Tue, 02 Sep 2014 12:48:45 -0700
8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
Summary: Add new C2 intrinsic for BigInteger::multiplyToLen() on x86 in 64-bit VM.
Reviewed-by: roland
duke@435 | 1 | /* |
mikael@6198 | 2 | * Copyright (c) 1999, 2013, Oracle and/or its affiliates. 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 | * |
trims@1907 | 19 | * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
trims@1907 | 20 | * or visit www.oracle.com if you need additional information or have any |
trims@1907 | 21 | * questions. |
duke@435 | 22 | * |
duke@435 | 23 | */ |
duke@435 | 24 | |
stefank@2314 | 25 | #include "precompiled.hpp" |
stefank@2314 | 26 | #include "memory/allocation.inline.hpp" |
stefank@2314 | 27 | #include "opto/addnode.hpp" |
stefank@2314 | 28 | #include "opto/connode.hpp" |
stefank@2314 | 29 | #include "opto/divnode.hpp" |
stefank@2314 | 30 | #include "opto/loopnode.hpp" |
kvn@3243 | 31 | #include "opto/matcher.hpp" |
stefank@2314 | 32 | #include "opto/mulnode.hpp" |
stefank@2314 | 33 | #include "opto/rootnode.hpp" |
stefank@2314 | 34 | #include "opto/subnode.hpp" |
duke@435 | 35 | |
duke@435 | 36 | //============================================================================= |
duke@435 | 37 | //------------------------------split_thru_phi--------------------------------- |
duke@435 | 38 | // Split Node 'n' through merge point if there is enough win. |
duke@435 | 39 | Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) { |
never@504 | 40 | if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) { |
never@504 | 41 | // ConvI2L may have type information on it which is unsafe to push up |
never@504 | 42 | // so disable this for now |
never@504 | 43 | return NULL; |
never@504 | 44 | } |
rbackman@6101 | 45 | |
duke@435 | 46 | int wins = 0; |
kvn@2665 | 47 | assert(!n->is_CFG(), ""); |
kvn@2665 | 48 | assert(region->is_Region(), ""); |
kvn@499 | 49 | |
kvn@499 | 50 | const Type* type = n->bottom_type(); |
kvn@499 | 51 | const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr(); |
kvn@499 | 52 | Node *phi; |
kvn@2665 | 53 | if (t_oop != NULL && t_oop->is_known_instance_field()) { |
kvn@499 | 54 | int iid = t_oop->instance_id(); |
kvn@499 | 55 | int index = C->get_alias_index(t_oop); |
kvn@499 | 56 | int offset = t_oop->offset(); |
kvn@4115 | 57 | phi = new (C) PhiNode(region, type, NULL, iid, index, offset); |
kvn@499 | 58 | } else { |
kvn@1689 | 59 | phi = PhiNode::make_blank(region, n); |
kvn@499 | 60 | } |
duke@435 | 61 | uint old_unique = C->unique(); |
kvn@2665 | 62 | for (uint i = 1; i < region->req(); i++) { |
duke@435 | 63 | Node *x; |
duke@435 | 64 | Node* the_clone = NULL; |
kvn@2665 | 65 | if (region->in(i) == C->top()) { |
duke@435 | 66 | x = C->top(); // Dead path? Use a dead data op |
duke@435 | 67 | } else { |
duke@435 | 68 | x = n->clone(); // Else clone up the data op |
duke@435 | 69 | the_clone = x; // Remember for possible deletion. |
duke@435 | 70 | // Alter data node to use pre-phi inputs |
kvn@2665 | 71 | if (n->in(0) == region) |
duke@435 | 72 | x->set_req( 0, region->in(i) ); |
kvn@2665 | 73 | for (uint j = 1; j < n->req(); j++) { |
duke@435 | 74 | Node *in = n->in(j); |
kvn@2665 | 75 | if (in->is_Phi() && in->in(0) == region) |
duke@435 | 76 | x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone |
duke@435 | 77 | } |
duke@435 | 78 | } |
duke@435 | 79 | // Check for a 'win' on some paths |
duke@435 | 80 | const Type *t = x->Value(&_igvn); |
duke@435 | 81 | |
duke@435 | 82 | bool singleton = t->singleton(); |
duke@435 | 83 | |
duke@435 | 84 | // A TOP singleton indicates that there are no possible values incoming |
duke@435 | 85 | // along a particular edge. In most cases, this is OK, and the Phi will |
duke@435 | 86 | // be eliminated later in an Ideal call. However, we can't allow this to |
duke@435 | 87 | // happen if the singleton occurs on loop entry, as the elimination of |
duke@435 | 88 | // the PhiNode may cause the resulting node to migrate back to a previous |
duke@435 | 89 | // loop iteration. |
kvn@2665 | 90 | if (singleton && t == Type::TOP) { |
duke@435 | 91 | // Is_Loop() == false does not confirm the absence of a loop (e.g., an |
duke@435 | 92 | // irreducible loop may not be indicated by an affirmative is_Loop()); |
duke@435 | 93 | // therefore, the only top we can split thru a phi is on a backedge of |
duke@435 | 94 | // a loop. |
duke@435 | 95 | singleton &= region->is_Loop() && (i != LoopNode::EntryControl); |
duke@435 | 96 | } |
duke@435 | 97 | |
kvn@2665 | 98 | if (singleton) { |
duke@435 | 99 | wins++; |
duke@435 | 100 | x = ((PhaseGVN&)_igvn).makecon(t); |
duke@435 | 101 | } else { |
duke@435 | 102 | // We now call Identity to try to simplify the cloned node. |
duke@435 | 103 | // Note that some Identity methods call phase->type(this). |
duke@435 | 104 | // Make sure that the type array is big enough for |
duke@435 | 105 | // our new node, even though we may throw the node away. |
duke@435 | 106 | // (Note: This tweaking with igvn only works because x is a new node.) |
duke@435 | 107 | _igvn.set_type(x, t); |
kvn@728 | 108 | // If x is a TypeNode, capture any more-precise type permanently into Node |
twisti@1040 | 109 | // otherwise it will be not updated during igvn->transform since |
kvn@728 | 110 | // igvn->type(x) is set to x->Value() already. |
kvn@728 | 111 | x->raise_bottom_type(t); |
duke@435 | 112 | Node *y = x->Identity(&_igvn); |
kvn@2665 | 113 | if (y != x) { |
duke@435 | 114 | wins++; |
duke@435 | 115 | x = y; |
duke@435 | 116 | } else { |
duke@435 | 117 | y = _igvn.hash_find(x); |
kvn@2665 | 118 | if (y) { |
duke@435 | 119 | wins++; |
duke@435 | 120 | x = y; |
duke@435 | 121 | } else { |
duke@435 | 122 | // Else x is a new node we are keeping |
duke@435 | 123 | // We do not need register_new_node_with_optimizer |
duke@435 | 124 | // because set_type has already been called. |
duke@435 | 125 | _igvn._worklist.push(x); |
duke@435 | 126 | } |
duke@435 | 127 | } |
duke@435 | 128 | } |
duke@435 | 129 | if (x != the_clone && the_clone != NULL) |
duke@435 | 130 | _igvn.remove_dead_node(the_clone); |
duke@435 | 131 | phi->set_req( i, x ); |
duke@435 | 132 | } |
duke@435 | 133 | // Too few wins? |
kvn@2665 | 134 | if (wins <= policy) { |
duke@435 | 135 | _igvn.remove_dead_node(phi); |
duke@435 | 136 | return NULL; |
duke@435 | 137 | } |
duke@435 | 138 | |
duke@435 | 139 | // Record Phi |
duke@435 | 140 | register_new_node( phi, region ); |
duke@435 | 141 | |
kvn@2665 | 142 | for (uint i2 = 1; i2 < phi->req(); i2++) { |
duke@435 | 143 | Node *x = phi->in(i2); |
duke@435 | 144 | // If we commoned up the cloned 'x' with another existing Node, |
duke@435 | 145 | // the existing Node picks up a new use. We need to make the |
duke@435 | 146 | // existing Node occur higher up so it dominates its uses. |
duke@435 | 147 | Node *old_ctrl; |
duke@435 | 148 | IdealLoopTree *old_loop; |
duke@435 | 149 | |
kvn@2665 | 150 | if (x->is_Con()) { |
kvn@2665 | 151 | // Constant's control is always root. |
kvn@2665 | 152 | set_ctrl(x, C->root()); |
kvn@2665 | 153 | continue; |
kvn@2665 | 154 | } |
duke@435 | 155 | // The occasional new node |
kvn@2665 | 156 | if (x->_idx >= old_unique) { // Found a new, unplaced node? |
kvn@2665 | 157 | old_ctrl = NULL; |
kvn@2665 | 158 | old_loop = NULL; // Not in any prior loop |
duke@435 | 159 | } else { |
kvn@2665 | 160 | old_ctrl = get_ctrl(x); |
duke@435 | 161 | old_loop = get_loop(old_ctrl); // Get prior loop |
duke@435 | 162 | } |
duke@435 | 163 | // New late point must dominate new use |
kvn@2665 | 164 | Node *new_ctrl = dom_lca(old_ctrl, region->in(i2)); |
kvn@2665 | 165 | if (new_ctrl == old_ctrl) // Nothing is changed |
kvn@2665 | 166 | continue; |
kvn@2665 | 167 | |
kvn@2665 | 168 | IdealLoopTree *new_loop = get_loop(new_ctrl); |
kvn@2665 | 169 | |
kvn@2665 | 170 | // Don't move x into a loop if its uses are |
kvn@2665 | 171 | // outside of loop. Otherwise x will be cloned |
kvn@2665 | 172 | // for each use outside of this loop. |
kvn@2665 | 173 | IdealLoopTree *use_loop = get_loop(region); |
kvn@2665 | 174 | if (!new_loop->is_member(use_loop) && |
kvn@2665 | 175 | (old_loop == NULL || !new_loop->is_member(old_loop))) { |
kvn@2665 | 176 | // Take early control, later control will be recalculated |
kvn@2665 | 177 | // during next iteration of loop optimizations. |
kvn@2665 | 178 | new_ctrl = get_early_ctrl(x); |
kvn@2665 | 179 | new_loop = get_loop(new_ctrl); |
kvn@2665 | 180 | } |
duke@435 | 181 | // Set new location |
duke@435 | 182 | set_ctrl(x, new_ctrl); |
duke@435 | 183 | // If changing loop bodies, see if we need to collect into new body |
kvn@2665 | 184 | if (old_loop != new_loop) { |
kvn@2665 | 185 | if (old_loop && !old_loop->_child) |
duke@435 | 186 | old_loop->_body.yank(x); |
kvn@2665 | 187 | if (!new_loop->_child) |
duke@435 | 188 | new_loop->_body.push(x); // Collect body info |
duke@435 | 189 | } |
duke@435 | 190 | } |
duke@435 | 191 | |
duke@435 | 192 | return phi; |
duke@435 | 193 | } |
duke@435 | 194 | |
duke@435 | 195 | //------------------------------dominated_by------------------------------------ |
duke@435 | 196 | // Replace the dominated test with an obvious true or false. Place it on the |
duke@435 | 197 | // IGVN worklist for later cleanup. Move control-dependent data Nodes on the |
duke@435 | 198 | // live path up to the dominating control. |
kvn@3038 | 199 | void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff, bool flip, bool exclude_loop_predicate ) { |
duke@435 | 200 | #ifndef PRODUCT |
kvn@2665 | 201 | if (VerifyLoopOptimizations && PrintOpto) tty->print_cr("dominating test"); |
duke@435 | 202 | #endif |
duke@435 | 203 | |
duke@435 | 204 | |
duke@435 | 205 | // prevdom is the dominating projection of the dominating test. |
duke@435 | 206 | assert( iff->is_If(), "" ); |
duke@435 | 207 | assert( iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added"); |
duke@435 | 208 | int pop = prevdom->Opcode(); |
duke@435 | 209 | assert( pop == Op_IfFalse || pop == Op_IfTrue, "" ); |
kvn@2665 | 210 | if (flip) { |
kvn@2665 | 211 | if (pop == Op_IfTrue) |
kvn@2665 | 212 | pop = Op_IfFalse; |
kvn@2665 | 213 | else |
kvn@2665 | 214 | pop = Op_IfTrue; |
kvn@2665 | 215 | } |
duke@435 | 216 | // 'con' is set to true or false to kill the dominated test. |
duke@435 | 217 | Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO); |
duke@435 | 218 | set_ctrl(con, C->root()); // Constant gets a new use |
duke@435 | 219 | // Hack the dominated test |
kvn@3847 | 220 | _igvn.replace_input_of(iff, 1, con); |
duke@435 | 221 | |
duke@435 | 222 | // If I dont have a reachable TRUE and FALSE path following the IfNode then |
duke@435 | 223 | // I can assume this path reaches an infinite loop. In this case it's not |
duke@435 | 224 | // important to optimize the data Nodes - either the whole compilation will |
duke@435 | 225 | // be tossed or this path (and all data Nodes) will go dead. |
kvn@2665 | 226 | if (iff->outcnt() != 2) return; |
duke@435 | 227 | |
duke@435 | 228 | // Make control-dependent data Nodes on the live path (path that will remain |
duke@435 | 229 | // once the dominated IF is removed) become control-dependent on the |
duke@435 | 230 | // dominating projection. |
kvn@3038 | 231 | Node* dp = iff->as_If()->proj_out(pop == Op_IfTrue); |
kvn@3038 | 232 | |
kvn@3038 | 233 | // Loop predicates may have depending checks which should not |
kvn@3038 | 234 | // be skipped. For example, range check predicate has two checks |
kvn@3038 | 235 | // for lower and upper bounds. |
morris@4774 | 236 | if (dp == NULL) |
morris@4774 | 237 | return; |
morris@4774 | 238 | |
morris@4774 | 239 | ProjNode* dp_proj = dp->as_Proj(); |
morris@4774 | 240 | ProjNode* unc_proj = iff->as_If()->proj_out(1 - dp_proj->_con)->as_Proj(); |
kvn@3038 | 241 | if (exclude_loop_predicate && |
roland@5981 | 242 | unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate)) |
kvn@3038 | 243 | return; // Let IGVN transformation change control dependence. |
kvn@3038 | 244 | |
duke@435 | 245 | IdealLoopTree *old_loop = get_loop(dp); |
duke@435 | 246 | |
duke@435 | 247 | for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { |
duke@435 | 248 | Node* cd = dp->fast_out(i); // Control-dependent node |
kvn@2665 | 249 | if (cd->depends_only_on_test()) { |
kvn@2665 | 250 | assert(cd->in(0) == dp, ""); |
kvn@3847 | 251 | _igvn.replace_input_of(cd, 0, prevdom); |
kvn@2665 | 252 | set_early_ctrl(cd); |
duke@435 | 253 | IdealLoopTree *new_loop = get_loop(get_ctrl(cd)); |
kvn@2665 | 254 | if (old_loop != new_loop) { |
kvn@2665 | 255 | if (!old_loop->_child) old_loop->_body.yank(cd); |
kvn@2665 | 256 | if (!new_loop->_child) new_loop->_body.push(cd); |
duke@435 | 257 | } |
duke@435 | 258 | --i; |
duke@435 | 259 | --imax; |
duke@435 | 260 | } |
duke@435 | 261 | } |
duke@435 | 262 | } |
duke@435 | 263 | |
duke@435 | 264 | //------------------------------has_local_phi_input---------------------------- |
duke@435 | 265 | // Return TRUE if 'n' has Phi inputs from its local block and no other |
duke@435 | 266 | // block-local inputs (all non-local-phi inputs come from earlier blocks) |
duke@435 | 267 | Node *PhaseIdealLoop::has_local_phi_input( Node *n ) { |
duke@435 | 268 | Node *n_ctrl = get_ctrl(n); |
duke@435 | 269 | // See if some inputs come from a Phi in this block, or from before |
duke@435 | 270 | // this block. |
duke@435 | 271 | uint i; |
duke@435 | 272 | for( i = 1; i < n->req(); i++ ) { |
duke@435 | 273 | Node *phi = n->in(i); |
duke@435 | 274 | if( phi->is_Phi() && phi->in(0) == n_ctrl ) |
duke@435 | 275 | break; |
duke@435 | 276 | } |
duke@435 | 277 | if( i >= n->req() ) |
duke@435 | 278 | return NULL; // No Phi inputs; nowhere to clone thru |
duke@435 | 279 | |
duke@435 | 280 | // Check for inputs created between 'n' and the Phi input. These |
duke@435 | 281 | // must split as well; they have already been given the chance |
duke@435 | 282 | // (courtesy of a post-order visit) and since they did not we must |
duke@435 | 283 | // recover the 'cost' of splitting them by being very profitable |
duke@435 | 284 | // when splitting 'n'. Since this is unlikely we simply give up. |
duke@435 | 285 | for( i = 1; i < n->req(); i++ ) { |
duke@435 | 286 | Node *m = n->in(i); |
duke@435 | 287 | if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) { |
duke@435 | 288 | // We allow the special case of AddP's with no local inputs. |
duke@435 | 289 | // This allows us to split-up address expressions. |
duke@435 | 290 | if (m->is_AddP() && |
duke@435 | 291 | get_ctrl(m->in(2)) != n_ctrl && |
duke@435 | 292 | get_ctrl(m->in(3)) != n_ctrl) { |
duke@435 | 293 | // Move the AddP up to dominating point |
duke@435 | 294 | set_ctrl_and_loop(m, find_non_split_ctrl(idom(n_ctrl))); |
duke@435 | 295 | continue; |
duke@435 | 296 | } |
duke@435 | 297 | return NULL; |
duke@435 | 298 | } |
duke@435 | 299 | } |
duke@435 | 300 | |
duke@435 | 301 | return n_ctrl; |
duke@435 | 302 | } |
duke@435 | 303 | |
duke@435 | 304 | //------------------------------remix_address_expressions---------------------- |
duke@435 | 305 | // Rework addressing expressions to get the most loop-invariant stuff |
duke@435 | 306 | // moved out. We'd like to do all associative operators, but it's especially |
duke@435 | 307 | // important (common) to do address expressions. |
duke@435 | 308 | Node *PhaseIdealLoop::remix_address_expressions( Node *n ) { |
duke@435 | 309 | if (!has_ctrl(n)) return NULL; |
duke@435 | 310 | Node *n_ctrl = get_ctrl(n); |
duke@435 | 311 | IdealLoopTree *n_loop = get_loop(n_ctrl); |
duke@435 | 312 | |
duke@435 | 313 | // See if 'n' mixes loop-varying and loop-invariant inputs and |
duke@435 | 314 | // itself is loop-varying. |
duke@435 | 315 | |
duke@435 | 316 | // Only interested in binary ops (and AddP) |
duke@435 | 317 | if( n->req() < 3 || n->req() > 4 ) return NULL; |
duke@435 | 318 | |
duke@435 | 319 | Node *n1_ctrl = get_ctrl(n->in( 1)); |
duke@435 | 320 | Node *n2_ctrl = get_ctrl(n->in( 2)); |
duke@435 | 321 | Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3)); |
duke@435 | 322 | IdealLoopTree *n1_loop = get_loop( n1_ctrl ); |
duke@435 | 323 | IdealLoopTree *n2_loop = get_loop( n2_ctrl ); |
duke@435 | 324 | IdealLoopTree *n3_loop = get_loop( n3_ctrl ); |
duke@435 | 325 | |
duke@435 | 326 | // Does one of my inputs spin in a tighter loop than self? |
duke@435 | 327 | if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) || |
duke@435 | 328 | (n_loop->is_member( n2_loop ) && n_loop != n2_loop) || |
duke@435 | 329 | (n_loop->is_member( n3_loop ) && n_loop != n3_loop) ) |
duke@435 | 330 | return NULL; // Leave well enough alone |
duke@435 | 331 | |
duke@435 | 332 | // Is at least one of my inputs loop-invariant? |
duke@435 | 333 | if( n1_loop == n_loop && |
duke@435 | 334 | n2_loop == n_loop && |
duke@435 | 335 | n3_loop == n_loop ) |
duke@435 | 336 | return NULL; // No loop-invariant inputs |
duke@435 | 337 | |
duke@435 | 338 | |
duke@435 | 339 | int n_op = n->Opcode(); |
duke@435 | 340 | |
duke@435 | 341 | // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2). |
duke@435 | 342 | if( n_op == Op_LShiftI ) { |
duke@435 | 343 | // Scale is loop invariant |
duke@435 | 344 | Node *scale = n->in(2); |
duke@435 | 345 | Node *scale_ctrl = get_ctrl(scale); |
duke@435 | 346 | IdealLoopTree *scale_loop = get_loop(scale_ctrl ); |
duke@435 | 347 | if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) ) |
duke@435 | 348 | return NULL; |
duke@435 | 349 | const TypeInt *scale_t = scale->bottom_type()->isa_int(); |
duke@435 | 350 | if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 ) |
duke@435 | 351 | return NULL; // Dont bother with byte/short masking |
duke@435 | 352 | // Add must vary with loop (else shift would be loop-invariant) |
duke@435 | 353 | Node *add = n->in(1); |
duke@435 | 354 | Node *add_ctrl = get_ctrl(add); |
duke@435 | 355 | IdealLoopTree *add_loop = get_loop(add_ctrl); |
duke@435 | 356 | //assert( n_loop == add_loop, "" ); |
duke@435 | 357 | if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops |
duke@435 | 358 | |
duke@435 | 359 | // Convert I-V into I+ (0-V); same for V-I |
duke@435 | 360 | if( add->Opcode() == Op_SubI && |
duke@435 | 361 | _igvn.type( add->in(1) ) != TypeInt::ZERO ) { |
duke@435 | 362 | Node *zero = _igvn.intcon(0); |
duke@435 | 363 | set_ctrl(zero, C->root()); |
kvn@4115 | 364 | Node *neg = new (C) SubINode( _igvn.intcon(0), add->in(2) ); |
duke@435 | 365 | register_new_node( neg, get_ctrl(add->in(2) ) ); |
kvn@4115 | 366 | add = new (C) AddINode( add->in(1), neg ); |
duke@435 | 367 | register_new_node( add, add_ctrl ); |
duke@435 | 368 | } |
duke@435 | 369 | if( add->Opcode() != Op_AddI ) return NULL; |
duke@435 | 370 | // See if one add input is loop invariant |
duke@435 | 371 | Node *add_var = add->in(1); |
duke@435 | 372 | Node *add_var_ctrl = get_ctrl(add_var); |
duke@435 | 373 | IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); |
duke@435 | 374 | Node *add_invar = add->in(2); |
duke@435 | 375 | Node *add_invar_ctrl = get_ctrl(add_invar); |
duke@435 | 376 | IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl ); |
duke@435 | 377 | if( add_var_loop == n_loop ) { |
duke@435 | 378 | } else if( add_invar_loop == n_loop ) { |
duke@435 | 379 | // Swap to find the invariant part |
duke@435 | 380 | add_invar = add_var; |
duke@435 | 381 | add_invar_ctrl = add_var_ctrl; |
duke@435 | 382 | add_invar_loop = add_var_loop; |
duke@435 | 383 | add_var = add->in(2); |
duke@435 | 384 | Node *add_var_ctrl = get_ctrl(add_var); |
duke@435 | 385 | IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); |
duke@435 | 386 | } else // Else neither input is loop invariant |
duke@435 | 387 | return NULL; |
duke@435 | 388 | if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) ) |
duke@435 | 389 | return NULL; // No invariant part of the add? |
duke@435 | 390 | |
duke@435 | 391 | // Yes! Reshape address expression! |
kvn@4115 | 392 | Node *inv_scale = new (C) LShiftINode( add_invar, scale ); |
never@1285 | 393 | Node *inv_scale_ctrl = |
never@1285 | 394 | dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ? |
never@1285 | 395 | add_invar_ctrl : scale_ctrl; |
never@1285 | 396 | register_new_node( inv_scale, inv_scale_ctrl ); |
kvn@4115 | 397 | Node *var_scale = new (C) LShiftINode( add_var, scale ); |
duke@435 | 398 | register_new_node( var_scale, n_ctrl ); |
kvn@4115 | 399 | Node *var_add = new (C) AddINode( var_scale, inv_scale ); |
duke@435 | 400 | register_new_node( var_add, n_ctrl ); |
kvn@1976 | 401 | _igvn.replace_node( n, var_add ); |
duke@435 | 402 | return var_add; |
duke@435 | 403 | } |
duke@435 | 404 | |
duke@435 | 405 | // Replace (I+V) with (V+I) |
duke@435 | 406 | if( n_op == Op_AddI || |
duke@435 | 407 | n_op == Op_AddL || |
duke@435 | 408 | n_op == Op_AddF || |
duke@435 | 409 | n_op == Op_AddD || |
duke@435 | 410 | n_op == Op_MulI || |
duke@435 | 411 | n_op == Op_MulL || |
duke@435 | 412 | n_op == Op_MulF || |
duke@435 | 413 | n_op == Op_MulD ) { |
duke@435 | 414 | if( n2_loop == n_loop ) { |
duke@435 | 415 | assert( n1_loop != n_loop, "" ); |
duke@435 | 416 | n->swap_edges(1, 2); |
duke@435 | 417 | } |
duke@435 | 418 | } |
duke@435 | 419 | |
duke@435 | 420 | // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V), |
duke@435 | 421 | // but not if I2 is a constant. |
duke@435 | 422 | if( n_op == Op_AddP ) { |
duke@435 | 423 | if( n2_loop == n_loop && n3_loop != n_loop ) { |
duke@435 | 424 | if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) { |
duke@435 | 425 | Node *n22_ctrl = get_ctrl(n->in(2)->in(2)); |
duke@435 | 426 | Node *n23_ctrl = get_ctrl(n->in(2)->in(3)); |
duke@435 | 427 | IdealLoopTree *n22loop = get_loop( n22_ctrl ); |
duke@435 | 428 | IdealLoopTree *n23_loop = get_loop( n23_ctrl ); |
duke@435 | 429 | if( n22loop != n_loop && n22loop->is_member(n_loop) && |
duke@435 | 430 | n23_loop == n_loop ) { |
kvn@4115 | 431 | Node *add1 = new (C) AddPNode( n->in(1), n->in(2)->in(2), n->in(3) ); |
duke@435 | 432 | // Stuff new AddP in the loop preheader |
duke@435 | 433 | register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) ); |
kvn@4115 | 434 | Node *add2 = new (C) AddPNode( n->in(1), add1, n->in(2)->in(3) ); |
duke@435 | 435 | register_new_node( add2, n_ctrl ); |
kvn@1976 | 436 | _igvn.replace_node( n, add2 ); |
duke@435 | 437 | return add2; |
duke@435 | 438 | } |
duke@435 | 439 | } |
duke@435 | 440 | } |
duke@435 | 441 | |
duke@435 | 442 | // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V) |
duke@435 | 443 | if( n2_loop != n_loop && n3_loop == n_loop ) { |
duke@435 | 444 | if( n->in(3)->Opcode() == Op_AddI ) { |
duke@435 | 445 | Node *V = n->in(3)->in(1); |
duke@435 | 446 | Node *I = n->in(3)->in(2); |
duke@435 | 447 | if( is_member(n_loop,get_ctrl(V)) ) { |
duke@435 | 448 | } else { |
duke@435 | 449 | Node *tmp = V; V = I; I = tmp; |
duke@435 | 450 | } |
duke@435 | 451 | if( !is_member(n_loop,get_ctrl(I)) ) { |
kvn@4115 | 452 | Node *add1 = new (C) AddPNode( n->in(1), n->in(2), I ); |
duke@435 | 453 | // Stuff new AddP in the loop preheader |
duke@435 | 454 | register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) ); |
kvn@4115 | 455 | Node *add2 = new (C) AddPNode( n->in(1), add1, V ); |
duke@435 | 456 | register_new_node( add2, n_ctrl ); |
kvn@1976 | 457 | _igvn.replace_node( n, add2 ); |
duke@435 | 458 | return add2; |
duke@435 | 459 | } |
duke@435 | 460 | } |
duke@435 | 461 | } |
duke@435 | 462 | } |
duke@435 | 463 | |
duke@435 | 464 | return NULL; |
duke@435 | 465 | } |
duke@435 | 466 | |
duke@435 | 467 | //------------------------------conditional_move------------------------------- |
duke@435 | 468 | // Attempt to replace a Phi with a conditional move. We have some pretty |
duke@435 | 469 | // strict profitability requirements. All Phis at the merge point must |
duke@435 | 470 | // be converted, so we can remove the control flow. We need to limit the |
duke@435 | 471 | // number of c-moves to a small handful. All code that was in the side-arms |
duke@435 | 472 | // of the CFG diamond is now speculatively executed. This code has to be |
duke@435 | 473 | // "cheap enough". We are pretty much limited to CFG diamonds that merge |
duke@435 | 474 | // 1 or 2 items with a total of 1 or 2 ops executed speculatively. |
duke@435 | 475 | Node *PhaseIdealLoop::conditional_move( Node *region ) { |
duke@435 | 476 | |
kvn@3243 | 477 | assert(region->is_Region(), "sanity check"); |
kvn@3243 | 478 | if (region->req() != 3) return NULL; |
duke@435 | 479 | |
duke@435 | 480 | // Check for CFG diamond |
duke@435 | 481 | Node *lp = region->in(1); |
duke@435 | 482 | Node *rp = region->in(2); |
kvn@3243 | 483 | if (!lp || !rp) return NULL; |
duke@435 | 484 | Node *lp_c = lp->in(0); |
kvn@3243 | 485 | if (lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If()) return NULL; |
duke@435 | 486 | IfNode *iff = lp_c->as_If(); |
duke@435 | 487 | |
duke@435 | 488 | // Check for ops pinned in an arm of the diamond. |
duke@435 | 489 | // Can't remove the control flow in this case |
kvn@3243 | 490 | if (lp->outcnt() > 1) return NULL; |
kvn@3243 | 491 | if (rp->outcnt() > 1) return NULL; |
kvn@3243 | 492 | |
kvn@3243 | 493 | IdealLoopTree* r_loop = get_loop(region); |
kvn@3243 | 494 | assert(r_loop == get_loop(iff), "sanity"); |
kvn@3243 | 495 | // Always convert to CMOVE if all results are used only outside this loop. |
kvn@3243 | 496 | bool used_inside_loop = (r_loop == _ltree_root); |
duke@435 | 497 | |
duke@435 | 498 | // Check profitability |
duke@435 | 499 | int cost = 0; |
kvn@470 | 500 | int phis = 0; |
duke@435 | 501 | for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { |
duke@435 | 502 | Node *out = region->fast_out(i); |
kvn@3243 | 503 | if (!out->is_Phi()) continue; // Ignore other control edges, etc |
kvn@470 | 504 | phis++; |
duke@435 | 505 | PhiNode* phi = out->as_Phi(); |
kvn@3243 | 506 | BasicType bt = phi->type()->basic_type(); |
kvn@3243 | 507 | switch (bt) { |
kvn@3243 | 508 | case T_FLOAT: |
kvn@3243 | 509 | case T_DOUBLE: { |
kvn@3243 | 510 | cost += Matcher::float_cmove_cost(); // Could be very expensive |
kvn@3243 | 511 | break; |
kvn@3243 | 512 | } |
kvn@3243 | 513 | case T_LONG: { |
kvn@3243 | 514 | cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's |
kvn@3243 | 515 | } |
duke@435 | 516 | case T_INT: // These all CMOV fine |
kvn@3243 | 517 | case T_ADDRESS: { // (RawPtr) |
duke@435 | 518 | cost++; |
duke@435 | 519 | break; |
kvn@3243 | 520 | } |
kvn@728 | 521 | case T_NARROWOOP: // Fall through |
duke@435 | 522 | case T_OBJECT: { // Base oops are OK, but not derived oops |
kvn@728 | 523 | const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr(); |
duke@435 | 524 | // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a |
duke@435 | 525 | // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus |
duke@435 | 526 | // CMOVE'ing a derived pointer requires we also CMOVE the base. If we |
duke@435 | 527 | // have a Phi for the base here that we convert to a CMOVE all is well |
duke@435 | 528 | // and good. But if the base is dead, we'll not make a CMOVE. Later |
duke@435 | 529 | // the allocator will have to produce a base by creating a CMOVE of the |
duke@435 | 530 | // relevant bases. This puts the allocator in the business of |
duke@435 | 531 | // manufacturing expensive instructions, generally a bad plan. |
duke@435 | 532 | // Just Say No to Conditionally-Moved Derived Pointers. |
kvn@3243 | 533 | if (tp && tp->offset() != 0) |
duke@435 | 534 | return NULL; |
duke@435 | 535 | cost++; |
duke@435 | 536 | break; |
duke@435 | 537 | } |
duke@435 | 538 | default: |
duke@435 | 539 | return NULL; // In particular, can't do memory or I/O |
duke@435 | 540 | } |
duke@435 | 541 | // Add in cost any speculative ops |
kvn@3243 | 542 | for (uint j = 1; j < region->req(); j++) { |
duke@435 | 543 | Node *proj = region->in(j); |
duke@435 | 544 | Node *inp = phi->in(j); |
duke@435 | 545 | if (get_ctrl(inp) == proj) { // Found local op |
duke@435 | 546 | cost++; |
duke@435 | 547 | // Check for a chain of dependent ops; these will all become |
duke@435 | 548 | // speculative in a CMOV. |
kvn@3243 | 549 | for (uint k = 1; k < inp->req(); k++) |
duke@435 | 550 | if (get_ctrl(inp->in(k)) == proj) |
kvn@3243 | 551 | cost += ConditionalMoveLimit; // Too much speculative goo |
duke@435 | 552 | } |
duke@435 | 553 | } |
kvn@728 | 554 | // See if the Phi is used by a Cmp or Narrow oop Decode/Encode. |
kvn@728 | 555 | // This will likely Split-If, a higher-payoff operation. |
duke@435 | 556 | for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) { |
duke@435 | 557 | Node* use = phi->fast_out(k); |
roland@4159 | 558 | if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr()) |
kvn@3243 | 559 | cost += ConditionalMoveLimit; |
kvn@3243 | 560 | // Is there a use inside the loop? |
kvn@3243 | 561 | // Note: check only basic types since CMoveP is pinned. |
kvn@3243 | 562 | if (!used_inside_loop && is_java_primitive(bt)) { |
kvn@3243 | 563 | IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use); |
kvn@3243 | 564 | if (r_loop == u_loop || r_loop->is_member(u_loop)) { |
kvn@3243 | 565 | used_inside_loop = true; |
kvn@3243 | 566 | } |
kvn@3243 | 567 | } |
duke@435 | 568 | } |
duke@435 | 569 | } |
kvn@470 | 570 | Node* bol = iff->in(1); |
kvn@3243 | 571 | assert(bol->Opcode() == Op_Bool, ""); |
kvn@470 | 572 | int cmp_op = bol->in(1)->Opcode(); |
kvn@470 | 573 | // It is expensive to generate flags from a float compare. |
kvn@470 | 574 | // Avoid duplicated float compare. |
kvn@3243 | 575 | if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL; |
kvn@3243 | 576 | |
kvn@3243 | 577 | float infrequent_prob = PROB_UNLIKELY_MAG(3); |
kvn@3243 | 578 | // Ignore cost and blocks frequency if CMOVE can be moved outside the loop. |
kvn@3243 | 579 | if (used_inside_loop) { |
kvn@3243 | 580 | if (cost >= ConditionalMoveLimit) return NULL; // Too much goo |
kvn@3243 | 581 | |
kvn@3243 | 582 | // BlockLayoutByFrequency optimization moves infrequent branch |
kvn@3243 | 583 | // from hot path. No point in CMOV'ing in such case (110 is used |
kvn@3243 | 584 | // instead of 100 to take into account not exactness of float value). |
kvn@3243 | 585 | if (BlockLayoutByFrequency) { |
kvn@3243 | 586 | infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f); |
kvn@3243 | 587 | } |
kvn@3243 | 588 | } |
kvn@3243 | 589 | // Check for highly predictable branch. No point in CMOV'ing if |
kvn@3243 | 590 | // we are going to predict accurately all the time. |
kvn@3243 | 591 | if (iff->_prob < infrequent_prob || |
kvn@3243 | 592 | iff->_prob > (1.0f - infrequent_prob)) |
kvn@3243 | 593 | return NULL; |
duke@435 | 594 | |
duke@435 | 595 | // -------------- |
duke@435 | 596 | // Now replace all Phis with CMOV's |
duke@435 | 597 | Node *cmov_ctrl = iff->in(0); |
duke@435 | 598 | uint flip = (lp->Opcode() == Op_IfTrue); |
kvn@3243 | 599 | while (1) { |
duke@435 | 600 | PhiNode* phi = NULL; |
duke@435 | 601 | for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { |
duke@435 | 602 | Node *out = region->fast_out(i); |
duke@435 | 603 | if (out->is_Phi()) { |
duke@435 | 604 | phi = out->as_Phi(); |
duke@435 | 605 | break; |
duke@435 | 606 | } |
duke@435 | 607 | } |
duke@435 | 608 | if (phi == NULL) break; |
duke@435 | 609 | #ifndef PRODUCT |
kvn@3243 | 610 | if (PrintOpto && VerifyLoopOptimizations) tty->print_cr("CMOV"); |
duke@435 | 611 | #endif |
duke@435 | 612 | // Move speculative ops |
kvn@3243 | 613 | for (uint j = 1; j < region->req(); j++) { |
duke@435 | 614 | Node *proj = region->in(j); |
duke@435 | 615 | Node *inp = phi->in(j); |
duke@435 | 616 | if (get_ctrl(inp) == proj) { // Found local op |
duke@435 | 617 | #ifndef PRODUCT |
kvn@3243 | 618 | if (PrintOpto && VerifyLoopOptimizations) { |
duke@435 | 619 | tty->print(" speculate: "); |
duke@435 | 620 | inp->dump(); |
duke@435 | 621 | } |
duke@435 | 622 | #endif |
duke@435 | 623 | set_ctrl(inp, cmov_ctrl); |
duke@435 | 624 | } |
duke@435 | 625 | } |
duke@435 | 626 | Node *cmov = CMoveNode::make( C, cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi) ); |
duke@435 | 627 | register_new_node( cmov, cmov_ctrl ); |
kvn@1976 | 628 | _igvn.replace_node( phi, cmov ); |
duke@435 | 629 | #ifndef PRODUCT |
kvn@3243 | 630 | if (TraceLoopOpts) { |
kvn@3243 | 631 | tty->print("CMOV "); |
kvn@3243 | 632 | r_loop->dump_head(); |
kvn@3260 | 633 | if (Verbose) { |
kvn@3243 | 634 | bol->in(1)->dump(1); |
kvn@3243 | 635 | cmov->dump(1); |
kvn@3260 | 636 | } |
kvn@3243 | 637 | } |
kvn@3243 | 638 | if (VerifyLoopOptimizations) verify(); |
duke@435 | 639 | #endif |
duke@435 | 640 | } |
duke@435 | 641 | |
duke@435 | 642 | // The useless CFG diamond will fold up later; see the optimization in |
duke@435 | 643 | // RegionNode::Ideal. |
duke@435 | 644 | _igvn._worklist.push(region); |
duke@435 | 645 | |
duke@435 | 646 | return iff->in(1); |
duke@435 | 647 | } |
duke@435 | 648 | |
duke@435 | 649 | //------------------------------split_if_with_blocks_pre----------------------- |
duke@435 | 650 | // Do the real work in a non-recursive function. Data nodes want to be |
duke@435 | 651 | // cloned in the pre-order so they can feed each other nicely. |
duke@435 | 652 | Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) { |
duke@435 | 653 | // Cloning these guys is unlikely to win |
duke@435 | 654 | int n_op = n->Opcode(); |
duke@435 | 655 | if( n_op == Op_MergeMem ) return n; |
duke@435 | 656 | if( n->is_Proj() ) return n; |
duke@435 | 657 | // Do not clone-up CmpFXXX variations, as these are always |
duke@435 | 658 | // followed by a CmpI |
duke@435 | 659 | if( n->is_Cmp() ) return n; |
duke@435 | 660 | // Attempt to use a conditional move instead of a phi/branch |
duke@435 | 661 | if( ConditionalMoveLimit > 0 && n_op == Op_Region ) { |
duke@435 | 662 | Node *cmov = conditional_move( n ); |
duke@435 | 663 | if( cmov ) return cmov; |
duke@435 | 664 | } |
kvn@688 | 665 | if( n->is_CFG() || n->is_LoadStore() ) |
kvn@688 | 666 | return n; |
duke@435 | 667 | if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd |
duke@435 | 668 | n_op == Op_Opaque2 ) { |
duke@435 | 669 | if( !C->major_progress() ) // If chance of no more loop opts... |
duke@435 | 670 | _igvn._worklist.push(n); // maybe we'll remove them |
duke@435 | 671 | return n; |
duke@435 | 672 | } |
duke@435 | 673 | |
duke@435 | 674 | if( n->is_Con() ) return n; // No cloning for Con nodes |
duke@435 | 675 | |
duke@435 | 676 | Node *n_ctrl = get_ctrl(n); |
duke@435 | 677 | if( !n_ctrl ) return n; // Dead node |
duke@435 | 678 | |
duke@435 | 679 | // Attempt to remix address expressions for loop invariants |
duke@435 | 680 | Node *m = remix_address_expressions( n ); |
duke@435 | 681 | if( m ) return m; |
duke@435 | 682 | |
duke@435 | 683 | // Determine if the Node has inputs from some local Phi. |
duke@435 | 684 | // Returns the block to clone thru. |
duke@435 | 685 | Node *n_blk = has_local_phi_input( n ); |
duke@435 | 686 | if( !n_blk ) return n; |
duke@435 | 687 | // Do not clone the trip counter through on a CountedLoop |
duke@435 | 688 | // (messes up the canonical shape). |
duke@435 | 689 | if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n; |
duke@435 | 690 | |
duke@435 | 691 | // Check for having no control input; not pinned. Allow |
duke@435 | 692 | // dominating control. |
duke@435 | 693 | if( n->in(0) ) { |
duke@435 | 694 | Node *dom = idom(n_blk); |
duke@435 | 695 | if( dom_lca( n->in(0), dom ) != n->in(0) ) |
duke@435 | 696 | return n; |
duke@435 | 697 | } |
duke@435 | 698 | // Policy: when is it profitable. You must get more wins than |
duke@435 | 699 | // policy before it is considered profitable. Policy is usually 0, |
duke@435 | 700 | // so 1 win is considered profitable. Big merges will require big |
duke@435 | 701 | // cloning, so get a larger policy. |
duke@435 | 702 | int policy = n_blk->req() >> 2; |
duke@435 | 703 | |
duke@435 | 704 | // If the loop is a candidate for range check elimination, |
duke@435 | 705 | // delay splitting through it's phi until a later loop optimization |
duke@435 | 706 | if (n_blk->is_CountedLoop()) { |
duke@435 | 707 | IdealLoopTree *lp = get_loop(n_blk); |
duke@435 | 708 | if (lp && lp->_rce_candidate) { |
duke@435 | 709 | return n; |
duke@435 | 710 | } |
duke@435 | 711 | } |
duke@435 | 712 | |
duke@435 | 713 | // Use same limit as split_if_with_blocks_post |
duke@435 | 714 | if( C->unique() > 35000 ) return n; // Method too big |
duke@435 | 715 | |
duke@435 | 716 | // Split 'n' through the merge point if it is profitable |
duke@435 | 717 | Node *phi = split_thru_phi( n, n_blk, policy ); |
kvn@3243 | 718 | if (!phi) return n; |
duke@435 | 719 | |
duke@435 | 720 | // Found a Phi to split thru! |
duke@435 | 721 | // Replace 'n' with the new phi |
kvn@1976 | 722 | _igvn.replace_node( n, phi ); |
duke@435 | 723 | // Moved a load around the loop, 'en-registering' something. |
kvn@3243 | 724 | if (n_blk->is_Loop() && n->is_Load() && |
kvn@3243 | 725 | !phi->in(LoopNode::LoopBackControl)->is_Load()) |
duke@435 | 726 | C->set_major_progress(); |
duke@435 | 727 | |
duke@435 | 728 | return phi; |
duke@435 | 729 | } |
duke@435 | 730 | |
duke@435 | 731 | static bool merge_point_too_heavy(Compile* C, Node* region) { |
duke@435 | 732 | // Bail out if the region and its phis have too many users. |
duke@435 | 733 | int weight = 0; |
duke@435 | 734 | for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { |
duke@435 | 735 | weight += region->fast_out(i)->outcnt(); |
duke@435 | 736 | } |
bharadwaj@4315 | 737 | int nodes_left = MaxNodeLimit - C->live_nodes(); |
duke@435 | 738 | if (weight * 8 > nodes_left) { |
duke@435 | 739 | #ifndef PRODUCT |
duke@435 | 740 | if (PrintOpto) |
duke@435 | 741 | tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight); |
duke@435 | 742 | #endif |
duke@435 | 743 | return true; |
duke@435 | 744 | } else { |
duke@435 | 745 | return false; |
duke@435 | 746 | } |
duke@435 | 747 | } |
duke@435 | 748 | |
duke@435 | 749 | static bool merge_point_safe(Node* region) { |
duke@435 | 750 | // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode |
duke@435 | 751 | // having a PhiNode input. This sidesteps the dangerous case where the split |
duke@435 | 752 | // ConvI2LNode may become TOP if the input Value() does not |
duke@435 | 753 | // overlap the ConvI2L range, leaving a node which may not dominate its |
duke@435 | 754 | // uses. |
duke@435 | 755 | // A better fix for this problem can be found in the BugTraq entry, but |
duke@435 | 756 | // expediency for Mantis demands this hack. |
cfang@1269 | 757 | // 6855164: If the merge point has a FastLockNode with a PhiNode input, we stop |
cfang@1269 | 758 | // split_if_with_blocks from splitting a block because we could not move around |
cfang@1269 | 759 | // the FastLockNode. |
duke@435 | 760 | for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { |
duke@435 | 761 | Node* n = region->fast_out(i); |
duke@435 | 762 | if (n->is_Phi()) { |
duke@435 | 763 | for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { |
duke@435 | 764 | Node* m = n->fast_out(j); |
cfang@1269 | 765 | if (m->is_FastLock()) |
duke@435 | 766 | return false; |
cfang@1269 | 767 | #ifdef _LP64 |
cfang@1269 | 768 | if (m->Opcode() == Op_ConvI2L) |
cfang@1269 | 769 | return false; |
cfang@1269 | 770 | #endif |
duke@435 | 771 | } |
duke@435 | 772 | } |
duke@435 | 773 | } |
duke@435 | 774 | return true; |
duke@435 | 775 | } |
duke@435 | 776 | |
duke@435 | 777 | |
duke@435 | 778 | //------------------------------place_near_use--------------------------------- |
duke@435 | 779 | // Place some computation next to use but not inside inner loops. |
duke@435 | 780 | // For inner loop uses move it to the preheader area. |
duke@435 | 781 | Node *PhaseIdealLoop::place_near_use( Node *useblock ) const { |
duke@435 | 782 | IdealLoopTree *u_loop = get_loop( useblock ); |
duke@435 | 783 | return (u_loop->_irreducible || u_loop->_child) |
duke@435 | 784 | ? useblock |
duke@435 | 785 | : u_loop->_head->in(LoopNode::EntryControl); |
duke@435 | 786 | } |
duke@435 | 787 | |
duke@435 | 788 | |
duke@435 | 789 | //------------------------------split_if_with_blocks_post---------------------- |
duke@435 | 790 | // Do the real work in a non-recursive function. CFG hackery wants to be |
duke@435 | 791 | // in the post-order, so it can dirty the I-DOM info and not use the dirtied |
duke@435 | 792 | // info. |
duke@435 | 793 | void PhaseIdealLoop::split_if_with_blocks_post( Node *n ) { |
duke@435 | 794 | |
duke@435 | 795 | // Cloning Cmp through Phi's involves the split-if transform. |
duke@435 | 796 | // FastLock is not used by an If |
duke@435 | 797 | if( n->is_Cmp() && !n->is_FastLock() ) { |
duke@435 | 798 | if( C->unique() > 35000 ) return; // Method too big |
duke@435 | 799 | |
duke@435 | 800 | // Do not do 'split-if' if irreducible loops are present. |
duke@435 | 801 | if( _has_irreducible_loops ) |
duke@435 | 802 | return; |
duke@435 | 803 | |
duke@435 | 804 | Node *n_ctrl = get_ctrl(n); |
duke@435 | 805 | // Determine if the Node has inputs from some local Phi. |
duke@435 | 806 | // Returns the block to clone thru. |
duke@435 | 807 | Node *n_blk = has_local_phi_input( n ); |
duke@435 | 808 | if( n_blk != n_ctrl ) return; |
duke@435 | 809 | |
duke@435 | 810 | if( merge_point_too_heavy(C, n_ctrl) ) |
duke@435 | 811 | return; |
duke@435 | 812 | |
duke@435 | 813 | if( n->outcnt() != 1 ) return; // Multiple bool's from 1 compare? |
duke@435 | 814 | Node *bol = n->unique_out(); |
duke@435 | 815 | assert( bol->is_Bool(), "expect a bool here" ); |
duke@435 | 816 | if( bol->outcnt() != 1 ) return;// Multiple branches from 1 compare? |
duke@435 | 817 | Node *iff = bol->unique_out(); |
duke@435 | 818 | |
duke@435 | 819 | // Check some safety conditions |
duke@435 | 820 | if( iff->is_If() ) { // Classic split-if? |
duke@435 | 821 | if( iff->in(0) != n_ctrl ) return; // Compare must be in same blk as if |
duke@435 | 822 | } else if (iff->is_CMove()) { // Trying to split-up a CMOVE |
kvn@3408 | 823 | // Can't split CMove with different control edge. |
kvn@3408 | 824 | if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) return; |
duke@435 | 825 | if( get_ctrl(iff->in(2)) == n_ctrl || |
duke@435 | 826 | get_ctrl(iff->in(3)) == n_ctrl ) |
duke@435 | 827 | return; // Inputs not yet split-up |
duke@435 | 828 | if ( get_loop(n_ctrl) != get_loop(get_ctrl(iff)) ) { |
duke@435 | 829 | return; // Loop-invar test gates loop-varying CMOVE |
duke@435 | 830 | } |
duke@435 | 831 | } else { |
duke@435 | 832 | return; // some other kind of node, such as an Allocate |
duke@435 | 833 | } |
duke@435 | 834 | |
duke@435 | 835 | // Do not do 'split-if' if some paths are dead. First do dead code |
duke@435 | 836 | // elimination and then see if its still profitable. |
duke@435 | 837 | for( uint i = 1; i < n_ctrl->req(); i++ ) |
duke@435 | 838 | if( n_ctrl->in(i) == C->top() ) |
duke@435 | 839 | return; |
duke@435 | 840 | |
duke@435 | 841 | // When is split-if profitable? Every 'win' on means some control flow |
duke@435 | 842 | // goes dead, so it's almost always a win. |
duke@435 | 843 | int policy = 0; |
duke@435 | 844 | // If trying to do a 'Split-If' at the loop head, it is only |
duke@435 | 845 | // profitable if the cmp folds up on BOTH paths. Otherwise we |
duke@435 | 846 | // risk peeling a loop forever. |
duke@435 | 847 | |
duke@435 | 848 | // CNC - Disabled for now. Requires careful handling of loop |
duke@435 | 849 | // body selection for the cloned code. Also, make sure we check |
duke@435 | 850 | // for any input path not being in the same loop as n_ctrl. For |
duke@435 | 851 | // irreducible loops we cannot check for 'n_ctrl->is_Loop()' |
duke@435 | 852 | // because the alternative loop entry points won't be converted |
duke@435 | 853 | // into LoopNodes. |
duke@435 | 854 | IdealLoopTree *n_loop = get_loop(n_ctrl); |
duke@435 | 855 | for( uint j = 1; j < n_ctrl->req(); j++ ) |
duke@435 | 856 | if( get_loop(n_ctrl->in(j)) != n_loop ) |
duke@435 | 857 | return; |
duke@435 | 858 | |
duke@435 | 859 | // Check for safety of the merge point. |
duke@435 | 860 | if( !merge_point_safe(n_ctrl) ) { |
duke@435 | 861 | return; |
duke@435 | 862 | } |
duke@435 | 863 | |
duke@435 | 864 | // Split compare 'n' through the merge point if it is profitable |
duke@435 | 865 | Node *phi = split_thru_phi( n, n_ctrl, policy ); |
duke@435 | 866 | if( !phi ) return; |
duke@435 | 867 | |
duke@435 | 868 | // Found a Phi to split thru! |
duke@435 | 869 | // Replace 'n' with the new phi |
kvn@1976 | 870 | _igvn.replace_node( n, phi ); |
duke@435 | 871 | |
duke@435 | 872 | // Now split the bool up thru the phi |
duke@435 | 873 | Node *bolphi = split_thru_phi( bol, n_ctrl, -1 ); |
morris@4774 | 874 | guarantee(bolphi != NULL, "null boolean phi node"); |
morris@4774 | 875 | |
kvn@1976 | 876 | _igvn.replace_node( bol, bolphi ); |
duke@435 | 877 | assert( iff->in(1) == bolphi, "" ); |
morris@4774 | 878 | |
duke@435 | 879 | if( bolphi->Value(&_igvn)->singleton() ) |
duke@435 | 880 | return; |
duke@435 | 881 | |
duke@435 | 882 | // Conditional-move? Must split up now |
duke@435 | 883 | if( !iff->is_If() ) { |
duke@435 | 884 | Node *cmovphi = split_thru_phi( iff, n_ctrl, -1 ); |
kvn@1976 | 885 | _igvn.replace_node( iff, cmovphi ); |
duke@435 | 886 | return; |
duke@435 | 887 | } |
duke@435 | 888 | |
duke@435 | 889 | // Now split the IF |
duke@435 | 890 | do_split_if( iff ); |
duke@435 | 891 | return; |
duke@435 | 892 | } |
duke@435 | 893 | |
duke@435 | 894 | // Check for an IF ready to split; one that has its |
duke@435 | 895 | // condition codes input coming from a Phi at the block start. |
duke@435 | 896 | int n_op = n->Opcode(); |
duke@435 | 897 | |
duke@435 | 898 | // Check for an IF being dominated by another IF same test |
duke@435 | 899 | if( n_op == Op_If ) { |
duke@435 | 900 | Node *bol = n->in(1); |
duke@435 | 901 | uint max = bol->outcnt(); |
duke@435 | 902 | // Check for same test used more than once? |
duke@435 | 903 | if( n_op == Op_If && max > 1 && bol->is_Bool() ) { |
duke@435 | 904 | // Search up IDOMs to see if this IF is dominated. |
duke@435 | 905 | Node *cutoff = get_ctrl(bol); |
duke@435 | 906 | |
duke@435 | 907 | // Now search up IDOMs till cutoff, looking for a dominating test |
duke@435 | 908 | Node *prevdom = n; |
duke@435 | 909 | Node *dom = idom(prevdom); |
duke@435 | 910 | while( dom != cutoff ) { |
duke@435 | 911 | if( dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom ) { |
duke@435 | 912 | // Replace the dominated test with an obvious true or false. |
duke@435 | 913 | // Place it on the IGVN worklist for later cleanup. |
duke@435 | 914 | C->set_major_progress(); |
kvn@3038 | 915 | dominated_by( prevdom, n, false, true ); |
duke@435 | 916 | #ifndef PRODUCT |
duke@435 | 917 | if( VerifyLoopOptimizations ) verify(); |
duke@435 | 918 | #endif |
duke@435 | 919 | return; |
duke@435 | 920 | } |
duke@435 | 921 | prevdom = dom; |
duke@435 | 922 | dom = idom(prevdom); |
duke@435 | 923 | } |
duke@435 | 924 | } |
duke@435 | 925 | } |
duke@435 | 926 | |
duke@435 | 927 | // See if a shared loop-varying computation has no loop-varying uses. |
duke@435 | 928 | // Happens if something is only used for JVM state in uncommon trap exits, |
duke@435 | 929 | // like various versions of induction variable+offset. Clone the |
duke@435 | 930 | // computation per usage to allow it to sink out of the loop. |
duke@435 | 931 | if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about) |
duke@435 | 932 | Node *n_ctrl = get_ctrl(n); |
duke@435 | 933 | IdealLoopTree *n_loop = get_loop(n_ctrl); |
duke@435 | 934 | if( n_loop != _ltree_root ) { |
duke@435 | 935 | DUIterator_Fast imax, i = n->fast_outs(imax); |
duke@435 | 936 | for (; i < imax; i++) { |
duke@435 | 937 | Node* u = n->fast_out(i); |
duke@435 | 938 | if( !has_ctrl(u) ) break; // Found control user |
duke@435 | 939 | IdealLoopTree *u_loop = get_loop(get_ctrl(u)); |
duke@435 | 940 | if( u_loop == n_loop ) break; // Found loop-varying use |
duke@435 | 941 | if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop |
duke@435 | 942 | if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003 |
duke@435 | 943 | } |
duke@435 | 944 | bool did_break = (i < imax); // Did we break out of the previous loop? |
duke@435 | 945 | if (!did_break && n->outcnt() > 1) { // All uses in outer loops! |
kvn@3408 | 946 | Node *late_load_ctrl = NULL; |
duke@435 | 947 | if (n->is_Load()) { |
duke@435 | 948 | // If n is a load, get and save the result from get_late_ctrl(), |
duke@435 | 949 | // to be later used in calculating the control for n's clones. |
duke@435 | 950 | clear_dom_lca_tags(); |
duke@435 | 951 | late_load_ctrl = get_late_ctrl(n, n_ctrl); |
duke@435 | 952 | } |
duke@435 | 953 | // If n is a load, and the late control is the same as the current |
duke@435 | 954 | // control, then the cloning of n is a pointless exercise, because |
duke@435 | 955 | // GVN will ensure that we end up where we started. |
duke@435 | 956 | if (!n->is_Load() || late_load_ctrl != n_ctrl) { |
duke@435 | 957 | for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) { |
duke@435 | 958 | Node *u = n->last_out(j); // Clone private computation per use |
kvn@3847 | 959 | _igvn.rehash_node_delayed(u); |
duke@435 | 960 | Node *x = n->clone(); // Clone computation |
duke@435 | 961 | Node *x_ctrl = NULL; |
duke@435 | 962 | if( u->is_Phi() ) { |
duke@435 | 963 | // Replace all uses of normal nodes. Replace Phi uses |
twisti@1040 | 964 | // individually, so the separate Nodes can sink down |
duke@435 | 965 | // different paths. |
duke@435 | 966 | uint k = 1; |
duke@435 | 967 | while( u->in(k) != n ) k++; |
duke@435 | 968 | u->set_req( k, x ); |
duke@435 | 969 | // x goes next to Phi input path |
duke@435 | 970 | x_ctrl = u->in(0)->in(k); |
duke@435 | 971 | --j; |
duke@435 | 972 | } else { // Normal use |
duke@435 | 973 | // Replace all uses |
duke@435 | 974 | for( uint k = 0; k < u->req(); k++ ) { |
duke@435 | 975 | if( u->in(k) == n ) { |
duke@435 | 976 | u->set_req( k, x ); |
duke@435 | 977 | --j; |
duke@435 | 978 | } |
duke@435 | 979 | } |
duke@435 | 980 | x_ctrl = get_ctrl(u); |
duke@435 | 981 | } |
duke@435 | 982 | |
duke@435 | 983 | // Find control for 'x' next to use but not inside inner loops. |
duke@435 | 984 | // For inner loop uses get the preheader area. |
duke@435 | 985 | x_ctrl = place_near_use(x_ctrl); |
duke@435 | 986 | |
duke@435 | 987 | if (n->is_Load()) { |
duke@435 | 988 | // For loads, add a control edge to a CFG node outside of the loop |
duke@435 | 989 | // to force them to not combine and return back inside the loop |
duke@435 | 990 | // during GVN optimization (4641526). |
duke@435 | 991 | // |
duke@435 | 992 | // Because we are setting the actual control input, factor in |
duke@435 | 993 | // the result from get_late_ctrl() so we respect any |
duke@435 | 994 | // anti-dependences. (6233005). |
duke@435 | 995 | x_ctrl = dom_lca(late_load_ctrl, x_ctrl); |
duke@435 | 996 | |
duke@435 | 997 | // Don't allow the control input to be a CFG splitting node. |
duke@435 | 998 | // Such nodes should only have ProjNodes as outs, e.g. IfNode |
duke@435 | 999 | // should only have IfTrueNode and IfFalseNode (4985384). |
duke@435 | 1000 | x_ctrl = find_non_split_ctrl(x_ctrl); |
duke@435 | 1001 | assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone"); |
duke@435 | 1002 | |
duke@435 | 1003 | x->set_req(0, x_ctrl); |
duke@435 | 1004 | } |
duke@435 | 1005 | register_new_node(x, x_ctrl); |
duke@435 | 1006 | |
duke@435 | 1007 | // Some institutional knowledge is needed here: 'x' is |
duke@435 | 1008 | // yanked because if the optimizer runs GVN on it all the |
duke@435 | 1009 | // cloned x's will common up and undo this optimization and |
duke@435 | 1010 | // be forced back in the loop. This is annoying because it |
duke@435 | 1011 | // makes +VerifyOpto report false-positives on progress. I |
duke@435 | 1012 | // tried setting control edges on the x's to force them to |
duke@435 | 1013 | // not combine, but the matching gets worried when it tries |
duke@435 | 1014 | // to fold a StoreP and an AddP together (as part of an |
duke@435 | 1015 | // address expression) and the AddP and StoreP have |
duke@435 | 1016 | // different controls. |
roland@4159 | 1017 | if (!x->is_Load() && !x->is_DecodeNarrowPtr()) _igvn._worklist.yank(x); |
duke@435 | 1018 | } |
duke@435 | 1019 | _igvn.remove_dead_node(n); |
duke@435 | 1020 | } |
duke@435 | 1021 | } |
duke@435 | 1022 | } |
duke@435 | 1023 | } |
duke@435 | 1024 | |
duke@435 | 1025 | // Check for Opaque2's who's loop has disappeared - who's input is in the |
duke@435 | 1026 | // same loop nest as their output. Remove 'em, they are no longer useful. |
duke@435 | 1027 | if( n_op == Op_Opaque2 && |
duke@435 | 1028 | n->in(1) != NULL && |
duke@435 | 1029 | get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) { |
kvn@1976 | 1030 | _igvn.replace_node( n, n->in(1) ); |
duke@435 | 1031 | } |
duke@435 | 1032 | } |
duke@435 | 1033 | |
duke@435 | 1034 | //------------------------------split_if_with_blocks--------------------------- |
duke@435 | 1035 | // Check for aggressive application of 'split-if' optimization, |
duke@435 | 1036 | // using basic block level info. |
duke@435 | 1037 | void PhaseIdealLoop::split_if_with_blocks( VectorSet &visited, Node_Stack &nstack ) { |
duke@435 | 1038 | Node *n = C->root(); |
duke@435 | 1039 | visited.set(n->_idx); // first, mark node as visited |
duke@435 | 1040 | // Do pre-visit work for root |
duke@435 | 1041 | n = split_if_with_blocks_pre( n ); |
duke@435 | 1042 | uint cnt = n->outcnt(); |
duke@435 | 1043 | uint i = 0; |
duke@435 | 1044 | while (true) { |
duke@435 | 1045 | // Visit all children |
duke@435 | 1046 | if (i < cnt) { |
duke@435 | 1047 | Node* use = n->raw_out(i); |
duke@435 | 1048 | ++i; |
duke@435 | 1049 | if (use->outcnt() != 0 && !visited.test_set(use->_idx)) { |
duke@435 | 1050 | // Now do pre-visit work for this use |
duke@435 | 1051 | use = split_if_with_blocks_pre( use ); |
duke@435 | 1052 | nstack.push(n, i); // Save parent and next use's index. |
duke@435 | 1053 | n = use; // Process all children of current use. |
duke@435 | 1054 | cnt = use->outcnt(); |
duke@435 | 1055 | i = 0; |
duke@435 | 1056 | } |
duke@435 | 1057 | } |
duke@435 | 1058 | else { |
duke@435 | 1059 | // All of n's children have been processed, complete post-processing. |
duke@435 | 1060 | if (cnt != 0 && !n->is_Con()) { |
duke@435 | 1061 | assert(has_node(n), "no dead nodes"); |
duke@435 | 1062 | split_if_with_blocks_post( n ); |
duke@435 | 1063 | } |
duke@435 | 1064 | if (nstack.is_empty()) { |
duke@435 | 1065 | // Finished all nodes on stack. |
duke@435 | 1066 | break; |
duke@435 | 1067 | } |
duke@435 | 1068 | // Get saved parent node and next use's index. Visit the rest of uses. |
duke@435 | 1069 | n = nstack.node(); |
duke@435 | 1070 | cnt = n->outcnt(); |
duke@435 | 1071 | i = nstack.index(); |
duke@435 | 1072 | nstack.pop(); |
duke@435 | 1073 | } |
duke@435 | 1074 | } |
duke@435 | 1075 | } |
duke@435 | 1076 | |
duke@435 | 1077 | |
duke@435 | 1078 | //============================================================================= |
duke@435 | 1079 | // |
duke@435 | 1080 | // C L O N E A L O O P B O D Y |
duke@435 | 1081 | // |
duke@435 | 1082 | |
duke@435 | 1083 | //------------------------------clone_iff-------------------------------------- |
duke@435 | 1084 | // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. |
duke@435 | 1085 | // "Nearly" because all Nodes have been cloned from the original in the loop, |
duke@435 | 1086 | // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs |
duke@435 | 1087 | // through the Phi recursively, and return a Bool. |
duke@435 | 1088 | BoolNode *PhaseIdealLoop::clone_iff( PhiNode *phi, IdealLoopTree *loop ) { |
duke@435 | 1089 | |
duke@435 | 1090 | // Convert this Phi into a Phi merging Bools |
duke@435 | 1091 | uint i; |
duke@435 | 1092 | for( i = 1; i < phi->req(); i++ ) { |
duke@435 | 1093 | Node *b = phi->in(i); |
duke@435 | 1094 | if( b->is_Phi() ) { |
kvn@3847 | 1095 | _igvn.replace_input_of(phi, i, clone_iff( b->as_Phi(), loop )); |
duke@435 | 1096 | } else { |
duke@435 | 1097 | assert( b->is_Bool(), "" ); |
duke@435 | 1098 | } |
duke@435 | 1099 | } |
duke@435 | 1100 | |
duke@435 | 1101 | Node *sample_bool = phi->in(1); |
duke@435 | 1102 | Node *sample_cmp = sample_bool->in(1); |
duke@435 | 1103 | |
duke@435 | 1104 | // Make Phis to merge the Cmp's inputs. |
kvn@4115 | 1105 | PhiNode *phi1 = new (C) PhiNode( phi->in(0), Type::TOP ); |
kvn@4115 | 1106 | PhiNode *phi2 = new (C) PhiNode( phi->in(0), Type::TOP ); |
duke@435 | 1107 | for( i = 1; i < phi->req(); i++ ) { |
duke@435 | 1108 | Node *n1 = phi->in(i)->in(1)->in(1); |
duke@435 | 1109 | Node *n2 = phi->in(i)->in(1)->in(2); |
duke@435 | 1110 | phi1->set_req( i, n1 ); |
duke@435 | 1111 | phi2->set_req( i, n2 ); |
roland@6313 | 1112 | phi1->set_type( phi1->type()->meet_speculative(n1->bottom_type())); |
roland@6313 | 1113 | phi2->set_type( phi2->type()->meet_speculative(n2->bottom_type())); |
duke@435 | 1114 | } |
duke@435 | 1115 | // See if these Phis have been made before. |
duke@435 | 1116 | // Register with optimizer |
duke@435 | 1117 | Node *hit1 = _igvn.hash_find_insert(phi1); |
duke@435 | 1118 | if( hit1 ) { // Hit, toss just made Phi |
duke@435 | 1119 | _igvn.remove_dead_node(phi1); // Remove new phi |
duke@435 | 1120 | assert( hit1->is_Phi(), "" ); |
duke@435 | 1121 | phi1 = (PhiNode*)hit1; // Use existing phi |
duke@435 | 1122 | } else { // Miss |
duke@435 | 1123 | _igvn.register_new_node_with_optimizer(phi1); |
duke@435 | 1124 | } |
duke@435 | 1125 | Node *hit2 = _igvn.hash_find_insert(phi2); |
duke@435 | 1126 | if( hit2 ) { // Hit, toss just made Phi |
duke@435 | 1127 | _igvn.remove_dead_node(phi2); // Remove new phi |
duke@435 | 1128 | assert( hit2->is_Phi(), "" ); |
duke@435 | 1129 | phi2 = (PhiNode*)hit2; // Use existing phi |
duke@435 | 1130 | } else { // Miss |
duke@435 | 1131 | _igvn.register_new_node_with_optimizer(phi2); |
duke@435 | 1132 | } |
duke@435 | 1133 | // Register Phis with loop/block info |
duke@435 | 1134 | set_ctrl(phi1, phi->in(0)); |
duke@435 | 1135 | set_ctrl(phi2, phi->in(0)); |
duke@435 | 1136 | // Make a new Cmp |
duke@435 | 1137 | Node *cmp = sample_cmp->clone(); |
duke@435 | 1138 | cmp->set_req( 1, phi1 ); |
duke@435 | 1139 | cmp->set_req( 2, phi2 ); |
duke@435 | 1140 | _igvn.register_new_node_with_optimizer(cmp); |
duke@435 | 1141 | set_ctrl(cmp, phi->in(0)); |
duke@435 | 1142 | |
duke@435 | 1143 | // Make a new Bool |
duke@435 | 1144 | Node *b = sample_bool->clone(); |
duke@435 | 1145 | b->set_req(1,cmp); |
duke@435 | 1146 | _igvn.register_new_node_with_optimizer(b); |
duke@435 | 1147 | set_ctrl(b, phi->in(0)); |
duke@435 | 1148 | |
duke@435 | 1149 | assert( b->is_Bool(), "" ); |
duke@435 | 1150 | return (BoolNode*)b; |
duke@435 | 1151 | } |
duke@435 | 1152 | |
duke@435 | 1153 | //------------------------------clone_bool------------------------------------- |
duke@435 | 1154 | // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. |
duke@435 | 1155 | // "Nearly" because all Nodes have been cloned from the original in the loop, |
duke@435 | 1156 | // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs |
duke@435 | 1157 | // through the Phi recursively, and return a Bool. |
duke@435 | 1158 | CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) { |
duke@435 | 1159 | uint i; |
duke@435 | 1160 | // Convert this Phi into a Phi merging Bools |
duke@435 | 1161 | for( i = 1; i < phi->req(); i++ ) { |
duke@435 | 1162 | Node *b = phi->in(i); |
duke@435 | 1163 | if( b->is_Phi() ) { |
kvn@3847 | 1164 | _igvn.replace_input_of(phi, i, clone_bool( b->as_Phi(), loop )); |
duke@435 | 1165 | } else { |
duke@435 | 1166 | assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" ); |
duke@435 | 1167 | } |
duke@435 | 1168 | } |
duke@435 | 1169 | |
duke@435 | 1170 | Node *sample_cmp = phi->in(1); |
duke@435 | 1171 | |
duke@435 | 1172 | // Make Phis to merge the Cmp's inputs. |
kvn@4115 | 1173 | PhiNode *phi1 = new (C) PhiNode( phi->in(0), Type::TOP ); |
kvn@4115 | 1174 | PhiNode *phi2 = new (C) PhiNode( phi->in(0), Type::TOP ); |
duke@435 | 1175 | for( uint j = 1; j < phi->req(); j++ ) { |
duke@435 | 1176 | Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP |
duke@435 | 1177 | Node *n1, *n2; |
duke@435 | 1178 | if( cmp_top->is_Cmp() ) { |
duke@435 | 1179 | n1 = cmp_top->in(1); |
duke@435 | 1180 | n2 = cmp_top->in(2); |
duke@435 | 1181 | } else { |
duke@435 | 1182 | n1 = n2 = cmp_top; |
duke@435 | 1183 | } |
duke@435 | 1184 | phi1->set_req( j, n1 ); |
duke@435 | 1185 | phi2->set_req( j, n2 ); |
roland@6313 | 1186 | phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type())); |
roland@6313 | 1187 | phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type())); |
duke@435 | 1188 | } |
duke@435 | 1189 | |
duke@435 | 1190 | // See if these Phis have been made before. |
duke@435 | 1191 | // Register with optimizer |
duke@435 | 1192 | Node *hit1 = _igvn.hash_find_insert(phi1); |
duke@435 | 1193 | if( hit1 ) { // Hit, toss just made Phi |
duke@435 | 1194 | _igvn.remove_dead_node(phi1); // Remove new phi |
duke@435 | 1195 | assert( hit1->is_Phi(), "" ); |
duke@435 | 1196 | phi1 = (PhiNode*)hit1; // Use existing phi |
duke@435 | 1197 | } else { // Miss |
duke@435 | 1198 | _igvn.register_new_node_with_optimizer(phi1); |
duke@435 | 1199 | } |
duke@435 | 1200 | Node *hit2 = _igvn.hash_find_insert(phi2); |
duke@435 | 1201 | if( hit2 ) { // Hit, toss just made Phi |
duke@435 | 1202 | _igvn.remove_dead_node(phi2); // Remove new phi |
duke@435 | 1203 | assert( hit2->is_Phi(), "" ); |
duke@435 | 1204 | phi2 = (PhiNode*)hit2; // Use existing phi |
duke@435 | 1205 | } else { // Miss |
duke@435 | 1206 | _igvn.register_new_node_with_optimizer(phi2); |
duke@435 | 1207 | } |
duke@435 | 1208 | // Register Phis with loop/block info |
duke@435 | 1209 | set_ctrl(phi1, phi->in(0)); |
duke@435 | 1210 | set_ctrl(phi2, phi->in(0)); |
duke@435 | 1211 | // Make a new Cmp |
duke@435 | 1212 | Node *cmp = sample_cmp->clone(); |
duke@435 | 1213 | cmp->set_req( 1, phi1 ); |
duke@435 | 1214 | cmp->set_req( 2, phi2 ); |
duke@435 | 1215 | _igvn.register_new_node_with_optimizer(cmp); |
duke@435 | 1216 | set_ctrl(cmp, phi->in(0)); |
duke@435 | 1217 | |
duke@435 | 1218 | assert( cmp->is_Cmp(), "" ); |
duke@435 | 1219 | return (CmpNode*)cmp; |
duke@435 | 1220 | } |
duke@435 | 1221 | |
duke@435 | 1222 | //------------------------------sink_use--------------------------------------- |
duke@435 | 1223 | // If 'use' was in the loop-exit block, it now needs to be sunk |
duke@435 | 1224 | // below the post-loop merge point. |
duke@435 | 1225 | void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) { |
duke@435 | 1226 | if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) { |
duke@435 | 1227 | set_ctrl(use, post_loop); |
duke@435 | 1228 | for (DUIterator j = use->outs(); use->has_out(j); j++) |
duke@435 | 1229 | sink_use(use->out(j), post_loop); |
duke@435 | 1230 | } |
duke@435 | 1231 | } |
duke@435 | 1232 | |
duke@435 | 1233 | //------------------------------clone_loop------------------------------------- |
duke@435 | 1234 | // |
duke@435 | 1235 | // C L O N E A L O O P B O D Y |
duke@435 | 1236 | // |
duke@435 | 1237 | // This is the basic building block of the loop optimizations. It clones an |
duke@435 | 1238 | // entire loop body. It makes an old_new loop body mapping; with this mapping |
duke@435 | 1239 | // you can find the new-loop equivalent to an old-loop node. All new-loop |
duke@435 | 1240 | // nodes are exactly equal to their old-loop counterparts, all edges are the |
duke@435 | 1241 | // same. All exits from the old-loop now have a RegionNode that merges the |
duke@435 | 1242 | // equivalent new-loop path. This is true even for the normal "loop-exit" |
duke@435 | 1243 | // condition. All uses of loop-invariant old-loop values now come from (one |
duke@435 | 1244 | // or more) Phis that merge their new-loop equivalents. |
duke@435 | 1245 | // |
duke@435 | 1246 | // This operation leaves the graph in an illegal state: there are two valid |
duke@435 | 1247 | // control edges coming from the loop pre-header to both loop bodies. I'll |
duke@435 | 1248 | // definitely have to hack the graph after running this transform. |
duke@435 | 1249 | // |
duke@435 | 1250 | // From this building block I will further edit edges to perform loop peeling |
duke@435 | 1251 | // or loop unrolling or iteration splitting (Range-Check-Elimination), etc. |
duke@435 | 1252 | // |
duke@435 | 1253 | // Parameter side_by_size_idom: |
duke@435 | 1254 | // When side_by_size_idom is NULL, the dominator tree is constructed for |
duke@435 | 1255 | // the clone loop to dominate the original. Used in construction of |
duke@435 | 1256 | // pre-main-post loop sequence. |
duke@435 | 1257 | // When nonnull, the clone and original are side-by-side, both are |
duke@435 | 1258 | // dominated by the side_by_side_idom node. Used in construction of |
duke@435 | 1259 | // unswitched loops. |
duke@435 | 1260 | void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd, |
duke@435 | 1261 | Node* side_by_side_idom) { |
duke@435 | 1262 | |
duke@435 | 1263 | // Step 1: Clone the loop body. Make the old->new mapping. |
duke@435 | 1264 | uint i; |
duke@435 | 1265 | for( i = 0; i < loop->_body.size(); i++ ) { |
duke@435 | 1266 | Node *old = loop->_body.at(i); |
duke@435 | 1267 | Node *nnn = old->clone(); |
duke@435 | 1268 | old_new.map( old->_idx, nnn ); |
duke@435 | 1269 | _igvn.register_new_node_with_optimizer(nnn); |
duke@435 | 1270 | } |
duke@435 | 1271 | |
duke@435 | 1272 | |
duke@435 | 1273 | // Step 2: Fix the edges in the new body. If the old input is outside the |
duke@435 | 1274 | // loop use it. If the old input is INside the loop, use the corresponding |
duke@435 | 1275 | // new node instead. |
duke@435 | 1276 | for( i = 0; i < loop->_body.size(); i++ ) { |
duke@435 | 1277 | Node *old = loop->_body.at(i); |
duke@435 | 1278 | Node *nnn = old_new[old->_idx]; |
duke@435 | 1279 | // Fix CFG/Loop controlling the new node |
duke@435 | 1280 | if (has_ctrl(old)) { |
duke@435 | 1281 | set_ctrl(nnn, old_new[get_ctrl(old)->_idx]); |
duke@435 | 1282 | } else { |
duke@435 | 1283 | set_loop(nnn, loop->_parent); |
duke@435 | 1284 | if (old->outcnt() > 0) { |
duke@435 | 1285 | set_idom( nnn, old_new[idom(old)->_idx], dd ); |
duke@435 | 1286 | } |
duke@435 | 1287 | } |
duke@435 | 1288 | // Correct edges to the new node |
duke@435 | 1289 | for( uint j = 0; j < nnn->req(); j++ ) { |
duke@435 | 1290 | Node *n = nnn->in(j); |
duke@435 | 1291 | if( n ) { |
duke@435 | 1292 | IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n ); |
duke@435 | 1293 | if( loop->is_member( old_in_loop ) ) |
duke@435 | 1294 | nnn->set_req(j, old_new[n->_idx]); |
duke@435 | 1295 | } |
duke@435 | 1296 | } |
duke@435 | 1297 | _igvn.hash_find_insert(nnn); |
duke@435 | 1298 | } |
duke@435 | 1299 | Node *newhead = old_new[loop->_head->_idx]; |
duke@435 | 1300 | set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); |
duke@435 | 1301 | |
duke@435 | 1302 | |
duke@435 | 1303 | // Step 3: Now fix control uses. Loop varying control uses have already |
duke@435 | 1304 | // been fixed up (as part of all input edges in Step 2). Loop invariant |
duke@435 | 1305 | // control uses must be either an IfFalse or an IfTrue. Make a merge |
duke@435 | 1306 | // point to merge the old and new IfFalse/IfTrue nodes; make the use |
duke@435 | 1307 | // refer to this. |
duke@435 | 1308 | ResourceArea *area = Thread::current()->resource_area(); |
duke@435 | 1309 | Node_List worklist(area); |
duke@435 | 1310 | uint new_counter = C->unique(); |
duke@435 | 1311 | for( i = 0; i < loop->_body.size(); i++ ) { |
duke@435 | 1312 | Node* old = loop->_body.at(i); |
duke@435 | 1313 | if( !old->is_CFG() ) continue; |
duke@435 | 1314 | Node* nnn = old_new[old->_idx]; |
duke@435 | 1315 | |
duke@435 | 1316 | // Copy uses to a worklist, so I can munge the def-use info |
duke@435 | 1317 | // with impunity. |
duke@435 | 1318 | for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) |
duke@435 | 1319 | worklist.push(old->fast_out(j)); |
duke@435 | 1320 | |
duke@435 | 1321 | while( worklist.size() ) { // Visit all uses |
duke@435 | 1322 | Node *use = worklist.pop(); |
duke@435 | 1323 | if (!has_node(use)) continue; // Ignore dead nodes |
duke@435 | 1324 | IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); |
duke@435 | 1325 | if( !loop->is_member( use_loop ) && use->is_CFG() ) { |
duke@435 | 1326 | // Both OLD and USE are CFG nodes here. |
duke@435 | 1327 | assert( use->is_Proj(), "" ); |
duke@435 | 1328 | |
duke@435 | 1329 | // Clone the loop exit control projection |
duke@435 | 1330 | Node *newuse = use->clone(); |
duke@435 | 1331 | newuse->set_req(0,nnn); |
duke@435 | 1332 | _igvn.register_new_node_with_optimizer(newuse); |
duke@435 | 1333 | set_loop(newuse, use_loop); |
duke@435 | 1334 | set_idom(newuse, nnn, dom_depth(nnn) + 1 ); |
duke@435 | 1335 | |
duke@435 | 1336 | // We need a Region to merge the exit from the peeled body and the |
duke@435 | 1337 | // exit from the old loop body. |
kvn@4115 | 1338 | RegionNode *r = new (C) RegionNode(3); |
duke@435 | 1339 | // Map the old use to the new merge point |
duke@435 | 1340 | old_new.map( use->_idx, r ); |
duke@435 | 1341 | uint dd_r = MIN2(dom_depth(newuse),dom_depth(use)); |
duke@435 | 1342 | assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" ); |
duke@435 | 1343 | |
duke@435 | 1344 | // The original user of 'use' uses 'r' instead. |
duke@435 | 1345 | for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) { |
duke@435 | 1346 | Node* useuse = use->last_out(l); |
kvn@3847 | 1347 | _igvn.rehash_node_delayed(useuse); |
duke@435 | 1348 | uint uses_found = 0; |
duke@435 | 1349 | if( useuse->in(0) == use ) { |
duke@435 | 1350 | useuse->set_req(0, r); |
duke@435 | 1351 | uses_found++; |
duke@435 | 1352 | if( useuse->is_CFG() ) { |
duke@435 | 1353 | assert( dom_depth(useuse) > dd_r, "" ); |
duke@435 | 1354 | set_idom(useuse, r, dom_depth(useuse)); |
duke@435 | 1355 | } |
duke@435 | 1356 | } |
duke@435 | 1357 | for( uint k = 1; k < useuse->req(); k++ ) { |
duke@435 | 1358 | if( useuse->in(k) == use ) { |
duke@435 | 1359 | useuse->set_req(k, r); |
duke@435 | 1360 | uses_found++; |
duke@435 | 1361 | } |
duke@435 | 1362 | } |
duke@435 | 1363 | l -= uses_found; // we deleted 1 or more copies of this edge |
duke@435 | 1364 | } |
duke@435 | 1365 | |
duke@435 | 1366 | // Now finish up 'r' |
duke@435 | 1367 | r->set_req( 1, newuse ); |
duke@435 | 1368 | r->set_req( 2, use ); |
duke@435 | 1369 | _igvn.register_new_node_with_optimizer(r); |
duke@435 | 1370 | set_loop(r, use_loop); |
duke@435 | 1371 | set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r); |
duke@435 | 1372 | } // End of if a loop-exit test |
duke@435 | 1373 | } |
duke@435 | 1374 | } |
duke@435 | 1375 | |
duke@435 | 1376 | // Step 4: If loop-invariant use is not control, it must be dominated by a |
duke@435 | 1377 | // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region |
duke@435 | 1378 | // there if needed. Make a Phi there merging old and new used values. |
duke@435 | 1379 | Node_List *split_if_set = NULL; |
duke@435 | 1380 | Node_List *split_bool_set = NULL; |
duke@435 | 1381 | Node_List *split_cex_set = NULL; |
duke@435 | 1382 | for( i = 0; i < loop->_body.size(); i++ ) { |
duke@435 | 1383 | Node* old = loop->_body.at(i); |
duke@435 | 1384 | Node* nnn = old_new[old->_idx]; |
duke@435 | 1385 | // Copy uses to a worklist, so I can munge the def-use info |
duke@435 | 1386 | // with impunity. |
duke@435 | 1387 | for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) |
duke@435 | 1388 | worklist.push(old->fast_out(j)); |
duke@435 | 1389 | |
duke@435 | 1390 | while( worklist.size() ) { |
duke@435 | 1391 | Node *use = worklist.pop(); |
duke@435 | 1392 | if (!has_node(use)) continue; // Ignore dead nodes |
duke@435 | 1393 | if (use->in(0) == C->top()) continue; |
duke@435 | 1394 | IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); |
duke@435 | 1395 | // Check for data-use outside of loop - at least one of OLD or USE |
duke@435 | 1396 | // must not be a CFG node. |
duke@435 | 1397 | if( !loop->is_member( use_loop ) && (!old->is_CFG() || !use->is_CFG())) { |
duke@435 | 1398 | |
duke@435 | 1399 | // If the Data use is an IF, that means we have an IF outside of the |
duke@435 | 1400 | // loop that is switching on a condition that is set inside of the |
duke@435 | 1401 | // loop. Happens if people set a loop-exit flag; then test the flag |
duke@435 | 1402 | // in the loop to break the loop, then test is again outside of the |
duke@435 | 1403 | // loop to determine which way the loop exited. |
kvn@6721 | 1404 | // Loop predicate If node connects to Bool node through Opaque1 node. |
kvn@6721 | 1405 | if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use)) { |
duke@435 | 1406 | // Since this code is highly unlikely, we lazily build the worklist |
duke@435 | 1407 | // of such Nodes to go split. |
duke@435 | 1408 | if( !split_if_set ) |
duke@435 | 1409 | split_if_set = new Node_List(area); |
duke@435 | 1410 | split_if_set->push(use); |
duke@435 | 1411 | } |
duke@435 | 1412 | if( use->is_Bool() ) { |
duke@435 | 1413 | if( !split_bool_set ) |
duke@435 | 1414 | split_bool_set = new Node_List(area); |
duke@435 | 1415 | split_bool_set->push(use); |
duke@435 | 1416 | } |
duke@435 | 1417 | if( use->Opcode() == Op_CreateEx ) { |
duke@435 | 1418 | if( !split_cex_set ) |
duke@435 | 1419 | split_cex_set = new Node_List(area); |
duke@435 | 1420 | split_cex_set->push(use); |
duke@435 | 1421 | } |
duke@435 | 1422 | |
duke@435 | 1423 | |
duke@435 | 1424 | // Get "block" use is in |
duke@435 | 1425 | uint idx = 0; |
duke@435 | 1426 | while( use->in(idx) != old ) idx++; |
duke@435 | 1427 | Node *prev = use->is_CFG() ? use : get_ctrl(use); |
duke@435 | 1428 | assert( !loop->is_member( get_loop( prev ) ), "" ); |
duke@435 | 1429 | Node *cfg = prev->_idx >= new_counter |
duke@435 | 1430 | ? prev->in(2) |
duke@435 | 1431 | : idom(prev); |
duke@435 | 1432 | if( use->is_Phi() ) // Phi use is in prior block |
duke@435 | 1433 | cfg = prev->in(idx); // NOT in block of Phi itself |
duke@435 | 1434 | if (cfg->is_top()) { // Use is dead? |
kvn@3847 | 1435 | _igvn.replace_input_of(use, idx, C->top()); |
duke@435 | 1436 | continue; |
duke@435 | 1437 | } |
duke@435 | 1438 | |
duke@435 | 1439 | while( !loop->is_member( get_loop( cfg ) ) ) { |
duke@435 | 1440 | prev = cfg; |
duke@435 | 1441 | cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg); |
duke@435 | 1442 | } |
duke@435 | 1443 | // If the use occurs after merging several exits from the loop, then |
duke@435 | 1444 | // old value must have dominated all those exits. Since the same old |
duke@435 | 1445 | // value was used on all those exits we did not need a Phi at this |
duke@435 | 1446 | // merge point. NOW we do need a Phi here. Each loop exit value |
duke@435 | 1447 | // is now merged with the peeled body exit; each exit gets its own |
duke@435 | 1448 | // private Phi and those Phis need to be merged here. |
duke@435 | 1449 | Node *phi; |
duke@435 | 1450 | if( prev->is_Region() ) { |
duke@435 | 1451 | if( idx == 0 ) { // Updating control edge? |
duke@435 | 1452 | phi = prev; // Just use existing control |
duke@435 | 1453 | } else { // Else need a new Phi |
duke@435 | 1454 | phi = PhiNode::make( prev, old ); |
duke@435 | 1455 | // Now recursively fix up the new uses of old! |
duke@435 | 1456 | for( uint i = 1; i < prev->req(); i++ ) { |
duke@435 | 1457 | worklist.push(phi); // Onto worklist once for each 'old' input |
duke@435 | 1458 | } |
duke@435 | 1459 | } |
duke@435 | 1460 | } else { |
duke@435 | 1461 | // Get new RegionNode merging old and new loop exits |
duke@435 | 1462 | prev = old_new[prev->_idx]; |
duke@435 | 1463 | assert( prev, "just made this in step 7" ); |
duke@435 | 1464 | if( idx == 0 ) { // Updating control edge? |
duke@435 | 1465 | phi = prev; // Just use existing control |
duke@435 | 1466 | } else { // Else need a new Phi |
duke@435 | 1467 | // Make a new Phi merging data values properly |
duke@435 | 1468 | phi = PhiNode::make( prev, old ); |
duke@435 | 1469 | phi->set_req( 1, nnn ); |
duke@435 | 1470 | } |
duke@435 | 1471 | } |
duke@435 | 1472 | // If inserting a new Phi, check for prior hits |
duke@435 | 1473 | if( idx != 0 ) { |
duke@435 | 1474 | Node *hit = _igvn.hash_find_insert(phi); |
duke@435 | 1475 | if( hit == NULL ) { |
duke@435 | 1476 | _igvn.register_new_node_with_optimizer(phi); // Register new phi |
duke@435 | 1477 | } else { // or |
duke@435 | 1478 | // Remove the new phi from the graph and use the hit |
duke@435 | 1479 | _igvn.remove_dead_node(phi); |
duke@435 | 1480 | phi = hit; // Use existing phi |
duke@435 | 1481 | } |
duke@435 | 1482 | set_ctrl(phi, prev); |
duke@435 | 1483 | } |
duke@435 | 1484 | // Make 'use' use the Phi instead of the old loop body exit value |
kvn@3847 | 1485 | _igvn.replace_input_of(use, idx, phi); |
duke@435 | 1486 | if( use->_idx >= new_counter ) { // If updating new phis |
duke@435 | 1487 | // Not needed for correctness, but prevents a weak assert |
duke@435 | 1488 | // in AddPNode from tripping (when we end up with different |
duke@435 | 1489 | // base & derived Phis that will become the same after |
duke@435 | 1490 | // IGVN does CSE). |
duke@435 | 1491 | Node *hit = _igvn.hash_find_insert(use); |
duke@435 | 1492 | if( hit ) // Go ahead and re-hash for hits. |
kvn@1976 | 1493 | _igvn.replace_node( use, hit ); |
duke@435 | 1494 | } |
duke@435 | 1495 | |
duke@435 | 1496 | // If 'use' was in the loop-exit block, it now needs to be sunk |
duke@435 | 1497 | // below the post-loop merge point. |
duke@435 | 1498 | sink_use( use, prev ); |
duke@435 | 1499 | } |
duke@435 | 1500 | } |
duke@435 | 1501 | } |
duke@435 | 1502 | |
duke@435 | 1503 | // Check for IFs that need splitting/cloning. Happens if an IF outside of |
duke@435 | 1504 | // the loop uses a condition set in the loop. The original IF probably |
duke@435 | 1505 | // takes control from one or more OLD Regions (which in turn get from NEW |
duke@435 | 1506 | // Regions). In any case, there will be a set of Phis for each merge point |
duke@435 | 1507 | // from the IF up to where the original BOOL def exists the loop. |
duke@435 | 1508 | if( split_if_set ) { |
duke@435 | 1509 | while( split_if_set->size() ) { |
duke@435 | 1510 | Node *iff = split_if_set->pop(); |
duke@435 | 1511 | if( iff->in(1)->is_Phi() ) { |
duke@435 | 1512 | BoolNode *b = clone_iff( iff->in(1)->as_Phi(), loop ); |
kvn@3847 | 1513 | _igvn.replace_input_of(iff, 1, b); |
duke@435 | 1514 | } |
duke@435 | 1515 | } |
duke@435 | 1516 | } |
duke@435 | 1517 | if( split_bool_set ) { |
duke@435 | 1518 | while( split_bool_set->size() ) { |
duke@435 | 1519 | Node *b = split_bool_set->pop(); |
duke@435 | 1520 | Node *phi = b->in(1); |
duke@435 | 1521 | assert( phi->is_Phi(), "" ); |
duke@435 | 1522 | CmpNode *cmp = clone_bool( (PhiNode*)phi, loop ); |
kvn@3847 | 1523 | _igvn.replace_input_of(b, 1, cmp); |
duke@435 | 1524 | } |
duke@435 | 1525 | } |
duke@435 | 1526 | if( split_cex_set ) { |
duke@435 | 1527 | while( split_cex_set->size() ) { |
duke@435 | 1528 | Node *b = split_cex_set->pop(); |
duke@435 | 1529 | assert( b->in(0)->is_Region(), "" ); |
duke@435 | 1530 | assert( b->in(1)->is_Phi(), "" ); |
duke@435 | 1531 | assert( b->in(0)->in(0) == b->in(1)->in(0), "" ); |
duke@435 | 1532 | split_up( b, b->in(0), NULL ); |
duke@435 | 1533 | } |
duke@435 | 1534 | } |
duke@435 | 1535 | |
duke@435 | 1536 | } |
duke@435 | 1537 | |
duke@435 | 1538 | |
duke@435 | 1539 | //---------------------- stride_of_possible_iv ------------------------------------- |
duke@435 | 1540 | // Looks for an iff/bool/comp with one operand of the compare |
duke@435 | 1541 | // being a cycle involving an add and a phi, |
duke@435 | 1542 | // with an optional truncation (left-shift followed by a right-shift) |
duke@435 | 1543 | // of the add. Returns zero if not an iv. |
duke@435 | 1544 | int PhaseIdealLoop::stride_of_possible_iv(Node* iff) { |
duke@435 | 1545 | Node* trunc1 = NULL; |
duke@435 | 1546 | Node* trunc2 = NULL; |
duke@435 | 1547 | const TypeInt* ttype = NULL; |
duke@435 | 1548 | if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) { |
duke@435 | 1549 | return 0; |
duke@435 | 1550 | } |
duke@435 | 1551 | BoolNode* bl = iff->in(1)->as_Bool(); |
duke@435 | 1552 | Node* cmp = bl->in(1); |
duke@435 | 1553 | if (!cmp || cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU) { |
duke@435 | 1554 | return 0; |
duke@435 | 1555 | } |
duke@435 | 1556 | // Must have an invariant operand |
duke@435 | 1557 | if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) { |
duke@435 | 1558 | return 0; |
duke@435 | 1559 | } |
duke@435 | 1560 | Node* add2 = NULL; |
duke@435 | 1561 | Node* cmp1 = cmp->in(1); |
duke@435 | 1562 | if (cmp1->is_Phi()) { |
duke@435 | 1563 | // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) ))) |
duke@435 | 1564 | Node* phi = cmp1; |
duke@435 | 1565 | for (uint i = 1; i < phi->req(); i++) { |
duke@435 | 1566 | Node* in = phi->in(i); |
duke@435 | 1567 | Node* add = CountedLoopNode::match_incr_with_optional_truncation(in, |
duke@435 | 1568 | &trunc1, &trunc2, &ttype); |
duke@435 | 1569 | if (add && add->in(1) == phi) { |
duke@435 | 1570 | add2 = add->in(2); |
duke@435 | 1571 | break; |
duke@435 | 1572 | } |
duke@435 | 1573 | } |
duke@435 | 1574 | } else { |
duke@435 | 1575 | // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) ))) |
duke@435 | 1576 | Node* addtrunc = cmp1; |
duke@435 | 1577 | Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc, |
duke@435 | 1578 | &trunc1, &trunc2, &ttype); |
duke@435 | 1579 | if (add && add->in(1)->is_Phi()) { |
duke@435 | 1580 | Node* phi = add->in(1); |
duke@435 | 1581 | for (uint i = 1; i < phi->req(); i++) { |
duke@435 | 1582 | if (phi->in(i) == addtrunc) { |
duke@435 | 1583 | add2 = add->in(2); |
duke@435 | 1584 | break; |
duke@435 | 1585 | } |
duke@435 | 1586 | } |
duke@435 | 1587 | } |
duke@435 | 1588 | } |
duke@435 | 1589 | if (add2 != NULL) { |
duke@435 | 1590 | const TypeInt* add2t = _igvn.type(add2)->is_int(); |
duke@435 | 1591 | if (add2t->is_con()) { |
duke@435 | 1592 | return add2t->get_con(); |
duke@435 | 1593 | } |
duke@435 | 1594 | } |
duke@435 | 1595 | return 0; |
duke@435 | 1596 | } |
duke@435 | 1597 | |
duke@435 | 1598 | |
duke@435 | 1599 | //---------------------- stay_in_loop ------------------------------------- |
duke@435 | 1600 | // Return the (unique) control output node that's in the loop (if it exists.) |
duke@435 | 1601 | Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) { |
duke@435 | 1602 | Node* unique = NULL; |
duke@435 | 1603 | if (!n) return NULL; |
duke@435 | 1604 | for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { |
duke@435 | 1605 | Node* use = n->fast_out(i); |
duke@435 | 1606 | if (!has_ctrl(use) && loop->is_member(get_loop(use))) { |
duke@435 | 1607 | if (unique != NULL) { |
duke@435 | 1608 | return NULL; |
duke@435 | 1609 | } |
duke@435 | 1610 | unique = use; |
duke@435 | 1611 | } |
duke@435 | 1612 | } |
duke@435 | 1613 | return unique; |
duke@435 | 1614 | } |
duke@435 | 1615 | |
duke@435 | 1616 | //------------------------------ register_node ------------------------------------- |
duke@435 | 1617 | // Utility to register node "n" with PhaseIdealLoop |
duke@435 | 1618 | void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) { |
duke@435 | 1619 | _igvn.register_new_node_with_optimizer(n); |
duke@435 | 1620 | loop->_body.push(n); |
duke@435 | 1621 | if (n->is_CFG()) { |
duke@435 | 1622 | set_loop(n, loop); |
duke@435 | 1623 | set_idom(n, pred, ddepth); |
duke@435 | 1624 | } else { |
duke@435 | 1625 | set_ctrl(n, pred); |
duke@435 | 1626 | } |
duke@435 | 1627 | } |
duke@435 | 1628 | |
duke@435 | 1629 | //------------------------------ proj_clone ------------------------------------- |
duke@435 | 1630 | // Utility to create an if-projection |
duke@435 | 1631 | ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) { |
duke@435 | 1632 | ProjNode* c = p->clone()->as_Proj(); |
duke@435 | 1633 | c->set_req(0, iff); |
duke@435 | 1634 | return c; |
duke@435 | 1635 | } |
duke@435 | 1636 | |
duke@435 | 1637 | //------------------------------ short_circuit_if ------------------------------------- |
duke@435 | 1638 | // Force the iff control output to be the live_proj |
duke@435 | 1639 | Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) { |
morris@4774 | 1640 | guarantee(live_proj != NULL, "null projection"); |
duke@435 | 1641 | int proj_con = live_proj->_con; |
duke@435 | 1642 | assert(proj_con == 0 || proj_con == 1, "false or true projection"); |
duke@435 | 1643 | Node *con = _igvn.intcon(proj_con); |
duke@435 | 1644 | set_ctrl(con, C->root()); |
duke@435 | 1645 | if (iff) { |
duke@435 | 1646 | iff->set_req(1, con); |
duke@435 | 1647 | } |
duke@435 | 1648 | return con; |
duke@435 | 1649 | } |
duke@435 | 1650 | |
duke@435 | 1651 | //------------------------------ insert_if_before_proj ------------------------------------- |
duke@435 | 1652 | // Insert a new if before an if projection (* - new node) |
duke@435 | 1653 | // |
duke@435 | 1654 | // before |
duke@435 | 1655 | // if(test) |
duke@435 | 1656 | // / \ |
duke@435 | 1657 | // v v |
duke@435 | 1658 | // other-proj proj (arg) |
duke@435 | 1659 | // |
duke@435 | 1660 | // after |
duke@435 | 1661 | // if(test) |
duke@435 | 1662 | // / \ |
duke@435 | 1663 | // / v |
duke@435 | 1664 | // | * proj-clone |
duke@435 | 1665 | // v | |
duke@435 | 1666 | // other-proj v |
duke@435 | 1667 | // * new_if(relop(cmp[IU](left,right))) |
duke@435 | 1668 | // / \ |
duke@435 | 1669 | // v v |
duke@435 | 1670 | // * new-proj proj |
duke@435 | 1671 | // (returned) |
duke@435 | 1672 | // |
duke@435 | 1673 | ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) { |
duke@435 | 1674 | IfNode* iff = proj->in(0)->as_If(); |
duke@435 | 1675 | IdealLoopTree *loop = get_loop(proj); |
duke@435 | 1676 | ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); |
duke@435 | 1677 | int ddepth = dom_depth(proj); |
duke@435 | 1678 | |
kvn@3847 | 1679 | _igvn.rehash_node_delayed(iff); |
kvn@3847 | 1680 | _igvn.rehash_node_delayed(proj); |
duke@435 | 1681 | |
duke@435 | 1682 | proj->set_req(0, NULL); // temporary disconnect |
duke@435 | 1683 | ProjNode* proj2 = proj_clone(proj, iff); |
duke@435 | 1684 | register_node(proj2, loop, iff, ddepth); |
duke@435 | 1685 | |
kvn@4115 | 1686 | Node* cmp = Signed ? (Node*) new (C)CmpINode(left, right) : (Node*) new (C)CmpUNode(left, right); |
duke@435 | 1687 | register_node(cmp, loop, proj2, ddepth); |
duke@435 | 1688 | |
kvn@4115 | 1689 | BoolNode* bol = new (C)BoolNode(cmp, relop); |
duke@435 | 1690 | register_node(bol, loop, proj2, ddepth); |
duke@435 | 1691 | |
kvn@4115 | 1692 | IfNode* new_if = new (C)IfNode(proj2, bol, iff->_prob, iff->_fcnt); |
duke@435 | 1693 | register_node(new_if, loop, proj2, ddepth); |
duke@435 | 1694 | |
duke@435 | 1695 | proj->set_req(0, new_if); // reattach |
duke@435 | 1696 | set_idom(proj, new_if, ddepth); |
duke@435 | 1697 | |
duke@435 | 1698 | ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj(); |
morris@4774 | 1699 | guarantee(new_exit != NULL, "null exit node"); |
duke@435 | 1700 | register_node(new_exit, get_loop(other_proj), new_if, ddepth); |
duke@435 | 1701 | |
duke@435 | 1702 | return new_exit; |
duke@435 | 1703 | } |
duke@435 | 1704 | |
duke@435 | 1705 | //------------------------------ insert_region_before_proj ------------------------------------- |
duke@435 | 1706 | // Insert a region before an if projection (* - new node) |
duke@435 | 1707 | // |
duke@435 | 1708 | // before |
duke@435 | 1709 | // if(test) |
duke@435 | 1710 | // / | |
duke@435 | 1711 | // v | |
duke@435 | 1712 | // proj v |
duke@435 | 1713 | // other-proj |
duke@435 | 1714 | // |
duke@435 | 1715 | // after |
duke@435 | 1716 | // if(test) |
duke@435 | 1717 | // / | |
duke@435 | 1718 | // v | |
duke@435 | 1719 | // * proj-clone v |
duke@435 | 1720 | // | other-proj |
duke@435 | 1721 | // v |
duke@435 | 1722 | // * new-region |
duke@435 | 1723 | // | |
duke@435 | 1724 | // v |
duke@435 | 1725 | // * dum_if |
duke@435 | 1726 | // / \ |
duke@435 | 1727 | // v \ |
duke@435 | 1728 | // * dum-proj v |
duke@435 | 1729 | // proj |
duke@435 | 1730 | // |
duke@435 | 1731 | RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) { |
duke@435 | 1732 | IfNode* iff = proj->in(0)->as_If(); |
duke@435 | 1733 | IdealLoopTree *loop = get_loop(proj); |
duke@435 | 1734 | ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); |
duke@435 | 1735 | int ddepth = dom_depth(proj); |
duke@435 | 1736 | |
kvn@3847 | 1737 | _igvn.rehash_node_delayed(iff); |
kvn@3847 | 1738 | _igvn.rehash_node_delayed(proj); |
duke@435 | 1739 | |
duke@435 | 1740 | proj->set_req(0, NULL); // temporary disconnect |
duke@435 | 1741 | ProjNode* proj2 = proj_clone(proj, iff); |
duke@435 | 1742 | register_node(proj2, loop, iff, ddepth); |
duke@435 | 1743 | |
kvn@4115 | 1744 | RegionNode* reg = new (C)RegionNode(2); |
duke@435 | 1745 | reg->set_req(1, proj2); |
duke@435 | 1746 | register_node(reg, loop, iff, ddepth); |
duke@435 | 1747 | |
kvn@4115 | 1748 | IfNode* dum_if = new (C)IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt); |
duke@435 | 1749 | register_node(dum_if, loop, reg, ddepth); |
duke@435 | 1750 | |
duke@435 | 1751 | proj->set_req(0, dum_if); // reattach |
duke@435 | 1752 | set_idom(proj, dum_if, ddepth); |
duke@435 | 1753 | |
duke@435 | 1754 | ProjNode* dum_proj = proj_clone(other_proj, dum_if); |
duke@435 | 1755 | register_node(dum_proj, loop, dum_if, ddepth); |
duke@435 | 1756 | |
duke@435 | 1757 | return reg; |
duke@435 | 1758 | } |
duke@435 | 1759 | |
duke@435 | 1760 | //------------------------------ insert_cmpi_loop_exit ------------------------------------- |
duke@435 | 1761 | // Clone a signed compare loop exit from an unsigned compare and |
duke@435 | 1762 | // insert it before the unsigned cmp on the stay-in-loop path. |
duke@435 | 1763 | // All new nodes inserted in the dominator tree between the original |
duke@435 | 1764 | // if and it's projections. The original if test is replaced with |
duke@435 | 1765 | // a constant to force the stay-in-loop path. |
duke@435 | 1766 | // |
duke@435 | 1767 | // This is done to make sure that the original if and it's projections |
duke@435 | 1768 | // still dominate the same set of control nodes, that the ctrl() relation |
duke@435 | 1769 | // from data nodes to them is preserved, and that their loop nesting is |
duke@435 | 1770 | // preserved. |
duke@435 | 1771 | // |
duke@435 | 1772 | // before |
duke@435 | 1773 | // if(i <u limit) unsigned compare loop exit |
duke@435 | 1774 | // / | |
duke@435 | 1775 | // v v |
duke@435 | 1776 | // exit-proj stay-in-loop-proj |
duke@435 | 1777 | // |
duke@435 | 1778 | // after |
duke@435 | 1779 | // if(stay-in-loop-const) original if |
duke@435 | 1780 | // / | |
duke@435 | 1781 | // / v |
duke@435 | 1782 | // / if(i < limit) new signed test |
duke@435 | 1783 | // / / | |
duke@435 | 1784 | // / / v |
duke@435 | 1785 | // / / if(i <u limit) new cloned unsigned test |
duke@435 | 1786 | // / / / | |
duke@435 | 1787 | // v v v | |
duke@435 | 1788 | // region | |
duke@435 | 1789 | // | | |
duke@435 | 1790 | // dum-if | |
duke@435 | 1791 | // / | | |
duke@435 | 1792 | // ether | | |
duke@435 | 1793 | // v v |
duke@435 | 1794 | // exit-proj stay-in-loop-proj |
duke@435 | 1795 | // |
duke@435 | 1796 | IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) { |
duke@435 | 1797 | const bool Signed = true; |
duke@435 | 1798 | const bool Unsigned = false; |
duke@435 | 1799 | |
duke@435 | 1800 | BoolNode* bol = if_cmpu->in(1)->as_Bool(); |
duke@435 | 1801 | if (bol->_test._test != BoolTest::lt) return NULL; |
duke@435 | 1802 | CmpNode* cmpu = bol->in(1)->as_Cmp(); |
duke@435 | 1803 | if (cmpu->Opcode() != Op_CmpU) return NULL; |
duke@435 | 1804 | int stride = stride_of_possible_iv(if_cmpu); |
duke@435 | 1805 | if (stride == 0) return NULL; |
duke@435 | 1806 | |
morris@4774 | 1807 | Node* lp_proj = stay_in_loop(if_cmpu, loop); |
morris@4774 | 1808 | guarantee(lp_proj != NULL, "null loop node"); |
morris@4774 | 1809 | |
morris@4774 | 1810 | ProjNode* lp_continue = lp_proj->as_Proj(); |
duke@435 | 1811 | ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj(); |
duke@435 | 1812 | |
duke@435 | 1813 | Node* limit = NULL; |
duke@435 | 1814 | if (stride > 0) { |
duke@435 | 1815 | limit = cmpu->in(2); |
duke@435 | 1816 | } else { |
duke@435 | 1817 | limit = _igvn.makecon(TypeInt::ZERO); |
duke@435 | 1818 | set_ctrl(limit, C->root()); |
duke@435 | 1819 | } |
duke@435 | 1820 | // Create a new region on the exit path |
duke@435 | 1821 | RegionNode* reg = insert_region_before_proj(lp_exit); |
morris@4774 | 1822 | guarantee(reg != NULL, "null region node"); |
duke@435 | 1823 | |
duke@435 | 1824 | // Clone the if-cmpu-true-false using a signed compare |
duke@435 | 1825 | BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge; |
duke@435 | 1826 | ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue); |
duke@435 | 1827 | reg->add_req(cmpi_exit); |
duke@435 | 1828 | |
duke@435 | 1829 | // Clone the if-cmpu-true-false |
duke@435 | 1830 | BoolTest::mask rel_u = bol->_test._test; |
duke@435 | 1831 | ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue); |
duke@435 | 1832 | reg->add_req(cmpu_exit); |
duke@435 | 1833 | |
duke@435 | 1834 | // Force original if to stay in loop. |
duke@435 | 1835 | short_circuit_if(if_cmpu, lp_continue); |
duke@435 | 1836 | |
duke@435 | 1837 | return cmpi_exit->in(0)->as_If(); |
duke@435 | 1838 | } |
duke@435 | 1839 | |
duke@435 | 1840 | //------------------------------ remove_cmpi_loop_exit ------------------------------------- |
duke@435 | 1841 | // Remove a previously inserted signed compare loop exit. |
duke@435 | 1842 | void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) { |
duke@435 | 1843 | Node* lp_proj = stay_in_loop(if_cmp, loop); |
duke@435 | 1844 | assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI && |
duke@435 | 1845 | stay_in_loop(lp_proj, loop)->is_If() && |
duke@435 | 1846 | stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu"); |
duke@435 | 1847 | Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO); |
duke@435 | 1848 | set_ctrl(con, C->root()); |
duke@435 | 1849 | if_cmp->set_req(1, con); |
duke@435 | 1850 | } |
duke@435 | 1851 | |
duke@435 | 1852 | //------------------------------ scheduled_nodelist ------------------------------------- |
duke@435 | 1853 | // Create a post order schedule of nodes that are in the |
duke@435 | 1854 | // "member" set. The list is returned in "sched". |
duke@435 | 1855 | // The first node in "sched" is the loop head, followed by |
duke@435 | 1856 | // nodes which have no inputs in the "member" set, and then |
duke@435 | 1857 | // followed by the nodes that have an immediate input dependence |
duke@435 | 1858 | // on a node in "sched". |
duke@435 | 1859 | void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) { |
duke@435 | 1860 | |
duke@435 | 1861 | assert(member.test(loop->_head->_idx), "loop head must be in member set"); |
duke@435 | 1862 | Arena *a = Thread::current()->resource_area(); |
duke@435 | 1863 | VectorSet visited(a); |
duke@435 | 1864 | Node_Stack nstack(a, loop->_body.size()); |
duke@435 | 1865 | |
duke@435 | 1866 | Node* n = loop->_head; // top of stack is cached in "n" |
duke@435 | 1867 | uint idx = 0; |
duke@435 | 1868 | visited.set(n->_idx); |
duke@435 | 1869 | |
duke@435 | 1870 | // Initially push all with no inputs from within member set |
duke@435 | 1871 | for(uint i = 0; i < loop->_body.size(); i++ ) { |
duke@435 | 1872 | Node *elt = loop->_body.at(i); |
duke@435 | 1873 | if (member.test(elt->_idx)) { |
duke@435 | 1874 | bool found = false; |
duke@435 | 1875 | for (uint j = 0; j < elt->req(); j++) { |
duke@435 | 1876 | Node* def = elt->in(j); |
duke@435 | 1877 | if (def && member.test(def->_idx) && def != elt) { |
duke@435 | 1878 | found = true; |
duke@435 | 1879 | break; |
duke@435 | 1880 | } |
duke@435 | 1881 | } |
duke@435 | 1882 | if (!found && elt != loop->_head) { |
duke@435 | 1883 | nstack.push(n, idx); |
duke@435 | 1884 | n = elt; |
duke@435 | 1885 | assert(!visited.test(n->_idx), "not seen yet"); |
duke@435 | 1886 | visited.set(n->_idx); |
duke@435 | 1887 | } |
duke@435 | 1888 | } |
duke@435 | 1889 | } |
duke@435 | 1890 | |
duke@435 | 1891 | // traverse out's that are in the member set |
duke@435 | 1892 | while (true) { |
duke@435 | 1893 | if (idx < n->outcnt()) { |
duke@435 | 1894 | Node* use = n->raw_out(idx); |
duke@435 | 1895 | idx++; |
duke@435 | 1896 | if (!visited.test_set(use->_idx)) { |
duke@435 | 1897 | if (member.test(use->_idx)) { |
duke@435 | 1898 | nstack.push(n, idx); |
duke@435 | 1899 | n = use; |
duke@435 | 1900 | idx = 0; |
duke@435 | 1901 | } |
duke@435 | 1902 | } |
duke@435 | 1903 | } else { |
duke@435 | 1904 | // All outputs processed |
duke@435 | 1905 | sched.push(n); |
duke@435 | 1906 | if (nstack.is_empty()) break; |
duke@435 | 1907 | n = nstack.node(); |
duke@435 | 1908 | idx = nstack.index(); |
duke@435 | 1909 | nstack.pop(); |
duke@435 | 1910 | } |
duke@435 | 1911 | } |
duke@435 | 1912 | } |
duke@435 | 1913 | |
duke@435 | 1914 | |
duke@435 | 1915 | //------------------------------ has_use_in_set ------------------------------------- |
duke@435 | 1916 | // Has a use in the vector set |
duke@435 | 1917 | bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) { |
duke@435 | 1918 | for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { |
duke@435 | 1919 | Node* use = n->fast_out(j); |
duke@435 | 1920 | if (vset.test(use->_idx)) { |
duke@435 | 1921 | return true; |
duke@435 | 1922 | } |
duke@435 | 1923 | } |
duke@435 | 1924 | return false; |
duke@435 | 1925 | } |
duke@435 | 1926 | |
duke@435 | 1927 | |
duke@435 | 1928 | //------------------------------ has_use_internal_to_set ------------------------------------- |
duke@435 | 1929 | // Has use internal to the vector set (ie. not in a phi at the loop head) |
duke@435 | 1930 | bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) { |
duke@435 | 1931 | Node* head = loop->_head; |
duke@435 | 1932 | for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { |
duke@435 | 1933 | Node* use = n->fast_out(j); |
duke@435 | 1934 | if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) { |
duke@435 | 1935 | return true; |
duke@435 | 1936 | } |
duke@435 | 1937 | } |
duke@435 | 1938 | return false; |
duke@435 | 1939 | } |
duke@435 | 1940 | |
duke@435 | 1941 | |
duke@435 | 1942 | //------------------------------ clone_for_use_outside_loop ------------------------------------- |
duke@435 | 1943 | // clone "n" for uses that are outside of loop |
kvn@5154 | 1944 | int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) { |
kvn@5154 | 1945 | int cloned = 0; |
duke@435 | 1946 | assert(worklist.size() == 0, "should be empty"); |
duke@435 | 1947 | for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { |
duke@435 | 1948 | Node* use = n->fast_out(j); |
duke@435 | 1949 | if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) { |
duke@435 | 1950 | worklist.push(use); |
duke@435 | 1951 | } |
duke@435 | 1952 | } |
duke@435 | 1953 | while( worklist.size() ) { |
duke@435 | 1954 | Node *use = worklist.pop(); |
duke@435 | 1955 | if (!has_node(use) || use->in(0) == C->top()) continue; |
duke@435 | 1956 | uint j; |
duke@435 | 1957 | for (j = 0; j < use->req(); j++) { |
duke@435 | 1958 | if (use->in(j) == n) break; |
duke@435 | 1959 | } |
duke@435 | 1960 | assert(j < use->req(), "must be there"); |
duke@435 | 1961 | |
duke@435 | 1962 | // clone "n" and insert it between the inputs of "n" and the use outside the loop |
duke@435 | 1963 | Node* n_clone = n->clone(); |
kvn@3847 | 1964 | _igvn.replace_input_of(use, j, n_clone); |
kvn@5154 | 1965 | cloned++; |
never@686 | 1966 | Node* use_c; |
duke@435 | 1967 | if (!use->is_Phi()) { |
never@686 | 1968 | use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0); |
duke@435 | 1969 | } else { |
duke@435 | 1970 | // Use in a phi is considered a use in the associated predecessor block |
never@686 | 1971 | use_c = use->in(0)->in(j); |
duke@435 | 1972 | } |
never@686 | 1973 | set_ctrl(n_clone, use_c); |
never@686 | 1974 | assert(!loop->is_member(get_loop(use_c)), "should be outside loop"); |
never@686 | 1975 | get_loop(use_c)->_body.push(n_clone); |
duke@435 | 1976 | _igvn.register_new_node_with_optimizer(n_clone); |
duke@435 | 1977 | #if !defined(PRODUCT) |
duke@435 | 1978 | if (TracePartialPeeling) { |
duke@435 | 1979 | tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx); |
duke@435 | 1980 | } |
duke@435 | 1981 | #endif |
duke@435 | 1982 | } |
kvn@5154 | 1983 | return cloned; |
duke@435 | 1984 | } |
duke@435 | 1985 | |
duke@435 | 1986 | |
duke@435 | 1987 | //------------------------------ clone_for_special_use_inside_loop ------------------------------------- |
duke@435 | 1988 | // clone "n" for special uses that are in the not_peeled region. |
duke@435 | 1989 | // If these def-uses occur in separate blocks, the code generator |
duke@435 | 1990 | // marks the method as not compilable. For example, if a "BoolNode" |
duke@435 | 1991 | // is in a different basic block than the "IfNode" that uses it, then |
duke@435 | 1992 | // the compilation is aborted in the code generator. |
duke@435 | 1993 | void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, |
duke@435 | 1994 | VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) { |
duke@435 | 1995 | if (n->is_Phi() || n->is_Load()) { |
duke@435 | 1996 | return; |
duke@435 | 1997 | } |
duke@435 | 1998 | assert(worklist.size() == 0, "should be empty"); |
duke@435 | 1999 | for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { |
duke@435 | 2000 | Node* use = n->fast_out(j); |
duke@435 | 2001 | if ( not_peel.test(use->_idx) && |
duke@435 | 2002 | (use->is_If() || use->is_CMove() || use->is_Bool()) && |
duke@435 | 2003 | use->in(1) == n) { |
duke@435 | 2004 | worklist.push(use); |
duke@435 | 2005 | } |
duke@435 | 2006 | } |
duke@435 | 2007 | if (worklist.size() > 0) { |
duke@435 | 2008 | // clone "n" and insert it between inputs of "n" and the use |
duke@435 | 2009 | Node* n_clone = n->clone(); |
duke@435 | 2010 | loop->_body.push(n_clone); |
duke@435 | 2011 | _igvn.register_new_node_with_optimizer(n_clone); |
duke@435 | 2012 | set_ctrl(n_clone, get_ctrl(n)); |
duke@435 | 2013 | sink_list.push(n_clone); |
duke@435 | 2014 | not_peel <<= n_clone->_idx; // add n_clone to not_peel set. |
duke@435 | 2015 | #if !defined(PRODUCT) |
duke@435 | 2016 | if (TracePartialPeeling) { |
duke@435 | 2017 | tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx); |
duke@435 | 2018 | } |
duke@435 | 2019 | #endif |
duke@435 | 2020 | while( worklist.size() ) { |
duke@435 | 2021 | Node *use = worklist.pop(); |
kvn@3847 | 2022 | _igvn.rehash_node_delayed(use); |
duke@435 | 2023 | for (uint j = 1; j < use->req(); j++) { |
duke@435 | 2024 | if (use->in(j) == n) { |
duke@435 | 2025 | use->set_req(j, n_clone); |
duke@435 | 2026 | } |
duke@435 | 2027 | } |
duke@435 | 2028 | } |
duke@435 | 2029 | } |
duke@435 | 2030 | } |
duke@435 | 2031 | |
duke@435 | 2032 | |
duke@435 | 2033 | //------------------------------ insert_phi_for_loop ------------------------------------- |
duke@435 | 2034 | // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist |
duke@435 | 2035 | void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) { |
duke@435 | 2036 | Node *phi = PhiNode::make(lp, back_edge_val); |
duke@435 | 2037 | phi->set_req(LoopNode::EntryControl, lp_entry_val); |
duke@435 | 2038 | // Use existing phi if it already exists |
duke@435 | 2039 | Node *hit = _igvn.hash_find_insert(phi); |
duke@435 | 2040 | if( hit == NULL ) { |
duke@435 | 2041 | _igvn.register_new_node_with_optimizer(phi); |
duke@435 | 2042 | set_ctrl(phi, lp); |
duke@435 | 2043 | } else { |
duke@435 | 2044 | // Remove the new phi from the graph and use the hit |
duke@435 | 2045 | _igvn.remove_dead_node(phi); |
duke@435 | 2046 | phi = hit; |
duke@435 | 2047 | } |
kvn@3847 | 2048 | _igvn.replace_input_of(use, idx, phi); |
duke@435 | 2049 | } |
duke@435 | 2050 | |
duke@435 | 2051 | #ifdef ASSERT |
duke@435 | 2052 | //------------------------------ is_valid_loop_partition ------------------------------------- |
duke@435 | 2053 | // Validate the loop partition sets: peel and not_peel |
duke@435 | 2054 | bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, |
duke@435 | 2055 | VectorSet& not_peel ) { |
duke@435 | 2056 | uint i; |
duke@435 | 2057 | // Check that peel_list entries are in the peel set |
duke@435 | 2058 | for (i = 0; i < peel_list.size(); i++) { |
duke@435 | 2059 | if (!peel.test(peel_list.at(i)->_idx)) { |
duke@435 | 2060 | return false; |
duke@435 | 2061 | } |
duke@435 | 2062 | } |
duke@435 | 2063 | // Check at loop members are in one of peel set or not_peel set |
duke@435 | 2064 | for (i = 0; i < loop->_body.size(); i++ ) { |
duke@435 | 2065 | Node *def = loop->_body.at(i); |
duke@435 | 2066 | uint di = def->_idx; |
duke@435 | 2067 | // Check that peel set elements are in peel_list |
duke@435 | 2068 | if (peel.test(di)) { |
duke@435 | 2069 | if (not_peel.test(di)) { |
duke@435 | 2070 | return false; |
duke@435 | 2071 | } |
duke@435 | 2072 | // Must be in peel_list also |
duke@435 | 2073 | bool found = false; |
duke@435 | 2074 | for (uint j = 0; j < peel_list.size(); j++) { |
duke@435 | 2075 | if (peel_list.at(j)->_idx == di) { |
duke@435 | 2076 | found = true; |
duke@435 | 2077 | break; |
duke@435 | 2078 | } |
duke@435 | 2079 | } |
duke@435 | 2080 | if (!found) { |
duke@435 | 2081 | return false; |
duke@435 | 2082 | } |
duke@435 | 2083 | } else if (not_peel.test(di)) { |
duke@435 | 2084 | if (peel.test(di)) { |
duke@435 | 2085 | return false; |
duke@435 | 2086 | } |
duke@435 | 2087 | } else { |
duke@435 | 2088 | return false; |
duke@435 | 2089 | } |
duke@435 | 2090 | } |
duke@435 | 2091 | return true; |
duke@435 | 2092 | } |
duke@435 | 2093 | |
duke@435 | 2094 | //------------------------------ is_valid_clone_loop_exit_use ------------------------------------- |
duke@435 | 2095 | // Ensure a use outside of loop is of the right form |
duke@435 | 2096 | bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) { |
duke@435 | 2097 | Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; |
duke@435 | 2098 | return (use->is_Phi() && |
duke@435 | 2099 | use_c->is_Region() && use_c->req() == 3 && |
duke@435 | 2100 | (use_c->in(exit_idx)->Opcode() == Op_IfTrue || |
duke@435 | 2101 | use_c->in(exit_idx)->Opcode() == Op_IfFalse || |
duke@435 | 2102 | use_c->in(exit_idx)->Opcode() == Op_JumpProj) && |
duke@435 | 2103 | loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) ); |
duke@435 | 2104 | } |
duke@435 | 2105 | |
duke@435 | 2106 | //------------------------------ is_valid_clone_loop_form ------------------------------------- |
duke@435 | 2107 | // Ensure that all uses outside of loop are of the right form |
duke@435 | 2108 | bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, |
duke@435 | 2109 | uint orig_exit_idx, uint clone_exit_idx) { |
duke@435 | 2110 | uint len = peel_list.size(); |
duke@435 | 2111 | for (uint i = 0; i < len; i++) { |
duke@435 | 2112 | Node *def = peel_list.at(i); |
duke@435 | 2113 | |
duke@435 | 2114 | for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { |
duke@435 | 2115 | Node *use = def->fast_out(j); |
duke@435 | 2116 | Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; |
duke@435 | 2117 | if (!loop->is_member(get_loop(use_c))) { |
duke@435 | 2118 | // use is not in the loop, check for correct structure |
duke@435 | 2119 | if (use->in(0) == def) { |
duke@435 | 2120 | // Okay |
duke@435 | 2121 | } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) { |
duke@435 | 2122 | return false; |
duke@435 | 2123 | } |
duke@435 | 2124 | } |
duke@435 | 2125 | } |
duke@435 | 2126 | } |
duke@435 | 2127 | return true; |
duke@435 | 2128 | } |
duke@435 | 2129 | #endif |
duke@435 | 2130 | |
duke@435 | 2131 | //------------------------------ partial_peel ------------------------------------- |
duke@435 | 2132 | // Partially peel (aka loop rotation) the top portion of a loop (called |
duke@435 | 2133 | // the peel section below) by cloning it and placing one copy just before |
duke@435 | 2134 | // the new loop head and the other copy at the bottom of the new loop. |
duke@435 | 2135 | // |
duke@435 | 2136 | // before after where it came from |
duke@435 | 2137 | // |
duke@435 | 2138 | // stmt1 stmt1 |
duke@435 | 2139 | // loop: stmt2 clone |
duke@435 | 2140 | // stmt2 if condA goto exitA clone |
duke@435 | 2141 | // if condA goto exitA new_loop: new |
duke@435 | 2142 | // stmt3 stmt3 clone |
duke@435 | 2143 | // if !condB goto loop if condB goto exitB clone |
duke@435 | 2144 | // exitB: stmt2 orig |
duke@435 | 2145 | // stmt4 if !condA goto new_loop orig |
duke@435 | 2146 | // exitA: goto exitA |
duke@435 | 2147 | // exitB: |
duke@435 | 2148 | // stmt4 |
duke@435 | 2149 | // exitA: |
duke@435 | 2150 | // |
duke@435 | 2151 | // Step 1: find the cut point: an exit test on probable |
duke@435 | 2152 | // induction variable. |
duke@435 | 2153 | // Step 2: schedule (with cloning) operations in the peel |
duke@435 | 2154 | // section that can be executed after the cut into |
duke@435 | 2155 | // the section that is not peeled. This may need |
duke@435 | 2156 | // to clone operations into exit blocks. For |
duke@435 | 2157 | // instance, a reference to A[i] in the not-peel |
duke@435 | 2158 | // section and a reference to B[i] in an exit block |
duke@435 | 2159 | // may cause a left-shift of i by 2 to be placed |
duke@435 | 2160 | // in the peel block. This step will clone the left |
duke@435 | 2161 | // shift into the exit block and sink the left shift |
duke@435 | 2162 | // from the peel to the not-peel section. |
duke@435 | 2163 | // Step 3: clone the loop, retarget the control, and insert |
duke@435 | 2164 | // phis for values that are live across the new loop |
duke@435 | 2165 | // head. This is very dependent on the graph structure |
duke@435 | 2166 | // from clone_loop. It creates region nodes for |
duke@435 | 2167 | // exit control and associated phi nodes for values |
duke@435 | 2168 | // flow out of the loop through that exit. The region |
duke@435 | 2169 | // node is dominated by the clone's control projection. |
duke@435 | 2170 | // So the clone's peel section is placed before the |
duke@435 | 2171 | // new loop head, and the clone's not-peel section is |
duke@435 | 2172 | // forms the top part of the new loop. The original |
duke@435 | 2173 | // peel section forms the tail of the new loop. |
duke@435 | 2174 | // Step 4: update the dominator tree and recompute the |
duke@435 | 2175 | // dominator depth. |
duke@435 | 2176 | // |
duke@435 | 2177 | // orig |
duke@435 | 2178 | // |
kvn@2727 | 2179 | // stmt1 |
kvn@2727 | 2180 | // | |
kvn@2727 | 2181 | // v |
kvn@2727 | 2182 | // loop predicate |
kvn@2727 | 2183 | // | |
kvn@2727 | 2184 | // v |
duke@435 | 2185 | // loop<----+ |
duke@435 | 2186 | // | | |
duke@435 | 2187 | // stmt2 | |
duke@435 | 2188 | // | | |
duke@435 | 2189 | // v | |
duke@435 | 2190 | // ifA | |
duke@435 | 2191 | // / | | |
duke@435 | 2192 | // v v | |
duke@435 | 2193 | // false true ^ <-- last_peel |
duke@435 | 2194 | // / | | |
duke@435 | 2195 | // / ===|==cut | |
duke@435 | 2196 | // / stmt3 | <-- first_not_peel |
duke@435 | 2197 | // / | | |
duke@435 | 2198 | // | v | |
duke@435 | 2199 | // v ifB | |
duke@435 | 2200 | // exitA: / \ | |
duke@435 | 2201 | // / \ | |
duke@435 | 2202 | // v v | |
duke@435 | 2203 | // false true | |
duke@435 | 2204 | // / \ | |
duke@435 | 2205 | // / ----+ |
duke@435 | 2206 | // | |
duke@435 | 2207 | // v |
duke@435 | 2208 | // exitB: |
duke@435 | 2209 | // stmt4 |
duke@435 | 2210 | // |
duke@435 | 2211 | // |
duke@435 | 2212 | // after clone loop |
duke@435 | 2213 | // |
duke@435 | 2214 | // stmt1 |
kvn@2727 | 2215 | // | |
kvn@2727 | 2216 | // v |
kvn@2727 | 2217 | // loop predicate |
duke@435 | 2218 | // / \ |
duke@435 | 2219 | // clone / \ orig |
duke@435 | 2220 | // / \ |
duke@435 | 2221 | // / \ |
duke@435 | 2222 | // v v |
duke@435 | 2223 | // +---->loop loop<----+ |
duke@435 | 2224 | // | | | | |
duke@435 | 2225 | // | stmt2 stmt2 | |
duke@435 | 2226 | // | | | | |
duke@435 | 2227 | // | v v | |
duke@435 | 2228 | // | ifA ifA | |
duke@435 | 2229 | // | | \ / | | |
duke@435 | 2230 | // | v v v v | |
duke@435 | 2231 | // ^ true false false true ^ <-- last_peel |
duke@435 | 2232 | // | | ^ \ / | | |
duke@435 | 2233 | // | cut==|== \ \ / ===|==cut | |
duke@435 | 2234 | // | stmt3 \ \ / stmt3 | <-- first_not_peel |
duke@435 | 2235 | // | | dom | | | | |
duke@435 | 2236 | // | v \ 1v v2 v | |
duke@435 | 2237 | // | ifB regionA ifB | |
duke@435 | 2238 | // | / \ | / \ | |
duke@435 | 2239 | // | / \ v / \ | |
duke@435 | 2240 | // | v v exitA: v v | |
duke@435 | 2241 | // | true false false true | |
duke@435 | 2242 | // | / ^ \ / \ | |
duke@435 | 2243 | // +---- \ \ / ----+ |
duke@435 | 2244 | // dom \ / |
duke@435 | 2245 | // \ 1v v2 |
duke@435 | 2246 | // regionB |
duke@435 | 2247 | // | |
duke@435 | 2248 | // v |
duke@435 | 2249 | // exitB: |
duke@435 | 2250 | // stmt4 |
duke@435 | 2251 | // |
duke@435 | 2252 | // |
duke@435 | 2253 | // after partial peel |
duke@435 | 2254 | // |
duke@435 | 2255 | // stmt1 |
kvn@2727 | 2256 | // | |
kvn@2727 | 2257 | // v |
kvn@2727 | 2258 | // loop predicate |
duke@435 | 2259 | // / |
duke@435 | 2260 | // clone / orig |
duke@435 | 2261 | // / TOP |
duke@435 | 2262 | // / \ |
duke@435 | 2263 | // v v |
kvn@2727 | 2264 | // TOP->loop loop----+ |
duke@435 | 2265 | // | | | |
duke@435 | 2266 | // stmt2 stmt2 | |
duke@435 | 2267 | // | | | |
duke@435 | 2268 | // v v | |
duke@435 | 2269 | // ifA ifA | |
duke@435 | 2270 | // | \ / | | |
duke@435 | 2271 | // v v v v | |
duke@435 | 2272 | // true false false true | <-- last_peel |
duke@435 | 2273 | // | ^ \ / +------|---+ |
duke@435 | 2274 | // +->newloop \ \ / === ==cut | | |
duke@435 | 2275 | // | stmt3 \ \ / TOP | | |
duke@435 | 2276 | // | | dom | | stmt3 | | <-- first_not_peel |
duke@435 | 2277 | // | v \ 1v v2 v | | |
duke@435 | 2278 | // | ifB regionA ifB ^ v |
duke@435 | 2279 | // | / \ | / \ | | |
duke@435 | 2280 | // | / \ v / \ | | |
duke@435 | 2281 | // | v v exitA: v v | | |
duke@435 | 2282 | // | true false false true | | |
duke@435 | 2283 | // | / ^ \ / \ | | |
duke@435 | 2284 | // | | \ \ / v | | |
duke@435 | 2285 | // | | dom \ / TOP | | |
duke@435 | 2286 | // | | \ 1v v2 | | |
duke@435 | 2287 | // ^ v regionB | | |
duke@435 | 2288 | // | | | | | |
duke@435 | 2289 | // | | v ^ v |
duke@435 | 2290 | // | | exitB: | | |
duke@435 | 2291 | // | | stmt4 | | |
duke@435 | 2292 | // | +------------>-----------------+ | |
duke@435 | 2293 | // | | |
duke@435 | 2294 | // +-----------------<---------------------+ |
duke@435 | 2295 | // |
duke@435 | 2296 | // |
duke@435 | 2297 | // final graph |
duke@435 | 2298 | // |
duke@435 | 2299 | // stmt1 |
duke@435 | 2300 | // | |
duke@435 | 2301 | // v |
kvn@2813 | 2302 | // loop predicate |
kvn@2813 | 2303 | // | |
kvn@2813 | 2304 | // v |
kvn@2727 | 2305 | // stmt2 clone |
kvn@2727 | 2306 | // | |
kvn@2727 | 2307 | // v |
duke@435 | 2308 | // ........> ifA clone |
duke@435 | 2309 | // : / | |
duke@435 | 2310 | // dom / | |
duke@435 | 2311 | // : v v |
duke@435 | 2312 | // : false true |
duke@435 | 2313 | // : | | |
kvn@2727 | 2314 | // : | v |
duke@435 | 2315 | // : | newloop<-----+ |
duke@435 | 2316 | // : | | | |
duke@435 | 2317 | // : | stmt3 clone | |
duke@435 | 2318 | // : | | | |
duke@435 | 2319 | // : | v | |
duke@435 | 2320 | // : | ifB | |
duke@435 | 2321 | // : | / \ | |
duke@435 | 2322 | // : | v v | |
duke@435 | 2323 | // : | false true | |
duke@435 | 2324 | // : | | | | |
duke@435 | 2325 | // : | v stmt2 | |
duke@435 | 2326 | // : | exitB: | | |
duke@435 | 2327 | // : | stmt4 v | |
duke@435 | 2328 | // : | ifA orig | |
duke@435 | 2329 | // : | / \ | |
duke@435 | 2330 | // : | / \ | |
duke@435 | 2331 | // : | v v | |
duke@435 | 2332 | // : | false true | |
duke@435 | 2333 | // : | / \ | |
duke@435 | 2334 | // : v v -----+ |
duke@435 | 2335 | // RegionA |
duke@435 | 2336 | // | |
duke@435 | 2337 | // v |
duke@435 | 2338 | // exitA |
duke@435 | 2339 | // |
duke@435 | 2340 | bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) { |
duke@435 | 2341 | |
kvn@2727 | 2342 | assert(!loop->_head->is_CountedLoop(), "Non-counted loop only"); |
rasbold@543 | 2343 | if (!loop->_head->is_Loop()) { |
rasbold@543 | 2344 | return false; } |
rasbold@543 | 2345 | |
duke@435 | 2346 | LoopNode *head = loop->_head->as_Loop(); |
duke@435 | 2347 | |
duke@435 | 2348 | if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) { |
duke@435 | 2349 | return false; |
duke@435 | 2350 | } |
duke@435 | 2351 | |
duke@435 | 2352 | // Check for complex exit control |
duke@435 | 2353 | for(uint ii = 0; ii < loop->_body.size(); ii++ ) { |
duke@435 | 2354 | Node *n = loop->_body.at(ii); |
duke@435 | 2355 | int opc = n->Opcode(); |
duke@435 | 2356 | if (n->is_Call() || |
duke@435 | 2357 | opc == Op_Catch || |
duke@435 | 2358 | opc == Op_CatchProj || |
duke@435 | 2359 | opc == Op_Jump || |
rbackman@6375 | 2360 | opc == Op_JumpProj) { |
duke@435 | 2361 | #if !defined(PRODUCT) |
duke@435 | 2362 | if (TracePartialPeeling) { |
duke@435 | 2363 | tty->print_cr("\nExit control too complex: lp: %d", head->_idx); |
duke@435 | 2364 | } |
duke@435 | 2365 | #endif |
duke@435 | 2366 | return false; |
duke@435 | 2367 | } |
duke@435 | 2368 | } |
duke@435 | 2369 | |
duke@435 | 2370 | int dd = dom_depth(head); |
duke@435 | 2371 | |
duke@435 | 2372 | // Step 1: find cut point |
duke@435 | 2373 | |
duke@435 | 2374 | // Walk up dominators to loop head looking for first loop exit |
duke@435 | 2375 | // which is executed on every path thru loop. |
duke@435 | 2376 | IfNode *peel_if = NULL; |
duke@435 | 2377 | IfNode *peel_if_cmpu = NULL; |
duke@435 | 2378 | |
duke@435 | 2379 | Node *iff = loop->tail(); |
duke@435 | 2380 | while( iff != head ) { |
duke@435 | 2381 | if( iff->is_If() ) { |
duke@435 | 2382 | Node *ctrl = get_ctrl(iff->in(1)); |
duke@435 | 2383 | if (ctrl->is_top()) return false; // Dead test on live IF. |
duke@435 | 2384 | // If loop-varying exit-test, check for induction variable |
duke@435 | 2385 | if( loop->is_member(get_loop(ctrl)) && |
duke@435 | 2386 | loop->is_loop_exit(iff) && |
duke@435 | 2387 | is_possible_iv_test(iff)) { |
duke@435 | 2388 | Node* cmp = iff->in(1)->in(1); |
duke@435 | 2389 | if (cmp->Opcode() == Op_CmpI) { |
duke@435 | 2390 | peel_if = iff->as_If(); |
duke@435 | 2391 | } else { |
duke@435 | 2392 | assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU"); |
duke@435 | 2393 | peel_if_cmpu = iff->as_If(); |
duke@435 | 2394 | } |
duke@435 | 2395 | } |
duke@435 | 2396 | } |
duke@435 | 2397 | iff = idom(iff); |
duke@435 | 2398 | } |
duke@435 | 2399 | // Prefer signed compare over unsigned compare. |
duke@435 | 2400 | IfNode* new_peel_if = NULL; |
duke@435 | 2401 | if (peel_if == NULL) { |
duke@435 | 2402 | if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) { |
duke@435 | 2403 | return false; // No peel point found |
duke@435 | 2404 | } |
duke@435 | 2405 | new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop); |
duke@435 | 2406 | if (new_peel_if == NULL) { |
duke@435 | 2407 | return false; // No peel point found |
duke@435 | 2408 | } |
duke@435 | 2409 | peel_if = new_peel_if; |
duke@435 | 2410 | } |
duke@435 | 2411 | Node* last_peel = stay_in_loop(peel_if, loop); |
duke@435 | 2412 | Node* first_not_peeled = stay_in_loop(last_peel, loop); |
duke@435 | 2413 | if (first_not_peeled == NULL || first_not_peeled == head) { |
duke@435 | 2414 | return false; |
duke@435 | 2415 | } |
duke@435 | 2416 | |
duke@435 | 2417 | #if !defined(PRODUCT) |
kvn@2665 | 2418 | if (TraceLoopOpts) { |
kvn@2665 | 2419 | tty->print("PartialPeel "); |
kvn@2665 | 2420 | loop->dump_head(); |
kvn@2665 | 2421 | } |
kvn@2665 | 2422 | |
duke@435 | 2423 | if (TracePartialPeeling) { |
duke@435 | 2424 | tty->print_cr("before partial peel one iteration"); |
duke@435 | 2425 | Node_List wl; |
duke@435 | 2426 | Node* t = head->in(2); |
duke@435 | 2427 | while (true) { |
duke@435 | 2428 | wl.push(t); |
duke@435 | 2429 | if (t == head) break; |
duke@435 | 2430 | t = idom(t); |
duke@435 | 2431 | } |
duke@435 | 2432 | while (wl.size() > 0) { |
duke@435 | 2433 | Node* tt = wl.pop(); |
duke@435 | 2434 | tt->dump(); |
duke@435 | 2435 | if (tt == last_peel) tty->print_cr("-- cut --"); |
duke@435 | 2436 | } |
duke@435 | 2437 | } |
duke@435 | 2438 | #endif |
duke@435 | 2439 | ResourceArea *area = Thread::current()->resource_area(); |
duke@435 | 2440 | VectorSet peel(area); |
duke@435 | 2441 | VectorSet not_peel(area); |
duke@435 | 2442 | Node_List peel_list(area); |
duke@435 | 2443 | Node_List worklist(area); |
duke@435 | 2444 | Node_List sink_list(area); |
duke@435 | 2445 | |
duke@435 | 2446 | // Set of cfg nodes to peel are those that are executable from |
duke@435 | 2447 | // the head through last_peel. |
duke@435 | 2448 | assert(worklist.size() == 0, "should be empty"); |
duke@435 | 2449 | worklist.push(head); |
duke@435 | 2450 | peel.set(head->_idx); |
duke@435 | 2451 | while (worklist.size() > 0) { |
duke@435 | 2452 | Node *n = worklist.pop(); |
duke@435 | 2453 | if (n != last_peel) { |
duke@435 | 2454 | for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { |
duke@435 | 2455 | Node* use = n->fast_out(j); |
duke@435 | 2456 | if (use->is_CFG() && |
duke@435 | 2457 | loop->is_member(get_loop(use)) && |
duke@435 | 2458 | !peel.test_set(use->_idx)) { |
duke@435 | 2459 | worklist.push(use); |
duke@435 | 2460 | } |
duke@435 | 2461 | } |
duke@435 | 2462 | } |
duke@435 | 2463 | } |
duke@435 | 2464 | |
duke@435 | 2465 | // Set of non-cfg nodes to peel are those that are control |
duke@435 | 2466 | // dependent on the cfg nodes. |
duke@435 | 2467 | uint i; |
duke@435 | 2468 | for(i = 0; i < loop->_body.size(); i++ ) { |
duke@435 | 2469 | Node *n = loop->_body.at(i); |
duke@435 | 2470 | Node *n_c = has_ctrl(n) ? get_ctrl(n) : n; |
duke@435 | 2471 | if (peel.test(n_c->_idx)) { |
duke@435 | 2472 | peel.set(n->_idx); |
duke@435 | 2473 | } else { |
duke@435 | 2474 | not_peel.set(n->_idx); |
duke@435 | 2475 | } |
duke@435 | 2476 | } |
duke@435 | 2477 | |
duke@435 | 2478 | // Step 2: move operations from the peeled section down into the |
duke@435 | 2479 | // not-peeled section |
duke@435 | 2480 | |
duke@435 | 2481 | // Get a post order schedule of nodes in the peel region |
duke@435 | 2482 | // Result in right-most operand. |
duke@435 | 2483 | scheduled_nodelist(loop, peel, peel_list ); |
duke@435 | 2484 | |
duke@435 | 2485 | assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); |
duke@435 | 2486 | |
duke@435 | 2487 | // For future check for too many new phis |
duke@435 | 2488 | uint old_phi_cnt = 0; |
duke@435 | 2489 | for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { |
duke@435 | 2490 | Node* use = head->fast_out(j); |
duke@435 | 2491 | if (use->is_Phi()) old_phi_cnt++; |
duke@435 | 2492 | } |
duke@435 | 2493 | |
duke@435 | 2494 | #if !defined(PRODUCT) |
duke@435 | 2495 | if (TracePartialPeeling) { |
duke@435 | 2496 | tty->print_cr("\npeeled list"); |
duke@435 | 2497 | } |
duke@435 | 2498 | #endif |
duke@435 | 2499 | |
duke@435 | 2500 | // Evacuate nodes in peel region into the not_peeled region if possible |
duke@435 | 2501 | uint new_phi_cnt = 0; |
kvn@5154 | 2502 | uint cloned_for_outside_use = 0; |
duke@435 | 2503 | for (i = 0; i < peel_list.size();) { |
duke@435 | 2504 | Node* n = peel_list.at(i); |
duke@435 | 2505 | #if !defined(PRODUCT) |
duke@435 | 2506 | if (TracePartialPeeling) n->dump(); |
duke@435 | 2507 | #endif |
duke@435 | 2508 | bool incr = true; |
duke@435 | 2509 | if ( !n->is_CFG() ) { |
duke@435 | 2510 | |
duke@435 | 2511 | if ( has_use_in_set(n, not_peel) ) { |
duke@435 | 2512 | |
duke@435 | 2513 | // If not used internal to the peeled region, |
duke@435 | 2514 | // move "n" from peeled to not_peeled region. |
duke@435 | 2515 | |
duke@435 | 2516 | if ( !has_use_internal_to_set(n, peel, loop) ) { |
duke@435 | 2517 | |
duke@435 | 2518 | // if not pinned and not a load (which maybe anti-dependent on a store) |
duke@435 | 2519 | // and not a CMove (Matcher expects only bool->cmove). |
duke@435 | 2520 | if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) { |
kvn@5154 | 2521 | cloned_for_outside_use += clone_for_use_outside_loop( loop, n, worklist ); |
duke@435 | 2522 | sink_list.push(n); |
duke@435 | 2523 | peel >>= n->_idx; // delete n from peel set. |
duke@435 | 2524 | not_peel <<= n->_idx; // add n to not_peel set. |
duke@435 | 2525 | peel_list.remove(i); |
duke@435 | 2526 | incr = false; |
duke@435 | 2527 | #if !defined(PRODUCT) |
duke@435 | 2528 | if (TracePartialPeeling) { |
duke@435 | 2529 | tty->print_cr("sink to not_peeled region: %d newbb: %d", |
duke@435 | 2530 | n->_idx, get_ctrl(n)->_idx); |
duke@435 | 2531 | } |
duke@435 | 2532 | #endif |
duke@435 | 2533 | } |
duke@435 | 2534 | } else { |
duke@435 | 2535 | // Otherwise check for special def-use cases that span |
duke@435 | 2536 | // the peel/not_peel boundary such as bool->if |
duke@435 | 2537 | clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist ); |
duke@435 | 2538 | new_phi_cnt++; |
duke@435 | 2539 | } |
duke@435 | 2540 | } |
duke@435 | 2541 | } |
duke@435 | 2542 | if (incr) i++; |
duke@435 | 2543 | } |
duke@435 | 2544 | |
duke@435 | 2545 | if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) { |
duke@435 | 2546 | #if !defined(PRODUCT) |
duke@435 | 2547 | if (TracePartialPeeling) { |
duke@435 | 2548 | tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c", |
duke@435 | 2549 | new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F'); |
duke@435 | 2550 | } |
duke@435 | 2551 | #endif |
duke@435 | 2552 | if (new_peel_if != NULL) { |
duke@435 | 2553 | remove_cmpi_loop_exit(new_peel_if, loop); |
duke@435 | 2554 | } |
duke@435 | 2555 | // Inhibit more partial peeling on this loop |
duke@435 | 2556 | assert(!head->is_partial_peel_loop(), "not partial peeled"); |
duke@435 | 2557 | head->mark_partial_peel_failed(); |
kvn@5154 | 2558 | if (cloned_for_outside_use > 0) { |
kvn@5154 | 2559 | // Terminate this round of loop opts because |
kvn@5154 | 2560 | // the graph outside this loop was changed. |
kvn@5154 | 2561 | C->set_major_progress(); |
kvn@5154 | 2562 | return true; |
kvn@5154 | 2563 | } |
duke@435 | 2564 | return false; |
duke@435 | 2565 | } |
duke@435 | 2566 | |
duke@435 | 2567 | // Step 3: clone loop, retarget control, and insert new phis |
duke@435 | 2568 | |
duke@435 | 2569 | // Create new loop head for new phis and to hang |
duke@435 | 2570 | // the nodes being moved (sinked) from the peel region. |
kvn@4115 | 2571 | LoopNode* new_head = new (C) LoopNode(last_peel, last_peel); |
kvn@2665 | 2572 | new_head->set_unswitch_count(head->unswitch_count()); // Preserve |
duke@435 | 2573 | _igvn.register_new_node_with_optimizer(new_head); |
duke@435 | 2574 | assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled"); |
duke@435 | 2575 | first_not_peeled->set_req(0, new_head); |
duke@435 | 2576 | set_loop(new_head, loop); |
duke@435 | 2577 | loop->_body.push(new_head); |
duke@435 | 2578 | not_peel.set(new_head->_idx); |
duke@435 | 2579 | set_idom(new_head, last_peel, dom_depth(first_not_peeled)); |
duke@435 | 2580 | set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled)); |
duke@435 | 2581 | |
duke@435 | 2582 | while (sink_list.size() > 0) { |
duke@435 | 2583 | Node* n = sink_list.pop(); |
duke@435 | 2584 | set_ctrl(n, new_head); |
duke@435 | 2585 | } |
duke@435 | 2586 | |
duke@435 | 2587 | assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); |
duke@435 | 2588 | |
duke@435 | 2589 | clone_loop( loop, old_new, dd ); |
duke@435 | 2590 | |
duke@435 | 2591 | const uint clone_exit_idx = 1; |
duke@435 | 2592 | const uint orig_exit_idx = 2; |
duke@435 | 2593 | assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop"); |
duke@435 | 2594 | |
duke@435 | 2595 | Node* head_clone = old_new[head->_idx]; |
duke@435 | 2596 | LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop(); |
duke@435 | 2597 | Node* orig_tail_clone = head_clone->in(2); |
duke@435 | 2598 | |
duke@435 | 2599 | // Add phi if "def" node is in peel set and "use" is not |
duke@435 | 2600 | |
duke@435 | 2601 | for(i = 0; i < peel_list.size(); i++ ) { |
duke@435 | 2602 | Node *def = peel_list.at(i); |
duke@435 | 2603 | if (!def->is_CFG()) { |
duke@435 | 2604 | for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { |
duke@435 | 2605 | Node *use = def->fast_out(j); |
duke@435 | 2606 | if (has_node(use) && use->in(0) != C->top() && |
duke@435 | 2607 | (!peel.test(use->_idx) || |
duke@435 | 2608 | (use->is_Phi() && use->in(0) == head)) ) { |
duke@435 | 2609 | worklist.push(use); |
duke@435 | 2610 | } |
duke@435 | 2611 | } |
duke@435 | 2612 | while( worklist.size() ) { |
duke@435 | 2613 | Node *use = worklist.pop(); |
duke@435 | 2614 | for (uint j = 1; j < use->req(); j++) { |
duke@435 | 2615 | Node* n = use->in(j); |
duke@435 | 2616 | if (n == def) { |
duke@435 | 2617 | |
duke@435 | 2618 | // "def" is in peel set, "use" is not in peel set |
duke@435 | 2619 | // or "use" is in the entry boundary (a phi) of the peel set |
duke@435 | 2620 | |
duke@435 | 2621 | Node* use_c = has_ctrl(use) ? get_ctrl(use) : use; |
duke@435 | 2622 | |
duke@435 | 2623 | if ( loop->is_member(get_loop( use_c )) ) { |
duke@435 | 2624 | // use is in loop |
duke@435 | 2625 | if (old_new[use->_idx] != NULL) { // null for dead code |
duke@435 | 2626 | Node* use_clone = old_new[use->_idx]; |
kvn@3847 | 2627 | _igvn.replace_input_of(use, j, C->top()); |
duke@435 | 2628 | insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone ); |
duke@435 | 2629 | } |
duke@435 | 2630 | } else { |
duke@435 | 2631 | assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format"); |
duke@435 | 2632 | // use is not in the loop, check if the live range includes the cut |
duke@435 | 2633 | Node* lp_if = use_c->in(orig_exit_idx)->in(0); |
duke@435 | 2634 | if (not_peel.test(lp_if->_idx)) { |
duke@435 | 2635 | assert(j == orig_exit_idx, "use from original loop"); |
duke@435 | 2636 | insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone ); |
duke@435 | 2637 | } |
duke@435 | 2638 | } |
duke@435 | 2639 | } |
duke@435 | 2640 | } |
duke@435 | 2641 | } |
duke@435 | 2642 | } |
duke@435 | 2643 | } |
duke@435 | 2644 | |
duke@435 | 2645 | // Step 3b: retarget control |
duke@435 | 2646 | |
duke@435 | 2647 | // Redirect control to the new loop head if a cloned node in |
duke@435 | 2648 | // the not_peeled region has control that points into the peeled region. |
duke@435 | 2649 | // This necessary because the cloned peeled region will be outside |
duke@435 | 2650 | // the loop. |
duke@435 | 2651 | // from to |
duke@435 | 2652 | // cloned-peeled <---+ |
duke@435 | 2653 | // new_head_clone: | <--+ |
duke@435 | 2654 | // cloned-not_peeled in(0) in(0) |
duke@435 | 2655 | // orig-peeled |
duke@435 | 2656 | |
duke@435 | 2657 | for(i = 0; i < loop->_body.size(); i++ ) { |
duke@435 | 2658 | Node *n = loop->_body.at(i); |
duke@435 | 2659 | if (!n->is_CFG() && n->in(0) != NULL && |
duke@435 | 2660 | not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) { |
duke@435 | 2661 | Node* n_clone = old_new[n->_idx]; |
kvn@3847 | 2662 | _igvn.replace_input_of(n_clone, 0, new_head_clone); |
duke@435 | 2663 | } |
duke@435 | 2664 | } |
duke@435 | 2665 | |
duke@435 | 2666 | // Backedge of the surviving new_head (the clone) is original last_peel |
kvn@3847 | 2667 | _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel); |
duke@435 | 2668 | |
duke@435 | 2669 | // Cut first node in original not_peel set |
kvn@3847 | 2670 | _igvn.rehash_node_delayed(new_head); // Multiple edge updates: |
kvn@3847 | 2671 | new_head->set_req(LoopNode::EntryControl, C->top()); // use rehash_node_delayed / set_req instead of |
kvn@3847 | 2672 | new_head->set_req(LoopNode::LoopBackControl, C->top()); // multiple replace_input_of calls |
duke@435 | 2673 | |
duke@435 | 2674 | // Copy head_clone back-branch info to original head |
duke@435 | 2675 | // and remove original head's loop entry and |
duke@435 | 2676 | // clone head's back-branch |
kvn@3847 | 2677 | _igvn.rehash_node_delayed(head); // Multiple edge updates |
kvn@3847 | 2678 | head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl)); |
duke@435 | 2679 | head->set_req(LoopNode::LoopBackControl, C->top()); |
kvn@3847 | 2680 | _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top()); |
duke@435 | 2681 | |
duke@435 | 2682 | // Similarly modify the phis |
duke@435 | 2683 | for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) { |
duke@435 | 2684 | Node* use = head->fast_out(k); |
duke@435 | 2685 | if (use->is_Phi() && use->outcnt() > 0) { |
duke@435 | 2686 | Node* use_clone = old_new[use->_idx]; |
kvn@3847 | 2687 | _igvn.rehash_node_delayed(use); // Multiple edge updates |
kvn@3847 | 2688 | use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl)); |
duke@435 | 2689 | use->set_req(LoopNode::LoopBackControl, C->top()); |
kvn@3847 | 2690 | _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top()); |
duke@435 | 2691 | } |
duke@435 | 2692 | } |
duke@435 | 2693 | |
duke@435 | 2694 | // Step 4: update dominator tree and dominator depth |
duke@435 | 2695 | |
duke@435 | 2696 | set_idom(head, orig_tail_clone, dd); |
duke@435 | 2697 | recompute_dom_depth(); |
duke@435 | 2698 | |
duke@435 | 2699 | // Inhibit more partial peeling on this loop |
duke@435 | 2700 | new_head_clone->set_partial_peel_loop(); |
duke@435 | 2701 | C->set_major_progress(); |
kvn@6630 | 2702 | loop->record_for_igvn(); |
duke@435 | 2703 | |
duke@435 | 2704 | #if !defined(PRODUCT) |
duke@435 | 2705 | if (TracePartialPeeling) { |
duke@435 | 2706 | tty->print_cr("\nafter partial peel one iteration"); |
duke@435 | 2707 | Node_List wl(area); |
duke@435 | 2708 | Node* t = last_peel; |
duke@435 | 2709 | while (true) { |
duke@435 | 2710 | wl.push(t); |
duke@435 | 2711 | if (t == head_clone) break; |
duke@435 | 2712 | t = idom(t); |
duke@435 | 2713 | } |
duke@435 | 2714 | while (wl.size() > 0) { |
duke@435 | 2715 | Node* tt = wl.pop(); |
duke@435 | 2716 | if (tt == head) tty->print_cr("orig head"); |
duke@435 | 2717 | else if (tt == new_head_clone) tty->print_cr("new head"); |
duke@435 | 2718 | else if (tt == head_clone) tty->print_cr("clone head"); |
duke@435 | 2719 | tt->dump(); |
duke@435 | 2720 | } |
duke@435 | 2721 | } |
duke@435 | 2722 | #endif |
duke@435 | 2723 | return true; |
duke@435 | 2724 | } |
duke@435 | 2725 | |
duke@435 | 2726 | //------------------------------reorg_offsets---------------------------------- |
duke@435 | 2727 | // Reorganize offset computations to lower register pressure. Mostly |
duke@435 | 2728 | // prevent loop-fallout uses of the pre-incremented trip counter (which are |
duke@435 | 2729 | // then alive with the post-incremented trip counter forcing an extra |
duke@435 | 2730 | // register move) |
kvn@2665 | 2731 | void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) { |
kvn@2665 | 2732 | // Perform it only for canonical counted loops. |
kvn@2665 | 2733 | // Loop's shape could be messed up by iteration_split_impl. |
kvn@2665 | 2734 | if (!loop->_head->is_CountedLoop()) |
kvn@2665 | 2735 | return; |
kvn@2665 | 2736 | if (!loop->_head->as_Loop()->is_valid_counted_loop()) |
kvn@2665 | 2737 | return; |
duke@435 | 2738 | |
duke@435 | 2739 | CountedLoopNode *cl = loop->_head->as_CountedLoop(); |
duke@435 | 2740 | CountedLoopEndNode *cle = cl->loopexit(); |
duke@435 | 2741 | Node *exit = cle->proj_out(false); |
kvn@2665 | 2742 | Node *phi = cl->phi(); |
duke@435 | 2743 | |
duke@435 | 2744 | // Check for the special case of folks using the pre-incremented |
duke@435 | 2745 | // trip-counter on the fall-out path (forces the pre-incremented |
duke@435 | 2746 | // and post-incremented trip counter to be live at the same time). |
duke@435 | 2747 | // Fix this by adjusting to use the post-increment trip counter. |
never@802 | 2748 | |
duke@435 | 2749 | bool progress = true; |
duke@435 | 2750 | while (progress) { |
duke@435 | 2751 | progress = false; |
duke@435 | 2752 | for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) { |
duke@435 | 2753 | Node* use = phi->fast_out(i); // User of trip-counter |
duke@435 | 2754 | if (!has_ctrl(use)) continue; |
duke@435 | 2755 | Node *u_ctrl = get_ctrl(use); |
kvn@2665 | 2756 | if (use->is_Phi()) { |
duke@435 | 2757 | u_ctrl = NULL; |
kvn@2665 | 2758 | for (uint j = 1; j < use->req(); j++) |
kvn@2665 | 2759 | if (use->in(j) == phi) |
kvn@2665 | 2760 | u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j)); |
duke@435 | 2761 | } |
duke@435 | 2762 | IdealLoopTree *u_loop = get_loop(u_ctrl); |
duke@435 | 2763 | // Look for loop-invariant use |
kvn@2665 | 2764 | if (u_loop == loop) continue; |
kvn@2665 | 2765 | if (loop->is_member(u_loop)) continue; |
duke@435 | 2766 | // Check that use is live out the bottom. Assuming the trip-counter |
duke@435 | 2767 | // update is right at the bottom, uses of of the loop middle are ok. |
kvn@2665 | 2768 | if (dom_lca(exit, u_ctrl) != exit) continue; |
duke@435 | 2769 | // Hit! Refactor use to use the post-incremented tripcounter. |
duke@435 | 2770 | // Compute a post-increment tripcounter. |
kvn@4115 | 2771 | Node *opaq = new (C) Opaque2Node( C, cle->incr() ); |
kvn@6728 | 2772 | register_new_node(opaq, exit); |
duke@435 | 2773 | Node *neg_stride = _igvn.intcon(-cle->stride_con()); |
duke@435 | 2774 | set_ctrl(neg_stride, C->root()); |
kvn@4115 | 2775 | Node *post = new (C) AddINode( opaq, neg_stride); |
kvn@6728 | 2776 | register_new_node(post, exit); |
kvn@3847 | 2777 | _igvn.rehash_node_delayed(use); |
kvn@2665 | 2778 | for (uint j = 1; j < use->req(); j++) { |
kvn@2665 | 2779 | if (use->in(j) == phi) |
duke@435 | 2780 | use->set_req(j, post); |
kvn@2665 | 2781 | } |
duke@435 | 2782 | // Since DU info changed, rerun loop |
duke@435 | 2783 | progress = true; |
duke@435 | 2784 | break; |
duke@435 | 2785 | } |
duke@435 | 2786 | } |
duke@435 | 2787 | |
duke@435 | 2788 | } |