1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/src/share/vm/opto/loopopts.cpp Sat Dec 01 00:00:00 2007 +0000
1.3 @@ -0,0 +1,2677 @@
1.4 +/*
1.5 + * Copyright 1999-2006 Sun Microsystems, Inc. All Rights Reserved.
1.6 + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
1.7 + *
1.8 + * This code is free software; you can redistribute it and/or modify it
1.9 + * under the terms of the GNU General Public License version 2 only, as
1.10 + * published by the Free Software Foundation.
1.11 + *
1.12 + * This code is distributed in the hope that it will be useful, but WITHOUT
1.13 + * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
1.14 + * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
1.15 + * version 2 for more details (a copy is included in the LICENSE file that
1.16 + * accompanied this code).
1.17 + *
1.18 + * You should have received a copy of the GNU General Public License version
1.19 + * 2 along with this work; if not, write to the Free Software Foundation,
1.20 + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
1.21 + *
1.22 + * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
1.23 + * CA 95054 USA or visit www.sun.com if you need additional information or
1.24 + * have any questions.
1.25 + *
1.26 + */
1.27 +
1.28 +#include "incls/_precompiled.incl"
1.29 +#include "incls/_loopopts.cpp.incl"
1.30 +
1.31 +//=============================================================================
1.32 +//------------------------------split_thru_phi---------------------------------
1.33 +// Split Node 'n' through merge point if there is enough win.
1.34 +Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) {
1.35 + int wins = 0;
1.36 + assert( !n->is_CFG(), "" );
1.37 + assert( region->is_Region(), "" );
1.38 + Node *phi = new (C, region->req()) PhiNode( region, n->bottom_type() );
1.39 + uint old_unique = C->unique();
1.40 + for( uint i = 1; i < region->req(); i++ ) {
1.41 + Node *x;
1.42 + Node* the_clone = NULL;
1.43 + if( region->in(i) == C->top() ) {
1.44 + x = C->top(); // Dead path? Use a dead data op
1.45 + } else {
1.46 + x = n->clone(); // Else clone up the data op
1.47 + the_clone = x; // Remember for possible deletion.
1.48 + // Alter data node to use pre-phi inputs
1.49 + if( n->in(0) == region )
1.50 + x->set_req( 0, region->in(i) );
1.51 + for( uint j = 1; j < n->req(); j++ ) {
1.52 + Node *in = n->in(j);
1.53 + if( in->is_Phi() && in->in(0) == region )
1.54 + x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone
1.55 + }
1.56 + }
1.57 + // Check for a 'win' on some paths
1.58 + const Type *t = x->Value(&_igvn);
1.59 +
1.60 + bool singleton = t->singleton();
1.61 +
1.62 + // A TOP singleton indicates that there are no possible values incoming
1.63 + // along a particular edge. In most cases, this is OK, and the Phi will
1.64 + // be eliminated later in an Ideal call. However, we can't allow this to
1.65 + // happen if the singleton occurs on loop entry, as the elimination of
1.66 + // the PhiNode may cause the resulting node to migrate back to a previous
1.67 + // loop iteration.
1.68 + if( singleton && t == Type::TOP ) {
1.69 + // Is_Loop() == false does not confirm the absence of a loop (e.g., an
1.70 + // irreducible loop may not be indicated by an affirmative is_Loop());
1.71 + // therefore, the only top we can split thru a phi is on a backedge of
1.72 + // a loop.
1.73 + singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1.74 + }
1.75 +
1.76 + if( singleton ) {
1.77 + wins++;
1.78 + x = ((PhaseGVN&)_igvn).makecon(t);
1.79 + } else {
1.80 + // We now call Identity to try to simplify the cloned node.
1.81 + // Note that some Identity methods call phase->type(this).
1.82 + // Make sure that the type array is big enough for
1.83 + // our new node, even though we may throw the node away.
1.84 + // (Note: This tweaking with igvn only works because x is a new node.)
1.85 + _igvn.set_type(x, t);
1.86 + Node *y = x->Identity(&_igvn);
1.87 + if( y != x ) {
1.88 + wins++;
1.89 + x = y;
1.90 + } else {
1.91 + y = _igvn.hash_find(x);
1.92 + if( y ) {
1.93 + wins++;
1.94 + x = y;
1.95 + } else {
1.96 + // Else x is a new node we are keeping
1.97 + // We do not need register_new_node_with_optimizer
1.98 + // because set_type has already been called.
1.99 + _igvn._worklist.push(x);
1.100 + }
1.101 + }
1.102 + }
1.103 + if (x != the_clone && the_clone != NULL)
1.104 + _igvn.remove_dead_node(the_clone);
1.105 + phi->set_req( i, x );
1.106 + }
1.107 + // Too few wins?
1.108 + if( wins <= policy ) {
1.109 + _igvn.remove_dead_node(phi);
1.110 + return NULL;
1.111 + }
1.112 +
1.113 + // Record Phi
1.114 + register_new_node( phi, region );
1.115 +
1.116 + for( uint i2 = 1; i2 < phi->req(); i2++ ) {
1.117 + Node *x = phi->in(i2);
1.118 + // If we commoned up the cloned 'x' with another existing Node,
1.119 + // the existing Node picks up a new use. We need to make the
1.120 + // existing Node occur higher up so it dominates its uses.
1.121 + Node *old_ctrl;
1.122 + IdealLoopTree *old_loop;
1.123 +
1.124 + // The occasional new node
1.125 + if( x->_idx >= old_unique ) { // Found a new, unplaced node?
1.126 + old_ctrl = x->is_Con() ? C->root() : NULL;
1.127 + old_loop = NULL; // Not in any prior loop
1.128 + } else {
1.129 + old_ctrl = x->is_Con() ? C->root() : get_ctrl(x);
1.130 + old_loop = get_loop(old_ctrl); // Get prior loop
1.131 + }
1.132 + // New late point must dominate new use
1.133 + Node *new_ctrl = dom_lca( old_ctrl, region->in(i2) );
1.134 + // Set new location
1.135 + set_ctrl(x, new_ctrl);
1.136 + IdealLoopTree *new_loop = get_loop( new_ctrl );
1.137 + // If changing loop bodies, see if we need to collect into new body
1.138 + if( old_loop != new_loop ) {
1.139 + if( old_loop && !old_loop->_child )
1.140 + old_loop->_body.yank(x);
1.141 + if( !new_loop->_child )
1.142 + new_loop->_body.push(x); // Collect body info
1.143 + }
1.144 + }
1.145 +
1.146 + return phi;
1.147 +}
1.148 +
1.149 +//------------------------------dominated_by------------------------------------
1.150 +// Replace the dominated test with an obvious true or false. Place it on the
1.151 +// IGVN worklist for later cleanup. Move control-dependent data Nodes on the
1.152 +// live path up to the dominating control.
1.153 +void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff ) {
1.154 +#ifndef PRODUCT
1.155 + if( VerifyLoopOptimizations && PrintOpto ) tty->print_cr("dominating test");
1.156 +#endif
1.157 +
1.158 +
1.159 + // prevdom is the dominating projection of the dominating test.
1.160 + assert( iff->is_If(), "" );
1.161 + assert( iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added");
1.162 + int pop = prevdom->Opcode();
1.163 + assert( pop == Op_IfFalse || pop == Op_IfTrue, "" );
1.164 + // 'con' is set to true or false to kill the dominated test.
1.165 + Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO);
1.166 + set_ctrl(con, C->root()); // Constant gets a new use
1.167 + // Hack the dominated test
1.168 + _igvn.hash_delete(iff);
1.169 + iff->set_req(1, con);
1.170 + _igvn._worklist.push(iff);
1.171 +
1.172 + // If I dont have a reachable TRUE and FALSE path following the IfNode then
1.173 + // I can assume this path reaches an infinite loop. In this case it's not
1.174 + // important to optimize the data Nodes - either the whole compilation will
1.175 + // be tossed or this path (and all data Nodes) will go dead.
1.176 + if( iff->outcnt() != 2 ) return;
1.177 +
1.178 + // Make control-dependent data Nodes on the live path (path that will remain
1.179 + // once the dominated IF is removed) become control-dependent on the
1.180 + // dominating projection.
1.181 + Node* dp = ((IfNode*)iff)->proj_out(pop == Op_IfTrue);
1.182 + IdealLoopTree *old_loop = get_loop(dp);
1.183 +
1.184 + for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
1.185 + Node* cd = dp->fast_out(i); // Control-dependent node
1.186 + if( cd->depends_only_on_test() ) {
1.187 + assert( cd->in(0) == dp, "" );
1.188 + _igvn.hash_delete( cd );
1.189 + cd->set_req(0, prevdom);
1.190 + set_early_ctrl( cd );
1.191 + _igvn._worklist.push(cd);
1.192 + IdealLoopTree *new_loop = get_loop(get_ctrl(cd));
1.193 + if( old_loop != new_loop ) {
1.194 + if( !old_loop->_child ) old_loop->_body.yank(cd);
1.195 + if( !new_loop->_child ) new_loop->_body.push(cd);
1.196 + }
1.197 + --i;
1.198 + --imax;
1.199 + }
1.200 + }
1.201 +}
1.202 +
1.203 +//------------------------------has_local_phi_input----------------------------
1.204 +// Return TRUE if 'n' has Phi inputs from its local block and no other
1.205 +// block-local inputs (all non-local-phi inputs come from earlier blocks)
1.206 +Node *PhaseIdealLoop::has_local_phi_input( Node *n ) {
1.207 + Node *n_ctrl = get_ctrl(n);
1.208 + // See if some inputs come from a Phi in this block, or from before
1.209 + // this block.
1.210 + uint i;
1.211 + for( i = 1; i < n->req(); i++ ) {
1.212 + Node *phi = n->in(i);
1.213 + if( phi->is_Phi() && phi->in(0) == n_ctrl )
1.214 + break;
1.215 + }
1.216 + if( i >= n->req() )
1.217 + return NULL; // No Phi inputs; nowhere to clone thru
1.218 +
1.219 + // Check for inputs created between 'n' and the Phi input. These
1.220 + // must split as well; they have already been given the chance
1.221 + // (courtesy of a post-order visit) and since they did not we must
1.222 + // recover the 'cost' of splitting them by being very profitable
1.223 + // when splitting 'n'. Since this is unlikely we simply give up.
1.224 + for( i = 1; i < n->req(); i++ ) {
1.225 + Node *m = n->in(i);
1.226 + if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) {
1.227 + // We allow the special case of AddP's with no local inputs.
1.228 + // This allows us to split-up address expressions.
1.229 + if (m->is_AddP() &&
1.230 + get_ctrl(m->in(2)) != n_ctrl &&
1.231 + get_ctrl(m->in(3)) != n_ctrl) {
1.232 + // Move the AddP up to dominating point
1.233 + set_ctrl_and_loop(m, find_non_split_ctrl(idom(n_ctrl)));
1.234 + continue;
1.235 + }
1.236 + return NULL;
1.237 + }
1.238 + }
1.239 +
1.240 + return n_ctrl;
1.241 +}
1.242 +
1.243 +//------------------------------remix_address_expressions----------------------
1.244 +// Rework addressing expressions to get the most loop-invariant stuff
1.245 +// moved out. We'd like to do all associative operators, but it's especially
1.246 +// important (common) to do address expressions.
1.247 +Node *PhaseIdealLoop::remix_address_expressions( Node *n ) {
1.248 + if (!has_ctrl(n)) return NULL;
1.249 + Node *n_ctrl = get_ctrl(n);
1.250 + IdealLoopTree *n_loop = get_loop(n_ctrl);
1.251 +
1.252 + // See if 'n' mixes loop-varying and loop-invariant inputs and
1.253 + // itself is loop-varying.
1.254 +
1.255 + // Only interested in binary ops (and AddP)
1.256 + if( n->req() < 3 || n->req() > 4 ) return NULL;
1.257 +
1.258 + Node *n1_ctrl = get_ctrl(n->in( 1));
1.259 + Node *n2_ctrl = get_ctrl(n->in( 2));
1.260 + Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3));
1.261 + IdealLoopTree *n1_loop = get_loop( n1_ctrl );
1.262 + IdealLoopTree *n2_loop = get_loop( n2_ctrl );
1.263 + IdealLoopTree *n3_loop = get_loop( n3_ctrl );
1.264 +
1.265 + // Does one of my inputs spin in a tighter loop than self?
1.266 + if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) ||
1.267 + (n_loop->is_member( n2_loop ) && n_loop != n2_loop) ||
1.268 + (n_loop->is_member( n3_loop ) && n_loop != n3_loop) )
1.269 + return NULL; // Leave well enough alone
1.270 +
1.271 + // Is at least one of my inputs loop-invariant?
1.272 + if( n1_loop == n_loop &&
1.273 + n2_loop == n_loop &&
1.274 + n3_loop == n_loop )
1.275 + return NULL; // No loop-invariant inputs
1.276 +
1.277 +
1.278 + int n_op = n->Opcode();
1.279 +
1.280 + // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2).
1.281 + if( n_op == Op_LShiftI ) {
1.282 + // Scale is loop invariant
1.283 + Node *scale = n->in(2);
1.284 + Node *scale_ctrl = get_ctrl(scale);
1.285 + IdealLoopTree *scale_loop = get_loop(scale_ctrl );
1.286 + if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) )
1.287 + return NULL;
1.288 + const TypeInt *scale_t = scale->bottom_type()->isa_int();
1.289 + if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 )
1.290 + return NULL; // Dont bother with byte/short masking
1.291 + // Add must vary with loop (else shift would be loop-invariant)
1.292 + Node *add = n->in(1);
1.293 + Node *add_ctrl = get_ctrl(add);
1.294 + IdealLoopTree *add_loop = get_loop(add_ctrl);
1.295 + //assert( n_loop == add_loop, "" );
1.296 + if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops
1.297 +
1.298 + // Convert I-V into I+ (0-V); same for V-I
1.299 + if( add->Opcode() == Op_SubI &&
1.300 + _igvn.type( add->in(1) ) != TypeInt::ZERO ) {
1.301 + Node *zero = _igvn.intcon(0);
1.302 + set_ctrl(zero, C->root());
1.303 + Node *neg = new (C, 3) SubINode( _igvn.intcon(0), add->in(2) );
1.304 + register_new_node( neg, get_ctrl(add->in(2) ) );
1.305 + add = new (C, 3) AddINode( add->in(1), neg );
1.306 + register_new_node( add, add_ctrl );
1.307 + }
1.308 + if( add->Opcode() != Op_AddI ) return NULL;
1.309 + // See if one add input is loop invariant
1.310 + Node *add_var = add->in(1);
1.311 + Node *add_var_ctrl = get_ctrl(add_var);
1.312 + IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
1.313 + Node *add_invar = add->in(2);
1.314 + Node *add_invar_ctrl = get_ctrl(add_invar);
1.315 + IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl );
1.316 + if( add_var_loop == n_loop ) {
1.317 + } else if( add_invar_loop == n_loop ) {
1.318 + // Swap to find the invariant part
1.319 + add_invar = add_var;
1.320 + add_invar_ctrl = add_var_ctrl;
1.321 + add_invar_loop = add_var_loop;
1.322 + add_var = add->in(2);
1.323 + Node *add_var_ctrl = get_ctrl(add_var);
1.324 + IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
1.325 + } else // Else neither input is loop invariant
1.326 + return NULL;
1.327 + if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) )
1.328 + return NULL; // No invariant part of the add?
1.329 +
1.330 + // Yes! Reshape address expression!
1.331 + Node *inv_scale = new (C, 3) LShiftINode( add_invar, scale );
1.332 + register_new_node( inv_scale, add_invar_ctrl );
1.333 + Node *var_scale = new (C, 3) LShiftINode( add_var, scale );
1.334 + register_new_node( var_scale, n_ctrl );
1.335 + Node *var_add = new (C, 3) AddINode( var_scale, inv_scale );
1.336 + register_new_node( var_add, n_ctrl );
1.337 + _igvn.hash_delete( n );
1.338 + _igvn.subsume_node( n, var_add );
1.339 + return var_add;
1.340 + }
1.341 +
1.342 + // Replace (I+V) with (V+I)
1.343 + if( n_op == Op_AddI ||
1.344 + n_op == Op_AddL ||
1.345 + n_op == Op_AddF ||
1.346 + n_op == Op_AddD ||
1.347 + n_op == Op_MulI ||
1.348 + n_op == Op_MulL ||
1.349 + n_op == Op_MulF ||
1.350 + n_op == Op_MulD ) {
1.351 + if( n2_loop == n_loop ) {
1.352 + assert( n1_loop != n_loop, "" );
1.353 + n->swap_edges(1, 2);
1.354 + }
1.355 + }
1.356 +
1.357 + // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V),
1.358 + // but not if I2 is a constant.
1.359 + if( n_op == Op_AddP ) {
1.360 + if( n2_loop == n_loop && n3_loop != n_loop ) {
1.361 + if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) {
1.362 + Node *n22_ctrl = get_ctrl(n->in(2)->in(2));
1.363 + Node *n23_ctrl = get_ctrl(n->in(2)->in(3));
1.364 + IdealLoopTree *n22loop = get_loop( n22_ctrl );
1.365 + IdealLoopTree *n23_loop = get_loop( n23_ctrl );
1.366 + if( n22loop != n_loop && n22loop->is_member(n_loop) &&
1.367 + n23_loop == n_loop ) {
1.368 + Node *add1 = new (C, 4) AddPNode( n->in(1), n->in(2)->in(2), n->in(3) );
1.369 + // Stuff new AddP in the loop preheader
1.370 + register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) );
1.371 + Node *add2 = new (C, 4) AddPNode( n->in(1), add1, n->in(2)->in(3) );
1.372 + register_new_node( add2, n_ctrl );
1.373 + _igvn.hash_delete( n );
1.374 + _igvn.subsume_node( n, add2 );
1.375 + return add2;
1.376 + }
1.377 + }
1.378 + }
1.379 +
1.380 + // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V)
1.381 + if( n2_loop != n_loop && n3_loop == n_loop ) {
1.382 + if( n->in(3)->Opcode() == Op_AddI ) {
1.383 + Node *V = n->in(3)->in(1);
1.384 + Node *I = n->in(3)->in(2);
1.385 + if( is_member(n_loop,get_ctrl(V)) ) {
1.386 + } else {
1.387 + Node *tmp = V; V = I; I = tmp;
1.388 + }
1.389 + if( !is_member(n_loop,get_ctrl(I)) ) {
1.390 + Node *add1 = new (C, 4) AddPNode( n->in(1), n->in(2), I );
1.391 + // Stuff new AddP in the loop preheader
1.392 + register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) );
1.393 + Node *add2 = new (C, 4) AddPNode( n->in(1), add1, V );
1.394 + register_new_node( add2, n_ctrl );
1.395 + _igvn.hash_delete( n );
1.396 + _igvn.subsume_node( n, add2 );
1.397 + return add2;
1.398 + }
1.399 + }
1.400 + }
1.401 + }
1.402 +
1.403 + return NULL;
1.404 +}
1.405 +
1.406 +//------------------------------conditional_move-------------------------------
1.407 +// Attempt to replace a Phi with a conditional move. We have some pretty
1.408 +// strict profitability requirements. All Phis at the merge point must
1.409 +// be converted, so we can remove the control flow. We need to limit the
1.410 +// number of c-moves to a small handful. All code that was in the side-arms
1.411 +// of the CFG diamond is now speculatively executed. This code has to be
1.412 +// "cheap enough". We are pretty much limited to CFG diamonds that merge
1.413 +// 1 or 2 items with a total of 1 or 2 ops executed speculatively.
1.414 +Node *PhaseIdealLoop::conditional_move( Node *region ) {
1.415 +
1.416 + assert( region->is_Region(), "sanity check" );
1.417 + if( region->req() != 3 ) return NULL;
1.418 +
1.419 + // Check for CFG diamond
1.420 + Node *lp = region->in(1);
1.421 + Node *rp = region->in(2);
1.422 + if( !lp || !rp ) return NULL;
1.423 + Node *lp_c = lp->in(0);
1.424 + if( lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If() ) return NULL;
1.425 + IfNode *iff = lp_c->as_If();
1.426 +
1.427 + // Check for highly predictable branch. No point in CMOV'ing if
1.428 + // we are going to predict accurately all the time.
1.429 + // %%% This hides patterns produced by utility methods like Math.min.
1.430 + if( iff->_prob < PROB_UNLIKELY_MAG(3) ||
1.431 + iff->_prob > PROB_LIKELY_MAG(3) )
1.432 + return NULL;
1.433 +
1.434 + // Check for ops pinned in an arm of the diamond.
1.435 + // Can't remove the control flow in this case
1.436 + if( lp->outcnt() > 1 ) return NULL;
1.437 + if( rp->outcnt() > 1 ) return NULL;
1.438 +
1.439 + // Check profitability
1.440 + int cost = 0;
1.441 + for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1.442 + Node *out = region->fast_out(i);
1.443 + if( !out->is_Phi() ) continue; // Ignore other control edges, etc
1.444 + PhiNode* phi = out->as_Phi();
1.445 + switch (phi->type()->basic_type()) {
1.446 + case T_LONG:
1.447 + cost++; // Probably encodes as 2 CMOV's
1.448 + case T_INT: // These all CMOV fine
1.449 + case T_FLOAT:
1.450 + case T_DOUBLE:
1.451 + case T_ADDRESS: // (RawPtr)
1.452 + cost++;
1.453 + break;
1.454 + case T_OBJECT: { // Base oops are OK, but not derived oops
1.455 + const TypeOopPtr *tp = phi->type()->isa_oopptr();
1.456 + // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a
1.457 + // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus
1.458 + // CMOVE'ing a derived pointer requires we also CMOVE the base. If we
1.459 + // have a Phi for the base here that we convert to a CMOVE all is well
1.460 + // and good. But if the base is dead, we'll not make a CMOVE. Later
1.461 + // the allocator will have to produce a base by creating a CMOVE of the
1.462 + // relevant bases. This puts the allocator in the business of
1.463 + // manufacturing expensive instructions, generally a bad plan.
1.464 + // Just Say No to Conditionally-Moved Derived Pointers.
1.465 + if( tp && tp->offset() != 0 )
1.466 + return NULL;
1.467 + cost++;
1.468 + break;
1.469 + }
1.470 + default:
1.471 + return NULL; // In particular, can't do memory or I/O
1.472 + }
1.473 + // Add in cost any speculative ops
1.474 + for( uint j = 1; j < region->req(); j++ ) {
1.475 + Node *proj = region->in(j);
1.476 + Node *inp = phi->in(j);
1.477 + if (get_ctrl(inp) == proj) { // Found local op
1.478 + cost++;
1.479 + // Check for a chain of dependent ops; these will all become
1.480 + // speculative in a CMOV.
1.481 + for( uint k = 1; k < inp->req(); k++ )
1.482 + if (get_ctrl(inp->in(k)) == proj)
1.483 + return NULL; // Too much speculative goo
1.484 + }
1.485 + }
1.486 + // See if the Phi is used by a Cmp. This will likely Split-If, a
1.487 + // higher-payoff operation.
1.488 + for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) {
1.489 + Node* use = phi->fast_out(k);
1.490 + if( use->is_Cmp() )
1.491 + return NULL;
1.492 + }
1.493 + }
1.494 + if( cost >= ConditionalMoveLimit ) return NULL; // Too much goo
1.495 +
1.496 + // --------------
1.497 + // Now replace all Phis with CMOV's
1.498 + Node *cmov_ctrl = iff->in(0);
1.499 + uint flip = (lp->Opcode() == Op_IfTrue);
1.500 + while( 1 ) {
1.501 + PhiNode* phi = NULL;
1.502 + for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1.503 + Node *out = region->fast_out(i);
1.504 + if (out->is_Phi()) {
1.505 + phi = out->as_Phi();
1.506 + break;
1.507 + }
1.508 + }
1.509 + if (phi == NULL) break;
1.510 +#ifndef PRODUCT
1.511 + if( PrintOpto && VerifyLoopOptimizations ) tty->print_cr("CMOV");
1.512 +#endif
1.513 + // Move speculative ops
1.514 + for( uint j = 1; j < region->req(); j++ ) {
1.515 + Node *proj = region->in(j);
1.516 + Node *inp = phi->in(j);
1.517 + if (get_ctrl(inp) == proj) { // Found local op
1.518 +#ifndef PRODUCT
1.519 + if( PrintOpto && VerifyLoopOptimizations ) {
1.520 + tty->print(" speculate: ");
1.521 + inp->dump();
1.522 + }
1.523 +#endif
1.524 + set_ctrl(inp, cmov_ctrl);
1.525 + }
1.526 + }
1.527 + Node *cmov = CMoveNode::make( C, cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi) );
1.528 + register_new_node( cmov, cmov_ctrl );
1.529 + _igvn.hash_delete(phi);
1.530 + _igvn.subsume_node( phi, cmov );
1.531 +#ifndef PRODUCT
1.532 + if( VerifyLoopOptimizations ) verify();
1.533 +#endif
1.534 + }
1.535 +
1.536 + // The useless CFG diamond will fold up later; see the optimization in
1.537 + // RegionNode::Ideal.
1.538 + _igvn._worklist.push(region);
1.539 +
1.540 + return iff->in(1);
1.541 +}
1.542 +
1.543 +//------------------------------split_if_with_blocks_pre-----------------------
1.544 +// Do the real work in a non-recursive function. Data nodes want to be
1.545 +// cloned in the pre-order so they can feed each other nicely.
1.546 +Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
1.547 + // Cloning these guys is unlikely to win
1.548 + int n_op = n->Opcode();
1.549 + if( n_op == Op_MergeMem ) return n;
1.550 + if( n->is_Proj() ) return n;
1.551 + // Do not clone-up CmpFXXX variations, as these are always
1.552 + // followed by a CmpI
1.553 + if( n->is_Cmp() ) return n;
1.554 + // Attempt to use a conditional move instead of a phi/branch
1.555 + if( ConditionalMoveLimit > 0 && n_op == Op_Region ) {
1.556 + Node *cmov = conditional_move( n );
1.557 + if( cmov ) return cmov;
1.558 + }
1.559 + if( n->is_CFG() || n_op == Op_StorePConditional || n_op == Op_StoreLConditional || n_op == Op_CompareAndSwapI || n_op == Op_CompareAndSwapL ||n_op == Op_CompareAndSwapP) return n;
1.560 + if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd
1.561 + n_op == Op_Opaque2 ) {
1.562 + if( !C->major_progress() ) // If chance of no more loop opts...
1.563 + _igvn._worklist.push(n); // maybe we'll remove them
1.564 + return n;
1.565 + }
1.566 +
1.567 + if( n->is_Con() ) return n; // No cloning for Con nodes
1.568 +
1.569 + Node *n_ctrl = get_ctrl(n);
1.570 + if( !n_ctrl ) return n; // Dead node
1.571 +
1.572 + // Attempt to remix address expressions for loop invariants
1.573 + Node *m = remix_address_expressions( n );
1.574 + if( m ) return m;
1.575 +
1.576 + // Determine if the Node has inputs from some local Phi.
1.577 + // Returns the block to clone thru.
1.578 + Node *n_blk = has_local_phi_input( n );
1.579 + if( !n_blk ) return n;
1.580 + // Do not clone the trip counter through on a CountedLoop
1.581 + // (messes up the canonical shape).
1.582 + if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n;
1.583 +
1.584 + // Check for having no control input; not pinned. Allow
1.585 + // dominating control.
1.586 + if( n->in(0) ) {
1.587 + Node *dom = idom(n_blk);
1.588 + if( dom_lca( n->in(0), dom ) != n->in(0) )
1.589 + return n;
1.590 + }
1.591 + // Policy: when is it profitable. You must get more wins than
1.592 + // policy before it is considered profitable. Policy is usually 0,
1.593 + // so 1 win is considered profitable. Big merges will require big
1.594 + // cloning, so get a larger policy.
1.595 + int policy = n_blk->req() >> 2;
1.596 +
1.597 + // If the loop is a candidate for range check elimination,
1.598 + // delay splitting through it's phi until a later loop optimization
1.599 + if (n_blk->is_CountedLoop()) {
1.600 + IdealLoopTree *lp = get_loop(n_blk);
1.601 + if (lp && lp->_rce_candidate) {
1.602 + return n;
1.603 + }
1.604 + }
1.605 +
1.606 + // Use same limit as split_if_with_blocks_post
1.607 + if( C->unique() > 35000 ) return n; // Method too big
1.608 +
1.609 + // Split 'n' through the merge point if it is profitable
1.610 + Node *phi = split_thru_phi( n, n_blk, policy );
1.611 + if( !phi ) return n;
1.612 +
1.613 + // Found a Phi to split thru!
1.614 + // Replace 'n' with the new phi
1.615 + _igvn.hash_delete(n);
1.616 + _igvn.subsume_node( n, phi );
1.617 + // Moved a load around the loop, 'en-registering' something.
1.618 + if( n_blk->Opcode() == Op_Loop && n->is_Load() &&
1.619 + !phi->in(LoopNode::LoopBackControl)->is_Load() )
1.620 + C->set_major_progress();
1.621 +
1.622 + return phi;
1.623 +}
1.624 +
1.625 +static bool merge_point_too_heavy(Compile* C, Node* region) {
1.626 + // Bail out if the region and its phis have too many users.
1.627 + int weight = 0;
1.628 + for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1.629 + weight += region->fast_out(i)->outcnt();
1.630 + }
1.631 + int nodes_left = MaxNodeLimit - C->unique();
1.632 + if (weight * 8 > nodes_left) {
1.633 +#ifndef PRODUCT
1.634 + if (PrintOpto)
1.635 + tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight);
1.636 +#endif
1.637 + return true;
1.638 + } else {
1.639 + return false;
1.640 + }
1.641 +}
1.642 +
1.643 +#ifdef _LP64
1.644 +static bool merge_point_safe(Node* region) {
1.645 + // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode
1.646 + // having a PhiNode input. This sidesteps the dangerous case where the split
1.647 + // ConvI2LNode may become TOP if the input Value() does not
1.648 + // overlap the ConvI2L range, leaving a node which may not dominate its
1.649 + // uses.
1.650 + // A better fix for this problem can be found in the BugTraq entry, but
1.651 + // expediency for Mantis demands this hack.
1.652 + for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1.653 + Node* n = region->fast_out(i);
1.654 + if (n->is_Phi()) {
1.655 + for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1.656 + Node* m = n->fast_out(j);
1.657 + if (m->Opcode() == Op_ConvI2L) {
1.658 + return false;
1.659 + }
1.660 + }
1.661 + }
1.662 + }
1.663 + return true;
1.664 +}
1.665 +#endif
1.666 +
1.667 +
1.668 +//------------------------------place_near_use---------------------------------
1.669 +// Place some computation next to use but not inside inner loops.
1.670 +// For inner loop uses move it to the preheader area.
1.671 +Node *PhaseIdealLoop::place_near_use( Node *useblock ) const {
1.672 + IdealLoopTree *u_loop = get_loop( useblock );
1.673 + return (u_loop->_irreducible || u_loop->_child)
1.674 + ? useblock
1.675 + : u_loop->_head->in(LoopNode::EntryControl);
1.676 +}
1.677 +
1.678 +
1.679 +//------------------------------split_if_with_blocks_post----------------------
1.680 +// Do the real work in a non-recursive function. CFG hackery wants to be
1.681 +// in the post-order, so it can dirty the I-DOM info and not use the dirtied
1.682 +// info.
1.683 +void PhaseIdealLoop::split_if_with_blocks_post( Node *n ) {
1.684 +
1.685 + // Cloning Cmp through Phi's involves the split-if transform.
1.686 + // FastLock is not used by an If
1.687 + if( n->is_Cmp() && !n->is_FastLock() ) {
1.688 + if( C->unique() > 35000 ) return; // Method too big
1.689 +
1.690 + // Do not do 'split-if' if irreducible loops are present.
1.691 + if( _has_irreducible_loops )
1.692 + return;
1.693 +
1.694 + Node *n_ctrl = get_ctrl(n);
1.695 + // Determine if the Node has inputs from some local Phi.
1.696 + // Returns the block to clone thru.
1.697 + Node *n_blk = has_local_phi_input( n );
1.698 + if( n_blk != n_ctrl ) return;
1.699 +
1.700 + if( merge_point_too_heavy(C, n_ctrl) )
1.701 + return;
1.702 +
1.703 + if( n->outcnt() != 1 ) return; // Multiple bool's from 1 compare?
1.704 + Node *bol = n->unique_out();
1.705 + assert( bol->is_Bool(), "expect a bool here" );
1.706 + if( bol->outcnt() != 1 ) return;// Multiple branches from 1 compare?
1.707 + Node *iff = bol->unique_out();
1.708 +
1.709 + // Check some safety conditions
1.710 + if( iff->is_If() ) { // Classic split-if?
1.711 + if( iff->in(0) != n_ctrl ) return; // Compare must be in same blk as if
1.712 + } else if (iff->is_CMove()) { // Trying to split-up a CMOVE
1.713 + if( get_ctrl(iff->in(2)) == n_ctrl ||
1.714 + get_ctrl(iff->in(3)) == n_ctrl )
1.715 + return; // Inputs not yet split-up
1.716 + if ( get_loop(n_ctrl) != get_loop(get_ctrl(iff)) ) {
1.717 + return; // Loop-invar test gates loop-varying CMOVE
1.718 + }
1.719 + } else {
1.720 + return; // some other kind of node, such as an Allocate
1.721 + }
1.722 +
1.723 + // Do not do 'split-if' if some paths are dead. First do dead code
1.724 + // elimination and then see if its still profitable.
1.725 + for( uint i = 1; i < n_ctrl->req(); i++ )
1.726 + if( n_ctrl->in(i) == C->top() )
1.727 + return;
1.728 +
1.729 + // When is split-if profitable? Every 'win' on means some control flow
1.730 + // goes dead, so it's almost always a win.
1.731 + int policy = 0;
1.732 + // If trying to do a 'Split-If' at the loop head, it is only
1.733 + // profitable if the cmp folds up on BOTH paths. Otherwise we
1.734 + // risk peeling a loop forever.
1.735 +
1.736 + // CNC - Disabled for now. Requires careful handling of loop
1.737 + // body selection for the cloned code. Also, make sure we check
1.738 + // for any input path not being in the same loop as n_ctrl. For
1.739 + // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
1.740 + // because the alternative loop entry points won't be converted
1.741 + // into LoopNodes.
1.742 + IdealLoopTree *n_loop = get_loop(n_ctrl);
1.743 + for( uint j = 1; j < n_ctrl->req(); j++ )
1.744 + if( get_loop(n_ctrl->in(j)) != n_loop )
1.745 + return;
1.746 +
1.747 +#ifdef _LP64
1.748 + // Check for safety of the merge point.
1.749 + if( !merge_point_safe(n_ctrl) ) {
1.750 + return;
1.751 + }
1.752 +#endif
1.753 +
1.754 + // Split compare 'n' through the merge point if it is profitable
1.755 + Node *phi = split_thru_phi( n, n_ctrl, policy );
1.756 + if( !phi ) return;
1.757 +
1.758 + // Found a Phi to split thru!
1.759 + // Replace 'n' with the new phi
1.760 + _igvn.hash_delete(n);
1.761 + _igvn.subsume_node( n, phi );
1.762 +
1.763 + // Now split the bool up thru the phi
1.764 + Node *bolphi = split_thru_phi( bol, n_ctrl, -1 );
1.765 + _igvn.hash_delete(bol);
1.766 + _igvn.subsume_node( bol, bolphi );
1.767 + assert( iff->in(1) == bolphi, "" );
1.768 + if( bolphi->Value(&_igvn)->singleton() )
1.769 + return;
1.770 +
1.771 + // Conditional-move? Must split up now
1.772 + if( !iff->is_If() ) {
1.773 + Node *cmovphi = split_thru_phi( iff, n_ctrl, -1 );
1.774 + _igvn.hash_delete(iff);
1.775 + _igvn.subsume_node( iff, cmovphi );
1.776 + return;
1.777 + }
1.778 +
1.779 + // Now split the IF
1.780 + do_split_if( iff );
1.781 + return;
1.782 + }
1.783 +
1.784 + // Check for an IF ready to split; one that has its
1.785 + // condition codes input coming from a Phi at the block start.
1.786 + int n_op = n->Opcode();
1.787 +
1.788 + // Check for an IF being dominated by another IF same test
1.789 + if( n_op == Op_If ) {
1.790 + Node *bol = n->in(1);
1.791 + uint max = bol->outcnt();
1.792 + // Check for same test used more than once?
1.793 + if( n_op == Op_If && max > 1 && bol->is_Bool() ) {
1.794 + // Search up IDOMs to see if this IF is dominated.
1.795 + Node *cutoff = get_ctrl(bol);
1.796 +
1.797 + // Now search up IDOMs till cutoff, looking for a dominating test
1.798 + Node *prevdom = n;
1.799 + Node *dom = idom(prevdom);
1.800 + while( dom != cutoff ) {
1.801 + if( dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom ) {
1.802 + // Replace the dominated test with an obvious true or false.
1.803 + // Place it on the IGVN worklist for later cleanup.
1.804 + C->set_major_progress();
1.805 + dominated_by( prevdom, n );
1.806 +#ifndef PRODUCT
1.807 + if( VerifyLoopOptimizations ) verify();
1.808 +#endif
1.809 + return;
1.810 + }
1.811 + prevdom = dom;
1.812 + dom = idom(prevdom);
1.813 + }
1.814 + }
1.815 + }
1.816 +
1.817 + // See if a shared loop-varying computation has no loop-varying uses.
1.818 + // Happens if something is only used for JVM state in uncommon trap exits,
1.819 + // like various versions of induction variable+offset. Clone the
1.820 + // computation per usage to allow it to sink out of the loop.
1.821 + if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about)
1.822 + Node *n_ctrl = get_ctrl(n);
1.823 + IdealLoopTree *n_loop = get_loop(n_ctrl);
1.824 + if( n_loop != _ltree_root ) {
1.825 + DUIterator_Fast imax, i = n->fast_outs(imax);
1.826 + for (; i < imax; i++) {
1.827 + Node* u = n->fast_out(i);
1.828 + if( !has_ctrl(u) ) break; // Found control user
1.829 + IdealLoopTree *u_loop = get_loop(get_ctrl(u));
1.830 + if( u_loop == n_loop ) break; // Found loop-varying use
1.831 + if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop
1.832 + if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003
1.833 + }
1.834 + bool did_break = (i < imax); // Did we break out of the previous loop?
1.835 + if (!did_break && n->outcnt() > 1) { // All uses in outer loops!
1.836 + Node *late_load_ctrl;
1.837 + if (n->is_Load()) {
1.838 + // If n is a load, get and save the result from get_late_ctrl(),
1.839 + // to be later used in calculating the control for n's clones.
1.840 + clear_dom_lca_tags();
1.841 + late_load_ctrl = get_late_ctrl(n, n_ctrl);
1.842 + }
1.843 + // If n is a load, and the late control is the same as the current
1.844 + // control, then the cloning of n is a pointless exercise, because
1.845 + // GVN will ensure that we end up where we started.
1.846 + if (!n->is_Load() || late_load_ctrl != n_ctrl) {
1.847 + for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) {
1.848 + Node *u = n->last_out(j); // Clone private computation per use
1.849 + _igvn.hash_delete(u);
1.850 + _igvn._worklist.push(u);
1.851 + Node *x = n->clone(); // Clone computation
1.852 + Node *x_ctrl = NULL;
1.853 + if( u->is_Phi() ) {
1.854 + // Replace all uses of normal nodes. Replace Phi uses
1.855 + // individually, so the seperate Nodes can sink down
1.856 + // different paths.
1.857 + uint k = 1;
1.858 + while( u->in(k) != n ) k++;
1.859 + u->set_req( k, x );
1.860 + // x goes next to Phi input path
1.861 + x_ctrl = u->in(0)->in(k);
1.862 + --j;
1.863 + } else { // Normal use
1.864 + // Replace all uses
1.865 + for( uint k = 0; k < u->req(); k++ ) {
1.866 + if( u->in(k) == n ) {
1.867 + u->set_req( k, x );
1.868 + --j;
1.869 + }
1.870 + }
1.871 + x_ctrl = get_ctrl(u);
1.872 + }
1.873 +
1.874 + // Find control for 'x' next to use but not inside inner loops.
1.875 + // For inner loop uses get the preheader area.
1.876 + x_ctrl = place_near_use(x_ctrl);
1.877 +
1.878 + if (n->is_Load()) {
1.879 + // For loads, add a control edge to a CFG node outside of the loop
1.880 + // to force them to not combine and return back inside the loop
1.881 + // during GVN optimization (4641526).
1.882 + //
1.883 + // Because we are setting the actual control input, factor in
1.884 + // the result from get_late_ctrl() so we respect any
1.885 + // anti-dependences. (6233005).
1.886 + x_ctrl = dom_lca(late_load_ctrl, x_ctrl);
1.887 +
1.888 + // Don't allow the control input to be a CFG splitting node.
1.889 + // Such nodes should only have ProjNodes as outs, e.g. IfNode
1.890 + // should only have IfTrueNode and IfFalseNode (4985384).
1.891 + x_ctrl = find_non_split_ctrl(x_ctrl);
1.892 + assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
1.893 +
1.894 + x->set_req(0, x_ctrl);
1.895 + }
1.896 + register_new_node(x, x_ctrl);
1.897 +
1.898 + // Some institutional knowledge is needed here: 'x' is
1.899 + // yanked because if the optimizer runs GVN on it all the
1.900 + // cloned x's will common up and undo this optimization and
1.901 + // be forced back in the loop. This is annoying because it
1.902 + // makes +VerifyOpto report false-positives on progress. I
1.903 + // tried setting control edges on the x's to force them to
1.904 + // not combine, but the matching gets worried when it tries
1.905 + // to fold a StoreP and an AddP together (as part of an
1.906 + // address expression) and the AddP and StoreP have
1.907 + // different controls.
1.908 + if( !x->is_Load() ) _igvn._worklist.yank(x);
1.909 + }
1.910 + _igvn.remove_dead_node(n);
1.911 + }
1.912 + }
1.913 + }
1.914 + }
1.915 +
1.916 + // Check for Opaque2's who's loop has disappeared - who's input is in the
1.917 + // same loop nest as their output. Remove 'em, they are no longer useful.
1.918 + if( n_op == Op_Opaque2 &&
1.919 + n->in(1) != NULL &&
1.920 + get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) {
1.921 + _igvn.add_users_to_worklist(n);
1.922 + _igvn.hash_delete(n);
1.923 + _igvn.subsume_node( n, n->in(1) );
1.924 + }
1.925 +}
1.926 +
1.927 +//------------------------------split_if_with_blocks---------------------------
1.928 +// Check for aggressive application of 'split-if' optimization,
1.929 +// using basic block level info.
1.930 +void PhaseIdealLoop::split_if_with_blocks( VectorSet &visited, Node_Stack &nstack ) {
1.931 + Node *n = C->root();
1.932 + visited.set(n->_idx); // first, mark node as visited
1.933 + // Do pre-visit work for root
1.934 + n = split_if_with_blocks_pre( n );
1.935 + uint cnt = n->outcnt();
1.936 + uint i = 0;
1.937 + while (true) {
1.938 + // Visit all children
1.939 + if (i < cnt) {
1.940 + Node* use = n->raw_out(i);
1.941 + ++i;
1.942 + if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
1.943 + // Now do pre-visit work for this use
1.944 + use = split_if_with_blocks_pre( use );
1.945 + nstack.push(n, i); // Save parent and next use's index.
1.946 + n = use; // Process all children of current use.
1.947 + cnt = use->outcnt();
1.948 + i = 0;
1.949 + }
1.950 + }
1.951 + else {
1.952 + // All of n's children have been processed, complete post-processing.
1.953 + if (cnt != 0 && !n->is_Con()) {
1.954 + assert(has_node(n), "no dead nodes");
1.955 + split_if_with_blocks_post( n );
1.956 + }
1.957 + if (nstack.is_empty()) {
1.958 + // Finished all nodes on stack.
1.959 + break;
1.960 + }
1.961 + // Get saved parent node and next use's index. Visit the rest of uses.
1.962 + n = nstack.node();
1.963 + cnt = n->outcnt();
1.964 + i = nstack.index();
1.965 + nstack.pop();
1.966 + }
1.967 + }
1.968 +}
1.969 +
1.970 +
1.971 +//=============================================================================
1.972 +//
1.973 +// C L O N E A L O O P B O D Y
1.974 +//
1.975 +
1.976 +//------------------------------clone_iff--------------------------------------
1.977 +// Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1.978 +// "Nearly" because all Nodes have been cloned from the original in the loop,
1.979 +// but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1.980 +// through the Phi recursively, and return a Bool.
1.981 +BoolNode *PhaseIdealLoop::clone_iff( PhiNode *phi, IdealLoopTree *loop ) {
1.982 +
1.983 + // Convert this Phi into a Phi merging Bools
1.984 + uint i;
1.985 + for( i = 1; i < phi->req(); i++ ) {
1.986 + Node *b = phi->in(i);
1.987 + if( b->is_Phi() ) {
1.988 + _igvn.hash_delete(phi);
1.989 + _igvn._worklist.push(phi);
1.990 + phi->set_req(i, clone_iff( b->as_Phi(), loop ));
1.991 + } else {
1.992 + assert( b->is_Bool(), "" );
1.993 + }
1.994 + }
1.995 +
1.996 + Node *sample_bool = phi->in(1);
1.997 + Node *sample_cmp = sample_bool->in(1);
1.998 +
1.999 + // Make Phis to merge the Cmp's inputs.
1.1000 + int size = phi->in(0)->req();
1.1001 + PhiNode *phi1 = new (C, size) PhiNode( phi->in(0), Type::TOP );
1.1002 + PhiNode *phi2 = new (C, size) PhiNode( phi->in(0), Type::TOP );
1.1003 + for( i = 1; i < phi->req(); i++ ) {
1.1004 + Node *n1 = phi->in(i)->in(1)->in(1);
1.1005 + Node *n2 = phi->in(i)->in(1)->in(2);
1.1006 + phi1->set_req( i, n1 );
1.1007 + phi2->set_req( i, n2 );
1.1008 + phi1->set_type( phi1->type()->meet(n1->bottom_type()) );
1.1009 + phi2->set_type( phi2->type()->meet(n2->bottom_type()) );
1.1010 + }
1.1011 + // See if these Phis have been made before.
1.1012 + // Register with optimizer
1.1013 + Node *hit1 = _igvn.hash_find_insert(phi1);
1.1014 + if( hit1 ) { // Hit, toss just made Phi
1.1015 + _igvn.remove_dead_node(phi1); // Remove new phi
1.1016 + assert( hit1->is_Phi(), "" );
1.1017 + phi1 = (PhiNode*)hit1; // Use existing phi
1.1018 + } else { // Miss
1.1019 + _igvn.register_new_node_with_optimizer(phi1);
1.1020 + }
1.1021 + Node *hit2 = _igvn.hash_find_insert(phi2);
1.1022 + if( hit2 ) { // Hit, toss just made Phi
1.1023 + _igvn.remove_dead_node(phi2); // Remove new phi
1.1024 + assert( hit2->is_Phi(), "" );
1.1025 + phi2 = (PhiNode*)hit2; // Use existing phi
1.1026 + } else { // Miss
1.1027 + _igvn.register_new_node_with_optimizer(phi2);
1.1028 + }
1.1029 + // Register Phis with loop/block info
1.1030 + set_ctrl(phi1, phi->in(0));
1.1031 + set_ctrl(phi2, phi->in(0));
1.1032 + // Make a new Cmp
1.1033 + Node *cmp = sample_cmp->clone();
1.1034 + cmp->set_req( 1, phi1 );
1.1035 + cmp->set_req( 2, phi2 );
1.1036 + _igvn.register_new_node_with_optimizer(cmp);
1.1037 + set_ctrl(cmp, phi->in(0));
1.1038 +
1.1039 + // Make a new Bool
1.1040 + Node *b = sample_bool->clone();
1.1041 + b->set_req(1,cmp);
1.1042 + _igvn.register_new_node_with_optimizer(b);
1.1043 + set_ctrl(b, phi->in(0));
1.1044 +
1.1045 + assert( b->is_Bool(), "" );
1.1046 + return (BoolNode*)b;
1.1047 +}
1.1048 +
1.1049 +//------------------------------clone_bool-------------------------------------
1.1050 +// Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1.1051 +// "Nearly" because all Nodes have been cloned from the original in the loop,
1.1052 +// but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1.1053 +// through the Phi recursively, and return a Bool.
1.1054 +CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) {
1.1055 + uint i;
1.1056 + // Convert this Phi into a Phi merging Bools
1.1057 + for( i = 1; i < phi->req(); i++ ) {
1.1058 + Node *b = phi->in(i);
1.1059 + if( b->is_Phi() ) {
1.1060 + _igvn.hash_delete(phi);
1.1061 + _igvn._worklist.push(phi);
1.1062 + phi->set_req(i, clone_bool( b->as_Phi(), loop ));
1.1063 + } else {
1.1064 + assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
1.1065 + }
1.1066 + }
1.1067 +
1.1068 + Node *sample_cmp = phi->in(1);
1.1069 +
1.1070 + // Make Phis to merge the Cmp's inputs.
1.1071 + int size = phi->in(0)->req();
1.1072 + PhiNode *phi1 = new (C, size) PhiNode( phi->in(0), Type::TOP );
1.1073 + PhiNode *phi2 = new (C, size) PhiNode( phi->in(0), Type::TOP );
1.1074 + for( uint j = 1; j < phi->req(); j++ ) {
1.1075 + Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
1.1076 + Node *n1, *n2;
1.1077 + if( cmp_top->is_Cmp() ) {
1.1078 + n1 = cmp_top->in(1);
1.1079 + n2 = cmp_top->in(2);
1.1080 + } else {
1.1081 + n1 = n2 = cmp_top;
1.1082 + }
1.1083 + phi1->set_req( j, n1 );
1.1084 + phi2->set_req( j, n2 );
1.1085 + phi1->set_type( phi1->type()->meet(n1->bottom_type()) );
1.1086 + phi2->set_type( phi2->type()->meet(n2->bottom_type()) );
1.1087 + }
1.1088 +
1.1089 + // See if these Phis have been made before.
1.1090 + // Register with optimizer
1.1091 + Node *hit1 = _igvn.hash_find_insert(phi1);
1.1092 + if( hit1 ) { // Hit, toss just made Phi
1.1093 + _igvn.remove_dead_node(phi1); // Remove new phi
1.1094 + assert( hit1->is_Phi(), "" );
1.1095 + phi1 = (PhiNode*)hit1; // Use existing phi
1.1096 + } else { // Miss
1.1097 + _igvn.register_new_node_with_optimizer(phi1);
1.1098 + }
1.1099 + Node *hit2 = _igvn.hash_find_insert(phi2);
1.1100 + if( hit2 ) { // Hit, toss just made Phi
1.1101 + _igvn.remove_dead_node(phi2); // Remove new phi
1.1102 + assert( hit2->is_Phi(), "" );
1.1103 + phi2 = (PhiNode*)hit2; // Use existing phi
1.1104 + } else { // Miss
1.1105 + _igvn.register_new_node_with_optimizer(phi2);
1.1106 + }
1.1107 + // Register Phis with loop/block info
1.1108 + set_ctrl(phi1, phi->in(0));
1.1109 + set_ctrl(phi2, phi->in(0));
1.1110 + // Make a new Cmp
1.1111 + Node *cmp = sample_cmp->clone();
1.1112 + cmp->set_req( 1, phi1 );
1.1113 + cmp->set_req( 2, phi2 );
1.1114 + _igvn.register_new_node_with_optimizer(cmp);
1.1115 + set_ctrl(cmp, phi->in(0));
1.1116 +
1.1117 + assert( cmp->is_Cmp(), "" );
1.1118 + return (CmpNode*)cmp;
1.1119 +}
1.1120 +
1.1121 +//------------------------------sink_use---------------------------------------
1.1122 +// If 'use' was in the loop-exit block, it now needs to be sunk
1.1123 +// below the post-loop merge point.
1.1124 +void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) {
1.1125 + if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) {
1.1126 + set_ctrl(use, post_loop);
1.1127 + for (DUIterator j = use->outs(); use->has_out(j); j++)
1.1128 + sink_use(use->out(j), post_loop);
1.1129 + }
1.1130 +}
1.1131 +
1.1132 +//------------------------------clone_loop-------------------------------------
1.1133 +//
1.1134 +// C L O N E A L O O P B O D Y
1.1135 +//
1.1136 +// This is the basic building block of the loop optimizations. It clones an
1.1137 +// entire loop body. It makes an old_new loop body mapping; with this mapping
1.1138 +// you can find the new-loop equivalent to an old-loop node. All new-loop
1.1139 +// nodes are exactly equal to their old-loop counterparts, all edges are the
1.1140 +// same. All exits from the old-loop now have a RegionNode that merges the
1.1141 +// equivalent new-loop path. This is true even for the normal "loop-exit"
1.1142 +// condition. All uses of loop-invariant old-loop values now come from (one
1.1143 +// or more) Phis that merge their new-loop equivalents.
1.1144 +//
1.1145 +// This operation leaves the graph in an illegal state: there are two valid
1.1146 +// control edges coming from the loop pre-header to both loop bodies. I'll
1.1147 +// definitely have to hack the graph after running this transform.
1.1148 +//
1.1149 +// From this building block I will further edit edges to perform loop peeling
1.1150 +// or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
1.1151 +//
1.1152 +// Parameter side_by_size_idom:
1.1153 +// When side_by_size_idom is NULL, the dominator tree is constructed for
1.1154 +// the clone loop to dominate the original. Used in construction of
1.1155 +// pre-main-post loop sequence.
1.1156 +// When nonnull, the clone and original are side-by-side, both are
1.1157 +// dominated by the side_by_side_idom node. Used in construction of
1.1158 +// unswitched loops.
1.1159 +void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
1.1160 + Node* side_by_side_idom) {
1.1161 +
1.1162 + // Step 1: Clone the loop body. Make the old->new mapping.
1.1163 + uint i;
1.1164 + for( i = 0; i < loop->_body.size(); i++ ) {
1.1165 + Node *old = loop->_body.at(i);
1.1166 + Node *nnn = old->clone();
1.1167 + old_new.map( old->_idx, nnn );
1.1168 + _igvn.register_new_node_with_optimizer(nnn);
1.1169 + }
1.1170 +
1.1171 +
1.1172 + // Step 2: Fix the edges in the new body. If the old input is outside the
1.1173 + // loop use it. If the old input is INside the loop, use the corresponding
1.1174 + // new node instead.
1.1175 + for( i = 0; i < loop->_body.size(); i++ ) {
1.1176 + Node *old = loop->_body.at(i);
1.1177 + Node *nnn = old_new[old->_idx];
1.1178 + // Fix CFG/Loop controlling the new node
1.1179 + if (has_ctrl(old)) {
1.1180 + set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
1.1181 + } else {
1.1182 + set_loop(nnn, loop->_parent);
1.1183 + if (old->outcnt() > 0) {
1.1184 + set_idom( nnn, old_new[idom(old)->_idx], dd );
1.1185 + }
1.1186 + }
1.1187 + // Correct edges to the new node
1.1188 + for( uint j = 0; j < nnn->req(); j++ ) {
1.1189 + Node *n = nnn->in(j);
1.1190 + if( n ) {
1.1191 + IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n );
1.1192 + if( loop->is_member( old_in_loop ) )
1.1193 + nnn->set_req(j, old_new[n->_idx]);
1.1194 + }
1.1195 + }
1.1196 + _igvn.hash_find_insert(nnn);
1.1197 + }
1.1198 + Node *newhead = old_new[loop->_head->_idx];
1.1199 + set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
1.1200 +
1.1201 +
1.1202 + // Step 3: Now fix control uses. Loop varying control uses have already
1.1203 + // been fixed up (as part of all input edges in Step 2). Loop invariant
1.1204 + // control uses must be either an IfFalse or an IfTrue. Make a merge
1.1205 + // point to merge the old and new IfFalse/IfTrue nodes; make the use
1.1206 + // refer to this.
1.1207 + ResourceArea *area = Thread::current()->resource_area();
1.1208 + Node_List worklist(area);
1.1209 + uint new_counter = C->unique();
1.1210 + for( i = 0; i < loop->_body.size(); i++ ) {
1.1211 + Node* old = loop->_body.at(i);
1.1212 + if( !old->is_CFG() ) continue;
1.1213 + Node* nnn = old_new[old->_idx];
1.1214 +
1.1215 + // Copy uses to a worklist, so I can munge the def-use info
1.1216 + // with impunity.
1.1217 + for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
1.1218 + worklist.push(old->fast_out(j));
1.1219 +
1.1220 + while( worklist.size() ) { // Visit all uses
1.1221 + Node *use = worklist.pop();
1.1222 + if (!has_node(use)) continue; // Ignore dead nodes
1.1223 + IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
1.1224 + if( !loop->is_member( use_loop ) && use->is_CFG() ) {
1.1225 + // Both OLD and USE are CFG nodes here.
1.1226 + assert( use->is_Proj(), "" );
1.1227 +
1.1228 + // Clone the loop exit control projection
1.1229 + Node *newuse = use->clone();
1.1230 + newuse->set_req(0,nnn);
1.1231 + _igvn.register_new_node_with_optimizer(newuse);
1.1232 + set_loop(newuse, use_loop);
1.1233 + set_idom(newuse, nnn, dom_depth(nnn) + 1 );
1.1234 +
1.1235 + // We need a Region to merge the exit from the peeled body and the
1.1236 + // exit from the old loop body.
1.1237 + RegionNode *r = new (C, 3) RegionNode(3);
1.1238 + // Map the old use to the new merge point
1.1239 + old_new.map( use->_idx, r );
1.1240 + uint dd_r = MIN2(dom_depth(newuse),dom_depth(use));
1.1241 + assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" );
1.1242 +
1.1243 + // The original user of 'use' uses 'r' instead.
1.1244 + for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
1.1245 + Node* useuse = use->last_out(l);
1.1246 + _igvn.hash_delete(useuse);
1.1247 + _igvn._worklist.push(useuse);
1.1248 + uint uses_found = 0;
1.1249 + if( useuse->in(0) == use ) {
1.1250 + useuse->set_req(0, r);
1.1251 + uses_found++;
1.1252 + if( useuse->is_CFG() ) {
1.1253 + assert( dom_depth(useuse) > dd_r, "" );
1.1254 + set_idom(useuse, r, dom_depth(useuse));
1.1255 + }
1.1256 + }
1.1257 + for( uint k = 1; k < useuse->req(); k++ ) {
1.1258 + if( useuse->in(k) == use ) {
1.1259 + useuse->set_req(k, r);
1.1260 + uses_found++;
1.1261 + }
1.1262 + }
1.1263 + l -= uses_found; // we deleted 1 or more copies of this edge
1.1264 + }
1.1265 +
1.1266 + // Now finish up 'r'
1.1267 + r->set_req( 1, newuse );
1.1268 + r->set_req( 2, use );
1.1269 + _igvn.register_new_node_with_optimizer(r);
1.1270 + set_loop(r, use_loop);
1.1271 + set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r);
1.1272 + } // End of if a loop-exit test
1.1273 + }
1.1274 + }
1.1275 +
1.1276 + // Step 4: If loop-invariant use is not control, it must be dominated by a
1.1277 + // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region
1.1278 + // there if needed. Make a Phi there merging old and new used values.
1.1279 + Node_List *split_if_set = NULL;
1.1280 + Node_List *split_bool_set = NULL;
1.1281 + Node_List *split_cex_set = NULL;
1.1282 + for( i = 0; i < loop->_body.size(); i++ ) {
1.1283 + Node* old = loop->_body.at(i);
1.1284 + Node* nnn = old_new[old->_idx];
1.1285 + // Copy uses to a worklist, so I can munge the def-use info
1.1286 + // with impunity.
1.1287 + for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
1.1288 + worklist.push(old->fast_out(j));
1.1289 +
1.1290 + while( worklist.size() ) {
1.1291 + Node *use = worklist.pop();
1.1292 + if (!has_node(use)) continue; // Ignore dead nodes
1.1293 + if (use->in(0) == C->top()) continue;
1.1294 + IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
1.1295 + // Check for data-use outside of loop - at least one of OLD or USE
1.1296 + // must not be a CFG node.
1.1297 + if( !loop->is_member( use_loop ) && (!old->is_CFG() || !use->is_CFG())) {
1.1298 +
1.1299 + // If the Data use is an IF, that means we have an IF outside of the
1.1300 + // loop that is switching on a condition that is set inside of the
1.1301 + // loop. Happens if people set a loop-exit flag; then test the flag
1.1302 + // in the loop to break the loop, then test is again outside of the
1.1303 + // loop to determine which way the loop exited.
1.1304 + if( use->is_If() || use->is_CMove() ) {
1.1305 + // Since this code is highly unlikely, we lazily build the worklist
1.1306 + // of such Nodes to go split.
1.1307 + if( !split_if_set )
1.1308 + split_if_set = new Node_List(area);
1.1309 + split_if_set->push(use);
1.1310 + }
1.1311 + if( use->is_Bool() ) {
1.1312 + if( !split_bool_set )
1.1313 + split_bool_set = new Node_List(area);
1.1314 + split_bool_set->push(use);
1.1315 + }
1.1316 + if( use->Opcode() == Op_CreateEx ) {
1.1317 + if( !split_cex_set )
1.1318 + split_cex_set = new Node_List(area);
1.1319 + split_cex_set->push(use);
1.1320 + }
1.1321 +
1.1322 +
1.1323 + // Get "block" use is in
1.1324 + uint idx = 0;
1.1325 + while( use->in(idx) != old ) idx++;
1.1326 + Node *prev = use->is_CFG() ? use : get_ctrl(use);
1.1327 + assert( !loop->is_member( get_loop( prev ) ), "" );
1.1328 + Node *cfg = prev->_idx >= new_counter
1.1329 + ? prev->in(2)
1.1330 + : idom(prev);
1.1331 + if( use->is_Phi() ) // Phi use is in prior block
1.1332 + cfg = prev->in(idx); // NOT in block of Phi itself
1.1333 + if (cfg->is_top()) { // Use is dead?
1.1334 + _igvn.hash_delete(use);
1.1335 + _igvn._worklist.push(use);
1.1336 + use->set_req(idx, C->top());
1.1337 + continue;
1.1338 + }
1.1339 +
1.1340 + while( !loop->is_member( get_loop( cfg ) ) ) {
1.1341 + prev = cfg;
1.1342 + cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg);
1.1343 + }
1.1344 + // If the use occurs after merging several exits from the loop, then
1.1345 + // old value must have dominated all those exits. Since the same old
1.1346 + // value was used on all those exits we did not need a Phi at this
1.1347 + // merge point. NOW we do need a Phi here. Each loop exit value
1.1348 + // is now merged with the peeled body exit; each exit gets its own
1.1349 + // private Phi and those Phis need to be merged here.
1.1350 + Node *phi;
1.1351 + if( prev->is_Region() ) {
1.1352 + if( idx == 0 ) { // Updating control edge?
1.1353 + phi = prev; // Just use existing control
1.1354 + } else { // Else need a new Phi
1.1355 + phi = PhiNode::make( prev, old );
1.1356 + // Now recursively fix up the new uses of old!
1.1357 + for( uint i = 1; i < prev->req(); i++ ) {
1.1358 + worklist.push(phi); // Onto worklist once for each 'old' input
1.1359 + }
1.1360 + }
1.1361 + } else {
1.1362 + // Get new RegionNode merging old and new loop exits
1.1363 + prev = old_new[prev->_idx];
1.1364 + assert( prev, "just made this in step 7" );
1.1365 + if( idx == 0 ) { // Updating control edge?
1.1366 + phi = prev; // Just use existing control
1.1367 + } else { // Else need a new Phi
1.1368 + // Make a new Phi merging data values properly
1.1369 + phi = PhiNode::make( prev, old );
1.1370 + phi->set_req( 1, nnn );
1.1371 + }
1.1372 + }
1.1373 + // If inserting a new Phi, check for prior hits
1.1374 + if( idx != 0 ) {
1.1375 + Node *hit = _igvn.hash_find_insert(phi);
1.1376 + if( hit == NULL ) {
1.1377 + _igvn.register_new_node_with_optimizer(phi); // Register new phi
1.1378 + } else { // or
1.1379 + // Remove the new phi from the graph and use the hit
1.1380 + _igvn.remove_dead_node(phi);
1.1381 + phi = hit; // Use existing phi
1.1382 + }
1.1383 + set_ctrl(phi, prev);
1.1384 + }
1.1385 + // Make 'use' use the Phi instead of the old loop body exit value
1.1386 + _igvn.hash_delete(use);
1.1387 + _igvn._worklist.push(use);
1.1388 + use->set_req(idx, phi);
1.1389 + if( use->_idx >= new_counter ) { // If updating new phis
1.1390 + // Not needed for correctness, but prevents a weak assert
1.1391 + // in AddPNode from tripping (when we end up with different
1.1392 + // base & derived Phis that will become the same after
1.1393 + // IGVN does CSE).
1.1394 + Node *hit = _igvn.hash_find_insert(use);
1.1395 + if( hit ) // Go ahead and re-hash for hits.
1.1396 + _igvn.subsume_node( use, hit );
1.1397 + }
1.1398 +
1.1399 + // If 'use' was in the loop-exit block, it now needs to be sunk
1.1400 + // below the post-loop merge point.
1.1401 + sink_use( use, prev );
1.1402 + }
1.1403 + }
1.1404 + }
1.1405 +
1.1406 + // Check for IFs that need splitting/cloning. Happens if an IF outside of
1.1407 + // the loop uses a condition set in the loop. The original IF probably
1.1408 + // takes control from one or more OLD Regions (which in turn get from NEW
1.1409 + // Regions). In any case, there will be a set of Phis for each merge point
1.1410 + // from the IF up to where the original BOOL def exists the loop.
1.1411 + if( split_if_set ) {
1.1412 + while( split_if_set->size() ) {
1.1413 + Node *iff = split_if_set->pop();
1.1414 + if( iff->in(1)->is_Phi() ) {
1.1415 + BoolNode *b = clone_iff( iff->in(1)->as_Phi(), loop );
1.1416 + _igvn.hash_delete(iff);
1.1417 + _igvn._worklist.push(iff);
1.1418 + iff->set_req(1, b);
1.1419 + }
1.1420 + }
1.1421 + }
1.1422 + if( split_bool_set ) {
1.1423 + while( split_bool_set->size() ) {
1.1424 + Node *b = split_bool_set->pop();
1.1425 + Node *phi = b->in(1);
1.1426 + assert( phi->is_Phi(), "" );
1.1427 + CmpNode *cmp = clone_bool( (PhiNode*)phi, loop );
1.1428 + _igvn.hash_delete(b);
1.1429 + _igvn._worklist.push(b);
1.1430 + b->set_req(1, cmp);
1.1431 + }
1.1432 + }
1.1433 + if( split_cex_set ) {
1.1434 + while( split_cex_set->size() ) {
1.1435 + Node *b = split_cex_set->pop();
1.1436 + assert( b->in(0)->is_Region(), "" );
1.1437 + assert( b->in(1)->is_Phi(), "" );
1.1438 + assert( b->in(0)->in(0) == b->in(1)->in(0), "" );
1.1439 + split_up( b, b->in(0), NULL );
1.1440 + }
1.1441 + }
1.1442 +
1.1443 +}
1.1444 +
1.1445 +
1.1446 +//---------------------- stride_of_possible_iv -------------------------------------
1.1447 +// Looks for an iff/bool/comp with one operand of the compare
1.1448 +// being a cycle involving an add and a phi,
1.1449 +// with an optional truncation (left-shift followed by a right-shift)
1.1450 +// of the add. Returns zero if not an iv.
1.1451 +int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
1.1452 + Node* trunc1 = NULL;
1.1453 + Node* trunc2 = NULL;
1.1454 + const TypeInt* ttype = NULL;
1.1455 + if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) {
1.1456 + return 0;
1.1457 + }
1.1458 + BoolNode* bl = iff->in(1)->as_Bool();
1.1459 + Node* cmp = bl->in(1);
1.1460 + if (!cmp || cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU) {
1.1461 + return 0;
1.1462 + }
1.1463 + // Must have an invariant operand
1.1464 + if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) {
1.1465 + return 0;
1.1466 + }
1.1467 + Node* add2 = NULL;
1.1468 + Node* cmp1 = cmp->in(1);
1.1469 + if (cmp1->is_Phi()) {
1.1470 + // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
1.1471 + Node* phi = cmp1;
1.1472 + for (uint i = 1; i < phi->req(); i++) {
1.1473 + Node* in = phi->in(i);
1.1474 + Node* add = CountedLoopNode::match_incr_with_optional_truncation(in,
1.1475 + &trunc1, &trunc2, &ttype);
1.1476 + if (add && add->in(1) == phi) {
1.1477 + add2 = add->in(2);
1.1478 + break;
1.1479 + }
1.1480 + }
1.1481 + } else {
1.1482 + // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
1.1483 + Node* addtrunc = cmp1;
1.1484 + Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc,
1.1485 + &trunc1, &trunc2, &ttype);
1.1486 + if (add && add->in(1)->is_Phi()) {
1.1487 + Node* phi = add->in(1);
1.1488 + for (uint i = 1; i < phi->req(); i++) {
1.1489 + if (phi->in(i) == addtrunc) {
1.1490 + add2 = add->in(2);
1.1491 + break;
1.1492 + }
1.1493 + }
1.1494 + }
1.1495 + }
1.1496 + if (add2 != NULL) {
1.1497 + const TypeInt* add2t = _igvn.type(add2)->is_int();
1.1498 + if (add2t->is_con()) {
1.1499 + return add2t->get_con();
1.1500 + }
1.1501 + }
1.1502 + return 0;
1.1503 +}
1.1504 +
1.1505 +
1.1506 +//---------------------- stay_in_loop -------------------------------------
1.1507 +// Return the (unique) control output node that's in the loop (if it exists.)
1.1508 +Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
1.1509 + Node* unique = NULL;
1.1510 + if (!n) return NULL;
1.1511 + for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1.1512 + Node* use = n->fast_out(i);
1.1513 + if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
1.1514 + if (unique != NULL) {
1.1515 + return NULL;
1.1516 + }
1.1517 + unique = use;
1.1518 + }
1.1519 + }
1.1520 + return unique;
1.1521 +}
1.1522 +
1.1523 +//------------------------------ register_node -------------------------------------
1.1524 +// Utility to register node "n" with PhaseIdealLoop
1.1525 +void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) {
1.1526 + _igvn.register_new_node_with_optimizer(n);
1.1527 + loop->_body.push(n);
1.1528 + if (n->is_CFG()) {
1.1529 + set_loop(n, loop);
1.1530 + set_idom(n, pred, ddepth);
1.1531 + } else {
1.1532 + set_ctrl(n, pred);
1.1533 + }
1.1534 +}
1.1535 +
1.1536 +//------------------------------ proj_clone -------------------------------------
1.1537 +// Utility to create an if-projection
1.1538 +ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
1.1539 + ProjNode* c = p->clone()->as_Proj();
1.1540 + c->set_req(0, iff);
1.1541 + return c;
1.1542 +}
1.1543 +
1.1544 +//------------------------------ short_circuit_if -------------------------------------
1.1545 +// Force the iff control output to be the live_proj
1.1546 +Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
1.1547 + int proj_con = live_proj->_con;
1.1548 + assert(proj_con == 0 || proj_con == 1, "false or true projection");
1.1549 + Node *con = _igvn.intcon(proj_con);
1.1550 + set_ctrl(con, C->root());
1.1551 + if (iff) {
1.1552 + iff->set_req(1, con);
1.1553 + }
1.1554 + return con;
1.1555 +}
1.1556 +
1.1557 +//------------------------------ insert_if_before_proj -------------------------------------
1.1558 +// Insert a new if before an if projection (* - new node)
1.1559 +//
1.1560 +// before
1.1561 +// if(test)
1.1562 +// / \
1.1563 +// v v
1.1564 +// other-proj proj (arg)
1.1565 +//
1.1566 +// after
1.1567 +// if(test)
1.1568 +// / \
1.1569 +// / v
1.1570 +// | * proj-clone
1.1571 +// v |
1.1572 +// other-proj v
1.1573 +// * new_if(relop(cmp[IU](left,right)))
1.1574 +// / \
1.1575 +// v v
1.1576 +// * new-proj proj
1.1577 +// (returned)
1.1578 +//
1.1579 +ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
1.1580 + IfNode* iff = proj->in(0)->as_If();
1.1581 + IdealLoopTree *loop = get_loop(proj);
1.1582 + ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
1.1583 + int ddepth = dom_depth(proj);
1.1584 +
1.1585 + _igvn.hash_delete(iff);
1.1586 + _igvn._worklist.push(iff);
1.1587 + _igvn.hash_delete(proj);
1.1588 + _igvn._worklist.push(proj);
1.1589 +
1.1590 + proj->set_req(0, NULL); // temporary disconnect
1.1591 + ProjNode* proj2 = proj_clone(proj, iff);
1.1592 + register_node(proj2, loop, iff, ddepth);
1.1593 +
1.1594 + Node* cmp = Signed ? (Node*) new (C,3)CmpINode(left, right) : (Node*) new (C,3)CmpUNode(left, right);
1.1595 + register_node(cmp, loop, proj2, ddepth);
1.1596 +
1.1597 + BoolNode* bol = new (C,2)BoolNode(cmp, relop);
1.1598 + register_node(bol, loop, proj2, ddepth);
1.1599 +
1.1600 + IfNode* new_if = new (C,2)IfNode(proj2, bol, iff->_prob, iff->_fcnt);
1.1601 + register_node(new_if, loop, proj2, ddepth);
1.1602 +
1.1603 + proj->set_req(0, new_if); // reattach
1.1604 + set_idom(proj, new_if, ddepth);
1.1605 +
1.1606 + ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
1.1607 + register_node(new_exit, get_loop(other_proj), new_if, ddepth);
1.1608 +
1.1609 + return new_exit;
1.1610 +}
1.1611 +
1.1612 +//------------------------------ insert_region_before_proj -------------------------------------
1.1613 +// Insert a region before an if projection (* - new node)
1.1614 +//
1.1615 +// before
1.1616 +// if(test)
1.1617 +// / |
1.1618 +// v |
1.1619 +// proj v
1.1620 +// other-proj
1.1621 +//
1.1622 +// after
1.1623 +// if(test)
1.1624 +// / |
1.1625 +// v |
1.1626 +// * proj-clone v
1.1627 +// | other-proj
1.1628 +// v
1.1629 +// * new-region
1.1630 +// |
1.1631 +// v
1.1632 +// * dum_if
1.1633 +// / \
1.1634 +// v \
1.1635 +// * dum-proj v
1.1636 +// proj
1.1637 +//
1.1638 +RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
1.1639 + IfNode* iff = proj->in(0)->as_If();
1.1640 + IdealLoopTree *loop = get_loop(proj);
1.1641 + ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
1.1642 + int ddepth = dom_depth(proj);
1.1643 +
1.1644 + _igvn.hash_delete(iff);
1.1645 + _igvn._worklist.push(iff);
1.1646 + _igvn.hash_delete(proj);
1.1647 + _igvn._worklist.push(proj);
1.1648 +
1.1649 + proj->set_req(0, NULL); // temporary disconnect
1.1650 + ProjNode* proj2 = proj_clone(proj, iff);
1.1651 + register_node(proj2, loop, iff, ddepth);
1.1652 +
1.1653 + RegionNode* reg = new (C,2)RegionNode(2);
1.1654 + reg->set_req(1, proj2);
1.1655 + register_node(reg, loop, iff, ddepth);
1.1656 +
1.1657 + IfNode* dum_if = new (C,2)IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt);
1.1658 + register_node(dum_if, loop, reg, ddepth);
1.1659 +
1.1660 + proj->set_req(0, dum_if); // reattach
1.1661 + set_idom(proj, dum_if, ddepth);
1.1662 +
1.1663 + ProjNode* dum_proj = proj_clone(other_proj, dum_if);
1.1664 + register_node(dum_proj, loop, dum_if, ddepth);
1.1665 +
1.1666 + return reg;
1.1667 +}
1.1668 +
1.1669 +//------------------------------ insert_cmpi_loop_exit -------------------------------------
1.1670 +// Clone a signed compare loop exit from an unsigned compare and
1.1671 +// insert it before the unsigned cmp on the stay-in-loop path.
1.1672 +// All new nodes inserted in the dominator tree between the original
1.1673 +// if and it's projections. The original if test is replaced with
1.1674 +// a constant to force the stay-in-loop path.
1.1675 +//
1.1676 +// This is done to make sure that the original if and it's projections
1.1677 +// still dominate the same set of control nodes, that the ctrl() relation
1.1678 +// from data nodes to them is preserved, and that their loop nesting is
1.1679 +// preserved.
1.1680 +//
1.1681 +// before
1.1682 +// if(i <u limit) unsigned compare loop exit
1.1683 +// / |
1.1684 +// v v
1.1685 +// exit-proj stay-in-loop-proj
1.1686 +//
1.1687 +// after
1.1688 +// if(stay-in-loop-const) original if
1.1689 +// / |
1.1690 +// / v
1.1691 +// / if(i < limit) new signed test
1.1692 +// / / |
1.1693 +// / / v
1.1694 +// / / if(i <u limit) new cloned unsigned test
1.1695 +// / / / |
1.1696 +// v v v |
1.1697 +// region |
1.1698 +// | |
1.1699 +// dum-if |
1.1700 +// / | |
1.1701 +// ether | |
1.1702 +// v v
1.1703 +// exit-proj stay-in-loop-proj
1.1704 +//
1.1705 +IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) {
1.1706 + const bool Signed = true;
1.1707 + const bool Unsigned = false;
1.1708 +
1.1709 + BoolNode* bol = if_cmpu->in(1)->as_Bool();
1.1710 + if (bol->_test._test != BoolTest::lt) return NULL;
1.1711 + CmpNode* cmpu = bol->in(1)->as_Cmp();
1.1712 + if (cmpu->Opcode() != Op_CmpU) return NULL;
1.1713 + int stride = stride_of_possible_iv(if_cmpu);
1.1714 + if (stride == 0) return NULL;
1.1715 +
1.1716 + ProjNode* lp_continue = stay_in_loop(if_cmpu, loop)->as_Proj();
1.1717 + ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
1.1718 +
1.1719 + Node* limit = NULL;
1.1720 + if (stride > 0) {
1.1721 + limit = cmpu->in(2);
1.1722 + } else {
1.1723 + limit = _igvn.makecon(TypeInt::ZERO);
1.1724 + set_ctrl(limit, C->root());
1.1725 + }
1.1726 + // Create a new region on the exit path
1.1727 + RegionNode* reg = insert_region_before_proj(lp_exit);
1.1728 +
1.1729 + // Clone the if-cmpu-true-false using a signed compare
1.1730 + BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
1.1731 + ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue);
1.1732 + reg->add_req(cmpi_exit);
1.1733 +
1.1734 + // Clone the if-cmpu-true-false
1.1735 + BoolTest::mask rel_u = bol->_test._test;
1.1736 + ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
1.1737 + reg->add_req(cmpu_exit);
1.1738 +
1.1739 + // Force original if to stay in loop.
1.1740 + short_circuit_if(if_cmpu, lp_continue);
1.1741 +
1.1742 + return cmpi_exit->in(0)->as_If();
1.1743 +}
1.1744 +
1.1745 +//------------------------------ remove_cmpi_loop_exit -------------------------------------
1.1746 +// Remove a previously inserted signed compare loop exit.
1.1747 +void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
1.1748 + Node* lp_proj = stay_in_loop(if_cmp, loop);
1.1749 + assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
1.1750 + stay_in_loop(lp_proj, loop)->is_If() &&
1.1751 + stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
1.1752 + Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
1.1753 + set_ctrl(con, C->root());
1.1754 + if_cmp->set_req(1, con);
1.1755 +}
1.1756 +
1.1757 +//------------------------------ scheduled_nodelist -------------------------------------
1.1758 +// Create a post order schedule of nodes that are in the
1.1759 +// "member" set. The list is returned in "sched".
1.1760 +// The first node in "sched" is the loop head, followed by
1.1761 +// nodes which have no inputs in the "member" set, and then
1.1762 +// followed by the nodes that have an immediate input dependence
1.1763 +// on a node in "sched".
1.1764 +void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
1.1765 +
1.1766 + assert(member.test(loop->_head->_idx), "loop head must be in member set");
1.1767 + Arena *a = Thread::current()->resource_area();
1.1768 + VectorSet visited(a);
1.1769 + Node_Stack nstack(a, loop->_body.size());
1.1770 +
1.1771 + Node* n = loop->_head; // top of stack is cached in "n"
1.1772 + uint idx = 0;
1.1773 + visited.set(n->_idx);
1.1774 +
1.1775 + // Initially push all with no inputs from within member set
1.1776 + for(uint i = 0; i < loop->_body.size(); i++ ) {
1.1777 + Node *elt = loop->_body.at(i);
1.1778 + if (member.test(elt->_idx)) {
1.1779 + bool found = false;
1.1780 + for (uint j = 0; j < elt->req(); j++) {
1.1781 + Node* def = elt->in(j);
1.1782 + if (def && member.test(def->_idx) && def != elt) {
1.1783 + found = true;
1.1784 + break;
1.1785 + }
1.1786 + }
1.1787 + if (!found && elt != loop->_head) {
1.1788 + nstack.push(n, idx);
1.1789 + n = elt;
1.1790 + assert(!visited.test(n->_idx), "not seen yet");
1.1791 + visited.set(n->_idx);
1.1792 + }
1.1793 + }
1.1794 + }
1.1795 +
1.1796 + // traverse out's that are in the member set
1.1797 + while (true) {
1.1798 + if (idx < n->outcnt()) {
1.1799 + Node* use = n->raw_out(idx);
1.1800 + idx++;
1.1801 + if (!visited.test_set(use->_idx)) {
1.1802 + if (member.test(use->_idx)) {
1.1803 + nstack.push(n, idx);
1.1804 + n = use;
1.1805 + idx = 0;
1.1806 + }
1.1807 + }
1.1808 + } else {
1.1809 + // All outputs processed
1.1810 + sched.push(n);
1.1811 + if (nstack.is_empty()) break;
1.1812 + n = nstack.node();
1.1813 + idx = nstack.index();
1.1814 + nstack.pop();
1.1815 + }
1.1816 + }
1.1817 +}
1.1818 +
1.1819 +
1.1820 +//------------------------------ has_use_in_set -------------------------------------
1.1821 +// Has a use in the vector set
1.1822 +bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
1.1823 + for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1.1824 + Node* use = n->fast_out(j);
1.1825 + if (vset.test(use->_idx)) {
1.1826 + return true;
1.1827 + }
1.1828 + }
1.1829 + return false;
1.1830 +}
1.1831 +
1.1832 +
1.1833 +//------------------------------ has_use_internal_to_set -------------------------------------
1.1834 +// Has use internal to the vector set (ie. not in a phi at the loop head)
1.1835 +bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
1.1836 + Node* head = loop->_head;
1.1837 + for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1.1838 + Node* use = n->fast_out(j);
1.1839 + if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
1.1840 + return true;
1.1841 + }
1.1842 + }
1.1843 + return false;
1.1844 +}
1.1845 +
1.1846 +
1.1847 +//------------------------------ clone_for_use_outside_loop -------------------------------------
1.1848 +// clone "n" for uses that are outside of loop
1.1849 +void PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
1.1850 +
1.1851 + assert(worklist.size() == 0, "should be empty");
1.1852 + for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1.1853 + Node* use = n->fast_out(j);
1.1854 + if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
1.1855 + worklist.push(use);
1.1856 + }
1.1857 + }
1.1858 + while( worklist.size() ) {
1.1859 + Node *use = worklist.pop();
1.1860 + if (!has_node(use) || use->in(0) == C->top()) continue;
1.1861 + uint j;
1.1862 + for (j = 0; j < use->req(); j++) {
1.1863 + if (use->in(j) == n) break;
1.1864 + }
1.1865 + assert(j < use->req(), "must be there");
1.1866 +
1.1867 + // clone "n" and insert it between the inputs of "n" and the use outside the loop
1.1868 + Node* n_clone = n->clone();
1.1869 + _igvn.hash_delete(use);
1.1870 + use->set_req(j, n_clone);
1.1871 + _igvn._worklist.push(use);
1.1872 + if (!use->is_Phi()) {
1.1873 + Node* use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
1.1874 + set_ctrl(n_clone, use_c);
1.1875 + assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
1.1876 + get_loop(use_c)->_body.push(n_clone);
1.1877 + } else {
1.1878 + // Use in a phi is considered a use in the associated predecessor block
1.1879 + Node *prevbb = use->in(0)->in(j);
1.1880 + set_ctrl(n_clone, prevbb);
1.1881 + assert(!loop->is_member(get_loop(prevbb)), "should be outside loop");
1.1882 + get_loop(prevbb)->_body.push(n_clone);
1.1883 + }
1.1884 + _igvn.register_new_node_with_optimizer(n_clone);
1.1885 +#if !defined(PRODUCT)
1.1886 + if (TracePartialPeeling) {
1.1887 + tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
1.1888 + }
1.1889 +#endif
1.1890 + }
1.1891 +}
1.1892 +
1.1893 +
1.1894 +//------------------------------ clone_for_special_use_inside_loop -------------------------------------
1.1895 +// clone "n" for special uses that are in the not_peeled region.
1.1896 +// If these def-uses occur in separate blocks, the code generator
1.1897 +// marks the method as not compilable. For example, if a "BoolNode"
1.1898 +// is in a different basic block than the "IfNode" that uses it, then
1.1899 +// the compilation is aborted in the code generator.
1.1900 +void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
1.1901 + VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
1.1902 + if (n->is_Phi() || n->is_Load()) {
1.1903 + return;
1.1904 + }
1.1905 + assert(worklist.size() == 0, "should be empty");
1.1906 + for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1.1907 + Node* use = n->fast_out(j);
1.1908 + if ( not_peel.test(use->_idx) &&
1.1909 + (use->is_If() || use->is_CMove() || use->is_Bool()) &&
1.1910 + use->in(1) == n) {
1.1911 + worklist.push(use);
1.1912 + }
1.1913 + }
1.1914 + if (worklist.size() > 0) {
1.1915 + // clone "n" and insert it between inputs of "n" and the use
1.1916 + Node* n_clone = n->clone();
1.1917 + loop->_body.push(n_clone);
1.1918 + _igvn.register_new_node_with_optimizer(n_clone);
1.1919 + set_ctrl(n_clone, get_ctrl(n));
1.1920 + sink_list.push(n_clone);
1.1921 + not_peel <<= n_clone->_idx; // add n_clone to not_peel set.
1.1922 +#if !defined(PRODUCT)
1.1923 + if (TracePartialPeeling) {
1.1924 + tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
1.1925 + }
1.1926 +#endif
1.1927 + while( worklist.size() ) {
1.1928 + Node *use = worklist.pop();
1.1929 + _igvn.hash_delete(use);
1.1930 + _igvn._worklist.push(use);
1.1931 + for (uint j = 1; j < use->req(); j++) {
1.1932 + if (use->in(j) == n) {
1.1933 + use->set_req(j, n_clone);
1.1934 + }
1.1935 + }
1.1936 + }
1.1937 + }
1.1938 +}
1.1939 +
1.1940 +
1.1941 +//------------------------------ insert_phi_for_loop -------------------------------------
1.1942 +// Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
1.1943 +void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
1.1944 + Node *phi = PhiNode::make(lp, back_edge_val);
1.1945 + phi->set_req(LoopNode::EntryControl, lp_entry_val);
1.1946 + // Use existing phi if it already exists
1.1947 + Node *hit = _igvn.hash_find_insert(phi);
1.1948 + if( hit == NULL ) {
1.1949 + _igvn.register_new_node_with_optimizer(phi);
1.1950 + set_ctrl(phi, lp);
1.1951 + } else {
1.1952 + // Remove the new phi from the graph and use the hit
1.1953 + _igvn.remove_dead_node(phi);
1.1954 + phi = hit;
1.1955 + }
1.1956 + _igvn.hash_delete(use);
1.1957 + _igvn._worklist.push(use);
1.1958 + use->set_req(idx, phi);
1.1959 +}
1.1960 +
1.1961 +#ifdef ASSERT
1.1962 +//------------------------------ is_valid_loop_partition -------------------------------------
1.1963 +// Validate the loop partition sets: peel and not_peel
1.1964 +bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
1.1965 + VectorSet& not_peel ) {
1.1966 + uint i;
1.1967 + // Check that peel_list entries are in the peel set
1.1968 + for (i = 0; i < peel_list.size(); i++) {
1.1969 + if (!peel.test(peel_list.at(i)->_idx)) {
1.1970 + return false;
1.1971 + }
1.1972 + }
1.1973 + // Check at loop members are in one of peel set or not_peel set
1.1974 + for (i = 0; i < loop->_body.size(); i++ ) {
1.1975 + Node *def = loop->_body.at(i);
1.1976 + uint di = def->_idx;
1.1977 + // Check that peel set elements are in peel_list
1.1978 + if (peel.test(di)) {
1.1979 + if (not_peel.test(di)) {
1.1980 + return false;
1.1981 + }
1.1982 + // Must be in peel_list also
1.1983 + bool found = false;
1.1984 + for (uint j = 0; j < peel_list.size(); j++) {
1.1985 + if (peel_list.at(j)->_idx == di) {
1.1986 + found = true;
1.1987 + break;
1.1988 + }
1.1989 + }
1.1990 + if (!found) {
1.1991 + return false;
1.1992 + }
1.1993 + } else if (not_peel.test(di)) {
1.1994 + if (peel.test(di)) {
1.1995 + return false;
1.1996 + }
1.1997 + } else {
1.1998 + return false;
1.1999 + }
1.2000 + }
1.2001 + return true;
1.2002 +}
1.2003 +
1.2004 +//------------------------------ is_valid_clone_loop_exit_use -------------------------------------
1.2005 +// Ensure a use outside of loop is of the right form
1.2006 +bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
1.2007 + Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
1.2008 + return (use->is_Phi() &&
1.2009 + use_c->is_Region() && use_c->req() == 3 &&
1.2010 + (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
1.2011 + use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
1.2012 + use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
1.2013 + loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
1.2014 +}
1.2015 +
1.2016 +//------------------------------ is_valid_clone_loop_form -------------------------------------
1.2017 +// Ensure that all uses outside of loop are of the right form
1.2018 +bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
1.2019 + uint orig_exit_idx, uint clone_exit_idx) {
1.2020 + uint len = peel_list.size();
1.2021 + for (uint i = 0; i < len; i++) {
1.2022 + Node *def = peel_list.at(i);
1.2023 +
1.2024 + for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1.2025 + Node *use = def->fast_out(j);
1.2026 + Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
1.2027 + if (!loop->is_member(get_loop(use_c))) {
1.2028 + // use is not in the loop, check for correct structure
1.2029 + if (use->in(0) == def) {
1.2030 + // Okay
1.2031 + } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
1.2032 + return false;
1.2033 + }
1.2034 + }
1.2035 + }
1.2036 + }
1.2037 + return true;
1.2038 +}
1.2039 +#endif
1.2040 +
1.2041 +//------------------------------ partial_peel -------------------------------------
1.2042 +// Partially peel (aka loop rotation) the top portion of a loop (called
1.2043 +// the peel section below) by cloning it and placing one copy just before
1.2044 +// the new loop head and the other copy at the bottom of the new loop.
1.2045 +//
1.2046 +// before after where it came from
1.2047 +//
1.2048 +// stmt1 stmt1
1.2049 +// loop: stmt2 clone
1.2050 +// stmt2 if condA goto exitA clone
1.2051 +// if condA goto exitA new_loop: new
1.2052 +// stmt3 stmt3 clone
1.2053 +// if !condB goto loop if condB goto exitB clone
1.2054 +// exitB: stmt2 orig
1.2055 +// stmt4 if !condA goto new_loop orig
1.2056 +// exitA: goto exitA
1.2057 +// exitB:
1.2058 +// stmt4
1.2059 +// exitA:
1.2060 +//
1.2061 +// Step 1: find the cut point: an exit test on probable
1.2062 +// induction variable.
1.2063 +// Step 2: schedule (with cloning) operations in the peel
1.2064 +// section that can be executed after the cut into
1.2065 +// the section that is not peeled. This may need
1.2066 +// to clone operations into exit blocks. For
1.2067 +// instance, a reference to A[i] in the not-peel
1.2068 +// section and a reference to B[i] in an exit block
1.2069 +// may cause a left-shift of i by 2 to be placed
1.2070 +// in the peel block. This step will clone the left
1.2071 +// shift into the exit block and sink the left shift
1.2072 +// from the peel to the not-peel section.
1.2073 +// Step 3: clone the loop, retarget the control, and insert
1.2074 +// phis for values that are live across the new loop
1.2075 +// head. This is very dependent on the graph structure
1.2076 +// from clone_loop. It creates region nodes for
1.2077 +// exit control and associated phi nodes for values
1.2078 +// flow out of the loop through that exit. The region
1.2079 +// node is dominated by the clone's control projection.
1.2080 +// So the clone's peel section is placed before the
1.2081 +// new loop head, and the clone's not-peel section is
1.2082 +// forms the top part of the new loop. The original
1.2083 +// peel section forms the tail of the new loop.
1.2084 +// Step 4: update the dominator tree and recompute the
1.2085 +// dominator depth.
1.2086 +//
1.2087 +// orig
1.2088 +//
1.2089 +// stmt1
1.2090 +// |
1.2091 +// v
1.2092 +// loop<----+
1.2093 +// | |
1.2094 +// stmt2 |
1.2095 +// | |
1.2096 +// v |
1.2097 +// ifA |
1.2098 +// / | |
1.2099 +// v v |
1.2100 +// false true ^ <-- last_peel
1.2101 +// / | |
1.2102 +// / ===|==cut |
1.2103 +// / stmt3 | <-- first_not_peel
1.2104 +// / | |
1.2105 +// | v |
1.2106 +// v ifB |
1.2107 +// exitA: / \ |
1.2108 +// / \ |
1.2109 +// v v |
1.2110 +// false true |
1.2111 +// / \ |
1.2112 +// / ----+
1.2113 +// |
1.2114 +// v
1.2115 +// exitB:
1.2116 +// stmt4
1.2117 +//
1.2118 +//
1.2119 +// after clone loop
1.2120 +//
1.2121 +// stmt1
1.2122 +// / \
1.2123 +// clone / \ orig
1.2124 +// / \
1.2125 +// / \
1.2126 +// v v
1.2127 +// +---->loop loop<----+
1.2128 +// | | | |
1.2129 +// | stmt2 stmt2 |
1.2130 +// | | | |
1.2131 +// | v v |
1.2132 +// | ifA ifA |
1.2133 +// | | \ / | |
1.2134 +// | v v v v |
1.2135 +// ^ true false false true ^ <-- last_peel
1.2136 +// | | ^ \ / | |
1.2137 +// | cut==|== \ \ / ===|==cut |
1.2138 +// | stmt3 \ \ / stmt3 | <-- first_not_peel
1.2139 +// | | dom | | | |
1.2140 +// | v \ 1v v2 v |
1.2141 +// | ifB regionA ifB |
1.2142 +// | / \ | / \ |
1.2143 +// | / \ v / \ |
1.2144 +// | v v exitA: v v |
1.2145 +// | true false false true |
1.2146 +// | / ^ \ / \ |
1.2147 +// +---- \ \ / ----+
1.2148 +// dom \ /
1.2149 +// \ 1v v2
1.2150 +// regionB
1.2151 +// |
1.2152 +// v
1.2153 +// exitB:
1.2154 +// stmt4
1.2155 +//
1.2156 +//
1.2157 +// after partial peel
1.2158 +//
1.2159 +// stmt1
1.2160 +// /
1.2161 +// clone / orig
1.2162 +// / TOP
1.2163 +// / \
1.2164 +// v v
1.2165 +// TOP->region region----+
1.2166 +// | | |
1.2167 +// stmt2 stmt2 |
1.2168 +// | | |
1.2169 +// v v |
1.2170 +// ifA ifA |
1.2171 +// | \ / | |
1.2172 +// v v v v |
1.2173 +// true false false true | <-- last_peel
1.2174 +// | ^ \ / +------|---+
1.2175 +// +->newloop \ \ / === ==cut | |
1.2176 +// | stmt3 \ \ / TOP | |
1.2177 +// | | dom | | stmt3 | | <-- first_not_peel
1.2178 +// | v \ 1v v2 v | |
1.2179 +// | ifB regionA ifB ^ v
1.2180 +// | / \ | / \ | |
1.2181 +// | / \ v / \ | |
1.2182 +// | v v exitA: v v | |
1.2183 +// | true false false true | |
1.2184 +// | / ^ \ / \ | |
1.2185 +// | | \ \ / v | |
1.2186 +// | | dom \ / TOP | |
1.2187 +// | | \ 1v v2 | |
1.2188 +// ^ v regionB | |
1.2189 +// | | | | |
1.2190 +// | | v ^ v
1.2191 +// | | exitB: | |
1.2192 +// | | stmt4 | |
1.2193 +// | +------------>-----------------+ |
1.2194 +// | |
1.2195 +// +-----------------<---------------------+
1.2196 +//
1.2197 +//
1.2198 +// final graph
1.2199 +//
1.2200 +// stmt1
1.2201 +// |
1.2202 +// v
1.2203 +// ........> ifA clone
1.2204 +// : / |
1.2205 +// dom / |
1.2206 +// : v v
1.2207 +// : false true
1.2208 +// : | |
1.2209 +// : | stmt2 clone
1.2210 +// : | |
1.2211 +// : | v
1.2212 +// : | newloop<-----+
1.2213 +// : | | |
1.2214 +// : | stmt3 clone |
1.2215 +// : | | |
1.2216 +// : | v |
1.2217 +// : | ifB |
1.2218 +// : | / \ |
1.2219 +// : | v v |
1.2220 +// : | false true |
1.2221 +// : | | | |
1.2222 +// : | v stmt2 |
1.2223 +// : | exitB: | |
1.2224 +// : | stmt4 v |
1.2225 +// : | ifA orig |
1.2226 +// : | / \ |
1.2227 +// : | / \ |
1.2228 +// : | v v |
1.2229 +// : | false true |
1.2230 +// : | / \ |
1.2231 +// : v v -----+
1.2232 +// RegionA
1.2233 +// |
1.2234 +// v
1.2235 +// exitA
1.2236 +//
1.2237 +bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
1.2238 +
1.2239 + LoopNode *head = loop->_head->as_Loop();
1.2240 +
1.2241 + if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
1.2242 + return false;
1.2243 + }
1.2244 +
1.2245 + // Check for complex exit control
1.2246 + for(uint ii = 0; ii < loop->_body.size(); ii++ ) {
1.2247 + Node *n = loop->_body.at(ii);
1.2248 + int opc = n->Opcode();
1.2249 + if (n->is_Call() ||
1.2250 + opc == Op_Catch ||
1.2251 + opc == Op_CatchProj ||
1.2252 + opc == Op_Jump ||
1.2253 + opc == Op_JumpProj) {
1.2254 +#if !defined(PRODUCT)
1.2255 + if (TracePartialPeeling) {
1.2256 + tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
1.2257 + }
1.2258 +#endif
1.2259 + return false;
1.2260 + }
1.2261 + }
1.2262 +
1.2263 + int dd = dom_depth(head);
1.2264 +
1.2265 + // Step 1: find cut point
1.2266 +
1.2267 + // Walk up dominators to loop head looking for first loop exit
1.2268 + // which is executed on every path thru loop.
1.2269 + IfNode *peel_if = NULL;
1.2270 + IfNode *peel_if_cmpu = NULL;
1.2271 +
1.2272 + Node *iff = loop->tail();
1.2273 + while( iff != head ) {
1.2274 + if( iff->is_If() ) {
1.2275 + Node *ctrl = get_ctrl(iff->in(1));
1.2276 + if (ctrl->is_top()) return false; // Dead test on live IF.
1.2277 + // If loop-varying exit-test, check for induction variable
1.2278 + if( loop->is_member(get_loop(ctrl)) &&
1.2279 + loop->is_loop_exit(iff) &&
1.2280 + is_possible_iv_test(iff)) {
1.2281 + Node* cmp = iff->in(1)->in(1);
1.2282 + if (cmp->Opcode() == Op_CmpI) {
1.2283 + peel_if = iff->as_If();
1.2284 + } else {
1.2285 + assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
1.2286 + peel_if_cmpu = iff->as_If();
1.2287 + }
1.2288 + }
1.2289 + }
1.2290 + iff = idom(iff);
1.2291 + }
1.2292 + // Prefer signed compare over unsigned compare.
1.2293 + IfNode* new_peel_if = NULL;
1.2294 + if (peel_if == NULL) {
1.2295 + if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) {
1.2296 + return false; // No peel point found
1.2297 + }
1.2298 + new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
1.2299 + if (new_peel_if == NULL) {
1.2300 + return false; // No peel point found
1.2301 + }
1.2302 + peel_if = new_peel_if;
1.2303 + }
1.2304 + Node* last_peel = stay_in_loop(peel_if, loop);
1.2305 + Node* first_not_peeled = stay_in_loop(last_peel, loop);
1.2306 + if (first_not_peeled == NULL || first_not_peeled == head) {
1.2307 + return false;
1.2308 + }
1.2309 +
1.2310 +#if !defined(PRODUCT)
1.2311 + if (TracePartialPeeling) {
1.2312 + tty->print_cr("before partial peel one iteration");
1.2313 + Node_List wl;
1.2314 + Node* t = head->in(2);
1.2315 + while (true) {
1.2316 + wl.push(t);
1.2317 + if (t == head) break;
1.2318 + t = idom(t);
1.2319 + }
1.2320 + while (wl.size() > 0) {
1.2321 + Node* tt = wl.pop();
1.2322 + tt->dump();
1.2323 + if (tt == last_peel) tty->print_cr("-- cut --");
1.2324 + }
1.2325 + }
1.2326 +#endif
1.2327 + ResourceArea *area = Thread::current()->resource_area();
1.2328 + VectorSet peel(area);
1.2329 + VectorSet not_peel(area);
1.2330 + Node_List peel_list(area);
1.2331 + Node_List worklist(area);
1.2332 + Node_List sink_list(area);
1.2333 +
1.2334 + // Set of cfg nodes to peel are those that are executable from
1.2335 + // the head through last_peel.
1.2336 + assert(worklist.size() == 0, "should be empty");
1.2337 + worklist.push(head);
1.2338 + peel.set(head->_idx);
1.2339 + while (worklist.size() > 0) {
1.2340 + Node *n = worklist.pop();
1.2341 + if (n != last_peel) {
1.2342 + for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1.2343 + Node* use = n->fast_out(j);
1.2344 + if (use->is_CFG() &&
1.2345 + loop->is_member(get_loop(use)) &&
1.2346 + !peel.test_set(use->_idx)) {
1.2347 + worklist.push(use);
1.2348 + }
1.2349 + }
1.2350 + }
1.2351 + }
1.2352 +
1.2353 + // Set of non-cfg nodes to peel are those that are control
1.2354 + // dependent on the cfg nodes.
1.2355 + uint i;
1.2356 + for(i = 0; i < loop->_body.size(); i++ ) {
1.2357 + Node *n = loop->_body.at(i);
1.2358 + Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
1.2359 + if (peel.test(n_c->_idx)) {
1.2360 + peel.set(n->_idx);
1.2361 + } else {
1.2362 + not_peel.set(n->_idx);
1.2363 + }
1.2364 + }
1.2365 +
1.2366 + // Step 2: move operations from the peeled section down into the
1.2367 + // not-peeled section
1.2368 +
1.2369 + // Get a post order schedule of nodes in the peel region
1.2370 + // Result in right-most operand.
1.2371 + scheduled_nodelist(loop, peel, peel_list );
1.2372 +
1.2373 + assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
1.2374 +
1.2375 + // For future check for too many new phis
1.2376 + uint old_phi_cnt = 0;
1.2377 + for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
1.2378 + Node* use = head->fast_out(j);
1.2379 + if (use->is_Phi()) old_phi_cnt++;
1.2380 + }
1.2381 +
1.2382 +#if !defined(PRODUCT)
1.2383 + if (TracePartialPeeling) {
1.2384 + tty->print_cr("\npeeled list");
1.2385 + }
1.2386 +#endif
1.2387 +
1.2388 + // Evacuate nodes in peel region into the not_peeled region if possible
1.2389 + uint new_phi_cnt = 0;
1.2390 + for (i = 0; i < peel_list.size();) {
1.2391 + Node* n = peel_list.at(i);
1.2392 +#if !defined(PRODUCT)
1.2393 + if (TracePartialPeeling) n->dump();
1.2394 +#endif
1.2395 + bool incr = true;
1.2396 + if ( !n->is_CFG() ) {
1.2397 +
1.2398 + if ( has_use_in_set(n, not_peel) ) {
1.2399 +
1.2400 + // If not used internal to the peeled region,
1.2401 + // move "n" from peeled to not_peeled region.
1.2402 +
1.2403 + if ( !has_use_internal_to_set(n, peel, loop) ) {
1.2404 +
1.2405 + // if not pinned and not a load (which maybe anti-dependent on a store)
1.2406 + // and not a CMove (Matcher expects only bool->cmove).
1.2407 + if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) {
1.2408 + clone_for_use_outside_loop( loop, n, worklist );
1.2409 +
1.2410 + sink_list.push(n);
1.2411 + peel >>= n->_idx; // delete n from peel set.
1.2412 + not_peel <<= n->_idx; // add n to not_peel set.
1.2413 + peel_list.remove(i);
1.2414 + incr = false;
1.2415 +#if !defined(PRODUCT)
1.2416 + if (TracePartialPeeling) {
1.2417 + tty->print_cr("sink to not_peeled region: %d newbb: %d",
1.2418 + n->_idx, get_ctrl(n)->_idx);
1.2419 + }
1.2420 +#endif
1.2421 + }
1.2422 + } else {
1.2423 + // Otherwise check for special def-use cases that span
1.2424 + // the peel/not_peel boundary such as bool->if
1.2425 + clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist );
1.2426 + new_phi_cnt++;
1.2427 + }
1.2428 + }
1.2429 + }
1.2430 + if (incr) i++;
1.2431 + }
1.2432 +
1.2433 + if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) {
1.2434 +#if !defined(PRODUCT)
1.2435 + if (TracePartialPeeling) {
1.2436 + tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c",
1.2437 + new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F');
1.2438 + }
1.2439 +#endif
1.2440 + if (new_peel_if != NULL) {
1.2441 + remove_cmpi_loop_exit(new_peel_if, loop);
1.2442 + }
1.2443 + // Inhibit more partial peeling on this loop
1.2444 + assert(!head->is_partial_peel_loop(), "not partial peeled");
1.2445 + head->mark_partial_peel_failed();
1.2446 + return false;
1.2447 + }
1.2448 +
1.2449 + // Step 3: clone loop, retarget control, and insert new phis
1.2450 +
1.2451 + // Create new loop head for new phis and to hang
1.2452 + // the nodes being moved (sinked) from the peel region.
1.2453 + LoopNode* new_head = new (C, 3) LoopNode(last_peel, last_peel);
1.2454 + _igvn.register_new_node_with_optimizer(new_head);
1.2455 + assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
1.2456 + first_not_peeled->set_req(0, new_head);
1.2457 + set_loop(new_head, loop);
1.2458 + loop->_body.push(new_head);
1.2459 + not_peel.set(new_head->_idx);
1.2460 + set_idom(new_head, last_peel, dom_depth(first_not_peeled));
1.2461 + set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
1.2462 +
1.2463 + while (sink_list.size() > 0) {
1.2464 + Node* n = sink_list.pop();
1.2465 + set_ctrl(n, new_head);
1.2466 + }
1.2467 +
1.2468 + assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
1.2469 +
1.2470 + clone_loop( loop, old_new, dd );
1.2471 +
1.2472 + const uint clone_exit_idx = 1;
1.2473 + const uint orig_exit_idx = 2;
1.2474 + assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop");
1.2475 +
1.2476 + Node* head_clone = old_new[head->_idx];
1.2477 + LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop();
1.2478 + Node* orig_tail_clone = head_clone->in(2);
1.2479 +
1.2480 + // Add phi if "def" node is in peel set and "use" is not
1.2481 +
1.2482 + for(i = 0; i < peel_list.size(); i++ ) {
1.2483 + Node *def = peel_list.at(i);
1.2484 + if (!def->is_CFG()) {
1.2485 + for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
1.2486 + Node *use = def->fast_out(j);
1.2487 + if (has_node(use) && use->in(0) != C->top() &&
1.2488 + (!peel.test(use->_idx) ||
1.2489 + (use->is_Phi() && use->in(0) == head)) ) {
1.2490 + worklist.push(use);
1.2491 + }
1.2492 + }
1.2493 + while( worklist.size() ) {
1.2494 + Node *use = worklist.pop();
1.2495 + for (uint j = 1; j < use->req(); j++) {
1.2496 + Node* n = use->in(j);
1.2497 + if (n == def) {
1.2498 +
1.2499 + // "def" is in peel set, "use" is not in peel set
1.2500 + // or "use" is in the entry boundary (a phi) of the peel set
1.2501 +
1.2502 + Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
1.2503 +
1.2504 + if ( loop->is_member(get_loop( use_c )) ) {
1.2505 + // use is in loop
1.2506 + if (old_new[use->_idx] != NULL) { // null for dead code
1.2507 + Node* use_clone = old_new[use->_idx];
1.2508 + _igvn.hash_delete(use);
1.2509 + use->set_req(j, C->top());
1.2510 + _igvn._worklist.push(use);
1.2511 + insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
1.2512 + }
1.2513 + } else {
1.2514 + assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
1.2515 + // use is not in the loop, check if the live range includes the cut
1.2516 + Node* lp_if = use_c->in(orig_exit_idx)->in(0);
1.2517 + if (not_peel.test(lp_if->_idx)) {
1.2518 + assert(j == orig_exit_idx, "use from original loop");
1.2519 + insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
1.2520 + }
1.2521 + }
1.2522 + }
1.2523 + }
1.2524 + }
1.2525 + }
1.2526 + }
1.2527 +
1.2528 + // Step 3b: retarget control
1.2529 +
1.2530 + // Redirect control to the new loop head if a cloned node in
1.2531 + // the not_peeled region has control that points into the peeled region.
1.2532 + // This necessary because the cloned peeled region will be outside
1.2533 + // the loop.
1.2534 + // from to
1.2535 + // cloned-peeled <---+
1.2536 + // new_head_clone: | <--+
1.2537 + // cloned-not_peeled in(0) in(0)
1.2538 + // orig-peeled
1.2539 +
1.2540 + for(i = 0; i < loop->_body.size(); i++ ) {
1.2541 + Node *n = loop->_body.at(i);
1.2542 + if (!n->is_CFG() && n->in(0) != NULL &&
1.2543 + not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
1.2544 + Node* n_clone = old_new[n->_idx];
1.2545 + _igvn.hash_delete(n_clone);
1.2546 + n_clone->set_req(0, new_head_clone);
1.2547 + _igvn._worklist.push(n_clone);
1.2548 + }
1.2549 + }
1.2550 +
1.2551 + // Backedge of the surviving new_head (the clone) is original last_peel
1.2552 + _igvn.hash_delete(new_head_clone);
1.2553 + new_head_clone->set_req(LoopNode::LoopBackControl, last_peel);
1.2554 + _igvn._worklist.push(new_head_clone);
1.2555 +
1.2556 + // Cut first node in original not_peel set
1.2557 + _igvn.hash_delete(new_head);
1.2558 + new_head->set_req(LoopNode::EntryControl, C->top());
1.2559 + new_head->set_req(LoopNode::LoopBackControl, C->top());
1.2560 + _igvn._worklist.push(new_head);
1.2561 +
1.2562 + // Copy head_clone back-branch info to original head
1.2563 + // and remove original head's loop entry and
1.2564 + // clone head's back-branch
1.2565 + _igvn.hash_delete(head);
1.2566 + _igvn.hash_delete(head_clone);
1.2567 + head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl));
1.2568 + head->set_req(LoopNode::LoopBackControl, C->top());
1.2569 + head_clone->set_req(LoopNode::LoopBackControl, C->top());
1.2570 + _igvn._worklist.push(head);
1.2571 + _igvn._worklist.push(head_clone);
1.2572 +
1.2573 + // Similarly modify the phis
1.2574 + for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
1.2575 + Node* use = head->fast_out(k);
1.2576 + if (use->is_Phi() && use->outcnt() > 0) {
1.2577 + Node* use_clone = old_new[use->_idx];
1.2578 + _igvn.hash_delete(use);
1.2579 + _igvn.hash_delete(use_clone);
1.2580 + use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl));
1.2581 + use->set_req(LoopNode::LoopBackControl, C->top());
1.2582 + use_clone->set_req(LoopNode::LoopBackControl, C->top());
1.2583 + _igvn._worklist.push(use);
1.2584 + _igvn._worklist.push(use_clone);
1.2585 + }
1.2586 + }
1.2587 +
1.2588 + // Step 4: update dominator tree and dominator depth
1.2589 +
1.2590 + set_idom(head, orig_tail_clone, dd);
1.2591 + recompute_dom_depth();
1.2592 +
1.2593 + // Inhibit more partial peeling on this loop
1.2594 + new_head_clone->set_partial_peel_loop();
1.2595 + C->set_major_progress();
1.2596 +
1.2597 +#if !defined(PRODUCT)
1.2598 + if (TracePartialPeeling) {
1.2599 + tty->print_cr("\nafter partial peel one iteration");
1.2600 + Node_List wl(area);
1.2601 + Node* t = last_peel;
1.2602 + while (true) {
1.2603 + wl.push(t);
1.2604 + if (t == head_clone) break;
1.2605 + t = idom(t);
1.2606 + }
1.2607 + while (wl.size() > 0) {
1.2608 + Node* tt = wl.pop();
1.2609 + if (tt == head) tty->print_cr("orig head");
1.2610 + else if (tt == new_head_clone) tty->print_cr("new head");
1.2611 + else if (tt == head_clone) tty->print_cr("clone head");
1.2612 + tt->dump();
1.2613 + }
1.2614 + }
1.2615 +#endif
1.2616 + return true;
1.2617 +}
1.2618 +
1.2619 +//------------------------------reorg_offsets----------------------------------
1.2620 +// Reorganize offset computations to lower register pressure. Mostly
1.2621 +// prevent loop-fallout uses of the pre-incremented trip counter (which are
1.2622 +// then alive with the post-incremented trip counter forcing an extra
1.2623 +// register move)
1.2624 +void PhaseIdealLoop::reorg_offsets( IdealLoopTree *loop ) {
1.2625 +
1.2626 + CountedLoopNode *cl = loop->_head->as_CountedLoop();
1.2627 + CountedLoopEndNode *cle = cl->loopexit();
1.2628 + if( !cle ) return; // The occasional dead loop
1.2629 + // Find loop exit control
1.2630 + Node *exit = cle->proj_out(false);
1.2631 + assert( exit->Opcode() == Op_IfFalse, "" );
1.2632 +
1.2633 + // Check for the special case of folks using the pre-incremented
1.2634 + // trip-counter on the fall-out path (forces the pre-incremented
1.2635 + // and post-incremented trip counter to be live at the same time).
1.2636 + // Fix this by adjusting to use the post-increment trip counter.
1.2637 + Node *phi = cl->phi();
1.2638 + if( !phi ) return; // Dead infinite loop
1.2639 + bool progress = true;
1.2640 + while (progress) {
1.2641 + progress = false;
1.2642 + for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
1.2643 + Node* use = phi->fast_out(i); // User of trip-counter
1.2644 + if (!has_ctrl(use)) continue;
1.2645 + Node *u_ctrl = get_ctrl(use);
1.2646 + if( use->is_Phi() ) {
1.2647 + u_ctrl = NULL;
1.2648 + for( uint j = 1; j < use->req(); j++ )
1.2649 + if( use->in(j) == phi )
1.2650 + u_ctrl = dom_lca( u_ctrl, use->in(0)->in(j) );
1.2651 + }
1.2652 + IdealLoopTree *u_loop = get_loop(u_ctrl);
1.2653 + // Look for loop-invariant use
1.2654 + if( u_loop == loop ) continue;
1.2655 + if( loop->is_member( u_loop ) ) continue;
1.2656 + // Check that use is live out the bottom. Assuming the trip-counter
1.2657 + // update is right at the bottom, uses of of the loop middle are ok.
1.2658 + if( dom_lca( exit, u_ctrl ) != exit ) continue;
1.2659 + // protect against stride not being a constant
1.2660 + if( !cle->stride_is_con() ) continue;
1.2661 + // Hit! Refactor use to use the post-incremented tripcounter.
1.2662 + // Compute a post-increment tripcounter.
1.2663 + Node *opaq = new (C, 2) Opaque2Node( cle->incr() );
1.2664 + register_new_node( opaq, u_ctrl );
1.2665 + Node *neg_stride = _igvn.intcon(-cle->stride_con());
1.2666 + set_ctrl(neg_stride, C->root());
1.2667 + Node *post = new (C, 3) AddINode( opaq, neg_stride);
1.2668 + register_new_node( post, u_ctrl );
1.2669 + _igvn.hash_delete(use);
1.2670 + _igvn._worklist.push(use);
1.2671 + for( uint j = 1; j < use->req(); j++ )
1.2672 + if( use->in(j) == phi )
1.2673 + use->set_req(j, post);
1.2674 + // Since DU info changed, rerun loop
1.2675 + progress = true;
1.2676 + break;
1.2677 + }
1.2678 + }
1.2679 +
1.2680 +}
