duke@435: /* xdono@631: * Copyright 1999-2008 Sun Microsystems, Inc. All Rights Reserved. duke@435: * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. duke@435: * duke@435: * This code is free software; you can redistribute it and/or modify it duke@435: * under the terms of the GNU General Public License version 2 only, as duke@435: * published by the Free Software Foundation. duke@435: * duke@435: * This code is distributed in the hope that it will be useful, but WITHOUT duke@435: * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or duke@435: * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License duke@435: * version 2 for more details (a copy is included in the LICENSE file that duke@435: * accompanied this code). duke@435: * duke@435: * You should have received a copy of the GNU General Public License version duke@435: * 2 along with this work; if not, write to the Free Software Foundation, duke@435: * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. duke@435: * duke@435: * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, duke@435: * CA 95054 USA or visit www.sun.com if you need additional information or duke@435: * have any questions. duke@435: * duke@435: */ duke@435: duke@435: #include "incls/_precompiled.incl" duke@435: #include "incls/_loopopts.cpp.incl" duke@435: duke@435: //============================================================================= duke@435: //------------------------------split_thru_phi--------------------------------- duke@435: // Split Node 'n' through merge point if there is enough win. duke@435: Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) { never@504: if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) { never@504: // ConvI2L may have type information on it which is unsafe to push up never@504: // so disable this for now never@504: return NULL; never@504: } duke@435: int wins = 0; duke@435: assert( !n->is_CFG(), "" ); duke@435: assert( region->is_Region(), "" ); kvn@499: kvn@499: const Type* type = n->bottom_type(); kvn@499: const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr(); kvn@499: Node *phi; kvn@658: if( t_oop != NULL && t_oop->is_known_instance_field() ) { kvn@499: int iid = t_oop->instance_id(); kvn@499: int index = C->get_alias_index(t_oop); kvn@499: int offset = t_oop->offset(); kvn@499: phi = new (C,region->req()) PhiNode(region, type, NULL, iid, index, offset); kvn@499: } else { kvn@499: phi = new (C,region->req()) PhiNode(region, type); kvn@499: } duke@435: uint old_unique = C->unique(); duke@435: for( uint i = 1; i < region->req(); i++ ) { duke@435: Node *x; duke@435: Node* the_clone = NULL; duke@435: if( region->in(i) == C->top() ) { duke@435: x = C->top(); // Dead path? Use a dead data op duke@435: } else { duke@435: x = n->clone(); // Else clone up the data op duke@435: the_clone = x; // Remember for possible deletion. duke@435: // Alter data node to use pre-phi inputs duke@435: if( n->in(0) == region ) duke@435: x->set_req( 0, region->in(i) ); duke@435: for( uint j = 1; j < n->req(); j++ ) { duke@435: Node *in = n->in(j); duke@435: if( in->is_Phi() && in->in(0) == region ) duke@435: x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone duke@435: } duke@435: } duke@435: // Check for a 'win' on some paths duke@435: const Type *t = x->Value(&_igvn); duke@435: duke@435: bool singleton = t->singleton(); duke@435: duke@435: // A TOP singleton indicates that there are no possible values incoming duke@435: // along a particular edge. In most cases, this is OK, and the Phi will duke@435: // be eliminated later in an Ideal call. However, we can't allow this to duke@435: // happen if the singleton occurs on loop entry, as the elimination of duke@435: // the PhiNode may cause the resulting node to migrate back to a previous duke@435: // loop iteration. duke@435: if( singleton && t == Type::TOP ) { duke@435: // Is_Loop() == false does not confirm the absence of a loop (e.g., an duke@435: // irreducible loop may not be indicated by an affirmative is_Loop()); duke@435: // therefore, the only top we can split thru a phi is on a backedge of duke@435: // a loop. duke@435: singleton &= region->is_Loop() && (i != LoopNode::EntryControl); duke@435: } duke@435: duke@435: if( singleton ) { duke@435: wins++; duke@435: x = ((PhaseGVN&)_igvn).makecon(t); duke@435: } else { duke@435: // We now call Identity to try to simplify the cloned node. duke@435: // Note that some Identity methods call phase->type(this). duke@435: // Make sure that the type array is big enough for duke@435: // our new node, even though we may throw the node away. duke@435: // (Note: This tweaking with igvn only works because x is a new node.) duke@435: _igvn.set_type(x, t); kvn@728: // If x is a TypeNode, capture any more-precise type permanently into Node twisti@1040: // otherwise it will be not updated during igvn->transform since kvn@728: // igvn->type(x) is set to x->Value() already. kvn@728: x->raise_bottom_type(t); duke@435: Node *y = x->Identity(&_igvn); duke@435: if( y != x ) { duke@435: wins++; duke@435: x = y; duke@435: } else { duke@435: y = _igvn.hash_find(x); duke@435: if( y ) { duke@435: wins++; duke@435: x = y; duke@435: } else { duke@435: // Else x is a new node we are keeping duke@435: // We do not need register_new_node_with_optimizer duke@435: // because set_type has already been called. duke@435: _igvn._worklist.push(x); duke@435: } duke@435: } duke@435: } duke@435: if (x != the_clone && the_clone != NULL) duke@435: _igvn.remove_dead_node(the_clone); duke@435: phi->set_req( i, x ); duke@435: } duke@435: // Too few wins? duke@435: if( wins <= policy ) { duke@435: _igvn.remove_dead_node(phi); duke@435: return NULL; duke@435: } duke@435: duke@435: // Record Phi duke@435: register_new_node( phi, region ); duke@435: duke@435: for( uint i2 = 1; i2 < phi->req(); i2++ ) { duke@435: Node *x = phi->in(i2); duke@435: // If we commoned up the cloned 'x' with another existing Node, duke@435: // the existing Node picks up a new use. We need to make the duke@435: // existing Node occur higher up so it dominates its uses. duke@435: Node *old_ctrl; duke@435: IdealLoopTree *old_loop; duke@435: duke@435: // The occasional new node duke@435: if( x->_idx >= old_unique ) { // Found a new, unplaced node? duke@435: old_ctrl = x->is_Con() ? C->root() : NULL; duke@435: old_loop = NULL; // Not in any prior loop duke@435: } else { duke@435: old_ctrl = x->is_Con() ? C->root() : get_ctrl(x); duke@435: old_loop = get_loop(old_ctrl); // Get prior loop duke@435: } duke@435: // New late point must dominate new use duke@435: Node *new_ctrl = dom_lca( old_ctrl, region->in(i2) ); duke@435: // Set new location duke@435: set_ctrl(x, new_ctrl); duke@435: IdealLoopTree *new_loop = get_loop( new_ctrl ); duke@435: // If changing loop bodies, see if we need to collect into new body duke@435: if( old_loop != new_loop ) { duke@435: if( old_loop && !old_loop->_child ) duke@435: old_loop->_body.yank(x); duke@435: if( !new_loop->_child ) duke@435: new_loop->_body.push(x); // Collect body info duke@435: } duke@435: } duke@435: duke@435: return phi; duke@435: } duke@435: duke@435: //------------------------------dominated_by------------------------------------ duke@435: // Replace the dominated test with an obvious true or false. Place it on the duke@435: // IGVN worklist for later cleanup. Move control-dependent data Nodes on the duke@435: // live path up to the dominating control. duke@435: void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff ) { duke@435: #ifndef PRODUCT duke@435: if( VerifyLoopOptimizations && PrintOpto ) tty->print_cr("dominating test"); duke@435: #endif duke@435: duke@435: duke@435: // prevdom is the dominating projection of the dominating test. duke@435: assert( iff->is_If(), "" ); duke@435: assert( iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added"); duke@435: int pop = prevdom->Opcode(); duke@435: assert( pop == Op_IfFalse || pop == Op_IfTrue, "" ); duke@435: // 'con' is set to true or false to kill the dominated test. duke@435: Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO); duke@435: set_ctrl(con, C->root()); // Constant gets a new use duke@435: // Hack the dominated test duke@435: _igvn.hash_delete(iff); duke@435: iff->set_req(1, con); duke@435: _igvn._worklist.push(iff); duke@435: duke@435: // If I dont have a reachable TRUE and FALSE path following the IfNode then duke@435: // I can assume this path reaches an infinite loop. In this case it's not duke@435: // important to optimize the data Nodes - either the whole compilation will duke@435: // be tossed or this path (and all data Nodes) will go dead. duke@435: if( iff->outcnt() != 2 ) return; duke@435: duke@435: // Make control-dependent data Nodes on the live path (path that will remain duke@435: // once the dominated IF is removed) become control-dependent on the duke@435: // dominating projection. duke@435: Node* dp = ((IfNode*)iff)->proj_out(pop == Op_IfTrue); duke@435: IdealLoopTree *old_loop = get_loop(dp); duke@435: duke@435: for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { duke@435: Node* cd = dp->fast_out(i); // Control-dependent node duke@435: if( cd->depends_only_on_test() ) { duke@435: assert( cd->in(0) == dp, "" ); duke@435: _igvn.hash_delete( cd ); duke@435: cd->set_req(0, prevdom); duke@435: set_early_ctrl( cd ); duke@435: _igvn._worklist.push(cd); duke@435: IdealLoopTree *new_loop = get_loop(get_ctrl(cd)); duke@435: if( old_loop != new_loop ) { duke@435: if( !old_loop->_child ) old_loop->_body.yank(cd); duke@435: if( !new_loop->_child ) new_loop->_body.push(cd); duke@435: } duke@435: --i; duke@435: --imax; duke@435: } duke@435: } duke@435: } duke@435: duke@435: //------------------------------has_local_phi_input---------------------------- duke@435: // Return TRUE if 'n' has Phi inputs from its local block and no other duke@435: // block-local inputs (all non-local-phi inputs come from earlier blocks) duke@435: Node *PhaseIdealLoop::has_local_phi_input( Node *n ) { duke@435: Node *n_ctrl = get_ctrl(n); duke@435: // See if some inputs come from a Phi in this block, or from before duke@435: // this block. duke@435: uint i; duke@435: for( i = 1; i < n->req(); i++ ) { duke@435: Node *phi = n->in(i); duke@435: if( phi->is_Phi() && phi->in(0) == n_ctrl ) duke@435: break; duke@435: } duke@435: if( i >= n->req() ) duke@435: return NULL; // No Phi inputs; nowhere to clone thru duke@435: duke@435: // Check for inputs created between 'n' and the Phi input. These duke@435: // must split as well; they have already been given the chance duke@435: // (courtesy of a post-order visit) and since they did not we must duke@435: // recover the 'cost' of splitting them by being very profitable duke@435: // when splitting 'n'. Since this is unlikely we simply give up. duke@435: for( i = 1; i < n->req(); i++ ) { duke@435: Node *m = n->in(i); duke@435: if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) { duke@435: // We allow the special case of AddP's with no local inputs. duke@435: // This allows us to split-up address expressions. duke@435: if (m->is_AddP() && duke@435: get_ctrl(m->in(2)) != n_ctrl && duke@435: get_ctrl(m->in(3)) != n_ctrl) { duke@435: // Move the AddP up to dominating point duke@435: set_ctrl_and_loop(m, find_non_split_ctrl(idom(n_ctrl))); duke@435: continue; duke@435: } duke@435: return NULL; duke@435: } duke@435: } duke@435: duke@435: return n_ctrl; duke@435: } duke@435: duke@435: //------------------------------remix_address_expressions---------------------- duke@435: // Rework addressing expressions to get the most loop-invariant stuff duke@435: // moved out. We'd like to do all associative operators, but it's especially duke@435: // important (common) to do address expressions. duke@435: Node *PhaseIdealLoop::remix_address_expressions( Node *n ) { duke@435: if (!has_ctrl(n)) return NULL; duke@435: Node *n_ctrl = get_ctrl(n); duke@435: IdealLoopTree *n_loop = get_loop(n_ctrl); duke@435: duke@435: // See if 'n' mixes loop-varying and loop-invariant inputs and duke@435: // itself is loop-varying. duke@435: duke@435: // Only interested in binary ops (and AddP) duke@435: if( n->req() < 3 || n->req() > 4 ) return NULL; duke@435: duke@435: Node *n1_ctrl = get_ctrl(n->in( 1)); duke@435: Node *n2_ctrl = get_ctrl(n->in( 2)); duke@435: Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3)); duke@435: IdealLoopTree *n1_loop = get_loop( n1_ctrl ); duke@435: IdealLoopTree *n2_loop = get_loop( n2_ctrl ); duke@435: IdealLoopTree *n3_loop = get_loop( n3_ctrl ); duke@435: duke@435: // Does one of my inputs spin in a tighter loop than self? duke@435: if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) || duke@435: (n_loop->is_member( n2_loop ) && n_loop != n2_loop) || duke@435: (n_loop->is_member( n3_loop ) && n_loop != n3_loop) ) duke@435: return NULL; // Leave well enough alone duke@435: duke@435: // Is at least one of my inputs loop-invariant? duke@435: if( n1_loop == n_loop && duke@435: n2_loop == n_loop && duke@435: n3_loop == n_loop ) duke@435: return NULL; // No loop-invariant inputs duke@435: duke@435: duke@435: int n_op = n->Opcode(); duke@435: duke@435: // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2). duke@435: if( n_op == Op_LShiftI ) { duke@435: // Scale is loop invariant duke@435: Node *scale = n->in(2); duke@435: Node *scale_ctrl = get_ctrl(scale); duke@435: IdealLoopTree *scale_loop = get_loop(scale_ctrl ); duke@435: if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) ) duke@435: return NULL; duke@435: const TypeInt *scale_t = scale->bottom_type()->isa_int(); duke@435: if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 ) duke@435: return NULL; // Dont bother with byte/short masking duke@435: // Add must vary with loop (else shift would be loop-invariant) duke@435: Node *add = n->in(1); duke@435: Node *add_ctrl = get_ctrl(add); duke@435: IdealLoopTree *add_loop = get_loop(add_ctrl); duke@435: //assert( n_loop == add_loop, "" ); duke@435: if( n_loop != add_loop ) return NULL; // happens w/ evil ZKM loops duke@435: duke@435: // Convert I-V into I+ (0-V); same for V-I duke@435: if( add->Opcode() == Op_SubI && duke@435: _igvn.type( add->in(1) ) != TypeInt::ZERO ) { duke@435: Node *zero = _igvn.intcon(0); duke@435: set_ctrl(zero, C->root()); duke@435: Node *neg = new (C, 3) SubINode( _igvn.intcon(0), add->in(2) ); duke@435: register_new_node( neg, get_ctrl(add->in(2) ) ); duke@435: add = new (C, 3) AddINode( add->in(1), neg ); duke@435: register_new_node( add, add_ctrl ); duke@435: } duke@435: if( add->Opcode() != Op_AddI ) return NULL; duke@435: // See if one add input is loop invariant duke@435: Node *add_var = add->in(1); duke@435: Node *add_var_ctrl = get_ctrl(add_var); duke@435: IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); duke@435: Node *add_invar = add->in(2); duke@435: Node *add_invar_ctrl = get_ctrl(add_invar); duke@435: IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl ); duke@435: if( add_var_loop == n_loop ) { duke@435: } else if( add_invar_loop == n_loop ) { duke@435: // Swap to find the invariant part duke@435: add_invar = add_var; duke@435: add_invar_ctrl = add_var_ctrl; duke@435: add_invar_loop = add_var_loop; duke@435: add_var = add->in(2); duke@435: Node *add_var_ctrl = get_ctrl(add_var); duke@435: IdealLoopTree *add_var_loop = get_loop(add_var_ctrl ); duke@435: } else // Else neither input is loop invariant duke@435: return NULL; duke@435: if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) ) duke@435: return NULL; // No invariant part of the add? duke@435: duke@435: // Yes! Reshape address expression! duke@435: Node *inv_scale = new (C, 3) LShiftINode( add_invar, scale ); duke@435: register_new_node( inv_scale, add_invar_ctrl ); duke@435: Node *var_scale = new (C, 3) LShiftINode( add_var, scale ); duke@435: register_new_node( var_scale, n_ctrl ); duke@435: Node *var_add = new (C, 3) AddINode( var_scale, inv_scale ); duke@435: register_new_node( var_add, n_ctrl ); duke@435: _igvn.hash_delete( n ); duke@435: _igvn.subsume_node( n, var_add ); duke@435: return var_add; duke@435: } duke@435: duke@435: // Replace (I+V) with (V+I) duke@435: if( n_op == Op_AddI || duke@435: n_op == Op_AddL || duke@435: n_op == Op_AddF || duke@435: n_op == Op_AddD || duke@435: n_op == Op_MulI || duke@435: n_op == Op_MulL || duke@435: n_op == Op_MulF || duke@435: n_op == Op_MulD ) { duke@435: if( n2_loop == n_loop ) { duke@435: assert( n1_loop != n_loop, "" ); duke@435: n->swap_edges(1, 2); duke@435: } duke@435: } duke@435: duke@435: // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V), duke@435: // but not if I2 is a constant. duke@435: if( n_op == Op_AddP ) { duke@435: if( n2_loop == n_loop && n3_loop != n_loop ) { duke@435: if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) { duke@435: Node *n22_ctrl = get_ctrl(n->in(2)->in(2)); duke@435: Node *n23_ctrl = get_ctrl(n->in(2)->in(3)); duke@435: IdealLoopTree *n22loop = get_loop( n22_ctrl ); duke@435: IdealLoopTree *n23_loop = get_loop( n23_ctrl ); duke@435: if( n22loop != n_loop && n22loop->is_member(n_loop) && duke@435: n23_loop == n_loop ) { duke@435: Node *add1 = new (C, 4) AddPNode( n->in(1), n->in(2)->in(2), n->in(3) ); duke@435: // Stuff new AddP in the loop preheader duke@435: register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) ); duke@435: Node *add2 = new (C, 4) AddPNode( n->in(1), add1, n->in(2)->in(3) ); duke@435: register_new_node( add2, n_ctrl ); duke@435: _igvn.hash_delete( n ); duke@435: _igvn.subsume_node( n, add2 ); duke@435: return add2; duke@435: } duke@435: } duke@435: } duke@435: duke@435: // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V) duke@435: if( n2_loop != n_loop && n3_loop == n_loop ) { duke@435: if( n->in(3)->Opcode() == Op_AddI ) { duke@435: Node *V = n->in(3)->in(1); duke@435: Node *I = n->in(3)->in(2); duke@435: if( is_member(n_loop,get_ctrl(V)) ) { duke@435: } else { duke@435: Node *tmp = V; V = I; I = tmp; duke@435: } duke@435: if( !is_member(n_loop,get_ctrl(I)) ) { duke@435: Node *add1 = new (C, 4) AddPNode( n->in(1), n->in(2), I ); duke@435: // Stuff new AddP in the loop preheader duke@435: register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) ); duke@435: Node *add2 = new (C, 4) AddPNode( n->in(1), add1, V ); duke@435: register_new_node( add2, n_ctrl ); duke@435: _igvn.hash_delete( n ); duke@435: _igvn.subsume_node( n, add2 ); duke@435: return add2; duke@435: } duke@435: } duke@435: } duke@435: } duke@435: duke@435: return NULL; duke@435: } duke@435: duke@435: //------------------------------conditional_move------------------------------- duke@435: // Attempt to replace a Phi with a conditional move. We have some pretty duke@435: // strict profitability requirements. All Phis at the merge point must duke@435: // be converted, so we can remove the control flow. We need to limit the duke@435: // number of c-moves to a small handful. All code that was in the side-arms duke@435: // of the CFG diamond is now speculatively executed. This code has to be duke@435: // "cheap enough". We are pretty much limited to CFG diamonds that merge duke@435: // 1 or 2 items with a total of 1 or 2 ops executed speculatively. duke@435: Node *PhaseIdealLoop::conditional_move( Node *region ) { duke@435: duke@435: assert( region->is_Region(), "sanity check" ); duke@435: if( region->req() != 3 ) return NULL; duke@435: duke@435: // Check for CFG diamond duke@435: Node *lp = region->in(1); duke@435: Node *rp = region->in(2); duke@435: if( !lp || !rp ) return NULL; duke@435: Node *lp_c = lp->in(0); duke@435: if( lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If() ) return NULL; duke@435: IfNode *iff = lp_c->as_If(); duke@435: duke@435: // Check for highly predictable branch. No point in CMOV'ing if duke@435: // we are going to predict accurately all the time. duke@435: // %%% This hides patterns produced by utility methods like Math.min. duke@435: if( iff->_prob < PROB_UNLIKELY_MAG(3) || duke@435: iff->_prob > PROB_LIKELY_MAG(3) ) duke@435: return NULL; duke@435: duke@435: // Check for ops pinned in an arm of the diamond. duke@435: // Can't remove the control flow in this case duke@435: if( lp->outcnt() > 1 ) return NULL; duke@435: if( rp->outcnt() > 1 ) return NULL; duke@435: duke@435: // Check profitability duke@435: int cost = 0; kvn@470: int phis = 0; duke@435: for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { duke@435: Node *out = region->fast_out(i); duke@435: if( !out->is_Phi() ) continue; // Ignore other control edges, etc kvn@470: phis++; duke@435: PhiNode* phi = out->as_Phi(); duke@435: switch (phi->type()->basic_type()) { duke@435: case T_LONG: duke@435: cost++; // Probably encodes as 2 CMOV's duke@435: case T_INT: // These all CMOV fine duke@435: case T_FLOAT: duke@435: case T_DOUBLE: duke@435: case T_ADDRESS: // (RawPtr) duke@435: cost++; duke@435: break; kvn@728: case T_NARROWOOP: // Fall through duke@435: case T_OBJECT: { // Base oops are OK, but not derived oops kvn@728: const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr(); duke@435: // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a duke@435: // CMOVE'd derived pointer? It's a CMOVE'd derived base. Thus duke@435: // CMOVE'ing a derived pointer requires we also CMOVE the base. If we duke@435: // have a Phi for the base here that we convert to a CMOVE all is well duke@435: // and good. But if the base is dead, we'll not make a CMOVE. Later duke@435: // the allocator will have to produce a base by creating a CMOVE of the duke@435: // relevant bases. This puts the allocator in the business of duke@435: // manufacturing expensive instructions, generally a bad plan. duke@435: // Just Say No to Conditionally-Moved Derived Pointers. duke@435: if( tp && tp->offset() != 0 ) duke@435: return NULL; duke@435: cost++; duke@435: break; duke@435: } duke@435: default: duke@435: return NULL; // In particular, can't do memory or I/O duke@435: } duke@435: // Add in cost any speculative ops duke@435: for( uint j = 1; j < region->req(); j++ ) { duke@435: Node *proj = region->in(j); duke@435: Node *inp = phi->in(j); duke@435: if (get_ctrl(inp) == proj) { // Found local op duke@435: cost++; duke@435: // Check for a chain of dependent ops; these will all become duke@435: // speculative in a CMOV. duke@435: for( uint k = 1; k < inp->req(); k++ ) duke@435: if (get_ctrl(inp->in(k)) == proj) duke@435: return NULL; // Too much speculative goo duke@435: } duke@435: } kvn@728: // See if the Phi is used by a Cmp or Narrow oop Decode/Encode. kvn@728: // This will likely Split-If, a higher-payoff operation. duke@435: for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) { duke@435: Node* use = phi->fast_out(k); kvn@728: if( use->is_Cmp() || use->is_DecodeN() || use->is_EncodeP() ) duke@435: return NULL; duke@435: } duke@435: } duke@435: if( cost >= ConditionalMoveLimit ) return NULL; // Too much goo kvn@470: Node* bol = iff->in(1); kvn@470: assert( bol->Opcode() == Op_Bool, "" ); kvn@470: int cmp_op = bol->in(1)->Opcode(); kvn@470: // It is expensive to generate flags from a float compare. kvn@470: // Avoid duplicated float compare. kvn@470: if( phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL; duke@435: duke@435: // -------------- duke@435: // Now replace all Phis with CMOV's duke@435: Node *cmov_ctrl = iff->in(0); duke@435: uint flip = (lp->Opcode() == Op_IfTrue); duke@435: while( 1 ) { duke@435: PhiNode* phi = NULL; duke@435: for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { duke@435: Node *out = region->fast_out(i); duke@435: if (out->is_Phi()) { duke@435: phi = out->as_Phi(); duke@435: break; duke@435: } duke@435: } duke@435: if (phi == NULL) break; duke@435: #ifndef PRODUCT duke@435: if( PrintOpto && VerifyLoopOptimizations ) tty->print_cr("CMOV"); duke@435: #endif duke@435: // Move speculative ops duke@435: for( uint j = 1; j < region->req(); j++ ) { duke@435: Node *proj = region->in(j); duke@435: Node *inp = phi->in(j); duke@435: if (get_ctrl(inp) == proj) { // Found local op duke@435: #ifndef PRODUCT duke@435: if( PrintOpto && VerifyLoopOptimizations ) { duke@435: tty->print(" speculate: "); duke@435: inp->dump(); duke@435: } duke@435: #endif duke@435: set_ctrl(inp, cmov_ctrl); duke@435: } duke@435: } duke@435: Node *cmov = CMoveNode::make( C, cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi) ); duke@435: register_new_node( cmov, cmov_ctrl ); duke@435: _igvn.hash_delete(phi); duke@435: _igvn.subsume_node( phi, cmov ); duke@435: #ifndef PRODUCT duke@435: if( VerifyLoopOptimizations ) verify(); duke@435: #endif duke@435: } duke@435: duke@435: // The useless CFG diamond will fold up later; see the optimization in duke@435: // RegionNode::Ideal. duke@435: _igvn._worklist.push(region); duke@435: duke@435: return iff->in(1); duke@435: } duke@435: duke@435: //------------------------------split_if_with_blocks_pre----------------------- duke@435: // Do the real work in a non-recursive function. Data nodes want to be duke@435: // cloned in the pre-order so they can feed each other nicely. duke@435: Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) { duke@435: // Cloning these guys is unlikely to win duke@435: int n_op = n->Opcode(); duke@435: if( n_op == Op_MergeMem ) return n; duke@435: if( n->is_Proj() ) return n; duke@435: // Do not clone-up CmpFXXX variations, as these are always duke@435: // followed by a CmpI duke@435: if( n->is_Cmp() ) return n; duke@435: // Attempt to use a conditional move instead of a phi/branch duke@435: if( ConditionalMoveLimit > 0 && n_op == Op_Region ) { duke@435: Node *cmov = conditional_move( n ); duke@435: if( cmov ) return cmov; duke@435: } kvn@688: if( n->is_CFG() || n->is_LoadStore() ) kvn@688: return n; duke@435: if( n_op == Op_Opaque1 || // Opaque nodes cannot be mod'd duke@435: n_op == Op_Opaque2 ) { duke@435: if( !C->major_progress() ) // If chance of no more loop opts... duke@435: _igvn._worklist.push(n); // maybe we'll remove them duke@435: return n; duke@435: } duke@435: duke@435: if( n->is_Con() ) return n; // No cloning for Con nodes duke@435: duke@435: Node *n_ctrl = get_ctrl(n); duke@435: if( !n_ctrl ) return n; // Dead node duke@435: duke@435: // Attempt to remix address expressions for loop invariants duke@435: Node *m = remix_address_expressions( n ); duke@435: if( m ) return m; duke@435: duke@435: // Determine if the Node has inputs from some local Phi. duke@435: // Returns the block to clone thru. duke@435: Node *n_blk = has_local_phi_input( n ); duke@435: if( !n_blk ) return n; duke@435: // Do not clone the trip counter through on a CountedLoop duke@435: // (messes up the canonical shape). duke@435: if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n; duke@435: duke@435: // Check for having no control input; not pinned. Allow duke@435: // dominating control. duke@435: if( n->in(0) ) { duke@435: Node *dom = idom(n_blk); duke@435: if( dom_lca( n->in(0), dom ) != n->in(0) ) duke@435: return n; duke@435: } duke@435: // Policy: when is it profitable. You must get more wins than duke@435: // policy before it is considered profitable. Policy is usually 0, duke@435: // so 1 win is considered profitable. Big merges will require big duke@435: // cloning, so get a larger policy. duke@435: int policy = n_blk->req() >> 2; duke@435: duke@435: // If the loop is a candidate for range check elimination, duke@435: // delay splitting through it's phi until a later loop optimization duke@435: if (n_blk->is_CountedLoop()) { duke@435: IdealLoopTree *lp = get_loop(n_blk); duke@435: if (lp && lp->_rce_candidate) { duke@435: return n; duke@435: } duke@435: } duke@435: duke@435: // Use same limit as split_if_with_blocks_post duke@435: if( C->unique() > 35000 ) return n; // Method too big duke@435: duke@435: // Split 'n' through the merge point if it is profitable duke@435: Node *phi = split_thru_phi( n, n_blk, policy ); duke@435: if( !phi ) return n; duke@435: duke@435: // Found a Phi to split thru! duke@435: // Replace 'n' with the new phi duke@435: _igvn.hash_delete(n); duke@435: _igvn.subsume_node( n, phi ); duke@435: // Moved a load around the loop, 'en-registering' something. duke@435: if( n_blk->Opcode() == Op_Loop && n->is_Load() && duke@435: !phi->in(LoopNode::LoopBackControl)->is_Load() ) duke@435: C->set_major_progress(); duke@435: duke@435: return phi; duke@435: } duke@435: duke@435: static bool merge_point_too_heavy(Compile* C, Node* region) { duke@435: // Bail out if the region and its phis have too many users. duke@435: int weight = 0; duke@435: for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { duke@435: weight += region->fast_out(i)->outcnt(); duke@435: } duke@435: int nodes_left = MaxNodeLimit - C->unique(); duke@435: if (weight * 8 > nodes_left) { duke@435: #ifndef PRODUCT duke@435: if (PrintOpto) duke@435: tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight); duke@435: #endif duke@435: return true; duke@435: } else { duke@435: return false; duke@435: } duke@435: } duke@435: duke@435: #ifdef _LP64 duke@435: static bool merge_point_safe(Node* region) { duke@435: // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode duke@435: // having a PhiNode input. This sidesteps the dangerous case where the split duke@435: // ConvI2LNode may become TOP if the input Value() does not duke@435: // overlap the ConvI2L range, leaving a node which may not dominate its duke@435: // uses. duke@435: // A better fix for this problem can be found in the BugTraq entry, but duke@435: // expediency for Mantis demands this hack. duke@435: for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { duke@435: Node* n = region->fast_out(i); duke@435: if (n->is_Phi()) { duke@435: for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { duke@435: Node* m = n->fast_out(j); duke@435: if (m->Opcode() == Op_ConvI2L) { duke@435: return false; duke@435: } duke@435: } duke@435: } duke@435: } duke@435: return true; duke@435: } duke@435: #endif duke@435: duke@435: duke@435: //------------------------------place_near_use--------------------------------- duke@435: // Place some computation next to use but not inside inner loops. duke@435: // For inner loop uses move it to the preheader area. duke@435: Node *PhaseIdealLoop::place_near_use( Node *useblock ) const { duke@435: IdealLoopTree *u_loop = get_loop( useblock ); duke@435: return (u_loop->_irreducible || u_loop->_child) duke@435: ? useblock duke@435: : u_loop->_head->in(LoopNode::EntryControl); duke@435: } duke@435: duke@435: duke@435: //------------------------------split_if_with_blocks_post---------------------- duke@435: // Do the real work in a non-recursive function. CFG hackery wants to be duke@435: // in the post-order, so it can dirty the I-DOM info and not use the dirtied duke@435: // info. duke@435: void PhaseIdealLoop::split_if_with_blocks_post( Node *n ) { duke@435: duke@435: // Cloning Cmp through Phi's involves the split-if transform. duke@435: // FastLock is not used by an If duke@435: if( n->is_Cmp() && !n->is_FastLock() ) { duke@435: if( C->unique() > 35000 ) return; // Method too big duke@435: duke@435: // Do not do 'split-if' if irreducible loops are present. duke@435: if( _has_irreducible_loops ) duke@435: return; duke@435: duke@435: Node *n_ctrl = get_ctrl(n); duke@435: // Determine if the Node has inputs from some local Phi. duke@435: // Returns the block to clone thru. duke@435: Node *n_blk = has_local_phi_input( n ); duke@435: if( n_blk != n_ctrl ) return; duke@435: duke@435: if( merge_point_too_heavy(C, n_ctrl) ) duke@435: return; duke@435: duke@435: if( n->outcnt() != 1 ) return; // Multiple bool's from 1 compare? duke@435: Node *bol = n->unique_out(); duke@435: assert( bol->is_Bool(), "expect a bool here" ); duke@435: if( bol->outcnt() != 1 ) return;// Multiple branches from 1 compare? duke@435: Node *iff = bol->unique_out(); duke@435: duke@435: // Check some safety conditions duke@435: if( iff->is_If() ) { // Classic split-if? duke@435: if( iff->in(0) != n_ctrl ) return; // Compare must be in same blk as if duke@435: } else if (iff->is_CMove()) { // Trying to split-up a CMOVE duke@435: if( get_ctrl(iff->in(2)) == n_ctrl || duke@435: get_ctrl(iff->in(3)) == n_ctrl ) duke@435: return; // Inputs not yet split-up duke@435: if ( get_loop(n_ctrl) != get_loop(get_ctrl(iff)) ) { duke@435: return; // Loop-invar test gates loop-varying CMOVE duke@435: } duke@435: } else { duke@435: return; // some other kind of node, such as an Allocate duke@435: } duke@435: duke@435: // Do not do 'split-if' if some paths are dead. First do dead code duke@435: // elimination and then see if its still profitable. duke@435: for( uint i = 1; i < n_ctrl->req(); i++ ) duke@435: if( n_ctrl->in(i) == C->top() ) duke@435: return; duke@435: duke@435: // When is split-if profitable? Every 'win' on means some control flow duke@435: // goes dead, so it's almost always a win. duke@435: int policy = 0; duke@435: // If trying to do a 'Split-If' at the loop head, it is only duke@435: // profitable if the cmp folds up on BOTH paths. Otherwise we duke@435: // risk peeling a loop forever. duke@435: duke@435: // CNC - Disabled for now. Requires careful handling of loop duke@435: // body selection for the cloned code. Also, make sure we check duke@435: // for any input path not being in the same loop as n_ctrl. For duke@435: // irreducible loops we cannot check for 'n_ctrl->is_Loop()' duke@435: // because the alternative loop entry points won't be converted duke@435: // into LoopNodes. duke@435: IdealLoopTree *n_loop = get_loop(n_ctrl); duke@435: for( uint j = 1; j < n_ctrl->req(); j++ ) duke@435: if( get_loop(n_ctrl->in(j)) != n_loop ) duke@435: return; duke@435: duke@435: #ifdef _LP64 duke@435: // Check for safety of the merge point. duke@435: if( !merge_point_safe(n_ctrl) ) { duke@435: return; duke@435: } duke@435: #endif duke@435: duke@435: // Split compare 'n' through the merge point if it is profitable duke@435: Node *phi = split_thru_phi( n, n_ctrl, policy ); duke@435: if( !phi ) return; duke@435: duke@435: // Found a Phi to split thru! duke@435: // Replace 'n' with the new phi duke@435: _igvn.hash_delete(n); duke@435: _igvn.subsume_node( n, phi ); duke@435: duke@435: // Now split the bool up thru the phi duke@435: Node *bolphi = split_thru_phi( bol, n_ctrl, -1 ); duke@435: _igvn.hash_delete(bol); duke@435: _igvn.subsume_node( bol, bolphi ); duke@435: assert( iff->in(1) == bolphi, "" ); duke@435: if( bolphi->Value(&_igvn)->singleton() ) duke@435: return; duke@435: duke@435: // Conditional-move? Must split up now duke@435: if( !iff->is_If() ) { duke@435: Node *cmovphi = split_thru_phi( iff, n_ctrl, -1 ); duke@435: _igvn.hash_delete(iff); duke@435: _igvn.subsume_node( iff, cmovphi ); duke@435: return; duke@435: } duke@435: duke@435: // Now split the IF duke@435: do_split_if( iff ); duke@435: return; duke@435: } duke@435: duke@435: // Check for an IF ready to split; one that has its duke@435: // condition codes input coming from a Phi at the block start. duke@435: int n_op = n->Opcode(); duke@435: duke@435: // Check for an IF being dominated by another IF same test duke@435: if( n_op == Op_If ) { duke@435: Node *bol = n->in(1); duke@435: uint max = bol->outcnt(); duke@435: // Check for same test used more than once? duke@435: if( n_op == Op_If && max > 1 && bol->is_Bool() ) { duke@435: // Search up IDOMs to see if this IF is dominated. duke@435: Node *cutoff = get_ctrl(bol); duke@435: duke@435: // Now search up IDOMs till cutoff, looking for a dominating test duke@435: Node *prevdom = n; duke@435: Node *dom = idom(prevdom); duke@435: while( dom != cutoff ) { duke@435: if( dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom ) { duke@435: // Replace the dominated test with an obvious true or false. duke@435: // Place it on the IGVN worklist for later cleanup. duke@435: C->set_major_progress(); duke@435: dominated_by( prevdom, n ); duke@435: #ifndef PRODUCT duke@435: if( VerifyLoopOptimizations ) verify(); duke@435: #endif duke@435: return; duke@435: } duke@435: prevdom = dom; duke@435: dom = idom(prevdom); duke@435: } duke@435: } duke@435: } duke@435: duke@435: // See if a shared loop-varying computation has no loop-varying uses. duke@435: // Happens if something is only used for JVM state in uncommon trap exits, duke@435: // like various versions of induction variable+offset. Clone the duke@435: // computation per usage to allow it to sink out of the loop. duke@435: if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about) duke@435: Node *n_ctrl = get_ctrl(n); duke@435: IdealLoopTree *n_loop = get_loop(n_ctrl); duke@435: if( n_loop != _ltree_root ) { duke@435: DUIterator_Fast imax, i = n->fast_outs(imax); duke@435: for (; i < imax; i++) { duke@435: Node* u = n->fast_out(i); duke@435: if( !has_ctrl(u) ) break; // Found control user duke@435: IdealLoopTree *u_loop = get_loop(get_ctrl(u)); duke@435: if( u_loop == n_loop ) break; // Found loop-varying use duke@435: if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop duke@435: if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003 duke@435: } duke@435: bool did_break = (i < imax); // Did we break out of the previous loop? duke@435: if (!did_break && n->outcnt() > 1) { // All uses in outer loops! duke@435: Node *late_load_ctrl; duke@435: if (n->is_Load()) { duke@435: // If n is a load, get and save the result from get_late_ctrl(), duke@435: // to be later used in calculating the control for n's clones. duke@435: clear_dom_lca_tags(); duke@435: late_load_ctrl = get_late_ctrl(n, n_ctrl); duke@435: } duke@435: // If n is a load, and the late control is the same as the current duke@435: // control, then the cloning of n is a pointless exercise, because duke@435: // GVN will ensure that we end up where we started. duke@435: if (!n->is_Load() || late_load_ctrl != n_ctrl) { duke@435: for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) { duke@435: Node *u = n->last_out(j); // Clone private computation per use duke@435: _igvn.hash_delete(u); duke@435: _igvn._worklist.push(u); duke@435: Node *x = n->clone(); // Clone computation duke@435: Node *x_ctrl = NULL; duke@435: if( u->is_Phi() ) { duke@435: // Replace all uses of normal nodes. Replace Phi uses twisti@1040: // individually, so the separate Nodes can sink down duke@435: // different paths. duke@435: uint k = 1; duke@435: while( u->in(k) != n ) k++; duke@435: u->set_req( k, x ); duke@435: // x goes next to Phi input path duke@435: x_ctrl = u->in(0)->in(k); duke@435: --j; duke@435: } else { // Normal use duke@435: // Replace all uses duke@435: for( uint k = 0; k < u->req(); k++ ) { duke@435: if( u->in(k) == n ) { duke@435: u->set_req( k, x ); duke@435: --j; duke@435: } duke@435: } duke@435: x_ctrl = get_ctrl(u); duke@435: } duke@435: duke@435: // Find control for 'x' next to use but not inside inner loops. duke@435: // For inner loop uses get the preheader area. duke@435: x_ctrl = place_near_use(x_ctrl); duke@435: duke@435: if (n->is_Load()) { duke@435: // For loads, add a control edge to a CFG node outside of the loop duke@435: // to force them to not combine and return back inside the loop duke@435: // during GVN optimization (4641526). duke@435: // duke@435: // Because we are setting the actual control input, factor in duke@435: // the result from get_late_ctrl() so we respect any duke@435: // anti-dependences. (6233005). duke@435: x_ctrl = dom_lca(late_load_ctrl, x_ctrl); duke@435: duke@435: // Don't allow the control input to be a CFG splitting node. duke@435: // Such nodes should only have ProjNodes as outs, e.g. IfNode duke@435: // should only have IfTrueNode and IfFalseNode (4985384). duke@435: x_ctrl = find_non_split_ctrl(x_ctrl); duke@435: assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone"); duke@435: duke@435: x->set_req(0, x_ctrl); duke@435: } duke@435: register_new_node(x, x_ctrl); duke@435: duke@435: // Some institutional knowledge is needed here: 'x' is duke@435: // yanked because if the optimizer runs GVN on it all the duke@435: // cloned x's will common up and undo this optimization and duke@435: // be forced back in the loop. This is annoying because it duke@435: // makes +VerifyOpto report false-positives on progress. I duke@435: // tried setting control edges on the x's to force them to duke@435: // not combine, but the matching gets worried when it tries duke@435: // to fold a StoreP and an AddP together (as part of an duke@435: // address expression) and the AddP and StoreP have duke@435: // different controls. never@753: if( !x->is_Load() && !x->is_DecodeN() ) _igvn._worklist.yank(x); duke@435: } duke@435: _igvn.remove_dead_node(n); duke@435: } duke@435: } duke@435: } duke@435: } duke@435: duke@435: // Check for Opaque2's who's loop has disappeared - who's input is in the duke@435: // same loop nest as their output. Remove 'em, they are no longer useful. duke@435: if( n_op == Op_Opaque2 && duke@435: n->in(1) != NULL && duke@435: get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) { duke@435: _igvn.add_users_to_worklist(n); duke@435: _igvn.hash_delete(n); duke@435: _igvn.subsume_node( n, n->in(1) ); duke@435: } duke@435: } duke@435: duke@435: //------------------------------split_if_with_blocks--------------------------- duke@435: // Check for aggressive application of 'split-if' optimization, duke@435: // using basic block level info. duke@435: void PhaseIdealLoop::split_if_with_blocks( VectorSet &visited, Node_Stack &nstack ) { duke@435: Node *n = C->root(); duke@435: visited.set(n->_idx); // first, mark node as visited duke@435: // Do pre-visit work for root duke@435: n = split_if_with_blocks_pre( n ); duke@435: uint cnt = n->outcnt(); duke@435: uint i = 0; duke@435: while (true) { duke@435: // Visit all children duke@435: if (i < cnt) { duke@435: Node* use = n->raw_out(i); duke@435: ++i; duke@435: if (use->outcnt() != 0 && !visited.test_set(use->_idx)) { duke@435: // Now do pre-visit work for this use duke@435: use = split_if_with_blocks_pre( use ); duke@435: nstack.push(n, i); // Save parent and next use's index. duke@435: n = use; // Process all children of current use. duke@435: cnt = use->outcnt(); duke@435: i = 0; duke@435: } duke@435: } duke@435: else { duke@435: // All of n's children have been processed, complete post-processing. duke@435: if (cnt != 0 && !n->is_Con()) { duke@435: assert(has_node(n), "no dead nodes"); duke@435: split_if_with_blocks_post( n ); duke@435: } duke@435: if (nstack.is_empty()) { duke@435: // Finished all nodes on stack. duke@435: break; duke@435: } duke@435: // Get saved parent node and next use's index. Visit the rest of uses. duke@435: n = nstack.node(); duke@435: cnt = n->outcnt(); duke@435: i = nstack.index(); duke@435: nstack.pop(); duke@435: } duke@435: } duke@435: } duke@435: duke@435: duke@435: //============================================================================= duke@435: // duke@435: // C L O N E A L O O P B O D Y duke@435: // duke@435: duke@435: //------------------------------clone_iff-------------------------------------- duke@435: // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. duke@435: // "Nearly" because all Nodes have been cloned from the original in the loop, duke@435: // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs duke@435: // through the Phi recursively, and return a Bool. duke@435: BoolNode *PhaseIdealLoop::clone_iff( PhiNode *phi, IdealLoopTree *loop ) { duke@435: duke@435: // Convert this Phi into a Phi merging Bools duke@435: uint i; duke@435: for( i = 1; i < phi->req(); i++ ) { duke@435: Node *b = phi->in(i); duke@435: if( b->is_Phi() ) { duke@435: _igvn.hash_delete(phi); duke@435: _igvn._worklist.push(phi); duke@435: phi->set_req(i, clone_iff( b->as_Phi(), loop )); duke@435: } else { duke@435: assert( b->is_Bool(), "" ); duke@435: } duke@435: } duke@435: duke@435: Node *sample_bool = phi->in(1); duke@435: Node *sample_cmp = sample_bool->in(1); duke@435: duke@435: // Make Phis to merge the Cmp's inputs. duke@435: int size = phi->in(0)->req(); duke@435: PhiNode *phi1 = new (C, size) PhiNode( phi->in(0), Type::TOP ); duke@435: PhiNode *phi2 = new (C, size) PhiNode( phi->in(0), Type::TOP ); duke@435: for( i = 1; i < phi->req(); i++ ) { duke@435: Node *n1 = phi->in(i)->in(1)->in(1); duke@435: Node *n2 = phi->in(i)->in(1)->in(2); duke@435: phi1->set_req( i, n1 ); duke@435: phi2->set_req( i, n2 ); duke@435: phi1->set_type( phi1->type()->meet(n1->bottom_type()) ); duke@435: phi2->set_type( phi2->type()->meet(n2->bottom_type()) ); duke@435: } duke@435: // See if these Phis have been made before. duke@435: // Register with optimizer duke@435: Node *hit1 = _igvn.hash_find_insert(phi1); duke@435: if( hit1 ) { // Hit, toss just made Phi duke@435: _igvn.remove_dead_node(phi1); // Remove new phi duke@435: assert( hit1->is_Phi(), "" ); duke@435: phi1 = (PhiNode*)hit1; // Use existing phi duke@435: } else { // Miss duke@435: _igvn.register_new_node_with_optimizer(phi1); duke@435: } duke@435: Node *hit2 = _igvn.hash_find_insert(phi2); duke@435: if( hit2 ) { // Hit, toss just made Phi duke@435: _igvn.remove_dead_node(phi2); // Remove new phi duke@435: assert( hit2->is_Phi(), "" ); duke@435: phi2 = (PhiNode*)hit2; // Use existing phi duke@435: } else { // Miss duke@435: _igvn.register_new_node_with_optimizer(phi2); duke@435: } duke@435: // Register Phis with loop/block info duke@435: set_ctrl(phi1, phi->in(0)); duke@435: set_ctrl(phi2, phi->in(0)); duke@435: // Make a new Cmp duke@435: Node *cmp = sample_cmp->clone(); duke@435: cmp->set_req( 1, phi1 ); duke@435: cmp->set_req( 2, phi2 ); duke@435: _igvn.register_new_node_with_optimizer(cmp); duke@435: set_ctrl(cmp, phi->in(0)); duke@435: duke@435: // Make a new Bool duke@435: Node *b = sample_bool->clone(); duke@435: b->set_req(1,cmp); duke@435: _igvn.register_new_node_with_optimizer(b); duke@435: set_ctrl(b, phi->in(0)); duke@435: duke@435: assert( b->is_Bool(), "" ); duke@435: return (BoolNode*)b; duke@435: } duke@435: duke@435: //------------------------------clone_bool------------------------------------- duke@435: // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. duke@435: // "Nearly" because all Nodes have been cloned from the original in the loop, duke@435: // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs duke@435: // through the Phi recursively, and return a Bool. duke@435: CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) { duke@435: uint i; duke@435: // Convert this Phi into a Phi merging Bools duke@435: for( i = 1; i < phi->req(); i++ ) { duke@435: Node *b = phi->in(i); duke@435: if( b->is_Phi() ) { duke@435: _igvn.hash_delete(phi); duke@435: _igvn._worklist.push(phi); duke@435: phi->set_req(i, clone_bool( b->as_Phi(), loop )); duke@435: } else { duke@435: assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" ); duke@435: } duke@435: } duke@435: duke@435: Node *sample_cmp = phi->in(1); duke@435: duke@435: // Make Phis to merge the Cmp's inputs. duke@435: int size = phi->in(0)->req(); duke@435: PhiNode *phi1 = new (C, size) PhiNode( phi->in(0), Type::TOP ); duke@435: PhiNode *phi2 = new (C, size) PhiNode( phi->in(0), Type::TOP ); duke@435: for( uint j = 1; j < phi->req(); j++ ) { duke@435: Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP duke@435: Node *n1, *n2; duke@435: if( cmp_top->is_Cmp() ) { duke@435: n1 = cmp_top->in(1); duke@435: n2 = cmp_top->in(2); duke@435: } else { duke@435: n1 = n2 = cmp_top; duke@435: } duke@435: phi1->set_req( j, n1 ); duke@435: phi2->set_req( j, n2 ); duke@435: phi1->set_type( phi1->type()->meet(n1->bottom_type()) ); duke@435: phi2->set_type( phi2->type()->meet(n2->bottom_type()) ); duke@435: } duke@435: duke@435: // See if these Phis have been made before. duke@435: // Register with optimizer duke@435: Node *hit1 = _igvn.hash_find_insert(phi1); duke@435: if( hit1 ) { // Hit, toss just made Phi duke@435: _igvn.remove_dead_node(phi1); // Remove new phi duke@435: assert( hit1->is_Phi(), "" ); duke@435: phi1 = (PhiNode*)hit1; // Use existing phi duke@435: } else { // Miss duke@435: _igvn.register_new_node_with_optimizer(phi1); duke@435: } duke@435: Node *hit2 = _igvn.hash_find_insert(phi2); duke@435: if( hit2 ) { // Hit, toss just made Phi duke@435: _igvn.remove_dead_node(phi2); // Remove new phi duke@435: assert( hit2->is_Phi(), "" ); duke@435: phi2 = (PhiNode*)hit2; // Use existing phi duke@435: } else { // Miss duke@435: _igvn.register_new_node_with_optimizer(phi2); duke@435: } duke@435: // Register Phis with loop/block info duke@435: set_ctrl(phi1, phi->in(0)); duke@435: set_ctrl(phi2, phi->in(0)); duke@435: // Make a new Cmp duke@435: Node *cmp = sample_cmp->clone(); duke@435: cmp->set_req( 1, phi1 ); duke@435: cmp->set_req( 2, phi2 ); duke@435: _igvn.register_new_node_with_optimizer(cmp); duke@435: set_ctrl(cmp, phi->in(0)); duke@435: duke@435: assert( cmp->is_Cmp(), "" ); duke@435: return (CmpNode*)cmp; duke@435: } duke@435: duke@435: //------------------------------sink_use--------------------------------------- duke@435: // If 'use' was in the loop-exit block, it now needs to be sunk duke@435: // below the post-loop merge point. duke@435: void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) { duke@435: if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) { duke@435: set_ctrl(use, post_loop); duke@435: for (DUIterator j = use->outs(); use->has_out(j); j++) duke@435: sink_use(use->out(j), post_loop); duke@435: } duke@435: } duke@435: duke@435: //------------------------------clone_loop------------------------------------- duke@435: // duke@435: // C L O N E A L O O P B O D Y duke@435: // duke@435: // This is the basic building block of the loop optimizations. It clones an duke@435: // entire loop body. It makes an old_new loop body mapping; with this mapping duke@435: // you can find the new-loop equivalent to an old-loop node. All new-loop duke@435: // nodes are exactly equal to their old-loop counterparts, all edges are the duke@435: // same. All exits from the old-loop now have a RegionNode that merges the duke@435: // equivalent new-loop path. This is true even for the normal "loop-exit" duke@435: // condition. All uses of loop-invariant old-loop values now come from (one duke@435: // or more) Phis that merge their new-loop equivalents. duke@435: // duke@435: // This operation leaves the graph in an illegal state: there are two valid duke@435: // control edges coming from the loop pre-header to both loop bodies. I'll duke@435: // definitely have to hack the graph after running this transform. duke@435: // duke@435: // From this building block I will further edit edges to perform loop peeling duke@435: // or loop unrolling or iteration splitting (Range-Check-Elimination), etc. duke@435: // duke@435: // Parameter side_by_size_idom: duke@435: // When side_by_size_idom is NULL, the dominator tree is constructed for duke@435: // the clone loop to dominate the original. Used in construction of duke@435: // pre-main-post loop sequence. duke@435: // When nonnull, the clone and original are side-by-side, both are duke@435: // dominated by the side_by_side_idom node. Used in construction of duke@435: // unswitched loops. duke@435: void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd, duke@435: Node* side_by_side_idom) { duke@435: duke@435: // Step 1: Clone the loop body. Make the old->new mapping. duke@435: uint i; duke@435: for( i = 0; i < loop->_body.size(); i++ ) { duke@435: Node *old = loop->_body.at(i); duke@435: Node *nnn = old->clone(); duke@435: old_new.map( old->_idx, nnn ); duke@435: _igvn.register_new_node_with_optimizer(nnn); duke@435: } duke@435: duke@435: duke@435: // Step 2: Fix the edges in the new body. If the old input is outside the duke@435: // loop use it. If the old input is INside the loop, use the corresponding duke@435: // new node instead. duke@435: for( i = 0; i < loop->_body.size(); i++ ) { duke@435: Node *old = loop->_body.at(i); duke@435: Node *nnn = old_new[old->_idx]; duke@435: // Fix CFG/Loop controlling the new node duke@435: if (has_ctrl(old)) { duke@435: set_ctrl(nnn, old_new[get_ctrl(old)->_idx]); duke@435: } else { duke@435: set_loop(nnn, loop->_parent); duke@435: if (old->outcnt() > 0) { duke@435: set_idom( nnn, old_new[idom(old)->_idx], dd ); duke@435: } duke@435: } duke@435: // Correct edges to the new node duke@435: for( uint j = 0; j < nnn->req(); j++ ) { duke@435: Node *n = nnn->in(j); duke@435: if( n ) { duke@435: IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n ); duke@435: if( loop->is_member( old_in_loop ) ) duke@435: nnn->set_req(j, old_new[n->_idx]); duke@435: } duke@435: } duke@435: _igvn.hash_find_insert(nnn); duke@435: } duke@435: Node *newhead = old_new[loop->_head->_idx]; duke@435: set_idom(newhead, newhead->in(LoopNode::EntryControl), dd); duke@435: duke@435: duke@435: // Step 3: Now fix control uses. Loop varying control uses have already duke@435: // been fixed up (as part of all input edges in Step 2). Loop invariant duke@435: // control uses must be either an IfFalse or an IfTrue. Make a merge duke@435: // point to merge the old and new IfFalse/IfTrue nodes; make the use duke@435: // refer to this. duke@435: ResourceArea *area = Thread::current()->resource_area(); duke@435: Node_List worklist(area); duke@435: uint new_counter = C->unique(); duke@435: for( i = 0; i < loop->_body.size(); i++ ) { duke@435: Node* old = loop->_body.at(i); duke@435: if( !old->is_CFG() ) continue; duke@435: Node* nnn = old_new[old->_idx]; duke@435: duke@435: // Copy uses to a worklist, so I can munge the def-use info duke@435: // with impunity. duke@435: for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) duke@435: worklist.push(old->fast_out(j)); duke@435: duke@435: while( worklist.size() ) { // Visit all uses duke@435: Node *use = worklist.pop(); duke@435: if (!has_node(use)) continue; // Ignore dead nodes duke@435: IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); duke@435: if( !loop->is_member( use_loop ) && use->is_CFG() ) { duke@435: // Both OLD and USE are CFG nodes here. duke@435: assert( use->is_Proj(), "" ); duke@435: duke@435: // Clone the loop exit control projection duke@435: Node *newuse = use->clone(); duke@435: newuse->set_req(0,nnn); duke@435: _igvn.register_new_node_with_optimizer(newuse); duke@435: set_loop(newuse, use_loop); duke@435: set_idom(newuse, nnn, dom_depth(nnn) + 1 ); duke@435: duke@435: // We need a Region to merge the exit from the peeled body and the duke@435: // exit from the old loop body. duke@435: RegionNode *r = new (C, 3) RegionNode(3); duke@435: // Map the old use to the new merge point duke@435: old_new.map( use->_idx, r ); duke@435: uint dd_r = MIN2(dom_depth(newuse),dom_depth(use)); duke@435: assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" ); duke@435: duke@435: // The original user of 'use' uses 'r' instead. duke@435: for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) { duke@435: Node* useuse = use->last_out(l); duke@435: _igvn.hash_delete(useuse); duke@435: _igvn._worklist.push(useuse); duke@435: uint uses_found = 0; duke@435: if( useuse->in(0) == use ) { duke@435: useuse->set_req(0, r); duke@435: uses_found++; duke@435: if( useuse->is_CFG() ) { duke@435: assert( dom_depth(useuse) > dd_r, "" ); duke@435: set_idom(useuse, r, dom_depth(useuse)); duke@435: } duke@435: } duke@435: for( uint k = 1; k < useuse->req(); k++ ) { duke@435: if( useuse->in(k) == use ) { duke@435: useuse->set_req(k, r); duke@435: uses_found++; duke@435: } duke@435: } duke@435: l -= uses_found; // we deleted 1 or more copies of this edge duke@435: } duke@435: duke@435: // Now finish up 'r' duke@435: r->set_req( 1, newuse ); duke@435: r->set_req( 2, use ); duke@435: _igvn.register_new_node_with_optimizer(r); duke@435: set_loop(r, use_loop); duke@435: set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r); duke@435: } // End of if a loop-exit test duke@435: } duke@435: } duke@435: duke@435: // Step 4: If loop-invariant use is not control, it must be dominated by a duke@435: // loop exit IfFalse/IfTrue. Find "proper" loop exit. Make a Region duke@435: // there if needed. Make a Phi there merging old and new used values. duke@435: Node_List *split_if_set = NULL; duke@435: Node_List *split_bool_set = NULL; duke@435: Node_List *split_cex_set = NULL; duke@435: for( i = 0; i < loop->_body.size(); i++ ) { duke@435: Node* old = loop->_body.at(i); duke@435: Node* nnn = old_new[old->_idx]; duke@435: // Copy uses to a worklist, so I can munge the def-use info duke@435: // with impunity. duke@435: for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++) duke@435: worklist.push(old->fast_out(j)); duke@435: duke@435: while( worklist.size() ) { duke@435: Node *use = worklist.pop(); duke@435: if (!has_node(use)) continue; // Ignore dead nodes duke@435: if (use->in(0) == C->top()) continue; duke@435: IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use ); duke@435: // Check for data-use outside of loop - at least one of OLD or USE duke@435: // must not be a CFG node. duke@435: if( !loop->is_member( use_loop ) && (!old->is_CFG() || !use->is_CFG())) { duke@435: duke@435: // If the Data use is an IF, that means we have an IF outside of the duke@435: // loop that is switching on a condition that is set inside of the duke@435: // loop. Happens if people set a loop-exit flag; then test the flag duke@435: // in the loop to break the loop, then test is again outside of the duke@435: // loop to determine which way the loop exited. duke@435: if( use->is_If() || use->is_CMove() ) { duke@435: // Since this code is highly unlikely, we lazily build the worklist duke@435: // of such Nodes to go split. duke@435: if( !split_if_set ) duke@435: split_if_set = new Node_List(area); duke@435: split_if_set->push(use); duke@435: } duke@435: if( use->is_Bool() ) { duke@435: if( !split_bool_set ) duke@435: split_bool_set = new Node_List(area); duke@435: split_bool_set->push(use); duke@435: } duke@435: if( use->Opcode() == Op_CreateEx ) { duke@435: if( !split_cex_set ) duke@435: split_cex_set = new Node_List(area); duke@435: split_cex_set->push(use); duke@435: } duke@435: duke@435: duke@435: // Get "block" use is in duke@435: uint idx = 0; duke@435: while( use->in(idx) != old ) idx++; duke@435: Node *prev = use->is_CFG() ? use : get_ctrl(use); duke@435: assert( !loop->is_member( get_loop( prev ) ), "" ); duke@435: Node *cfg = prev->_idx >= new_counter duke@435: ? prev->in(2) duke@435: : idom(prev); duke@435: if( use->is_Phi() ) // Phi use is in prior block duke@435: cfg = prev->in(idx); // NOT in block of Phi itself duke@435: if (cfg->is_top()) { // Use is dead? duke@435: _igvn.hash_delete(use); duke@435: _igvn._worklist.push(use); duke@435: use->set_req(idx, C->top()); duke@435: continue; duke@435: } duke@435: duke@435: while( !loop->is_member( get_loop( cfg ) ) ) { duke@435: prev = cfg; duke@435: cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg); duke@435: } duke@435: // If the use occurs after merging several exits from the loop, then duke@435: // old value must have dominated all those exits. Since the same old duke@435: // value was used on all those exits we did not need a Phi at this duke@435: // merge point. NOW we do need a Phi here. Each loop exit value duke@435: // is now merged with the peeled body exit; each exit gets its own duke@435: // private Phi and those Phis need to be merged here. duke@435: Node *phi; duke@435: if( prev->is_Region() ) { duke@435: if( idx == 0 ) { // Updating control edge? duke@435: phi = prev; // Just use existing control duke@435: } else { // Else need a new Phi duke@435: phi = PhiNode::make( prev, old ); duke@435: // Now recursively fix up the new uses of old! duke@435: for( uint i = 1; i < prev->req(); i++ ) { duke@435: worklist.push(phi); // Onto worklist once for each 'old' input duke@435: } duke@435: } duke@435: } else { duke@435: // Get new RegionNode merging old and new loop exits duke@435: prev = old_new[prev->_idx]; duke@435: assert( prev, "just made this in step 7" ); duke@435: if( idx == 0 ) { // Updating control edge? duke@435: phi = prev; // Just use existing control duke@435: } else { // Else need a new Phi duke@435: // Make a new Phi merging data values properly duke@435: phi = PhiNode::make( prev, old ); duke@435: phi->set_req( 1, nnn ); duke@435: } duke@435: } duke@435: // If inserting a new Phi, check for prior hits duke@435: if( idx != 0 ) { duke@435: Node *hit = _igvn.hash_find_insert(phi); duke@435: if( hit == NULL ) { duke@435: _igvn.register_new_node_with_optimizer(phi); // Register new phi duke@435: } else { // or duke@435: // Remove the new phi from the graph and use the hit duke@435: _igvn.remove_dead_node(phi); duke@435: phi = hit; // Use existing phi duke@435: } duke@435: set_ctrl(phi, prev); duke@435: } duke@435: // Make 'use' use the Phi instead of the old loop body exit value duke@435: _igvn.hash_delete(use); duke@435: _igvn._worklist.push(use); duke@435: use->set_req(idx, phi); duke@435: if( use->_idx >= new_counter ) { // If updating new phis duke@435: // Not needed for correctness, but prevents a weak assert duke@435: // in AddPNode from tripping (when we end up with different duke@435: // base & derived Phis that will become the same after duke@435: // IGVN does CSE). duke@435: Node *hit = _igvn.hash_find_insert(use); duke@435: if( hit ) // Go ahead and re-hash for hits. duke@435: _igvn.subsume_node( use, hit ); duke@435: } duke@435: duke@435: // If 'use' was in the loop-exit block, it now needs to be sunk duke@435: // below the post-loop merge point. duke@435: sink_use( use, prev ); duke@435: } duke@435: } duke@435: } duke@435: duke@435: // Check for IFs that need splitting/cloning. Happens if an IF outside of duke@435: // the loop uses a condition set in the loop. The original IF probably duke@435: // takes control from one or more OLD Regions (which in turn get from NEW duke@435: // Regions). In any case, there will be a set of Phis for each merge point duke@435: // from the IF up to where the original BOOL def exists the loop. duke@435: if( split_if_set ) { duke@435: while( split_if_set->size() ) { duke@435: Node *iff = split_if_set->pop(); duke@435: if( iff->in(1)->is_Phi() ) { duke@435: BoolNode *b = clone_iff( iff->in(1)->as_Phi(), loop ); duke@435: _igvn.hash_delete(iff); duke@435: _igvn._worklist.push(iff); duke@435: iff->set_req(1, b); duke@435: } duke@435: } duke@435: } duke@435: if( split_bool_set ) { duke@435: while( split_bool_set->size() ) { duke@435: Node *b = split_bool_set->pop(); duke@435: Node *phi = b->in(1); duke@435: assert( phi->is_Phi(), "" ); duke@435: CmpNode *cmp = clone_bool( (PhiNode*)phi, loop ); duke@435: _igvn.hash_delete(b); duke@435: _igvn._worklist.push(b); duke@435: b->set_req(1, cmp); duke@435: } duke@435: } duke@435: if( split_cex_set ) { duke@435: while( split_cex_set->size() ) { duke@435: Node *b = split_cex_set->pop(); duke@435: assert( b->in(0)->is_Region(), "" ); duke@435: assert( b->in(1)->is_Phi(), "" ); duke@435: assert( b->in(0)->in(0) == b->in(1)->in(0), "" ); duke@435: split_up( b, b->in(0), NULL ); duke@435: } duke@435: } duke@435: duke@435: } duke@435: duke@435: duke@435: //---------------------- stride_of_possible_iv ------------------------------------- duke@435: // Looks for an iff/bool/comp with one operand of the compare duke@435: // being a cycle involving an add and a phi, duke@435: // with an optional truncation (left-shift followed by a right-shift) duke@435: // of the add. Returns zero if not an iv. duke@435: int PhaseIdealLoop::stride_of_possible_iv(Node* iff) { duke@435: Node* trunc1 = NULL; duke@435: Node* trunc2 = NULL; duke@435: const TypeInt* ttype = NULL; duke@435: if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) { duke@435: return 0; duke@435: } duke@435: BoolNode* bl = iff->in(1)->as_Bool(); duke@435: Node* cmp = bl->in(1); duke@435: if (!cmp || cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU) { duke@435: return 0; duke@435: } duke@435: // Must have an invariant operand duke@435: if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) { duke@435: return 0; duke@435: } duke@435: Node* add2 = NULL; duke@435: Node* cmp1 = cmp->in(1); duke@435: if (cmp1->is_Phi()) { duke@435: // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) ))) duke@435: Node* phi = cmp1; duke@435: for (uint i = 1; i < phi->req(); i++) { duke@435: Node* in = phi->in(i); duke@435: Node* add = CountedLoopNode::match_incr_with_optional_truncation(in, duke@435: &trunc1, &trunc2, &ttype); duke@435: if (add && add->in(1) == phi) { duke@435: add2 = add->in(2); duke@435: break; duke@435: } duke@435: } duke@435: } else { duke@435: // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) ))) duke@435: Node* addtrunc = cmp1; duke@435: Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc, duke@435: &trunc1, &trunc2, &ttype); duke@435: if (add && add->in(1)->is_Phi()) { duke@435: Node* phi = add->in(1); duke@435: for (uint i = 1; i < phi->req(); i++) { duke@435: if (phi->in(i) == addtrunc) { duke@435: add2 = add->in(2); duke@435: break; duke@435: } duke@435: } duke@435: } duke@435: } duke@435: if (add2 != NULL) { duke@435: const TypeInt* add2t = _igvn.type(add2)->is_int(); duke@435: if (add2t->is_con()) { duke@435: return add2t->get_con(); duke@435: } duke@435: } duke@435: return 0; duke@435: } duke@435: duke@435: duke@435: //---------------------- stay_in_loop ------------------------------------- duke@435: // Return the (unique) control output node that's in the loop (if it exists.) duke@435: Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) { duke@435: Node* unique = NULL; duke@435: if (!n) return NULL; duke@435: for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { duke@435: Node* use = n->fast_out(i); duke@435: if (!has_ctrl(use) && loop->is_member(get_loop(use))) { duke@435: if (unique != NULL) { duke@435: return NULL; duke@435: } duke@435: unique = use; duke@435: } duke@435: } duke@435: return unique; duke@435: } duke@435: duke@435: //------------------------------ register_node ------------------------------------- duke@435: // Utility to register node "n" with PhaseIdealLoop duke@435: void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) { duke@435: _igvn.register_new_node_with_optimizer(n); duke@435: loop->_body.push(n); duke@435: if (n->is_CFG()) { duke@435: set_loop(n, loop); duke@435: set_idom(n, pred, ddepth); duke@435: } else { duke@435: set_ctrl(n, pred); duke@435: } duke@435: } duke@435: duke@435: //------------------------------ proj_clone ------------------------------------- duke@435: // Utility to create an if-projection duke@435: ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) { duke@435: ProjNode* c = p->clone()->as_Proj(); duke@435: c->set_req(0, iff); duke@435: return c; duke@435: } duke@435: duke@435: //------------------------------ short_circuit_if ------------------------------------- duke@435: // Force the iff control output to be the live_proj duke@435: Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) { duke@435: int proj_con = live_proj->_con; duke@435: assert(proj_con == 0 || proj_con == 1, "false or true projection"); duke@435: Node *con = _igvn.intcon(proj_con); duke@435: set_ctrl(con, C->root()); duke@435: if (iff) { duke@435: iff->set_req(1, con); duke@435: } duke@435: return con; duke@435: } duke@435: duke@435: //------------------------------ insert_if_before_proj ------------------------------------- duke@435: // Insert a new if before an if projection (* - new node) duke@435: // duke@435: // before duke@435: // if(test) duke@435: // / \ duke@435: // v v duke@435: // other-proj proj (arg) duke@435: // duke@435: // after duke@435: // if(test) duke@435: // / \ duke@435: // / v duke@435: // | * proj-clone duke@435: // v | duke@435: // other-proj v duke@435: // * new_if(relop(cmp[IU](left,right))) duke@435: // / \ duke@435: // v v duke@435: // * new-proj proj duke@435: // (returned) duke@435: // duke@435: ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) { duke@435: IfNode* iff = proj->in(0)->as_If(); duke@435: IdealLoopTree *loop = get_loop(proj); duke@435: ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); duke@435: int ddepth = dom_depth(proj); duke@435: duke@435: _igvn.hash_delete(iff); duke@435: _igvn._worklist.push(iff); duke@435: _igvn.hash_delete(proj); duke@435: _igvn._worklist.push(proj); duke@435: duke@435: proj->set_req(0, NULL); // temporary disconnect duke@435: ProjNode* proj2 = proj_clone(proj, iff); duke@435: register_node(proj2, loop, iff, ddepth); duke@435: duke@435: Node* cmp = Signed ? (Node*) new (C,3)CmpINode(left, right) : (Node*) new (C,3)CmpUNode(left, right); duke@435: register_node(cmp, loop, proj2, ddepth); duke@435: duke@435: BoolNode* bol = new (C,2)BoolNode(cmp, relop); duke@435: register_node(bol, loop, proj2, ddepth); duke@435: duke@435: IfNode* new_if = new (C,2)IfNode(proj2, bol, iff->_prob, iff->_fcnt); duke@435: register_node(new_if, loop, proj2, ddepth); duke@435: duke@435: proj->set_req(0, new_if); // reattach duke@435: set_idom(proj, new_if, ddepth); duke@435: duke@435: ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj(); duke@435: register_node(new_exit, get_loop(other_proj), new_if, ddepth); duke@435: duke@435: return new_exit; duke@435: } duke@435: duke@435: //------------------------------ insert_region_before_proj ------------------------------------- duke@435: // Insert a region before an if projection (* - new node) duke@435: // duke@435: // before duke@435: // if(test) duke@435: // / | duke@435: // v | duke@435: // proj v duke@435: // other-proj duke@435: // duke@435: // after duke@435: // if(test) duke@435: // / | duke@435: // v | duke@435: // * proj-clone v duke@435: // | other-proj duke@435: // v duke@435: // * new-region duke@435: // | duke@435: // v duke@435: // * dum_if duke@435: // / \ duke@435: // v \ duke@435: // * dum-proj v duke@435: // proj duke@435: // duke@435: RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) { duke@435: IfNode* iff = proj->in(0)->as_If(); duke@435: IdealLoopTree *loop = get_loop(proj); duke@435: ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj(); duke@435: int ddepth = dom_depth(proj); duke@435: duke@435: _igvn.hash_delete(iff); duke@435: _igvn._worklist.push(iff); duke@435: _igvn.hash_delete(proj); duke@435: _igvn._worklist.push(proj); duke@435: duke@435: proj->set_req(0, NULL); // temporary disconnect duke@435: ProjNode* proj2 = proj_clone(proj, iff); duke@435: register_node(proj2, loop, iff, ddepth); duke@435: duke@435: RegionNode* reg = new (C,2)RegionNode(2); duke@435: reg->set_req(1, proj2); duke@435: register_node(reg, loop, iff, ddepth); duke@435: duke@435: IfNode* dum_if = new (C,2)IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt); duke@435: register_node(dum_if, loop, reg, ddepth); duke@435: duke@435: proj->set_req(0, dum_if); // reattach duke@435: set_idom(proj, dum_if, ddepth); duke@435: duke@435: ProjNode* dum_proj = proj_clone(other_proj, dum_if); duke@435: register_node(dum_proj, loop, dum_if, ddepth); duke@435: duke@435: return reg; duke@435: } duke@435: duke@435: //------------------------------ insert_cmpi_loop_exit ------------------------------------- duke@435: // Clone a signed compare loop exit from an unsigned compare and duke@435: // insert it before the unsigned cmp on the stay-in-loop path. duke@435: // All new nodes inserted in the dominator tree between the original duke@435: // if and it's projections. The original if test is replaced with duke@435: // a constant to force the stay-in-loop path. duke@435: // duke@435: // This is done to make sure that the original if and it's projections duke@435: // still dominate the same set of control nodes, that the ctrl() relation duke@435: // from data nodes to them is preserved, and that their loop nesting is duke@435: // preserved. duke@435: // duke@435: // before duke@435: // if(i in(1)->as_Bool(); duke@435: if (bol->_test._test != BoolTest::lt) return NULL; duke@435: CmpNode* cmpu = bol->in(1)->as_Cmp(); duke@435: if (cmpu->Opcode() != Op_CmpU) return NULL; duke@435: int stride = stride_of_possible_iv(if_cmpu); duke@435: if (stride == 0) return NULL; duke@435: duke@435: ProjNode* lp_continue = stay_in_loop(if_cmpu, loop)->as_Proj(); duke@435: ProjNode* lp_exit = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj(); duke@435: duke@435: Node* limit = NULL; duke@435: if (stride > 0) { duke@435: limit = cmpu->in(2); duke@435: } else { duke@435: limit = _igvn.makecon(TypeInt::ZERO); duke@435: set_ctrl(limit, C->root()); duke@435: } duke@435: // Create a new region on the exit path duke@435: RegionNode* reg = insert_region_before_proj(lp_exit); duke@435: duke@435: // Clone the if-cmpu-true-false using a signed compare duke@435: BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge; duke@435: ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue); duke@435: reg->add_req(cmpi_exit); duke@435: duke@435: // Clone the if-cmpu-true-false duke@435: BoolTest::mask rel_u = bol->_test._test; duke@435: ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue); duke@435: reg->add_req(cmpu_exit); duke@435: duke@435: // Force original if to stay in loop. duke@435: short_circuit_if(if_cmpu, lp_continue); duke@435: duke@435: return cmpi_exit->in(0)->as_If(); duke@435: } duke@435: duke@435: //------------------------------ remove_cmpi_loop_exit ------------------------------------- duke@435: // Remove a previously inserted signed compare loop exit. duke@435: void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) { duke@435: Node* lp_proj = stay_in_loop(if_cmp, loop); duke@435: assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI && duke@435: stay_in_loop(lp_proj, loop)->is_If() && duke@435: stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu"); duke@435: Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO); duke@435: set_ctrl(con, C->root()); duke@435: if_cmp->set_req(1, con); duke@435: } duke@435: duke@435: //------------------------------ scheduled_nodelist ------------------------------------- duke@435: // Create a post order schedule of nodes that are in the duke@435: // "member" set. The list is returned in "sched". duke@435: // The first node in "sched" is the loop head, followed by duke@435: // nodes which have no inputs in the "member" set, and then duke@435: // followed by the nodes that have an immediate input dependence duke@435: // on a node in "sched". duke@435: void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) { duke@435: duke@435: assert(member.test(loop->_head->_idx), "loop head must be in member set"); duke@435: Arena *a = Thread::current()->resource_area(); duke@435: VectorSet visited(a); duke@435: Node_Stack nstack(a, loop->_body.size()); duke@435: duke@435: Node* n = loop->_head; // top of stack is cached in "n" duke@435: uint idx = 0; duke@435: visited.set(n->_idx); duke@435: duke@435: // Initially push all with no inputs from within member set duke@435: for(uint i = 0; i < loop->_body.size(); i++ ) { duke@435: Node *elt = loop->_body.at(i); duke@435: if (member.test(elt->_idx)) { duke@435: bool found = false; duke@435: for (uint j = 0; j < elt->req(); j++) { duke@435: Node* def = elt->in(j); duke@435: if (def && member.test(def->_idx) && def != elt) { duke@435: found = true; duke@435: break; duke@435: } duke@435: } duke@435: if (!found && elt != loop->_head) { duke@435: nstack.push(n, idx); duke@435: n = elt; duke@435: assert(!visited.test(n->_idx), "not seen yet"); duke@435: visited.set(n->_idx); duke@435: } duke@435: } duke@435: } duke@435: duke@435: // traverse out's that are in the member set duke@435: while (true) { duke@435: if (idx < n->outcnt()) { duke@435: Node* use = n->raw_out(idx); duke@435: idx++; duke@435: if (!visited.test_set(use->_idx)) { duke@435: if (member.test(use->_idx)) { duke@435: nstack.push(n, idx); duke@435: n = use; duke@435: idx = 0; duke@435: } duke@435: } duke@435: } else { duke@435: // All outputs processed duke@435: sched.push(n); duke@435: if (nstack.is_empty()) break; duke@435: n = nstack.node(); duke@435: idx = nstack.index(); duke@435: nstack.pop(); duke@435: } duke@435: } duke@435: } duke@435: duke@435: duke@435: //------------------------------ has_use_in_set ------------------------------------- duke@435: // Has a use in the vector set duke@435: bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) { duke@435: for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { duke@435: Node* use = n->fast_out(j); duke@435: if (vset.test(use->_idx)) { duke@435: return true; duke@435: } duke@435: } duke@435: return false; duke@435: } duke@435: duke@435: duke@435: //------------------------------ has_use_internal_to_set ------------------------------------- duke@435: // Has use internal to the vector set (ie. not in a phi at the loop head) duke@435: bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) { duke@435: Node* head = loop->_head; duke@435: for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { duke@435: Node* use = n->fast_out(j); duke@435: if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) { duke@435: return true; duke@435: } duke@435: } duke@435: return false; duke@435: } duke@435: duke@435: duke@435: //------------------------------ clone_for_use_outside_loop ------------------------------------- duke@435: // clone "n" for uses that are outside of loop duke@435: void PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) { duke@435: duke@435: assert(worklist.size() == 0, "should be empty"); duke@435: for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { duke@435: Node* use = n->fast_out(j); duke@435: if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) { duke@435: worklist.push(use); duke@435: } duke@435: } duke@435: while( worklist.size() ) { duke@435: Node *use = worklist.pop(); duke@435: if (!has_node(use) || use->in(0) == C->top()) continue; duke@435: uint j; duke@435: for (j = 0; j < use->req(); j++) { duke@435: if (use->in(j) == n) break; duke@435: } duke@435: assert(j < use->req(), "must be there"); duke@435: duke@435: // clone "n" and insert it between the inputs of "n" and the use outside the loop duke@435: Node* n_clone = n->clone(); duke@435: _igvn.hash_delete(use); duke@435: use->set_req(j, n_clone); duke@435: _igvn._worklist.push(use); never@686: Node* use_c; duke@435: if (!use->is_Phi()) { never@686: use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0); duke@435: } else { duke@435: // Use in a phi is considered a use in the associated predecessor block never@686: use_c = use->in(0)->in(j); duke@435: } never@686: set_ctrl(n_clone, use_c); never@686: assert(!loop->is_member(get_loop(use_c)), "should be outside loop"); never@686: get_loop(use_c)->_body.push(n_clone); duke@435: _igvn.register_new_node_with_optimizer(n_clone); duke@435: #if !defined(PRODUCT) duke@435: if (TracePartialPeeling) { duke@435: tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx); duke@435: } duke@435: #endif duke@435: } duke@435: } duke@435: duke@435: duke@435: //------------------------------ clone_for_special_use_inside_loop ------------------------------------- duke@435: // clone "n" for special uses that are in the not_peeled region. duke@435: // If these def-uses occur in separate blocks, the code generator duke@435: // marks the method as not compilable. For example, if a "BoolNode" duke@435: // is in a different basic block than the "IfNode" that uses it, then duke@435: // the compilation is aborted in the code generator. duke@435: void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, duke@435: VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) { duke@435: if (n->is_Phi() || n->is_Load()) { duke@435: return; duke@435: } duke@435: assert(worklist.size() == 0, "should be empty"); duke@435: for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { duke@435: Node* use = n->fast_out(j); duke@435: if ( not_peel.test(use->_idx) && duke@435: (use->is_If() || use->is_CMove() || use->is_Bool()) && duke@435: use->in(1) == n) { duke@435: worklist.push(use); duke@435: } duke@435: } duke@435: if (worklist.size() > 0) { duke@435: // clone "n" and insert it between inputs of "n" and the use duke@435: Node* n_clone = n->clone(); duke@435: loop->_body.push(n_clone); duke@435: _igvn.register_new_node_with_optimizer(n_clone); duke@435: set_ctrl(n_clone, get_ctrl(n)); duke@435: sink_list.push(n_clone); duke@435: not_peel <<= n_clone->_idx; // add n_clone to not_peel set. duke@435: #if !defined(PRODUCT) duke@435: if (TracePartialPeeling) { duke@435: tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx); duke@435: } duke@435: #endif duke@435: while( worklist.size() ) { duke@435: Node *use = worklist.pop(); duke@435: _igvn.hash_delete(use); duke@435: _igvn._worklist.push(use); duke@435: for (uint j = 1; j < use->req(); j++) { duke@435: if (use->in(j) == n) { duke@435: use->set_req(j, n_clone); duke@435: } duke@435: } duke@435: } duke@435: } duke@435: } duke@435: duke@435: duke@435: //------------------------------ insert_phi_for_loop ------------------------------------- duke@435: // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist duke@435: void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) { duke@435: Node *phi = PhiNode::make(lp, back_edge_val); duke@435: phi->set_req(LoopNode::EntryControl, lp_entry_val); duke@435: // Use existing phi if it already exists duke@435: Node *hit = _igvn.hash_find_insert(phi); duke@435: if( hit == NULL ) { duke@435: _igvn.register_new_node_with_optimizer(phi); duke@435: set_ctrl(phi, lp); duke@435: } else { duke@435: // Remove the new phi from the graph and use the hit duke@435: _igvn.remove_dead_node(phi); duke@435: phi = hit; duke@435: } duke@435: _igvn.hash_delete(use); duke@435: _igvn._worklist.push(use); duke@435: use->set_req(idx, phi); duke@435: } duke@435: duke@435: #ifdef ASSERT duke@435: //------------------------------ is_valid_loop_partition ------------------------------------- duke@435: // Validate the loop partition sets: peel and not_peel duke@435: bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, duke@435: VectorSet& not_peel ) { duke@435: uint i; duke@435: // Check that peel_list entries are in the peel set duke@435: for (i = 0; i < peel_list.size(); i++) { duke@435: if (!peel.test(peel_list.at(i)->_idx)) { duke@435: return false; duke@435: } duke@435: } duke@435: // Check at loop members are in one of peel set or not_peel set duke@435: for (i = 0; i < loop->_body.size(); i++ ) { duke@435: Node *def = loop->_body.at(i); duke@435: uint di = def->_idx; duke@435: // Check that peel set elements are in peel_list duke@435: if (peel.test(di)) { duke@435: if (not_peel.test(di)) { duke@435: return false; duke@435: } duke@435: // Must be in peel_list also duke@435: bool found = false; duke@435: for (uint j = 0; j < peel_list.size(); j++) { duke@435: if (peel_list.at(j)->_idx == di) { duke@435: found = true; duke@435: break; duke@435: } duke@435: } duke@435: if (!found) { duke@435: return false; duke@435: } duke@435: } else if (not_peel.test(di)) { duke@435: if (peel.test(di)) { duke@435: return false; duke@435: } duke@435: } else { duke@435: return false; duke@435: } duke@435: } duke@435: return true; duke@435: } duke@435: duke@435: //------------------------------ is_valid_clone_loop_exit_use ------------------------------------- duke@435: // Ensure a use outside of loop is of the right form duke@435: bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) { duke@435: Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; duke@435: return (use->is_Phi() && duke@435: use_c->is_Region() && use_c->req() == 3 && duke@435: (use_c->in(exit_idx)->Opcode() == Op_IfTrue || duke@435: use_c->in(exit_idx)->Opcode() == Op_IfFalse || duke@435: use_c->in(exit_idx)->Opcode() == Op_JumpProj) && duke@435: loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) ); duke@435: } duke@435: duke@435: //------------------------------ is_valid_clone_loop_form ------------------------------------- duke@435: // Ensure that all uses outside of loop are of the right form duke@435: bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, duke@435: uint orig_exit_idx, uint clone_exit_idx) { duke@435: uint len = peel_list.size(); duke@435: for (uint i = 0; i < len; i++) { duke@435: Node *def = peel_list.at(i); duke@435: duke@435: for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { duke@435: Node *use = def->fast_out(j); duke@435: Node *use_c = has_ctrl(use) ? get_ctrl(use) : use; duke@435: if (!loop->is_member(get_loop(use_c))) { duke@435: // use is not in the loop, check for correct structure duke@435: if (use->in(0) == def) { duke@435: // Okay duke@435: } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) { duke@435: return false; duke@435: } duke@435: } duke@435: } duke@435: } duke@435: return true; duke@435: } duke@435: #endif duke@435: duke@435: //------------------------------ partial_peel ------------------------------------- duke@435: // Partially peel (aka loop rotation) the top portion of a loop (called duke@435: // the peel section below) by cloning it and placing one copy just before duke@435: // the new loop head and the other copy at the bottom of the new loop. duke@435: // duke@435: // before after where it came from duke@435: // duke@435: // stmt1 stmt1 duke@435: // loop: stmt2 clone duke@435: // stmt2 if condA goto exitA clone duke@435: // if condA goto exitA new_loop: new duke@435: // stmt3 stmt3 clone duke@435: // if !condB goto loop if condB goto exitB clone duke@435: // exitB: stmt2 orig duke@435: // stmt4 if !condA goto new_loop orig duke@435: // exitA: goto exitA duke@435: // exitB: duke@435: // stmt4 duke@435: // exitA: duke@435: // duke@435: // Step 1: find the cut point: an exit test on probable duke@435: // induction variable. duke@435: // Step 2: schedule (with cloning) operations in the peel duke@435: // section that can be executed after the cut into duke@435: // the section that is not peeled. This may need duke@435: // to clone operations into exit blocks. For duke@435: // instance, a reference to A[i] in the not-peel duke@435: // section and a reference to B[i] in an exit block duke@435: // may cause a left-shift of i by 2 to be placed duke@435: // in the peel block. This step will clone the left duke@435: // shift into the exit block and sink the left shift duke@435: // from the peel to the not-peel section. duke@435: // Step 3: clone the loop, retarget the control, and insert duke@435: // phis for values that are live across the new loop duke@435: // head. This is very dependent on the graph structure duke@435: // from clone_loop. It creates region nodes for duke@435: // exit control and associated phi nodes for values duke@435: // flow out of the loop through that exit. The region duke@435: // node is dominated by the clone's control projection. duke@435: // So the clone's peel section is placed before the duke@435: // new loop head, and the clone's not-peel section is duke@435: // forms the top part of the new loop. The original duke@435: // peel section forms the tail of the new loop. duke@435: // Step 4: update the dominator tree and recompute the duke@435: // dominator depth. duke@435: // duke@435: // orig duke@435: // duke@435: // stmt1 duke@435: // | duke@435: // v duke@435: // loop<----+ duke@435: // | | duke@435: // stmt2 | duke@435: // | | duke@435: // v | duke@435: // ifA | duke@435: // / | | duke@435: // v v | duke@435: // false true ^ <-- last_peel duke@435: // / | | duke@435: // / ===|==cut | duke@435: // / stmt3 | <-- first_not_peel duke@435: // / | | duke@435: // | v | duke@435: // v ifB | duke@435: // exitA: / \ | duke@435: // / \ | duke@435: // v v | duke@435: // false true | duke@435: // / \ | duke@435: // / ----+ duke@435: // | duke@435: // v duke@435: // exitB: duke@435: // stmt4 duke@435: // duke@435: // duke@435: // after clone loop duke@435: // duke@435: // stmt1 duke@435: // / \ duke@435: // clone / \ orig duke@435: // / \ duke@435: // / \ duke@435: // v v duke@435: // +---->loop loop<----+ duke@435: // | | | | duke@435: // | stmt2 stmt2 | duke@435: // | | | | duke@435: // | v v | duke@435: // | ifA ifA | duke@435: // | | \ / | | duke@435: // | v v v v | duke@435: // ^ true false false true ^ <-- last_peel duke@435: // | | ^ \ / | | duke@435: // | cut==|== \ \ / ===|==cut | duke@435: // | stmt3 \ \ / stmt3 | <-- first_not_peel duke@435: // | | dom | | | | duke@435: // | v \ 1v v2 v | duke@435: // | ifB regionA ifB | duke@435: // | / \ | / \ | duke@435: // | / \ v / \ | duke@435: // | v v exitA: v v | duke@435: // | true false false true | duke@435: // | / ^ \ / \ | duke@435: // +---- \ \ / ----+ duke@435: // dom \ / duke@435: // \ 1v v2 duke@435: // regionB duke@435: // | duke@435: // v duke@435: // exitB: duke@435: // stmt4 duke@435: // duke@435: // duke@435: // after partial peel duke@435: // duke@435: // stmt1 duke@435: // / duke@435: // clone / orig duke@435: // / TOP duke@435: // / \ duke@435: // v v duke@435: // TOP->region region----+ duke@435: // | | | duke@435: // stmt2 stmt2 | duke@435: // | | | duke@435: // v v | duke@435: // ifA ifA | duke@435: // | \ / | | duke@435: // v v v v | duke@435: // true false false true | <-- last_peel duke@435: // | ^ \ / +------|---+ duke@435: // +->newloop \ \ / === ==cut | | duke@435: // | stmt3 \ \ / TOP | | duke@435: // | | dom | | stmt3 | | <-- first_not_peel duke@435: // | v \ 1v v2 v | | duke@435: // | ifB regionA ifB ^ v duke@435: // | / \ | / \ | | duke@435: // | / \ v / \ | | duke@435: // | v v exitA: v v | | duke@435: // | true false false true | | duke@435: // | / ^ \ / \ | | duke@435: // | | \ \ / v | | duke@435: // | | dom \ / TOP | | duke@435: // | | \ 1v v2 | | duke@435: // ^ v regionB | | duke@435: // | | | | | duke@435: // | | v ^ v duke@435: // | | exitB: | | duke@435: // | | stmt4 | | duke@435: // | +------------>-----------------+ | duke@435: // | | duke@435: // +-----------------<---------------------+ duke@435: // duke@435: // duke@435: // final graph duke@435: // duke@435: // stmt1 duke@435: // | duke@435: // v duke@435: // ........> ifA clone duke@435: // : / | duke@435: // dom / | duke@435: // : v v duke@435: // : false true duke@435: // : | | duke@435: // : | stmt2 clone duke@435: // : | | duke@435: // : | v duke@435: // : | newloop<-----+ duke@435: // : | | | duke@435: // : | stmt3 clone | duke@435: // : | | | duke@435: // : | v | duke@435: // : | ifB | duke@435: // : | / \ | duke@435: // : | v v | duke@435: // : | false true | duke@435: // : | | | | duke@435: // : | v stmt2 | duke@435: // : | exitB: | | duke@435: // : | stmt4 v | duke@435: // : | ifA orig | duke@435: // : | / \ | duke@435: // : | / \ | duke@435: // : | v v | duke@435: // : | false true | duke@435: // : | / \ | duke@435: // : v v -----+ duke@435: // RegionA duke@435: // | duke@435: // v duke@435: // exitA duke@435: // duke@435: bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) { duke@435: rasbold@543: if (!loop->_head->is_Loop()) { rasbold@543: return false; } rasbold@543: duke@435: LoopNode *head = loop->_head->as_Loop(); duke@435: duke@435: if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) { duke@435: return false; duke@435: } duke@435: duke@435: // Check for complex exit control duke@435: for(uint ii = 0; ii < loop->_body.size(); ii++ ) { duke@435: Node *n = loop->_body.at(ii); duke@435: int opc = n->Opcode(); duke@435: if (n->is_Call() || duke@435: opc == Op_Catch || duke@435: opc == Op_CatchProj || duke@435: opc == Op_Jump || duke@435: opc == Op_JumpProj) { duke@435: #if !defined(PRODUCT) duke@435: if (TracePartialPeeling) { duke@435: tty->print_cr("\nExit control too complex: lp: %d", head->_idx); duke@435: } duke@435: #endif duke@435: return false; duke@435: } duke@435: } duke@435: duke@435: int dd = dom_depth(head); duke@435: duke@435: // Step 1: find cut point duke@435: duke@435: // Walk up dominators to loop head looking for first loop exit duke@435: // which is executed on every path thru loop. duke@435: IfNode *peel_if = NULL; duke@435: IfNode *peel_if_cmpu = NULL; duke@435: duke@435: Node *iff = loop->tail(); duke@435: while( iff != head ) { duke@435: if( iff->is_If() ) { duke@435: Node *ctrl = get_ctrl(iff->in(1)); duke@435: if (ctrl->is_top()) return false; // Dead test on live IF. duke@435: // If loop-varying exit-test, check for induction variable duke@435: if( loop->is_member(get_loop(ctrl)) && duke@435: loop->is_loop_exit(iff) && duke@435: is_possible_iv_test(iff)) { duke@435: Node* cmp = iff->in(1)->in(1); duke@435: if (cmp->Opcode() == Op_CmpI) { duke@435: peel_if = iff->as_If(); duke@435: } else { duke@435: assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU"); duke@435: peel_if_cmpu = iff->as_If(); duke@435: } duke@435: } duke@435: } duke@435: iff = idom(iff); duke@435: } duke@435: // Prefer signed compare over unsigned compare. duke@435: IfNode* new_peel_if = NULL; duke@435: if (peel_if == NULL) { duke@435: if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) { duke@435: return false; // No peel point found duke@435: } duke@435: new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop); duke@435: if (new_peel_if == NULL) { duke@435: return false; // No peel point found duke@435: } duke@435: peel_if = new_peel_if; duke@435: } duke@435: Node* last_peel = stay_in_loop(peel_if, loop); duke@435: Node* first_not_peeled = stay_in_loop(last_peel, loop); duke@435: if (first_not_peeled == NULL || first_not_peeled == head) { duke@435: return false; duke@435: } duke@435: duke@435: #if !defined(PRODUCT) duke@435: if (TracePartialPeeling) { duke@435: tty->print_cr("before partial peel one iteration"); duke@435: Node_List wl; duke@435: Node* t = head->in(2); duke@435: while (true) { duke@435: wl.push(t); duke@435: if (t == head) break; duke@435: t = idom(t); duke@435: } duke@435: while (wl.size() > 0) { duke@435: Node* tt = wl.pop(); duke@435: tt->dump(); duke@435: if (tt == last_peel) tty->print_cr("-- cut --"); duke@435: } duke@435: } duke@435: #endif duke@435: ResourceArea *area = Thread::current()->resource_area(); duke@435: VectorSet peel(area); duke@435: VectorSet not_peel(area); duke@435: Node_List peel_list(area); duke@435: Node_List worklist(area); duke@435: Node_List sink_list(area); duke@435: duke@435: // Set of cfg nodes to peel are those that are executable from duke@435: // the head through last_peel. duke@435: assert(worklist.size() == 0, "should be empty"); duke@435: worklist.push(head); duke@435: peel.set(head->_idx); duke@435: while (worklist.size() > 0) { duke@435: Node *n = worklist.pop(); duke@435: if (n != last_peel) { duke@435: for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { duke@435: Node* use = n->fast_out(j); duke@435: if (use->is_CFG() && duke@435: loop->is_member(get_loop(use)) && duke@435: !peel.test_set(use->_idx)) { duke@435: worklist.push(use); duke@435: } duke@435: } duke@435: } duke@435: } duke@435: duke@435: // Set of non-cfg nodes to peel are those that are control duke@435: // dependent on the cfg nodes. duke@435: uint i; duke@435: for(i = 0; i < loop->_body.size(); i++ ) { duke@435: Node *n = loop->_body.at(i); duke@435: Node *n_c = has_ctrl(n) ? get_ctrl(n) : n; duke@435: if (peel.test(n_c->_idx)) { duke@435: peel.set(n->_idx); duke@435: } else { duke@435: not_peel.set(n->_idx); duke@435: } duke@435: } duke@435: duke@435: // Step 2: move operations from the peeled section down into the duke@435: // not-peeled section duke@435: duke@435: // Get a post order schedule of nodes in the peel region duke@435: // Result in right-most operand. duke@435: scheduled_nodelist(loop, peel, peel_list ); duke@435: duke@435: assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); duke@435: duke@435: // For future check for too many new phis duke@435: uint old_phi_cnt = 0; duke@435: for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { duke@435: Node* use = head->fast_out(j); duke@435: if (use->is_Phi()) old_phi_cnt++; duke@435: } duke@435: duke@435: #if !defined(PRODUCT) duke@435: if (TracePartialPeeling) { duke@435: tty->print_cr("\npeeled list"); duke@435: } duke@435: #endif duke@435: duke@435: // Evacuate nodes in peel region into the not_peeled region if possible duke@435: uint new_phi_cnt = 0; duke@435: for (i = 0; i < peel_list.size();) { duke@435: Node* n = peel_list.at(i); duke@435: #if !defined(PRODUCT) duke@435: if (TracePartialPeeling) n->dump(); duke@435: #endif duke@435: bool incr = true; duke@435: if ( !n->is_CFG() ) { duke@435: duke@435: if ( has_use_in_set(n, not_peel) ) { duke@435: duke@435: // If not used internal to the peeled region, duke@435: // move "n" from peeled to not_peeled region. duke@435: duke@435: if ( !has_use_internal_to_set(n, peel, loop) ) { duke@435: duke@435: // if not pinned and not a load (which maybe anti-dependent on a store) duke@435: // and not a CMove (Matcher expects only bool->cmove). duke@435: if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) { duke@435: clone_for_use_outside_loop( loop, n, worklist ); duke@435: duke@435: sink_list.push(n); duke@435: peel >>= n->_idx; // delete n from peel set. duke@435: not_peel <<= n->_idx; // add n to not_peel set. duke@435: peel_list.remove(i); duke@435: incr = false; duke@435: #if !defined(PRODUCT) duke@435: if (TracePartialPeeling) { duke@435: tty->print_cr("sink to not_peeled region: %d newbb: %d", duke@435: n->_idx, get_ctrl(n)->_idx); duke@435: } duke@435: #endif duke@435: } duke@435: } else { duke@435: // Otherwise check for special def-use cases that span duke@435: // the peel/not_peel boundary such as bool->if duke@435: clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist ); duke@435: new_phi_cnt++; duke@435: } duke@435: } duke@435: } duke@435: if (incr) i++; duke@435: } duke@435: duke@435: if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) { duke@435: #if !defined(PRODUCT) duke@435: if (TracePartialPeeling) { duke@435: tty->print_cr("\nToo many new phis: %d old %d new cmpi: %c", duke@435: new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F'); duke@435: } duke@435: #endif duke@435: if (new_peel_if != NULL) { duke@435: remove_cmpi_loop_exit(new_peel_if, loop); duke@435: } duke@435: // Inhibit more partial peeling on this loop duke@435: assert(!head->is_partial_peel_loop(), "not partial peeled"); duke@435: head->mark_partial_peel_failed(); duke@435: return false; duke@435: } duke@435: duke@435: // Step 3: clone loop, retarget control, and insert new phis duke@435: duke@435: // Create new loop head for new phis and to hang duke@435: // the nodes being moved (sinked) from the peel region. duke@435: LoopNode* new_head = new (C, 3) LoopNode(last_peel, last_peel); duke@435: _igvn.register_new_node_with_optimizer(new_head); duke@435: assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled"); duke@435: first_not_peeled->set_req(0, new_head); duke@435: set_loop(new_head, loop); duke@435: loop->_body.push(new_head); duke@435: not_peel.set(new_head->_idx); duke@435: set_idom(new_head, last_peel, dom_depth(first_not_peeled)); duke@435: set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled)); duke@435: duke@435: while (sink_list.size() > 0) { duke@435: Node* n = sink_list.pop(); duke@435: set_ctrl(n, new_head); duke@435: } duke@435: duke@435: assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition"); duke@435: duke@435: clone_loop( loop, old_new, dd ); duke@435: duke@435: const uint clone_exit_idx = 1; duke@435: const uint orig_exit_idx = 2; duke@435: assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop"); duke@435: duke@435: Node* head_clone = old_new[head->_idx]; duke@435: LoopNode* new_head_clone = old_new[new_head->_idx]->as_Loop(); duke@435: Node* orig_tail_clone = head_clone->in(2); duke@435: duke@435: // Add phi if "def" node is in peel set and "use" is not duke@435: duke@435: for(i = 0; i < peel_list.size(); i++ ) { duke@435: Node *def = peel_list.at(i); duke@435: if (!def->is_CFG()) { duke@435: for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) { duke@435: Node *use = def->fast_out(j); duke@435: if (has_node(use) && use->in(0) != C->top() && duke@435: (!peel.test(use->_idx) || duke@435: (use->is_Phi() && use->in(0) == head)) ) { duke@435: worklist.push(use); duke@435: } duke@435: } duke@435: while( worklist.size() ) { duke@435: Node *use = worklist.pop(); duke@435: for (uint j = 1; j < use->req(); j++) { duke@435: Node* n = use->in(j); duke@435: if (n == def) { duke@435: duke@435: // "def" is in peel set, "use" is not in peel set duke@435: // or "use" is in the entry boundary (a phi) of the peel set duke@435: duke@435: Node* use_c = has_ctrl(use) ? get_ctrl(use) : use; duke@435: duke@435: if ( loop->is_member(get_loop( use_c )) ) { duke@435: // use is in loop duke@435: if (old_new[use->_idx] != NULL) { // null for dead code duke@435: Node* use_clone = old_new[use->_idx]; duke@435: _igvn.hash_delete(use); duke@435: use->set_req(j, C->top()); duke@435: _igvn._worklist.push(use); duke@435: insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone ); duke@435: } duke@435: } else { duke@435: assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format"); duke@435: // use is not in the loop, check if the live range includes the cut duke@435: Node* lp_if = use_c->in(orig_exit_idx)->in(0); duke@435: if (not_peel.test(lp_if->_idx)) { duke@435: assert(j == orig_exit_idx, "use from original loop"); duke@435: insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone ); duke@435: } duke@435: } duke@435: } duke@435: } duke@435: } duke@435: } duke@435: } duke@435: duke@435: // Step 3b: retarget control duke@435: duke@435: // Redirect control to the new loop head if a cloned node in duke@435: // the not_peeled region has control that points into the peeled region. duke@435: // This necessary because the cloned peeled region will be outside duke@435: // the loop. duke@435: // from to duke@435: // cloned-peeled <---+ duke@435: // new_head_clone: | <--+ duke@435: // cloned-not_peeled in(0) in(0) duke@435: // orig-peeled duke@435: duke@435: for(i = 0; i < loop->_body.size(); i++ ) { duke@435: Node *n = loop->_body.at(i); duke@435: if (!n->is_CFG() && n->in(0) != NULL && duke@435: not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) { duke@435: Node* n_clone = old_new[n->_idx]; duke@435: _igvn.hash_delete(n_clone); duke@435: n_clone->set_req(0, new_head_clone); duke@435: _igvn._worklist.push(n_clone); duke@435: } duke@435: } duke@435: duke@435: // Backedge of the surviving new_head (the clone) is original last_peel duke@435: _igvn.hash_delete(new_head_clone); duke@435: new_head_clone->set_req(LoopNode::LoopBackControl, last_peel); duke@435: _igvn._worklist.push(new_head_clone); duke@435: duke@435: // Cut first node in original not_peel set duke@435: _igvn.hash_delete(new_head); duke@435: new_head->set_req(LoopNode::EntryControl, C->top()); duke@435: new_head->set_req(LoopNode::LoopBackControl, C->top()); duke@435: _igvn._worklist.push(new_head); duke@435: duke@435: // Copy head_clone back-branch info to original head duke@435: // and remove original head's loop entry and duke@435: // clone head's back-branch duke@435: _igvn.hash_delete(head); duke@435: _igvn.hash_delete(head_clone); duke@435: head->set_req(LoopNode::EntryControl, head_clone->in(LoopNode::LoopBackControl)); duke@435: head->set_req(LoopNode::LoopBackControl, C->top()); duke@435: head_clone->set_req(LoopNode::LoopBackControl, C->top()); duke@435: _igvn._worklist.push(head); duke@435: _igvn._worklist.push(head_clone); duke@435: duke@435: // Similarly modify the phis duke@435: for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) { duke@435: Node* use = head->fast_out(k); duke@435: if (use->is_Phi() && use->outcnt() > 0) { duke@435: Node* use_clone = old_new[use->_idx]; duke@435: _igvn.hash_delete(use); duke@435: _igvn.hash_delete(use_clone); duke@435: use->set_req(LoopNode::EntryControl, use_clone->in(LoopNode::LoopBackControl)); duke@435: use->set_req(LoopNode::LoopBackControl, C->top()); duke@435: use_clone->set_req(LoopNode::LoopBackControl, C->top()); duke@435: _igvn._worklist.push(use); duke@435: _igvn._worklist.push(use_clone); duke@435: } duke@435: } duke@435: duke@435: // Step 4: update dominator tree and dominator depth duke@435: duke@435: set_idom(head, orig_tail_clone, dd); duke@435: recompute_dom_depth(); duke@435: duke@435: // Inhibit more partial peeling on this loop duke@435: new_head_clone->set_partial_peel_loop(); duke@435: C->set_major_progress(); duke@435: duke@435: #if !defined(PRODUCT) duke@435: if (TracePartialPeeling) { duke@435: tty->print_cr("\nafter partial peel one iteration"); duke@435: Node_List wl(area); duke@435: Node* t = last_peel; duke@435: while (true) { duke@435: wl.push(t); duke@435: if (t == head_clone) break; duke@435: t = idom(t); duke@435: } duke@435: while (wl.size() > 0) { duke@435: Node* tt = wl.pop(); duke@435: if (tt == head) tty->print_cr("orig head"); duke@435: else if (tt == new_head_clone) tty->print_cr("new head"); duke@435: else if (tt == head_clone) tty->print_cr("clone head"); duke@435: tt->dump(); duke@435: } duke@435: } duke@435: #endif duke@435: return true; duke@435: } duke@435: duke@435: //------------------------------reorg_offsets---------------------------------- duke@435: // Reorganize offset computations to lower register pressure. Mostly duke@435: // prevent loop-fallout uses of the pre-incremented trip counter (which are duke@435: // then alive with the post-incremented trip counter forcing an extra duke@435: // register move) duke@435: void PhaseIdealLoop::reorg_offsets( IdealLoopTree *loop ) { duke@435: duke@435: CountedLoopNode *cl = loop->_head->as_CountedLoop(); duke@435: CountedLoopEndNode *cle = cl->loopexit(); duke@435: if( !cle ) return; // The occasional dead loop duke@435: // Find loop exit control duke@435: Node *exit = cle->proj_out(false); duke@435: assert( exit->Opcode() == Op_IfFalse, "" ); duke@435: duke@435: // Check for the special case of folks using the pre-incremented duke@435: // trip-counter on the fall-out path (forces the pre-incremented duke@435: // and post-incremented trip counter to be live at the same time). duke@435: // Fix this by adjusting to use the post-increment trip counter. duke@435: Node *phi = cl->phi(); duke@435: if( !phi ) return; // Dead infinite loop never@802: never@802: // Shape messed up, probably by iteration_split_impl never@802: if (phi->in(LoopNode::LoopBackControl) != cl->incr()) return; never@802: duke@435: bool progress = true; duke@435: while (progress) { duke@435: progress = false; duke@435: for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) { duke@435: Node* use = phi->fast_out(i); // User of trip-counter duke@435: if (!has_ctrl(use)) continue; duke@435: Node *u_ctrl = get_ctrl(use); duke@435: if( use->is_Phi() ) { duke@435: u_ctrl = NULL; duke@435: for( uint j = 1; j < use->req(); j++ ) duke@435: if( use->in(j) == phi ) duke@435: u_ctrl = dom_lca( u_ctrl, use->in(0)->in(j) ); duke@435: } duke@435: IdealLoopTree *u_loop = get_loop(u_ctrl); duke@435: // Look for loop-invariant use duke@435: if( u_loop == loop ) continue; duke@435: if( loop->is_member( u_loop ) ) continue; duke@435: // Check that use is live out the bottom. Assuming the trip-counter duke@435: // update is right at the bottom, uses of of the loop middle are ok. duke@435: if( dom_lca( exit, u_ctrl ) != exit ) continue; duke@435: // protect against stride not being a constant duke@435: if( !cle->stride_is_con() ) continue; duke@435: // Hit! Refactor use to use the post-incremented tripcounter. duke@435: // Compute a post-increment tripcounter. kvn@651: Node *opaq = new (C, 2) Opaque2Node( C, cle->incr() ); duke@435: register_new_node( opaq, u_ctrl ); duke@435: Node *neg_stride = _igvn.intcon(-cle->stride_con()); duke@435: set_ctrl(neg_stride, C->root()); duke@435: Node *post = new (C, 3) AddINode( opaq, neg_stride); duke@435: register_new_node( post, u_ctrl ); duke@435: _igvn.hash_delete(use); duke@435: _igvn._worklist.push(use); duke@435: for( uint j = 1; j < use->req(); j++ ) duke@435: if( use->in(j) == phi ) duke@435: use->set_req(j, post); duke@435: // Since DU info changed, rerun loop duke@435: progress = true; duke@435: break; duke@435: } duke@435: } duke@435: duke@435: }