aoqi@0: /*
aoqi@0:  * Copyright (c) 2011, 2014, Oracle and/or its affiliates. All rights reserved.
aoqi@0:  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
aoqi@0:  *
aoqi@0:  * This code is free software; you can redistribute it and/or modify it
aoqi@0:  * under the terms of the GNU General Public License version 2 only, as
aoqi@0:  * published by the Free Software Foundation.
aoqi@0:  *
aoqi@0:  * This code is distributed in the hope that it will be useful, but WITHOUT
aoqi@0:  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
aoqi@0:  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
aoqi@0:  * version 2 for more details (a copy is included in the LICENSE file that
aoqi@0:  * accompanied this code).
aoqi@0:  *
aoqi@0:  * You should have received a copy of the GNU General Public License version
aoqi@0:  * 2 along with this work; if not, write to the Free Software Foundation,
aoqi@0:  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
aoqi@0:  *
aoqi@0:  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
aoqi@0:  * or visit www.oracle.com if you need additional information or have any
aoqi@0:  * questions.
aoqi@0:  *
aoqi@0:  */
aoqi@0: 
aoqi@0: #include "precompiled.hpp"
aoqi@0: #include "opto/loopnode.hpp"
aoqi@0: #include "opto/addnode.hpp"
aoqi@0: #include "opto/callnode.hpp"
aoqi@0: #include "opto/connode.hpp"
aoqi@0: #include "opto/loopnode.hpp"
aoqi@0: #include "opto/mulnode.hpp"
aoqi@0: #include "opto/rootnode.hpp"
aoqi@0: #include "opto/subnode.hpp"
aoqi@0: 
aoqi@0: /*
aoqi@0:  * The general idea of Loop Predication is to insert a predicate on the entry
aoqi@0:  * path to a loop, and raise a uncommon trap if the check of the condition fails.
aoqi@0:  * The condition checks are promoted from inside the loop body, and thus
aoqi@0:  * the checks inside the loop could be eliminated. Currently, loop predication
aoqi@0:  * optimization has been applied to remove array range check and loop invariant
aoqi@0:  * checks (such as null checks).
aoqi@0: */
aoqi@0: 
aoqi@0: //-------------------------------register_control-------------------------
aoqi@0: void PhaseIdealLoop::register_control(Node* n, IdealLoopTree *loop, Node* pred) {
aoqi@0:   assert(n->is_CFG(), "must be control node");
aoqi@0:   _igvn.register_new_node_with_optimizer(n);
aoqi@0:   loop->_body.push(n);
aoqi@0:   set_loop(n, loop);
aoqi@0:   // When called from beautify_loops() idom is not constructed yet.
aoqi@0:   if (_idom != NULL) {
aoqi@0:     set_idom(n, pred, dom_depth(pred));
aoqi@0:   }
aoqi@0: }
aoqi@0: 
aoqi@0: //------------------------------create_new_if_for_predicate------------------------
aoqi@0: // create a new if above the uct_if_pattern for the predicate to be promoted.
aoqi@0: //
aoqi@0: //          before                                after
aoqi@0: //        ----------                           ----------
aoqi@0: //           ctrl                                 ctrl
aoqi@0: //            |                                     |
aoqi@0: //            |                                     |
aoqi@0: //            v                                     v
aoqi@0: //           iff                                 new_iff
aoqi@0: //          /    \                                /      \
aoqi@0: //         /      \                              /        \
aoqi@0: //        v        v                            v          v
aoqi@0: //  uncommon_proj cont_proj                   if_uct     if_cont
aoqi@0: // \      |        |                           |          |
aoqi@0: //  \     |        |                           |          |
aoqi@0: //   v    v        v                           |          v
aoqi@0: //     rgn       loop                          |         iff
aoqi@0: //      |                                      |        /     \
aoqi@0: //      |                                      |       /       \
aoqi@0: //      v                                      |      v         v
aoqi@0: // uncommon_trap                               | uncommon_proj cont_proj
aoqi@0: //                                           \  \    |           |
aoqi@0: //                                            \  \   |           |
aoqi@0: //                                             v  v  v           v
aoqi@0: //                                               rgn           loop
aoqi@0: //                                                |
aoqi@0: //                                                |
aoqi@0: //                                                v
aoqi@0: //                                           uncommon_trap
aoqi@0: //
aoqi@0: //
aoqi@0: // We will create a region to guard the uct call if there is no one there.
aoqi@0: // The true projecttion (if_cont) of the new_iff is returned.
aoqi@0: // This code is also used to clone predicates to clonned loops.
aoqi@0: ProjNode* PhaseIdealLoop::create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry,
aoqi@0:                                                       Deoptimization::DeoptReason reason) {
aoqi@0:   assert(cont_proj->is_uncommon_trap_if_pattern(reason), "must be a uct if pattern!");
aoqi@0:   IfNode* iff = cont_proj->in(0)->as_If();
aoqi@0: 
aoqi@0:   ProjNode *uncommon_proj = iff->proj_out(1 - cont_proj->_con);
aoqi@0:   Node     *rgn   = uncommon_proj->unique_ctrl_out();
aoqi@0:   assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct");
aoqi@0: 
aoqi@0:   uint proj_index = 1; // region's edge corresponding to uncommon_proj
aoqi@0:   if (!rgn->is_Region()) { // create a region to guard the call
aoqi@0:     assert(rgn->is_Call(), "must be call uct");
aoqi@0:     CallNode* call = rgn->as_Call();
aoqi@0:     IdealLoopTree* loop = get_loop(call);
aoqi@0:     rgn = new (C) RegionNode(1);
aoqi@0:     rgn->add_req(uncommon_proj);
aoqi@0:     register_control(rgn, loop, uncommon_proj);
aoqi@0:     _igvn.hash_delete(call);
aoqi@0:     call->set_req(0, rgn);
aoqi@0:     // When called from beautify_loops() idom is not constructed yet.
aoqi@0:     if (_idom != NULL) {
aoqi@0:       set_idom(call, rgn, dom_depth(rgn));
aoqi@0:     }
aoqi@0:   } else {
aoqi@0:     // Find region's edge corresponding to uncommon_proj
aoqi@0:     for (; proj_index < rgn->req(); proj_index++)
aoqi@0:       if (rgn->in(proj_index) == uncommon_proj) break;
aoqi@0:     assert(proj_index < rgn->req(), "sanity");
aoqi@0:   }
aoqi@0: 
aoqi@0:   Node* entry = iff->in(0);
aoqi@0:   if (new_entry != NULL) {
aoqi@0:     // Clonning the predicate to new location.
aoqi@0:     entry = new_entry;
aoqi@0:   }
aoqi@0:   // Create new_iff
aoqi@0:   IdealLoopTree* lp = get_loop(entry);
aoqi@0:   IfNode *new_iff = iff->clone()->as_If();
aoqi@0:   new_iff->set_req(0, entry);
aoqi@0:   register_control(new_iff, lp, entry);
aoqi@0:   Node *if_cont = new (C) IfTrueNode(new_iff);
aoqi@0:   Node *if_uct  = new (C) IfFalseNode(new_iff);
aoqi@0:   if (cont_proj->is_IfFalse()) {
aoqi@0:     // Swap
aoqi@0:     Node* tmp = if_uct; if_uct = if_cont; if_cont = tmp;
aoqi@0:   }
aoqi@0:   register_control(if_cont, lp, new_iff);
aoqi@0:   register_control(if_uct, get_loop(rgn), new_iff);
aoqi@0: 
aoqi@0:   // if_uct to rgn
aoqi@0:   _igvn.hash_delete(rgn);
aoqi@0:   rgn->add_req(if_uct);
aoqi@0:   // When called from beautify_loops() idom is not constructed yet.
aoqi@0:   if (_idom != NULL) {
aoqi@0:     Node* ridom = idom(rgn);
aoqi@0:     Node* nrdom = dom_lca(ridom, new_iff);
aoqi@0:     set_idom(rgn, nrdom, dom_depth(rgn));
aoqi@0:   }
aoqi@0: 
aoqi@0:   // If rgn has phis add new edges which has the same
aoqi@0:   // value as on original uncommon_proj pass.
aoqi@0:   assert(rgn->in(rgn->req() -1) == if_uct, "new edge should be last");
aoqi@0:   bool has_phi = false;
aoqi@0:   for (DUIterator_Fast imax, i = rgn->fast_outs(imax); i < imax; i++) {
aoqi@0:     Node* use = rgn->fast_out(i);
aoqi@0:     if (use->is_Phi() && use->outcnt() > 0) {
aoqi@0:       assert(use->in(0) == rgn, "");
aoqi@0:       _igvn.rehash_node_delayed(use);
aoqi@0:       use->add_req(use->in(proj_index));
aoqi@0:       has_phi = true;
aoqi@0:     }
aoqi@0:   }
aoqi@0:   assert(!has_phi || rgn->req() > 3, "no phis when region is created");
aoqi@0: 
aoqi@0:   if (new_entry == NULL) {
aoqi@0:     // Attach if_cont to iff
aoqi@0:     _igvn.hash_delete(iff);
aoqi@0:     iff->set_req(0, if_cont);
aoqi@0:     if (_idom != NULL) {
aoqi@0:       set_idom(iff, if_cont, dom_depth(iff));
aoqi@0:     }
aoqi@0:   }
aoqi@0:   return if_cont->as_Proj();
aoqi@0: }
aoqi@0: 
aoqi@0: //------------------------------create_new_if_for_predicate------------------------
aoqi@0: // Create a new if below new_entry for the predicate to be cloned (IGVN optimization)
aoqi@0: ProjNode* PhaseIterGVN::create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry,
aoqi@0:                                                     Deoptimization::DeoptReason reason) {
aoqi@0:   assert(new_entry != 0, "only used for clone predicate");
aoqi@0:   assert(cont_proj->is_uncommon_trap_if_pattern(reason), "must be a uct if pattern!");
aoqi@0:   IfNode* iff = cont_proj->in(0)->as_If();
aoqi@0: 
aoqi@0:   ProjNode *uncommon_proj = iff->proj_out(1 - cont_proj->_con);
aoqi@0:   Node     *rgn   = uncommon_proj->unique_ctrl_out();
aoqi@0:   assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct");
aoqi@0: 
aoqi@0:   uint proj_index = 1; // region's edge corresponding to uncommon_proj
aoqi@0:   if (!rgn->is_Region()) { // create a region to guard the call
aoqi@0:     assert(rgn->is_Call(), "must be call uct");
aoqi@0:     CallNode* call = rgn->as_Call();
aoqi@0:     rgn = new (C) RegionNode(1);
aoqi@0:     register_new_node_with_optimizer(rgn);
aoqi@0:     rgn->add_req(uncommon_proj);
aoqi@0:     hash_delete(call);
aoqi@0:     call->set_req(0, rgn);
aoqi@0:   } else {
aoqi@0:     // Find region's edge corresponding to uncommon_proj
aoqi@0:     for (; proj_index < rgn->req(); proj_index++)
aoqi@0:       if (rgn->in(proj_index) == uncommon_proj) break;
aoqi@0:     assert(proj_index < rgn->req(), "sanity");
aoqi@0:   }
aoqi@0: 
aoqi@0:   // Create new_iff in new location.
aoqi@0:   IfNode *new_iff = iff->clone()->as_If();
aoqi@0:   new_iff->set_req(0, new_entry);
aoqi@0: 
aoqi@0:   register_new_node_with_optimizer(new_iff);
aoqi@0:   Node *if_cont = new (C) IfTrueNode(new_iff);
aoqi@0:   Node *if_uct  = new (C) IfFalseNode(new_iff);
aoqi@0:   if (cont_proj->is_IfFalse()) {
aoqi@0:     // Swap
aoqi@0:     Node* tmp = if_uct; if_uct = if_cont; if_cont = tmp;
aoqi@0:   }
aoqi@0:   register_new_node_with_optimizer(if_cont);
aoqi@0:   register_new_node_with_optimizer(if_uct);
aoqi@0: 
aoqi@0:   // if_uct to rgn
aoqi@0:   hash_delete(rgn);
aoqi@0:   rgn->add_req(if_uct);
aoqi@0: 
aoqi@0:   // If rgn has phis add corresponding new edges which has the same
aoqi@0:   // value as on original uncommon_proj pass.
aoqi@0:   assert(rgn->in(rgn->req() -1) == if_uct, "new edge should be last");
aoqi@0:   bool has_phi = false;
aoqi@0:   for (DUIterator_Fast imax, i = rgn->fast_outs(imax); i < imax; i++) {
aoqi@0:     Node* use = rgn->fast_out(i);
aoqi@0:     if (use->is_Phi() && use->outcnt() > 0) {
aoqi@0:       rehash_node_delayed(use);
aoqi@0:       use->add_req(use->in(proj_index));
aoqi@0:       has_phi = true;
aoqi@0:     }
aoqi@0:   }
aoqi@0:   assert(!has_phi || rgn->req() > 3, "no phis when region is created");
aoqi@0: 
aoqi@0:   return if_cont->as_Proj();
aoqi@0: }
aoqi@0: 
aoqi@0: //--------------------------clone_predicate-----------------------
aoqi@0: ProjNode* PhaseIdealLoop::clone_predicate(ProjNode* predicate_proj, Node* new_entry,
aoqi@0:                                           Deoptimization::DeoptReason reason,
aoqi@0:                                           PhaseIdealLoop* loop_phase,
aoqi@0:                                           PhaseIterGVN* igvn) {
aoqi@0:   ProjNode* new_predicate_proj;
aoqi@0:   if (loop_phase != NULL) {
aoqi@0:     new_predicate_proj = loop_phase->create_new_if_for_predicate(predicate_proj, new_entry, reason);
aoqi@0:   } else {
aoqi@0:     new_predicate_proj =       igvn->create_new_if_for_predicate(predicate_proj, new_entry, reason);
aoqi@0:   }
aoqi@0:   IfNode* iff = new_predicate_proj->in(0)->as_If();
aoqi@0:   Node* ctrl  = iff->in(0);
aoqi@0: 
aoqi@0:   // Match original condition since predicate's projections could be swapped.
aoqi@0:   assert(predicate_proj->in(0)->in(1)->in(1)->Opcode()==Op_Opaque1, "must be");
aoqi@0:   Node* opq = new (igvn->C) Opaque1Node(igvn->C, predicate_proj->in(0)->in(1)->in(1)->in(1));
aoqi@0:   igvn->C->add_predicate_opaq(opq);
aoqi@0: 
aoqi@0:   Node* bol = new (igvn->C) Conv2BNode(opq);
aoqi@0:   if (loop_phase != NULL) {
aoqi@0:     loop_phase->register_new_node(opq, ctrl);
aoqi@0:     loop_phase->register_new_node(bol, ctrl);
aoqi@0:   } else {
aoqi@0:     igvn->register_new_node_with_optimizer(opq);
aoqi@0:     igvn->register_new_node_with_optimizer(bol);
aoqi@0:   }
aoqi@0:   igvn->hash_delete(iff);
aoqi@0:   iff->set_req(1, bol);
aoqi@0:   return new_predicate_proj;
aoqi@0: }
aoqi@0: 
aoqi@0: 
aoqi@0: //--------------------------clone_loop_predicates-----------------------
aoqi@0: // Interface from IGVN
aoqi@0: Node* PhaseIterGVN::clone_loop_predicates(Node* old_entry, Node* new_entry, bool clone_limit_check) {
aoqi@0:   return PhaseIdealLoop::clone_loop_predicates(old_entry, new_entry, clone_limit_check, NULL, this);
aoqi@0: }
aoqi@0: 
aoqi@0: // Interface from PhaseIdealLoop
aoqi@0: Node* PhaseIdealLoop::clone_loop_predicates(Node* old_entry, Node* new_entry, bool clone_limit_check) {
aoqi@0:   return clone_loop_predicates(old_entry, new_entry, clone_limit_check, this, &this->_igvn);
aoqi@0: }
aoqi@0: 
aoqi@0: // Clone loop predicates to cloned loops (peeled, unswitched, split_if).
aoqi@0: Node* PhaseIdealLoop::clone_loop_predicates(Node* old_entry, Node* new_entry,
aoqi@0:                                                 bool clone_limit_check,
aoqi@0:                                                 PhaseIdealLoop* loop_phase,
aoqi@0:                                                 PhaseIterGVN* igvn) {
aoqi@0: #ifdef ASSERT
aoqi@0:   if (new_entry == NULL || !(new_entry->is_Proj() || new_entry->is_Region() || new_entry->is_SafePoint())) {
aoqi@0:     if (new_entry != NULL)
aoqi@0:       new_entry->dump();
aoqi@0:     assert(false, "not IfTrue, IfFalse, Region or SafePoint");
aoqi@0:   }
aoqi@0: #endif
aoqi@0:   // Search original predicates
aoqi@0:   Node* entry = old_entry;
aoqi@0:   ProjNode* limit_check_proj = NULL;
aoqi@0:   if (LoopLimitCheck) {
aoqi@0:     limit_check_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
aoqi@0:     if (limit_check_proj != NULL) {
aoqi@0:       entry = entry->in(0)->in(0);
aoqi@0:     }
aoqi@0:   }
aoqi@0:   if (UseLoopPredicate) {
aoqi@0:     ProjNode* predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
aoqi@0:     if (predicate_proj != NULL) { // right pattern that can be used by loop predication
aoqi@0:       // clone predicate
aoqi@0:       new_entry = clone_predicate(predicate_proj, new_entry,
aoqi@0:                                   Deoptimization::Reason_predicate,
aoqi@0:                                   loop_phase, igvn);
aoqi@0:       assert(new_entry != NULL && new_entry->is_Proj(), "IfTrue or IfFalse after clone predicate");
aoqi@0:       if (TraceLoopPredicate) {
aoqi@0:         tty->print("Loop Predicate cloned: ");
aoqi@0:         debug_only( new_entry->in(0)->dump(); )
aoqi@0:       }
aoqi@0:     }
aoqi@0:   }
aoqi@0:   if (limit_check_proj != NULL && clone_limit_check) {
aoqi@0:     // Clone loop limit check last to insert it before loop.
aoqi@0:     // Don't clone a limit check which was already finalized
aoqi@0:     // for this counted loop (only one limit check is needed).
aoqi@0:     new_entry = clone_predicate(limit_check_proj, new_entry,
aoqi@0:                                 Deoptimization::Reason_loop_limit_check,
aoqi@0:                                 loop_phase, igvn);
aoqi@0:     assert(new_entry != NULL && new_entry->is_Proj(), "IfTrue or IfFalse after clone limit check");
aoqi@0:     if (TraceLoopLimitCheck) {
aoqi@0:       tty->print("Loop Limit Check cloned: ");
aoqi@0:       debug_only( new_entry->in(0)->dump(); )
aoqi@0:     }
aoqi@0:   }
aoqi@0:   return new_entry;
aoqi@0: }
aoqi@0: 
aoqi@0: //--------------------------skip_loop_predicates------------------------------
aoqi@0: // Skip related predicates.
aoqi@0: Node* PhaseIdealLoop::skip_loop_predicates(Node* entry) {
aoqi@0:   Node* predicate = NULL;
aoqi@0:   if (LoopLimitCheck) {
aoqi@0:     predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
aoqi@0:     if (predicate != NULL) {
aoqi@0:       entry = entry->in(0)->in(0);
aoqi@0:     }
aoqi@0:   }
aoqi@0:   if (UseLoopPredicate) {
aoqi@0:     predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
aoqi@0:     if (predicate != NULL) { // right pattern that can be used by loop predication
aoqi@0:       IfNode* iff = entry->in(0)->as_If();
aoqi@0:       ProjNode* uncommon_proj = iff->proj_out(1 - entry->as_Proj()->_con);
aoqi@0:       Node* rgn = uncommon_proj->unique_ctrl_out();
aoqi@0:       assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct");
aoqi@0:       entry = entry->in(0)->in(0);
aoqi@0:       while (entry != NULL && entry->is_Proj() && entry->in(0)->is_If()) {
aoqi@0:         uncommon_proj = entry->in(0)->as_If()->proj_out(1 - entry->as_Proj()->_con);
aoqi@0:         if (uncommon_proj->unique_ctrl_out() != rgn)
aoqi@0:           break;
aoqi@0:         entry = entry->in(0)->in(0);
aoqi@0:       }
aoqi@0:     }
aoqi@0:   }
aoqi@0:   return entry;
aoqi@0: }
aoqi@0: 
aoqi@0: //--------------------------find_predicate_insertion_point-------------------
aoqi@0: // Find a good location to insert a predicate
aoqi@0: ProjNode* PhaseIdealLoop::find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason) {
aoqi@0:   if (start_c == NULL || !start_c->is_Proj())
aoqi@0:     return NULL;
aoqi@0:   if (start_c->as_Proj()->is_uncommon_trap_if_pattern(reason)) {
aoqi@0:     return start_c->as_Proj();
aoqi@0:   }
aoqi@0:   return NULL;
aoqi@0: }
aoqi@0: 
aoqi@0: //--------------------------find_predicate------------------------------------
aoqi@0: // Find a predicate
aoqi@0: Node* PhaseIdealLoop::find_predicate(Node* entry) {
aoqi@0:   Node* predicate = NULL;
aoqi@0:   if (LoopLimitCheck) {
aoqi@0:     predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
aoqi@0:     if (predicate != NULL) { // right pattern that can be used by loop predication
aoqi@0:       return entry;
aoqi@0:     }
aoqi@0:   }
aoqi@0:   if (UseLoopPredicate) {
aoqi@0:     predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
aoqi@0:     if (predicate != NULL) { // right pattern that can be used by loop predication
aoqi@0:       return entry;
aoqi@0:     }
aoqi@0:   }
aoqi@0:   return NULL;
aoqi@0: }
aoqi@0: 
aoqi@0: //------------------------------Invariance-----------------------------------
aoqi@0: // Helper class for loop_predication_impl to compute invariance on the fly and
aoqi@0: // clone invariants.
aoqi@0: class Invariance : public StackObj {
aoqi@0:   VectorSet _visited, _invariant;
aoqi@0:   Node_Stack _stack;
aoqi@0:   VectorSet _clone_visited;
aoqi@0:   Node_List _old_new; // map of old to new (clone)
aoqi@0:   IdealLoopTree* _lpt;
aoqi@0:   PhaseIdealLoop* _phase;
aoqi@0: 
aoqi@0:   // Helper function to set up the invariance for invariance computation
aoqi@0:   // If n is a known invariant, set up directly. Otherwise, look up the
aoqi@0:   // the possibility to push n onto the stack for further processing.
aoqi@0:   void visit(Node* use, Node* n) {
aoqi@0:     if (_lpt->is_invariant(n)) { // known invariant
aoqi@0:       _invariant.set(n->_idx);
aoqi@0:     } else if (!n->is_CFG()) {
aoqi@0:       Node *n_ctrl = _phase->ctrl_or_self(n);
aoqi@0:       Node *u_ctrl = _phase->ctrl_or_self(use); // self if use is a CFG
aoqi@0:       if (_phase->is_dominator(n_ctrl, u_ctrl)) {
aoqi@0:         _stack.push(n, n->in(0) == NULL ? 1 : 0);
aoqi@0:       }
aoqi@0:     }
aoqi@0:   }
aoqi@0: 
aoqi@0:   // Compute invariance for "the_node" and (possibly) all its inputs recursively
aoqi@0:   // on the fly
aoqi@0:   void compute_invariance(Node* n) {
aoqi@0:     assert(_visited.test(n->_idx), "must be");
aoqi@0:     visit(n, n);
aoqi@0:     while (_stack.is_nonempty()) {
aoqi@0:       Node*  n = _stack.node();
aoqi@0:       uint idx = _stack.index();
aoqi@0:       if (idx == n->req()) { // all inputs are processed
aoqi@0:         _stack.pop();
aoqi@0:         // n is invariant if it's inputs are all invariant
aoqi@0:         bool all_inputs_invariant = true;
aoqi@0:         for (uint i = 0; i < n->req(); i++) {
aoqi@0:           Node* in = n->in(i);
aoqi@0:           if (in == NULL) continue;
aoqi@0:           assert(_visited.test(in->_idx), "must have visited input");
aoqi@0:           if (!_invariant.test(in->_idx)) { // bad guy
aoqi@0:             all_inputs_invariant = false;
aoqi@0:             break;
aoqi@0:           }
aoqi@0:         }
aoqi@0:         if (all_inputs_invariant) {
aoqi@0:           _invariant.set(n->_idx); // I am a invariant too
aoqi@0:         }
aoqi@0:       } else { // process next input
aoqi@0:         _stack.set_index(idx + 1);
aoqi@0:         Node* m = n->in(idx);
aoqi@0:         if (m != NULL && !_visited.test_set(m->_idx)) {
aoqi@0:           visit(n, m);
aoqi@0:         }
aoqi@0:       }
aoqi@0:     }
aoqi@0:   }
aoqi@0: 
aoqi@0:   // Helper function to set up _old_new map for clone_nodes.
aoqi@0:   // If n is a known invariant, set up directly ("clone" of n == n).
aoqi@0:   // Otherwise, push n onto the stack for real cloning.
aoqi@0:   void clone_visit(Node* n) {
aoqi@0:     assert(_invariant.test(n->_idx), "must be invariant");
aoqi@0:     if (_lpt->is_invariant(n)) { // known invariant
aoqi@0:       _old_new.map(n->_idx, n);
aoqi@0:     } else { // to be cloned
aoqi@0:       assert(!n->is_CFG(), "should not see CFG here");
aoqi@0:       _stack.push(n, n->in(0) == NULL ? 1 : 0);
aoqi@0:     }
aoqi@0:   }
aoqi@0: 
aoqi@0:   // Clone "n" and (possibly) all its inputs recursively
aoqi@0:   void clone_nodes(Node* n, Node* ctrl) {
aoqi@0:     clone_visit(n);
aoqi@0:     while (_stack.is_nonempty()) {
aoqi@0:       Node*  n = _stack.node();
aoqi@0:       uint idx = _stack.index();
aoqi@0:       if (idx == n->req()) { // all inputs processed, clone n!
aoqi@0:         _stack.pop();
aoqi@0:         // clone invariant node
aoqi@0:         Node* n_cl = n->clone();
aoqi@0:         _old_new.map(n->_idx, n_cl);
aoqi@0:         _phase->register_new_node(n_cl, ctrl);
aoqi@0:         for (uint i = 0; i < n->req(); i++) {
aoqi@0:           Node* in = n_cl->in(i);
aoqi@0:           if (in == NULL) continue;
aoqi@0:           n_cl->set_req(i, _old_new[in->_idx]);
aoqi@0:         }
aoqi@0:       } else { // process next input
aoqi@0:         _stack.set_index(idx + 1);
aoqi@0:         Node* m = n->in(idx);
aoqi@0:         if (m != NULL && !_clone_visited.test_set(m->_idx)) {
aoqi@0:           clone_visit(m); // visit the input
aoqi@0:         }
aoqi@0:       }
aoqi@0:     }
aoqi@0:   }
aoqi@0: 
aoqi@0:  public:
aoqi@0:   Invariance(Arena* area, IdealLoopTree* lpt) :
aoqi@0:     _lpt(lpt), _phase(lpt->_phase),
aoqi@0:     _visited(area), _invariant(area), _stack(area, 10 /* guess */),
aoqi@0:     _clone_visited(area), _old_new(area)
aoqi@0:   {}
aoqi@0: 
aoqi@0:   // Map old to n for invariance computation and clone
aoqi@0:   void map_ctrl(Node* old, Node* n) {
aoqi@0:     assert(old->is_CFG() && n->is_CFG(), "must be");
aoqi@0:     _old_new.map(old->_idx, n); // "clone" of old is n
aoqi@0:     _invariant.set(old->_idx);  // old is invariant
aoqi@0:     _clone_visited.set(old->_idx);
aoqi@0:   }
aoqi@0: 
aoqi@0:   // Driver function to compute invariance
aoqi@0:   bool is_invariant(Node* n) {
aoqi@0:     if (!_visited.test_set(n->_idx))
aoqi@0:       compute_invariance(n);
aoqi@0:     return (_invariant.test(n->_idx) != 0);
aoqi@0:   }
aoqi@0: 
aoqi@0:   // Driver function to clone invariant
aoqi@0:   Node* clone(Node* n, Node* ctrl) {
aoqi@0:     assert(ctrl->is_CFG(), "must be");
aoqi@0:     assert(_invariant.test(n->_idx), "must be an invariant");
aoqi@0:     if (!_clone_visited.test(n->_idx))
aoqi@0:       clone_nodes(n, ctrl);
aoqi@0:     return _old_new[n->_idx];
aoqi@0:   }
aoqi@0: };
aoqi@0: 
aoqi@0: //------------------------------is_range_check_if -----------------------------------
aoqi@0: // Returns true if the predicate of iff is in "scale*iv + offset u< load_range(ptr)" format
aoqi@0: // Note: this function is particularly designed for loop predication. We require load_range
aoqi@0: //       and offset to be loop invariant computed on the fly by "invar"
aoqi@0: bool IdealLoopTree::is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const {
aoqi@0:   if (!is_loop_exit(iff)) {
aoqi@0:     return false;
aoqi@0:   }
aoqi@0:   if (!iff->in(1)->is_Bool()) {
aoqi@0:     return false;
aoqi@0:   }
aoqi@0:   const BoolNode *bol = iff->in(1)->as_Bool();
aoqi@0:   if (bol->_test._test != BoolTest::lt) {
aoqi@0:     return false;
aoqi@0:   }
aoqi@0:   if (!bol->in(1)->is_Cmp()) {
aoqi@0:     return false;
aoqi@0:   }
aoqi@0:   const CmpNode *cmp = bol->in(1)->as_Cmp();
aoqi@0:   if (cmp->Opcode() != Op_CmpU) {
aoqi@0:     return false;
aoqi@0:   }
aoqi@0:   Node* range = cmp->in(2);
aoqi@0:   if (range->Opcode() != Op_LoadRange) {
aoqi@0:     const TypeInt* tint = phase->_igvn.type(range)->isa_int();
aoqi@0:     if (tint == NULL || tint->empty() || tint->_lo < 0) {
aoqi@0:       // Allow predication on positive values that aren't LoadRanges.
aoqi@0:       // This allows optimization of loops where the length of the
aoqi@0:       // array is a known value and doesn't need to be loaded back
aoqi@0:       // from the array.
aoqi@0:       return false;
aoqi@0:     }
aoqi@0:   }
aoqi@0:   if (!invar.is_invariant(range)) {
aoqi@0:     return false;
aoqi@0:   }
aoqi@0:   Node *iv     = _head->as_CountedLoop()->phi();
aoqi@0:   int   scale  = 0;
aoqi@0:   Node *offset = NULL;
aoqi@0:   if (!phase->is_scaled_iv_plus_offset(cmp->in(1), iv, &scale, &offset)) {
aoqi@0:     return false;
aoqi@0:   }
aoqi@0:   if (offset && !invar.is_invariant(offset)) { // offset must be invariant
aoqi@0:     return false;
aoqi@0:   }
aoqi@0:   return true;
aoqi@0: }
aoqi@0: 
aoqi@0: //------------------------------rc_predicate-----------------------------------
aoqi@0: // Create a range check predicate
aoqi@0: //
aoqi@0: // for (i = init; i < limit; i += stride) {
aoqi@0: //    a[scale*i+offset]
aoqi@0: // }
aoqi@0: //
aoqi@0: // Compute max(scale*i + offset) for init <= i < limit and build the predicate
aoqi@0: // as "max(scale*i + offset) u< a.length".
aoqi@0: //
aoqi@0: // There are two cases for max(scale*i + offset):
aoqi@0: // (1) stride*scale > 0
aoqi@0: //   max(scale*i + offset) = scale*(limit-stride) + offset
aoqi@0: // (2) stride*scale < 0
aoqi@0: //   max(scale*i + offset) = scale*init + offset
aoqi@0: BoolNode* PhaseIdealLoop::rc_predicate(IdealLoopTree *loop, Node* ctrl,
aoqi@0:                                        int scale, Node* offset,
aoqi@0:                                        Node* init, Node* limit, Node* stride,
aoqi@0:                                        Node* range, bool upper) {
aoqi@0:   stringStream* predString = NULL;
aoqi@0:   if (TraceLoopPredicate) {
aoqi@0:     predString = new stringStream();
aoqi@0:     predString->print("rc_predicate ");
aoqi@0:   }
aoqi@0: 
aoqi@0:   Node* max_idx_expr  = init;
aoqi@0:   int stride_con = stride->get_int();
aoqi@0:   if ((stride_con > 0) == (scale > 0) == upper) {
aoqi@0:     if (LoopLimitCheck) {
aoqi@0:       // With LoopLimitCheck limit is not exact.
aoqi@0:       // Calculate exact limit here.
aoqi@0:       // Note, counted loop's test is '<' or '>'.
aoqi@0:       limit = exact_limit(loop);
aoqi@0:       max_idx_expr = new (C) SubINode(limit, stride);
aoqi@0:       register_new_node(max_idx_expr, ctrl);
aoqi@0:       if (TraceLoopPredicate) predString->print("(limit - stride) ");
aoqi@0:     } else {
aoqi@0:       max_idx_expr = new (C) SubINode(limit, stride);
aoqi@0:       register_new_node(max_idx_expr, ctrl);
aoqi@0:       if (TraceLoopPredicate) predString->print("(limit - stride) ");
aoqi@0:     }
aoqi@0:   } else {
aoqi@0:     if (TraceLoopPredicate) predString->print("init ");
aoqi@0:   }
aoqi@0: 
aoqi@0:   if (scale != 1) {
aoqi@0:     ConNode* con_scale = _igvn.intcon(scale);
aoqi@0:     max_idx_expr = new (C) MulINode(max_idx_expr, con_scale);
aoqi@0:     register_new_node(max_idx_expr, ctrl);
aoqi@0:     if (TraceLoopPredicate) predString->print("* %d ", scale);
aoqi@0:   }
aoqi@0: 
aoqi@0:   if (offset && (!offset->is_Con() || offset->get_int() != 0)){
aoqi@0:     max_idx_expr = new (C) AddINode(max_idx_expr, offset);
aoqi@0:     register_new_node(max_idx_expr, ctrl);
aoqi@0:     if (TraceLoopPredicate)
aoqi@0:       if (offset->is_Con()) predString->print("+ %d ", offset->get_int());
aoqi@0:       else predString->print("+ offset ");
aoqi@0:   }
aoqi@0: 
aoqi@0:   CmpUNode* cmp = new (C) CmpUNode(max_idx_expr, range);
aoqi@0:   register_new_node(cmp, ctrl);
aoqi@0:   BoolNode* bol = new (C) BoolNode(cmp, BoolTest::lt);
aoqi@0:   register_new_node(bol, ctrl);
aoqi@0: 
aoqi@0:   if (TraceLoopPredicate) {
aoqi@0:     predString->print_cr("<u range");
aoqi@0:     tty->print("%s", predString->as_string());
aoqi@0:   }
aoqi@0:   return bol;
aoqi@0: }
aoqi@0: 
aoqi@0: //------------------------------ loop_predication_impl--------------------------
aoqi@0: // Insert loop predicates for null checks and range checks
aoqi@0: bool PhaseIdealLoop::loop_predication_impl(IdealLoopTree *loop) {
aoqi@0:   if (!UseLoopPredicate) return false;
aoqi@0: 
aoqi@0:   if (!loop->_head->is_Loop()) {
aoqi@0:     // Could be a simple region when irreducible loops are present.
aoqi@0:     return false;
aoqi@0:   }
aoqi@0:   LoopNode* head = loop->_head->as_Loop();
aoqi@0: 
aoqi@0:   if (head->unique_ctrl_out()->Opcode() == Op_NeverBranch) {
aoqi@0:     // do nothing for infinite loops
aoqi@0:     return false;
aoqi@0:   }
aoqi@0: 
aoqi@0:   CountedLoopNode *cl = NULL;
aoqi@0:   if (head->is_valid_counted_loop()) {
aoqi@0:     cl = head->as_CountedLoop();
aoqi@0:     // do nothing for iteration-splitted loops
aoqi@0:     if (!cl->is_normal_loop()) return false;
aoqi@0:     // Avoid RCE if Counted loop's test is '!='.
aoqi@0:     BoolTest::mask bt = cl->loopexit()->test_trip();
aoqi@0:     if (bt != BoolTest::lt && bt != BoolTest::gt)
aoqi@0:       cl = NULL;
aoqi@0:   }
aoqi@0: 
aoqi@0:   Node* entry = head->in(LoopNode::EntryControl);
aoqi@0:   ProjNode *predicate_proj = NULL;
aoqi@0:   // Loop limit check predicate should be near the loop.
aoqi@0:   if (LoopLimitCheck) {
aoqi@0:     predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
aoqi@0:     if (predicate_proj != NULL)
aoqi@0:       entry = predicate_proj->in(0)->in(0);
aoqi@0:   }
aoqi@0: 
aoqi@0:   predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
aoqi@0:   if (!predicate_proj) {
aoqi@0: #ifndef PRODUCT
aoqi@0:     if (TraceLoopPredicate) {
aoqi@0:       tty->print("missing predicate:");
aoqi@0:       loop->dump_head();
aoqi@0:       head->dump(1);
aoqi@0:     }
aoqi@0: #endif
aoqi@0:     return false;
aoqi@0:   }
aoqi@0:   ConNode* zero = _igvn.intcon(0);
aoqi@0:   set_ctrl(zero, C->root());
aoqi@0: 
aoqi@0:   ResourceArea *area = Thread::current()->resource_area();
aoqi@0:   Invariance invar(area, loop);
aoqi@0: 
aoqi@0:   // Create list of if-projs such that a newer proj dominates all older
aoqi@0:   // projs in the list, and they all dominate loop->tail()
aoqi@0:   Node_List if_proj_list(area);
aoqi@0:   Node *current_proj = loop->tail(); //start from tail
aoqi@0:   while (current_proj != head) {
aoqi@0:     if (loop == get_loop(current_proj) && // still in the loop ?
aoqi@0:         current_proj->is_Proj()        && // is a projection  ?
aoqi@0:         current_proj->in(0)->Opcode() == Op_If) { // is a if projection ?
aoqi@0:       if_proj_list.push(current_proj);
aoqi@0:     }
aoqi@0:     current_proj = idom(current_proj);
aoqi@0:   }
aoqi@0: 
aoqi@0:   bool hoisted = false; // true if at least one proj is promoted
aoqi@0:   while (if_proj_list.size() > 0) {
aoqi@0:     // Following are changed to nonnull when a predicate can be hoisted
aoqi@0:     ProjNode* new_predicate_proj = NULL;
aoqi@0: 
aoqi@0:     ProjNode* proj = if_proj_list.pop()->as_Proj();
aoqi@0:     IfNode*   iff  = proj->in(0)->as_If();
aoqi@0: 
aoqi@0:     if (!proj->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) {
aoqi@0:       if (loop->is_loop_exit(iff)) {
aoqi@0:         // stop processing the remaining projs in the list because the execution of them
aoqi@0:         // depends on the condition of "iff" (iff->in(1)).
aoqi@0:         break;
aoqi@0:       } else {
aoqi@0:         // Both arms are inside the loop. There are two cases:
aoqi@0:         // (1) there is one backward branch. In this case, any remaining proj
aoqi@0:         //     in the if_proj list post-dominates "iff". So, the condition of "iff"
aoqi@0:         //     does not determine the execution the remining projs directly, and we
aoqi@0:         //     can safely continue.
aoqi@0:         // (2) both arms are forwarded, i.e. a diamond shape. In this case, "proj"
aoqi@0:         //     does not dominate loop->tail(), so it can not be in the if_proj list.
aoqi@0:         continue;
aoqi@0:       }
aoqi@0:     }
aoqi@0: 
aoqi@0:     Node*     test = iff->in(1);
aoqi@0:     if (!test->is_Bool()){ //Conv2B, ...
aoqi@0:       continue;
aoqi@0:     }
aoqi@0:     BoolNode* bol = test->as_Bool();
aoqi@0:     if (invar.is_invariant(bol)) {
aoqi@0:       // Invariant test
aoqi@0:       new_predicate_proj = create_new_if_for_predicate(predicate_proj, NULL,
aoqi@0:                                                        Deoptimization::Reason_predicate);
aoqi@0:       Node* ctrl = new_predicate_proj->in(0)->as_If()->in(0);
aoqi@0:       BoolNode* new_predicate_bol = invar.clone(bol, ctrl)->as_Bool();
aoqi@0: 
aoqi@0:       // Negate test if necessary
aoqi@0:       bool negated = false;
aoqi@0:       if (proj->_con != predicate_proj->_con) {
aoqi@0:         new_predicate_bol = new (C) BoolNode(new_predicate_bol->in(1), new_predicate_bol->_test.negate());
aoqi@0:         register_new_node(new_predicate_bol, ctrl);
aoqi@0:         negated = true;
aoqi@0:       }
aoqi@0:       IfNode* new_predicate_iff = new_predicate_proj->in(0)->as_If();
aoqi@0:       _igvn.hash_delete(new_predicate_iff);
aoqi@0:       new_predicate_iff->set_req(1, new_predicate_bol);
aoqi@0: #ifndef PRODUCT
aoqi@0:       if (TraceLoopPredicate) {
aoqi@0:         tty->print("Predicate invariant if%s: %d ", negated ? " negated" : "", new_predicate_iff->_idx);
aoqi@0:         loop->dump_head();
aoqi@0:       } else if (TraceLoopOpts) {
aoqi@0:         tty->print("Predicate IC ");
aoqi@0:         loop->dump_head();
aoqi@0:       }
aoqi@0: #endif
aoqi@0:     } else if ((cl != NULL) && (proj->_con == predicate_proj->_con) &&
aoqi@0:                loop->is_range_check_if(iff, this, invar)) {
aoqi@0: 
aoqi@0:       // Range check for counted loops
aoqi@0:       const Node*    cmp    = bol->in(1)->as_Cmp();
aoqi@0:       Node*          idx    = cmp->in(1);
aoqi@0:       assert(!invar.is_invariant(idx), "index is variant");
aoqi@0:       Node* rng = cmp->in(2);
aoqi@0:       assert(rng->Opcode() == Op_LoadRange || _igvn.type(rng)->is_int() >= 0, "must be");
aoqi@0:       assert(invar.is_invariant(rng), "range must be invariant");
aoqi@0:       int scale    = 1;
aoqi@0:       Node* offset = zero;
aoqi@0:       bool ok = is_scaled_iv_plus_offset(idx, cl->phi(), &scale, &offset);
aoqi@0:       assert(ok, "must be index expression");
aoqi@0: 
aoqi@0:       Node* init    = cl->init_trip();
aoqi@0:       Node* limit   = cl->limit();
aoqi@0:       Node* stride  = cl->stride();
aoqi@0: 
aoqi@0:       // Build if's for the upper and lower bound tests.  The
aoqi@0:       // lower_bound test will dominate the upper bound test and all
aoqi@0:       // cloned or created nodes will use the lower bound test as
aoqi@0:       // their declared control.
aoqi@0:       ProjNode* lower_bound_proj = create_new_if_for_predicate(predicate_proj, NULL, Deoptimization::Reason_predicate);
aoqi@0:       ProjNode* upper_bound_proj = create_new_if_for_predicate(predicate_proj, NULL, Deoptimization::Reason_predicate);
aoqi@0:       assert(upper_bound_proj->in(0)->as_If()->in(0) == lower_bound_proj, "should dominate");
aoqi@0:       Node *ctrl = lower_bound_proj->in(0)->as_If()->in(0);
aoqi@0: 
aoqi@0:       // Perform cloning to keep Invariance state correct since the
aoqi@0:       // late schedule will place invariant things in the loop.
aoqi@0:       rng = invar.clone(rng, ctrl);
aoqi@0:       if (offset && offset != zero) {
aoqi@0:         assert(invar.is_invariant(offset), "offset must be loop invariant");
aoqi@0:         offset = invar.clone(offset, ctrl);
aoqi@0:       }
aoqi@0: 
aoqi@0:       // Test the lower bound
aoqi@0:       Node*  lower_bound_bol = rc_predicate(loop, ctrl, scale, offset, init, limit, stride, rng, false);
aoqi@0:       IfNode* lower_bound_iff = lower_bound_proj->in(0)->as_If();
aoqi@0:       _igvn.hash_delete(lower_bound_iff);
aoqi@0:       lower_bound_iff->set_req(1, lower_bound_bol);
aoqi@0:       if (TraceLoopPredicate) tty->print_cr("lower bound check if: %d", lower_bound_iff->_idx);
aoqi@0: 
aoqi@0:       // Test the upper bound
aoqi@0:       Node* upper_bound_bol = rc_predicate(loop, lower_bound_proj, scale, offset, init, limit, stride, rng, true);
aoqi@0:       IfNode* upper_bound_iff = upper_bound_proj->in(0)->as_If();
aoqi@0:       _igvn.hash_delete(upper_bound_iff);
aoqi@0:       upper_bound_iff->set_req(1, upper_bound_bol);
aoqi@0:       if (TraceLoopPredicate) tty->print_cr("upper bound check if: %d", lower_bound_iff->_idx);
aoqi@0: 
aoqi@0:       // Fall through into rest of the clean up code which will move
aoqi@0:       // any dependent nodes onto the upper bound test.
aoqi@0:       new_predicate_proj = upper_bound_proj;
aoqi@0: 
aoqi@0: #ifndef PRODUCT
aoqi@0:       if (TraceLoopOpts && !TraceLoopPredicate) {
aoqi@0:         tty->print("Predicate RC ");
aoqi@0:         loop->dump_head();
aoqi@0:       }
aoqi@0: #endif
aoqi@0:     } else {
aoqi@0:       // Loop variant check (for example, range check in non-counted loop)
aoqi@0:       // with uncommon trap.
aoqi@0:       continue;
aoqi@0:     }
aoqi@0:     assert(new_predicate_proj != NULL, "sanity");
aoqi@0:     // Success - attach condition (new_predicate_bol) to predicate if
aoqi@0:     invar.map_ctrl(proj, new_predicate_proj); // so that invariance test can be appropriate
aoqi@0: 
aoqi@0:     // Eliminate the old If in the loop body
aoqi@0:     dominated_by( new_predicate_proj, iff, proj->_con != new_predicate_proj->_con );
aoqi@0: 
aoqi@0:     hoisted = true;
aoqi@0:     C->set_major_progress();
aoqi@0:   } // end while
aoqi@0: 
aoqi@0: #ifndef PRODUCT
aoqi@0:   // report that the loop predication has been actually performed
aoqi@0:   // for this loop
aoqi@0:   if (TraceLoopPredicate && hoisted) {
aoqi@0:     tty->print("Loop Predication Performed:");
aoqi@0:     loop->dump_head();
aoqi@0:   }
aoqi@0: #endif
aoqi@0: 
aoqi@0:   return hoisted;
aoqi@0: }
aoqi@0: 
aoqi@0: //------------------------------loop_predication--------------------------------
aoqi@0: // driver routine for loop predication optimization
aoqi@0: bool IdealLoopTree::loop_predication( PhaseIdealLoop *phase) {
aoqi@0:   bool hoisted = false;
aoqi@0:   // Recursively promote predicates
aoqi@0:   if (_child) {
aoqi@0:     hoisted = _child->loop_predication( phase);
aoqi@0:   }
aoqi@0: 
aoqi@0:   // self
aoqi@0:   if (!_irreducible && !tail()->is_top()) {
aoqi@0:     hoisted |= phase->loop_predication_impl(this);
aoqi@0:   }
aoqi@0: 
aoqi@0:   if (_next) { //sibling
aoqi@0:     hoisted |= _next->loop_predication( phase);
aoqi@0:   }
aoqi@0: 
aoqi@0:   return hoisted;
aoqi@0: }