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 #include "utilities/checkedCast.hpp" 34 35 class CmpNode; 36 class BaseCountedLoopEndNode; 37 class CountedLoopNode; 38 class IdealLoopTree; 39 class LoopNode; 40 class Node; 41 class OuterStripMinedLoopEndNode; 42 class PredicateBlock; 43 class PathFrequency; 44 class PhaseIdealLoop; 45 class VectorSet; 46 class Invariance; 47 struct small_cache; 48 49 // 50 // I D E A L I Z E D L O O P S 51 // 52 // Idealized loops are the set of loops I perform more interesting 53 // transformations on, beyond simple hoisting. 54 55 //------------------------------LoopNode--------------------------------------- 56 // Simple loop header. Fall in path on left, loop-back path on right. 57 class LoopNode : public RegionNode { 58 // Size is bigger to hold the flags. However, the flags do not change 59 // the semantics so it does not appear in the hash & cmp functions. 60 virtual uint size_of() const { return sizeof(*this); } 61 protected: 62 uint _loop_flags; 63 // Names for flag bitfields 64 enum { Normal=0, Pre=1, Main=2, Post=3, PreMainPostFlagsMask=3, 65 MainHasNoPreLoop = 1<<2, 66 HasExactTripCount = 1<<3, 67 InnerLoop = 1<<4, 68 PartialPeelLoop = 1<<5, 69 PartialPeelFailed = 1<<6, 70 WasSlpAnalyzed = 1<<7, 71 PassedSlpAnalysis = 1<<8, 72 DoUnrollOnly = 1<<9, 73 VectorizedLoop = 1<<10, 74 HasAtomicPostLoop = 1<<11, 75 StripMined = 1<<12, 76 SubwordLoop = 1<<13, 77 ProfileTripFailed = 1<<14, 78 LoopNestInnerLoop = 1<<15, 79 LoopNestLongOuterLoop = 1<<16, 80 FlatArrays = 1<<17}; 81 char _unswitch_count; 82 enum { _unswitch_max=3 }; 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 bool is_inner_loop() const { return _loop_flags & InnerLoop; } 92 void set_inner_loop() { _loop_flags |= InnerLoop; } 93 94 bool is_vectorized_loop() const { return _loop_flags & VectorizedLoop; } 95 bool is_partial_peel_loop() const { return _loop_flags & PartialPeelLoop; } 96 void set_partial_peel_loop() { _loop_flags |= PartialPeelLoop; } 97 bool partial_peel_has_failed() const { return _loop_flags & PartialPeelFailed; } 98 bool is_strip_mined() const { return _loop_flags & StripMined; } 99 bool is_profile_trip_failed() const { return _loop_flags & ProfileTripFailed; } 100 bool is_subword_loop() const { return _loop_flags & SubwordLoop; } 101 bool is_loop_nest_inner_loop() const { return _loop_flags & LoopNestInnerLoop; } 102 bool is_loop_nest_outer_loop() const { return _loop_flags & LoopNestLongOuterLoop; } 103 bool is_flat_arrays() const { return _loop_flags & FlatArrays; } 104 105 void mark_partial_peel_failed() { _loop_flags |= PartialPeelFailed; } 106 void mark_was_slp() { _loop_flags |= WasSlpAnalyzed; } 107 void mark_passed_slp() { _loop_flags |= PassedSlpAnalysis; } 108 void mark_do_unroll_only() { _loop_flags |= DoUnrollOnly; } 109 void mark_loop_vectorized() { _loop_flags |= VectorizedLoop; } 110 void mark_has_atomic_post_loop() { _loop_flags |= HasAtomicPostLoop; } 111 void mark_strip_mined() { _loop_flags |= StripMined; } 112 void clear_strip_mined() { _loop_flags &= ~StripMined; } 113 void mark_profile_trip_failed() { _loop_flags |= ProfileTripFailed; } 114 void mark_subword_loop() { _loop_flags |= SubwordLoop; } 115 void mark_loop_nest_inner_loop() { _loop_flags |= LoopNestInnerLoop; } 116 void mark_loop_nest_outer_loop() { _loop_flags |= LoopNestLongOuterLoop; } 117 void mark_flat_arrays() { _loop_flags |= FlatArrays; } 118 119 int unswitch_max() { return _unswitch_max; } 120 int unswitch_count() { return _unswitch_count; } 121 122 void set_unswitch_count(int val) { 123 assert (val <= unswitch_max(), "too many unswitches"); 124 _unswitch_count = val; 125 } 126 127 void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; } 128 float profile_trip_cnt() { return _profile_trip_cnt; } 129 130 LoopNode(Node *entry, Node *backedge) 131 : RegionNode(3), _loop_flags(0), _unswitch_count(0), 132 _profile_trip_cnt(COUNT_UNKNOWN) { 133 init_class_id(Class_Loop); 134 init_req(EntryControl, entry); 135 init_req(LoopBackControl, backedge); 136 } 137 138 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); 139 virtual int Opcode() const; 140 bool can_be_counted_loop(PhaseValues* phase) const { 141 return req() == 3 && in(0) != nullptr && 142 in(1) != nullptr && phase->type(in(1)) != Type::TOP && 143 in(2) != nullptr && phase->type(in(2)) != Type::TOP; 144 } 145 bool is_valid_counted_loop(BasicType bt) const; 146 #ifndef PRODUCT 147 virtual void dump_spec(outputStream *st) const; 148 #endif 149 150 void verify_strip_mined(int expect_skeleton) const NOT_DEBUG_RETURN; 151 virtual LoopNode* skip_strip_mined(int expect_skeleton = 1) { return this; } 152 virtual IfTrueNode* outer_loop_tail() const { ShouldNotReachHere(); return nullptr; } 153 virtual OuterStripMinedLoopEndNode* outer_loop_end() const { ShouldNotReachHere(); return nullptr; } 154 virtual IfFalseNode* outer_loop_exit() const { ShouldNotReachHere(); return nullptr; } 155 virtual SafePointNode* outer_safepoint() const { ShouldNotReachHere(); return nullptr; } 156 }; 157 158 //------------------------------Counted Loops---------------------------------- 159 // Counted loops are all trip-counted loops, with exactly 1 trip-counter exit 160 // path (and maybe some other exit paths). The trip-counter exit is always 161 // last in the loop. The trip-counter have to stride by a constant; 162 // the exit value is also loop invariant. 163 164 // CountedLoopNodes and CountedLoopEndNodes come in matched pairs. The 165 // CountedLoopNode has the incoming loop control and the loop-back-control 166 // which is always the IfTrue before the matching CountedLoopEndNode. The 167 // CountedLoopEndNode has an incoming control (possibly not the 168 // CountedLoopNode if there is control flow in the loop), the post-increment 169 // trip-counter value, and the limit. The trip-counter value is always of 170 // the form (Op old-trip-counter stride). The old-trip-counter is produced 171 // by a Phi connected to the CountedLoopNode. The stride is constant. 172 // The Op is any commutable opcode, including Add, Mul, Xor. The 173 // CountedLoopEndNode also takes in the loop-invariant limit value. 174 175 // From a CountedLoopNode I can reach the matching CountedLoopEndNode via the 176 // loop-back control. From CountedLoopEndNodes I can reach CountedLoopNodes 177 // via the old-trip-counter from the Op node. 178 179 //------------------------------CountedLoopNode-------------------------------- 180 // CountedLoopNodes head simple counted loops. CountedLoopNodes have as 181 // inputs the incoming loop-start control and the loop-back control, so they 182 // act like RegionNodes. They also take in the initial trip counter, the 183 // loop-invariant stride and the loop-invariant limit value. CountedLoopNodes 184 // produce a loop-body control and the trip counter value. Since 185 // CountedLoopNodes behave like RegionNodes I still have a standard CFG model. 186 187 class BaseCountedLoopNode : public LoopNode { 188 public: 189 BaseCountedLoopNode(Node *entry, Node *backedge) 190 : LoopNode(entry, backedge) { 191 } 192 193 Node *init_control() const { return in(EntryControl); } 194 Node *back_control() const { return in(LoopBackControl); } 195 196 Node* init_trip() const; 197 Node* stride() const; 198 bool stride_is_con() const; 199 Node* limit() const; 200 Node* incr() const; 201 Node* phi() const; 202 203 BaseCountedLoopEndNode* loopexit_or_null() const; 204 BaseCountedLoopEndNode* loopexit() const; 205 206 virtual BasicType bt() const = 0; 207 208 jlong stride_con() const; 209 210 static BaseCountedLoopNode* make(Node* entry, Node* backedge, BasicType bt); 211 }; 212 213 214 class CountedLoopNode : public BaseCountedLoopNode { 215 // Size is bigger to hold _main_idx. However, _main_idx does not change 216 // the semantics so it does not appear in the hash & cmp functions. 217 virtual uint size_of() const { return sizeof(*this); } 218 219 // For Pre- and Post-loops during debugging ONLY, this holds the index of 220 // the Main CountedLoop. Used to assert that we understand the graph shape. 221 node_idx_t _main_idx; 222 223 // Known trip count calculated by compute_exact_trip_count() 224 uint _trip_count; 225 226 // Log2 of original loop bodies in unrolled loop 227 int _unrolled_count_log2; 228 229 // Node count prior to last unrolling - used to decide if 230 // unroll,optimize,unroll,optimize,... is making progress 231 int _node_count_before_unroll; 232 233 // If slp analysis is performed we record the maximum 234 // vector mapped unroll factor here 235 int _slp_maximum_unroll_factor; 236 237 // Cached CountedLoopEndNode of pre loop for main loops 238 CountedLoopEndNode* _pre_loop_end; 239 240 public: 241 CountedLoopNode(Node *entry, Node *backedge) 242 : BaseCountedLoopNode(entry, backedge), _main_idx(0), _trip_count(max_juint), 243 _unrolled_count_log2(0), _node_count_before_unroll(0), 244 _slp_maximum_unroll_factor(0), _pre_loop_end(nullptr) { 245 init_class_id(Class_CountedLoop); 246 // Initialize _trip_count to the largest possible value. 247 // Will be reset (lower) if the loop's trip count is known. 248 } 249 250 virtual int Opcode() const; 251 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); 252 253 CountedLoopEndNode* loopexit_or_null() const { return (CountedLoopEndNode*) BaseCountedLoopNode::loopexit_or_null(); } 254 CountedLoopEndNode* loopexit() const { return (CountedLoopEndNode*) BaseCountedLoopNode::loopexit(); } 255 int stride_con() const; 256 257 // Match increment with optional truncation 258 static Node* 259 match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2, const TypeInteger** trunc_type, 260 BasicType bt); 261 262 // A 'main' loop has a pre-loop and a post-loop. The 'main' loop 263 // can run short a few iterations and may start a few iterations in. 264 // It will be RCE'd and unrolled and aligned. 265 266 // A following 'post' loop will run any remaining iterations. Used 267 // during Range Check Elimination, the 'post' loop will do any final 268 // iterations with full checks. Also used by Loop Unrolling, where 269 // the 'post' loop will do any epilog iterations needed. Basically, 270 // a 'post' loop can not profitably be further unrolled or RCE'd. 271 272 // A preceding 'pre' loop will run at least 1 iteration (to do peeling), 273 // it may do under-flow checks for RCE and may do alignment iterations 274 // so the following main loop 'knows' that it is striding down cache 275 // lines. 276 277 // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or 278 // Aligned, may be missing it's pre-loop. 279 bool is_normal_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Normal; } 280 bool is_pre_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Pre; } 281 bool is_main_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Main; } 282 bool is_post_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Post; } 283 bool was_slp_analyzed () const { return (_loop_flags&WasSlpAnalyzed) == WasSlpAnalyzed; } 284 bool has_passed_slp () const { return (_loop_flags&PassedSlpAnalysis) == PassedSlpAnalysis; } 285 bool is_unroll_only () const { return (_loop_flags&DoUnrollOnly) == DoUnrollOnly; } 286 bool is_main_no_pre_loop() const { return _loop_flags & MainHasNoPreLoop; } 287 bool has_atomic_post_loop () const { return (_loop_flags & HasAtomicPostLoop) == HasAtomicPostLoop; } 288 void set_main_no_pre_loop() { _loop_flags |= MainHasNoPreLoop; } 289 290 int main_idx() const { return _main_idx; } 291 292 293 void set_pre_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; } 294 void set_main_loop ( ) { assert(is_normal_loop(),""); _loop_flags |= Main; } 295 void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; } 296 void set_normal_loop( ) { _loop_flags &= ~PreMainPostFlagsMask; } 297 298 void set_trip_count(uint tc) { _trip_count = tc; } 299 uint trip_count() { return _trip_count; } 300 301 bool has_exact_trip_count() const { return (_loop_flags & HasExactTripCount) != 0; } 302 void set_exact_trip_count(uint tc) { 303 _trip_count = tc; 304 _loop_flags |= HasExactTripCount; 305 } 306 void set_nonexact_trip_count() { 307 _loop_flags &= ~HasExactTripCount; 308 } 309 void set_notpassed_slp() { 310 _loop_flags &= ~PassedSlpAnalysis; 311 } 312 313 void double_unrolled_count() { _unrolled_count_log2++; } 314 int unrolled_count() { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); } 315 316 void set_node_count_before_unroll(int ct) { _node_count_before_unroll = ct; } 317 int node_count_before_unroll() { return _node_count_before_unroll; } 318 void set_slp_max_unroll(int unroll_factor) { _slp_maximum_unroll_factor = unroll_factor; } 319 int slp_max_unroll() const { return _slp_maximum_unroll_factor; } 320 321 virtual LoopNode* skip_strip_mined(int expect_skeleton = 1); 322 OuterStripMinedLoopNode* outer_loop() const; 323 virtual IfTrueNode* outer_loop_tail() const; 324 virtual OuterStripMinedLoopEndNode* outer_loop_end() const; 325 virtual IfFalseNode* outer_loop_exit() const; 326 virtual SafePointNode* outer_safepoint() const; 327 328 Node* skip_assertion_predicates_with_halt(); 329 330 virtual BasicType bt() const { 331 return T_INT; 332 } 333 334 Node* is_canonical_loop_entry(); 335 CountedLoopEndNode* find_pre_loop_end(); 336 CountedLoopNode* pre_loop_head() const; 337 CountedLoopEndNode* pre_loop_end(); 338 void set_pre_loop_end(CountedLoopEndNode* pre_loop_end); 339 340 #ifndef PRODUCT 341 virtual void dump_spec(outputStream *st) const; 342 #endif 343 }; 344 345 class LongCountedLoopNode : public BaseCountedLoopNode { 346 public: 347 LongCountedLoopNode(Node *entry, Node *backedge) 348 : BaseCountedLoopNode(entry, backedge) { 349 init_class_id(Class_LongCountedLoop); 350 } 351 352 virtual int Opcode() const; 353 354 virtual BasicType bt() const { 355 return T_LONG; 356 } 357 358 LongCountedLoopEndNode* loopexit_or_null() const { return (LongCountedLoopEndNode*) BaseCountedLoopNode::loopexit_or_null(); } 359 LongCountedLoopEndNode* loopexit() const { return (LongCountedLoopEndNode*) BaseCountedLoopNode::loopexit(); } 360 }; 361 362 363 //------------------------------CountedLoopEndNode----------------------------- 364 // CountedLoopEndNodes end simple trip counted loops. They act much like 365 // IfNodes. 366 367 class BaseCountedLoopEndNode : public IfNode { 368 public: 369 enum { TestControl, TestValue }; 370 BaseCountedLoopEndNode(Node *control, Node *test, float prob, float cnt) 371 : IfNode(control, test, prob, cnt) { 372 init_class_id(Class_BaseCountedLoopEnd); 373 } 374 375 Node *cmp_node() const { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : nullptr; } 376 Node* incr() const { Node* tmp = cmp_node(); return (tmp && tmp->req() == 3) ? tmp->in(1) : nullptr; } 377 Node* limit() const { Node* tmp = cmp_node(); return (tmp && tmp->req() == 3) ? tmp->in(2) : nullptr; } 378 Node* stride() const { Node* tmp = incr(); return (tmp && tmp->req() == 3) ? tmp->in(2) : nullptr; } 379 Node* init_trip() const { Node* tmp = phi(); return (tmp && tmp->req() == 3) ? tmp->in(1) : nullptr; } 380 bool stride_is_con() const { Node *tmp = stride(); return (tmp != nullptr && tmp->is_Con()); } 381 382 PhiNode* phi() const { 383 Node* tmp = incr(); 384 if (tmp && tmp->req() == 3) { 385 Node* phi = tmp->in(1); 386 if (phi->is_Phi()) { 387 return phi->as_Phi(); 388 } 389 } 390 return nullptr; 391 } 392 393 BaseCountedLoopNode* loopnode() const { 394 // The CountedLoopNode that goes with this CountedLoopEndNode may 395 // have been optimized out by the IGVN so be cautious with the 396 // pattern matching on the graph 397 PhiNode* iv_phi = phi(); 398 if (iv_phi == nullptr) { 399 return nullptr; 400 } 401 Node* ln = iv_phi->in(0); 402 if (!ln->is_BaseCountedLoop() || ln->as_BaseCountedLoop()->loopexit_or_null() != this) { 403 return nullptr; 404 } 405 if (ln->as_BaseCountedLoop()->bt() != bt()) { 406 return nullptr; 407 } 408 return ln->as_BaseCountedLoop(); 409 } 410 411 BoolTest::mask test_trip() const { return in(TestValue)->as_Bool()->_test._test; } 412 413 jlong stride_con() const; 414 virtual BasicType bt() const = 0; 415 416 static BaseCountedLoopEndNode* make(Node* control, Node* test, float prob, float cnt, BasicType bt); 417 }; 418 419 class CountedLoopEndNode : public BaseCountedLoopEndNode { 420 public: 421 422 CountedLoopEndNode(Node *control, Node *test, float prob, float cnt) 423 : BaseCountedLoopEndNode(control, test, prob, cnt) { 424 init_class_id(Class_CountedLoopEnd); 425 } 426 virtual int Opcode() const; 427 428 CountedLoopNode* loopnode() const { 429 return (CountedLoopNode*) BaseCountedLoopEndNode::loopnode(); 430 } 431 432 virtual BasicType bt() const { 433 return T_INT; 434 } 435 436 #ifndef PRODUCT 437 virtual void dump_spec(outputStream *st) const; 438 #endif 439 }; 440 441 class LongCountedLoopEndNode : public BaseCountedLoopEndNode { 442 public: 443 LongCountedLoopEndNode(Node *control, Node *test, float prob, float cnt) 444 : BaseCountedLoopEndNode(control, test, prob, cnt) { 445 init_class_id(Class_LongCountedLoopEnd); 446 } 447 448 LongCountedLoopNode* loopnode() const { 449 return (LongCountedLoopNode*) BaseCountedLoopEndNode::loopnode(); 450 } 451 452 virtual int Opcode() const; 453 454 virtual BasicType bt() const { 455 return T_LONG; 456 } 457 }; 458 459 460 inline BaseCountedLoopEndNode* BaseCountedLoopNode::loopexit_or_null() const { 461 Node* bctrl = back_control(); 462 if (bctrl == nullptr) return nullptr; 463 464 Node* lexit = bctrl->in(0); 465 if (!lexit->is_BaseCountedLoopEnd()) { 466 return nullptr; 467 } 468 BaseCountedLoopEndNode* result = lexit->as_BaseCountedLoopEnd(); 469 if (result->bt() != bt()) { 470 return nullptr; 471 } 472 return result; 473 } 474 475 inline BaseCountedLoopEndNode* BaseCountedLoopNode::loopexit() const { 476 BaseCountedLoopEndNode* cle = loopexit_or_null(); 477 assert(cle != nullptr, "loopexit is null"); 478 return cle; 479 } 480 481 inline Node* BaseCountedLoopNode::init_trip() const { 482 BaseCountedLoopEndNode* cle = loopexit_or_null(); 483 return cle != nullptr ? cle->init_trip() : nullptr; 484 } 485 inline Node* BaseCountedLoopNode::stride() const { 486 BaseCountedLoopEndNode* cle = loopexit_or_null(); 487 return cle != nullptr ? cle->stride() : nullptr; 488 } 489 490 inline bool BaseCountedLoopNode::stride_is_con() const { 491 BaseCountedLoopEndNode* cle = loopexit_or_null(); 492 return cle != nullptr && cle->stride_is_con(); 493 } 494 inline Node* BaseCountedLoopNode::limit() const { 495 BaseCountedLoopEndNode* cle = loopexit_or_null(); 496 return cle != nullptr ? cle->limit() : nullptr; 497 } 498 inline Node* BaseCountedLoopNode::incr() const { 499 BaseCountedLoopEndNode* cle = loopexit_or_null(); 500 return cle != nullptr ? cle->incr() : nullptr; 501 } 502 inline Node* BaseCountedLoopNode::phi() const { 503 BaseCountedLoopEndNode* cle = loopexit_or_null(); 504 return cle != nullptr ? cle->phi() : nullptr; 505 } 506 507 inline jlong BaseCountedLoopNode::stride_con() const { 508 BaseCountedLoopEndNode* cle = loopexit_or_null(); 509 return cle != nullptr ? cle->stride_con() : 0; 510 } 511 512 513 //------------------------------LoopLimitNode----------------------------- 514 // Counted Loop limit node which represents exact final iterator value: 515 // trip_count = (limit - init_trip + stride - 1)/stride 516 // final_value= trip_count * stride + init_trip. 517 // Use HW instructions to calculate it when it can overflow in integer. 518 // Note, final_value should fit into integer since counted loop has 519 // limit check: limit <= max_int-stride. 520 class LoopLimitNode : public Node { 521 enum { Init=1, Limit=2, Stride=3 }; 522 public: 523 LoopLimitNode( Compile* C, Node *init, Node *limit, Node *stride ) : Node(0,init,limit,stride) { 524 // Put it on the Macro nodes list to optimize during macro nodes expansion. 525 init_flags(Flag_is_macro); 526 C->add_macro_node(this); 527 } 528 virtual int Opcode() const; 529 virtual const Type *bottom_type() const { return TypeInt::INT; } 530 virtual uint ideal_reg() const { return Op_RegI; } 531 virtual const Type* Value(PhaseGVN* phase) const; 532 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); 533 virtual Node* Identity(PhaseGVN* phase); 534 }; 535 536 // Support for strip mining 537 class OuterStripMinedLoopNode : public LoopNode { 538 private: 539 static void fix_sunk_stores(CountedLoopEndNode* inner_cle, LoopNode* inner_cl, PhaseIterGVN* igvn, PhaseIdealLoop* iloop); 540 541 public: 542 OuterStripMinedLoopNode(Compile* C, Node *entry, Node *backedge) 543 : LoopNode(entry, backedge) { 544 init_class_id(Class_OuterStripMinedLoop); 545 init_flags(Flag_is_macro); 546 C->add_macro_node(this); 547 } 548 549 virtual int Opcode() const; 550 551 virtual IfTrueNode* outer_loop_tail() const; 552 virtual OuterStripMinedLoopEndNode* outer_loop_end() const; 553 virtual IfFalseNode* outer_loop_exit() const; 554 virtual SafePointNode* outer_safepoint() const; 555 void adjust_strip_mined_loop(PhaseIterGVN* igvn); 556 557 void remove_outer_loop_and_safepoint(PhaseIterGVN* igvn) const; 558 559 void transform_to_counted_loop(PhaseIterGVN* igvn, PhaseIdealLoop* iloop); 560 561 static Node* register_new_node(Node* node, LoopNode* ctrl, PhaseIterGVN* igvn, PhaseIdealLoop* iloop); 562 563 Node* register_control(Node* node, Node* loop, Node* idom, PhaseIterGVN* igvn, 564 PhaseIdealLoop* iloop); 565 }; 566 567 class OuterStripMinedLoopEndNode : public IfNode { 568 public: 569 OuterStripMinedLoopEndNode(Node *control, Node *test, float prob, float cnt) 570 : IfNode(control, test, prob, cnt) { 571 init_class_id(Class_OuterStripMinedLoopEnd); 572 } 573 574 virtual int Opcode() const; 575 576 virtual const Type* Value(PhaseGVN* phase) const; 577 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); 578 579 bool is_expanded(PhaseGVN *phase) const; 580 }; 581 582 // -----------------------------IdealLoopTree---------------------------------- 583 class IdealLoopTree : public ResourceObj { 584 public: 585 IdealLoopTree *_parent; // Parent in loop tree 586 IdealLoopTree *_next; // Next sibling in loop tree 587 IdealLoopTree *_child; // First child in loop tree 588 589 // The head-tail backedge defines the loop. 590 // If a loop has multiple backedges, this is addressed during cleanup where 591 // we peel off the multiple backedges, merging all edges at the bottom and 592 // ensuring that one proper backedge flow into the loop. 593 Node *_head; // Head of loop 594 Node *_tail; // Tail of loop 595 inline Node *tail(); // Handle lazy update of _tail field 596 inline Node *head(); // Handle lazy update of _head field 597 PhaseIdealLoop* _phase; 598 int _local_loop_unroll_limit; 599 int _local_loop_unroll_factor; 600 601 Node_List _body; // Loop body for inner loops 602 603 uint16_t _nest; // Nesting depth 604 uint8_t _irreducible:1, // True if irreducible 605 _has_call:1, // True if has call safepoint 606 _has_sfpt:1, // True if has non-call safepoint 607 _rce_candidate:1, // True if candidate for range check elimination 608 _has_range_checks:1, 609 _has_range_checks_computed:1; 610 611 Node_List* _safepts; // List of safepoints in this loop 612 Node_List* _required_safept; // A inner loop cannot delete these safepts; 613 bool _allow_optimizations; // Allow loop optimizations 614 615 IdealLoopTree( PhaseIdealLoop* phase, Node *head, Node *tail ) 616 : _parent(0), _next(0), _child(0), 617 _head(head), _tail(tail), 618 _phase(phase), 619 _local_loop_unroll_limit(0), _local_loop_unroll_factor(0), 620 _nest(0), _irreducible(0), _has_call(0), _has_sfpt(0), _rce_candidate(0), 621 _has_range_checks(0), _has_range_checks_computed(0), 622 _safepts(nullptr), 623 _required_safept(nullptr), 624 _allow_optimizations(true) 625 { 626 precond(_head != nullptr); 627 precond(_tail != nullptr); 628 } 629 630 // Is 'l' a member of 'this'? 631 bool is_member(const IdealLoopTree *l) const; // Test for nested membership 632 633 // Set loop nesting depth. Accumulate has_call bits. 634 int set_nest( uint depth ); 635 636 // Split out multiple fall-in edges from the loop header. Move them to a 637 // private RegionNode before the loop. This becomes the loop landing pad. 638 void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ); 639 640 // Split out the outermost loop from this shared header. 641 void split_outer_loop( PhaseIdealLoop *phase ); 642 643 // Merge all the backedges from the shared header into a private Region. 644 // Feed that region as the one backedge to this loop. 645 void merge_many_backedges( PhaseIdealLoop *phase ); 646 647 // Split shared headers and insert loop landing pads. 648 // Insert a LoopNode to replace the RegionNode. 649 // Returns TRUE if loop tree is structurally changed. 650 bool beautify_loops( PhaseIdealLoop *phase ); 651 652 // Perform optimization to use the loop predicates for null checks and range checks. 653 // Applies to any loop level (not just the innermost one) 654 bool loop_predication( PhaseIdealLoop *phase); 655 bool can_apply_loop_predication(); 656 657 // Perform iteration-splitting on inner loops. Split iterations to 658 // avoid range checks or one-shot null checks. Returns false if the 659 // current round of loop opts should stop. 660 bool iteration_split( PhaseIdealLoop *phase, Node_List &old_new ); 661 662 // Driver for various flavors of iteration splitting. Returns false 663 // if the current round of loop opts should stop. 664 bool iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ); 665 666 // Given dominators, try to find loops with calls that must always be 667 // executed (call dominates loop tail). These loops do not need non-call 668 // safepoints (ncsfpt). 669 void check_safepts(VectorSet &visited, Node_List &stack); 670 671 // Allpaths backwards scan from loop tail, terminating each path at first safepoint 672 // encountered. 673 void allpaths_check_safepts(VectorSet &visited, Node_List &stack); 674 675 // Remove safepoints from loop. Optionally keeping one. 676 void remove_safepoints(PhaseIdealLoop* phase, bool keep_one); 677 678 // Convert to counted loops where possible 679 void counted_loop( PhaseIdealLoop *phase ); 680 681 // Check for Node being a loop-breaking test 682 Node *is_loop_exit(Node *iff) const; 683 684 // Remove simplistic dead code from loop body 685 void DCE_loop_body(); 686 687 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage. 688 // Replace with a 1-in-10 exit guess. 689 void adjust_loop_exit_prob( PhaseIdealLoop *phase ); 690 691 // Return TRUE or FALSE if the loop should never be RCE'd or aligned. 692 // Useful for unrolling loops with NO array accesses. 693 bool policy_peel_only( PhaseIdealLoop *phase ) const; 694 695 // Return TRUE or FALSE if the loop should be unswitched -- clone 696 // loop with an invariant test 697 bool policy_unswitching( PhaseIdealLoop *phase ) const; 698 699 // Micro-benchmark spamming. Remove empty loops. 700 bool do_remove_empty_loop( PhaseIdealLoop *phase ); 701 702 // Convert one iteration loop into normal code. 703 bool do_one_iteration_loop( PhaseIdealLoop *phase ); 704 705 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can 706 // move some loop-invariant test (usually a null-check) before the loop. 707 bool policy_peeling(PhaseIdealLoop *phase); 708 709 uint estimate_peeling(PhaseIdealLoop *phase); 710 711 // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any 712 // known trip count in the counted loop node. 713 bool policy_maximally_unroll(PhaseIdealLoop *phase) const; 714 715 // Return TRUE or FALSE if the loop should be unrolled or not. Apply unroll 716 // if the loop is a counted loop and the loop body is small enough. 717 bool policy_unroll(PhaseIdealLoop *phase); 718 719 // Loop analyses to map to a maximal superword unrolling for vectorization. 720 void policy_unroll_slp_analysis(CountedLoopNode *cl, PhaseIdealLoop *phase, int future_unroll_ct); 721 722 // Return TRUE or FALSE if the loop should be range-check-eliminated. 723 // Gather a list of IF tests that are dominated by iteration splitting; 724 // also gather the end of the first split and the start of the 2nd split. 725 bool policy_range_check(PhaseIdealLoop* phase, bool provisional, BasicType bt) const; 726 727 // Return TRUE if "iff" is a range check. 728 bool is_range_check_if(IfProjNode* if_success_proj, PhaseIdealLoop* phase, Invariance& invar DEBUG_ONLY(COMMA ProjNode* predicate_proj)) const; 729 bool is_range_check_if(IfProjNode* if_success_proj, PhaseIdealLoop* phase, BasicType bt, Node* iv, Node*& range, Node*& offset, 730 jlong& scale) const; 731 732 // Estimate the number of nodes required when cloning a loop (body). 733 uint est_loop_clone_sz(uint factor) const; 734 // Estimate the number of nodes required when unrolling a loop (body). 735 uint est_loop_unroll_sz(uint factor) const; 736 737 // Compute loop trip count if possible 738 void compute_trip_count(PhaseIdealLoop* phase); 739 740 // Compute loop trip count from profile data 741 float compute_profile_trip_cnt_helper(Node* n); 742 void compute_profile_trip_cnt( PhaseIdealLoop *phase ); 743 744 // Reassociate invariant expressions. 745 void reassociate_invariants(PhaseIdealLoop *phase); 746 // Reassociate invariant binary expressions. 747 Node* reassociate(Node* n1, PhaseIdealLoop *phase); 748 // Reassociate invariant add and subtract expressions. 749 Node* reassociate_add_sub(Node* n1, int inv1_idx, int inv2_idx, PhaseIdealLoop *phase); 750 // Return nonzero index of invariant operand if invariant and variant 751 // are combined with an associative binary. Helper for reassociate_invariants. 752 int find_invariant(Node* n, PhaseIdealLoop *phase); 753 // Return TRUE if "n" is associative. 754 bool is_associative(Node* n, Node* base=nullptr); 755 756 // Return true if n is invariant 757 bool is_invariant(Node* n) const; 758 759 // Put loop body on igvn work list 760 void record_for_igvn(); 761 762 bool is_root() { return _parent == nullptr; } 763 // A proper/reducible loop w/o any (occasional) dead back-edge. 764 bool is_loop() { return !_irreducible && !tail()->is_top(); } 765 bool is_counted() { return is_loop() && _head->is_CountedLoop(); } 766 bool is_innermost() { return is_loop() && _child == nullptr; } 767 768 void remove_main_post_loops(CountedLoopNode *cl, PhaseIdealLoop *phase); 769 770 bool compute_has_range_checks() const; 771 bool range_checks_present() { 772 if (!_has_range_checks_computed) { 773 if (compute_has_range_checks()) { 774 _has_range_checks = 1; 775 } 776 _has_range_checks_computed = 1; 777 } 778 return _has_range_checks; 779 } 780 781 #ifndef PRODUCT 782 void dump_head(); // Dump loop head only 783 void dump(); // Dump this loop recursively 784 #endif 785 786 #ifdef ASSERT 787 GrowableArray<IdealLoopTree*> collect_sorted_children() const; 788 bool verify_tree(IdealLoopTree* loop_verify) const; 789 #endif 790 791 private: 792 enum { EMPTY_LOOP_SIZE = 7 }; // Number of nodes in an empty loop. 793 794 // Estimate the number of nodes resulting from control and data flow merge. 795 uint est_loop_flow_merge_sz() const; 796 797 // Check if the number of residual iterations is large with unroll_cnt. 798 // Return true if the residual iterations are more than 10% of the trip count. 799 bool is_residual_iters_large(int unroll_cnt, CountedLoopNode *cl) const { 800 return (unroll_cnt - 1) * (100.0 / LoopPercentProfileLimit) > cl->profile_trip_cnt(); 801 } 802 803 void collect_loop_core_nodes(PhaseIdealLoop* phase, Unique_Node_List& wq) const; 804 805 bool empty_loop_with_data_nodes(PhaseIdealLoop* phase) const; 806 807 void enqueue_data_nodes(PhaseIdealLoop* phase, Unique_Node_List& empty_loop_nodes, Unique_Node_List& wq) const; 808 809 bool process_safepoint(PhaseIdealLoop* phase, Unique_Node_List& empty_loop_nodes, Unique_Node_List& wq, 810 Node* sfpt) const; 811 812 bool empty_loop_candidate(PhaseIdealLoop* phase) const; 813 814 bool empty_loop_with_extra_nodes_candidate(PhaseIdealLoop* phase) const; 815 }; 816 817 // -----------------------------PhaseIdealLoop--------------------------------- 818 // Computes the mapping from Nodes to IdealLoopTrees. Organizes IdealLoopTrees 819 // into a loop tree. Drives the loop-based transformations on the ideal graph. 820 class PhaseIdealLoop : public PhaseTransform { 821 friend class IdealLoopTree; 822 friend class SuperWord; 823 friend class ShenandoahBarrierC2Support; 824 friend class AutoNodeBudget; 825 826 // Map loop membership for CFG nodes, and ctrl for non-CFG nodes. 827 Node_List _loop_or_ctrl; 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(const Node* n) const { return ((intptr_t)_loop_or_ctrl[n->_idx]) & 1; } 891 892 private: 893 // clear out dead code after build_loop_late 894 Node_List _deadlist; 895 Node_List _zero_trip_guard_opaque_nodes; 896 897 // Support for faster execution of get_late_ctrl()/dom_lca() 898 // when a node has many uses and dominator depth is deep. 899 GrowableArray<jlong> _dom_lca_tags; 900 uint _dom_lca_tags_round; 901 void init_dom_lca_tags(); 902 903 // Helper for debugging bad dominance relationships 904 bool verify_dominance(Node* n, Node* use, Node* LCA, Node* early); 905 906 Node* compute_lca_of_uses(Node* n, Node* early, bool verify = false); 907 908 // Inline wrapper for frequent cases: 909 // 1) only one use 910 // 2) a use is the same as the current LCA passed as 'n1' 911 Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) { 912 assert( n->is_CFG(), "" ); 913 // Fast-path null lca 914 if( lca != nullptr && lca != n ) { 915 assert( lca->is_CFG(), "" ); 916 // find LCA of all uses 917 n = dom_lca_for_get_late_ctrl_internal( lca, n, tag ); 918 } 919 return find_non_split_ctrl(n); 920 } 921 Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag ); 922 923 // Helper function for directing control inputs away from CFG split points. 924 Node *find_non_split_ctrl( Node *ctrl ) const { 925 if (ctrl != nullptr) { 926 if (ctrl->is_MultiBranch()) { 927 ctrl = ctrl->in(0); 928 } 929 assert(ctrl->is_CFG(), "CFG"); 930 } 931 return ctrl; 932 } 933 934 Node* cast_incr_before_loop(Node* incr, Node* ctrl, Node* loop); 935 936 #ifdef ASSERT 937 void ensure_zero_trip_guard_proj(Node* node, bool is_main_loop); 938 #endif 939 void copy_assertion_predicates_to_main_loop_helper(const PredicateBlock* predicate_block, Node* init, Node* stride, 940 IdealLoopTree* outer_loop, LoopNode* outer_main_head, 941 uint dd_main_head, uint idx_before_pre_post, 942 uint idx_after_post_before_pre, Node* zero_trip_guard_proj_main, 943 Node* zero_trip_guard_proj_post, const Node_List &old_new); 944 void copy_assertion_predicates_to_main_loop(CountedLoopNode* pre_head, Node* init, Node* stride, IdealLoopTree* outer_loop, 945 LoopNode* outer_main_head, uint dd_main_head, uint idx_before_pre_post, 946 uint idx_after_post_before_pre, Node* zero_trip_guard_proj_main, 947 Node* zero_trip_guard_proj_post, const Node_List& old_new); 948 Node* clone_assertion_predicate_and_initialize(Node* iff, Node* new_init, Node* new_stride, Node* predicate, 949 Node* uncommon_proj, Node* control, IdealLoopTree* outer_loop, 950 Node* input_proj); 951 static void count_opaque_loop_nodes(Node* n, uint& init, uint& stride); 952 static bool subgraph_has_opaque(Node* n); 953 Node* create_bool_from_template_assertion_predicate(Node* template_assertion_predicate, Node* new_init, Node* new_stride, 954 Node* control); 955 static bool assertion_predicate_has_loop_opaque_node(IfNode* iff); 956 static void get_assertion_predicates(Node* predicate, Unique_Node_List& list, bool get_opaque = false); 957 void update_main_loop_assertion_predicates(Node* ctrl, CountedLoopNode* loop_head, Node* init, int stride_con); 958 void copy_assertion_predicates_to_post_loop(LoopNode* main_loop_head, CountedLoopNode* post_loop_head, Node* init, 959 Node* stride); 960 void initialize_assertion_predicates_for_peeled_loop(const PredicateBlock* predicate_block, LoopNode* outer_loop_head, 961 int dd_outer_loop_head, Node* init, Node* stride, 962 IdealLoopTree* outer_loop, uint idx_before_clone, 963 const Node_List& old_new); 964 void insert_loop_limit_check_predicate(ParsePredicateSuccessProj* loop_limit_check_parse_proj, Node* cmp_limit, 965 Node* bol); 966 #ifdef ASSERT 967 bool only_has_infinite_loops(); 968 #endif 969 970 void log_loop_tree(); 971 972 public: 973 974 PhaseIterGVN &igvn() const { return _igvn; } 975 976 bool has_node(const Node* n) const { 977 guarantee(n != nullptr, "No Node."); 978 return _loop_or_ctrl[n->_idx] != nullptr; 979 } 980 // check if transform created new nodes that need _ctrl recorded 981 Node *get_late_ctrl( Node *n, Node *early ); 982 Node *get_early_ctrl( Node *n ); 983 Node *get_early_ctrl_for_expensive(Node *n, Node* earliest); 984 void set_early_ctrl(Node* n, bool update_body); 985 void set_subtree_ctrl(Node* n, bool update_body); 986 void set_ctrl( Node *n, Node *ctrl ) { 987 assert( !has_node(n) || has_ctrl(n), "" ); 988 assert( ctrl->in(0), "cannot set dead control node" ); 989 assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" ); 990 _loop_or_ctrl.map(n->_idx, (Node*)((intptr_t)ctrl + 1)); 991 } 992 // Set control and update loop membership 993 void set_ctrl_and_loop(Node* n, Node* ctrl) { 994 IdealLoopTree* old_loop = get_loop(get_ctrl(n)); 995 IdealLoopTree* new_loop = get_loop(ctrl); 996 if (old_loop != new_loop) { 997 if (old_loop->_child == nullptr) old_loop->_body.yank(n); 998 if (new_loop->_child == nullptr) new_loop->_body.push(n); 999 } 1000 set_ctrl(n, ctrl); 1001 } 1002 // Control nodes can be replaced or subsumed. During this pass they 1003 // get their replacement Node in slot 1. Instead of updating the block 1004 // location of all Nodes in the subsumed block, we lazily do it. As we 1005 // pull such a subsumed block out of the array, we write back the final 1006 // correct block. 1007 Node* get_ctrl(const Node* i) { 1008 assert(has_node(i), ""); 1009 Node *n = get_ctrl_no_update(i); 1010 _loop_or_ctrl.map(i->_idx, (Node*)((intptr_t)n + 1)); 1011 assert(has_node(i) && has_ctrl(i), ""); 1012 assert(n == find_non_split_ctrl(n), "must return legal ctrl" ); 1013 return n; 1014 } 1015 // true if CFG node d dominates CFG node n 1016 bool is_dominator(Node *d, Node *n); 1017 // return get_ctrl for a data node and self(n) for a CFG node 1018 Node* ctrl_or_self(Node* n) { 1019 if (has_ctrl(n)) 1020 return get_ctrl(n); 1021 else { 1022 assert (n->is_CFG(), "must be a CFG node"); 1023 return n; 1024 } 1025 } 1026 1027 Node* get_ctrl_no_update_helper(const Node* i) const { 1028 assert(has_ctrl(i), "should be control, not loop"); 1029 return (Node*)(((intptr_t)_loop_or_ctrl[i->_idx]) & ~1); 1030 } 1031 1032 Node* get_ctrl_no_update(const Node* i) const { 1033 assert( has_ctrl(i), "" ); 1034 Node *n = get_ctrl_no_update_helper(i); 1035 if (!n->in(0)) { 1036 // Skip dead CFG nodes 1037 do { 1038 n = get_ctrl_no_update_helper(n); 1039 } while (!n->in(0)); 1040 n = find_non_split_ctrl(n); 1041 } 1042 return n; 1043 } 1044 1045 // Check for loop being set 1046 // "n" must be a control node. Returns true if "n" is known to be in a loop. 1047 bool has_loop( Node *n ) const { 1048 assert(!has_node(n) || !has_ctrl(n), ""); 1049 return has_node(n); 1050 } 1051 // Set loop 1052 void set_loop( Node *n, IdealLoopTree *loop ) { 1053 _loop_or_ctrl.map(n->_idx, (Node*)loop); 1054 } 1055 // Lazy-dazy update of 'get_ctrl' and 'idom_at' mechanisms. Replace 1056 // the 'old_node' with 'new_node'. Kill old-node. Add a reference 1057 // from old_node to new_node to support the lazy update. Reference 1058 // replaces loop reference, since that is not needed for dead node. 1059 void lazy_update(Node *old_node, Node *new_node) { 1060 assert(old_node != new_node, "no cycles please"); 1061 // Re-use the side array slot for this node to provide the 1062 // forwarding pointer. 1063 _loop_or_ctrl.map(old_node->_idx, (Node*)((intptr_t)new_node + 1)); 1064 } 1065 void lazy_replace(Node *old_node, Node *new_node) { 1066 _igvn.replace_node(old_node, new_node); 1067 lazy_update(old_node, new_node); 1068 } 1069 1070 private: 1071 1072 // Place 'n' in some loop nest, where 'n' is a CFG node 1073 void build_loop_tree(); 1074 int build_loop_tree_impl( Node *n, int pre_order ); 1075 // Insert loop into the existing loop tree. 'innermost' is a leaf of the 1076 // loop tree, not the root. 1077 IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost ); 1078 1079 #ifdef ASSERT 1080 // verify that regions in irreducible loops are marked is_in_irreducible_loop 1081 void verify_regions_in_irreducible_loops(); 1082 bool is_in_irreducible_loop(RegionNode* region); 1083 #endif 1084 1085 // Place Data nodes in some loop nest 1086 void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ); 1087 void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ); 1088 void build_loop_late_post_work(Node* n, bool pinned); 1089 void build_loop_late_post(Node* n); 1090 void verify_strip_mined_scheduling(Node *n, Node* least); 1091 1092 // Array of immediate dominance info for each CFG node indexed by node idx 1093 private: 1094 uint _idom_size; 1095 Node **_idom; // Array of immediate dominators 1096 uint *_dom_depth; // Used for fast LCA test 1097 GrowableArray<uint>* _dom_stk; // For recomputation of dom depth 1098 LoopOptsMode _mode; 1099 1100 // build the loop tree and perform any requested optimizations 1101 void build_and_optimize(); 1102 1103 // Dominators for the sea of nodes 1104 void Dominators(); 1105 1106 // Compute the Ideal Node to Loop mapping 1107 PhaseIdealLoop(PhaseIterGVN& igvn, LoopOptsMode mode) : 1108 PhaseTransform(Ideal_Loop), 1109 _igvn(igvn), 1110 _verify_me(nullptr), 1111 _verify_only(false), 1112 _mode(mode), 1113 _nodes_required(UINT_MAX) { 1114 assert(mode != LoopOptsVerify, "wrong constructor to verify IdealLoop"); 1115 build_and_optimize(); 1116 } 1117 1118 #ifndef PRODUCT 1119 // Verify that verify_me made the same decisions as a fresh run 1120 // or only verify that the graph is valid if verify_me is null. 1121 PhaseIdealLoop(PhaseIterGVN& igvn, const PhaseIdealLoop* verify_me = nullptr) : 1122 PhaseTransform(Ideal_Loop), 1123 _igvn(igvn), 1124 _verify_me(verify_me), 1125 _verify_only(verify_me == nullptr), 1126 _mode(LoopOptsVerify), 1127 _nodes_required(UINT_MAX) { 1128 build_and_optimize(); 1129 } 1130 #endif 1131 1132 public: 1133 Node* idom_no_update(Node* d) const { 1134 return idom_no_update(d->_idx); 1135 } 1136 1137 Node* idom_no_update(uint didx) const { 1138 assert(didx < _idom_size, "oob"); 1139 Node* n = _idom[didx]; 1140 assert(n != nullptr,"Bad immediate dominator info."); 1141 while (n->in(0) == nullptr) { // Skip dead CFG nodes 1142 n = (Node*)(((intptr_t)_loop_or_ctrl[n->_idx]) & ~1); 1143 assert(n != nullptr,"Bad immediate dominator info."); 1144 } 1145 return n; 1146 } 1147 1148 Node *idom(Node* d) const { 1149 return idom(d->_idx); 1150 } 1151 1152 Node *idom(uint didx) const { 1153 Node *n = idom_no_update(didx); 1154 _idom[didx] = n; // Lazily remove dead CFG nodes from table. 1155 return n; 1156 } 1157 1158 uint dom_depth(Node* d) const { 1159 guarantee(d != nullptr, "Null dominator info."); 1160 guarantee(d->_idx < _idom_size, ""); 1161 return _dom_depth[d->_idx]; 1162 } 1163 void set_idom(Node* d, Node* n, uint dom_depth); 1164 // Locally compute IDOM using dom_lca call 1165 Node *compute_idom( Node *region ) const; 1166 // Recompute dom_depth 1167 void recompute_dom_depth(); 1168 1169 // Is safept not required by an outer loop? 1170 bool is_deleteable_safept(Node* sfpt); 1171 1172 // Replace parallel induction variable (parallel to trip counter) 1173 void replace_parallel_iv(IdealLoopTree *loop); 1174 1175 Node *dom_lca( Node *n1, Node *n2 ) const { 1176 return find_non_split_ctrl(dom_lca_internal(n1, n2)); 1177 } 1178 Node *dom_lca_internal( Node *n1, Node *n2 ) const; 1179 1180 // Build and verify the loop tree without modifying the graph. This 1181 // is useful to verify that all inputs properly dominate their uses. 1182 static void verify(PhaseIterGVN& igvn) { 1183 #ifdef ASSERT 1184 ResourceMark rm; 1185 Compile::TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 1186 PhaseIdealLoop v(igvn); 1187 #endif 1188 } 1189 1190 // Recommended way to use PhaseIdealLoop. 1191 // Run PhaseIdealLoop in some mode and allocates a local scope for memory allocations. 1192 static void optimize(PhaseIterGVN &igvn, LoopOptsMode mode) { 1193 ResourceMark rm; 1194 PhaseIdealLoop v(igvn, mode); 1195 1196 Compile* C = Compile::current(); 1197 if (!C->failing()) { 1198 // Cleanup any modified bits 1199 igvn.optimize(); 1200 if (C->failing()) { return; } 1201 v.log_loop_tree(); 1202 } 1203 } 1204 1205 // True if the method has at least 1 irreducible loop 1206 bool _has_irreducible_loops; 1207 1208 // Per-Node transform 1209 virtual Node* transform(Node* n) { return nullptr; } 1210 1211 Node* loop_exit_control(Node* x, IdealLoopTree* loop); 1212 Node* loop_exit_test(Node* back_control, IdealLoopTree* loop, Node*& incr, Node*& limit, BoolTest::mask& bt, float& cl_prob); 1213 Node* loop_iv_incr(Node* incr, Node* x, IdealLoopTree* loop, Node*& phi_incr); 1214 Node* loop_iv_stride(Node* incr, IdealLoopTree* loop, Node*& xphi); 1215 PhiNode* loop_iv_phi(Node* xphi, Node* phi_incr, Node* x, IdealLoopTree* loop); 1216 1217 bool is_counted_loop(Node* x, IdealLoopTree*&loop, BasicType iv_bt); 1218 1219 Node* loop_nest_replace_iv(Node* iv_to_replace, Node* inner_iv, Node* outer_phi, Node* inner_head, BasicType bt); 1220 bool create_loop_nest(IdealLoopTree* loop, Node_List &old_new); 1221 #ifdef ASSERT 1222 bool convert_to_long_loop(Node* cmp, Node* phi, IdealLoopTree* loop); 1223 #endif 1224 void add_parse_predicate(Deoptimization::DeoptReason reason, Node* inner_head, IdealLoopTree* loop, SafePointNode* sfpt); 1225 SafePointNode* find_safepoint(Node* back_control, Node* x, IdealLoopTree* loop); 1226 IdealLoopTree* insert_outer_loop(IdealLoopTree* loop, LoopNode* outer_l, Node* outer_ift); 1227 IdealLoopTree* create_outer_strip_mined_loop(BoolNode *test, Node *cmp, Node *init_control, 1228 IdealLoopTree* loop, float cl_prob, float le_fcnt, 1229 Node*& entry_control, Node*& iffalse); 1230 1231 Node* exact_limit( IdealLoopTree *loop ); 1232 1233 // Return a post-walked LoopNode 1234 IdealLoopTree *get_loop( Node *n ) const { 1235 // Dead nodes have no loop, so return the top level loop instead 1236 if (!has_node(n)) return _ltree_root; 1237 assert(!has_ctrl(n), ""); 1238 return (IdealLoopTree*)_loop_or_ctrl[n->_idx]; 1239 } 1240 1241 IdealLoopTree* ltree_root() const { return _ltree_root; } 1242 1243 // Is 'n' a (nested) member of 'loop'? 1244 int is_member( const IdealLoopTree *loop, Node *n ) const { 1245 return loop->is_member(get_loop(n)); } 1246 1247 // This is the basic building block of the loop optimizations. It clones an 1248 // entire loop body. It makes an old_new loop body mapping; with this 1249 // mapping you can find the new-loop equivalent to an old-loop node. All 1250 // new-loop nodes are exactly equal to their old-loop counterparts, all 1251 // edges are the same. All exits from the old-loop now have a RegionNode 1252 // that merges the equivalent new-loop path. This is true even for the 1253 // normal "loop-exit" condition. All uses of loop-invariant old-loop values 1254 // now come from (one or more) Phis that merge their new-loop equivalents. 1255 // Parameter side_by_side_idom: 1256 // When side_by_size_idom is null, the dominator tree is constructed for 1257 // the clone loop to dominate the original. Used in construction of 1258 // pre-main-post loop sequence. 1259 // When nonnull, the clone and original are side-by-side, both are 1260 // dominated by the passed in side_by_side_idom node. Used in 1261 // construction of unswitched loops. 1262 enum CloneLoopMode { 1263 IgnoreStripMined = 0, // Only clone inner strip mined loop 1264 CloneIncludesStripMined = 1, // clone both inner and outer strip mined loops 1265 ControlAroundStripMined = 2 // Only clone inner strip mined loop, 1266 // result control flow branches 1267 // either to inner clone or outer 1268 // strip mined loop. 1269 }; 1270 void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth, 1271 CloneLoopMode mode, Node* side_by_side_idom = nullptr); 1272 void clone_loop_handle_data_uses(Node* old, Node_List &old_new, 1273 IdealLoopTree* loop, IdealLoopTree* companion_loop, 1274 Node_List*& split_if_set, Node_List*& split_bool_set, 1275 Node_List*& split_cex_set, Node_List& worklist, 1276 uint new_counter, CloneLoopMode mode); 1277 void clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop, 1278 IdealLoopTree* outer_loop, int dd, Node_List &old_new, 1279 Node_List& extra_data_nodes); 1280 1281 // If we got the effect of peeling, either by actually peeling or by 1282 // making a pre-loop which must execute at least once, we can remove 1283 // all loop-invariant dominated tests in the main body. 1284 void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new ); 1285 1286 // Generate code to do a loop peel for the given loop (and body). 1287 // old_new is a temp array. 1288 void do_peeling( IdealLoopTree *loop, Node_List &old_new ); 1289 1290 // Add pre and post loops around the given loop. These loops are used 1291 // during RCE, unrolling and aligning loops. 1292 void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ); 1293 1294 // Add post loop after the given loop. 1295 Node *insert_post_loop(IdealLoopTree* loop, Node_List& old_new, 1296 CountedLoopNode* main_head, CountedLoopEndNode* main_end, 1297 Node*& incr, Node* limit, CountedLoopNode*& post_head); 1298 1299 // Add a vector post loop between a vector main loop and the current post loop 1300 void insert_vector_post_loop(IdealLoopTree *loop, Node_List &old_new); 1301 // If Node n lives in the back_ctrl block, we clone a private version of n 1302 // in preheader_ctrl block and return that, otherwise return n. 1303 Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones ); 1304 1305 // Take steps to maximally unroll the loop. Peel any odd iterations, then 1306 // unroll to do double iterations. The next round of major loop transforms 1307 // will repeat till the doubled loop body does all remaining iterations in 1 1308 // pass. 1309 void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ); 1310 1311 // Unroll the loop body one step - make each trip do 2 iterations. 1312 void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ); 1313 1314 // Return true if exp is a constant times an induction var 1315 bool is_scaled_iv(Node* exp, Node* iv, BasicType bt, jlong* p_scale, bool* p_short_scale, int depth = 0); 1316 1317 bool is_iv(Node* exp, Node* iv, BasicType bt); 1318 1319 // Return true if exp is a scaled induction var plus (or minus) constant 1320 bool is_scaled_iv_plus_offset(Node* exp, Node* iv, BasicType bt, jlong* p_scale, Node** p_offset, bool* p_short_scale = nullptr, int depth = 0); 1321 bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset) { 1322 jlong long_scale; 1323 if (is_scaled_iv_plus_offset(exp, iv, T_INT, &long_scale, p_offset)) { 1324 int int_scale = checked_cast<int>(long_scale); 1325 if (p_scale != nullptr) { 1326 *p_scale = int_scale; 1327 } 1328 return true; 1329 } 1330 return false; 1331 } 1332 // Helper for finding more complex matches to is_scaled_iv_plus_offset. 1333 bool is_scaled_iv_plus_extra_offset(Node* exp1, Node* offset2, Node* iv, 1334 BasicType bt, 1335 jlong* p_scale, Node** p_offset, 1336 bool* p_short_scale, int depth); 1337 1338 // Create a new if above the uncommon_trap_if_pattern for the predicate to be promoted 1339 IfProjNode* create_new_if_for_predicate(ParsePredicateSuccessProj* parse_predicate_proj, Node* new_entry, 1340 Deoptimization::DeoptReason reason, int opcode, 1341 bool rewire_uncommon_proj_phi_inputs = false); 1342 1343 private: 1344 // Helper functions for create_new_if_for_predicate() 1345 void set_ctrl_of_nodes_with_same_ctrl(Node* node, ProjNode* old_ctrl, Node* new_ctrl); 1346 Unique_Node_List find_nodes_with_same_ctrl(Node* node, const ProjNode* ctrl); 1347 Node* clone_nodes_with_same_ctrl(Node* node, ProjNode* old_ctrl, Node* new_ctrl); 1348 Dict clone_nodes(const Node_List& list_to_clone); 1349 void rewire_cloned_nodes_to_ctrl(const ProjNode* old_ctrl, Node* new_ctrl, const Node_List& nodes_with_same_ctrl, 1350 const Dict& old_new_mapping); 1351 void rewire_inputs_of_clones_to_clones(Node* new_ctrl, Node* clone, const Dict& old_new_mapping, const Node* next); 1352 1353 public: 1354 void register_control(Node* n, IdealLoopTree *loop, Node* pred, bool update_body = true); 1355 1356 // Construct a range check for a predicate if 1357 BoolNode* rc_predicate(IdealLoopTree* loop, Node* ctrl, int scale, Node* offset, Node* init, Node* limit, 1358 jint stride, Node* range, bool upper, bool& overflow); 1359 1360 // Implementation of the loop predication to promote checks outside the loop 1361 bool loop_predication_impl(IdealLoopTree *loop); 1362 1363 private: 1364 bool loop_predication_impl_helper(IdealLoopTree* loop, IfProjNode* if_success_proj, 1365 ParsePredicateSuccessProj* parse_predicate_proj, CountedLoopNode* cl, ConNode* zero, 1366 Invariance& invar, Deoptimization::DeoptReason reason); 1367 bool can_create_loop_predicates(const PredicateBlock* profiled_loop_predicate_block) const; 1368 bool loop_predication_should_follow_branches(IdealLoopTree* loop, float& loop_trip_cnt); 1369 void loop_predication_follow_branches(Node *c, IdealLoopTree *loop, float loop_trip_cnt, 1370 PathFrequency& pf, Node_Stack& stack, VectorSet& seen, 1371 Node_List& if_proj_list); 1372 IfProjNode* add_template_assertion_predicate(IfNode* iff, IdealLoopTree* loop, IfProjNode* if_proj, 1373 ParsePredicateSuccessProj* parse_predicate_proj, 1374 IfProjNode* upper_bound_proj, int scale, Node* offset, Node* init, Node* limit, 1375 jint stride, Node* rng, bool& overflow, Deoptimization::DeoptReason reason); 1376 Node* add_range_check_elimination_assertion_predicate(IdealLoopTree* loop, Node* predicate_proj, int scale_con, 1377 Node* offset, Node* limit, jint stride_con, Node* value); 1378 1379 // Helper function to collect predicate for eliminating the useless ones 1380 void eliminate_useless_predicates(); 1381 1382 void eliminate_useless_parse_predicates(); 1383 void mark_all_parse_predicates_useless() const; 1384 void mark_loop_associated_parse_predicates_useful(); 1385 static void mark_useful_parse_predicates_for_loop(IdealLoopTree* loop); 1386 void add_useless_parse_predicates_to_igvn_worklist(); 1387 1388 void eliminate_useless_template_assertion_predicates(); 1389 void collect_useful_template_assertion_predicates(Unique_Node_List& useful_predicates); 1390 static void collect_useful_template_assertion_predicates_for_loop(IdealLoopTree* loop, Unique_Node_List& useful_predicates); 1391 void eliminate_useless_template_assertion_predicates(Unique_Node_List& useful_predicates); 1392 1393 void eliminate_useless_zero_trip_guard(); 1394 1395 bool has_control_dependencies_from_predicates(LoopNode* head) const; 1396 void verify_fast_loop(LoopNode* head, const ProjNode* proj_true) const NOT_DEBUG_RETURN; 1397 public: 1398 // Change the control input of expensive nodes to allow commoning by 1399 // IGVN when it is guaranteed to not result in a more frequent 1400 // execution of the expensive node. Return true if progress. 1401 bool process_expensive_nodes(); 1402 1403 // Check whether node has become unreachable 1404 bool is_node_unreachable(Node *n) const { 1405 return !has_node(n) || n->is_unreachable(_igvn); 1406 } 1407 1408 // Eliminate range-checks and other trip-counter vs loop-invariant tests. 1409 void do_range_check(IdealLoopTree *loop, Node_List &old_new); 1410 1411 // Create a slow version of the loop by cloning the loop 1412 // and inserting an if to select fast-slow versions. 1413 // Return the inserted if. 1414 IfNode* create_slow_version_of_loop(IdealLoopTree *loop, 1415 Node_List &old_new, 1416 Node_List &unswitch_iffs, 1417 CloneLoopMode mode); 1418 1419 // Clone loop with an invariant test (that does not exit) and 1420 // insert a clone of the test that selects which version to 1421 // execute. 1422 void do_unswitching (IdealLoopTree *loop, Node_List &old_new); 1423 1424 // Find candidate "if" for unswitching 1425 IfNode* find_unswitching_candidate(const IdealLoopTree *loop, Node_List& unswitch_iffs) const; 1426 1427 // Range Check Elimination uses this function! 1428 // Constrain the main loop iterations so the affine function: 1429 // low_limit <= scale_con * I + offset < upper_limit 1430 // always holds true. That is, either increase the number of iterations in 1431 // the pre-loop or the post-loop until the condition holds true in the main 1432 // loop. Scale_con, offset and limit are all loop invariant. 1433 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); 1434 // Helper function for add_constraint(). 1435 Node* adjust_limit(bool reduce, Node* scale, Node* offset, Node* rc_limit, Node* old_limit, Node* pre_ctrl, bool round); 1436 1437 // Partially peel loop up through last_peel node. 1438 bool partial_peel( IdealLoopTree *loop, Node_List &old_new ); 1439 bool duplicate_loop_backedge(IdealLoopTree *loop, Node_List &old_new); 1440 1441 // Move UnorderedReduction out of loop if possible 1442 void move_unordered_reduction_out_of_loop(IdealLoopTree* loop); 1443 1444 // Create a scheduled list of nodes control dependent on ctrl set. 1445 void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched ); 1446 // Has a use in the vector set 1447 bool has_use_in_set( Node* n, VectorSet& vset ); 1448 // Has use internal to the vector set (ie. not in a phi at the loop head) 1449 bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ); 1450 // clone "n" for uses that are outside of loop 1451 int clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ); 1452 // clone "n" for special uses that are in the not_peeled region 1453 void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, 1454 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ); 1455 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist 1456 void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ); 1457 #ifdef ASSERT 1458 // Validate the loop partition sets: peel and not_peel 1459 bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel ); 1460 // Ensure that uses outside of loop are of the right form 1461 bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, 1462 uint orig_exit_idx, uint clone_exit_idx); 1463 bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx); 1464 #endif 1465 1466 // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.) 1467 int stride_of_possible_iv( Node* iff ); 1468 bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; } 1469 // Return the (unique) control output node that's in the loop (if it exists.) 1470 Node* stay_in_loop( Node* n, IdealLoopTree *loop); 1471 // Insert a signed compare loop exit cloned from an unsigned compare. 1472 IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop); 1473 void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop); 1474 // Utility to register node "n" with PhaseIdealLoop 1475 void register_node(Node* n, IdealLoopTree* loop, Node* pred, uint ddepth); 1476 // Utility to create an if-projection 1477 ProjNode* proj_clone(ProjNode* p, IfNode* iff); 1478 // Force the iff control output to be the live_proj 1479 Node* short_circuit_if(IfNode* iff, ProjNode* live_proj); 1480 // Insert a region before an if projection 1481 RegionNode* insert_region_before_proj(ProjNode* proj); 1482 // Insert a new if before an if projection 1483 ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj); 1484 1485 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. 1486 // "Nearly" because all Nodes have been cloned from the original in the loop, 1487 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs 1488 // through the Phi recursively, and return a Bool. 1489 Node* clone_iff(PhiNode* phi); 1490 CmpNode* clone_bool(PhiNode* phi); 1491 1492 1493 // Rework addressing expressions to get the most loop-invariant stuff 1494 // moved out. We'd like to do all associative operators, but it's especially 1495 // important (common) to do address expressions. 1496 Node* remix_address_expressions(Node* n); 1497 Node* remix_address_expressions_add_left_shift(Node* n, IdealLoopTree* n_loop, Node* n_ctrl, BasicType bt); 1498 1499 // Convert add to muladd to generate MuladdS2I under certain criteria 1500 Node * convert_add_to_muladd(Node * n); 1501 1502 // Attempt to use a conditional move instead of a phi/branch 1503 Node *conditional_move( Node *n ); 1504 1505 // Check for aggressive application of 'split-if' optimization, 1506 // using basic block level info. 1507 void split_if_with_blocks ( VectorSet &visited, Node_Stack &nstack); 1508 Node *split_if_with_blocks_pre ( Node *n ); 1509 void split_if_with_blocks_post( Node *n ); 1510 Node *has_local_phi_input( Node *n ); 1511 // Mark an IfNode as being dominated by a prior test, 1512 // without actually altering the CFG (and hence IDOM info). 1513 void dominated_by(IfProjNode* prevdom, IfNode* iff, bool flip = false, bool exclude_loop_predicate = false); 1514 1515 // Split Node 'n' through merge point 1516 RegionNode* split_thru_region(Node* n, RegionNode* region); 1517 // Split Node 'n' through merge point if there is enough win. 1518 Node *split_thru_phi( Node *n, Node *region, int policy ); 1519 // Found an If getting its condition-code input from a Phi in the 1520 // same block. Split thru the Region. 1521 void do_split_if(Node *iff, RegionNode** new_false_region = nullptr, RegionNode** new_true_region = nullptr); 1522 1523 // Conversion of fill/copy patterns into intrinsic versions 1524 bool do_intrinsify_fill(); 1525 bool intrinsify_fill(IdealLoopTree* lpt); 1526 bool match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value, 1527 Node*& shift, Node*& offset); 1528 1529 private: 1530 // Return a type based on condition control flow 1531 const TypeInt* filtered_type( Node *n, Node* n_ctrl); 1532 const TypeInt* filtered_type( Node *n ) { return filtered_type(n, nullptr); } 1533 // Helpers for filtered type 1534 const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl); 1535 1536 // Helper functions 1537 Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache ); 1538 Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true ); 1539 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 ); 1540 bool split_up( Node *n, Node *blk1, Node *blk2 ); 1541 1542 Node* place_outside_loop(Node* useblock, IdealLoopTree* loop) const; 1543 Node* try_move_store_before_loop(Node* n, Node *n_ctrl); 1544 void try_move_store_after_loop(Node* n); 1545 void move_flat_array_check_out_of_loop(Node* n); 1546 bool identical_backtoback_ifs(Node *n); 1547 bool flat_array_element_type_check(Node *n); 1548 bool can_split_if(Node *n_ctrl); 1549 bool cannot_split_division(const Node* n, const Node* region) const; 1550 static bool is_divisor_counted_loop_phi(const Node* divisor, const Node* loop); 1551 bool loop_phi_backedge_type_contains_zero(const Node* phi_divisor, const Type* zero) const; 1552 1553 // Determine if a method is too big for a/another round of split-if, based on 1554 // a magic (approximate) ratio derived from the equally magic constant 35000, 1555 // previously used for this purpose (but without relating to the node limit). 1556 bool must_throttle_split_if() { 1557 uint threshold = C->max_node_limit() * 2 / 5; 1558 return C->live_nodes() > threshold; 1559 } 1560 1561 // A simplistic node request tracking mechanism, where 1562 // = UINT_MAX Request not valid or made final. 1563 // < UINT_MAX Nodes currently requested (estimate). 1564 uint _nodes_required; 1565 1566 enum { REQUIRE_MIN = 70 }; 1567 1568 uint nodes_required() const { return _nodes_required; } 1569 1570 // Given the _currently_ available number of nodes, check whether there is 1571 // "room" for an additional request or not, considering the already required 1572 // number of nodes. Return TRUE if the new request is exceeding the node 1573 // budget limit, otherwise return FALSE. Note that this interpretation will 1574 // act pessimistic on additional requests when new nodes have already been 1575 // generated since the 'begin'. This behaviour fits with the intention that 1576 // node estimates/requests should be made upfront. 1577 bool exceeding_node_budget(uint required = 0) { 1578 assert(C->live_nodes() < C->max_node_limit(), "sanity"); 1579 uint available = C->max_node_limit() - C->live_nodes(); 1580 return available < required + _nodes_required + REQUIRE_MIN; 1581 } 1582 1583 uint require_nodes(uint require, uint minreq = REQUIRE_MIN) { 1584 precond(require > 0); 1585 _nodes_required += MAX2(require, minreq); 1586 return _nodes_required; 1587 } 1588 1589 bool may_require_nodes(uint require, uint minreq = REQUIRE_MIN) { 1590 return !exceeding_node_budget(require) && require_nodes(require, minreq) > 0; 1591 } 1592 1593 uint require_nodes_begin() { 1594 assert(_nodes_required == UINT_MAX, "Bad state (begin)."); 1595 _nodes_required = 0; 1596 return C->live_nodes(); 1597 } 1598 1599 // When a node request is final, optionally check that the requested number 1600 // of nodes was reasonably correct with respect to the number of new nodes 1601 // introduced since the last 'begin'. Always check that we have not exceeded 1602 // the maximum node limit. 1603 void require_nodes_final(uint live_at_begin, bool check_estimate) { 1604 assert(_nodes_required < UINT_MAX, "Bad state (final)."); 1605 1606 #ifdef ASSERT 1607 if (check_estimate) { 1608 // Check that the node budget request was not off by too much (x2). 1609 // Should this be the case we _surely_ need to improve the estimates 1610 // used in our budget calculations. 1611 if (C->live_nodes() - live_at_begin > 2 * _nodes_required) { 1612 log_info(compilation)("Bad node estimate: actual = %d >> request = %d", 1613 C->live_nodes() - live_at_begin, _nodes_required); 1614 } 1615 } 1616 #endif 1617 // Assert that we have stayed within the node budget limit. 1618 assert(C->live_nodes() < C->max_node_limit(), 1619 "Exceeding node budget limit: %d + %d > %d (request = %d)", 1620 C->live_nodes() - live_at_begin, live_at_begin, 1621 C->max_node_limit(), _nodes_required); 1622 1623 _nodes_required = UINT_MAX; 1624 } 1625 1626 // Clone Parse Predicates to slow and fast loop when unswitching a loop 1627 void clone_parse_and_assertion_predicates_to_unswitched_loop(IdealLoopTree* loop, Node_List& old_new, 1628 IfProjNode*& iffast_pred, IfProjNode*& ifslow_pred); 1629 void clone_loop_predication_predicates_to_unswitched_loop(IdealLoopTree* loop, const Node_List& old_new, 1630 const PredicateBlock* predicate_block, 1631 Deoptimization::DeoptReason reason, IfProjNode*& iffast_pred, 1632 IfProjNode*& ifslow_pred); 1633 void clone_parse_predicate_to_unswitched_loops(const PredicateBlock* predicate_block, Deoptimization::DeoptReason reason, 1634 IfProjNode*& iffast_pred, IfProjNode*& ifslow_pred); 1635 IfProjNode* clone_parse_predicate_to_unswitched_loop(ParsePredicateSuccessProj* parse_predicate_proj, Node* new_entry, 1636 Deoptimization::DeoptReason reason, bool slow_loop); 1637 void clone_assertion_predicates_to_unswitched_loop(IdealLoopTree* loop, const Node_List& old_new, 1638 Deoptimization::DeoptReason reason, IfProjNode* old_predicate_proj, 1639 ParsePredicateSuccessProj* fast_loop_parse_predicate_proj, 1640 ParsePredicateSuccessProj* slow_loop_parse_predicate_proj); 1641 IfProjNode* clone_assertion_predicate_for_unswitched_loops(Node* iff, IfProjNode* predicate, 1642 Deoptimization::DeoptReason reason, 1643 ParsePredicateSuccessProj* parse_predicate_proj); 1644 static void check_cloned_parse_predicate_for_unswitching(const Node* new_entry, bool is_fast_loop) PRODUCT_RETURN; 1645 1646 bool _created_loop_node; 1647 DEBUG_ONLY(void dump_idoms(Node* early, Node* wrong_lca);) 1648 NOT_PRODUCT(void dump_idoms_in_reverse(const Node* n, const Node_List& idom_list) const;) 1649 1650 public: 1651 void set_created_loop_node() { _created_loop_node = true; } 1652 bool created_loop_node() { return _created_loop_node; } 1653 void register_new_node(Node* n, Node* blk); 1654 1655 #ifdef ASSERT 1656 void dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA); 1657 #endif 1658 1659 #ifndef PRODUCT 1660 void dump() const; 1661 void dump_idom(Node* n) const { dump_idom(n, 1000); } // For debugging 1662 void dump_idom(Node* n, uint count) const; 1663 void get_idoms(Node* n, uint count, Unique_Node_List& idoms) const; 1664 void dump(IdealLoopTree* loop, uint rpo_idx, Node_List &rpo_list) const; 1665 IdealLoopTree* get_loop_idx(Node* n) const { 1666 // Dead nodes have no loop, so return the top level loop instead 1667 return _loop_or_ctrl[n->_idx] ? (IdealLoopTree*)_loop_or_ctrl[n->_idx] : _ltree_root; 1668 } 1669 // Print some stats 1670 static void print_statistics(); 1671 static int _loop_invokes; // Count of PhaseIdealLoop invokes 1672 static int _loop_work; // Sum of PhaseIdealLoop x _unique 1673 static volatile int _long_loop_candidates; 1674 static volatile int _long_loop_nests; 1675 static volatile int _long_loop_counted_loops; 1676 #endif 1677 1678 #ifdef ASSERT 1679 void verify() const; 1680 bool verify_idom_and_nodes(Node* root, const PhaseIdealLoop* phase_verify) const; 1681 bool verify_idom(Node* n, const PhaseIdealLoop* phase_verify) const; 1682 bool verify_loop_ctrl(Node* n, const PhaseIdealLoop* phase_verify) const; 1683 #endif 1684 1685 void rpo(Node* start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list) const; 1686 1687 void check_counted_loop_shape(IdealLoopTree* loop, Node* x, BasicType bt) NOT_DEBUG_RETURN; 1688 1689 LoopNode* create_inner_head(IdealLoopTree* loop, BaseCountedLoopNode* head, IfNode* exit_test); 1690 1691 1692 int extract_long_range_checks(const IdealLoopTree* loop, jlong stride_con, int iters_limit, PhiNode* phi, 1693 Node_List &range_checks); 1694 1695 void transform_long_range_checks(int stride_con, const Node_List &range_checks, Node* outer_phi, 1696 Node* inner_iters_actual_int, Node* inner_phi, 1697 Node* iv_add, LoopNode* inner_head); 1698 1699 Node* get_late_ctrl_with_anti_dep(LoadNode* n, Node* early, Node* LCA); 1700 1701 bool ctrl_of_use_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop, Node* ctrl); 1702 1703 bool ctrl_of_all_uses_out_of_loop(const Node* n, Node* n_ctrl, IdealLoopTree* n_loop); 1704 1705 Node* compute_early_ctrl(Node* n, Node* n_ctrl); 1706 1707 void try_sink_out_of_loop(Node* n); 1708 1709 Node* clamp(Node* R, Node* L, Node* H); 1710 1711 bool safe_for_if_replacement(const Node* dom) const; 1712 1713 void push_pinned_nodes_thru_region(IfNode* dom_if, Node* region); 1714 1715 bool try_merge_identical_ifs(Node* n); 1716 1717 void clone_loop_body(const Node_List& body, Node_List &old_new, CloneMap* cm); 1718 1719 void fix_body_edges(const Node_List &body, IdealLoopTree* loop, const Node_List &old_new, int dd, 1720 IdealLoopTree* parent, bool partial); 1721 1722 void fix_ctrl_uses(const Node_List& body, const IdealLoopTree* loop, Node_List &old_new, CloneLoopMode mode, 1723 Node* side_by_side_idom, CloneMap* cm, Node_List &worklist); 1724 1725 void fix_data_uses(Node_List& body, IdealLoopTree* loop, CloneLoopMode mode, IdealLoopTree* outer_loop, 1726 uint new_counter, Node_List& old_new, Node_List& worklist, Node_List*& split_if_set, 1727 Node_List*& split_bool_set, Node_List*& split_cex_set); 1728 1729 void finish_clone_loop(Node_List* split_if_set, Node_List* split_bool_set, Node_List* split_cex_set); 1730 1731 bool clone_cmp_down(Node* n, const Node* blk1, const Node* blk2); 1732 1733 void clone_loadklass_nodes_at_cmp_index(const Node* n, Node* cmp, int i); 1734 1735 bool clone_cmp_loadklass_down(Node* n, const Node* blk1, const Node* blk2); 1736 1737 bool at_relevant_ctrl(Node* n, const Node* blk1, const Node* blk2); 1738 1739 1740 Node* similar_subtype_check(const Node* x, Node* r_in); 1741 1742 void update_addp_chain_base(Node* x, Node* old_base, Node* new_base); 1743 1744 bool can_move_to_inner_loop(Node* n, LoopNode* n_loop, Node* x); 1745 }; 1746 1747 1748 class AutoNodeBudget : public StackObj 1749 { 1750 public: 1751 enum budget_check_t { BUDGET_CHECK, NO_BUDGET_CHECK }; 1752 1753 AutoNodeBudget(PhaseIdealLoop* phase, budget_check_t chk = BUDGET_CHECK) 1754 : _phase(phase), 1755 _check_at_final(chk == BUDGET_CHECK), 1756 _nodes_at_begin(0) 1757 { 1758 precond(_phase != nullptr); 1759 1760 _nodes_at_begin = _phase->require_nodes_begin(); 1761 } 1762 1763 ~AutoNodeBudget() { 1764 #ifndef PRODUCT 1765 if (TraceLoopOpts) { 1766 uint request = _phase->nodes_required(); 1767 uint delta = _phase->C->live_nodes() - _nodes_at_begin; 1768 1769 if (request < delta) { 1770 tty->print_cr("Exceeding node budget: %d < %d", request, delta); 1771 } else { 1772 uint const REQUIRE_MIN = PhaseIdealLoop::REQUIRE_MIN; 1773 // Identify the worst estimates as "poor" ones. 1774 if (request > REQUIRE_MIN && delta > 0) { 1775 if ((delta > REQUIRE_MIN && request > 3 * delta) || 1776 (delta <= REQUIRE_MIN && request > 10 * delta)) { 1777 tty->print_cr("Poor node estimate: %d >> %d", request, delta); 1778 } 1779 } 1780 } 1781 } 1782 #endif // PRODUCT 1783 _phase->require_nodes_final(_nodes_at_begin, _check_at_final); 1784 } 1785 1786 private: 1787 PhaseIdealLoop* _phase; 1788 bool _check_at_final; 1789 uint _nodes_at_begin; 1790 }; 1791 1792 inline Node* IdealLoopTree::tail() { 1793 // Handle lazy update of _tail field. 1794 if (_tail->in(0) == nullptr) { 1795 _tail = _phase->get_ctrl(_tail); 1796 } 1797 return _tail; 1798 } 1799 1800 inline Node* IdealLoopTree::head() { 1801 // Handle lazy update of _head field. 1802 if (_head->in(0) == nullptr) { 1803 _head = _phase->get_ctrl(_head); 1804 } 1805 return _head; 1806 } 1807 1808 // Iterate over the loop tree using a preorder, left-to-right traversal. 1809 // 1810 // Example that visits all counted loops from within PhaseIdealLoop 1811 // 1812 // for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { 1813 // IdealLoopTree* lpt = iter.current(); 1814 // if (!lpt->is_counted()) continue; 1815 // ... 1816 class LoopTreeIterator : public StackObj { 1817 private: 1818 IdealLoopTree* _root; 1819 IdealLoopTree* _curnt; 1820 1821 public: 1822 LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {} 1823 1824 bool done() { return _curnt == nullptr; } // Finished iterating? 1825 1826 void next(); // Advance to next loop tree 1827 1828 IdealLoopTree* current() { return _curnt; } // Return current value of iterator. 1829 }; 1830 1831 // Compute probability of reaching some CFG node from a fixed 1832 // dominating CFG node 1833 class PathFrequency { 1834 private: 1835 Node* _dom; // frequencies are computed relative to this node 1836 Node_Stack _stack; 1837 GrowableArray<float> _freqs_stack; // keep track of intermediate result at regions 1838 GrowableArray<float> _freqs; // cache frequencies 1839 PhaseIdealLoop* _phase; 1840 1841 float check_and_truncate_frequency(float f) { 1842 assert(f >= 0, "Incorrect frequency"); 1843 // We do not perform an exact (f <= 1) check 1844 // this would be error prone with rounding of floats. 1845 // Performing a check like (f <= 1+eps) would be of benefit, 1846 // however, it is not evident how to determine such an eps, 1847 // given that an arbitrary number of add/mul operations 1848 // are performed on these frequencies. 1849 return (f > 1) ? 1 : f; 1850 } 1851 1852 public: 1853 PathFrequency(Node* dom, PhaseIdealLoop* phase) 1854 : _dom(dom), _stack(0), _phase(phase) { 1855 } 1856 1857 float to(Node* n); 1858 }; 1859 1860 #endif // SHARE_OPTO_LOOPNODE_HPP