Wed, 09 Dec 2009 16:40:45 -0800
6895383: JCK test throws NPE for method compiled with Escape Analysis
Summary: Add missing checks for MemBar nodes in EA.
Reviewed-by: never
duke@435 | 1 | /* |
never@1356 | 2 | * Copyright 1998-2009 Sun Microsystems, Inc. All Rights Reserved. |
duke@435 | 3 | * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
duke@435 | 4 | * |
duke@435 | 5 | * This code is free software; you can redistribute it and/or modify it |
duke@435 | 6 | * under the terms of the GNU General Public License version 2 only, as |
duke@435 | 7 | * published by the Free Software Foundation. |
duke@435 | 8 | * |
duke@435 | 9 | * This code is distributed in the hope that it will be useful, but WITHOUT |
duke@435 | 10 | * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or |
duke@435 | 11 | * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
duke@435 | 12 | * version 2 for more details (a copy is included in the LICENSE file that |
duke@435 | 13 | * accompanied this code). |
duke@435 | 14 | * |
duke@435 | 15 | * You should have received a copy of the GNU General Public License version |
duke@435 | 16 | * 2 along with this work; if not, write to the Free Software Foundation, |
duke@435 | 17 | * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. |
duke@435 | 18 | * |
duke@435 | 19 | * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, |
duke@435 | 20 | * CA 95054 USA or visit www.sun.com if you need additional information or |
duke@435 | 21 | * have any questions. |
duke@435 | 22 | * |
duke@435 | 23 | */ |
duke@435 | 24 | |
duke@435 | 25 | class CmpNode; |
duke@435 | 26 | class CountedLoopEndNode; |
duke@435 | 27 | class CountedLoopNode; |
duke@435 | 28 | class IdealLoopTree; |
duke@435 | 29 | class LoopNode; |
duke@435 | 30 | class Node; |
duke@435 | 31 | class PhaseIdealLoop; |
duke@435 | 32 | class VectorSet; |
duke@435 | 33 | struct small_cache; |
duke@435 | 34 | |
duke@435 | 35 | // |
duke@435 | 36 | // I D E A L I Z E D L O O P S |
duke@435 | 37 | // |
duke@435 | 38 | // Idealized loops are the set of loops I perform more interesting |
duke@435 | 39 | // transformations on, beyond simple hoisting. |
duke@435 | 40 | |
duke@435 | 41 | //------------------------------LoopNode--------------------------------------- |
duke@435 | 42 | // Simple loop header. Fall in path on left, loop-back path on right. |
duke@435 | 43 | class LoopNode : public RegionNode { |
duke@435 | 44 | // Size is bigger to hold the flags. However, the flags do not change |
duke@435 | 45 | // the semantics so it does not appear in the hash & cmp functions. |
duke@435 | 46 | virtual uint size_of() const { return sizeof(*this); } |
duke@435 | 47 | protected: |
duke@435 | 48 | short _loop_flags; |
duke@435 | 49 | // Names for flag bitfields |
duke@435 | 50 | enum { pre_post_main=0, inner_loop=8, partial_peel_loop=16, partial_peel_failed=32 }; |
duke@435 | 51 | char _unswitch_count; |
duke@435 | 52 | enum { _unswitch_max=3 }; |
duke@435 | 53 | |
duke@435 | 54 | public: |
duke@435 | 55 | // Names for edge indices |
duke@435 | 56 | enum { Self=0, EntryControl, LoopBackControl }; |
duke@435 | 57 | |
duke@435 | 58 | int is_inner_loop() const { return _loop_flags & inner_loop; } |
duke@435 | 59 | void set_inner_loop() { _loop_flags |= inner_loop; } |
duke@435 | 60 | |
duke@435 | 61 | int is_partial_peel_loop() const { return _loop_flags & partial_peel_loop; } |
duke@435 | 62 | void set_partial_peel_loop() { _loop_flags |= partial_peel_loop; } |
duke@435 | 63 | int partial_peel_has_failed() const { return _loop_flags & partial_peel_failed; } |
duke@435 | 64 | void mark_partial_peel_failed() { _loop_flags |= partial_peel_failed; } |
duke@435 | 65 | |
duke@435 | 66 | int unswitch_max() { return _unswitch_max; } |
duke@435 | 67 | int unswitch_count() { return _unswitch_count; } |
duke@435 | 68 | void set_unswitch_count(int val) { |
duke@435 | 69 | assert (val <= unswitch_max(), "too many unswitches"); |
duke@435 | 70 | _unswitch_count = val; |
duke@435 | 71 | } |
duke@435 | 72 | |
duke@435 | 73 | LoopNode( Node *entry, Node *backedge ) : RegionNode(3), _loop_flags(0), _unswitch_count(0) { |
duke@435 | 74 | init_class_id(Class_Loop); |
duke@435 | 75 | init_req(EntryControl, entry); |
duke@435 | 76 | init_req(LoopBackControl, backedge); |
duke@435 | 77 | } |
duke@435 | 78 | |
duke@435 | 79 | virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); |
duke@435 | 80 | virtual int Opcode() const; |
duke@435 | 81 | bool can_be_counted_loop(PhaseTransform* phase) const { |
duke@435 | 82 | return req() == 3 && in(0) != NULL && |
duke@435 | 83 | in(1) != NULL && phase->type(in(1)) != Type::TOP && |
duke@435 | 84 | in(2) != NULL && phase->type(in(2)) != Type::TOP; |
duke@435 | 85 | } |
duke@435 | 86 | #ifndef PRODUCT |
duke@435 | 87 | virtual void dump_spec(outputStream *st) const; |
duke@435 | 88 | #endif |
duke@435 | 89 | }; |
duke@435 | 90 | |
duke@435 | 91 | //------------------------------Counted Loops---------------------------------- |
duke@435 | 92 | // Counted loops are all trip-counted loops, with exactly 1 trip-counter exit |
duke@435 | 93 | // path (and maybe some other exit paths). The trip-counter exit is always |
duke@435 | 94 | // last in the loop. The trip-counter does not have to stride by a constant, |
duke@435 | 95 | // but it does have to stride by a loop-invariant amount; the exit value is |
duke@435 | 96 | // also loop invariant. |
duke@435 | 97 | |
duke@435 | 98 | // CountedLoopNodes and CountedLoopEndNodes come in matched pairs. The |
duke@435 | 99 | // CountedLoopNode has the incoming loop control and the loop-back-control |
duke@435 | 100 | // which is always the IfTrue before the matching CountedLoopEndNode. The |
duke@435 | 101 | // CountedLoopEndNode has an incoming control (possibly not the |
duke@435 | 102 | // CountedLoopNode if there is control flow in the loop), the post-increment |
duke@435 | 103 | // trip-counter value, and the limit. The trip-counter value is always of |
duke@435 | 104 | // the form (Op old-trip-counter stride). The old-trip-counter is produced |
duke@435 | 105 | // by a Phi connected to the CountedLoopNode. The stride is loop invariant. |
duke@435 | 106 | // The Op is any commutable opcode, including Add, Mul, Xor. The |
duke@435 | 107 | // CountedLoopEndNode also takes in the loop-invariant limit value. |
duke@435 | 108 | |
duke@435 | 109 | // From a CountedLoopNode I can reach the matching CountedLoopEndNode via the |
duke@435 | 110 | // loop-back control. From CountedLoopEndNodes I can reach CountedLoopNodes |
duke@435 | 111 | // via the old-trip-counter from the Op node. |
duke@435 | 112 | |
duke@435 | 113 | //------------------------------CountedLoopNode-------------------------------- |
duke@435 | 114 | // CountedLoopNodes head simple counted loops. CountedLoopNodes have as |
duke@435 | 115 | // inputs the incoming loop-start control and the loop-back control, so they |
duke@435 | 116 | // act like RegionNodes. They also take in the initial trip counter, the |
duke@435 | 117 | // loop-invariant stride and the loop-invariant limit value. CountedLoopNodes |
duke@435 | 118 | // produce a loop-body control and the trip counter value. Since |
duke@435 | 119 | // CountedLoopNodes behave like RegionNodes I still have a standard CFG model. |
duke@435 | 120 | |
duke@435 | 121 | class CountedLoopNode : public LoopNode { |
duke@435 | 122 | // Size is bigger to hold _main_idx. However, _main_idx does not change |
duke@435 | 123 | // the semantics so it does not appear in the hash & cmp functions. |
duke@435 | 124 | virtual uint size_of() const { return sizeof(*this); } |
duke@435 | 125 | |
duke@435 | 126 | // For Pre- and Post-loops during debugging ONLY, this holds the index of |
duke@435 | 127 | // the Main CountedLoop. Used to assert that we understand the graph shape. |
duke@435 | 128 | node_idx_t _main_idx; |
duke@435 | 129 | |
duke@435 | 130 | // Known trip count calculated by policy_maximally_unroll |
duke@435 | 131 | int _trip_count; |
duke@435 | 132 | |
duke@435 | 133 | // Expected trip count from profile data |
duke@435 | 134 | float _profile_trip_cnt; |
duke@435 | 135 | |
duke@435 | 136 | // Log2 of original loop bodies in unrolled loop |
duke@435 | 137 | int _unrolled_count_log2; |
duke@435 | 138 | |
duke@435 | 139 | // Node count prior to last unrolling - used to decide if |
duke@435 | 140 | // unroll,optimize,unroll,optimize,... is making progress |
duke@435 | 141 | int _node_count_before_unroll; |
duke@435 | 142 | |
duke@435 | 143 | public: |
duke@435 | 144 | CountedLoopNode( Node *entry, Node *backedge ) |
duke@435 | 145 | : LoopNode(entry, backedge), _trip_count(max_jint), |
duke@435 | 146 | _profile_trip_cnt(COUNT_UNKNOWN), _unrolled_count_log2(0), |
duke@435 | 147 | _node_count_before_unroll(0) { |
duke@435 | 148 | init_class_id(Class_CountedLoop); |
duke@435 | 149 | // Initialize _trip_count to the largest possible value. |
duke@435 | 150 | // Will be reset (lower) if the loop's trip count is known. |
duke@435 | 151 | } |
duke@435 | 152 | |
duke@435 | 153 | virtual int Opcode() const; |
duke@435 | 154 | virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); |
duke@435 | 155 | |
duke@435 | 156 | Node *init_control() const { return in(EntryControl); } |
duke@435 | 157 | Node *back_control() const { return in(LoopBackControl); } |
duke@435 | 158 | CountedLoopEndNode *loopexit() const; |
duke@435 | 159 | Node *init_trip() const; |
duke@435 | 160 | Node *stride() const; |
duke@435 | 161 | int stride_con() const; |
duke@435 | 162 | bool stride_is_con() const; |
duke@435 | 163 | Node *limit() const; |
duke@435 | 164 | Node *incr() const; |
duke@435 | 165 | Node *phi() const; |
duke@435 | 166 | |
duke@435 | 167 | // Match increment with optional truncation |
duke@435 | 168 | static Node* match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type); |
duke@435 | 169 | |
duke@435 | 170 | // A 'main' loop has a pre-loop and a post-loop. The 'main' loop |
duke@435 | 171 | // can run short a few iterations and may start a few iterations in. |
duke@435 | 172 | // It will be RCE'd and unrolled and aligned. |
duke@435 | 173 | |
duke@435 | 174 | // A following 'post' loop will run any remaining iterations. Used |
duke@435 | 175 | // during Range Check Elimination, the 'post' loop will do any final |
duke@435 | 176 | // iterations with full checks. Also used by Loop Unrolling, where |
duke@435 | 177 | // the 'post' loop will do any epilog iterations needed. Basically, |
duke@435 | 178 | // a 'post' loop can not profitably be further unrolled or RCE'd. |
duke@435 | 179 | |
duke@435 | 180 | // A preceding 'pre' loop will run at least 1 iteration (to do peeling), |
duke@435 | 181 | // it may do under-flow checks for RCE and may do alignment iterations |
duke@435 | 182 | // so the following main loop 'knows' that it is striding down cache |
duke@435 | 183 | // lines. |
duke@435 | 184 | |
duke@435 | 185 | // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or |
duke@435 | 186 | // Aligned, may be missing it's pre-loop. |
duke@435 | 187 | enum { Normal=0, Pre=1, Main=2, Post=3, PrePostFlagsMask=3, Main_Has_No_Pre_Loop=4 }; |
duke@435 | 188 | int is_normal_loop() const { return (_loop_flags&PrePostFlagsMask) == Normal; } |
duke@435 | 189 | int is_pre_loop () const { return (_loop_flags&PrePostFlagsMask) == Pre; } |
duke@435 | 190 | int is_main_loop () const { return (_loop_flags&PrePostFlagsMask) == Main; } |
duke@435 | 191 | int is_post_loop () const { return (_loop_flags&PrePostFlagsMask) == Post; } |
duke@435 | 192 | int is_main_no_pre_loop() const { return _loop_flags & Main_Has_No_Pre_Loop; } |
duke@435 | 193 | void set_main_no_pre_loop() { _loop_flags |= Main_Has_No_Pre_Loop; } |
duke@435 | 194 | |
never@802 | 195 | int main_idx() const { return _main_idx; } |
never@802 | 196 | |
duke@435 | 197 | |
duke@435 | 198 | void set_pre_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; } |
duke@435 | 199 | void set_main_loop ( ) { assert(is_normal_loop(),""); _loop_flags |= Main; } |
duke@435 | 200 | void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; } |
duke@435 | 201 | void set_normal_loop( ) { _loop_flags &= ~PrePostFlagsMask; } |
duke@435 | 202 | |
duke@435 | 203 | void set_trip_count(int tc) { _trip_count = tc; } |
duke@435 | 204 | int trip_count() { return _trip_count; } |
duke@435 | 205 | |
duke@435 | 206 | void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; } |
duke@435 | 207 | float profile_trip_cnt() { return _profile_trip_cnt; } |
duke@435 | 208 | |
duke@435 | 209 | void double_unrolled_count() { _unrolled_count_log2++; } |
duke@435 | 210 | int unrolled_count() { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); } |
duke@435 | 211 | |
duke@435 | 212 | void set_node_count_before_unroll(int ct) { _node_count_before_unroll = ct; } |
duke@435 | 213 | int node_count_before_unroll() { return _node_count_before_unroll; } |
duke@435 | 214 | |
duke@435 | 215 | #ifndef PRODUCT |
duke@435 | 216 | virtual void dump_spec(outputStream *st) const; |
duke@435 | 217 | #endif |
duke@435 | 218 | }; |
duke@435 | 219 | |
duke@435 | 220 | //------------------------------CountedLoopEndNode----------------------------- |
duke@435 | 221 | // CountedLoopEndNodes end simple trip counted loops. They act much like |
duke@435 | 222 | // IfNodes. |
duke@435 | 223 | class CountedLoopEndNode : public IfNode { |
duke@435 | 224 | public: |
duke@435 | 225 | enum { TestControl, TestValue }; |
duke@435 | 226 | |
duke@435 | 227 | CountedLoopEndNode( Node *control, Node *test, float prob, float cnt ) |
duke@435 | 228 | : IfNode( control, test, prob, cnt) { |
duke@435 | 229 | init_class_id(Class_CountedLoopEnd); |
duke@435 | 230 | } |
duke@435 | 231 | virtual int Opcode() const; |
duke@435 | 232 | |
duke@435 | 233 | Node *cmp_node() const { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : NULL; } |
duke@435 | 234 | Node *incr() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; } |
duke@435 | 235 | Node *limit() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; } |
duke@435 | 236 | Node *stride() const { Node *tmp = incr (); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; } |
duke@435 | 237 | Node *phi() const { Node *tmp = incr (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; } |
duke@435 | 238 | Node *init_trip() const { Node *tmp = phi (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; } |
duke@435 | 239 | int stride_con() const; |
duke@435 | 240 | bool stride_is_con() const { Node *tmp = stride (); return (tmp != NULL && tmp->is_Con()); } |
duke@435 | 241 | BoolTest::mask test_trip() const { return in(TestValue)->as_Bool()->_test._test; } |
duke@435 | 242 | CountedLoopNode *loopnode() const { |
duke@435 | 243 | Node *ln = phi()->in(0); |
duke@435 | 244 | assert( ln->Opcode() == Op_CountedLoop, "malformed loop" ); |
duke@435 | 245 | return (CountedLoopNode*)ln; } |
duke@435 | 246 | |
duke@435 | 247 | #ifndef PRODUCT |
duke@435 | 248 | virtual void dump_spec(outputStream *st) const; |
duke@435 | 249 | #endif |
duke@435 | 250 | }; |
duke@435 | 251 | |
duke@435 | 252 | |
duke@435 | 253 | inline CountedLoopEndNode *CountedLoopNode::loopexit() const { |
duke@435 | 254 | Node *bc = back_control(); |
duke@435 | 255 | if( bc == NULL ) return NULL; |
duke@435 | 256 | Node *le = bc->in(0); |
duke@435 | 257 | if( le->Opcode() != Op_CountedLoopEnd ) |
duke@435 | 258 | return NULL; |
duke@435 | 259 | return (CountedLoopEndNode*)le; |
duke@435 | 260 | } |
duke@435 | 261 | inline Node *CountedLoopNode::init_trip() const { return loopexit() ? loopexit()->init_trip() : NULL; } |
duke@435 | 262 | inline Node *CountedLoopNode::stride() const { return loopexit() ? loopexit()->stride() : NULL; } |
duke@435 | 263 | inline int CountedLoopNode::stride_con() const { return loopexit() ? loopexit()->stride_con() : 0; } |
duke@435 | 264 | inline bool CountedLoopNode::stride_is_con() const { return loopexit() && loopexit()->stride_is_con(); } |
duke@435 | 265 | inline Node *CountedLoopNode::limit() const { return loopexit() ? loopexit()->limit() : NULL; } |
duke@435 | 266 | inline Node *CountedLoopNode::incr() const { return loopexit() ? loopexit()->incr() : NULL; } |
duke@435 | 267 | inline Node *CountedLoopNode::phi() const { return loopexit() ? loopexit()->phi() : NULL; } |
duke@435 | 268 | |
duke@435 | 269 | |
duke@435 | 270 | // -----------------------------IdealLoopTree---------------------------------- |
duke@435 | 271 | class IdealLoopTree : public ResourceObj { |
duke@435 | 272 | public: |
duke@435 | 273 | IdealLoopTree *_parent; // Parent in loop tree |
duke@435 | 274 | IdealLoopTree *_next; // Next sibling in loop tree |
duke@435 | 275 | IdealLoopTree *_child; // First child in loop tree |
duke@435 | 276 | |
duke@435 | 277 | // The head-tail backedge defines the loop. |
duke@435 | 278 | // If tail is NULL then this loop has multiple backedges as part of the |
duke@435 | 279 | // same loop. During cleanup I'll peel off the multiple backedges; merge |
duke@435 | 280 | // them at the loop bottom and flow 1 real backedge into the loop. |
duke@435 | 281 | Node *_head; // Head of loop |
duke@435 | 282 | Node *_tail; // Tail of loop |
duke@435 | 283 | inline Node *tail(); // Handle lazy update of _tail field |
duke@435 | 284 | PhaseIdealLoop* _phase; |
duke@435 | 285 | |
duke@435 | 286 | Node_List _body; // Loop body for inner loops |
duke@435 | 287 | |
duke@435 | 288 | uint8 _nest; // Nesting depth |
duke@435 | 289 | uint8 _irreducible:1, // True if irreducible |
duke@435 | 290 | _has_call:1, // True if has call safepoint |
duke@435 | 291 | _has_sfpt:1, // True if has non-call safepoint |
duke@435 | 292 | _rce_candidate:1; // True if candidate for range check elimination |
duke@435 | 293 | |
kvn@474 | 294 | Node_List* _required_safept; // A inner loop cannot delete these safepts; |
kvn@474 | 295 | bool _allow_optimizations; // Allow loop optimizations |
duke@435 | 296 | |
duke@435 | 297 | IdealLoopTree( PhaseIdealLoop* phase, Node *head, Node *tail ) |
duke@435 | 298 | : _parent(0), _next(0), _child(0), |
duke@435 | 299 | _head(head), _tail(tail), |
duke@435 | 300 | _phase(phase), |
duke@435 | 301 | _required_safept(NULL), |
kvn@474 | 302 | _allow_optimizations(true), |
duke@435 | 303 | _nest(0), _irreducible(0), _has_call(0), _has_sfpt(0), _rce_candidate(0) |
duke@435 | 304 | { } |
duke@435 | 305 | |
duke@435 | 306 | // Is 'l' a member of 'this'? |
duke@435 | 307 | int is_member( const IdealLoopTree *l ) const; // Test for nested membership |
duke@435 | 308 | |
duke@435 | 309 | // Set loop nesting depth. Accumulate has_call bits. |
duke@435 | 310 | int set_nest( uint depth ); |
duke@435 | 311 | |
duke@435 | 312 | // Split out multiple fall-in edges from the loop header. Move them to a |
duke@435 | 313 | // private RegionNode before the loop. This becomes the loop landing pad. |
duke@435 | 314 | void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ); |
duke@435 | 315 | |
duke@435 | 316 | // Split out the outermost loop from this shared header. |
duke@435 | 317 | void split_outer_loop( PhaseIdealLoop *phase ); |
duke@435 | 318 | |
duke@435 | 319 | // Merge all the backedges from the shared header into a private Region. |
duke@435 | 320 | // Feed that region as the one backedge to this loop. |
duke@435 | 321 | void merge_many_backedges( PhaseIdealLoop *phase ); |
duke@435 | 322 | |
duke@435 | 323 | // Split shared headers and insert loop landing pads. |
duke@435 | 324 | // Insert a LoopNode to replace the RegionNode. |
duke@435 | 325 | // Returns TRUE if loop tree is structurally changed. |
duke@435 | 326 | bool beautify_loops( PhaseIdealLoop *phase ); |
duke@435 | 327 | |
never@836 | 328 | // Perform iteration-splitting on inner loops. Split iterations to |
never@836 | 329 | // avoid range checks or one-shot null checks. Returns false if the |
never@836 | 330 | // current round of loop opts should stop. |
never@836 | 331 | bool iteration_split( PhaseIdealLoop *phase, Node_List &old_new ); |
duke@435 | 332 | |
never@836 | 333 | // Driver for various flavors of iteration splitting. Returns false |
never@836 | 334 | // if the current round of loop opts should stop. |
never@836 | 335 | bool iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ); |
duke@435 | 336 | |
duke@435 | 337 | // Given dominators, try to find loops with calls that must always be |
duke@435 | 338 | // executed (call dominates loop tail). These loops do not need non-call |
duke@435 | 339 | // safepoints (ncsfpt). |
duke@435 | 340 | void check_safepts(VectorSet &visited, Node_List &stack); |
duke@435 | 341 | |
duke@435 | 342 | // Allpaths backwards scan from loop tail, terminating each path at first safepoint |
duke@435 | 343 | // encountered. |
duke@435 | 344 | void allpaths_check_safepts(VectorSet &visited, Node_List &stack); |
duke@435 | 345 | |
duke@435 | 346 | // Convert to counted loops where possible |
duke@435 | 347 | void counted_loop( PhaseIdealLoop *phase ); |
duke@435 | 348 | |
duke@435 | 349 | // Check for Node being a loop-breaking test |
duke@435 | 350 | Node *is_loop_exit(Node *iff) const; |
duke@435 | 351 | |
duke@435 | 352 | // Returns true if ctrl is executed on every complete iteration |
duke@435 | 353 | bool dominates_backedge(Node* ctrl); |
duke@435 | 354 | |
duke@435 | 355 | // Remove simplistic dead code from loop body |
duke@435 | 356 | void DCE_loop_body(); |
duke@435 | 357 | |
duke@435 | 358 | // Look for loop-exit tests with my 50/50 guesses from the Parsing stage. |
duke@435 | 359 | // Replace with a 1-in-10 exit guess. |
duke@435 | 360 | void adjust_loop_exit_prob( PhaseIdealLoop *phase ); |
duke@435 | 361 | |
duke@435 | 362 | // Return TRUE or FALSE if the loop should never be RCE'd or aligned. |
duke@435 | 363 | // Useful for unrolling loops with NO array accesses. |
duke@435 | 364 | bool policy_peel_only( PhaseIdealLoop *phase ) const; |
duke@435 | 365 | |
duke@435 | 366 | // Return TRUE or FALSE if the loop should be unswitched -- clone |
duke@435 | 367 | // loop with an invariant test |
duke@435 | 368 | bool policy_unswitching( PhaseIdealLoop *phase ) const; |
duke@435 | 369 | |
duke@435 | 370 | // Micro-benchmark spamming. Remove empty loops. |
duke@435 | 371 | bool policy_do_remove_empty_loop( PhaseIdealLoop *phase ); |
duke@435 | 372 | |
duke@435 | 373 | // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can |
duke@435 | 374 | // make some loop-invariant test (usually a null-check) happen before the |
duke@435 | 375 | // loop. |
duke@435 | 376 | bool policy_peeling( PhaseIdealLoop *phase ) const; |
duke@435 | 377 | |
duke@435 | 378 | // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any |
duke@435 | 379 | // known trip count in the counted loop node. |
duke@435 | 380 | bool policy_maximally_unroll( PhaseIdealLoop *phase ) const; |
duke@435 | 381 | |
duke@435 | 382 | // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if |
duke@435 | 383 | // the loop is a CountedLoop and the body is small enough. |
duke@435 | 384 | bool policy_unroll( PhaseIdealLoop *phase ) const; |
duke@435 | 385 | |
duke@435 | 386 | // Return TRUE or FALSE if the loop should be range-check-eliminated. |
duke@435 | 387 | // Gather a list of IF tests that are dominated by iteration splitting; |
duke@435 | 388 | // also gather the end of the first split and the start of the 2nd split. |
duke@435 | 389 | bool policy_range_check( PhaseIdealLoop *phase ) const; |
duke@435 | 390 | |
duke@435 | 391 | // Return TRUE or FALSE if the loop should be cache-line aligned. |
duke@435 | 392 | // Gather the expression that does the alignment. Note that only |
twisti@1040 | 393 | // one array base can be aligned in a loop (unless the VM guarantees |
duke@435 | 394 | // mutual alignment). Note that if we vectorize short memory ops |
duke@435 | 395 | // into longer memory ops, we may want to increase alignment. |
duke@435 | 396 | bool policy_align( PhaseIdealLoop *phase ) const; |
duke@435 | 397 | |
duke@435 | 398 | // Compute loop trip count from profile data |
duke@435 | 399 | void compute_profile_trip_cnt( PhaseIdealLoop *phase ); |
duke@435 | 400 | |
duke@435 | 401 | // Reassociate invariant expressions. |
duke@435 | 402 | void reassociate_invariants(PhaseIdealLoop *phase); |
duke@435 | 403 | // Reassociate invariant add and subtract expressions. |
duke@435 | 404 | Node* reassociate_add_sub(Node* n1, PhaseIdealLoop *phase); |
duke@435 | 405 | // Return nonzero index of invariant operand if invariant and variant |
twisti@1040 | 406 | // are combined with an Add or Sub. Helper for reassociate_invariants. |
duke@435 | 407 | int is_invariant_addition(Node* n, PhaseIdealLoop *phase); |
duke@435 | 408 | |
duke@435 | 409 | // Return true if n is invariant |
duke@435 | 410 | bool is_invariant(Node* n) const; |
duke@435 | 411 | |
duke@435 | 412 | // Put loop body on igvn work list |
duke@435 | 413 | void record_for_igvn(); |
duke@435 | 414 | |
duke@435 | 415 | bool is_loop() { return !_irreducible && _tail && !_tail->is_top(); } |
duke@435 | 416 | bool is_inner() { return is_loop() && _child == NULL; } |
duke@435 | 417 | bool is_counted() { return is_loop() && _head != NULL && _head->is_CountedLoop(); } |
duke@435 | 418 | |
duke@435 | 419 | #ifndef PRODUCT |
duke@435 | 420 | void dump_head( ) const; // Dump loop head only |
duke@435 | 421 | void dump() const; // Dump this loop recursively |
duke@435 | 422 | void verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const; |
duke@435 | 423 | #endif |
duke@435 | 424 | |
duke@435 | 425 | }; |
duke@435 | 426 | |
duke@435 | 427 | // -----------------------------PhaseIdealLoop--------------------------------- |
duke@435 | 428 | // Computes the mapping from Nodes to IdealLoopTrees. Organizes IdealLoopTrees into a |
duke@435 | 429 | // loop tree. Drives the loop-based transformations on the ideal graph. |
duke@435 | 430 | class PhaseIdealLoop : public PhaseTransform { |
duke@435 | 431 | friend class IdealLoopTree; |
duke@435 | 432 | friend class SuperWord; |
duke@435 | 433 | // Pre-computed def-use info |
duke@435 | 434 | PhaseIterGVN &_igvn; |
duke@435 | 435 | |
duke@435 | 436 | // Head of loop tree |
duke@435 | 437 | IdealLoopTree *_ltree_root; |
duke@435 | 438 | |
duke@435 | 439 | // Array of pre-order numbers, plus post-visited bit. |
duke@435 | 440 | // ZERO for not pre-visited. EVEN for pre-visited but not post-visited. |
duke@435 | 441 | // ODD for post-visited. Other bits are the pre-order number. |
duke@435 | 442 | uint *_preorders; |
duke@435 | 443 | uint _max_preorder; |
duke@435 | 444 | |
never@1356 | 445 | const PhaseIdealLoop* _verify_me; |
never@1356 | 446 | bool _verify_only; |
never@1356 | 447 | |
duke@435 | 448 | // Allocate _preorders[] array |
duke@435 | 449 | void allocate_preorders() { |
duke@435 | 450 | _max_preorder = C->unique()+8; |
duke@435 | 451 | _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder); |
duke@435 | 452 | memset(_preorders, 0, sizeof(uint) * _max_preorder); |
duke@435 | 453 | } |
duke@435 | 454 | |
duke@435 | 455 | // Allocate _preorders[] array |
duke@435 | 456 | void reallocate_preorders() { |
duke@435 | 457 | if ( _max_preorder < C->unique() ) { |
duke@435 | 458 | _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique()); |
duke@435 | 459 | _max_preorder = C->unique(); |
duke@435 | 460 | } |
duke@435 | 461 | memset(_preorders, 0, sizeof(uint) * _max_preorder); |
duke@435 | 462 | } |
duke@435 | 463 | |
duke@435 | 464 | // Check to grow _preorders[] array for the case when build_loop_tree_impl() |
duke@435 | 465 | // adds new nodes. |
duke@435 | 466 | void check_grow_preorders( ) { |
duke@435 | 467 | if ( _max_preorder < C->unique() ) { |
duke@435 | 468 | uint newsize = _max_preorder<<1; // double size of array |
duke@435 | 469 | _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize); |
duke@435 | 470 | memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder)); |
duke@435 | 471 | _max_preorder = newsize; |
duke@435 | 472 | } |
duke@435 | 473 | } |
duke@435 | 474 | // Check for pre-visited. Zero for NOT visited; non-zero for visited. |
duke@435 | 475 | int is_visited( Node *n ) const { return _preorders[n->_idx]; } |
duke@435 | 476 | // Pre-order numbers are written to the Nodes array as low-bit-set values. |
duke@435 | 477 | void set_preorder_visited( Node *n, int pre_order ) { |
duke@435 | 478 | assert( !is_visited( n ), "already set" ); |
duke@435 | 479 | _preorders[n->_idx] = (pre_order<<1); |
duke@435 | 480 | }; |
duke@435 | 481 | // Return pre-order number. |
duke@435 | 482 | int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; } |
duke@435 | 483 | |
duke@435 | 484 | // Check for being post-visited. |
duke@435 | 485 | // Should be previsited already (checked with assert(is_visited(n))). |
duke@435 | 486 | int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; } |
duke@435 | 487 | |
duke@435 | 488 | // Mark as post visited |
duke@435 | 489 | void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; } |
duke@435 | 490 | |
duke@435 | 491 | // Set/get control node out. Set lower bit to distinguish from IdealLoopTree |
duke@435 | 492 | // Returns true if "n" is a data node, false if it's a control node. |
duke@435 | 493 | bool has_ctrl( Node *n ) const { return ((intptr_t)_nodes[n->_idx]) & 1; } |
duke@435 | 494 | |
duke@435 | 495 | // clear out dead code after build_loop_late |
duke@435 | 496 | Node_List _deadlist; |
duke@435 | 497 | |
duke@435 | 498 | // Support for faster execution of get_late_ctrl()/dom_lca() |
duke@435 | 499 | // when a node has many uses and dominator depth is deep. |
duke@435 | 500 | Node_Array _dom_lca_tags; |
duke@435 | 501 | void init_dom_lca_tags(); |
duke@435 | 502 | void clear_dom_lca_tags(); |
never@1356 | 503 | |
never@1356 | 504 | // Helper for debugging bad dominance relationships |
never@1356 | 505 | bool verify_dominance(Node* n, Node* use, Node* LCA, Node* early); |
never@1356 | 506 | |
never@1356 | 507 | Node* compute_lca_of_uses(Node* n, Node* early, bool verify = false); |
never@1356 | 508 | |
duke@435 | 509 | // Inline wrapper for frequent cases: |
duke@435 | 510 | // 1) only one use |
duke@435 | 511 | // 2) a use is the same as the current LCA passed as 'n1' |
duke@435 | 512 | Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) { |
duke@435 | 513 | assert( n->is_CFG(), "" ); |
duke@435 | 514 | // Fast-path NULL lca |
duke@435 | 515 | if( lca != NULL && lca != n ) { |
duke@435 | 516 | assert( lca->is_CFG(), "" ); |
duke@435 | 517 | // find LCA of all uses |
duke@435 | 518 | n = dom_lca_for_get_late_ctrl_internal( lca, n, tag ); |
duke@435 | 519 | } |
duke@435 | 520 | return find_non_split_ctrl(n); |
duke@435 | 521 | } |
duke@435 | 522 | Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag ); |
never@1356 | 523 | |
duke@435 | 524 | // true if CFG node d dominates CFG node n |
duke@435 | 525 | bool is_dominator(Node *d, Node *n); |
duke@435 | 526 | |
duke@435 | 527 | // Helper function for directing control inputs away from CFG split |
duke@435 | 528 | // points. |
duke@435 | 529 | Node *find_non_split_ctrl( Node *ctrl ) const { |
duke@435 | 530 | if (ctrl != NULL) { |
duke@435 | 531 | if (ctrl->is_MultiBranch()) { |
duke@435 | 532 | ctrl = ctrl->in(0); |
duke@435 | 533 | } |
duke@435 | 534 | assert(ctrl->is_CFG(), "CFG"); |
duke@435 | 535 | } |
duke@435 | 536 | return ctrl; |
duke@435 | 537 | } |
duke@435 | 538 | |
duke@435 | 539 | public: |
duke@435 | 540 | bool has_node( Node* n ) const { return _nodes[n->_idx] != NULL; } |
duke@435 | 541 | // check if transform created new nodes that need _ctrl recorded |
duke@435 | 542 | Node *get_late_ctrl( Node *n, Node *early ); |
duke@435 | 543 | Node *get_early_ctrl( Node *n ); |
duke@435 | 544 | void set_early_ctrl( Node *n ); |
duke@435 | 545 | void set_subtree_ctrl( Node *root ); |
duke@435 | 546 | void set_ctrl( Node *n, Node *ctrl ) { |
duke@435 | 547 | assert( !has_node(n) || has_ctrl(n), "" ); |
duke@435 | 548 | assert( ctrl->in(0), "cannot set dead control node" ); |
duke@435 | 549 | assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" ); |
duke@435 | 550 | _nodes.map( n->_idx, (Node*)((intptr_t)ctrl + 1) ); |
duke@435 | 551 | } |
duke@435 | 552 | // Set control and update loop membership |
duke@435 | 553 | void set_ctrl_and_loop(Node* n, Node* ctrl) { |
duke@435 | 554 | IdealLoopTree* old_loop = get_loop(get_ctrl(n)); |
duke@435 | 555 | IdealLoopTree* new_loop = get_loop(ctrl); |
duke@435 | 556 | if (old_loop != new_loop) { |
duke@435 | 557 | if (old_loop->_child == NULL) old_loop->_body.yank(n); |
duke@435 | 558 | if (new_loop->_child == NULL) new_loop->_body.push(n); |
duke@435 | 559 | } |
duke@435 | 560 | set_ctrl(n, ctrl); |
duke@435 | 561 | } |
duke@435 | 562 | // Control nodes can be replaced or subsumed. During this pass they |
duke@435 | 563 | // get their replacement Node in slot 1. Instead of updating the block |
duke@435 | 564 | // location of all Nodes in the subsumed block, we lazily do it. As we |
duke@435 | 565 | // pull such a subsumed block out of the array, we write back the final |
duke@435 | 566 | // correct block. |
duke@435 | 567 | Node *get_ctrl( Node *i ) { |
duke@435 | 568 | assert(has_node(i), ""); |
duke@435 | 569 | Node *n = get_ctrl_no_update(i); |
duke@435 | 570 | _nodes.map( i->_idx, (Node*)((intptr_t)n + 1) ); |
duke@435 | 571 | assert(has_node(i) && has_ctrl(i), ""); |
duke@435 | 572 | assert(n == find_non_split_ctrl(n), "must return legal ctrl" ); |
duke@435 | 573 | return n; |
duke@435 | 574 | } |
duke@435 | 575 | |
duke@435 | 576 | private: |
duke@435 | 577 | Node *get_ctrl_no_update( Node *i ) const { |
duke@435 | 578 | assert( has_ctrl(i), "" ); |
duke@435 | 579 | Node *n = (Node*)(((intptr_t)_nodes[i->_idx]) & ~1); |
duke@435 | 580 | if (!n->in(0)) { |
duke@435 | 581 | // Skip dead CFG nodes |
duke@435 | 582 | do { |
duke@435 | 583 | n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1); |
duke@435 | 584 | } while (!n->in(0)); |
duke@435 | 585 | n = find_non_split_ctrl(n); |
duke@435 | 586 | } |
duke@435 | 587 | return n; |
duke@435 | 588 | } |
duke@435 | 589 | |
duke@435 | 590 | // Check for loop being set |
duke@435 | 591 | // "n" must be a control node. Returns true if "n" is known to be in a loop. |
duke@435 | 592 | bool has_loop( Node *n ) const { |
duke@435 | 593 | assert(!has_node(n) || !has_ctrl(n), ""); |
duke@435 | 594 | return has_node(n); |
duke@435 | 595 | } |
duke@435 | 596 | // Set loop |
duke@435 | 597 | void set_loop( Node *n, IdealLoopTree *loop ) { |
duke@435 | 598 | _nodes.map(n->_idx, (Node*)loop); |
duke@435 | 599 | } |
duke@435 | 600 | // Lazy-dazy update of 'get_ctrl' and 'idom_at' mechanisms. Replace |
duke@435 | 601 | // the 'old_node' with 'new_node'. Kill old-node. Add a reference |
duke@435 | 602 | // from old_node to new_node to support the lazy update. Reference |
duke@435 | 603 | // replaces loop reference, since that is not neede for dead node. |
duke@435 | 604 | public: |
duke@435 | 605 | void lazy_update( Node *old_node, Node *new_node ) { |
duke@435 | 606 | assert( old_node != new_node, "no cycles please" ); |
duke@435 | 607 | //old_node->set_req( 1, new_node /*NO DU INFO*/ ); |
duke@435 | 608 | // Nodes always have DU info now, so re-use the side array slot |
duke@435 | 609 | // for this node to provide the forwarding pointer. |
duke@435 | 610 | _nodes.map( old_node->_idx, (Node*)((intptr_t)new_node + 1) ); |
duke@435 | 611 | } |
duke@435 | 612 | void lazy_replace( Node *old_node, Node *new_node ) { |
duke@435 | 613 | _igvn.hash_delete(old_node); |
duke@435 | 614 | _igvn.subsume_node( old_node, new_node ); |
duke@435 | 615 | lazy_update( old_node, new_node ); |
duke@435 | 616 | } |
duke@435 | 617 | void lazy_replace_proj( Node *old_node, Node *new_node ) { |
duke@435 | 618 | assert( old_node->req() == 1, "use this for Projs" ); |
duke@435 | 619 | _igvn.hash_delete(old_node); // Must hash-delete before hacking edges |
duke@435 | 620 | old_node->add_req( NULL ); |
duke@435 | 621 | lazy_replace( old_node, new_node ); |
duke@435 | 622 | } |
duke@435 | 623 | |
duke@435 | 624 | private: |
duke@435 | 625 | |
duke@435 | 626 | // Place 'n' in some loop nest, where 'n' is a CFG node |
duke@435 | 627 | void build_loop_tree(); |
duke@435 | 628 | int build_loop_tree_impl( Node *n, int pre_order ); |
duke@435 | 629 | // Insert loop into the existing loop tree. 'innermost' is a leaf of the |
duke@435 | 630 | // loop tree, not the root. |
duke@435 | 631 | IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost ); |
duke@435 | 632 | |
duke@435 | 633 | // Place Data nodes in some loop nest |
never@1356 | 634 | void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ); |
never@1356 | 635 | void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ); |
never@1356 | 636 | void build_loop_late_post ( Node* n ); |
duke@435 | 637 | |
duke@435 | 638 | // Array of immediate dominance info for each CFG node indexed by node idx |
duke@435 | 639 | private: |
duke@435 | 640 | uint _idom_size; |
duke@435 | 641 | Node **_idom; // Array of immediate dominators |
duke@435 | 642 | uint *_dom_depth; // Used for fast LCA test |
duke@435 | 643 | GrowableArray<uint>* _dom_stk; // For recomputation of dom depth |
duke@435 | 644 | |
duke@435 | 645 | Node* idom_no_update(Node* d) const { |
duke@435 | 646 | assert(d->_idx < _idom_size, "oob"); |
duke@435 | 647 | Node* n = _idom[d->_idx]; |
duke@435 | 648 | assert(n != NULL,"Bad immediate dominator info."); |
duke@435 | 649 | while (n->in(0) == NULL) { // Skip dead CFG nodes |
duke@435 | 650 | //n = n->in(1); |
duke@435 | 651 | n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1); |
duke@435 | 652 | assert(n != NULL,"Bad immediate dominator info."); |
duke@435 | 653 | } |
duke@435 | 654 | return n; |
duke@435 | 655 | } |
duke@435 | 656 | Node *idom(Node* d) const { |
duke@435 | 657 | uint didx = d->_idx; |
duke@435 | 658 | Node *n = idom_no_update(d); |
duke@435 | 659 | _idom[didx] = n; // Lazily remove dead CFG nodes from table. |
duke@435 | 660 | return n; |
duke@435 | 661 | } |
duke@435 | 662 | uint dom_depth(Node* d) const { |
duke@435 | 663 | assert(d->_idx < _idom_size, ""); |
duke@435 | 664 | return _dom_depth[d->_idx]; |
duke@435 | 665 | } |
duke@435 | 666 | void set_idom(Node* d, Node* n, uint dom_depth); |
duke@435 | 667 | // Locally compute IDOM using dom_lca call |
duke@435 | 668 | Node *compute_idom( Node *region ) const; |
duke@435 | 669 | // Recompute dom_depth |
duke@435 | 670 | void recompute_dom_depth(); |
duke@435 | 671 | |
duke@435 | 672 | // Is safept not required by an outer loop? |
duke@435 | 673 | bool is_deleteable_safept(Node* sfpt); |
duke@435 | 674 | |
never@1356 | 675 | // Perform verification that the graph is valid. |
never@1356 | 676 | PhaseIdealLoop( PhaseIterGVN &igvn) : |
never@1356 | 677 | PhaseTransform(Ideal_Loop), |
never@1356 | 678 | _igvn(igvn), |
never@1356 | 679 | _dom_lca_tags(C->comp_arena()), |
never@1356 | 680 | _verify_me(NULL), |
never@1356 | 681 | _verify_only(true) { |
never@1356 | 682 | build_and_optimize(false); |
never@1356 | 683 | } |
never@1356 | 684 | |
never@1356 | 685 | // build the loop tree and perform any requested optimizations |
never@1356 | 686 | void build_and_optimize(bool do_split_if); |
never@1356 | 687 | |
duke@435 | 688 | public: |
duke@435 | 689 | // Dominators for the sea of nodes |
duke@435 | 690 | void Dominators(); |
duke@435 | 691 | Node *dom_lca( Node *n1, Node *n2 ) const { |
duke@435 | 692 | return find_non_split_ctrl(dom_lca_internal(n1, n2)); |
duke@435 | 693 | } |
duke@435 | 694 | Node *dom_lca_internal( Node *n1, Node *n2 ) const; |
duke@435 | 695 | |
duke@435 | 696 | // Compute the Ideal Node to Loop mapping |
never@1356 | 697 | PhaseIdealLoop( PhaseIterGVN &igvn, bool do_split_ifs) : |
never@1356 | 698 | PhaseTransform(Ideal_Loop), |
never@1356 | 699 | _igvn(igvn), |
never@1356 | 700 | _dom_lca_tags(C->comp_arena()), |
never@1356 | 701 | _verify_me(NULL), |
never@1356 | 702 | _verify_only(false) { |
never@1356 | 703 | build_and_optimize(do_split_ifs); |
never@1356 | 704 | } |
never@1356 | 705 | |
never@1356 | 706 | // Verify that verify_me made the same decisions as a fresh run. |
never@1356 | 707 | PhaseIdealLoop( PhaseIterGVN &igvn, const PhaseIdealLoop *verify_me) : |
never@1356 | 708 | PhaseTransform(Ideal_Loop), |
never@1356 | 709 | _igvn(igvn), |
never@1356 | 710 | _dom_lca_tags(C->comp_arena()), |
never@1356 | 711 | _verify_me(verify_me), |
never@1356 | 712 | _verify_only(false) { |
never@1356 | 713 | build_and_optimize(false); |
never@1356 | 714 | } |
never@1356 | 715 | |
never@1356 | 716 | // Build and verify the loop tree without modifying the graph. This |
never@1356 | 717 | // is useful to verify that all inputs properly dominate their uses. |
never@1356 | 718 | static void verify(PhaseIterGVN& igvn) { |
never@1356 | 719 | #ifdef ASSERT |
never@1356 | 720 | PhaseIdealLoop v(igvn); |
never@1356 | 721 | #endif |
never@1356 | 722 | } |
duke@435 | 723 | |
duke@435 | 724 | // True if the method has at least 1 irreducible loop |
duke@435 | 725 | bool _has_irreducible_loops; |
duke@435 | 726 | |
duke@435 | 727 | // Per-Node transform |
duke@435 | 728 | virtual Node *transform( Node *a_node ) { return 0; } |
duke@435 | 729 | |
duke@435 | 730 | Node *is_counted_loop( Node *x, IdealLoopTree *loop ); |
duke@435 | 731 | |
duke@435 | 732 | // Return a post-walked LoopNode |
duke@435 | 733 | IdealLoopTree *get_loop( Node *n ) const { |
duke@435 | 734 | // Dead nodes have no loop, so return the top level loop instead |
duke@435 | 735 | if (!has_node(n)) return _ltree_root; |
duke@435 | 736 | assert(!has_ctrl(n), ""); |
duke@435 | 737 | return (IdealLoopTree*)_nodes[n->_idx]; |
duke@435 | 738 | } |
duke@435 | 739 | |
duke@435 | 740 | // Is 'n' a (nested) member of 'loop'? |
duke@435 | 741 | int is_member( const IdealLoopTree *loop, Node *n ) const { |
duke@435 | 742 | return loop->is_member(get_loop(n)); } |
duke@435 | 743 | |
duke@435 | 744 | // This is the basic building block of the loop optimizations. It clones an |
duke@435 | 745 | // entire loop body. It makes an old_new loop body mapping; with this |
duke@435 | 746 | // mapping you can find the new-loop equivalent to an old-loop node. All |
duke@435 | 747 | // new-loop nodes are exactly equal to their old-loop counterparts, all |
duke@435 | 748 | // edges are the same. All exits from the old-loop now have a RegionNode |
duke@435 | 749 | // that merges the equivalent new-loop path. This is true even for the |
duke@435 | 750 | // normal "loop-exit" condition. All uses of loop-invariant old-loop values |
duke@435 | 751 | // now come from (one or more) Phis that merge their new-loop equivalents. |
duke@435 | 752 | // Parameter side_by_side_idom: |
duke@435 | 753 | // When side_by_size_idom is NULL, the dominator tree is constructed for |
duke@435 | 754 | // the clone loop to dominate the original. Used in construction of |
duke@435 | 755 | // pre-main-post loop sequence. |
duke@435 | 756 | // When nonnull, the clone and original are side-by-side, both are |
duke@435 | 757 | // dominated by the passed in side_by_side_idom node. Used in |
duke@435 | 758 | // construction of unswitched loops. |
duke@435 | 759 | void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth, |
duke@435 | 760 | Node* side_by_side_idom = NULL); |
duke@435 | 761 | |
duke@435 | 762 | // If we got the effect of peeling, either by actually peeling or by |
duke@435 | 763 | // making a pre-loop which must execute at least once, we can remove |
duke@435 | 764 | // all loop-invariant dominated tests in the main body. |
duke@435 | 765 | void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new ); |
duke@435 | 766 | |
duke@435 | 767 | // Generate code to do a loop peel for the given loop (and body). |
duke@435 | 768 | // old_new is a temp array. |
duke@435 | 769 | void do_peeling( IdealLoopTree *loop, Node_List &old_new ); |
duke@435 | 770 | |
duke@435 | 771 | // Add pre and post loops around the given loop. These loops are used |
duke@435 | 772 | // during RCE, unrolling and aligning loops. |
duke@435 | 773 | void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ); |
duke@435 | 774 | // If Node n lives in the back_ctrl block, we clone a private version of n |
duke@435 | 775 | // in preheader_ctrl block and return that, otherwise return n. |
duke@435 | 776 | Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n ); |
duke@435 | 777 | |
duke@435 | 778 | // Take steps to maximally unroll the loop. Peel any odd iterations, then |
duke@435 | 779 | // unroll to do double iterations. The next round of major loop transforms |
duke@435 | 780 | // will repeat till the doubled loop body does all remaining iterations in 1 |
duke@435 | 781 | // pass. |
duke@435 | 782 | void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ); |
duke@435 | 783 | |
duke@435 | 784 | // Unroll the loop body one step - make each trip do 2 iterations. |
duke@435 | 785 | void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ); |
duke@435 | 786 | |
duke@435 | 787 | // Return true if exp is a constant times an induction var |
duke@435 | 788 | bool is_scaled_iv(Node* exp, Node* iv, int* p_scale); |
duke@435 | 789 | |
duke@435 | 790 | // Return true if exp is a scaled induction var plus (or minus) constant |
duke@435 | 791 | bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth = 0); |
duke@435 | 792 | |
duke@435 | 793 | // Eliminate range-checks and other trip-counter vs loop-invariant tests. |
duke@435 | 794 | void do_range_check( IdealLoopTree *loop, Node_List &old_new ); |
duke@435 | 795 | |
duke@435 | 796 | // Create a slow version of the loop by cloning the loop |
duke@435 | 797 | // and inserting an if to select fast-slow versions. |
duke@435 | 798 | ProjNode* create_slow_version_of_loop(IdealLoopTree *loop, |
duke@435 | 799 | Node_List &old_new); |
duke@435 | 800 | |
duke@435 | 801 | // Clone loop with an invariant test (that does not exit) and |
duke@435 | 802 | // insert a clone of the test that selects which version to |
duke@435 | 803 | // execute. |
duke@435 | 804 | void do_unswitching (IdealLoopTree *loop, Node_List &old_new); |
duke@435 | 805 | |
duke@435 | 806 | // Find candidate "if" for unswitching |
duke@435 | 807 | IfNode* find_unswitching_candidate(const IdealLoopTree *loop) const; |
duke@435 | 808 | |
duke@435 | 809 | // Range Check Elimination uses this function! |
duke@435 | 810 | // Constrain the main loop iterations so the affine function: |
duke@435 | 811 | // scale_con * I + offset < limit |
duke@435 | 812 | // always holds true. That is, either increase the number of iterations in |
duke@435 | 813 | // the pre-loop or the post-loop until the condition holds true in the main |
duke@435 | 814 | // loop. Scale_con, offset and limit are all loop invariant. |
duke@435 | 815 | void add_constraint( int stride_con, int scale_con, Node *offset, Node *limit, Node *pre_ctrl, Node **pre_limit, Node **main_limit ); |
duke@435 | 816 | |
duke@435 | 817 | // Partially peel loop up through last_peel node. |
duke@435 | 818 | bool partial_peel( IdealLoopTree *loop, Node_List &old_new ); |
duke@435 | 819 | |
duke@435 | 820 | // Create a scheduled list of nodes control dependent on ctrl set. |
duke@435 | 821 | void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched ); |
duke@435 | 822 | // Has a use in the vector set |
duke@435 | 823 | bool has_use_in_set( Node* n, VectorSet& vset ); |
duke@435 | 824 | // Has use internal to the vector set (ie. not in a phi at the loop head) |
duke@435 | 825 | bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ); |
duke@435 | 826 | // clone "n" for uses that are outside of loop |
duke@435 | 827 | void clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ); |
duke@435 | 828 | // clone "n" for special uses that are in the not_peeled region |
duke@435 | 829 | void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, |
duke@435 | 830 | VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ); |
duke@435 | 831 | // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist |
duke@435 | 832 | void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ); |
duke@435 | 833 | #ifdef ASSERT |
duke@435 | 834 | // Validate the loop partition sets: peel and not_peel |
duke@435 | 835 | bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel ); |
duke@435 | 836 | // Ensure that uses outside of loop are of the right form |
duke@435 | 837 | bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, |
duke@435 | 838 | uint orig_exit_idx, uint clone_exit_idx); |
duke@435 | 839 | bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx); |
duke@435 | 840 | #endif |
duke@435 | 841 | |
duke@435 | 842 | // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.) |
duke@435 | 843 | int stride_of_possible_iv( Node* iff ); |
duke@435 | 844 | bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; } |
duke@435 | 845 | // Return the (unique) control output node that's in the loop (if it exists.) |
duke@435 | 846 | Node* stay_in_loop( Node* n, IdealLoopTree *loop); |
duke@435 | 847 | // Insert a signed compare loop exit cloned from an unsigned compare. |
duke@435 | 848 | IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop); |
duke@435 | 849 | void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop); |
duke@435 | 850 | // Utility to register node "n" with PhaseIdealLoop |
duke@435 | 851 | void register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth); |
duke@435 | 852 | // Utility to create an if-projection |
duke@435 | 853 | ProjNode* proj_clone(ProjNode* p, IfNode* iff); |
duke@435 | 854 | // Force the iff control output to be the live_proj |
duke@435 | 855 | Node* short_circuit_if(IfNode* iff, ProjNode* live_proj); |
duke@435 | 856 | // Insert a region before an if projection |
duke@435 | 857 | RegionNode* insert_region_before_proj(ProjNode* proj); |
duke@435 | 858 | // Insert a new if before an if projection |
duke@435 | 859 | ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj); |
duke@435 | 860 | |
duke@435 | 861 | // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. |
duke@435 | 862 | // "Nearly" because all Nodes have been cloned from the original in the loop, |
duke@435 | 863 | // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs |
duke@435 | 864 | // through the Phi recursively, and return a Bool. |
duke@435 | 865 | BoolNode *clone_iff( PhiNode *phi, IdealLoopTree *loop ); |
duke@435 | 866 | CmpNode *clone_bool( PhiNode *phi, IdealLoopTree *loop ); |
duke@435 | 867 | |
duke@435 | 868 | |
duke@435 | 869 | // Rework addressing expressions to get the most loop-invariant stuff |
duke@435 | 870 | // moved out. We'd like to do all associative operators, but it's especially |
duke@435 | 871 | // important (common) to do address expressions. |
duke@435 | 872 | Node *remix_address_expressions( Node *n ); |
duke@435 | 873 | |
duke@435 | 874 | // Attempt to use a conditional move instead of a phi/branch |
duke@435 | 875 | Node *conditional_move( Node *n ); |
duke@435 | 876 | |
duke@435 | 877 | // Reorganize offset computations to lower register pressure. |
duke@435 | 878 | // Mostly prevent loop-fallout uses of the pre-incremented trip counter |
duke@435 | 879 | // (which are then alive with the post-incremented trip counter |
duke@435 | 880 | // forcing an extra register move) |
duke@435 | 881 | void reorg_offsets( IdealLoopTree *loop ); |
duke@435 | 882 | |
duke@435 | 883 | // Check for aggressive application of 'split-if' optimization, |
duke@435 | 884 | // using basic block level info. |
duke@435 | 885 | void split_if_with_blocks ( VectorSet &visited, Node_Stack &nstack ); |
duke@435 | 886 | Node *split_if_with_blocks_pre ( Node *n ); |
duke@435 | 887 | void split_if_with_blocks_post( Node *n ); |
duke@435 | 888 | Node *has_local_phi_input( Node *n ); |
duke@435 | 889 | // Mark an IfNode as being dominated by a prior test, |
duke@435 | 890 | // without actually altering the CFG (and hence IDOM info). |
duke@435 | 891 | void dominated_by( Node *prevdom, Node *iff ); |
duke@435 | 892 | |
duke@435 | 893 | // Split Node 'n' through merge point |
duke@435 | 894 | Node *split_thru_region( Node *n, Node *region ); |
duke@435 | 895 | // Split Node 'n' through merge point if there is enough win. |
duke@435 | 896 | Node *split_thru_phi( Node *n, Node *region, int policy ); |
duke@435 | 897 | // Found an If getting its condition-code input from a Phi in the |
duke@435 | 898 | // same block. Split thru the Region. |
duke@435 | 899 | void do_split_if( Node *iff ); |
duke@435 | 900 | |
duke@435 | 901 | private: |
duke@435 | 902 | // Return a type based on condition control flow |
duke@435 | 903 | const TypeInt* filtered_type( Node *n, Node* n_ctrl); |
duke@435 | 904 | const TypeInt* filtered_type( Node *n ) { return filtered_type(n, NULL); } |
duke@435 | 905 | // Helpers for filtered type |
duke@435 | 906 | const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl); |
duke@435 | 907 | |
duke@435 | 908 | // Helper functions |
duke@435 | 909 | void register_new_node( Node *n, Node *blk ); |
duke@435 | 910 | Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache ); |
duke@435 | 911 | Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true ); |
duke@435 | 912 | 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 | 913 | bool split_up( Node *n, Node *blk1, Node *blk2 ); |
duke@435 | 914 | void sink_use( Node *use, Node *post_loop ); |
duke@435 | 915 | Node *place_near_use( Node *useblock ) const; |
duke@435 | 916 | |
duke@435 | 917 | bool _created_loop_node; |
duke@435 | 918 | public: |
duke@435 | 919 | void set_created_loop_node() { _created_loop_node = true; } |
duke@435 | 920 | bool created_loop_node() { return _created_loop_node; } |
duke@435 | 921 | |
duke@435 | 922 | #ifndef PRODUCT |
duke@435 | 923 | void dump( ) const; |
duke@435 | 924 | void dump( IdealLoopTree *loop, uint rpo_idx, Node_List &rpo_list ) const; |
duke@435 | 925 | void rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const; |
duke@435 | 926 | void verify() const; // Major slow :-) |
duke@435 | 927 | void verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const; |
duke@435 | 928 | IdealLoopTree *get_loop_idx(Node* n) const { |
duke@435 | 929 | // Dead nodes have no loop, so return the top level loop instead |
duke@435 | 930 | return _nodes[n->_idx] ? (IdealLoopTree*)_nodes[n->_idx] : _ltree_root; |
duke@435 | 931 | } |
duke@435 | 932 | // Print some stats |
duke@435 | 933 | static void print_statistics(); |
duke@435 | 934 | static int _loop_invokes; // Count of PhaseIdealLoop invokes |
duke@435 | 935 | static int _loop_work; // Sum of PhaseIdealLoop x _unique |
duke@435 | 936 | #endif |
duke@435 | 937 | }; |
duke@435 | 938 | |
duke@435 | 939 | inline Node* IdealLoopTree::tail() { |
duke@435 | 940 | // Handle lazy update of _tail field |
duke@435 | 941 | Node *n = _tail; |
duke@435 | 942 | //while( !n->in(0) ) // Skip dead CFG nodes |
duke@435 | 943 | //n = n->in(1); |
duke@435 | 944 | if (n->in(0) == NULL) |
duke@435 | 945 | n = _phase->get_ctrl(n); |
duke@435 | 946 | _tail = n; |
duke@435 | 947 | return n; |
duke@435 | 948 | } |
duke@435 | 949 | |
duke@435 | 950 | |
duke@435 | 951 | // Iterate over the loop tree using a preorder, left-to-right traversal. |
duke@435 | 952 | // |
duke@435 | 953 | // Example that visits all counted loops from within PhaseIdealLoop |
duke@435 | 954 | // |
duke@435 | 955 | // for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { |
duke@435 | 956 | // IdealLoopTree* lpt = iter.current(); |
duke@435 | 957 | // if (!lpt->is_counted()) continue; |
duke@435 | 958 | // ... |
duke@435 | 959 | class LoopTreeIterator : public StackObj { |
duke@435 | 960 | private: |
duke@435 | 961 | IdealLoopTree* _root; |
duke@435 | 962 | IdealLoopTree* _curnt; |
duke@435 | 963 | |
duke@435 | 964 | public: |
duke@435 | 965 | LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {} |
duke@435 | 966 | |
duke@435 | 967 | bool done() { return _curnt == NULL; } // Finished iterating? |
duke@435 | 968 | |
duke@435 | 969 | void next(); // Advance to next loop tree |
duke@435 | 970 | |
duke@435 | 971 | IdealLoopTree* current() { return _curnt; } // Return current value of iterator. |
duke@435 | 972 | }; |