loopnode.hpp revision 2230
1008N/A/*
1008N/A * Copyright (c) 1998, 2011, Oracle and/or its affiliates. All rights reserved.
1008N/A * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
1008N/A *
1008N/A * This code is free software; you can redistribute it and/or modify it
1008N/A * under the terms of the GNU General Public License version 2 only, as
1008N/A * published by the Free Software Foundation.
1008N/A *
1008N/A * This code is distributed in the hope that it will be useful, but WITHOUT
1008N/A * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
1008N/A * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
1008N/A * version 2 for more details (a copy is included in the LICENSE file that
1008N/A * accompanied this code).
1008N/A *
1008N/A * You should have received a copy of the GNU General Public License version
1008N/A * 2 along with this work; if not, write to the Free Software Foundation,
1008N/A * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
1008N/A *
1008N/A * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
1008N/A * or visit www.oracle.com if you need additional information or have any
1008N/A * questions.
1008N/A *
1008N/A */
1008N/A
1008N/A#ifndef SHARE_VM_OPTO_LOOPNODE_HPP
1008N/A#define SHARE_VM_OPTO_LOOPNODE_HPP
1008N/A
1008N/A#include "opto/cfgnode.hpp"
1008N/A#include "opto/multnode.hpp"
1008N/A#include "opto/phaseX.hpp"
1008N/A#include "opto/subnode.hpp"
1008N/A#include "opto/type.hpp"
1008N/A
1008N/Aclass CmpNode;
1008N/Aclass CountedLoopEndNode;
1008N/Aclass CountedLoopNode;
1008N/Aclass IdealLoopTree;
1008N/Aclass LoopNode;
1008N/Aclass Node;
1008N/Aclass PhaseIdealLoop;
1008N/Aclass VectorSet;
1008N/Aclass Invariance;
1008N/Astruct small_cache;
1008N/A
1008N/A//
1008N/A// I D E A L I Z E D L O O P S
1008N/A//
1008N/A// Idealized loops are the set of loops I perform more interesting
1008N/A// transformations on, beyond simple hoisting.
1008N/A
1008N/A//------------------------------LoopNode---------------------------------------
1008N/A// Simple loop header. Fall in path on left, loop-back path on right.
1008N/Aclass LoopNode : public RegionNode {
1008N/A // Size is bigger to hold the flags. However, the flags do not change
1008N/A // the semantics so it does not appear in the hash & cmp functions.
1008N/A virtual uint size_of() const { return sizeof(*this); }
1008N/Aprotected:
1008N/A short _loop_flags;
1008N/A // Names for flag bitfields
1008N/A enum { pre_post_main=0, inner_loop=8, partial_peel_loop=16, partial_peel_failed=32 };
1008N/A char _unswitch_count;
1008N/A enum { _unswitch_max=3 };
1008N/A
1008N/Apublic:
1008N/A // Names for edge indices
1008N/A enum { Self=0, EntryControl, LoopBackControl };
1008N/A
1008N/A int is_inner_loop() const { return _loop_flags & inner_loop; }
1008N/A void set_inner_loop() { _loop_flags |= inner_loop; }
1008N/A
1008N/A int is_partial_peel_loop() const { return _loop_flags & partial_peel_loop; }
1008N/A void set_partial_peel_loop() { _loop_flags |= partial_peel_loop; }
1008N/A int partial_peel_has_failed() const { return _loop_flags & partial_peel_failed; }
1008N/A void mark_partial_peel_failed() { _loop_flags |= partial_peel_failed; }
1008N/A
1008N/A int unswitch_max() { return _unswitch_max; }
1008N/A int unswitch_count() { return _unswitch_count; }
1008N/A void set_unswitch_count(int val) {
1008N/A assert (val <= unswitch_max(), "too many unswitches");
1008N/A _unswitch_count = val;
1008N/A }
1008N/A
1008N/A LoopNode( Node *entry, Node *backedge ) : RegionNode(3), _loop_flags(0), _unswitch_count(0) {
1008N/A init_class_id(Class_Loop);
1008N/A init_req(EntryControl, entry);
1008N/A init_req(LoopBackControl, backedge);
1008N/A }
1008N/A
1008N/A virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
1008N/A virtual int Opcode() const;
1008N/A bool can_be_counted_loop(PhaseTransform* phase) const {
1008N/A return req() == 3 && in(0) != NULL &&
1008N/A in(1) != NULL && phase->type(in(1)) != Type::TOP &&
1008N/A in(2) != NULL && phase->type(in(2)) != Type::TOP;
1008N/A }
1008N/A bool is_valid_counted_loop() const;
1008N/A#ifndef PRODUCT
1008N/A virtual void dump_spec(outputStream *st) const;
1008N/A#endif
1008N/A};
1008N/A
1008N/A//------------------------------Counted Loops----------------------------------
1008N/A// Counted loops are all trip-counted loops, with exactly 1 trip-counter exit
1008N/A// path (and maybe some other exit paths). The trip-counter exit is always
1008N/A// last in the loop. The trip-counter have to stride by a constant;
1008N/A// the exit value is also loop invariant.
1008N/A
1008N/A// CountedLoopNodes and CountedLoopEndNodes come in matched pairs. The
1008N/A// CountedLoopNode has the incoming loop control and the loop-back-control
1008N/A// which is always the IfTrue before the matching CountedLoopEndNode. The
1008N/A// CountedLoopEndNode has an incoming control (possibly not the
1008N/A// CountedLoopNode if there is control flow in the loop), the post-increment
1008N/A// trip-counter value, and the limit. The trip-counter value is always of
1008N/A// the form (Op old-trip-counter stride). The old-trip-counter is produced
1008N/A// by a Phi connected to the CountedLoopNode. The stride is constant.
1008N/A// The Op is any commutable opcode, including Add, Mul, Xor. The
1008N/A// CountedLoopEndNode also takes in the loop-invariant limit value.
1008N/A
1008N/A// From a CountedLoopNode I can reach the matching CountedLoopEndNode via the
1008N/A// loop-back control. From CountedLoopEndNodes I can reach CountedLoopNodes
1008N/A// via the old-trip-counter from the Op node.
1008N/A
1008N/A//------------------------------CountedLoopNode--------------------------------
1008N/A// CountedLoopNodes head simple counted loops. CountedLoopNodes have as
1008N/A// inputs the incoming loop-start control and the loop-back control, so they
1008N/A// act like RegionNodes. They also take in the initial trip counter, the
1008N/A// loop-invariant stride and the loop-invariant limit value. CountedLoopNodes
1008N/A// produce a loop-body control and the trip counter value. Since
1008N/A// CountedLoopNodes behave like RegionNodes I still have a standard CFG model.
1008N/A
1008N/Aclass CountedLoopNode : public LoopNode {
1008N/A // Size is bigger to hold _main_idx. However, _main_idx does not change
1008N/A // the semantics so it does not appear in the hash & cmp functions.
1008N/A virtual uint size_of() const { return sizeof(*this); }
1008N/A
1008N/A // For Pre- and Post-loops during debugging ONLY, this holds the index of
1008N/A // the Main CountedLoop. Used to assert that we understand the graph shape.
1008N/A node_idx_t _main_idx;
1008N/A
1008N/A // Known trip count calculated by policy_maximally_unroll
1008N/A int _trip_count;
1008N/A
1008N/A // Expected trip count from profile data
1008N/A float _profile_trip_cnt;
1008N/A
1008N/A // Log2 of original loop bodies in unrolled loop
1008N/A int _unrolled_count_log2;
1008N/A
1008N/A // Node count prior to last unrolling - used to decide if
1008N/A // unroll,optimize,unroll,optimize,... is making progress
1008N/A int _node_count_before_unroll;
1008N/A
1008N/Apublic:
1008N/A CountedLoopNode( Node *entry, Node *backedge )
1008N/A : LoopNode(entry, backedge), _trip_count(max_jint),
1008N/A _profile_trip_cnt(COUNT_UNKNOWN), _unrolled_count_log2(0),
1008N/A _node_count_before_unroll(0) {
1008N/A init_class_id(Class_CountedLoop);
1008N/A // Initialize _trip_count to the largest possible value.
1008N/A // Will be reset (lower) if the loop's trip count is known.
1008N/A }
1008N/A
1008N/A virtual int Opcode() const;
1008N/A virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
1008N/A
1008N/A Node *init_control() const { return in(EntryControl); }
1008N/A Node *back_control() const { return in(LoopBackControl); }
1008N/A CountedLoopEndNode *loopexit() const;
1008N/A Node *init_trip() const;
1008N/A Node *stride() const;
1008N/A int stride_con() const;
1008N/A bool stride_is_con() const;
1008N/A Node *limit() const;
1008N/A Node *incr() const;
1008N/A Node *phi() const;
1008N/A
1008N/A // Match increment with optional truncation
1008N/A static Node* match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type);
1008N/A
1008N/A // A 'main' loop has a pre-loop and a post-loop. The 'main' loop
1008N/A // can run short a few iterations and may start a few iterations in.
1008N/A // It will be RCE'd and unrolled and aligned.
1008N/A
1008N/A // A following 'post' loop will run any remaining iterations. Used
1008N/A // during Range Check Elimination, the 'post' loop will do any final
1008N/A // iterations with full checks. Also used by Loop Unrolling, where
1008N/A // the 'post' loop will do any epilog iterations needed. Basically,
1008N/A // a 'post' loop can not profitably be further unrolled or RCE'd.
1008N/A
1008N/A // A preceding 'pre' loop will run at least 1 iteration (to do peeling),
1008N/A // it may do under-flow checks for RCE and may do alignment iterations
1008N/A // so the following main loop 'knows' that it is striding down cache
1008N/A // lines.
1008N/A
1008N/A // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or
1008N/A // Aligned, may be missing it's pre-loop.
1008N/A enum { Normal=0, Pre=1, Main=2, Post=3, PrePostFlagsMask=3, Main_Has_No_Pre_Loop=4 };
1008N/A int is_normal_loop() const { return (_loop_flags&PrePostFlagsMask) == Normal; }
1008N/A int is_pre_loop () const { return (_loop_flags&PrePostFlagsMask) == Pre; }
1008N/A int is_main_loop () const { return (_loop_flags&PrePostFlagsMask) == Main; }
1008N/A int is_post_loop () const { return (_loop_flags&PrePostFlagsMask) == Post; }
1008N/A int is_main_no_pre_loop() const { return _loop_flags & Main_Has_No_Pre_Loop; }
1008N/A void set_main_no_pre_loop() { _loop_flags |= Main_Has_No_Pre_Loop; }
1008N/A
1008N/A int main_idx() const { return _main_idx; }
1008N/A
1008N/A
1008N/A void set_pre_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; }
1008N/A void set_main_loop ( ) { assert(is_normal_loop(),""); _loop_flags |= Main; }
1008N/A void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; }
1008N/A void set_normal_loop( ) { _loop_flags &= ~PrePostFlagsMask; }
1008N/A
1008N/A void set_trip_count(int tc) { _trip_count = tc; }
1008N/A int trip_count() { return _trip_count; }
1008N/A
1008N/A void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; }
1008N/A float profile_trip_cnt() { return _profile_trip_cnt; }
1008N/A
1008N/A void double_unrolled_count() { _unrolled_count_log2++; }
1008N/A int unrolled_count() { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); }
1008N/A
1008N/A void set_node_count_before_unroll(int ct) { _node_count_before_unroll = ct; }
1008N/A int node_count_before_unroll() { return _node_count_before_unroll; }
1008N/A
1008N/A#ifndef PRODUCT
1008N/A virtual void dump_spec(outputStream *st) const;
1008N/A#endif
1008N/A};
1008N/A
1008N/A//------------------------------CountedLoopEndNode-----------------------------
1008N/A// CountedLoopEndNodes end simple trip counted loops. They act much like
1008N/A// IfNodes.
1008N/Aclass CountedLoopEndNode : public IfNode {
1008N/Apublic:
1008N/A enum { TestControl, TestValue };
1008N/A
1008N/A CountedLoopEndNode( Node *control, Node *test, float prob, float cnt )
1008N/A : IfNode( control, test, prob, cnt) {
1008N/A init_class_id(Class_CountedLoopEnd);
1008N/A }
1008N/A virtual int Opcode() const;
1008N/A
1008N/A Node *cmp_node() const { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : NULL; }
1008N/A Node *incr() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
1008N/A Node *limit() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; }
1008N/A Node *stride() const { Node *tmp = incr (); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; }
1008N/A Node *phi() const { Node *tmp = incr (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
1008N/A Node *init_trip() const { Node *tmp = phi (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
1008N/A int stride_con() const;
1008N/A bool stride_is_con() const { Node *tmp = stride (); return (tmp != NULL && tmp->is_Con()); }
1008N/A BoolTest::mask test_trip() const { return in(TestValue)->as_Bool()->_test._test; }
1008N/A CountedLoopNode *loopnode() const {
1008N/A Node *ln = phi()->in(0);
1142N/A assert( ln->Opcode() == Op_CountedLoop, "malformed loop" );
1008N/A return (CountedLoopNode*)ln; }
1008N/A
1008N/A#ifndef PRODUCT
1008N/A virtual void dump_spec(outputStream *st) const;
1008N/A#endif
1008N/A};
1008N/A
1008N/A
1008N/Ainline CountedLoopEndNode *CountedLoopNode::loopexit() const {
1142N/A Node *bc = back_control();
1008N/A if( bc == NULL ) return NULL;
1008N/A Node *le = bc->in(0);
1008N/A if( le->Opcode() != Op_CountedLoopEnd )
1008N/A return NULL;
1008N/A return (CountedLoopEndNode*)le;
1008N/A}
1008N/Ainline Node *CountedLoopNode::init_trip() const { return loopexit() ? loopexit()->init_trip() : NULL; }
1008N/Ainline Node *CountedLoopNode::stride() const { return loopexit() ? loopexit()->stride() : NULL; }
1008N/Ainline int CountedLoopNode::stride_con() const { return loopexit() ? loopexit()->stride_con() : 0; }
1008N/Ainline bool CountedLoopNode::stride_is_con() const { return loopexit() && loopexit()->stride_is_con(); }
1008N/Ainline Node *CountedLoopNode::limit() const { return loopexit() ? loopexit()->limit() : NULL; }
1008N/Ainline Node *CountedLoopNode::incr() const { return loopexit() ? loopexit()->incr() : NULL; }
1008N/Ainline Node *CountedLoopNode::phi() const { return loopexit() ? loopexit()->phi() : NULL; }
1008N/A
1008N/A
1008N/A// -----------------------------IdealLoopTree----------------------------------
1008N/Aclass IdealLoopTree : public ResourceObj {
1008N/Apublic:
1008N/A IdealLoopTree *_parent; // Parent in loop tree
1008N/A IdealLoopTree *_next; // Next sibling in loop tree
1008N/A IdealLoopTree *_child; // First child in loop tree
1008N/A
1008N/A // The head-tail backedge defines the loop.
1008N/A // If tail is NULL then this loop has multiple backedges as part of the
1008N/A // same loop. During cleanup I'll peel off the multiple backedges; merge
1008N/A // them at the loop bottom and flow 1 real backedge into the loop.
1008N/A Node *_head; // Head of loop
1008N/A Node *_tail; // Tail of loop
1008N/A inline Node *tail(); // Handle lazy update of _tail field
1008N/A PhaseIdealLoop* _phase;
1008N/A
1008N/A Node_List _body; // Loop body for inner loops
1008N/A
1008N/A uint8 _nest; // Nesting depth
1008N/A uint8 _irreducible:1, // True if irreducible
1008N/A _has_call:1, // True if has call safepoint
1008N/A _has_sfpt:1, // True if has non-call safepoint
1008N/A _rce_candidate:1; // True if candidate for range check elimination
1008N/A
1008N/A Node_List* _required_safept; // A inner loop cannot delete these safepts;
1008N/A bool _allow_optimizations; // Allow loop optimizations
1008N/A
1008N/A IdealLoopTree( PhaseIdealLoop* phase, Node *head, Node *tail )
1008N/A : _parent(0), _next(0), _child(0),
1008N/A _head(head), _tail(tail),
1008N/A _phase(phase),
1008N/A _required_safept(NULL),
1008N/A _allow_optimizations(true),
1008N/A _nest(0), _irreducible(0), _has_call(0), _has_sfpt(0), _rce_candidate(0)
1008N/A { }
1008N/A
1008N/A // Is 'l' a member of 'this'?
1008N/A int is_member( const IdealLoopTree *l ) const; // Test for nested membership
1008N/A
1008N/A // Set loop nesting depth. Accumulate has_call bits.
1008N/A int set_nest( uint depth );
1008N/A
1008N/A // Split out multiple fall-in edges from the loop header. Move them to a
1008N/A // private RegionNode before the loop. This becomes the loop landing pad.
1008N/A void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt );
1008N/A
1008N/A // Split out the outermost loop from this shared header.
1089N/A void split_outer_loop( PhaseIdealLoop *phase );
1008N/A
1008N/A // Merge all the backedges from the shared header into a private Region.
1008N/A // Feed that region as the one backedge to this loop.
1008N/A void merge_many_backedges( PhaseIdealLoop *phase );
1008N/A
1008N/A // Split shared headers and insert loop landing pads.
1008N/A // Insert a LoopNode to replace the RegionNode.
1008N/A // Returns TRUE if loop tree is structurally changed.
1008N/A bool beautify_loops( PhaseIdealLoop *phase );
1008N/A
1008N/A // Perform optimization to use the loop predicates for null checks and range checks.
1008N/A // Applies to any loop level (not just the innermost one)
1008N/A bool loop_predication( PhaseIdealLoop *phase);
1008N/A
1008N/A // Perform iteration-splitting on inner loops. Split iterations to
1008N/A // avoid range checks or one-shot null checks. Returns false if the
1008N/A // current round of loop opts should stop.
1008N/A bool iteration_split( PhaseIdealLoop *phase, Node_List &old_new );
1008N/A
1008N/A // Driver for various flavors of iteration splitting. Returns false
1008N/A // if the current round of loop opts should stop.
1008N/A bool iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new );
1008N/A
1008N/A // Given dominators, try to find loops with calls that must always be
1008N/A // executed (call dominates loop tail). These loops do not need non-call
1008N/A // safepoints (ncsfpt).
1008N/A void check_safepts(VectorSet &visited, Node_List &stack);
1008N/A
1008N/A // Allpaths backwards scan from loop tail, terminating each path at first safepoint
1008N/A // encountered.
1008N/A void allpaths_check_safepts(VectorSet &visited, Node_List &stack);
1008N/A
1008N/A // Convert to counted loops where possible
1008N/A void counted_loop( PhaseIdealLoop *phase );
1008N/A
1008N/A // Check for Node being a loop-breaking test
1008N/A Node *is_loop_exit(Node *iff) const;
1008N/A
1008N/A // Returns true if ctrl is executed on every complete iteration
1008N/A bool dominates_backedge(Node* ctrl);
1008N/A
1008N/A // Remove simplistic dead code from loop body
1008N/A void DCE_loop_body();
1008N/A
1008N/A // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
1008N/A // Replace with a 1-in-10 exit guess.
1008N/A void adjust_loop_exit_prob( PhaseIdealLoop *phase );
1008N/A
1008N/A // Return TRUE or FALSE if the loop should never be RCE'd or aligned.
1008N/A // Useful for unrolling loops with NO array accesses.
1008N/A bool policy_peel_only( PhaseIdealLoop *phase ) const;
1008N/A
1008N/A // Return TRUE or FALSE if the loop should be unswitched -- clone
1008N/A // loop with an invariant test
1008N/A bool policy_unswitching( PhaseIdealLoop *phase ) const;
1008N/A
1008N/A // Micro-benchmark spamming. Remove empty loops.
1008N/A bool policy_do_remove_empty_loop( PhaseIdealLoop *phase );
1008N/A
1008N/A // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
1008N/A // make some loop-invariant test (usually a null-check) happen before the
1008N/A // loop.
1418N/A bool policy_peeling( PhaseIdealLoop *phase ) const;
1418N/A
1389N/A // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any
1389N/A // known trip count in the counted loop node.
1008N/A bool policy_maximally_unroll( PhaseIdealLoop *phase ) const;
1008N/A
1008N/A // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if
1008N/A // the loop is a CountedLoop and the body is small enough.
1008N/A bool policy_unroll( PhaseIdealLoop *phase ) const;
1008N/A
1008N/A // Return TRUE or FALSE if the loop should be range-check-eliminated.
1008N/A // Gather a list of IF tests that are dominated by iteration splitting;
1008N/A // also gather the end of the first split and the start of the 2nd split.
1008N/A bool policy_range_check( PhaseIdealLoop *phase ) const;
1008N/A
1008N/A // Return TRUE or FALSE if the loop should be cache-line aligned.
1008N/A // Gather the expression that does the alignment. Note that only
1008N/A // one array base can be aligned in a loop (unless the VM guarantees
1008N/A // mutual alignment). Note that if we vectorize short memory ops
1008N/A // into longer memory ops, we may want to increase alignment.
1008N/A bool policy_align( PhaseIdealLoop *phase ) const;
1008N/A
1008N/A // Return TRUE if "iff" is a range check.
1008N/A bool is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const;
1008N/A
1008N/A // Compute loop trip count from profile data
1008N/A void compute_profile_trip_cnt( PhaseIdealLoop *phase );
1008N/A
1008N/A // Reassociate invariant expressions.
1008N/A void reassociate_invariants(PhaseIdealLoop *phase);
1008N/A // Reassociate invariant add and subtract expressions.
1008N/A Node* reassociate_add_sub(Node* n1, PhaseIdealLoop *phase);
1008N/A // Return nonzero index of invariant operand if invariant and variant
1008N/A // are combined with an Add or Sub. Helper for reassociate_invariants.
1008N/A int is_invariant_addition(Node* n, PhaseIdealLoop *phase);
1008N/A
1008N/A // Return true if n is invariant
1008N/A bool is_invariant(Node* n) const;
1008N/A
1008N/A // Put loop body on igvn work list
1008N/A void record_for_igvn();
1008N/A
1008N/A bool is_loop() { return !_irreducible && _tail && !_tail->is_top(); }
1008N/A bool is_inner() { return is_loop() && _child == NULL; }
1089N/A bool is_counted() { return is_loop() && _head != NULL && _head->is_CountedLoop(); }
1008N/A
1008N/A#ifndef PRODUCT
1008N/A void dump_head( ) const; // Dump loop head only
1008N/A void dump() const; // Dump this loop recursively
1008N/A void verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const;
1008N/A#endif
1008N/A
1008N/A};
1008N/A
1008N/A// -----------------------------PhaseIdealLoop---------------------------------
1008N/A// Computes the mapping from Nodes to IdealLoopTrees. Organizes IdealLoopTrees into a
1008N/A// loop tree. Drives the loop-based transformations on the ideal graph.
1008N/Aclass PhaseIdealLoop : public PhaseTransform {
1008N/A friend class IdealLoopTree;
1008N/A friend class SuperWord;
1008N/A // Pre-computed def-use info
1008N/A PhaseIterGVN &_igvn;
1008N/A
1008N/A // Head of loop tree
1008N/A IdealLoopTree *_ltree_root;
1008N/A
1008N/A // Array of pre-order numbers, plus post-visited bit.
1008N/A // ZERO for not pre-visited. EVEN for pre-visited but not post-visited.
1008N/A // ODD for post-visited. Other bits are the pre-order number.
1008N/A uint *_preorders;
1008N/A uint _max_preorder;
1008N/A
1008N/A const PhaseIdealLoop* _verify_me;
1008N/A bool _verify_only;
1008N/A
1008N/A // Allocate _preorders[] array
1008N/A void allocate_preorders() {
1008N/A _max_preorder = C->unique()+8;
1008N/A _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder);
1008N/A memset(_preorders, 0, sizeof(uint) * _max_preorder);
1008N/A }
1008N/A
1008N/A // Allocate _preorders[] array
1008N/A void reallocate_preorders() {
1008N/A if ( _max_preorder < C->unique() ) {
1008N/A _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique());
1008N/A _max_preorder = C->unique();
1008N/A }
1008N/A memset(_preorders, 0, sizeof(uint) * _max_preorder);
1008N/A }
1008N/A
1008N/A // Check to grow _preorders[] array for the case when build_loop_tree_impl()
1008N/A // adds new nodes.
1008N/A void check_grow_preorders( ) {
1008N/A if ( _max_preorder < C->unique() ) {
1008N/A uint newsize = _max_preorder<<1; // double size of array
1008N/A _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize);
1008N/A memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder));
1008N/A _max_preorder = newsize;
1008N/A }
1008N/A }
1008N/A // Check for pre-visited. Zero for NOT visited; non-zero for visited.
1008N/A int is_visited( Node *n ) const { return _preorders[n->_idx]; }
1008N/A // Pre-order numbers are written to the Nodes array as low-bit-set values.
1008N/A void set_preorder_visited( Node *n, int pre_order ) {
1008N/A assert( !is_visited( n ), "already set" );
1008N/A _preorders[n->_idx] = (pre_order<<1);
1008N/A };
1008N/A // Return pre-order number.
1008N/A int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; }
1008N/A
1008N/A // Check for being post-visited.
1008N/A // Should be previsited already (checked with assert(is_visited(n))).
1008N/A int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; }
1008N/A
1008N/A // Mark as post visited
1008N/A void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; }
1008N/A
1008N/A // Set/get control node out. Set lower bit to distinguish from IdealLoopTree
1008N/A // Returns true if "n" is a data node, false if it's a control node.
1008N/A bool has_ctrl( Node *n ) const { return ((intptr_t)_nodes[n->_idx]) & 1; }
1008N/A
1008N/A // clear out dead code after build_loop_late
1008N/A Node_List _deadlist;
1008N/A
1008N/A // Support for faster execution of get_late_ctrl()/dom_lca()
1008N/A // when a node has many uses and dominator depth is deep.
1008N/A Node_Array _dom_lca_tags;
1008N/A void init_dom_lca_tags();
1008N/A void clear_dom_lca_tags();
1008N/A
1008N/A // Helper for debugging bad dominance relationships
1008N/A bool verify_dominance(Node* n, Node* use, Node* LCA, Node* early);
1008N/A
1008N/A Node* compute_lca_of_uses(Node* n, Node* early, bool verify = false);
1008N/A
1008N/A // Inline wrapper for frequent cases:
1008N/A // 1) only one use
1008N/A // 2) a use is the same as the current LCA passed as 'n1'
1008N/A Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) {
1008N/A assert( n->is_CFG(), "" );
1008N/A // Fast-path NULL lca
1008N/A if( lca != NULL && lca != n ) {
1008N/A assert( lca->is_CFG(), "" );
1008N/A // find LCA of all uses
1008N/A n = dom_lca_for_get_late_ctrl_internal( lca, n, tag );
1008N/A }
1008N/A return find_non_split_ctrl(n);
1008N/A }
1008N/A Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag );
1008N/A
1008N/A // Helper function for directing control inputs away from CFG split
1008N/A // points.
1008N/A Node *find_non_split_ctrl( Node *ctrl ) const {
1008N/A if (ctrl != NULL) {
1008N/A if (ctrl->is_MultiBranch()) {
1008N/A ctrl = ctrl->in(0);
1008N/A }
1008N/A assert(ctrl->is_CFG(), "CFG");
1008N/A }
1008N/A return ctrl;
1008N/A }
1008N/A
1008N/Apublic:
1008N/A bool has_node( Node* n ) const { return _nodes[n->_idx] != NULL; }
1008N/A // check if transform created new nodes that need _ctrl recorded
1008N/A Node *get_late_ctrl( Node *n, Node *early );
1008N/A Node *get_early_ctrl( Node *n );
1008N/A void set_early_ctrl( Node *n );
1008N/A void set_subtree_ctrl( Node *root );
1008N/A void set_ctrl( Node *n, Node *ctrl ) {
1008N/A assert( !has_node(n) || has_ctrl(n), "" );
1008N/A assert( ctrl->in(0), "cannot set dead control node" );
1008N/A assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" );
1008N/A _nodes.map( n->_idx, (Node*)((intptr_t)ctrl + 1) );
1140N/A }
1008N/A // Set control and update loop membership
1008N/A void set_ctrl_and_loop(Node* n, Node* ctrl) {
1008N/A IdealLoopTree* old_loop = get_loop(get_ctrl(n));
1140N/A IdealLoopTree* new_loop = get_loop(ctrl);
1008N/A if (old_loop != new_loop) {
1008N/A if (old_loop->_child == NULL) old_loop->_body.yank(n);
1418N/A if (new_loop->_child == NULL) new_loop->_body.push(n);
1008N/A }
1008N/A set_ctrl(n, ctrl);
1008N/A }
1418N/A // Control nodes can be replaced or subsumed. During this pass they
1008N/A // get their replacement Node in slot 1. Instead of updating the block
1008N/A // location of all Nodes in the subsumed block, we lazily do it. As we
1008N/A // pull such a subsumed block out of the array, we write back the final
1418N/A // correct block.
1418N/A Node *get_ctrl( Node *i ) {
1418N/A assert(has_node(i), "");
1418N/A Node *n = get_ctrl_no_update(i);
1008N/A _nodes.map( i->_idx, (Node*)((intptr_t)n + 1) );
1008N/A assert(has_node(i) && has_ctrl(i), "");
1008N/A assert(n == find_non_split_ctrl(n), "must return legal ctrl" );
1008N/A return n;
1008N/A }
1418N/A // true if CFG node d dominates CFG node n
1008N/A bool is_dominator(Node *d, Node *n);
1008N/A // return get_ctrl for a data node and self(n) for a CFG node
1008N/A Node* ctrl_or_self(Node* n) {
1418N/A if (has_ctrl(n))
1008N/A return get_ctrl(n);
1008N/A else {
1008N/A assert (n->is_CFG(), "must be a CFG node");
1008N/A return n;
1418N/A }
1008N/A }
1008N/A
1008N/Aprivate:
1008N/A Node *get_ctrl_no_update( Node *i ) const {
1008N/A assert( has_ctrl(i), "" );
1008N/A Node *n = (Node*)(((intptr_t)_nodes[i->_idx]) & ~1);
1008N/A if (!n->in(0)) {
1008N/A // Skip dead CFG nodes
1008N/A do {
1008N/A n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1);
1008N/A } while (!n->in(0));
1008N/A n = find_non_split_ctrl(n);
1008N/A }
1008N/A return n;
1008N/A }
1008N/A
1008N/A // Check for loop being set
1008N/A // "n" must be a control node. Returns true if "n" is known to be in a loop.
1140N/A bool has_loop( Node *n ) const {
1008N/A assert(!has_node(n) || !has_ctrl(n), "");
1008N/A return has_node(n);
1008N/A }
1008N/A // Set loop
1008N/A void set_loop( Node *n, IdealLoopTree *loop ) {
1008N/A _nodes.map(n->_idx, (Node*)loop);
1008N/A }
1008N/A // Lazy-dazy update of 'get_ctrl' and 'idom_at' mechanisms. Replace
1008N/A // the 'old_node' with 'new_node'. Kill old-node. Add a reference
1008N/A // from old_node to new_node to support the lazy update. Reference
1008N/A // replaces loop reference, since that is not needed for dead node.
1008N/Apublic:
1008N/A void lazy_update( Node *old_node, Node *new_node ) {
1008N/A assert( old_node != new_node, "no cycles please" );
1008N/A //old_node->set_req( 1, new_node /*NO DU INFO*/ );
1008N/A // Nodes always have DU info now, so re-use the side array slot
1008N/A // for this node to provide the forwarding pointer.
1008N/A _nodes.map( old_node->_idx, (Node*)((intptr_t)new_node + 1) );
1008N/A }
1418N/A void lazy_replace( Node *old_node, Node *new_node ) {
1418N/A _igvn.replace_node( old_node, new_node );
1418N/A lazy_update( old_node, new_node );
1418N/A }
1418N/A void lazy_replace_proj( Node *old_node, Node *new_node ) {
1418N/A assert( old_node->req() == 1, "use this for Projs" );
1418N/A _igvn.hash_delete(old_node); // Must hash-delete before hacking edges
1418N/A old_node->add_req( NULL );
1418N/A lazy_replace( old_node, new_node );
1418N/A }
1418N/A
1418N/Aprivate:
1418N/A
1418N/A // Place 'n' in some loop nest, where 'n' is a CFG node
1418N/A void build_loop_tree();
1418N/A int build_loop_tree_impl( Node *n, int pre_order );
1418N/A // Insert loop into the existing loop tree. 'innermost' is a leaf of the
1008N/A // loop tree, not the root.
1008N/A IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost );
1008N/A
1418N/A // Place Data nodes in some loop nest
1418N/A void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
1418N/A void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
1418N/A void build_loop_late_post ( Node* n );
1418N/A
1418N/A // Array of immediate dominance info for each CFG node indexed by node idx
1418N/Aprivate:
1418N/A uint _idom_size;
1418N/A Node **_idom; // Array of immediate dominators
1418N/A uint *_dom_depth; // Used for fast LCA test
1418N/A GrowableArray<uint>* _dom_stk; // For recomputation of dom depth
1008N/A
1008N/A Node* idom_no_update(Node* d) const {
1008N/A assert(d->_idx < _idom_size, "oob");
1008N/A Node* n = _idom[d->_idx];
1008N/A assert(n != NULL,"Bad immediate dominator info.");
1008N/A while (n->in(0) == NULL) { // Skip dead CFG nodes
1008N/A //n = n->in(1);
1008N/A n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1);
1008N/A assert(n != NULL,"Bad immediate dominator info.");
1008N/A }
1008N/A return n;
1008N/A }
1008N/A Node *idom(Node* d) const {
1008N/A uint didx = d->_idx;
1008N/A Node *n = idom_no_update(d);
1008N/A _idom[didx] = n; // Lazily remove dead CFG nodes from table.
1008N/A return n;
1008N/A }
1008N/A uint dom_depth(Node* d) const {
1008N/A assert(d->_idx < _idom_size, "");
1008N/A return _dom_depth[d->_idx];
1008N/A }
1008N/A void set_idom(Node* d, Node* n, uint dom_depth);
1008N/A // Locally compute IDOM using dom_lca call
1008N/A Node *compute_idom( Node *region ) const;
1008N/A // Recompute dom_depth
1008N/A void recompute_dom_depth();
1008N/A
1008N/A // Is safept not required by an outer loop?
1008N/A bool is_deleteable_safept(Node* sfpt);
1008N/A
1008N/A // Replace parallel induction variable (parallel to trip counter)
1008N/A void replace_parallel_iv(IdealLoopTree *loop);
1008N/A
1008N/A // Perform verification that the graph is valid.
1008N/A PhaseIdealLoop( PhaseIterGVN &igvn) :
1008N/A PhaseTransform(Ideal_Loop),
1008N/A _igvn(igvn),
1008N/A _dom_lca_tags(arena()), // Thread::resource_area
1008N/A _verify_me(NULL),
1008N/A _verify_only(true) {
1008N/A build_and_optimize(false, false);
1008N/A }
1008N/A
1008N/A // build the loop tree and perform any requested optimizations
1008N/A void build_and_optimize(bool do_split_if, bool do_loop_pred);
1008N/A
1008N/Apublic:
1008N/A // Dominators for the sea of nodes
1008N/A void Dominators();
1008N/A Node *dom_lca( Node *n1, Node *n2 ) const {
1008N/A return find_non_split_ctrl(dom_lca_internal(n1, n2));
1008N/A }
1008N/A Node *dom_lca_internal( Node *n1, Node *n2 ) const;
1008N/A
1008N/A // Compute the Ideal Node to Loop mapping
1008N/A PhaseIdealLoop( PhaseIterGVN &igvn, bool do_split_ifs, bool do_loop_pred) :
1008N/A PhaseTransform(Ideal_Loop),
1008N/A _igvn(igvn),
1008N/A _dom_lca_tags(arena()), // Thread::resource_area
1008N/A _verify_me(NULL),
1008N/A _verify_only(false) {
1008N/A build_and_optimize(do_split_ifs, do_loop_pred);
1008N/A }
1008N/A
1008N/A // Verify that verify_me made the same decisions as a fresh run.
1008N/A PhaseIdealLoop( PhaseIterGVN &igvn, const PhaseIdealLoop *verify_me) :
1008N/A PhaseTransform(Ideal_Loop),
1008N/A _igvn(igvn),
1008N/A _dom_lca_tags(arena()), // Thread::resource_area
1008N/A _verify_me(verify_me),
1008N/A _verify_only(false) {
1008N/A build_and_optimize(false, false);
1008N/A }
1008N/A
1008N/A // Build and verify the loop tree without modifying the graph. This
1008N/A // is useful to verify that all inputs properly dominate their uses.
1008N/A static void verify(PhaseIterGVN& igvn) {
1008N/A#ifdef ASSERT
1008N/A PhaseIdealLoop v(igvn);
1008N/A#endif
1008N/A }
1008N/A
1008N/A // True if the method has at least 1 irreducible loop
1008N/A bool _has_irreducible_loops;
1008N/A
1008N/A // Per-Node transform
1008N/A virtual Node *transform( Node *a_node ) { return 0; }
1008N/A
1008N/A bool is_counted_loop( Node *x, IdealLoopTree *loop );
1008N/A
1008N/A // Return a post-walked LoopNode
1008N/A IdealLoopTree *get_loop( Node *n ) const {
1008N/A // Dead nodes have no loop, so return the top level loop instead
1008N/A if (!has_node(n)) return _ltree_root;
1008N/A assert(!has_ctrl(n), "");
1008N/A return (IdealLoopTree*)_nodes[n->_idx];
1008N/A }
1008N/A
1008N/A // Is 'n' a (nested) member of 'loop'?
1008N/A int is_member( const IdealLoopTree *loop, Node *n ) const {
1008N/A return loop->is_member(get_loop(n)); }
1008N/A
1008N/A // This is the basic building block of the loop optimizations. It clones an
1008N/A // entire loop body. It makes an old_new loop body mapping; with this
1008N/A // mapping you can find the new-loop equivalent to an old-loop node. All
1008N/A // new-loop nodes are exactly equal to their old-loop counterparts, all
1008N/A // edges are the same. All exits from the old-loop now have a RegionNode
1008N/A // that merges the equivalent new-loop path. This is true even for the
1008N/A // normal "loop-exit" condition. All uses of loop-invariant old-loop values
1008N/A // now come from (one or more) Phis that merge their new-loop equivalents.
1008N/A // Parameter side_by_side_idom:
1008N/A // When side_by_size_idom is NULL, the dominator tree is constructed for
1008N/A // the clone loop to dominate the original. Used in construction of
1008N/A // pre-main-post loop sequence.
1008N/A // When nonnull, the clone and original are side-by-side, both are
1008N/A // dominated by the passed in side_by_side_idom node. Used in
1008N/A // construction of unswitched loops.
1008N/A void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth,
1008N/A Node* side_by_side_idom = NULL);
1008N/A
1008N/A // If we got the effect of peeling, either by actually peeling or by
1008N/A // making a pre-loop which must execute at least once, we can remove
1008N/A // all loop-invariant dominated tests in the main body.
1008N/A void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new );
1008N/A
1008N/A // Generate code to do a loop peel for the given loop (and body).
1008N/A // old_new is a temp array.
1008N/A void do_peeling( IdealLoopTree *loop, Node_List &old_new );
1008N/A
1008N/A // Add pre and post loops around the given loop. These loops are used
1008N/A // during RCE, unrolling and aligning loops.
1008N/A void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only );
1008N/A // If Node n lives in the back_ctrl block, we clone a private version of n
1008N/A // in preheader_ctrl block and return that, otherwise return n.
1008N/A Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n );
1008N/A
1008N/A // Take steps to maximally unroll the loop. Peel any odd iterations, then
1008N/A // unroll to do double iterations. The next round of major loop transforms
1008N/A // will repeat till the doubled loop body does all remaining iterations in 1
1008N/A // pass.
1008N/A void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new );
1008N/A
1008N/A // Unroll the loop body one step - make each trip do 2 iterations.
1008N/A void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip );
1008N/A
1008N/A // Return true if exp is a constant times an induction var
1008N/A bool is_scaled_iv(Node* exp, Node* iv, int* p_scale);
1008N/A
1008N/A // Return true if exp is a scaled induction var plus (or minus) constant
1008N/A bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth = 0);
1008N/A
1008N/A // Return true if proj is for "proj->[region->..]call_uct"
1008N/A // Return true if proj is for "proj->[region->..]call_uct"
1008N/A static bool is_uncommon_trap_proj(ProjNode* proj, Deoptimization::DeoptReason reason);
1008N/A // Return true for "if(test)-> proj -> ...
1008N/A // |
1008N/A // V
1008N/A // other_proj->[region->..]call_uct"
1008N/A static bool is_uncommon_trap_if_pattern(ProjNode* proj, Deoptimization::DeoptReason reason);
1008N/A // Create a new if above the uncommon_trap_if_pattern for the predicate to be promoted
1008N/A ProjNode* create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry,
1008N/A Deoptimization::DeoptReason reason);
1008N/A void register_control(Node* n, IdealLoopTree *loop, Node* pred);
1008N/A
1008N/A // Find a good location to insert a predicate
1008N/A static ProjNode* find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason);
1008N/A // Find a predicate
1008N/A static Node* find_predicate(Node* entry);
1008N/A // Construct a range check for a predicate if
1008N/A BoolNode* rc_predicate(Node* ctrl,
1008N/A int scale, Node* offset,
1008N/A Node* init, Node* limit, Node* stride,
1008N/A Node* range, bool upper);
1008N/A
1008N/A // Implementation of the loop predication to promote checks outside the loop
1008N/A bool loop_predication_impl(IdealLoopTree *loop);
1008N/A
1008N/A // Helper function to collect predicate for eliminating the useless ones
1008N/A void collect_potentially_useful_predicates(IdealLoopTree *loop, Unique_Node_List &predicate_opaque1);
1008N/A void eliminate_useless_predicates();
1008N/A
1008N/A // Eliminate range-checks and other trip-counter vs loop-invariant tests.
1008N/A void do_range_check( IdealLoopTree *loop, Node_List &old_new );
1008N/A
1008N/A // Create a slow version of the loop by cloning the loop
1008N/A // and inserting an if to select fast-slow versions.
1008N/A ProjNode* create_slow_version_of_loop(IdealLoopTree *loop,
1008N/A Node_List &old_new);
1008N/A
1008N/A // Clone loop with an invariant test (that does not exit) and
1008N/A // insert a clone of the test that selects which version to
1008N/A // execute.
1008N/A void do_unswitching (IdealLoopTree *loop, Node_List &old_new);
1008N/A
1008N/A // Find candidate "if" for unswitching
1008N/A IfNode* find_unswitching_candidate(const IdealLoopTree *loop) const;
1008N/A
1008N/A // Range Check Elimination uses this function!
1008N/A // Constrain the main loop iterations so the affine function:
1008N/A // scale_con * I + offset < limit
1008N/A // always holds true. That is, either increase the number of iterations in
1008N/A // the pre-loop or the post-loop until the condition holds true in the main
1008N/A // loop. Scale_con, offset and limit are all loop invariant.
1008N/A void add_constraint( int stride_con, int scale_con, Node *offset, Node *limit, Node *pre_ctrl, Node **pre_limit, Node **main_limit );
1008N/A
1008N/A // Partially peel loop up through last_peel node.
1008N/A bool partial_peel( IdealLoopTree *loop, Node_List &old_new );
1008N/A
1008N/A // Create a scheduled list of nodes control dependent on ctrl set.
1008N/A void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched );
1008N/A // Has a use in the vector set
1008N/A bool has_use_in_set( Node* n, VectorSet& vset );
1008N/A // Has use internal to the vector set (ie. not in a phi at the loop head)
1008N/A bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop );
1008N/A // clone "n" for uses that are outside of loop
1008N/A void clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist );
1008N/A // clone "n" for special uses that are in the not_peeled region
1008N/A void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
1008N/A VectorSet& not_peel, Node_List& sink_list, Node_List& worklist );
1008N/A // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
1008N/A void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp );
1008N/A#ifdef ASSERT
1008N/A // Validate the loop partition sets: peel and not_peel
1008N/A bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel );
1008N/A // Ensure that uses outside of loop are of the right form
1008N/A bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
1008N/A uint orig_exit_idx, uint clone_exit_idx);
1008N/A bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx);
1008N/A#endif
1008N/A
1008N/A // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.)
1008N/A int stride_of_possible_iv( Node* iff );
1008N/A bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; }
1008N/A // Return the (unique) control output node that's in the loop (if it exists.)
1008N/A Node* stay_in_loop( Node* n, IdealLoopTree *loop);
1008N/A // Insert a signed compare loop exit cloned from an unsigned compare.
1008N/A IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop);
1008N/A void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop);
1008N/A // Utility to register node "n" with PhaseIdealLoop
1008N/A void register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth);
1008N/A // Utility to create an if-projection
1008N/A ProjNode* proj_clone(ProjNode* p, IfNode* iff);
1008N/A // Force the iff control output to be the live_proj
1008N/A Node* short_circuit_if(IfNode* iff, ProjNode* live_proj);
1008N/A // Insert a region before an if projection
1008N/A RegionNode* insert_region_before_proj(ProjNode* proj);
1008N/A // Insert a new if before an if projection
1008N/A ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj);
1008N/A
1008N/A // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1008N/A // "Nearly" because all Nodes have been cloned from the original in the loop,
1008N/A // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1008N/A // through the Phi recursively, and return a Bool.
1008N/A BoolNode *clone_iff( PhiNode *phi, IdealLoopTree *loop );
1008N/A CmpNode *clone_bool( PhiNode *phi, IdealLoopTree *loop );
1008N/A
1008N/A
1008N/A // Rework addressing expressions to get the most loop-invariant stuff
1008N/A // moved out. We'd like to do all associative operators, but it's especially
1008N/A // important (common) to do address expressions.
1008N/A Node *remix_address_expressions( Node *n );
1008N/A
1008N/A // Attempt to use a conditional move instead of a phi/branch
1008N/A Node *conditional_move( Node *n );
1008N/A
1008N/A // Reorganize offset computations to lower register pressure.
1008N/A // Mostly prevent loop-fallout uses of the pre-incremented trip counter
1008N/A // (which are then alive with the post-incremented trip counter
1008N/A // forcing an extra register move)
1008N/A void reorg_offsets( IdealLoopTree *loop );
1008N/A
1008N/A // Check for aggressive application of 'split-if' optimization,
1008N/A // using basic block level info.
1008N/A void split_if_with_blocks ( VectorSet &visited, Node_Stack &nstack );
1008N/A Node *split_if_with_blocks_pre ( Node *n );
1008N/A void split_if_with_blocks_post( Node *n );
1008N/A Node *has_local_phi_input( Node *n );
1008N/A // Mark an IfNode as being dominated by a prior test,
1008N/A // without actually altering the CFG (and hence IDOM info).
1008N/A void dominated_by( Node *prevdom, Node *iff, bool flip = false );
1008N/A
1008N/A // Split Node 'n' through merge point
1008N/A Node *split_thru_region( Node *n, Node *region );
1008N/A // Split Node 'n' through merge point if there is enough win.
1008N/A Node *split_thru_phi( Node *n, Node *region, int policy );
1008N/A // Found an If getting its condition-code input from a Phi in the
1008N/A // same block. Split thru the Region.
1008N/A void do_split_if( Node *iff );
1008N/A
1008N/A // Conversion of fill/copy patterns into intrisic versions
1008N/A bool do_intrinsify_fill();
1008N/A bool intrinsify_fill(IdealLoopTree* lpt);
1008N/A bool match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
1008N/A Node*& shift, Node*& offset);
1008N/A
1008N/Aprivate:
1008N/A // Return a type based on condition control flow
1008N/A const TypeInt* filtered_type( Node *n, Node* n_ctrl);
1008N/A const TypeInt* filtered_type( Node *n ) { return filtered_type(n, NULL); }
1008N/A // Helpers for filtered type
1008N/A const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl);
1008N/A
1008N/A // Helper functions
1008N/A Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache );
1008N/A Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true );
1008N/A void handle_use( Node *use, Node *def, small_cache *cache, Node *region_dom, Node *new_false, Node *new_true, Node *old_false, Node *old_true );
1008N/A bool split_up( Node *n, Node *blk1, Node *blk2 );
1008N/A void sink_use( Node *use, Node *post_loop );
1008N/A Node *place_near_use( Node *useblock ) const;
1008N/A
1008N/A bool _created_loop_node;
1008N/Apublic:
1008N/A void set_created_loop_node() { _created_loop_node = true; }
1008N/A bool created_loop_node() { return _created_loop_node; }
1008N/A void register_new_node( Node *n, Node *blk );
1008N/A
1008N/A#ifndef PRODUCT
1089N/A void dump( ) const;
1008N/A void dump( IdealLoopTree *loop, uint rpo_idx, Node_List &rpo_list ) const;
1008N/A void rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const;
1008N/A void verify() const; // Major slow :-)
1008N/A void verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const;
1008N/A IdealLoopTree *get_loop_idx(Node* n) const {
1008N/A // Dead nodes have no loop, so return the top level loop instead
1008N/A return _nodes[n->_idx] ? (IdealLoopTree*)_nodes[n->_idx] : _ltree_root;
1089N/A }
1008N/A // Print some stats
1008N/A static void print_statistics();
1008N/A static int _loop_invokes; // Count of PhaseIdealLoop invokes
1008N/A static int _loop_work; // Sum of PhaseIdealLoop x _unique
1008N/A#endif
1008N/A};
1008N/A
1008N/Ainline Node* IdealLoopTree::tail() {
1008N/A// Handle lazy update of _tail field
1008N/A Node *n = _tail;
1008N/A //while( !n->in(0) ) // Skip dead CFG nodes
1008N/A //n = n->in(1);
1008N/A if (n->in(0) == NULL)
1008N/A n = _phase->get_ctrl(n);
1008N/A _tail = n;
1008N/A return n;
1008N/A}
1008N/A
1008N/A
1008N/A// Iterate over the loop tree using a preorder, left-to-right traversal.
1008N/A//
1008N/A// Example that visits all counted loops from within PhaseIdealLoop
1008N/A//
1008N/A// for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
1008N/A// IdealLoopTree* lpt = iter.current();
1008N/A// if (!lpt->is_counted()) continue;
1008N/A// ...
1008N/Aclass LoopTreeIterator : public StackObj {
1008N/Aprivate:
1008N/A IdealLoopTree* _root;
1008N/A IdealLoopTree* _curnt;
1008N/A
1008N/Apublic:
1008N/A LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {}
1008N/A
1008N/A bool done() { return _curnt == NULL; } // Finished iterating?
1008N/A
1008N/A void next(); // Advance to next loop tree
1008N/A
1008N/A IdealLoopTree* current() { return _curnt; } // Return current value of iterator.
1008N/A};
1008N/A
1008N/A#endif // SHARE_VM_OPTO_LOOPNODE_HPP
1008N/A