1 /* 2 * Copyright (c) 1998, 2023, 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_OPTO_LOOPNODE_HPP 26 #define SHARE_OPTO_LOOPNODE_HPP 27 28 #include "opto/cfgnode.hpp" 29 #include "opto/multnode.hpp" 30 #include "opto/phaseX.hpp" 31 #include "opto/subnode.hpp" 32 #include "opto/type.hpp" 33 34 class CmpNode; 35 class BaseCountedLoopEndNode; 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 = 1<<2, 65 HasExactTripCount = 1<<3, 66 InnerLoop = 1<<4, 67 PartialPeelLoop = 1<<5, 68 PartialPeelFailed = 1<<6, 69 HasReductions = 1<<7, 70 WasSlpAnalyzed = 1<<8, 71 PassedSlpAnalysis = 1<<9, 72 DoUnrollOnly = 1<<10, 73 VectorizedLoop = 1<<11, 74 HasAtomicPostLoop = 1<<12, 75 HasRangeChecks = 1<<13, 76 IsMultiversioned = 1<<14, 77 StripMined = 1<<15, 78 SubwordLoop = 1<<16, 79 ProfileTripFailed = 1<<17, 80 LoopNestInnerLoop = 1<< 18, 81 LoopNestLongOuterLoop = 1<< 19, 82 FlattenedArrays = 1<<20}; 83 84 char _unswitch_count; 85 enum { _unswitch_max=3 }; 86 char _postloop_flags; 87 enum { LoopNotRCEChecked = 0, LoopRCEChecked = 1, RCEPostLoop = 2 }; 88 89 // Expected trip count from profile data 90 float _profile_trip_cnt; 91 92 public: 93 // Names for edge indices 94 enum { Self=0, EntryControl, LoopBackControl }; 95 96 bool is_inner_loop() const { return _loop_flags & InnerLoop; } 97 void set_inner_loop() { _loop_flags |= InnerLoop; } 98 99 bool range_checks_present() const { return _loop_flags & HasRangeChecks; } 100 bool is_multiversioned() const { return _loop_flags & IsMultiversioned; } 101 bool is_vectorized_loop() const { return _loop_flags & VectorizedLoop; } 102 bool is_partial_peel_loop() const { return _loop_flags & PartialPeelLoop; } 103 void set_partial_peel_loop() { _loop_flags |= PartialPeelLoop; } 104 bool partial_peel_has_failed() const { return _loop_flags & PartialPeelFailed; } 105 bool is_strip_mined() const { return _loop_flags & StripMined; } 106 bool is_profile_trip_failed() const { return _loop_flags & ProfileTripFailed; } 107 bool is_subword_loop() const { return _loop_flags & SubwordLoop; } 108 bool is_loop_nest_inner_loop() const { return _loop_flags & LoopNestInnerLoop; } 109 bool is_loop_nest_outer_loop() const { return _loop_flags & LoopNestLongOuterLoop; } 110 bool is_flattened_arrays() const { return _loop_flags & FlattenedArrays; } 111 112 void mark_partial_peel_failed() { _loop_flags |= PartialPeelFailed; } 113 void mark_has_reductions() { _loop_flags |= HasReductions; } 114 void mark_was_slp() { _loop_flags |= WasSlpAnalyzed; } 115 void mark_passed_slp() { _loop_flags |= PassedSlpAnalysis; } 116 void mark_do_unroll_only() { _loop_flags |= DoUnrollOnly; } 117 void mark_loop_vectorized() { _loop_flags |= VectorizedLoop; } 118 void mark_has_atomic_post_loop() { _loop_flags |= HasAtomicPostLoop; } 119 void mark_has_range_checks() { _loop_flags |= HasRangeChecks; } 120 void clear_has_range_checks() { _loop_flags &= ~HasRangeChecks; } 121 void mark_is_multiversioned() { _loop_flags |= IsMultiversioned; } 122 void mark_strip_mined() { _loop_flags |= StripMined; } 123 void clear_strip_mined() { _loop_flags &= ~StripMined; } 124 void mark_profile_trip_failed() { _loop_flags |= ProfileTripFailed; } 125 void mark_subword_loop() { _loop_flags |= SubwordLoop; } 126 void mark_loop_nest_inner_loop() { _loop_flags |= LoopNestInnerLoop; } 127 void mark_loop_nest_outer_loop() { _loop_flags |= LoopNestLongOuterLoop; } 128 void mark_flattened_arrays() { _loop_flags |= FlattenedArrays; } 129 130 int unswitch_max() { return _unswitch_max; } 131 int unswitch_count() { return _unswitch_count; } 132 133 int has_been_range_checked() const { return _postloop_flags & LoopRCEChecked; } 134 void set_has_been_range_checked() { _postloop_flags |= LoopRCEChecked; } 135 int is_rce_post_loop() const { return _postloop_flags & RCEPostLoop; } 136 void set_is_rce_post_loop() { _postloop_flags |= RCEPostLoop; } 137 138 void set_unswitch_count(int val) { 139 assert (val <= unswitch_max(), "too many unswitches"); 140 _unswitch_count = val; 141 } 142 143 void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; } 144 float profile_trip_cnt() { return _profile_trip_cnt; } 145 146 LoopNode(Node *entry, Node *backedge) 147 : RegionNode(3), _loop_flags(0), _unswitch_count(0), 148 _postloop_flags(0), _profile_trip_cnt(COUNT_UNKNOWN) { 149 init_class_id(Class_Loop); 150 init_req(EntryControl, entry); 151 init_req(LoopBackControl, backedge); 152 } 153 154 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); 155 virtual int Opcode() const; 156 bool can_be_counted_loop(PhaseTransform* phase) const { 157 return req() == 3 && in(0) != NULL && 158 in(1) != NULL && phase->type(in(1)) != Type::TOP && 159 in(2) != NULL && phase->type(in(2)) != Type::TOP; 160 } 161 bool is_valid_counted_loop(BasicType bt) const; 162 #ifndef PRODUCT 163 virtual void dump_spec(outputStream *st) const; 164 #endif 165 166 void verify_strip_mined(int expect_skeleton) const NOT_DEBUG_RETURN; 167 virtual LoopNode* skip_strip_mined(int expect_skeleton = 1) { return this; } 168 virtual IfTrueNode* outer_loop_tail() const { ShouldNotReachHere(); return NULL; } 169 virtual OuterStripMinedLoopEndNode* outer_loop_end() const { ShouldNotReachHere(); return NULL; } 170 virtual IfFalseNode* outer_loop_exit() const { ShouldNotReachHere(); return NULL; } 171 virtual SafePointNode* outer_safepoint() const { ShouldNotReachHere(); return NULL; } 172 }; 173 174 //------------------------------Counted Loops---------------------------------- 175 // Counted loops are all trip-counted loops, with exactly 1 trip-counter exit 176 // path (and maybe some other exit paths). The trip-counter exit is always 177 // last in the loop. The trip-counter have to stride by a constant; 178 // the exit value is also loop invariant. 179 180 // CountedLoopNodes and CountedLoopEndNodes come in matched pairs. The 181 // CountedLoopNode has the incoming loop control and the loop-back-control 182 // which is always the IfTrue before the matching CountedLoopEndNode. The 183 // CountedLoopEndNode has an incoming control (possibly not the 184 // CountedLoopNode if there is control flow in the loop), the post-increment 185 // trip-counter value, and the limit. The trip-counter value is always of 186 // the form (Op old-trip-counter stride). The old-trip-counter is produced 187 // by a Phi connected to the CountedLoopNode. The stride is constant. 188 // The Op is any commutable opcode, including Add, Mul, Xor. The 189 // CountedLoopEndNode also takes in the loop-invariant limit value. 190 191 // From a CountedLoopNode I can reach the matching CountedLoopEndNode via the 192 // loop-back control. From CountedLoopEndNodes I can reach CountedLoopNodes 193 // via the old-trip-counter from the Op node. 194 195 //------------------------------CountedLoopNode-------------------------------- 196 // CountedLoopNodes head simple counted loops. CountedLoopNodes have as 197 // inputs the incoming loop-start control and the loop-back control, so they 198 // act like RegionNodes. They also take in the initial trip counter, the 199 // loop-invariant stride and the loop-invariant limit value. CountedLoopNodes 200 // produce a loop-body control and the trip counter value. Since 201 // CountedLoopNodes behave like RegionNodes I still have a standard CFG model. 202 203 class BaseCountedLoopNode : public LoopNode { 204 public: 205 BaseCountedLoopNode(Node *entry, Node *backedge) 206 : LoopNode(entry, backedge) { 207 } 208 209 Node *init_control() const { return in(EntryControl); } 210 Node *back_control() const { return in(LoopBackControl); } 211 212 Node* init_trip() const; 213 Node* stride() const; 214 bool stride_is_con() const; 215 Node* limit() const; 216 Node* incr() const; 217 Node* phi() const; 218 219 BaseCountedLoopEndNode* loopexit_or_null() const; 220 BaseCountedLoopEndNode* loopexit() const; 221 222 virtual BasicType bt() const = 0; 223 224 jlong stride_con() const; 225 226 static BaseCountedLoopNode* make(Node* entry, Node* backedge, BasicType bt); 227 }; 228 229 230 class CountedLoopNode : public BaseCountedLoopNode { 231 // Size is bigger to hold _main_idx. However, _main_idx does not change 232 // the semantics so it does not appear in the hash & cmp functions. 233 virtual uint size_of() const { return sizeof(*this); } 234 235 // For Pre- and Post-loops during debugging ONLY, this holds the index of 236 // the Main CountedLoop. Used to assert that we understand the graph shape. 237 node_idx_t _main_idx; 238 239 // Known trip count calculated by compute_exact_trip_count() 240 uint _trip_count; 241 242 // Log2 of original loop bodies in unrolled loop 243 int _unrolled_count_log2; 244 245 // Node count prior to last unrolling - used to decide if 246 // unroll,optimize,unroll,optimize,... is making progress 247 int _node_count_before_unroll; 248 249 // If slp analysis is performed we record the maximum 250 // vector mapped unroll factor here 251 int _slp_maximum_unroll_factor; 252 253 // The eventual count of vectorizable packs in slp 254 int _slp_vector_pack_count; 255 256 public: 257 CountedLoopNode(Node *entry, Node *backedge) 258 : BaseCountedLoopNode(entry, backedge), _main_idx(0), _trip_count(max_juint), 259 _unrolled_count_log2(0), _node_count_before_unroll(0), 260 _slp_maximum_unroll_factor(0), _slp_vector_pack_count(0) { 261 init_class_id(Class_CountedLoop); 262 // Initialize _trip_count to the largest possible value. 263 // Will be reset (lower) if the loop's trip count is known. 264 } 265 266 virtual int Opcode() const; 267 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); 268 269 CountedLoopEndNode* loopexit_or_null() const { return (CountedLoopEndNode*) BaseCountedLoopNode::loopexit_or_null(); } 270 CountedLoopEndNode* loopexit() const { return (CountedLoopEndNode*) BaseCountedLoopNode::loopexit(); } 271 int stride_con() const; 272 273 // Match increment with optional truncation 274 static Node* 275 match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2, const TypeInteger** trunc_type, 276 BasicType bt); 277 278 // A 'main' loop has a pre-loop and a post-loop. The 'main' loop 279 // can run short a few iterations and may start a few iterations in. 280 // It will be RCE'd and unrolled and aligned. 281 282 // A following 'post' loop will run any remaining iterations. Used 283 // during Range Check Elimination, the 'post' loop will do any final 284 // iterations with full checks. Also used by Loop Unrolling, where 285 // the 'post' loop will do any epilog iterations needed. Basically, 286 // a 'post' loop can not profitably be further unrolled or RCE'd. 287 288 // A preceding 'pre' loop will run at least 1 iteration (to do peeling), 289 // it may do under-flow checks for RCE and may do alignment iterations 290 // so the following main loop 'knows' that it is striding down cache 291 // lines. 292 293 // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or 294 // Aligned, may be missing it's pre-loop. 295 bool is_normal_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Normal; } 296 bool is_pre_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Pre; } 297 bool is_main_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Main; } 298 bool is_post_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Post; } 299 bool is_reduction_loop() const { return (_loop_flags&HasReductions) == HasReductions; } 300 bool was_slp_analyzed () const { return (_loop_flags&WasSlpAnalyzed) == WasSlpAnalyzed; } 301 bool has_passed_slp () const { return (_loop_flags&PassedSlpAnalysis) == PassedSlpAnalysis; } 302 bool is_unroll_only () const { return (_loop_flags&DoUnrollOnly) == DoUnrollOnly; } 303 bool is_main_no_pre_loop() const { return _loop_flags & MainHasNoPreLoop; } 304 bool has_atomic_post_loop () const { return (_loop_flags & HasAtomicPostLoop) == HasAtomicPostLoop; } 305 void set_main_no_pre_loop() { _loop_flags |= MainHasNoPreLoop; } 306 307 int main_idx() const { return _main_idx; } 308 309 310 void set_pre_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; } 311 void set_main_loop ( ) { assert(is_normal_loop(),""); _loop_flags |= Main; } 312 void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; } 313 void set_normal_loop( ) { _loop_flags &= ~PreMainPostFlagsMask; } 314 315 void set_trip_count(uint tc) { _trip_count = tc; } 316 uint trip_count() { return _trip_count; } 317 318 bool has_exact_trip_count() const { return (_loop_flags & HasExactTripCount) != 0; } 319 void set_exact_trip_count(uint tc) { 320 _trip_count = tc; 321 _loop_flags |= HasExactTripCount; 322 } 323 void set_nonexact_trip_count() { 324 _loop_flags &= ~HasExactTripCount; 325 } 326 void set_notpassed_slp() { 327 _loop_flags &= ~PassedSlpAnalysis; 328 } 329 330 void double_unrolled_count() { _unrolled_count_log2++; } 331 int unrolled_count() { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); } 332 333 void set_node_count_before_unroll(int ct) { _node_count_before_unroll = ct; } 334 int node_count_before_unroll() { return _node_count_before_unroll; } 335 void set_slp_max_unroll(int unroll_factor) { _slp_maximum_unroll_factor = unroll_factor; } 336 int slp_max_unroll() const { return _slp_maximum_unroll_factor; } 337 void set_slp_pack_count(int pack_count) { _slp_vector_pack_count = pack_count; } 338 int slp_pack_count() const { return _slp_vector_pack_count; } 339 340 virtual LoopNode* skip_strip_mined(int expect_skeleton = 1); 341 OuterStripMinedLoopNode* outer_loop() const; 342 virtual IfTrueNode* outer_loop_tail() const; 343 virtual OuterStripMinedLoopEndNode* outer_loop_end() const; 344 virtual IfFalseNode* outer_loop_exit() const; 345 virtual SafePointNode* outer_safepoint() const; 346 347 // If this is a main loop in a pre/main/post loop nest, walk over 348 // the predicates that were inserted by 349 // duplicate_predicates()/add_range_check_predicate() 350 static Node* skip_predicates_from_entry(Node* ctrl); 351 Node* skip_predicates(); 352 353 virtual BasicType bt() const { 354 return T_INT; 355 } 356 357 Node* is_canonical_loop_entry(); 358 359 #ifndef PRODUCT 360 virtual void dump_spec(outputStream *st) const; 361 #endif 362 363 static bool is_zero_trip_guard_if(const IfNode* iff); 364 }; 365 366 class LongCountedLoopNode : public BaseCountedLoopNode { 367 public: 368 LongCountedLoopNode(Node *entry, Node *backedge) 369 : BaseCountedLoopNode(entry, backedge) { 370 init_class_id(Class_LongCountedLoop); 371 } 372 373 virtual int Opcode() const; 374 375 virtual BasicType bt() const { 376 return T_LONG; 377 } 378 379 LongCountedLoopEndNode* loopexit_or_null() const { return (LongCountedLoopEndNode*) BaseCountedLoopNode::loopexit_or_null(); } 380 LongCountedLoopEndNode* loopexit() const { return (LongCountedLoopEndNode*) BaseCountedLoopNode::loopexit(); } 381 }; 382 383 384 //------------------------------CountedLoopEndNode----------------------------- 385 // CountedLoopEndNodes end simple trip counted loops. They act much like 386 // IfNodes. 387 388 class BaseCountedLoopEndNode : public IfNode { 389 public: 390 enum { TestControl, TestValue }; 391 BaseCountedLoopEndNode(Node *control, Node *test, float prob, float cnt) 392 : IfNode(control, test, prob, cnt) { 393 init_class_id(Class_BaseCountedLoopEnd); 394 } 395 396 Node *cmp_node() const { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : NULL; } 397 Node* incr() const { Node* tmp = cmp_node(); return (tmp && tmp->req() == 3) ? tmp->in(1) : NULL; } 398 Node* limit() const { Node* tmp = cmp_node(); return (tmp && tmp->req() == 3) ? tmp->in(2) : NULL; } 399 Node* stride() const { Node* tmp = incr(); return (tmp && tmp->req() == 3) ? tmp->in(2) : NULL; } 400 Node* init_trip() const { Node* tmp = phi(); return (tmp && tmp->req() == 3) ? tmp->in(1) : NULL; } 401 bool stride_is_con() const { Node *tmp = stride(); return (tmp != NULL && tmp->is_Con()); } 402 403 PhiNode* phi() const { 404 Node* tmp = incr(); 405 if (tmp && tmp->req() == 3) { 406 Node* phi = tmp->in(1); 407 if (phi->is_Phi()) { 408 return phi->as_Phi(); 409 } 410 } 411 return NULL; 412 } 413 414 BaseCountedLoopNode* loopnode() const { 415 // The CountedLoopNode that goes with this CountedLoopEndNode may 416 // have been optimized out by the IGVN so be cautious with the 417 // pattern matching on the graph 418 PhiNode* iv_phi = phi(); 419 if (iv_phi == NULL) { 420 return NULL; 421 } 422 Node* ln = iv_phi->in(0); 423 if (!ln->is_BaseCountedLoop() || ln->as_BaseCountedLoop()->loopexit_or_null() != this) { 424 return NULL; 425 } 426 if (ln->as_BaseCountedLoop()->bt() != bt()) { 427 return NULL; 428 } 429 return ln->as_BaseCountedLoop(); 430 } 431 432 BoolTest::mask test_trip() const { return in(TestValue)->as_Bool()->_test._test; } 433 434 jlong stride_con() const; 435 virtual BasicType bt() const = 0; 436 437 static BaseCountedLoopEndNode* make(Node* control, Node* test, float prob, float cnt, BasicType bt); 438 }; 439 440 class CountedLoopEndNode : public BaseCountedLoopEndNode { 441 public: 442 443 CountedLoopEndNode(Node *control, Node *test, float prob, float cnt) 444 : BaseCountedLoopEndNode(control, test, prob, cnt) { 445 init_class_id(Class_CountedLoopEnd); 446 } 447 virtual int Opcode() const; 448 449 CountedLoopNode* loopnode() const { 450 return (CountedLoopNode*) BaseCountedLoopEndNode::loopnode(); 451 } 452 453 virtual BasicType bt() const { 454 return T_INT; 455 } 456 457 #ifndef PRODUCT 458 virtual void dump_spec(outputStream *st) const; 459 #endif 460 }; 461 462 class LongCountedLoopEndNode : public BaseCountedLoopEndNode { 463 public: 464 LongCountedLoopEndNode(Node *control, Node *test, float prob, float cnt) 465 : BaseCountedLoopEndNode(control, test, prob, cnt) { 466 init_class_id(Class_LongCountedLoopEnd); 467 } 468 469 LongCountedLoopNode* loopnode() const { 470 return (LongCountedLoopNode*) BaseCountedLoopEndNode::loopnode(); 471 } 472 473 virtual int Opcode() const; 474 475 virtual BasicType bt() const { 476 return T_LONG; 477 } 478 }; 479 480 481 inline BaseCountedLoopEndNode* BaseCountedLoopNode::loopexit_or_null() const { 482 Node* bctrl = back_control(); 483 if (bctrl == NULL) return NULL; 484 485 Node* lexit = bctrl->in(0); 486 if (!lexit->is_BaseCountedLoopEnd()) { 487 return NULL; 488 } 489 BaseCountedLoopEndNode* result = lexit->as_BaseCountedLoopEnd(); 490 if (result->bt() != bt()) { 491 return NULL; 492 } 493 return result; 494 } 495 496 inline BaseCountedLoopEndNode* BaseCountedLoopNode::loopexit() const { 497 BaseCountedLoopEndNode* cle = loopexit_or_null(); 498 assert(cle != NULL, "loopexit is NULL"); 499 return cle; 500 } 501 502 inline Node* BaseCountedLoopNode::init_trip() const { 503 BaseCountedLoopEndNode* cle = loopexit_or_null(); 504 return cle != NULL ? cle->init_trip() : NULL; 505 } 506 inline Node* BaseCountedLoopNode::stride() const { 507 BaseCountedLoopEndNode* cle = loopexit_or_null(); 508 return cle != NULL ? cle->stride() : NULL; 509 } 510 511 inline bool BaseCountedLoopNode::stride_is_con() const { 512 BaseCountedLoopEndNode* cle = loopexit_or_null(); 513 return cle != NULL && cle->stride_is_con(); 514 } 515 inline Node* BaseCountedLoopNode::limit() const { 516 BaseCountedLoopEndNode* cle = loopexit_or_null(); 517 return cle != NULL ? cle->limit() : NULL; 518 } 519 inline Node* BaseCountedLoopNode::incr() const { 520 BaseCountedLoopEndNode* cle = loopexit_or_null(); 521 return cle != NULL ? cle->incr() : NULL; 522 } 523 inline Node* BaseCountedLoopNode::phi() const { 524 BaseCountedLoopEndNode* cle = loopexit_or_null(); 525 return cle != NULL ? cle->phi() : NULL; 526 } 527 528 inline jlong BaseCountedLoopNode::stride_con() const { 529 BaseCountedLoopEndNode* cle = loopexit_or_null(); 530 return cle != NULL ? cle->stride_con() : 0; 531 } 532 533 534 //------------------------------LoopLimitNode----------------------------- 535 // Counted Loop limit node which represents exact final iterator value: 536 // trip_count = (limit - init_trip + stride - 1)/stride 537 // final_value= trip_count * stride + init_trip. 538 // Use HW instructions to calculate it when it can overflow in integer. 539 // Note, final_value should fit into integer since counted loop has 540 // limit check: limit <= max_int-stride. 541 class LoopLimitNode : public Node { 542 enum { Init=1, Limit=2, Stride=3 }; 543 public: 544 LoopLimitNode( Compile* C, Node *init, Node *limit, Node *stride ) : Node(0,init,limit,stride) { 545 // Put it on the Macro nodes list to optimize during macro nodes expansion. 546 init_flags(Flag_is_macro); 547 C->add_macro_node(this); 548 } 549 virtual int Opcode() const; 550 virtual const Type *bottom_type() const { return TypeInt::INT; } 551 virtual uint ideal_reg() const { return Op_RegI; } 552 virtual const Type* Value(PhaseGVN* phase) const; 553 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); 554 virtual Node* Identity(PhaseGVN* phase); 555 }; 556 557 // Support for strip mining 558 class OuterStripMinedLoopNode : public LoopNode { 559 private: 560 static void fix_sunk_stores(CountedLoopEndNode* inner_cle, LoopNode* inner_cl, PhaseIterGVN* igvn, PhaseIdealLoop* iloop); 561 562 public: 563 OuterStripMinedLoopNode(Compile* C, Node *entry, Node *backedge) 564 : LoopNode(entry, backedge) { 565 init_class_id(Class_OuterStripMinedLoop); 566 init_flags(Flag_is_macro); 567 C->add_macro_node(this); 568 } 569 570 virtual int Opcode() const; 571 572 virtual IfTrueNode* outer_loop_tail() const; 573 virtual OuterStripMinedLoopEndNode* outer_loop_end() const; 574 virtual IfFalseNode* outer_loop_exit() const; 575 virtual SafePointNode* outer_safepoint() const; 576 void adjust_strip_mined_loop(PhaseIterGVN* igvn); 577 578 void remove_outer_loop_and_safepoint(PhaseIterGVN* igvn) const; 579 580 void transform_to_counted_loop(PhaseIterGVN* igvn, PhaseIdealLoop* iloop); 581 582 static Node* register_new_node(Node* node, LoopNode* ctrl, PhaseIterGVN* igvn, PhaseIdealLoop* iloop); 583 584 Node* register_control(Node* node, Node* loop, Node* idom, PhaseIterGVN* igvn, 585 PhaseIdealLoop* iloop); 586 }; 587 588 class OuterStripMinedLoopEndNode : public IfNode { 589 public: 590 OuterStripMinedLoopEndNode(Node *control, Node *test, float prob, float cnt) 591 : IfNode(control, test, prob, cnt) { 592 init_class_id(Class_OuterStripMinedLoopEnd); 593 } 594 595 virtual int Opcode() const; 596 597 virtual const Type* Value(PhaseGVN* phase) const; 598 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); 599 600 bool is_expanded(PhaseGVN *phase) const; 601 }; 602 603 // -----------------------------IdealLoopTree---------------------------------- 604 class IdealLoopTree : public ResourceObj { 605 public: 606 IdealLoopTree *_parent; // Parent in loop tree 607 IdealLoopTree *_next; // Next sibling in loop tree 608 IdealLoopTree *_child; // First child in loop tree 609 610 // The head-tail backedge defines the loop. 611 // If a loop has multiple backedges, this is addressed during cleanup where 612 // we peel off the multiple backedges, merging all edges at the bottom and 613 // ensuring that one proper backedge flow into the loop. 614 Node *_head; // Head of loop 615 Node *_tail; // Tail of loop 616 inline Node *tail(); // Handle lazy update of _tail field 617 inline Node *head(); // Handle lazy update of _head field 618 PhaseIdealLoop* _phase; 619 int _local_loop_unroll_limit; 620 int _local_loop_unroll_factor; 621 622 Node_List _body; // Loop body for inner loops 623 624 uint16_t _nest; // Nesting depth 625 uint8_t _irreducible:1, // True if irreducible 626 _has_call:1, // True if has call safepoint 627 _has_sfpt:1, // True if has non-call safepoint 628 _rce_candidate:1; // True if candidate for range check elimination 629 630 Node_List* _safepts; // List of safepoints in this loop 631 Node_List* _required_safept; // A inner loop cannot delete these safepts; 632 bool _allow_optimizations; // Allow loop optimizations 633 634 IdealLoopTree( PhaseIdealLoop* phase, Node *head, Node *tail ) 635 : _parent(0), _next(0), _child(0), 636 _head(head), _tail(tail), 637 _phase(phase), 638 _local_loop_unroll_limit(0), _local_loop_unroll_factor(0), 639 _nest(0), _irreducible(0), _has_call(0), _has_sfpt(0), _rce_candidate(0), 640 _safepts(NULL), 641 _required_safept(NULL), 642 _allow_optimizations(true) 643 { 644 precond(_head != NULL); 645 precond(_tail != NULL); 646 } 647 648 // Is 'l' a member of 'this'? 649 bool is_member(const IdealLoopTree *l) const; // Test for nested membership 650 651 // Set loop nesting depth. Accumulate has_call bits. 652 int set_nest( uint depth ); 653 654 // Split out multiple fall-in edges from the loop header. Move them to a 655 // private RegionNode before the loop. This becomes the loop landing pad. 656 void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ); 657 658 // Split out the outermost loop from this shared header. 659 void split_outer_loop( PhaseIdealLoop *phase ); 660 661 // Merge all the backedges from the shared header into a private Region. 662 // Feed that region as the one backedge to this loop. 663 void merge_many_backedges( PhaseIdealLoop *phase ); 664 665 // Split shared headers and insert loop landing pads. 666 // Insert a LoopNode to replace the RegionNode. 667 // Returns TRUE if loop tree is structurally changed. 668 bool beautify_loops( PhaseIdealLoop *phase ); 669 670 // Perform optimization to use the loop predicates for null checks and range checks. 671 // Applies to any loop level (not just the innermost one) 672 bool loop_predication( PhaseIdealLoop *phase); 673 674 // Perform iteration-splitting on inner loops. Split iterations to 675 // avoid range checks or one-shot null checks. Returns false if the 676 // current round of loop opts should stop. 677 bool iteration_split( PhaseIdealLoop *phase, Node_List &old_new ); 678 679 // Driver for various flavors of iteration splitting. Returns false 680 // if the current round of loop opts should stop. 681 bool iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ); 682 683 // Given dominators, try to find loops with calls that must always be 684 // executed (call dominates loop tail). These loops do not need non-call 685 // safepoints (ncsfpt). 686 void check_safepts(VectorSet &visited, Node_List &stack); 687 688 // Allpaths backwards scan from loop tail, terminating each path at first safepoint 689 // encountered. 690 void allpaths_check_safepts(VectorSet &visited, Node_List &stack); 691 692 // Remove safepoints from loop. Optionally keeping one. 693 void remove_safepoints(PhaseIdealLoop* phase, bool keep_one); 694 695 // Convert to counted loops where possible 696 void counted_loop( PhaseIdealLoop *phase ); 697 698 // Check for Node being a loop-breaking test 699 Node *is_loop_exit(Node *iff) const; 700 701 // Remove simplistic dead code from loop body 702 void DCE_loop_body(); 703 704 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage. 705 // Replace with a 1-in-10 exit guess. 706 void adjust_loop_exit_prob( PhaseIdealLoop *phase ); 707 708 // Return TRUE or FALSE if the loop should never be RCE'd or aligned. 709 // Useful for unrolling loops with NO array accesses. 710 bool policy_peel_only( PhaseIdealLoop *phase ) const; 711 712 // Return TRUE or FALSE if the loop should be unswitched -- clone 713 // loop with an invariant test 714 bool policy_unswitching( PhaseIdealLoop *phase ) const; 715 716 // Micro-benchmark spamming. Remove empty loops. 717 bool do_remove_empty_loop( PhaseIdealLoop *phase ); 718 719 // Convert one iteration loop into normal code. 720 bool do_one_iteration_loop( PhaseIdealLoop *phase ); 721 722 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can 723 // move some loop-invariant test (usually a null-check) before the loop. 724 bool policy_peeling(PhaseIdealLoop *phase); 725 726 uint estimate_peeling(PhaseIdealLoop *phase); 727 728 // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any 729 // known trip count in the counted loop node. 730 bool policy_maximally_unroll(PhaseIdealLoop *phase) const; 731 732 // Return TRUE or FALSE if the loop should be unrolled or not. Apply unroll 733 // if the loop is a counted loop and the loop body is small enough. 734 bool policy_unroll(PhaseIdealLoop *phase); 735 736 // Loop analyses to map to a maximal superword unrolling for vectorization. 737 void policy_unroll_slp_analysis(CountedLoopNode *cl, PhaseIdealLoop *phase, int future_unroll_ct); 738 739 // Return TRUE or FALSE if the loop should be range-check-eliminated. 740 // Gather a list of IF tests that are dominated by iteration splitting; 741 // also gather the end of the first split and the start of the 2nd split. 742 bool policy_range_check(PhaseIdealLoop* phase, bool provisional, BasicType bt) const; 743 744 // Return TRUE if "iff" is a range check. 745 bool is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar DEBUG_ONLY(COMMA ProjNode *predicate_proj)) const; 746 bool is_range_check_if(IfNode* iff, PhaseIdealLoop* phase, BasicType bt, Node* iv, Node*& range, Node*& offset, 747 jlong& scale) const; 748 749 // Estimate the number of nodes required when cloning a loop (body). 750 uint est_loop_clone_sz(uint factor) const; 751 // Estimate the number of nodes required when unrolling a loop (body). 752 uint est_loop_unroll_sz(uint factor) const; 753 754 // Compute loop trip count if possible 755 void compute_trip_count(PhaseIdealLoop* phase); 756 757 // Compute loop trip count from profile data 758 float compute_profile_trip_cnt_helper(Node* n); 759 void compute_profile_trip_cnt( PhaseIdealLoop *phase ); 760 761 // Reassociate invariant expressions. 762 void reassociate_invariants(PhaseIdealLoop *phase); 763 // Reassociate invariant binary expressions. 764 Node* reassociate(Node* n1, PhaseIdealLoop *phase); 765 // Reassociate invariant add and subtract expressions. 766 Node* reassociate_add_sub(Node* n1, int inv1_idx, int inv2_idx, PhaseIdealLoop *phase); 767 // Return nonzero index of invariant operand if invariant and variant 768 // are combined with an associative binary. Helper for reassociate_invariants. 769 int find_invariant(Node* n, PhaseIdealLoop *phase); 770 // Return TRUE if "n" is associative. 771 bool is_associative(Node* n, Node* base=NULL); 772 773 // Return true if n is invariant 774 bool is_invariant(Node* n) const; 775 776 // Put loop body on igvn work list 777 void record_for_igvn(); 778 779 bool is_root() { return _parent == NULL; } 780 // A proper/reducible loop w/o any (occasional) dead back-edge. 781 bool is_loop() { return !_irreducible && !tail()->is_top(); } 782 bool is_counted() { return is_loop() && _head->is_CountedLoop(); } 783 bool is_innermost() { return is_loop() && _child == NULL; } 784 785 void remove_main_post_loops(CountedLoopNode *cl, PhaseIdealLoop *phase); 786 787 #ifndef PRODUCT 788 void dump_head() const; // Dump loop head only 789 void dump() const; // Dump this loop recursively 790 void verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const; 791 #endif 792 793 private: 794 enum { EMPTY_LOOP_SIZE = 7 }; // Number of nodes in an empty loop. 795 796 // Estimate the number of nodes resulting from control and data flow merge. 797 uint est_loop_flow_merge_sz() const; 798 799 // Check if the number of residual iterations is large with unroll_cnt. 800 // Return true if the residual iterations are more than 10% of the trip count. 801 bool is_residual_iters_large(int unroll_cnt, CountedLoopNode *cl) const { 802 return (unroll_cnt - 1) * (100.0 / LoopPercentProfileLimit) > cl->profile_trip_cnt(); 803 } 804 805 void collect_loop_core_nodes(PhaseIdealLoop* phase, Unique_Node_List& wq) const; 806 807 bool empty_loop_with_data_nodes(PhaseIdealLoop* phase) const; 808 809 void enqueue_data_nodes(PhaseIdealLoop* phase, Unique_Node_List& empty_loop_nodes, Unique_Node_List& wq) const; 810 811 bool process_safepoint(PhaseIdealLoop* phase, Unique_Node_List& empty_loop_nodes, Unique_Node_List& wq, 812 Node* sfpt) const; 813 814 bool empty_loop_candidate(PhaseIdealLoop* phase) const; 815 816 bool empty_loop_with_extra_nodes_candidate(PhaseIdealLoop* phase) const; 817 }; 818 819 // -----------------------------PhaseIdealLoop--------------------------------- 820 // Computes the mapping from Nodes to IdealLoopTrees. Organizes IdealLoopTrees 821 // into a loop tree. Drives the loop-based transformations on the ideal graph. 822 class PhaseIdealLoop : public PhaseTransform { 823 friend class IdealLoopTree; 824 friend class SuperWord; 825 friend class CountedLoopReserveKit; 826 friend class ShenandoahBarrierC2Support; 827 friend class AutoNodeBudget; 828 829 // Pre-computed def-use info 830 PhaseIterGVN &_igvn; 831 832 // Head of loop tree 833 IdealLoopTree* _ltree_root; 834 835 // Array of pre-order numbers, plus post-visited bit. 836 // ZERO for not pre-visited. EVEN for pre-visited but not post-visited. 837 // ODD for post-visited. Other bits are the pre-order number. 838 uint *_preorders; 839 uint _max_preorder; 840 841 const PhaseIdealLoop* _verify_me; 842 bool _verify_only; 843 844 // Allocate _preorders[] array 845 void allocate_preorders() { 846 _max_preorder = C->unique()+8; 847 _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder); 848 memset(_preorders, 0, sizeof(uint) * _max_preorder); 849 } 850 851 // Allocate _preorders[] array 852 void reallocate_preorders() { 853 if ( _max_preorder < C->unique() ) { 854 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique()); 855 _max_preorder = C->unique(); 856 } 857 memset(_preorders, 0, sizeof(uint) * _max_preorder); 858 } 859 860 // Check to grow _preorders[] array for the case when build_loop_tree_impl() 861 // adds new nodes. 862 void check_grow_preorders( ) { 863 if ( _max_preorder < C->unique() ) { 864 uint newsize = _max_preorder<<1; // double size of array 865 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize); 866 memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder)); 867 _max_preorder = newsize; 868 } 869 } 870 // Check for pre-visited. Zero for NOT visited; non-zero for visited. 871 int is_visited( Node *n ) const { return _preorders[n->_idx]; } 872 // Pre-order numbers are written to the Nodes array as low-bit-set values. 873 void set_preorder_visited( Node *n, int pre_order ) { 874 assert( !is_visited( n ), "already set" ); 875 _preorders[n->_idx] = (pre_order<<1); 876 }; 877 // Return pre-order number. 878 int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; } 879 880 // Check for being post-visited. 881 // Should be previsited already (checked with assert(is_visited(n))). 882 int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; } 883 884 // Mark as post visited 885 void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; } 886 887 public: 888 // Set/get control node out. Set lower bit to distinguish from IdealLoopTree 889 // Returns true if "n" is a data node, false if it's a control node. 890 bool has_ctrl( Node *n ) const { return ((intptr_t)_nodes[n->_idx]) & 1; } 891 892 private: 893 // clear out dead code after build_loop_late 894 Node_List _deadlist; 895 896 // Support for faster execution of get_late_ctrl()/dom_lca() 897 // when a node has many uses and dominator depth is deep. 898 GrowableArray<jlong> _dom_lca_tags; 899 uint _dom_lca_tags_round; 900 void init_dom_lca_tags(); 901 902 // Helper for debugging bad dominance relationships 903 bool verify_dominance(Node* n, Node* use, Node* LCA, Node* early); 904 905 Node* compute_lca_of_uses(Node* n, Node* early, bool verify = false); 906 907 // Inline wrapper for frequent cases: 908 // 1) only one use 909 // 2) a use is the same as the current LCA passed as 'n1' 910 Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) { 911 assert( n->is_CFG(), "" ); 912 // Fast-path NULL lca 913 if( lca != NULL && lca != n ) { 914 assert( lca->is_CFG(), "" ); 915 // find LCA of all uses 916 n = dom_lca_for_get_late_ctrl_internal( lca, n, tag ); 917 } 918 return find_non_split_ctrl(n); 919 } 920 Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag ); 921 922 // Helper function for directing control inputs away from CFG split points. 923 Node *find_non_split_ctrl( Node *ctrl ) const { 924 if (ctrl != NULL) { 925 if (ctrl->is_MultiBranch()) { 926 ctrl = ctrl->in(0); 927 } 928 assert(ctrl->is_CFG(), "CFG"); 929 } 930 return ctrl; 931 } 932 933 Node* cast_incr_before_loop(Node* incr, Node* ctrl, Node* loop); 934 935 #ifdef ASSERT 936 void ensure_zero_trip_guard_proj(Node* node, bool is_main_loop); 937 #endif 938 void copy_skeleton_predicates_to_main_loop_helper(Node* predicate, Node* init, Node* stride, IdealLoopTree* outer_loop, LoopNode* outer_main_head, 939 uint dd_main_head, const uint idx_before_pre_post, const uint idx_after_post_before_pre, 940 Node* zero_trip_guard_proj_main, Node* zero_trip_guard_proj_post, const Node_List &old_new); 941 void copy_skeleton_predicates_to_main_loop(CountedLoopNode* pre_head, Node* init, Node* stride, IdealLoopTree* outer_loop, LoopNode* outer_main_head, 942 uint dd_main_head, const uint idx_before_pre_post, const uint idx_after_post_before_pre, 943 Node* zero_trip_guard_proj_main, Node* zero_trip_guard_proj_post, const Node_List &old_new); 944 Node* clone_skeleton_predicate_and_initialize(Node* iff, Node* new_init, Node* new_stride, Node* predicate, Node* uncommon_proj, Node* control, 945 IdealLoopTree* outer_loop, Node* input_proj); 946 Node* clone_skeleton_predicate_bool(Node* iff, Node* new_init, Node* new_stride, Node* control); 947 static bool skeleton_predicate_has_opaque(IfNode* iff); 948 static void count_opaque_loop_nodes(Node* n, uint& init, uint& stride); 949 static bool subgraph_has_opaque(Node* n); 950 static void get_skeleton_predicates(Node* predicate, Unique_Node_List& list, bool get_opaque = false); 951 void update_main_loop_skeleton_predicates(Node* ctrl, CountedLoopNode* loop_head, Node* init, int stride_con); 952 void copy_skeleton_predicates_to_post_loop(LoopNode* main_loop_head, CountedLoopNode* post_loop_head, Node* init, Node* stride); 953 void initialize_skeleton_predicates_for_peeled_loop(ProjNode* predicate, LoopNode* outer_loop_head, int dd_outer_loop_head, 954 Node* init, Node* stride, IdealLoopTree* outer_loop, 955 const uint idx_before_clone, const Node_List& old_new); 956 void insert_loop_limit_check(ProjNode* limit_check_proj, Node* cmp_limit, Node* bol); 957 #ifdef ASSERT 958 bool only_has_infinite_loops(); 959 #endif 960 961 void log_loop_tree(); 962 963 public: 964 965 PhaseIterGVN &igvn() const { return _igvn; } 966 967 bool has_node( Node* n ) const { 968 guarantee(n != NULL, "No Node."); 969 return _nodes[n->_idx] != NULL; 970 } 971 // check if transform created new nodes that need _ctrl recorded 972 Node *get_late_ctrl( Node *n, Node *early ); 973 Node *get_early_ctrl( Node *n ); 974 Node *get_early_ctrl_for_expensive(Node *n, Node* earliest); 975 void set_early_ctrl(Node* n, bool update_body); 976 void set_subtree_ctrl(Node* n, bool update_body); 977 void set_ctrl( Node *n, Node *ctrl ) { 978 assert( !has_node(n) || has_ctrl(n), "" ); 979 assert( ctrl->in(0), "cannot set dead control node" ); 980 assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" ); 981 _nodes.map( n->_idx, (Node*)((intptr_t)ctrl + 1) ); 982 } 983 // Set control and update loop membership 984 void set_ctrl_and_loop(Node* n, Node* ctrl) { 985 IdealLoopTree* old_loop = get_loop(get_ctrl(n)); 986 IdealLoopTree* new_loop = get_loop(ctrl); 987 if (old_loop != new_loop) { 988 if (old_loop->_child == NULL) old_loop->_body.yank(n); 989 if (new_loop->_child == NULL) new_loop->_body.push(n); 990 } 991 set_ctrl(n, ctrl); 992 } 993 // Control nodes can be replaced or subsumed. During this pass they 994 // get their replacement Node in slot 1. Instead of updating the block 995 // location of all Nodes in the subsumed block, we lazily do it. As we 996 // pull such a subsumed block out of the array, we write back the final 997 // correct block. 998 Node *get_ctrl( Node *i ) { 999 1000 assert(has_node(i), ""); 1001 Node *n = get_ctrl_no_update(i); 1002 _nodes.map( i->_idx, (Node*)((intptr_t)n + 1) ); 1003 assert(has_node(i) && has_ctrl(i), ""); 1004 assert(n == find_non_split_ctrl(n), "must return legal ctrl" ); 1005 return n; 1006 } 1007 // true if CFG node d dominates CFG node n 1008 bool is_dominator(Node *d, Node *n); 1009 // return get_ctrl for a data node and self(n) for a CFG node 1010 Node* ctrl_or_self(Node* n) { 1011 if (has_ctrl(n)) 1012 return get_ctrl(n); 1013 else { 1014 assert (n->is_CFG(), "must be a CFG node"); 1015 return n; 1016 } 1017 } 1018 1019 Node *get_ctrl_no_update_helper(Node *i) const { 1020 assert(has_ctrl(i), "should be control, not loop"); 1021 return (Node*)(((intptr_t)_nodes[i->_idx]) & ~1); 1022 } 1023 1024 Node *get_ctrl_no_update(Node *i) const { 1025 assert( has_ctrl(i), "" ); 1026 Node *n = get_ctrl_no_update_helper(i); 1027 if (!n->in(0)) { 1028 // Skip dead CFG nodes 1029 do { 1030 n = get_ctrl_no_update_helper(n); 1031 } while (!n->in(0)); 1032 n = find_non_split_ctrl(n); 1033 } 1034 return n; 1035 } 1036 1037 // Check for loop being set 1038 // "n" must be a control node. Returns true if "n" is known to be in a loop. 1039 bool has_loop( Node *n ) const { 1040 assert(!has_node(n) || !has_ctrl(n), ""); 1041 return has_node(n); 1042 } 1043 // Set loop 1044 void set_loop( Node *n, IdealLoopTree *loop ) { 1045 _nodes.map(n->_idx, (Node*)loop); 1046 } 1047 // Lazy-dazy update of 'get_ctrl' and 'idom_at' mechanisms. Replace 1048 // the 'old_node' with 'new_node'. Kill old-node. Add a reference 1049 // from old_node to new_node to support the lazy update. Reference 1050 // replaces loop reference, since that is not needed for dead node. 1051 void lazy_update(Node *old_node, Node *new_node) { 1052 assert(old_node != new_node, "no cycles please"); 1053 // Re-use the side array slot for this node to provide the 1054 // forwarding pointer. 1055 _nodes.map(old_node->_idx, (Node*)((intptr_t)new_node + 1)); 1056 } 1057 void lazy_replace(Node *old_node, Node *new_node) { 1058 _igvn.replace_node(old_node, new_node); 1059 lazy_update(old_node, new_node); 1060 } 1061 1062 private: 1063 1064 // Place 'n' in some loop nest, where 'n' is a CFG node 1065 void build_loop_tree(); 1066 int build_loop_tree_impl( Node *n, int pre_order ); 1067 // Insert loop into the existing loop tree. 'innermost' is a leaf of the 1068 // loop tree, not the root. 1069 IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost ); 1070 1071 // Place Data nodes in some loop nest 1072 void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ); 1073 void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ); 1074 void build_loop_late_post_work(Node* n, bool pinned); 1075 void build_loop_late_post(Node* n); 1076 void verify_strip_mined_scheduling(Node *n, Node* least); 1077 1078 // Array of immediate dominance info for each CFG node indexed by node idx 1079 private: 1080 uint _idom_size; 1081 Node **_idom; // Array of immediate dominators 1082 uint *_dom_depth; // Used for fast LCA test 1083 GrowableArray<uint>* _dom_stk; // For recomputation of dom depth 1084 LoopOptsMode _mode; 1085 1086 // build the loop tree and perform any requested optimizations 1087 void build_and_optimize(); 1088 1089 // Dominators for the sea of nodes 1090 void Dominators(); 1091 1092 // Compute the Ideal Node to Loop mapping 1093 PhaseIdealLoop(PhaseIterGVN& igvn, LoopOptsMode mode) : 1094 PhaseTransform(Ideal_Loop), 1095 _igvn(igvn), 1096 _verify_me(nullptr), 1097 _verify_only(false), 1098 _mode(mode), 1099 _nodes_required(UINT_MAX) { 1100 assert(mode != LoopOptsVerify, "wrong constructor to verify IdealLoop"); 1101 build_and_optimize(); 1102 } 1103 1104 #ifndef PRODUCT 1105 // Verify that verify_me made the same decisions as a fresh run 1106 // or only verify that the graph is valid if verify_me is null. 1107 PhaseIdealLoop(PhaseIterGVN& igvn, const PhaseIdealLoop* verify_me = nullptr) : 1108 PhaseTransform(Ideal_Loop), 1109 _igvn(igvn), 1110 _verify_me(verify_me), 1111 _verify_only(verify_me == nullptr), 1112 _mode(LoopOptsVerify), 1113 _nodes_required(UINT_MAX) { 1114 build_and_optimize(); 1115 } 1116 #endif 1117 1118 public: 1119 Node* idom_no_update(Node* d) const { 1120 return idom_no_update(d->_idx); 1121 } 1122 1123 Node* idom_no_update(uint didx) const { 1124 assert(didx < _idom_size, "oob"); 1125 Node* n = _idom[didx]; 1126 assert(n != NULL,"Bad immediate dominator info."); 1127 while (n->in(0) == NULL) { // Skip dead CFG nodes 1128 n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1); 1129 assert(n != NULL,"Bad immediate dominator info."); 1130 } 1131 return n; 1132 } 1133 1134 Node *idom(Node* d) const { 1135 return idom(d->_idx); 1136 } 1137 1138 Node *idom(uint didx) const { 1139 Node *n = idom_no_update(didx); 1140 _idom[didx] = n; // Lazily remove dead CFG nodes from table. 1141 return n; 1142 } 1143 1144 uint dom_depth(Node* d) const { 1145 guarantee(d != NULL, "Null dominator info."); 1146 guarantee(d->_idx < _idom_size, ""); 1147 return _dom_depth[d->_idx]; 1148 } 1149 void set_idom(Node* d, Node* n, uint dom_depth); 1150 // Locally compute IDOM using dom_lca call 1151 Node *compute_idom( Node *region ) const; 1152 // Recompute dom_depth 1153 void recompute_dom_depth(); 1154 1155 // Is safept not required by an outer loop? 1156 bool is_deleteable_safept(Node* sfpt); 1157 1158 // Replace parallel induction variable (parallel to trip counter) 1159 void replace_parallel_iv(IdealLoopTree *loop); 1160 1161 Node *dom_lca( Node *n1, Node *n2 ) const { 1162 return find_non_split_ctrl(dom_lca_internal(n1, n2)); 1163 } 1164 Node *dom_lca_internal( Node *n1, Node *n2 ) const; 1165 1166 // Build and verify the loop tree without modifying the graph. This 1167 // is useful to verify that all inputs properly dominate their uses. 1168 static void verify(PhaseIterGVN& igvn) { 1169 #ifdef ASSERT 1170 ResourceMark rm; 1171 Compile::TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 1172 PhaseIdealLoop v(igvn); 1173 #endif 1174 } 1175 1176 // Recommended way to use PhaseIdealLoop. 1177 // Run PhaseIdealLoop in some mode and allocates a local scope for memory allocations. 1178 static void optimize(PhaseIterGVN &igvn, LoopOptsMode mode) { 1179 ResourceMark rm; 1180 PhaseIdealLoop v(igvn, mode); 1181 1182 Compile* C = Compile::current(); 1183 if (!C->failing()) { 1184 // Cleanup any modified bits 1185 igvn.optimize(); 1186 1187 v.log_loop_tree(); 1188 } 1189 } 1190 1191 // True if the method has at least 1 irreducible loop 1192 bool _has_irreducible_loops; 1193 1194 // Per-Node transform 1195 virtual Node* transform(Node* n) { return NULL; } 1196 1197 Node* loop_exit_control(Node* x, IdealLoopTree* loop); 1198 Node* loop_exit_test(Node* back_control, IdealLoopTree* loop, Node*& incr, Node*& limit, BoolTest::mask& bt, float& cl_prob); 1199 Node* loop_iv_incr(Node* incr, Node* x, IdealLoopTree* loop, Node*& phi_incr); 1200 Node* loop_iv_stride(Node* incr, IdealLoopTree* loop, Node*& xphi); 1201 PhiNode* loop_iv_phi(Node* xphi, Node* phi_incr, Node* x, IdealLoopTree* loop); 1202 1203 bool is_counted_loop(Node* x, IdealLoopTree*&loop, BasicType iv_bt); 1204 1205 Node* loop_nest_replace_iv(Node* iv_to_replace, Node* inner_iv, Node* outer_phi, Node* inner_head, BasicType bt); 1206 bool create_loop_nest(IdealLoopTree* loop, Node_List &old_new); 1207 #ifdef ASSERT 1208 bool convert_to_long_loop(Node* cmp, Node* phi, IdealLoopTree* loop); 1209 #endif 1210 void add_empty_predicate(Deoptimization::DeoptReason reason, Node* inner_head, IdealLoopTree* loop, SafePointNode* sfpt); 1211 SafePointNode* find_safepoint(Node* back_control, Node* x, IdealLoopTree* loop); 1212 IdealLoopTree* insert_outer_loop(IdealLoopTree* loop, LoopNode* outer_l, Node* outer_ift); 1213 IdealLoopTree* create_outer_strip_mined_loop(BoolNode *test, Node *cmp, Node *init_control, 1214 IdealLoopTree* loop, float cl_prob, float le_fcnt, 1215 Node*& entry_control, Node*& iffalse); 1216 1217 Node* exact_limit( IdealLoopTree *loop ); 1218 1219 // Return a post-walked LoopNode 1220 IdealLoopTree *get_loop( Node *n ) const { 1221 // Dead nodes have no loop, so return the top level loop instead 1222 if (!has_node(n)) return _ltree_root; 1223 assert(!has_ctrl(n), ""); 1224 return (IdealLoopTree*)_nodes[n->_idx]; 1225 } 1226 1227 IdealLoopTree* ltree_root() const { return _ltree_root; } 1228 1229 // Is 'n' a (nested) member of 'loop'? 1230 int is_member( const IdealLoopTree *loop, Node *n ) const { 1231 return loop->is_member(get_loop(n)); } 1232 1233 // This is the basic building block of the loop optimizations. It clones an 1234 // entire loop body. It makes an old_new loop body mapping; with this 1235 // mapping you can find the new-loop equivalent to an old-loop node. All 1236 // new-loop nodes are exactly equal to their old-loop counterparts, all 1237 // edges are the same. All exits from the old-loop now have a RegionNode 1238 // that merges the equivalent new-loop path. This is true even for the 1239 // normal "loop-exit" condition. All uses of loop-invariant old-loop values 1240 // now come from (one or more) Phis that merge their new-loop equivalents. 1241 // Parameter side_by_side_idom: 1242 // When side_by_size_idom is NULL, the dominator tree is constructed for 1243 // the clone loop to dominate the original. Used in construction of 1244 // pre-main-post loop sequence. 1245 // When nonnull, the clone and original are side-by-side, both are 1246 // dominated by the passed in side_by_side_idom node. Used in 1247 // construction of unswitched loops. 1248 enum CloneLoopMode { 1249 IgnoreStripMined = 0, // Only clone inner strip mined loop 1250 CloneIncludesStripMined = 1, // clone both inner and outer strip mined loops 1251 ControlAroundStripMined = 2 // Only clone inner strip mined loop, 1252 // result control flow branches 1253 // either to inner clone or outer 1254 // strip mined loop. 1255 }; 1256 void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth, 1257 CloneLoopMode mode, Node* side_by_side_idom = NULL); 1258 void clone_loop_handle_data_uses(Node* old, Node_List &old_new, 1259 IdealLoopTree* loop, IdealLoopTree* companion_loop, 1260 Node_List*& split_if_set, Node_List*& split_bool_set, 1261 Node_List*& split_cex_set, Node_List& worklist, 1262 uint new_counter, CloneLoopMode mode); 1263 void clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop, 1264 IdealLoopTree* outer_loop, int dd, Node_List &old_new, 1265 Node_List& extra_data_nodes); 1266 1267 // If we got the effect of peeling, either by actually peeling or by 1268 // making a pre-loop which must execute at least once, we can remove 1269 // all loop-invariant dominated tests in the main body. 1270 void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new ); 1271 1272 // Generate code to do a loop peel for the given loop (and body). 1273 // old_new is a temp array. 1274 void do_peeling( IdealLoopTree *loop, Node_List &old_new ); 1275 1276 // Add pre and post loops around the given loop. These loops are used 1277 // during RCE, unrolling and aligning loops. 1278 void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ); 1279 1280 // Add post loop after the given loop. 1281 Node *insert_post_loop(IdealLoopTree* loop, Node_List& old_new, 1282 CountedLoopNode* main_head, CountedLoopEndNode* main_end, 1283 Node*& incr, Node* limit, CountedLoopNode*& post_head); 1284 1285 // Add an RCE'd post loop which we will multi-version adapt for run time test path usage 1286 void insert_scalar_rced_post_loop( IdealLoopTree *loop, Node_List &old_new ); 1287 1288 // Add a vector post loop between a vector main loop and the current post loop 1289 void insert_vector_post_loop(IdealLoopTree *loop, Node_List &old_new); 1290 // If Node n lives in the back_ctrl block, we clone a private version of n 1291 // in preheader_ctrl block and return that, otherwise return n. 1292 Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones ); 1293 1294 // Take steps to maximally unroll the loop. Peel any odd iterations, then 1295 // unroll to do double iterations. The next round of major loop transforms 1296 // will repeat till the doubled loop body does all remaining iterations in 1 1297 // pass. 1298 void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ); 1299 1300 // Unroll the loop body one step - make each trip do 2 iterations. 1301 void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ); 1302 1303 // Mark vector reduction candidates before loop unrolling 1304 void mark_reductions( IdealLoopTree *loop ); 1305 1306 // Return true if exp is a constant times an induction var 1307 bool is_scaled_iv(Node* exp, Node* iv, BasicType bt, jlong* p_scale, bool* p_short_scale, int depth = 0); 1308 1309 bool is_iv(Node* exp, Node* iv, BasicType bt); 1310 1311 // Return true if exp is a scaled induction var plus (or minus) constant 1312 bool is_scaled_iv_plus_offset(Node* exp, Node* iv, BasicType bt, jlong* p_scale, Node** p_offset, bool* p_short_scale = NULL, int depth = 0); 1313 bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset) { 1314 jlong long_scale; 1315 if (is_scaled_iv_plus_offset(exp, iv, T_INT, &long_scale, p_offset)) { 1316 int int_scale = checked_cast<int>(long_scale); 1317 if (p_scale != NULL) { 1318 *p_scale = int_scale; 1319 } 1320 return true; 1321 } 1322 return false; 1323 } 1324 // Helper for finding more complex matches to is_scaled_iv_plus_offset. 1325 bool is_scaled_iv_plus_extra_offset(Node* exp1, Node* offset2, Node* iv, 1326 BasicType bt, 1327 jlong* p_scale, Node** p_offset, 1328 bool* p_short_scale, int depth); 1329 1330 // Create a new if above the uncommon_trap_if_pattern for the predicate to be promoted 1331 ProjNode* create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry, Deoptimization::DeoptReason reason, 1332 int opcode, bool rewire_uncommon_proj_phi_inputs = false, 1333 bool if_cont_is_true_proj = true); 1334 1335 private: 1336 // Helper functions for create_new_if_for_predicate() 1337 void set_ctrl_of_nodes_with_same_ctrl(Node* node, ProjNode* old_ctrl, Node* new_ctrl); 1338 Unique_Node_List find_nodes_with_same_ctrl(Node* node, const ProjNode* ctrl); 1339 Node* clone_nodes_with_same_ctrl(Node* node, ProjNode* old_ctrl, Node* new_ctrl); 1340 Dict clone_nodes(const Node_List& list_to_clone); 1341 void rewire_cloned_nodes_to_ctrl(const ProjNode* old_ctrl, Node* new_ctrl, const Node_List& nodes_with_same_ctrl, 1342 const Dict& old_new_mapping); 1343 void rewire_inputs_of_clones_to_clones(Node* new_ctrl, Node* clone, const Dict& old_new_mapping, const Node* next); 1344 1345 public: 1346 void register_control(Node* n, IdealLoopTree *loop, Node* pred, bool update_body = true); 1347 1348 static Node* skip_all_loop_predicates(Node* entry); 1349 static Node* skip_loop_predicates(Node* entry); 1350 static ProjNode* next_predicate(ProjNode* predicate); 1351 1352 // Find a good location to insert a predicate 1353 static ProjNode* find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason); 1354 1355 class Predicates { 1356 public: 1357 // given loop entry, find all predicates above loop 1358 Predicates(Node* entry); 1359 1360 // Proj of empty loop limit check predicate 1361 ProjNode* loop_limit_check() { 1362 return _loop_limit_check; 1363 } 1364 1365 // Proj of empty profile predicate 1366 ProjNode* profile_predicate() { 1367 return _profile_predicate; 1368 } 1369 1370 // Proj of empty predicate 1371 ProjNode* predicate() { 1372 return _predicate; 1373 } 1374 1375 // First control node above all predicates 1376 Node* skip_all() { 1377 return _entry_to_all_predicates; 1378 } 1379 1380 private: 1381 ProjNode*_loop_limit_check = nullptr; 1382 ProjNode* _profile_predicate = nullptr; 1383 ProjNode* _predicate = nullptr; 1384 Node* _entry_to_all_predicates = nullptr; 1385 }; 1386 1387 // Find a predicate 1388 static Node* find_predicate(Node* entry); 1389 // Construct a range check for a predicate if 1390 BoolNode* rc_predicate(IdealLoopTree *loop, Node* ctrl, 1391 int scale, Node* offset, 1392 Node* init, Node* limit, jint stride, 1393 Node* range, bool upper, bool &overflow, 1394 bool negate); 1395 1396 // Implementation of the loop predication to promote checks outside the loop 1397 bool loop_predication_impl(IdealLoopTree *loop); 1398 bool loop_predication_impl_helper(IdealLoopTree *loop, ProjNode* proj, ProjNode *predicate_proj, 1399 CountedLoopNode *cl, ConNode* zero, Invariance& invar, 1400 Deoptimization::DeoptReason reason); 1401 bool loop_predication_should_follow_branches(IdealLoopTree *loop, ProjNode *predicate_proj, float& loop_trip_cnt); 1402 void loop_predication_follow_branches(Node *c, IdealLoopTree *loop, float loop_trip_cnt, 1403 PathFrequency& pf, Node_Stack& stack, VectorSet& seen, 1404 Node_List& if_proj_list); 1405 ProjNode* insert_initial_skeleton_predicate(IfNode* iff, IdealLoopTree *loop, 1406 ProjNode* proj, ProjNode *predicate_proj, 1407 ProjNode* upper_bound_proj, 1408 int scale, Node* offset, 1409 Node* init, Node* limit, jint stride, 1410 Node* rng, bool& overflow, 1411 Deoptimization::DeoptReason reason); 1412 Node* add_range_check_predicate(IdealLoopTree* loop, CountedLoopNode* cl, 1413 Node* predicate_proj, int scale_con, Node* offset, 1414 Node* limit, jint stride_con, Node* value); 1415 1416 // Helper function to collect predicate for eliminating the useless ones 1417 void collect_potentially_useful_predicates(IdealLoopTree *loop, Unique_Node_List &predicate_opaque1); 1418 void eliminate_useless_predicates(); 1419 1420 // Change the control input of expensive nodes to allow commoning by 1421 // IGVN when it is guaranteed to not result in a more frequent 1422 // execution of the expensive node. Return true if progress. 1423 bool process_expensive_nodes(); 1424 1425 // Check whether node has become unreachable 1426 bool is_node_unreachable(Node *n) const { 1427 return !has_node(n) || n->is_unreachable(_igvn); 1428 } 1429 1430 // Eliminate range-checks and other trip-counter vs loop-invariant tests. 1431 int do_range_check( IdealLoopTree *loop, Node_List &old_new ); 1432 1433 // Check to see if do_range_check(...) cleaned the main loop of range-checks 1434 void has_range_checks(IdealLoopTree *loop); 1435 1436 // Process post loops which have range checks and try to build a multi-version 1437 // guard to safely determine if we can execute the post loop which was RCE'd. 1438 bool multi_version_post_loops(IdealLoopTree *rce_loop, IdealLoopTree *legacy_loop); 1439 1440 // Cause the rce'd post loop to optimized away, this happens if we cannot complete multiverioning 1441 void poison_rce_post_loop(IdealLoopTree *rce_loop); 1442 1443 // Create a slow version of the loop by cloning the loop 1444 // and inserting an if to select fast-slow versions. 1445 // Return the inserted if. 1446 IfNode* create_slow_version_of_loop(IdealLoopTree *loop, 1447 Node_List &old_new, 1448 Node_List &unswitch_iffs, 1449 CloneLoopMode mode); 1450 1451 // Clone a loop and return the clone head (clone_loop_head). 1452 // Added nodes include int(1), int(0) - disconnected, If, IfTrue, IfFalse, 1453 // This routine was created for usage in CountedLoopReserveKit. 1454 // 1455 // int(1) -> If -> IfTrue -> original_loop_head 1456 // | 1457 // V 1458 // IfFalse -> clone_loop_head (returned by function pointer) 1459 // 1460 LoopNode* create_reserve_version_of_loop(IdealLoopTree *loop, CountedLoopReserveKit* lk); 1461 // Clone loop with an invariant test (that does not exit) and 1462 // insert a clone of the test that selects which version to 1463 // execute. 1464 void do_unswitching (IdealLoopTree *loop, Node_List &old_new); 1465 1466 // Find candidate "if" for unswitching 1467 IfNode* find_unswitching_candidate(const IdealLoopTree *loop, Node_List& unswitch_iffs) const; 1468 1469 // Range Check Elimination uses this function! 1470 // Constrain the main loop iterations so the affine function: 1471 // low_limit <= scale_con * I + offset < upper_limit 1472 // always holds true. That is, either increase the number of iterations in 1473 // the pre-loop or the post-loop until the condition holds true in the main 1474 // loop. Scale_con, offset and limit are all loop invariant. 1475 void add_constraint(jlong stride_con, jlong scale_con, Node* offset, Node* low_limit, Node* upper_limit, Node* pre_ctrl, Node** pre_limit, Node** main_limit); 1476 // Helper function for add_constraint(). 1477 Node* adjust_limit(bool reduce, Node* scale, Node* offset, Node* rc_limit, Node* old_limit, Node* pre_ctrl, bool round); 1478 1479 // Partially peel loop up through last_peel node. 1480 bool partial_peel( IdealLoopTree *loop, Node_List &old_new ); 1481 bool duplicate_loop_backedge(IdealLoopTree *loop, Node_List &old_new); 1482 1483 // Create a scheduled list of nodes control dependent on ctrl set. 1484 void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched ); 1485 // Has a use in the vector set 1486 bool has_use_in_set( Node* n, VectorSet& vset ); 1487 // Has use internal to the vector set (ie. not in a phi at the loop head) 1488 bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ); 1489 // clone "n" for uses that are outside of loop 1490 int clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ); 1491 // clone "n" for special uses that are in the not_peeled region 1492 void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 1493 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ); 1494 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 1495 void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ); 1496 #ifdef ASSERT 1497 // Validate the loop partition sets: peel and not_peel 1498 bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel ); 1499 // Ensure that uses outside of loop are of the right form 1500 bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 1501 uint orig_exit_idx, uint clone_exit_idx); 1502 bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx); 1503 #endif 1504 1505 // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.) 1506 int stride_of_possible_iv( Node* iff ); 1507 bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; } 1508 // Return the (unique) control output node that's in the loop (if it exists.) 1509 Node* stay_in_loop( Node* n, IdealLoopTree *loop); 1510 // Insert a signed compare loop exit cloned from an unsigned compare. 1511 IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop); 1512 void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop); 1513 // Utility to register node "n" with PhaseIdealLoop 1514 void register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth); 1515 // Utility to create an if-projection 1516 ProjNode* proj_clone(ProjNode* p, IfNode* iff); 1517 // Force the iff control output to be the live_proj 1518 Node* short_circuit_if(IfNode* iff, ProjNode* live_proj); 1519 // Insert a region before an if projection 1520 RegionNode* insert_region_before_proj(ProjNode* proj); 1521 // Insert a new if before an if projection 1522 ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj); 1523 1524 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1525 // "Nearly" because all Nodes have been cloned from the original in the loop, 1526 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1527 // through the Phi recursively, and return a Bool. 1528 Node* clone_iff(PhiNode* phi); 1529 CmpNode* clone_bool(PhiNode* phi); 1530 1531 1532 // Rework addressing expressions to get the most loop-invariant stuff 1533 // moved out. We'd like to do all associative operators, but it's especially 1534 // important (common) to do address expressions. 1535 Node* remix_address_expressions(Node* n); 1536 Node* remix_address_expressions_add_left_shift(Node* n, IdealLoopTree* n_loop, Node* n_ctrl, BasicType bt); 1537 1538 // Convert add to muladd to generate MuladdS2I under certain criteria 1539 Node * convert_add_to_muladd(Node * n); 1540 1541 // Attempt to use a conditional move instead of a phi/branch 1542 Node *conditional_move( Node *n ); 1543 1544 // Check for aggressive application of 'split-if' optimization, 1545 // using basic block level info. 1546 void split_if_with_blocks ( VectorSet &visited, Node_Stack &nstack); 1547 Node *split_if_with_blocks_pre ( Node *n ); 1548 void split_if_with_blocks_post( Node *n ); 1549 Node *has_local_phi_input( Node *n ); 1550 // Mark an IfNode as being dominated by a prior test, 1551 // without actually altering the CFG (and hence IDOM info). 1552 void dominated_by(IfProjNode* prevdom, IfNode* iff, bool flip = false, bool exclude_loop_predicate = false); 1553 1554 // Split Node 'n' through merge point 1555 RegionNode* split_thru_region(Node* n, RegionNode* region); 1556 // Split Node 'n' through merge point if there is enough win. 1557 Node *split_thru_phi( Node *n, Node *region, int policy ); 1558 // Found an If getting its condition-code input from a Phi in the 1559 // same block. Split thru the Region. 1560 void do_split_if(Node *iff, RegionNode** new_false_region = NULL, RegionNode** new_true_region = NULL); 1561 1562 // Conversion of fill/copy patterns into intrinsic versions 1563 bool do_intrinsify_fill(); 1564 bool intrinsify_fill(IdealLoopTree* lpt); 1565 bool match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value, 1566 Node*& shift, Node*& offset); 1567 1568 private: 1569 // Return a type based on condition control flow 1570 const TypeInt* filtered_type( Node *n, Node* n_ctrl); 1571 const TypeInt* filtered_type( Node *n ) { return filtered_type(n, NULL); } 1572 // Helpers for filtered type 1573 const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl); 1574 1575 // Helper functions 1576 Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache ); 1577 Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true ); 1578 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 ); 1579 bool split_up( Node *n, Node *blk1, Node *blk2 ); 1580 void sink_use( Node *use, Node *post_loop ); 1581 Node* place_outside_loop(Node* useblock, IdealLoopTree* loop) const; 1582 Node* try_move_store_before_loop(Node* n, Node *n_ctrl); 1583 void try_move_store_after_loop(Node* n); 1584 void move_flat_array_check_out_of_loop(Node* n); 1585 bool identical_backtoback_ifs(Node *n); 1586 bool flatten_array_element_type_check(Node *n); 1587 bool can_split_if(Node *n_ctrl); 1588 bool cannot_split_division(const Node* n, const Node* region) const; 1589 static bool is_divisor_counted_loop_phi(const Node* divisor, const Node* loop); 1590 bool loop_phi_backedge_type_contains_zero(const Node* phi_divisor, const Type* zero) const; 1591 1592 // Determine if a method is too big for a/another round of split-if, based on 1593 // a magic (approximate) ratio derived from the equally magic constant 35000, 1594 // previously used for this purpose (but without relating to the node limit). 1595 bool must_throttle_split_if() { 1596 uint threshold = C->max_node_limit() * 2 / 5; 1597 return C->live_nodes() > threshold; 1598 } 1599 1600 // A simplistic node request tracking mechanism, where 1601 // = UINT_MAX Request not valid or made final. 1602 // < UINT_MAX Nodes currently requested (estimate). 1603 uint _nodes_required; 1604 1605 enum { REQUIRE_MIN = 70 }; 1606 1607 uint nodes_required() const { return _nodes_required; } 1608 1609 // Given the _currently_ available number of nodes, check whether there is 1610 // "room" for an additional request or not, considering the already required 1611 // number of nodes. Return TRUE if the new request is exceeding the node 1612 // budget limit, otherwise return FALSE. Note that this interpretation will 1613 // act pessimistic on additional requests when new nodes have already been 1614 // generated since the 'begin'. This behaviour fits with the intention that 1615 // node estimates/requests should be made upfront. 1616 bool exceeding_node_budget(uint required = 0) { 1617 assert(C->live_nodes() < C->max_node_limit(), "sanity"); 1618 uint available = C->max_node_limit() - C->live_nodes(); 1619 return available < required + _nodes_required + REQUIRE_MIN; 1620 } 1621 1622 uint require_nodes(uint require, uint minreq = REQUIRE_MIN) { 1623 precond(require > 0); 1624 _nodes_required += MAX2(require, minreq); 1625 return _nodes_required; 1626 } 1627 1628 bool may_require_nodes(uint require, uint minreq = REQUIRE_MIN) { 1629 return !exceeding_node_budget(require) && require_nodes(require, minreq) > 0; 1630 } 1631 1632 uint require_nodes_begin() { 1633 assert(_nodes_required == UINT_MAX, "Bad state (begin)."); 1634 _nodes_required = 0; 1635 return C->live_nodes(); 1636 } 1637 1638 // When a node request is final, optionally check that the requested number 1639 // of nodes was reasonably correct with respect to the number of new nodes 1640 // introduced since the last 'begin'. Always check that we have not exceeded 1641 // the maximum node limit. 1642 void require_nodes_final(uint live_at_begin, bool check_estimate) { 1643 assert(_nodes_required < UINT_MAX, "Bad state (final)."); 1644 1645 #ifdef ASSERT 1646 if (check_estimate) { 1647 // Check that the node budget request was not off by too much (x2). 1648 // Should this be the case we _surely_ need to improve the estimates 1649 // used in our budget calculations. 1650 if (C->live_nodes() - live_at_begin > 2 * _nodes_required) { 1651 log_info(compilation)("Bad node estimate: actual = %d >> request = %d", 1652 C->live_nodes() - live_at_begin, _nodes_required); 1653 } 1654 } 1655 #endif 1656 // Assert that we have stayed within the node budget limit. 1657 assert(C->live_nodes() < C->max_node_limit(), 1658 "Exceeding node budget limit: %d + %d > %d (request = %d)", 1659 C->live_nodes() - live_at_begin, live_at_begin, 1660 C->max_node_limit(), _nodes_required); 1661 1662 _nodes_required = UINT_MAX; 1663 } 1664 1665 // Clone loop predicates to slow and fast loop when unswitching a loop 1666 void clone_predicates_to_unswitched_loop(IdealLoopTree* loop, Node_List& old_new, ProjNode*& iffast_pred, ProjNode*& ifslow_pred); 1667 ProjNode* clone_predicate_to_unswitched_loop(ProjNode* predicate_proj, Node* new_entry, 1668 Deoptimization::DeoptReason reason, bool slow_loop); 1669 void clone_skeleton_predicates_to_unswitched_loop(IdealLoopTree* loop, const Node_List& old_new, Deoptimization::DeoptReason reason, 1670 ProjNode* old_predicate_proj, ProjNode* iffast_pred, ProjNode* ifslow_pred); 1671 ProjNode* clone_skeleton_predicate_for_unswitched_loops(Node* iff, ProjNode* predicate, 1672 Deoptimization::DeoptReason reason, 1673 ProjNode* output_proj); 1674 static void check_created_predicate_for_unswitching(const Node* new_entry) PRODUCT_RETURN; 1675 1676 bool _created_loop_node; 1677 DEBUG_ONLY(void dump_idoms(Node* early, Node* wrong_lca);) 1678 NOT_PRODUCT(void dump_idoms_in_reverse(const Node* n, const Node_List& idom_list) const;) 1679 1680 public: 1681 void set_created_loop_node() { _created_loop_node = true; } 1682 bool created_loop_node() { return _created_loop_node; } 1683 void register_new_node(Node* n, Node* blk); 1684 1685 #ifdef ASSERT 1686 void dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA); 1687 #endif 1688 1689 #ifndef PRODUCT 1690 void dump() const; 1691 void dump_idom(Node* n) const { dump_idom(n, 1000); } // For debugging 1692 void dump_idom(Node* n, uint count) const; 1693 void get_idoms(Node* n, uint count, Unique_Node_List& idoms) const; 1694 void dump(IdealLoopTree* loop, uint rpo_idx, Node_List &rpo_list) const; 1695 void verify() const; // Major slow :-) 1696 void verify_compare(Node* n, const PhaseIdealLoop* loop_verify, VectorSet &visited) const; 1697 IdealLoopTree* get_loop_idx(Node* n) const { 1698 // Dead nodes have no loop, so return the top level loop instead 1699 return _nodes[n->_idx] ? (IdealLoopTree*)_nodes[n->_idx] : _ltree_root; 1700 } 1701 // Print some stats 1702 static void print_statistics(); 1703 static int _loop_invokes; // Count of PhaseIdealLoop invokes 1704 static int _loop_work; // Sum of PhaseIdealLoop x _unique 1705 static volatile int _long_loop_candidates; 1706 static volatile int _long_loop_nests; 1707 static volatile int _long_loop_counted_loops; 1708 #endif 1709 1710 void rpo(Node* start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list) const; 1711 1712 void check_counted_loop_shape(IdealLoopTree* loop, Node* x, BasicType bt) NOT_DEBUG_RETURN; 1713 1714 LoopNode* create_inner_head(IdealLoopTree* loop, BaseCountedLoopNode* head, IfNode* exit_test); 1715 1716 1717 int extract_long_range_checks(const IdealLoopTree* loop, jlong stride_con, int iters_limit, PhiNode* phi, 1718 Node_List &range_checks); 1719 1720 void transform_long_range_checks(int stride_con, const Node_List &range_checks, Node* outer_phi, 1721 Node* inner_iters_actual_int, Node* inner_phi, 1722 Node* iv_add, LoopNode* inner_head); 1723 1724 Node* get_late_ctrl_with_anti_dep(LoadNode* n, Node* early, Node* LCA); 1725 1726 bool ctrl_of_use_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop, Node* ctrl); 1727 1728 bool ctrl_of_all_uses_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop); 1729 1730 Node* compute_early_ctrl(Node* n, Node* n_ctrl); 1731 1732 void try_sink_out_of_loop(Node* n); 1733 1734 Node* clamp(Node* R, Node* L, Node* H); 1735 1736 bool safe_for_if_replacement(const Node* dom) const; 1737 1738 void push_pinned_nodes_thru_region(IfNode* dom_if, Node* region); 1739 1740 bool try_merge_identical_ifs(Node* n); 1741 1742 void clone_loop_body(const Node_List& body, Node_List &old_new, CloneMap* cm); 1743 1744 void fix_body_edges(const Node_List &body, IdealLoopTree* loop, const Node_List &old_new, int dd, 1745 IdealLoopTree* parent, bool partial); 1746 1747 void fix_ctrl_uses(const Node_List& body, const IdealLoopTree* loop, Node_List &old_new, CloneLoopMode mode, 1748 Node* side_by_side_idom, CloneMap* cm, Node_List &worklist); 1749 1750 void fix_data_uses(Node_List& body, IdealLoopTree* loop, CloneLoopMode mode, IdealLoopTree* outer_loop, 1751 uint new_counter, Node_List& old_new, Node_List& worklist, Node_List*& split_if_set, 1752 Node_List*& split_bool_set, Node_List*& split_cex_set); 1753 1754 void finish_clone_loop(Node_List* split_if_set, Node_List* split_bool_set, Node_List* split_cex_set); 1755 1756 bool clone_cmp_down(Node* n, const Node* blk1, const Node* blk2); 1757 1758 void clone_loadklass_nodes_at_cmp_index(const Node* n, Node* cmp, int i); 1759 1760 bool clone_cmp_loadklass_down(Node* n, const Node* blk1, const Node* blk2); 1761 1762 bool at_relevant_ctrl(Node* n, const Node* blk1, const Node* blk2); 1763 }; 1764 1765 1766 class AutoNodeBudget : public StackObj 1767 { 1768 public: 1769 enum budget_check_t { BUDGET_CHECK, NO_BUDGET_CHECK }; 1770 1771 AutoNodeBudget(PhaseIdealLoop* phase, budget_check_t chk = BUDGET_CHECK) 1772 : _phase(phase), 1773 _check_at_final(chk == BUDGET_CHECK), 1774 _nodes_at_begin(0) 1775 { 1776 precond(_phase != NULL); 1777 1778 _nodes_at_begin = _phase->require_nodes_begin(); 1779 } 1780 1781 ~AutoNodeBudget() { 1782 #ifndef PRODUCT 1783 if (TraceLoopOpts) { 1784 uint request = _phase->nodes_required(); 1785 uint delta = _phase->C->live_nodes() - _nodes_at_begin; 1786 1787 if (request < delta) { 1788 tty->print_cr("Exceeding node budget: %d < %d", request, delta); 1789 } else { 1790 uint const REQUIRE_MIN = PhaseIdealLoop::REQUIRE_MIN; 1791 // Identify the worst estimates as "poor" ones. 1792 if (request > REQUIRE_MIN && delta > 0) { 1793 if ((delta > REQUIRE_MIN && request > 3 * delta) || 1794 (delta <= REQUIRE_MIN && request > 10 * delta)) { 1795 tty->print_cr("Poor node estimate: %d >> %d", request, delta); 1796 } 1797 } 1798 } 1799 } 1800 #endif // PRODUCT 1801 _phase->require_nodes_final(_nodes_at_begin, _check_at_final); 1802 } 1803 1804 private: 1805 PhaseIdealLoop* _phase; 1806 bool _check_at_final; 1807 uint _nodes_at_begin; 1808 }; 1809 1810 1811 // This kit may be used for making of a reserved copy of a loop before this loop 1812 // goes under non-reversible changes. 1813 // 1814 // Function create_reserve() creates a reserved copy (clone) of the loop. 1815 // The reserved copy is created by calling 1816 // PhaseIdealLoop::create_reserve_version_of_loop - see there how 1817 // the original and reserved loops are connected in the outer graph. 1818 // If create_reserve succeeded, it returns 'true' and _has_reserved is set to 'true'. 1819 // 1820 // By default the reserved copy (clone) of the loop is created as dead code - it is 1821 // dominated in the outer loop by this node chain: 1822 // intcon(1)->If->IfFalse->reserved_copy. 1823 // The original loop is dominated by the same node chain but IfTrue projection: 1824 // intcon(0)->If->IfTrue->original_loop. 1825 // 1826 // In this implementation of CountedLoopReserveKit the ctor includes create_reserve() 1827 // and the dtor, checks _use_new value. 1828 // If _use_new == false, it "switches" control to reserved copy of the loop 1829 // by simple replacing of node intcon(1) with node intcon(0). 1830 // 1831 // Here is a proposed example of usage (see also SuperWord::output in superword.cpp). 1832 // 1833 // void CountedLoopReserveKit_example() 1834 // { 1835 // CountedLoopReserveKit lrk((phase, lpt, DoReserveCopy = true); // create local object 1836 // if (DoReserveCopy && !lrk.has_reserved()) { 1837 // return; //failed to create reserved loop copy 1838 // } 1839 // ... 1840 // //something is wrong, switch to original loop 1841 /// if(something_is_wrong) return; // ~CountedLoopReserveKit makes the switch 1842 // ... 1843 // //everything worked ok, return with the newly modified loop 1844 // lrk.use_new(); 1845 // return; // ~CountedLoopReserveKit does nothing once use_new() was called 1846 // } 1847 // 1848 // Keep in mind, that by default if create_reserve() is not followed by use_new() 1849 // the dtor will "switch to the original" loop. 1850 // NOTE. You you modify outside of the original loop this class is no help. 1851 // 1852 class CountedLoopReserveKit { 1853 private: 1854 PhaseIdealLoop* _phase; 1855 IdealLoopTree* _lpt; 1856 LoopNode* _lp; 1857 IfNode* _iff; 1858 LoopNode* _lp_reserved; 1859 bool _has_reserved; 1860 bool _use_new; 1861 const bool _active; //may be set to false in ctor, then the object is dummy 1862 1863 public: 1864 CountedLoopReserveKit(PhaseIdealLoop* phase, IdealLoopTree *loop, bool active); 1865 ~CountedLoopReserveKit(); 1866 void use_new() {_use_new = true;} 1867 void set_iff(IfNode* x) {_iff = x;} 1868 bool has_reserved() const { return _active && _has_reserved;} 1869 private: 1870 bool create_reserve(); 1871 };// class CountedLoopReserveKit 1872 1873 inline Node* IdealLoopTree::tail() { 1874 // Handle lazy update of _tail field. 1875 if (_tail->in(0) == NULL) { 1876 _tail = _phase->get_ctrl(_tail); 1877 } 1878 return _tail; 1879 } 1880 1881 inline Node* IdealLoopTree::head() { 1882 // Handle lazy update of _head field. 1883 if (_head->in(0) == NULL) { 1884 _head = _phase->get_ctrl(_head); 1885 } 1886 return _head; 1887 } 1888 1889 // Iterate over the loop tree using a preorder, left-to-right traversal. 1890 // 1891 // Example that visits all counted loops from within PhaseIdealLoop 1892 // 1893 // for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { 1894 // IdealLoopTree* lpt = iter.current(); 1895 // if (!lpt->is_counted()) continue; 1896 // ... 1897 class LoopTreeIterator : public StackObj { 1898 private: 1899 IdealLoopTree* _root; 1900 IdealLoopTree* _curnt; 1901 1902 public: 1903 LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {} 1904 1905 bool done() { return _curnt == NULL; } // Finished iterating? 1906 1907 void next(); // Advance to next loop tree 1908 1909 IdealLoopTree* current() { return _curnt; } // Return current value of iterator. 1910 }; 1911 1912 // Compute probability of reaching some CFG node from a fixed 1913 // dominating CFG node 1914 class PathFrequency { 1915 private: 1916 Node* _dom; // frequencies are computed relative to this node 1917 Node_Stack _stack; 1918 GrowableArray<float> _freqs_stack; // keep track of intermediate result at regions 1919 GrowableArray<float> _freqs; // cache frequencies 1920 PhaseIdealLoop* _phase; 1921 1922 float check_and_truncate_frequency(float f) { 1923 assert(f >= 0, "Incorrect frequency"); 1924 // We do not perform an exact (f <= 1) check 1925 // this would be error prone with rounding of floats. 1926 // Performing a check like (f <= 1+eps) would be of benefit, 1927 // however, it is not evident how to determine such an eps, 1928 // given that an arbitrary number of add/mul operations 1929 // are performed on these frequencies. 1930 return (f > 1) ? 1 : f; 1931 } 1932 1933 public: 1934 PathFrequency(Node* dom, PhaseIdealLoop* phase) 1935 : _dom(dom), _stack(0), _phase(phase) { 1936 } 1937 1938 float to(Node* n); 1939 }; 1940 1941 #endif // SHARE_OPTO_LOOPNODE_HPP