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 LoopNode(Node *entry, Node *backedge)
137 : RegionNode(3), _loop_flags(0), _unswitch_count(0),
138 _profile_trip_cnt(COUNT_UNKNOWN) {
139 init_class_id(Class_Loop);
140 init_req(EntryControl, entry);
141 init_req(LoopBackControl, backedge);
142 }
143
709 // Convert to counted loops where possible
710 void counted_loop( PhaseIdealLoop *phase );
711
712 // Check for Node being a loop-breaking test
713 Node *is_loop_exit(Node *iff) const;
714
715 // Remove simplistic dead code from loop body
716 void DCE_loop_body();
717
718 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
719 // Replace with a 1-in-10 exit guess.
720 void adjust_loop_exit_prob( PhaseIdealLoop *phase );
721
722 // Return TRUE or FALSE if the loop should never be RCE'd or aligned.
723 // Useful for unrolling loops with NO array accesses.
724 bool policy_peel_only( PhaseIdealLoop *phase ) const;
725
726 // Return TRUE or FALSE if the loop should be unswitched -- clone
727 // loop with an invariant test
728 bool policy_unswitching( PhaseIdealLoop *phase ) const;
729
730 // Micro-benchmark spamming. Remove empty loops.
731 bool do_remove_empty_loop( PhaseIdealLoop *phase );
732
733 // Convert one iteration loop into normal code.
734 bool do_one_iteration_loop( PhaseIdealLoop *phase );
735
736 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
737 // move some loop-invariant test (usually a null-check) before the loop.
738 bool policy_peeling(PhaseIdealLoop *phase);
739
740 uint estimate_peeling(PhaseIdealLoop *phase);
741
742 // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any
743 // known trip count in the counted loop node.
744 bool policy_maximally_unroll(PhaseIdealLoop *phase) const;
745
746 // Return TRUE or FALSE if the loop should be unrolled or not. Apply unroll
747 // if the loop is a counted loop and the loop body is small enough.
748 bool policy_unroll(PhaseIdealLoop *phase);
1454
1455 public:
1456 // Change the control input of expensive nodes to allow commoning by
1457 // IGVN when it is guaranteed to not result in a more frequent
1458 // execution of the expensive node. Return true if progress.
1459 bool process_expensive_nodes();
1460
1461 // Check whether node has become unreachable
1462 bool is_node_unreachable(Node *n) const {
1463 return !has_node(n) || n->is_unreachable(_igvn);
1464 }
1465
1466 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
1467 void do_range_check(IdealLoopTree* loop);
1468
1469 // Clone loop with an invariant test (that does not exit) and
1470 // insert a clone of the test that selects which version to
1471 // execute.
1472 void do_unswitching(IdealLoopTree* loop, Node_List& old_new);
1473
1474 IfNode* find_unswitch_candidate(const IdealLoopTree* loop) const;
1475
1476 private:
1477 static bool has_control_dependencies_from_predicates(LoopNode* head);
1478 static void revert_to_normal_loop(const LoopNode* loop_head);
1479
1480 void hoist_invariant_check_casts(const IdealLoopTree* loop, const Node_List& old_new,
1481 const UnswitchedLoopSelector& unswitched_loop_selector);
1482 void add_unswitched_loop_version_bodies_to_igvn(IdealLoopTree* loop, const Node_List& old_new);
1483 static void increment_unswitch_counts(LoopNode* original_head, LoopNode* new_head);
1484 void remove_unswitch_candidate_from_loops(const Node_List& old_new, const UnswitchedLoopSelector& unswitched_loop_selector);
1485 #ifndef PRODUCT
1486 static void trace_loop_unswitching_count(IdealLoopTree* loop, LoopNode* original_head);
1487 static void trace_loop_unswitching_impossible(const LoopNode* original_head);
1488 static void trace_loop_unswitching_result(const UnswitchedLoopSelector& unswitched_loop_selector,
1489 const LoopNode* original_head, const LoopNode* new_head);
1490 static void trace_loop_multiversioning_result(const LoopSelector& loop_selector,
1491 const LoopNode* original_head, const LoopNode* new_head);
1492 #endif
1493
1494 public:
1495
1496 // Range Check Elimination uses this function!
1497 // Constrain the main loop iterations so the affine function:
1498 // low_limit <= scale_con * I + offset < upper_limit
1499 // always holds true. That is, either increase the number of iterations in
1500 // the pre-loop or the post-loop until the condition holds true in the main
1501 // loop. Scale_con, offset and limit are all loop invariant.
1502 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);
1503 // Helper function for add_constraint().
1504 Node* adjust_limit(bool reduce, Node* scale, Node* offset, Node* rc_limit, Node* old_limit, Node* pre_ctrl, bool round);
1505
1506 // Partially peel loop up through last_peel node.
1507 bool partial_peel( IdealLoopTree *loop, Node_List &old_new );
1508 bool duplicate_loop_backedge(IdealLoopTree *loop, Node_List &old_new);
1608 bool intrinsify_fill(IdealLoopTree* lpt);
1609 bool match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
1610 Node*& shift, Node*& offset);
1611
1612 private:
1613 // Return a type based on condition control flow
1614 const TypeInt* filtered_type( Node *n, Node* n_ctrl);
1615 const TypeInt* filtered_type( Node *n ) { return filtered_type(n, nullptr); }
1616 // Helpers for filtered type
1617 const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl);
1618
1619 // Helper functions
1620 Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache );
1621 Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true );
1622 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 );
1623 bool split_up( Node *n, Node *blk1, Node *blk2 );
1624
1625 Node* place_outside_loop(Node* useblock, IdealLoopTree* loop) const;
1626 Node* try_move_store_before_loop(Node* n, Node *n_ctrl);
1627 void try_move_store_after_loop(Node* n);
1628 bool identical_backtoback_ifs(Node *n);
1629 bool can_split_if(Node *n_ctrl);
1630 bool cannot_split_division(const Node* n, const Node* region) const;
1631 static bool is_divisor_loop_phi(const Node* divisor, const Node* loop);
1632 bool loop_phi_backedge_type_contains_zero(const Node* phi_divisor, const Type* zero) const;
1633
1634 // Determine if a method is too big for a/another round of split-if, based on
1635 // a magic (approximate) ratio derived from the equally magic constant 35000,
1636 // previously used for this purpose (but without relating to the node limit).
1637 bool must_throttle_split_if() {
1638 uint threshold = C->max_node_limit() * 2 / 5;
1639 return C->live_nodes() > threshold;
1640 }
1641
1642 // A simplistic node request tracking mechanism, where
1643 // = UINT_MAX Request not valid or made final.
1644 // < UINT_MAX Nodes currently requested (estimate).
1645 uint _nodes_required;
1646
1647 enum { REQUIRE_MIN = 70 };
1648
1798 uint new_counter, Node_List& old_new, Node_List& worklist, Node_List*& split_if_set,
1799 Node_List*& split_bool_set, Node_List*& split_cex_set);
1800
1801 void finish_clone_loop(Node_List* split_if_set, Node_List* split_bool_set, Node_List* split_cex_set);
1802
1803 bool at_relevant_ctrl(Node* n, const Node* blk1, const Node* blk2);
1804
1805 bool clone_cmp_loadklass_down(Node* n, const Node* blk1, const Node* blk2);
1806 void clone_loadklass_nodes_at_cmp_index(const Node* n, Node* cmp, int i);
1807 bool clone_cmp_down(Node* n, const Node* blk1, const Node* blk2);
1808 void clone_template_assertion_expression_down(Node* node);
1809
1810 Node* similar_subtype_check(const Node* x, Node* r_in);
1811
1812 void update_addp_chain_base(Node* x, Node* old_base, Node* new_base);
1813
1814 bool can_move_to_inner_loop(Node* n, LoopNode* n_loop, Node* x);
1815
1816 void pin_array_access_nodes_dependent_on(Node* ctrl);
1817
1818 Node* ensure_node_and_inputs_are_above_pre_end(CountedLoopEndNode* pre_end, Node* node);
1819
1820 ConINode* intcon(jint i);
1821
1822 ConLNode* longcon(jlong i);
1823
1824 ConNode* makecon(const Type* t);
1825
1826 ConNode* integercon(jlong l, BasicType bt);
1827
1828 ConNode* zerocon(BasicType bt);
1829 };
1830
1831
1832 class AutoNodeBudget : public StackObj
1833 {
1834 public:
1835 enum budget_check_t { BUDGET_CHECK, NO_BUDGET_CHECK };
1836
1837 AutoNodeBudget(PhaseIdealLoop* phase, budget_check_t chk = BUDGET_CHECK)
|
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 LoopNode(Node *entry, Node *backedge)
140 : RegionNode(3), _loop_flags(0), _unswitch_count(0),
141 _profile_trip_cnt(COUNT_UNKNOWN) {
142 init_class_id(Class_Loop);
143 init_req(EntryControl, entry);
144 init_req(LoopBackControl, backedge);
145 }
146
712 // Convert to counted loops where possible
713 void counted_loop( PhaseIdealLoop *phase );
714
715 // Check for Node being a loop-breaking test
716 Node *is_loop_exit(Node *iff) const;
717
718 // Remove simplistic dead code from loop body
719 void DCE_loop_body();
720
721 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
722 // Replace with a 1-in-10 exit guess.
723 void adjust_loop_exit_prob( PhaseIdealLoop *phase );
724
725 // Return TRUE or FALSE if the loop should never be RCE'd or aligned.
726 // Useful for unrolling loops with NO array accesses.
727 bool policy_peel_only( PhaseIdealLoop *phase ) const;
728
729 // Return TRUE or FALSE if the loop should be unswitched -- clone
730 // loop with an invariant test
731 bool policy_unswitching( PhaseIdealLoop *phase ) const;
732 bool no_unswitch_candidate() const;
733
734 // Micro-benchmark spamming. Remove empty loops.
735 bool do_remove_empty_loop( PhaseIdealLoop *phase );
736
737 // Convert one iteration loop into normal code.
738 bool do_one_iteration_loop( PhaseIdealLoop *phase );
739
740 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
741 // move some loop-invariant test (usually a null-check) before the loop.
742 bool policy_peeling(PhaseIdealLoop *phase);
743
744 uint estimate_peeling(PhaseIdealLoop *phase);
745
746 // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any
747 // known trip count in the counted loop node.
748 bool policy_maximally_unroll(PhaseIdealLoop *phase) const;
749
750 // Return TRUE or FALSE if the loop should be unrolled or not. Apply unroll
751 // if the loop is a counted loop and the loop body is small enough.
752 bool policy_unroll(PhaseIdealLoop *phase);
1458
1459 public:
1460 // Change the control input of expensive nodes to allow commoning by
1461 // IGVN when it is guaranteed to not result in a more frequent
1462 // execution of the expensive node. Return true if progress.
1463 bool process_expensive_nodes();
1464
1465 // Check whether node has become unreachable
1466 bool is_node_unreachable(Node *n) const {
1467 return !has_node(n) || n->is_unreachable(_igvn);
1468 }
1469
1470 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
1471 void do_range_check(IdealLoopTree* loop);
1472
1473 // Clone loop with an invariant test (that does not exit) and
1474 // insert a clone of the test that selects which version to
1475 // execute.
1476 void do_unswitching(IdealLoopTree* loop, Node_List& old_new);
1477
1478 IfNode* find_unswitch_candidates(const IdealLoopTree* loop, Node_List& flat_array_checks) const;
1479 IfNode* find_unswitch_candidate_from_idoms(const IdealLoopTree* loop) const;
1480
1481 private:
1482 static bool has_control_dependencies_from_predicates(LoopNode* head);
1483 static void revert_to_normal_loop(const LoopNode* loop_head);
1484
1485 void hoist_invariant_check_casts(const IdealLoopTree* loop, const Node_List& old_new,
1486 const UnswitchCandidate& unswitch_candidate, const IfNode* loop_selector);
1487 void add_unswitched_loop_version_bodies_to_igvn(IdealLoopTree* loop, const Node_List& old_new);
1488 static void increment_unswitch_counts(LoopNode* original_head, LoopNode* new_head);
1489 void remove_unswitch_candidate_from_loops(const Node_List& old_new, const UnswitchedLoopSelector& unswitched_loop_selector);
1490 #ifndef PRODUCT
1491 static void trace_loop_unswitching_count(IdealLoopTree* loop, LoopNode* original_head);
1492 static void trace_loop_unswitching_impossible(const LoopNode* original_head);
1493 static void trace_loop_unswitching_result(const UnswitchedLoopSelector& unswitched_loop_selector,
1494 const UnswitchCandidate& unswitch_candidate,
1495 const LoopNode* original_head, const LoopNode* new_head);
1496 static void trace_loop_multiversioning_result(const LoopSelector& loop_selector,
1497 const LoopNode* original_head, const LoopNode* new_head);
1498 #endif
1499
1500 public:
1501
1502 // Range Check Elimination uses this function!
1503 // Constrain the main loop iterations so the affine function:
1504 // low_limit <= scale_con * I + offset < upper_limit
1505 // always holds true. That is, either increase the number of iterations in
1506 // the pre-loop or the post-loop until the condition holds true in the main
1507 // loop. Scale_con, offset and limit are all loop invariant.
1508 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);
1509 // Helper function for add_constraint().
1510 Node* adjust_limit(bool reduce, Node* scale, Node* offset, Node* rc_limit, Node* old_limit, Node* pre_ctrl, bool round);
1511
1512 // Partially peel loop up through last_peel node.
1513 bool partial_peel( IdealLoopTree *loop, Node_List &old_new );
1514 bool duplicate_loop_backedge(IdealLoopTree *loop, Node_List &old_new);
1614 bool intrinsify_fill(IdealLoopTree* lpt);
1615 bool match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
1616 Node*& shift, Node*& offset);
1617
1618 private:
1619 // Return a type based on condition control flow
1620 const TypeInt* filtered_type( Node *n, Node* n_ctrl);
1621 const TypeInt* filtered_type( Node *n ) { return filtered_type(n, nullptr); }
1622 // Helpers for filtered type
1623 const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl);
1624
1625 // Helper functions
1626 Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache );
1627 Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true );
1628 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 );
1629 bool split_up( Node *n, Node *blk1, Node *blk2 );
1630
1631 Node* place_outside_loop(Node* useblock, IdealLoopTree* loop) const;
1632 Node* try_move_store_before_loop(Node* n, Node *n_ctrl);
1633 void try_move_store_after_loop(Node* n);
1634 void move_flat_array_check_out_of_loop(Node* n);
1635 bool identical_backtoback_ifs(Node *n);
1636 bool flat_array_element_type_check(Node *n);
1637 bool can_split_if(Node *n_ctrl);
1638 bool cannot_split_division(const Node* n, const Node* region) const;
1639 static bool is_divisor_loop_phi(const Node* divisor, const Node* loop);
1640 bool loop_phi_backedge_type_contains_zero(const Node* phi_divisor, const Type* zero) const;
1641
1642 // Determine if a method is too big for a/another round of split-if, based on
1643 // a magic (approximate) ratio derived from the equally magic constant 35000,
1644 // previously used for this purpose (but without relating to the node limit).
1645 bool must_throttle_split_if() {
1646 uint threshold = C->max_node_limit() * 2 / 5;
1647 return C->live_nodes() > threshold;
1648 }
1649
1650 // A simplistic node request tracking mechanism, where
1651 // = UINT_MAX Request not valid or made final.
1652 // < UINT_MAX Nodes currently requested (estimate).
1653 uint _nodes_required;
1654
1655 enum { REQUIRE_MIN = 70 };
1656
1806 uint new_counter, Node_List& old_new, Node_List& worklist, Node_List*& split_if_set,
1807 Node_List*& split_bool_set, Node_List*& split_cex_set);
1808
1809 void finish_clone_loop(Node_List* split_if_set, Node_List* split_bool_set, Node_List* split_cex_set);
1810
1811 bool at_relevant_ctrl(Node* n, const Node* blk1, const Node* blk2);
1812
1813 bool clone_cmp_loadklass_down(Node* n, const Node* blk1, const Node* blk2);
1814 void clone_loadklass_nodes_at_cmp_index(const Node* n, Node* cmp, int i);
1815 bool clone_cmp_down(Node* n, const Node* blk1, const Node* blk2);
1816 void clone_template_assertion_expression_down(Node* node);
1817
1818 Node* similar_subtype_check(const Node* x, Node* r_in);
1819
1820 void update_addp_chain_base(Node* x, Node* old_base, Node* new_base);
1821
1822 bool can_move_to_inner_loop(Node* n, LoopNode* n_loop, Node* x);
1823
1824 void pin_array_access_nodes_dependent_on(Node* ctrl);
1825
1826 void collect_flat_array_checks(const IdealLoopTree* loop, Node_List& flat_array_checks) const;
1827
1828 Node* ensure_node_and_inputs_are_above_pre_end(CountedLoopEndNode* pre_end, Node* node);
1829
1830 ConINode* intcon(jint i);
1831
1832 ConLNode* longcon(jlong i);
1833
1834 ConNode* makecon(const Type* t);
1835
1836 ConNode* integercon(jlong l, BasicType bt);
1837
1838 ConNode* zerocon(BasicType bt);
1839 };
1840
1841
1842 class AutoNodeBudget : public StackObj
1843 {
1844 public:
1845 enum budget_check_t { BUDGET_CHECK, NO_BUDGET_CHECK };
1846
1847 AutoNodeBudget(PhaseIdealLoop* phase, budget_check_t chk = BUDGET_CHECK)
|