435 | 2365 | } |
duke@435 | 2366 | } |
duke@435 | 2367 | } |
duke@435 | 2368 | |
duke@435 | 2369 | // Set of non-cfg nodes to peel are those that are control |
duke@435 | 2370 | // dependent on the cfg nodes. |
duke@435 | 2371 | uint i; |
duke@435 | 2372 | for(i = 0; i < loop->_body.size(); i++ ) { |
duke@435 | 2373 | Node *n = loop->_body.at(i); |
duke@435 | 2374 | Node *n_c = has_ctrl(n) ? get_ctrl(n) : n; |
duke@435 | 2375 | if (peel.test(n_c->_idx)) { |
duke@435 | 2376 | peel.set(n->_idx); |
duke@435 | 2377 | } else { |
duke@435 | 2378 | not_peel.set(n->_idx); |
duke@435 | 2379 | } |
duke@435 | 2380 | } |
duke@435 | 2381 | |
duke@435 | 2382 | // Step 2: move operations from the peeled section down into the |
duke@435 | 2383 | // not-peeled section |
duke@435 | 2384 | |
duke@435 | 2385 | // Get a post order schedule of nodes in the peel region |
duke@435 | 2386 | // Result in right-most operand. |
duke@435 | 2387 | scheduled_nodelist(loop, peel, peel_list ); |
duke@435 | 2388 | |
duke@435 | 2389 | assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); |
duke@435 | 2390 | |
duke@435 | 2391 | // For future check for too many new phis |
duke@435 | 2392 | uint old_phi_cnt = 0; |
duke@435 | 2393 | for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { |
duke@435 | 2394 | Node* use = head->fast_out(j); |
duke@435 | 2395 | if (use->is_Phi()) old_phi_cnt++; |
duke@435 | 2396 | } |
duke@435 | 2397 | |
duke@435 | 2398 | #if !defined(PRODUCT) |
duke@435 | 2399 | if (TracePartialPeeling) { |
duke@435 | 2400 | tty->print_cr("\npeeled list"); |
duke@435 | 2401 | } |
duke@435 | 2402 | #endif |
duke@435 | 2403 | |
duke@435 | 2404 | // Evacuate nodes in peel region into the not_peeled region if possible |
duke@435 | 2405 | uint new_phi_cnt = 0; |
duke@435 | 2406 | for (i = 0; i < peel_list.size();) { |
duke@435 | 2407 | Node* n = peel_list.at(i); |
duke@435 | 2408 | #if !defined(PRODUCT) |
duke@435 | 2409 | if (TracePartialPeeling) n->dump(); |
duke@435 | 2410 | #endif |
duke@435 | 2411 | bool incr = true; |
duke@435 | 2412 | if ( !n->is_CFG() ) { |
duke@435 | 2413 | |
duke@435 | 2414 | if ( has_use_in_set(n, not_peel) ) { |
duke@435 | 2415 | |
duke@435 | 2416 | // If not used internal to the peeled region, |
duke@435 | 2417 | // move "n" from peeled to not_peeled region. |
duke@435 | 2418 | |
duke@435 | 2419 | if ( !has_use_internal_to_set(n, peel, loop) ) { |
duke@435 | 2420 | |
duke@435 | 2421 | // if not pinned and not a load (which maybe anti-dependent on a store) |
duke@435 | 2422 | // and not a CMove (Matcher expects only bool->cmove). |
duke@435 | 2423 | if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) { |
duke@435 | 2424 | clone_for_use_outside_loop( loop, n, worklist ); |
duke@435 | 2425 | |
duke@435 | 2426 | sink_list.push(n); |
duke@435 | 2427 | peel >>= n->_idx; // delete n from peel set. |
duke@435 | 2428 | not_peel <<= n->_idx; // add n to not_peel set. |
duke@435 | 2429 | peel_list.remove(i); |
duke@435 | 2430 | incr = false; |
duke@435 | 2431 | #if !defined(PRODUCT) |
duke@435 | 2432 | if (TracePartialPeeling) { |
duke@435 | 2433 | tty->print_cr("sink to not_peeled region: %d newbb: %d", |
duke@435 | 2434 | n->_idx, get_ctrl(n)->_idx); |
duke@435 | 2435 | } |
duke@435 | 2436 | #endif |
duke@435 | 2437 | } |
duke@435 | 2438 | } else { |
duke@435 | 2439 | // Otherwise check for special def-use cases that span |
duke@435 | 2440 | // the peel/not_peel boundary such as bool->if |
duke@435 | 2441 | clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist ); |
duke@435 | 2442 | new_phi_cnt++; |
duke@435 | 2443 | } |
duke@435 | 2444 | } |
duke@435 | 2445 | } |
duke@435 | 2446 | if (incr) i++; |
duke@435 | 2447 | } |
duke@435 | 2448 | |
duke@435 | 2449 | if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) { |
duke@435 | 2450 | #if !defined(PRODUCT) |
duke@435 | 2451 | if (TracePartialPeeling) { |
duke@435 | 2452 | tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c", |
duke@435 | 2453 | new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F'); |
duke@435 | 2454 | } |
duke@435 | 2455 | #endif |
duke@435 | 2456 | if (new_peel_if != NULL) { |
duke@435 | 2457 | remove_cmpi_loop_exit(new_peel_if, loop); |
duke@435 | 2458 | } |
duke@435 | 2459 | // Inhibit more partial peeling on this loop |
duke@435 | 2460 | assert(!head->is_partial_peel_loop(), "not partial peeled"); |
duke@435 | 2461 | head->mark_partial_peel_failed(); |
duke@435 | 2462 | return false; |
duke@435 | 2463 | } |
duke@435 | 2464 | |
duke@435 | 2465 | // Step 3: clone loop, retarget control, and insert new phis |
duke@435 | 2466 | |
duke@435 | 2467 | // Create new loop head for new phis and to hang |
duke@435 | 2468 | // the nodes being moved (sinked) from the peel region. |
duke@435 | 2469 | LoopNode* new_head = new (C, 3) LoopNode(last_peel, last_peel); |
duke@435 | 2470 | _igvn.register_new_node_with_optimizer(new_head); |
duke@435 | 2471 | assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled"); |
duke@435 | 2472 | first_not_peeled->set_req(0, new_head); |
duke@435 | 2473 | set_loop(new_head, loop); |
duke@435 | 2474 | loop->_body.push(new_head); |
duke@435 | 2475 | not_peel.set(new_head->_idx); |
duke@435 | 2476 | set_idom(new_head, last_peel, dom_depth(first_not_peeled)); |
duke@435 | 2477 | set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled)); |
duke@435 | 2478 | |
duke@435 | 2479 | while (sink_list.size() > 0) { |
duke@435 | 2480 | Node* n = sink_list.pop(); |
duke@435 | 2481 | set_ctrl(n, new_head); |
duke@435 | 2482 | } |
duke@435 | 2483 | |
duke@435 | 2484 | assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); |
duke@435 | 2485 | |
duke@435 | 2486 | clone_loop( loop, old_new, dd ); |
duke@435 | 2487 | |
duke@435 | 2488 | const uint clone_exit_idx = 1; |
duke@435 | 2489 | const uint orig_exit_idx = 2; |
duke@435 | 2490 | assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop"); |
duke@435 | 2491 | |
duke@435 | 2492 | Node* head_clone = old_new[head->_idx]; |
duke@435 | 2493 | LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop(); |
duke@435 | 2494 | Node* orig_tail_clone = head_clone->in(2); |
duke@435 | 2495 | |
duke@435 | 2496 | // Add phi if "def" node is in peel set and "use" is not |
duke@435 | 2497 | |
duke@435 | 2498 | for(i = 0; i < peel_list.size(); i++ ) { |
duke@435 | 2499 | Node *def = peel_list.at(i); |
duke@435 | 2500 | if (!def->is_CFG()) { |
duke@435 | 2501 | for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { |
duke@435 | 2502 | Node *use = def->fast_out(j); |
duke@435 | 2503 | if (has_node(use) && use->in(0) != C->top() && |
duke@435 | 2504 | (!peel.test(use->_idx) || |
duke@435 | 2505 | (use->is_Phi() && use->in(0) == head)) ) { |
duke@435 | 2506 | worklist.push(use); |
duke@435 | 2507 | } |
duke@435 | 2508 | } |
duke@435 | 2509 | while( worklist.size() ) { |
duke@435 | 2510 | Node *use = worklist.pop(); |
duke@435 | 2511 | for (uint j = 1; j < use->req(); j++) { |
duke@435 | 2512 | Node* n = use->in(j); |
duke@435 | 2513 | if (n == def) { |
duke@435 | 2514 | |
duke@435 | 2515 | // "def" is in peel set, "use" is not in peel set |
duke@435 | 2516 | // or "use" is in the entry boundary (a phi) of the peel set |
duke@435 | 2517 | |
duke@435 | 2518 | Node* use_c = has_ctrl(use) ? get_ctrl(use) : use; |
duke@435 | 2519 | |
duke@435 | 2520 | if ( loop->is_member(get_loop( use_c )) ) { |
duke@435 | 2521 | // use is in loop |
duke@435 | 2522 | if (old_new[use->_idx] != NULL) { // null for dead code |
duke@435 | 2523 | Node* use_clone = old_new[use->_idx]; |
duke@435 | 2524 | _igvn.hash_delete(use); |
duke@435 | 2525 | use->set_req(j, C->top()); |
duke@435 | 2526 | _igvn._worklist.push(use); |
duke@435 | 2527 | insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone ); |
duke@435 | 2528 | } |
duke@435 | 2529 | } else { |
duke@435 | 2530 | assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format"); |
duke@435 | 2531 | // use is not in the loop, check if the live range includes the cut |
duke@435 | 2532 | Node* lp_if = use_c->in(orig_exit_idx)->in(0); |
duke@435 | 2533 | if (not_peel.test(lp_if->_idx)) { |
duke@435 | 2534 | assert(j == orig_exit_idx, "use from original loop"); |
duke@435 | 2535 | insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone ); |
duke@435 | 2536 | } |
duke@435 | 2537 | } |
duke@435 | 2538 | } |
duke@435 | 2539 | } |
duke@435 | 2540 | } |
duke@435 | 2541 | } |
duke@435 | 2542 | } |
duke@435 | 2543 | |
duke@435 | 2544 | // Step 3b: retarget control |
duke@435 | 2545 | |
duke@435 | 2546 | // Redirect control to the new loop head if a cloned node in |
duke@435 | 2547 | // the not_peeled region has control that points into the peeled region. |
duke@435 | 2548 | // This necessary because the cloned peeled region will be outside |
duke@435 | 2549 | // the loop. |
duke@435 | 2550 | // from to |
duke@435 | 2551 | // cloned-peeled <---+ |
duke@435 | 2552 | // new_head_clone: | <--+ |
duke@435 | 2553 | // cloned-not_peeled in(0) in(0) |
duke@435 | 2554 | // orig-peeled |
duke@435 | 2555 | |
duke@435 | 2556 | for(i = 0; i < loop->_body.size(); i++ ) { |
duke@435 | 2557 | Node *n = loop->_body.at(i); |
duke@435 | 2558 | if (!n->is_CFG() && n->in(0) != NULL && |
duke@435 | 2559 | not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) { |
duke@435 | 2560 | Node* n_clone = old_new[n->_idx]; |
duke@435 | 2561 | _igvn.hash_delete(n_clone); |
duke@435 | 2562 | n_clone->set_req(0, new_head_clone); |
duke@435 | 2563 | _igvn._worklist.push(n_clone); |
duke@435 | 2564 | } |
duke@435 | 2565 | } |
duke@435 | 2566 | |
duke@435 | 2567 | // Backedge of the surviving new_head (the clone) is original last_peel |
duke@435 | 2568 | _igvn.hash_delete(new_head_clone); |
duke@435 | 2569 | new_head_clone->set_req(LoopNode::LoopBackControl, last_peel); |
duke@435 | 2570 | _igvn._worklist.push(new_head_clone); |
duke@435 | 2571 | |
duke@435 | 2572 | // Cut first node in original not_peel set |
duke@435 | 2573 | _igvn.hash_delete(new_head); |
duke@435 | 2574 | new_head->set_req(LoopNode::EntryControl, C->top()); |
duke@435 | 2575 | new_head->set_req(LoopNode::LoopBackControl, C->top()); |
duke@435 | 2576 | _igvn._worklist.push(new_head); |
duke@435 | 2577 | |
duke@435 | 2578 | // Copy head_clone back-branch info to original head |
duke@435 | 2579 | // and remove original head's loop entry and |
duke@435 | 2580 | // clone head's back-branch |
duke@435 | 2581 | _igvn.hash_delete(head); |
duke@435 | 2582 | _igvn.hash_delete(head_clone); |
duke@435 | 2583 | head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl)); |
duke@435 | 2584 | head->set_req(LoopNode::LoopBackControl, C->top()); |
duke@435 | 2585 | head_clone->set_req(LoopNode::LoopBackControl, C->top()); |
duke@435 | 2586 | _igvn._worklist.push(head); |
duke@435 | 2587 | _igvn._worklist.push(head_clone); |
duke@435 | 2588 | |
duke@435 | 2589 | // Similarly modify the phis |
duke@435 | 2590 | for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) { |
duke@435 | 2591 | Node* use = head->fast_out(k); |
duke@435 | 2592 | if (use->is_Phi() && use->outcnt() > 0) { |
duke@435 | 2593 | Node* use_clone = old_new[use->_idx]; |
duke@435 | 2594 | _igvn.hash_delete(use); |
duke@435 | 2595 | _igvn.hash_delete(use_clone); |
duke@435 | 2596 | use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl)); |
duke@435 | 2597 | use->set_req(LoopNode::LoopBackControl, C->top()); |
duke@435 | 2598 | use_clone->set_req(LoopNode::LoopBackControl, C->top()); |
duke@435 | 2599 | _igvn._worklist.push(use); |
duke@435 | 2600 | _igvn._worklist.push(use_clone); |
duke@435 | 2601 | } |
duke@435 | 2602 | } |
duke@435 | 2603 | |
duke@435 | 2604 | // Step 4: update dominator tree and dominator depth |
duke@435 | 2605 | |
duke@435 | 2606 | set_idom(head, orig_tail_clone, dd); |
duke@435 | 2607 | recompute_dom_depth(); |
duke@435 | 2608 | |
duke@435 | 2609 | // Inhibit more partial peeling on this loop |
duke@435 | 2610 | new_head_clone->set_partial_peel_loop(); |
duke@435 | 2611 | C->set_major_progress(); |
duke@435 | 2612 | |
duke@435 | 2613 | #if !defined(PRODUCT) |
duke@435 | 2614 | if (TracePartialPeeling) { |
duke@435 | 2615 | tty->print_cr("\nafter partial peel one iteration"); |
duke@435 | 2616 | Node_List wl(area); |
duke@435 | 2617 | Node* t = last_peel; |
duke@435 | 2618 | while (true) { |
duke@435 | 2619 | wl.push(t); |
duke@435 | 2620 | if (t == head_clone) break; |
duke@435 | 2621 | t = idom(t); |
duke@435 | 2622 | } |
duke@435 | 2623 | while (wl.size() > 0) { |
duke@435 | 2624 | Node* tt = wl.pop(); |
duke@435 | 2625 | if (tt == head) tty->print_cr("orig head"); |
duke@435 | 2626 | else if (tt == new_head_clone) tty->print_cr("new head"); |
duke@435 | 2627 | else if (tt == head_clone) tty->print_cr("clone head"); |
duke@435 | 2628 | tt->dump(); |
duke@435 | 2629 | } |
duke@435 | 2630 | } |
duke@435 | 2631 | #endif |
duke@435 | 2632 | return true; |
duke@435 | 2633 | } |
duke@435 | 2634 | |
duke@435 | 2635 | //------------------------------reorg_offsets---------------------------------- |
duke@435 | 2636 | // Reorganize offset computations to lower register pressure. Mostly |
duke@435 | 2637 | // prevent loop-fallout uses of the pre-incremented trip counter (which are |
duke@435 | 2638 | // then alive with the post-incremented trip counter forcing an extra |
duke@435 | 2639 | // register move) |
duke@435 | 2640 | void PhaseIdealLoop::reorg_offsets( IdealLoopTree *loop ) { |
duke@435 | 2641 | |
duke@435 | 2642 | CountedLoopNode *cl = loop->_head->as_CountedLoop(); |
duke@435 | 2643 | CountedLoopEndNode *cle = cl->loopexit(); |
duke@435 | 2644 | if( !cle ) return; // The occasional dead loop |
duke@435 | 2645 | // Find loop exit control |
duke@435 | 2646 | Node *exit = cle->proj_out(false); |
duke@435 | 2647 | assert( exit->Opcode() == Op_IfFalse, "" ); |
duke@435 | 2648 | |
duke@435 | 2649 | // Check for the special case of folks using the pre-incremented |
duke@435 | 2650 | // trip-counter on the fall-out path (forces the pre-incremented |
duke@435 | 2651 | // and post-incremented trip counter to be live at the same time). |
duke@435 | 2652 | // Fix this by adjusting to use the post-increment trip counter. |
duke@435 | 2653 | Node *phi = cl->phi(); |
duke@435 | 2654 | if( !phi ) return; // Dead infinite loop |
duke@435 | 2655 | bool progress = true; |
duke@435 | 2656 | while (progress) { |
duke@435 | 2657 | progress = false; |
duke@435 | 2658 | for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) { |
duke@435 | 2659 | Node* use = phi->fast_out(i); // User of trip-counter |
duke@435 | 2660 | if (!has_ctrl(use)) continue; |
duke@435 | 2661 | Node *u_ctrl = get_ctrl(use); |
duke@435 | 2662 | if( use->is_Phi() ) { |
duke@435 | 2663 | u_ctrl = NULL; |
duke@435 | 2664 | for( uint j = 1; j < use->req(); j++ ) |
duke@435 | 2665 | if( use->in(j) == phi ) |
duke@435 | 2666 | u_ctrl = dom_lca( u_ctrl, use->in(0)->in(j) ); |
duke@435 | 2667 | } |
duke@435 | 2668 | IdealLoopTree *u_loop = get_loop(u_ctrl); |
duke@435 | 2669 | // Look for loop-invariant use |
duke@435 | 2670 | if( u_loop == loop ) continue; |
duke@435 | 2671 | if( loop->is_member( u_loop ) ) continue; |
duke@435 | 2672 | // Check that use is live out the bottom. Assuming the trip-counter |
duke@435 | 2673 | // update is right at the bottom, uses of of the loop middle are ok. |
duke@435 | 2674 | if( dom_lca( exit, u_ctrl ) != exit ) continue; |
duke@435 | 2675 | // protect against stride not being a constant |
duke@435 | 2676 | if( !cle->stride_is_con() ) continue; |
duke@435 | 2677 | // Hit! Refactor use to use the post-incremented tripcounter. |
duke@435 | 2678 | // Compute a post-increment tripcounter. |
duke@435 | 2679 | Node *opaq = new (C, 2) Opaque2Node( cle->incr() ); |
duke@435 | 2680 | register_new_node( opaq, u_ctrl ); |
duke@435 | 2681 | Node *neg_stride = _igvn.intcon(-cle->stride_con()); |
duke@435 | 2682 | set_ctrl(neg_stride, C->root()); |
duke@435 | 2683 | Node *post = new (C, 3) AddINode( opaq, neg_stride); |
duke@435 | 2684 | register_new_node( post, u_ctrl ); |
duke@435 | 2685 | _igvn.hash_delete(use); |
duke@435 | 2686 | _igvn._worklist.push(use); |
duke@435 | 2687 | for( uint j = 1; j < use->req(); j++ ) |
duke@435 | 2688 | if( use->in(j) == phi ) |
duke@435 | 2689 | use->set_req(j, post); |
duke@435 | 2690 | // Since DU info changed, rerun loop |
duke@435 | 2691 | progress = true; |
duke@435 | 2692 | break; |
duke@435 | 2693 | } |
duke@435 | 2694 | } |
duke@435 | 2695 | |
duke@435 | 2696 | } |