1 /* 2 * Copyright (c) 1998, 2018, Oracle and/or its affiliates. All rights reserved. 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 Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #ifndef SHARE_VM_OPTO_LOOPNODE_HPP 26 #define SHARE_VM_OPTO_LOOPNODE_HPP 27 28 #include "opto/cfgnode.hpp" 29 #include "opto/multnode.hpp" 30 #include "opto/phaseX.hpp" 31 #include "opto/subnode.hpp" 32 #include "opto/type.hpp" 33 34 class CmpNode; 35 class CountedLoopEndNode; 36 class CountedLoopNode; 37 class IdealLoopTree; 38 class LoopNode; 39 class Node; 40 class OuterStripMinedLoopEndNode; 41 class PathFrequency; 42 class PhaseIdealLoop; 43 class CountedLoopReserveKit; 44 class VectorSet; 45 class Invariance; 46 struct small_cache; 47 48 // 49 // I D E A L I Z E D L O O P S 50 // 51 // Idealized loops are the set of loops I perform more interesting 52 // transformations on, beyond simple hoisting. 53 54 //------------------------------LoopNode--------------------------------------- 55 // Simple loop header. Fall in path on left, loop-back path on right. 56 class LoopNode : public RegionNode { 57 // Size is bigger to hold the flags. However, the flags do not change 58 // the semantics so it does not appear in the hash & cmp functions. 59 virtual uint size_of() const { return sizeof(*this); } 60 protected: 61 uint _loop_flags; 62 // Names for flag bitfields 63 enum { Normal=0, Pre=1, Main=2, Post=3, PreMainPostFlagsMask=3, 64 MainHasNoPreLoop=4, 65 HasExactTripCount=8, 66 InnerLoop=16, 67 PartialPeelLoop=32, 68 PartialPeelFailed=64, 69 HasReductions=128, 70 WasSlpAnalyzed=256, 71 PassedSlpAnalysis=512, 72 DoUnrollOnly=1024, 73 VectorizedLoop=2048, 74 HasAtomicPostLoop=4096, 75 HasRangeChecks=8192, 76 IsMultiversioned=16384, 77 StripMined=32768, 78 ProfileTripFailed=65536}; 79 char _unswitch_count; 80 enum { _unswitch_max=3 }; 81 char _postloop_flags; 82 enum { LoopNotRCEChecked = 0, LoopRCEChecked = 1, RCEPostLoop = 2 }; 83 84 // Expected trip count from profile data 85 float _profile_trip_cnt; 86 87 public: 88 // Names for edge indices 89 enum { Self=0, EntryControl, LoopBackControl }; 90 91 uint is_inner_loop() const { return _loop_flags & InnerLoop; } 92 void set_inner_loop() { _loop_flags |= InnerLoop; } 93 94 uint range_checks_present() const { return _loop_flags & HasRangeChecks; } 95 uint is_multiversioned() const { return _loop_flags & IsMultiversioned; } 96 uint is_vectorized_loop() const { return _loop_flags & VectorizedLoop; } 97 uint is_partial_peel_loop() const { return _loop_flags & PartialPeelLoop; } 98 void set_partial_peel_loop() { _loop_flags |= PartialPeelLoop; } 99 uint partial_peel_has_failed() const { return _loop_flags & PartialPeelFailed; } 100 uint is_strip_mined() const { return _loop_flags & StripMined; } 101 uint is_profile_trip_failed() const { return _loop_flags & ProfileTripFailed; } 102 103 void mark_partial_peel_failed() { _loop_flags |= PartialPeelFailed; } 104 void mark_has_reductions() { _loop_flags |= HasReductions; } 105 void mark_was_slp() { _loop_flags |= WasSlpAnalyzed; } 106 void mark_passed_slp() { _loop_flags |= PassedSlpAnalysis; } 107 void mark_do_unroll_only() { _loop_flags |= DoUnrollOnly; } 108 void mark_loop_vectorized() { _loop_flags |= VectorizedLoop; } 109 void mark_has_atomic_post_loop() { _loop_flags |= HasAtomicPostLoop; } 110 void mark_has_range_checks() { _loop_flags |= HasRangeChecks; } 111 void mark_is_multiversioned() { _loop_flags |= IsMultiversioned; } 112 void mark_strip_mined() { _loop_flags |= StripMined; } 113 void clear_strip_mined() { _loop_flags &= ~StripMined; } 114 void mark_profile_trip_failed() { _loop_flags |= ProfileTripFailed; } 115 116 int unswitch_max() { return _unswitch_max; } 117 int unswitch_count() { return _unswitch_count; } 118 119 int has_been_range_checked() const { return _postloop_flags & LoopRCEChecked; } 120 void set_has_been_range_checked() { _postloop_flags |= LoopRCEChecked; } 121 int is_rce_post_loop() const { return _postloop_flags & RCEPostLoop; } 122 void set_is_rce_post_loop() { _postloop_flags |= RCEPostLoop; } 123 124 void set_unswitch_count(int val) { 125 assert (val <= unswitch_max(), "too many unswitches"); 126 _unswitch_count = val; 127 } 128 129 void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; } 130 float profile_trip_cnt() { return _profile_trip_cnt; } 131 132 LoopNode(Node *entry, Node *backedge) 133 : RegionNode(3), _loop_flags(0), _unswitch_count(0), 134 _postloop_flags(0), _profile_trip_cnt(COUNT_UNKNOWN) { 135 init_class_id(Class_Loop); 136 init_req(EntryControl, entry); 137 init_req(LoopBackControl, backedge); 138 } 139 140 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); 141 virtual int Opcode() const; 142 bool can_be_counted_loop(PhaseTransform* phase) const { 143 return req() == 3 && in(0) != NULL && 144 in(1) != NULL && phase->type(in(1)) != Type::TOP && 145 in(2) != NULL && phase->type(in(2)) != Type::TOP; 146 } 147 bool is_valid_counted_loop() const; 148 #ifndef PRODUCT 149 virtual void dump_spec(outputStream *st) const; 150 #endif 151 152 void verify_strip_mined(int expect_skeleton) const; 153 virtual LoopNode* skip_strip_mined(int expect_skeleton = 1) { return this; } 154 virtual IfTrueNode* outer_loop_tail() const { ShouldNotReachHere(); return NULL; } 155 virtual OuterStripMinedLoopEndNode* outer_loop_end() const { ShouldNotReachHere(); return NULL; } 156 virtual IfFalseNode* outer_loop_exit() const { ShouldNotReachHere(); return NULL; } 157 virtual SafePointNode* outer_safepoint() const { ShouldNotReachHere(); return NULL; } 158 }; 159 160 //------------------------------Counted Loops---------------------------------- 161 // Counted loops are all trip-counted loops, with exactly 1 trip-counter exit 162 // path (and maybe some other exit paths). The trip-counter exit is always 163 // last in the loop. The trip-counter have to stride by a constant; 164 // the exit value is also loop invariant. 165 166 // CountedLoopNodes and CountedLoopEndNodes come in matched pairs. The 167 // CountedLoopNode has the incoming loop control and the loop-back-control 168 // which is always the IfTrue before the matching CountedLoopEndNode. The 169 // CountedLoopEndNode has an incoming control (possibly not the 170 // CountedLoopNode if there is control flow in the loop), the post-increment 171 // trip-counter value, and the limit. The trip-counter value is always of 172 // the form (Op old-trip-counter stride). The old-trip-counter is produced 173 // by a Phi connected to the CountedLoopNode. The stride is constant. 174 // The Op is any commutable opcode, including Add, Mul, Xor. The 175 // CountedLoopEndNode also takes in the loop-invariant limit value. 176 177 // From a CountedLoopNode I can reach the matching CountedLoopEndNode via the 178 // loop-back control. From CountedLoopEndNodes I can reach CountedLoopNodes 179 // via the old-trip-counter from the Op node. 180 181 //------------------------------CountedLoopNode-------------------------------- 182 // CountedLoopNodes head simple counted loops. CountedLoopNodes have as 183 // inputs the incoming loop-start control and the loop-back control, so they 184 // act like RegionNodes. They also take in the initial trip counter, the 185 // loop-invariant stride and the loop-invariant limit value. CountedLoopNodes 186 // produce a loop-body control and the trip counter value. Since 187 // CountedLoopNodes behave like RegionNodes I still have a standard CFG model. 188 189 class CountedLoopNode : public LoopNode { 190 // Size is bigger to hold _main_idx. However, _main_idx does not change 191 // the semantics so it does not appear in the hash & cmp functions. 192 virtual uint size_of() const { return sizeof(*this); } 193 194 // For Pre- and Post-loops during debugging ONLY, this holds the index of 195 // the Main CountedLoop. Used to assert that we understand the graph shape. 196 node_idx_t _main_idx; 197 198 // Known trip count calculated by compute_exact_trip_count() 199 uint _trip_count; 200 201 // Log2 of original loop bodies in unrolled loop 202 int _unrolled_count_log2; 203 204 // Node count prior to last unrolling - used to decide if 205 // unroll,optimize,unroll,optimize,... is making progress 206 int _node_count_before_unroll; 207 208 // If slp analysis is performed we record the maximum 209 // vector mapped unroll factor here 210 int _slp_maximum_unroll_factor; 211 212 public: 213 CountedLoopNode( Node *entry, Node *backedge ) 214 : LoopNode(entry, backedge), _main_idx(0), _trip_count(max_juint), 215 _unrolled_count_log2(0), _node_count_before_unroll(0), 216 _slp_maximum_unroll_factor(0) { 217 init_class_id(Class_CountedLoop); 218 // Initialize _trip_count to the largest possible value. 219 // Will be reset (lower) if the loop's trip count is known. 220 } 221 222 virtual int Opcode() const; 223 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); 224 225 Node *init_control() const { return in(EntryControl); } 226 Node *back_control() const { return in(LoopBackControl); } 227 CountedLoopEndNode *loopexit_or_null() const; 228 CountedLoopEndNode *loopexit() const; 229 Node *init_trip() const; 230 Node *stride() const; 231 int stride_con() const; 232 bool stride_is_con() const; 233 Node *limit() const; 234 Node *incr() const; 235 Node *phi() const; 236 237 // Match increment with optional truncation 238 static Node* match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type); 239 240 // A 'main' loop has a pre-loop and a post-loop. The 'main' loop 241 // can run short a few iterations and may start a few iterations in. 242 // It will be RCE'd and unrolled and aligned. 243 244 // A following 'post' loop will run any remaining iterations. Used 245 // during Range Check Elimination, the 'post' loop will do any final 246 // iterations with full checks. Also used by Loop Unrolling, where 247 // the 'post' loop will do any epilog iterations needed. Basically, 248 // a 'post' loop can not profitably be further unrolled or RCE'd. 249 250 // A preceding 'pre' loop will run at least 1 iteration (to do peeling), 251 // it may do under-flow checks for RCE and may do alignment iterations 252 // so the following main loop 'knows' that it is striding down cache 253 // lines. 254 255 // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or 256 // Aligned, may be missing it's pre-loop. 257 uint is_normal_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Normal; } 258 uint is_pre_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Pre; } 259 uint is_main_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Main; } 260 uint is_post_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Post; } 261 uint is_reduction_loop() const { return (_loop_flags&HasReductions) == HasReductions; } 262 uint was_slp_analyzed () const { return (_loop_flags&WasSlpAnalyzed) == WasSlpAnalyzed; } 263 uint has_passed_slp () const { return (_loop_flags&PassedSlpAnalysis) == PassedSlpAnalysis; } 264 uint do_unroll_only () const { return (_loop_flags&DoUnrollOnly) == DoUnrollOnly; } 265 uint is_main_no_pre_loop() const { return _loop_flags & MainHasNoPreLoop; } 266 uint has_atomic_post_loop () const { return (_loop_flags & HasAtomicPostLoop) == HasAtomicPostLoop; } 267 void set_main_no_pre_loop() { _loop_flags |= MainHasNoPreLoop; } 268 269 int main_idx() const { return _main_idx; } 270 271 272 void set_pre_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; } 273 void set_main_loop ( ) { assert(is_normal_loop(),""); _loop_flags |= Main; } 274 void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; } 275 void set_normal_loop( ) { _loop_flags &= ~PreMainPostFlagsMask; } 276 277 void set_trip_count(uint tc) { _trip_count = tc; } 278 uint trip_count() { return _trip_count; } 279 280 bool has_exact_trip_count() const { return (_loop_flags & HasExactTripCount) != 0; } 281 void set_exact_trip_count(uint tc) { 282 _trip_count = tc; 283 _loop_flags |= HasExactTripCount; 284 } 285 void set_nonexact_trip_count() { 286 _loop_flags &= ~HasExactTripCount; 287 } 288 void set_notpassed_slp() { 289 _loop_flags &= ~PassedSlpAnalysis; 290 } 291 292 void double_unrolled_count() { _unrolled_count_log2++; } 293 int unrolled_count() { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); } 294 295 void set_node_count_before_unroll(int ct) { _node_count_before_unroll = ct; } 296 int node_count_before_unroll() { return _node_count_before_unroll; } 297 void set_slp_max_unroll(int unroll_factor) { _slp_maximum_unroll_factor = unroll_factor; } 298 int slp_max_unroll() const { return _slp_maximum_unroll_factor; } 299 300 virtual LoopNode* skip_strip_mined(int expect_opaq = 1); 301 OuterStripMinedLoopNode* outer_loop() const; 302 virtual IfTrueNode* outer_loop_tail() const; 303 virtual OuterStripMinedLoopEndNode* outer_loop_end() const; 304 virtual IfFalseNode* outer_loop_exit() const; 305 virtual SafePointNode* outer_safepoint() const; 306 307 // If this is a main loop in a pre/main/post loop nest, walk over 308 // the predicates that were inserted by 309 // duplicate_predicates()/add_range_check_predicate() 310 static Node* skip_predicates_from_entry(Node* ctrl); 311 Node* skip_predicates(); 312 313 #ifndef PRODUCT 314 virtual void dump_spec(outputStream *st) const; 315 #endif 316 }; 317 318 //------------------------------CountedLoopEndNode----------------------------- 319 // CountedLoopEndNodes end simple trip counted loops. They act much like 320 // IfNodes. 321 class CountedLoopEndNode : public IfNode { 322 public: 323 enum { TestControl, TestValue }; 324 325 CountedLoopEndNode( Node *control, Node *test, float prob, float cnt ) 326 : IfNode( control, test, prob, cnt) { 327 init_class_id(Class_CountedLoopEnd); 328 } 329 virtual int Opcode() const; 330 331 Node *cmp_node() const { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : NULL; } 332 Node *incr() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; } 333 Node *limit() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; } 334 Node *stride() const { Node *tmp = incr (); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; } 335 Node *init_trip() const { Node *tmp = phi (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; } 336 int stride_con() const; 337 bool stride_is_con() const { Node *tmp = stride (); return (tmp != NULL && tmp->is_Con()); } 338 BoolTest::mask test_trip() const { return in(TestValue)->as_Bool()->_test._test; } 339 PhiNode *phi() const { 340 Node *tmp = incr(); 341 if (tmp && tmp->req() == 3) { 342 Node* phi = tmp->in(1); 343 if (phi->is_Phi()) { 344 return phi->as_Phi(); 345 } 346 } 347 return NULL; 348 } 349 CountedLoopNode *loopnode() const { 350 // The CountedLoopNode that goes with this CountedLoopEndNode may 351 // have been optimized out by the IGVN so be cautious with the 352 // pattern matching on the graph 353 PhiNode* iv_phi = phi(); 354 if (iv_phi == NULL) { 355 return NULL; 356 } 357 Node *ln = iv_phi->in(0); 358 if (ln->is_CountedLoop() && ln->as_CountedLoop()->loopexit_or_null() == this) { 359 return (CountedLoopNode*)ln; 360 } 361 return NULL; 362 } 363 364 #ifndef PRODUCT 365 virtual void dump_spec(outputStream *st) const; 366 #endif 367 }; 368 369 370 inline CountedLoopEndNode *CountedLoopNode::loopexit_or_null() const { 371 Node *bc = back_control(); 372 if( bc == NULL ) return NULL; 373 Node *le = bc->in(0); 374 if( le->Opcode() != Op_CountedLoopEnd ) 375 return NULL; 376 return (CountedLoopEndNode*)le; 377 } 378 inline CountedLoopEndNode *CountedLoopNode::loopexit() const { 379 CountedLoopEndNode* cle = loopexit_or_null(); 380 assert(cle != NULL, "loopexit is NULL"); 381 return cle; 382 } 383 inline Node *CountedLoopNode::init_trip() const { return loopexit_or_null() ? loopexit()->init_trip() : NULL; } 384 inline Node *CountedLoopNode::stride() const { return loopexit_or_null() ? loopexit()->stride() : NULL; } 385 inline int CountedLoopNode::stride_con() const { return loopexit_or_null() ? loopexit()->stride_con() : 0; } 386 inline bool CountedLoopNode::stride_is_con() const { return loopexit_or_null() && loopexit()->stride_is_con(); } 387 inline Node *CountedLoopNode::limit() const { return loopexit_or_null() ? loopexit()->limit() : NULL; } 388 inline Node *CountedLoopNode::incr() const { return loopexit_or_null() ? loopexit()->incr() : NULL; } 389 inline Node *CountedLoopNode::phi() const { return loopexit_or_null() ? loopexit()->phi() : NULL; } 390 391 //------------------------------LoopLimitNode----------------------------- 392 // Counted Loop limit node which represents exact final iterator value: 393 // trip_count = (limit - init_trip + stride - 1)/stride 394 // final_value= trip_count * stride + init_trip. 395 // Use HW instructions to calculate it when it can overflow in integer. 396 // Note, final_value should fit into integer since counted loop has 397 // limit check: limit <= max_int-stride. 398 class LoopLimitNode : public Node { 399 enum { Init=1, Limit=2, Stride=3 }; 400 public: 401 LoopLimitNode( Compile* C, Node *init, Node *limit, Node *stride ) : Node(0,init,limit,stride) { 402 // Put it on the Macro nodes list to optimize during macro nodes expansion. 403 init_flags(Flag_is_macro); 404 C->add_macro_node(this); 405 } 406 virtual int Opcode() const; 407 virtual const Type *bottom_type() const { return TypeInt::INT; } 408 virtual uint ideal_reg() const { return Op_RegI; } 409 virtual const Type* Value(PhaseGVN* phase) const; 410 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); 411 virtual Node* Identity(PhaseGVN* phase); 412 }; 413 414 // Support for strip mining 415 class OuterStripMinedLoopNode : public LoopNode { 416 private: 417 CountedLoopNode* inner_loop() const; 418 public: 419 OuterStripMinedLoopNode(Compile* C, Node *entry, Node *backedge) 420 : LoopNode(entry, backedge) { 421 init_class_id(Class_OuterStripMinedLoop); 422 init_flags(Flag_is_macro); 423 C->add_macro_node(this); 424 } 425 426 virtual int Opcode() const; 427 428 virtual IfTrueNode* outer_loop_tail() const; 429 virtual OuterStripMinedLoopEndNode* outer_loop_end() const; 430 virtual IfFalseNode* outer_loop_exit() const; 431 virtual SafePointNode* outer_safepoint() const; 432 void adjust_strip_mined_loop(PhaseIterGVN* igvn); 433 }; 434 435 class OuterStripMinedLoopEndNode : public IfNode { 436 public: 437 OuterStripMinedLoopEndNode(Node *control, Node *test, float prob, float cnt) 438 : IfNode(control, test, prob, cnt) { 439 init_class_id(Class_OuterStripMinedLoopEnd); 440 } 441 442 virtual int Opcode() const; 443 444 virtual const Type* Value(PhaseGVN* phase) const; 445 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); 446 }; 447 448 // -----------------------------IdealLoopTree---------------------------------- 449 class IdealLoopTree : public ResourceObj { 450 public: 451 IdealLoopTree *_parent; // Parent in loop tree 452 IdealLoopTree *_next; // Next sibling in loop tree 453 IdealLoopTree *_child; // First child in loop tree 454 455 // The head-tail backedge defines the loop. 456 // If tail is NULL then this loop has multiple backedges as part of the 457 // same loop. During cleanup I'll peel off the multiple backedges; merge 458 // them at the loop bottom and flow 1 real backedge into the loop. 459 Node *_head; // Head of loop 460 Node *_tail; // Tail of loop 461 inline Node *tail(); // Handle lazy update of _tail field 462 PhaseIdealLoop* _phase; 463 int _local_loop_unroll_limit; 464 int _local_loop_unroll_factor; 465 466 Node_List _body; // Loop body for inner loops 467 468 uint8_t _nest; // Nesting depth 469 uint8_t _irreducible:1, // True if irreducible 470 _has_call:1, // True if has call safepoint 471 _has_sfpt:1, // True if has non-call safepoint 472 _rce_candidate:1; // True if candidate for range check elimination 473 474 Node_List* _safepts; // List of safepoints in this loop 475 Node_List* _required_safept; // A inner loop cannot delete these safepts; 476 bool _allow_optimizations; // Allow loop optimizations 477 478 IdealLoopTree( PhaseIdealLoop* phase, Node *head, Node *tail ) 479 : _parent(0), _next(0), _child(0), 480 _head(head), _tail(tail), 481 _phase(phase), 482 _safepts(NULL), 483 _required_safept(NULL), 484 _allow_optimizations(true), 485 _nest(0), _irreducible(0), _has_call(0), _has_sfpt(0), _rce_candidate(0), 486 _local_loop_unroll_limit(0), _local_loop_unroll_factor(0) 487 { } 488 489 // Is 'l' a member of 'this'? 490 bool is_member(const IdealLoopTree *l) const; // Test for nested membership 491 492 // Set loop nesting depth. Accumulate has_call bits. 493 int set_nest( uint depth ); 494 495 // Split out multiple fall-in edges from the loop header. Move them to a 496 // private RegionNode before the loop. This becomes the loop landing pad. 497 void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ); 498 499 // Split out the outermost loop from this shared header. 500 void split_outer_loop( PhaseIdealLoop *phase ); 501 502 // Merge all the backedges from the shared header into a private Region. 503 // Feed that region as the one backedge to this loop. 504 void merge_many_backedges( PhaseIdealLoop *phase ); 505 506 // Split shared headers and insert loop landing pads. 507 // Insert a LoopNode to replace the RegionNode. 508 // Returns TRUE if loop tree is structurally changed. 509 bool beautify_loops( PhaseIdealLoop *phase ); 510 511 // Perform optimization to use the loop predicates for null checks and range checks. 512 // Applies to any loop level (not just the innermost one) 513 bool loop_predication( PhaseIdealLoop *phase); 514 515 // Perform iteration-splitting on inner loops. Split iterations to 516 // avoid range checks or one-shot null checks. Returns false if the 517 // current round of loop opts should stop. 518 bool iteration_split( PhaseIdealLoop *phase, Node_List &old_new ); 519 520 // Driver for various flavors of iteration splitting. Returns false 521 // if the current round of loop opts should stop. 522 bool iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ); 523 524 // Given dominators, try to find loops with calls that must always be 525 // executed (call dominates loop tail). These loops do not need non-call 526 // safepoints (ncsfpt). 527 void check_safepts(VectorSet &visited, Node_List &stack); 528 529 // Allpaths backwards scan from loop tail, terminating each path at first safepoint 530 // encountered. 531 void allpaths_check_safepts(VectorSet &visited, Node_List &stack); 532 533 // Remove safepoints from loop. Optionally keeping one. 534 void remove_safepoints(PhaseIdealLoop* phase, bool keep_one); 535 536 // Convert to counted loops where possible 537 void counted_loop( PhaseIdealLoop *phase ); 538 539 // Check for Node being a loop-breaking test 540 Node *is_loop_exit(Node *iff) const; 541 542 // Remove simplistic dead code from loop body 543 void DCE_loop_body(); 544 545 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage. 546 // Replace with a 1-in-10 exit guess. 547 void adjust_loop_exit_prob( PhaseIdealLoop *phase ); 548 549 // Return TRUE or FALSE if the loop should never be RCE'd or aligned. 550 // Useful for unrolling loops with NO array accesses. 551 bool policy_peel_only( PhaseIdealLoop *phase ) const; 552 553 // Return TRUE or FALSE if the loop should be unswitched -- clone 554 // loop with an invariant test 555 bool policy_unswitching( PhaseIdealLoop *phase ) const; 556 557 // Micro-benchmark spamming. Remove empty loops. 558 bool policy_do_remove_empty_loop( PhaseIdealLoop *phase ); 559 560 // Convert one iteration loop into normal code. 561 bool policy_do_one_iteration_loop( PhaseIdealLoop *phase ); 562 563 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can 564 // make some loop-invariant test (usually a null-check) happen before the 565 // loop. 566 bool policy_peeling( PhaseIdealLoop *phase ) const; 567 568 // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any 569 // known trip count in the counted loop node. 570 bool policy_maximally_unroll( PhaseIdealLoop *phase ) const; 571 572 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if 573 // the loop is a CountedLoop and the body is small enough. 574 bool policy_unroll(PhaseIdealLoop *phase); 575 576 // Loop analyses to map to a maximal superword unrolling for vectorization. 577 void policy_unroll_slp_analysis(CountedLoopNode *cl, PhaseIdealLoop *phase, int future_unroll_ct); 578 579 // Return TRUE or FALSE if the loop should be range-check-eliminated. 580 // Gather a list of IF tests that are dominated by iteration splitting; 581 // also gather the end of the first split and the start of the 2nd split. 582 bool policy_range_check( PhaseIdealLoop *phase ) const; 583 584 // Return TRUE or FALSE if the loop should be cache-line aligned. 585 // Gather the expression that does the alignment. Note that only 586 // one array base can be aligned in a loop (unless the VM guarantees 587 // mutual alignment). Note that if we vectorize short memory ops 588 // into longer memory ops, we may want to increase alignment. 589 bool policy_align( PhaseIdealLoop *phase ) const; 590 591 // Return TRUE if "iff" is a range check. 592 bool is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const; 593 594 // Compute loop trip count if possible 595 void compute_trip_count(PhaseIdealLoop* phase); 596 597 // Compute loop trip count from profile data 598 float compute_profile_trip_cnt_helper(Node* n); 599 void compute_profile_trip_cnt( PhaseIdealLoop *phase ); 600 601 // Reassociate invariant expressions. 602 void reassociate_invariants(PhaseIdealLoop *phase); 603 // Reassociate invariant add and subtract expressions. 604 Node* reassociate_add_sub(Node* n1, PhaseIdealLoop *phase); 605 // Return nonzero index of invariant operand if invariant and variant 606 // are combined with an Add or Sub. Helper for reassociate_invariants. 607 int is_invariant_addition(Node* n, PhaseIdealLoop *phase); 608 609 // Return true if n is invariant 610 bool is_invariant(Node* n) const; 611 612 // Put loop body on igvn work list 613 void record_for_igvn(); 614 615 bool is_loop() { return !_irreducible && _tail && !_tail->is_top(); } 616 bool is_inner() { return is_loop() && _child == NULL; } 617 bool is_counted() { return is_loop() && _head != NULL && _head->is_CountedLoop(); } 618 619 void remove_main_post_loops(CountedLoopNode *cl, PhaseIdealLoop *phase); 620 621 #ifndef PRODUCT 622 void dump_head( ) const; // Dump loop head only 623 void dump() const; // Dump this loop recursively 624 void verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const; 625 #endif 626 627 }; 628 629 // -----------------------------PhaseIdealLoop--------------------------------- 630 // Computes the mapping from Nodes to IdealLoopTrees. Organizes IdealLoopTrees into a 631 // loop tree. Drives the loop-based transformations on the ideal graph. 632 class PhaseIdealLoop : public PhaseTransform { 633 friend class IdealLoopTree; 634 friend class SuperWord; 635 friend class CountedLoopReserveKit; 636 637 // Pre-computed def-use info 638 PhaseIterGVN &_igvn; 639 640 // Head of loop tree 641 IdealLoopTree *_ltree_root; 642 643 // Array of pre-order numbers, plus post-visited bit. 644 // ZERO for not pre-visited. EVEN for pre-visited but not post-visited. 645 // ODD for post-visited. Other bits are the pre-order number. 646 uint *_preorders; 647 uint _max_preorder; 648 649 const PhaseIdealLoop* _verify_me; 650 bool _verify_only; 651 652 // Allocate _preorders[] array 653 void allocate_preorders() { 654 _max_preorder = C->unique()+8; 655 _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder); 656 memset(_preorders, 0, sizeof(uint) * _max_preorder); 657 } 658 659 // Allocate _preorders[] array 660 void reallocate_preorders() { 661 if ( _max_preorder < C->unique() ) { 662 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique()); 663 _max_preorder = C->unique(); 664 } 665 memset(_preorders, 0, sizeof(uint) * _max_preorder); 666 } 667 668 // Check to grow _preorders[] array for the case when build_loop_tree_impl() 669 // adds new nodes. 670 void check_grow_preorders( ) { 671 if ( _max_preorder < C->unique() ) { 672 uint newsize = _max_preorder<<1; // double size of array 673 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize); 674 memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder)); 675 _max_preorder = newsize; 676 } 677 } 678 // Check for pre-visited. Zero for NOT visited; non-zero for visited. 679 int is_visited( Node *n ) const { return _preorders[n->_idx]; } 680 // Pre-order numbers are written to the Nodes array as low-bit-set values. 681 void set_preorder_visited( Node *n, int pre_order ) { 682 assert( !is_visited( n ), "already set" ); 683 _preorders[n->_idx] = (pre_order<<1); 684 }; 685 // Return pre-order number. 686 int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; } 687 688 // Check for being post-visited. 689 // Should be previsited already (checked with assert(is_visited(n))). 690 int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; } 691 692 // Mark as post visited 693 void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; } 694 695 // Set/get control node out. Set lower bit to distinguish from IdealLoopTree 696 // Returns true if "n" is a data node, false if it's a control node. 697 bool has_ctrl( Node *n ) const { return ((intptr_t)_nodes[n->_idx]) & 1; } 698 699 // clear out dead code after build_loop_late 700 Node_List _deadlist; 701 702 // Support for faster execution of get_late_ctrl()/dom_lca() 703 // when a node has many uses and dominator depth is deep. 704 Node_Array _dom_lca_tags; 705 void init_dom_lca_tags(); 706 void clear_dom_lca_tags(); 707 708 // Helper for debugging bad dominance relationships 709 bool verify_dominance(Node* n, Node* use, Node* LCA, Node* early); 710 711 Node* compute_lca_of_uses(Node* n, Node* early, bool verify = false); 712 713 // Inline wrapper for frequent cases: 714 // 1) only one use 715 // 2) a use is the same as the current LCA passed as 'n1' 716 Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) { 717 assert( n->is_CFG(), "" ); 718 // Fast-path NULL lca 719 if( lca != NULL && lca != n ) { 720 assert( lca->is_CFG(), "" ); 721 // find LCA of all uses 722 n = dom_lca_for_get_late_ctrl_internal( lca, n, tag ); 723 } 724 return find_non_split_ctrl(n); 725 } 726 Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag ); 727 728 // Helper function for directing control inputs away from CFG split 729 // points. 730 Node *find_non_split_ctrl( Node *ctrl ) const { 731 if (ctrl != NULL) { 732 if (ctrl->is_MultiBranch()) { 733 ctrl = ctrl->in(0); 734 } 735 assert(ctrl->is_CFG(), "CFG"); 736 } 737 return ctrl; 738 } 739 740 Node* cast_incr_before_loop(Node* incr, Node* ctrl, Node* loop); 741 void duplicate_predicates_helper(Node* predicate, Node* castii, IdealLoopTree* outer_loop, 742 LoopNode* outer_main_head, uint dd_main_head); 743 void duplicate_predicates(CountedLoopNode* pre_head, Node* castii, IdealLoopTree* outer_loop, 744 LoopNode* outer_main_head, uint dd_main_head); 745 746 public: 747 748 static bool is_canonical_loop_entry(CountedLoopNode* cl); 749 750 bool has_node( Node* n ) const { 751 guarantee(n != NULL, "No Node."); 752 return _nodes[n->_idx] != NULL; 753 } 754 // check if transform created new nodes that need _ctrl recorded 755 Node *get_late_ctrl( Node *n, Node *early ); 756 Node *get_early_ctrl( Node *n ); 757 Node *get_early_ctrl_for_expensive(Node *n, Node* earliest); 758 void set_early_ctrl( Node *n ); 759 void set_subtree_ctrl( Node *root ); 760 void set_ctrl( Node *n, Node *ctrl ) { 761 assert( !has_node(n) || has_ctrl(n), "" ); 762 assert( ctrl->in(0), "cannot set dead control node" ); 763 assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" ); 764 _nodes.map( n->_idx, (Node*)((intptr_t)ctrl + 1) ); 765 } 766 // Set control and update loop membership 767 void set_ctrl_and_loop(Node* n, Node* ctrl) { 768 IdealLoopTree* old_loop = get_loop(get_ctrl(n)); 769 IdealLoopTree* new_loop = get_loop(ctrl); 770 if (old_loop != new_loop) { 771 if (old_loop->_child == NULL) old_loop->_body.yank(n); 772 if (new_loop->_child == NULL) new_loop->_body.push(n); 773 } 774 set_ctrl(n, ctrl); 775 } 776 // Control nodes can be replaced or subsumed. During this pass they 777 // get their replacement Node in slot 1. Instead of updating the block 778 // location of all Nodes in the subsumed block, we lazily do it. As we 779 // pull such a subsumed block out of the array, we write back the final 780 // correct block. 781 Node *get_ctrl( Node *i ) { 782 assert(has_node(i), ""); 783 Node *n = get_ctrl_no_update(i); 784 _nodes.map( i->_idx, (Node*)((intptr_t)n + 1) ); 785 assert(has_node(i) && has_ctrl(i), ""); 786 assert(n == find_non_split_ctrl(n), "must return legal ctrl" ); 787 return n; 788 } 789 // true if CFG node d dominates CFG node n 790 bool is_dominator(Node *d, Node *n); 791 // return get_ctrl for a data node and self(n) for a CFG node 792 Node* ctrl_or_self(Node* n) { 793 if (has_ctrl(n)) 794 return get_ctrl(n); 795 else { 796 assert (n->is_CFG(), "must be a CFG node"); 797 return n; 798 } 799 } 800 801 private: 802 Node *get_ctrl_no_update_helper(Node *i) const { 803 assert(has_ctrl(i), "should be control, not loop"); 804 return (Node*)(((intptr_t)_nodes[i->_idx]) & ~1); 805 } 806 807 Node *get_ctrl_no_update(Node *i) const { 808 assert( has_ctrl(i), "" ); 809 Node *n = get_ctrl_no_update_helper(i); 810 if (!n->in(0)) { 811 // Skip dead CFG nodes 812 do { 813 n = get_ctrl_no_update_helper(n); 814 } while (!n->in(0)); 815 n = find_non_split_ctrl(n); 816 } 817 return n; 818 } 819 820 // Check for loop being set 821 // "n" must be a control node. Returns true if "n" is known to be in a loop. 822 bool has_loop( Node *n ) const { 823 assert(!has_node(n) || !has_ctrl(n), ""); 824 return has_node(n); 825 } 826 // Set loop 827 void set_loop( Node *n, IdealLoopTree *loop ) { 828 _nodes.map(n->_idx, (Node*)loop); 829 } 830 // Lazy-dazy update of 'get_ctrl' and 'idom_at' mechanisms. Replace 831 // the 'old_node' with 'new_node'. Kill old-node. Add a reference 832 // from old_node to new_node to support the lazy update. Reference 833 // replaces loop reference, since that is not needed for dead node. 834 public: 835 void lazy_update(Node *old_node, Node *new_node) { 836 assert(old_node != new_node, "no cycles please"); 837 // Re-use the side array slot for this node to provide the 838 // forwarding pointer. 839 _nodes.map(old_node->_idx, (Node*)((intptr_t)new_node + 1)); 840 } 841 void lazy_replace(Node *old_node, Node *new_node) { 842 _igvn.replace_node(old_node, new_node); 843 lazy_update(old_node, new_node); 844 } 845 846 private: 847 848 // Place 'n' in some loop nest, where 'n' is a CFG node 849 void build_loop_tree(); 850 int build_loop_tree_impl( Node *n, int pre_order ); 851 // Insert loop into the existing loop tree. 'innermost' is a leaf of the 852 // loop tree, not the root. 853 IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost ); 854 855 // Place Data nodes in some loop nest 856 void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ); 857 void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ); 858 void build_loop_late_post ( Node* n ); 859 void verify_strip_mined_scheduling(Node *n, Node* least); 860 861 // Array of immediate dominance info for each CFG node indexed by node idx 862 private: 863 uint _idom_size; 864 Node **_idom; // Array of immediate dominators 865 uint *_dom_depth; // Used for fast LCA test 866 GrowableArray<uint>* _dom_stk; // For recomputation of dom depth 867 868 Node* idom_no_update(Node* d) const { 869 assert(d->_idx < _idom_size, "oob"); 870 Node* n = _idom[d->_idx]; 871 assert(n != NULL,"Bad immediate dominator info."); 872 while (n->in(0) == NULL) { // Skip dead CFG nodes 873 //n = n->in(1); 874 n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1); 875 assert(n != NULL,"Bad immediate dominator info."); 876 } 877 return n; 878 } 879 Node *idom(Node* d) const { 880 uint didx = d->_idx; 881 Node *n = idom_no_update(d); 882 _idom[didx] = n; // Lazily remove dead CFG nodes from table. 883 return n; 884 } 885 uint dom_depth(Node* d) const { 886 guarantee(d != NULL, "Null dominator info."); 887 guarantee(d->_idx < _idom_size, ""); 888 return _dom_depth[d->_idx]; 889 } 890 void set_idom(Node* d, Node* n, uint dom_depth); 891 // Locally compute IDOM using dom_lca call 892 Node *compute_idom( Node *region ) const; 893 // Recompute dom_depth 894 void recompute_dom_depth(); 895 896 // Is safept not required by an outer loop? 897 bool is_deleteable_safept(Node* sfpt); 898 899 // Replace parallel induction variable (parallel to trip counter) 900 void replace_parallel_iv(IdealLoopTree *loop); 901 902 // Perform verification that the graph is valid. 903 PhaseIdealLoop( PhaseIterGVN &igvn) : 904 PhaseTransform(Ideal_Loop), 905 _igvn(igvn), 906 _dom_lca_tags(arena()), // Thread::resource_area 907 _verify_me(NULL), 908 _verify_only(true) { 909 build_and_optimize(false, false); 910 } 911 912 // build the loop tree and perform any requested optimizations 913 void build_and_optimize(bool do_split_if, bool skip_loop_opts); 914 915 public: 916 // Dominators for the sea of nodes 917 void Dominators(); 918 Node *dom_lca( Node *n1, Node *n2 ) const { 919 return find_non_split_ctrl(dom_lca_internal(n1, n2)); 920 } 921 Node *dom_lca_internal( Node *n1, Node *n2 ) const; 922 923 // Compute the Ideal Node to Loop mapping 924 PhaseIdealLoop( PhaseIterGVN &igvn, bool do_split_ifs, bool skip_loop_opts = false) : 925 PhaseTransform(Ideal_Loop), 926 _igvn(igvn), 927 _dom_lca_tags(arena()), // Thread::resource_area 928 _verify_me(NULL), 929 _verify_only(false) { 930 build_and_optimize(do_split_ifs, skip_loop_opts); 931 } 932 933 // Verify that verify_me made the same decisions as a fresh run. 934 PhaseIdealLoop( PhaseIterGVN &igvn, const PhaseIdealLoop *verify_me) : 935 PhaseTransform(Ideal_Loop), 936 _igvn(igvn), 937 _dom_lca_tags(arena()), // Thread::resource_area 938 _verify_me(verify_me), 939 _verify_only(false) { 940 build_and_optimize(false, false); 941 } 942 943 // Build and verify the loop tree without modifying the graph. This 944 // is useful to verify that all inputs properly dominate their uses. 945 static void verify(PhaseIterGVN& igvn) { 946 #ifdef ASSERT 947 PhaseIdealLoop v(igvn); 948 #endif 949 } 950 951 // True if the method has at least 1 irreducible loop 952 bool _has_irreducible_loops; 953 954 // Per-Node transform 955 virtual Node *transform( Node *a_node ) { return 0; } 956 957 bool is_counted_loop(Node* x, IdealLoopTree*& loop); 958 IdealLoopTree* create_outer_strip_mined_loop(BoolNode *test, Node *cmp, Node *init_control, 959 IdealLoopTree* loop, float cl_prob, float le_fcnt, 960 Node*& entry_control, Node*& iffalse); 961 962 Node* exact_limit( IdealLoopTree *loop ); 963 964 // Return a post-walked LoopNode 965 IdealLoopTree *get_loop( Node *n ) const { 966 // Dead nodes have no loop, so return the top level loop instead 967 if (!has_node(n)) return _ltree_root; 968 assert(!has_ctrl(n), ""); 969 return (IdealLoopTree*)_nodes[n->_idx]; 970 } 971 972 // Is 'n' a (nested) member of 'loop'? 973 int is_member( const IdealLoopTree *loop, Node *n ) const { 974 return loop->is_member(get_loop(n)); } 975 976 // This is the basic building block of the loop optimizations. It clones an 977 // entire loop body. It makes an old_new loop body mapping; with this 978 // mapping you can find the new-loop equivalent to an old-loop node. All 979 // new-loop nodes are exactly equal to their old-loop counterparts, all 980 // edges are the same. All exits from the old-loop now have a RegionNode 981 // that merges the equivalent new-loop path. This is true even for the 982 // normal "loop-exit" condition. All uses of loop-invariant old-loop values 983 // now come from (one or more) Phis that merge their new-loop equivalents. 984 // Parameter side_by_side_idom: 985 // When side_by_size_idom is NULL, the dominator tree is constructed for 986 // the clone loop to dominate the original. Used in construction of 987 // pre-main-post loop sequence. 988 // When nonnull, the clone and original are side-by-side, both are 989 // dominated by the passed in side_by_side_idom node. Used in 990 // construction of unswitched loops. 991 enum CloneLoopMode { 992 IgnoreStripMined = 0, // Only clone inner strip mined loop 993 CloneIncludesStripMined = 1, // clone both inner and outer strip mined loops 994 ControlAroundStripMined = 2 // Only clone inner strip mined loop, 995 // result control flow branches 996 // either to inner clone or outer 997 // strip mined loop. 998 }; 999 void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth, 1000 CloneLoopMode mode, Node* side_by_side_idom = NULL); 1001 void clone_loop_handle_data_uses(Node* old, Node_List &old_new, 1002 IdealLoopTree* loop, IdealLoopTree* companion_loop, 1003 Node_List*& split_if_set, Node_List*& split_bool_set, 1004 Node_List*& split_cex_set, Node_List& worklist, 1005 uint new_counter, CloneLoopMode mode); 1006 void clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop, 1007 IdealLoopTree* outer_loop, int dd, Node_List &old_new, 1008 Node_List& extra_data_nodes); 1009 1010 // If we got the effect of peeling, either by actually peeling or by 1011 // making a pre-loop which must execute at least once, we can remove 1012 // all loop-invariant dominated tests in the main body. 1013 void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new ); 1014 1015 // Generate code to do a loop peel for the given loop (and body). 1016 // old_new is a temp array. 1017 void do_peeling( IdealLoopTree *loop, Node_List &old_new ); 1018 1019 // Add pre and post loops around the given loop. These loops are used 1020 // during RCE, unrolling and aligning loops. 1021 void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ); 1022 1023 // Add post loop after the given loop. 1024 Node *insert_post_loop(IdealLoopTree *loop, Node_List &old_new, 1025 CountedLoopNode *main_head, CountedLoopEndNode *main_end, 1026 Node *incr, Node *limit, CountedLoopNode *&post_head); 1027 1028 // Add an RCE'd post loop which we will multi-version adapt for run time test path usage 1029 void insert_scalar_rced_post_loop( IdealLoopTree *loop, Node_List &old_new ); 1030 1031 // Add a vector post loop between a vector main loop and the current post loop 1032 void insert_vector_post_loop(IdealLoopTree *loop, Node_List &old_new); 1033 // If Node n lives in the back_ctrl block, we clone a private version of n 1034 // in preheader_ctrl block and return that, otherwise return n. 1035 Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones ); 1036 1037 // Take steps to maximally unroll the loop. Peel any odd iterations, then 1038 // unroll to do double iterations. The next round of major loop transforms 1039 // will repeat till the doubled loop body does all remaining iterations in 1 1040 // pass. 1041 void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ); 1042 1043 // Unroll the loop body one step - make each trip do 2 iterations. 1044 void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ); 1045 1046 // Mark vector reduction candidates before loop unrolling 1047 void mark_reductions( IdealLoopTree *loop ); 1048 1049 // Return true if exp is a constant times an induction var 1050 bool is_scaled_iv(Node* exp, Node* iv, int* p_scale); 1051 1052 // Return true if exp is a scaled induction var plus (or minus) constant 1053 bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth = 0); 1054 1055 // Create a new if above the uncommon_trap_if_pattern for the predicate to be promoted 1056 ProjNode* create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry, 1057 Deoptimization::DeoptReason reason, 1058 int opcode); 1059 void register_control(Node* n, IdealLoopTree *loop, Node* pred); 1060 1061 // Clone loop predicates to cloned loops (peeled, unswitched) 1062 static ProjNode* clone_predicate(ProjNode* predicate_proj, Node* new_entry, 1063 Deoptimization::DeoptReason reason, 1064 PhaseIdealLoop* loop_phase, 1065 PhaseIterGVN* igvn); 1066 1067 static Node* clone_loop_predicates(Node* old_entry, Node* new_entry, 1068 bool clone_limit_check, 1069 PhaseIdealLoop* loop_phase, 1070 PhaseIterGVN* igvn); 1071 Node* clone_loop_predicates(Node* old_entry, Node* new_entry, bool clone_limit_check); 1072 1073 static Node* skip_all_loop_predicates(Node* entry); 1074 static Node* skip_loop_predicates(Node* entry); 1075 1076 // Find a good location to insert a predicate 1077 static ProjNode* find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason); 1078 // Find a predicate 1079 static Node* find_predicate(Node* entry); 1080 // Construct a range check for a predicate if 1081 BoolNode* rc_predicate(IdealLoopTree *loop, Node* ctrl, 1082 int scale, Node* offset, 1083 Node* init, Node* limit, jint stride, 1084 Node* range, bool upper, bool &overflow); 1085 1086 // Implementation of the loop predication to promote checks outside the loop 1087 bool loop_predication_impl(IdealLoopTree *loop); 1088 bool loop_predication_impl_helper(IdealLoopTree *loop, ProjNode* proj, ProjNode *predicate_proj, 1089 CountedLoopNode *cl, ConNode* zero, Invariance& invar, 1090 Deoptimization::DeoptReason reason); 1091 bool loop_predication_should_follow_branches(IdealLoopTree *loop, ProjNode *predicate_proj, float& loop_trip_cnt); 1092 void loop_predication_follow_branches(Node *c, IdealLoopTree *loop, float loop_trip_cnt, 1093 PathFrequency& pf, Node_Stack& stack, VectorSet& seen, 1094 Node_List& if_proj_list); 1095 ProjNode* insert_skeleton_predicate(IfNode* iff, IdealLoopTree *loop, 1096 ProjNode* proj, ProjNode *predicate_proj, 1097 ProjNode* upper_bound_proj, 1098 int scale, Node* offset, 1099 Node* init, Node* limit, jint stride, 1100 Node* rng, bool& overflow, 1101 Deoptimization::DeoptReason reason); 1102 Node* add_range_check_predicate(IdealLoopTree* loop, CountedLoopNode* cl, 1103 Node* predicate_proj, int scale_con, Node* offset, 1104 Node* limit, jint stride_con); 1105 1106 // Helper function to collect predicate for eliminating the useless ones 1107 void collect_potentially_useful_predicates(IdealLoopTree *loop, Unique_Node_List &predicate_opaque1); 1108 void eliminate_useless_predicates(); 1109 1110 // Change the control input of expensive nodes to allow commoning by 1111 // IGVN when it is guaranteed to not result in a more frequent 1112 // execution of the expensive node. Return true if progress. 1113 bool process_expensive_nodes(); 1114 1115 // Check whether node has become unreachable 1116 bool is_node_unreachable(Node *n) const { 1117 return !has_node(n) || n->is_unreachable(_igvn); 1118 } 1119 1120 // Eliminate range-checks and other trip-counter vs loop-invariant tests. 1121 int do_range_check( IdealLoopTree *loop, Node_List &old_new ); 1122 1123 // Check to see if do_range_check(...) cleaned the main loop of range-checks 1124 void has_range_checks(IdealLoopTree *loop); 1125 1126 // Process post loops which have range checks and try to build a multi-version 1127 // guard to safely determine if we can execute the post loop which was RCE'd. 1128 bool multi_version_post_loops(IdealLoopTree *rce_loop, IdealLoopTree *legacy_loop); 1129 1130 // Cause the rce'd post loop to optimized away, this happens if we cannot complete multiverioning 1131 void poison_rce_post_loop(IdealLoopTree *rce_loop); 1132 1133 // Create a slow version of the loop by cloning the loop 1134 // and inserting an if to select fast-slow versions. 1135 ProjNode* create_slow_version_of_loop(IdealLoopTree *loop, 1136 Node_List &old_new, 1137 int opcode, 1138 CloneLoopMode mode); 1139 1140 // Clone a loop and return the clone head (clone_loop_head). 1141 // Added nodes include int(1), int(0) - disconnected, If, IfTrue, IfFalse, 1142 // This routine was created for usage in CountedLoopReserveKit. 1143 // 1144 // int(1) -> If -> IfTrue -> original_loop_head 1145 // | 1146 // V 1147 // IfFalse -> clone_loop_head (returned by function pointer) 1148 // 1149 LoopNode* create_reserve_version_of_loop(IdealLoopTree *loop, CountedLoopReserveKit* lk); 1150 // Clone loop with an invariant test (that does not exit) and 1151 // insert a clone of the test that selects which version to 1152 // execute. 1153 void do_unswitching (IdealLoopTree *loop, Node_List &old_new); 1154 1155 // Find candidate "if" for unswitching 1156 IfNode* find_unswitching_candidate(const IdealLoopTree *loop) const; 1157 1158 // Range Check Elimination uses this function! 1159 // Constrain the main loop iterations so the affine function: 1160 // low_limit <= scale_con * I + offset < upper_limit 1161 // always holds true. That is, either increase the number of iterations in 1162 // the pre-loop or the post-loop until the condition holds true in the main 1163 // loop. Scale_con, offset and limit are all loop invariant. 1164 void add_constraint( int stride_con, int scale_con, Node *offset, Node *low_limit, Node *upper_limit, Node *pre_ctrl, Node **pre_limit, Node **main_limit ); 1165 // Helper function for add_constraint(). 1166 Node* adjust_limit( int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl ); 1167 1168 // Partially peel loop up through last_peel node. 1169 bool partial_peel( IdealLoopTree *loop, Node_List &old_new ); 1170 1171 // Create a scheduled list of nodes control dependent on ctrl set. 1172 void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched ); 1173 // Has a use in the vector set 1174 bool has_use_in_set( Node* n, VectorSet& vset ); 1175 // Has use internal to the vector set (ie. not in a phi at the loop head) 1176 bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ); 1177 // clone "n" for uses that are outside of loop 1178 int clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ); 1179 // clone "n" for special uses that are in the not_peeled region 1180 void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 1181 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ); 1182 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 1183 void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ); 1184 #ifdef ASSERT 1185 // Validate the loop partition sets: peel and not_peel 1186 bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel ); 1187 // Ensure that uses outside of loop are of the right form 1188 bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 1189 uint orig_exit_idx, uint clone_exit_idx); 1190 bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx); 1191 #endif 1192 1193 // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.) 1194 int stride_of_possible_iv( Node* iff ); 1195 bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; } 1196 // Return the (unique) control output node that's in the loop (if it exists.) 1197 Node* stay_in_loop( Node* n, IdealLoopTree *loop); 1198 // Insert a signed compare loop exit cloned from an unsigned compare. 1199 IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop); 1200 void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop); 1201 // Utility to register node "n" with PhaseIdealLoop 1202 void register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth); 1203 // Utility to create an if-projection 1204 ProjNode* proj_clone(ProjNode* p, IfNode* iff); 1205 // Force the iff control output to be the live_proj 1206 Node* short_circuit_if(IfNode* iff, ProjNode* live_proj); 1207 // Insert a region before an if projection 1208 RegionNode* insert_region_before_proj(ProjNode* proj); 1209 // Insert a new if before an if projection 1210 ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj); 1211 1212 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1213 // "Nearly" because all Nodes have been cloned from the original in the loop, 1214 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1215 // through the Phi recursively, and return a Bool. 1216 Node *clone_iff( PhiNode *phi, IdealLoopTree *loop ); 1217 CmpNode *clone_bool( PhiNode *phi, IdealLoopTree *loop ); 1218 1219 1220 // Rework addressing expressions to get the most loop-invariant stuff 1221 // moved out. We'd like to do all associative operators, but it's especially 1222 // important (common) to do address expressions. 1223 Node *remix_address_expressions( Node *n ); 1224 1225 // Attempt to use a conditional move instead of a phi/branch 1226 Node *conditional_move( Node *n ); 1227 1228 // Reorganize offset computations to lower register pressure. 1229 // Mostly prevent loop-fallout uses of the pre-incremented trip counter 1230 // (which are then alive with the post-incremented trip counter 1231 // forcing an extra register move) 1232 void reorg_offsets( IdealLoopTree *loop ); 1233 1234 // Check for aggressive application of 'split-if' optimization, 1235 // using basic block level info. 1236 void split_if_with_blocks ( VectorSet &visited, Node_Stack &nstack ); 1237 Node *split_if_with_blocks_pre ( Node *n ); 1238 void split_if_with_blocks_post( Node *n ); 1239 Node *has_local_phi_input( Node *n ); 1240 // Mark an IfNode as being dominated by a prior test, 1241 // without actually altering the CFG (and hence IDOM info). 1242 void dominated_by( Node *prevdom, Node *iff, bool flip = false, bool exclude_loop_predicate = false ); 1243 1244 // Split Node 'n' through merge point 1245 Node *split_thru_region( Node *n, Node *region ); 1246 // Split Node 'n' through merge point if there is enough win. 1247 Node *split_thru_phi( Node *n, Node *region, int policy ); 1248 // Found an If getting its condition-code input from a Phi in the 1249 // same block. Split thru the Region. 1250 void do_split_if( Node *iff ); 1251 1252 // Conversion of fill/copy patterns into intrisic versions 1253 bool do_intrinsify_fill(); 1254 bool intrinsify_fill(IdealLoopTree* lpt); 1255 bool match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value, 1256 Node*& shift, Node*& offset); 1257 1258 private: 1259 // Return a type based on condition control flow 1260 const TypeInt* filtered_type( Node *n, Node* n_ctrl); 1261 const TypeInt* filtered_type( Node *n ) { return filtered_type(n, NULL); } 1262 // Helpers for filtered type 1263 const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl); 1264 1265 // Helper functions 1266 Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache ); 1267 Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true ); 1268 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 ); 1269 bool split_up( Node *n, Node *blk1, Node *blk2 ); 1270 void sink_use( Node *use, Node *post_loop ); 1271 Node *place_near_use( Node *useblock ) const; 1272 Node* try_move_store_before_loop(Node* n, Node *n_ctrl); 1273 void try_move_store_after_loop(Node* n); 1274 bool identical_backtoback_ifs(Node *n); 1275 bool can_split_if(Node *n_ctrl); 1276 1277 bool _created_loop_node; 1278 public: 1279 void set_created_loop_node() { _created_loop_node = true; } 1280 bool created_loop_node() { return _created_loop_node; } 1281 void register_new_node( Node *n, Node *blk ); 1282 1283 #ifdef ASSERT 1284 void dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA); 1285 #endif 1286 1287 #ifndef PRODUCT 1288 void dump( ) const; 1289 void dump( IdealLoopTree *loop, uint rpo_idx, Node_List &rpo_list ) const; 1290 void rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const; 1291 void verify() const; // Major slow :-) 1292 void verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const; 1293 IdealLoopTree *get_loop_idx(Node* n) const { 1294 // Dead nodes have no loop, so return the top level loop instead 1295 return _nodes[n->_idx] ? (IdealLoopTree*)_nodes[n->_idx] : _ltree_root; 1296 } 1297 // Print some stats 1298 static void print_statistics(); 1299 static int _loop_invokes; // Count of PhaseIdealLoop invokes 1300 static int _loop_work; // Sum of PhaseIdealLoop x _unique 1301 #endif 1302 }; 1303 1304 // This kit may be used for making of a reserved copy of a loop before this loop 1305 // goes under non-reversible changes. 1306 // 1307 // Function create_reserve() creates a reserved copy (clone) of the loop. 1308 // The reserved copy is created by calling 1309 // PhaseIdealLoop::create_reserve_version_of_loop - see there how 1310 // the original and reserved loops are connected in the outer graph. 1311 // If create_reserve succeeded, it returns 'true' and _has_reserved is set to 'true'. 1312 // 1313 // By default the reserved copy (clone) of the loop is created as dead code - it is 1314 // dominated in the outer loop by this node chain: 1315 // intcon(1)->If->IfFalse->reserved_copy. 1316 // The original loop is dominated by the the same node chain but IfTrue projection: 1317 // intcon(0)->If->IfTrue->original_loop. 1318 // 1319 // In this implementation of CountedLoopReserveKit the ctor includes create_reserve() 1320 // and the dtor, checks _use_new value. 1321 // If _use_new == false, it "switches" control to reserved copy of the loop 1322 // by simple replacing of node intcon(1) with node intcon(0). 1323 // 1324 // Here is a proposed example of usage (see also SuperWord::output in superword.cpp). 1325 // 1326 // void CountedLoopReserveKit_example() 1327 // { 1328 // CountedLoopReserveKit lrk((phase, lpt, DoReserveCopy = true); // create local object 1329 // if (DoReserveCopy && !lrk.has_reserved()) { 1330 // return; //failed to create reserved loop copy 1331 // } 1332 // ... 1333 // //something is wrong, switch to original loop 1334 /// if(something_is_wrong) return; // ~CountedLoopReserveKit makes the switch 1335 // ... 1336 // //everything worked ok, return with the newly modified loop 1337 // lrk.use_new(); 1338 // return; // ~CountedLoopReserveKit does nothing once use_new() was called 1339 // } 1340 // 1341 // Keep in mind, that by default if create_reserve() is not followed by use_new() 1342 // the dtor will "switch to the original" loop. 1343 // NOTE. You you modify outside of the original loop this class is no help. 1344 // 1345 class CountedLoopReserveKit { 1346 private: 1347 PhaseIdealLoop* _phase; 1348 IdealLoopTree* _lpt; 1349 LoopNode* _lp; 1350 IfNode* _iff; 1351 LoopNode* _lp_reserved; 1352 bool _has_reserved; 1353 bool _use_new; 1354 const bool _active; //may be set to false in ctor, then the object is dummy 1355 1356 public: 1357 CountedLoopReserveKit(PhaseIdealLoop* phase, IdealLoopTree *loop, bool active); 1358 ~CountedLoopReserveKit(); 1359 void use_new() {_use_new = true;} 1360 void set_iff(IfNode* x) {_iff = x;} 1361 bool has_reserved() const { return _active && _has_reserved;} 1362 private: 1363 bool create_reserve(); 1364 };// class CountedLoopReserveKit 1365 1366 inline Node* IdealLoopTree::tail() { 1367 // Handle lazy update of _tail field 1368 Node *n = _tail; 1369 //while( !n->in(0) ) // Skip dead CFG nodes 1370 //n = n->in(1); 1371 if (n->in(0) == NULL) 1372 n = _phase->get_ctrl(n); 1373 _tail = n; 1374 return n; 1375 } 1376 1377 1378 // Iterate over the loop tree using a preorder, left-to-right traversal. 1379 // 1380 // Example that visits all counted loops from within PhaseIdealLoop 1381 // 1382 // for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { 1383 // IdealLoopTree* lpt = iter.current(); 1384 // if (!lpt->is_counted()) continue; 1385 // ... 1386 class LoopTreeIterator : public StackObj { 1387 private: 1388 IdealLoopTree* _root; 1389 IdealLoopTree* _curnt; 1390 1391 public: 1392 LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {} 1393 1394 bool done() { return _curnt == NULL; } // Finished iterating? 1395 1396 void next(); // Advance to next loop tree 1397 1398 IdealLoopTree* current() { return _curnt; } // Return current value of iterator. 1399 }; 1400 1401 #endif // SHARE_VM_OPTO_LOOPNODE_HPP