LLVM  14.0.0git
LoopVectorize.cpp
Go to the documentation of this file.
1 //===- LoopVectorize.cpp - A Loop Vectorizer ------------------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This is the LLVM loop vectorizer. This pass modifies 'vectorizable' loops
10 // and generates target-independent LLVM-IR.
11 // The vectorizer uses the TargetTransformInfo analysis to estimate the costs
12 // of instructions in order to estimate the profitability of vectorization.
13 //
14 // The loop vectorizer combines consecutive loop iterations into a single
15 // 'wide' iteration. After this transformation the index is incremented
16 // by the SIMD vector width, and not by one.
17 //
18 // This pass has three parts:
19 // 1. The main loop pass that drives the different parts.
20 // 2. LoopVectorizationLegality - A unit that checks for the legality
21 // of the vectorization.
22 // 3. InnerLoopVectorizer - A unit that performs the actual
23 // widening of instructions.
24 // 4. LoopVectorizationCostModel - A unit that checks for the profitability
25 // of vectorization. It decides on the optimal vector width, which
26 // can be one, if vectorization is not profitable.
27 //
28 // There is a development effort going on to migrate loop vectorizer to the
29 // VPlan infrastructure and to introduce outer loop vectorization support (see
30 // docs/Proposal/VectorizationPlan.rst and
31 // http://lists.llvm.org/pipermail/llvm-dev/2017-December/119523.html). For this
32 // purpose, we temporarily introduced the VPlan-native vectorization path: an
33 // alternative vectorization path that is natively implemented on top of the
34 // VPlan infrastructure. See EnableVPlanNativePath for enabling.
35 //
36 //===----------------------------------------------------------------------===//
37 //
38 // The reduction-variable vectorization is based on the paper:
39 // D. Nuzman and R. Henderson. Multi-platform Auto-vectorization.
40 //
41 // Variable uniformity checks are inspired by:
42 // Karrenberg, R. and Hack, S. Whole Function Vectorization.
43 //
44 // The interleaved access vectorization is based on the paper:
45 // Dorit Nuzman, Ira Rosen and Ayal Zaks. Auto-Vectorization of Interleaved
46 // Data for SIMD
47 //
48 // Other ideas/concepts are from:
49 // A. Zaks and D. Nuzman. Autovectorization in GCC-two years later.
50 //
51 // S. Maleki, Y. Gao, M. Garzaran, T. Wong and D. Padua. An Evaluation of
52 // Vectorizing Compilers.
53 //
54 //===----------------------------------------------------------------------===//
55 
58 #include "VPRecipeBuilder.h"
59 #include "VPlan.h"
60 #include "VPlanHCFGBuilder.h"
61 #include "VPlanPredicator.h"
62 #include "VPlanTransforms.h"
63 #include "llvm/ADT/APInt.h"
64 #include "llvm/ADT/ArrayRef.h"
65 #include "llvm/ADT/DenseMap.h"
66 #include "llvm/ADT/DenseMapInfo.h"
67 #include "llvm/ADT/Hashing.h"
68 #include "llvm/ADT/MapVector.h"
69 #include "llvm/ADT/None.h"
70 #include "llvm/ADT/Optional.h"
71 #include "llvm/ADT/STLExtras.h"
72 #include "llvm/ADT/SmallPtrSet.h"
73 #include "llvm/ADT/SmallSet.h"
74 #include "llvm/ADT/SmallVector.h"
75 #include "llvm/ADT/Statistic.h"
76 #include "llvm/ADT/StringRef.h"
77 #include "llvm/ADT/Twine.h"
82 #include "llvm/Analysis/CFG.h"
88 #include "llvm/Analysis/LoopInfo.h"
97 #include "llvm/IR/Attributes.h"
98 #include "llvm/IR/BasicBlock.h"
99 #include "llvm/IR/CFG.h"
100 #include "llvm/IR/Constant.h"
101 #include "llvm/IR/Constants.h"
102 #include "llvm/IR/DataLayout.h"
104 #include "llvm/IR/DebugLoc.h"
105 #include "llvm/IR/DerivedTypes.h"
106 #include "llvm/IR/DiagnosticInfo.h"
107 #include "llvm/IR/Dominators.h"
108 #include "llvm/IR/Function.h"
109 #include "llvm/IR/IRBuilder.h"
110 #include "llvm/IR/InstrTypes.h"
111 #include "llvm/IR/Instruction.h"
112 #include "llvm/IR/Instructions.h"
113 #include "llvm/IR/IntrinsicInst.h"
114 #include "llvm/IR/Intrinsics.h"
115 #include "llvm/IR/LLVMContext.h"
116 #include "llvm/IR/Metadata.h"
117 #include "llvm/IR/Module.h"
118 #include "llvm/IR/Operator.h"
119 #include "llvm/IR/PatternMatch.h"
120 #include "llvm/IR/Type.h"
121 #include "llvm/IR/Use.h"
122 #include "llvm/IR/User.h"
123 #include "llvm/IR/Value.h"
124 #include "llvm/IR/ValueHandle.h"
125 #include "llvm/IR/Verifier.h"
126 #include "llvm/InitializePasses.h"
127 #include "llvm/Pass.h"
128 #include "llvm/Support/Casting.h"
130 #include "llvm/Support/Compiler.h"
131 #include "llvm/Support/Debug.h"
134 #include "llvm/Support/MathExtras.h"
144 #include <algorithm>
145 #include <cassert>
146 #include <cstdint>
147 #include <cstdlib>
148 #include <functional>
149 #include <iterator>
150 #include <limits>
151 #include <memory>
152 #include <string>
153 #include <tuple>
154 #include <utility>
155 
156 using namespace llvm;
157 
158 #define LV_NAME "loop-vectorize"
159 #define DEBUG_TYPE LV_NAME
160 
161 #ifndef NDEBUG
162 const char VerboseDebug[] = DEBUG_TYPE "-verbose";
163 #endif
164 
165 /// @{
166 /// Metadata attribute names
167 const char LLVMLoopVectorizeFollowupAll[] = "llvm.loop.vectorize.followup_all";
169  "llvm.loop.vectorize.followup_vectorized";
171  "llvm.loop.vectorize.followup_epilogue";
172 /// @}
173 
174 STATISTIC(LoopsVectorized, "Number of loops vectorized");
175 STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization");
176 STATISTIC(LoopsEpilogueVectorized, "Number of epilogues vectorized");
177 
179  "enable-epilogue-vectorization", cl::init(true), cl::Hidden,
180  cl::desc("Enable vectorization of epilogue loops."));
181 
183  "epilogue-vectorization-force-VF", cl::init(1), cl::Hidden,
184  cl::desc("When epilogue vectorization is enabled, and a value greater than "
185  "1 is specified, forces the given VF for all applicable epilogue "
186  "loops."));
187 
189  "epilogue-vectorization-minimum-VF", cl::init(16), cl::Hidden,
190  cl::desc("Only loops with vectorization factor equal to or larger than "
191  "the specified value are considered for epilogue vectorization."));
192 
193 /// Loops with a known constant trip count below this number are vectorized only
194 /// if no scalar iteration overheads are incurred.
196  "vectorizer-min-trip-count", cl::init(16), cl::Hidden,
197  cl::desc("Loops with a constant trip count that is smaller than this "
198  "value are vectorized only if no scalar iteration overheads "
199  "are incurred."));
200 
202  "pragma-vectorize-memory-check-threshold", cl::init(128), cl::Hidden,
203  cl::desc("The maximum allowed number of runtime memory checks with a "
204  "vectorize(enable) pragma."));
205 
206 // Option prefer-predicate-over-epilogue indicates that an epilogue is undesired,
207 // that predication is preferred, and this lists all options. I.e., the
208 // vectorizer will try to fold the tail-loop (epilogue) into the vector body
209 // and predicate the instructions accordingly. If tail-folding fails, there are
210 // different fallback strategies depending on these values:
211 namespace PreferPredicateTy {
212  enum Option {
216  };
217 } // namespace PreferPredicateTy
218 
220  "prefer-predicate-over-epilogue",
222  cl::Hidden,
223  cl::desc("Tail-folding and predication preferences over creating a scalar "
224  "epilogue loop."),
226  "scalar-epilogue",
227  "Don't tail-predicate loops, create scalar epilogue"),
229  "predicate-else-scalar-epilogue",
230  "prefer tail-folding, create scalar epilogue if tail "
231  "folding fails."),
233  "predicate-dont-vectorize",
234  "prefers tail-folding, don't attempt vectorization if "
235  "tail-folding fails.")));
236 
238  "vectorizer-maximize-bandwidth", cl::init(false), cl::Hidden,
239  cl::desc("Maximize bandwidth when selecting vectorization factor which "
240  "will be determined by the smallest type in loop."));
241 
243  "enable-interleaved-mem-accesses", cl::init(false), cl::Hidden,
244  cl::desc("Enable vectorization on interleaved memory accesses in a loop"));
245 
246 /// An interleave-group may need masking if it resides in a block that needs
247 /// predication, or in order to mask away gaps.
249  "enable-masked-interleaved-mem-accesses", cl::init(false), cl::Hidden,
250  cl::desc("Enable vectorization on masked interleaved memory accesses in a loop"));
251 
253  "tiny-trip-count-interleave-threshold", cl::init(128), cl::Hidden,
254  cl::desc("We don't interleave loops with a estimated constant trip count "
255  "below this number"));
256 
258  "force-target-num-scalar-regs", cl::init(0), cl::Hidden,
259  cl::desc("A flag that overrides the target's number of scalar registers."));
260 
262  "force-target-num-vector-regs", cl::init(0), cl::Hidden,
263  cl::desc("A flag that overrides the target's number of vector registers."));
264 
266  "force-target-max-scalar-interleave", cl::init(0), cl::Hidden,
267  cl::desc("A flag that overrides the target's max interleave factor for "
268  "scalar loops."));
269 
271  "force-target-max-vector-interleave", cl::init(0), cl::Hidden,
272  cl::desc("A flag that overrides the target's max interleave factor for "
273  "vectorized loops."));
274 
276  "force-target-instruction-cost", cl::init(0), cl::Hidden,
277  cl::desc("A flag that overrides the target's expected cost for "
278  "an instruction to a single constant value. Mostly "
279  "useful for getting consistent testing."));
280 
282  "force-target-supports-scalable-vectors", cl::init(false), cl::Hidden,
283  cl::desc(
284  "Pretend that scalable vectors are supported, even if the target does "
285  "not support them. This flag should only be used for testing."));
286 
288  "small-loop-cost", cl::init(20), cl::Hidden,
289  cl::desc(
290  "The cost of a loop that is considered 'small' by the interleaver."));
291 
293  "loop-vectorize-with-block-frequency", cl::init(true), cl::Hidden,
294  cl::desc("Enable the use of the block frequency analysis to access PGO "
295  "heuristics minimizing code growth in cold regions and being more "
296  "aggressive in hot regions."));
297 
298 // Runtime interleave loops for load/store throughput.
300  "enable-loadstore-runtime-interleave", cl::init(true), cl::Hidden,
301  cl::desc(
302  "Enable runtime interleaving until load/store ports are saturated"));
303 
304 /// Interleave small loops with scalar reductions.
306  "interleave-small-loop-scalar-reduction", cl::init(false), cl::Hidden,
307  cl::desc("Enable interleaving for loops with small iteration counts that "
308  "contain scalar reductions to expose ILP."));
309 
310 /// The number of stores in a loop that are allowed to need predication.
312  "vectorize-num-stores-pred", cl::init(1), cl::Hidden,
313  cl::desc("Max number of stores to be predicated behind an if."));
314 
316  "enable-ind-var-reg-heur", cl::init(true), cl::Hidden,
317  cl::desc("Count the induction variable only once when interleaving"));
318 
320  "enable-cond-stores-vec", cl::init(true), cl::Hidden,
321  cl::desc("Enable if predication of stores during vectorization."));
322 
324  "max-nested-scalar-reduction-interleave", cl::init(2), cl::Hidden,
325  cl::desc("The maximum interleave count to use when interleaving a scalar "
326  "reduction in a nested loop."));
327 
328 static cl::opt<bool>
329  PreferInLoopReductions("prefer-inloop-reductions", cl::init(false),
330  cl::Hidden,
331  cl::desc("Prefer in-loop vector reductions, "
332  "overriding the targets preference."));
333 
335  "force-ordered-reductions", cl::init(false), cl::Hidden,
336  cl::desc("Enable the vectorisation of loops with in-order (strict) "
337  "FP reductions"));
338 
340  "prefer-predicated-reduction-select", cl::init(false), cl::Hidden,
341  cl::desc(
342  "Prefer predicating a reduction operation over an after loop select."));
343 
345  "enable-vplan-native-path", cl::init(false), cl::Hidden,
346  cl::desc("Enable VPlan-native vectorization path with "
347  "support for outer loop vectorization."));
348 
349 // FIXME: Remove this switch once we have divergence analysis. Currently we
350 // assume divergent non-backedge branches when this switch is true.
352  "enable-vplan-predication", cl::init(false), cl::Hidden,
353  cl::desc("Enable VPlan-native vectorization path predicator with "
354  "support for outer loop vectorization."));
355 
356 // This flag enables the stress testing of the VPlan H-CFG construction in the
357 // VPlan-native vectorization path. It must be used in conjuction with
358 // -enable-vplan-native-path. -vplan-verify-hcfg can also be used to enable the
359 // verification of the H-CFGs built.
361  "vplan-build-stress-test", cl::init(false), cl::Hidden,
362  cl::desc(
363  "Build VPlan for every supported loop nest in the function and bail "
364  "out right after the build (stress test the VPlan H-CFG construction "
365  "in the VPlan-native vectorization path)."));
366 
368  "interleave-loops", cl::init(true), cl::Hidden,
369  cl::desc("Enable loop interleaving in Loop vectorization passes"));
371  "vectorize-loops", cl::init(true), cl::Hidden,
372  cl::desc("Run the Loop vectorization passes"));
373 
375  "vplan-print-in-dot-format", cl::init(false), cl::Hidden,
376  cl::desc("Use dot format instead of plain text when dumping VPlans"));
377 
378 /// A helper function that returns true if the given type is irregular. The
379 /// type is irregular if its allocated size doesn't equal the store size of an
380 /// element of the corresponding vector type.
381 static bool hasIrregularType(Type *Ty, const DataLayout &DL) {
382  // Determine if an array of N elements of type Ty is "bitcast compatible"
383  // with a <N x Ty> vector.
384  // This is only true if there is no padding between the array elements.
385  return DL.getTypeAllocSizeInBits(Ty) != DL.getTypeSizeInBits(Ty);
386 }
387 
388 /// A helper function that returns the reciprocal of the block probability of
389 /// predicated blocks. If we return X, we are assuming the predicated block
390 /// will execute once for every X iterations of the loop header.
391 ///
392 /// TODO: We should use actual block probability here, if available. Currently,
393 /// we always assume predicated blocks have a 50% chance of executing.
394 static unsigned getReciprocalPredBlockProb() { return 2; }
395 
396 /// A helper function that returns an integer or floating-point constant with
397 /// value C.
398 static Constant *getSignedIntOrFpConstant(Type *Ty, int64_t C) {
399  return Ty->isIntegerTy() ? ConstantInt::getSigned(Ty, C)
400  : ConstantFP::get(Ty, C);
401 }
402 
403 /// Returns "best known" trip count for the specified loop \p L as defined by
404 /// the following procedure:
405 /// 1) Returns exact trip count if it is known.
406 /// 2) Returns expected trip count according to profile data if any.
407 /// 3) Returns upper bound estimate if it is known.
408 /// 4) Returns None if all of the above failed.
410  // Check if exact trip count is known.
411  if (unsigned ExpectedTC = SE.getSmallConstantTripCount(L))
412  return ExpectedTC;
413 
414  // Check if there is an expected trip count available from profile data.
416  if (auto EstimatedTC = getLoopEstimatedTripCount(L))
417  return EstimatedTC;
418 
419  // Check if upper bound estimate is known.
420  if (unsigned ExpectedTC = SE.getSmallConstantMaxTripCount(L))
421  return ExpectedTC;
422 
423  return None;
424 }
425 
426 // Forward declare GeneratedRTChecks.
427 class GeneratedRTChecks;
428 
429 namespace llvm {
430 
431 /// InnerLoopVectorizer vectorizes loops which contain only one basic
432 /// block to a specified vectorization factor (VF).
433 /// This class performs the widening of scalars into vectors, or multiple
434 /// scalars. This class also implements the following features:
435 /// * It inserts an epilogue loop for handling loops that don't have iteration
436 /// counts that are known to be a multiple of the vectorization factor.
437 /// * It handles the code generation for reduction variables.
438 /// * Scalarization (implementation using scalars) of un-vectorizable
439 /// instructions.
440 /// InnerLoopVectorizer does not perform any vectorization-legality
441 /// checks, and relies on the caller to check for the different legality
442 /// aspects. The InnerLoopVectorizer relies on the
443 /// LoopVectorizationLegality class to provide information about the induction
444 /// and reduction variables that were found to a given vectorization factor.
446 public:
449  const TargetLibraryInfo *TLI,
452  unsigned UnrollFactor, LoopVectorizationLegality *LVL,
455  : OrigLoop(OrigLoop), PSE(PSE), LI(LI), DT(DT), TLI(TLI), TTI(TTI),
456  AC(AC), ORE(ORE), VF(VecWidth), UF(UnrollFactor),
457  Builder(PSE.getSE()->getContext()), Legal(LVL), Cost(CM), BFI(BFI),
458  PSI(PSI), RTChecks(RTChecks) {
459  // Query this against the original loop and save it here because the profile
460  // of the original loop header may change as the transformation happens.
463  }
464 
465  virtual ~InnerLoopVectorizer() = default;
466 
467  /// Create a new empty loop that will contain vectorized instructions later
468  /// on, while the old loop will be used as the scalar remainder. Control flow
469  /// is generated around the vectorized (and scalar epilogue) loops consisting
470  /// of various checks and bypasses. Return the pre-header block of the new
471  /// loop.
472  /// In the case of epilogue vectorization, this function is overriden to
473  /// handle the more complex control flow around the loops.
475 
476  /// Widen a single instruction within the innermost loop.
478  VPTransformState &State);
479 
480  /// Widen a single call instruction within the innermost loop.
481  void widenCallInstruction(CallInst &I, VPValue *Def, VPUser &ArgOperands,
482  VPTransformState &State);
483 
484  /// Widen a single select instruction within the innermost loop.
486  bool InvariantCond, VPTransformState &State);
487 
488  /// Fix the vectorized code, taking care of header phi's, live-outs, and more.
489  void fixVectorizedLoop(VPTransformState &State);
490 
491  // Return true if any runtime check is added.
493 
494  /// A type for vectorized values in the new loop. Each value from the
495  /// original loop, when vectorized, is represented by UF vector values in the
496  /// new unrolled loop, where UF is the unroll factor.
498 
499  /// Vectorize a single GetElementPtrInst based on information gathered and
500  /// decisions taken during planning.
501  void widenGEP(GetElementPtrInst *GEP, VPValue *VPDef, VPUser &Indices,
502  unsigned UF, ElementCount VF, bool IsPtrLoopInvariant,
503  SmallBitVector &IsIndexLoopInvariant, VPTransformState &State);
504 
505  /// Vectorize a single first-order recurrence or pointer induction PHINode in
506  /// a block. This method handles the induction variable canonicalization. It
507  /// supports both VF = 1 for unrolled loops and arbitrary length vectors.
509  VPTransformState &State);
510 
511  /// A helper function to scalarize a single Instruction in the innermost loop.
512  /// Generates a sequence of scalar instances for each lane between \p MinLane
513  /// and \p MaxLane, times each part between \p MinPart and \p MaxPart,
514  /// inclusive. Uses the VPValue operands from \p Operands instead of \p
515  /// Instr's operands.
517  const VPIteration &Instance, bool IfPredicateInstr,
518  VPTransformState &State);
519 
520  /// Widen an integer or floating-point induction variable \p IV. If \p Trunc
521  /// is provided, the integer induction variable will first be truncated to
522  /// the corresponding type.
523  void widenIntOrFpInduction(PHINode *IV, Value *Start, TruncInst *Trunc,
524  VPValue *Def, VPValue *CastDef,
525  VPTransformState &State);
526 
527  /// Construct the vector value of a scalarized value \p V one lane at a time.
528  void packScalarIntoVectorValue(VPValue *Def, const VPIteration &Instance,
529  VPTransformState &State);
530 
531  /// Try to vectorize interleaved access group \p Group with the base address
532  /// given in \p Addr, optionally masking the vector operations if \p
533  /// BlockInMask is non-null. Use \p State to translate given VPValues to IR
534  /// values in the vectorized loop.
536  ArrayRef<VPValue *> VPDefs,
537  VPTransformState &State, VPValue *Addr,
538  ArrayRef<VPValue *> StoredValues,
539  VPValue *BlockInMask = nullptr);
540 
541  /// Vectorize Load and Store instructions with the base address given in \p
542  /// Addr, optionally masking the vector operations if \p BlockInMask is
543  /// non-null. Use \p State to translate given VPValues to IR values in the
544  /// vectorized loop.
546  VPValue *Def, VPValue *Addr,
547  VPValue *StoredValue, VPValue *BlockInMask,
548  bool ConsecutiveStride, bool Reverse);
549 
550  /// Set the debug location in the builder \p Ptr using the debug location in
551  /// \p V. If \p Ptr is None then it uses the class member's Builder.
552  void setDebugLocFromInst(const Value *V,
553  Optional<IRBuilder<> *> CustomBuilder = None);
554 
555  /// Fix the non-induction PHIs in the OrigPHIsToFix vector.
557 
558  /// Returns true if the reordering of FP operations is not allowed, but we are
559  /// able to vectorize with strict in-order reductions for the given RdxDesc.
561 
562  /// Create a broadcast instruction. This method generates a broadcast
563  /// instruction (shuffle) for loop invariant values and for the induction
564  /// value. If this is the induction variable then we extend it to N, N+1, ...
565  /// this is needed because each iteration in the loop corresponds to a SIMD
566  /// element.
567  virtual Value *getBroadcastInstrs(Value *V);
568 
569 protected:
571 
572  /// A small list of PHINodes.
574 
575  /// A type for scalarized values in the new loop. Each value from the
576  /// original loop, when scalarized, is represented by UF x VF scalar values
577  /// in the new unrolled loop, where UF is the unroll factor and VF is the
578  /// vectorization factor.
580 
581  /// Set up the values of the IVs correctly when exiting the vector loop.
582  void fixupIVUsers(PHINode *OrigPhi, const InductionDescriptor &II,
583  Value *CountRoundDown, Value *EndValue,
584  BasicBlock *MiddleBlock);
585 
586  /// Create a new induction variable inside L.
587  PHINode *createInductionVariable(Loop *L, Value *Start, Value *End,
588  Value *Step, Instruction *DL);
589 
590  /// Handle all cross-iteration phis in the header.
592 
593  /// Create the exit value of first order recurrences in the middle block and
594  /// update their users.
596 
597  /// Create code for the loop exit value of the reduction.
599 
600  /// Clear NSW/NUW flags from reduction instructions if necessary.
601  void clearReductionWrapFlags(const RecurrenceDescriptor &RdxDesc,
602  VPTransformState &State);
603 
604  /// Fixup the LCSSA phi nodes in the unique exit block. This simply
605  /// means we need to add the appropriate incoming value from the middle
606  /// block as exiting edges from the scalar epilogue loop (if present) are
607  /// already in place, and we exit the vector loop exclusively to the middle
608  /// block.
609  void fixLCSSAPHIs(VPTransformState &State);
610 
611  /// Iteratively sink the scalarized operands of a predicated instruction into
612  /// the block that was created for it.
613  void sinkScalarOperands(Instruction *PredInst);
614 
615  /// Shrinks vector element sizes to the smallest bitwidth they can be legally
616  /// represented as.
618 
619  /// This function adds
620  /// (StartIdx * Step, (StartIdx + 1) * Step, (StartIdx + 2) * Step, ...)
621  /// to each vector element of Val. The sequence starts at StartIndex.
622  /// \p Opcode is relevant for FP induction variable.
623  virtual Value *getStepVector(Value *Val, int StartIdx, Value *Step,
624  Instruction::BinaryOps Opcode =
625  Instruction::BinaryOpsEnd);
626 
627  /// Compute scalar induction steps. \p ScalarIV is the scalar induction
628  /// variable on which to base the steps, \p Step is the size of the step, and
629  /// \p EntryVal is the value from the original loop that maps to the steps.
630  /// Note that \p EntryVal doesn't have to be an induction variable - it
631  /// can also be a truncate instruction.
632  void buildScalarSteps(Value *ScalarIV, Value *Step, Instruction *EntryVal,
634  VPValue *CastDef, VPTransformState &State);
635 
636  /// Create a vector induction phi node based on an existing scalar one. \p
637  /// EntryVal is the value from the original loop that maps to the vector phi
638  /// node, and \p Step is the loop-invariant step. If \p EntryVal is a
639  /// truncate instruction, instead of widening the original IV, we widen a
640  /// version of the IV truncated to \p EntryVal's type.
642  Value *Step, Value *Start,
643  Instruction *EntryVal, VPValue *Def,
644  VPValue *CastDef,
645  VPTransformState &State);
646 
647  /// Returns true if an instruction \p I should be scalarized instead of
648  /// vectorized for the chosen vectorization factor.
650 
651  /// Returns true if we should generate a scalar version of \p IV.
652  bool needsScalarInduction(Instruction *IV) const;
653 
654  /// If there is a cast involved in the induction variable \p ID, which should
655  /// be ignored in the vectorized loop body, this function records the
656  /// VectorLoopValue of the respective Phi also as the VectorLoopValue of the
657  /// cast. We had already proved that the casted Phi is equal to the uncasted
658  /// Phi in the vectorized loop (under a runtime guard), and therefore
659  /// there is no need to vectorize the cast - the same value can be used in the
660  /// vector loop for both the Phi and the cast.
661  /// If \p VectorLoopValue is a scalarized value, \p Lane is also specified,
662  /// Otherwise, \p VectorLoopValue is a widened/vectorized value.
663  ///
664  /// \p EntryVal is the value from the original loop that maps to the vector
665  /// phi node and is used to distinguish what is the IV currently being
666  /// processed - original one (if \p EntryVal is a phi corresponding to the
667  /// original IV) or the "newly-created" one based on the proof mentioned above
668  /// (see also buildScalarSteps() and createVectorIntOrFPInductionPHI()). In the
669  /// latter case \p EntryVal is a TruncInst and we must not record anything for
670  /// that IV, but it's error-prone to expect callers of this routine to care
671  /// about that, hence this explicit parameter.
673  const InductionDescriptor &ID, const Instruction *EntryVal,
674  Value *VectorLoopValue, VPValue *CastDef, VPTransformState &State,
675  unsigned Part, unsigned Lane = UINT_MAX);
676 
677  /// Generate a shuffle sequence that will reverse the vector Vec.
678  virtual Value *reverseVector(Value *Vec);
679 
680  /// Returns (and creates if needed) the original loop trip count.
681  Value *getOrCreateTripCount(Loop *NewLoop);
682 
683  /// Returns (and creates if needed) the trip count of the widened loop.
685 
686  /// Returns a bitcasted value to the requested vector type.
687  /// Also handles bitcasts of vector<float> <-> vector<pointer> types.
689  const DataLayout &DL);
690 
691  /// Emit a bypass check to see if the vector trip count is zero, including if
692  /// it overflows.
694 
695  /// Emit a bypass check to see if all of the SCEV assumptions we've
696  /// had to make are correct. Returns the block containing the checks or
697  /// nullptr if no checks have been added.
699 
700  /// Emit bypass checks to check any memory assumptions we may have made.
701  /// Returns the block containing the checks or nullptr if no checks have been
702  /// added.
704 
705  /// Compute the transformed value of Index at offset StartValue using step
706  /// StepValue.
707  /// For integer induction, returns StartValue + Index * StepValue.
708  /// For pointer induction, returns StartValue[Index * StepValue].
709  /// FIXME: The newly created binary instructions should contain nsw/nuw
710  /// flags, which can be found from the original scalar operations.
712  const DataLayout &DL,
713  const InductionDescriptor &ID) const;
714 
715  /// Emit basic blocks (prefixed with \p Prefix) for the iteration check,
716  /// vector loop preheader, middle block and scalar preheader. Also
717  /// allocate a loop object for the new vector loop and return it.
719 
720  /// Create new phi nodes for the induction variables to resume iteration count
721  /// in the scalar epilogue, from where the vectorized loop left off (given by
722  /// \p VectorTripCount).
723  /// In cases where the loop skeleton is more complicated (eg. epilogue
724  /// vectorization) and the resume values can come from an additional bypass
725  /// block, the \p AdditionalBypass pair provides information about the bypass
726  /// block and the end value on the edge from bypass to this loop.
729  std::pair<BasicBlock *, Value *> AdditionalBypass = {nullptr, nullptr});
730 
731  /// Complete the loop skeleton by adding debug MDs, creating appropriate
732  /// conditional branches in the middle block, preparing the builder and
733  /// running the verifier. Take in the vector loop \p L as argument, and return
734  /// the preheader of the completed vector loop.
735  BasicBlock *completeLoopSkeleton(Loop *L, MDNode *OrigLoopID);
736 
737  /// Add additional metadata to \p To that was not present on \p Orig.
738  ///
739  /// Currently this is used to add the noalias annotations based on the
740  /// inserted memchecks. Use this for instructions that are *cloned* into the
741  /// vector loop.
742  void addNewMetadata(Instruction *To, const Instruction *Orig);
743 
744  /// Add metadata from one instruction to another.
745  ///
746  /// This includes both the original MDs from \p From and additional ones (\see
747  /// addNewMetadata). Use this for *newly created* instructions in the vector
748  /// loop.
750 
751  /// Similar to the previous function but it adds the metadata to a
752  /// vector of instructions.
754 
755  /// Allow subclasses to override and print debug traces before/after vplan
756  /// execution, when trace information is requested.
757  virtual void printDebugTracesAtStart(){};
758  virtual void printDebugTracesAtEnd(){};
759 
760  /// The original loop.
761  Loop *OrigLoop;
762 
763  /// A wrapper around ScalarEvolution used to add runtime SCEV checks. Applies
764  /// dynamic knowledge to simplify SCEV expressions and converts them to a
765  /// more usable form.
767 
768  /// Loop Info.
770 
771  /// Dominator Tree.
773 
774  /// Alias Analysis.
776 
777  /// Target Library Info.
779 
780  /// Target Transform Info.
782 
783  /// Assumption Cache.
785 
786  /// Interface to emit optimization remarks.
788 
789  /// LoopVersioning. It's only set up (non-null) if memchecks were
790  /// used.
791  ///
792  /// This is currently only used to add no-alias metadata based on the
793  /// memchecks. The actually versioning is performed manually.
794  std::unique_ptr<LoopVersioning> LVer;
795 
796  /// The vectorization SIMD factor to use. Each vector will have this many
797  /// vector elements.
799 
800  /// The vectorization unroll factor to use. Each scalar is vectorized to this
801  /// many different vector instructions.
802  unsigned UF;
803 
804  /// The builder that we use
806 
807  // --- Vectorization state ---
808 
809  /// The vector-loop preheader.
811 
812  /// The scalar-loop preheader.
814 
815  /// Middle Block between the vector and the scalar.
817 
818  /// The unique ExitBlock of the scalar loop if one exists. Note that
819  /// there can be multiple exiting edges reaching this block.
821 
822  /// The vector loop body.
824 
825  /// The scalar loop body.
827 
828  /// A list of all bypass blocks. The first block is the entry of the loop.
830 
831  /// The new Induction variable which was added to the new block.
832  PHINode *Induction = nullptr;
833 
834  /// The induction variable of the old basic block.
835  PHINode *OldInduction = nullptr;
836 
837  /// Store instructions that were predicated.
839 
840  /// Trip count of the original loop.
841  Value *TripCount = nullptr;
842 
843  /// Trip count of the widened loop (TripCount - TripCount % (VF*UF))
844  Value *VectorTripCount = nullptr;
845 
846  /// The legality analysis.
848 
849  /// The profitablity analysis.
851 
852  // Record whether runtime checks are added.
853  bool AddedSafetyChecks = false;
854 
855  // Holds the end values for each induction variable. We save the end values
856  // so we can later fix-up the external users of the induction variables.
858 
859  // Vector of original scalar PHIs whose corresponding widened PHIs need to be
860  // fixed up at the end of vector code generation.
862 
863  /// BFI and PSI are used to check for profile guided size optimizations.
866 
867  // Whether this loop should be optimized for size based on profile guided size
868  // optimizatios.
870 
871  /// Structure to hold information about generated runtime checks, responsible
872  /// for cleaning the checks, if vectorization turns out unprofitable.
874 };
875 
877 public:
880  const TargetLibraryInfo *TLI,
882  OptimizationRemarkEmitter *ORE, unsigned UnrollFactor,
887  ElementCount::getFixed(1), UnrollFactor, LVL, CM,
888  BFI, PSI, Check) {}
889 
890 private:
891  Value *getBroadcastInstrs(Value *V) override;
892  Value *getStepVector(Value *Val, int StartIdx, Value *Step,
893  Instruction::BinaryOps Opcode =
894  Instruction::BinaryOpsEnd) override;
895  Value *reverseVector(Value *Vec) override;
896 };
897 
898 /// Encapsulate information regarding vectorization of a loop and its epilogue.
899 /// This information is meant to be updated and used across two stages of
900 /// epilogue vectorization.
903  unsigned MainLoopUF = 0;
905  unsigned EpilogueUF = 0;
910  Value *TripCount = nullptr;
911  Value *VectorTripCount = nullptr;
912 
914  ElementCount EVF, unsigned EUF)
915  : MainLoopVF(MVF), MainLoopUF(MUF), EpilogueVF(EVF), EpilogueUF(EUF) {
916  assert(EUF == 1 &&
917  "A high UF for the epilogue loop is likely not beneficial.");
918  }
919 };
920 
921 /// An extension of the inner loop vectorizer that creates a skeleton for a
922 /// vectorized loop that has its epilogue (residual) also vectorized.
923 /// The idea is to run the vplan on a given loop twice, firstly to setup the
924 /// skeleton and vectorize the main loop, and secondly to complete the skeleton
925 /// from the first step and vectorize the epilogue. This is achieved by
926 /// deriving two concrete strategy classes from this base class and invoking
927 /// them in succession from the loop vectorizer planner.
929 public:
937  GeneratedRTChecks &Checks)
939  EPI.MainLoopVF, EPI.MainLoopUF, LVL, CM, BFI, PSI,
940  Checks),
941  EPI(EPI) {}
942 
943  // Override this function to handle the more complex control flow around the
944  // three loops.
947  }
948 
949  /// The interface for creating a vectorized skeleton using one of two
950  /// different strategies, each corresponding to one execution of the vplan
951  /// as described above.
953 
954  /// Holds and updates state information required to vectorize the main loop
955  /// and its epilogue in two separate passes. This setup helps us avoid
956  /// regenerating and recomputing runtime safety checks. It also helps us to
957  /// shorten the iteration-count-check path length for the cases where the
958  /// iteration count of the loop is so small that the main vector loop is
959  /// completely skipped.
961 };
962 
963 /// A specialized derived class of inner loop vectorizer that performs
964 /// vectorization of *main* loops in the process of vectorizing loops and their
965 /// epilogues.
967 public:
977  EPI, LVL, CM, BFI, PSI, Check) {}
978  /// Implements the interface for creating a vectorized skeleton using the
979  /// *main loop* strategy (ie the first pass of vplan execution).
981 
982 protected:
983  /// Emits an iteration count bypass check once for the main loop (when \p
984  /// ForEpilogue is false) and once for the epilogue loop (when \p
985  /// ForEpilogue is true).
987  bool ForEpilogue);
988  void printDebugTracesAtStart() override;
989  void printDebugTracesAtEnd() override;
990 };
991 
992 // A specialized derived class of inner loop vectorizer that performs
993 // vectorization of *epilogue* loops in the process of vectorizing loops and
994 // their epilogues.
996 public:
1004  GeneratedRTChecks &Checks)
1006  EPI, LVL, CM, BFI, PSI, Checks) {}
1007  /// Implements the interface for creating a vectorized skeleton using the
1008  /// *epilogue loop* strategy (ie the second pass of vplan execution).
1010 
1011 protected:
1012  /// Emits an iteration count bypass check after the main vector loop has
1013  /// finished to see if there are any iterations left to execute by either
1014  /// the vector epilogue or the scalar epilogue.
1015  BasicBlock *emitMinimumVectorEpilogueIterCountCheck(Loop *L,
1016  BasicBlock *Bypass,
1017  BasicBlock *Insert);
1018  void printDebugTracesAtStart() override;
1019  void printDebugTracesAtEnd() override;
1020 };
1021 } // end namespace llvm
1022 
1023 /// Look for a meaningful debug location on the instruction or it's
1024 /// operands.
1026  if (!I)
1027  return I;
1028 
1029  DebugLoc Empty;
1030  if (I->getDebugLoc() != Empty)
1031  return I;
1032 
1033  for (Use &Op : I->operands()) {
1034  if (Instruction *OpInst = dyn_cast<Instruction>(Op))
1035  if (OpInst->getDebugLoc() != Empty)
1036  return OpInst;
1037  }
1038 
1039  return I;
1040 }
1041 
1043  const Value *V, Optional<IRBuilder<> *> CustomBuilder) {
1044  IRBuilder<> *B = (CustomBuilder == None) ? &Builder : *CustomBuilder;
1045  if (const Instruction *Inst = dyn_cast_or_null<Instruction>(V)) {
1046  const DILocation *DIL = Inst->getDebugLoc();
1047 
1048  // When a FSDiscriminator is enabled, we don't need to add the multiply
1049  // factors to the discriminators.
1050  if (DIL && Inst->getFunction()->isDebugInfoForProfiling() &&
1051  !isa<DbgInfoIntrinsic>(Inst) && !EnableFSDiscriminator) {
1052  // FIXME: For scalable vectors, assume vscale=1.
1053  auto NewDIL =
1055  if (NewDIL)
1056  B->SetCurrentDebugLocation(NewDIL.getValue());
1057  else
1058  LLVM_DEBUG(dbgs()
1059  << "Failed to create new discriminator: "
1060  << DIL->getFilename() << " Line: " << DIL->getLine());
1061  } else
1062  B->SetCurrentDebugLocation(DIL);
1063  } else
1064  B->SetCurrentDebugLocation(DebugLoc());
1065 }
1066 
1067 /// Write a \p DebugMsg about vectorization to the debug output stream. If \p I
1068 /// is passed, the message relates to that particular instruction.
1069 #ifndef NDEBUG
1071  const StringRef DebugMsg,
1072  Instruction *I) {
1073  dbgs() << "LV: " << Prefix << DebugMsg;
1074  if (I != nullptr)
1075  dbgs() << " " << *I;
1076  else
1077  dbgs() << '.';
1078  dbgs() << '\n';
1079 }
1080 #endif
1081 
1082 /// Create an analysis remark that explains why vectorization failed
1083 ///
1084 /// \p PassName is the name of the pass (e.g. can be AlwaysPrint). \p
1085 /// RemarkName is the identifier for the remark. If \p I is passed it is an
1086 /// instruction that prevents vectorization. Otherwise \p TheLoop is used for
1087 /// the location of the remark. \return the remark object that can be
1088 /// streamed to.
1090  StringRef RemarkName, Loop *TheLoop, Instruction *I) {
1091  Value *CodeRegion = TheLoop->getHeader();
1092  DebugLoc DL = TheLoop->getStartLoc();
1093 
1094  if (I) {
1095  CodeRegion = I->getParent();
1096  // If there is no debug location attached to the instruction, revert back to
1097  // using the loop's.
1098  if (I->getDebugLoc())
1099  DL = I->getDebugLoc();
1100  }
1101 
1102  return OptimizationRemarkAnalysis(PassName, RemarkName, DL, CodeRegion);
1103 }
1104 
1105 /// Return a value for Step multiplied by VF.
1107  assert(isa<ConstantInt>(Step) && "Expected an integer step");
1108  Constant *StepVal = ConstantInt::get(
1109  Step->getType(),
1110  cast<ConstantInt>(Step)->getSExtValue() * VF.getKnownMinValue());
1111  return VF.isScalable() ? B.CreateVScale(StepVal) : StepVal;
1112 }
1113 
1114 namespace llvm {
1115 
1116 /// Return the runtime value for VF.
1119  return VF.isScalable() ? B.CreateVScale(EC) : EC;
1120 }
1121 
1123  const StringRef OREMsg, const StringRef ORETag,
1124  OptimizationRemarkEmitter *ORE, Loop *TheLoop,
1125  Instruction *I) {
1126  LLVM_DEBUG(debugVectorizationMessage("Not vectorizing: ", DebugMsg, I));
1127  LoopVectorizeHints Hints(TheLoop, true /* doesn't matter */, *ORE);
1128  ORE->emit(
1129  createLVAnalysis(Hints.vectorizeAnalysisPassName(), ORETag, TheLoop, I)
1130  << "loop not vectorized: " << OREMsg);
1131 }
1132 
1133 void reportVectorizationInfo(const StringRef Msg, const StringRef ORETag,
1134  OptimizationRemarkEmitter *ORE, Loop *TheLoop,
1135  Instruction *I) {
1137  LoopVectorizeHints Hints(TheLoop, true /* doesn't matter */, *ORE);
1138  ORE->emit(
1139  createLVAnalysis(Hints.vectorizeAnalysisPassName(), ORETag, TheLoop, I)
1140  << Msg);
1141 }
1142 
1143 } // end namespace llvm
1144 
1145 #ifndef NDEBUG
1146 /// \return string containing a file name and a line # for the given loop.
1147 static std::string getDebugLocString(const Loop *L) {
1148  std::string Result;
1149  if (L) {
1150  raw_string_ostream OS(Result);
1151  if (const DebugLoc LoopDbgLoc = L->getStartLoc())
1152  LoopDbgLoc.print(OS);
1153  else
1154  // Just print the module name.
1155  OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier();
1156  OS.flush();
1157  }
1158  return Result;
1159 }
1160 #endif
1161 
1163  const Instruction *Orig) {
1164  // If the loop was versioned with memchecks, add the corresponding no-alias
1165  // metadata.
1166  if (LVer && (isa<LoadInst>(Orig) || isa<StoreInst>(Orig)))
1167  LVer->annotateInstWithNoAlias(To, Orig);
1168 }
1169 
1171  Instruction *From) {
1172  propagateMetadata(To, From);
1173  addNewMetadata(To, From);
1174 }
1175 
1177  Instruction *From) {
1178  for (Value *V : To) {
1179  if (Instruction *I = dyn_cast<Instruction>(V))
1180  addMetadata(I, From);
1181  }
1182 }
1183 
1184 namespace llvm {
1185 
1186 // Loop vectorization cost-model hints how the scalar epilogue loop should be
1187 // lowered.
1189 
1190  // The default: allowing scalar epilogues.
1192 
1193  // Vectorization with OptForSize: don't allow epilogues.
1195 
1196  // A special case of vectorisation with OptForSize: loops with a very small
1197  // trip count are considered for vectorization under OptForSize, thereby
1198  // making sure the cost of their loop body is dominant, free of runtime
1199  // guards and scalar iteration overheads.
1201 
1202  // Loop hint predicate indicating an epilogue is undesired.
1204 
1205  // Directive indicating we must either tail fold or not vectorize
1207 };
1208 
1209 /// ElementCountComparator creates a total ordering for ElementCount
1210 /// for the purposes of using it in a set structure.
1212  bool operator()(const ElementCount &LHS, const ElementCount &RHS) const {
1213  return std::make_tuple(LHS.isScalable(), LHS.getKnownMinValue()) <
1214  std::make_tuple(RHS.isScalable(), RHS.getKnownMinValue());
1215  }
1216 };
1218 
1219 /// LoopVectorizationCostModel - estimates the expected speedups due to
1220 /// vectorization.
1221 /// In many cases vectorization is not profitable. This can happen because of
1222 /// a number of reasons. In this class we mainly attempt to predict the
1223 /// expected speedup/slowdowns due to the supported instruction set. We use the
1224 /// TargetTransformInfo to query the different backends for the cost of
1225 /// different operations.
1227 public:
1231  const TargetTransformInfo &TTI,
1232  const TargetLibraryInfo *TLI, DemandedBits *DB,
1235  const LoopVectorizeHints *Hints,
1236  InterleavedAccessInfo &IAI)
1237  : ScalarEpilogueStatus(SEL), TheLoop(L), PSE(PSE), LI(LI), Legal(Legal),
1238  TTI(TTI), TLI(TLI), DB(DB), AC(AC), ORE(ORE), TheFunction(F),
1239  Hints(Hints), InterleaveInfo(IAI) {}
1240 
1241  /// \return An upper bound for the vectorization factors (both fixed and
1242  /// scalable). If the factors are 0, vectorization and interleaving should be
1243  /// avoided up front.
1244  FixedScalableVFPair computeMaxVF(ElementCount UserVF, unsigned UserIC);
1245 
1246  /// \return True if runtime checks are required for vectorization, and false
1247  /// otherwise.
1248  bool runtimeChecksRequired();
1249 
1250  /// \return The most profitable vectorization factor and the cost of that VF.
1251  /// This method checks every VF in \p CandidateVFs. If UserVF is not ZERO
1252  /// then this vectorization factor will be selected if vectorization is
1253  /// possible.
1255  selectVectorizationFactor(const ElementCountSet &CandidateVFs);
1256 
1258  selectEpilogueVectorizationFactor(const ElementCount MaxVF,
1259  const LoopVectorizationPlanner &LVP);
1260 
1261  /// Setup cost-based decisions for user vectorization factor.
1262  /// \return true if the UserVF is a feasible VF to be chosen.
1264  collectUniformsAndScalars(UserVF);
1265  collectInstsToScalarize(UserVF);
1266  return expectedCost(UserVF).first.isValid();
1267  }
1268 
1269  /// \return The size (in bits) of the smallest and widest types in the code
1270  /// that needs to be vectorized. We ignore values that remain scalar such as
1271  /// 64 bit loop indices.
1272  std::pair<unsigned, unsigned> getSmallestAndWidestTypes();
1273 
1274  /// \return The desired interleave count.
1275  /// If interleave count has been specified by metadata it will be returned.
1276  /// Otherwise, the interleave count is computed and returned. VF and LoopCost
1277  /// are the selected vectorization factor and the cost of the selected VF.
1278  unsigned selectInterleaveCount(ElementCount VF, unsigned LoopCost);
1279 
1280  /// Memory access instruction may be vectorized in more than one way.
1281  /// Form of instruction after vectorization depends on cost.
1282  /// This function takes cost-based decisions for Load/Store instructions
1283  /// and collects them in a map. This decisions map is used for building
1284  /// the lists of loop-uniform and loop-scalar instructions.
1285  /// The calculated cost is saved with widening decision in order to
1286  /// avoid redundant calculations.
1287  void setCostBasedWideningDecision(ElementCount VF);
1288 
1289  /// A struct that represents some properties of the register usage
1290  /// of a loop.
1291  struct RegisterUsage {
1292  /// Holds the number of loop invariant values that are used in the loop.
1293  /// The key is ClassID of target-provided register class.
1295  /// Holds the maximum number of concurrent live intervals in the loop.
1296  /// The key is ClassID of target-provided register class.
1298  };
1299 
1300  /// \return Returns information about the register usages of the loop for the
1301  /// given vectorization factors.
1303  calculateRegisterUsage(ArrayRef<ElementCount> VFs);
1304 
1305  /// Collect values we want to ignore in the cost model.
1306  void collectValuesToIgnore();
1307 
1308  /// Collect all element types in the loop for which widening is needed.
1309  void collectElementTypesForWidening();
1310 
1311  /// Split reductions into those that happen in the loop, and those that happen
1312  /// outside. In loop reductions are collected into InLoopReductionChains.
1313  void collectInLoopReductions();
1314 
1315  /// Returns true if we should use strict in-order reductions for the given
1316  /// RdxDesc. This is true if the -enable-strict-reductions flag is passed,
1317  /// the IsOrdered flag of RdxDesc is set and we do not allow reordering
1318  /// of FP operations.
1320  return !Hints->allowReordering() && RdxDesc.isOrdered();
1321  }
1322 
1323  /// \returns The smallest bitwidth each instruction can be represented with.
1324  /// The vector equivalents of these instructions should be truncated to this
1325  /// type.
1327  return MinBWs;
1328  }
1329 
1330  /// \returns True if it is more profitable to scalarize instruction \p I for
1331  /// vectorization factor \p VF.
1333  assert(VF.isVector() &&
1334  "Profitable to scalarize relevant only for VF > 1.");
1335 
1336  // Cost model is not run in the VPlan-native path - return conservative
1337  // result until this changes.
1339  return false;
1340 
1341  auto Scalars = InstsToScalarize.find(VF);
1342  assert(Scalars != InstsToScalarize.end() &&
1343  "VF not yet analyzed for scalarization profitability");
1344  return Scalars->second.find(I) != Scalars->second.end();
1345  }
1346 
1347  /// Returns true if \p I is known to be uniform after vectorization.
1349  if (VF.isScalar())
1350  return true;
1351 
1352  // Cost model is not run in the VPlan-native path - return conservative
1353  // result until this changes.
1355  return false;
1356 
1357  auto UniformsPerVF = Uniforms.find(VF);
1358  assert(UniformsPerVF != Uniforms.end() &&
1359  "VF not yet analyzed for uniformity");
1360  return UniformsPerVF->second.count(I);
1361  }
1362 
1363  /// Returns true if \p I is known to be scalar after vectorization.
1365  if (VF.isScalar())
1366  return true;
1367 
1368  // Cost model is not run in the VPlan-native path - return conservative
1369  // result until this changes.
1371  return false;
1372 
1373  auto ScalarsPerVF = Scalars.find(VF);
1374  assert(ScalarsPerVF != Scalars.end() &&
1375  "Scalar values are not calculated for VF");
1376  return ScalarsPerVF->second.count(I);
1377  }
1378 
1379  /// \returns True if instruction \p I can be truncated to a smaller bitwidth
1380  /// for vectorization factor \p VF.
1382  return VF.isVector() && MinBWs.find(I) != MinBWs.end() &&
1383  !isProfitableToScalarize(I, VF) &&
1384  !isScalarAfterVectorization(I, VF);
1385  }
1386 
1387  /// Decision that was taken during cost calculation for memory instruction.
1390  CM_Widen, // For consecutive accesses with stride +1.
1391  CM_Widen_Reverse, // For consecutive accesses with stride -1.
1394  CM_Scalarize
1395  };
1396 
1397  /// Save vectorization decision \p W and \p Cost taken by the cost model for
1398  /// instruction \p I and vector width \p VF.
1401  assert(VF.isVector() && "Expected VF >=2");
1402  WideningDecisions[std::make_pair(I, VF)] = std::make_pair(W, Cost);
1403  }
1404 
1405  /// Save vectorization decision \p W and \p Cost taken by the cost model for
1406  /// interleaving group \p Grp and vector width \p VF.
1410  assert(VF.isVector() && "Expected VF >=2");
1411  /// Broadcast this decicion to all instructions inside the group.
1412  /// But the cost will be assigned to one instruction only.
1413  for (unsigned i = 0; i < Grp->getFactor(); ++i) {
1414  if (auto *I = Grp->getMember(i)) {
1415  if (Grp->getInsertPos() == I)
1416  WideningDecisions[std::make_pair(I, VF)] = std::make_pair(W, Cost);
1417  else
1418  WideningDecisions[std::make_pair(I, VF)] = std::make_pair(W, 0);
1419  }
1420  }
1421  }
1422 
1423  /// Return the cost model decision for the given instruction \p I and vector
1424  /// width \p VF. Return CM_Unknown if this instruction did not pass
1425  /// through the cost modeling.
1427  assert(VF.isVector() && "Expected VF to be a vector VF");
1428  // Cost model is not run in the VPlan-native path - return conservative
1429  // result until this changes.
1431  return CM_GatherScatter;
1432 
1433  std::pair<Instruction *, ElementCount> InstOnVF = std::make_pair(I, VF);
1434  auto Itr = WideningDecisions.find(InstOnVF);
1435  if (Itr == WideningDecisions.end())
1436  return CM_Unknown;
1437  return Itr->second.first;
1438  }
1439 
1440  /// Return the vectorization cost for the given instruction \p I and vector
1441  /// width \p VF.
1443  assert(VF.isVector() && "Expected VF >=2");
1444  std::pair<Instruction *, ElementCount> InstOnVF = std::make_pair(I, VF);
1445  assert(WideningDecisions.find(InstOnVF) != WideningDecisions.end() &&
1446  "The cost is not calculated");
1447  return WideningDecisions[InstOnVF].second;
1448  }
1449 
1450  /// Return True if instruction \p I is an optimizable truncate whose operand
1451  /// is an induction variable. Such a truncate will be removed by adding a new
1452  /// induction variable with the destination type.
1454  // If the instruction is not a truncate, return false.
1455  auto *Trunc = dyn_cast<TruncInst>(I);
1456  if (!Trunc)
1457  return false;
1458 
1459  // Get the source and destination types of the truncate.
1460  Type *SrcTy = ToVectorTy(cast<CastInst>(I)->getSrcTy(), VF);
1461  Type *DestTy = ToVectorTy(cast<CastInst>(I)->getDestTy(), VF);
1462 
1463  // If the truncate is free for the given types, return false. Replacing a
1464  // free truncate with an induction variable would add an induction variable
1465  // update instruction to each iteration of the loop. We exclude from this
1466  // check the primary induction variable since it will need an update
1467  // instruction regardless.
1468  Value *Op = Trunc->getOperand(0);
1469  if (Op != Legal->getPrimaryInduction() && TTI.isTruncateFree(SrcTy, DestTy))
1470  return false;
1471 
1472  // If the truncated value is not an induction variable, return false.
1473  return Legal->isInductionPhi(Op);
1474  }
1475 
1476  /// Collects the instructions to scalarize for each predicated instruction in
1477  /// the loop.
1478  void collectInstsToScalarize(ElementCount VF);
1479 
1480  /// Collect Uniform and Scalar values for the given \p VF.
1481  /// The sets depend on CM decision for Load/Store instructions
1482  /// that may be vectorized as interleave, gather-scatter or scalarized.
1484  // Do the analysis once.
1485  if (VF.isScalar() || Uniforms.find(VF) != Uniforms.end())
1486  return;
1487  setCostBasedWideningDecision(VF);
1488  collectLoopUniforms(VF);
1489  collectLoopScalars(VF);
1490  }
1491 
1492  /// Returns true if the target machine supports masked store operation
1493  /// for the given \p DataType and kind of access to \p Ptr.
1494  bool isLegalMaskedStore(Type *DataType, Value *Ptr, Align Alignment) const {
1495  return Legal->isConsecutivePtr(DataType, Ptr) &&
1496  TTI.isLegalMaskedStore(DataType, Alignment);
1497  }
1498 
1499  /// Returns true if the target machine supports masked load operation
1500  /// for the given \p DataType and kind of access to \p Ptr.
1501  bool isLegalMaskedLoad(Type *DataType, Value *Ptr, Align Alignment) const {
1502  return Legal->isConsecutivePtr(DataType, Ptr) &&
1503  TTI.isLegalMaskedLoad(DataType, Alignment);
1504  }
1505 
1506  /// Returns true if the target machine can represent \p V as a masked gather
1507  /// or scatter operation.
1509  bool LI = isa<LoadInst>(V);
1510  bool SI = isa<StoreInst>(V);
1511  if (!LI && !SI)
1512  return false;
1513  auto *Ty = getLoadStoreType(V);
1515  return (LI && TTI.isLegalMaskedGather(Ty, Align)) ||
1516  (SI && TTI.isLegalMaskedScatter(Ty, Align));
1517  }
1518 
1519  /// Returns true if the target machine supports all of the reduction
1520  /// variables found for the given VF.
1522  return (all_of(Legal->getReductionVars(), [&](auto &Reduction) -> bool {
1523  const RecurrenceDescriptor &RdxDesc = Reduction.second;
1524  return TTI.isLegalToVectorizeReduction(RdxDesc, VF);
1525  }));
1526  }
1527 
1528  /// Returns true if \p I is an instruction that will be scalarized with
1529  /// predication. Such instructions include conditional stores and
1530  /// instructions that may divide by zero.
1531  /// If a non-zero VF has been calculated, we check if I will be scalarized
1532  /// predication for that VF.
1533  bool isScalarWithPredication(Instruction *I) const;
1534 
1535  // Returns true if \p I is an instruction that will be predicated either
1536  // through scalar predication or masked load/store or masked gather/scatter.
1537  // Superset of instructions that return true for isScalarWithPredication.
1539  if (!blockNeedsPredication(I->getParent()))
1540  return false;
1541  // Loads and stores that need some form of masked operation are predicated
1542  // instructions.
1543  if (isa<LoadInst>(I) || isa<StoreInst>(I))
1544  return Legal->isMaskRequired(I);
1545  return isScalarWithPredication(I);
1546  }
1547 
1548  /// Returns true if \p I is a memory instruction with consecutive memory
1549  /// access that can be widened.
1550  bool
1551  memoryInstructionCanBeWidened(Instruction *I,
1553 
1554  /// Returns true if \p I is a memory instruction in an interleaved-group
1555  /// of memory accesses that can be vectorized with wide vector loads/stores
1556  /// and shuffles.
1557  bool
1558  interleavedAccessCanBeWidened(Instruction *I,
1560 
1561  /// Check if \p Instr belongs to any interleaved access group.
1563  return InterleaveInfo.isInterleaved(Instr);
1564  }
1565 
1566  /// Get the interleaved access group that \p Instr belongs to.
1569  return InterleaveInfo.getInterleaveGroup(Instr);
1570  }
1571 
1572  /// Returns true if we're required to use a scalar epilogue for at least
1573  /// the final iteration of the original loop.
1575  if (!isScalarEpilogueAllowed())
1576  return false;
1577  // If we might exit from anywhere but the latch, must run the exiting
1578  // iteration in scalar form.
1579  if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch())
1580  return true;
1581  return VF.isVector() && InterleaveInfo.requiresScalarEpilogue();
1582  }
1583 
1584  /// Returns true if a scalar epilogue is not allowed due to optsize or a
1585  /// loop hint annotation.
1587  return ScalarEpilogueStatus == CM_ScalarEpilogueAllowed;
1588  }
1589 
1590  /// Returns true if all loop blocks should be masked to fold tail loop.
1591  bool foldTailByMasking() const { return FoldTailByMasking; }
1592 
1594  return foldTailByMasking() || Legal->blockNeedsPredication(BB);
1595  }
1596 
1597  /// A SmallMapVector to store the InLoop reduction op chains, mapping phi
1598  /// nodes to the chain of instructions representing the reductions. Uses a
1599  /// MapVector to ensure deterministic iteration order.
1600  using ReductionChainMap =
1602 
1603  /// Return the chain of instructions representing an inloop reduction.
1605  return InLoopReductionChains;
1606  }
1607 
1608  /// Returns true if the Phi is part of an inloop reduction.
1609  bool isInLoopReduction(PHINode *Phi) const {
1610  return InLoopReductionChains.count(Phi);
1611  }
1612 
1613  /// Estimate cost of an intrinsic call instruction CI if it were vectorized
1614  /// with factor VF. Return the cost of the instruction, including
1615  /// scalarization overhead if it's needed.
1616  InstructionCost getVectorIntrinsicCost(CallInst *CI, ElementCount VF) const;
1617 
1618  /// Estimate cost of a call instruction CI if it were vectorized with factor
1619  /// VF. Return the cost of the instruction, including scalarization overhead
1620  /// if it's needed. The flag NeedToScalarize shows if the call needs to be
1621  /// scalarized -
1622  /// i.e. either vector version isn't available, or is too expensive.
1623  InstructionCost getVectorCallCost(CallInst *CI, ElementCount VF,
1624  bool &NeedToScalarize) const;
1625 
1626  /// Returns true if the per-lane cost of VectorizationFactor A is lower than
1627  /// that of B.
1628  bool isMoreProfitable(const VectorizationFactor &A,
1629  const VectorizationFactor &B) const;
1630 
1631  /// Invalidates decisions already taken by the cost model.
1633  WideningDecisions.clear();
1634  Uniforms.clear();
1635  Scalars.clear();
1636  }
1637 
1638 private:
1639  unsigned NumPredStores = 0;
1640 
1641  /// \return An upper bound for the vectorization factors for both
1642  /// fixed and scalable vectorization, where the minimum-known number of
1643  /// elements is a power-of-2 larger than zero. If scalable vectorization is
1644  /// disabled or unsupported, then the scalable part will be equal to
1645  /// ElementCount::getScalable(0).
1646  FixedScalableVFPair computeFeasibleMaxVF(unsigned ConstTripCount,
1647  ElementCount UserVF);
1648 
1649  /// \return the maximized element count based on the targets vector
1650  /// registers and the loop trip-count, but limited to a maximum safe VF.
1651  /// This is a helper function of computeFeasibleMaxVF.
1652  /// FIXME: MaxSafeVF is currently passed by reference to avoid some obscure
1653  /// issue that occurred on one of the buildbots which cannot be reproduced
1654  /// without having access to the properietary compiler (see comments on
1655  /// D98509). The issue is currently under investigation and this workaround
1656  /// will be removed as soon as possible.
1657  ElementCount getMaximizedVFForTarget(unsigned ConstTripCount,
1658  unsigned SmallestType,
1659  unsigned WidestType,
1660  const ElementCount &MaxSafeVF);
1661 
1662  /// \return the maximum legal scalable VF, based on the safe max number
1663  /// of elements.
1664  ElementCount getMaxLegalScalableVF(unsigned MaxSafeElements);
1665 
1666  /// The vectorization cost is a combination of the cost itself and a boolean
1667  /// indicating whether any of the contributing operations will actually
1668  /// operate on vector values after type legalization in the backend. If this
1669  /// latter value is false, then all operations will be scalarized (i.e. no
1670  /// vectorization has actually taken place).
1671  using VectorizationCostTy = std::pair<InstructionCost, bool>;
1672 
1673  /// Returns the expected execution cost. The unit of the cost does
1674  /// not matter because we use the 'cost' units to compare different
1675  /// vector widths. The cost that is returned is *not* normalized by
1676  /// the factor width. If \p Invalid is not nullptr, this function
1677  /// will add a pair(Instruction*, ElementCount) to \p Invalid for
1678  /// each instruction that has an Invalid cost for the given VF.
1679  using InstructionVFPair = std::pair<Instruction *, ElementCount>;
1680  VectorizationCostTy
1681  expectedCost(ElementCount VF,
1682  SmallVectorImpl<InstructionVFPair> *Invalid = nullptr);
1683 
1684  /// Returns the execution time cost of an instruction for a given vector
1685  /// width. Vector width of one means scalar.
1686  VectorizationCostTy getInstructionCost(Instruction *I, ElementCount VF);
1687 
1688  /// The cost-computation logic from getInstructionCost which provides
1689  /// the vector type as an output parameter.
1690  InstructionCost getInstructionCost(Instruction *I, ElementCount VF,
1691  Type *&VectorTy);
1692 
1693  /// Return the cost of instructions in an inloop reduction pattern, if I is
1694  /// part of that pattern.
1696  getReductionPatternCost(Instruction *I, ElementCount VF, Type *VectorTy,
1698 
1699  /// Calculate vectorization cost of memory instruction \p I.
1700  InstructionCost getMemoryInstructionCost(Instruction *I, ElementCount VF);
1701 
1702  /// The cost computation for scalarized memory instruction.
1703  InstructionCost getMemInstScalarizationCost(Instruction *I, ElementCount VF);
1704 
1705  /// The cost computation for interleaving group of memory instructions.
1706  InstructionCost getInterleaveGroupCost(Instruction *I, ElementCount VF);
1707 
1708  /// The cost computation for Gather/Scatter instruction.
1709  InstructionCost getGatherScatterCost(Instruction *I, ElementCount VF);
1710 
1711  /// The cost computation for widening instruction \p I with consecutive
1712  /// memory access.
1713  InstructionCost getConsecutiveMemOpCost(Instruction *I, ElementCount VF);
1714 
1715  /// The cost calculation for Load/Store instruction \p I with uniform pointer -
1716  /// Load: scalar load + broadcast.
1717  /// Store: scalar store + (loop invariant value stored? 0 : extract of last
1718  /// element)
1719  InstructionCost getUniformMemOpCost(Instruction *I, ElementCount VF);
1720 
1721  /// Estimate the overhead of scalarizing an instruction. This is a
1722  /// convenience wrapper for the type-based getScalarizationOverhead API.
1723  InstructionCost getScalarizationOverhead(Instruction *I,
1724  ElementCount VF) const;
1725 
1726  /// Returns whether the instruction is a load or store and will be a emitted
1727  /// as a vector operation.
1728  bool isConsecutiveLoadOrStore(Instruction *I);
1729 
1730  /// Returns true if an artificially high cost for emulated masked memrefs
1731  /// should be used.
1732  bool useEmulatedMaskMemRefHack(Instruction *I);
1733 
1734  /// Map of scalar integer values to the smallest bitwidth they can be legally
1735  /// represented as. The vector equivalents of these values should be truncated
1736  /// to this type.
1738 
1739  /// A type representing the costs for instructions if they were to be
1740  /// scalarized rather than vectorized. The entries are Instruction-Cost
1741  /// pairs.
1742  using ScalarCostsTy = DenseMap<Instruction *, InstructionCost>;
1743 
1744  /// A set containing all BasicBlocks that are known to present after
1745  /// vectorization as a predicated block.
1746  SmallPtrSet<BasicBlock *, 4> PredicatedBBsAfterVectorization;
1747 
1748  /// Records whether it is allowed to have the original scalar loop execute at
1749  /// least once. This may be needed as a fallback loop in case runtime
1750  /// aliasing/dependence checks fail, or to handle the tail/remainder
1751  /// iterations when the trip count is unknown or doesn't divide by the VF,
1752  /// or as a peel-loop to handle gaps in interleave-groups.
1753  /// Under optsize and when the trip count is very small we don't allow any
1754  /// iterations to execute in the scalar loop.
1755  ScalarEpilogueLowering ScalarEpilogueStatus = CM_ScalarEpilogueAllowed;
1756 
1757  /// All blocks of loop are to be masked to fold tail of scalar iterations.
1758  bool FoldTailByMasking = false;
1759 
1760  /// A map holding scalar costs for different vectorization factors. The
1761  /// presence of a cost for an instruction in the mapping indicates that the
1762  /// instruction will be scalarized when vectorizing with the associated
1763  /// vectorization factor. The entries are VF-ScalarCostTy pairs.
1764  DenseMap<ElementCount, ScalarCostsTy> InstsToScalarize;
1765 
1766  /// Holds the instructions known to be uniform after vectorization.
1767  /// The data is collected per VF.
1769 
1770  /// Holds the instructions known to be scalar after vectorization.
1771  /// The data is collected per VF.
1773 
1774  /// Holds the instructions (address computations) that are forced to be
1775  /// scalarized.
1777 
1778  /// PHINodes of the reductions that should be expanded in-loop along with
1779  /// their associated chains of reduction operations, in program order from top
1780  /// (PHI) to bottom
1781  ReductionChainMap InLoopReductionChains;
1782 
1783  /// A Map of inloop reduction operations and their immediate chain operand.
1784  /// FIXME: This can be removed once reductions can be costed correctly in
1785  /// vplan. This was added to allow quick lookup to the inloop operations,
1786  /// without having to loop through InLoopReductionChains.
1787  DenseMap<Instruction *, Instruction *> InLoopReductionImmediateChains;
1788 
1789  /// Returns the expected difference in cost from scalarizing the expression
1790  /// feeding a predicated instruction \p PredInst. The instructions to
1791  /// scalarize and their scalar costs are collected in \p ScalarCosts. A
1792  /// non-negative return value implies the expression will be scalarized.
1793  /// Currently, only single-use chains are considered for scalarization.
1794  int computePredInstDiscount(Instruction *PredInst, ScalarCostsTy &ScalarCosts,
1795  ElementCount VF);
1796 
1797  /// Collect the instructions that are uniform after vectorization. An
1798  /// instruction is uniform if we represent it with a single scalar value in
1799  /// the vectorized loop corresponding to each vector iteration. Examples of
1800  /// uniform instructions include pointer operands of consecutive or
1801  /// interleaved memory accesses. Note that although uniformity implies an
1802  /// instruction will be scalar, the reverse is not true. In general, a
1803  /// scalarized instruction will be represented by VF scalar values in the
1804  /// vectorized loop, each corresponding to an iteration of the original
1805  /// scalar loop.
1806  void collectLoopUniforms(ElementCount VF);
1807 
1808  /// Collect the instructions that are scalar after vectorization. An
1809  /// instruction is scalar if it is known to be uniform or will be scalarized
1810  /// during vectorization. Non-uniform scalarized instructions will be
1811  /// represented by VF values in the vectorized loop, each corresponding to an
1812  /// iteration of the original scalar loop.
1813  void collectLoopScalars(ElementCount VF);
1814 
1815  /// Keeps cost model vectorization decision and cost for instructions.
1816  /// Right now it is used for memory instructions only.
1818  std::pair<InstWidening, InstructionCost>>;
1819 
1820  DecisionList WideningDecisions;
1821 
1822  /// Returns true if \p V is expected to be vectorized and it needs to be
1823  /// extracted.
1824  bool needsExtract(Value *V, ElementCount VF) const {
1825  Instruction *I = dyn_cast<Instruction>(V);
1826  if (VF.isScalar() || !I || !TheLoop->contains(I) ||
1827  TheLoop->isLoopInvariant(I))
1828  return false;
1829 
1830  // Assume we can vectorize V (and hence we need extraction) if the
1831  // scalars are not computed yet. This can happen, because it is called
1832  // via getScalarizationOverhead from setCostBasedWideningDecision, before
1833  // the scalars are collected. That should be a safe assumption in most
1834  // cases, because we check if the operands have vectorizable types
1835  // beforehand in LoopVectorizationLegality.
1836  return Scalars.find(VF) == Scalars.end() ||
1837  !isScalarAfterVectorization(I, VF);
1838  };
1839 
1840  /// Returns a range containing only operands needing to be extracted.
1841  SmallVector<Value *, 4> filterExtractingOperands(Instruction::op_range Ops,
1842  ElementCount VF) const {
1844  Ops, [this, VF](Value *V) { return this->needsExtract(V, VF); }));
1845  }
1846 
1847  /// Determines if we have the infrastructure to vectorize loop \p L and its
1848  /// epilogue, assuming the main loop is vectorized by \p VF.
1849  bool isCandidateForEpilogueVectorization(const Loop &L,
1850  const ElementCount VF) const;
1851 
1852  /// Returns true if epilogue vectorization is considered profitable, and
1853  /// false otherwise.
1854  /// \p VF is the vectorization factor chosen for the original loop.
1855  bool isEpilogueVectorizationProfitable(const ElementCount VF) const;
1856 
1857 public:
1858  /// The loop that we evaluate.
1860 
1861  /// Predicated scalar evolution analysis.
1863 
1864  /// Loop Info analysis.
1866 
1867  /// Vectorization legality.
1869 
1870  /// Vector target information.
1872 
1873  /// Target Library Info.
1875 
1876  /// Demanded bits analysis.
1878 
1879  /// Assumption cache.
1881 
1882  /// Interface to emit optimization remarks.
1884 
1886 
1887  /// Loop Vectorize Hint.
1889 
1890  /// The interleave access information contains groups of interleaved accesses
1891  /// with the same stride and close to each other.
1893 
1894  /// Values to ignore in the cost model.
1896 
1897  /// Values to ignore in the cost model when VF > 1.
1899 
1900  /// All element types found in the loop.
1902 
1903  /// Profitable vector factors.
1905 };
1906 } // end namespace llvm
1907 
1908 /// Helper struct to manage generating runtime checks for vectorization.
1909 ///
1910 /// The runtime checks are created up-front in temporary blocks to allow better
1911 /// estimating the cost and un-linked from the existing IR. After deciding to
1912 /// vectorize, the checks are moved back. If deciding not to vectorize, the
1913 /// temporary blocks are completely removed.
1915  /// Basic block which contains the generated SCEV checks, if any.
1916  BasicBlock *SCEVCheckBlock = nullptr;
1917 
1918  /// The value representing the result of the generated SCEV checks. If it is
1919  /// nullptr, either no SCEV checks have been generated or they have been used.
1920  Value *SCEVCheckCond = nullptr;
1921 
1922  /// Basic block which contains the generated memory runtime checks, if any.
1923  BasicBlock *MemCheckBlock = nullptr;
1924 
1925  /// The value representing the result of the generated memory runtime checks.
1926  /// If it is nullptr, either no memory runtime checks have been generated or
1927  /// they have been used.
1928  Value *MemRuntimeCheckCond = nullptr;
1929 
1930  DominatorTree *DT;
1931  LoopInfo *LI;
1932 
1933  SCEVExpander SCEVExp;
1934  SCEVExpander MemCheckExp;
1935 
1936 public:
1938  const DataLayout &DL)
1939  : DT(DT), LI(LI), SCEVExp(SE, DL, "scev.check"),
1940  MemCheckExp(SE, DL, "scev.check") {}
1941 
1942  /// Generate runtime checks in SCEVCheckBlock and MemCheckBlock, so we can
1943  /// accurately estimate the cost of the runtime checks. The blocks are
1944  /// un-linked from the IR and is added back during vector code generation. If
1945  /// there is no vector code generation, the check blocks are removed
1946  /// completely.
1947  void Create(Loop *L, const LoopAccessInfo &LAI,
1948  const SCEVUnionPredicate &UnionPred) {
1949 
1950  BasicBlock *LoopHeader = L->getHeader();
1951  BasicBlock *Preheader = L->getLoopPreheader();
1952 
1953  // Use SplitBlock to create blocks for SCEV & memory runtime checks to
1954  // ensure the blocks are properly added to LoopInfo & DominatorTree. Those
1955  // may be used by SCEVExpander. The blocks will be un-linked from their
1956  // predecessors and removed from LI & DT at the end of the function.
1957  if (!UnionPred.isAlwaysTrue()) {
1958  SCEVCheckBlock = SplitBlock(Preheader, Preheader->getTerminator(), DT, LI,
1959  nullptr, "vector.scevcheck");
1960 
1961  SCEVCheckCond = SCEVExp.expandCodeForPredicate(
1962  &UnionPred, SCEVCheckBlock->getTerminator());
1963  }
1964 
1965  const auto &RtPtrChecking = *LAI.getRuntimePointerChecking();
1966  if (RtPtrChecking.Need) {
1967  auto *Pred = SCEVCheckBlock ? SCEVCheckBlock : Preheader;
1968  MemCheckBlock = SplitBlock(Pred, Pred->getTerminator(), DT, LI, nullptr,
1969  "vector.memcheck");
1970 
1971  MemRuntimeCheckCond =
1972  addRuntimeChecks(MemCheckBlock->getTerminator(), L,
1973  RtPtrChecking.getChecks(), MemCheckExp);
1974  assert(MemRuntimeCheckCond &&
1975  "no RT checks generated although RtPtrChecking "
1976  "claimed checks are required");
1977  }
1978 
1979  if (!MemCheckBlock && !SCEVCheckBlock)
1980  return;
1981 
1982  // Unhook the temporary block with the checks, update various places
1983  // accordingly.
1984  if (SCEVCheckBlock)
1985  SCEVCheckBlock->replaceAllUsesWith(Preheader);
1986  if (MemCheckBlock)
1987  MemCheckBlock->replaceAllUsesWith(Preheader);
1988 
1989  if (SCEVCheckBlock) {
1990  SCEVCheckBlock->getTerminator()->moveBefore(Preheader->getTerminator());
1991  new UnreachableInst(Preheader->getContext(), SCEVCheckBlock);
1992  Preheader->getTerminator()->eraseFromParent();
1993  }
1994  if (MemCheckBlock) {
1995  MemCheckBlock->getTerminator()->moveBefore(Preheader->getTerminator());
1996  new UnreachableInst(Preheader->getContext(), MemCheckBlock);
1997  Preheader->getTerminator()->eraseFromParent();
1998  }
1999 
2000  DT->changeImmediateDominator(LoopHeader, Preheader);
2001  if (MemCheckBlock) {
2002  DT->eraseNode(MemCheckBlock);
2003  LI->removeBlock(MemCheckBlock);
2004  }
2005  if (SCEVCheckBlock) {
2006  DT->eraseNode(SCEVCheckBlock);
2007  LI->removeBlock(SCEVCheckBlock);
2008  }
2009  }
2010 
2011  /// Remove the created SCEV & memory runtime check blocks & instructions, if
2012  /// unused.
2014  SCEVExpanderCleaner SCEVCleaner(SCEVExp, *DT);
2015  SCEVExpanderCleaner MemCheckCleaner(MemCheckExp, *DT);
2016  if (!SCEVCheckCond)
2017  SCEVCleaner.markResultUsed();
2018 
2019  if (!MemRuntimeCheckCond)
2020  MemCheckCleaner.markResultUsed();
2021 
2022  if (MemRuntimeCheckCond) {
2023  auto &SE = *MemCheckExp.getSE();
2024  // Memory runtime check generation creates compares that use expanded
2025  // values. Remove them before running the SCEVExpanderCleaners.
2026  for (auto &I : make_early_inc_range(reverse(*MemCheckBlock))) {
2027  if (MemCheckExp.isInsertedInstruction(&I))
2028  continue;
2029  SE.forgetValue(&I);
2030  SE.eraseValueFromMap(&I);
2031  I.eraseFromParent();
2032  }
2033  }
2034  MemCheckCleaner.cleanup();
2035  SCEVCleaner.cleanup();
2036 
2037  if (SCEVCheckCond)
2038  SCEVCheckBlock->eraseFromParent();
2039  if (MemRuntimeCheckCond)
2040  MemCheckBlock->eraseFromParent();
2041  }
2042 
2043  /// Adds the generated SCEVCheckBlock before \p LoopVectorPreHeader and
2044  /// adjusts the branches to branch to the vector preheader or \p Bypass,
2045  /// depending on the generated condition.
2049  if (!SCEVCheckCond)
2050  return nullptr;
2051  if (auto *C = dyn_cast<ConstantInt>(SCEVCheckCond))
2052  if (C->isZero())
2053  return nullptr;
2054 
2055  auto *Pred = LoopVectorPreHeader->getSinglePredecessor();
2056 
2057  BranchInst::Create(LoopVectorPreHeader, SCEVCheckBlock);
2058  // Create new preheader for vector loop.
2059  if (auto *PL = LI->getLoopFor(LoopVectorPreHeader))
2060  PL->addBasicBlockToLoop(SCEVCheckBlock, *LI);
2061 
2062  SCEVCheckBlock->getTerminator()->eraseFromParent();
2063  SCEVCheckBlock->moveBefore(LoopVectorPreHeader);
2064  Pred->getTerminator()->replaceSuccessorWith(LoopVectorPreHeader,
2065  SCEVCheckBlock);
2066 
2067  DT->addNewBlock(SCEVCheckBlock, Pred);
2069 
2071  SCEVCheckBlock->getTerminator(),
2072  BranchInst::Create(Bypass, LoopVectorPreHeader, SCEVCheckCond));
2073  // Mark the check as used, to prevent it from being removed during cleanup.
2074  SCEVCheckCond = nullptr;
2075  return SCEVCheckBlock;
2076  }
2077 
2078  /// Adds the generated MemCheckBlock before \p LoopVectorPreHeader and adjusts
2079  /// the branches to branch to the vector preheader or \p Bypass, depending on
2080  /// the generated condition.
2083  // Check if we generated code that checks in runtime if arrays overlap.
2084  if (!MemRuntimeCheckCond)
2085  return nullptr;
2086 
2087  auto *Pred = LoopVectorPreHeader->getSinglePredecessor();
2089  MemCheckBlock);
2090 
2091  DT->addNewBlock(MemCheckBlock, Pred);
2093  MemCheckBlock->moveBefore(LoopVectorPreHeader);
2094 
2095  if (auto *PL = LI->getLoopFor(LoopVectorPreHeader))
2096  PL->addBasicBlockToLoop(MemCheckBlock, *LI);
2097 
2099  MemCheckBlock->getTerminator(),
2100  BranchInst::Create(Bypass, LoopVectorPreHeader, MemRuntimeCheckCond));
2101  MemCheckBlock->getTerminator()->setDebugLoc(
2102  Pred->getTerminator()->getDebugLoc());
2103 
2104  // Mark the check as used, to prevent it from being removed during cleanup.
2105  MemRuntimeCheckCond = nullptr;
2106  return MemCheckBlock;
2107  }
2108 };
2109 
2110 // Return true if \p OuterLp is an outer loop annotated with hints for explicit
2111 // vectorization. The loop needs to be annotated with #pragma omp simd
2112 // simdlen(#) or #pragma clang vectorize(enable) vectorize_width(#). If the
2113 // vector length information is not provided, vectorization is not considered
2114 // explicit. Interleave hints are not allowed either. These limitations will be
2115 // relaxed in the future.
2116 // Please, note that we are currently forced to abuse the pragma 'clang
2117 // vectorize' semantics. This pragma provides *auto-vectorization hints*
2118 // (i.e., LV must check that vectorization is legal) whereas pragma 'omp simd'
2119 // provides *explicit vectorization hints* (LV can bypass legal checks and
2120 // assume that vectorization is legal). However, both hints are implemented
2121 // using the same metadata (llvm.loop.vectorize, processed by
2122 // LoopVectorizeHints). This will be fixed in the future when the native IR
2123 // representation for pragma 'omp simd' is introduced.
2124 static bool isExplicitVecOuterLoop(Loop *OuterLp,
2126  assert(!OuterLp->isInnermost() && "This is not an outer loop");
2127  LoopVectorizeHints Hints(OuterLp, true /*DisableInterleaving*/, *ORE);
2128 
2129  // Only outer loops with an explicit vectorization hint are supported.
2130  // Unannotated outer loops are ignored.
2132  return false;
2133 
2134  Function *Fn = OuterLp->getHeader()->getParent();
2135  if (!Hints.allowVectorization(Fn, OuterLp,
2136  true /*VectorizeOnlyWhenForced*/)) {
2137  LLVM_DEBUG(dbgs() << "LV: Loop hints prevent outer loop vectorization.\n");
2138  return false;
2139  }
2140 
2141  if (Hints.getInterleave() > 1) {
2142  // TODO: Interleave support is future work.
2143  LLVM_DEBUG(dbgs() << "LV: Not vectorizing: Interleave is not supported for "
2144  "outer loops.\n");
2145  Hints.emitRemarkWithHints();
2146  return false;
2147  }
2148 
2149  return true;
2150 }
2151 
2155  // Collect inner loops and outer loops without irreducible control flow. For
2156  // now, only collect outer loops that have explicit vectorization hints. If we
2157  // are stress testing the VPlan H-CFG construction, we collect the outermost
2158  // loop of every loop nest.
2159  if (L.isInnermost() || VPlanBuildStressTest ||
2161  LoopBlocksRPO RPOT(&L);
2162  RPOT.perform(LI);
2163  if (!containsIrreducibleCFG<const BasicBlock *>(RPOT, *LI)) {
2164  V.push_back(&L);
2165  // TODO: Collect inner loops inside marked outer loops in case
2166  // vectorization fails for the outer loop. Do not invoke
2167  // 'containsIrreducibleCFG' again for inner loops when the outer loop is
2168  // already known to be reducible. We can use an inherited attribute for
2169  // that.
2170  return;
2171  }
2172  }
2173  for (Loop *InnerL : L)
2174  collectSupportedLoops(*InnerL, LI, ORE, V);
2175 }
2176 
2177 namespace {
2178 
2179 /// The LoopVectorize Pass.
2180 struct LoopVectorize : public FunctionPass {
2181  /// Pass identification, replacement for typeid
2182  static char ID;
2183 
2184  LoopVectorizePass Impl;
2185 
2186  explicit LoopVectorize(bool InterleaveOnlyWhenForced = false,
2187  bool VectorizeOnlyWhenForced = false)
2188  : FunctionPass(ID),
2189  Impl({InterleaveOnlyWhenForced, VectorizeOnlyWhenForced}) {
2191  }
2192 
2193  bool runOnFunction(Function &F) override {
2194  if (skipFunction(F))
2195  return false;
2196 
2197  auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
2198  auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
2199  auto *TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
2200  auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2201  auto *BFI = &getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI();
2202  auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
2203  auto *TLI = TLIP ? &TLIP->getTLI(F) : nullptr;
2204  auto *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
2205  auto *AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
2206  auto *LAA = &getAnalysis<LoopAccessLegacyAnalysis>();
2207  auto *DB = &getAnalysis<DemandedBitsWrapperPass>().getDemandedBits();
2208  auto *ORE = &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE();
2209  auto *PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
2210 
2211  std::function<const LoopAccessInfo &(Loop &)> GetLAA =
2212  [&](Loop &L) -> const LoopAccessInfo & { return LAA->getInfo(&L); };
2213 
2214  return Impl.runImpl(F, *SE, *LI, *TTI, *DT, *BFI, TLI, *DB, *AA, *AC,
2215  GetLAA, *ORE, PSI).MadeAnyChange;
2216  }
2217 
2218  void getAnalysisUsage(AnalysisUsage &AU) const override {
2230 
2231  // We currently do not preserve loopinfo/dominator analyses with outer loop
2232  // vectorization. Until this is addressed, mark these analyses as preserved
2233  // only for non-VPlan-native path.
2234  // TODO: Preserve Loop and Dominator analyses for VPlan-native path.
2235  if (!EnableVPlanNativePath) {
2238  }
2239 
2243  }
2244 };
2245 
2246 } // end anonymous namespace
2247 
2248 //===----------------------------------------------------------------------===//
2249 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
2250 // LoopVectorizationCostModel and LoopVectorizationPlanner.
2251 //===----------------------------------------------------------------------===//
2252 
2254  // We need to place the broadcast of invariant variables outside the loop,
2255  // but only if it's proven safe to do so. Else, broadcast will be inside
2256  // vector loop body.
2257  Instruction *Instr = dyn_cast<Instruction>(V);
2258  bool SafeToHoist = OrigLoop->isLoopInvariant(V) &&
2259  (!Instr ||
2261  // Place the code for broadcasting invariant variables in the new preheader.
2263  if (SafeToHoist)
2265 
2266  // Broadcast the scalar into all locations in the vector.
2267  Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
2268 
2269  return Shuf;
2270 }
2271 
2273  const InductionDescriptor &II, Value *Step, Value *Start,
2274  Instruction *EntryVal, VPValue *Def, VPValue *CastDef,
2275  VPTransformState &State) {
2276  assert((isa<PHINode>(EntryVal) || isa<TruncInst>(EntryVal)) &&
2277  "Expected either an induction phi-node or a truncate of it!");
2278 
2279  // Construct the initial value of the vector IV in the vector loop preheader
2280  auto CurrIP = Builder.saveIP();
2282  if (isa<TruncInst>(EntryVal)) {
2283  assert(Start->getType()->isIntegerTy() &&
2284  "Truncation requires an integer type");
2285  auto *TruncType = cast<IntegerType>(EntryVal->getType());
2286  Step = Builder.CreateTrunc(Step, TruncType);
2287  Start = Builder.CreateCast(Instruction::Trunc, Start, TruncType);
2288  }
2289  Value *SplatStart = Builder.CreateVectorSplat(VF, Start);
2290  Value *SteppedStart =
2291  getStepVector(SplatStart, 0, Step, II.getInductionOpcode());
2292 
2293  // We create vector phi nodes for both integer and floating-point induction
2294  // variables. Here, we determine the kind of arithmetic we will perform.
2295  Instruction::BinaryOps AddOp;
2296  Instruction::BinaryOps MulOp;
2297  if (Step->getType()->isIntegerTy()) {
2298  AddOp = Instruction::Add;
2299  MulOp = Instruction::Mul;
2300  } else {
2301  AddOp = II.getInductionOpcode();
2302  MulOp = Instruction::FMul;
2303  }
2304 
2305  // Multiply the vectorization factor by the step using integer or
2306  // floating-point arithmetic as appropriate.
2307  Type *StepType = Step->getType();
2308  if (Step->getType()->isFloatingPointTy())
2309  StepType = IntegerType::get(StepType->getContext(),
2310  StepType->getScalarSizeInBits());
2311  Value *RuntimeVF = getRuntimeVF(Builder, StepType, VF);
2312  if (Step->getType()->isFloatingPointTy())
2313  RuntimeVF = Builder.CreateSIToFP(RuntimeVF, Step->getType());
2314  Value *Mul = Builder.CreateBinOp(MulOp, Step, RuntimeVF);
2315 
2316  // Create a vector splat to use in the induction update.
2317  //
2318  // FIXME: If the step is non-constant, we create the vector splat with
2319  // IRBuilder. IRBuilder can constant-fold the multiply, but it doesn't
2320  // handle a constant vector splat.
2321  Value *SplatVF = isa<Constant>(Mul)
2322  ? ConstantVector::getSplat(VF, cast<Constant>(Mul))
2324  Builder.restoreIP(CurrIP);
2325 
2326  // We may need to add the step a number of times, depending on the unroll
2327  // factor. The last of those goes into the PHI.
2328  PHINode *VecInd = PHINode::Create(SteppedStart->getType(), 2, "vec.ind",
2330  VecInd->setDebugLoc(EntryVal->getDebugLoc());
2331  Instruction *LastInduction = VecInd;
2332  for (unsigned Part = 0; Part < UF; ++Part) {
2333  State.set(Def, LastInduction, Part);
2334 
2335  if (isa<TruncInst>(EntryVal))
2336  addMetadata(LastInduction, EntryVal);
2337  recordVectorLoopValueForInductionCast(II, EntryVal, LastInduction, CastDef,
2338  State, Part);
2339 
2340  LastInduction = cast<Instruction>(
2341  Builder.CreateBinOp(AddOp, LastInduction, SplatVF, "step.add"));
2342  LastInduction->setDebugLoc(EntryVal->getDebugLoc());
2343  }
2344 
2345  // Move the last step to the end of the latch block. This ensures consistent
2346  // placement of all induction updates.
2347  auto *LoopVectorLatch = LI->getLoopFor(LoopVectorBody)->getLoopLatch();
2348  auto *Br = cast<BranchInst>(LoopVectorLatch->getTerminator());
2349  auto *ICmp = cast<Instruction>(Br->getCondition());
2350  LastInduction->moveBefore(ICmp);
2351  LastInduction->setName("vec.ind.next");
2352 
2353  VecInd->addIncoming(SteppedStart, LoopVectorPreHeader);
2354  VecInd->addIncoming(LastInduction, LoopVectorLatch);
2355 }
2356 
2358  return Cost->isScalarAfterVectorization(I, VF) ||
2360 }
2361 
2364  return true;
2365  auto isScalarInst = [&](User *U) -> bool {
2366  auto *I = cast<Instruction>(U);
2368  };
2369  return llvm::any_of(IV->users(), isScalarInst);
2370 }
2371 
2373  const InductionDescriptor &ID, const Instruction *EntryVal,
2374  Value *VectorLoopVal, VPValue *CastDef, VPTransformState &State,
2375  unsigned Part, unsigned Lane) {
2376  assert((isa<PHINode>(EntryVal) || isa<TruncInst>(EntryVal)) &&
2377  "Expected either an induction phi-node or a truncate of it!");
2378 
2379  // This induction variable is not the phi from the original loop but the
2380  // newly-created IV based on the proof that casted Phi is equal to the
2381  // uncasted Phi in the vectorized loop (under a runtime guard possibly). It
2382  // re-uses the same InductionDescriptor that original IV uses but we don't
2383  // have to do any recording in this case - that is done when original IV is
2384  // processed.
2385  if (isa<TruncInst>(EntryVal))
2386  return;
2387 
2388  const SmallVectorImpl<Instruction *> &Casts = ID.getCastInsts();
2389  if (Casts.empty())
2390  return;
2391  // Only the first Cast instruction in the Casts vector is of interest.
2392  // The rest of the Casts (if exist) have no uses outside the
2393  // induction update chain itself.
2394  if (Lane < UINT_MAX)
2395  State.set(CastDef, VectorLoopVal, VPIteration(Part, Lane));
2396  else
2397  State.set(CastDef, VectorLoopVal, Part);
2398 }
2399 
2401  TruncInst *Trunc, VPValue *Def,
2402  VPValue *CastDef,
2403  VPTransformState &State) {
2404  assert((IV->getType()->isIntegerTy() || IV != OldInduction) &&
2405  "Primary induction variable must have an integer type");
2406 
2407  auto II = Legal->getInductionVars().find(IV);
2408  assert(II != Legal->getInductionVars().end() && "IV is not an induction");
2409 
2410  auto ID = II->second;
2411  assert(IV->getType() == ID.getStartValue()->getType() && "Types must match");
2412 
2413  // The value from the original loop to which we are mapping the new induction
2414  // variable.
2415  Instruction *EntryVal = Trunc ? cast<Instruction>(Trunc) : IV;
2416 
2417  auto &DL = OrigLoop->getHeader()->getModule()->getDataLayout();
2418 
2419  // Generate code for the induction step. Note that induction steps are
2420  // required to be loop-invariant
2421  auto CreateStepValue = [&](const SCEV *Step) -> Value * {
2422  assert(PSE.getSE()->isLoopInvariant(Step, OrigLoop) &&
2423  "Induction step should be loop invariant");
2424  if (PSE.getSE()->isSCEVable(IV->getType())) {
2425  SCEVExpander Exp(*PSE.getSE(), DL, "induction");
2426  return Exp.expandCodeFor(Step, Step->getType(),
2428  }
2429  return cast<SCEVUnknown>(Step)->getValue();
2430  };
2431 
2432  // The scalar value to broadcast. This is derived from the canonical
2433  // induction variable. If a truncation type is given, truncate the canonical
2434  // induction variable and step. Otherwise, derive these values from the
2435  // induction descriptor.
2436  auto CreateScalarIV = [&](Value *&Step) -> Value * {
2437  Value *ScalarIV = Induction;
2438  if (IV != OldInduction) {
2439  ScalarIV = IV->getType()->isIntegerTy()
2441  : Builder.CreateCast(Instruction::SIToFP, Induction,
2442  IV->getType());
2443  ScalarIV = emitTransformedIndex(Builder, ScalarIV, PSE.getSE(), DL, ID);
2444  ScalarIV->setName("offset.idx");
2445  }
2446  if (Trunc) {
2447  auto *TruncType = cast<IntegerType>(Trunc->getType());
2448  assert(Step->getType()->isIntegerTy() &&
2449  "Truncation requires an integer step");
2450  ScalarIV = Builder.CreateTrunc(ScalarIV, TruncType);
2451  Step = Builder.CreateTrunc(Step, TruncType);
2452  }
2453  return ScalarIV;
2454  };
2455 
2456  // Create the vector values from the scalar IV, in the absence of creating a
2457  // vector IV.
2458  auto CreateSplatIV = [&](Value *ScalarIV, Value *Step) {
2459  Value *Broadcasted = getBroadcastInstrs(ScalarIV);
2460  for (unsigned Part = 0; Part < UF; ++Part) {
2461  assert(!VF.isScalable() && "scalable vectors not yet supported.");
2462  Value *EntryPart =
2463  getStepVector(Broadcasted, VF.getKnownMinValue() * Part, Step,
2464  ID.getInductionOpcode());
2465  State.set(Def, EntryPart, Part);
2466  if (Trunc)
2467  addMetadata(EntryPart, Trunc);
2468  recordVectorLoopValueForInductionCast(ID, EntryVal, EntryPart, CastDef,
2469  State, Part);
2470  }
2471  };
2472 
2473  // Fast-math-flags propagate from the original induction instruction.
2475  if (ID.getInductionBinOp() && isa<FPMathOperator>(ID.getInductionBinOp()))
2476  Builder.setFastMathFlags(ID.getInductionBinOp()->getFastMathFlags());
2477 
2478  // Now do the actual transformations, and start with creating the step value.
2479  Value *Step = CreateStepValue(ID.getStep());
2480  if (VF.isZero() || VF.isScalar()) {
2481  Value *ScalarIV = CreateScalarIV(Step);
2482  CreateSplatIV(ScalarIV, Step);
2483  return;
2484  }
2485 
2486  // Determine if we want a scalar version of the induction variable. This is
2487  // true if the induction variable itself is not widened, or if it has at
2488  // least one user in the loop that is not widened.
2489  auto NeedsScalarIV = needsScalarInduction(EntryVal);
2490  if (!NeedsScalarIV) {
2491  createVectorIntOrFpInductionPHI(ID, Step, Start, EntryVal, Def, CastDef,
2492  State);
2493  return;
2494  }
2495 
2496  // Try to create a new independent vector induction variable. If we can't
2497  // create the phi node, we will splat the scalar induction variable in each
2498  // loop iteration.
2499  if (!shouldScalarizeInstruction(EntryVal)) {
2500  createVectorIntOrFpInductionPHI(ID, Step, Start, EntryVal, Def, CastDef,
2501  State);
2502  Value *ScalarIV = CreateScalarIV(Step);
2503  // Create scalar steps that can be used by instructions we will later
2504  // scalarize. Note that the addition of the scalar steps will not increase
2505  // the number of instructions in the loop in the common case prior to
2506  // InstCombine. We will be trading one vector extract for each scalar step.
2507  buildScalarSteps(ScalarIV, Step, EntryVal, ID, Def, CastDef, State);
2508  return;
2509  }
2510 
2511  // All IV users are scalar instructions, so only emit a scalar IV, not a
2512  // vectorised IV. Except when we tail-fold, then the splat IV feeds the
2513  // predicate used by the masked loads/stores.
2514  Value *ScalarIV = CreateScalarIV(Step);
2515  if (!Cost->isScalarEpilogueAllowed())
2516  CreateSplatIV(ScalarIV, Step);
2517  buildScalarSteps(ScalarIV, Step, EntryVal, ID, Def, CastDef, State);
2518 }
2519 
2521  Instruction::BinaryOps BinOp) {
2522  // Create and check the types.
2523  auto *ValVTy = cast<VectorType>(Val->getType());
2524  ElementCount VLen = ValVTy->getElementCount();
2525 
2526  Type *STy = Val->getType()->getScalarType();
2527  assert((STy->isIntegerTy() || STy->isFloatingPointTy()) &&
2528  "Induction Step must be an integer or FP");
2529  assert(Step->getType() == STy && "Step has wrong type");
2530 
2532 
2533  // Create a vector of consecutive numbers from zero to VF.
2534  VectorType *InitVecValVTy = ValVTy;
2535  Type *InitVecValSTy = STy;
2536  if (STy->isFloatingPointTy()) {
2537  InitVecValSTy =
2539  InitVecValVTy = VectorType::get(InitVecValSTy, VLen);
2540  }
2541  Value *InitVec = Builder.CreateStepVector(InitVecValVTy);
2542 
2543  // Add on StartIdx
2544  Value *StartIdxSplat = Builder.CreateVectorSplat(
2545  VLen, ConstantInt::get(InitVecValSTy, StartIdx));
2546  InitVec = Builder.CreateAdd(InitVec, StartIdxSplat);
2547 
2548  if (STy->isIntegerTy()) {
2549  Step = Builder.CreateVectorSplat(VLen, Step);
2550  assert(Step->getType() == Val->getType() && "Invalid step vec");
2551  // FIXME: The newly created binary instructions should contain nsw/nuw flags,
2552  // which can be found from the original scalar operations.
2553  Step = Builder.CreateMul(InitVec, Step);
2554  return Builder.CreateAdd(Val, Step, "induction");
2555  }
2556 
2557  // Floating point induction.
2558  assert((BinOp == Instruction::FAdd || BinOp == Instruction::FSub) &&
2559  "Binary Opcode should be specified for FP induction");
2560  InitVec = Builder.CreateUIToFP(InitVec, ValVTy);
2561  Step = Builder.CreateVectorSplat(VLen, Step);
2562  Value *MulOp = Builder.CreateFMul(InitVec, Step);
2563  return Builder.CreateBinOp(BinOp, Val, MulOp, "induction");
2564 }
2565 
2567  Instruction *EntryVal,
2568  const InductionDescriptor &ID,
2569  VPValue *Def, VPValue *CastDef,
2570  VPTransformState &State) {
2571  // We shouldn't have to build scalar steps if we aren't vectorizing.
2572  assert(VF.isVector() && "VF should be greater than one");
2573  // Get the value type and ensure it and the step have the same integer type.
2574  Type *ScalarIVTy = ScalarIV->getType()->getScalarType();
2575  assert(ScalarIVTy == Step->getType() &&
2576  "Val and Step should have the same type");
2577 
2578  // We build scalar steps for both integer and floating-point induction
2579  // variables. Here, we determine the kind of arithmetic we will perform.
2580  Instruction::BinaryOps AddOp;
2581  Instruction::BinaryOps MulOp;
2582  if (ScalarIVTy->isIntegerTy()) {
2583  AddOp = Instruction::Add;
2584  MulOp = Instruction::Mul;
2585  } else {
2586  AddOp = ID.getInductionOpcode();
2587  MulOp = Instruction::FMul;
2588  }
2589 
2590  // Determine the number of scalars we need to generate for each unroll
2591  // iteration. If EntryVal is uniform, we only need to generate the first
2592  // lane. Otherwise, we generate all VF values.
2593  bool IsUniform =
2594  Cost->isUniformAfterVectorization(cast<Instruction>(EntryVal), VF);
2595  unsigned Lanes = IsUniform ? 1 : VF.getKnownMinValue();
2596  // Compute the scalar steps and save the results in State.
2597  Type *IntStepTy = IntegerType::get(ScalarIVTy->getContext(),
2598  ScalarIVTy->getScalarSizeInBits());
2599  Type *VecIVTy = nullptr;
2600  Value *UnitStepVec = nullptr, *SplatStep = nullptr, *SplatIV = nullptr;
2601  if (!IsUniform && VF.isScalable()) {
2602  VecIVTy = VectorType::get(ScalarIVTy, VF);
2603  UnitStepVec = Builder.CreateStepVector(VectorType::get(IntStepTy, VF));
2604  SplatStep = Builder.CreateVectorSplat(VF, Step);
2605  SplatIV = Builder.CreateVectorSplat(VF, ScalarIV);
2606  }
2607 
2608  for (unsigned Part = 0; Part < UF; ++Part) {
2609  Value *StartIdx0 =
2610  createStepForVF(Builder, ConstantInt::get(IntStepTy, Part), VF);
2611 
2612  if (!IsUniform && VF.isScalable()) {
2613  auto *SplatStartIdx = Builder.CreateVectorSplat(VF, StartIdx0);
2614  auto *InitVec = Builder.CreateAdd(SplatStartIdx, UnitStepVec);
2615  if (ScalarIVTy->isFloatingPointTy())
2616  InitVec = Builder.CreateSIToFP(InitVec, VecIVTy);
2617  auto *Mul = Builder.CreateBinOp(MulOp, InitVec, SplatStep);
2618  auto *Add = Builder.CreateBinOp(AddOp, SplatIV, Mul);
2619  State.set(Def, Add, Part);
2620  recordVectorLoopValueForInductionCast(ID, EntryVal, Add, CastDef, State,
2621  Part);
2622  // It's useful to record the lane values too for the known minimum number
2623  // of elements so we do those below. This improves the code quality when
2624  // trying to extract the first element, for example.
2625  }
2626 
2627  if (ScalarIVTy->isFloatingPointTy())
2628  StartIdx0 = Builder.CreateSIToFP(StartIdx0, ScalarIVTy);
2629 
2630  for (unsigned Lane = 0; Lane < Lanes; ++Lane) {
2631  Value *StartIdx = Builder.CreateBinOp(
2632  AddOp, StartIdx0, getSignedIntOrFpConstant(ScalarIVTy, Lane));
2633  // The step returned by `createStepForVF` is a runtime-evaluated value
2634  // when VF is scalable. Otherwise, it should be folded into a Constant.
2635  assert((VF.isScalable() || isa<Constant>(StartIdx)) &&
2636  "Expected StartIdx to be folded to a constant when VF is not "
2637  "scalable");
2638  auto *Mul = Builder.CreateBinOp(MulOp, StartIdx, Step);
2639  auto *Add = Builder.CreateBinOp(AddOp, ScalarIV, Mul);
2640  State.set(Def, Add, VPIteration(Part, Lane));
2641  recordVectorLoopValueForInductionCast(ID, EntryVal, Add, CastDef, State,
2642  Part, Lane);
2643  }
2644  }
2645 }
2646 
2648  const VPIteration &Instance,
2649  VPTransformState &State) {
2650  Value *ScalarInst = State.get(Def, Instance);
2651  Value *VectorValue = State.get(Def, Instance.Part);
2652  VectorValue = Builder.CreateInsertElement(
2653  VectorValue, ScalarInst,
2654  Instance.Lane.getAsRuntimeExpr(State.Builder, VF));
2655  State.set(Def, VectorValue, Instance.Part);
2656 }
2657 
2659  assert(Vec->getType()->isVectorTy() && "Invalid type");
2660  return Builder.CreateVectorReverse(Vec, "reverse");
2661 }
2662 
2663 // Return whether we allow using masked interleave-groups (for dealing with
2664 // strided loads/stores that reside in predicated blocks, or for dealing
2665 // with gaps).
2667  // If an override option has been passed in for interleaved accesses, use it.
2670 
2672 }
2673 
2674 // Try to vectorize the interleave group that \p Instr belongs to.
2675 //
2676 // E.g. Translate following interleaved load group (factor = 3):
2677 // for (i = 0; i < N; i+=3) {
2678 // R = Pic[i]; // Member of index 0
2679 // G = Pic[i+1]; // Member of index 1
2680 // B = Pic[i+2]; // Member of index 2
2681 // ... // do something to R, G, B
2682 // }
2683 // To:
2684 // %wide.vec = load <12 x i32> ; Read 4 tuples of R,G,B
2685 // %R.vec = shuffle %wide.vec, poison, <0, 3, 6, 9> ; R elements
2686 // %G.vec = shuffle %wide.vec, poison, <1, 4, 7, 10> ; G elements
2687 // %B.vec = shuffle %wide.vec, poison, <2, 5, 8, 11> ; B elements
2688 //
2689 // Or translate following interleaved store group (factor = 3):
2690 // for (i = 0; i < N; i+=3) {
2691 // ... do something to R, G, B
2692 // Pic[i] = R; // Member of index 0
2693 // Pic[i+1] = G; // Member of index 1
2694 // Pic[i+2] = B; // Member of index 2
2695 // }
2696 // To:
2697 // %R_G.vec = shuffle %R.vec, %G.vec, <0, 1, 2, ..., 7>
2698 // %B_U.vec = shuffle %B.vec, poison, <0, 1, 2, 3, u, u, u, u>
2699 // %interleaved.vec = shuffle %R_G.vec, %B_U.vec,
2700 // <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11> ; Interleave R,G,B elements
2701 // store <12 x i32> %interleaved.vec ; Write 4 tuples of R,G,B
2704  VPTransformState &State, VPValue *Addr, ArrayRef<VPValue *> StoredValues,
2705  VPValue *BlockInMask) {
2706  Instruction *Instr = Group->getInsertPos();
2707  const DataLayout &DL = Instr->getModule()->getDataLayout();
2708 
2709  // Prepare for the vector type of the interleaved load/store.
2710  Type *ScalarTy = getLoadStoreType(Instr);
2711  unsigned InterleaveFactor = Group->getFactor();
2712  assert(!VF.isScalable() && "scalable vectors not yet supported.");
2713  auto *VecTy = VectorType::get(ScalarTy, VF * InterleaveFactor);
2714 
2715  // Prepare for the new pointers.
2716  SmallVector<Value *, 2> AddrParts;
2717  unsigned Index = Group->getIndex(Instr);
2718 
2719  // TODO: extend the masked interleaved-group support to reversed access.
2720  assert((!BlockInMask || !Group->isReverse()) &&
2721  "Reversed masked interleave-group not supported.");
2722 
2723  // If the group is reverse, adjust the index to refer to the last vector lane
2724  // instead of the first. We adjust the index from the first vector lane,
2725  // rather than directly getting the pointer for lane VF - 1, because the
2726  // pointer operand of the interleaved access is supposed to be uniform. For
2727  // uniform instructions, we're only required to generate a value for the
2728  // first vector lane in each unroll iteration.
2729  if (Group->isReverse())
2730  Index += (VF.getKnownMinValue() - 1) * Group->getFactor();
2731 
2732  for (unsigned Part = 0; Part < UF; Part++) {
2733  Value *AddrPart = State.get(Addr, VPIteration(Part, 0));
2734  setDebugLocFromInst(AddrPart);
2735 
2736  // Notice current instruction could be any index. Need to adjust the address
2737  // to the member of index 0.
2738  //
2739  // E.g. a = A[i+1]; // Member of index 1 (Current instruction)
2740  // b = A[i]; // Member of index 0
2741  // Current pointer is pointed to A[i+1], adjust it to A[i].
2742  //
2743  // E.g. A[i+1] = a; // Member of index 1
2744  // A[i] = b; // Member of index 0
2745  // A[i+2] = c; // Member of index 2 (Current instruction)
2746  // Current pointer is pointed to A[i+2], adjust it to A[i].
2747 
2748  bool InBounds = false;
2749  if (auto *gep = dyn_cast<GetElementPtrInst>(AddrPart->stripPointerCasts()))
2750  InBounds = gep->isInBounds();
2751  AddrPart = Builder.CreateGEP(ScalarTy, AddrPart, Builder.getInt32(-Index));
2752  cast<GetElementPtrInst>(AddrPart)->setIsInBounds(InBounds);
2753 
2754  // Cast to the vector pointer type.
2755  unsigned AddressSpace = AddrPart->getType()->getPointerAddressSpace();
2756  Type *PtrTy = VecTy->getPointerTo(AddressSpace);
2757  AddrParts.push_back(Builder.CreateBitCast(AddrPart, PtrTy));
2758  }
2759 
2760  setDebugLocFromInst(Instr);
2761  Value *PoisonVec = PoisonValue::get(VecTy);
2762 
2763  Value *MaskForGaps = nullptr;
2764  if (Group->requiresScalarEpilogue() && !Cost->isScalarEpilogueAllowed()) {
2765  MaskForGaps = createBitMaskForGaps(Builder, VF.getKnownMinValue(), *Group);
2766  assert(MaskForGaps && "Mask for Gaps is required but it is null");
2767  }
2768 
2769  // Vectorize the interleaved load group.
2770  if (isa<LoadInst>(Instr)) {
2771  // For each unroll part, create a wide load for the group.
2772  SmallVector<Value *, 2> NewLoads;
2773  for (unsigned Part = 0; Part < UF; Part++) {
2774  Instruction *NewLoad;
2775  if (BlockInMask || MaskForGaps) {
2777  "masked interleaved groups are not allowed.");
2778  Value *GroupMask = MaskForGaps;
2779  if (BlockInMask) {
2780  Value *BlockInMaskPart = State.get(BlockInMask, Part);
2781  Value *ShuffledMask = Builder.CreateShuffleVector(
2782  BlockInMaskPart,
2783  createReplicatedMask(InterleaveFactor, VF.getKnownMinValue()),
2784  "interleaved.mask");
2785  GroupMask = MaskForGaps
2786  ? Builder.CreateBinOp(Instruction::And, ShuffledMask,
2787  MaskForGaps)
2788  : ShuffledMask;
2789  }
2790  NewLoad =
2791  Builder.CreateMaskedLoad(VecTy, AddrParts[Part], Group->getAlign(),
2792  GroupMask, PoisonVec, "wide.masked.vec");
2793  }
2794  else
2795  NewLoad = Builder.CreateAlignedLoad(VecTy, AddrParts[Part],
2796  Group->getAlign(), "wide.vec");
2797  Group->addMetadata(NewLoad);
2798  NewLoads.push_back(NewLoad);
2799  }
2800 
2801  // For each member in the group, shuffle out the appropriate data from the
2802  // wide loads.
2803  unsigned J = 0;
2804  for (unsigned I = 0; I < InterleaveFactor; ++I) {
2805  Instruction *Member = Group->getMember(I);
2806 
2807  // Skip the gaps in the group.
2808  if (!Member)
2809  continue;
2810 
2811  auto StrideMask =
2812  createStrideMask(I, InterleaveFactor, VF.getKnownMinValue());
2813  for (unsigned Part = 0; Part < UF; Part++) {
2814  Value *StridedVec = Builder.CreateShuffleVector(
2815  NewLoads[Part], StrideMask, "strided.vec");
2816 
2817  // If this member has different type, cast the result type.
2818  if (Member->getType() != ScalarTy) {
2819  assert(!VF.isScalable() && "VF is assumed to be non scalable.");
2820  VectorType *OtherVTy = VectorType::get(Member->getType(), VF);
2821  StridedVec = createBitOrPointerCast(StridedVec, OtherVTy, DL);
2822  }
2823 
2824  if (Group->isReverse())
2825  StridedVec = reverseVector(StridedVec);
2826 
2827  State.set(VPDefs[J], StridedVec, Part);
2828  }
2829  ++J;
2830  }
2831  return;
2832  }
2833 
2834  // The sub vector type for current instruction.
2835  auto *SubVT = VectorType::get(ScalarTy, VF);
2836 
2837  // Vectorize the interleaved store group.
2838  MaskForGaps = createBitMaskForGaps(Builder, VF.getKnownMinValue(), *Group);
2839  assert((!MaskForGaps || useMaskedInterleavedAccesses(*TTI)) &&
2840  "masked interleaved groups are not allowed.");
2841  assert((!MaskForGaps || !VF.isScalable()) &&
2842  "masking gaps for scalable vectors is not yet supported.");
2843  for (unsigned Part = 0; Part < UF; Part++) {
2844  // Collect the stored vector from each member.
2845  SmallVector<Value *, 4> StoredVecs;
2846  for (unsigned i = 0; i < InterleaveFactor; i++) {
2847  assert((Group->getMember(i) || MaskForGaps) &&
2848  "Fail to get a member from an interleaved store group");
2849  Instruction *Member = Group->getMember(i);
2850 
2851  // Skip the gaps in the group.
2852  if (!Member) {
2853  Value *Undef = PoisonValue::get(SubVT);
2854  StoredVecs.push_back(Undef);
2855  continue;
2856  }
2857 
2858  Value *StoredVec = State.get(StoredValues[i], Part);
2859 
2860  if (Group->isReverse())
2861  StoredVec = reverseVector(StoredVec);
2862 
2863  // If this member has different type, cast it to a unified type.
2864 
2865  if (StoredVec->getType() != SubVT)
2866  StoredVec = createBitOrPointerCast(StoredVec, SubVT, DL);
2867 
2868  StoredVecs.push_back(StoredVec);
2869  }
2870 
2871  // Concatenate all vectors into a wide vector.
2872  Value *WideVec = concatenateVectors(Builder, StoredVecs);
2873 
2874  // Interleave the elements in the wide vector.
2876  WideVec, createInterleaveMask(VF.getKnownMinValue(), InterleaveFactor),
2877  "interleaved.vec");
2878 
2879  Instruction *NewStoreInstr;
2880  if (BlockInMask || MaskForGaps) {
2881  Value *GroupMask = MaskForGaps;
2882  if (BlockInMask) {
2883  Value *BlockInMaskPart = State.get(BlockInMask, Part);
2884  Value *ShuffledMask = Builder.CreateShuffleVector(
2885  BlockInMaskPart,
2886  createReplicatedMask(InterleaveFactor, VF.getKnownMinValue()),
2887  "interleaved.mask");
2888  GroupMask = MaskForGaps ? Builder.CreateBinOp(Instruction::And,
2889  ShuffledMask, MaskForGaps)
2890  : ShuffledMask;
2891  }
2892  NewStoreInstr = Builder.CreateMaskedStore(IVec, AddrParts[Part],
2893  Group->getAlign(), GroupMask);
2894  } else
2895  NewStoreInstr =
2896  Builder.CreateAlignedStore(IVec, AddrParts[Part], Group->getAlign());
2897 
2898  Group->addMetadata(NewStoreInstr);
2899  }
2900 }
2901 
2903  Instruction *Instr, VPTransformState &State, VPValue *Def, VPValue *Addr,
2904  VPValue *StoredValue, VPValue *BlockInMask, bool ConsecutiveStride,
2905  bool Reverse) {
2906  // Attempt to issue a wide load.
2907  LoadInst *LI = dyn_cast<LoadInst>(Instr);
2908  StoreInst *SI = dyn_cast<StoreInst>(Instr);
2909 
2910  assert((LI || SI) && "Invalid Load/Store instruction");
2911  assert((!SI || StoredValue) && "No stored value provided for widened store");
2912  assert((!LI || !StoredValue) && "Stored value provided for widened load");
2913 
2914  Type *ScalarDataTy = getLoadStoreType(Instr);
2915 
2916  auto *DataTy = VectorType::get(ScalarDataTy, VF);
2917  const Align Alignment = getLoadStoreAlignment(Instr);
2918  bool CreateGatherScatter = !ConsecutiveStride;
2919 
2920  VectorParts BlockInMaskParts(UF);
2921  bool isMaskRequired = BlockInMask;
2922  if (isMaskRequired)
2923  for (unsigned Part = 0; Part < UF; ++Part)
2924  BlockInMaskParts[Part] = State.get(BlockInMask, Part);
2925 
2926  const auto CreateVecPtr = [&](unsigned Part, Value *Ptr) -> Value * {
2927  // Calculate the pointer for the specific unroll-part.
2928  GetElementPtrInst *PartPtr = nullptr;
2929 
2930  bool InBounds = false;
2931  if (auto *gep = dyn_cast<GetElementPtrInst>(Ptr->stripPointerCasts()))
2932  InBounds = gep->isInBounds();
2933  if (Reverse) {
2934  // If the address is consecutive but reversed, then the
2935  // wide store needs to start at the last vector element.
2936  // RunTimeVF = VScale * VF.getKnownMinValue()
2937  // For fixed-width VScale is 1, then RunTimeVF = VF.getKnownMinValue()
2938  Value *RunTimeVF = getRuntimeVF(Builder, Builder.getInt32Ty(), VF);
2939  // NumElt = -Part * RunTimeVF
2940  Value *NumElt = Builder.CreateMul(Builder.getInt32(-Part), RunTimeVF);
2941  // LastLane = 1 - RunTimeVF
2942  Value *LastLane = Builder.CreateSub(Builder.getInt32(1), RunTimeVF);
2943  PartPtr =
2944  cast<GetElementPtrInst>(Builder.CreateGEP(ScalarDataTy, Ptr, NumElt));
2945  PartPtr->setIsInBounds(InBounds);
2946  PartPtr = cast<GetElementPtrInst>(
2947  Builder.CreateGEP(ScalarDataTy, PartPtr, LastLane));
2948  PartPtr->setIsInBounds(InBounds);
2949  if (isMaskRequired) // Reverse of a null all-one mask is a null mask.
2950  BlockInMaskParts[Part] = reverseVector(BlockInMaskParts[Part]);
2951  } else {
2952  Value *Increment = createStepForVF(Builder, Builder.getInt32(Part), VF);
2953  PartPtr = cast<GetElementPtrInst>(
2954  Builder.CreateGEP(ScalarDataTy, Ptr, Increment));
2955  PartPtr->setIsInBounds(InBounds);
2956  }
2957 
2958  unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
2959  return Builder.CreateBitCast(PartPtr, DataTy->getPointerTo(AddressSpace));
2960  };
2961 
2962  // Handle Stores:
2963  if (SI) {
2965 
2966  for (unsigned Part = 0; Part < UF; ++Part) {
2967  Instruction *NewSI = nullptr;
2968  Value *StoredVal = State.get(StoredValue, Part);
2969  if (CreateGatherScatter) {
2970  Value *MaskPart = isMaskRequired ? BlockInMaskParts[Part] : nullptr;
2971  Value *VectorGep = State.get(Addr, Part);
2972  NewSI = Builder.CreateMaskedScatter(StoredVal, VectorGep, Alignment,
2973  MaskPart);
2974  } else {
2975  if (Reverse) {
2976  // If we store to reverse consecutive memory locations, then we need
2977  // to reverse the order of elements in the stored value.
2978  StoredVal = reverseVector(StoredVal);
2979  // We don't want to update the value in the map as it might be used in
2980  // another expression. So don't call resetVectorValue(StoredVal).
2981  }
2982  auto *VecPtr = CreateVecPtr(Part, State.get(Addr, VPIteration(0, 0)));
2983  if (isMaskRequired)
2984  NewSI = Builder.CreateMaskedStore(StoredVal, VecPtr, Alignment,
2985  BlockInMaskParts[Part]);
2986  else
2987  NewSI = Builder.CreateAlignedStore(StoredVal, VecPtr, Alignment);
2988  }
2989  addMetadata(NewSI, SI);
2990  }
2991  return;
2992  }
2993 
2994  // Handle loads.
2995  assert(LI && "Must have a load instruction");
2997  for (unsigned Part = 0; Part < UF; ++Part) {
2998  Value *NewLI;
2999  if (CreateGatherScatter) {
3000  Value *MaskPart = isMaskRequired ? BlockInMaskParts[Part] : nullptr;
3001  Value *VectorGep = State.get(Addr, Part);
3002  NewLI = Builder.CreateMaskedGather(DataTy, VectorGep, Alignment, MaskPart,
3003  nullptr, "wide.masked.gather");
3004  addMetadata(NewLI, LI);
3005  } else {
3006  auto *VecPtr = CreateVecPtr(Part, State.get(Addr, VPIteration(0, 0)));
3007  if (isMaskRequired)
3008  NewLI = Builder.CreateMaskedLoad(
3009  DataTy, VecPtr, Alignment, BlockInMaskParts[Part],
3010  PoisonValue::get(DataTy), "wide.masked.load");
3011  else
3012  NewLI =
3013  Builder.CreateAlignedLoad(DataTy, VecPtr, Alignment, "wide.load");
3014 
3015  // Add metadata to the load, but setVectorValue to the reverse shuffle.
3016  addMetadata(NewLI, LI);
3017  if (Reverse)
3018  NewLI = reverseVector(NewLI);
3019  }
3020 
3021  State.set(Def, NewLI, Part);
3022  }
3023 }
3024 
3026  VPUser &User,
3027  const VPIteration &Instance,
3028  bool IfPredicateInstr,
3029  VPTransformState &State) {
3030  assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
3031 
3032  // llvm.experimental.noalias.scope.decl intrinsics must only be duplicated for
3033  // the first lane and part.
3034  if (isa<NoAliasScopeDeclInst>(Instr))
3035  if (!Instance.isFirstIteration())
3036  return;
3037 
3038  setDebugLocFromInst(Instr);
3039 
3040  // Does this instruction return a value ?
3041  bool IsVoidRetTy = Instr->getType()->isVoidTy();
3042 
3043  Instruction *Cloned = Instr->clone();
3044  if (!IsVoidRetTy)
3045  Cloned->setName(Instr->getName() + ".cloned");
3046 
3049  // Replace the operands of the cloned instructions with their scalar
3050  // equivalents in the new loop.
3051  for (unsigned op = 0, e = User.getNumOperands(); op != e; ++op) {
3052  auto *Operand = dyn_cast<Instruction>(Instr->getOperand(op));
3053  auto InputInstance = Instance;
3054  if (!Operand || !OrigLoop->contains(Operand) ||
3055  (Cost->isUniformAfterVectorization(Operand, State.VF)))
3056  InputInstance.Lane = VPLane::getFirstLane();
3057  auto *NewOp = State.get(User.getOperand(op), InputInstance);
3058  Cloned->setOperand(op, NewOp);
3059  }
3060  addNewMetadata(Cloned, Instr);
3061 
3062  // Place the cloned scalar in the new loop.
3063  Builder.Insert(Cloned);
3064 
3065  State.set(Def, Cloned, Instance);
3066 
3067  // If we just cloned a new assumption, add it the assumption cache.
3068  if (auto *II = dyn_cast<AssumeInst>(Cloned))
3069  AC->registerAssumption(II);
3070 
3071  // End if-block.
3072  if (IfPredicateInstr)
3073  PredicatedInstructions.push_back(Cloned);
3074 }
3075 
3077  Value *End, Value *Step,
3078  Instruction *DL) {
3079  BasicBlock *Header = L->getHeader();
3080  BasicBlock *Latch = L->getLoopLatch();
3081  // As we're just creating this loop, it's possible no latch exists
3082  // yet. If so, use the header as this will be a single block loop.
3083  if (!Latch)
3084  Latch = Header;
3085 
3086  IRBuilder<> B(&*Header->getFirstInsertionPt());
3088  setDebugLocFromInst(OldInst, &B);
3089  auto *Induction = B.CreatePHI(Start->getType(), 2, "index");
3090 
3091  B.SetInsertPoint(Latch->getTerminator());
3092  setDebugLocFromInst(OldInst, &B);
3093 
3094  // Create i+1 and fill the PHINode.
3095  //
3096  // If the tail is not folded, we know that End - Start >= Step (either
3097  // statically or through the minimum iteration checks). We also know that both
3098  // Start % Step == 0 and End % Step == 0. We exit the vector loop if %IV +
3099  // %Step == %End. Hence we must exit the loop before %IV + %Step unsigned
3100  // overflows and we can mark the induction increment as NUW.
3101  Value *Next = B.CreateAdd(Induction, Step, "index.next",
3102  /*NUW=*/!Cost->foldTailByMasking(), /*NSW=*/false);
3103  Induction->addIncoming(Start, L->getLoopPreheader());
3104  Induction->addIncoming(Next, Latch);
3105  // Create the compare.
3106  Value *ICmp = B.CreateICmpEQ(Next, End);
3107  B.CreateCondBr(ICmp, L->getUniqueExitBlock(), Header);
3108 
3109  // Now we have two terminators. Remove the old one from the block.
3110  Latch->getTerminator()->eraseFromParent();
3111 
3112  return Induction;
3113 }
3114 
3116  if (TripCount)
3117  return TripCount;
3118 
3119  assert(L && "Create Trip Count for null loop.");
3121  // Find the loop boundaries.
3122  ScalarEvolution *SE = PSE.getSE();
3123  const SCEV *BackedgeTakenCount = PSE.getBackedgeTakenCount();
3124  assert(!isa<SCEVCouldNotCompute>(BackedgeTakenCount) &&
3125  "Invalid loop count");
3126 
3127  Type *IdxTy = Legal->getWidestInductionType();
3128  assert(IdxTy && "No type for induction");
3129 
3130  // The exit count might have the type of i64 while the phi is i32. This can
3131  // happen if we have an induction variable that is sign extended before the
3132  // compare. The only way that we get a backedge taken count is that the
3133  // induction variable was signed and as such will not overflow. In such a case
3134  // truncation is legal.
3135  if (SE->getTypeSizeInBits(BackedgeTakenCount->getType()) >
3136  IdxTy->getPrimitiveSizeInBits())
3137  BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount, IdxTy);
3138  BackedgeTakenCount = SE->getNoopOrZeroExtend(BackedgeTakenCount, IdxTy);
3139 
3140  // Get the total trip count from the count by adding 1.
3141  const SCEV *ExitCount = SE->getAddExpr(
3142  BackedgeTakenCount, SE->getOne(BackedgeTakenCount->getType()));
3143 
3144  const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
3145 
3146  // Expand the trip count and place the new instructions in the preheader.
3147  // Notice that the pre-header does not change, only the loop body.
3148  SCEVExpander Exp(*SE, DL, "induction");
3149 
3150  // Count holds the overall loop count (N).
3151  TripCount = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
3153 
3154  if (TripCount->getType()->isPointerTy())
3155  TripCount =
3156  CastInst::CreatePointerCast(TripCount, IdxTy, "exitcount.ptrcnt.to.int",
3158 
3159  return TripCount;
3160 }
3161 
3163  if (VectorTripCount)
3164  return VectorTripCount;
3165 
3166  Value *TC = getOrCreateTripCount(L);
3168 
3169  Type *Ty = TC->getType();
3170  // This is where we can make the step a runtime constant.
3172 
3173  // If the tail is to be folded by masking, round the number of iterations N
3174  // up to a multiple of Step instead of rounding down. This is done by first
3175  // adding Step-1 and then rounding down. Note that it's ok if this addition
3176  // overflows: the vector induction variable will eventually wrap to zero given
3177  // that it starts at zero and its Step is a power of two; the loop will then
3178  // exit, with the last early-exit vector comparison also producing all-true.
3179  if (Cost->foldTailByMasking()) {
3181  "VF*UF must be a power of 2 when folding tail by masking");
3182  assert(!VF.isScalable() &&
3183  "Tail folding not yet supported for scalable vectors");
3184  TC = Builder.CreateAdd(
3185  TC, ConstantInt::get(Ty, VF.getKnownMinValue() * UF - 1), "n.rnd.up");
3186  }
3187 
3188  // Now we need to generate the expression for the part of the loop that the
3189  // vectorized body will execute. This is equal to N - (N % Step) if scalar
3190  // iterations are not required for correctness, or N - Step, otherwise. Step
3191  // is equal to the vectorization factor (number of SIMD elements) times the
3192  // unroll factor (number of SIMD instructions).
3193  Value *R = Builder.CreateURem(TC, Step, "n.mod.vf");
3194 
3195  // There are cases where we *must* run at least one iteration in the remainder
3196  // loop. See the cost model for when this can happen. If the step evenly
3197  // divides the trip count, we set the remainder to be equal to the step. If
3198  // the step does not evenly divide the trip count, no adjustment is necessary
3199  // since there will already be scalar iterations. Note that the minimum
3200  // iterations check ensures that N >= Step.
3201  if (Cost->requiresScalarEpilogue(VF)) {
3202  auto *IsZero = Builder.CreateICmpEQ(R, ConstantInt::get(R->getType(), 0));
3203  R = Builder.CreateSelect(IsZero, Step, R);
3204  }
3205 
3206  VectorTripCount = Builder.CreateSub(TC, R, "n.vec");
3207 
3208  return VectorTripCount;
3209 }
3210 
3212  const DataLayout &DL) {
3213  // Verify that V is a vector type with same number of elements as DstVTy.
3214  auto *DstFVTy = cast<FixedVectorType>(DstVTy);
3215  unsigned VF = DstFVTy->getNumElements();
3216  auto *SrcVecTy = cast<FixedVectorType>(V->getType());
3217  assert((VF == SrcVecTy->getNumElements()) && "Vector dimensions do not match");
3218  Type *SrcElemTy = SrcVecTy->getElementType();
3219  Type *DstElemTy = DstFVTy->getElementType();
3220  assert((DL.getTypeSizeInBits(SrcElemTy) == DL.getTypeSizeInBits(DstElemTy)) &&
3221  "Vector elements must have same size");
3222 
3223  // Do a direct cast if element types are castable.
3224  if (CastInst::isBitOrNoopPointerCastable(SrcElemTy, DstElemTy, DL)) {
3225  return Builder.CreateBitOrPointerCast(V, DstFVTy);
3226  }
3227  // V cannot be directly casted to desired vector type.
3228  // May happen when V is a floating point vector but DstVTy is a vector of
3229  // pointers or vice-versa. Handle this using a two-step bitcast using an
3230  // intermediate Integer type for the bitcast i.e. Ptr <-> Int <-> Float.
3231  assert((DstElemTy->isPointerTy() != SrcElemTy->isPointerTy()) &&
3232  "Only one type should be a pointer type");
3233  assert((DstElemTy->isFloatingPointTy() != SrcElemTy->isFloatingPointTy()) &&
3234  "Only one type should be a floating point type");
3235  Type *IntTy =
3236  IntegerType::getIntNTy(V->getContext(), DL.getTypeSizeInBits(SrcElemTy));
3237  auto *VecIntTy = FixedVectorType::get(IntTy, VF);
3238  Value *CastVal = Builder.CreateBitOrPointerCast(V, VecIntTy);
3239  return Builder.CreateBitOrPointerCast(CastVal, DstFVTy);
3240 }
3241 
3243  BasicBlock *Bypass) {
3244  Value *Count = getOrCreateTripCount(L);
3245  // Reuse existing vector loop preheader for TC checks.
3246  // Note that new preheader block is generated for vector loop.
3247  BasicBlock *const TCCheckBlock = LoopVectorPreHeader;
3248  IRBuilder<> Builder(TCCheckBlock->getTerminator());
3249 
3250  // Generate code to check if the loop's trip count is less than VF * UF, or
3251  // equal to it in case a scalar epilogue is required; this implies that the
3252  // vector trip count is zero. This check also covers the case where adding one
3253  // to the backedge-taken count overflowed leading to an incorrect trip count
3254  // of zero. In this case we will also jump to the scalar loop.
3257 
3258  // If tail is to be folded, vector loop takes care of all iterations.
3259  Value *CheckMinIters = Builder.getFalse();
3260  if (!Cost->foldTailByMasking()) {
3261  Value *Step =
3263  CheckMinIters = Builder.CreateICmp(P, Count, Step, "min.iters.check");
3264  }
3265  // Create new preheader for vector loop.
3267  SplitBlock(TCCheckBlock, TCCheckBlock->getTerminator(), DT, LI, nullptr,
3268  "vector.ph");
3269 
3270  assert(DT->properlyDominates(DT->getNode(TCCheckBlock),
3271  DT->getNode(Bypass)->getIDom()) &&
3272  "TC check is expected to dominate Bypass");
3273 
3274  // Update dominator for Bypass & LoopExit (if needed).
3275  DT->changeImmediateDominator(Bypass, TCCheckBlock);
3277  // If there is an epilogue which must run, there's no edge from the
3278  // middle block to exit blocks and thus no need to update the immediate
3279  // dominator of the exit blocks.
3280  DT->changeImmediateDominator(LoopExitBlock, TCCheckBlock);
3281 
3283  TCCheckBlock->getTerminator(),
3284  BranchInst::Create(Bypass, LoopVectorPreHeader, CheckMinIters));
3285  LoopBypassBlocks.push_back(TCCheckBlock);
3286 }
3287 
3289 
3290  BasicBlock *const SCEVCheckBlock =
3292  if (!SCEVCheckBlock)
3293  return nullptr;
3294 
3295  assert(!(SCEVCheckBlock->getParent()->hasOptSize() ||
3298  "Cannot SCEV check stride or overflow when optimizing for size");
3299 
3300 
3301  // Update dominator only if this is first RT check.
3302  if (LoopBypassBlocks.empty()) {
3303  DT->changeImmediateDominator(Bypass, SCEVCheckBlock);
3305  // If there is an epilogue which must run, there's no edge from the
3306  // middle block to exit blocks and thus no need to update the immediate
3307  // dominator of the exit blocks.
3308  DT->changeImmediateDominator(LoopExitBlock, SCEVCheckBlock);
3309  }
3310 
3311  LoopBypassBlocks.push_back(SCEVCheckBlock);
3312  AddedSafetyChecks = true;
3313  return SCEVCheckBlock;
3314 }
3315 
3317  BasicBlock *Bypass) {
3318  // VPlan-native path does not do any analysis for runtime checks currently.
3320  return nullptr;
3321 
3322  BasicBlock *const MemCheckBlock =
3324 
3325  // Check if we generated code that checks in runtime if arrays overlap. We put
3326  // the checks into a separate block to make the more common case of few
3327  // elements faster.
3328  if (!MemCheckBlock)
3329  return nullptr;
3330 
3331  if (MemCheckBlock->getParent()->hasOptSize() || OptForSizeBasedOnProfile) {
3333  "Cannot emit memory checks when optimizing for size, unless forced "
3334  "to vectorize.");
3335  ORE->emit([&]() {
3336  return OptimizationRemarkAnalysis(DEBUG_TYPE, "VectorizationCodeSize",
3337  L->getStartLoc(), L->getHeader())
3338  << "Code-size may be reduced by not forcing "
3339  "vectorization, or by source-code modifications "
3340  "eliminating the need for runtime checks "
3341  "(e.g., adding 'restrict').";
3342  });
3343  }
3344 
3345  LoopBypassBlocks.push_back(MemCheckBlock);
3346 
3347  AddedSafetyChecks = true;
3348 
3349  // We currently don't use LoopVersioning for the actual loop cloning but we
3350  // still use it to add the noalias metadata.
3351  LVer = std::make_unique<LoopVersioning>(
3352  *Legal->getLAI(),
3354  DT, PSE.getSE());
3355  LVer->prepareNoAliasMetadata();
3356  return MemCheckBlock;
3357 }
3358 
3361  const InductionDescriptor &ID) const {
3362 
3363  SCEVExpander Exp(*SE, DL, "induction");
3364  auto Step = ID.getStep();
3365  auto StartValue = ID.getStartValue();
3366  assert(Index->getType()->getScalarType() == Step->getType() &&
3367  "Index scalar type does not match StepValue type");
3368 
3369  // Note: the IR at this point is broken. We cannot use SE to create any new
3370  // SCEV and then expand it, hoping that SCEV's simplification will give us
3371  // a more optimal code. Unfortunately, attempt of doing so on invalid IR may
3372  // lead to various SCEV crashes. So all we can do is to use builder and rely
3373  // on InstCombine for future simplifications. Here we handle some trivial
3374  // cases only.
3375  auto CreateAdd = [&B](Value *X, Value *Y) {
3376  assert(X->getType() == Y->getType() && "Types don't match!");
3377  if (auto *CX = dyn_cast<ConstantInt>(X))
3378  if (CX->isZero())
3379  return Y;
3380  if (auto *CY = dyn_cast<ConstantInt>(Y))
3381  if (CY->isZero())
3382  return X;
3383  return B.CreateAdd(X, Y);
3384  };
3385 
3386  // We allow X to be a vector type, in which case Y will potentially be
3387  // splatted into a vector with the same element count.
3388  auto CreateMul = [&B](Value *X, Value *Y) {
3389  assert(X->getType()->getScalarType() == Y->getType() &&
3390  "Types don't match!");
3391  if (auto *CX = dyn_cast<ConstantInt>(X))
3392  if (CX->isOne())
3393  return Y;
3394  if (auto *CY = dyn_cast<ConstantInt>(Y))
3395  if (CY->isOne())
3396  return X;
3397  VectorType *XVTy = dyn_cast<VectorType>(X->getType());
3398  if (XVTy && !isa<VectorType>(Y->getType()))
3399  Y = B.CreateVectorSplat(XVTy->getElementCount(), Y);
3400  return B.CreateMul(X, Y);
3401  };
3402 
3403  // Get a suitable insert point for SCEV expansion. For blocks in the vector
3404  // loop, choose the end of the vector loop header (=LoopVectorBody), because
3405  // the DomTree is not kept up-to-date for additional blocks generated in the
3406  // vector loop. By using the header as insertion point, we guarantee that the
3407  // expanded instructions dominate all their uses.
3408  auto GetInsertPoint = [this, &B]() {
3409  BasicBlock *InsertBB = B.GetInsertPoint()->getParent();
3410  if (InsertBB != LoopVectorBody &&
3411  LI->getLoopFor(LoopVectorBody) == LI->getLoopFor(InsertBB))
3412  return LoopVectorBody->getTerminator();
3413  return &*B.GetInsertPoint();
3414  };
3415 
3416  switch (ID.getKind()) {
3418  assert(!isa<VectorType>(Index->getType()) &&
3419  "Vector indices not supported for integer inductions yet");
3420  assert(Index->getType() == StartValue->getType() &&
3421  "Index type does not match StartValue type");
3422  if (ID.getConstIntStepValue() && ID.getConstIntStepValue()->isMinusOne())
3423  return B.CreateSub(StartValue, Index);
3424  auto *Offset = CreateMul(
3425  Index, Exp.expandCodeFor(Step, Index->getType(), GetInsertPoint()));
3426  return CreateAdd(StartValue, Offset);
3427  }
3429  assert(isa<SCEVConstant>(Step) &&
3430  "Expected constant step for pointer induction");
3431  return B.CreateGEP(
3432  ID.getElementType(), StartValue,
3433  CreateMul(Index,
3434  Exp.expandCodeFor(Step, Index->getType()->getScalarType(),
3435  GetInsertPoint())));
3436  }
3438  assert(!isa<VectorType>(Index->getType()) &&
3439  "Vector indices not supported for FP inductions yet");
3440  assert(Step->getType()->isFloatingPointTy() && "Expected FP Step value");
3441  auto InductionBinOp = ID.getInductionBinOp();
3442  assert(InductionBinOp &&
3443  (InductionBinOp->getOpcode() == Instruction::FAdd ||
3444  InductionBinOp->getOpcode() == Instruction::FSub) &&
3445  "Original bin op should be defined for FP induction");
3446 
3447  Value *StepValue = cast<SCEVUnknown>(Step)->getValue();
3448  Value *MulExp = B.CreateFMul(StepValue, Index);
3449  return B.CreateBinOp(InductionBinOp->getOpcode(), StartValue, MulExp,
3450  "induction");
3451  }
3453  return nullptr;
3454  }
3455  llvm_unreachable("invalid enum");
3456 }
3457 
3461  assert(LoopVectorPreHeader && "Invalid loop structure");
3462  LoopExitBlock = OrigLoop->getUniqueExitBlock(); // may be nullptr
3464  "multiple exit loop without required epilogue?");
3465 
3466  LoopMiddleBlock =
3468  LI, nullptr, Twine(Prefix) + "middle.block");
3471  nullptr, Twine(Prefix) + "scalar.ph");
3472 
3473  auto *ScalarLatchTerm = OrigLoop->getLoopLatch()->getTerminator();
3474 
3475  // Set up the middle block terminator. Two cases:
3476  // 1) If we know that we must execute the scalar epilogue, emit an
3477  // unconditional branch.
3478  // 2) Otherwise, we must have a single unique exit block (due to how we
3479  // implement the multiple exit case). In this case, set up a conditonal
3480  // branch from the middle block to the loop scalar preheader, and the
3481  // exit block. completeLoopSkeleton will update the condition to use an
3482  // iteration check, if required to decide whether to execute the remainder.
3486  Builder.getTrue());
3487  BrInst->setDebugLoc(ScalarLatchTerm->getDebugLoc());
3489 
3490  // We intentionally don't let SplitBlock to update LoopInfo since
3491  // LoopVectorBody should belong to another loop than LoopVectorPreHeader.
3492  // LoopVectorBody is explicitly added to the correct place few lines later.
3493  LoopVectorBody =
3495  nullptr, nullptr, Twine(Prefix) + "vector.body");
3496 
3497  // Update dominator for loop exit.
3499  // If there is an epilogue which must run, there's no edge from the
3500  // middle block to exit blocks and thus no need to update the immediate
3501  // dominator of the exit blocks.
3503 
3504  // Create and register the new vector loop.
3505  Loop *Lp = LI->AllocateLoop();
3506  Loop *ParentLoop = OrigLoop->getParentLoop();
3507 
3508  // Insert the new loop into the loop nest and register the new basic blocks
3509  // before calling any utilities such as SCEV that require valid LoopInfo.
3510  if (ParentLoop) {
3511  ParentLoop->addChildLoop(Lp);
3512  } else {
3513  LI->addTopLevelLoop(Lp);
3514  }
3516  return Lp;
3517 }
3518 
3520  Loop *L, Value *VectorTripCount,
3521  std::pair<BasicBlock *, Value *> AdditionalBypass) {
3522  assert(VectorTripCount && L && "Expected valid arguments");
3523  assert(((AdditionalBypass.first && AdditionalBypass.second) ||
3524  (!AdditionalBypass.first && !AdditionalBypass.second)) &&
3525  "Inconsistent information about additional bypass.");
3526  // We are going to resume the execution of the scalar loop.
3527  // Go over all of the induction variables that we found and fix the
3528  // PHIs that are left in the scalar version of the loop.
3529  // The starting values of PHI nodes depend on the counter of the last
3530  // iteration in the vectorized loop.
3531  // If we come from a bypass edge then we need to start from the original
3532  // start value.
3533  for (auto &InductionEntry : Legal->getInductionVars()) {
3534  PHINode *OrigPhi = InductionEntry.first;
3535  InductionDescriptor II = InductionEntry.second;
3536 
3537  // Create phi nodes to merge from the backedge-taken check block.
3538  PHINode *BCResumeVal =
3539  PHINode::Create(OrigPhi->getType(), 3, "bc.resume.val",
3541  // Copy original phi DL over to the new one.
3542  BCResumeVal->setDebugLoc(OrigPhi->getDebugLoc());
3543  Value *&EndValue = IVEndValues[OrigPhi];
3544  Value *EndValueFromAdditionalBypass = AdditionalBypass.second;
3545  if (OrigPhi == OldInduction) {
3546  // We know what the end value is.
3547  EndValue = VectorTripCount;
3548  } else {
3550 
3551  // Fast-math-flags propagate from the original induction instruction.
3552  if (II.getInductionBinOp() && isa<FPMathOperator>(II.getInductionBinOp()))
3553  B.setFastMathFlags(II.getInductionBinOp()->getFastMathFlags());
3554 
3555  Type *StepType = II.getStep()->getType();
3556  Instruction::CastOps CastOp =
3557  CastInst::getCastOpcode(VectorTripCount, true, StepType, true);
3558  Value *CRD = B.CreateCast(CastOp, VectorTripCount, StepType, "cast.crd");
3560  EndValue = emitTransformedIndex(B, CRD, PSE.getSE(), DL, II);
3561  EndValue->setName("ind.end");
3562 
3563  // Compute the end value for the additional bypass (if applicable).
3564  if (AdditionalBypass.first) {
3565  B.SetInsertPoint(&(*AdditionalBypass.first->getFirstInsertionPt()));
3566  CastOp = CastInst::getCastOpcode(AdditionalBypass.second, true,
3567  StepType, true);
3568  CRD =
3569  B.CreateCast(CastOp, AdditionalBypass.second, StepType, "cast.crd");
3570  EndValueFromAdditionalBypass =
3571  emitTransformedIndex(B, CRD, PSE.getSE(), DL, II);
3572  EndValueFromAdditionalBypass->setName("ind.end");
3573  }
3574  }
3575  // The new PHI merges the original incoming value, in case of a bypass,
3576  // or the value at the end of the vectorized loop.
3577  BCResumeVal->addIncoming(EndValue, LoopMiddleBlock);
3578 
3579  // Fix the scalar body counter (PHI node).
3580  // The old induction's phi node in the scalar body needs the truncated
3581  // value.
3582  for (BasicBlock *BB : LoopBypassBlocks)
3583  BCResumeVal->addIncoming(II.getStartValue(), BB);
3584 
3585  if (AdditionalBypass.first)
3586  BCResumeVal->setIncomingValueForBlock(AdditionalBypass.first,
3587  EndValueFromAdditionalBypass);
3588 
3589  OrigPhi->setIncomingValueForBlock(LoopScalarPreHeader, BCResumeVal);
3590  }
3591 }
3592 
3594  MDNode *OrigLoopID) {
3595  assert(L && "Expected valid loop.");
3596 
3597  // The trip counts should be cached by now.
3598  Value *Count = getOrCreateTripCount(L);
3600 
3601  auto *ScalarLatchTerm = OrigLoop->getLoopLatch()->getTerminator();
3602 
3603  // Add a check in the middle block to see if we have completed
3604  // all of the iterations in the first vector loop. Three cases:
3605  // 1) If we require a scalar epilogue, there is no conditional branch as
3606  // we unconditionally branch to the scalar preheader. Do nothing.
3607  // 2) If (N - N%VF) == N, then we *don't* need to run the remainder.
3608  // Thus if tail is to be folded, we know we don't need to run the
3609  // remainder and we can use the previous value for the condition (true).
3610  // 3) Otherwise, construct a runtime check.
3612  Instruction *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ,
3613  Count, VectorTripCount, "cmp.n",
3615 
3616  // Here we use the same DebugLoc as the scalar loop latch terminator instead
3617  // of the corresponding compare because they may have ended up with
3618  // different line numbers and we want to avoid awkward line stepping while
3619  // debugging. Eg. if the compare has got a line number inside the loop.
3620  CmpN->setDebugLoc(ScalarLatchTerm->getDebugLoc());
3621  cast<BranchInst>(LoopMiddleBlock->getTerminator())->setCondition(CmpN);
3622  }
3623 
3624  // Get ready to start creating new instructions into the vectorized body.
3626  "Inconsistent vector loop preheader");
3628 
3629  Optional<MDNode *> VectorizedLoopID =
3632  if (VectorizedLoopID.hasValue()) {
3633  L->setLoopID(VectorizedLoopID.getValue());
3634 
3635  // Do not setAlreadyVectorized if loop attributes have been defined
3636  // explicitly.
3637  return LoopVectorPreHeader;
3638  }
3639 
3640  // Keep all loop hints from the original loop on the vector loop (we'll
3641  // replace the vectorizer-specific hints below).
3642  if (MDNode *LID = OrigLoop->getLoopID())
3643  L->setLoopID(LID);
3644 
3645  LoopVectorizeHints Hints(L, true, *ORE);
3646  Hints.setAlreadyVectorized();
3647 
3648 #ifdef EXPENSIVE_CHECKS
3650  LI->verify(*DT);
3651 #endif
3652 
3653  return LoopVectorPreHeader;
3654 }
3655 
3657  /*
3658  In this function we generate a new loop. The new loop will contain
3659  the vectorized instructions while the old loop will continue to run the
3660  scalar remainder.
3661 
3662  [ ] <-- loop iteration number check.
3663  / |
3664  / v
3665  | [ ] <-- vector loop bypass (may consist of multiple blocks).
3666  | / |
3667  | / v
3668  || [ ] <-- vector pre header.
3669  |/ |
3670  | v
3671  | [ ] \
3672  | [ ]_| <-- vector loop.
3673  | |
3674  | v
3675  \ -[ ] <--- middle-block.
3676  \/ |
3677  /\ v
3678  | ->[ ] <--- new preheader.
3679  | |
3680  (opt) v <-- edge from middle to exit iff epilogue is not required.
3681  | [ ] \
3682  | [ ]_| <-- old scalar loop to handle remainder (scalar epilogue).
3683  \ |
3684  \ v
3685  >[ ] <-- exit block(s).
3686  ...
3687  */
3688 
3689  // Get the metadata of the original loop before it gets modified.
3690  MDNode *OrigLoopID = OrigLoop->getLoopID();
3691 
3692  // Workaround! Compute the trip count of the original loop and cache it
3693  // before we start modifying the CFG. This code has a systemic problem
3694  // wherein it tries to run analysis over partially constructed IR; this is
3695  // wrong, and not simply for SCEV. The trip count of the original loop
3696  // simply happens to be prone to hitting this in practice. In theory, we
3697  // can hit the same issue for any SCEV, or ValueTracking query done during
3698  // mutation. See PR49900.
3700 
3701  // Create an empty vector loop, and prepare basic blocks for the runtime
3702  // checks.
3703  Loop *Lp = createVectorLoopSkeleton("");
3704 
3705  // Now, compare the new count to zero. If it is zero skip the vector loop and
3706  // jump to the scalar loop. This check also covers the case where the
3707  // backedge-taken count is uint##_max: adding one to it will overflow leading
3708  // to an incorrect trip count of zero. In this (rare) case we will also jump
3709  // to the scalar loop.
3711 
3712  // Generate the code to check any assumptions that we've made for SCEV
3713  // expressions.
3715 
3716  // Generate the code that checks in runtime if arrays overlap. We put the
3717  // checks into a separate block to make the more common case of few elements
3718  // faster.
3720 
3721  // Some loops have a single integer induction variable, while other loops
3722  // don't. One example is c++ iterators that often have multiple pointer
3723  // induction variables. In the code below we also support a case where we
3724  // don't have a single induction variable.
3725  //
3726  // We try to obtain an induction variable from the original loop as hard
3727  // as possible. However if we don't find one that:
3728  // - is an integer
3729  // - counts from zero, stepping by one
3730  // - is the size of the widest induction variable type
3731  // then we create a new one.
3733  Type *IdxTy = Legal->getWidestInductionType();
3734  Value *StartIdx = ConstantInt::get(IdxTy, 0);
3735  // The loop step is equal to the vectorization factor (num of SIMD elements)
3736  // times the unroll factor (num of SIMD instructions).
3738  Value *Step = createStepForVF(Builder, ConstantInt::get(IdxTy, UF), VF);
3739  Value *CountRoundDown = getOrCreateVectorTripCount(Lp);
3740  Induction =
3741  createInductionVariable(Lp, StartIdx, CountRoundDown, Step,
3743 
3744  // Emit phis for the new starting index of the scalar loop.
3745  createInductionResumeValues(Lp, CountRoundDown);
3746 
3747  return completeLoopSkeleton(Lp, OrigLoopID);
3748 }
3749 
3750 // Fix up external users of the induction variable. At this point, we are
3751 // in LCSSA form, with all external PHIs that use the IV having one input value,
3752 // coming from the remainder loop. We need those PHIs to also have a correct
3753 // value for the IV when arriving directly from the middle block.
3755  const InductionDescriptor &II,
3756  Value *CountRoundDown, Value *EndValue,
3757  BasicBlock *MiddleBlock) {
3758  // There are two kinds of external IV usages - those that use the value
3759  // computed in the last iteration (the PHI) and those that use the penultimate
3760  // value (the value that feeds into the phi from the loop latch).
3761  // We allow both, but they, obviously, have different values.
3762 
3763  assert(OrigLoop->getUniqueExitBlock() && "Expected a single exit block");
3764 
3765  DenseMap<Value *, Value *> MissingVals;
3766 
3767  // An external user of the last iteration's value should see the value that
3768  // the remainder loop uses to initialize its own IV.
3770  for (User *U : PostInc->users()) {
3771  Instruction *UI = cast<Instruction>(U);
3772  if (!OrigLoop->contains(UI)) {
3773  assert(isa<PHINode>(UI) && "Expected LCSSA form");
3774  MissingVals[UI] = EndValue;
3775  }
3776  }
3777 
3778  // An external user of the penultimate value need to see EndValue - Step.
3779  // The simplest way to get this is to recompute it from the constituent SCEVs,
3780  // that is Start + (Step * (CRD - 1)).
3781  for (User *U : OrigPhi->users()) {
3782  auto *UI = cast<Instruction>(U);
3783  if (!OrigLoop->contains(UI)) {
3784  const DataLayout &DL =
3786  assert(isa<PHINode>(UI) && "Expected LCSSA form");
3787 
3788  IRBuilder<> B(MiddleBlock->getTerminator());
3789 
3790  // Fast-math-flags propagate from the original induction instruction.
3791  if (II.getInductionBinOp() && isa<FPMathOperator>(II.getInductionBinOp()))
3792  B.setFastMathFlags(II.getInductionBinOp()->getFastMathFlags());
3793 
3794  Value *CountMinusOne = B.CreateSub(
3795  CountRoundDown, ConstantInt::get(CountRoundDown->getType(), 1));
3796  Value *CMO =
3797  !II.getStep()->getType()->isIntegerTy()
3798  ? B.CreateCast(Instruction::SIToFP, CountMinusOne,
3799  II.getStep()->getType())
3800  : B.CreateSExtOrTrunc(CountMinusOne, II.getStep()->getType());
3801  CMO->setName("cast.cmo");
3802  Value *Escape = emitTransformedIndex(B, CMO, PSE.getSE(), DL, II);
3803  Escape->setName("ind.escape");
3804  MissingVals[UI] = Escape;
3805  }
3806  }
3807 
3808  for (auto &I : MissingVals) {
3809  PHINode *PHI = cast<PHINode>(I.first);
3810  // One corner case we have to handle is two IVs "chasing" each-other,
3811  // that is %IV2 = phi [...], [ %IV1, %latch ]
3812  // In this case, if IV1 has an external use, we need to avoid adding both
3813  // "last value of IV1" and "penultimate value of IV2". So, verify that we
3814  // don't already have an incoming value for the middle block.
3815  if (PHI->getBasicBlockIndex(MiddleBlock) == -1)
3816  PHI->addIncoming(I.second, MiddleBlock);
3817  }
3818 }
3819 
3820 namespace {
3821 
3822 struct CSEDenseMapInfo {
3823  static bool canHandle(const Instruction *I) {
3824  return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
3825  isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
3826  }
3827 
3828  static inline Instruction *getEmptyKey() {
3830  }
3831 
3832  static inline Instruction *getTombstoneKey() {
3834  }
3835 
3836  static unsigned getHashValue(const Instruction *I) {
3837  assert(canHandle(I) && "Unknown instruction!");
3838  return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
3839  I->value_op_end()));
3840  }
3841 
3842  static bool isEqual(const Instruction *LHS, const Instruction *RHS) {
3843  if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
3844  LHS == getTombstoneKey() || RHS == getTombstoneKey())
3845  return LHS == RHS;
3846  return LHS->isIdenticalTo(RHS);
3847  }
3848 };
3849 
3850 } // end anonymous namespace
3851 
3852 ///Perform cse of induction variable instructions.
3853 static void cse(BasicBlock *BB) {
3854  // Perform simple cse.
3857  if (!CSEDenseMapInfo::canHandle(&In))
3858  continue;
3859 
3860  // Check if we can replace this instruction with any of the
3861  // visited instructions.
3862  if (Instruction *V = CSEMap.lookup(&In)) {
3863  In.replaceAllUsesWith(V);
3864  In.eraseFromParent();
3865  continue;
3866  }
3867 
3868  CSEMap[&In] = &In;
3869  }
3870 }
3871 
3874  bool &NeedToScalarize) const {
3875  Function *F = CI->getCalledFunction();
3876  Type *ScalarRetTy = CI->getType();
3877  SmallVector<Type *, 4> Tys, ScalarTys;
3878  for (auto &ArgOp : CI->args())
3879  ScalarTys.push_back(ArgOp->getType());
3880 
3881  // Estimate cost of scalarized vector call. The source operands are assumed
3882  // to be vectors, so we need to extract individual elements from there,
3883  // execute VF scalar calls, and then gather the result into the vector return
3884  // value.
3885  InstructionCost ScalarCallCost =
3886  TTI.getCallInstrCost(F, ScalarRetTy, ScalarTys, TTI::TCK_RecipThroughput);
3887  if (VF.isScalar())
3888  return ScalarCallCost;
3889 
3890  // Compute corresponding vector type for return value and arguments.
3891  Type *RetTy = ToVectorTy(ScalarRetTy, VF);
3892  for (Type *ScalarTy : ScalarTys)
3893  Tys.push_back(ToVectorTy(ScalarTy, VF));
3894 
3895  // Compute costs of unpacking argument values for the scalar calls and
3896  // packing the return values to a vector.
3897  InstructionCost ScalarizationCost = getScalarizationOverhead(CI, VF);
3898 
3900  ScalarCallCost * VF.getKnownMinValue() + ScalarizationCost;
3901 
3902  // If we can't emit a vector call for this function, then the currently found
3903  // cost is the cost we need to return.
3904  NeedToScalarize = true;
3905  VFShape Shape = VFShape::get(*CI, VF, false /*HasGlobalPred*/);
3906  Function *VecFunc = VFDatabase(*CI).getVectorizedFunction(Shape);
3907 
3908  if (!TLI || CI->isNoBuiltin() || !VecFunc)
3909  return Cost;
3910 
3911  // If the corresponding vector cost is cheaper, return its cost.
3912  InstructionCost VectorCallCost =
3913  TTI.getCallInstrCost(nullptr, RetTy, Tys, TTI::TCK_RecipThroughput);
3914  if (VectorCallCost < Cost) {
3915  NeedToScalarize = false;
3916  Cost = VectorCallCost;
3917  }
3918  return Cost;
3919 }
3920 
3922  if (VF.isScalar() || (!Elt->isIntOrPtrTy() && !Elt->isFloatingPointTy()))
3923  return Elt;
3924  return VectorType::get(Elt, VF);
3925 }
3926 
3929  ElementCount VF) const {
3931  assert(ID && "Expected intrinsic call!");
3932  Type *RetTy = MaybeVectorizeType(CI->getType(), VF);
3933  FastMathFlags FMF;
3934  if (auto *FPMO = dyn_cast<FPMathOperator>(CI))
3935  FMF = FPMO->getFastMathFlags();
3936 
3939  SmallVector<Type *> ParamTys;
3940  std::transform(FTy->param_begin(), FTy->param_end(),
3941  std::back_inserter(ParamTys),
3942  [&](Type *Ty) { return MaybeVectorizeType(Ty, VF); });
3943 
3944  IntrinsicCostAttributes CostAttrs(ID, RetTy, Arguments, ParamTys, FMF,
3945  dyn_cast<IntrinsicInst>(CI));
3946  return TTI.getIntrinsicInstrCost(CostAttrs,
3948 }
3949 
3951  auto *I1 = cast<IntegerType>(cast<VectorType>(T1)->getElementType());
3952  auto *I2 = cast<IntegerType>(cast<VectorType>(T2)->getElementType());
3953  return I1->getBitWidth() < I2->getBitWidth() ? T1 : T2;
3954 }
3955 
3957  auto *I1 = cast<IntegerType>(cast<VectorType>(T1)->getElementType());
3958  auto *I2 = cast<IntegerType>(cast<VectorType>(T2)->getElementType());
3959  return I1->getBitWidth() > I2->getBitWidth() ? T1 : T2;
3960 }
3961 
3963  // For every instruction `I` in MinBWs, truncate the operands, create a
3964  // truncated version of `I` and reextend its result. InstCombine runs
3965  // later and will remove any ext/trunc pairs.
3966  SmallPtrSet<Value *, 4> Erased;
3967  for (const auto &KV : Cost->getMinimalBitwidths()) {
3968  // If the value wasn't vectorized, we must maintain the original scalar
3969  // type. The absence of the value from State indicates that it
3970  // wasn't vectorized.
3971  // FIXME: Should not rely on getVPValue at this point.
3972  VPValue *Def = State.Plan->getVPValue(KV.first, true);
3973  if (!State.hasAnyVectorValue(Def))
3974  continue;
3975  for (unsigned Part = 0; Part < UF; ++Part) {
3976  Value *I = State.get(Def, Part);
3977  if (Erased.count(I) || I->use_empty() || !isa<Instruction>(I))
3978  continue;
3979  Type *OriginalTy = I->getType();
3980  Type *ScalarTruncatedTy =
3981  IntegerType::get(OriginalTy->getContext(), KV.second);
3982  auto *TruncatedTy = VectorType::get(
3983  ScalarTruncatedTy, cast<VectorType>(OriginalTy)->getElementCount());
3984  if (TruncatedTy == OriginalTy)
3985  continue;
3986 
3987  IRBuilder<> B(cast<Instruction>(I));
3988  auto ShrinkOperand = [&](Value *V) -> Value * {
3989  if (auto *ZI = dyn_cast<ZExtInst>(V))
3990  if (ZI->getSrcTy() == TruncatedTy)
3991  return ZI->getOperand(0);
3992  return B.CreateZExtOrTrunc(V, TruncatedTy);
3993  };
3994 
3995  // The actual instruction modification depends on the instruction type,
3996  // unfortunately.
3997  Value *NewI = nullptr;
3998  if (auto *BO = dyn_cast<BinaryOperator>(I)) {
3999  NewI = B.CreateBinOp(BO->getOpcode(), ShrinkOperand(BO->getOperand(0)),
4000  ShrinkOperand(BO->getOperand(1)));
4001 
4002  // Any wrapping introduced by shrinking this operation shouldn't be
4003  // considered undefined behavior. So, we can't unconditionally copy
4004  // arithmetic wrapping flags to NewI.
4005  cast<BinaryOperator>(NewI)->copyIRFlags(I, /*IncludeWrapFlags=*/false);
4006  } else if (auto *CI = dyn_cast<ICmpInst>(I)) {
4007  NewI =
4008  B.CreateICmp(CI->getPredicate(), ShrinkOperand(CI->getOperand(0)),
4009  ShrinkOperand(CI->getOperand(1)));
4010  } else if (auto *SI = dyn_cast<SelectInst>(I)) {
4011  NewI = B.CreateSelect(SI->getCondition(),
4012  ShrinkOperand(SI->getTrueValue()),
4013  ShrinkOperand(SI->getFalseValue()));
4014  } else if (auto *CI = dyn_cast<CastInst>(I)) {
4015  switch (CI->getOpcode()) {
4016  default:
4017  llvm_unreachable("Unhandled cast!");
4018  case Instruction::Trunc:
4019  NewI = ShrinkOperand(CI->getOperand(0));
4020  break;
4021  case Instruction::SExt:
4022  NewI = B.CreateSExtOrTrunc(
4023  CI->getOperand(0),
4024  smallestIntegerVectorType(OriginalTy, TruncatedTy));
4025  break;
4026  case Instruction::ZExt:
4027  NewI = B.CreateZExtOrTrunc(
4028  CI->getOperand(0),
4029  smallestIntegerVectorType(OriginalTy, TruncatedTy));
4030  break;
4031  }
4032  } else if (auto *SI = dyn_cast<ShuffleVectorInst>(I)) {
4033  auto Elements0 =
4034  cast<VectorType>(SI->getOperand(0)->getType())->getElementCount();
4035  auto *O0 = B.CreateZExtOrTrunc(
4036  SI->getOperand(0), VectorType::get(ScalarTruncatedTy, Elements0));
4037  auto Elements1 =
4038  cast<VectorType>(SI->getOperand(1)->getType())->getElementCount();
4039  auto *O1 = B.CreateZExtOrTrunc(
4040  SI->getOperand(1), VectorType::get(ScalarTruncatedTy, Elements1));
4041 
4042  NewI = B.CreateShuffleVector(O0, O1, SI->getShuffleMask());
4043  } else if (isa<LoadInst>(I) || isa<PHINode>(I)) {
4044  // Don't do anything with the operands, just extend the result.
4045  continue;
4046  } else if (auto *IE = dyn_cast<InsertElementInst>(I)) {
4047  auto Elements =
4048  cast<VectorType>(IE->getOperand(0)->getType())->getElementCount();
4049  auto *O0 = B.CreateZExtOrTrunc(
4050  IE->getOperand(0), VectorType::get(ScalarTruncatedTy, Elements));
4051  auto *O1 = B.CreateZExtOrTrunc(IE->getOperand(1), ScalarTruncatedTy);
4052  NewI = B.CreateInsertElement(O0, O1, IE->getOperand(2));
4053  } else if (auto *EE = dyn_cast<ExtractElementInst>(I)) {
4054  auto Elements =
4055  cast<VectorType>(EE->getOperand(0)->getType())->getElementCount();
4056  auto *O0 = B.CreateZExtOrTrunc(
4057  EE->getOperand(0), VectorType::get(ScalarTruncatedTy, Elements));
4058  NewI = B.CreateExtractElement(O0, EE->getOperand(2));
4059  } else {
4060  // If we don't know what to do, be conservative and don't do anything.
4061  continue;
4062  }
4063 
4064  // Lastly, extend the result.
4065  NewI->takeName(cast<Instruction>(I));
4066  Value *Res = B.CreateZExtOrTrunc(NewI, OriginalTy);
4067  I->replaceAllUsesWith(Res);
4068  cast<Instruction>(I)->eraseFromParent();
4069  Erased.insert(I);
4070  State.reset(Def, Res, Part);
4071  }
4072  }
4073 
4074  // We'll have created a bunch of ZExts that are now parentless. Clean up.
4075  for (const auto &KV : Cost->getMinimalBitwidths()) {
4076  // If the value wasn't vectorized, we must maintain the original scalar
4077  // type. The absence of the value from State indicates that it
4078  // wasn't vectorized.
4079  // FIXME: Should not rely on getVPValue at this point.
4080  VPValue *Def = State.Plan->getVPValue(KV.first, true);
4081  if (!State.hasAnyVectorValue(Def))
4082  continue;
4083  for (unsigned Part = 0; Part < UF; ++Part) {
4084  Value *I = State.get(Def, Part);
4085  ZExtInst *Inst = dyn_cast<ZExtInst>(I);
4086  if (Inst && Inst->use_empty()) {
4087  Value *NewI = Inst->getOperand(0);
4088  Inst->eraseFromParent();
4089  State.reset(Def, NewI, Part);
4090  }
4091  }
4092  }
4093 }
4094 
4096  // Insert truncates and extends for any truncated instructions as hints to
4097  // InstCombine.
4098  if (VF.isVector())
4100 
4101  // Fix widened non-induction PHIs by setting up the PHI operands.
4102  if (OrigPHIsToFix.size()) {
4104  "Unexpected non-induction PHIs for fixup in non VPlan-native path");
4105  fixNonInductionPHIs(State);
4106  }
4107 
4108  // At this point every instruction in the original loop is widened to a
4109  // vector form. Now we need to fix the recurrences in the loop. These PHI
4110  // nodes are currently empty because we did not want to introduce cycles.
4111  // This is the second stage of vectorizing recurrences.
4112  fixCrossIterationPHIs(State);
4113 
4114  // Forget the original basic block.
4116 
4117  // If we inserted an edge from the middle block to the unique exit block,
4118  // update uses outside the loop (phis) to account for the newly inserted
4119  // edge.
4120  if (!Cost->requiresScalarEpilogue(VF)) {
4121  // Fix-up external users of the induction variables.
4122  for (auto &Entry : Legal->getInductionVars())
4123  fixupIVUsers(Entry.first, Entry.second,
4125  IVEndValues[Entry.first], LoopMiddleBlock);
4126 
4127  fixLCSSAPHIs(State);
4128  }
4129 
4131  sinkScalarOperands(&*PI);
4132 
4133  // Remove redundant induction instructions.
4135 
4136  // Set/update profile weights for the vector and remainder loops as original
4137  // loop iterations are now distributed among them. Note that original loop
4138  // represented by LoopScalarBody becomes remainder loop after vectorization.
4139  //
4140  // For cases like foldTailByMasking() and requiresScalarEpiloque() we may
4141  // end up getting slightly roughened result but that should be OK since
4142  // profile is not inherently precise anyway. Note also possible bypass of
4143  // vector code caused by legality checks is ignored, assigning all the weight
4144  // to the vector loop, optimistically.
4145  //
4146  // For scalable vectorization we can't know at compile time how many iterations
4147  // of the loop are handled in one vector iteration, so instead assume a pessimistic
4148  // vscale of '1'.
4152 }
4153 
4155  // In order to support recurrences we need to be able to vectorize Phi nodes.
4156  // Phi nodes have cycles, so we need to vectorize them in two stages. This is
4157  // stage #2: We now need to fix the recurrences by adding incoming edges to
4158  // the currently empty PHI nodes. At this point every instruction in the
4159  // original loop is widened to a vector form so we can use them to construct
4160  // the incoming edges.
4161  VPBasicBlock *Header = State.Plan->getEntry()->getEntryBasicBlock();
4162  for (VPRecipeBase &R : Header->phis()) {
4163  if (auto *ReductionPhi = dyn_cast<VPReductionPHIRecipe>(&R))
4164  fixReduction(ReductionPhi, State);
4165  else if (auto *FOR = dyn_cast<VPFirstOrderRecurrencePHIRecipe>(&R))
4166  fixFirstOrderRecurrence(FOR, State);
4167  }
4168 }
4169 
4171  VPTransformState &State) {
4172  // This is the second phase of vectorizing first-order recurrences. An
4173  // overview of the transformation is described below. Suppose we have the
4174  // following loop.
4175  //
4176  // for (int i = 0; i < n; ++i)
4177  // b[i] = a[i] - a[i - 1];
4178  //
4179  // There is a first-order recurrence on "a". For this loop, the shorthand
4180  // scalar IR looks like:
4181  //
4182  // scalar.ph:
4183  // s_init = a[-1]
4184  // br scalar.body
4185  //
4186  // scalar.body:
4187  // i = phi [0, scalar.ph], [i+1, scalar.body]
4188  // s1 = phi [s_init, scalar.ph], [s2, scalar.body]
4189  // s2 = a[i]
4190  // b[i] = s2 - s1
4191  // br cond, scalar.body, ...
4192  //
4193  // In this example, s1 is a recurrence because it's value depends on the
4194  // previous iteration. In the first phase of vectorization, we created a
4195  // vector phi v1 for s1. We now complete the vectorization and produce the
4196  // shorthand vector IR shown below (for VF = 4, UF = 1).
4197  //
4198  // vector.ph:
4199  // v_init = vector(..., ..., ..., a[-1])
4200  // br vector.body
4201  //
4202  // vector.body
4203  // i = phi [0, vector.ph], [i+4, vector.body]
4204  // v1 = phi [v_init, vector.ph], [v2, vector.body]
4205  // v2 = a[i, i+1, i+2, i+3];
4206  // v3 = vector(v1(3), v2(0, 1, 2))
4207  // b[i, i+1, i+2, i+3] = v2 - v3
4208  // br cond, vector.body, middle.block
4209  //
4210  // middle.block:
4211  // x = v2(3)
4212  // br scalar.ph
4213  //
4214  // scalar.ph:
4215  // s_init = phi [x, middle.block], [a[-1], otherwise]
4216  // br scalar.body
4217  //
4218  // After execution completes the vector loop, we extract the next value of
4219  // the recurrence (x) to use as the initial value in the scalar loop.
4220 
4221  // Extract the last vector element in the middle block. This will be the
4222  // initial value for the recurrence when jumping to the scalar loop.
4223  VPValue *PreviousDef = PhiR->getBackedgeValue();
4224  Value *Incoming = State.get(PreviousDef, UF - 1);
4225  auto *ExtractForScalar = Incoming;
4226  auto *IdxTy = Builder.getInt32Ty();
4227  if (VF.isVector()) {
4228  auto *One = ConstantInt::get(IdxTy, 1);
4230  auto *RuntimeVF = getRuntimeVF(Builder, IdxTy, VF);
4231  auto *LastIdx = Builder.CreateSub(RuntimeVF, One);
4232  ExtractForScalar = Builder.CreateExtractElement(ExtractForScalar, LastIdx,
4233  "vector.recur.extract");
4234  }
4235  // Extract the second last element in the middle block if the
4236  // Phi is used outside the loop. We need to extract the phi itself
4237  // and not the last element (the phi update in the current iteration). This
4238  // will be the value when jumping to the exit block from the LoopMiddleBlock,
4239  // when the scalar loop is not run at all.
4240  Value *ExtractForPhiUsedOutsideLoop = nullptr;
4241  if (VF.isVector()) {
4242  auto *RuntimeVF = getRuntimeVF(Builder, IdxTy, VF);
4243  auto *Idx = Builder.CreateSub(RuntimeVF, ConstantInt::get(IdxTy, 2));
4244  ExtractForPhiUsedOutsideLoop = Builder.CreateExtractElement(
4245  Incoming, Idx, "vector.recur.extract.for.phi");
4246  } else if (UF > 1)
4247  // When loop is unrolled without vectorizing, initialize
4248  // ExtractForPhiUsedOutsideLoop with the value just prior to unrolled value
4249  // of `Incoming`. This is analogous to the vectorized case above: extracting
4250  // the second last element when VF > 1.
4251  ExtractForPhiUsedOutsideLoop = State.get(PreviousDef, UF - 2);
4252 
4253  // Fix the initial value of the original recurrence in the scalar loop.
4255  PHINode *Phi = cast<PHINode>(PhiR->getUnderlyingValue());
4256  auto *Start = Builder.CreatePHI(Phi->getType(), 2, "scalar.recur.init");
4257  auto *ScalarInit = PhiR->getStartValue()->getLiveInIRValue();
4258  for (auto *BB : predecessors(LoopScalarPreHeader)) {
4259  auto *Incoming = BB == LoopMiddleBlock ? ExtractForScalar : ScalarInit;
4260  Start->addIncoming(Incoming, BB);
4261  }
4262 
4264  Phi->setName("scalar.recur");
4265 
4266  // Finally, fix users of the recurrence outside the loop. The users will need
4267  // either the last value of the scalar recurrence or the last value of the
4268  // vector recurrence we extracted in the middle block. Since the loop is in
4269  // LCSSA form, we just need to find all the phi nodes for the original scalar
4270  // recurrence in the exit block, and then add an edge for the middle block.
4271  // Note that LCSSA does not imply single entry when the original scalar loop
4272  // had multiple exiting edges (as we always run the last iteration in the
4273  // scalar epilogue); in that case, there is no edge from middle to exit and
4274  // and thus no phis which needed updated.
4276  for (PHINode &LCSSAPhi : LoopExitBlock->phis())
4277  if (llvm::is_contained(LCSSAPhi.incoming_values(), Phi))
4278  LCSSAPhi.addIncoming(ExtractForPhiUsedOutsideLoop, LoopMiddleBlock);
4279 }
4280 
4282  VPTransformState &State) {
4283  PHINode *OrigPhi = cast<PHINode>(PhiR->getUnderlyingValue());
4284  // Get it's reduction variable descriptor.
4285  assert(Legal->isReductionVariable(OrigPhi) &&
4286  "Unable to find the reduction variable");
4287  const RecurrenceDescriptor &RdxDesc = PhiR->getRecurrenceDescriptor();
4288 
4289  RecurKind RK = RdxDesc.getRecurrenceKind();
4290  TrackingVH<Value> ReductionStartValue = RdxDesc.getRecurrenceStartValue();
4291  Instruction *LoopExitInst = RdxDesc.getLoopExitInstr();
4292  setDebugLocFromInst(ReductionStartValue);
4293 
4294  VPValue *LoopExitInstDef = PhiR->getBackedgeValue();
4295  // This is the vector-clone of the value that leaves the loop.
4296  Type *VecTy = State.get(LoopExitInstDef, 0)->getType();
4297 
4298  // Wrap flags are in general invalid after vectorization, clear them.
4299  clearReductionWrapFlags(RdxDesc, State);
4300 
4301  // Before each round, move the insertion point right between
4302  // the PHIs and the values we are going to write.
4303  // This allows us to write both PHINodes and the extractelement
4304  // instructions.
4306 
4307  setDebugLocFromInst(LoopExitInst);
4308 
4309  Type *PhiTy = OrigPhi->getType();
4310  // If tail is folded by masking, the vector value to leave the loop should be
4311  // a Select choosing between the vectorized LoopExitInst and vectorized Phi,
4312  // instead of the former. For an inloop reduction the reduction will already
4313  // be predicated, and does not need to be handled here.
4314  if (Cost->foldTailByMasking() && !PhiR->isInLoop()) {
4315  for (unsigned Part = 0; Part < UF; ++Part) {
4316  Value *VecLoopExitInst = State.get(LoopExitInstDef, Part);
4317  Value *Sel = nullptr;
4318  for (User *U : VecLoopExitInst->users()) {
4319  if (isa<SelectInst>(U)) {
4320  assert(!Sel && "Reduction exit feeding two selects");
4321  Sel = U;
4322  } else
4323  assert(isa<PHINode>(U) && "Reduction exit must feed Phi's or select");
4324  }
4325  assert(Sel && "Reduction exit feeds no select");
4326  State.reset(LoopExitInstDef, Sel, Part);
4327 
4328  // If the target can create a predicated operator for the reduction at no
4329  // extra cost in the loop (for example a predicated vadd), it can be
4330  // cheaper for the select to remain in the loop than be sunk out of it,
4331  // and so use the select value for the phi instead of the old
4332  // LoopExitValue.
4335  RdxDesc.getOpcode(), PhiTy,
4337  auto *VecRdxPhi =
4338  cast<PHINode>(State.get(PhiR->getVPSingleValue(), Part));
4339  VecRdxPhi->setIncomingValueForBlock(
4341  }
4342  }
4343  }
4344 
4345  // If the vector reduction can be performed in a smaller type, we truncate
4346  // then extend the loop exit value to enable InstCombine to evaluate the
4347  // entire expression in the smaller type.
4348  if (VF.isVector() && PhiTy != RdxDesc.getRecurrenceType()) {
4349  assert(!PhiR->isInLoop() && "Unexpected truncated inloop reduction!");
4350  Type *RdxVecTy = VectorType::get(RdxDesc.getRecurrenceType(), VF);
4353  VectorParts RdxParts(UF);
4354  for (unsigned Part = 0; Part < UF; ++Part) {
4355  RdxParts[Part] = State.get(LoopExitInstDef, Part);
4356  Value *Trunc = Builder.CreateTrunc(RdxParts[Part], RdxVecTy);
4357  Value *Extnd = RdxDesc.isSigned() ? Builder.CreateSExt(Trunc, VecTy)
4358  : Builder.CreateZExt(Trunc, VecTy);
4359  for (Value::user_iterator UI = RdxParts[Part]->user_begin();
4360  UI != RdxParts[Part]->user_end();)
4361  if (*UI != Trunc) {
4362  (*UI++)->replaceUsesOfWith(RdxParts[Part], Extnd);
4363  RdxParts[Part] = Extnd;
4364  } else {
4365  ++UI;
4366  }
4367  }
4369  for (unsigned Part = 0; Part < UF; ++Part) {
4370  RdxParts[Part] = Builder.CreateTrunc(RdxParts[Part], RdxVecTy);
4371  State.reset(LoopExitInstDef, RdxParts[Part], Part);
4372  }
4373  }
4374 
4375  // Reduce all of the unrolled parts into a single vector.
4376  Value *ReducedPartRdx = State.get(LoopExitInstDef, 0);
4377  unsigned Op = RecurrenceDescriptor::getOpcode(RK);
4378 
4379  // The middle block terminator has already been assigned a DebugLoc here (the
4380  // OrigLoop's single latch terminator). We want the whole middle block to
4381  // appear to execute on this line because: (a) it is all compiler generated,
4382  // (b) these instructions are always executed after evaluating the latch
4383  // conditional branch, and (c) other passes may add new predecessors which
4384  // terminate on this line. This is the easiest way to ensure we don't
4385  // accidentally cause an extra step back into the loop while debugging.
4387  if (PhiR->isOrdered())
4388  ReducedPartRdx = State.get(LoopExitInstDef, UF - 1);
4389  else {
4390  // Floating-point operations should have some FMF to enable the reduction.
4393  for (unsigned Part = 1; Part < UF; ++Part) {
4394  Value *RdxPart = State.get(LoopExitInstDef, Part);
4395  if (Op != Instruction::ICmp && Op != Instruction::FCmp) {
4396  ReducedPartRdx = Builder.CreateBinOp(
4397  (Instruction::BinaryOps)Op, RdxPart, ReducedPartRdx, "bin.rdx");
4399  ReducedPartRdx = createSelectCmpOp(Builder, ReductionStartValue, RK,
4400  ReducedPartRdx, RdxPart);
4401  else
4402  ReducedPartRdx = createMinMaxOp(Builder, RK, ReducedPartRdx, RdxPart);
4403  }
4404  }
4405 
4406  // Create the reduction after the loop. Note that inloop reductions create the
4407  // target reduction in the loop using a Reduction recipe.
4408  if (VF.isVector() && !PhiR->isInLoop()) {
4409  ReducedPartRdx =
4410  createTargetReduction(Builder, TTI, RdxDesc, ReducedPartRdx, OrigPhi);
4411  // If the reduction can be performed in a smaller type, we need to extend
4412  // the reduction to the wider type before we branch to the original loop.
4413  if (PhiTy != RdxDesc.getRecurrenceType())
4414  ReducedPartRdx = RdxDesc.isSigned()
4415  ? Builder.CreateSExt(ReducedPartRdx, PhiTy)
4416  : Builder.CreateZExt(ReducedPartRdx, PhiTy);
4417  }
4418 
4419  // Create a phi node that merges control-flow from the backedge-taken check
4420  // block and the middle block.
4421  PHINode *BCBlockPhi = PHINode::Create(PhiTy, 2, "bc.merge.rdx",
4423  for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
4424  BCBlockPhi->addIncoming(ReductionStartValue, LoopBypassBlocks[I]);
4425  BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
4426 
4427  // Now, we need to fix the users of the reduction variable
4428  // inside and outside of the scalar remainder loop.
4429 
4430  // We know that the loop is in LCSSA form. We need to update the PHI nodes
4431  // in the exit blocks. See comment on analogous loop in
4432  // fixFirstOrderRecurrence for a more complete explaination of the logic.
4434  for (PHINode &LCSSAPhi : LoopExitBlock->phis())
4435  if (llvm::is_contained(LCSSAPhi.incoming_values(), LoopExitInst))
4436  LCSSAPhi.addIncoming(ReducedPartRdx, LoopMiddleBlock);
4437 
4438  // Fix the scalar loop reduction variable with the incoming reduction sum
4439  // from the vector body and from the backedge value.
4440  int IncomingEdgeBlockIdx =
4442  assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
4443  // Pick the other block.
4444  int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
4445  OrigPhi->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
4446  OrigPhi->setIncomingValue(IncomingEdgeBlockIdx, LoopExitInst);
4447 }
4448 
4450  VPTransformState &State) {
4451  RecurKind RK = RdxDesc.getRecurrenceKind();
4452  if (RK != RecurKind::Add && RK != RecurKind::Mul)
4453  return;
4454 
4455  Instruction *LoopExitInstr = RdxDesc.getLoopExitInstr();
4456  assert(LoopExitInstr && "null loop exit instruction");
4459  Worklist.push_back(LoopExitInstr);
4460  Visited.insert(LoopExitInstr);
4461 
4462  while (!Worklist.empty()) {
4463  Instruction *Cur = Worklist.pop_back_val();
4464  if (isa<OverflowingBinaryOperator>(Cur))
4465  for (unsigned Part = 0; Part < UF; ++Part) {
4466  // FIXME: Should not rely on getVPValue at this point.
4467  Value *V = State.get(State.Plan->getVPValue(Cur, true), Part);
4468  cast<Instruction>(V)->dropPoisonGeneratingFlags();
4469  }
4470 
4471  for (User *U : Cur->users()) {
4472  Instruction *UI = cast<Instruction>(U);
4473  if ((Cur != LoopExitInstr || OrigLoop->contains(UI->getParent())) &&
4474  Visited.insert(UI).second)
4475  Worklist.push_back(UI);
4476  }
4477  }
4478 }
4479 
4481  for (PHINode &LCSSAPhi : LoopExitBlock->phis()) {
4482  if (LCSSAPhi.getBasicBlockIndex(LoopMiddleBlock) != -1)
4483  // Some phis were already hand updated by the reduction and recurrence
4484  // code above, leave them alone.
4485  continue;
4486 
4487  auto *IncomingValue = LCSSAPhi.getIncomingValue(0);
4488  // Non-instruction incoming values will have only one value.
4489 
4490  VPLane Lane = VPLane::getFirstLane();
4491  if (isa<Instruction>(IncomingValue) &&
4492  !Cost->isUniformAfterVectorization(cast<Instruction>(IncomingValue),
4493  VF))
4494  Lane = VPLane::getLastLaneForVF(VF);
4495 
4496  // Can be a loop invariant incoming value or the last scalar value to be
4497  // extracted from the vectorized loop.
4498  // FIXME: Should not rely on getVPValue at this point.
4500  Value *lastIncomingValue =
4501  OrigLoop->isLoopInvariant(IncomingValue)
4502  ? IncomingValue
4503  : State.get(State.Plan->getVPValue(IncomingValue, true),
4504  VPIteration(UF - 1, Lane));
4505  LCSSAPhi.addIncoming(lastIncomingValue, LoopMiddleBlock);
4506  }
4507 }
4508 
4510  // The basic block and loop containing the predicated instruction.
4511  auto *PredBB = PredInst->getParent();
4512  auto *VectorLoop = LI->getLoopFor(PredBB);
4513 
4514  // Initialize a worklist with the operands of the predicated instruction.
4515  SetVector<Value *> Worklist(PredInst->op_begin(), PredInst->op_end());
4516 
4517  // Holds instructions that we need to analyze again. An instruction may be
4518  // reanalyzed if we don't yet know if we can sink it or not.
4519  SmallVector<Instruction *, 8> InstsToReanalyze;
4520 
4521  // Returns true if a given use occurs in the predicated block. Phi nodes use
4522  // their operands in their corresponding predecessor blocks.
4523  auto isBlockOfUsePredicated = [&](Use &U) -> bool {
4524  auto *I = cast<Instruction>(U.getUser());
4525  BasicBlock *BB = I->getParent();
4526  if (auto *Phi = dyn_cast<PHINode>(I))
4527  BB = Phi->getIncomingBlock(
4528  PHINode::getIncomingValueNumForOperand(U.getOperandNo()));
4529  return BB == PredBB;
4530  };
4531 
4532  // Iteratively sink the scalarized operands of the predicated instruction
4533  // into the block we created for it. When an instruction is sunk, it's
4534  // operands are then added to the worklist. The algorithm ends after one pass
4535  // through the worklist doesn't sink a single instruction.
4536  bool Changed;
4537  do {
4538  // Add the instructions that need to be reanalyzed to the worklist, and
4539  // reset the changed indicator.
4540  Worklist.insert(InstsToReanalyze.begin(), InstsToReanalyze.end());
4541  InstsToReanalyze.clear();
4542  Changed = false;
4543 
4544  while (!Worklist.empty()) {
4545  auto *I = dyn_cast<Instruction>(Worklist.pop_back_val());
4546 
4547  // We can't sink an instruction if it is a phi node, is not in the loop,
4548  // or may have side effects.
4549  if (!I || isa<PHINode>(I) || !VectorLoop->contains(I) ||
4550  I->mayHaveSideEffects())
4551  continue;
4552 
4553  // If the instruction is already in PredBB, check if we can sink its
4554  // operands. In that case, VPlan's sinkScalarOperands() succeeded in
4555  // sinking the scalar instruction I, hence it appears in PredBB; but it
4556  // may have failed to sink I's operands (recursively), which we try
4557  // (again) here.
4558  if (I->getParent() == PredBB) {
4559  Worklist.insert(I->op_begin(), I->op_end());
4560  continue;
4561  }
4562 
4563  // It's legal to sink the instruction if all its uses occur in the
4564  // predicated block. Otherwise, there's nothing to do yet, and we may
4565  // need to reanalyze the instruction.
4566  if (!llvm::all_of(I->uses(), isBlockOfUsePredicated)) {
4567  InstsToReanalyze.push_back(I);
4568  continue;
4569  }
4570 
4571  // Move the instruction to the beginning of the predicated block, and add
4572  // it's operands to the worklist.
4573  I->moveBefore(&*PredBB->getFirstInsertionPt());
4574  Wo