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