1 #ifdef USE_PRAGMA_IDENT_HDR 2 #pragma ident "@(#)loopnode.hpp 1.146 07/10/23 13:12:55 JVM" 3 #endif 4 /* 5 * Copyright 1998-2008 Sun Microsystems, Inc. All Rights Reserved. 6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 7 * 8 * This code is free software; you can redistribute it and/or modify it 9 * under the terms of the GNU General Public License version 2 only, as 10 * published by the Free Software Foundation. 11 * 12 * This code is distributed in the hope that it will be useful, but WITHOUT 13 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 14 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 15 * version 2 for more details (a copy is included in the LICENSE file that 16 * accompanied this code). 17 * 18 * You should have received a copy of the GNU General Public License version 19 * 2 along with this work; if not, write to the Free Software Foundation, 20 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 21 * 22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, 23 * CA 95054 USA or visit www.sun.com if you need additional information or 24 * have any questions. 25 * 26 */ 27 28 class CmpNode; 29 class CountedLoopEndNode; 30 class CountedLoopNode; 31 class IdealLoopTree; 32 class LoopNode; 33 class Node; 34 class PhaseIdealLoop; 35 class VectorSet; 36 struct small_cache; 37 38 // 39 // I D E A L I Z E D L O O P S 40 // 41 // Idealized loops are the set of loops I perform more interesting 42 // transformations on, beyond simple hoisting. 43 44 //------------------------------LoopNode--------------------------------------- 45 // Simple loop header. Fall in path on left, loop-back path on right. 46 class LoopNode : public RegionNode { 47 // Size is bigger to hold the flags. However, the flags do not change 48 // the semantics so it does not appear in the hash & cmp functions. 49 virtual uint size_of() const { return sizeof(*this); } 50 protected: 51 short _loop_flags; 52 // Names for flag bitfields 53 enum { pre_post_main=0, inner_loop=8, partial_peel_loop=16, partial_peel_failed=32 }; 54 char _unswitch_count; 55 enum { _unswitch_max=3 }; 56 57 public: 58 // Names for edge indices 59 enum { Self=0, EntryControl, LoopBackControl }; 60 61 int is_inner_loop() const { return _loop_flags & inner_loop; } 62 void set_inner_loop() { _loop_flags |= inner_loop; } 63 64 int is_partial_peel_loop() const { return _loop_flags & partial_peel_loop; } 65 void set_partial_peel_loop() { _loop_flags |= partial_peel_loop; } 66 int partial_peel_has_failed() const { return _loop_flags & partial_peel_failed; } 67 void mark_partial_peel_failed() { _loop_flags |= partial_peel_failed; } 68 69 int unswitch_max() { return _unswitch_max; } 70 int unswitch_count() { return _unswitch_count; } 71 void set_unswitch_count(int val) { 72 assert (val <= unswitch_max(), "too many unswitches"); 73 _unswitch_count = val; 74 } 75 76 LoopNode( Node *entry, Node *backedge ) : RegionNode(3), _loop_flags(0), _unswitch_count(0) { 77 init_class_id(Class_Loop); 78 init_req(EntryControl, entry); 79 init_req(LoopBackControl, backedge); 80 } 81 82 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); 83 virtual int Opcode() const; 84 bool can_be_counted_loop(PhaseTransform* phase) const { 85 return req() == 3 && in(0) != NULL && 86 in(1) != NULL && phase->type(in(1)) != Type::TOP && 87 in(2) != NULL && phase->type(in(2)) != Type::TOP; 88 } 89 #ifndef PRODUCT 90 virtual void dump_spec(outputStream *st) const; 91 #endif 92 }; 93 94 //------------------------------Counted Loops---------------------------------- 95 // Counted loops are all trip-counted loops, with exactly 1 trip-counter exit 96 // path (and maybe some other exit paths). The trip-counter exit is always 97 // last in the loop. The trip-counter does not have to stride by a constant, 98 // but it does have to stride by a loop-invariant amount; the exit value is 99 // also loop invariant. 100 101 // CountedLoopNodes and CountedLoopEndNodes come in matched pairs. The 102 // CountedLoopNode has the incoming loop control and the loop-back-control 103 // which is always the IfTrue before the matching CountedLoopEndNode. The 104 // CountedLoopEndNode has an incoming control (possibly not the 105 // CountedLoopNode if there is control flow in the loop), the post-increment 106 // trip-counter value, and the limit. The trip-counter value is always of 107 // the form (Op old-trip-counter stride). The old-trip-counter is produced 108 // by a Phi connected to the CountedLoopNode. The stride is loop invariant. 109 // The Op is any commutable opcode, including Add, Mul, Xor. The 110 // CountedLoopEndNode also takes in the loop-invariant limit value. 111 112 // From a CountedLoopNode I can reach the matching CountedLoopEndNode via the 113 // loop-back control. From CountedLoopEndNodes I can reach CountedLoopNodes 114 // via the old-trip-counter from the Op node. 115 116 //------------------------------CountedLoopNode-------------------------------- 117 // CountedLoopNodes head simple counted loops. CountedLoopNodes have as 118 // inputs the incoming loop-start control and the loop-back control, so they 119 // act like RegionNodes. They also take in the initial trip counter, the 120 // loop-invariant stride and the loop-invariant limit value. CountedLoopNodes 121 // produce a loop-body control and the trip counter value. Since 122 // CountedLoopNodes behave like RegionNodes I still have a standard CFG model. 123 124 class CountedLoopNode : public LoopNode { 125 // Size is bigger to hold _main_idx. However, _main_idx does not change 126 // the semantics so it does not appear in the hash & cmp functions. 127 virtual uint size_of() const { return sizeof(*this); } 128 129 // For Pre- and Post-loops during debugging ONLY, this holds the index of 130 // the Main CountedLoop. Used to assert that we understand the graph shape. 131 node_idx_t _main_idx; 132 133 // Known trip count calculated by policy_maximally_unroll 134 int _trip_count; 135 136 // Expected trip count from profile data 137 float _profile_trip_cnt; 138 139 // Log2 of original loop bodies in unrolled loop 140 int _unrolled_count_log2; 141 142 // Node count prior to last unrolling - used to decide if 143 // unroll,optimize,unroll,optimize,... is making progress 144 int _node_count_before_unroll; 145 146 public: 147 CountedLoopNode( Node *entry, Node *backedge ) 148 : LoopNode(entry, backedge), _trip_count(max_jint), 149 _profile_trip_cnt(COUNT_UNKNOWN), _unrolled_count_log2(0), 150 _node_count_before_unroll(0) { 151 init_class_id(Class_CountedLoop); 152 // Initialize _trip_count to the largest possible value. 153 // Will be reset (lower) if the loop's trip count is known. 154 } 155 156 virtual int Opcode() const; 157 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); 158 159 Node *init_control() const { return in(EntryControl); } 160 Node *back_control() const { return in(LoopBackControl); } 161 CountedLoopEndNode *loopexit() const; 162 Node *init_trip() const; 163 Node *stride() const; 164 int stride_con() const; 165 bool stride_is_con() const; 166 Node *limit() const; 167 Node *incr() const; 168 Node *phi() const; 169 170 // Match increment with optional truncation 171 static Node* match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type); 172 173 // A 'main' loop has a pre-loop and a post-loop. The 'main' loop 174 // can run short a few iterations and may start a few iterations in. 175 // It will be RCE'd and unrolled and aligned. 176 177 // A following 'post' loop will run any remaining iterations. Used 178 // during Range Check Elimination, the 'post' loop will do any final 179 // iterations with full checks. Also used by Loop Unrolling, where 180 // the 'post' loop will do any epilog iterations needed. Basically, 181 // a 'post' loop can not profitably be further unrolled or RCE'd. 182 183 // A preceding 'pre' loop will run at least 1 iteration (to do peeling), 184 // it may do under-flow checks for RCE and may do alignment iterations 185 // so the following main loop 'knows' that it is striding down cache 186 // lines. 187 188 // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or 189 // Aligned, may be missing it's pre-loop. 190 enum { Normal=0, Pre=1, Main=2, Post=3, PrePostFlagsMask=3, Main_Has_No_Pre_Loop=4 }; 191 int is_normal_loop() const { return (_loop_flags&PrePostFlagsMask) == Normal; } 192 int is_pre_loop () const { return (_loop_flags&PrePostFlagsMask) == Pre; } 193 int is_main_loop () const { return (_loop_flags&PrePostFlagsMask) == Main; } 194 int is_post_loop () const { return (_loop_flags&PrePostFlagsMask) == Post; } 195 int is_main_no_pre_loop() const { return _loop_flags & Main_Has_No_Pre_Loop; } 196 void set_main_no_pre_loop() { _loop_flags |= Main_Has_No_Pre_Loop; } 197 198 int main_idx() const { return _main_idx; } 199 200 201 void set_pre_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; } 202 void set_main_loop ( ) { assert(is_normal_loop(),""); _loop_flags |= Main; } 203 void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; } 204 void set_normal_loop( ) { _loop_flags &= ~PrePostFlagsMask; } 205 206 void set_trip_count(int tc) { _trip_count = tc; } 207 int trip_count() { return _trip_count; } 208 209 void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; } 210 float profile_trip_cnt() { return _profile_trip_cnt; } 211 212 void double_unrolled_count() { _unrolled_count_log2++; } 213 int unrolled_count() { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); } 214 215 void set_node_count_before_unroll(int ct) { _node_count_before_unroll = ct; } 216 int node_count_before_unroll() { return _node_count_before_unroll; } 217 218 #ifndef PRODUCT 219 virtual void dump_spec(outputStream *st) const; 220 #endif 221 }; 222 223 //------------------------------CountedLoopEndNode----------------------------- 224 // CountedLoopEndNodes end simple trip counted loops. They act much like 225 // IfNodes. 226 class CountedLoopEndNode : public IfNode { 227 public: 228 enum { TestControl, TestValue }; 229 230 CountedLoopEndNode( Node *control, Node *test, float prob, float cnt ) 231 : IfNode( control, test, prob, cnt) { 232 init_class_id(Class_CountedLoopEnd); 233 } 234 virtual int Opcode() const; 235 236 Node *cmp_node() const { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : NULL; } 237 Node *incr() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; } 238 Node *limit() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; } 239 Node *stride() const { Node *tmp = incr (); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; } 240 Node *phi() const { Node *tmp = incr (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; } 241 Node *init_trip() const { Node *tmp = phi (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; } 242 int stride_con() const; 243 bool stride_is_con() const { Node *tmp = stride (); return (tmp != NULL && tmp->is_Con()); } 244 BoolTest::mask test_trip() const { return in(TestValue)->as_Bool()->_test._test; } 245 CountedLoopNode *loopnode() const { 246 Node *ln = phi()->in(0); 247 assert( ln->Opcode() == Op_CountedLoop, "malformed loop" ); 248 return (CountedLoopNode*)ln; } 249 250 #ifndef PRODUCT 251 virtual void dump_spec(outputStream *st) const; 252 #endif 253 }; 254 255 256 inline CountedLoopEndNode *CountedLoopNode::loopexit() const { 257 Node *bc = back_control(); 258 if( bc == NULL ) return NULL; 259 Node *le = bc->in(0); 260 if( le->Opcode() != Op_CountedLoopEnd ) 261 return NULL; 262 return (CountedLoopEndNode*)le; 263 } 264 inline Node *CountedLoopNode::init_trip() const { return loopexit() ? loopexit()->init_trip() : NULL; } 265 inline Node *CountedLoopNode::stride() const { return loopexit() ? loopexit()->stride() : NULL; } 266 inline int CountedLoopNode::stride_con() const { return loopexit() ? loopexit()->stride_con() : 0; } 267 inline bool CountedLoopNode::stride_is_con() const { return loopexit() && loopexit()->stride_is_con(); } 268 inline Node *CountedLoopNode::limit() const { return loopexit() ? loopexit()->limit() : NULL; } 269 inline Node *CountedLoopNode::incr() const { return loopexit() ? loopexit()->incr() : NULL; } 270 inline Node *CountedLoopNode::phi() const { return loopexit() ? loopexit()->phi() : NULL; } 271 272 273 // -----------------------------IdealLoopTree---------------------------------- 274 class IdealLoopTree : public ResourceObj { 275 public: 276 IdealLoopTree *_parent; // Parent in loop tree 277 IdealLoopTree *_next; // Next sibling in loop tree 278 IdealLoopTree *_child; // First child in loop tree 279 280 // The head-tail backedge defines the loop. 281 // If tail is NULL then this loop has multiple backedges as part of the 282 // same loop. During cleanup I'll peel off the multiple backedges; merge 283 // them at the loop bottom and flow 1 real backedge into the loop. 284 Node *_head; // Head of loop 285 Node *_tail; // Tail of loop 286 inline Node *tail(); // Handle lazy update of _tail field 287 PhaseIdealLoop* _phase; 288 289 Node_List _body; // Loop body for inner loops 290 291 uint8 _nest; // Nesting depth 292 uint8 _irreducible:1, // True if irreducible 293 _has_call:1, // True if has call safepoint 294 _has_sfpt:1, // True if has non-call safepoint 295 _rce_candidate:1; // True if candidate for range check elimination 296 297 Node_List* _required_safept; // A inner loop cannot delete these safepts; 298 bool _allow_optimizations; // Allow loop optimizations 299 300 IdealLoopTree( PhaseIdealLoop* phase, Node *head, Node *tail ) 301 : _parent(0), _next(0), _child(0), 302 _head(head), _tail(tail), 303 _phase(phase), 304 _required_safept(NULL), 305 _allow_optimizations(true), 306 _nest(0), _irreducible(0), _has_call(0), _has_sfpt(0), _rce_candidate(0) 307 { } 308 309 // Is 'l' a member of 'this'? 310 int is_member( const IdealLoopTree *l ) const; // Test for nested membership 311 312 // Set loop nesting depth. Accumulate has_call bits. 313 int set_nest( uint depth ); 314 315 // Split out multiple fall-in edges from the loop header. Move them to a 316 // private RegionNode before the loop. This becomes the loop landing pad. 317 void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ); 318 319 // Split out the outermost loop from this shared header. 320 void split_outer_loop( PhaseIdealLoop *phase ); 321 322 // Merge all the backedges from the shared header into a private Region. 323 // Feed that region as the one backedge to this loop. 324 void merge_many_backedges( PhaseIdealLoop *phase ); 325 326 // Split shared headers and insert loop landing pads. 327 // Insert a LoopNode to replace the RegionNode. 328 // Returns TRUE if loop tree is structurally changed. 329 bool beautify_loops( PhaseIdealLoop *phase ); 330 331 // Perform iteration-splitting on inner loops. Split iterations to 332 // avoid range checks or one-shot null checks. Returns false if the 333 // current round of loop opts should stop. 334 bool iteration_split( PhaseIdealLoop *phase, Node_List &old_new ); 335 336 // Driver for various flavors of iteration splitting. Returns false 337 // if the current round of loop opts should stop. 338 bool iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ); 339 340 // Given dominators, try to find loops with calls that must always be 341 // executed (call dominates loop tail). These loops do not need non-call 342 // safepoints (ncsfpt). 343 void check_safepts(VectorSet &visited, Node_List &stack); 344 345 // Allpaths backwards scan from loop tail, terminating each path at first safepoint 346 // encountered. 347 void allpaths_check_safepts(VectorSet &visited, Node_List &stack); 348 349 // Convert to counted loops where possible 350 void counted_loop( PhaseIdealLoop *phase ); 351 352 // Check for Node being a loop-breaking test 353 Node *is_loop_exit(Node *iff) const; 354 355 // Returns true if ctrl is executed on every complete iteration 356 bool dominates_backedge(Node* ctrl); 357 358 // Remove simplistic dead code from loop body 359 void DCE_loop_body(); 360 361 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage. 362 // Replace with a 1-in-10 exit guess. 363 void adjust_loop_exit_prob( PhaseIdealLoop *phase ); 364 365 // Return TRUE or FALSE if the loop should never be RCE'd or aligned. 366 // Useful for unrolling loops with NO array accesses. 367 bool policy_peel_only( PhaseIdealLoop *phase ) const; 368 369 // Return TRUE or FALSE if the loop should be unswitched -- clone 370 // loop with an invariant test 371 bool policy_unswitching( PhaseIdealLoop *phase ) const; 372 373 // Micro-benchmark spamming. Remove empty loops. 374 bool policy_do_remove_empty_loop( PhaseIdealLoop *phase ); 375 376 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can 377 // make some loop-invariant test (usually a null-check) happen before the 378 // loop. 379 bool policy_peeling( PhaseIdealLoop *phase ) const; 380 381 // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any 382 // known trip count in the counted loop node. 383 bool policy_maximally_unroll( PhaseIdealLoop *phase ) const; 384 385 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if 386 // the loop is a CountedLoop and the body is small enough. 387 bool policy_unroll( PhaseIdealLoop *phase ) const; 388 389 // Return TRUE or FALSE if the loop should be range-check-eliminated. 390 // Gather a list of IF tests that are dominated by iteration splitting; 391 // also gather the end of the first split and the start of the 2nd split. 392 bool policy_range_check( PhaseIdealLoop *phase ) const; 393 394 // Return TRUE or FALSE if the loop should be cache-line aligned. 395 // Gather the expression that does the alignment. Note that only 396 // one array base can be aligned in a loop (unless the VM guarentees 397 // mutual alignment). Note that if we vectorize short memory ops 398 // into longer memory ops, we may want to increase alignment. 399 bool policy_align( PhaseIdealLoop *phase ) const; 400 401 // Compute loop trip count from profile data 402 void compute_profile_trip_cnt( PhaseIdealLoop *phase ); 403 404 // Reassociate invariant expressions. 405 void reassociate_invariants(PhaseIdealLoop *phase); 406 // Reassociate invariant add and subtract expressions. 407 Node* reassociate_add_sub(Node* n1, PhaseIdealLoop *phase); 408 // Return nonzero index of invariant operand if invariant and variant 409 // are combined with an Add or Sub. Helper for reassoicate_invariants. 410 int is_invariant_addition(Node* n, PhaseIdealLoop *phase); 411 412 // Return true if n is invariant 413 bool is_invariant(Node* n) const; 414 415 // Put loop body on igvn work list 416 void record_for_igvn(); 417 418 bool is_loop() { return !_irreducible && _tail && !_tail->is_top(); } 419 bool is_inner() { return is_loop() && _child == NULL; } 420 bool is_counted() { return is_loop() && _head != NULL && _head->is_CountedLoop(); } 421 422 #ifndef PRODUCT 423 void dump_head( ) const; // Dump loop head only 424 void dump() const; // Dump this loop recursively 425 void verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const; 426 #endif 427 428 }; 429 430 // -----------------------------PhaseIdealLoop--------------------------------- 431 // Computes the mapping from Nodes to IdealLoopTrees. Organizes IdealLoopTrees into a 432 // loop tree. Drives the loop-based transformations on the ideal graph. 433 class PhaseIdealLoop : public PhaseTransform { 434 friend class IdealLoopTree; 435 friend class SuperWord; 436 // Pre-computed def-use info 437 PhaseIterGVN &_igvn; 438 439 // Head of loop tree 440 IdealLoopTree *_ltree_root; 441 442 // Array of pre-order numbers, plus post-visited bit. 443 // ZERO for not pre-visited. EVEN for pre-visited but not post-visited. 444 // ODD for post-visited. Other bits are the pre-order number. 445 uint *_preorders; 446 uint _max_preorder; 447 448 // Allocate _preorders[] array 449 void allocate_preorders() { 450 _max_preorder = C->unique()+8; 451 _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder); 452 memset(_preorders, 0, sizeof(uint) * _max_preorder); 453 } 454 455 // Allocate _preorders[] array 456 void reallocate_preorders() { 457 if ( _max_preorder < C->unique() ) { 458 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique()); 459 _max_preorder = C->unique(); 460 } 461 memset(_preorders, 0, sizeof(uint) * _max_preorder); 462 } 463 464 // Check to grow _preorders[] array for the case when build_loop_tree_impl() 465 // adds new nodes. 466 void check_grow_preorders( ) { 467 if ( _max_preorder < C->unique() ) { 468 uint newsize = _max_preorder<<1; // double size of array 469 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize); 470 memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder)); 471 _max_preorder = newsize; 472 } 473 } 474 // Check for pre-visited. Zero for NOT visited; non-zero for visited. 475 int is_visited( Node *n ) const { return _preorders[n->_idx]; } 476 // Pre-order numbers are written to the Nodes array as low-bit-set values. 477 void set_preorder_visited( Node *n, int pre_order ) { 478 assert( !is_visited( n ), "already set" ); 479 _preorders[n->_idx] = (pre_order<<1); 480 }; 481 // Return pre-order number. 482 int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; } 483 484 // Check for being post-visited. 485 // Should be previsited already (checked with assert(is_visited(n))). 486 int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; } 487 488 // Mark as post visited 489 void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; } 490 491 // Set/get control node out. Set lower bit to distinguish from IdealLoopTree 492 // Returns true if "n" is a data node, false if it's a control node. 493 bool has_ctrl( Node *n ) const { return ((intptr_t)_nodes[n->_idx]) & 1; } 494 495 // clear out dead code after build_loop_late 496 Node_List _deadlist; 497 498 // Support for faster execution of get_late_ctrl()/dom_lca() 499 // when a node has many uses and dominator depth is deep. 500 Node_Array _dom_lca_tags; 501 void init_dom_lca_tags(); 502 void clear_dom_lca_tags(); 503 // Inline wrapper for frequent cases: 504 // 1) only one use 505 // 2) a use is the same as the current LCA passed as 'n1' 506 Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) { 507 assert( n->is_CFG(), "" ); 508 // Fast-path NULL lca 509 if( lca != NULL && lca != n ) { 510 assert( lca->is_CFG(), "" ); 511 // find LCA of all uses 512 n = dom_lca_for_get_late_ctrl_internal( lca, n, tag ); 513 } 514 return find_non_split_ctrl(n); 515 } 516 Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag ); 517 // true if CFG node d dominates CFG node n 518 bool is_dominator(Node *d, Node *n); 519 520 // Helper function for directing control inputs away from CFG split 521 // points. 522 Node *find_non_split_ctrl( Node *ctrl ) const { 523 if (ctrl != NULL) { 524 if (ctrl->is_MultiBranch()) { 525 ctrl = ctrl->in(0); 526 } 527 assert(ctrl->is_CFG(), "CFG"); 528 } 529 return ctrl; 530 } 531 532 public: 533 bool has_node( Node* n ) const { return _nodes[n->_idx] != NULL; } 534 // check if transform created new nodes that need _ctrl recorded 535 Node *get_late_ctrl( Node *n, Node *early ); 536 Node *get_early_ctrl( Node *n ); 537 void set_early_ctrl( Node *n ); 538 void set_subtree_ctrl( Node *root ); 539 void set_ctrl( Node *n, Node *ctrl ) { 540 assert( !has_node(n) || has_ctrl(n), "" ); 541 assert( ctrl->in(0), "cannot set dead control node" ); 542 assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" ); 543 _nodes.map( n->_idx, (Node*)((intptr_t)ctrl + 1) ); 544 } 545 // Set control and update loop membership 546 void set_ctrl_and_loop(Node* n, Node* ctrl) { 547 IdealLoopTree* old_loop = get_loop(get_ctrl(n)); 548 IdealLoopTree* new_loop = get_loop(ctrl); 549 if (old_loop != new_loop) { 550 if (old_loop->_child == NULL) old_loop->_body.yank(n); 551 if (new_loop->_child == NULL) new_loop->_body.push(n); 552 } 553 set_ctrl(n, ctrl); 554 } 555 // Control nodes can be replaced or subsumed. During this pass they 556 // get their replacement Node in slot 1. Instead of updating the block 557 // location of all Nodes in the subsumed block, we lazily do it. As we 558 // pull such a subsumed block out of the array, we write back the final 559 // correct block. 560 Node *get_ctrl( Node *i ) { 561 assert(has_node(i), ""); 562 Node *n = get_ctrl_no_update(i); 563 _nodes.map( i->_idx, (Node*)((intptr_t)n + 1) ); 564 assert(has_node(i) && has_ctrl(i), ""); 565 assert(n == find_non_split_ctrl(n), "must return legal ctrl" ); 566 return n; 567 } 568 569 private: 570 Node *get_ctrl_no_update( Node *i ) const { 571 assert( has_ctrl(i), "" ); 572 Node *n = (Node*)(((intptr_t)_nodes[i->_idx]) & ~1); 573 if (!n->in(0)) { 574 // Skip dead CFG nodes 575 do { 576 n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1); 577 } while (!n->in(0)); 578 n = find_non_split_ctrl(n); 579 } 580 return n; 581 } 582 583 // Check for loop being set 584 // "n" must be a control node. Returns true if "n" is known to be in a loop. 585 bool has_loop( Node *n ) const { 586 assert(!has_node(n) || !has_ctrl(n), ""); 587 return has_node(n); 588 } 589 // Set loop 590 void set_loop( Node *n, IdealLoopTree *loop ) { 591 _nodes.map(n->_idx, (Node*)loop); 592 } 593 // Lazy-dazy update of 'get_ctrl' and 'idom_at' mechanisms. Replace 594 // the 'old_node' with 'new_node'. Kill old-node. Add a reference 595 // from old_node to new_node to support the lazy update. Reference 596 // replaces loop reference, since that is not neede for dead node. 597 public: 598 void lazy_update( Node *old_node, Node *new_node ) { 599 assert( old_node != new_node, "no cycles please" ); 600 //old_node->set_req( 1, new_node /*NO DU INFO*/ ); 601 // Nodes always have DU info now, so re-use the side array slot 602 // for this node to provide the forwarding pointer. 603 _nodes.map( old_node->_idx, (Node*)((intptr_t)new_node + 1) ); 604 } 605 void lazy_replace( Node *old_node, Node *new_node ) { 606 _igvn.hash_delete(old_node); 607 _igvn.subsume_node( old_node, new_node ); 608 lazy_update( old_node, new_node ); 609 } 610 void lazy_replace_proj( Node *old_node, Node *new_node ) { 611 assert( old_node->req() == 1, "use this for Projs" ); 612 _igvn.hash_delete(old_node); // Must hash-delete before hacking edges 613 old_node->add_req( NULL ); 614 lazy_replace( old_node, new_node ); 615 } 616 617 private: 618 619 // Place 'n' in some loop nest, where 'n' is a CFG node 620 void build_loop_tree(); 621 int build_loop_tree_impl( Node *n, int pre_order ); 622 // Insert loop into the existing loop tree. 'innermost' is a leaf of the 623 // loop tree, not the root. 624 IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost ); 625 626 // Place Data nodes in some loop nest 627 void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack, const PhaseIdealLoop *verify_me ); 628 void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack, const PhaseIdealLoop *verify_me ); 629 void build_loop_late_post ( Node* n, const PhaseIdealLoop *verify_me ); 630 631 // Array of immediate dominance info for each CFG node indexed by node idx 632 private: 633 uint _idom_size; 634 Node **_idom; // Array of immediate dominators 635 uint *_dom_depth; // Used for fast LCA test 636 GrowableArray<uint>* _dom_stk; // For recomputation of dom depth 637 638 Node* idom_no_update(Node* d) const { 639 assert(d->_idx < _idom_size, "oob"); 640 Node* n = _idom[d->_idx]; 641 assert(n != NULL,"Bad immediate dominator info."); 642 while (n->in(0) == NULL) { // Skip dead CFG nodes 643 //n = n->in(1); 644 n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1); 645 assert(n != NULL,"Bad immediate dominator info."); 646 } 647 return n; 648 } 649 Node *idom(Node* d) const { 650 uint didx = d->_idx; 651 Node *n = idom_no_update(d); 652 _idom[didx] = n; // Lazily remove dead CFG nodes from table. 653 return n; 654 } 655 uint dom_depth(Node* d) const { 656 assert(d->_idx < _idom_size, ""); 657 return _dom_depth[d->_idx]; 658 } 659 void set_idom(Node* d, Node* n, uint dom_depth); 660 // Locally compute IDOM using dom_lca call 661 Node *compute_idom( Node *region ) const; 662 // Recompute dom_depth 663 void recompute_dom_depth(); 664 665 // Is safept not required by an outer loop? 666 bool is_deleteable_safept(Node* sfpt); 667 668 public: 669 // Dominators for the sea of nodes 670 void Dominators(); 671 Node *dom_lca( Node *n1, Node *n2 ) const { 672 return find_non_split_ctrl(dom_lca_internal(n1, n2)); 673 } 674 Node *dom_lca_internal( Node *n1, Node *n2 ) const; 675 676 // Compute the Ideal Node to Loop mapping 677 PhaseIdealLoop( PhaseIterGVN &igvn, const PhaseIdealLoop *verify_me, bool do_split_ifs ); 678 679 // True if the method has at least 1 irreducible loop 680 bool _has_irreducible_loops; 681 682 // Per-Node transform 683 virtual Node *transform( Node *a_node ) { return 0; } 684 685 Node *is_counted_loop( Node *x, IdealLoopTree *loop ); 686 687 // Return a post-walked LoopNode 688 IdealLoopTree *get_loop( Node *n ) const { 689 // Dead nodes have no loop, so return the top level loop instead 690 if (!has_node(n)) return _ltree_root; 691 assert(!has_ctrl(n), ""); 692 return (IdealLoopTree*)_nodes[n->_idx]; 693 } 694 695 // Is 'n' a (nested) member of 'loop'? 696 int is_member( const IdealLoopTree *loop, Node *n ) const { 697 return loop->is_member(get_loop(n)); } 698 699 // This is the basic building block of the loop optimizations. It clones an 700 // entire loop body. It makes an old_new loop body mapping; with this 701 // mapping you can find the new-loop equivalent to an old-loop node. All 702 // new-loop nodes are exactly equal to their old-loop counterparts, all 703 // edges are the same. All exits from the old-loop now have a RegionNode 704 // that merges the equivalent new-loop path. This is true even for the 705 // normal "loop-exit" condition. All uses of loop-invariant old-loop values 706 // now come from (one or more) Phis that merge their new-loop equivalents. 707 // Parameter side_by_side_idom: 708 // When side_by_size_idom is NULL, the dominator tree is constructed for 709 // the clone loop to dominate the original. Used in construction of 710 // pre-main-post loop sequence. 711 // When nonnull, the clone and original are side-by-side, both are 712 // dominated by the passed in side_by_side_idom node. Used in 713 // construction of unswitched loops. 714 void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth, 715 Node* side_by_side_idom = NULL); 716 717 // If we got the effect of peeling, either by actually peeling or by 718 // making a pre-loop which must execute at least once, we can remove 719 // all loop-invariant dominated tests in the main body. 720 void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new ); 721 722 // Generate code to do a loop peel for the given loop (and body). 723 // old_new is a temp array. 724 void do_peeling( IdealLoopTree *loop, Node_List &old_new ); 725 726 // Add pre and post loops around the given loop. These loops are used 727 // during RCE, unrolling and aligning loops. 728 void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ); 729 // If Node n lives in the back_ctrl block, we clone a private version of n 730 // in preheader_ctrl block and return that, otherwise return n. 731 Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n ); 732 733 // Take steps to maximally unroll the loop. Peel any odd iterations, then 734 // unroll to do double iterations. The next round of major loop transforms 735 // will repeat till the doubled loop body does all remaining iterations in 1 736 // pass. 737 void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ); 738 739 // Unroll the loop body one step - make each trip do 2 iterations. 740 void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ); 741 742 // Return true if exp is a constant times an induction var 743 bool is_scaled_iv(Node* exp, Node* iv, int* p_scale); 744 745 // Return true if exp is a scaled induction var plus (or minus) constant 746 bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth = 0); 747 748 // Eliminate range-checks and other trip-counter vs loop-invariant tests. 749 void do_range_check( IdealLoopTree *loop, Node_List &old_new ); 750 751 // Create a slow version of the loop by cloning the loop 752 // and inserting an if to select fast-slow versions. 753 ProjNode* create_slow_version_of_loop(IdealLoopTree *loop, 754 Node_List &old_new); 755 756 // Clone loop with an invariant test (that does not exit) and 757 // insert a clone of the test that selects which version to 758 // execute. 759 void do_unswitching (IdealLoopTree *loop, Node_List &old_new); 760 761 // Find candidate "if" for unswitching 762 IfNode* find_unswitching_candidate(const IdealLoopTree *loop) const; 763 764 // Range Check Elimination uses this function! 765 // Constrain the main loop iterations so the affine function: 766 // scale_con * I + offset < limit 767 // always holds true. That is, either increase the number of iterations in 768 // the pre-loop or the post-loop until the condition holds true in the main 769 // loop. Scale_con, offset and limit are all loop invariant. 770 void add_constraint( int stride_con, int scale_con, Node *offset, Node *limit, Node *pre_ctrl, Node **pre_limit, Node **main_limit ); 771 772 // Partially peel loop up through last_peel node. 773 bool partial_peel( IdealLoopTree *loop, Node_List &old_new ); 774 775 // Create a scheduled list of nodes control dependent on ctrl set. 776 void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched ); 777 // Has a use in the vector set 778 bool has_use_in_set( Node* n, VectorSet& vset ); 779 // Has use internal to the vector set (ie. not in a phi at the loop head) 780 bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ); 781 // clone "n" for uses that are outside of loop 782 void clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ); 783 // clone "n" for special uses that are in the not_peeled region 784 void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 785 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ); 786 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 787 void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ); 788 #ifdef ASSERT 789 // Validate the loop partition sets: peel and not_peel 790 bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel ); 791 // Ensure that uses outside of loop are of the right form 792 bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 793 uint orig_exit_idx, uint clone_exit_idx); 794 bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx); 795 #endif 796 797 // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.) 798 int stride_of_possible_iv( Node* iff ); 799 bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; } 800 // Return the (unique) control output node that's in the loop (if it exists.) 801 Node* stay_in_loop( Node* n, IdealLoopTree *loop); 802 // Insert a signed compare loop exit cloned from an unsigned compare. 803 IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop); 804 void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop); 805 // Utility to register node "n" with PhaseIdealLoop 806 void register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth); 807 // Utility to create an if-projection 808 ProjNode* proj_clone(ProjNode* p, IfNode* iff); 809 // Force the iff control output to be the live_proj 810 Node* short_circuit_if(IfNode* iff, ProjNode* live_proj); 811 // Insert a region before an if projection 812 RegionNode* insert_region_before_proj(ProjNode* proj); 813 // Insert a new if before an if projection 814 ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj); 815 816 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 817 // "Nearly" because all Nodes have been cloned from the original in the loop, 818 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 819 // through the Phi recursively, and return a Bool. 820 BoolNode *clone_iff( PhiNode *phi, IdealLoopTree *loop ); 821 CmpNode *clone_bool( PhiNode *phi, IdealLoopTree *loop ); 822 823 824 // Rework addressing expressions to get the most loop-invariant stuff 825 // moved out. We'd like to do all associative operators, but it's especially 826 // important (common) to do address expressions. 827 Node *remix_address_expressions( Node *n ); 828 829 // Attempt to use a conditional move instead of a phi/branch 830 Node *conditional_move( Node *n ); 831 832 // Reorganize offset computations to lower register pressure. 833 // Mostly prevent loop-fallout uses of the pre-incremented trip counter 834 // (which are then alive with the post-incremented trip counter 835 // forcing an extra register move) 836 void reorg_offsets( IdealLoopTree *loop ); 837 838 // Check for aggressive application of 'split-if' optimization, 839 // using basic block level info. 840 void split_if_with_blocks ( VectorSet &visited, Node_Stack &nstack ); 841 Node *split_if_with_blocks_pre ( Node *n ); 842 void split_if_with_blocks_post( Node *n ); 843 Node *has_local_phi_input( Node *n ); 844 // Mark an IfNode as being dominated by a prior test, 845 // without actually altering the CFG (and hence IDOM info). 846 void dominated_by( Node *prevdom, Node *iff ); 847 848 // Split Node 'n' through merge point 849 Node *split_thru_region( Node *n, Node *region ); 850 // Split Node 'n' through merge point if there is enough win. 851 Node *split_thru_phi( Node *n, Node *region, int policy ); 852 // Found an If getting its condition-code input from a Phi in the 853 // same block. Split thru the Region. 854 void do_split_if( Node *iff ); 855 856 private: 857 // Return a type based on condition control flow 858 const TypeInt* filtered_type( Node *n, Node* n_ctrl); 859 const TypeInt* filtered_type( Node *n ) { return filtered_type(n, NULL); } 860 // Helpers for filtered type 861 const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl); 862 863 // Helper functions 864 void register_new_node( Node *n, Node *blk ); 865 Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache ); 866 Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true ); 867 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 ); 868 bool split_up( Node *n, Node *blk1, Node *blk2 ); 869 void sink_use( Node *use, Node *post_loop ); 870 Node *place_near_use( Node *useblock ) const; 871 872 bool _created_loop_node; 873 public: 874 void set_created_loop_node() { _created_loop_node = true; } 875 bool created_loop_node() { return _created_loop_node; } 876 877 #ifndef PRODUCT 878 void dump( ) const; 879 void dump( IdealLoopTree *loop, uint rpo_idx, Node_List &rpo_list ) const; 880 void rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const; 881 void verify() const; // Major slow :-) 882 void verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const; 883 IdealLoopTree *get_loop_idx(Node* n) const { 884 // Dead nodes have no loop, so return the top level loop instead 885 return _nodes[n->_idx] ? (IdealLoopTree*)_nodes[n->_idx] : _ltree_root; 886 } 887 // Print some stats 888 static void print_statistics(); 889 static int _loop_invokes; // Count of PhaseIdealLoop invokes 890 static int _loop_work; // Sum of PhaseIdealLoop x _unique 891 #endif 892 }; 893 894 inline Node* IdealLoopTree::tail() { 895 // Handle lazy update of _tail field 896 Node *n = _tail; 897 //while( !n->in(0) ) // Skip dead CFG nodes 898 //n = n->in(1); 899 if (n->in(0) == NULL) 900 n = _phase->get_ctrl(n); 901 _tail = n; 902 return n; 903 } 904 905 906 // Iterate over the loop tree using a preorder, left-to-right traversal. 907 // 908 // Example that visits all counted loops from within PhaseIdealLoop 909 // 910 // for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { 911 // IdealLoopTree* lpt = iter.current(); 912 // if (!lpt->is_counted()) continue; 913 // ... 914 class LoopTreeIterator : public StackObj { 915 private: 916 IdealLoopTree* _root; 917 IdealLoopTree* _curnt; 918 919 public: 920 LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {} 921 922 bool done() { return _curnt == NULL; } // Finished iterating? 923 924 void next(); // Advance to next loop tree 925 926 IdealLoopTree* current() { return _curnt; } // Return current value of iterator. 927 }; 928