LLVM  6.0.0svn
LoopVectorize.cpp
Go to the documentation of this file.
1 //===- LoopVectorize.cpp - A Loop Vectorizer ------------------------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This is the LLVM loop vectorizer. This pass modifies 'vectorizable' loops
11 // and generates target-independent LLVM-IR.
12 // The vectorizer uses the TargetTransformInfo analysis to estimate the costs
13 // of instructions in order to estimate the profitability of vectorization.
14 //
15 // The loop vectorizer combines consecutive loop iterations into a single
16 // 'wide' iteration. After this transformation the index is incremented
17 // by the SIMD vector width, and not by one.
18 //
19 // This pass has three parts:
20 // 1. The main loop pass that drives the different parts.
21 // 2. LoopVectorizationLegality - A unit that checks for the legality
22 // of the vectorization.
23 // 3. InnerLoopVectorizer - A unit that performs the actual
24 // widening of instructions.
25 // 4. LoopVectorizationCostModel - A unit that checks for the profitability
26 // of vectorization. It decides on the optimal vector width, which
27 // can be one, if vectorization is not profitable.
28 //
29 //===----------------------------------------------------------------------===//
30 //
31 // The reduction-variable vectorization is based on the paper:
32 // D. Nuzman and R. Henderson. Multi-platform Auto-vectorization.
33 //
34 // Variable uniformity checks are inspired by:
35 // Karrenberg, R. and Hack, S. Whole Function Vectorization.
36 //
37 // The interleaved access vectorization is based on the paper:
38 // Dorit Nuzman, Ira Rosen and Ayal Zaks. Auto-Vectorization of Interleaved
39 // Data for SIMD
40 //
41 // Other ideas/concepts are from:
42 // A. Zaks and D. Nuzman. Autovectorization in GCC-two years later.
43 //
44 // S. Maleki, Y. Gao, M. Garzaran, T. Wong and D. Padua. An Evaluation of
45 // Vectorizing Compilers.
46 //
47 //===----------------------------------------------------------------------===//
48 
50 #include "VPlan.h"
51 #include "VPlanBuilder.h"
52 #include "llvm/ADT/APInt.h"
53 #include "llvm/ADT/ArrayRef.h"
54 #include "llvm/ADT/DenseMap.h"
55 #include "llvm/ADT/DenseMapInfo.h"
56 #include "llvm/ADT/Hashing.h"
57 #include "llvm/ADT/MapVector.h"
58 #include "llvm/ADT/None.h"
59 #include "llvm/ADT/Optional.h"
60 #include "llvm/ADT/SCCIterator.h"
61 #include "llvm/ADT/STLExtras.h"
62 #include "llvm/ADT/SetVector.h"
63 #include "llvm/ADT/SmallPtrSet.h"
64 #include "llvm/ADT/SmallSet.h"
65 #include "llvm/ADT/SmallVector.h"
66 #include "llvm/ADT/Statistic.h"
67 #include "llvm/ADT/StringRef.h"
68 #include "llvm/ADT/Twine.h"
78 #include "llvm/Analysis/LoopInfo.h"
87 #include "llvm/IR/Attributes.h"
88 #include "llvm/IR/BasicBlock.h"
89 #include "llvm/IR/CFG.h"
90 #include "llvm/IR/Constant.h"
91 #include "llvm/IR/Constants.h"
92 #include "llvm/IR/DataLayout.h"
94 #include "llvm/IR/DebugLoc.h"
95 #include "llvm/IR/DerivedTypes.h"
96 #include "llvm/IR/DiagnosticInfo.h"
97 #include "llvm/IR/Dominators.h"
98 #include "llvm/IR/Function.h"
99 #include "llvm/IR/IRBuilder.h"
100 #include "llvm/IR/InstrTypes.h"
101 #include "llvm/IR/Instruction.h"
102 #include "llvm/IR/Instructions.h"
103 #include "llvm/IR/IntrinsicInst.h"
104 #include "llvm/IR/Intrinsics.h"
105 #include "llvm/IR/LLVMContext.h"
106 #include "llvm/IR/Metadata.h"
107 #include "llvm/IR/Module.h"
108 #include "llvm/IR/Operator.h"
109 #include "llvm/IR/Type.h"
110 #include "llvm/IR/Use.h"
111 #include "llvm/IR/User.h"
112 #include "llvm/IR/Value.h"
113 #include "llvm/IR/ValueHandle.h"
114 #include "llvm/IR/Verifier.h"
115 #include "llvm/Pass.h"
116 #include "llvm/Support/Casting.h"
118 #include "llvm/Support/Compiler.h"
119 #include "llvm/Support/Debug.h"
121 #include "llvm/Support/MathExtras.h"
127 #include <algorithm>
128 #include <cassert>
129 #include <cstdint>
130 #include <cstdlib>
131 #include <functional>
132 #include <iterator>
133 #include <limits>
134 #include <memory>
135 #include <string>
136 #include <tuple>
137 #include <utility>
138 #include <vector>
139 
140 using namespace llvm;
141 
142 #define LV_NAME "loop-vectorize"
143 #define DEBUG_TYPE LV_NAME
144 
145 STATISTIC(LoopsVectorized, "Number of loops vectorized");
146 STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization");
147 
148 static cl::opt<bool>
149  EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
150  cl::desc("Enable if-conversion during vectorization."));
151 
152 /// Loops with a known constant trip count below this number are vectorized only
153 /// if no scalar iteration overheads are incurred.
155  "vectorizer-min-trip-count", cl::init(16), cl::Hidden,
156  cl::desc("Loops with a constant trip count that is smaller than this "
157  "value are vectorized only if no scalar iteration overheads "
158  "are incurred."));
159 
161  "vectorizer-maximize-bandwidth", cl::init(false), cl::Hidden,
162  cl::desc("Maximize bandwidth when selecting vectorization factor which "
163  "will be determined by the smallest type in loop."));
164 
166  "enable-interleaved-mem-accesses", cl::init(false), cl::Hidden,
167  cl::desc("Enable vectorization on interleaved memory accesses in a loop"));
168 
169 /// Maximum factor for an interleaved memory access.
171  "max-interleave-group-factor", cl::Hidden,
172  cl::desc("Maximum factor for an interleaved access group (default = 8)"),
173  cl::init(8));
174 
175 /// We don't interleave loops with a known constant trip count below this
176 /// number.
177 static const unsigned TinyTripCountInterleaveThreshold = 128;
178 
180  "force-target-num-scalar-regs", cl::init(0), cl::Hidden,
181  cl::desc("A flag that overrides the target's number of scalar registers."));
182 
184  "force-target-num-vector-regs", cl::init(0), cl::Hidden,
185  cl::desc("A flag that overrides the target's number of vector registers."));
186 
187 /// Maximum vectorization interleave count.
188 static const unsigned MaxInterleaveFactor = 16;
189 
191  "force-target-max-scalar-interleave", cl::init(0), cl::Hidden,
192  cl::desc("A flag that overrides the target's max interleave factor for "
193  "scalar loops."));
194 
196  "force-target-max-vector-interleave", cl::init(0), cl::Hidden,
197  cl::desc("A flag that overrides the target's max interleave factor for "
198  "vectorized loops."));
199 
201  "force-target-instruction-cost", cl::init(0), cl::Hidden,
202  cl::desc("A flag that overrides the target's expected cost for "
203  "an instruction to a single constant value. Mostly "
204  "useful for getting consistent testing."));
205 
207  "small-loop-cost", cl::init(20), cl::Hidden,
208  cl::desc(
209  "The cost of a loop that is considered 'small' by the interleaver."));
210 
212  "loop-vectorize-with-block-frequency", cl::init(false), cl::Hidden,
213  cl::desc("Enable the use of the block frequency analysis to access PGO "
214  "heuristics minimizing code growth in cold regions and being more "
215  "aggressive in hot regions."));
216 
217 // Runtime interleave loops for load/store throughput.
219  "enable-loadstore-runtime-interleave", cl::init(true), cl::Hidden,
220  cl::desc(
221  "Enable runtime interleaving until load/store ports are saturated"));
222 
223 /// The number of stores in a loop that are allowed to need predication.
225  "vectorize-num-stores-pred", cl::init(1), cl::Hidden,
226  cl::desc("Max number of stores to be predicated behind an if."));
227 
229  "enable-ind-var-reg-heur", cl::init(true), cl::Hidden,
230  cl::desc("Count the induction variable only once when interleaving"));
231 
233  "enable-cond-stores-vec", cl::init(true), cl::Hidden,
234  cl::desc("Enable if predication of stores during vectorization."));
235 
237  "max-nested-scalar-reduction-interleave", cl::init(2), cl::Hidden,
238  cl::desc("The maximum interleave count to use when interleaving a scalar "
239  "reduction in a nested loop."));
240 
242  "pragma-vectorize-memory-check-threshold", cl::init(128), cl::Hidden,
243  cl::desc("The maximum allowed number of runtime memory checks with a "
244  "vectorize(enable) pragma."));
245 
247  "vectorize-scev-check-threshold", cl::init(16), cl::Hidden,
248  cl::desc("The maximum number of SCEV checks allowed."));
249 
251  "pragma-vectorize-scev-check-threshold", cl::init(128), cl::Hidden,
252  cl::desc("The maximum number of SCEV checks allowed with a "
253  "vectorize(enable) pragma"));
254 
255 /// Create an analysis remark that explains why vectorization failed
256 ///
257 /// \p PassName is the name of the pass (e.g. can be AlwaysPrint). \p
258 /// RemarkName is the identifier for the remark. If \p I is passed it is an
259 /// instruction that prevents vectorization. Otherwise \p TheLoop is used for
260 /// the location of the remark. \return the remark object that can be
261 /// streamed to.
263 createMissedAnalysis(const char *PassName, StringRef RemarkName, Loop *TheLoop,
264  Instruction *I = nullptr) {
265  Value *CodeRegion = TheLoop->getHeader();
266  DebugLoc DL = TheLoop->getStartLoc();
267 
268  if (I) {
269  CodeRegion = I->getParent();
270  // If there is no debug location attached to the instruction, revert back to
271  // using the loop's.
272  if (I->getDebugLoc())
273  DL = I->getDebugLoc();
274  }
275 
276  OptimizationRemarkAnalysis R(PassName, RemarkName, DL, CodeRegion);
277  R << "loop not vectorized: ";
278  return R;
279 }
280 
281 namespace {
282 
283 class LoopVectorizationLegality;
284 class LoopVectorizationCostModel;
285 class LoopVectorizationRequirements;
286 
287 } // end anonymous namespace
288 
289 /// Returns true if the given loop body has a cycle, excluding the loop
290 /// itself.
291 static bool hasCyclesInLoopBody(const Loop &L) {
292  if (!L.empty())
293  return true;
294 
295  for (const auto &SCC :
298  if (SCC.size() > 1) {
299  DEBUG(dbgs() << "LVL: Detected a cycle in the loop body:\n");
300  DEBUG(L.dump());
301  return true;
302  }
303  }
304  return false;
305 }
306 
307 /// A helper function for converting Scalar types to vector types.
308 /// If the incoming type is void, we return void. If the VF is 1, we return
309 /// the scalar type.
310 static Type *ToVectorTy(Type *Scalar, unsigned VF) {
311  if (Scalar->isVoidTy() || VF == 1)
312  return Scalar;
313  return VectorType::get(Scalar, VF);
314 }
315 
316 // FIXME: The following helper functions have multiple implementations
317 // in the project. They can be effectively organized in a common Load/Store
318 // utilities unit.
319 
320 /// A helper function that returns the pointer operand of a load or store
321 /// instruction.
323  if (auto *LI = dyn_cast<LoadInst>(I))
324  return LI->getPointerOperand();
325  if (auto *SI = dyn_cast<StoreInst>(I))
326  return SI->getPointerOperand();
327  return nullptr;
328 }
329 
330 /// A helper function that returns the type of loaded or stored value.
332  assert((isa<LoadInst>(I) || isa<StoreInst>(I)) &&
333  "Expected Load or Store instruction");
334  if (auto *LI = dyn_cast<LoadInst>(I))
335  return LI->getType();
336  return cast<StoreInst>(I)->getValueOperand()->getType();
337 }
338 
339 /// A helper function that returns the alignment of load or store instruction.
340 static unsigned getMemInstAlignment(Value *I) {
341  assert((isa<LoadInst>(I) || isa<StoreInst>(I)) &&
342  "Expected Load or Store instruction");
343  if (auto *LI = dyn_cast<LoadInst>(I))
344  return LI->getAlignment();
345  return cast<StoreInst>(I)->getAlignment();
346 }
347 
348 /// A helper function that returns the address space of the pointer operand of
349 /// load or store instruction.
350 static unsigned getMemInstAddressSpace(Value *I) {
351  assert((isa<LoadInst>(I) || isa<StoreInst>(I)) &&
352  "Expected Load or Store instruction");
353  if (auto *LI = dyn_cast<LoadInst>(I))
354  return LI->getPointerAddressSpace();
355  return cast<StoreInst>(I)->getPointerAddressSpace();
356 }
357 
358 /// A helper function that returns true if the given type is irregular. The
359 /// type is irregular if its allocated size doesn't equal the store size of an
360 /// element of the corresponding vector type at the given vectorization factor.
361 static bool hasIrregularType(Type *Ty, const DataLayout &DL, unsigned VF) {
362  // Determine if an array of VF elements of type Ty is "bitcast compatible"
363  // with a <VF x Ty> vector.
364  if (VF > 1) {
365  auto *VectorTy = VectorType::get(Ty, VF);
366  return VF * DL.getTypeAllocSize(Ty) != DL.getTypeStoreSize(VectorTy);
367  }
368 
369  // If the vectorization factor is one, we just check if an array of type Ty
370  // requires padding between elements.
371  return DL.getTypeAllocSizeInBits(Ty) != DL.getTypeSizeInBits(Ty);
372 }
373 
374 /// A helper function that returns the reciprocal of the block probability of
375 /// predicated blocks. If we return X, we are assuming the predicated block
376 /// will execute once for for every X iterations of the loop header.
377 ///
378 /// TODO: We should use actual block probability here, if available. Currently,
379 /// we always assume predicated blocks have a 50% chance of executing.
380 static unsigned getReciprocalPredBlockProb() { return 2; }
381 
382 /// A helper function that adds a 'fast' flag to floating-point operations.
384  if (isa<FPMathOperator>(V)) {
385  FastMathFlags Flags;
386  Flags.setFast();
387  cast<Instruction>(V)->setFastMathFlags(Flags);
388  }
389  return V;
390 }
391 
392 /// A helper function that returns an integer or floating-point constant with
393 /// value C.
394 static Constant *getSignedIntOrFpConstant(Type *Ty, int64_t C) {
395  return Ty->isIntegerTy() ? ConstantInt::getSigned(Ty, C)
396  : ConstantFP::get(Ty, C);
397 }
398 
399 namespace llvm {
400 
401 /// InnerLoopVectorizer vectorizes loops which contain only one basic
402 /// block to a specified vectorization factor (VF).
403 /// This class performs the widening of scalars into vectors, or multiple
404 /// scalars. This class also implements the following features:
405 /// * It inserts an epilogue loop for handling loops that don't have iteration
406 /// counts that are known to be a multiple of the vectorization factor.
407 /// * It handles the code generation for reduction variables.
408 /// * Scalarization (implementation using scalars) of un-vectorizable
409 /// instructions.
410 /// InnerLoopVectorizer does not perform any vectorization-legality
411 /// checks, and relies on the caller to check for the different legality
412 /// aspects. The InnerLoopVectorizer relies on the
413 /// LoopVectorizationLegality class to provide information about the induction
414 /// and reduction variables that were found to a given vectorization factor.
416 public:
419  const TargetLibraryInfo *TLI,
421  OptimizationRemarkEmitter *ORE, unsigned VecWidth,
422  unsigned UnrollFactor, LoopVectorizationLegality *LVL,
423  LoopVectorizationCostModel *CM)
424  : OrigLoop(OrigLoop), PSE(PSE), LI(LI), DT(DT), TLI(TLI), TTI(TTI),
425  AC(AC), ORE(ORE), VF(VecWidth), UF(UnrollFactor),
426  Builder(PSE.getSE()->getContext()),
427  VectorLoopValueMap(UnrollFactor, VecWidth), Legal(LVL), Cost(CM) {}
428  virtual ~InnerLoopVectorizer() = default;
429 
430  /// Create a new empty loop. Unlink the old loop and connect the new one.
431  /// Return the pre-header block of the new loop.
433 
434  /// Widen a single instruction within the innermost loop.
436 
437  /// Fix the vectorized code, taking care of header phi's, live-outs, and more.
438  void fixVectorizedLoop();
439 
440  // Return true if any runtime check is added.
442 
443  /// A type for vectorized values in the new loop. Each value from the
444  /// original loop, when vectorized, is represented by UF vector values in the
445  /// new unrolled loop, where UF is the unroll factor.
447 
448  /// Vectorize a single PHINode in a block. This method handles the induction
449  /// variable canonicalization. It supports both VF = 1 for unrolled loops and
450  /// arbitrary length vectors.
451  void widenPHIInstruction(Instruction *PN, unsigned UF, unsigned VF);
452 
453  /// A helper function to scalarize a single Instruction in the innermost loop.
454  /// Generates a sequence of scalar instances for each lane between \p MinLane
455  /// and \p MaxLane, times each part between \p MinPart and \p MaxPart,
456  /// inclusive..
457  void scalarizeInstruction(Instruction *Instr, const VPIteration &Instance,
458  bool IfPredicateInstr);
459 
460  /// Widen an integer or floating-point induction variable \p IV. If \p Trunc
461  /// is provided, the integer induction variable will first be truncated to
462  /// the corresponding type.
463  void widenIntOrFpInduction(PHINode *IV, TruncInst *Trunc = nullptr);
464 
465  /// getOrCreateVectorValue and getOrCreateScalarValue coordinate to generate a
466  /// vector or scalar value on-demand if one is not yet available. When
467  /// vectorizing a loop, we visit the definition of an instruction before its
468  /// uses. When visiting the definition, we either vectorize or scalarize the
469  /// instruction, creating an entry for it in the corresponding map. (In some
470  /// cases, such as induction variables, we will create both vector and scalar
471  /// entries.) Then, as we encounter uses of the definition, we derive values
472  /// for each scalar or vector use unless such a value is already available.
473  /// For example, if we scalarize a definition and one of its uses is vector,
474  /// we build the required vector on-demand with an insertelement sequence
475  /// when visiting the use. Otherwise, if the use is scalar, we can use the
476  /// existing scalar definition.
477  ///
478  /// Return a value in the new loop corresponding to \p V from the original
479  /// loop at unroll index \p Part. If the value has already been vectorized,
480  /// the corresponding vector entry in VectorLoopValueMap is returned. If,
481  /// however, the value has a scalar entry in VectorLoopValueMap, we construct
482  /// a new vector value on-demand by inserting the scalar values into a vector
483  /// with an insertelement sequence. If the value has been neither vectorized
484  /// nor scalarized, it must be loop invariant, so we simply broadcast the
485  /// value into a vector.
486  Value *getOrCreateVectorValue(Value *V, unsigned Part);
487 
488  /// Return a value in the new loop corresponding to \p V from the original
489  /// loop at unroll and vector indices \p Instance. If the value has been
490  /// vectorized but not scalarized, the necessary extractelement instruction
491  /// will be generated.
492  Value *getOrCreateScalarValue(Value *V, const VPIteration &Instance);
493 
494  /// Construct the vector value of a scalarized value \p V one lane at a time.
495  void packScalarIntoVectorValue(Value *V, const VPIteration &Instance);
496 
497  /// Try to vectorize the interleaved access group that \p Instr belongs to.
499 
500  /// Vectorize Load and Store instructions, optionally masking the vector
501  /// operations if \p BlockInMask is non-null.
503  VectorParts *BlockInMask = nullptr);
504 
505  /// \brief Set the debug location in the builder using the debug location in
506  /// the instruction.
507  void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr);
508 
509 protected:
511 
512  /// A small list of PHINodes.
514 
515  /// A type for scalarized values in the new loop. Each value from the
516  /// original loop, when scalarized, is represented by UF x VF scalar values
517  /// in the new unrolled loop, where UF is the unroll factor and VF is the
518  /// vectorization factor.
520 
521  /// Set up the values of the IVs correctly when exiting the vector loop.
522  void fixupIVUsers(PHINode *OrigPhi, const InductionDescriptor &II,
523  Value *CountRoundDown, Value *EndValue,
524  BasicBlock *MiddleBlock);
525 
526  /// Create a new induction variable inside L.
528  Value *Step, Instruction *DL);
529 
530  /// Handle all cross-iteration phis in the header.
531  void fixCrossIterationPHIs();
532 
533  /// Fix a first-order recurrence. This is the second phase of vectorizing
534  /// this phi node.
535  void fixFirstOrderRecurrence(PHINode *Phi);
536 
537  /// Fix a reduction cross-iteration phi. This is the second phase of
538  /// vectorizing this phi node.
539  void fixReduction(PHINode *Phi);
540 
541  /// \brief The Loop exit block may have single value PHI nodes with some
542  /// incoming value. While vectorizing we only handled real values
543  /// that were defined inside the loop and we should have one value for
544  /// each predecessor of its parent basic block. See PR14725.
545  void fixLCSSAPHIs();
546 
547  /// Iteratively sink the scalarized operands of a predicated instruction into
548  /// the block that was created for it.
549  void sinkScalarOperands(Instruction *PredInst);
550 
551  /// Shrinks vector element sizes to the smallest bitwidth they can be legally
552  /// represented as.
554 
555  /// Insert the new loop to the loop hierarchy and pass manager
556  /// and update the analysis passes.
557  void updateAnalysis();
558 
559  /// Create a broadcast instruction. This method generates a broadcast
560  /// instruction (shuffle) for loop invariant values and for the induction
561  /// value. If this is the induction variable then we extend it to N, N+1, ...
562  /// this is needed because each iteration in the loop corresponds to a SIMD
563  /// element.
564  virtual Value *getBroadcastInstrs(Value *V);
565 
566  /// This function adds (StartIdx, StartIdx + Step, StartIdx + 2*Step, ...)
567  /// to each vector element of Val. The sequence starts at StartIndex.
568  /// \p Opcode is relevant for FP induction variable.
569  virtual Value *getStepVector(Value *Val, int StartIdx, Value *Step,
570  Instruction::BinaryOps Opcode =
571  Instruction::BinaryOpsEnd);
572 
573  /// Compute scalar induction steps. \p ScalarIV is the scalar induction
574  /// variable on which to base the steps, \p Step is the size of the step, and
575  /// \p EntryVal is the value from the original loop that maps to the steps.
576  /// Note that \p EntryVal doesn't have to be an induction variable (e.g., it
577  /// can be a truncate instruction).
578  void buildScalarSteps(Value *ScalarIV, Value *Step, Value *EntryVal,
579  const InductionDescriptor &ID);
580 
581  /// Create a vector induction phi node based on an existing scalar one. \p
582  /// EntryVal is the value from the original loop that maps to the vector phi
583  /// node, and \p Step is the loop-invariant step. If \p EntryVal is a
584  /// truncate instruction, instead of widening the original IV, we widen a
585  /// version of the IV truncated to \p EntryVal's type.
587  Value *Step, Instruction *EntryVal);
588 
589  /// Returns true if an instruction \p I should be scalarized instead of
590  /// vectorized for the chosen vectorization factor.
592 
593  /// Returns true if we should generate a scalar version of \p IV.
594  bool needsScalarInduction(Instruction *IV) const;
595 
596  /// If there is a cast involved in the induction variable \p ID, which should
597  /// be ignored in the vectorized loop body, this function records the
598  /// VectorLoopValue of the respective Phi also as the VectorLoopValue of the
599  /// cast. We had already proved that the casted Phi is equal to the uncasted
600  /// Phi in the vectorized loop (under a runtime guard), and therefore
601  /// there is no need to vectorize the cast - the same value can be used in the
602  /// vector loop for both the Phi and the cast.
603  /// If \p VectorLoopValue is a scalarized value, \p Lane is also specified,
604  /// Otherwise, \p VectorLoopValue is a widened/vectorized value.
606  Value *VectorLoopValue,
607  unsigned Part,
608  unsigned Lane = UINT_MAX);
609 
610  /// Generate a shuffle sequence that will reverse the vector Vec.
611  virtual Value *reverseVector(Value *Vec);
612 
613  /// Returns (and creates if needed) the original loop trip count.
614  Value *getOrCreateTripCount(Loop *NewLoop);
615 
616  /// Returns (and creates if needed) the trip count of the widened loop.
618 
619  /// Returns a bitcasted value to the requested vector type.
620  /// Also handles bitcasts of vector<float> <-> vector<pointer> types.
622  const DataLayout &DL);
623 
624  /// Emit a bypass check to see if the vector trip count is zero, including if
625  /// it overflows.
627 
628  /// Emit a bypass check to see if all of the SCEV assumptions we've
629  /// had to make are correct.
630  void emitSCEVChecks(Loop *L, BasicBlock *Bypass);
631 
632  /// Emit bypass checks to check any memory assumptions we may have made.
633  void emitMemRuntimeChecks(Loop *L, BasicBlock *Bypass);
634 
635  /// Add additional metadata to \p To that was not present on \p Orig.
636  ///
637  /// Currently this is used to add the noalias annotations based on the
638  /// inserted memchecks. Use this for instructions that are *cloned* into the
639  /// vector loop.
640  void addNewMetadata(Instruction *To, const Instruction *Orig);
641 
642  /// Add metadata from one instruction to another.
643  ///
644  /// This includes both the original MDs from \p From and additional ones (\see
645  /// addNewMetadata). Use this for *newly created* instructions in the vector
646  /// loop.
647  void addMetadata(Instruction *To, Instruction *From);
648 
649  /// \brief Similar to the previous function but it adds the metadata to a
650  /// vector of instructions.
651  void addMetadata(ArrayRef<Value *> To, Instruction *From);
652 
653  /// The original loop.
655 
656  /// A wrapper around ScalarEvolution used to add runtime SCEV checks. Applies
657  /// dynamic knowledge to simplify SCEV expressions and converts them to a
658  /// more usable form.
660 
661  /// Loop Info.
663 
664  /// Dominator Tree.
666 
667  /// Alias Analysis.
669 
670  /// Target Library Info.
672 
673  /// Target Transform Info.
675 
676  /// Assumption Cache.
678 
679  /// Interface to emit optimization remarks.
681 
682  /// \brief LoopVersioning. It's only set up (non-null) if memchecks were
683  /// used.
684  ///
685  /// This is currently only used to add no-alias metadata based on the
686  /// memchecks. The actually versioning is performed manually.
687  std::unique_ptr<LoopVersioning> LVer;
688 
689  /// The vectorization SIMD factor to use. Each vector will have this many
690  /// vector elements.
691  unsigned VF;
692 
693  /// The vectorization unroll factor to use. Each scalar is vectorized to this
694  /// many different vector instructions.
695  unsigned UF;
696 
697  /// The builder that we use
699 
700  // --- Vectorization state ---
701 
702  /// The vector-loop preheader.
704 
705  /// The scalar-loop preheader.
707 
708  /// Middle Block between the vector and the scalar.
710 
711  /// The ExitBlock of the scalar loop.
713 
714  /// The vector loop body.
716 
717  /// The scalar loop body.
719 
720  /// A list of all bypass blocks. The first block is the entry of the loop.
722 
723  /// The new Induction variable which was added to the new block.
724  PHINode *Induction = nullptr;
725 
726  /// The induction variable of the old basic block.
727  PHINode *OldInduction = nullptr;
728 
729  /// Maps values from the original loop to their corresponding values in the
730  /// vectorized loop. A key value can map to either vector values, scalar
731  /// values or both kinds of values, depending on whether the key was
732  /// vectorized and scalarized.
734 
735  /// Store instructions that were predicated.
737 
738  /// Trip count of the original loop.
739  Value *TripCount = nullptr;
740 
741  /// Trip count of the widened loop (TripCount - TripCount % (VF*UF))
742  Value *VectorTripCount = nullptr;
743 
744  /// The legality analysis.
745  LoopVectorizationLegality *Legal;
746 
747  /// The profitablity analysis.
748  LoopVectorizationCostModel *Cost;
749 
750  // Record whether runtime checks are added.
751  bool AddedSafetyChecks = false;
752 
753  // Holds the end values for each induction variable. We save the end values
754  // so we can later fix-up the external users of the induction variables.
756 };
757 
759 public:
762  const TargetLibraryInfo *TLI,
764  OptimizationRemarkEmitter *ORE, unsigned UnrollFactor,
765  LoopVectorizationLegality *LVL,
766  LoopVectorizationCostModel *CM)
767  : InnerLoopVectorizer(OrigLoop, PSE, LI, DT, TLI, TTI, AC, ORE, 1,
768  UnrollFactor, LVL, CM) {}
769 
770 private:
771  Value *getBroadcastInstrs(Value *V) override;
772  Value *getStepVector(Value *Val, int StartIdx, Value *Step,
773  Instruction::BinaryOps Opcode =
774  Instruction::BinaryOpsEnd) override;
775  Value *reverseVector(Value *Vec) override;
776 };
777 
778 } // end namespace llvm
779 
780 /// \brief Look for a meaningful debug location on the instruction or it's
781 /// operands.
783  if (!I)
784  return I;
785 
786  DebugLoc Empty;
787  if (I->getDebugLoc() != Empty)
788  return I;
789 
790  for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
791  if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
792  if (OpInst->getDebugLoc() != Empty)
793  return OpInst;
794  }
795 
796  return I;
797 }
798 
800  if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr)) {
801  const DILocation *DIL = Inst->getDebugLoc();
802  if (DIL && Inst->getFunction()->isDebugInfoForProfiling() &&
803  !isa<DbgInfoIntrinsic>(Inst))
805  else
807  } else
809 }
810 
811 #ifndef NDEBUG
812 /// \return string containing a file name and a line # for the given loop.
813 static std::string getDebugLocString(const Loop *L) {
814  std::string Result;
815  if (L) {
816  raw_string_ostream OS(Result);
817  if (const DebugLoc LoopDbgLoc = L->getStartLoc())
818  LoopDbgLoc.print(OS);
819  else
820  // Just print the module name.
821  OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier();
822  OS.flush();
823  }
824  return Result;
825 }
826 #endif
827 
829  const Instruction *Orig) {
830  // If the loop was versioned with memchecks, add the corresponding no-alias
831  // metadata.
832  if (LVer && (isa<LoadInst>(Orig) || isa<StoreInst>(Orig)))
833  LVer->annotateInstWithNoAlias(To, Orig);
834 }
835 
837  Instruction *From) {
838  propagateMetadata(To, From);
839  addNewMetadata(To, From);
840 }
841 
843  Instruction *From) {
844  for (Value *V : To) {
845  if (Instruction *I = dyn_cast<Instruction>(V))
846  addMetadata(I, From);
847  }
848 }
849 
850 namespace llvm {
851 
852 /// \brief The group of interleaved loads/stores sharing the same stride and
853 /// close to each other.
854 ///
855 /// Each member in this group has an index starting from 0, and the largest
856 /// index should be less than interleaved factor, which is equal to the absolute
857 /// value of the access's stride.
858 ///
859 /// E.g. An interleaved load group of factor 4:
860 /// for (unsigned i = 0; i < 1024; i+=4) {
861 /// a = A[i]; // Member of index 0
862 /// b = A[i+1]; // Member of index 1
863 /// d = A[i+3]; // Member of index 3
864 /// ...
865 /// }
866 ///
867 /// An interleaved store group of factor 4:
868 /// for (unsigned i = 0; i < 1024; i+=4) {
869 /// ...
870 /// A[i] = a; // Member of index 0
871 /// A[i+1] = b; // Member of index 1
872 /// A[i+2] = c; // Member of index 2
873 /// A[i+3] = d; // Member of index 3
874 /// }
875 ///
876 /// Note: the interleaved load group could have gaps (missing members), but
877 /// the interleaved store group doesn't allow gaps.
879 public:
880  InterleaveGroup(Instruction *Instr, int Stride, unsigned Align)
881  : Align(Align), InsertPos(Instr) {
882  assert(Align && "The alignment should be non-zero");
883 
884  Factor = std::abs(Stride);
885  assert(Factor > 1 && "Invalid interleave factor");
886 
887  Reverse = Stride < 0;
888  Members[0] = Instr;
889  }
890 
891  bool isReverse() const { return Reverse; }
892  unsigned getFactor() const { return Factor; }
893  unsigned getAlignment() const { return Align; }
894  unsigned getNumMembers() const { return Members.size(); }
895 
896  /// \brief Try to insert a new member \p Instr with index \p Index and
897  /// alignment \p NewAlign. The index is related to the leader and it could be
898  /// negative if it is the new leader.
899  ///
900  /// \returns false if the instruction doesn't belong to the group.
901  bool insertMember(Instruction *Instr, int Index, unsigned NewAlign) {
902  assert(NewAlign && "The new member's alignment should be non-zero");
903 
904  int Key = Index + SmallestKey;
905 
906  // Skip if there is already a member with the same index.
907  if (Members.count(Key))
908  return false;
909 
910  if (Key > LargestKey) {
911  // The largest index is always less than the interleave factor.
912  if (Index >= static_cast<int>(Factor))
913  return false;
914 
915  LargestKey = Key;
916  } else if (Key < SmallestKey) {
917  // The largest index is always less than the interleave factor.
918  if (LargestKey - Key >= static_cast<int>(Factor))
919  return false;
920 
921  SmallestKey = Key;
922  }
923 
924  // It's always safe to select the minimum alignment.
925  Align = std::min(Align, NewAlign);
926  Members[Key] = Instr;
927  return true;
928  }
929 
930  /// \brief Get the member with the given index \p Index
931  ///
932  /// \returns nullptr if contains no such member.
933  Instruction *getMember(unsigned Index) const {
934  int Key = SmallestKey + Index;
935  if (!Members.count(Key))
936  return nullptr;
937 
938  return Members.find(Key)->second;
939  }
940 
941  /// \brief Get the index for the given member. Unlike the key in the member
942  /// map, the index starts from 0.
943  unsigned getIndex(Instruction *Instr) const {
944  for (auto I : Members)
945  if (I.second == Instr)
946  return I.first - SmallestKey;
947 
948  llvm_unreachable("InterleaveGroup contains no such member");
949  }
950 
951  Instruction *getInsertPos() const { return InsertPos; }
952  void setInsertPos(Instruction *Inst) { InsertPos = Inst; }
953 
954  /// Add metadata (e.g. alias info) from the instructions in this group to \p
955  /// NewInst.
956  ///
957  /// FIXME: this function currently does not add noalias metadata a'la
958  /// addNewMedata. To do that we need to compute the intersection of the
959  /// noalias info from all members.
960  void addMetadata(Instruction *NewInst) const {
962  std::transform(Members.begin(), Members.end(), std::back_inserter(VL),
963  [](std::pair<int, Instruction *> p) { return p.second; });
964  propagateMetadata(NewInst, VL);
965  }
966 
967 private:
968  unsigned Factor; // Interleave Factor.
969  bool Reverse;
970  unsigned Align;
972  int SmallestKey = 0;
973  int LargestKey = 0;
974 
975  // To avoid breaking dependences, vectorized instructions of an interleave
976  // group should be inserted at either the first load or the last store in
977  // program order.
978  //
979  // E.g. %even = load i32 // Insert Position
980  // %add = add i32 %even // Use of %even
981  // %odd = load i32
982  //
983  // store i32 %even
984  // %odd = add i32 // Def of %odd
985  // store i32 %odd // Insert Position
986  Instruction *InsertPos;
987 };
988 } // end namespace llvm
989 
990 namespace {
991 
992 /// \brief Drive the analysis of interleaved memory accesses in the loop.
993 ///
994 /// Use this class to analyze interleaved accesses only when we can vectorize
995 /// a loop. Otherwise it's meaningless to do analysis as the vectorization
996 /// on interleaved accesses is unsafe.
997 ///
998 /// The analysis collects interleave groups and records the relationships
999 /// between the member and the group in a map.
1000 class InterleavedAccessInfo {
1001 public:
1002  InterleavedAccessInfo(PredicatedScalarEvolution &PSE, Loop *L,
1004  : PSE(PSE), TheLoop(L), DT(DT), LI(LI) {}
1005 
1006  ~InterleavedAccessInfo() {
1008  // Avoid releasing a pointer twice.
1009  for (auto &I : InterleaveGroupMap)
1010  DelSet.insert(I.second);
1011  for (auto *Ptr : DelSet)
1012  delete Ptr;
1013  }
1014 
1015  /// \brief Analyze the interleaved accesses and collect them in interleave
1016  /// groups. Substitute symbolic strides using \p Strides.
1017  void analyzeInterleaving(const ValueToValueMap &Strides);
1018 
1019  /// \brief Check if \p Instr belongs to any interleave group.
1020  bool isInterleaved(Instruction *Instr) const {
1021  return InterleaveGroupMap.count(Instr);
1022  }
1023 
1024  /// \brief Get the interleave group that \p Instr belongs to.
1025  ///
1026  /// \returns nullptr if doesn't have such group.
1027  InterleaveGroup *getInterleaveGroup(Instruction *Instr) const {
1028  if (InterleaveGroupMap.count(Instr))
1029  return InterleaveGroupMap.find(Instr)->second;
1030  return nullptr;
1031  }
1032 
1033  /// \brief Returns true if an interleaved group that may access memory
1034  /// out-of-bounds requires a scalar epilogue iteration for correctness.
1035  bool requiresScalarEpilogue() const { return RequiresScalarEpilogue; }
1036 
1037  /// \brief Initialize the LoopAccessInfo used for dependence checking.
1038  void setLAI(const LoopAccessInfo *Info) { LAI = Info; }
1039 
1040 private:
1041  /// A wrapper around ScalarEvolution, used to add runtime SCEV checks.
1042  /// Simplifies SCEV expressions in the context of existing SCEV assumptions.
1043  /// The interleaved access analysis can also add new predicates (for example
1044  /// by versioning strides of pointers).
1046 
1047  Loop *TheLoop;
1048  DominatorTree *DT;
1049  LoopInfo *LI;
1050  const LoopAccessInfo *LAI = nullptr;
1051 
1052  /// True if the loop may contain non-reversed interleaved groups with
1053  /// out-of-bounds accesses. We ensure we don't speculatively access memory
1054  /// out-of-bounds by executing at least one scalar epilogue iteration.
1055  bool RequiresScalarEpilogue = false;
1056 
1057  /// Holds the relationships between the members and the interleave group.
1058  DenseMap<Instruction *, InterleaveGroup *> InterleaveGroupMap;
1059 
1060  /// Holds dependences among the memory accesses in the loop. It maps a source
1061  /// access to a set of dependent sink accesses.
1063 
1064  /// \brief The descriptor for a strided memory access.
1065  struct StrideDescriptor {
1066  StrideDescriptor() = default;
1067  StrideDescriptor(int64_t Stride, const SCEV *Scev, uint64_t Size,
1068  unsigned Align)
1069  : Stride(Stride), Scev(Scev), Size(Size), Align(Align) {}
1070 
1071  // The access's stride. It is negative for a reverse access.
1072  int64_t Stride = 0;
1073 
1074  // The scalar expression of this access.
1075  const SCEV *Scev = nullptr;
1076 
1077  // The size of the memory object.
1078  uint64_t Size = 0;
1079 
1080  // The alignment of this access.
1081  unsigned Align = 0;
1082  };
1083 
1084  /// \brief A type for holding instructions and their stride descriptors.
1085  using StrideEntry = std::pair<Instruction *, StrideDescriptor>;
1086 
1087  /// \brief Create a new interleave group with the given instruction \p Instr,
1088  /// stride \p Stride and alignment \p Align.
1089  ///
1090  /// \returns the newly created interleave group.
1091  InterleaveGroup *createInterleaveGroup(Instruction *Instr, int Stride,
1092  unsigned Align) {
1093  assert(!InterleaveGroupMap.count(Instr) &&
1094  "Already in an interleaved access group");
1095  InterleaveGroupMap[Instr] = new InterleaveGroup(Instr, Stride, Align);
1096  return InterleaveGroupMap[Instr];
1097  }
1098 
1099  /// \brief Release the group and remove all the relationships.
1100  void releaseGroup(InterleaveGroup *Group) {
1101  for (unsigned i = 0; i < Group->getFactor(); i++)
1102  if (Instruction *Member = Group->getMember(i))
1103  InterleaveGroupMap.erase(Member);
1104 
1105  delete Group;
1106  }
1107 
1108  /// \brief Collect all the accesses with a constant stride in program order.
1109  void collectConstStrideAccesses(
1111  const ValueToValueMap &Strides);
1112 
1113  /// \brief Returns true if \p Stride is allowed in an interleaved group.
1114  static bool isStrided(int Stride) {
1115  unsigned Factor = std::abs(Stride);
1116  return Factor >= 2 && Factor <= MaxInterleaveGroupFactor;
1117  }
1118 
1119  /// \brief Returns true if \p BB is a predicated block.
1120  bool isPredicated(BasicBlock *BB) const {
1121  return LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
1122  }
1123 
1124  /// \brief Returns true if LoopAccessInfo can be used for dependence queries.
1125  bool areDependencesValid() const {
1126  return LAI && LAI->getDepChecker().getDependences();
1127  }
1128 
1129  /// \brief Returns true if memory accesses \p A and \p B can be reordered, if
1130  /// necessary, when constructing interleaved groups.
1131  ///
1132  /// \p A must precede \p B in program order. We return false if reordering is
1133  /// not necessary or is prevented because \p A and \p B may be dependent.
1134  bool canReorderMemAccessesForInterleavedGroups(StrideEntry *A,
1135  StrideEntry *B) const {
1136  // Code motion for interleaved accesses can potentially hoist strided loads
1137  // and sink strided stores. The code below checks the legality of the
1138  // following two conditions:
1139  //
1140  // 1. Potentially moving a strided load (B) before any store (A) that
1141  // precedes B, or
1142  //
1143  // 2. Potentially moving a strided store (A) after any load or store (B)
1144  // that A precedes.
1145  //
1146  // It's legal to reorder A and B if we know there isn't a dependence from A
1147  // to B. Note that this determination is conservative since some
1148  // dependences could potentially be reordered safely.
1149 
1150  // A is potentially the source of a dependence.
1151  auto *Src = A->first;
1152  auto SrcDes = A->second;
1153 
1154  // B is potentially the sink of a dependence.
1155  auto *Sink = B->first;
1156  auto SinkDes = B->second;
1157 
1158  // Code motion for interleaved accesses can't violate WAR dependences.
1159  // Thus, reordering is legal if the source isn't a write.
1160  if (!Src->mayWriteToMemory())
1161  return true;
1162 
1163  // At least one of the accesses must be strided.
1164  if (!isStrided(SrcDes.Stride) && !isStrided(SinkDes.Stride))
1165  return true;
1166 
1167  // If dependence information is not available from LoopAccessInfo,
1168  // conservatively assume the instructions can't be reordered.
1169  if (!areDependencesValid())
1170  return false;
1171 
1172  // If we know there is a dependence from source to sink, assume the
1173  // instructions can't be reordered. Otherwise, reordering is legal.
1174  return !Dependences.count(Src) || !Dependences.lookup(Src).count(Sink);
1175  }
1176 
1177  /// \brief Collect the dependences from LoopAccessInfo.
1178  ///
1179  /// We process the dependences once during the interleaved access analysis to
1180  /// enable constant-time dependence queries.
1181  void collectDependences() {
1182  if (!areDependencesValid())
1183  return;
1184  auto *Deps = LAI->getDepChecker().getDependences();
1185  for (auto Dep : *Deps)
1186  Dependences[Dep.getSource(*LAI)].insert(Dep.getDestination(*LAI));
1187  }
1188 };
1189 
1190 /// Utility class for getting and setting loop vectorizer hints in the form
1191 /// of loop metadata.
1192 /// This class keeps a number of loop annotations locally (as member variables)
1193 /// and can, upon request, write them back as metadata on the loop. It will
1194 /// initially scan the loop for existing metadata, and will update the local
1195 /// values based on information in the loop.
1196 /// We cannot write all values to metadata, as the mere presence of some info,
1197 /// for example 'force', means a decision has been made. So, we need to be
1198 /// careful NOT to add them if the user hasn't specifically asked so.
1199 class LoopVectorizeHints {
1200  enum HintKind { HK_WIDTH, HK_UNROLL, HK_FORCE, HK_ISVECTORIZED };
1201 
1202  /// Hint - associates name and validation with the hint value.
1203  struct Hint {
1204  const char *Name;
1205  unsigned Value; // This may have to change for non-numeric values.
1206  HintKind Kind;
1207 
1208  Hint(const char *Name, unsigned Value, HintKind Kind)
1209  : Name(Name), Value(Value), Kind(Kind) {}
1210 
1211  bool validate(unsigned Val) {
1212  switch (Kind) {
1213  case HK_WIDTH:
1214  return isPowerOf2_32(Val) && Val <= VectorizerParams::MaxVectorWidth;
1215  case HK_UNROLL:
1216  return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
1217  case HK_FORCE:
1218  return (Val <= 1);
1219  case HK_ISVECTORIZED:
1220  return (Val==0 || Val==1);
1221  }
1222  return false;
1223  }
1224  };
1225 
1226  /// Vectorization width.
1227  Hint Width;
1228 
1229  /// Vectorization interleave factor.
1230  Hint Interleave;
1231 
1232  /// Vectorization forced
1233  Hint Force;
1234 
1235  /// Already Vectorized
1236  Hint IsVectorized;
1237 
1238  /// Return the loop metadata prefix.
1239  static StringRef Prefix() { return "llvm.loop."; }
1240 
1241  /// True if there is any unsafe math in the loop.
1242  bool PotentiallyUnsafe = false;
1243 
1244 public:
1245  enum ForceKind {
1246  FK_Undefined = -1, ///< Not selected.
1247  FK_Disabled = 0, ///< Forcing disabled.
1248  FK_Enabled = 1, ///< Forcing enabled.
1249  };
1250 
1251  LoopVectorizeHints(const Loop *L, bool DisableInterleaving,
1253  : Width("vectorize.width", VectorizerParams::VectorizationFactor,
1254  HK_WIDTH),
1255  Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
1256  Force("vectorize.enable", FK_Undefined, HK_FORCE),
1257  IsVectorized("isvectorized", 0, HK_ISVECTORIZED), TheLoop(L), ORE(ORE) {
1258  // Populate values with existing loop metadata.
1259  getHintsFromMetadata();
1260 
1261  // force-vector-interleave overrides DisableInterleaving.
1263  Interleave.Value = VectorizerParams::VectorizationInterleave;
1264 
1265  if (IsVectorized.Value != 1)
1266  // If the vectorization width and interleaving count are both 1 then
1267  // consider the loop to have been already vectorized because there's
1268  // nothing more that we can do.
1269  IsVectorized.Value = Width.Value == 1 && Interleave.Value == 1;
1270  DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs()
1271  << "LV: Interleaving disabled by the pass manager\n");
1272  }
1273 
1274  /// Mark the loop L as already vectorized by setting the width to 1.
1275  void setAlreadyVectorized() {
1276  IsVectorized.Value = 1;
1277  Hint Hints[] = {IsVectorized};
1278  writeHintsToMetadata(Hints);
1279  }
1280 
1281  bool allowVectorization(Function *F, Loop *L, bool AlwaysVectorize) const {
1282  if (getForce() == LoopVectorizeHints::FK_Disabled) {
1283  DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
1284  emitRemarkWithHints();
1285  return false;
1286  }
1287 
1288  if (!AlwaysVectorize && getForce() != LoopVectorizeHints::FK_Enabled) {
1289  DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
1290  emitRemarkWithHints();
1291  return false;
1292  }
1293 
1294  if (getIsVectorized() == 1) {
1295  DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
1296  // FIXME: Add interleave.disable metadata. This will allow
1297  // vectorize.disable to be used without disabling the pass and errors
1298  // to differentiate between disabled vectorization and a width of 1.
1299  ORE.emit([&]() {
1300  return OptimizationRemarkAnalysis(vectorizeAnalysisPassName(),
1301  "AllDisabled", L->getStartLoc(),
1302  L->getHeader())
1303  << "loop not vectorized: vectorization and interleaving are "
1304  "explicitly disabled, or the loop has already been "
1305  "vectorized";
1306  });
1307  return false;
1308  }
1309 
1310  return true;
1311  }
1312 
1313  /// Dumps all the hint information.
1314  void emitRemarkWithHints() const {
1315  using namespace ore;
1316 
1317  ORE.emit([&]() {
1318  if (Force.Value == LoopVectorizeHints::FK_Disabled)
1319  return OptimizationRemarkMissed(LV_NAME, "MissedExplicitlyDisabled",
1320  TheLoop->getStartLoc(),
1321  TheLoop->getHeader())
1322  << "loop not vectorized: vectorization is explicitly disabled";
1323  else {
1324  OptimizationRemarkMissed R(LV_NAME, "MissedDetails",
1325  TheLoop->getStartLoc(),
1326  TheLoop->getHeader());
1327  R << "loop not vectorized";
1328  if (Force.Value == LoopVectorizeHints::FK_Enabled) {
1329  R << " (Force=" << NV("Force", true);
1330  if (Width.Value != 0)
1331  R << ", Vector Width=" << NV("VectorWidth", Width.Value);
1332  if (Interleave.Value != 0)
1333  R << ", Interleave Count="
1334  << NV("InterleaveCount", Interleave.Value);
1335  R << ")";
1336  }
1337  return R;
1338  }
1339  });
1340  }
1341 
1342  unsigned getWidth() const { return Width.Value; }
1343  unsigned getInterleave() const { return Interleave.Value; }
1344  unsigned getIsVectorized() const { return IsVectorized.Value; }
1345  enum ForceKind getForce() const { return (ForceKind)Force.Value; }
1346 
1347  /// \brief If hints are provided that force vectorization, use the AlwaysPrint
1348  /// pass name to force the frontend to print the diagnostic.
1349  const char *vectorizeAnalysisPassName() const {
1350  if (getWidth() == 1)
1351  return LV_NAME;
1352  if (getForce() == LoopVectorizeHints::FK_Disabled)
1353  return LV_NAME;
1354  if (getForce() == LoopVectorizeHints::FK_Undefined && getWidth() == 0)
1355  return LV_NAME;
1357  }
1358 
1359  bool allowReordering() const {
1360  // When enabling loop hints are provided we allow the vectorizer to change
1361  // the order of operations that is given by the scalar loop. This is not
1362  // enabled by default because can be unsafe or inefficient. For example,
1363  // reordering floating-point operations will change the way round-off
1364  // error accumulates in the loop.
1365  return getForce() == LoopVectorizeHints::FK_Enabled || getWidth() > 1;
1366  }
1367 
1368  bool isPotentiallyUnsafe() const {
1369  // Avoid FP vectorization if the target is unsure about proper support.
1370  // This may be related to the SIMD unit in the target not handling
1371  // IEEE 754 FP ops properly, or bad single-to-double promotions.
1372  // Otherwise, a sequence of vectorized loops, even without reduction,
1373  // could lead to different end results on the destination vectors.
1374  return getForce() != LoopVectorizeHints::FK_Enabled && PotentiallyUnsafe;
1375  }
1376 
1377  void setPotentiallyUnsafe() { PotentiallyUnsafe = true; }
1378 
1379 private:
1380  /// Find hints specified in the loop metadata and update local values.
1381  void getHintsFromMetadata() {
1382  MDNode *LoopID = TheLoop->getLoopID();
1383  if (!LoopID)
1384  return;
1385 
1386  // First operand should refer to the loop id itself.
1387  assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
1388  assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
1389 
1390  for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1391  const MDString *S = nullptr;
1393 
1394  // The expected hint is either a MDString or a MDNode with the first
1395  // operand a MDString.
1396  if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
1397  if (!MD || MD->getNumOperands() == 0)
1398  continue;
1399  S = dyn_cast<MDString>(MD->getOperand(0));
1400  for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
1401  Args.push_back(MD->getOperand(i));
1402  } else {
1403  S = dyn_cast<MDString>(LoopID->getOperand(i));
1404  assert(Args.size() == 0 && "too many arguments for MDString");
1405  }
1406 
1407  if (!S)
1408  continue;
1409 
1410  // Check if the hint starts with the loop metadata prefix.
1411  StringRef Name = S->getString();
1412  if (Args.size() == 1)
1413  setHint(Name, Args[0]);
1414  }
1415  }
1416 
1417  /// Checks string hint with one operand and set value if valid.
1418  void setHint(StringRef Name, Metadata *Arg) {
1419  if (!Name.startswith(Prefix()))
1420  return;
1421  Name = Name.substr(Prefix().size(), StringRef::npos);
1422 
1423  const ConstantInt *C = mdconst::dyn_extract<ConstantInt>(Arg);
1424  if (!C)
1425  return;
1426  unsigned Val = C->getZExtValue();
1427 
1428  Hint *Hints[] = {&Width, &Interleave, &Force, &IsVectorized};
1429  for (auto H : Hints) {
1430  if (Name == H->Name) {
1431  if (H->validate(Val))
1432  H->Value = Val;
1433  else
1434  DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n");
1435  break;
1436  }
1437  }
1438  }
1439 
1440  /// Create a new hint from name / value pair.
1441  MDNode *createHintMetadata(StringRef Name, unsigned V) const {
1442  LLVMContext &Context = TheLoop->getHeader()->getContext();
1443  Metadata *MDs[] = {MDString::get(Context, Name),
1445  ConstantInt::get(Type::getInt32Ty(Context), V))};
1446  return MDNode::get(Context, MDs);
1447  }
1448 
1449  /// Matches metadata with hint name.
1450  bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) {
1451  MDString *Name = dyn_cast<MDString>(Node->getOperand(0));
1452  if (!Name)
1453  return false;
1454 
1455  for (auto H : HintTypes)
1456  if (Name->getString().endswith(H.Name))
1457  return true;
1458  return false;
1459  }
1460 
1461  /// Sets current hints into loop metadata, keeping other values intact.
1462  void writeHintsToMetadata(ArrayRef<Hint> HintTypes) {
1463  if (HintTypes.empty())
1464  return;
1465 
1466  // Reserve the first element to LoopID (see below).
1468  // If the loop already has metadata, then ignore the existing operands.
1469  MDNode *LoopID = TheLoop->getLoopID();
1470  if (LoopID) {
1471  for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1472  MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
1473  // If node in update list, ignore old value.
1474  if (!matchesHintMetadataName(Node, HintTypes))
1475  MDs.push_back(Node);
1476  }
1477  }
1478 
1479  // Now, add the missing hints.
1480  for (auto H : HintTypes)
1481  MDs.push_back(createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
1482 
1483  // Replace current metadata node with new one.
1484  LLVMContext &Context = TheLoop->getHeader()->getContext();
1485  MDNode *NewLoopID = MDNode::get(Context, MDs);
1486  // Set operand 0 to refer to the loop id itself.
1487  NewLoopID->replaceOperandWith(0, NewLoopID);
1488 
1489  TheLoop->setLoopID(NewLoopID);
1490  }
1491 
1492  /// The loop these hints belong to.
1493  const Loop *TheLoop;
1494 
1495  /// Interface to emit optimization remarks.
1497 };
1498 
1499 } // end anonymous namespace
1500 
1502  const LoopVectorizeHints &LH,
1504  LH.emitRemarkWithHints();
1505 
1506  if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1507  if (LH.getWidth() != 1)
1509  DEBUG_TYPE, "FailedRequestedVectorization",
1510  L->getStartLoc(), L->getHeader())
1511  << "loop not vectorized: "
1512  << "failed explicitly specified loop vectorization");
1513  else if (LH.getInterleave() != 1)
1515  DEBUG_TYPE, "FailedRequestedInterleaving", L->getStartLoc(),
1516  L->getHeader())
1517  << "loop not interleaved: "
1518  << "failed explicitly specified loop interleaving");
1519  }
1520 }
1521 
1522 namespace {
1523 
1524 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
1525 /// to what vectorization factor.
1526 /// This class does not look at the profitability of vectorization, only the
1527 /// legality. This class has two main kinds of checks:
1528 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
1529 /// will change the order of memory accesses in a way that will change the
1530 /// correctness of the program.
1531 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
1532 /// checks for a number of different conditions, such as the availability of a
1533 /// single induction variable, that all types are supported and vectorize-able,
1534 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
1535 /// This class is also used by InnerLoopVectorizer for identifying
1536 /// induction variable and the different reduction variables.
1537 class LoopVectorizationLegality {
1538 public:
1539  LoopVectorizationLegality(
1542  const TargetTransformInfo *TTI,
1543  std::function<const LoopAccessInfo &(Loop &)> *GetLAA, LoopInfo *LI,
1544  OptimizationRemarkEmitter *ORE, LoopVectorizationRequirements *R,
1545  LoopVectorizeHints *H)
1546  : TheLoop(L), PSE(PSE), TLI(TLI), TTI(TTI), DT(DT), GetLAA(GetLAA),
1547  ORE(ORE), InterleaveInfo(PSE, L, DT, LI), Requirements(R), Hints(H) {}
1548 
1549  /// ReductionList contains the reduction descriptors for all
1550  /// of the reductions that were found in the loop.
1551  using ReductionList = DenseMap<PHINode *, RecurrenceDescriptor>;
1552 
1553  /// InductionList saves induction variables and maps them to the
1554  /// induction descriptor.
1555  using InductionList = MapVector<PHINode *, InductionDescriptor>;
1556 
1557  /// RecurrenceSet contains the phi nodes that are recurrences other than
1558  /// inductions and reductions.
1559  using RecurrenceSet = SmallPtrSet<const PHINode *, 8>;
1560 
1561  /// Returns true if it is legal to vectorize this loop.
1562  /// This does not mean that it is profitable to vectorize this
1563  /// loop, only that it is legal to do so.
1564  bool canVectorize();
1565 
1566  /// Returns the primary induction variable.
1567  PHINode *getPrimaryInduction() { return PrimaryInduction; }
1568 
1569  /// Returns the reduction variables found in the loop.
1570  ReductionList *getReductionVars() { return &Reductions; }
1571 
1572  /// Returns the induction variables found in the loop.
1573  InductionList *getInductionVars() { return &Inductions; }
1574 
1575  /// Return the first-order recurrences found in the loop.
1576  RecurrenceSet *getFirstOrderRecurrences() { return &FirstOrderRecurrences; }
1577 
1578  /// Return the set of instructions to sink to handle first-order recurrences.
1579  DenseMap<Instruction *, Instruction *> &getSinkAfter() { return SinkAfter; }
1580 
1581  /// Returns the widest induction type.
1582  Type *getWidestInductionType() { return WidestIndTy; }
1583 
1584  /// Returns True if V is a Phi node of an induction variable in this loop.
1585  bool isInductionPhi(const Value *V);
1586 
1587  /// Returns True if V is a cast that is part of an induction def-use chain,
1588  /// and had been proven to be redundant under a runtime guard (in other
1589  /// words, the cast has the same SCEV expression as the induction phi).
1590  bool isCastedInductionVariable(const Value *V);
1591 
1592  /// Returns True if V can be considered as an induction variable in this
1593  /// loop. V can be the induction phi, or some redundant cast in the def-use
1594  /// chain of the inducion phi.
1595  bool isInductionVariable(const Value *V);
1596 
1597  /// Returns True if PN is a reduction variable in this loop.
1598  bool isReductionVariable(PHINode *PN) { return Reductions.count(PN); }
1599 
1600  /// Returns True if Phi is a first-order recurrence in this loop.
1601  bool isFirstOrderRecurrence(const PHINode *Phi);
1602 
1603  /// Return true if the block BB needs to be predicated in order for the loop
1604  /// to be vectorized.
1605  bool blockNeedsPredication(BasicBlock *BB);
1606 
1607  /// Check if this pointer is consecutive when vectorizing. This happens
1608  /// when the last index of the GEP is the induction variable, or that the
1609  /// pointer itself is an induction variable.
1610  /// This check allows us to vectorize A[idx] into a wide load/store.
1611  /// Returns:
1612  /// 0 - Stride is unknown or non-consecutive.
1613  /// 1 - Address is consecutive.
1614  /// -1 - Address is consecutive, and decreasing.
1615  /// NOTE: This method must only be used before modifying the original scalar
1616  /// loop. Do not use after invoking 'createVectorizedLoopSkeleton' (PR34965).
1617  int isConsecutivePtr(Value *Ptr);
1618 
1619  /// Returns true if the value V is uniform within the loop.
1620  bool isUniform(Value *V);
1621 
1622  /// Returns the information that we collected about runtime memory check.
1623  const RuntimePointerChecking *getRuntimePointerChecking() const {
1624  return LAI->getRuntimePointerChecking();
1625  }
1626 
1627  const LoopAccessInfo *getLAI() const { return LAI; }
1628 
1629  /// \brief Check if \p Instr belongs to any interleaved access group.
1630  bool isAccessInterleaved(Instruction *Instr) {
1631  return InterleaveInfo.isInterleaved(Instr);
1632  }
1633 
1634  /// \brief Get the interleaved access group that \p Instr belongs to.
1635  const InterleaveGroup *getInterleavedAccessGroup(Instruction *Instr) {
1636  return InterleaveInfo.getInterleaveGroup(Instr);
1637  }
1638 
1639  /// \brief Returns true if an interleaved group requires a scalar iteration
1640  /// to handle accesses with gaps.
1641  bool requiresScalarEpilogue() const {
1642  return InterleaveInfo.requiresScalarEpilogue();
1643  }
1644 
1645  unsigned getMaxSafeDepDistBytes() { return LAI->getMaxSafeDepDistBytes(); }
1646 
1647  uint64_t getMaxSafeRegisterWidth() const {
1648  return LAI->getDepChecker().getMaxSafeRegisterWidth();
1649  }
1650 
1651  bool hasStride(Value *V) { return LAI->hasStride(V); }
1652 
1653  /// Returns true if the target machine supports masked store operation
1654  /// for the given \p DataType and kind of access to \p Ptr.
1655  bool isLegalMaskedStore(Type *DataType, Value *Ptr) {
1656  return isConsecutivePtr(Ptr) && TTI->isLegalMaskedStore(DataType);
1657  }
1658 
1659  /// Returns true if the target machine supports masked load operation
1660  /// for the given \p DataType and kind of access to \p Ptr.
1661  bool isLegalMaskedLoad(Type *DataType, Value *Ptr) {
1662  return isConsecutivePtr(Ptr) && TTI->isLegalMaskedLoad(DataType);
1663  }
1664 
1665  /// Returns true if the target machine supports masked scatter operation
1666  /// for the given \p DataType.
1667  bool isLegalMaskedScatter(Type *DataType) {
1668  return TTI->isLegalMaskedScatter(DataType);
1669  }
1670 
1671  /// Returns true if the target machine supports masked gather operation
1672  /// for the given \p DataType.
1673  bool isLegalMaskedGather(Type *DataType) {
1674  return TTI->isLegalMaskedGather(DataType);
1675  }
1676 
1677  /// Returns true if the target machine can represent \p V as a masked gather
1678  /// or scatter operation.
1679  bool isLegalGatherOrScatter(Value *V) {
1680  auto *LI = dyn_cast<LoadInst>(V);
1681  auto *SI = dyn_cast<StoreInst>(V);
1682  if (!LI && !SI)
1683  return false;
1684  auto *Ptr = getPointerOperand(V);
1685  auto *Ty = cast<PointerType>(Ptr->getType())->getElementType();
1686  return (LI && isLegalMaskedGather(Ty)) || (SI && isLegalMaskedScatter(Ty));
1687  }
1688 
1689  /// Returns true if vector representation of the instruction \p I
1690  /// requires mask.
1691  bool isMaskRequired(const Instruction *I) { return (MaskedOp.count(I) != 0); }
1692 
1693  unsigned getNumStores() const { return LAI->getNumStores(); }
1694  unsigned getNumLoads() const { return LAI->getNumLoads(); }
1695  unsigned getNumPredStores() const { return NumPredStores; }
1696 
1697  /// Returns true if \p I is an instruction that will be scalarized with
1698  /// predication. Such instructions include conditional stores and
1699  /// instructions that may divide by zero.
1700  bool isScalarWithPredication(Instruction *I);
1701 
1702  /// Returns true if \p I is a memory instruction with consecutive memory
1703  /// access that can be widened.
1704  bool memoryInstructionCanBeWidened(Instruction *I, unsigned VF = 1);
1705 
1706  // Returns true if the NoNaN attribute is set on the function.
1707  bool hasFunNoNaNAttr() const { return HasFunNoNaNAttr; }
1708 
1709 private:
1710  /// Check if a single basic block loop is vectorizable.
1711  /// At this point we know that this is a loop with a constant trip count
1712  /// and we only need to check individual instructions.
1713  bool canVectorizeInstrs();
1714 
1715  /// When we vectorize loops we may change the order in which
1716  /// we read and write from memory. This method checks if it is
1717  /// legal to vectorize the code, considering only memory constrains.
1718  /// Returns true if the loop is vectorizable
1719  bool canVectorizeMemory();
1720 
1721  /// Return true if we can vectorize this loop using the IF-conversion
1722  /// transformation.
1723  bool canVectorizeWithIfConvert();
1724 
1725  /// Return true if all of the instructions in the block can be speculatively
1726  /// executed. \p SafePtrs is a list of addresses that are known to be legal
1727  /// and we know that we can read from them without segfault.
1728  bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
1729 
1730  /// Updates the vectorization state by adding \p Phi to the inductions list.
1731  /// This can set \p Phi as the main induction of the loop if \p Phi is a
1732  /// better choice for the main induction than the existing one.
1733  void addInductionPhi(PHINode *Phi, const InductionDescriptor &ID,
1734  SmallPtrSetImpl<Value *> &AllowedExit);
1735 
1736  /// Create an analysis remark that explains why vectorization failed
1737  ///
1738  /// \p RemarkName is the identifier for the remark. If \p I is passed it is
1739  /// an instruction that prevents vectorization. Otherwise the loop is used
1740  /// for the location of the remark. \return the remark object that can be
1741  /// streamed to.
1743  createMissedAnalysis(StringRef RemarkName, Instruction *I = nullptr) const {
1744  return ::createMissedAnalysis(Hints->vectorizeAnalysisPassName(),
1745  RemarkName, TheLoop, I);
1746  }
1747 
1748  /// \brief If an access has a symbolic strides, this maps the pointer value to
1749  /// the stride symbol.
1750  const ValueToValueMap *getSymbolicStrides() {
1751  // FIXME: Currently, the set of symbolic strides is sometimes queried before
1752  // it's collected. This happens from canVectorizeWithIfConvert, when the
1753  // pointer is checked to reference consecutive elements suitable for a
1754  // masked access.
1755  return LAI ? &LAI->getSymbolicStrides() : nullptr;
1756  }
1757 
1758  unsigned NumPredStores = 0;
1759 
1760  /// The loop that we evaluate.
1761  Loop *TheLoop;
1762 
1763  /// A wrapper around ScalarEvolution used to add runtime SCEV checks.
1764  /// Applies dynamic knowledge to simplify SCEV expressions in the context
1765  /// of existing SCEV assumptions. The analysis will also add a minimal set
1766  /// of new predicates if this is required to enable vectorization and
1767  /// unrolling.
1769 
1770  /// Target Library Info.
1772 
1773  /// Target Transform Info
1774  const TargetTransformInfo *TTI;
1775 
1776  /// Dominator Tree.
1777  DominatorTree *DT;
1778 
1779  // LoopAccess analysis.
1780  std::function<const LoopAccessInfo &(Loop &)> *GetLAA;
1781 
1782  // And the loop-accesses info corresponding to this loop. This pointer is
1783  // null until canVectorizeMemory sets it up.
1784  const LoopAccessInfo *LAI = nullptr;
1785 
1786  /// Interface to emit optimization remarks.
1788 
1789  /// The interleave access information contains groups of interleaved accesses
1790  /// with the same stride and close to each other.
1791  InterleavedAccessInfo InterleaveInfo;
1792 
1793  // --- vectorization state --- //
1794 
1795  /// Holds the primary induction variable. This is the counter of the
1796  /// loop.
1797  PHINode *PrimaryInduction = nullptr;
1798 
1799  /// Holds the reduction variables.
1800  ReductionList Reductions;
1801 
1802  /// Holds all of the induction variables that we found in the loop.
1803  /// Notice that inductions don't need to start at zero and that induction
1804  /// variables can be pointers.
1805  InductionList Inductions;
1806 
1807  /// Holds all the casts that participate in the update chain of the induction
1808  /// variables, and that have been proven to be redundant (possibly under a
1809  /// runtime guard). These casts can be ignored when creating the vectorized
1810  /// loop body.
1811  SmallPtrSet<Instruction *, 4> InductionCastsToIgnore;
1812 
1813  /// Holds the phi nodes that are first-order recurrences.
1814  RecurrenceSet FirstOrderRecurrences;
1815 
1816  /// Holds instructions that need to sink past other instructions to handle
1817  /// first-order recurrences.
1819 
1820  /// Holds the widest induction type encountered.
1821  Type *WidestIndTy = nullptr;
1822 
1823  /// Allowed outside users. This holds the induction and reduction
1824  /// vars which can be accessed from outside the loop.
1825  SmallPtrSet<Value *, 4> AllowedExit;
1826 
1827  /// Can we assume the absence of NaNs.
1828  bool HasFunNoNaNAttr = false;
1829 
1830  /// Vectorization requirements that will go through late-evaluation.
1831  LoopVectorizationRequirements *Requirements;
1832 
1833  /// Used to emit an analysis of any legality issues.
1834  LoopVectorizeHints *Hints;
1835 
1836  /// While vectorizing these instructions we have to generate a
1837  /// call to the appropriate masked intrinsic
1839 };
1840 
1841 /// LoopVectorizationCostModel - estimates the expected speedups due to
1842 /// vectorization.
1843 /// In many cases vectorization is not profitable. This can happen because of
1844 /// a number of reasons. In this class we mainly attempt to predict the
1845 /// expected speedup/slowdowns due to the supported instruction set. We use the
1846 /// TargetTransformInfo to query the different backends for the cost of
1847 /// different operations.
1848 class LoopVectorizationCostModel {
1849 public:
1850  LoopVectorizationCostModel(Loop *L, PredicatedScalarEvolution &PSE,
1851  LoopInfo *LI, LoopVectorizationLegality *Legal,
1852  const TargetTransformInfo &TTI,
1853  const TargetLibraryInfo *TLI, DemandedBits *DB,
1856  const LoopVectorizeHints *Hints)
1857  : TheLoop(L), PSE(PSE), LI(LI), Legal(Legal), TTI(TTI), TLI(TLI), DB(DB),
1858  AC(AC), ORE(ORE), TheFunction(F), Hints(Hints) {}
1859 
1860  /// \return An upper bound for the vectorization factor, or None if
1861  /// vectorization should be avoided up front.
1862  Optional<unsigned> computeMaxVF(bool OptForSize);
1863 
1864  /// Information about vectorization costs
1865  struct VectorizationFactor {
1866  // Vector width with best cost
1867  unsigned Width;
1868 
1869  // Cost of the loop with that width
1870  unsigned Cost;
1871  };
1872 
1873  /// \return The most profitable vectorization factor and the cost of that VF.
1874  /// This method checks every power of two up to MaxVF. If UserVF is not ZERO
1875  /// then this vectorization factor will be selected if vectorization is
1876  /// possible.
1877  VectorizationFactor selectVectorizationFactor(unsigned MaxVF);
1878 
1879  /// Setup cost-based decisions for user vectorization factor.
1880  void selectUserVectorizationFactor(unsigned UserVF) {
1881  collectUniformsAndScalars(UserVF);
1882  collectInstsToScalarize(UserVF);
1883  }
1884 
1885  /// \return The size (in bits) of the smallest and widest types in the code
1886  /// that needs to be vectorized. We ignore values that remain scalar such as
1887  /// 64 bit loop indices.
1888  std::pair<unsigned, unsigned> getSmallestAndWidestTypes();
1889 
1890  /// \return The desired interleave count.
1891  /// If interleave count has been specified by metadata it will be returned.
1892  /// Otherwise, the interleave count is computed and returned. VF and LoopCost
1893  /// are the selected vectorization factor and the cost of the selected VF.
1894  unsigned selectInterleaveCount(bool OptForSize, unsigned VF,
1895  unsigned LoopCost);
1896 
1897  /// Memory access instruction may be vectorized in more than one way.
1898  /// Form of instruction after vectorization depends on cost.
1899  /// This function takes cost-based decisions for Load/Store instructions
1900  /// and collects them in a map. This decisions map is used for building
1901  /// the lists of loop-uniform and loop-scalar instructions.
1902  /// The calculated cost is saved with widening decision in order to
1903  /// avoid redundant calculations.
1904  void setCostBasedWideningDecision(unsigned VF);
1905 
1906  /// \brief A struct that represents some properties of the register usage
1907  /// of a loop.
1908  struct RegisterUsage {
1909  /// Holds the number of loop invariant values that are used in the loop.
1910  unsigned LoopInvariantRegs;
1911 
1912  /// Holds the maximum number of concurrent live intervals in the loop.
1913  unsigned MaxLocalUsers;
1914 
1915  /// Holds the number of instructions in the loop.
1916  unsigned NumInstructions;
1917  };
1918 
1919  /// \return Returns information about the register usages of the loop for the
1920  /// given vectorization factors.
1921  SmallVector<RegisterUsage, 8> calculateRegisterUsage(ArrayRef<unsigned> VFs);
1922 
1923  /// Collect values we want to ignore in the cost model.
1924  void collectValuesToIgnore();
1925 
1926  /// \returns The smallest bitwidth each instruction can be represented with.
1927  /// The vector equivalents of these instructions should be truncated to this
1928  /// type.
1929  const MapVector<Instruction *, uint64_t> &getMinimalBitwidths() const {
1930  return MinBWs;
1931  }
1932 
1933  /// \returns True if it is more profitable to scalarize instruction \p I for
1934  /// vectorization factor \p VF.
1935  bool isProfitableToScalarize(Instruction *I, unsigned VF) const {
1936  assert(VF > 1 && "Profitable to scalarize relevant only for VF > 1.");
1937  auto Scalars = InstsToScalarize.find(VF);
1938  assert(Scalars != InstsToScalarize.end() &&
1939  "VF not yet analyzed for scalarization profitability");
1940  return Scalars->second.count(I);
1941  }
1942 
1943  /// Returns true if \p I is known to be uniform after vectorization.
1944  bool isUniformAfterVectorization(Instruction *I, unsigned VF) const {
1945  if (VF == 1)
1946  return true;
1947  assert(Uniforms.count(VF) && "VF not yet analyzed for uniformity");
1948  auto UniformsPerVF = Uniforms.find(VF);
1949  return UniformsPerVF->second.count(I);
1950  }
1951 
1952  /// Returns true if \p I is known to be scalar after vectorization.
1953  bool isScalarAfterVectorization(Instruction *I, unsigned VF) const {
1954  if (VF == 1)
1955  return true;
1956  assert(Scalars.count(VF) && "Scalar values are not calculated for VF");
1957  auto ScalarsPerVF = Scalars.find(VF);
1958  return ScalarsPerVF->second.count(I);
1959  }
1960 
1961  /// \returns True if instruction \p I can be truncated to a smaller bitwidth
1962  /// for vectorization factor \p VF.
1963  bool canTruncateToMinimalBitwidth(Instruction *I, unsigned VF) const {
1964  return VF > 1 && MinBWs.count(I) && !isProfitableToScalarize(I, VF) &&
1965  !isScalarAfterVectorization(I, VF);
1966  }
1967 
1968  /// Decision that was taken during cost calculation for memory instruction.
1969  enum InstWidening {
1970  CM_Unknown,
1971  CM_Widen, // For consecutive accesses with stride +1.
1972  CM_Widen_Reverse, // For consecutive accesses with stride -1.
1973  CM_Interleave,
1974  CM_GatherScatter,
1975  CM_Scalarize
1976  };
1977 
1978  /// Save vectorization decision \p W and \p Cost taken by the cost model for
1979  /// instruction \p I and vector width \p VF.
1980  void setWideningDecision(Instruction *I, unsigned VF, InstWidening W,
1981  unsigned Cost) {
1982  assert(VF >= 2 && "Expected VF >=2");
1983  WideningDecisions[std::make_pair(I, VF)] = std::make_pair(W, Cost);
1984  }
1985 
1986  /// Save vectorization decision \p W and \p Cost taken by the cost model for
1987  /// interleaving group \p Grp and vector width \p VF.
1988  void setWideningDecision(const InterleaveGroup *Grp, unsigned VF,
1989  InstWidening W, unsigned Cost) {
1990  assert(VF >= 2 && "Expected VF >=2");
1991  /// Broadcast this decicion to all instructions inside the group.
1992  /// But the cost will be assigned to one instruction only.
1993  for (unsigned i = 0; i < Grp->getFactor(); ++i) {
1994  if (auto *I = Grp->getMember(i)) {
1995  if (Grp->getInsertPos() == I)
1996  WideningDecisions[std::make_pair(I, VF)] = std::make_pair(W, Cost);
1997  else
1998  WideningDecisions[std::make_pair(I, VF)] = std::make_pair(W, 0);
1999  }
2000  }
2001  }
2002 
2003  /// Return the cost model decision for the given instruction \p I and vector
2004  /// width \p VF. Return CM_Unknown if this instruction did not pass
2005  /// through the cost modeling.
2006  InstWidening getWideningDecision(Instruction *I, unsigned VF) {
2007  assert(VF >= 2 && "Expected VF >=2");
2008  std::pair<Instruction *, unsigned> InstOnVF = std::make_pair(I, VF);
2009  auto Itr = WideningDecisions.find(InstOnVF);
2010  if (Itr == WideningDecisions.end())
2011  return CM_Unknown;
2012  return Itr->second.first;
2013  }
2014 
2015  /// Return the vectorization cost for the given instruction \p I and vector
2016  /// width \p VF.
2017  unsigned getWideningCost(Instruction *I, unsigned VF) {
2018  assert(VF >= 2 && "Expected VF >=2");
2019  std::pair<Instruction *, unsigned> InstOnVF = std::make_pair(I, VF);
2020  assert(WideningDecisions.count(InstOnVF) && "The cost is not calculated");
2021  return WideningDecisions[InstOnVF].second;
2022  }
2023 
2024  /// Return True if instruction \p I is an optimizable truncate whose operand
2025  /// is an induction variable. Such a truncate will be removed by adding a new
2026  /// induction variable with the destination type.
2027  bool isOptimizableIVTruncate(Instruction *I, unsigned VF) {
2028  // If the instruction is not a truncate, return false.
2029  auto *Trunc = dyn_cast<TruncInst>(I);
2030  if (!Trunc)
2031  return false;
2032 
2033  // Get the source and destination types of the truncate.
2034  Type *SrcTy = ToVectorTy(cast<CastInst>(I)->getSrcTy(), VF);
2035  Type *DestTy = ToVectorTy(cast<CastInst>(I)->getDestTy(), VF);
2036 
2037  // If the truncate is free for the given types, return false. Replacing a
2038  // free truncate with an induction variable would add an induction variable
2039  // update instruction to each iteration of the loop. We exclude from this
2040  // check the primary induction variable since it will need an update
2041  // instruction regardless.
2042  Value *Op = Trunc->getOperand(0);
2043  if (Op != Legal->getPrimaryInduction() && TTI.isTruncateFree(SrcTy, DestTy))
2044  return false;
2045 
2046  // If the truncated value is not an induction variable, return false.
2047  return Legal->isInductionPhi(Op);
2048  }
2049 
2050  /// Collects the instructions to scalarize for each predicated instruction in
2051  /// the loop.
2052  void collectInstsToScalarize(unsigned VF);
2053 
2054  /// Collect Uniform and Scalar values for the given \p VF.
2055  /// The sets depend on CM decision for Load/Store instructions
2056  /// that may be vectorized as interleave, gather-scatter or scalarized.
2057  void collectUniformsAndScalars(unsigned VF) {
2058  // Do the analysis once.
2059  if (VF == 1 || Uniforms.count(VF))
2060  return;
2061  setCostBasedWideningDecision(VF);
2062  collectLoopUniforms(VF);
2063  collectLoopScalars(VF);
2064  }
2065 
2066 private:
2067  /// \return An upper bound for the vectorization factor, larger than zero.
2068  /// One is returned if vectorization should best be avoided due to cost.
2069  unsigned computeFeasibleMaxVF(bool OptForSize, unsigned ConstTripCount);
2070 
2071  /// The vectorization cost is a combination of the cost itself and a boolean
2072  /// indicating whether any of the contributing operations will actually
2073  /// operate on
2074  /// vector values after type legalization in the backend. If this latter value
2075  /// is
2076  /// false, then all operations will be scalarized (i.e. no vectorization has
2077  /// actually taken place).
2078  using VectorizationCostTy = std::pair<unsigned, bool>;
2079 
2080  /// Returns the expected execution cost. The unit of the cost does
2081  /// not matter because we use the 'cost' units to compare different
2082  /// vector widths. The cost that is returned is *not* normalized by
2083  /// the factor width.
2084  VectorizationCostTy expectedCost(unsigned VF);
2085 
2086  /// Returns the execution time cost of an instruction for a given vector
2087  /// width. Vector width of one means scalar.
2088  VectorizationCostTy getInstructionCost(Instruction *I, unsigned VF);
2089 
2090  /// The cost-computation logic from getInstructionCost which provides
2091  /// the vector type as an output parameter.
2092  unsigned getInstructionCost(Instruction *I, unsigned VF, Type *&VectorTy);
2093 
2094  /// Calculate vectorization cost of memory instruction \p I.
2095  unsigned getMemoryInstructionCost(Instruction *I, unsigned VF);
2096 
2097  /// The cost computation for scalarized memory instruction.
2098  unsigned getMemInstScalarizationCost(Instruction *I, unsigned VF);
2099 
2100  /// The cost computation for interleaving group of memory instructions.
2101  unsigned getInterleaveGroupCost(Instruction *I, unsigned VF);
2102 
2103  /// The cost computation for Gather/Scatter instruction.
2104  unsigned getGatherScatterCost(Instruction *I, unsigned VF);
2105 
2106  /// The cost computation for widening instruction \p I with consecutive
2107  /// memory access.
2108  unsigned getConsecutiveMemOpCost(Instruction *I, unsigned VF);
2109 
2110  /// The cost calculation for Load instruction \p I with uniform pointer -
2111  /// scalar load + broadcast.
2112  unsigned getUniformMemOpCost(Instruction *I, unsigned VF);
2113 
2114  /// Returns whether the instruction is a load or store and will be a emitted
2115  /// as a vector operation.
2116  bool isConsecutiveLoadOrStore(Instruction *I);
2117 
2118  /// Create an analysis remark that explains why vectorization failed
2119  ///
2120  /// \p RemarkName is the identifier for the remark. \return the remark object
2121  /// that can be streamed to.
2123  return ::createMissedAnalysis(Hints->vectorizeAnalysisPassName(),
2124  RemarkName, TheLoop);
2125  }
2126 
2127  /// Map of scalar integer values to the smallest bitwidth they can be legally
2128  /// represented as. The vector equivalents of these values should be truncated
2129  /// to this type.
2131 
2132  /// A type representing the costs for instructions if they were to be
2133  /// scalarized rather than vectorized. The entries are Instruction-Cost
2134  /// pairs.
2135  using ScalarCostsTy = DenseMap<Instruction *, unsigned>;
2136 
2137  /// A set containing all BasicBlocks that are known to present after
2138  /// vectorization as a predicated block.
2139  SmallPtrSet<BasicBlock *, 4> PredicatedBBsAfterVectorization;
2140 
2141  /// A map holding scalar costs for different vectorization factors. The
2142  /// presence of a cost for an instruction in the mapping indicates that the
2143  /// instruction will be scalarized when vectorizing with the associated
2144  /// vectorization factor. The entries are VF-ScalarCostTy pairs.
2145  DenseMap<unsigned, ScalarCostsTy> InstsToScalarize;
2146 
2147  /// Holds the instructions known to be uniform after vectorization.
2148  /// The data is collected per VF.
2150 
2151  /// Holds the instructions known to be scalar after vectorization.
2152  /// The data is collected per VF.
2154 
2155  /// Holds the instructions (address computations) that are forced to be
2156  /// scalarized.
2158 
2159  /// Returns the expected difference in cost from scalarizing the expression
2160  /// feeding a predicated instruction \p PredInst. The instructions to
2161  /// scalarize and their scalar costs are collected in \p ScalarCosts. A
2162  /// non-negative return value implies the expression will be scalarized.
2163  /// Currently, only single-use chains are considered for scalarization.
2164  int computePredInstDiscount(Instruction *PredInst, ScalarCostsTy &ScalarCosts,
2165  unsigned VF);
2166 
2167  /// Collect the instructions that are uniform after vectorization. An
2168  /// instruction is uniform if we represent it with a single scalar value in
2169  /// the vectorized loop corresponding to each vector iteration. Examples of
2170  /// uniform instructions include pointer operands of consecutive or
2171  /// interleaved memory accesses. Note that although uniformity implies an
2172  /// instruction will be scalar, the reverse is not true. In general, a
2173  /// scalarized instruction will be represented by VF scalar values in the
2174  /// vectorized loop, each corresponding to an iteration of the original
2175  /// scalar loop.
2176  void collectLoopUniforms(unsigned VF);
2177 
2178  /// Collect the instructions that are scalar after vectorization. An
2179  /// instruction is scalar if it is known to be uniform or will be scalarized
2180  /// during vectorization. Non-uniform scalarized instructions will be
2181  /// represented by VF values in the vectorized loop, each corresponding to an
2182  /// iteration of the original scalar loop.
2183  void collectLoopScalars(unsigned VF);
2184 
2185  /// Keeps cost model vectorization decision and cost for instructions.
2186  /// Right now it is used for memory instructions only.
2187  using DecisionList = DenseMap<std::pair<Instruction *, unsigned>,
2188  std::pair<InstWidening, unsigned>>;
2189 
2190  DecisionList WideningDecisions;
2191 
2192 public:
2193  /// The loop that we evaluate.
2194  Loop *TheLoop;
2195 
2196  /// Predicated scalar evolution analysis.
2198 
2199  /// Loop Info analysis.
2200  LoopInfo *LI;
2201 
2202  /// Vectorization legality.
2203  LoopVectorizationLegality *Legal;
2204 
2205  /// Vector target information.
2206  const TargetTransformInfo &TTI;
2207 
2208  /// Target Library Info.
2209  const TargetLibraryInfo *TLI;
2210 
2211  /// Demanded bits analysis.
2212  DemandedBits *DB;
2213 
2214  /// Assumption cache.
2216 
2217  /// Interface to emit optimization remarks.
2219 
2220  const Function *TheFunction;
2221 
2222  /// Loop Vectorize Hint.
2223  const LoopVectorizeHints *Hints;
2224 
2225  /// Values to ignore in the cost model.
2226  SmallPtrSet<const Value *, 16> ValuesToIgnore;
2227 
2228  /// Values to ignore in the cost model when VF > 1.
2229  SmallPtrSet<const Value *, 16> VecValuesToIgnore;
2230 };
2231 
2232 } // end anonymous namespace
2233 
2234 namespace llvm {
2235 
2236 /// InnerLoopVectorizer vectorizes loops which contain only one basic
2237 /// LoopVectorizationPlanner - drives the vectorization process after having
2238 /// passed Legality checks.
2239 /// The planner builds and optimizes the Vectorization Plans which record the
2240 /// decisions how to vectorize the given loop. In particular, represent the
2241 /// control-flow of the vectorized version, the replication of instructions that
2242 /// are to be scalarized, and interleave access groups.
2244  /// The loop that we evaluate.
2245  Loop *OrigLoop;
2246 
2247  /// Loop Info analysis.
2248  LoopInfo *LI;
2249 
2250  /// Target Library Info.
2251  const TargetLibraryInfo *TLI;
2252 
2253  /// Target Transform Info.
2254  const TargetTransformInfo *TTI;
2255 
2256  /// The legality analysis.
2257  LoopVectorizationLegality *Legal;
2258 
2259  /// The profitablity analysis.
2260  LoopVectorizationCostModel &CM;
2261 
2262  using VPlanPtr = std::unique_ptr<VPlan>;
2263 
2264  SmallVector<VPlanPtr, 4> VPlans;
2265 
2266  /// This class is used to enable the VPlan to invoke a method of ILV. This is
2267  /// needed until the method is refactored out of ILV and becomes reusable.
2268  struct VPCallbackILV : public VPCallback {
2269  InnerLoopVectorizer &ILV;
2270 
2271  VPCallbackILV(InnerLoopVectorizer &ILV) : ILV(ILV) {}
2272 
2273  Value *getOrCreateVectorValues(Value *V, unsigned Part) override {
2274  return ILV.getOrCreateVectorValue(V, Part);
2275  }
2276  };
2277 
2278  /// A builder used to construct the current plan.
2280 
2281  /// When we if-convert we need to create edge masks. We have to cache values
2282  /// so that we don't end up with exponential recursion/IR. Note that
2283  /// if-conversion currently takes place during VPlan-construction, so these
2284  /// caches are only used at that stage.
2285  using EdgeMaskCacheTy =
2288  EdgeMaskCacheTy EdgeMaskCache;
2289  BlockMaskCacheTy BlockMaskCache;
2290 
2291  unsigned BestVF = 0;
2292  unsigned BestUF = 0;
2293 
2294 public:
2296  const TargetTransformInfo *TTI,
2297  LoopVectorizationLegality *Legal,
2298  LoopVectorizationCostModel &CM)
2299  : OrigLoop(L), LI(LI), TLI(TLI), TTI(TTI), Legal(Legal), CM(CM) {}
2300 
2301  /// Plan how to best vectorize, return the best VF and its cost.
2303  unsigned UserVF);
2304 
2305  /// Finalize the best decision and dispose of all other VPlans.
2306  void setBestPlan(unsigned VF, unsigned UF);
2307 
2308  /// Generate the IR code for the body of the vectorized loop according to the
2309  /// best selected VPlan.
2310  void executePlan(InnerLoopVectorizer &LB, DominatorTree *DT);
2311 
2313  for (const auto &Plan : VPlans)
2314  O << *Plan;
2315  }
2316 
2317 protected:
2318  /// Collect the instructions from the original loop that would be trivially
2319  /// dead in the vectorized loop if generated.
2320  void collectTriviallyDeadInstructions(
2321  SmallPtrSetImpl<Instruction *> &DeadInstructions);
2322 
2323  /// A range of powers-of-2 vectorization factors with fixed start and
2324  /// adjustable end. The range includes start and excludes end, e.g.,:
2325  /// [1, 9) = {1, 2, 4, 8}
2326  struct VFRange {
2327  // A power of 2.
2328  const unsigned Start;
2329 
2330  // Need not be a power of 2. If End <= Start range is empty.
2331  unsigned End;
2332  };
2333 
2334  /// Test a \p Predicate on a \p Range of VF's. Return the value of applying
2335  /// \p Predicate on Range.Start, possibly decreasing Range.End such that the
2336  /// returned value holds for the entire \p Range.
2337  bool getDecisionAndClampRange(const std::function<bool(unsigned)> &Predicate,
2338  VFRange &Range);
2339 
2340  /// Build VPlans for power-of-2 VF's between \p MinVF and \p MaxVF inclusive,
2341  /// according to the information gathered by Legal when it checked if it is
2342  /// legal to vectorize the loop.
2343  void buildVPlans(unsigned MinVF, unsigned MaxVF);
2344 
2345 private:
2346  /// A helper function that computes the predicate of the block BB, assuming
2347  /// that the header block of the loop is set to True. It returns the *entry*
2348  /// mask for the block BB.
2349  VPValue *createBlockInMask(BasicBlock *BB, VPlanPtr &Plan);
2350 
2351  /// A helper function that computes the predicate of the edge between SRC
2352  /// and DST.
2353  VPValue *createEdgeMask(BasicBlock *Src, BasicBlock *Dst, VPlanPtr &Plan);
2354 
2355  /// Check if \I belongs to an Interleave Group within the given VF \p Range,
2356  /// \return true in the first returned value if so and false otherwise.
2357  /// Build a new VPInterleaveGroup Recipe if \I is the primary member of an IG
2358  /// for \p Range.Start, and provide it as the second returned value.
2359  /// Note that if \I is an adjunct member of an IG for \p Range.Start, the
2360  /// \return value is <true, nullptr>, as it is handled by another recipe.
2361  /// \p Range.End may be decreased to ensure same decision from \p Range.Start
2362  /// to \p Range.End.
2363  VPInterleaveRecipe *tryToInterleaveMemory(Instruction *I, VFRange &Range);
2364 
2365  // Check if \I is a memory instruction to be widened for \p Range.Start and
2366  // potentially masked. Such instructions are handled by a recipe that takes an
2367  // additional VPInstruction for the mask.
2368  VPWidenMemoryInstructionRecipe *tryToWidenMemory(Instruction *I,
2369  VFRange &Range,
2370  VPlanPtr &Plan);
2371 
2372  /// Check if an induction recipe should be constructed for \I within the given
2373  /// VF \p Range. If so build and return it. If not, return null. \p Range.End
2374  /// may be decreased to ensure same decision from \p Range.Start to
2375  /// \p Range.End.
2376  VPWidenIntOrFpInductionRecipe *tryToOptimizeInduction(Instruction *I,
2377  VFRange &Range);
2378 
2379  /// Handle non-loop phi nodes. Currently all such phi nodes are turned into
2380  /// a sequence of select instructions as the vectorizer currently performs
2381  /// full if-conversion.
2382  VPBlendRecipe *tryToBlend(Instruction *I, VPlanPtr &Plan);
2383 
2384  /// Check if \p I can be widened within the given VF \p Range. If \p I can be
2385  /// widened for \p Range.Start, check if the last recipe of \p VPBB can be
2386  /// extended to include \p I or else build a new VPWidenRecipe for it and
2387  /// append it to \p VPBB. Return true if \p I can be widened for Range.Start,
2388  /// false otherwise. Range.End may be decreased to ensure same decision from
2389  /// \p Range.Start to \p Range.End.
2390  bool tryToWiden(Instruction *I, VPBasicBlock *VPBB, VFRange &Range);
2391 
2392  /// Build a VPReplicationRecipe for \p I and enclose it within a Region if it
2393  /// is predicated. \return \p VPBB augmented with this new recipe if \p I is
2394  /// not predicated, otherwise \return a new VPBasicBlock that succeeds the new
2395  /// Region. Update the packing decision of predicated instructions if they
2396  /// feed \p I. Range.End may be decreased to ensure same recipe behavior from
2397  /// \p Range.Start to \p Range.End.
2398  VPBasicBlock *handleReplication(
2399  Instruction *I, VFRange &Range, VPBasicBlock *VPBB,
2401  VPlanPtr &Plan);
2402 
2403  /// Create a replicating region for instruction \p I that requires
2404  /// predication. \p PredRecipe is a VPReplicateRecipe holding \p I.
2405  VPRegionBlock *createReplicateRegion(Instruction *I, VPRecipeBase *PredRecipe,
2406  VPlanPtr &Plan);
2407 
2408  /// Build a VPlan according to the information gathered by Legal. \return a
2409  /// VPlan for vectorization factors \p Range.Start and up to \p Range.End
2410  /// exclusive, possibly decreasing \p Range.End.
2411  VPlanPtr buildVPlan(VFRange &Range,
2412  const SmallPtrSetImpl<Value *> &NeedDef);
2413 };
2414 
2415 } // end namespace llvm
2416 
2417 namespace {
2418 
2419 /// \brief This holds vectorization requirements that must be verified late in
2420 /// the process. The requirements are set by legalize and costmodel. Once
2421 /// vectorization has been determined to be possible and profitable the
2422 /// requirements can be verified by looking for metadata or compiler options.
2423 /// For example, some loops require FP commutativity which is only allowed if
2424 /// vectorization is explicitly specified or if the fast-math compiler option
2425 /// has been provided.
2426 /// Late evaluation of these requirements allows helpful diagnostics to be
2427 /// composed that tells the user what need to be done to vectorize the loop. For
2428 /// example, by specifying #pragma clang loop vectorize or -ffast-math. Late
2429 /// evaluation should be used only when diagnostics can generated that can be
2430 /// followed by a non-expert user.
2431 class LoopVectorizationRequirements {
2432 public:
2433  LoopVectorizationRequirements(OptimizationRemarkEmitter &ORE) : ORE(ORE) {}
2434 
2435  void addUnsafeAlgebraInst(Instruction *I) {
2436  // First unsafe algebra instruction.
2437  if (!UnsafeAlgebraInst)
2438  UnsafeAlgebraInst = I;
2439  }
2440 
2441  void addRuntimePointerChecks(unsigned Num) { NumRuntimePointerChecks = Num; }
2442 
2443  bool doesNotMeet(Function *F, Loop *L, const LoopVectorizeHints &Hints) {
2444  const char *PassName = Hints.vectorizeAnalysisPassName();
2445  bool Failed = false;
2446  if (UnsafeAlgebraInst && !Hints.allowReordering()) {
2447  ORE.emit([&]() {
2449  PassName, "CantReorderFPOps",
2450  UnsafeAlgebraInst->getDebugLoc(),
2451  UnsafeAlgebraInst->getParent())
2452  << "loop not vectorized: cannot prove it is safe to reorder "
2453  "floating-point operations";
2454  });
2455  Failed = true;
2456  }
2457 
2458  // Test if runtime memcheck thresholds are exceeded.
2459  bool PragmaThresholdReached =
2460  NumRuntimePointerChecks > PragmaVectorizeMemoryCheckThreshold;
2461  bool ThresholdReached =
2462  NumRuntimePointerChecks > VectorizerParams::RuntimeMemoryCheckThreshold;
2463  if ((ThresholdReached && !Hints.allowReordering()) ||
2464  PragmaThresholdReached) {
2465  ORE.emit([&]() {
2466  return OptimizationRemarkAnalysisAliasing(PassName, "CantReorderMemOps",
2467  L->getStartLoc(),
2468  L->getHeader())
2469  << "loop not vectorized: cannot prove it is safe to reorder "
2470  "memory operations";
2471  });
2472  DEBUG(dbgs() << "LV: Too many memory checks needed.\n");
2473  Failed = true;
2474  }
2475 
2476  return Failed;
2477  }
2478 
2479 private:
2480  unsigned NumRuntimePointerChecks = 0;
2481  Instruction *UnsafeAlgebraInst = nullptr;
2482 
2483  /// Interface to emit optimization remarks.
2485 };
2486 
2487 } // end anonymous namespace
2488 
2490  if (L.empty()) {
2491  if (!hasCyclesInLoopBody(L))
2492  V.push_back(&L);
2493  return;
2494  }
2495  for (Loop *InnerL : L)
2496  addAcyclicInnerLoop(*InnerL, V);
2497 }
2498 
2499 namespace {
2500 
2501 /// The LoopVectorize Pass.
2502 struct LoopVectorize : public FunctionPass {
2503  /// Pass identification, replacement for typeid
2504  static char ID;
2505 
2506  LoopVectorizePass Impl;
2507 
2508  explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
2509  : FunctionPass(ID) {
2510  Impl.DisableUnrolling = NoUnrolling;
2511  Impl.AlwaysVectorize = AlwaysVectorize;
2513  }
2514 
2515  bool runOnFunction(Function &F) override {
2516  if (skipFunction(F))
2517  return false;
2518 
2519  auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
2520  auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
2521  auto *TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
2522  auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2523  auto *BFI = &getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI();
2524  auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
2525  auto *TLI = TLIP ? &TLIP->getTLI() : nullptr;
2526  auto *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
2527  auto *AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
2528  auto *LAA = &getAnalysis<LoopAccessLegacyAnalysis>();
2529  auto *DB = &getAnalysis<DemandedBitsWrapperPass>().getDemandedBits();
2530  auto *ORE = &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE();
2531 
2532  std::function<const LoopAccessInfo &(Loop &)> GetLAA =
2533  [&](Loop &L) -> const LoopAccessInfo & { return LAA->getInfo(&L); };
2534 
2535  return Impl.runImpl(F, *SE, *LI, *TTI, *DT, *BFI, TLI, *DB, *AA, *AC,
2536  GetLAA, *ORE);
2537  }
2538 
2539  void getAnalysisUsage(AnalysisUsage &AU) const override {
2554  }
2555 };
2556 
2557 } // end anonymous namespace
2558 
2559 //===----------------------------------------------------------------------===//
2560 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
2561 // LoopVectorizationCostModel and LoopVectorizationPlanner.
2562 //===----------------------------------------------------------------------===//
2563 
2565  // We need to place the broadcast of invariant variables outside the loop.
2566  Instruction *Instr = dyn_cast<Instruction>(V);
2567  bool NewInstr = (Instr && Instr->getParent() == LoopVectorBody);
2568  bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
2569 
2570  // Place the code for broadcasting invariant variables in the new preheader.
2572  if (Invariant)
2574 
2575  // Broadcast the scalar into all locations in the vector.
2576  Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
2577 
2578  return Shuf;
2579 }
2580 
2582  const InductionDescriptor &II, Value *Step, Instruction *EntryVal) {
2583  Value *Start = II.getStartValue();
2584 
2585  // Construct the initial value of the vector IV in the vector loop preheader
2586  auto CurrIP = Builder.saveIP();
2588  if (isa<TruncInst>(EntryVal)) {
2589  assert(Start->getType()->isIntegerTy() &&
2590  "Truncation requires an integer type");
2591  auto *TruncType = cast<IntegerType>(EntryVal->getType());
2592  Step = Builder.CreateTrunc(Step, TruncType);
2593  Start = Builder.CreateCast(Instruction::Trunc, Start, TruncType);
2594  }
2595  Value *SplatStart = Builder.CreateVectorSplat(VF, Start);
2596  Value *SteppedStart =
2597  getStepVector(SplatStart, 0, Step, II.getInductionOpcode());
2598 
2599  // We create vector phi nodes for both integer and floating-point induction
2600  // variables. Here, we determine the kind of arithmetic we will perform.
2601  Instruction::BinaryOps AddOp;
2602  Instruction::BinaryOps MulOp;
2603  if (Step->getType()->isIntegerTy()) {
2604  AddOp = Instruction::Add;
2605  MulOp = Instruction::Mul;
2606  } else {
2607  AddOp = II.getInductionOpcode();
2608  MulOp = Instruction::FMul;
2609  }
2610 
2611  // Multiply the vectorization factor by the step using integer or
2612  // floating-point arithmetic as appropriate.
2613  Value *ConstVF = getSignedIntOrFpConstant(Step->getType(), VF);
2614  Value *Mul = addFastMathFlag(Builder.CreateBinOp(MulOp, Step, ConstVF));
2615 
2616  // Create a vector splat to use in the induction update.
2617  //
2618  // FIXME: If the step is non-constant, we create the vector splat with
2619  // IRBuilder. IRBuilder can constant-fold the multiply, but it doesn't
2620  // handle a constant vector splat.
2621  Value *SplatVF = isa<Constant>(Mul)
2622  ? ConstantVector::getSplat(VF, cast<Constant>(Mul))
2623  : Builder.CreateVectorSplat(VF, Mul);
2624  Builder.restoreIP(CurrIP);
2625 
2626  // We may need to add the step a number of times, depending on the unroll
2627  // factor. The last of those goes into the PHI.
2628  PHINode *VecInd = PHINode::Create(SteppedStart->getType(), 2, "vec.ind",
2630  Instruction *LastInduction = VecInd;
2631  for (unsigned Part = 0; Part < UF; ++Part) {
2632  VectorLoopValueMap.setVectorValue(EntryVal, Part, LastInduction);
2633  recordVectorLoopValueForInductionCast(II, LastInduction, Part);
2634  if (isa<TruncInst>(EntryVal))
2635  addMetadata(LastInduction, EntryVal);
2636  LastInduction = cast<Instruction>(addFastMathFlag(
2637  Builder.CreateBinOp(AddOp, LastInduction, SplatVF, "step.add")));
2638  }
2639 
2640  // Move the last step to the end of the latch block. This ensures consistent
2641  // placement of all induction updates.
2642  auto *LoopVectorLatch = LI->getLoopFor(LoopVectorBody)->getLoopLatch();
2643  auto *Br = cast<BranchInst>(LoopVectorLatch->getTerminator());
2644  auto *ICmp = cast<Instruction>(Br->getCondition());
2645  LastInduction->moveBefore(ICmp);
2646  LastInduction->setName("vec.ind.next");
2647 
2648  VecInd->addIncoming(SteppedStart, LoopVectorPreHeader);
2649  VecInd->addIncoming(LastInduction, LoopVectorLatch);
2650 }
2651 
2653  return Cost->isScalarAfterVectorization(I, VF) ||
2654  Cost->isProfitableToScalarize(I, VF);
2655 }
2656 
2659  return true;
2660  auto isScalarInst = [&](User *U) -> bool {
2661  auto *I = cast<Instruction>(U);
2662  return (OrigLoop->contains(I) && shouldScalarizeInstruction(I));
2663  };
2664  return llvm::any_of(IV->users(), isScalarInst);
2665 }
2666 
2668  const InductionDescriptor &ID, Value *VectorLoopVal, unsigned Part,
2669  unsigned Lane) {
2670  const SmallVectorImpl<Instruction *> &Casts = ID.getCastInsts();
2671  if (Casts.empty())
2672  return;
2673  // Only the first Cast instruction in the Casts vector is of interest.
2674  // The rest of the Casts (if exist) have no uses outside the
2675  // induction update chain itself.
2676  Instruction *CastInst = *Casts.begin();
2677  if (Lane < UINT_MAX)
2678  VectorLoopValueMap.setScalarValue(CastInst, {Part, Lane}, VectorLoopVal);
2679  else
2680  VectorLoopValueMap.setVectorValue(CastInst, Part, VectorLoopVal);
2681 }
2682 
2684  assert((IV->getType()->isIntegerTy() || IV != OldInduction) &&
2685  "Primary induction variable must have an integer type");
2686 
2687  auto II = Legal->getInductionVars()->find(IV);
2688  assert(II != Legal->getInductionVars()->end() && "IV is not an induction");
2689 
2690  auto ID = II->second;
2691  assert(IV->getType() == ID.getStartValue()->getType() && "Types must match");
2692 
2693  // The scalar value to broadcast. This will be derived from the canonical
2694  // induction variable.
2695  Value *ScalarIV = nullptr;
2696 
2697  // The value from the original loop to which we are mapping the new induction
2698  // variable.
2699  Instruction *EntryVal = Trunc ? cast<Instruction>(Trunc) : IV;
2700 
2701  // True if we have vectorized the induction variable.
2702  auto VectorizedIV = false;
2703 
2704  // Determine if we want a scalar version of the induction variable. This is
2705  // true if the induction variable itself is not widened, or if it has at
2706  // least one user in the loop that is not widened.
2707  auto NeedsScalarIV = VF > 1 && needsScalarInduction(EntryVal);
2708 
2709  // Generate code for the induction step. Note that induction steps are
2710  // required to be loop-invariant
2711  assert(PSE.getSE()->isLoopInvariant(ID.getStep(), OrigLoop) &&
2712  "Induction step should be loop invariant");
2713  auto &DL = OrigLoop->getHeader()->getModule()->getDataLayout();
2714  Value *Step = nullptr;
2715  if (PSE.getSE()->isSCEVable(IV->getType())) {
2716  SCEVExpander Exp(*PSE.getSE(), DL, "induction");
2717  Step = Exp.expandCodeFor(ID.getStep(), ID.getStep()->getType(),
2719  } else {
2720  Step = cast<SCEVUnknown>(ID.getStep())->getValue();
2721  }
2722 
2723  // Try to create a new independent vector induction variable. If we can't
2724  // create the phi node, we will splat the scalar induction variable in each
2725  // loop iteration.
2726  if (VF > 1 && !shouldScalarizeInstruction(EntryVal)) {
2727  createVectorIntOrFpInductionPHI(ID, Step, EntryVal);
2728  VectorizedIV = true;
2729  }
2730 
2731  // If we haven't yet vectorized the induction variable, or if we will create
2732  // a scalar one, we need to define the scalar induction variable and step
2733  // values. If we were given a truncation type, truncate the canonical
2734  // induction variable and step. Otherwise, derive these values from the
2735  // induction descriptor.
2736  if (!VectorizedIV || NeedsScalarIV) {
2737  ScalarIV = Induction;
2738  if (IV != OldInduction) {
2739  ScalarIV = IV->getType()->isIntegerTy()
2741  : Builder.CreateCast(Instruction::SIToFP, Induction,
2742  IV->getType());
2743  ScalarIV = ID.transform(Builder, ScalarIV, PSE.getSE(), DL);
2744  ScalarIV->setName("offset.idx");
2745  }
2746  if (Trunc) {
2747  auto *TruncType = cast<IntegerType>(Trunc->getType());
2748  assert(Step->getType()->isIntegerTy() &&
2749  "Truncation requires an integer step");
2750  ScalarIV = Builder.CreateTrunc(ScalarIV, TruncType);
2751  Step = Builder.CreateTrunc(Step, TruncType);
2752  }
2753  }
2754 
2755  // If we haven't yet vectorized the induction variable, splat the scalar
2756  // induction variable, and build the necessary step vectors.
2757  if (!VectorizedIV) {
2758  Value *Broadcasted = getBroadcastInstrs(ScalarIV);
2759  for (unsigned Part = 0; Part < UF; ++Part) {
2760  Value *EntryPart =
2761  getStepVector(Broadcasted, VF * Part, Step, ID.getInductionOpcode());
2762  VectorLoopValueMap.setVectorValue(EntryVal, Part, EntryPart);
2763  recordVectorLoopValueForInductionCast(ID, EntryPart, Part);
2764  if (Trunc)
2765  addMetadata(EntryPart, Trunc);
2766  }
2767  }
2768 
2769  // If an induction variable is only used for counting loop iterations or
2770  // calculating addresses, it doesn't need to be widened. Create scalar steps
2771  // that can be used by instructions we will later scalarize. Note that the
2772  // addition of the scalar steps will not increase the number of instructions
2773  // in the loop in the common case prior to InstCombine. We will be trading
2774  // one vector extract for each scalar step.
2775  if (NeedsScalarIV)
2776  buildScalarSteps(ScalarIV, Step, EntryVal, ID);
2777 }
2778 
2780  Instruction::BinaryOps BinOp) {
2781  // Create and check the types.
2782  assert(Val->getType()->isVectorTy() && "Must be a vector");
2783  int VLen = Val->getType()->getVectorNumElements();
2784 
2785  Type *STy = Val->getType()->getScalarType();
2786  assert((STy->isIntegerTy() || STy->isFloatingPointTy()) &&
2787  "Induction Step must be an integer or FP");
2788  assert(Step->getType() == STy && "Step has wrong type");
2789 
2791 
2792  if (STy->isIntegerTy()) {
2793  // Create a vector of consecutive numbers from zero to VF.
2794  for (int i = 0; i < VLen; ++i)
2795  Indices.push_back(ConstantInt::get(STy, StartIdx + i));
2796 
2797  // Add the consecutive indices to the vector value.
2798  Constant *Cv = ConstantVector::get(Indices);
2799  assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
2800  Step = Builder.CreateVectorSplat(VLen, Step);
2801  assert(Step->getType() == Val->getType() && "Invalid step vec");
2802  // FIXME: The newly created binary instructions should contain nsw/nuw flags,
2803  // which can be found from the original scalar operations.
2804  Step = Builder.CreateMul(Cv, Step);
2805  return Builder.CreateAdd(Val, Step, "induction");
2806  }
2807 
2808  // Floating point induction.
2809  assert((BinOp == Instruction::FAdd || BinOp == Instruction::FSub) &&
2810  "Binary Opcode should be specified for FP induction");
2811  // Create a vector of consecutive numbers from zero to VF.
2812  for (int i = 0; i < VLen; ++i)
2813  Indices.push_back(ConstantFP::get(STy, (double)(StartIdx + i)));
2814 
2815  // Add the consecutive indices to the vector value.
2816  Constant *Cv = ConstantVector::get(Indices);
2817 
2818  Step = Builder.CreateVectorSplat(VLen, Step);
2819 
2820  // Floating point operations had to be 'fast' to enable the induction.
2821  FastMathFlags Flags;
2822  Flags.setFast();
2823 
2824  Value *MulOp = Builder.CreateFMul(Cv, Step);
2825  if (isa<Instruction>(MulOp))
2826  // Have to check, MulOp may be a constant
2827  cast<Instruction>(MulOp)->setFastMathFlags(Flags);
2828 
2829  Value *BOp = Builder.CreateBinOp(BinOp, Val, MulOp, "induction");
2830  if (isa<Instruction>(BOp))
2831  cast<Instruction>(BOp)->setFastMathFlags(Flags);
2832  return BOp;
2833 }
2834 
2836  Value *EntryVal,
2837  const InductionDescriptor &ID) {
2838  // We shouldn't have to build scalar steps if we aren't vectorizing.
2839  assert(VF > 1 && "VF should be greater than one");
2840 
2841  // Get the value type and ensure it and the step have the same integer type.
2842  Type *ScalarIVTy = ScalarIV->getType()->getScalarType();
2843  assert(ScalarIVTy == Step->getType() &&
2844  "Val and Step should have the same type");
2845 
2846  // We build scalar steps for both integer and floating-point induction
2847  // variables. Here, we determine the kind of arithmetic we will perform.
2848  Instruction::BinaryOps AddOp;
2849  Instruction::BinaryOps MulOp;
2850  if (ScalarIVTy->isIntegerTy()) {
2851  AddOp = Instruction::Add;
2852  MulOp = Instruction::Mul;
2853  } else {
2854  AddOp = ID.getInductionOpcode();
2855  MulOp = Instruction::FMul;
2856  }
2857 
2858  // Determine the number of scalars we need to generate for each unroll
2859  // iteration. If EntryVal is uniform, we only need to generate the first
2860  // lane. Otherwise, we generate all VF values.
2861  unsigned Lanes =
2862  Cost->isUniformAfterVectorization(cast<Instruction>(EntryVal), VF) ? 1
2863  : VF;
2864  // Compute the scalar steps and save the results in VectorLoopValueMap.
2865  for (unsigned Part = 0; Part < UF; ++Part) {
2866  for (unsigned Lane = 0; Lane < Lanes; ++Lane) {
2867  auto *StartIdx = getSignedIntOrFpConstant(ScalarIVTy, VF * Part + Lane);
2868  auto *Mul = addFastMathFlag(Builder.CreateBinOp(MulOp, StartIdx, Step));
2869  auto *Add = addFastMathFlag(Builder.CreateBinOp(AddOp, ScalarIV, Mul));
2870  VectorLoopValueMap.setScalarValue(EntryVal, {Part, Lane}, Add);
2871  recordVectorLoopValueForInductionCast(ID, Add, Part, Lane);
2872  }
2873  }
2874 }
2875 
2876 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
2877  const ValueToValueMap &Strides = getSymbolicStrides() ? *getSymbolicStrides() :
2878  ValueToValueMap();
2879 
2880  int Stride = getPtrStride(PSE, Ptr, TheLoop, Strides, true, false);
2881  if (Stride == 1 || Stride == -1)
2882  return Stride;
2883  return 0;
2884 }
2885 
2886 bool LoopVectorizationLegality::isUniform(Value *V) {
2887  return LAI->isUniform(V);
2888 }
2889 
2891  assert(V != Induction && "The new induction variable should not be used.");
2892  assert(!V->getType()->isVectorTy() && "Can't widen a vector");
2893  assert(!V->getType()->isVoidTy() && "Type does not produce a value");
2894 
2895  // If we have a stride that is replaced by one, do it here.
2896  if (Legal->hasStride(V))
2897  V = ConstantInt::get(V->getType(), 1);
2898 
2899  // If we have a vector mapped to this value, return it.
2900  if (VectorLoopValueMap.hasVectorValue(V, Part))
2901  return VectorLoopValueMap.getVectorValue(V, Part);
2902 
2903  // If the value has not been vectorized, check if it has been scalarized
2904  // instead. If it has been scalarized, and we actually need the value in
2905  // vector form, we will construct the vector values on demand.
2907  Value *ScalarValue = VectorLoopValueMap.getScalarValue(V, {Part, 0});
2908 
2909  // If we've scalarized a value, that value should be an instruction.
2910  auto *I = cast<Instruction>(V);
2911 
2912  // If we aren't vectorizing, we can just copy the scalar map values over to
2913  // the vector map.
2914  if (VF == 1) {
2915  VectorLoopValueMap.setVectorValue(V, Part, ScalarValue);
2916  return ScalarValue;
2917  }
2918 
2919  // Get the last scalar instruction we generated for V and Part. If the value
2920  // is known to be uniform after vectorization, this corresponds to lane zero
2921  // of the Part unroll iteration. Otherwise, the last instruction is the one
2922  // we created for the last vector lane of the Part unroll iteration.
2923  unsigned LastLane = Cost->isUniformAfterVectorization(I, VF) ? 0 : VF - 1;
2924  auto *LastInst = cast<Instruction>(
2925  VectorLoopValueMap.getScalarValue(V, {Part, LastLane}));
2926 
2927  // Set the insert point after the last scalarized instruction. This ensures
2928  // the insertelement sequence will directly follow the scalar definitions.
2929  auto OldIP = Builder.saveIP();
2930  auto NewIP = std::next(BasicBlock::iterator(LastInst));
2931  Builder.SetInsertPoint(&*NewIP);
2932 
2933  // However, if we are vectorizing, we need to construct the vector values.
2934  // If the value is known to be uniform after vectorization, we can just
2935  // broadcast the scalar value corresponding to lane zero for each unroll
2936  // iteration. Otherwise, we construct the vector values using insertelement
2937  // instructions. Since the resulting vectors are stored in
2938  // VectorLoopValueMap, we will only generate the insertelements once.
2939  Value *VectorValue = nullptr;
2940  if (Cost->isUniformAfterVectorization(I, VF)) {
2941  VectorValue = getBroadcastInstrs(ScalarValue);
2942  VectorLoopValueMap.setVectorValue(V, Part, VectorValue);
2943  } else {
2944  // Initialize packing with insertelements to start from undef.
2946  VectorLoopValueMap.setVectorValue(V, Part, Undef);
2947  for (unsigned Lane = 0; Lane < VF; ++Lane)
2948  packScalarIntoVectorValue(V, {Part, Lane});
2949  VectorValue = VectorLoopValueMap.getVectorValue(V, Part);
2950  }
2951  Builder.restoreIP(OldIP);
2952  return VectorValue;
2953  }
2954 
2955  // If this scalar is unknown, assume that it is a constant or that it is
2956  // loop invariant. Broadcast V and save the value for future uses.
2957  Value *B = getBroadcastInstrs(V);
2958  VectorLoopValueMap.setVectorValue(V, Part, B);
2959  return B;
2960 }
2961 
2962 Value *
2964  const VPIteration &Instance) {
2965  // If the value is not an instruction contained in the loop, it should
2966  // already be scalar.
2967  if (OrigLoop->isLoopInvariant(V))
2968  return V;
2969 
2970  assert(Instance.Lane > 0
2971  ? !Cost->isUniformAfterVectorization(cast<Instruction>(V), VF)
2972  : true && "Uniform values only have lane zero");
2973 
2974  // If the value from the original loop has not been vectorized, it is
2975  // represented by UF x VF scalar values in the new loop. Return the requested
2976  // scalar value.
2977  if (VectorLoopValueMap.hasScalarValue(V, Instance))
2978  return VectorLoopValueMap.getScalarValue(V, Instance);
2979 
2980  // If the value has not been scalarized, get its entry in VectorLoopValueMap
2981  // for the given unroll part. If this entry is not a vector type (i.e., the
2982  // vectorization factor is one), there is no need to generate an
2983  // extractelement instruction.
2984  auto *U = getOrCreateVectorValue(V, Instance.Part);
2985  if (!U->getType()->isVectorTy()) {
2986  assert(VF == 1 && "Value not scalarized has non-vector type");
2987  return U;
2988  }
2989 
2990  // Otherwise, the value from the original loop has been vectorized and is
2991  // represented by UF vector values. Extract and return the requested scalar
2992  // value from the appropriate vector lane.
2993  return Builder.CreateExtractElement(U, Builder.getInt32(Instance.Lane));
2994 }
2995 
2997  Value *V, const VPIteration &Instance) {
2998  assert(V != Induction && "The new induction variable should not be used.");
2999  assert(!V->getType()->isVectorTy() && "Can't pack a vector");
3000  assert(!V->getType()->isVoidTy() && "Type does not produce a value");
3001 
3002  Value *ScalarInst = VectorLoopValueMap.getScalarValue(V, Instance);
3003  Value *VectorValue = VectorLoopValueMap.getVectorValue(V, Instance.Part);
3004  VectorValue = Builder.CreateInsertElement(VectorValue, ScalarInst,
3005  Builder.getInt32(Instance.Lane));
3006  VectorLoopValueMap.resetVectorValue(V, Instance.Part, VectorValue);
3007 }
3008 
3010  assert(Vec->getType()->isVectorTy() && "Invalid type");
3012  for (unsigned i = 0; i < VF; ++i)
3013  ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
3014 
3015  return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
3016  ConstantVector::get(ShuffleMask),
3017  "reverse");
3018 }
3019 
3020 // Try to vectorize the interleave group that \p Instr belongs to.
3021 //
3022 // E.g. Translate following interleaved load group (factor = 3):
3023 // for (i = 0; i < N; i+=3) {
3024 // R = Pic[i]; // Member of index 0
3025 // G = Pic[i+1]; // Member of index 1
3026 // B = Pic[i+2]; // Member of index 2
3027 // ... // do something to R, G, B
3028 // }
3029 // To:
3030 // %wide.vec = load <12 x i32> ; Read 4 tuples of R,G,B
3031 // %R.vec = shuffle %wide.vec, undef, <0, 3, 6, 9> ; R elements
3032 // %G.vec = shuffle %wide.vec, undef, <1, 4, 7, 10> ; G elements
3033 // %B.vec = shuffle %wide.vec, undef, <2, 5, 8, 11> ; B elements
3034 //
3035 // Or translate following interleaved store group (factor = 3):
3036 // for (i = 0; i < N; i+=3) {
3037 // ... do something to R, G, B
3038 // Pic[i] = R; // Member of index 0
3039 // Pic[i+1] = G; // Member of index 1
3040 // Pic[i+2] = B; // Member of index 2
3041 // }
3042 // To:
3043 // %R_G.vec = shuffle %R.vec, %G.vec, <0, 1, 2, ..., 7>
3044 // %B_U.vec = shuffle %B.vec, undef, <0, 1, 2, 3, u, u, u, u>
3045 // %interleaved.vec = shuffle %R_G.vec, %B_U.vec,
3046 // <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11> ; Interleave R,G,B elements
3047 // store <12 x i32> %interleaved.vec ; Write 4 tuples of R,G,B
3049  const InterleaveGroup *Group = Legal->getInterleavedAccessGroup(Instr);
3050  assert(Group && "Fail to get an interleaved access group.");
3051 
3052  // Skip if current instruction is not the insert position.
3053  if (Instr != Group->getInsertPos())
3054  return;
3055 
3056  const DataLayout &DL = Instr->getModule()->getDataLayout();
3057  Value *Ptr = getPointerOperand(Instr);
3058 
3059  // Prepare for the vector type of the interleaved load/store.
3060  Type *ScalarTy = getMemInstValueType(Instr);
3061  unsigned InterleaveFactor = Group->getFactor();
3062  Type *VecTy = VectorType::get(ScalarTy, InterleaveFactor * VF);
3063  Type *PtrTy = VecTy->getPointerTo(getMemInstAddressSpace(Instr));
3064 
3065  // Prepare for the new pointers.
3067  SmallVector<Value *, 2> NewPtrs;
3068  unsigned Index = Group->getIndex(Instr);
3069 
3070  // If the group is reverse, adjust the index to refer to the last vector lane
3071  // instead of the first. We adjust the index from the first vector lane,
3072  // rather than directly getting the pointer for lane VF - 1, because the
3073  // pointer operand of the interleaved access is supposed to be uniform. For
3074  // uniform instructions, we're only required to generate a value for the
3075  // first vector lane in each unroll iteration.
3076  if (Group->isReverse())
3077  Index += (VF - 1) * Group->getFactor();
3078 
3079  for (unsigned Part = 0; Part < UF; Part++) {
3080  Value *NewPtr = getOrCreateScalarValue(Ptr, {Part, 0});
3081 
3082  // Notice current instruction could be any index. Need to adjust the address
3083  // to the member of index 0.
3084  //
3085  // E.g. a = A[i+1]; // Member of index 1 (Current instruction)
3086  // b = A[i]; // Member of index 0
3087  // Current pointer is pointed to A[i+1], adjust it to A[i].
3088  //
3089  // E.g. A[i+1] = a; // Member of index 1
3090  // A[i] = b; // Member of index 0
3091  // A[i+2] = c; // Member of index 2 (Current instruction)
3092  // Current pointer is pointed to A[i+2], adjust it to A[i].
3093  NewPtr = Builder.CreateGEP(NewPtr, Builder.getInt32(-Index));
3094 
3095  // Cast to the vector pointer type.
3096  NewPtrs.push_back(Builder.CreateBitCast(NewPtr, PtrTy));
3097  }
3098 
3099  setDebugLocFromInst(Builder, Instr);
3100  Value *UndefVec = UndefValue::get(VecTy);
3101 
3102  // Vectorize the interleaved load group.
3103  if (isa<LoadInst>(Instr)) {
3104  // For each unroll part, create a wide load for the group.
3105  SmallVector<Value *, 2> NewLoads;
3106  for (unsigned Part = 0; Part < UF; Part++) {
3107  auto *NewLoad = Builder.CreateAlignedLoad(
3108  NewPtrs[Part], Group->getAlignment(), "wide.vec");
3109  Group->addMetadata(NewLoad);
3110  NewLoads.push_back(NewLoad);
3111  }
3112 
3113  // For each member in the group, shuffle out the appropriate data from the
3114  // wide loads.
3115  for (unsigned I = 0; I < InterleaveFactor; ++I) {
3116  Instruction *Member = Group->getMember(I);
3117 
3118  // Skip the gaps in the group.
3119  if (!Member)
3120  continue;
3121 
3122  Constant *StrideMask = createStrideMask(Builder, I, InterleaveFactor, VF);
3123  for (unsigned Part = 0; Part < UF; Part++) {
3124  Value *StridedVec = Builder.CreateShuffleVector(
3125  NewLoads[Part], UndefVec, StrideMask, "strided.vec");
3126 
3127  // If this member has different type, cast the result type.
3128  if (Member->getType() != ScalarTy) {
3129  VectorType *OtherVTy = VectorType::get(Member->getType(), VF);
3130  StridedVec = createBitOrPointerCast(StridedVec, OtherVTy, DL);
3131  }
3132 
3133  if (Group->isReverse())
3134  StridedVec = reverseVector(StridedVec);
3135 
3136  VectorLoopValueMap.setVectorValue(Member, Part, StridedVec);
3137  }
3138  }
3139  return;
3140  }
3141 
3142  // The sub vector type for current instruction.
3143  VectorType *SubVT = VectorType::get(ScalarTy, VF);
3144 
3145  // Vectorize the interleaved store group.
3146  for (unsigned Part = 0; Part < UF; Part++) {
3147  // Collect the stored vector from each member.
3148  SmallVector<Value *, 4> StoredVecs;
3149  for (unsigned i = 0; i < InterleaveFactor; i++) {
3150  // Interleaved store group doesn't allow a gap, so each index has a member
3151  Instruction *Member = Group->getMember(i);
3152  assert(Member && "Fail to get a member from an interleaved store group");
3153 
3154  Value *StoredVec = getOrCreateVectorValue(
3155  cast<StoreInst>(Member)->getValueOperand(), Part);
3156  if (Group->isReverse())
3157  StoredVec = reverseVector(StoredVec);
3158 
3159  // If this member has different type, cast it to a unified type.
3160 
3161  if (StoredVec->getType() != SubVT)
3162  StoredVec = createBitOrPointerCast(StoredVec, SubVT, DL);
3163 
3164  StoredVecs.push_back(StoredVec);
3165  }
3166 
3167  // Concatenate all vectors into a wide vector.
3168  Value *WideVec = concatenateVectors(Builder, StoredVecs);
3169 
3170  // Interleave the elements in the wide vector.
3171  Constant *IMask = createInterleaveMask(Builder, VF, InterleaveFactor);
3172  Value *IVec = Builder.CreateShuffleVector(WideVec, UndefVec, IMask,
3173  "interleaved.vec");
3174 
3175  Instruction *NewStoreInstr =
3176  Builder.CreateAlignedStore(IVec, NewPtrs[Part], Group->getAlignment());
3177 
3178  Group->addMetadata(NewStoreInstr);
3179  }
3180 }
3181 
3183  VectorParts *BlockInMask) {
3184  // Attempt to issue a wide load.
3185  LoadInst *LI = dyn_cast<LoadInst>(Instr);
3186  StoreInst *SI = dyn_cast<StoreInst>(Instr);
3187 
3188  assert((LI || SI) && "Invalid Load/Store instruction");
3189 
3190  LoopVectorizationCostModel::InstWidening Decision =
3191  Cost->getWideningDecision(Instr, VF);
3192  assert(Decision != LoopVectorizationCostModel::CM_Unknown &&
3193  "CM decision should be taken at this point");
3194  if (Decision == LoopVectorizationCostModel::CM_Interleave)
3195  return vectorizeInterleaveGroup(Instr);
3196 
3197  Type *ScalarDataTy = getMemInstValueType(Instr);
3198  Type *DataTy = VectorType::get(ScalarDataTy, VF);
3199  Value *Ptr = getPointerOperand(Instr);
3200  unsigned Alignment = getMemInstAlignment(Instr);
3201  // An alignment of 0 means target abi alignment. We need to use the scalar's
3202  // target abi alignment in such a case.
3203  const DataLayout &DL = Instr->getModule()->getDataLayout();
3204  if (!Alignment)
3205  Alignment = DL.getABITypeAlignment(ScalarDataTy);
3206  unsigned AddressSpace = getMemInstAddressSpace(Instr);
3207 
3208  // Determine if the pointer operand of the access is either consecutive or
3209  // reverse consecutive.
3210  bool Reverse = (Decision == LoopVectorizationCostModel::CM_Widen_Reverse);
3211  bool ConsecutiveStride =
3212  Reverse || (Decision == LoopVectorizationCostModel::CM_Widen);
3213  bool CreateGatherScatter =
3214  (Decision == LoopVectorizationCostModel::CM_GatherScatter);
3215 
3216  // Either Ptr feeds a vector load/store, or a vector GEP should feed a vector
3217  // gather/scatter. Otherwise Decision should have been to Scalarize.
3218  assert((ConsecutiveStride || CreateGatherScatter) &&
3219  "The instruction should be scalarized");
3220 
3221  // Handle consecutive loads/stores.
3222  if (ConsecutiveStride)
3223  Ptr = getOrCreateScalarValue(Ptr, {0, 0});
3224 
3225  VectorParts Mask;
3226  bool isMaskRequired = BlockInMask;
3227  if (isMaskRequired)
3228  Mask = *BlockInMask;
3229 
3230  // Handle Stores:
3231  if (SI) {
3232  assert(!Legal->isUniform(SI->getPointerOperand()) &&
3233  "We do not allow storing to uniform addresses");
3235 
3236  for (unsigned Part = 0; Part < UF; ++Part) {
3237  Instruction *NewSI = nullptr;
3238  Value *StoredVal = getOrCreateVectorValue(SI->getValueOperand(), Part);
3239  if (CreateGatherScatter) {
3240  Value *MaskPart = isMaskRequired ? Mask[Part] : nullptr;
3241  Value *VectorGep = getOrCreateVectorValue(Ptr, Part);
3242  NewSI = Builder.CreateMaskedScatter(StoredVal, VectorGep, Alignment,
3243  MaskPart);
3244  } else {
3245  // Calculate the pointer for the specific unroll-part.
3246  Value *PartPtr =
3247  Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF));
3248 
3249  if (Reverse) {
3250  // If we store to reverse consecutive memory locations, then we need
3251  // to reverse the order of elements in the stored value.
3252  StoredVal = reverseVector(StoredVal);
3253  // We don't want to update the value in the map as it might be used in
3254  // another expression. So don't call resetVectorValue(StoredVal).
3255 
3256  // If the address is consecutive but reversed, then the
3257  // wide store needs to start at the last vector element.
3258  PartPtr =
3259  Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF));
3260  PartPtr =
3261  Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF));
3262  if (isMaskRequired) // Reverse of a null all-one mask is a null mask.
3263  Mask[Part] = reverseVector(Mask[Part]);
3264  }
3265 
3266  Value *VecPtr =
3267  Builder.CreateBitCast(PartPtr, DataTy->getPointerTo(AddressSpace));
3268 
3269  if (isMaskRequired)
3270  NewSI = Builder.CreateMaskedStore(StoredVal, VecPtr, Alignment,
3271  Mask[Part]);
3272  else
3273  NewSI = Builder.CreateAlignedStore(StoredVal, VecPtr, Alignment);
3274  }
3275  addMetadata(NewSI, SI);
3276  }
3277  return;
3278  }
3279 
3280  // Handle loads.
3281  assert(LI && "Must have a load instruction");
3283  for (unsigned Part = 0; Part < UF; ++Part) {
3284  Value *NewLI;
3285  if (CreateGatherScatter) {
3286  Value *MaskPart = isMaskRequired ? Mask[Part] : nullptr;
3287  Value *VectorGep = getOrCreateVectorValue(Ptr, Part);
3288  NewLI = Builder.CreateMaskedGather(VectorGep, Alignment, MaskPart,
3289  nullptr, "wide.masked.gather");
3290  addMetadata(NewLI, LI);
3291  } else {
3292  // Calculate the pointer for the specific unroll-part.
3293  Value *PartPtr =
3294  Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF));
3295 
3296  if (Reverse) {
3297  // If the address is consecutive but reversed, then the
3298  // wide load needs to start at the last vector element.
3299  PartPtr = Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(-Part * VF));
3300  PartPtr = Builder.CreateGEP(nullptr, PartPtr, Builder.getInt32(1 - VF));
3301  if (isMaskRequired) // Reverse of a null all-one mask is a null mask.
3302  Mask[Part] = reverseVector(Mask[Part]);
3303  }
3304 
3305  Value *VecPtr =
3306  Builder.CreateBitCast(PartPtr, DataTy->getPointerTo(AddressSpace));
3307  if (isMaskRequired)
3308  NewLI = Builder.CreateMaskedLoad(VecPtr, Alignment, Mask[Part],
3309  UndefValue::get(DataTy),
3310  "wide.masked.load");
3311  else
3312  NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
3313 
3314  // Add metadata to the load, but setVectorValue to the reverse shuffle.
3315  addMetadata(NewLI, LI);
3316  if (Reverse)
3317  NewLI = reverseVector(NewLI);
3318  }
3319  VectorLoopValueMap.setVectorValue(Instr, Part, NewLI);
3320  }
3321 }
3322 
3324  const VPIteration &Instance,
3325  bool IfPredicateInstr) {
3326  assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
3327 
3328  setDebugLocFromInst(Builder, Instr);
3329 
3330  // Does this instruction return a value ?
3331  bool IsVoidRetTy = Instr->getType()->isVoidTy();
3332 
3333  Instruction *Cloned = Instr->clone();
3334  if (!IsVoidRetTy)
3335  Cloned->setName(Instr->getName() + ".cloned");
3336 
3337  // Replace the operands of the cloned instructions with their scalar
3338  // equivalents in the new loop.
3339  for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
3340  auto *NewOp = getOrCreateScalarValue(Instr->getOperand(op), Instance);
3341  Cloned->setOperand(op, NewOp);
3342  }
3343  addNewMetadata(Cloned, Instr);
3344 
3345  // Place the cloned scalar in the new loop.
3346  Builder.Insert(Cloned);
3347 
3348  // Add the cloned scalar to the scalar map entry.
3349  VectorLoopValueMap.setScalarValue(Instr, Instance, Cloned);
3350 
3351  // If we just cloned a new assumption, add it the assumption cache.
3352  if (auto *II = dyn_cast<IntrinsicInst>(Cloned))
3353  if (II->getIntrinsicID() == Intrinsic::assume)
3354  AC->registerAssumption(II);
3355 
3356  // End if-block.
3357  if (IfPredicateInstr)
3358  PredicatedInstructions.push_back(Cloned);
3359 }
3360 
3362  Value *End, Value *Step,
3363  Instruction *DL) {
3364  BasicBlock *Header = L->getHeader();
3365  BasicBlock *Latch = L->getLoopLatch();
3366  // As we're just creating this loop, it's possible no latch exists
3367  // yet. If so, use the header as this will be a single block loop.
3368  if (!Latch)
3369  Latch = Header;
3370 
3373  setDebugLocFromInst(Builder, OldInst);
3374  auto *Induction = Builder.CreatePHI(Start->getType(), 2, "index");
3375 
3377  setDebugLocFromInst(Builder, OldInst);
3378 
3379  // Create i+1 and fill the PHINode.
3380  Value *Next = Builder.CreateAdd(Induction, Step, "index.next");
3381  Induction->addIncoming(Start, L->getLoopPreheader());
3382  Induction->addIncoming(Next, Latch);
3383  // Create the compare.
3384  Value *ICmp = Builder.CreateICmpEQ(Next, End);
3385  Builder.CreateCondBr(ICmp, L->getExitBlock(), Header);
3386 
3387  // Now we have two terminators. Remove the old one from the block.
3388  Latch->getTerminator()->eraseFromParent();
3389 
3390  return Induction;
3391 }
3392 
3394  if (TripCount)
3395  return TripCount;
3396 
3398  // Find the loop boundaries.
3399  ScalarEvolution *SE = PSE.getSE();
3400  const SCEV *BackedgeTakenCount = PSE.getBackedgeTakenCount();
3401  assert(BackedgeTakenCount != SE->getCouldNotCompute() &&
3402  "Invalid loop count");
3403 
3404  Type *IdxTy = Legal->getWidestInductionType();
3405 
3406  // The exit count might have the type of i64 while the phi is i32. This can
3407  // happen if we have an induction variable that is sign extended before the
3408  // compare. The only way that we get a backedge taken count is that the
3409  // induction variable was signed and as such will not overflow. In such a case
3410  // truncation is legal.
3411  if (BackedgeTakenCount->getType()->getPrimitiveSizeInBits() >
3412  IdxTy->getPrimitiveSizeInBits())
3413  BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount, IdxTy);
3414  BackedgeTakenCount = SE->getNoopOrZeroExtend(BackedgeTakenCount, IdxTy);
3415 
3416  // Get the total trip count from the count by adding 1.
3417  const SCEV *ExitCount = SE->getAddExpr(
3418  BackedgeTakenCount, SE->getOne(BackedgeTakenCount->getType()));
3419 
3420  const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
3421 
3422  // Expand the trip count and place the new instructions in the preheader.
3423  // Notice that the pre-header does not change, only the loop body.
3424  SCEVExpander Exp(*SE, DL, "induction");
3425 
3426  // Count holds the overall loop count (N).
3427  TripCount = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
3429 
3430  if (TripCount->getType()->isPointerTy())
3431  TripCount =
3432  CastInst::CreatePointerCast(TripCount, IdxTy, "exitcount.ptrcnt.to.int",
3434 
3435  return TripCount;
3436 }
3437 
3439  if (VectorTripCount)
3440  return VectorTripCount;
3441 
3442  Value *TC = getOrCreateTripCount(L);
3444 
3445  // Now we need to generate the expression for the part of the loop that the
3446  // vectorized body will execute. This is equal to N - (N % Step) if scalar
3447  // iterations are not required for correctness, or N - Step, otherwise. Step
3448  // is equal to the vectorization factor (number of SIMD elements) times the
3449  // unroll factor (number of SIMD instructions).
3450  Constant *Step = ConstantInt::get(TC->getType(), VF * UF);
3451  Value *R = Builder.CreateURem(TC, Step, "n.mod.vf");
3452 
3453  // If there is a non-reversed interleaved group that may speculatively access
3454  // memory out-of-bounds, we need to ensure that there will be at least one
3455  // iteration of the scalar epilogue loop. Thus, if the step evenly divides
3456  // the trip count, we set the remainder to be equal to the step. If the step
3457  // does not evenly divide the trip count, no adjustment is necessary since
3458  // there will already be scalar iterations. Note that the minimum iterations
3459  // check ensures that N >= Step.
3460  if (VF > 1 && Legal->requiresScalarEpilogue()) {
3461  auto *IsZero = Builder.CreateICmpEQ(R, ConstantInt::get(R->getType(), 0));
3462  R = Builder.CreateSelect(IsZero, Step, R);
3463  }
3464 
3465  VectorTripCount = Builder.CreateSub(TC, R, "n.vec");
3466 
3467  return VectorTripCount;
3468 }
3469 
3471  const DataLayout &DL) {
3472  // Verify that V is a vector type with same number of elements as DstVTy.
3473  unsigned VF = DstVTy->getNumElements();
3474  VectorType *SrcVecTy = cast<VectorType>(V->getType());
3475  assert((VF == SrcVecTy->getNumElements()) && "Vector dimensions do not match");
3476  Type *SrcElemTy = SrcVecTy->getElementType();
3477  Type *DstElemTy = DstVTy->getElementType();
3478  assert((DL.getTypeSizeInBits(SrcElemTy) == DL.getTypeSizeInBits(DstElemTy)) &&
3479  "Vector elements must have same size");
3480 
3481  // Do a direct cast if element types are castable.
3482  if (CastInst::isBitOrNoopPointerCastable(SrcElemTy, DstElemTy, DL)) {
3483  return Builder.CreateBitOrPointerCast(V, DstVTy);
3484  }
3485  // V cannot be directly casted to desired vector type.
3486  // May happen when V is a floating point vector but DstVTy is a vector of
3487  // pointers or vice-versa. Handle this using a two-step bitcast using an
3488  // intermediate Integer type for the bitcast i.e. Ptr <-> Int <-> Float.
3489  assert((DstElemTy->isPointerTy() != SrcElemTy->isPointerTy()) &&
3490  "Only one type should be a pointer type");
3491  assert((DstElemTy->isFloatingPointTy() != SrcElemTy->isFloatingPointTy()) &&
3492  "Only one type should be a floating point type");
3493  Type *IntTy =
3495  VectorType *VecIntTy = VectorType::get(IntTy, VF);
3496  Value *CastVal = Builder.CreateBitOrPointerCast(V, VecIntTy);
3497  return Builder.CreateBitOrPointerCast(CastVal, DstVTy);
3498 }
3499 
3501  BasicBlock *Bypass) {
3502  Value *Count = getOrCreateTripCount(L);
3503  BasicBlock *BB = L->getLoopPreheader();
3505 
3506  // Generate code to check if the loop's trip count is less than VF * UF, or
3507  // equal to it in case a scalar epilogue is required; this implies that the
3508  // vector trip count is zero. This check also covers the case where adding one
3509  // to the backedge-taken count overflowed leading to an incorrect trip count
3510  // of zero. In this case we will also jump to the scalar loop.
3511  auto P = Legal->requiresScalarEpilogue() ? ICmpInst::ICMP_ULE
3513  Value *CheckMinIters = Builder.CreateICmp(
3514  P, Count, ConstantInt::get(Count->getType(), VF * UF), "min.iters.check");
3515 
3516  BasicBlock *NewBB = BB->splitBasicBlock(BB->getTerminator(), "vector.ph");
3517  // Update dominator tree immediately if the generated block is a
3518  // LoopBypassBlock because SCEV expansions to generate loop bypass
3519  // checks may query it before the current function is finished.
3520  DT->addNewBlock(NewBB, BB);
3521  if (L->getParentLoop())
3522  L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
3524  BranchInst::Create(Bypass, NewBB, CheckMinIters));
3525  LoopBypassBlocks.push_back(BB);
3526 }
3527 
3529  BasicBlock *BB = L->getLoopPreheader();
3530 
3531  // Generate the code to check that the SCEV assumptions that we made.
3532  // We want the new basic block to start at the first instruction in a
3533  // sequence of instructions that form a check.
3534  SCEVExpander Exp(*PSE.getSE(), Bypass->getModule()->getDataLayout(),
3535  "scev.check");
3536  Value *SCEVCheck =
3537  Exp.expandCodeForPredicate(&PSE.getUnionPredicate(), BB->getTerminator());
3538 
3539  if (auto *C = dyn_cast<ConstantInt>(SCEVCheck))
3540  if (C->isZero())
3541  return;
3542 
3543  // Create a new block containing the stride check.
3544  BB->setName("vector.scevcheck");
3545  auto *NewBB = BB->splitBasicBlock(BB->getTerminator(), "vector.ph");
3546  // Update dominator tree immediately if the generated block is a
3547  // LoopBypassBlock because SCEV expansions to generate loop bypass
3548  // checks may query it before the current function is finished.
3549  DT->addNewBlock(NewBB, BB);
3550  if (L->getParentLoop())
3551  L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
3553  BranchInst::Create(Bypass, NewBB, SCEVCheck));
3554  LoopBypassBlocks.push_back(BB);
3555  AddedSafetyChecks = true;
3556 }
3557 
3559  BasicBlock *BB = L->getLoopPreheader();
3560 
3561  // Generate the code that checks in runtime if arrays overlap. We put the
3562  // checks into a separate block to make the more common case of few elements
3563  // faster.
3564  Instruction *FirstCheckInst;
3565  Instruction *MemRuntimeCheck;
3566  std::tie(FirstCheckInst, MemRuntimeCheck) =
3567  Legal->getLAI()->addRuntimeChecks(BB->getTerminator());
3568  if (!MemRuntimeCheck)
3569  return;
3570 
3571  // Create a new block containing the memory check.
3572  BB->setName("vector.memcheck");
3573  auto *NewBB = BB->splitBasicBlock(BB->getTerminator(), "vector.ph");
3574  // Update dominator tree immediately if the generated block is a
3575  // LoopBypassBlock because SCEV expansions to generate loop bypass
3576  // checks may query it before the current function is finished.
3577  DT->addNewBlock(NewBB, BB);
3578  if (L->getParentLoop())
3579  L->getParentLoop()->addBasicBlockToLoop(NewBB, *LI);
3581  BranchInst::Create(Bypass, NewBB, MemRuntimeCheck));
3582  LoopBypassBlocks.push_back(BB);
3583  AddedSafetyChecks = true;
3584 
3585  // We currently don't use LoopVersioning for the actual loop cloning but we
3586  // still use it to add the noalias metadata.
3587  LVer = llvm::make_unique<LoopVersioning>(*Legal->getLAI(), OrigLoop, LI, DT,
3588  PSE.getSE());
3589  LVer->prepareNoAliasMetadata();
3590 }
3591 
3593  /*
3594  In this function we generate a new loop. The new loop will contain
3595  the vectorized instructions while the old loop will continue to run the
3596  scalar remainder.
3597 
3598  [ ] <-- loop iteration number check.
3599  / |
3600  / v
3601  | [ ] <-- vector loop bypass (may consist of multiple blocks).
3602  | / |
3603  | / v
3604  || [ ] <-- vector pre header.
3605  |/ |
3606  | v
3607  | [ ] \
3608  | [ ]_| <-- vector loop.
3609  | |
3610  | v
3611  | -[ ] <--- middle-block.
3612  | / |
3613  | / v
3614  -|- >[ ] <--- new preheader.
3615  | |
3616  | v
3617  | [ ] \
3618  | [ ]_| <-- old scalar loop to handle remainder.
3619  \ |
3620  \ v
3621  >[ ] <-- exit block.
3622  ...
3623  */
3624 
3625  BasicBlock *OldBasicBlock = OrigLoop->getHeader();
3626  BasicBlock *VectorPH = OrigLoop->getLoopPreheader();
3627  BasicBlock *ExitBlock = OrigLoop->getExitBlock();
3628  assert(VectorPH && "Invalid loop structure");
3629  assert(ExitBlock && "Must have an exit block");
3630 
3631  // Some loops have a single integer induction variable, while other loops
3632  // don't. One example is c++ iterators that often have multiple pointer
3633  // induction variables. In the code below we also support a case where we
3634  // don't have a single induction variable.
3635  //
3636  // We try to obtain an induction variable from the original loop as hard
3637  // as possible. However if we don't find one that:
3638  // - is an integer
3639  // - counts from zero, stepping by one
3640  // - is the size of the widest induction variable type
3641  // then we create a new one.
3642  OldInduction = Legal->getPrimaryInduction();
3643  Type *IdxTy = Legal->getWidestInductionType();
3644 
3645  // Split the single block loop into the two loop structure described above.
3646  BasicBlock *VecBody =
3647  VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
3648  BasicBlock *MiddleBlock =
3649  VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
3650  BasicBlock *ScalarPH =
3651  MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
3652 
3653  // Create and register the new vector loop.
3654  Loop *Lp = LI->AllocateLoop();
3655  Loop *ParentLoop = OrigLoop->getParentLoop();
3656 
3657  // Insert the new loop into the loop nest and register the new basic blocks
3658  // before calling any utilities such as SCEV that require valid LoopInfo.
3659  if (ParentLoop) {
3660  ParentLoop->addChildLoop(Lp);
3661  ParentLoop->addBasicBlockToLoop(ScalarPH, *LI);
3662  ParentLoop->addBasicBlockToLoop(MiddleBlock, *LI);
3663  } else {
3664  LI->addTopLevelLoop(Lp);
3665  }
3666  Lp->addBasicBlockToLoop(VecBody, *LI);
3667 
3668  // Find the loop boundaries.
3669  Value *Count = getOrCreateTripCount(Lp);
3670 
3671  Value *StartIdx = ConstantInt::get(IdxTy, 0);
3672 
3673  // Now, compare the new count to zero. If it is zero skip the vector loop and
3674  // jump to the scalar loop. This check also covers the case where the
3675  // backedge-taken count is uint##_max: adding one to it will overflow leading
3676  // to an incorrect trip count of zero. In this (rare) case we will also jump
3677  // to the scalar loop.
3678  emitMinimumIterationCountCheck(Lp, ScalarPH);
3679 
3680  // Generate the code to check any assumptions that we've made for SCEV
3681  // expressions.
3682  emitSCEVChecks(Lp, ScalarPH);
3683 
3684  // Generate the code that checks in runtime if arrays overlap. We put the
3685  // checks into a separate block to make the more common case of few elements
3686  // faster.
3687  emitMemRuntimeChecks(Lp, ScalarPH);
3688 
3689  // Generate the induction variable.
3690  // The loop step is equal to the vectorization factor (num of SIMD elements)
3691  // times the unroll factor (num of SIMD instructions).
3692  Value *CountRoundDown = getOrCreateVectorTripCount(Lp);
3693  Constant *Step = ConstantInt::get(IdxTy, VF * UF);
3694  Induction =
3695  createInductionVariable(Lp, StartIdx, CountRoundDown, Step,
3697 
3698  // We are going to resume the execution of the scalar loop.
3699  // Go over all of the induction variables that we found and fix the
3700  // PHIs that are left in the scalar version of the loop.
3701  // The starting values of PHI nodes depend on the counter of the last
3702  // iteration in the vectorized loop.
3703  // If we come from a bypass edge then we need to start from the original
3704  // start value.
3705 
3706  // This variable saves the new starting index for the scalar loop. It is used
3707  // to test if there are any tail iterations left once the vector loop has
3708  // completed.
3709  LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
3710  for (auto &InductionEntry : *List) {
3711  PHINode *OrigPhi = InductionEntry.first;
3712  InductionDescriptor II = InductionEntry.second;
3713 
3714  // Create phi nodes to merge from the backedge-taken check block.
3715  PHINode *BCResumeVal = PHINode::Create(
3716  OrigPhi->getType(), 3, "bc.resume.val", ScalarPH->getTerminator());
3717  Value *&EndValue = IVEndValues[OrigPhi];
3718  if (OrigPhi == OldInduction) {
3719  // We know what the end value is.
3720  EndValue = CountRoundDown;
3721  } else {
3723  Type *StepType = II.getStep()->getType();
3724  Instruction::CastOps CastOp =
3725  CastInst::getCastOpcode(CountRoundDown, true, StepType, true);
3726  Value *CRD = B.CreateCast(CastOp, CountRoundDown, StepType, "cast.crd");
3727  const DataLayout &DL = OrigLoop->getHeader()->getModule()->getDataLayout();
3728  EndValue = II.transform(B, CRD, PSE.getSE(), DL);
3729  EndValue->setName("ind.end");
3730  }
3731 
3732  // The new PHI merges the original incoming value, in case of a bypass,
3733  // or the value at the end of the vectorized loop.
3734  BCResumeVal->addIncoming(EndValue, MiddleBlock);
3735 
3736  // Fix the scalar body counter (PHI node).
3737  unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
3738 
3739  // The old induction's phi node in the scalar body needs the truncated
3740  // value.
3741  for (BasicBlock *BB : LoopBypassBlocks)
3742  BCResumeVal->addIncoming(II.getStartValue(), BB);
3743  OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
3744  }
3745 
3746  // Add a check in the middle block to see if we have completed
3747  // all of the iterations in the first vector loop.
3748  // If (N - N%VF) == N, then we *don't* need to run the remainder.
3749  Value *CmpN =
3750  CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, Count,
3751  CountRoundDown, "cmp.n", MiddleBlock->getTerminator());
3752  ReplaceInstWithInst(MiddleBlock->getTerminator(),
3753  BranchInst::Create(ExitBlock, ScalarPH, CmpN));
3754 
3755  // Get ready to start creating new instructions into the vectorized body.
3757 
3758  // Save the state.
3760  LoopScalarPreHeader = ScalarPH;
3761  LoopMiddleBlock = MiddleBlock;
3762  LoopExitBlock = ExitBlock;
3763  LoopVectorBody = VecBody;
3764  LoopScalarBody = OldBasicBlock;
3765 
3766  // Keep all loop hints from the original loop on the vector loop (we'll
3767  // replace the vectorizer-specific hints below).
3768  if (MDNode *LID = OrigLoop->getLoopID())
3769  Lp->setLoopID(LID);
3770 
3771  LoopVectorizeHints Hints(Lp, true, *ORE);
3772  Hints.setAlreadyVectorized();
3773 
3774  return LoopVectorPreHeader;
3775 }
3776 
3777 // Fix up external users of the induction variable. At this point, we are
3778 // in LCSSA form, with all external PHIs that use the IV having one input value,
3779 // coming from the remainder loop. We need those PHIs to also have a correct
3780 // value for the IV when arriving directly from the middle block.
3782  const InductionDescriptor &II,
3783  Value *CountRoundDown, Value *EndValue,
3784  BasicBlock *MiddleBlock) {
3785  // There are two kinds of external IV usages - those that use the value
3786  // computed in the last iteration (the PHI) and those that use the penultimate
3787  // value (the value that feeds into the phi from the loop latch).
3788  // We allow both, but they, obviously, have different values.
3789 
3790  assert(OrigLoop->getExitBlock() && "Expected a single exit block");
3791 
3792  DenseMap<Value *, Value *> MissingVals;
3793 
3794  // An external user of the last iteration's value should see the value that
3795  // the remainder loop uses to initialize its own IV.
3797  for (User *U : PostInc->users()) {
3798  Instruction *UI = cast<Instruction>(U);
3799  if (!OrigLoop->contains(UI)) {
3800  assert(isa<PHINode>(UI) && "Expected LCSSA form");
3801  MissingVals[UI] = EndValue;
3802  }
3803  }
3804 
3805  // An external user of the penultimate value need to see EndValue - Step.
3806  // The simplest way to get this is to recompute it from the constituent SCEVs,
3807  // that is Start + (Step * (CRD - 1)).
3808  for (User *U : OrigPhi->users()) {
3809  auto *UI = cast<Instruction>(U);
3810  if (!OrigLoop->contains(UI)) {
3811  const DataLayout &DL =
3813  assert(isa<PHINode>(UI) && "Expected LCSSA form");
3814 
3815  IRBuilder<> B(MiddleBlock->getTerminator());
3816  Value *CountMinusOne = B.CreateSub(
3817  CountRoundDown, ConstantInt::get(CountRoundDown->getType(), 1));
3818  Value *CMO =
3819  !II.getStep()->getType()->isIntegerTy()
3820  ? B.CreateCast(Instruction::SIToFP, CountMinusOne,
3821  II.getStep()->getType())
3822  : B.CreateSExtOrTrunc(CountMinusOne, II.getStep()->getType());
3823  CMO->setName("cast.cmo");
3824  Value *Escape = II.transform(B, CMO, PSE.getSE(), DL);
3825  Escape->setName("ind.escape");
3826  MissingVals[UI] = Escape;
3827  }
3828  }
3829 
3830  for (auto &I : MissingVals) {
3831  PHINode *PHI = cast<PHINode>(I.first);
3832  // One corner case we have to handle is two IVs "chasing" each-other,
3833  // that is %IV2 = phi [...], [ %IV1, %latch ]
3834  // In this case, if IV1 has an external use, we need to avoid adding both
3835  // "last value of IV1" and "penultimate value of IV2". So, verify that we
3836  // don't already have an incoming value for the middle block.
3837  if (PHI->getBasicBlockIndex(MiddleBlock) == -1)
3838  PHI->addIncoming(I.second, MiddleBlock);
3839  }
3840 }
3841 
3842 namespace {
3843 
3844 struct CSEDenseMapInfo {
3845  static bool canHandle(const Instruction *I) {
3846  return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
3847  isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
3848  }
3849 
3850  static inline Instruction *getEmptyKey() {
3852  }
3853 
3854  static inline Instruction *getTombstoneKey() {
3856  }
3857 
3858  static unsigned getHashValue(const Instruction *I) {
3859  assert(canHandle(I) && "Unknown instruction!");
3861  I->value_op_end()));
3862  }
3863 
3864  static bool isEqual(const Instruction *LHS, const Instruction *RHS) {
3865  if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
3866  LHS == getTombstoneKey() || RHS == getTombstoneKey())
3867  return LHS == RHS;
3868  return LHS->isIdenticalTo(RHS);
3869  }
3870 };
3871 
3872 } // end anonymous namespace
3873 
3874 ///\brief Perform cse of induction variable instructions.
3875 static void cse(BasicBlock *BB) {
3876  // Perform simple cse.
3878  for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
3879  Instruction *In = &*I++;
3880 
3881  if (!CSEDenseMapInfo::canHandle(In))
3882  continue;
3883 
3884  // Check if we can replace this instruction with any of the
3885  // visited instructions.
3886  if (Instruction *V = CSEMap.lookup(In)) {
3887  In->replaceAllUsesWith(V);
3888  In->eraseFromParent();
3889  continue;
3890  }
3891 
3892  CSEMap[In] = In;
3893  }
3894 }
3895 
3896 /// \brief Estimate the overhead of scalarizing an instruction. This is a
3897 /// convenience wrapper for the type-based getScalarizationOverhead API.
3898 static unsigned getScalarizationOverhead(Instruction *I, unsigned VF,
3899  const TargetTransformInfo &TTI) {
3900  if (VF == 1)
3901  return 0;
3902 
3903  unsigned Cost = 0;
3904  Type *RetTy = ToVectorTy(I->getType(), VF);
3905  if (!RetTy->isVoidTy() &&
3906  (!isa<LoadInst>(I) ||
3908  Cost += TTI.getScalarizationOverhead(RetTy, true, false);
3909 
3910  if (CallInst *CI = dyn_cast<CallInst>(I)) {
3911  SmallVector<const Value *, 4> Operands(CI->arg_operands());
3912  Cost += TTI.getOperandsScalarizationOverhead(Operands, VF);
3913  }
3914  else if (!isa<StoreInst>(I) ||
3917  Cost += TTI.getOperandsScalarizationOverhead(Operands, VF);
3918  }
3919 
3920  return Cost;
3921 }
3922 
3923 // Estimate cost of a call instruction CI if it were vectorized with factor VF.
3924 // Return the cost of the instruction, including scalarization overhead if it's
3925 // needed. The flag NeedToScalarize shows if the call needs to be scalarized -
3926 // i.e. either vector version isn't available, or is too expensive.
3927 static unsigned getVectorCallCost(CallInst *CI, unsigned VF,
3928  const TargetTransformInfo &TTI,
3929  const TargetLibraryInfo *TLI,
3930  bool &NeedToScalarize) {
3931  Function *F = CI->getCalledFunction();
3932  StringRef FnName = CI->getCalledFunction()->getName();
3933  Type *ScalarRetTy = CI->getType();
3934  SmallVector<Type *, 4> Tys, ScalarTys;
3935  for (auto &ArgOp : CI->arg_operands())
3936  ScalarTys.push_back(ArgOp->getType());
3937 
3938  // Estimate cost of scalarized vector call. The source operands are assumed
3939  // to be vectors, so we need to extract individual elements from there,
3940  // execute VF scalar calls, and then gather the result into the vector return
3941  // value.
3942  unsigned ScalarCallCost = TTI.getCallInstrCost(F, ScalarRetTy, ScalarTys);
3943  if (VF == 1)
3944  return ScalarCallCost;
3945 
3946  // Compute corresponding vector type for return value and arguments.
3947  Type *RetTy = ToVectorTy(ScalarRetTy, VF);
3948  for (Type *ScalarTy : ScalarTys)
3949  Tys.push_back(ToVectorTy(ScalarTy, VF));
3950 
3951  // Compute costs of unpacking argument values for the scalar calls and
3952  // packing the return values to a vector.
3953  unsigned ScalarizationCost = getScalarizationOverhead(CI, VF, TTI);
3954 
3955  unsigned Cost = ScalarCallCost * VF + ScalarizationCost;
3956 
3957  // If we can't emit a vector call for this function, then the currently found
3958  // cost is the cost we need to return.
3959  NeedToScalarize = true;
3960  if (!TLI || !TLI->isFunctionVectorizable(FnName, VF) || CI->isNoBuiltin())
3961  return Cost;
3962 
3963  // If the corresponding vector cost is cheaper, return its cost.
3964  unsigned VectorCallCost = TTI.getCallInstrCost(nullptr, RetTy, Tys);
3965  if (VectorCallCost < Cost) {
3966  NeedToScalarize = false;
3967  return VectorCallCost;
3968  }
3969  return Cost;
3970 }
3971 
3972 // Estimate cost of an intrinsic call instruction CI if it were vectorized with
3973 // factor VF. Return the cost of the instruction, including scalarization
3974 // overhead if it's needed.
3975 static unsigned getVectorIntrinsicCost(CallInst *CI, unsigned VF,
3976  const TargetTransformInfo &TTI,
3977  const TargetLibraryInfo *TLI) {
3979  assert(ID && "Expected intrinsic call!");
3980 
3981  FastMathFlags FMF;
3982  if (auto *FPMO = dyn_cast<FPMathOperator>(CI))
3983  FMF = FPMO->getFastMathFlags();
3984 
3985  SmallVector<Value *, 4> Operands(CI->arg_operands());
3986  return TTI.getIntrinsicInstrCost(ID, CI->getType(), Operands, FMF, VF);
3987 }
3988 
3990  auto *I1 = cast<IntegerType>(T1->getVectorElementType());
3991  auto *I2 = cast<IntegerType>(T2->getVectorElementType());
3992  return I1->getBitWidth() < I2->getBitWidth() ? T1 : T2;
3993 }
3995  auto *I1 = cast<IntegerType>(T1->getVectorElementType());
3996  auto *I2 = cast<IntegerType>(T2->getVectorElementType());
3997  return I1->getBitWidth() > I2->getBitWidth() ? T1 : T2;
3998 }
3999 
4001  // For every instruction `I` in MinBWs, truncate the operands, create a
4002  // truncated version of `I` and reextend its result. InstCombine runs
4003  // later and will remove any ext/trunc pairs.
4004  SmallPtrSet<Value *, 4> Erased;
4005  for (const auto &KV : Cost->getMinimalBitwidths()) {
4006  // If the value wasn't vectorized, we must maintain the original scalar
4007  // type. The absence of the value from VectorLoopValueMap indicates that it
4008  // wasn't vectorized.
4009  if (!VectorLoopValueMap.hasAnyVectorValue(KV.first))
4010  continue;
4011  for (unsigned Part = 0; Part < UF; ++Part) {
4012  Value *I = getOrCreateVectorValue(KV.first, Part);
4013  if (Erased.count(I) || I->use_empty() || !isa<Instruction>(I))
4014  continue;
4015  Type *OriginalTy = I->getType();
4016  Type *ScalarTruncatedTy =
4017  IntegerType::get(OriginalTy->getContext(), KV.second);
4018  Type *TruncatedTy = VectorType::get(ScalarTruncatedTy,
4019  OriginalTy->getVectorNumElements());
4020  if (TruncatedTy == OriginalTy)
4021  continue;
4022 
4023  IRBuilder<> B(cast<Instruction>(I));
4024  auto ShrinkOperand = [&](Value *V) -> Value * {
4025  if (auto *ZI = dyn_cast<ZExtInst>(V))
4026  if (ZI->getSrcTy() == TruncatedTy)
4027  return ZI->getOperand(0);
4028  return B.CreateZExtOrTrunc(V, TruncatedTy);
4029  };
4030 
4031  // The actual instruction modification depends on the instruction type,
4032  // unfortunately.
4033  Value *NewI = nullptr;
4034  if (auto *BO = dyn_cast<BinaryOperator>(I)) {
4035  NewI = B.CreateBinOp(BO->getOpcode(), ShrinkOperand(BO->getOperand(0)),
4036  ShrinkOperand(BO->getOperand(1)));
4037 
4038  // Any wrapping introduced by shrinking this operation shouldn't be
4039  // considered undefined behavior. So, we can't unconditionally copy
4040  // arithmetic wrapping flags to NewI.
4041  cast<BinaryOperator>(NewI)->copyIRFlags(I, /*IncludeWrapFlags=*/false);
4042  } else if (auto *CI = dyn_cast<ICmpInst>(I)) {
4043  NewI =
4044  B.CreateICmp(CI->getPredicate(), ShrinkOperand(CI->getOperand(0)),
4045  ShrinkOperand(CI->getOperand(1)));
4046  } else if (auto *SI = dyn_cast<SelectInst>(I)) {
4047  NewI = B.CreateSelect(SI->getCondition(),
4048  ShrinkOperand(SI->getTrueValue()),
4049  ShrinkOperand(SI->getFalseValue()));
4050  } else if (auto *CI = dyn_cast<CastInst>(I)) {
4051  switch (CI->getOpcode()) {
4052  default:
4053  llvm_unreachable("Unhandled cast!");
4054  case Instruction::Trunc:
4055  NewI = ShrinkOperand(CI->getOperand(0));
4056  break;
4057  case Instruction::SExt:
4058  NewI = B.CreateSExtOrTrunc(
4059  CI->getOperand(0),
4060  smallestIntegerVectorType(OriginalTy, TruncatedTy));
4061  break;
4062  case Instruction::ZExt:
4063  NewI = B.CreateZExtOrTrunc(
4064  CI->getOperand(0),
4065  smallestIntegerVectorType(OriginalTy, TruncatedTy));
4066  break;
4067  }
4068  } else if (auto *SI = dyn_cast<ShuffleVectorInst>(I)) {
4069  auto Elements0 = SI->getOperand(0)->getType()->getVectorNumElements();
4070  auto *O0 = B.CreateZExtOrTrunc(
4071  SI->getOperand(0), VectorType::get(ScalarTruncatedTy, Elements0));
4072  auto Elements1 = SI->getOperand(1)->getType()->getVectorNumElements();
4073  auto *O1 = B.CreateZExtOrTrunc(
4074  SI->getOperand(1), VectorType::get(ScalarTruncatedTy, Elements1));
4075 
4076  NewI = B.CreateShuffleVector(O0, O1, SI->getMask());
4077  } else if (isa<LoadInst>(I)) {
4078  // Don't do anything with the operands, just extend the result.
4079  continue;
4080  } else if (auto *IE = dyn_cast<InsertElementInst>(I)) {
4081  auto Elements = IE->getOperand(0)->getType()->getVectorNumElements();
4082  auto *O0 = B.CreateZExtOrTrunc(
4083  IE->getOperand(0), VectorType::get(ScalarTruncatedTy, Elements));
4084  auto *O1 = B.CreateZExtOrTrunc(IE->getOperand(1), ScalarTruncatedTy);
4085  NewI = B.CreateInsertElement(O0, O1, IE->getOperand(2));
4086  } else if (auto *EE = dyn_cast<ExtractElementInst>(I)) {
4087  auto Elements = EE->getOperand(0)->getType()->getVectorNumElements();
4088  auto *O0 = B.CreateZExtOrTrunc(
4089  EE->getOperand(0), VectorType::get(ScalarTruncatedTy, Elements));
4090  NewI = B.CreateExtractElement(O0, EE->getOperand(2));
4091  } else {
4092  llvm_unreachable("Unhandled instruction type!");
4093  }
4094 
4095  // Lastly, extend the result.
4096  NewI->takeName(cast<Instruction>(I));
4097  Value *Res = B.CreateZExtOrTrunc(NewI, OriginalTy);
4098  I->replaceAllUsesWith(Res);
4099  cast<Instruction>(I)->eraseFromParent();
4100  Erased.insert(I);
4101  VectorLoopValueMap.resetVectorValue(KV.first, Part, Res);
4102  }
4103  }
4104 
4105  // We'll have created a bunch of ZExts that are now parentless. Clean up.
4106  for (const auto &KV : Cost->getMinimalBitwidths()) {
4107  // If the value wasn't vectorized, we must maintain the original scalar
4108  // type. The absence of the value from VectorLoopValueMap indicates that it
4109  // wasn't vectorized.
4110  if (!VectorLoopValueMap.hasAnyVectorValue(KV.first))
4111  continue;
4112  for (unsigned Part = 0; Part < UF; ++Part) {
4113  Value *I = getOrCreateVectorValue(KV.first, Part);
4114  ZExtInst *Inst = dyn_cast<ZExtInst>(I);
4115  if (Inst && Inst->use_empty()) {
4116  Value *NewI = Inst->getOperand(0);
4117  Inst->eraseFromParent();
4118  VectorLoopValueMap.resetVectorValue(KV.first, Part, NewI);
4119  }
4120  }
4121  }
4122 }
4123 
4125  // Insert truncates and extends for any truncated instructions as hints to
4126  // InstCombine.
4127  if (VF > 1)
4129 
4130  // At this point every instruction in the original loop is widened to a
4131  // vector form. Now we need to fix the recurrences in the loop. These PHI
4132  // nodes are currently empty because we did not want to introduce cycles.
4133  // This is the second stage of vectorizing recurrences.
4135 
4136  // Update the dominator tree.
4137  //
4138  // FIXME: After creating the structure of the new loop, the dominator tree is
4139  // no longer up-to-date, and it remains that way until we update it
4140  // here. An out-of-date dominator tree is problematic for SCEV,
4141  // because SCEVExpander uses it to guide code generation. The
4142  // vectorizer use SCEVExpanders in several places. Instead, we should
4143  // keep the dominator tree up-to-date as we go.
4144  updateAnalysis();
4145 
4146  // Fix-up external users of the induction variables.
4147  for (auto &Entry : *Legal->getInductionVars())
4148  fixupIVUsers(Entry.first, Entry.second,
4150  IVEndValues[Entry.first], LoopMiddleBlock);
4151 
4152  fixLCSSAPHIs();
4154  sinkScalarOperands(&*PI);
4155 
4156  // Remove redundant induction instructions.
4158 }
4159 
4161  // In order to support recurrences we need to be able to vectorize Phi nodes.
4162  // Phi nodes have cycles, so we need to vectorize them in two stages. This is
4163  // stage #2: We now need to fix the recurrences by adding incoming edges to
4164  // the currently empty PHI nodes. At this point every instruction in the
4165  // original loop is widened to a vector form so we can use them to construct
4166  // the incoming edges.
4167  for (Instruction &I : *OrigLoop->getHeader()) {
4168  PHINode *Phi = dyn_cast<PHINode>(&I);
4169  if (!Phi)
4170  break;
4171  // Handle first-order recurrences and reductions that need to be fixed.
4172  if (Legal->isFirstOrderRecurrence(Phi))
4174  else if (Legal->isReductionVariable(Phi))
4175  fixReduction(Phi);
4176  }
4177 }
4178 
4180  // This is the second phase of vectorizing first-order recurrences. An
4181  // overview of the transformation is described below. Suppose we have the
4182  // following loop.
4183  //
4184  // for (int i = 0; i < n; ++i)
4185  // b[i] = a[i] - a[i - 1];
4186  //
4187  // There is a first-order recurrence on "a". For this loop, the shorthand
4188  // scalar IR looks like:
4189  //
4190  // scalar.ph:
4191  // s_init = a[-1]
4192  // br scalar.body
4193  //
4194  // scalar.body:
4195  // i = phi [0, scalar.ph], [i+1, scalar.body]
4196  // s1 = phi [s_init, scalar.ph], [s2, scalar.body]
4197  // s2 = a[i]
4198  // b[i] = s2 - s1
4199  // br cond, scalar.body, ...
4200  //
4201  // In this example, s1 is a recurrence because it's value depends on the
4202  // previous iteration. In the first phase of vectorization, we created a
4203  // temporary value for s1. We now complete the vectorization and produce the
4204  // shorthand vector IR shown below (for VF = 4, UF = 1).
4205  //
4206  // vector.ph:
4207  // v_init = vector(..., ..., ..., a[-1])
4208  // br vector.body
4209  //
4210  // vector.body
4211  // i = phi [0, vector.ph], [i+4, vector.body]
4212  // v1 = phi [v_init, vector.ph], [v2, vector.body]
4213  // v2 = a[i, i+1, i+2, i+3];
4214  // v3 = vector(v1(3), v2(0, 1, 2))
4215  // b[i, i+1, i+2, i+3] = v2 - v3
4216  // br cond, vector.body, middle.block
4217  //
4218  // middle.block:
4219  // x = v2(3)
4220  // br scalar.ph
4221  //
4222  // scalar.ph:
4223  // s_init = phi [x, middle.block], [a[-1], otherwise]
4224  // br scalar.body
4225  //
4226  // After execution completes the vector loop, we extract the next value of
4227  // the recurrence (x) to use as the initial value in the scalar loop.
4228 
4229  // Get the original loop preheader and single loop latch.
4230  auto *Preheader = OrigLoop->getLoopPreheader();
4231  auto *Latch = OrigLoop->getLoopLatch();
4232 
4233  // Get the initial and previous values of the scalar recurrence.
4234  auto *ScalarInit = Phi->getIncomingValueForBlock(Preheader);
4235  auto *Previous = Phi->getIncomingValueForBlock(Latch);
4236 
4237  // Create a vector from the initial value.
4238  auto *VectorInit = ScalarInit;
4239  if (VF > 1) {
4241  VectorInit = Builder.CreateInsertElement(
4242  UndefValue::get(VectorType::get(VectorInit->getType(), VF)), VectorInit,
4243  Builder.getInt32(VF - 1), "vector.recur.init");
4244  }
4245 
4246  // We constructed a temporary phi node in the first phase of vectorization.
4247  // This phi node will eventually be deleted.
4249  cast<Instruction>(VectorLoopValueMap.getVectorValue(Phi, 0)));
4250 
4251  // Create a phi node for the new recurrence. The current value will either be
4252  // the initial value inserted into a vector or loop-varying vector value.
4253  auto *VecPhi = Builder.CreatePHI(VectorInit->getType(), 2, "vector.recur");
4254  VecPhi->addIncoming(VectorInit, LoopVectorPreHeader);
4255 
4256  // Get the vectorized previous value of the last part UF - 1. It appears last
4257  // among all unrolled iterations, due to the order of their construction.
4258  Value *PreviousLastPart = getOrCreateVectorValue(Previous, UF - 1);
4259 
4260  // Set the insertion point after the previous value if it is an instruction.
4261  // Note that the previous value may have been constant-folded so it is not
4262  // guaranteed to be an instruction in the vector loop. Also, if the previous
4263  // value is a phi node, we should insert after all the phi nodes to avoid
4264  // breaking basic block verification.
4265  if (LI->getLoopFor(LoopVectorBody)->isLoopInvariant(PreviousLastPart) ||
4266  isa<PHINode>(PreviousLastPart))
4268  else
4270  &*++BasicBlock::iterator(cast<Instruction>(PreviousLastPart)));
4271 
4272  // We will construct a vector for the recurrence by combining the values for
4273  // the current and previous iterations. This is the required shuffle mask.
4275  ShuffleMask[0] = Builder.getInt32(VF - 1);
4276  for (unsigned I = 1; I < VF; ++I)
4277  ShuffleMask[I] = Builder.getInt32(I + VF - 1);
4278 
4279  // The vector from which to take the initial value for the current iteration
4280  // (actual or unrolled). Initially, this is the vector phi node.
4281  Value *Incoming = VecPhi;
4282 
4283  // Shuffle the current and previous vector and update the vector parts.
4284  for (unsigned Part = 0; Part < UF; ++Part) {
4285  Value *PreviousPart = getOrCreateVectorValue(Previous, Part);
4286  Value *PhiPart = VectorLoopValueMap.getVectorValue(Phi, Part);
4287  auto *Shuffle =
4288  VF > 1 ? Builder.CreateShuffleVector(Incoming, PreviousPart,
4289  ConstantVector::get(ShuffleMask))
4290  : Incoming;
4291  PhiPart->replaceAllUsesWith(Shuffle);
4292  cast<Instruction>(PhiPart)->eraseFromParent();
4293  VectorLoopValueMap.resetVectorValue(Phi, Part, Shuffle);
4294  Incoming = PreviousPart;
4295  }
4296 
4297  // Fix the latch value of the new recurrence in the vector loop.
4298  VecPhi->addIncoming(Incoming, LI->getLoopFor(LoopVectorBody)->getLoopLatch());
4299 
4300  // Extract the last vector element in the middle block. This will be the
4301  // initial value for the recurrence when jumping to the scalar loop.
4302  auto *ExtractForScalar = Incoming;
4303  if (VF > 1) {
4305  ExtractForScalar = Builder.CreateExtractElement(
4306  ExtractForScalar, Builder.getInt32(VF - 1), "vector.recur.extract");
4307  }
4308  // Extract the second last element in the middle block if the
4309  // Phi is used outside the loop. We need to extract the phi itself
4310  // and not the last element (the phi update in the current iteration). This
4311  // will be the value when jumping to the exit block from the LoopMiddleBlock,
4312  // when the scalar loop is not run at all.
4313  Value *ExtractForPhiUsedOutsideLoop = nullptr;
4314  if (VF > 1)
4315  ExtractForPhiUsedOutsideLoop = Builder.CreateExtractElement(
4316  Incoming, Builder.getInt32(VF - 2), "vector.recur.extract.for.phi");
4317  // When loop is unrolled without vectorizing, initialize
4318  // ExtractForPhiUsedOutsideLoop with the value just prior to unrolled value of
4319  // `Incoming`. This is analogous to the vectorized case above: extracting the
4320  // second last element when VF > 1.
4321  else if (UF > 1)
4322  ExtractForPhiUsedOutsideLoop = getOrCreateVectorValue(Previous, UF - 2);
4323 
4324  // Fix the initial value of the original recurrence in the scalar loop.
4326  auto *Start = Builder.CreatePHI(Phi->getType(), 2, "scalar.recur.init");
4327  for (auto *BB : predecessors(LoopScalarPreHeader)) {
4328  auto *Incoming = BB == LoopMiddleBlock ? ExtractForScalar : ScalarInit;
4329  Start->addIncoming(Incoming, BB);
4330  }
4331 
4333  Phi->setName("scalar.recur");
4334 
4335  // Finally, fix users of the recurrence outside the loop. The users will need
4336  // either the last value of the scalar recurrence or the last value of the
4337  // vector recurrence we extracted in the middle block. Since the loop is in
4338  // LCSSA form, we just need to find the phi node for the original scalar
4339  // recurrence in the exit block, and then add an edge for the middle block.
4340  for (auto &I : *LoopExitBlock) {
4341  auto *LCSSAPhi = dyn_cast<PHINode>(&I);
4342  if (!LCSSAPhi)
4343  break;
4344  if (LCSSAPhi->getIncomingValue(0) == Phi) {
4345  LCSSAPhi->addIncoming(ExtractForPhiUsedOutsideLoop, LoopMiddleBlock);
4346  break;
4347  }
4348  }
4349 }
4350 
4352  Constant *Zero = Builder.getInt32(0);
4353 
4354  // Get it's reduction variable descriptor.
4355  assert(Legal->isReductionVariable(Phi) &&
4356  "Unable to find the reduction variable");
4357  RecurrenceDescriptor RdxDesc = (*Legal->getReductionVars())[Phi];
4358 
4360  TrackingVH<Value> ReductionStartValue = RdxDesc.getRecurrenceStartValue();
4361  Instruction *LoopExitInst = RdxDesc.getLoopExitInstr();
4363  RdxDesc.getMinMaxRecurrenceKind();
4364  setDebugLocFromInst(Builder, ReductionStartValue);
4365 
4366  // We need to generate a reduction vector from the incoming scalar.
4367  // To do so, we need to generate the 'identity' vector and override
4368  // one of the elements with the incoming scalar reduction. We need
4369  // to do it in the vector-loop preheader.
4371 
4372  // This is the vector-clone of the value that leaves the loop.
4373  Type *VecTy = getOrCreateVectorValue(LoopExitInst, 0)->getType();
4374 
4375  // Find the reduction identity variable. Zero for addition, or, xor,
4376  // one for multiplication, -1 for And.
4377  Value *Identity;
4378  Value *VectorStart;
4381  // MinMax reduction have the start value as their identify.
4382  if (VF == 1) {
4383  VectorStart = Identity = ReductionStartValue;
4384  } else {
4385  VectorStart = Identity =
4386  Builder.CreateVectorSplat(VF, ReductionStartValue, "minmax.ident");
4387  }
4388  } else {
4389  // Handle other reduction kinds:
4391  RK, VecTy->getScalarType());
4392  if (VF == 1) {
4393  Identity = Iden;
4394  // This vector is the Identity vector where the first element is the
4395  // incoming scalar reduction.
4396  VectorStart = ReductionStartValue;
4397  } else {
4398  Identity = ConstantVector::getSplat(VF, Iden);
4399 
4400  // This vector is the Identity vector where the first element is the
4401  // incoming scalar reduction.
4402  VectorStart =
4403  Builder.CreateInsertElement(Identity, ReductionStartValue, Zero);
4404  }
4405  }
4406 
4407  // Fix the vector-loop phi.
4408 
4409  // Reductions do not have to start at zero. They can start with
4410  // any loop invariant values.
4411  BasicBlock *Latch = OrigLoop->getLoopLatch();
4412  Value *LoopVal = Phi->getIncomingValueForBlock(Latch);
4413  for (unsigned Part = 0; Part < UF; ++Part) {
4414  Value *VecRdxPhi = getOrCreateVectorValue(Phi, Part);
4415  Value *Val = getOrCreateVectorValue(LoopVal, Part);
4416  // Make sure to add the reduction stat value only to the
4417  // first unroll part.
4418  Value *StartVal = (Part == 0) ? VectorStart : Identity;
4419  cast<PHINode>(VecRdxPhi)->addIncoming(StartVal, LoopVectorPreHeader);
4420  cast<PHINode>(VecRdxPhi)
4421  ->addIncoming(Val, LI->getLoopFor(LoopVectorBody)->getLoopLatch());
4422  }
4423 
4424  // Before each round, move the insertion point right between
4425  // the PHIs and the values we are going to write.
4426  // This allows us to write both PHINodes and the extractelement
4427  // instructions.
4429 
4430  setDebugLocFromInst(Builder, LoopExitInst);
4431 
4432  // If the vector reduction can be performed in a smaller type, we truncate
4433  // then extend the loop exit value to enable InstCombine to evaluate the
4434  // entire expression in the smaller type.
4435  if (VF > 1 && Phi->getType() != RdxDesc.getRecurrenceType()) {
4436  Type *RdxVecTy = VectorType::get(RdxDesc.getRecurrenceType(), VF);
4439  VectorParts RdxParts(UF);
4440  for (unsigned Part = 0; Part < UF; ++Part) {
4441  RdxParts[Part] = VectorLoopValueMap.getVectorValue(LoopExitInst, Part);
4442  Value *Trunc = Builder.CreateTrunc(RdxParts[Part], RdxVecTy);
4443  Value *Extnd = RdxDesc.isSigned() ? Builder.CreateSExt(Trunc, VecTy)
4444  : Builder.CreateZExt(Trunc, VecTy);
4445  for (Value::user_iterator UI = RdxParts[Part]->user_begin();
4446  UI != RdxParts[Part]->user_end();)
4447  if (*UI != Trunc) {
4448  (*UI++)->replaceUsesOfWith(RdxParts[Part], Extnd);
4449  RdxParts[Part] = Extnd;
4450  } else {
4451  ++UI;
4452  }
4453  }
4455  for (unsigned Part = 0; Part < UF; ++Part) {
4456  RdxParts[Part] = Builder.CreateTrunc(RdxParts[Part], RdxVecTy);
4457  VectorLoopValueMap.resetVectorValue(LoopExitInst, Part, RdxParts[Part]);
4458  }
4459  }
4460 
4461  // Reduce all of the unrolled parts into a single vector.
4462  Value *ReducedPartRdx = VectorLoopValueMap.getVectorValue(LoopExitInst, 0);
4464  setDebugLocFromInst(Builder, ReducedPartRdx);
4465  for (unsigned Part = 1; Part < UF; ++Part) {
4466  Value *RdxPart = VectorLoopValueMap.getVectorValue(LoopExitInst, Part);
4467  if (Op != Instruction::ICmp && Op != Instruction::FCmp)
4468  // Floating point operations had to be 'fast' to enable the reduction.
4469  ReducedPartRdx = addFastMathFlag(
4471  ReducedPartRdx, "bin.rdx"));
4472  else
4473  ReducedPartRdx = RecurrenceDescriptor::createMinMaxOp(
4474  Builder, MinMaxKind, ReducedPartRdx, RdxPart);
4475  }
4476 
4477  if (VF > 1) {
4478  bool NoNaN = Legal->hasFunNoNaNAttr();
4479  ReducedPartRdx =
4480  createTargetReduction(Builder, TTI, RdxDesc, ReducedPartRdx, NoNaN);
4481  // If the reduction can be performed in a smaller type, we need to extend
4482  // the reduction to the wider type before we branch to the original loop.
4483  if (Phi->getType() != RdxDesc.getRecurrenceType())
4484  ReducedPartRdx =
4485  RdxDesc.isSigned()
4486  ? Builder.CreateSExt(ReducedPartRdx, Phi->getType())
4487  : Builder.CreateZExt(ReducedPartRdx, Phi->getType());
4488  }
4489 
4490  // Create a phi node that merges control-flow from the backedge-taken check
4491  // block and the middle block.
4492  PHINode *BCBlockPhi = PHINode::Create(Phi->getType(), 2, "bc.merge.rdx",
4494  for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
4495  BCBlockPhi->addIncoming(ReductionStartValue, LoopBypassBlocks[I]);
4496  BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
4497 
4498  // Now, we need to fix the users of the reduction variable
4499  // inside and outside of the scalar remainder loop.
4500  // We know that the loop is in LCSSA form. We need to update the
4501  // PHI nodes in the exit blocks.
4503  LEE = LoopExitBlock->end();
4504  LEI != LEE; ++LEI) {
4505  PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
4506  if (!LCSSAPhi)
4507  break;
4508 
4509  // All PHINodes need to have a single entry edge, or two if
4510  // we already fixed them.
4511  assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
4512 
4513  // We found a reduction value exit-PHI. Update it with the
4514  // incoming bypass edge.
4515  if (LCSSAPhi->getIncomingValue(0) == LoopExitInst)
4516  LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
4517  } // end of the LCSSA phi scan.
4518 
4519  // Fix the scalar loop reduction variable with the incoming reduction sum
4520  // from the vector body and from the backedge value.
4521  int IncomingEdgeBlockIdx =
4523  assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
4524  // Pick the other block.
4525  int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
4526  Phi->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
4527  Phi->setIncomingValue(IncomingEdgeBlockIdx, LoopExitInst);
4528 }
4529 
4531  for (Instruction &LEI : *LoopExitBlock) {
4532  auto *LCSSAPhi = dyn_cast<PHINode>(&LEI);
4533  if (!LCSSAPhi)
4534  break;
4535  if (LCSSAPhi->getNumIncomingValues() == 1) {
4536  assert(OrigLoop->isLoopInvariant(LCSSAPhi->getIncomingValue(0)) &&
4537  "Incoming value isn't loop invariant");
4538  LCSSAPhi->addIncoming(LCSSAPhi->getIncomingValue(0), LoopMiddleBlock);
4539  }
4540  }
4541 }
4542 
4544  // The basic block and loop containing the predicated instruction.
4545  auto *PredBB = PredInst->getParent();
4546  auto *VectorLoop = LI->getLoopFor(PredBB);
4547 
4548  // Initialize a worklist with the operands of the predicated instruction.
4549  SetVector<Value *> Worklist(PredInst->op_begin(), PredInst->op_end());
4550 
4551  // Holds instructions that we need to analyze again. An instruction may be
4552  // reanalyzed if we don't yet know if we can sink it or not.
4553  SmallVector<Instruction *, 8> InstsToReanalyze;
4554 
4555  // Returns true if a given use occurs in the predicated block. Phi nodes use
4556  // their operands in their corresponding predecessor blocks.
4557  auto isBlockOfUsePredicated = [&](Use &U) -> bool {
4558  auto *I = cast<Instruction>(U.getUser());
4559  BasicBlock *BB = I->getParent();
4560  if (auto *Phi = dyn_cast<PHINode>(I))
4561  BB = Phi->getIncomingBlock(
4562  PHINode::getIncomingValueNumForOperand(U.getOperandNo()));
4563  return BB == PredBB;
4564  };
4565 
4566  // Iteratively sink the scalarized operands of the predicated instruction
4567  // into the block we created for it. When an instruction is sunk, it's
4568  // operands are then added to the worklist. The algorithm ends after one pass
4569  // through the worklist doesn't sink a single instruction.
4570  bool Changed;
4571  do {
4572  // Add the instructions that need to be reanalyzed to the worklist, and
4573  // reset the changed indicator.
4574  Worklist.insert(InstsToReanalyze.begin(), InstsToReanalyze.end());
4575  InstsToReanalyze.clear();
4576  Changed = false;
4577 
4578  while (!Worklist.empty()) {
4579  auto *I = dyn_cast<Instruction>(Worklist.pop_back_val());
4580 
4581  // We can't sink an instruction if it is a phi node, is already in the
4582  // predicated block, is not in the loop, or may have side effects.
4583  if (!I || isa<PHINode>(I) || I->getParent() == PredBB ||
4584  !VectorLoop->contains(I) || I->mayHaveSideEffects())
4585  continue;
4586 
4587  // It's legal to sink the instruction if all its uses occur in the
4588  // predicated block. Otherwise, there's nothing to do yet, and we may
4589  // need to reanalyze the instruction.
4590  if (!llvm::all_of(I->uses(), isBlockOfUsePredicated)) {
4591  InstsToReanalyze.push_back(I);
4592  continue;
4593  }
4594 
4595  // Move the instruction to the beginning of the predicated block, and add
4596  // it's operands to the worklist.
4597  I->moveBefore(&*PredBB->getFirstInsertionPt());
4598  Worklist.insert(I->op_begin(), I->op_end());
4599 
4600  // The sinking may have enabled other instructions to be sunk, so we will
4601  // need to iterate.
4602  Changed = true;
4603  }
4604  } while (Changed);
4605 }
4606 
4608  unsigned VF) {
4609  assert(PN->getParent() == OrigLoop->getHeader() &&
4610  "Non-header phis should have been handled elsewhere");
4611 
4612  PHINode *P = cast<PHINode>(PN);
4613  // In order to support recurrences we need to be able to vectorize Phi nodes.
4614  // Phi nodes have cycles, so we need to vectorize them in two stages. This is
4615  // stage #1: We create a new vector PHI node with no incoming edges. We'll use
4616  // this value when we vectorize all of the instructions that use the PHI.
4617  if (Legal->isReductionVariable(P) || Legal->isFirstOrderRecurrence(P)) {
4618  for (unsigned Part = 0; Part < UF; ++Part) {
4619  // This is phase one of vectorizing PHIs.
4620  Type *VecTy =
4621  (VF == 1) ? PN->getType() : VectorType::get(PN->getType(), VF);
4622  Value *EntryPart = PHINode::Create(
4623  VecTy, 2, "vec.phi", &*LoopVectorBody->getFirstInsertionPt());
4624  VectorLoopValueMap.setVectorValue(P, Part, EntryPart);
4625  }
4626  return;
4627  }
4628 
4630 
4631  // This PHINode must be an induction variable.
4632  // Make sure that we know about it.
4633  assert(Legal->getInductionVars()->count(P) && "Not an induction variable");
4634 
4635  InductionDescriptor II = Legal->getInductionVars()->lookup(P);
4636  const DataLayout &DL = OrigLoop->getHeader()->getModule()->getDataLayout();
4637 
4638  // FIXME: The newly created binary instructions should contain nsw/nuw flags,
4639  // which can be found from the original scalar operations.
4640  switch (II.getKind()) {
4642  llvm_unreachable("Unknown induction");
4645  llvm_unreachable("Integer/fp induction is handled elsewhere.");
4647  // Handle the pointer induction variable case.
4648  assert(P->getType()->isPointerTy() && "Unexpected type.");
4649  // This is the normalized GEP that starts counting at zero.
4650  Value *PtrInd = Induction;
4651  PtrInd = Builder.CreateSExtOrTrunc(PtrInd, II.getStep()->getType());
4652  // Determine the number of scalars we need to generate for each unroll
4653  // iteration. If the instruction is uniform, we only need to generate the
4654  // first lane. Otherwise, we generate all VF values.
4655  unsigned Lanes = Cost->isUniformAfterVectorization(P, VF) ? 1 : VF;
4656  // These are the scalar results. Notice that we don't generate vector GEPs
4657  // because scalar GEPs result in better code.
4658  for (unsigned Part = 0; Part < UF; ++Part) {
4659  for (unsigned Lane = 0; Lane < Lanes; ++Lane) {
4660  Constant *Idx = ConstantInt::get(PtrInd->getType(), Lane + Part * VF);
4661  Value *GlobalIdx = Builder.CreateAdd(PtrInd, Idx);
4662  Value *SclrGep = II.transform(Builder, GlobalIdx, PSE.getSE(), DL);
4663  SclrGep->setName("next.gep");
4664  VectorLoopValueMap.setScalarValue(P, {Part, Lane}, SclrGep);
4665  }
4666  }
4667  return;
4668  }
4669  }
4670 }
4671 
4672 /// A helper function for checking whether an integer division-related
4673 /// instruction may divide by zero (in which case it must be predicated if
4674 /// executed conditionally in the scalar code).
4675 /// TODO: It may be worthwhile to generalize and check isKnownNonZero().
4676 /// Non-zero divisors that are non compile-time constants will not be
4677 /// converted into multiplication, so we will still end up scalarizing
4678 /// the division, but can do so w/o predication.
4679 static bool mayDivideByZero(Instruction &I) {
4680  assert((I.getOpcode() == Instruction::UDiv ||
4681  I.getOpcode() == Instruction::SDiv ||
4682  I.getOpcode() == Instruction::URem ||
4683  I.getOpcode() == Instruction::SRem) &&
4684  "Unexpected instruction");
4685  Value *Divisor = I.getOperand(1);
4686  auto *CInt = dyn_cast<ConstantInt>(Divisor);
4687  return !CInt || CInt->isZero();
4688 }
4689 
4691  switch (I.getOpcode()) {
4692  case Instruction::Br:
4693  case Instruction::PHI:
4694  llvm_unreachable("This instruction is handled by a different recipe.");
4695  case Instruction::GetElementPtr: {
4696  // Construct a vector GEP by widening the operands of the scalar GEP as
4697  // necessary. We mark the vector GEP 'inbounds' if appropriate. A GEP
4698  // results in a vector of pointers when at least one operand of the GEP
4699  // is vector-typed. Thus, to keep the representation compact, we only use
4700  // vector-typed operands for loop-varying values.
4701  auto *GEP = cast<GetElementPtrInst>(&I);
4702 
4703  if (VF > 1 && OrigLoop->hasLoopInvariantOperands(GEP)) {
4704  // If we are vectorizing, but the GEP has only loop-invariant operands,
4705  // the GEP we build (by only using vector-typed operands for
4706  // loop-varying values) would be a scalar pointer. Thus, to ensure we
4707  // produce a vector of pointers, we need to either arbitrarily pick an
4708  // operand to broadcast, or broadcast a clone of the original GEP.
4709  // Here, we broadcast a clone of the original.
4710  //
4711  // TODO: If at some point we decide to scalarize instructions having
4712  // loop-invariant operands, this special case will no longer be
4713  // required. We would add the scalarization decision to
4714  // collectLoopScalars() and teach getVectorValue() to broadcast
4715  // the lane-zero scalar value.
4716  auto *Clone = Builder.Insert(GEP->clone());
4717  for (unsigned Part = 0; Part < UF; ++Part) {
4718  Value *EntryPart = Builder.CreateVectorSplat(VF, Clone);
4719  VectorLoopValueMap.setVectorValue(&I, Part, EntryPart);
4720  addMetadata(EntryPart, GEP);
4721  }
4722  } else {
4723  // If the GEP has at least one loop-varying operand, we are sure to
4724  // produce a vector of pointers. But if we are only unrolling, we want
4725  // to produce a scalar GEP for each unroll part. Thus, the GEP we
4726  // produce with the code below will be scalar (if VF == 1) or vector
4727  // (otherwise). Note that for the unroll-only case, we still maintain
4728  // values in the vector mapping with initVector, as we do for other
4729  // instructions.
4730  for (unsigned Part = 0; Part < UF; ++Part) {
4731  // The pointer operand of the new GEP. If it's loop-invariant, we
4732  // won't broadcast it.
4733  auto *Ptr =
4734  OrigLoop->isLoopInvariant(GEP->getPointerOperand())
4735  ? GEP->getPointerOperand()
4736  : getOrCreateVectorValue(GEP->getPointerOperand(), Part);
4737 
4738  // Collect all the indices for the new GEP. If any index is
4739  // loop-invariant, we won't broadcast it.
4740  SmallVector<Value *, 4> Indices;
4741  for (auto &U : make_range(GEP->idx_begin(), GEP->idx_end())) {
4742  if (OrigLoop->isLoopInvariant(U.get()))
4743  Indices.push_back(U.get());
4744  else
4745  Indices.push_back(getOrCreateVectorValue(U.get(), Part));
4746  }
4747 
4748  // Create the new GEP. Note that this GEP may be a scalar if VF == 1,
4749  // but it should be a vector, otherwise.
4750  auto *NewGEP = GEP->isInBounds()
4751  ? Builder.CreateInBoundsGEP(Ptr, Indices)
4752  : Builder.CreateGEP(Ptr, Indices);
4753  assert((VF == 1 || NewGEP->getType()->isVectorTy()) &&
4754  "NewGEP is not a pointer vector");
4755  VectorLoopValueMap.setVectorValue(&I, Part, NewGEP);
4756  addMetadata(NewGEP, GEP);
4757  }
4758  }
4759 
4760  break;
4761  }
4762  case Instruction::UDiv:
4763  case Instruction::SDiv:
4764  case Instruction::SRem:
4765  case Instruction::URem:
4766  case Instruction::Add:
4767  case Instruction::FAdd:
4768  case Instruction::Sub:
4769  case Instruction::FSub:
4770  case Instruction::Mul:
4771  case Instruction::FMul:
4772  case Instruction::FDiv:
4773  case Instruction::FRem:
4774  case Instruction::Shl:
4775  case Instruction::LShr:
4776  case Instruction::AShr:
4777  case Instruction::And:
4778  case Instruction::Or:
4779  case Instruction::Xor: {
4780  // Just widen binops.
4781  auto *BinOp = cast<BinaryOperator>(&I);
4782  setDebugLocFromInst(Builder, BinOp);
4783 
4784  for (unsigned Part = 0; Part < UF; ++Part) {
4785  Value *A = getOrCreateVectorValue(BinOp->getOperand(0), Part);
4786  Value *B = getOrCreateVectorValue(BinOp->getOperand(1), Part);
4787  Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A, B);
4788 
4789  if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
4790  VecOp->copyIRFlags(BinOp);
4791 
4792  // Use this vector value for all users of the original instruction.
4793  VectorLoopValueMap.setVectorValue(&I, Part, V);
4794  addMetadata(V, BinOp);
4795  }
4796 
4797  break;
4798  }
4799  case Instruction::Select: {
4800  // Widen selects.
4801  // If the selector is loop invariant we can create a select
4802  // instruction with a scalar condition. Otherwise, use vector-select.
4803  auto *SE = PSE.getSE();
4804  bool InvariantCond =
4807 
4808  // The condition can be loop invariant but still defined inside the
4809  // loop. This means that we can't just use the original 'cond' value.
4810  // We have to take the 'vectorized' value and pick the first lane.
4811  // Instcombine will make this a no-op.
4812 
4813  auto *ScalarCond = getOrCreateScalarValue(I.getOperand(0), {0, 0});
4814 
4815  for (unsigned Part = 0; Part < UF; ++Part) {
4816  Value *Cond = getOrCreateVectorValue(I.getOperand(0), Part);
4817  Value *Op0 = getOrCreateVectorValue(I.getOperand(1), Part);
4818  Value *Op1 = getOrCreateVectorValue(I.getOperand(2), Part);
4819  Value *Sel =
4820  Builder.CreateSelect(InvariantCond ? ScalarCond : Cond, Op0, Op1);
4821  VectorLoopValueMap.setVectorValue(&I, Part, Sel);
4822  addMetadata(Sel, &I);
4823  }
4824 
4825  break;
4826  }
4827 
4828  case Instruction::ICmp:
4829  case Instruction::FCmp: {
4830  // Widen compares. Generate vector compares.
4831  bool FCmp = (I.getOpcode() == Instruction::FCmp);
4832  auto *Cmp = dyn_cast<CmpInst>(&I);
4834  for (unsigned Part = 0; Part < UF; ++Part) {
4835  Value *A = getOrCreateVectorValue(Cmp->getOperand(0), Part);
4836  Value *B = getOrCreateVectorValue(Cmp->getOperand(1), Part);
4837  Value *C = nullptr;
4838  if (FCmp) {
4839  // Propagate fast math flags.
4841  Builder.setFastMathFlags(Cmp->getFastMathFlags());
4842  C = Builder.CreateFCmp(Cmp->getPredicate(), A, B);
4843  } else {
4844  C = Builder.CreateICmp(Cmp->getPredicate(), A, B);
4845  }
4846  VectorLoopValueMap.setVectorValue(&I, Part, C);
4847  addMetadata(C, &I);
4848  }
4849 
4850  break;
4851  }
4852 
4853  case Instruction::ZExt:
4854  case Instruction::SExt:
4855  case Instruction::FPToUI:
4856  case Instruction::FPToSI:
4857  case Instruction::FPExt:
4858  case Instruction::PtrToInt:
4859  case Instruction::IntToPtr:
4860  case Instruction::SIToFP:
4861  case Instruction::UIToFP:
4862  case Instruction::Trunc:
4863  case Instruction::FPTrunc:
4864  case Instruction::BitCast: {
4865  auto *CI = dyn_cast<CastInst>(&I);
4867 
4868  /// Vectorize casts.
4869  Type *DestTy =
4870  (VF == 1) ? CI->getType() : VectorType::get(CI->getType(), VF);
4871 
4872  for (unsigned Part = 0; Part < UF; ++Part) {
4873  Value *A = getOrCreateVectorValue(CI->getOperand(0), Part);
4874  Value *Cast = Builder.CreateCast(CI->getOpcode(), A, DestTy);
4875  VectorLoopValueMap.setVectorValue(&I, Part, Cast);
4876  addMetadata(Cast, &I);
4877  }
4878  break;
4879  }
4880 
4881  case Instruction::Call: {
4882  // Ignore dbg intrinsics.
4883  if (isa<DbgInfoIntrinsic>(I))
4884  break;
4886 
4887  Module *M = I.getParent()->getParent()->getParent();
4888  auto *CI = cast<CallInst>(&I);
4889 
4890  StringRef FnName = CI->getCalledFunction()->getName();
4891  Function *F = CI->getCalledFunction();
4892  Type *RetTy = ToVectorTy(CI->getType(), VF);
4894  for (Value *ArgOperand : CI->arg_operands())
4895  Tys.push_back(ToVectorTy(ArgOperand->getType(), VF));
4896 
4898 
4899  // The flag shows whether we use Intrinsic or a usual Call for vectorized
4900  // version of the instruction.
4901  // Is it beneficial to perform intrinsic call compared to lib call?
4902  bool NeedToScalarize;
4903  unsigned CallCost = getVectorCallCost(CI, VF, *TTI, TLI, NeedToScalarize);
4904  bool UseVectorIntrinsic =
4905  ID && getVectorIntrinsicCost(CI, VF, *TTI, TLI) <= CallCost;
4906  assert((UseVectorIntrinsic || !NeedToScalarize) &&
4907  "Instruction should be scalarized elsewhere.");
4908 
4909  for (unsigned Part = 0; Part < UF; ++Part) {
4911  for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
4912  Value *Arg = CI->getArgOperand(i);
4913  // Some intrinsics have a scalar argument - don't replace it with a
4914  // vector.
4915  if (!UseVectorIntrinsic || !hasVectorInstrinsicScalarOpd(ID, i))
4916  Arg = getOrCreateVectorValue(CI->getArgOperand(i), Part);
4917  Args.push_back(Arg);
4918  }
4919 
4920  Function *VectorF;
4921  if (UseVectorIntrinsic) {
4922  // Use vector version of the intrinsic.
4923  Type *TysForDecl[] = {CI->getType()};
4924  if (VF > 1)
4925  TysForDecl[0] = VectorType::get(CI->getType()->getScalarType(), VF);
4926  VectorF = Intrinsic::getDeclaration(M, ID, TysForDecl);
4927  } else {
4928  // Use vector version of the library call.
4929  StringRef VFnName = TLI->getVectorizedFunction(FnName, VF);
4930  assert(!VFnName.empty() && "Vector function name is empty.");
4931  VectorF = M->getFunction(VFnName);
4932  if (!VectorF) {
4933  // Generate a declaration
4934  FunctionType *FTy = FunctionType::get(RetTy, Tys, false);
4935  VectorF =
4936  Function::Create(FTy, Function::ExternalLinkage, VFnName, M);
4937  VectorF->copyAttributesFrom(F);
4938  }
4939  }
4940  assert(VectorF && "Can't create vector function.");
4941 
4943  CI->getOperandBundlesAsDefs(OpBundles);
4944  CallInst *V = Builder.CreateCall(VectorF, Args, OpBundles);
4945 
4946  if (isa<FPMathOperator>(V))
4947  V->copyFastMathFlags(CI);
4948 
4949  VectorLoopValueMap.setVectorValue(&I, Part, V);
4950  addMetadata(V, &I);
4951  }
4952 
4953  break;
4954  }
4955 
4956  default:
4957  // This instruction is not vectorized by simple widening.
4958  DEBUG(dbgs() << "LV: Found an unhandled instruction: " << I);
4959  llvm_unreachable("Unhandled instruction!");
4960  } // end of switch.
4961 }
4962 
4964  // Forget the original basic block.
4966 
4967  // Update the dominator tree information.
4969  "Entry does not dominate exit.");
4970 
4976  DEBUG(DT->verifyDomTree());
4977 }
4978 
4979 /// \brief Check whether it is safe to if-convert this phi node.
4980 ///
4981 /// Phi nodes with constant expressions that can trap are not safe to if
4982 /// convert.
4984  for (Instruction &I : *BB) {
4985  auto *Phi = dyn_cast<PHINode>(&I);
4986  if (!Phi)
4987  return true;
4988  for (Value *V : Phi->incoming_values())
4989  if (auto *C = dyn_cast<Constant>(V))
4990  if (C->canTrap())
4991  return false;
4992  }
4993  return true;
4994 }
4995 
4996 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
4997  if (!EnableIfConversion) {
4998  ORE->emit(createMissedAnalysis("IfConversionDisabled")
4999  << "if-conversion is disabled");
5000  return false;
5001  }
5002 
5003  assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
5004 
5005  // A list of pointers that we can safely read and write to.
5006  SmallPtrSet<Value *, 8> SafePointes;
5007 
5008  // Collect safe addresses.
5009  for (BasicBlock *BB : TheLoop->blocks()) {
5010  if (blockNeedsPredication(BB))
5011  continue;
5012 
5013  for (Instruction &I : *BB)
5014  if (auto *Ptr = getPointerOperand(&I))
5015  SafePointes.insert(Ptr);
5016  }
5017 
5018  // Collect the blocks that need predication.
5019  BasicBlock *Header = TheLoop->getHeader();
5020  for (BasicBlock *BB : TheLoop->blocks()) {
5021  // We don't support switch statements inside loops.
5022  if (!isa<BranchInst>(BB->getTerminator())) {
5023  ORE->emit(createMissedAnalysis("LoopContainsSwitch", BB->getTerminator())
5024  << "loop contains a switch statement");
5025  return false;
5026  }
5027 
5028  // We must be able to predicate all blocks that need to be predicated.
5029  if (blockNeedsPredication(BB)) {
5030  if (!blockCanBePredicated(BB, SafePointes)) {
5031  ORE->emit(createMissedAnalysis("NoCFGForSelect", BB->getTerminator())
5032  << "control flow cannot be substituted for a select");
5033  return false;
5034  }
5035  } else if (BB != Header && !canIfConvertPHINodes(BB)) {
5036  ORE->emit(createMissedAnalysis("NoCFGForSelect", BB->getTerminator())
5037  << "control flow cannot be substituted for a select");
5038  return false;
5039  }
5040  }
5041 
5042  // We can if-convert this loop.
5043  return true;
5044 }
5045 
5046 bool LoopVectorizationLegality::canVectorize() {
5047  // Store the result and return it at the end instead of exiting early, in case
5048  // allowExtraAnalysis is used to report multiple reasons for not vectorizing.
5049  bool Result = true;
5050 
5051  bool DoExtraAnalysis = ORE->allowExtraAnalysis(DEBUG_TYPE);
5052  if (DoExtraAnalysis)
5053  // We must have a loop in canonical form. Loops with indirectbr in them cannot
5054  // be canonicalized.
5055  if (!TheLoop->getLoopPreheader()) {
5056  ORE->emit(createMissedAnalysis("CFGNotUnderstood")
5057  << "loop control flow is not understood by vectorizer");
5058  if (DoExtraAnalysis)
5059  Result = false;
5060  else
5061  return false;
5062  }
5063 
5064  // FIXME: The code is currently dead, since the loop gets sent to
5065  // LoopVectorizationLegality is already an innermost loop.
5066  //
5067  // We can only vectorize innermost loops.
5068  if (!TheLoop->empty()) {
5069  ORE->emit(createMissedAnalysis("NotInnermostLoop")
5070  << "loop is not the innermost loop");
5071  if (DoExtraAnalysis)
5072  Result = false;
5073  else
5074  return false;
5075  }
5076 
5077  // We must have a single backedge.
5078  if (TheLoop->getNumBackEdges() != 1) {
5079  ORE->emit(createMissedAnalysis("CFGNotUnderstood")
5080  << "loop control flow is not understood by vectorizer");
5081  if (DoExtraAnalysis)
5082  Result = false;
5083  else
5084  return false;
5085  }
5086 
5087  // We must have a single exiting block.
5088  if (!TheLoop->getExitingBlock()) {
5089  ORE->emit(createMissedAnalysis("CFGNotUnderstood")
5090  << "loop control flow is not understood by vectorizer");
5091  if (DoExtraAnalysis)
5092  Result = false;
5093  else
5094  return false;
5095  }
5096 
5097  // We only handle bottom-tested loops, i.e. loop in which the condition is
5098  // checked at the end of each iteration. With that we can assume that all
5099  // instructions in the loop are executed the same number of times.
5100  if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
5101  ORE->emit(createMissedAnalysis("CFGNotUnderstood")
5102  << "loop control flow is not understood by vectorizer");
5103  if (DoExtraAnalysis)
5104  Result = false;
5105  else
5106  return false;
5107  }
5108 
5109  // We need to have a loop header.
5110  DEBUG(dbgs() << "LV: Found a loop: " << TheLoop->getHeader()->getName()
5111  << '\n');
5112 
5113  // Check if we can if-convert non-single-bb loops.
5114  unsigned NumBlocks = TheLoop->getNumBlocks();
5115  if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
5116  DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
5117  if (DoExtraAnalysis)
5118  Result = false;
5119  else
5120  return false;
5121  }
5122 
5123  // Check if we can vectorize the instructions and CFG in this loop.
5124  if (!canVectorizeInstrs()) {
5125  DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
5126  if (DoExtraAnalysis)
5127  Result = false;
5128  else
5129  return false;
5130  }
5131 
5132  // Go over each instruction and look at memory deps.
5133  if (!canVectorizeMemory()) {
5134  DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
5135  if (DoExtraAnalysis)
5136  Result = false;
5137  else
5138  return false;
5139  }
5140 
5141  DEBUG(dbgs() << "LV: We can vectorize this loop"
5142  << (LAI->getRuntimePointerChecking()->Need
5143  ? " (with a runtime bound check)"
5144  : "")
5145  << "!\n");
5146 
5147  bool UseInterleaved = TTI->enableInterleavedAccessVectorization();
5148 
5149  // If an override option has been passed in for interleaved accesses, use it.
5150  if (EnableInterleavedMemAccesses.getNumOccurrences() > 0)
5151  UseInterleaved = EnableInterleavedMemAccesses;
5152 
5153  // Analyze interleaved memory accesses.
5154  if (UseInterleaved)
5155  InterleaveInfo.analyzeInterleaving(*getSymbolicStrides());
5156 
5157  unsigned SCEVThreshold = VectorizeSCEVCheckThreshold;
5158  if (Hints->getForce() == LoopVectorizeHints::FK_Enabled)
5159  SCEVThreshold = PragmaVectorizeSCEVCheckThreshold;
5160 
5161  if (PSE.getUnionPredicate().getComplexity() > SCEVThreshold) {
5162  ORE->emit(createMissedAnalysis("TooManySCEVRunTimeChecks")
5163  << "Too many SCEV assumptions need to be made and checked "
5164  << "at runtime");
5165  DEBUG(dbgs() << "