duke@435: /* kvn@2555: * Copyright (c) 1998, 2011, Oracle and/or its affiliates. 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: * trims@1907: * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA trims@1907: * or visit www.oracle.com if you need additional information or have any trims@1907: * questions. duke@435: * duke@435: */ duke@435: stefank@2314: #ifndef SHARE_VM_OPTO_LOOPNODE_HPP stefank@2314: #define SHARE_VM_OPTO_LOOPNODE_HPP stefank@2314: stefank@2314: #include "opto/cfgnode.hpp" stefank@2314: #include "opto/multnode.hpp" stefank@2314: #include "opto/phaseX.hpp" stefank@2314: #include "opto/subnode.hpp" stefank@2314: #include "opto/type.hpp" stefank@2314: duke@435: class CmpNode; duke@435: class CountedLoopEndNode; duke@435: class CountedLoopNode; duke@435: class IdealLoopTree; duke@435: class LoopNode; duke@435: class Node; duke@435: class PhaseIdealLoop; duke@435: class VectorSet; cfang@1607: class Invariance; duke@435: struct small_cache; duke@435: duke@435: // duke@435: // I D E A L I Z E D L O O P S duke@435: // duke@435: // Idealized loops are the set of loops I perform more interesting duke@435: // transformations on, beyond simple hoisting. duke@435: duke@435: //------------------------------LoopNode--------------------------------------- duke@435: // Simple loop header. Fall in path on left, loop-back path on right. duke@435: class LoopNode : public RegionNode { duke@435: // Size is bigger to hold the flags. However, the flags do not change duke@435: // the semantics so it does not appear in the hash & cmp functions. duke@435: virtual uint size_of() const { return sizeof(*this); } duke@435: protected: duke@435: short _loop_flags; duke@435: // Names for flag bitfields kvn@2747: enum { Normal=0, Pre=1, Main=2, Post=3, PreMainPostFlagsMask=3, kvn@2747: MainHasNoPreLoop=4, kvn@2747: HasExactTripCount=8, kvn@2747: InnerLoop=16, kvn@2747: PartialPeelLoop=32, kvn@2747: PartialPeelFailed=64 }; duke@435: char _unswitch_count; duke@435: enum { _unswitch_max=3 }; duke@435: duke@435: public: duke@435: // Names for edge indices duke@435: enum { Self=0, EntryControl, LoopBackControl }; duke@435: kvn@2747: int is_inner_loop() const { return _loop_flags & InnerLoop; } kvn@2747: void set_inner_loop() { _loop_flags |= InnerLoop; } duke@435: kvn@2747: int is_partial_peel_loop() const { return _loop_flags & PartialPeelLoop; } kvn@2747: void set_partial_peel_loop() { _loop_flags |= PartialPeelLoop; } kvn@2747: int partial_peel_has_failed() const { return _loop_flags & PartialPeelFailed; } kvn@2747: void mark_partial_peel_failed() { _loop_flags |= PartialPeelFailed; } duke@435: duke@435: int unswitch_max() { return _unswitch_max; } duke@435: int unswitch_count() { return _unswitch_count; } duke@435: void set_unswitch_count(int val) { duke@435: assert (val <= unswitch_max(), "too many unswitches"); duke@435: _unswitch_count = val; duke@435: } duke@435: duke@435: LoopNode( Node *entry, Node *backedge ) : RegionNode(3), _loop_flags(0), _unswitch_count(0) { duke@435: init_class_id(Class_Loop); duke@435: init_req(EntryControl, entry); duke@435: init_req(LoopBackControl, backedge); duke@435: } duke@435: duke@435: virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); duke@435: virtual int Opcode() const; duke@435: bool can_be_counted_loop(PhaseTransform* phase) const { duke@435: return req() == 3 && in(0) != NULL && duke@435: in(1) != NULL && phase->type(in(1)) != Type::TOP && duke@435: in(2) != NULL && phase->type(in(2)) != Type::TOP; duke@435: } kvn@2665: bool is_valid_counted_loop() const; duke@435: #ifndef PRODUCT duke@435: virtual void dump_spec(outputStream *st) const; duke@435: #endif duke@435: }; duke@435: duke@435: //------------------------------Counted Loops---------------------------------- duke@435: // Counted loops are all trip-counted loops, with exactly 1 trip-counter exit duke@435: // path (and maybe some other exit paths). The trip-counter exit is always kvn@2665: // last in the loop. The trip-counter have to stride by a constant; kvn@2665: // the exit value is also loop invariant. duke@435: duke@435: // CountedLoopNodes and CountedLoopEndNodes come in matched pairs. The duke@435: // CountedLoopNode has the incoming loop control and the loop-back-control duke@435: // which is always the IfTrue before the matching CountedLoopEndNode. The duke@435: // CountedLoopEndNode has an incoming control (possibly not the duke@435: // CountedLoopNode if there is control flow in the loop), the post-increment duke@435: // trip-counter value, and the limit. The trip-counter value is always of duke@435: // the form (Op old-trip-counter stride). The old-trip-counter is produced kvn@2665: // by a Phi connected to the CountedLoopNode. The stride is constant. duke@435: // The Op is any commutable opcode, including Add, Mul, Xor. The duke@435: // CountedLoopEndNode also takes in the loop-invariant limit value. duke@435: duke@435: // From a CountedLoopNode I can reach the matching CountedLoopEndNode via the duke@435: // loop-back control. From CountedLoopEndNodes I can reach CountedLoopNodes duke@435: // via the old-trip-counter from the Op node. duke@435: duke@435: //------------------------------CountedLoopNode-------------------------------- duke@435: // CountedLoopNodes head simple counted loops. CountedLoopNodes have as duke@435: // inputs the incoming loop-start control and the loop-back control, so they duke@435: // act like RegionNodes. They also take in the initial trip counter, the duke@435: // loop-invariant stride and the loop-invariant limit value. CountedLoopNodes duke@435: // produce a loop-body control and the trip counter value. Since duke@435: // CountedLoopNodes behave like RegionNodes I still have a standard CFG model. duke@435: duke@435: class CountedLoopNode : public LoopNode { duke@435: // Size is bigger to hold _main_idx. However, _main_idx does not change duke@435: // the semantics so it does not appear in the hash & cmp functions. duke@435: virtual uint size_of() const { return sizeof(*this); } duke@435: duke@435: // For Pre- and Post-loops during debugging ONLY, this holds the index of duke@435: // the Main CountedLoop. Used to assert that we understand the graph shape. duke@435: node_idx_t _main_idx; duke@435: kvn@2747: // Known trip count calculated by compute_exact_trip_count() kvn@2747: uint _trip_count; duke@435: duke@435: // Expected trip count from profile data duke@435: float _profile_trip_cnt; duke@435: duke@435: // Log2 of original loop bodies in unrolled loop duke@435: int _unrolled_count_log2; duke@435: duke@435: // Node count prior to last unrolling - used to decide if duke@435: // unroll,optimize,unroll,optimize,... is making progress duke@435: int _node_count_before_unroll; duke@435: duke@435: public: duke@435: CountedLoopNode( Node *entry, Node *backedge ) kvn@2747: : LoopNode(entry, backedge), _main_idx(0), _trip_count(max_juint), duke@435: _profile_trip_cnt(COUNT_UNKNOWN), _unrolled_count_log2(0), duke@435: _node_count_before_unroll(0) { duke@435: init_class_id(Class_CountedLoop); duke@435: // Initialize _trip_count to the largest possible value. duke@435: // Will be reset (lower) if the loop's trip count is known. duke@435: } duke@435: duke@435: virtual int Opcode() const; duke@435: virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); duke@435: duke@435: Node *init_control() const { return in(EntryControl); } duke@435: Node *back_control() const { return in(LoopBackControl); } duke@435: CountedLoopEndNode *loopexit() const; duke@435: Node *init_trip() const; duke@435: Node *stride() const; duke@435: int stride_con() const; duke@435: bool stride_is_con() const; duke@435: Node *limit() const; duke@435: Node *incr() const; duke@435: Node *phi() const; duke@435: duke@435: // Match increment with optional truncation duke@435: static Node* match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type); duke@435: duke@435: // A 'main' loop has a pre-loop and a post-loop. The 'main' loop duke@435: // can run short a few iterations and may start a few iterations in. duke@435: // It will be RCE'd and unrolled and aligned. duke@435: duke@435: // A following 'post' loop will run any remaining iterations. Used duke@435: // during Range Check Elimination, the 'post' loop will do any final duke@435: // iterations with full checks. Also used by Loop Unrolling, where duke@435: // the 'post' loop will do any epilog iterations needed. Basically, duke@435: // a 'post' loop can not profitably be further unrolled or RCE'd. duke@435: duke@435: // A preceding 'pre' loop will run at least 1 iteration (to do peeling), duke@435: // it may do under-flow checks for RCE and may do alignment iterations duke@435: // so the following main loop 'knows' that it is striding down cache duke@435: // lines. duke@435: duke@435: // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or duke@435: // Aligned, may be missing it's pre-loop. kvn@2747: int is_normal_loop() const { return (_loop_flags&PreMainPostFlagsMask) == Normal; } kvn@2747: int is_pre_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Pre; } kvn@2747: int is_main_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Main; } kvn@2747: int is_post_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Post; } kvn@2747: int is_main_no_pre_loop() const { return _loop_flags & MainHasNoPreLoop; } kvn@2747: void set_main_no_pre_loop() { _loop_flags |= MainHasNoPreLoop; } duke@435: never@802: int main_idx() const { return _main_idx; } never@802: duke@435: duke@435: void set_pre_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; } duke@435: void set_main_loop ( ) { assert(is_normal_loop(),""); _loop_flags |= Main; } duke@435: void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; } kvn@2747: void set_normal_loop( ) { _loop_flags &= ~PreMainPostFlagsMask; } duke@435: kvn@2747: void set_trip_count(uint tc) { _trip_count = tc; } kvn@2747: uint trip_count() { return _trip_count; } kvn@2747: kvn@2747: bool has_exact_trip_count() const { return (_loop_flags & HasExactTripCount) != 0; } kvn@2747: void set_exact_trip_count(uint tc) { kvn@2747: _trip_count = tc; kvn@2747: _loop_flags |= HasExactTripCount; kvn@2747: } kvn@2747: void set_nonexact_trip_count() { kvn@2747: _loop_flags &= ~HasExactTripCount; kvn@2747: } duke@435: duke@435: void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; } duke@435: float profile_trip_cnt() { return _profile_trip_cnt; } duke@435: duke@435: void double_unrolled_count() { _unrolled_count_log2++; } duke@435: int unrolled_count() { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); } duke@435: duke@435: void set_node_count_before_unroll(int ct) { _node_count_before_unroll = ct; } duke@435: int node_count_before_unroll() { return _node_count_before_unroll; } duke@435: duke@435: #ifndef PRODUCT duke@435: virtual void dump_spec(outputStream *st) const; duke@435: #endif duke@435: }; duke@435: duke@435: //------------------------------CountedLoopEndNode----------------------------- duke@435: // CountedLoopEndNodes end simple trip counted loops. They act much like duke@435: // IfNodes. duke@435: class CountedLoopEndNode : public IfNode { duke@435: public: duke@435: enum { TestControl, TestValue }; duke@435: duke@435: CountedLoopEndNode( Node *control, Node *test, float prob, float cnt ) duke@435: : IfNode( control, test, prob, cnt) { duke@435: init_class_id(Class_CountedLoopEnd); duke@435: } duke@435: virtual int Opcode() const; duke@435: duke@435: Node *cmp_node() const { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : NULL; } duke@435: Node *incr() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; } duke@435: Node *limit() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; } duke@435: Node *stride() const { Node *tmp = incr (); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; } duke@435: Node *phi() const { Node *tmp = incr (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; } duke@435: Node *init_trip() const { Node *tmp = phi (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; } duke@435: int stride_con() const; duke@435: bool stride_is_con() const { Node *tmp = stride (); return (tmp != NULL && tmp->is_Con()); } duke@435: BoolTest::mask test_trip() const { return in(TestValue)->as_Bool()->_test._test; } duke@435: CountedLoopNode *loopnode() const { duke@435: Node *ln = phi()->in(0); duke@435: assert( ln->Opcode() == Op_CountedLoop, "malformed loop" ); duke@435: return (CountedLoopNode*)ln; } duke@435: duke@435: #ifndef PRODUCT duke@435: virtual void dump_spec(outputStream *st) const; duke@435: #endif duke@435: }; duke@435: duke@435: duke@435: inline CountedLoopEndNode *CountedLoopNode::loopexit() const { duke@435: Node *bc = back_control(); duke@435: if( bc == NULL ) return NULL; duke@435: Node *le = bc->in(0); duke@435: if( le->Opcode() != Op_CountedLoopEnd ) duke@435: return NULL; duke@435: return (CountedLoopEndNode*)le; duke@435: } duke@435: inline Node *CountedLoopNode::init_trip() const { return loopexit() ? loopexit()->init_trip() : NULL; } duke@435: inline Node *CountedLoopNode::stride() const { return loopexit() ? loopexit()->stride() : NULL; } duke@435: inline int CountedLoopNode::stride_con() const { return loopexit() ? loopexit()->stride_con() : 0; } duke@435: inline bool CountedLoopNode::stride_is_con() const { return loopexit() && loopexit()->stride_is_con(); } duke@435: inline Node *CountedLoopNode::limit() const { return loopexit() ? loopexit()->limit() : NULL; } duke@435: inline Node *CountedLoopNode::incr() const { return loopexit() ? loopexit()->incr() : NULL; } duke@435: inline Node *CountedLoopNode::phi() const { return loopexit() ? loopexit()->phi() : NULL; } duke@435: kvn@2877: //------------------------------LoopLimitNode----------------------------- kvn@2877: // Counted Loop limit node which represents exact final iterator value: kvn@2877: // trip_count = (limit - init_trip + stride - 1)/stride kvn@2877: // final_value= trip_count * stride + init_trip. kvn@2877: // Use HW instructions to calculate it when it can overflow in integer. kvn@2877: // Note, final_value should fit into integer since counted loop has kvn@2877: // limit check: limit <= max_int-stride. kvn@2877: class LoopLimitNode : public Node { kvn@2877: enum { Init=1, Limit=2, Stride=3 }; kvn@2877: public: kvn@2877: LoopLimitNode( Compile* C, Node *init, Node *limit, Node *stride ) : Node(0,init,limit,stride) { kvn@2877: // Put it on the Macro nodes list to optimize during macro nodes expansion. kvn@2877: init_flags(Flag_is_macro); kvn@2877: C->add_macro_node(this); kvn@2877: } kvn@2877: virtual int Opcode() const; kvn@2877: virtual const Type *bottom_type() const { return TypeInt::INT; } kvn@2877: virtual uint ideal_reg() const { return Op_RegI; } kvn@2877: virtual const Type *Value( PhaseTransform *phase ) const; kvn@2877: virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); kvn@2877: virtual Node *Identity( PhaseTransform *phase ); kvn@2877: }; duke@435: duke@435: // -----------------------------IdealLoopTree---------------------------------- duke@435: class IdealLoopTree : public ResourceObj { duke@435: public: duke@435: IdealLoopTree *_parent; // Parent in loop tree duke@435: IdealLoopTree *_next; // Next sibling in loop tree duke@435: IdealLoopTree *_child; // First child in loop tree duke@435: duke@435: // The head-tail backedge defines the loop. duke@435: // If tail is NULL then this loop has multiple backedges as part of the duke@435: // same loop. During cleanup I'll peel off the multiple backedges; merge duke@435: // them at the loop bottom and flow 1 real backedge into the loop. duke@435: Node *_head; // Head of loop duke@435: Node *_tail; // Tail of loop duke@435: inline Node *tail(); // Handle lazy update of _tail field duke@435: PhaseIdealLoop* _phase; duke@435: duke@435: Node_List _body; // Loop body for inner loops duke@435: duke@435: uint8 _nest; // Nesting depth duke@435: uint8 _irreducible:1, // True if irreducible duke@435: _has_call:1, // True if has call safepoint duke@435: _has_sfpt:1, // True if has non-call safepoint duke@435: _rce_candidate:1; // True if candidate for range check elimination duke@435: kvn@474: Node_List* _required_safept; // A inner loop cannot delete these safepts; kvn@474: bool _allow_optimizations; // Allow loop optimizations duke@435: duke@435: IdealLoopTree( PhaseIdealLoop* phase, Node *head, Node *tail ) duke@435: : _parent(0), _next(0), _child(0), duke@435: _head(head), _tail(tail), duke@435: _phase(phase), duke@435: _required_safept(NULL), kvn@474: _allow_optimizations(true), duke@435: _nest(0), _irreducible(0), _has_call(0), _has_sfpt(0), _rce_candidate(0) duke@435: { } duke@435: duke@435: // Is 'l' a member of 'this'? duke@435: int is_member( const IdealLoopTree *l ) const; // Test for nested membership duke@435: duke@435: // Set loop nesting depth. Accumulate has_call bits. duke@435: int set_nest( uint depth ); duke@435: duke@435: // Split out multiple fall-in edges from the loop header. Move them to a duke@435: // private RegionNode before the loop. This becomes the loop landing pad. duke@435: void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ); duke@435: duke@435: // Split out the outermost loop from this shared header. duke@435: void split_outer_loop( PhaseIdealLoop *phase ); duke@435: duke@435: // Merge all the backedges from the shared header into a private Region. duke@435: // Feed that region as the one backedge to this loop. duke@435: void merge_many_backedges( PhaseIdealLoop *phase ); duke@435: duke@435: // Split shared headers and insert loop landing pads. duke@435: // Insert a LoopNode to replace the RegionNode. duke@435: // Returns TRUE if loop tree is structurally changed. duke@435: bool beautify_loops( PhaseIdealLoop *phase ); duke@435: cfang@1607: // Perform optimization to use the loop predicates for null checks and range checks. cfang@1607: // Applies to any loop level (not just the innermost one) cfang@1607: bool loop_predication( PhaseIdealLoop *phase); cfang@1607: never@836: // Perform iteration-splitting on inner loops. Split iterations to never@836: // avoid range checks or one-shot null checks. Returns false if the never@836: // current round of loop opts should stop. never@836: bool iteration_split( PhaseIdealLoop *phase, Node_List &old_new ); duke@435: never@836: // Driver for various flavors of iteration splitting. Returns false never@836: // if the current round of loop opts should stop. never@836: bool iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ); duke@435: duke@435: // Given dominators, try to find loops with calls that must always be duke@435: // executed (call dominates loop tail). These loops do not need non-call duke@435: // safepoints (ncsfpt). duke@435: void check_safepts(VectorSet &visited, Node_List &stack); duke@435: duke@435: // Allpaths backwards scan from loop tail, terminating each path at first safepoint duke@435: // encountered. duke@435: void allpaths_check_safepts(VectorSet &visited, Node_List &stack); duke@435: duke@435: // Convert to counted loops where possible duke@435: void counted_loop( PhaseIdealLoop *phase ); duke@435: duke@435: // Check for Node being a loop-breaking test duke@435: Node *is_loop_exit(Node *iff) const; duke@435: duke@435: // Returns true if ctrl is executed on every complete iteration duke@435: bool dominates_backedge(Node* ctrl); duke@435: duke@435: // Remove simplistic dead code from loop body duke@435: void DCE_loop_body(); duke@435: duke@435: // Look for loop-exit tests with my 50/50 guesses from the Parsing stage. duke@435: // Replace with a 1-in-10 exit guess. duke@435: void adjust_loop_exit_prob( PhaseIdealLoop *phase ); duke@435: duke@435: // Return TRUE or FALSE if the loop should never be RCE'd or aligned. duke@435: // Useful for unrolling loops with NO array accesses. duke@435: bool policy_peel_only( PhaseIdealLoop *phase ) const; duke@435: duke@435: // Return TRUE or FALSE if the loop should be unswitched -- clone duke@435: // loop with an invariant test duke@435: bool policy_unswitching( PhaseIdealLoop *phase ) const; duke@435: duke@435: // Micro-benchmark spamming. Remove empty loops. duke@435: bool policy_do_remove_empty_loop( PhaseIdealLoop *phase ); duke@435: kvn@2747: // Convert one iteration loop into normal code. kvn@2747: bool policy_do_one_iteration_loop( PhaseIdealLoop *phase ); kvn@2747: duke@435: // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can duke@435: // make some loop-invariant test (usually a null-check) happen before the duke@435: // loop. duke@435: bool policy_peeling( PhaseIdealLoop *phase ) const; duke@435: duke@435: // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any duke@435: // known trip count in the counted loop node. duke@435: bool policy_maximally_unroll( PhaseIdealLoop *phase ) const; duke@435: duke@435: // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if duke@435: // the loop is a CountedLoop and the body is small enough. duke@435: bool policy_unroll( PhaseIdealLoop *phase ) const; duke@435: duke@435: // Return TRUE or FALSE if the loop should be range-check-eliminated. duke@435: // Gather a list of IF tests that are dominated by iteration splitting; duke@435: // also gather the end of the first split and the start of the 2nd split. duke@435: bool policy_range_check( PhaseIdealLoop *phase ) const; duke@435: duke@435: // Return TRUE or FALSE if the loop should be cache-line aligned. duke@435: // Gather the expression that does the alignment. Note that only twisti@1040: // one array base can be aligned in a loop (unless the VM guarantees duke@435: // mutual alignment). Note that if we vectorize short memory ops duke@435: // into longer memory ops, we may want to increase alignment. duke@435: bool policy_align( PhaseIdealLoop *phase ) const; duke@435: cfang@1607: // Return TRUE if "iff" is a range check. cfang@1607: bool is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const; cfang@1607: kvn@2747: // Compute loop exact trip count if possible kvn@2747: void compute_exact_trip_count( PhaseIdealLoop *phase ); kvn@2747: duke@435: // Compute loop trip count from profile data duke@435: void compute_profile_trip_cnt( PhaseIdealLoop *phase ); duke@435: duke@435: // Reassociate invariant expressions. duke@435: void reassociate_invariants(PhaseIdealLoop *phase); duke@435: // Reassociate invariant add and subtract expressions. duke@435: Node* reassociate_add_sub(Node* n1, PhaseIdealLoop *phase); duke@435: // Return nonzero index of invariant operand if invariant and variant twisti@1040: // are combined with an Add or Sub. Helper for reassociate_invariants. duke@435: int is_invariant_addition(Node* n, PhaseIdealLoop *phase); duke@435: duke@435: // Return true if n is invariant duke@435: bool is_invariant(Node* n) const; duke@435: duke@435: // Put loop body on igvn work list duke@435: void record_for_igvn(); duke@435: duke@435: bool is_loop() { return !_irreducible && _tail && !_tail->is_top(); } duke@435: bool is_inner() { return is_loop() && _child == NULL; } duke@435: bool is_counted() { return is_loop() && _head != NULL && _head->is_CountedLoop(); } duke@435: duke@435: #ifndef PRODUCT duke@435: void dump_head( ) const; // Dump loop head only duke@435: void dump() const; // Dump this loop recursively duke@435: void verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const; duke@435: #endif duke@435: duke@435: }; duke@435: duke@435: // -----------------------------PhaseIdealLoop--------------------------------- duke@435: // Computes the mapping from Nodes to IdealLoopTrees. Organizes IdealLoopTrees into a duke@435: // loop tree. Drives the loop-based transformations on the ideal graph. duke@435: class PhaseIdealLoop : public PhaseTransform { duke@435: friend class IdealLoopTree; duke@435: friend class SuperWord; duke@435: // Pre-computed def-use info duke@435: PhaseIterGVN &_igvn; duke@435: duke@435: // Head of loop tree duke@435: IdealLoopTree *_ltree_root; duke@435: duke@435: // Array of pre-order numbers, plus post-visited bit. duke@435: // ZERO for not pre-visited. EVEN for pre-visited but not post-visited. duke@435: // ODD for post-visited. Other bits are the pre-order number. duke@435: uint *_preorders; duke@435: uint _max_preorder; duke@435: never@1356: const PhaseIdealLoop* _verify_me; never@1356: bool _verify_only; never@1356: duke@435: // Allocate _preorders[] array duke@435: void allocate_preorders() { duke@435: _max_preorder = C->unique()+8; duke@435: _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder); duke@435: memset(_preorders, 0, sizeof(uint) * _max_preorder); duke@435: } duke@435: duke@435: // Allocate _preorders[] array duke@435: void reallocate_preorders() { duke@435: if ( _max_preorder < C->unique() ) { duke@435: _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique()); duke@435: _max_preorder = C->unique(); duke@435: } duke@435: memset(_preorders, 0, sizeof(uint) * _max_preorder); duke@435: } duke@435: duke@435: // Check to grow _preorders[] array for the case when build_loop_tree_impl() duke@435: // adds new nodes. duke@435: void check_grow_preorders( ) { duke@435: if ( _max_preorder < C->unique() ) { duke@435: uint newsize = _max_preorder<<1; // double size of array duke@435: _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize); duke@435: memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder)); duke@435: _max_preorder = newsize; duke@435: } duke@435: } duke@435: // Check for pre-visited. Zero for NOT visited; non-zero for visited. duke@435: int is_visited( Node *n ) const { return _preorders[n->_idx]; } duke@435: // Pre-order numbers are written to the Nodes array as low-bit-set values. duke@435: void set_preorder_visited( Node *n, int pre_order ) { duke@435: assert( !is_visited( n ), "already set" ); duke@435: _preorders[n->_idx] = (pre_order<<1); duke@435: }; duke@435: // Return pre-order number. duke@435: int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; } duke@435: duke@435: // Check for being post-visited. duke@435: // Should be previsited already (checked with assert(is_visited(n))). duke@435: int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; } duke@435: duke@435: // Mark as post visited duke@435: void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; } duke@435: duke@435: // Set/get control node out. Set lower bit to distinguish from IdealLoopTree duke@435: // Returns true if "n" is a data node, false if it's a control node. duke@435: bool has_ctrl( Node *n ) const { return ((intptr_t)_nodes[n->_idx]) & 1; } duke@435: duke@435: // clear out dead code after build_loop_late duke@435: Node_List _deadlist; duke@435: duke@435: // Support for faster execution of get_late_ctrl()/dom_lca() duke@435: // when a node has many uses and dominator depth is deep. duke@435: Node_Array _dom_lca_tags; duke@435: void init_dom_lca_tags(); duke@435: void clear_dom_lca_tags(); never@1356: never@1356: // Helper for debugging bad dominance relationships never@1356: bool verify_dominance(Node* n, Node* use, Node* LCA, Node* early); never@1356: never@1356: Node* compute_lca_of_uses(Node* n, Node* early, bool verify = false); never@1356: duke@435: // Inline wrapper for frequent cases: duke@435: // 1) only one use duke@435: // 2) a use is the same as the current LCA passed as 'n1' duke@435: Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) { duke@435: assert( n->is_CFG(), "" ); duke@435: // Fast-path NULL lca duke@435: if( lca != NULL && lca != n ) { duke@435: assert( lca->is_CFG(), "" ); duke@435: // find LCA of all uses duke@435: n = dom_lca_for_get_late_ctrl_internal( lca, n, tag ); duke@435: } duke@435: return find_non_split_ctrl(n); duke@435: } duke@435: Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag ); never@1356: duke@435: // Helper function for directing control inputs away from CFG split duke@435: // points. duke@435: Node *find_non_split_ctrl( Node *ctrl ) const { duke@435: if (ctrl != NULL) { duke@435: if (ctrl->is_MultiBranch()) { duke@435: ctrl = ctrl->in(0); duke@435: } duke@435: assert(ctrl->is_CFG(), "CFG"); duke@435: } duke@435: return ctrl; duke@435: } duke@435: duke@435: public: duke@435: bool has_node( Node* n ) const { return _nodes[n->_idx] != NULL; } duke@435: // check if transform created new nodes that need _ctrl recorded duke@435: Node *get_late_ctrl( Node *n, Node *early ); duke@435: Node *get_early_ctrl( Node *n ); duke@435: void set_early_ctrl( Node *n ); duke@435: void set_subtree_ctrl( Node *root ); duke@435: void set_ctrl( Node *n, Node *ctrl ) { duke@435: assert( !has_node(n) || has_ctrl(n), "" ); duke@435: assert( ctrl->in(0), "cannot set dead control node" ); duke@435: assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" ); duke@435: _nodes.map( n->_idx, (Node*)((intptr_t)ctrl + 1) ); duke@435: } duke@435: // Set control and update loop membership duke@435: void set_ctrl_and_loop(Node* n, Node* ctrl) { duke@435: IdealLoopTree* old_loop = get_loop(get_ctrl(n)); duke@435: IdealLoopTree* new_loop = get_loop(ctrl); duke@435: if (old_loop != new_loop) { duke@435: if (old_loop->_child == NULL) old_loop->_body.yank(n); duke@435: if (new_loop->_child == NULL) new_loop->_body.push(n); duke@435: } duke@435: set_ctrl(n, ctrl); duke@435: } duke@435: // Control nodes can be replaced or subsumed. During this pass they duke@435: // get their replacement Node in slot 1. Instead of updating the block duke@435: // location of all Nodes in the subsumed block, we lazily do it. As we duke@435: // pull such a subsumed block out of the array, we write back the final duke@435: // correct block. duke@435: Node *get_ctrl( Node *i ) { duke@435: assert(has_node(i), ""); duke@435: Node *n = get_ctrl_no_update(i); duke@435: _nodes.map( i->_idx, (Node*)((intptr_t)n + 1) ); duke@435: assert(has_node(i) && has_ctrl(i), ""); duke@435: assert(n == find_non_split_ctrl(n), "must return legal ctrl" ); duke@435: return n; duke@435: } cfang@1607: // true if CFG node d dominates CFG node n cfang@1607: bool is_dominator(Node *d, Node *n); cfang@1607: // return get_ctrl for a data node and self(n) for a CFG node cfang@1607: Node* ctrl_or_self(Node* n) { cfang@1607: if (has_ctrl(n)) cfang@1607: return get_ctrl(n); cfang@1607: else { cfang@1607: assert (n->is_CFG(), "must be a CFG node"); cfang@1607: return n; cfang@1607: } cfang@1607: } duke@435: duke@435: private: duke@435: Node *get_ctrl_no_update( Node *i ) const { duke@435: assert( has_ctrl(i), "" ); duke@435: Node *n = (Node*)(((intptr_t)_nodes[i->_idx]) & ~1); duke@435: if (!n->in(0)) { duke@435: // Skip dead CFG nodes duke@435: do { duke@435: n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1); duke@435: } while (!n->in(0)); duke@435: n = find_non_split_ctrl(n); duke@435: } duke@435: return n; duke@435: } duke@435: duke@435: // Check for loop being set duke@435: // "n" must be a control node. Returns true if "n" is known to be in a loop. duke@435: bool has_loop( Node *n ) const { duke@435: assert(!has_node(n) || !has_ctrl(n), ""); duke@435: return has_node(n); duke@435: } duke@435: // Set loop duke@435: void set_loop( Node *n, IdealLoopTree *loop ) { duke@435: _nodes.map(n->_idx, (Node*)loop); duke@435: } duke@435: // Lazy-dazy update of 'get_ctrl' and 'idom_at' mechanisms. Replace duke@435: // the 'old_node' with 'new_node'. Kill old-node. Add a reference duke@435: // from old_node to new_node to support the lazy update. Reference cfang@1607: // replaces loop reference, since that is not needed for dead node. duke@435: public: duke@435: void lazy_update( Node *old_node, Node *new_node ) { duke@435: assert( old_node != new_node, "no cycles please" ); duke@435: //old_node->set_req( 1, new_node /*NO DU INFO*/ ); duke@435: // Nodes always have DU info now, so re-use the side array slot duke@435: // for this node to provide the forwarding pointer. duke@435: _nodes.map( old_node->_idx, (Node*)((intptr_t)new_node + 1) ); duke@435: } duke@435: void lazy_replace( Node *old_node, Node *new_node ) { kvn@1976: _igvn.replace_node( old_node, new_node ); duke@435: lazy_update( old_node, new_node ); duke@435: } duke@435: void lazy_replace_proj( Node *old_node, Node *new_node ) { duke@435: assert( old_node->req() == 1, "use this for Projs" ); duke@435: _igvn.hash_delete(old_node); // Must hash-delete before hacking edges duke@435: old_node->add_req( NULL ); duke@435: lazy_replace( old_node, new_node ); duke@435: } duke@435: duke@435: private: duke@435: duke@435: // Place 'n' in some loop nest, where 'n' is a CFG node duke@435: void build_loop_tree(); duke@435: int build_loop_tree_impl( Node *n, int pre_order ); duke@435: // Insert loop into the existing loop tree. 'innermost' is a leaf of the duke@435: // loop tree, not the root. duke@435: IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost ); duke@435: duke@435: // Place Data nodes in some loop nest never@1356: void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ); never@1356: void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ); never@1356: void build_loop_late_post ( Node* n ); duke@435: duke@435: // Array of immediate dominance info for each CFG node indexed by node idx duke@435: private: duke@435: uint _idom_size; duke@435: Node **_idom; // Array of immediate dominators duke@435: uint *_dom_depth; // Used for fast LCA test duke@435: GrowableArray* _dom_stk; // For recomputation of dom depth duke@435: duke@435: Node* idom_no_update(Node* d) const { duke@435: assert(d->_idx < _idom_size, "oob"); duke@435: Node* n = _idom[d->_idx]; duke@435: assert(n != NULL,"Bad immediate dominator info."); duke@435: while (n->in(0) == NULL) { // Skip dead CFG nodes duke@435: //n = n->in(1); duke@435: n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1); duke@435: assert(n != NULL,"Bad immediate dominator info."); duke@435: } duke@435: return n; duke@435: } duke@435: Node *idom(Node* d) const { duke@435: uint didx = d->_idx; duke@435: Node *n = idom_no_update(d); duke@435: _idom[didx] = n; // Lazily remove dead CFG nodes from table. duke@435: return n; duke@435: } duke@435: uint dom_depth(Node* d) const { duke@435: assert(d->_idx < _idom_size, ""); duke@435: return _dom_depth[d->_idx]; duke@435: } duke@435: void set_idom(Node* d, Node* n, uint dom_depth); duke@435: // Locally compute IDOM using dom_lca call duke@435: Node *compute_idom( Node *region ) const; duke@435: // Recompute dom_depth duke@435: void recompute_dom_depth(); duke@435: duke@435: // Is safept not required by an outer loop? duke@435: bool is_deleteable_safept(Node* sfpt); duke@435: kvn@2665: // Replace parallel induction variable (parallel to trip counter) kvn@2665: void replace_parallel_iv(IdealLoopTree *loop); kvn@2665: never@1356: // Perform verification that the graph is valid. never@1356: PhaseIdealLoop( PhaseIterGVN &igvn) : never@1356: PhaseTransform(Ideal_Loop), never@1356: _igvn(igvn), kvn@2555: _dom_lca_tags(arena()), // Thread::resource_area never@1356: _verify_me(NULL), never@1356: _verify_only(true) { kvn@3260: build_and_optimize(false, false); never@1356: } never@1356: never@1356: // build the loop tree and perform any requested optimizations kvn@3260: void build_and_optimize(bool do_split_if, bool skip_loop_opts); never@1356: duke@435: public: duke@435: // Dominators for the sea of nodes duke@435: void Dominators(); duke@435: Node *dom_lca( Node *n1, Node *n2 ) const { duke@435: return find_non_split_ctrl(dom_lca_internal(n1, n2)); duke@435: } duke@435: Node *dom_lca_internal( Node *n1, Node *n2 ) const; duke@435: duke@435: // Compute the Ideal Node to Loop mapping kvn@3260: PhaseIdealLoop( PhaseIterGVN &igvn, bool do_split_ifs, bool skip_loop_opts = false) : never@1356: PhaseTransform(Ideal_Loop), never@1356: _igvn(igvn), kvn@2555: _dom_lca_tags(arena()), // Thread::resource_area never@1356: _verify_me(NULL), never@1356: _verify_only(false) { kvn@3260: build_and_optimize(do_split_ifs, skip_loop_opts); never@1356: } never@1356: never@1356: // Verify that verify_me made the same decisions as a fresh run. never@1356: PhaseIdealLoop( PhaseIterGVN &igvn, const PhaseIdealLoop *verify_me) : never@1356: PhaseTransform(Ideal_Loop), never@1356: _igvn(igvn), kvn@2555: _dom_lca_tags(arena()), // Thread::resource_area never@1356: _verify_me(verify_me), never@1356: _verify_only(false) { kvn@3260: build_and_optimize(false, false); never@1356: } never@1356: never@1356: // Build and verify the loop tree without modifying the graph. This never@1356: // is useful to verify that all inputs properly dominate their uses. never@1356: static void verify(PhaseIterGVN& igvn) { never@1356: #ifdef ASSERT never@1356: PhaseIdealLoop v(igvn); never@1356: #endif never@1356: } duke@435: duke@435: // True if the method has at least 1 irreducible loop duke@435: bool _has_irreducible_loops; duke@435: duke@435: // Per-Node transform duke@435: virtual Node *transform( Node *a_node ) { return 0; } duke@435: kvn@2665: bool is_counted_loop( Node *x, IdealLoopTree *loop ); duke@435: kvn@2877: Node* exact_limit( IdealLoopTree *loop ); kvn@2877: duke@435: // Return a post-walked LoopNode duke@435: IdealLoopTree *get_loop( Node *n ) const { duke@435: // Dead nodes have no loop, so return the top level loop instead duke@435: if (!has_node(n)) return _ltree_root; duke@435: assert(!has_ctrl(n), ""); duke@435: return (IdealLoopTree*)_nodes[n->_idx]; duke@435: } duke@435: duke@435: // Is 'n' a (nested) member of 'loop'? duke@435: int is_member( const IdealLoopTree *loop, Node *n ) const { duke@435: return loop->is_member(get_loop(n)); } 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 duke@435: // mapping you can find the new-loop equivalent to an old-loop node. All duke@435: // new-loop nodes are exactly equal to their old-loop counterparts, all duke@435: // edges are the same. All exits from the old-loop now have a RegionNode duke@435: // that merges the equivalent new-loop path. This is true even for the duke@435: // normal "loop-exit" condition. All uses of loop-invariant old-loop values duke@435: // now come from (one or more) Phis that merge their new-loop equivalents. duke@435: // Parameter side_by_side_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 passed in side_by_side_idom node. Used in duke@435: // construction of unswitched loops. duke@435: void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth, duke@435: Node* side_by_side_idom = NULL); duke@435: duke@435: // If we got the effect of peeling, either by actually peeling or by duke@435: // making a pre-loop which must execute at least once, we can remove duke@435: // all loop-invariant dominated tests in the main body. duke@435: void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new ); duke@435: duke@435: // Generate code to do a loop peel for the given loop (and body). duke@435: // old_new is a temp array. duke@435: void do_peeling( IdealLoopTree *loop, Node_List &old_new ); duke@435: duke@435: // Add pre and post loops around the given loop. These loops are used duke@435: // during RCE, unrolling and aligning loops. duke@435: void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ); duke@435: // If Node n lives in the back_ctrl block, we clone a private version of n duke@435: // in preheader_ctrl block and return that, otherwise return n. kvn@2985: Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones ); duke@435: duke@435: // Take steps to maximally unroll the loop. Peel any odd iterations, then duke@435: // unroll to do double iterations. The next round of major loop transforms duke@435: // will repeat till the doubled loop body does all remaining iterations in 1 duke@435: // pass. duke@435: void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ); duke@435: duke@435: // Unroll the loop body one step - make each trip do 2 iterations. duke@435: void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ); duke@435: duke@435: // Return true if exp is a constant times an induction var duke@435: bool is_scaled_iv(Node* exp, Node* iv, int* p_scale); duke@435: duke@435: // Return true if exp is a scaled induction var plus (or minus) constant duke@435: bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth = 0); duke@435: cfang@1607: // Return true if proj is for "proj->[region->..]call_uct" kvn@2665: static bool is_uncommon_trap_proj(ProjNode* proj, Deoptimization::DeoptReason reason); cfang@1607: // Return true for "if(test)-> proj -> ... cfang@1607: // | cfang@1607: // V cfang@1607: // other_proj->[region->..]call_uct" kvn@2665: static bool is_uncommon_trap_if_pattern(ProjNode* proj, Deoptimization::DeoptReason reason); cfang@1607: // Create a new if above the uncommon_trap_if_pattern for the predicate to be promoted kvn@2665: ProjNode* create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry, kvn@2665: Deoptimization::DeoptReason reason); kvn@2665: void register_control(Node* n, IdealLoopTree *loop, Node* pred); kvn@2665: kvn@2727: // Clone loop predicates to cloned loops (peeled, unswitched) kvn@2727: static ProjNode* clone_predicate(ProjNode* predicate_proj, Node* new_entry, kvn@2727: Deoptimization::DeoptReason reason, kvn@2727: PhaseIdealLoop* loop_phase, kvn@2727: PhaseIterGVN* igvn); kvn@3043: kvn@2727: static Node* clone_loop_predicates(Node* old_entry, Node* new_entry, kvn@2877: bool clone_limit_check, kvn@2727: PhaseIdealLoop* loop_phase, kvn@2727: PhaseIterGVN* igvn); kvn@2877: Node* clone_loop_predicates(Node* old_entry, Node* new_entry, bool clone_limit_check); kvn@2727: kvn@2727: static Node* skip_loop_predicates(Node* entry); kvn@2727: kvn@2727: // Find a good location to insert a predicate kvn@2665: static ProjNode* find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason); kvn@2665: // Find a predicate kvn@2665: static Node* find_predicate(Node* entry); cfang@1607: // Construct a range check for a predicate if kvn@2877: BoolNode* rc_predicate(IdealLoopTree *loop, Node* ctrl, cfang@1607: int scale, Node* offset, cfang@1607: Node* init, Node* limit, Node* stride, never@1738: Node* range, bool upper); cfang@1607: cfang@1607: // Implementation of the loop predication to promote checks outside the loop cfang@1607: bool loop_predication_impl(IdealLoopTree *loop); cfang@1607: cfang@1607: // Helper function to collect predicate for eliminating the useless ones cfang@1607: void collect_potentially_useful_predicates(IdealLoopTree *loop, Unique_Node_List &predicate_opaque1); cfang@1607: void eliminate_useless_predicates(); cfang@1607: duke@435: // Eliminate range-checks and other trip-counter vs loop-invariant tests. duke@435: void do_range_check( IdealLoopTree *loop, Node_List &old_new ); duke@435: duke@435: // Create a slow version of the loop by cloning the loop duke@435: // and inserting an if to select fast-slow versions. duke@435: ProjNode* create_slow_version_of_loop(IdealLoopTree *loop, duke@435: Node_List &old_new); duke@435: duke@435: // Clone loop with an invariant test (that does not exit) and duke@435: // insert a clone of the test that selects which version to duke@435: // execute. duke@435: void do_unswitching (IdealLoopTree *loop, Node_List &old_new); duke@435: duke@435: // Find candidate "if" for unswitching duke@435: IfNode* find_unswitching_candidate(const IdealLoopTree *loop) const; duke@435: duke@435: // Range Check Elimination uses this function! duke@435: // Constrain the main loop iterations so the affine function: kvn@2877: // low_limit <= scale_con * I + offset < upper_limit duke@435: // always holds true. That is, either increase the number of iterations in duke@435: // the pre-loop or the post-loop until the condition holds true in the main duke@435: // loop. Scale_con, offset and limit are all loop invariant. kvn@2877: void add_constraint( int stride_con, int scale_con, Node *offset, Node *low_limit, Node *upper_limit, Node *pre_ctrl, Node **pre_limit, Node **main_limit ); kvn@2915: // Helper function for add_constraint(). kvn@2915: Node* adjust_limit( int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl ); duke@435: duke@435: // Partially peel loop up through last_peel node. duke@435: bool partial_peel( IdealLoopTree *loop, Node_List &old_new ); duke@435: duke@435: // Create a scheduled list of nodes control dependent on ctrl set. duke@435: void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched ); duke@435: // Has a use in the vector set duke@435: bool has_use_in_set( Node* n, VectorSet& vset ); duke@435: // Has use internal to the vector set (ie. not in a phi at the loop head) duke@435: bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ); duke@435: // clone "n" for uses that are outside of loop duke@435: void clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ); duke@435: // clone "n" for special uses that are in the not_peeled region duke@435: void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, duke@435: VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ); 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 insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ); duke@435: #ifdef ASSERT duke@435: // Validate the loop partition sets: peel and not_peel duke@435: bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel ); duke@435: // Ensure that uses outside of loop are of the right form duke@435: bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, duke@435: uint orig_exit_idx, uint clone_exit_idx); duke@435: bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx); duke@435: #endif duke@435: duke@435: // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.) duke@435: int stride_of_possible_iv( Node* iff ); duke@435: bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; } duke@435: // Return the (unique) control output node that's in the loop (if it exists.) duke@435: Node* stay_in_loop( Node* n, IdealLoopTree *loop); duke@435: // Insert a signed compare loop exit cloned from an unsigned compare. duke@435: IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop); duke@435: void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop); duke@435: // Utility to register node "n" with PhaseIdealLoop duke@435: void register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth); duke@435: // Utility to create an if-projection duke@435: ProjNode* proj_clone(ProjNode* p, IfNode* iff); duke@435: // Force the iff control output to be the live_proj duke@435: Node* short_circuit_if(IfNode* iff, ProjNode* live_proj); duke@435: // Insert a region before an if projection duke@435: RegionNode* insert_region_before_proj(ProjNode* proj); duke@435: // Insert a new if before an if projection duke@435: ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj); duke@435: 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 *clone_iff( PhiNode *phi, IdealLoopTree *loop ); duke@435: CmpNode *clone_bool( PhiNode *phi, IdealLoopTree *loop ); duke@435: duke@435: 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 *remix_address_expressions( Node *n ); duke@435: duke@435: // Attempt to use a conditional move instead of a phi/branch duke@435: Node *conditional_move( Node *n ); duke@435: duke@435: // Reorganize offset computations to lower register pressure. duke@435: // Mostly prevent loop-fallout uses of the pre-incremented trip counter duke@435: // (which are then alive with the post-incremented trip counter duke@435: // forcing an extra register move) duke@435: void reorg_offsets( IdealLoopTree *loop ); duke@435: duke@435: // Check for aggressive application of 'split-if' optimization, duke@435: // using basic block level info. duke@435: void split_if_with_blocks ( VectorSet &visited, Node_Stack &nstack ); duke@435: Node *split_if_with_blocks_pre ( Node *n ); duke@435: void split_if_with_blocks_post( Node *n ); duke@435: Node *has_local_phi_input( Node *n ); duke@435: // Mark an IfNode as being dominated by a prior test, duke@435: // without actually altering the CFG (and hence IDOM info). kvn@3038: void dominated_by( Node *prevdom, Node *iff, bool flip = false, bool exclude_loop_predicate = false ); duke@435: duke@435: // Split Node 'n' through merge point duke@435: Node *split_thru_region( Node *n, Node *region ); duke@435: // Split Node 'n' through merge point if there is enough win. duke@435: Node *split_thru_phi( Node *n, Node *region, int policy ); duke@435: // Found an If getting its condition-code input from a Phi in the duke@435: // same block. Split thru the Region. duke@435: void do_split_if( Node *iff ); duke@435: never@2118: // Conversion of fill/copy patterns into intrisic versions never@2118: bool do_intrinsify_fill(); never@2118: bool intrinsify_fill(IdealLoopTree* lpt); never@2118: bool match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value, never@2118: Node*& shift, Node*& offset); never@2118: duke@435: private: duke@435: // Return a type based on condition control flow duke@435: const TypeInt* filtered_type( Node *n, Node* n_ctrl); duke@435: const TypeInt* filtered_type( Node *n ) { return filtered_type(n, NULL); } duke@435: // Helpers for filtered type duke@435: const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl); duke@435: duke@435: // Helper functions duke@435: Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache ); duke@435: Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true ); duke@435: 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 ); duke@435: bool split_up( Node *n, Node *blk1, Node *blk2 ); duke@435: void sink_use( Node *use, Node *post_loop ); duke@435: Node *place_near_use( Node *useblock ) const; duke@435: duke@435: bool _created_loop_node; duke@435: public: duke@435: void set_created_loop_node() { _created_loop_node = true; } duke@435: bool created_loop_node() { return _created_loop_node; } cfang@1607: void register_new_node( Node *n, Node *blk ); duke@435: duke@435: #ifndef PRODUCT duke@435: void dump( ) const; duke@435: void dump( IdealLoopTree *loop, uint rpo_idx, Node_List &rpo_list ) const; duke@435: void rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const; duke@435: void verify() const; // Major slow :-) duke@435: void verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const; duke@435: IdealLoopTree *get_loop_idx(Node* n) const { duke@435: // Dead nodes have no loop, so return the top level loop instead duke@435: return _nodes[n->_idx] ? (IdealLoopTree*)_nodes[n->_idx] : _ltree_root; duke@435: } duke@435: // Print some stats duke@435: static void print_statistics(); duke@435: static int _loop_invokes; // Count of PhaseIdealLoop invokes duke@435: static int _loop_work; // Sum of PhaseIdealLoop x _unique duke@435: #endif duke@435: }; duke@435: duke@435: inline Node* IdealLoopTree::tail() { duke@435: // Handle lazy update of _tail field duke@435: Node *n = _tail; duke@435: //while( !n->in(0) ) // Skip dead CFG nodes duke@435: //n = n->in(1); duke@435: if (n->in(0) == NULL) duke@435: n = _phase->get_ctrl(n); duke@435: _tail = n; duke@435: return n; duke@435: } duke@435: duke@435: duke@435: // Iterate over the loop tree using a preorder, left-to-right traversal. duke@435: // duke@435: // Example that visits all counted loops from within PhaseIdealLoop duke@435: // duke@435: // for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { duke@435: // IdealLoopTree* lpt = iter.current(); duke@435: // if (!lpt->is_counted()) continue; duke@435: // ... duke@435: class LoopTreeIterator : public StackObj { duke@435: private: duke@435: IdealLoopTree* _root; duke@435: IdealLoopTree* _curnt; duke@435: duke@435: public: duke@435: LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {} duke@435: duke@435: bool done() { return _curnt == NULL; } // Finished iterating? duke@435: duke@435: void next(); // Advance to next loop tree duke@435: duke@435: IdealLoopTree* current() { return _curnt; } // Return current value of iterator. duke@435: }; stefank@2314: stefank@2314: #endif // SHARE_VM_OPTO_LOOPNODE_HPP