26 #define SHARE_OPTO_LOOPNODE_HPP
27
28 #include "opto/cfgnode.hpp"
29 #include "opto/multnode.hpp"
30 #include "opto/phaseX.hpp"
31 #include "opto/predicates.hpp"
32 #include "opto/subnode.hpp"
33 #include "opto/type.hpp"
34 #include "utilities/checkedCast.hpp"
35
36 class CmpNode;
37 class BaseCountedLoopEndNode;
38 class CountedLoopNode;
39 class IdealLoopTree;
40 class LoopNode;
41 class Node;
42 class OuterStripMinedLoopEndNode;
43 class PredicateBlock;
44 class PathFrequency;
45 class PhaseIdealLoop;
46 class LoopSelector;
47 class UnswitchedLoopSelector;
48 class VectorSet;
49 class VSharedData;
50 class Invariance;
51 struct small_cache;
52
53 //
54 // I D E A L I Z E D L O O P S
55 //
56 // Idealized loops are the set of loops I perform more interesting
57 // transformations on, beyond simple hoisting.
58
59 //------------------------------LoopNode---------------------------------------
60 // Simple loop header. Fall in path on left, loop-back path on right.
61 class LoopNode : public RegionNode {
62 // Size is bigger to hold the flags. However, the flags do not change
63 // the semantics so it does not appear in the hash & cmp functions.
64 virtual uint size_of() const { return sizeof(*this); }
65 protected:
68 enum { Normal=0, Pre=1, Main=2, Post=3, PreMainPostFlagsMask=3,
69 MainHasNoPreLoop = 1<<2,
70 HasExactTripCount = 1<<3,
71 InnerLoop = 1<<4,
72 PartialPeelLoop = 1<<5,
73 PartialPeelFailed = 1<<6,
74 WasSlpAnalyzed = 1<<7,
75 PassedSlpAnalysis = 1<<8,
76 DoUnrollOnly = 1<<9,
77 VectorizedLoop = 1<<10,
78 HasAtomicPostLoop = 1<<11,
79 StripMined = 1<<12,
80 SubwordLoop = 1<<13,
81 ProfileTripFailed = 1<<14,
82 LoopNestInnerLoop = 1<<15,
83 LoopNestLongOuterLoop = 1<<16,
84 MultiversionFastLoop = 1<<17,
85 MultiversionSlowLoop = 2<<17,
86 MultiversionDelayedSlowLoop = 3<<17,
87 MultiversionFlagsMask = 3<<17,
88 };
89 char _unswitch_count;
90 enum { _unswitch_max=3 };
91
92 // Expected trip count from profile data
93 float _profile_trip_cnt;
94
95 public:
96 // Names for edge indices
97 enum { Self=0, EntryControl, LoopBackControl };
98
99 bool is_inner_loop() const { return _loop_flags & InnerLoop; }
100 void set_inner_loop() { _loop_flags |= InnerLoop; }
101
102 bool is_vectorized_loop() const { return _loop_flags & VectorizedLoop; }
103 bool is_partial_peel_loop() const { return _loop_flags & PartialPeelLoop; }
104 void set_partial_peel_loop() { _loop_flags |= PartialPeelLoop; }
105 bool partial_peel_has_failed() const { return _loop_flags & PartialPeelFailed; }
106 bool is_strip_mined() const { return _loop_flags & StripMined; }
107 bool is_profile_trip_failed() const { return _loop_flags & ProfileTripFailed; }
108 bool is_subword_loop() const { return _loop_flags & SubwordLoop; }
109 bool is_loop_nest_inner_loop() const { return _loop_flags & LoopNestInnerLoop; }
110 bool is_loop_nest_outer_loop() const { return _loop_flags & LoopNestLongOuterLoop; }
111
112 void mark_partial_peel_failed() { _loop_flags |= PartialPeelFailed; }
113 void mark_was_slp() { _loop_flags |= WasSlpAnalyzed; }
114 void mark_passed_slp() { _loop_flags |= PassedSlpAnalysis; }
115 void mark_do_unroll_only() { _loop_flags |= DoUnrollOnly; }
116 void mark_loop_vectorized() { _loop_flags |= VectorizedLoop; }
117 void mark_has_atomic_post_loop() { _loop_flags |= HasAtomicPostLoop; }
118 void mark_strip_mined() { _loop_flags |= StripMined; }
119 void clear_strip_mined() { _loop_flags &= ~StripMined; }
120 void mark_profile_trip_failed() { _loop_flags |= ProfileTripFailed; }
121 void mark_subword_loop() { _loop_flags |= SubwordLoop; }
122 void mark_loop_nest_inner_loop() { _loop_flags |= LoopNestInnerLoop; }
123 void mark_loop_nest_outer_loop() { _loop_flags |= LoopNestLongOuterLoop; }
124
125 int unswitch_max() { return _unswitch_max; }
126 int unswitch_count() { return _unswitch_count; }
127
128 void set_unswitch_count(int val) {
129 assert (val <= unswitch_max(), "too many unswitches");
130 _unswitch_count = val;
131 }
132
133 void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; }
134 float profile_trip_cnt() { return _profile_trip_cnt; }
135
136 #ifndef PRODUCT
137 uint _stress_peeling_attempts = 0;
138 #endif
139
140 LoopNode(Node *entry, Node *backedge)
141 : RegionNode(3), _loop_flags(0), _unswitch_count(0),
142 _profile_trip_cnt(COUNT_UNKNOWN) {
143 init_class_id(Class_Loop);
717 // Convert to counted loops where possible
718 void counted_loop( PhaseIdealLoop *phase );
719
720 // Check for Node being a loop-breaking test
721 Node *is_loop_exit(Node *iff) const;
722
723 // Remove simplistic dead code from loop body
724 void DCE_loop_body();
725
726 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
727 // Replace with a 1-in-10 exit guess.
728 void adjust_loop_exit_prob( PhaseIdealLoop *phase );
729
730 // Return TRUE or FALSE if the loop should never be RCE'd or aligned.
731 // Useful for unrolling loops with NO array accesses.
732 bool policy_peel_only( PhaseIdealLoop *phase ) const;
733
734 // Return TRUE or FALSE if the loop should be unswitched -- clone
735 // loop with an invariant test
736 bool policy_unswitching( PhaseIdealLoop *phase ) const;
737
738 // Micro-benchmark spamming. Remove empty loops.
739 bool do_remove_empty_loop( PhaseIdealLoop *phase );
740
741 // Convert one-iteration loop into normal code.
742 bool do_one_iteration_loop( PhaseIdealLoop *phase );
743
744 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
745 // move some loop-invariant test (usually a null-check) before the loop.
746 bool policy_peeling(PhaseIdealLoop *phase);
747
748 uint estimate_peeling(PhaseIdealLoop *phase);
749
750 // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any
751 // known trip count in the counted loop node.
752 bool policy_maximally_unroll(PhaseIdealLoop *phase) const;
753
754 // Return TRUE or FALSE if the loop should be unrolled or not. Apply unroll
755 // if the loop is a counted loop and the loop body is small enough.
756 bool policy_unroll(PhaseIdealLoop *phase);
1620
1621 public:
1622 // Change the control input of expensive nodes to allow commoning by
1623 // IGVN when it is guaranteed to not result in a more frequent
1624 // execution of the expensive node. Return true if progress.
1625 bool process_expensive_nodes();
1626
1627 // Check whether node has become unreachable
1628 bool is_node_unreachable(Node *n) const {
1629 return !has_node(n) || n->is_unreachable(_igvn);
1630 }
1631
1632 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
1633 void do_range_check(IdealLoopTree* loop);
1634
1635 // Clone loop with an invariant test (that does not exit) and
1636 // insert a clone of the test that selects which version to
1637 // execute.
1638 void do_unswitching(IdealLoopTree* loop, Node_List& old_new);
1639
1640 IfNode* find_unswitch_candidate(const IdealLoopTree* loop) const;
1641
1642 private:
1643 static bool has_control_dependencies_from_predicates(LoopNode* head);
1644 static void revert_to_normal_loop(const LoopNode* loop_head);
1645
1646 void hoist_invariant_check_casts(const IdealLoopTree* loop, const Node_List& old_new,
1647 const UnswitchedLoopSelector& unswitched_loop_selector);
1648 void add_unswitched_loop_version_bodies_to_igvn(IdealLoopTree* loop, const Node_List& old_new);
1649 static void increment_unswitch_counts(LoopNode* original_head, LoopNode* new_head);
1650 void remove_unswitch_candidate_from_loops(const Node_List& old_new, const UnswitchedLoopSelector& unswitched_loop_selector);
1651 #ifndef PRODUCT
1652 static void trace_loop_unswitching_count(IdealLoopTree* loop, LoopNode* original_head);
1653 static void trace_loop_unswitching_impossible(const LoopNode* original_head);
1654 static void trace_loop_unswitching_result(const UnswitchedLoopSelector& unswitched_loop_selector,
1655 const LoopNode* original_head, const LoopNode* new_head);
1656 static void trace_loop_multiversioning_result(const LoopSelector& loop_selector,
1657 const LoopNode* original_head, const LoopNode* new_head);
1658 #endif
1659
1660 public:
1661
1662 // Range Check Elimination uses this function!
1663 // Constrain the main loop iterations so the affine function:
1664 // low_limit <= scale_con * I + offset < upper_limit
1665 // always holds true. That is, either increase the number of iterations in
1666 // the pre-loop or the post-loop until the condition holds true in the main
1667 // loop. Scale_con, offset and limit are all loop invariant.
1668 void add_constraint(jlong stride_con, jlong scale_con, Node* offset, Node* low_limit, Node* upper_limit, Node* pre_ctrl, Node** pre_limit, Node** main_limit);
1669 // Helper function for add_constraint().
1670 Node* adjust_limit(bool reduce, Node* scale, Node* offset, Node* rc_limit, Node* old_limit, Node* pre_ctrl, bool round);
1671
1672 // Partially peel loop up through last_peel node.
1673 bool partial_peel( IdealLoopTree *loop, Node_List &old_new );
1674 bool duplicate_loop_backedge(IdealLoopTree *loop, Node_List &old_new);
1826 void split_thru_phi_yank_old_nodes(Node* n, Node* region);
1827
1828 public:
1829
1830 // Conversion of fill/copy patterns into intrinsic versions
1831 bool do_intrinsify_fill();
1832 bool intrinsify_fill(IdealLoopTree* lpt);
1833 bool match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
1834 Node*& shift, Node*& offset);
1835
1836 private:
1837 // Helper functions
1838 Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache );
1839 Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true );
1840 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 );
1841 bool split_up( Node *n, Node *blk1, Node *blk2 );
1842
1843 Node* place_outside_loop(Node* useblock, IdealLoopTree* loop) const;
1844 Node* try_move_store_before_loop(Node* n, Node *n_ctrl);
1845 void try_move_store_after_loop(Node* n);
1846 bool identical_backtoback_ifs(Node *n);
1847 bool can_split_if(Node *n_ctrl);
1848 bool cannot_split_division(const Node* n, const Node* region) const;
1849 static bool is_divisor_loop_phi(const Node* divisor, const Node* loop);
1850 bool loop_phi_backedge_type_contains_zero(const Node* phi_divisor, const Type* zero) const;
1851
1852 // Determine if a method is too big for a/another round of split-if, based on
1853 // a magic (approximate) ratio derived from the equally magic constant 35000,
1854 // previously used for this purpose (but without relating to the node limit).
1855 bool must_throttle_split_if() {
1856 uint threshold = C->max_node_limit() * 2 / 5;
1857 return C->live_nodes() > threshold;
1858 }
1859
1860 // A simplistic node request tracking mechanism, where
1861 // = UINT_MAX Request not valid or made final.
1862 // < UINT_MAX Nodes currently requested (estimate).
1863 uint _nodes_required;
1864
1865 enum { REQUIRE_MIN = 70 };
1866
2017 uint new_counter, Node_List& old_new, Node_List& worklist, Node_List*& split_if_set,
2018 Node_List*& split_bool_set, Node_List*& split_cex_set);
2019
2020 void finish_clone_loop(Node_List* split_if_set, Node_List* split_bool_set, Node_List* split_cex_set);
2021
2022 bool at_relevant_ctrl(Node* n, const Node* blk1, const Node* blk2);
2023
2024 bool clone_cmp_loadklass_down(Node* n, const Node* blk1, const Node* blk2);
2025 void clone_loadklass_nodes_at_cmp_index(const Node* n, Node* cmp, int i);
2026 bool clone_cmp_down(Node* n, const Node* blk1, const Node* blk2);
2027 void clone_template_assertion_expression_down(Node* node);
2028
2029 Node* similar_subtype_check(const Node* x, Node* r_in);
2030
2031 void update_addp_chain_base(Node* x, Node* old_base, Node* new_base);
2032
2033 bool can_move_to_inner_loop(Node* n, LoopNode* n_loop, Node* x);
2034
2035 void pin_nodes_dependent_on(Node* ctrl, bool old_iff_is_rangecheck);
2036
2037 Node* ensure_node_and_inputs_are_above_pre_end(CountedLoopEndNode* pre_end, Node* node);
2038
2039 Node* new_assertion_predicate_opaque_init(Node* entry_control, Node* init, Node* int_zero);
2040
2041 bool try_make_short_running_loop(IdealLoopTree* loop, jint stride_con, const Node_List& range_checks, const uint iters_limit);
2042
2043 ConINode* intcon(jint i);
2044
2045 ConLNode* longcon(jlong i);
2046
2047 ConNode* makecon(const Type* t);
2048
2049 ConNode* integercon(jlong l, BasicType bt);
2050
2051 ConNode* zerocon(BasicType bt);
2052 };
2053
2054 class CountedLoopConverter {
2055 friend class PhaseIdealLoop;
2056
|
26 #define SHARE_OPTO_LOOPNODE_HPP
27
28 #include "opto/cfgnode.hpp"
29 #include "opto/multnode.hpp"
30 #include "opto/phaseX.hpp"
31 #include "opto/predicates.hpp"
32 #include "opto/subnode.hpp"
33 #include "opto/type.hpp"
34 #include "utilities/checkedCast.hpp"
35
36 class CmpNode;
37 class BaseCountedLoopEndNode;
38 class CountedLoopNode;
39 class IdealLoopTree;
40 class LoopNode;
41 class Node;
42 class OuterStripMinedLoopEndNode;
43 class PredicateBlock;
44 class PathFrequency;
45 class PhaseIdealLoop;
46 class UnswitchCandidate;
47 class LoopSelector;
48 class UnswitchedLoopSelector;
49 class VectorSet;
50 class VSharedData;
51 class Invariance;
52 struct small_cache;
53
54 //
55 // I D E A L I Z E D L O O P S
56 //
57 // Idealized loops are the set of loops I perform more interesting
58 // transformations on, beyond simple hoisting.
59
60 //------------------------------LoopNode---------------------------------------
61 // Simple loop header. Fall in path on left, loop-back path on right.
62 class LoopNode : public RegionNode {
63 // Size is bigger to hold the flags. However, the flags do not change
64 // the semantics so it does not appear in the hash & cmp functions.
65 virtual uint size_of() const { return sizeof(*this); }
66 protected:
69 enum { Normal=0, Pre=1, Main=2, Post=3, PreMainPostFlagsMask=3,
70 MainHasNoPreLoop = 1<<2,
71 HasExactTripCount = 1<<3,
72 InnerLoop = 1<<4,
73 PartialPeelLoop = 1<<5,
74 PartialPeelFailed = 1<<6,
75 WasSlpAnalyzed = 1<<7,
76 PassedSlpAnalysis = 1<<8,
77 DoUnrollOnly = 1<<9,
78 VectorizedLoop = 1<<10,
79 HasAtomicPostLoop = 1<<11,
80 StripMined = 1<<12,
81 SubwordLoop = 1<<13,
82 ProfileTripFailed = 1<<14,
83 LoopNestInnerLoop = 1<<15,
84 LoopNestLongOuterLoop = 1<<16,
85 MultiversionFastLoop = 1<<17,
86 MultiversionSlowLoop = 2<<17,
87 MultiversionDelayedSlowLoop = 3<<17,
88 MultiversionFlagsMask = 3<<17,
89 FlatArrays = 1<<18};
90 char _unswitch_count;
91 enum { _unswitch_max=3 };
92
93 // Expected trip count from profile data
94 float _profile_trip_cnt;
95
96 public:
97 // Names for edge indices
98 enum { Self=0, EntryControl, LoopBackControl };
99
100 bool is_inner_loop() const { return _loop_flags & InnerLoop; }
101 void set_inner_loop() { _loop_flags |= InnerLoop; }
102
103 bool is_vectorized_loop() const { return _loop_flags & VectorizedLoop; }
104 bool is_partial_peel_loop() const { return _loop_flags & PartialPeelLoop; }
105 void set_partial_peel_loop() { _loop_flags |= PartialPeelLoop; }
106 bool partial_peel_has_failed() const { return _loop_flags & PartialPeelFailed; }
107 bool is_strip_mined() const { return _loop_flags & StripMined; }
108 bool is_profile_trip_failed() const { return _loop_flags & ProfileTripFailed; }
109 bool is_subword_loop() const { return _loop_flags & SubwordLoop; }
110 bool is_loop_nest_inner_loop() const { return _loop_flags & LoopNestInnerLoop; }
111 bool is_loop_nest_outer_loop() const { return _loop_flags & LoopNestLongOuterLoop; }
112 bool is_flat_arrays() const { return _loop_flags & FlatArrays; }
113
114 void mark_partial_peel_failed() { _loop_flags |= PartialPeelFailed; }
115 void mark_was_slp() { _loop_flags |= WasSlpAnalyzed; }
116 void mark_passed_slp() { _loop_flags |= PassedSlpAnalysis; }
117 void mark_do_unroll_only() { _loop_flags |= DoUnrollOnly; }
118 void mark_loop_vectorized() { _loop_flags |= VectorizedLoop; }
119 void mark_has_atomic_post_loop() { _loop_flags |= HasAtomicPostLoop; }
120 void mark_strip_mined() { _loop_flags |= StripMined; }
121 void clear_strip_mined() { _loop_flags &= ~StripMined; }
122 void mark_profile_trip_failed() { _loop_flags |= ProfileTripFailed; }
123 void mark_subword_loop() { _loop_flags |= SubwordLoop; }
124 void mark_loop_nest_inner_loop() { _loop_flags |= LoopNestInnerLoop; }
125 void mark_loop_nest_outer_loop() { _loop_flags |= LoopNestLongOuterLoop; }
126 void mark_flat_arrays() { _loop_flags |= FlatArrays; }
127
128 int unswitch_max() { return _unswitch_max; }
129 int unswitch_count() { return _unswitch_count; }
130
131 void set_unswitch_count(int val) {
132 assert (val <= unswitch_max(), "too many unswitches");
133 _unswitch_count = val;
134 }
135
136 void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; }
137 float profile_trip_cnt() { return _profile_trip_cnt; }
138
139 #ifndef PRODUCT
140 uint _stress_peeling_attempts = 0;
141 #endif
142
143 LoopNode(Node *entry, Node *backedge)
144 : RegionNode(3), _loop_flags(0), _unswitch_count(0),
145 _profile_trip_cnt(COUNT_UNKNOWN) {
146 init_class_id(Class_Loop);
720 // Convert to counted loops where possible
721 void counted_loop( PhaseIdealLoop *phase );
722
723 // Check for Node being a loop-breaking test
724 Node *is_loop_exit(Node *iff) const;
725
726 // Remove simplistic dead code from loop body
727 void DCE_loop_body();
728
729 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
730 // Replace with a 1-in-10 exit guess.
731 void adjust_loop_exit_prob( PhaseIdealLoop *phase );
732
733 // Return TRUE or FALSE if the loop should never be RCE'd or aligned.
734 // Useful for unrolling loops with NO array accesses.
735 bool policy_peel_only( PhaseIdealLoop *phase ) const;
736
737 // Return TRUE or FALSE if the loop should be unswitched -- clone
738 // loop with an invariant test
739 bool policy_unswitching( PhaseIdealLoop *phase ) const;
740 bool no_unswitch_candidate() const;
741
742 // Micro-benchmark spamming. Remove empty loops.
743 bool do_remove_empty_loop( PhaseIdealLoop *phase );
744
745 // Convert one-iteration loop into normal code.
746 bool do_one_iteration_loop( PhaseIdealLoop *phase );
747
748 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
749 // move some loop-invariant test (usually a null-check) before the loop.
750 bool policy_peeling(PhaseIdealLoop *phase);
751
752 uint estimate_peeling(PhaseIdealLoop *phase);
753
754 // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any
755 // known trip count in the counted loop node.
756 bool policy_maximally_unroll(PhaseIdealLoop *phase) const;
757
758 // Return TRUE or FALSE if the loop should be unrolled or not. Apply unroll
759 // if the loop is a counted loop and the loop body is small enough.
760 bool policy_unroll(PhaseIdealLoop *phase);
1624
1625 public:
1626 // Change the control input of expensive nodes to allow commoning by
1627 // IGVN when it is guaranteed to not result in a more frequent
1628 // execution of the expensive node. Return true if progress.
1629 bool process_expensive_nodes();
1630
1631 // Check whether node has become unreachable
1632 bool is_node_unreachable(Node *n) const {
1633 return !has_node(n) || n->is_unreachable(_igvn);
1634 }
1635
1636 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
1637 void do_range_check(IdealLoopTree* loop);
1638
1639 // Clone loop with an invariant test (that does not exit) and
1640 // insert a clone of the test that selects which version to
1641 // execute.
1642 void do_unswitching(IdealLoopTree* loop, Node_List& old_new);
1643
1644 IfNode* find_unswitch_candidates(const IdealLoopTree* loop, Node_List& flat_array_checks) const;
1645 IfNode* find_unswitch_candidate_from_idoms(const IdealLoopTree* loop) const;
1646
1647 private:
1648 static bool has_control_dependencies_from_predicates(LoopNode* head);
1649 static void revert_to_normal_loop(const LoopNode* loop_head);
1650
1651 void hoist_invariant_check_casts(const IdealLoopTree* loop, const Node_List& old_new,
1652 const UnswitchCandidate& unswitch_candidate, const IfNode* loop_selector);
1653 void add_unswitched_loop_version_bodies_to_igvn(IdealLoopTree* loop, const Node_List& old_new);
1654 static void increment_unswitch_counts(LoopNode* original_head, LoopNode* new_head);
1655 void remove_unswitch_candidate_from_loops(const Node_List& old_new, const UnswitchedLoopSelector& unswitched_loop_selector);
1656 #ifndef PRODUCT
1657 static void trace_loop_unswitching_count(IdealLoopTree* loop, LoopNode* original_head);
1658 static void trace_loop_unswitching_impossible(const LoopNode* original_head);
1659 static void trace_loop_unswitching_result(const UnswitchedLoopSelector& unswitched_loop_selector,
1660 const UnswitchCandidate& unswitch_candidate,
1661 const LoopNode* original_head, const LoopNode* new_head);
1662 static void trace_loop_multiversioning_result(const LoopSelector& loop_selector,
1663 const LoopNode* original_head, const LoopNode* new_head);
1664 #endif
1665
1666 public:
1667
1668 // Range Check Elimination uses this function!
1669 // Constrain the main loop iterations so the affine function:
1670 // low_limit <= scale_con * I + offset < upper_limit
1671 // always holds true. That is, either increase the number of iterations in
1672 // the pre-loop or the post-loop until the condition holds true in the main
1673 // loop. Scale_con, offset and limit are all loop invariant.
1674 void add_constraint(jlong stride_con, jlong scale_con, Node* offset, Node* low_limit, Node* upper_limit, Node* pre_ctrl, Node** pre_limit, Node** main_limit);
1675 // Helper function for add_constraint().
1676 Node* adjust_limit(bool reduce, Node* scale, Node* offset, Node* rc_limit, Node* old_limit, Node* pre_ctrl, bool round);
1677
1678 // Partially peel loop up through last_peel node.
1679 bool partial_peel( IdealLoopTree *loop, Node_List &old_new );
1680 bool duplicate_loop_backedge(IdealLoopTree *loop, Node_List &old_new);
1832 void split_thru_phi_yank_old_nodes(Node* n, Node* region);
1833
1834 public:
1835
1836 // Conversion of fill/copy patterns into intrinsic versions
1837 bool do_intrinsify_fill();
1838 bool intrinsify_fill(IdealLoopTree* lpt);
1839 bool match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
1840 Node*& shift, Node*& offset);
1841
1842 private:
1843 // Helper functions
1844 Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache );
1845 Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true );
1846 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 );
1847 bool split_up( Node *n, Node *blk1, Node *blk2 );
1848
1849 Node* place_outside_loop(Node* useblock, IdealLoopTree* loop) const;
1850 Node* try_move_store_before_loop(Node* n, Node *n_ctrl);
1851 void try_move_store_after_loop(Node* n);
1852 void move_flat_array_check_out_of_loop(Node* n);
1853 bool identical_backtoback_ifs(Node *n);
1854 bool flat_array_element_type_check(Node *n);
1855 bool can_split_if(Node *n_ctrl);
1856 bool cannot_split_division(const Node* n, const Node* region) const;
1857 static bool is_divisor_loop_phi(const Node* divisor, const Node* loop);
1858 bool loop_phi_backedge_type_contains_zero(const Node* phi_divisor, const Type* zero) const;
1859
1860 // Determine if a method is too big for a/another round of split-if, based on
1861 // a magic (approximate) ratio derived from the equally magic constant 35000,
1862 // previously used for this purpose (but without relating to the node limit).
1863 bool must_throttle_split_if() {
1864 uint threshold = C->max_node_limit() * 2 / 5;
1865 return C->live_nodes() > threshold;
1866 }
1867
1868 // A simplistic node request tracking mechanism, where
1869 // = UINT_MAX Request not valid or made final.
1870 // < UINT_MAX Nodes currently requested (estimate).
1871 uint _nodes_required;
1872
1873 enum { REQUIRE_MIN = 70 };
1874
2025 uint new_counter, Node_List& old_new, Node_List& worklist, Node_List*& split_if_set,
2026 Node_List*& split_bool_set, Node_List*& split_cex_set);
2027
2028 void finish_clone_loop(Node_List* split_if_set, Node_List* split_bool_set, Node_List* split_cex_set);
2029
2030 bool at_relevant_ctrl(Node* n, const Node* blk1, const Node* blk2);
2031
2032 bool clone_cmp_loadklass_down(Node* n, const Node* blk1, const Node* blk2);
2033 void clone_loadklass_nodes_at_cmp_index(const Node* n, Node* cmp, int i);
2034 bool clone_cmp_down(Node* n, const Node* blk1, const Node* blk2);
2035 void clone_template_assertion_expression_down(Node* node);
2036
2037 Node* similar_subtype_check(const Node* x, Node* r_in);
2038
2039 void update_addp_chain_base(Node* x, Node* old_base, Node* new_base);
2040
2041 bool can_move_to_inner_loop(Node* n, LoopNode* n_loop, Node* x);
2042
2043 void pin_nodes_dependent_on(Node* ctrl, bool old_iff_is_rangecheck);
2044
2045 void collect_flat_array_checks(const IdealLoopTree* loop, Node_List& flat_array_checks) const;
2046
2047 Node* ensure_node_and_inputs_are_above_pre_end(CountedLoopEndNode* pre_end, Node* node);
2048
2049 Node* new_assertion_predicate_opaque_init(Node* entry_control, Node* init, Node* int_zero);
2050
2051 bool try_make_short_running_loop(IdealLoopTree* loop, jint stride_con, const Node_List& range_checks, const uint iters_limit);
2052
2053 ConINode* intcon(jint i);
2054
2055 ConLNode* longcon(jlong i);
2056
2057 ConNode* makecon(const Type* t);
2058
2059 ConNode* integercon(jlong l, BasicType bt);
2060
2061 ConNode* zerocon(BasicType bt);
2062 };
2063
2064 class CountedLoopConverter {
2065 friend class PhaseIdealLoop;
2066
|