Mercurial > hg > truffle
annotate src/share/vm/opto/loopnode.hpp @ 367:194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
Reviewed-by: rasbold, jrose
| author | never |
|---|---|
| date | Wed, 17 Sep 2008 12:59:52 -0700 |
| parents | d1605aabd0a1 |
| children | ee8f06bfb27c |
| rev | line source |
|---|---|
| 0 | 1 /* |
| 196 | 2 * Copyright 1998-2008 Sun Microsystems, Inc. All Rights Reserved. |
| 0 | 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
| 4 * | |
| 5 * This code is free software; you can redistribute it and/or modify it | |
| 6 * under the terms of the GNU General Public License version 2 only, as | |
| 7 * published by the Free Software Foundation. | |
| 8 * | |
| 9 * This code is distributed in the hope that it will be useful, but WITHOUT | |
| 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
| 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
| 12 * version 2 for more details (a copy is included in the LICENSE file that | |
| 13 * accompanied this code). | |
| 14 * | |
| 15 * You should have received a copy of the GNU General Public License version | |
| 16 * 2 along with this work; if not, write to the Free Software Foundation, | |
| 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. | |
| 18 * | |
| 19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, | |
| 20 * CA 95054 USA or visit www.sun.com if you need additional information or | |
| 21 * have any questions. | |
| 22 * | |
| 23 */ | |
| 24 | |
| 25 class CmpNode; | |
| 26 class CountedLoopEndNode; | |
| 27 class CountedLoopNode; | |
| 28 class IdealLoopTree; | |
| 29 class LoopNode; | |
| 30 class Node; | |
| 31 class PhaseIdealLoop; | |
| 32 class VectorSet; | |
| 33 struct small_cache; | |
| 34 | |
| 35 // | |
| 36 // I D E A L I Z E D L O O P S | |
| 37 // | |
| 38 // Idealized loops are the set of loops I perform more interesting | |
| 39 // transformations on, beyond simple hoisting. | |
| 40 | |
| 41 //------------------------------LoopNode--------------------------------------- | |
| 42 // Simple loop header. Fall in path on left, loop-back path on right. | |
| 43 class LoopNode : public RegionNode { | |
| 44 // Size is bigger to hold the flags. However, the flags do not change | |
| 45 // the semantics so it does not appear in the hash & cmp functions. | |
| 46 virtual uint size_of() const { return sizeof(*this); } | |
| 47 protected: | |
| 48 short _loop_flags; | |
| 49 // Names for flag bitfields | |
| 50 enum { pre_post_main=0, inner_loop=8, partial_peel_loop=16, partial_peel_failed=32 }; | |
| 51 char _unswitch_count; | |
| 52 enum { _unswitch_max=3 }; | |
| 53 | |
| 54 public: | |
| 55 // Names for edge indices | |
| 56 enum { Self=0, EntryControl, LoopBackControl }; | |
| 57 | |
| 58 int is_inner_loop() const { return _loop_flags & inner_loop; } | |
| 59 void set_inner_loop() { _loop_flags |= inner_loop; } | |
| 60 | |
| 61 int is_partial_peel_loop() const { return _loop_flags & partial_peel_loop; } | |
| 62 void set_partial_peel_loop() { _loop_flags |= partial_peel_loop; } | |
| 63 int partial_peel_has_failed() const { return _loop_flags & partial_peel_failed; } | |
| 64 void mark_partial_peel_failed() { _loop_flags |= partial_peel_failed; } | |
| 65 | |
| 66 int unswitch_max() { return _unswitch_max; } | |
| 67 int unswitch_count() { return _unswitch_count; } | |
| 68 void set_unswitch_count(int val) { | |
| 69 assert (val <= unswitch_max(), "too many unswitches"); | |
| 70 _unswitch_count = val; | |
| 71 } | |
| 72 | |
| 73 LoopNode( Node *entry, Node *backedge ) : RegionNode(3), _loop_flags(0), _unswitch_count(0) { | |
| 74 init_class_id(Class_Loop); | |
| 75 init_req(EntryControl, entry); | |
| 76 init_req(LoopBackControl, backedge); | |
| 77 } | |
| 78 | |
| 79 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); | |
| 80 virtual int Opcode() const; | |
| 81 bool can_be_counted_loop(PhaseTransform* phase) const { | |
| 82 return req() == 3 && in(0) != NULL && | |
| 83 in(1) != NULL && phase->type(in(1)) != Type::TOP && | |
| 84 in(2) != NULL && phase->type(in(2)) != Type::TOP; | |
| 85 } | |
| 86 #ifndef PRODUCT | |
| 87 virtual void dump_spec(outputStream *st) const; | |
| 88 #endif | |
| 89 }; | |
| 90 | |
| 91 //------------------------------Counted Loops---------------------------------- | |
| 92 // Counted loops are all trip-counted loops, with exactly 1 trip-counter exit | |
| 93 // path (and maybe some other exit paths). The trip-counter exit is always | |
| 94 // last in the loop. The trip-counter does not have to stride by a constant, | |
| 95 // but it does have to stride by a loop-invariant amount; the exit value is | |
| 96 // also loop invariant. | |
| 97 | |
| 98 // CountedLoopNodes and CountedLoopEndNodes come in matched pairs. The | |
| 99 // CountedLoopNode has the incoming loop control and the loop-back-control | |
| 100 // which is always the IfTrue before the matching CountedLoopEndNode. The | |
| 101 // CountedLoopEndNode has an incoming control (possibly not the | |
| 102 // CountedLoopNode if there is control flow in the loop), the post-increment | |
| 103 // trip-counter value, and the limit. The trip-counter value is always of | |
| 104 // the form (Op old-trip-counter stride). The old-trip-counter is produced | |
| 105 // by a Phi connected to the CountedLoopNode. The stride is loop invariant. | |
| 106 // The Op is any commutable opcode, including Add, Mul, Xor. The | |
| 107 // CountedLoopEndNode also takes in the loop-invariant limit value. | |
| 108 | |
| 109 // From a CountedLoopNode I can reach the matching CountedLoopEndNode via the | |
| 110 // loop-back control. From CountedLoopEndNodes I can reach CountedLoopNodes | |
| 111 // via the old-trip-counter from the Op node. | |
| 112 | |
| 113 //------------------------------CountedLoopNode-------------------------------- | |
| 114 // CountedLoopNodes head simple counted loops. CountedLoopNodes have as | |
| 115 // inputs the incoming loop-start control and the loop-back control, so they | |
| 116 // act like RegionNodes. They also take in the initial trip counter, the | |
| 117 // loop-invariant stride and the loop-invariant limit value. CountedLoopNodes | |
| 118 // produce a loop-body control and the trip counter value. Since | |
| 119 // CountedLoopNodes behave like RegionNodes I still have a standard CFG model. | |
| 120 | |
| 121 class CountedLoopNode : public LoopNode { | |
| 122 // Size is bigger to hold _main_idx. However, _main_idx does not change | |
| 123 // the semantics so it does not appear in the hash & cmp functions. | |
| 124 virtual uint size_of() const { return sizeof(*this); } | |
| 125 | |
| 126 // For Pre- and Post-loops during debugging ONLY, this holds the index of | |
| 127 // the Main CountedLoop. Used to assert that we understand the graph shape. | |
| 128 node_idx_t _main_idx; | |
| 129 | |
| 130 // Known trip count calculated by policy_maximally_unroll | |
| 131 int _trip_count; | |
| 132 | |
| 133 // Expected trip count from profile data | |
| 134 float _profile_trip_cnt; | |
| 135 | |
| 136 // Log2 of original loop bodies in unrolled loop | |
| 137 int _unrolled_count_log2; | |
| 138 | |
| 139 // Node count prior to last unrolling - used to decide if | |
| 140 // unroll,optimize,unroll,optimize,... is making progress | |
| 141 int _node_count_before_unroll; | |
| 142 | |
| 143 public: | |
| 144 CountedLoopNode( Node *entry, Node *backedge ) | |
| 145 : LoopNode(entry, backedge), _trip_count(max_jint), | |
| 146 _profile_trip_cnt(COUNT_UNKNOWN), _unrolled_count_log2(0), | |
| 147 _node_count_before_unroll(0) { | |
| 148 init_class_id(Class_CountedLoop); | |
| 149 // Initialize _trip_count to the largest possible value. | |
| 150 // Will be reset (lower) if the loop's trip count is known. | |
| 151 } | |
| 152 | |
| 153 virtual int Opcode() const; | |
| 154 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); | |
| 155 | |
| 156 Node *init_control() const { return in(EntryControl); } | |
| 157 Node *back_control() const { return in(LoopBackControl); } | |
| 158 CountedLoopEndNode *loopexit() const; | |
| 159 Node *init_trip() const; | |
| 160 Node *stride() const; | |
| 161 int stride_con() const; | |
| 162 bool stride_is_con() const; | |
| 163 Node *limit() const; | |
| 164 Node *incr() const; | |
| 165 Node *phi() const; | |
| 166 | |
| 167 // Match increment with optional truncation | |
| 168 static Node* match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type); | |
| 169 | |
| 170 // A 'main' loop has a pre-loop and a post-loop. The 'main' loop | |
| 171 // can run short a few iterations and may start a few iterations in. | |
| 172 // It will be RCE'd and unrolled and aligned. | |
| 173 | |
| 174 // A following 'post' loop will run any remaining iterations. Used | |
| 175 // during Range Check Elimination, the 'post' loop will do any final | |
| 176 // iterations with full checks. Also used by Loop Unrolling, where | |
| 177 // the 'post' loop will do any epilog iterations needed. Basically, | |
| 178 // a 'post' loop can not profitably be further unrolled or RCE'd. | |
| 179 | |
| 180 // A preceding 'pre' loop will run at least 1 iteration (to do peeling), | |
| 181 // it may do under-flow checks for RCE and may do alignment iterations | |
| 182 // so the following main loop 'knows' that it is striding down cache | |
| 183 // lines. | |
| 184 | |
| 185 // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or | |
| 186 // Aligned, may be missing it's pre-loop. | |
| 187 enum { Normal=0, Pre=1, Main=2, Post=3, PrePostFlagsMask=3, Main_Has_No_Pre_Loop=4 }; | |
| 188 int is_normal_loop() const { return (_loop_flags&PrePostFlagsMask) == Normal; } | |
| 189 int is_pre_loop () const { return (_loop_flags&PrePostFlagsMask) == Pre; } | |
| 190 int is_main_loop () const { return (_loop_flags&PrePostFlagsMask) == Main; } | |
| 191 int is_post_loop () const { return (_loop_flags&PrePostFlagsMask) == Post; } | |
| 192 int is_main_no_pre_loop() const { return _loop_flags & Main_Has_No_Pre_Loop; } | |
| 193 void set_main_no_pre_loop() { _loop_flags |= Main_Has_No_Pre_Loop; } | |
| 194 | |
|
367
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
196
diff
changeset
|
195 int main_idx() const { return _main_idx; } |
|
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
196
diff
changeset
|
196 |
| 0 | 197 |
| 198 void set_pre_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; } | |
| 199 void set_main_loop ( ) { assert(is_normal_loop(),""); _loop_flags |= Main; } | |
| 200 void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; } | |
| 201 void set_normal_loop( ) { _loop_flags &= ~PrePostFlagsMask; } | |
| 202 | |
| 203 void set_trip_count(int tc) { _trip_count = tc; } | |
| 204 int trip_count() { return _trip_count; } | |
| 205 | |
| 206 void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; } | |
| 207 float profile_trip_cnt() { return _profile_trip_cnt; } | |
| 208 | |
| 209 void double_unrolled_count() { _unrolled_count_log2++; } | |
| 210 int unrolled_count() { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); } | |
| 211 | |
| 212 void set_node_count_before_unroll(int ct) { _node_count_before_unroll = ct; } | |
| 213 int node_count_before_unroll() { return _node_count_before_unroll; } | |
| 214 | |
| 215 #ifndef PRODUCT | |
| 216 virtual void dump_spec(outputStream *st) const; | |
| 217 #endif | |
| 218 }; | |
| 219 | |
| 220 //------------------------------CountedLoopEndNode----------------------------- | |
| 221 // CountedLoopEndNodes end simple trip counted loops. They act much like | |
| 222 // IfNodes. | |
| 223 class CountedLoopEndNode : public IfNode { | |
| 224 public: | |
| 225 enum { TestControl, TestValue }; | |
| 226 | |
| 227 CountedLoopEndNode( Node *control, Node *test, float prob, float cnt ) | |
| 228 : IfNode( control, test, prob, cnt) { | |
| 229 init_class_id(Class_CountedLoopEnd); | |
| 230 } | |
| 231 virtual int Opcode() const; | |
| 232 | |
| 233 Node *cmp_node() const { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : NULL; } | |
| 234 Node *incr() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; } | |
| 235 Node *limit() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; } | |
| 236 Node *stride() const { Node *tmp = incr (); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; } | |
| 237 Node *phi() const { Node *tmp = incr (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; } | |
| 238 Node *init_trip() const { Node *tmp = phi (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; } | |
| 239 int stride_con() const; | |
| 240 bool stride_is_con() const { Node *tmp = stride (); return (tmp != NULL && tmp->is_Con()); } | |
| 241 BoolTest::mask test_trip() const { return in(TestValue)->as_Bool()->_test._test; } | |
| 242 CountedLoopNode *loopnode() const { | |
| 243 Node *ln = phi()->in(0); | |
| 244 assert( ln->Opcode() == Op_CountedLoop, "malformed loop" ); | |
| 245 return (CountedLoopNode*)ln; } | |
| 246 | |
| 247 #ifndef PRODUCT | |
| 248 virtual void dump_spec(outputStream *st) const; | |
| 249 #endif | |
| 250 }; | |
| 251 | |
| 252 | |
| 253 inline CountedLoopEndNode *CountedLoopNode::loopexit() const { | |
| 254 Node *bc = back_control(); | |
| 255 if( bc == NULL ) return NULL; | |
| 256 Node *le = bc->in(0); | |
| 257 if( le->Opcode() != Op_CountedLoopEnd ) | |
| 258 return NULL; | |
| 259 return (CountedLoopEndNode*)le; | |
| 260 } | |
| 261 inline Node *CountedLoopNode::init_trip() const { return loopexit() ? loopexit()->init_trip() : NULL; } | |
| 262 inline Node *CountedLoopNode::stride() const { return loopexit() ? loopexit()->stride() : NULL; } | |
| 263 inline int CountedLoopNode::stride_con() const { return loopexit() ? loopexit()->stride_con() : 0; } | |
| 264 inline bool CountedLoopNode::stride_is_con() const { return loopexit() && loopexit()->stride_is_con(); } | |
| 265 inline Node *CountedLoopNode::limit() const { return loopexit() ? loopexit()->limit() : NULL; } | |
| 266 inline Node *CountedLoopNode::incr() const { return loopexit() ? loopexit()->incr() : NULL; } | |
| 267 inline Node *CountedLoopNode::phi() const { return loopexit() ? loopexit()->phi() : NULL; } | |
| 268 | |
| 269 | |
| 270 // -----------------------------IdealLoopTree---------------------------------- | |
| 271 class IdealLoopTree : public ResourceObj { | |
| 272 public: | |
| 273 IdealLoopTree *_parent; // Parent in loop tree | |
| 274 IdealLoopTree *_next; // Next sibling in loop tree | |
| 275 IdealLoopTree *_child; // First child in loop tree | |
| 276 | |
| 277 // The head-tail backedge defines the loop. | |
| 278 // If tail is NULL then this loop has multiple backedges as part of the | |
| 279 // same loop. During cleanup I'll peel off the multiple backedges; merge | |
| 280 // them at the loop bottom and flow 1 real backedge into the loop. | |
| 281 Node *_head; // Head of loop | |
| 282 Node *_tail; // Tail of loop | |
| 283 inline Node *tail(); // Handle lazy update of _tail field | |
| 284 PhaseIdealLoop* _phase; | |
| 285 | |
| 286 Node_List _body; // Loop body for inner loops | |
| 287 | |
| 288 uint8 _nest; // Nesting depth | |
| 289 uint8 _irreducible:1, // True if irreducible | |
| 290 _has_call:1, // True if has call safepoint | |
| 291 _has_sfpt:1, // True if has non-call safepoint | |
| 292 _rce_candidate:1; // True if candidate for range check elimination | |
| 293 | |
|
39
76256d272075
6667612: (Escape Analysis) disable loop cloning if it has a scalar replaceable allocation
kvn
parents:
17
diff
changeset
|
294 Node_List* _required_safept; // A inner loop cannot delete these safepts; |
|
76256d272075
6667612: (Escape Analysis) disable loop cloning if it has a scalar replaceable allocation
kvn
parents:
17
diff
changeset
|
295 bool _allow_optimizations; // Allow loop optimizations |
| 0 | 296 |
| 297 IdealLoopTree( PhaseIdealLoop* phase, Node *head, Node *tail ) | |
| 298 : _parent(0), _next(0), _child(0), | |
| 299 _head(head), _tail(tail), | |
| 300 _phase(phase), | |
| 301 _required_safept(NULL), | |
|
39
76256d272075
6667612: (Escape Analysis) disable loop cloning if it has a scalar replaceable allocation
kvn
parents:
17
diff
changeset
|
302 _allow_optimizations(true), |
| 0 | 303 _nest(0), _irreducible(0), _has_call(0), _has_sfpt(0), _rce_candidate(0) |
| 304 { } | |
| 305 | |
| 306 // Is 'l' a member of 'this'? | |
| 307 int is_member( const IdealLoopTree *l ) const; // Test for nested membership | |
| 308 | |
| 309 // Set loop nesting depth. Accumulate has_call bits. | |
| 310 int set_nest( uint depth ); | |
| 311 | |
| 312 // Split out multiple fall-in edges from the loop header. Move them to a | |
| 313 // private RegionNode before the loop. This becomes the loop landing pad. | |
| 314 void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ); | |
| 315 | |
| 316 // Split out the outermost loop from this shared header. | |
| 317 void split_outer_loop( PhaseIdealLoop *phase ); | |
| 318 | |
| 319 // Merge all the backedges from the shared header into a private Region. | |
| 320 // Feed that region as the one backedge to this loop. | |
| 321 void merge_many_backedges( PhaseIdealLoop *phase ); | |
| 322 | |
| 323 // Split shared headers and insert loop landing pads. | |
| 324 // Insert a LoopNode to replace the RegionNode. | |
| 325 // Returns TRUE if loop tree is structurally changed. | |
| 326 bool beautify_loops( PhaseIdealLoop *phase ); | |
| 327 | |
| 328 // Perform iteration-splitting on inner loops. Split iterations to avoid | |
| 329 // range checks or one-shot null checks. | |
| 330 void iteration_split( PhaseIdealLoop *phase, Node_List &old_new ); | |
| 331 | |
| 332 // Driver for various flavors of iteration splitting | |
| 333 void iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ); | |
| 334 | |
| 335 // Given dominators, try to find loops with calls that must always be | |
| 336 // executed (call dominates loop tail). These loops do not need non-call | |
| 337 // safepoints (ncsfpt). | |
| 338 void check_safepts(VectorSet &visited, Node_List &stack); | |
| 339 | |
| 340 // Allpaths backwards scan from loop tail, terminating each path at first safepoint | |
| 341 // encountered. | |
| 342 void allpaths_check_safepts(VectorSet &visited, Node_List &stack); | |
| 343 | |
| 344 // Convert to counted loops where possible | |
| 345 void counted_loop( PhaseIdealLoop *phase ); | |
| 346 | |
| 347 // Check for Node being a loop-breaking test | |
| 348 Node *is_loop_exit(Node *iff) const; | |
| 349 | |
| 350 // Returns true if ctrl is executed on every complete iteration | |
| 351 bool dominates_backedge(Node* ctrl); | |
| 352 | |
| 353 // Remove simplistic dead code from loop body | |
| 354 void DCE_loop_body(); | |
| 355 | |
| 356 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage. | |
| 357 // Replace with a 1-in-10 exit guess. | |
| 358 void adjust_loop_exit_prob( PhaseIdealLoop *phase ); | |
| 359 | |
| 360 // Return TRUE or FALSE if the loop should never be RCE'd or aligned. | |
| 361 // Useful for unrolling loops with NO array accesses. | |
| 362 bool policy_peel_only( PhaseIdealLoop *phase ) const; | |
| 363 | |
| 364 // Return TRUE or FALSE if the loop should be unswitched -- clone | |
| 365 // loop with an invariant test | |
| 366 bool policy_unswitching( PhaseIdealLoop *phase ) const; | |
| 367 | |
| 368 // Micro-benchmark spamming. Remove empty loops. | |
| 369 bool policy_do_remove_empty_loop( PhaseIdealLoop *phase ); | |
| 370 | |
| 371 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can | |
| 372 // make some loop-invariant test (usually a null-check) happen before the | |
| 373 // loop. | |
| 374 bool policy_peeling( PhaseIdealLoop *phase ) const; | |
| 375 | |
| 376 // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any | |
| 377 // known trip count in the counted loop node. | |
| 378 bool policy_maximally_unroll( PhaseIdealLoop *phase ) const; | |
| 379 | |
| 380 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if | |
| 381 // the loop is a CountedLoop and the body is small enough. | |
| 382 bool policy_unroll( PhaseIdealLoop *phase ) const; | |
| 383 | |
| 384 // Return TRUE or FALSE if the loop should be range-check-eliminated. | |
| 385 // Gather a list of IF tests that are dominated by iteration splitting; | |
| 386 // also gather the end of the first split and the start of the 2nd split. | |
| 387 bool policy_range_check( PhaseIdealLoop *phase ) const; | |
| 388 | |
| 389 // Return TRUE or FALSE if the loop should be cache-line aligned. | |
| 390 // Gather the expression that does the alignment. Note that only | |
| 391 // one array base can be aligned in a loop (unless the VM guarentees | |
| 392 // mutual alignment). Note that if we vectorize short memory ops | |
| 393 // into longer memory ops, we may want to increase alignment. | |
| 394 bool policy_align( PhaseIdealLoop *phase ) const; | |
| 395 | |
| 396 // Compute loop trip count from profile data | |
| 397 void compute_profile_trip_cnt( PhaseIdealLoop *phase ); | |
| 398 | |
| 399 // Reassociate invariant expressions. | |
| 400 void reassociate_invariants(PhaseIdealLoop *phase); | |
| 401 // Reassociate invariant add and subtract expressions. | |
| 402 Node* reassociate_add_sub(Node* n1, PhaseIdealLoop *phase); | |
| 403 // Return nonzero index of invariant operand if invariant and variant | |
| 404 // are combined with an Add or Sub. Helper for reassoicate_invariants. | |
| 405 int is_invariant_addition(Node* n, PhaseIdealLoop *phase); | |
| 406 | |
| 407 // Return true if n is invariant | |
| 408 bool is_invariant(Node* n) const; | |
| 409 | |
| 410 // Put loop body on igvn work list | |
| 411 void record_for_igvn(); | |
| 412 | |
| 413 bool is_loop() { return !_irreducible && _tail && !_tail->is_top(); } | |
| 414 bool is_inner() { return is_loop() && _child == NULL; } | |
| 415 bool is_counted() { return is_loop() && _head != NULL && _head->is_CountedLoop(); } | |
| 416 | |
| 417 #ifndef PRODUCT | |
| 418 void dump_head( ) const; // Dump loop head only | |
| 419 void dump() const; // Dump this loop recursively | |
| 420 void verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const; | |
| 421 #endif | |
| 422 | |
| 423 }; | |
| 424 | |
| 425 // -----------------------------PhaseIdealLoop--------------------------------- | |
| 426 // Computes the mapping from Nodes to IdealLoopTrees. Organizes IdealLoopTrees into a | |
| 427 // loop tree. Drives the loop-based transformations on the ideal graph. | |
| 428 class PhaseIdealLoop : public PhaseTransform { | |
| 429 friend class IdealLoopTree; | |
| 430 friend class SuperWord; | |
| 431 // Pre-computed def-use info | |
| 432 PhaseIterGVN &_igvn; | |
| 433 | |
| 434 // Head of loop tree | |
| 435 IdealLoopTree *_ltree_root; | |
| 436 | |
| 437 // Array of pre-order numbers, plus post-visited bit. | |
| 438 // ZERO for not pre-visited. EVEN for pre-visited but not post-visited. | |
| 439 // ODD for post-visited. Other bits are the pre-order number. | |
| 440 uint *_preorders; | |
| 441 uint _max_preorder; | |
| 442 | |
| 443 // Allocate _preorders[] array | |
| 444 void allocate_preorders() { | |
| 445 _max_preorder = C->unique()+8; | |
| 446 _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder); | |
| 447 memset(_preorders, 0, sizeof(uint) * _max_preorder); | |
| 448 } | |
| 449 | |
| 450 // Allocate _preorders[] array | |
| 451 void reallocate_preorders() { | |
| 452 if ( _max_preorder < C->unique() ) { | |
| 453 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique()); | |
| 454 _max_preorder = C->unique(); | |
| 455 } | |
| 456 memset(_preorders, 0, sizeof(uint) * _max_preorder); | |
| 457 } | |
| 458 | |
| 459 // Check to grow _preorders[] array for the case when build_loop_tree_impl() | |
| 460 // adds new nodes. | |
| 461 void check_grow_preorders( ) { | |
| 462 if ( _max_preorder < C->unique() ) { | |
| 463 uint newsize = _max_preorder<<1; // double size of array | |
| 464 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize); | |
| 465 memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder)); | |
| 466 _max_preorder = newsize; | |
| 467 } | |
| 468 } | |
| 469 // Check for pre-visited. Zero for NOT visited; non-zero for visited. | |
| 470 int is_visited( Node *n ) const { return _preorders[n->_idx]; } | |
| 471 // Pre-order numbers are written to the Nodes array as low-bit-set values. | |
| 472 void set_preorder_visited( Node *n, int pre_order ) { | |
| 473 assert( !is_visited( n ), "already set" ); | |
| 474 _preorders[n->_idx] = (pre_order<<1); | |
| 475 }; | |
| 476 // Return pre-order number. | |
| 477 int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; } | |
| 478 | |
| 479 // Check for being post-visited. | |
| 480 // Should be previsited already (checked with assert(is_visited(n))). | |
| 481 int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; } | |
| 482 | |
| 483 // Mark as post visited | |
| 484 void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; } | |
| 485 | |
| 486 // Set/get control node out. Set lower bit to distinguish from IdealLoopTree | |
| 487 // Returns true if "n" is a data node, false if it's a control node. | |
| 488 bool has_ctrl( Node *n ) const { return ((intptr_t)_nodes[n->_idx]) & 1; } | |
| 489 | |
| 490 // clear out dead code after build_loop_late | |
| 491 Node_List _deadlist; | |
| 492 | |
| 493 // Support for faster execution of get_late_ctrl()/dom_lca() | |
| 494 // when a node has many uses and dominator depth is deep. | |
| 495 Node_Array _dom_lca_tags; | |
| 496 void init_dom_lca_tags(); | |
| 497 void clear_dom_lca_tags(); | |
| 498 // Inline wrapper for frequent cases: | |
| 499 // 1) only one use | |
| 500 // 2) a use is the same as the current LCA passed as 'n1' | |
| 501 Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) { | |
| 502 assert( n->is_CFG(), "" ); | |
| 503 // Fast-path NULL lca | |
| 504 if( lca != NULL && lca != n ) { | |
| 505 assert( lca->is_CFG(), "" ); | |
| 506 // find LCA of all uses | |
| 507 n = dom_lca_for_get_late_ctrl_internal( lca, n, tag ); | |
| 508 } | |
| 509 return find_non_split_ctrl(n); | |
| 510 } | |
| 511 Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag ); | |
| 512 // true if CFG node d dominates CFG node n | |
| 513 bool is_dominator(Node *d, Node *n); | |
| 514 | |
| 515 // Helper function for directing control inputs away from CFG split | |
| 516 // points. | |
| 517 Node *find_non_split_ctrl( Node *ctrl ) const { | |
| 518 if (ctrl != NULL) { | |
| 519 if (ctrl->is_MultiBranch()) { | |
| 520 ctrl = ctrl->in(0); | |
| 521 } | |
| 522 assert(ctrl->is_CFG(), "CFG"); | |
| 523 } | |
| 524 return ctrl; | |
| 525 } | |
| 526 | |
| 527 public: | |
| 528 bool has_node( Node* n ) const { return _nodes[n->_idx] != NULL; } | |
| 529 // check if transform created new nodes that need _ctrl recorded | |
| 530 Node *get_late_ctrl( Node *n, Node *early ); | |
| 531 Node *get_early_ctrl( Node *n ); | |
| 532 void set_early_ctrl( Node *n ); | |
| 533 void set_subtree_ctrl( Node *root ); | |
| 534 void set_ctrl( Node *n, Node *ctrl ) { | |
| 535 assert( !has_node(n) || has_ctrl(n), "" ); | |
| 536 assert( ctrl->in(0), "cannot set dead control node" ); | |
| 537 assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" ); | |
| 538 _nodes.map( n->_idx, (Node*)((intptr_t)ctrl + 1) ); | |
| 539 } | |
| 540 // Set control and update loop membership | |
| 541 void set_ctrl_and_loop(Node* n, Node* ctrl) { | |
| 542 IdealLoopTree* old_loop = get_loop(get_ctrl(n)); | |
| 543 IdealLoopTree* new_loop = get_loop(ctrl); | |
| 544 if (old_loop != new_loop) { | |
| 545 if (old_loop->_child == NULL) old_loop->_body.yank(n); | |
| 546 if (new_loop->_child == NULL) new_loop->_body.push(n); | |
| 547 } | |
| 548 set_ctrl(n, ctrl); | |
| 549 } | |
| 550 // Control nodes can be replaced or subsumed. During this pass they | |
| 551 // get their replacement Node in slot 1. Instead of updating the block | |
| 552 // location of all Nodes in the subsumed block, we lazily do it. As we | |
| 553 // pull such a subsumed block out of the array, we write back the final | |
| 554 // correct block. | |
| 555 Node *get_ctrl( Node *i ) { | |
| 556 assert(has_node(i), ""); | |
| 557 Node *n = get_ctrl_no_update(i); | |
| 558 _nodes.map( i->_idx, (Node*)((intptr_t)n + 1) ); | |
| 559 assert(has_node(i) && has_ctrl(i), ""); | |
| 560 assert(n == find_non_split_ctrl(n), "must return legal ctrl" ); | |
| 561 return n; | |
| 562 } | |
| 563 | |
| 564 private: | |
| 565 Node *get_ctrl_no_update( Node *i ) const { | |
| 566 assert( has_ctrl(i), "" ); | |
| 567 Node *n = (Node*)(((intptr_t)_nodes[i->_idx]) & ~1); | |
| 568 if (!n->in(0)) { | |
| 569 // Skip dead CFG nodes | |
| 570 do { | |
| 571 n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1); | |
| 572 } while (!n->in(0)); | |
| 573 n = find_non_split_ctrl(n); | |
| 574 } | |
| 575 return n; | |
| 576 } | |
| 577 | |
| 578 // Check for loop being set | |
| 579 // "n" must be a control node. Returns true if "n" is known to be in a loop. | |
| 580 bool has_loop( Node *n ) const { | |
| 581 assert(!has_node(n) || !has_ctrl(n), ""); | |
| 582 return has_node(n); | |
| 583 } | |
| 584 // Set loop | |
| 585 void set_loop( Node *n, IdealLoopTree *loop ) { | |
| 586 _nodes.map(n->_idx, (Node*)loop); | |
| 587 } | |
| 588 // Lazy-dazy update of 'get_ctrl' and 'idom_at' mechanisms. Replace | |
| 589 // the 'old_node' with 'new_node'. Kill old-node. Add a reference | |
| 590 // from old_node to new_node to support the lazy update. Reference | |
| 591 // replaces loop reference, since that is not neede for dead node. | |
| 592 public: | |
| 593 void lazy_update( Node *old_node, Node *new_node ) { | |
| 594 assert( old_node != new_node, "no cycles please" ); | |
| 595 //old_node->set_req( 1, new_node /*NO DU INFO*/ ); | |
| 596 // Nodes always have DU info now, so re-use the side array slot | |
| 597 // for this node to provide the forwarding pointer. | |
| 598 _nodes.map( old_node->_idx, (Node*)((intptr_t)new_node + 1) ); | |
| 599 } | |
| 600 void lazy_replace( Node *old_node, Node *new_node ) { | |
| 601 _igvn.hash_delete(old_node); | |
| 602 _igvn.subsume_node( old_node, new_node ); | |
| 603 lazy_update( old_node, new_node ); | |
| 604 } | |
| 605 void lazy_replace_proj( Node *old_node, Node *new_node ) { | |
| 606 assert( old_node->req() == 1, "use this for Projs" ); | |
| 607 _igvn.hash_delete(old_node); // Must hash-delete before hacking edges | |
| 608 old_node->add_req( NULL ); | |
| 609 lazy_replace( old_node, new_node ); | |
| 610 } | |
| 611 | |
| 612 private: | |
| 613 | |
| 614 // Place 'n' in some loop nest, where 'n' is a CFG node | |
| 615 void build_loop_tree(); | |
| 616 int build_loop_tree_impl( Node *n, int pre_order ); | |
| 617 // Insert loop into the existing loop tree. 'innermost' is a leaf of the | |
| 618 // loop tree, not the root. | |
| 619 IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost ); | |
| 620 | |
| 621 // Place Data nodes in some loop nest | |
| 622 void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack, const PhaseIdealLoop *verify_me ); | |
| 623 void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack, const PhaseIdealLoop *verify_me ); | |
| 624 void build_loop_late_post ( Node* n, const PhaseIdealLoop *verify_me ); | |
| 625 | |
| 626 // Array of immediate dominance info for each CFG node indexed by node idx | |
| 627 private: | |
| 628 uint _idom_size; | |
| 629 Node **_idom; // Array of immediate dominators | |
| 630 uint *_dom_depth; // Used for fast LCA test | |
| 631 GrowableArray<uint>* _dom_stk; // For recomputation of dom depth | |
| 632 | |
| 633 Node* idom_no_update(Node* d) const { | |
| 634 assert(d->_idx < _idom_size, "oob"); | |
| 635 Node* n = _idom[d->_idx]; | |
| 636 assert(n != NULL,"Bad immediate dominator info."); | |
| 637 while (n->in(0) == NULL) { // Skip dead CFG nodes | |
| 638 //n = n->in(1); | |
| 639 n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1); | |
| 640 assert(n != NULL,"Bad immediate dominator info."); | |
| 641 } | |
| 642 return n; | |
| 643 } | |
| 644 Node *idom(Node* d) const { | |
| 645 uint didx = d->_idx; | |
| 646 Node *n = idom_no_update(d); | |
| 647 _idom[didx] = n; // Lazily remove dead CFG nodes from table. | |
| 648 return n; | |
| 649 } | |
| 650 uint dom_depth(Node* d) const { | |
| 651 assert(d->_idx < _idom_size, ""); | |
| 652 return _dom_depth[d->_idx]; | |
| 653 } | |
| 654 void set_idom(Node* d, Node* n, uint dom_depth); | |
| 655 // Locally compute IDOM using dom_lca call | |
| 656 Node *compute_idom( Node *region ) const; | |
| 657 // Recompute dom_depth | |
| 658 void recompute_dom_depth(); | |
| 659 | |
| 660 // Is safept not required by an outer loop? | |
| 661 bool is_deleteable_safept(Node* sfpt); | |
| 662 | |
| 663 public: | |
| 664 // Dominators for the sea of nodes | |
| 665 void Dominators(); | |
| 666 Node *dom_lca( Node *n1, Node *n2 ) const { | |
| 667 return find_non_split_ctrl(dom_lca_internal(n1, n2)); | |
| 668 } | |
| 669 Node *dom_lca_internal( Node *n1, Node *n2 ) const; | |
| 670 | |
| 671 // Compute the Ideal Node to Loop mapping | |
| 672 PhaseIdealLoop( PhaseIterGVN &igvn, const PhaseIdealLoop *verify_me, bool do_split_ifs ); | |
| 673 | |
| 674 // True if the method has at least 1 irreducible loop | |
| 675 bool _has_irreducible_loops; | |
| 676 | |
| 677 // Per-Node transform | |
| 678 virtual Node *transform( Node *a_node ) { return 0; } | |
| 679 | |
| 680 Node *is_counted_loop( Node *x, IdealLoopTree *loop ); | |
| 681 | |
| 682 // Return a post-walked LoopNode | |
| 683 IdealLoopTree *get_loop( Node *n ) const { | |
| 684 // Dead nodes have no loop, so return the top level loop instead | |
| 685 if (!has_node(n)) return _ltree_root; | |
| 686 assert(!has_ctrl(n), ""); | |
| 687 return (IdealLoopTree*)_nodes[n->_idx]; | |
| 688 } | |
| 689 | |
| 690 // Is 'n' a (nested) member of 'loop'? | |
| 691 int is_member( const IdealLoopTree *loop, Node *n ) const { | |
| 692 return loop->is_member(get_loop(n)); } | |
| 693 | |
| 694 // This is the basic building block of the loop optimizations. It clones an | |
| 695 // entire loop body. It makes an old_new loop body mapping; with this | |
| 696 // mapping you can find the new-loop equivalent to an old-loop node. All | |
| 697 // new-loop nodes are exactly equal to their old-loop counterparts, all | |
| 698 // edges are the same. All exits from the old-loop now have a RegionNode | |
| 699 // that merges the equivalent new-loop path. This is true even for the | |
| 700 // normal "loop-exit" condition. All uses of loop-invariant old-loop values | |
| 701 // now come from (one or more) Phis that merge their new-loop equivalents. | |
| 702 // Parameter side_by_side_idom: | |
| 703 // When side_by_size_idom is NULL, the dominator tree is constructed for | |
| 704 // the clone loop to dominate the original. Used in construction of | |
| 705 // pre-main-post loop sequence. | |
| 706 // When nonnull, the clone and original are side-by-side, both are | |
| 707 // dominated by the passed in side_by_side_idom node. Used in | |
| 708 // construction of unswitched loops. | |
| 709 void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth, | |
| 710 Node* side_by_side_idom = NULL); | |
| 711 | |
| 712 // If we got the effect of peeling, either by actually peeling or by | |
| 713 // making a pre-loop which must execute at least once, we can remove | |
| 714 // all loop-invariant dominated tests in the main body. | |
| 715 void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new ); | |
| 716 | |
| 717 // Generate code to do a loop peel for the given loop (and body). | |
| 718 // old_new is a temp array. | |
| 719 void do_peeling( IdealLoopTree *loop, Node_List &old_new ); | |
| 720 | |
| 721 // Add pre and post loops around the given loop. These loops are used | |
| 722 // during RCE, unrolling and aligning loops. | |
| 723 void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ); | |
| 724 // If Node n lives in the back_ctrl block, we clone a private version of n | |
| 725 // in preheader_ctrl block and return that, otherwise return n. | |
| 726 Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n ); | |
| 727 | |
| 728 // Take steps to maximally unroll the loop. Peel any odd iterations, then | |
| 729 // unroll to do double iterations. The next round of major loop transforms | |
| 730 // will repeat till the doubled loop body does all remaining iterations in 1 | |
| 731 // pass. | |
| 732 void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ); | |
| 733 | |
| 734 // Unroll the loop body one step - make each trip do 2 iterations. | |
| 735 void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ); | |
| 736 | |
| 737 // Return true if exp is a constant times an induction var | |
| 738 bool is_scaled_iv(Node* exp, Node* iv, int* p_scale); | |
| 739 | |
| 740 // Return true if exp is a scaled induction var plus (or minus) constant | |
| 741 bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth = 0); | |
| 742 | |
| 743 // Eliminate range-checks and other trip-counter vs loop-invariant tests. | |
| 744 void do_range_check( IdealLoopTree *loop, Node_List &old_new ); | |
| 745 | |
| 746 // Create a slow version of the loop by cloning the loop | |
| 747 // and inserting an if to select fast-slow versions. | |
| 748 ProjNode* create_slow_version_of_loop(IdealLoopTree *loop, | |
| 749 Node_List &old_new); | |
| 750 | |
| 751 // Clone loop with an invariant test (that does not exit) and | |
| 752 // insert a clone of the test that selects which version to | |
| 753 // execute. | |
| 754 void do_unswitching (IdealLoopTree *loop, Node_List &old_new); | |
| 755 | |
| 756 // Find candidate "if" for unswitching | |
| 757 IfNode* find_unswitching_candidate(const IdealLoopTree *loop) const; | |
| 758 | |
| 759 // Range Check Elimination uses this function! | |
| 760 // Constrain the main loop iterations so the affine function: | |
| 761 // scale_con * I + offset < limit | |
| 762 // always holds true. That is, either increase the number of iterations in | |
| 763 // the pre-loop or the post-loop until the condition holds true in the main | |
| 764 // loop. Scale_con, offset and limit are all loop invariant. | |
| 765 void add_constraint( int stride_con, int scale_con, Node *offset, Node *limit, Node *pre_ctrl, Node **pre_limit, Node **main_limit ); | |
| 766 | |
| 767 // Partially peel loop up through last_peel node. | |
| 768 bool partial_peel( IdealLoopTree *loop, Node_List &old_new ); | |
| 769 | |
| 770 // Create a scheduled list of nodes control dependent on ctrl set. | |
| 771 void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched ); | |
| 772 // Has a use in the vector set | |
| 773 bool has_use_in_set( Node* n, VectorSet& vset ); | |
| 774 // Has use internal to the vector set (ie. not in a phi at the loop head) | |
| 775 bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ); | |
| 776 // clone "n" for uses that are outside of loop | |
| 777 void clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ); | |
| 778 // clone "n" for special uses that are in the not_peeled region | |
| 779 void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, | |
| 780 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ); | |
| 781 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist | |
| 782 void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ); | |
| 783 #ifdef ASSERT | |
| 784 // Validate the loop partition sets: peel and not_peel | |
| 785 bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel ); | |
| 786 // Ensure that uses outside of loop are of the right form | |
| 787 bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, | |
| 788 uint orig_exit_idx, uint clone_exit_idx); | |
| 789 bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx); | |
| 790 #endif | |
| 791 | |
| 792 // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.) | |
| 793 int stride_of_possible_iv( Node* iff ); | |
| 794 bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; } | |
| 795 // Return the (unique) control output node that's in the loop (if it exists.) | |
| 796 Node* stay_in_loop( Node* n, IdealLoopTree *loop); | |
| 797 // Insert a signed compare loop exit cloned from an unsigned compare. | |
| 798 IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop); | |
| 799 void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop); | |
| 800 // Utility to register node "n" with PhaseIdealLoop | |
| 801 void register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth); | |
| 802 // Utility to create an if-projection | |
| 803 ProjNode* proj_clone(ProjNode* p, IfNode* iff); | |
| 804 // Force the iff control output to be the live_proj | |
| 805 Node* short_circuit_if(IfNode* iff, ProjNode* live_proj); | |
| 806 // Insert a region before an if projection | |
| 807 RegionNode* insert_region_before_proj(ProjNode* proj); | |
| 808 // Insert a new if before an if projection | |
| 809 ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj); | |
| 810 | |
| 811 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. | |
| 812 // "Nearly" because all Nodes have been cloned from the original in the loop, | |
| 813 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs | |
| 814 // through the Phi recursively, and return a Bool. | |
| 815 BoolNode *clone_iff( PhiNode *phi, IdealLoopTree *loop ); | |
| 816 CmpNode *clone_bool( PhiNode *phi, IdealLoopTree *loop ); | |
| 817 | |
| 818 | |
| 819 // Rework addressing expressions to get the most loop-invariant stuff | |
| 820 // moved out. We'd like to do all associative operators, but it's especially | |
| 821 // important (common) to do address expressions. | |
| 822 Node *remix_address_expressions( Node *n ); | |
| 823 | |
| 824 // Attempt to use a conditional move instead of a phi/branch | |
| 825 Node *conditional_move( Node *n ); | |
| 826 | |
| 827 // Reorganize offset computations to lower register pressure. | |
| 828 // Mostly prevent loop-fallout uses of the pre-incremented trip counter | |
| 829 // (which are then alive with the post-incremented trip counter | |
| 830 // forcing an extra register move) | |
| 831 void reorg_offsets( IdealLoopTree *loop ); | |
| 832 | |
| 833 // Check for aggressive application of 'split-if' optimization, | |
| 834 // using basic block level info. | |
| 835 void split_if_with_blocks ( VectorSet &visited, Node_Stack &nstack ); | |
| 836 Node *split_if_with_blocks_pre ( Node *n ); | |
| 837 void split_if_with_blocks_post( Node *n ); | |
| 838 Node *has_local_phi_input( Node *n ); | |
| 839 // Mark an IfNode as being dominated by a prior test, | |
| 840 // without actually altering the CFG (and hence IDOM info). | |
| 841 void dominated_by( Node *prevdom, Node *iff ); | |
| 842 | |
| 843 // Split Node 'n' through merge point | |
| 844 Node *split_thru_region( Node *n, Node *region ); | |
| 845 // Split Node 'n' through merge point if there is enough win. | |
| 846 Node *split_thru_phi( Node *n, Node *region, int policy ); | |
| 847 // Found an If getting its condition-code input from a Phi in the | |
| 848 // same block. Split thru the Region. | |
| 849 void do_split_if( Node *iff ); | |
| 850 | |
| 851 private: | |
| 852 // Return a type based on condition control flow | |
| 853 const TypeInt* filtered_type( Node *n, Node* n_ctrl); | |
| 854 const TypeInt* filtered_type( Node *n ) { return filtered_type(n, NULL); } | |
| 855 // Helpers for filtered type | |
| 856 const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl); | |
| 857 | |
| 858 // Helper functions | |
| 859 void register_new_node( Node *n, Node *blk ); | |
| 860 Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache ); | |
| 861 Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true ); | |
| 862 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 ); | |
| 863 bool split_up( Node *n, Node *blk1, Node *blk2 ); | |
| 864 void sink_use( Node *use, Node *post_loop ); | |
| 865 Node *place_near_use( Node *useblock ) const; | |
| 866 | |
| 867 bool _created_loop_node; | |
| 868 public: | |
| 869 void set_created_loop_node() { _created_loop_node = true; } | |
| 870 bool created_loop_node() { return _created_loop_node; } | |
| 871 | |
| 872 #ifndef PRODUCT | |
| 873 void dump( ) const; | |
| 874 void dump( IdealLoopTree *loop, uint rpo_idx, Node_List &rpo_list ) const; | |
| 875 void rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const; | |
| 876 void verify() const; // Major slow :-) | |
| 877 void verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const; | |
| 878 IdealLoopTree *get_loop_idx(Node* n) const { | |
| 879 // Dead nodes have no loop, so return the top level loop instead | |
| 880 return _nodes[n->_idx] ? (IdealLoopTree*)_nodes[n->_idx] : _ltree_root; | |
| 881 } | |
| 882 // Print some stats | |
| 883 static void print_statistics(); | |
| 884 static int _loop_invokes; // Count of PhaseIdealLoop invokes | |
| 885 static int _loop_work; // Sum of PhaseIdealLoop x _unique | |
| 886 #endif | |
| 887 }; | |
| 888 | |
| 889 inline Node* IdealLoopTree::tail() { | |
| 890 // Handle lazy update of _tail field | |
| 891 Node *n = _tail; | |
| 892 //while( !n->in(0) ) // Skip dead CFG nodes | |
| 893 //n = n->in(1); | |
| 894 if (n->in(0) == NULL) | |
| 895 n = _phase->get_ctrl(n); | |
| 896 _tail = n; | |
| 897 return n; | |
| 898 } | |
| 899 | |
| 900 | |
| 901 // Iterate over the loop tree using a preorder, left-to-right traversal. | |
| 902 // | |
| 903 // Example that visits all counted loops from within PhaseIdealLoop | |
| 904 // | |
| 905 // for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { | |
| 906 // IdealLoopTree* lpt = iter.current(); | |
| 907 // if (!lpt->is_counted()) continue; | |
| 908 // ... | |
| 909 class LoopTreeIterator : public StackObj { | |
| 910 private: | |
| 911 IdealLoopTree* _root; | |
| 912 IdealLoopTree* _curnt; | |
| 913 | |
| 914 public: | |
| 915 LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {} | |
| 916 | |
| 917 bool done() { return _curnt == NULL; } // Finished iterating? | |
| 918 | |
| 919 void next(); // Advance to next loop tree | |
| 920 | |
| 921 IdealLoopTree* current() { return _curnt; } // Return current value of iterator. | |
| 922 }; |