1 //===- LoopVectorize.cpp - A Loop Vectorizer ------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
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.
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.
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.
29 //===----------------------------------------------------------------------===//
31 // The reduction-variable vectorization is based on the paper:
32 // D. Nuzman and R. Henderson. Multi-platform Auto-vectorization.
34 // Variable uniformity checks are inspired by:
35 // Karrenberg, R. and Hack, S. Whole Function Vectorization.
37 // Other ideas/concepts are from:
38 // A. Zaks and D. Nuzman. Autovectorization in GCC-two years later.
40 // S. Maleki, Y. Gao, M. Garzaran, T. Wong and D. Padua. An Evaluation of
41 // Vectorizing Compilers.
43 //===----------------------------------------------------------------------===//
45 #include "llvm/Transforms/Vectorize.h"
46 #include "llvm/ADT/DenseMap.h"
47 #include "llvm/ADT/EquivalenceClasses.h"
48 #include "llvm/ADT/Hashing.h"
49 #include "llvm/ADT/MapVector.h"
50 #include "llvm/ADT/SetVector.h"
51 #include "llvm/ADT/SmallPtrSet.h"
52 #include "llvm/ADT/SmallSet.h"
53 #include "llvm/ADT/SmallVector.h"
54 #include "llvm/ADT/Statistic.h"
55 #include "llvm/ADT/StringExtras.h"
56 #include "llvm/Analysis/AliasAnalysis.h"
57 #include "llvm/Analysis/AliasSetTracker.h"
58 #include "llvm/Analysis/AssumptionCache.h"
59 #include "llvm/Analysis/BlockFrequencyInfo.h"
60 #include "llvm/Analysis/CodeMetrics.h"
61 #include "llvm/Analysis/LoopAccessAnalysis.h"
62 #include "llvm/Analysis/LoopInfo.h"
63 #include "llvm/Analysis/LoopIterator.h"
64 #include "llvm/Analysis/LoopPass.h"
65 #include "llvm/Analysis/ScalarEvolution.h"
66 #include "llvm/Analysis/ScalarEvolutionExpander.h"
67 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
68 #include "llvm/Analysis/TargetTransformInfo.h"
69 #include "llvm/Analysis/ValueTracking.h"
70 #include "llvm/IR/Constants.h"
71 #include "llvm/IR/DataLayout.h"
72 #include "llvm/IR/DebugInfo.h"
73 #include "llvm/IR/DerivedTypes.h"
74 #include "llvm/IR/DiagnosticInfo.h"
75 #include "llvm/IR/Dominators.h"
76 #include "llvm/IR/Function.h"
77 #include "llvm/IR/IRBuilder.h"
78 #include "llvm/IR/Instructions.h"
79 #include "llvm/IR/IntrinsicInst.h"
80 #include "llvm/IR/LLVMContext.h"
81 #include "llvm/IR/Module.h"
82 #include "llvm/IR/PatternMatch.h"
83 #include "llvm/IR/Type.h"
84 #include "llvm/IR/Value.h"
85 #include "llvm/IR/ValueHandle.h"
86 #include "llvm/IR/Verifier.h"
87 #include "llvm/Pass.h"
88 #include "llvm/Support/BranchProbability.h"
89 #include "llvm/Support/CommandLine.h"
90 #include "llvm/Support/Debug.h"
91 #include "llvm/Support/raw_ostream.h"
92 #include "llvm/Transforms/Scalar.h"
93 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
94 #include "llvm/Transforms/Utils/Local.h"
95 #include "llvm/Transforms/Utils/VectorUtils.h"
100 using namespace llvm;
101 using namespace llvm::PatternMatch;
103 #define LV_NAME "loop-vectorize"
104 #define DEBUG_TYPE LV_NAME
106 STATISTIC(LoopsVectorized, "Number of loops vectorized");
107 STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization");
109 static cl::opt<unsigned, true>
110 VectorizationFactor("force-vector-width", cl::Hidden,
111 cl::desc("Sets the SIMD width. Zero is autoselect."),
112 cl::location(VectorizerParams::VectorizationFactor));
113 unsigned VectorizerParams::VectorizationFactor = 0;
115 static cl::opt<unsigned, true>
116 VectorizationInterleave("force-vector-interleave", cl::Hidden,
117 cl::desc("Sets the vectorization interleave count. "
118 "Zero is autoselect."),
120 VectorizerParams::VectorizationInterleave));
121 unsigned VectorizerParams::VectorizationInterleave = 0;
124 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
125 cl::desc("Enable if-conversion during vectorization."));
127 /// We don't vectorize loops with a known constant trip count below this number.
128 static cl::opt<unsigned>
129 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
131 cl::desc("Don't vectorize loops with a constant "
132 "trip count that is smaller than this "
135 /// This enables versioning on the strides of symbolically striding memory
136 /// accesses in code like the following.
137 /// for (i = 0; i < N; ++i)
138 /// A[i * Stride1] += B[i * Stride2] ...
140 /// Will be roughly translated to
141 /// if (Stride1 == 1 && Stride2 == 1) {
142 /// for (i = 0; i < N; i+=4)
146 static cl::opt<bool> EnableMemAccessVersioning(
147 "enable-mem-access-versioning", cl::init(true), cl::Hidden,
148 cl::desc("Enable symblic stride memory access versioning"));
150 /// We don't unroll loops with a known constant trip count below this number.
151 static const unsigned TinyTripCountUnrollThreshold = 128;
153 /// When performing memory disambiguation checks at runtime do not make more
154 /// than this number of comparisons.
155 const unsigned VectorizerParams::RuntimeMemoryCheckThreshold = 8;
157 /// Maximum simd width.
158 const unsigned VectorizerParams::MaxVectorWidth = 64;
160 static cl::opt<unsigned> ForceTargetNumScalarRegs(
161 "force-target-num-scalar-regs", cl::init(0), cl::Hidden,
162 cl::desc("A flag that overrides the target's number of scalar registers."));
164 static cl::opt<unsigned> ForceTargetNumVectorRegs(
165 "force-target-num-vector-regs", cl::init(0), cl::Hidden,
166 cl::desc("A flag that overrides the target's number of vector registers."));
168 /// Maximum vectorization interleave count.
169 static const unsigned MaxInterleaveFactor = 16;
171 static cl::opt<unsigned> ForceTargetMaxScalarInterleaveFactor(
172 "force-target-max-scalar-interleave", cl::init(0), cl::Hidden,
173 cl::desc("A flag that overrides the target's max interleave factor for "
176 static cl::opt<unsigned> ForceTargetMaxVectorInterleaveFactor(
177 "force-target-max-vector-interleave", cl::init(0), cl::Hidden,
178 cl::desc("A flag that overrides the target's max interleave factor for "
179 "vectorized loops."));
181 static cl::opt<unsigned> ForceTargetInstructionCost(
182 "force-target-instruction-cost", cl::init(0), cl::Hidden,
183 cl::desc("A flag that overrides the target's expected cost for "
184 "an instruction to a single constant value. Mostly "
185 "useful for getting consistent testing."));
187 static cl::opt<unsigned> SmallLoopCost(
188 "small-loop-cost", cl::init(20), cl::Hidden,
189 cl::desc("The cost of a loop that is considered 'small' by the unroller."));
191 static cl::opt<bool> LoopVectorizeWithBlockFrequency(
192 "loop-vectorize-with-block-frequency", cl::init(false), cl::Hidden,
193 cl::desc("Enable the use of the block frequency analysis to access PGO "
194 "heuristics minimizing code growth in cold regions and being more "
195 "aggressive in hot regions."));
197 // Runtime unroll loops for load/store throughput.
198 static cl::opt<bool> EnableLoadStoreRuntimeUnroll(
199 "enable-loadstore-runtime-unroll", cl::init(true), cl::Hidden,
200 cl::desc("Enable runtime unrolling until load/store ports are saturated"));
202 /// The number of stores in a loop that are allowed to need predication.
203 static cl::opt<unsigned> NumberOfStoresToPredicate(
204 "vectorize-num-stores-pred", cl::init(1), cl::Hidden,
205 cl::desc("Max number of stores to be predicated behind an if."));
207 static cl::opt<bool> EnableIndVarRegisterHeur(
208 "enable-ind-var-reg-heur", cl::init(true), cl::Hidden,
209 cl::desc("Count the induction variable only once when unrolling"));
211 static cl::opt<bool> EnableCondStoresVectorization(
212 "enable-cond-stores-vec", cl::init(false), cl::Hidden,
213 cl::desc("Enable if predication of stores during vectorization."));
215 static cl::opt<unsigned> MaxNestedScalarReductionUF(
216 "max-nested-scalar-reduction-unroll", cl::init(2), cl::Hidden,
217 cl::desc("The maximum unroll factor to use when unrolling a scalar "
218 "reduction in a nested loop."));
222 // Forward declarations.
223 class LoopVectorizationLegality;
224 class LoopVectorizationCostModel;
225 class LoopVectorizeHints;
227 /// InnerLoopVectorizer vectorizes loops which contain only one basic
228 /// block to a specified vectorization factor (VF).
229 /// This class performs the widening of scalars into vectors, or multiple
230 /// scalars. This class also implements the following features:
231 /// * It inserts an epilogue loop for handling loops that don't have iteration
232 /// counts that are known to be a multiple of the vectorization factor.
233 /// * It handles the code generation for reduction variables.
234 /// * Scalarization (implementation using scalars) of un-vectorizable
236 /// InnerLoopVectorizer does not perform any vectorization-legality
237 /// checks, and relies on the caller to check for the different legality
238 /// aspects. The InnerLoopVectorizer relies on the
239 /// LoopVectorizationLegality class to provide information about the induction
240 /// and reduction variables that were found to a given vectorization factor.
241 class InnerLoopVectorizer {
243 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
244 DominatorTree *DT, const DataLayout *DL,
245 const TargetLibraryInfo *TLI, unsigned VecWidth,
246 unsigned UnrollFactor)
247 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), TLI(TLI),
248 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()),
249 Induction(nullptr), OldInduction(nullptr), WidenMap(UnrollFactor),
252 // Perform the actual loop widening (vectorization).
253 void vectorize(LoopVectorizationLegality *L) {
255 // Create a new empty loop. Unlink the old loop and connect the new one.
257 // Widen each instruction in the old loop to a new one in the new loop.
258 // Use the Legality module to find the induction and reduction variables.
260 // Register the new loop and update the analysis passes.
264 virtual ~InnerLoopVectorizer() {}
267 /// A small list of PHINodes.
268 typedef SmallVector<PHINode*, 4> PhiVector;
269 /// When we unroll loops we have multiple vector values for each scalar.
270 /// This data structure holds the unrolled and vectorized values that
271 /// originated from one scalar instruction.
272 typedef SmallVector<Value*, 2> VectorParts;
274 // When we if-convert we need create edge masks. We have to cache values so
275 // that we don't end up with exponential recursion/IR.
276 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
277 VectorParts> EdgeMaskCache;
279 /// \brief Add checks for strides that where assumed to be 1.
281 /// Returns the last check instruction and the first check instruction in the
282 /// pair as (first, last).
283 std::pair<Instruction *, Instruction *> addStrideCheck(Instruction *Loc);
285 /// Create an empty loop, based on the loop ranges of the old loop.
286 void createEmptyLoop();
287 /// Copy and widen the instructions from the old loop.
288 virtual void vectorizeLoop();
290 /// \brief The Loop exit block may have single value PHI nodes where the
291 /// incoming value is 'Undef'. While vectorizing we only handled real values
292 /// that were defined inside the loop. Here we fix the 'undef case'.
296 /// A helper function that computes the predicate of the block BB, assuming
297 /// that the header block of the loop is set to True. It returns the *entry*
298 /// mask for the block BB.
299 VectorParts createBlockInMask(BasicBlock *BB);
300 /// A helper function that computes the predicate of the edge between SRC
302 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
304 /// A helper function to vectorize a single BB within the innermost loop.
305 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
307 /// Vectorize a single PHINode in a block. This method handles the induction
308 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
309 /// arbitrary length vectors.
310 void widenPHIInstruction(Instruction *PN, VectorParts &Entry,
311 unsigned UF, unsigned VF, PhiVector *PV);
313 /// Insert the new loop to the loop hierarchy and pass manager
314 /// and update the analysis passes.
315 void updateAnalysis();
317 /// This instruction is un-vectorizable. Implement it as a sequence
318 /// of scalars. If \p IfPredicateStore is true we need to 'hide' each
319 /// scalarized instruction behind an if block predicated on the control
320 /// dependence of the instruction.
321 virtual void scalarizeInstruction(Instruction *Instr,
322 bool IfPredicateStore=false);
324 /// Vectorize Load and Store instructions,
325 virtual void vectorizeMemoryInstruction(Instruction *Instr);
327 /// Create a broadcast instruction. This method generates a broadcast
328 /// instruction (shuffle) for loop invariant values and for the induction
329 /// value. If this is the induction variable then we extend it to N, N+1, ...
330 /// this is needed because each iteration in the loop corresponds to a SIMD
332 virtual Value *getBroadcastInstrs(Value *V);
334 /// This function adds (StartIdx, StartIdx + Step, StartIdx + 2*Step, ...)
335 /// to each vector element of Val. The sequence starts at StartIndex.
336 virtual Value *getStepVector(Value *Val, int StartIdx, Value *Step);
338 /// When we go over instructions in the basic block we rely on previous
339 /// values within the current basic block or on loop invariant values.
340 /// When we widen (vectorize) values we place them in the map. If the values
341 /// are not within the map, they have to be loop invariant, so we simply
342 /// broadcast them into a vector.
343 VectorParts &getVectorValue(Value *V);
345 /// Generate a shuffle sequence that will reverse the vector Vec.
346 virtual Value *reverseVector(Value *Vec);
348 /// This is a helper class that holds the vectorizer state. It maps scalar
349 /// instructions to vector instructions. When the code is 'unrolled' then
350 /// then a single scalar value is mapped to multiple vector parts. The parts
351 /// are stored in the VectorPart type.
353 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
355 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
357 /// \return True if 'Key' is saved in the Value Map.
358 bool has(Value *Key) const { return MapStorage.count(Key); }
360 /// Initializes a new entry in the map. Sets all of the vector parts to the
361 /// save value in 'Val'.
362 /// \return A reference to a vector with splat values.
363 VectorParts &splat(Value *Key, Value *Val) {
364 VectorParts &Entry = MapStorage[Key];
365 Entry.assign(UF, Val);
369 ///\return A reference to the value that is stored at 'Key'.
370 VectorParts &get(Value *Key) {
371 VectorParts &Entry = MapStorage[Key];
374 assert(Entry.size() == UF);
379 /// The unroll factor. Each entry in the map stores this number of vector
383 /// Map storage. We use std::map and not DenseMap because insertions to a
384 /// dense map invalidates its iterators.
385 std::map<Value *, VectorParts> MapStorage;
388 /// The original loop.
390 /// Scev analysis to use.
399 const DataLayout *DL;
400 /// Target Library Info.
401 const TargetLibraryInfo *TLI;
403 /// The vectorization SIMD factor to use. Each vector will have this many
408 /// The vectorization unroll factor to use. Each scalar is vectorized to this
409 /// many different vector instructions.
412 /// The builder that we use
415 // --- Vectorization state ---
417 /// The vector-loop preheader.
418 BasicBlock *LoopVectorPreHeader;
419 /// The scalar-loop preheader.
420 BasicBlock *LoopScalarPreHeader;
421 /// Middle Block between the vector and the scalar.
422 BasicBlock *LoopMiddleBlock;
423 ///The ExitBlock of the scalar loop.
424 BasicBlock *LoopExitBlock;
425 ///The vector loop body.
426 SmallVector<BasicBlock *, 4> LoopVectorBody;
427 ///The scalar loop body.
428 BasicBlock *LoopScalarBody;
429 /// A list of all bypass blocks. The first block is the entry of the loop.
430 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
432 /// The new Induction variable which was added to the new block.
434 /// The induction variable of the old basic block.
435 PHINode *OldInduction;
436 /// Holds the extended (to the widest induction type) start index.
438 /// Maps scalars to widened vectors.
440 EdgeMaskCache MaskCache;
442 LoopVectorizationLegality *Legal;
445 class InnerLoopUnroller : public InnerLoopVectorizer {
447 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
448 DominatorTree *DT, const DataLayout *DL,
449 const TargetLibraryInfo *TLI, unsigned UnrollFactor) :
450 InnerLoopVectorizer(OrigLoop, SE, LI, DT, DL, TLI, 1, UnrollFactor) { }
453 void scalarizeInstruction(Instruction *Instr,
454 bool IfPredicateStore = false) override;
455 void vectorizeMemoryInstruction(Instruction *Instr) override;
456 Value *getBroadcastInstrs(Value *V) override;
457 Value *getStepVector(Value *Val, int StartIdx, Value *Step) override;
458 Value *reverseVector(Value *Vec) override;
461 /// \brief Look for a meaningful debug location on the instruction or it's
463 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
468 if (I->getDebugLoc() != Empty)
471 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
472 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
473 if (OpInst->getDebugLoc() != Empty)
480 /// \brief Set the debug location in the builder using the debug location in the
482 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
483 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
484 B.SetCurrentDebugLocation(Inst->getDebugLoc());
486 B.SetCurrentDebugLocation(DebugLoc());
490 /// \return string containing a file name and a line # for the given loop.
491 static std::string getDebugLocString(const Loop *L) {
494 raw_string_ostream OS(Result);
495 const DebugLoc LoopDbgLoc = L->getStartLoc();
496 if (!LoopDbgLoc.isUnknown())
497 LoopDbgLoc.print(L->getHeader()->getContext(), OS);
499 // Just print the module name.
500 OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier();
507 /// \brief Propagate known metadata from one instruction to another.
508 static void propagateMetadata(Instruction *To, const Instruction *From) {
509 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
510 From->getAllMetadataOtherThanDebugLoc(Metadata);
512 for (auto M : Metadata) {
513 unsigned Kind = M.first;
515 // These are safe to transfer (this is safe for TBAA, even when we
516 // if-convert, because should that metadata have had a control dependency
517 // on the condition, and thus actually aliased with some other
518 // non-speculated memory access when the condition was false, this would be
519 // caught by the runtime overlap checks).
520 if (Kind != LLVMContext::MD_tbaa &&
521 Kind != LLVMContext::MD_alias_scope &&
522 Kind != LLVMContext::MD_noalias &&
523 Kind != LLVMContext::MD_fpmath)
526 To->setMetadata(Kind, M.second);
530 /// \brief Propagate known metadata from one instruction to a vector of others.
531 static void propagateMetadata(SmallVectorImpl<Value *> &To, const Instruction *From) {
533 if (Instruction *I = dyn_cast<Instruction>(V))
534 propagateMetadata(I, From);
537 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
538 /// to what vectorization factor.
539 /// This class does not look at the profitability of vectorization, only the
540 /// legality. This class has two main kinds of checks:
541 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
542 /// will change the order of memory accesses in a way that will change the
543 /// correctness of the program.
544 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
545 /// checks for a number of different conditions, such as the availability of a
546 /// single induction variable, that all types are supported and vectorize-able,
547 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
548 /// This class is also used by InnerLoopVectorizer for identifying
549 /// induction variable and the different reduction variables.
550 class LoopVectorizationLegality {
552 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, const DataLayout *DL,
553 DominatorTree *DT, TargetLibraryInfo *TLI,
554 AliasAnalysis *AA, Function *F,
555 const TargetTransformInfo *TTI,
556 LoopAccessAnalysis *LAA)
557 : NumPredStores(0), TheLoop(L), SE(SE), DL(DL),
558 TLI(TLI), TheFunction(F), TTI(TTI), DT(DT), LAA(LAA), LAI(nullptr),
559 Induction(nullptr), WidestIndTy(nullptr), HasFunNoNaNAttr(false) {}
561 /// This enum represents the kinds of reductions that we support.
563 RK_NoReduction, ///< Not a reduction.
564 RK_IntegerAdd, ///< Sum of integers.
565 RK_IntegerMult, ///< Product of integers.
566 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
567 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
568 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
569 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
570 RK_FloatAdd, ///< Sum of floats.
571 RK_FloatMult, ///< Product of floats.
572 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()).
575 /// This enum represents the kinds of inductions that we support.
577 IK_NoInduction, ///< Not an induction variable.
578 IK_IntInduction, ///< Integer induction variable. Step = C.
579 IK_PtrInduction ///< Pointer induction var. Step = C / sizeof(elem).
582 // This enum represents the kind of minmax reduction.
583 enum MinMaxReductionKind {
593 /// This struct holds information about reduction variables.
594 struct ReductionDescriptor {
595 ReductionDescriptor() : StartValue(nullptr), LoopExitInstr(nullptr),
596 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
598 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
599 MinMaxReductionKind MK)
600 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
602 // The starting value of the reduction.
603 // It does not have to be zero!
604 TrackingVH<Value> StartValue;
605 // The instruction who's value is used outside the loop.
606 Instruction *LoopExitInstr;
607 // The kind of the reduction.
609 // If this a min/max reduction the kind of reduction.
610 MinMaxReductionKind MinMaxKind;
613 /// This POD struct holds information about a potential reduction operation.
614 struct ReductionInstDesc {
615 ReductionInstDesc(bool IsRedux, Instruction *I) :
616 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
618 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
619 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
621 // Is this instruction a reduction candidate.
623 // The last instruction in a min/max pattern (select of the select(icmp())
624 // pattern), or the current reduction instruction otherwise.
625 Instruction *PatternLastInst;
626 // If this is a min/max pattern the comparison predicate.
627 MinMaxReductionKind MinMaxKind;
630 /// A struct for saving information about induction variables.
631 struct InductionInfo {
632 InductionInfo(Value *Start, InductionKind K, ConstantInt *Step)
633 : StartValue(Start), IK(K), StepValue(Step) {
634 assert(IK != IK_NoInduction && "Not an induction");
635 assert(StartValue && "StartValue is null");
636 assert(StepValue && !StepValue->isZero() && "StepValue is zero");
637 assert((IK != IK_PtrInduction || StartValue->getType()->isPointerTy()) &&
638 "StartValue is not a pointer for pointer induction");
639 assert((IK != IK_IntInduction || StartValue->getType()->isIntegerTy()) &&
640 "StartValue is not an integer for integer induction");
641 assert(StepValue->getType()->isIntegerTy() &&
642 "StepValue is not an integer");
645 : StartValue(nullptr), IK(IK_NoInduction), StepValue(nullptr) {}
647 /// Get the consecutive direction. Returns:
648 /// 0 - unknown or non-consecutive.
649 /// 1 - consecutive and increasing.
650 /// -1 - consecutive and decreasing.
651 int getConsecutiveDirection() const {
652 if (StepValue && (StepValue->isOne() || StepValue->isMinusOne()))
653 return StepValue->getSExtValue();
657 /// Compute the transformed value of Index at offset StartValue using step
659 /// For integer induction, returns StartValue + Index * StepValue.
660 /// For pointer induction, returns StartValue[Index * StepValue].
661 /// FIXME: The newly created binary instructions should contain nsw/nuw
662 /// flags, which can be found from the original scalar operations.
663 Value *transform(IRBuilder<> &B, Value *Index) const {
665 case IK_IntInduction:
666 assert(Index->getType() == StartValue->getType() &&
667 "Index type does not match StartValue type");
668 if (StepValue->isMinusOne())
669 return B.CreateSub(StartValue, Index);
670 if (!StepValue->isOne())
671 Index = B.CreateMul(Index, StepValue);
672 return B.CreateAdd(StartValue, Index);
674 case IK_PtrInduction:
675 if (StepValue->isMinusOne())
676 Index = B.CreateNeg(Index);
677 else if (!StepValue->isOne())
678 Index = B.CreateMul(Index, StepValue);
679 return B.CreateGEP(StartValue, Index);
684 llvm_unreachable("invalid enum");
688 TrackingVH<Value> StartValue;
692 ConstantInt *StepValue;
695 /// ReductionList contains the reduction descriptors for all
696 /// of the reductions that were found in the loop.
697 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
699 /// InductionList saves induction variables and maps them to the
700 /// induction descriptor.
701 typedef MapVector<PHINode*, InductionInfo> InductionList;
703 /// Returns true if it is legal to vectorize this loop.
704 /// This does not mean that it is profitable to vectorize this
705 /// loop, only that it is legal to do so.
708 /// Returns the Induction variable.
709 PHINode *getInduction() { return Induction; }
711 /// Returns the reduction variables found in the loop.
712 ReductionList *getReductionVars() { return &Reductions; }
714 /// Returns the induction variables found in the loop.
715 InductionList *getInductionVars() { return &Inductions; }
717 /// Returns the widest induction type.
718 Type *getWidestInductionType() { return WidestIndTy; }
720 /// Returns True if V is an induction variable in this loop.
721 bool isInductionVariable(const Value *V);
723 /// Return true if the block BB needs to be predicated in order for the loop
724 /// to be vectorized.
725 bool blockNeedsPredication(BasicBlock *BB);
727 /// Check if this pointer is consecutive when vectorizing. This happens
728 /// when the last index of the GEP is the induction variable, or that the
729 /// pointer itself is an induction variable.
730 /// This check allows us to vectorize A[idx] into a wide load/store.
732 /// 0 - Stride is unknown or non-consecutive.
733 /// 1 - Address is consecutive.
734 /// -1 - Address is consecutive, and decreasing.
735 int isConsecutivePtr(Value *Ptr);
737 /// Returns true if the value V is uniform within the loop.
738 bool isUniform(Value *V);
740 /// Returns true if this instruction will remain scalar after vectorization.
741 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
743 /// Returns the information that we collected about runtime memory check.
744 LoopAccessInfo::RuntimePointerCheck *getRuntimePointerCheck() {
745 return LAI->getRuntimePointerCheck();
748 LoopAccessInfo *getLAI() {
752 /// This function returns the identity element (or neutral element) for
754 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
756 unsigned getMaxSafeDepDistBytes() { return LAI->getMaxSafeDepDistBytes(); }
758 bool hasStride(Value *V) { return StrideSet.count(V); }
759 bool mustCheckStrides() { return !StrideSet.empty(); }
760 SmallPtrSet<Value *, 8>::iterator strides_begin() {
761 return StrideSet.begin();
763 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
765 /// Returns true if the target machine supports masked store operation
766 /// for the given \p DataType and kind of access to \p Ptr.
767 bool isLegalMaskedStore(Type *DataType, Value *Ptr) {
768 return TTI->isLegalMaskedStore(DataType, isConsecutivePtr(Ptr));
770 /// Returns true if the target machine supports masked load operation
771 /// for the given \p DataType and kind of access to \p Ptr.
772 bool isLegalMaskedLoad(Type *DataType, Value *Ptr) {
773 return TTI->isLegalMaskedLoad(DataType, isConsecutivePtr(Ptr));
775 /// Returns true if vector representation of the instruction \p I
777 bool isMaskRequired(const Instruction* I) {
778 return (MaskedOp.count(I) != 0);
780 unsigned getNumStores() const {
781 return LAI->getNumStores();
783 unsigned getNumLoads() const {
784 return LAI->getNumLoads();
786 unsigned getNumPredStores() const {
787 return NumPredStores;
790 /// Check if a single basic block loop is vectorizable.
791 /// At this point we know that this is a loop with a constant trip count
792 /// and we only need to check individual instructions.
793 bool canVectorizeInstrs();
795 /// When we vectorize loops we may change the order in which
796 /// we read and write from memory. This method checks if it is
797 /// legal to vectorize the code, considering only memory constrains.
798 /// Returns true if the loop is vectorizable
799 bool canVectorizeMemory();
801 /// Return true if we can vectorize this loop using the IF-conversion
803 bool canVectorizeWithIfConvert();
805 /// Collect the variables that need to stay uniform after vectorization.
806 void collectLoopUniforms();
808 /// Return true if all of the instructions in the block can be speculatively
809 /// executed. \p SafePtrs is a list of addresses that are known to be legal
810 /// and we know that we can read from them without segfault.
811 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
813 /// Returns True, if 'Phi' is the kind of reduction variable for type
814 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
815 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
816 /// Returns a struct describing if the instruction 'I' can be a reduction
817 /// variable of type 'Kind'. If the reduction is a min/max pattern of
818 /// select(icmp()) this function advances the instruction pointer 'I' from the
819 /// compare instruction to the select instruction and stores this pointer in
820 /// 'PatternLastInst' member of the returned struct.
821 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
822 ReductionInstDesc &Desc);
823 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
824 /// pattern corresponding to a min(X, Y) or max(X, Y).
825 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
826 ReductionInstDesc &Prev);
827 /// Returns the induction kind of Phi and record the step. This function may
828 /// return NoInduction if the PHI is not an induction variable.
829 InductionKind isInductionVariable(PHINode *Phi, ConstantInt *&StepValue);
831 /// \brief Collect memory access with loop invariant strides.
833 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
835 void collectStridedAccess(Value *LoadOrStoreInst);
837 /// Report an analysis message to assist the user in diagnosing loops that are
839 void emitAnalysis(VectorizationReport &Message) {
840 VectorizationReport::emitAnalysis(Message, TheFunction, TheLoop, LV_NAME);
843 unsigned NumPredStores;
845 /// The loop that we evaluate.
849 /// DataLayout analysis.
850 const DataLayout *DL;
851 /// Target Library Info.
852 TargetLibraryInfo *TLI;
854 Function *TheFunction;
855 /// Target Transform Info
856 const TargetTransformInfo *TTI;
859 // LoopAccess analysis.
860 LoopAccessAnalysis *LAA;
861 // And the loop-accesses info corresponding to this loop. This pointer is
862 // null until canVectorizeMemory sets it up.
865 // --- vectorization state --- //
867 /// Holds the integer induction variable. This is the counter of the
870 /// Holds the reduction variables.
871 ReductionList Reductions;
872 /// Holds all of the induction variables that we found in the loop.
873 /// Notice that inductions don't need to start at zero and that induction
874 /// variables can be pointers.
875 InductionList Inductions;
876 /// Holds the widest induction type encountered.
879 /// Allowed outside users. This holds the reduction
880 /// vars which can be accessed from outside the loop.
881 SmallPtrSet<Value*, 4> AllowedExit;
882 /// This set holds the variables which are known to be uniform after
884 SmallPtrSet<Instruction*, 4> Uniforms;
886 /// Can we assume the absence of NaNs.
887 bool HasFunNoNaNAttr;
889 ValueToValueMap Strides;
890 SmallPtrSet<Value *, 8> StrideSet;
892 /// While vectorizing these instructions we have to generate a
893 /// call to the appropriate masked intrinsic
894 SmallPtrSet<const Instruction*, 8> MaskedOp;
897 /// LoopVectorizationCostModel - estimates the expected speedups due to
899 /// In many cases vectorization is not profitable. This can happen because of
900 /// a number of reasons. In this class we mainly attempt to predict the
901 /// expected speedup/slowdowns due to the supported instruction set. We use the
902 /// TargetTransformInfo to query the different backends for the cost of
903 /// different operations.
904 class LoopVectorizationCostModel {
906 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
907 LoopVectorizationLegality *Legal,
908 const TargetTransformInfo &TTI,
909 const DataLayout *DL, const TargetLibraryInfo *TLI,
910 AssumptionCache *AC, const Function *F,
911 const LoopVectorizeHints *Hints)
912 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI),
913 TheFunction(F), Hints(Hints) {
914 CodeMetrics::collectEphemeralValues(L, AC, EphValues);
917 /// Information about vectorization costs
918 struct VectorizationFactor {
919 unsigned Width; // Vector width with best cost
920 unsigned Cost; // Cost of the loop with that width
922 /// \return The most profitable vectorization factor and the cost of that VF.
923 /// This method checks every power of two up to VF. If UserVF is not ZERO
924 /// then this vectorization factor will be selected if vectorization is
926 VectorizationFactor selectVectorizationFactor(bool OptForSize);
928 /// \return The size (in bits) of the widest type in the code that
929 /// needs to be vectorized. We ignore values that remain scalar such as
930 /// 64 bit loop indices.
931 unsigned getWidestType();
933 /// \return The most profitable unroll factor.
934 /// If UserUF is non-zero then this method finds the best unroll-factor
935 /// based on register pressure and other parameters.
936 /// VF and LoopCost are the selected vectorization factor and the cost of the
938 unsigned selectUnrollFactor(bool OptForSize, unsigned VF, unsigned LoopCost);
940 /// \brief A struct that represents some properties of the register usage
942 struct RegisterUsage {
943 /// Holds the number of loop invariant values that are used in the loop.
944 unsigned LoopInvariantRegs;
945 /// Holds the maximum number of concurrent live intervals in the loop.
946 unsigned MaxLocalUsers;
947 /// Holds the number of instructions in the loop.
948 unsigned NumInstructions;
951 /// \return information about the register usage of the loop.
952 RegisterUsage calculateRegisterUsage();
955 /// Returns the expected execution cost. The unit of the cost does
956 /// not matter because we use the 'cost' units to compare different
957 /// vector widths. The cost that is returned is *not* normalized by
958 /// the factor width.
959 unsigned expectedCost(unsigned VF);
961 /// Returns the execution time cost of an instruction for a given vector
962 /// width. Vector width of one means scalar.
963 unsigned getInstructionCost(Instruction *I, unsigned VF);
965 /// A helper function for converting Scalar types to vector types.
966 /// If the incoming type is void, we return void. If the VF is 1, we return
968 static Type* ToVectorTy(Type *Scalar, unsigned VF);
970 /// Returns whether the instruction is a load or store and will be a emitted
971 /// as a vector operation.
972 bool isConsecutiveLoadOrStore(Instruction *I);
974 /// Report an analysis message to assist the user in diagnosing loops that are
976 void emitAnalysis(VectorizationReport &Message) {
977 VectorizationReport::emitAnalysis(Message, TheFunction, TheLoop, LV_NAME);
980 /// Values used only by @llvm.assume calls.
981 SmallPtrSet<const Value *, 32> EphValues;
983 /// The loop that we evaluate.
987 /// Loop Info analysis.
989 /// Vectorization legality.
990 LoopVectorizationLegality *Legal;
991 /// Vector target information.
992 const TargetTransformInfo &TTI;
993 /// Target data layout information.
994 const DataLayout *DL;
995 /// Target Library Info.
996 const TargetLibraryInfo *TLI;
997 const Function *TheFunction;
998 // Loop Vectorize Hint.
999 const LoopVectorizeHints *Hints;
1002 /// Utility class for getting and setting loop vectorizer hints in the form
1003 /// of loop metadata.
1004 /// This class keeps a number of loop annotations locally (as member variables)
1005 /// and can, upon request, write them back as metadata on the loop. It will
1006 /// initially scan the loop for existing metadata, and will update the local
1007 /// values based on information in the loop.
1008 /// We cannot write all values to metadata, as the mere presence of some info,
1009 /// for example 'force', means a decision has been made. So, we need to be
1010 /// careful NOT to add them if the user hasn't specifically asked so.
1011 class LoopVectorizeHints {
1018 /// Hint - associates name and validation with the hint value.
1021 unsigned Value; // This may have to change for non-numeric values.
1024 Hint(const char * Name, unsigned Value, HintKind Kind)
1025 : Name(Name), Value(Value), Kind(Kind) { }
1027 bool validate(unsigned Val) {
1030 return isPowerOf2_32(Val) && Val <= VectorizerParams::MaxVectorWidth;
1032 return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
1040 /// Vectorization width.
1042 /// Vectorization interleave factor.
1044 /// Vectorization forced
1047 /// Return the loop metadata prefix.
1048 static StringRef Prefix() { return "llvm.loop."; }
1052 FK_Undefined = -1, ///< Not selected.
1053 FK_Disabled = 0, ///< Forcing disabled.
1054 FK_Enabled = 1, ///< Forcing enabled.
1057 LoopVectorizeHints(const Loop *L, bool DisableInterleaving)
1058 : Width("vectorize.width", VectorizationFactor, HK_WIDTH),
1059 Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
1060 Force("vectorize.enable", FK_Undefined, HK_FORCE),
1062 // Populate values with existing loop metadata.
1063 getHintsFromMetadata();
1065 // force-vector-interleave overrides DisableInterleaving.
1066 if (VectorizationInterleave.getNumOccurrences() > 0)
1067 Interleave.Value = VectorizationInterleave;
1069 DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs()
1070 << "LV: Interleaving disabled by the pass manager\n");
1073 /// Mark the loop L as already vectorized by setting the width to 1.
1074 void setAlreadyVectorized() {
1075 Width.Value = Interleave.Value = 1;
1076 Hint Hints[] = {Width, Interleave};
1077 writeHintsToMetadata(Hints);
1080 /// Dumps all the hint information.
1081 std::string emitRemark() const {
1082 VectorizationReport R;
1083 if (Force.Value == LoopVectorizeHints::FK_Disabled)
1084 R << "vectorization is explicitly disabled";
1086 R << "use -Rpass-analysis=loop-vectorize for more info";
1087 if (Force.Value == LoopVectorizeHints::FK_Enabled) {
1088 R << " (Force=true";
1089 if (Width.Value != 0)
1090 R << ", Vector Width=" << Width.Value;
1091 if (Interleave.Value != 0)
1092 R << ", Interleave Count=" << Interleave.Value;
1100 unsigned getWidth() const { return Width.Value; }
1101 unsigned getInterleave() const { return Interleave.Value; }
1102 enum ForceKind getForce() const { return (ForceKind)Force.Value; }
1105 /// Find hints specified in the loop metadata and update local values.
1106 void getHintsFromMetadata() {
1107 MDNode *LoopID = TheLoop->getLoopID();
1111 // First operand should refer to the loop id itself.
1112 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
1113 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
1115 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1116 const MDString *S = nullptr;
1117 SmallVector<Metadata *, 4> Args;
1119 // The expected hint is either a MDString or a MDNode with the first
1120 // operand a MDString.
1121 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
1122 if (!MD || MD->getNumOperands() == 0)
1124 S = dyn_cast<MDString>(MD->getOperand(0));
1125 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
1126 Args.push_back(MD->getOperand(i));
1128 S = dyn_cast<MDString>(LoopID->getOperand(i));
1129 assert(Args.size() == 0 && "too many arguments for MDString");
1135 // Check if the hint starts with the loop metadata prefix.
1136 StringRef Name = S->getString();
1137 if (Args.size() == 1)
1138 setHint(Name, Args[0]);
1142 /// Checks string hint with one operand and set value if valid.
1143 void setHint(StringRef Name, Metadata *Arg) {
1144 if (!Name.startswith(Prefix()))
1146 Name = Name.substr(Prefix().size(), StringRef::npos);
1148 const ConstantInt *C = mdconst::dyn_extract<ConstantInt>(Arg);
1150 unsigned Val = C->getZExtValue();
1152 Hint *Hints[] = {&Width, &Interleave, &Force};
1153 for (auto H : Hints) {
1154 if (Name == H->Name) {
1155 if (H->validate(Val))
1158 DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n");
1164 /// Create a new hint from name / value pair.
1165 MDNode *createHintMetadata(StringRef Name, unsigned V) const {
1166 LLVMContext &Context = TheLoop->getHeader()->getContext();
1167 Metadata *MDs[] = {MDString::get(Context, Name),
1168 ConstantAsMetadata::get(
1169 ConstantInt::get(Type::getInt32Ty(Context), V))};
1170 return MDNode::get(Context, MDs);
1173 /// Matches metadata with hint name.
1174 bool matchesHintMetadataName(MDNode *Node, ArrayRef<Hint> HintTypes) {
1175 MDString* Name = dyn_cast<MDString>(Node->getOperand(0));
1179 for (auto H : HintTypes)
1180 if (Name->getString().endswith(H.Name))
1185 /// Sets current hints into loop metadata, keeping other values intact.
1186 void writeHintsToMetadata(ArrayRef<Hint> HintTypes) {
1187 if (HintTypes.size() == 0)
1190 // Reserve the first element to LoopID (see below).
1191 SmallVector<Metadata *, 4> MDs(1);
1192 // If the loop already has metadata, then ignore the existing operands.
1193 MDNode *LoopID = TheLoop->getLoopID();
1195 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1196 MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
1197 // If node in update list, ignore old value.
1198 if (!matchesHintMetadataName(Node, HintTypes))
1199 MDs.push_back(Node);
1203 // Now, add the missing hints.
1204 for (auto H : HintTypes)
1205 MDs.push_back(createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
1207 // Replace current metadata node with new one.
1208 LLVMContext &Context = TheLoop->getHeader()->getContext();
1209 MDNode *NewLoopID = MDNode::get(Context, MDs);
1210 // Set operand 0 to refer to the loop id itself.
1211 NewLoopID->replaceOperandWith(0, NewLoopID);
1213 TheLoop->setLoopID(NewLoopID);
1216 /// The loop these hints belong to.
1217 const Loop *TheLoop;
1220 static void emitMissedWarning(Function *F, Loop *L,
1221 const LoopVectorizeHints &LH) {
1222 emitOptimizationRemarkMissed(F->getContext(), DEBUG_TYPE, *F,
1223 L->getStartLoc(), LH.emitRemark());
1225 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1226 if (LH.getWidth() != 1)
1227 emitLoopVectorizeWarning(
1228 F->getContext(), *F, L->getStartLoc(),
1229 "failed explicitly specified loop vectorization");
1230 else if (LH.getInterleave() != 1)
1231 emitLoopInterleaveWarning(
1232 F->getContext(), *F, L->getStartLoc(),
1233 "failed explicitly specified loop interleaving");
1237 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1239 return V.push_back(&L);
1241 for (Loop *InnerL : L)
1242 addInnerLoop(*InnerL, V);
1245 /// The LoopVectorize Pass.
1246 struct LoopVectorize : public FunctionPass {
1247 /// Pass identification, replacement for typeid
1250 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1252 DisableUnrolling(NoUnrolling),
1253 AlwaysVectorize(AlwaysVectorize) {
1254 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1257 ScalarEvolution *SE;
1258 const DataLayout *DL;
1260 TargetTransformInfo *TTI;
1262 BlockFrequencyInfo *BFI;
1263 TargetLibraryInfo *TLI;
1265 AssumptionCache *AC;
1266 LoopAccessAnalysis *LAA;
1267 bool DisableUnrolling;
1268 bool AlwaysVectorize;
1270 BlockFrequency ColdEntryFreq;
1272 bool runOnFunction(Function &F) override {
1273 SE = &getAnalysis<ScalarEvolution>();
1274 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1275 DL = DLP ? &DLP->getDataLayout() : nullptr;
1276 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1277 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1278 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1279 BFI = &getAnalysis<BlockFrequencyInfo>();
1280 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
1281 TLI = TLIP ? &TLIP->getTLI() : nullptr;
1282 AA = &getAnalysis<AliasAnalysis>();
1283 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1284 LAA = &getAnalysis<LoopAccessAnalysis>();
1286 // Compute some weights outside of the loop over the loops. Compute this
1287 // using a BranchProbability to re-use its scaling math.
1288 const BranchProbability ColdProb(1, 5); // 20%
1289 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1291 // If the target claims to have no vector registers don't attempt
1293 if (!TTI->getNumberOfRegisters(true))
1297 DEBUG(dbgs() << "\nLV: Not vectorizing " << F.getName()
1298 << ": Missing data layout\n");
1302 // Build up a worklist of inner-loops to vectorize. This is necessary as
1303 // the act of vectorizing or partially unrolling a loop creates new loops
1304 // and can invalidate iterators across the loops.
1305 SmallVector<Loop *, 8> Worklist;
1308 addInnerLoop(*L, Worklist);
1310 LoopsAnalyzed += Worklist.size();
1312 // Now walk the identified inner loops.
1313 bool Changed = false;
1314 while (!Worklist.empty())
1315 Changed |= processLoop(Worklist.pop_back_val());
1317 // Process each loop nest in the function.
1321 bool processLoop(Loop *L) {
1322 assert(L->empty() && "Only process inner loops.");
1325 const std::string DebugLocStr = getDebugLocString(L);
1328 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1329 << L->getHeader()->getParent()->getName() << "\" from "
1330 << DebugLocStr << "\n");
1332 LoopVectorizeHints Hints(L, DisableUnrolling);
1334 DEBUG(dbgs() << "LV: Loop hints:"
1336 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
1338 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
1340 : "?")) << " width=" << Hints.getWidth()
1341 << " unroll=" << Hints.getInterleave() << "\n");
1343 // Function containing loop
1344 Function *F = L->getHeader()->getParent();
1346 // Looking at the diagnostic output is the only way to determine if a loop
1347 // was vectorized (other than looking at the IR or machine code), so it
1348 // is important to generate an optimization remark for each loop. Most of
1349 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
1350 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
1351 // less verbose reporting vectorized loops and unvectorized loops that may
1352 // benefit from vectorization, respectively.
1354 if (Hints.getForce() == LoopVectorizeHints::FK_Disabled) {
1355 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
1356 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1357 L->getStartLoc(), Hints.emitRemark());
1361 if (!AlwaysVectorize && Hints.getForce() != LoopVectorizeHints::FK_Enabled) {
1362 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
1363 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1364 L->getStartLoc(), Hints.emitRemark());
1368 if (Hints.getWidth() == 1 && Hints.getInterleave() == 1) {
1369 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
1370 emitOptimizationRemarkAnalysis(
1371 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1372 "loop not vectorized: vector width and interleave count are "
1373 "explicitly set to 1");
1377 // Check the loop for a trip count threshold:
1378 // do not vectorize loops with a tiny trip count.
1379 const unsigned TC = SE->getSmallConstantTripCount(L);
1380 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1381 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
1382 << "This loop is not worth vectorizing.");
1383 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
1384 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
1386 DEBUG(dbgs() << "\n");
1387 emitOptimizationRemarkAnalysis(
1388 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1389 "vectorization is not beneficial and is not explicitly forced");
1394 // Check if it is legal to vectorize the loop.
1395 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI, AA, F, TTI, LAA);
1396 if (!LVL.canVectorize()) {
1397 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1398 emitMissedWarning(F, L, Hints);
1402 // Use the cost model.
1403 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI, AC, F,
1406 // Check the function attributes to find out if this function should be
1407 // optimized for size.
1408 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1409 F->hasFnAttribute(Attribute::OptimizeForSize);
1411 // Compute the weighted frequency of this loop being executed and see if it
1412 // is less than 20% of the function entry baseline frequency. Note that we
1413 // always have a canonical loop here because we think we *can* vectoriez.
1414 // FIXME: This is hidden behind a flag due to pervasive problems with
1415 // exactly what block frequency models.
1416 if (LoopVectorizeWithBlockFrequency) {
1417 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1418 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1419 LoopEntryFreq < ColdEntryFreq)
1423 // Check the function attributes to see if implicit floats are allowed.a
1424 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1425 // an integer loop and the vector instructions selected are purely integer
1426 // vector instructions?
1427 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1428 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1429 "attribute is used.\n");
1430 emitOptimizationRemarkAnalysis(
1431 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1432 "loop not vectorized due to NoImplicitFloat attribute");
1433 emitMissedWarning(F, L, Hints);
1437 // Select the optimal vectorization factor.
1438 const LoopVectorizationCostModel::VectorizationFactor VF =
1439 CM.selectVectorizationFactor(OptForSize);
1441 // Select the unroll factor.
1443 CM.selectUnrollFactor(OptForSize, VF.Width, VF.Cost);
1445 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1446 << DebugLocStr << '\n');
1447 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n');
1449 if (VF.Width == 1) {
1450 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial\n");
1453 emitOptimizationRemarkAnalysis(
1454 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1455 "not beneficial to vectorize and user disabled interleaving");
1458 DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n");
1460 // Report the unrolling decision.
1461 emitOptimizationRemark(F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1462 Twine("unrolled with interleaving factor " +
1464 " (vectorization not beneficial)"));
1466 // We decided not to vectorize, but we may want to unroll.
1468 InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
1469 Unroller.vectorize(&LVL);
1471 // If we decided that it is *legal* to vectorize the loop then do it.
1472 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
1476 // Report the vectorization decision.
1477 emitOptimizationRemark(
1478 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1479 Twine("vectorized loop (vectorization factor: ") + Twine(VF.Width) +
1480 ", unrolling interleave factor: " + Twine(UF) + ")");
1483 // Mark the loop as already vectorized to avoid vectorizing again.
1484 Hints.setAlreadyVectorized();
1486 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1490 void getAnalysisUsage(AnalysisUsage &AU) const override {
1491 AU.addRequired<AssumptionCacheTracker>();
1492 AU.addRequiredID(LoopSimplifyID);
1493 AU.addRequiredID(LCSSAID);
1494 AU.addRequired<BlockFrequencyInfo>();
1495 AU.addRequired<DominatorTreeWrapperPass>();
1496 AU.addRequired<LoopInfoWrapperPass>();
1497 AU.addRequired<ScalarEvolution>();
1498 AU.addRequired<TargetTransformInfoWrapperPass>();
1499 AU.addRequired<AliasAnalysis>();
1500 AU.addRequired<LoopAccessAnalysis>();
1501 AU.addPreserved<LoopInfoWrapperPass>();
1502 AU.addPreserved<DominatorTreeWrapperPass>();
1503 AU.addPreserved<AliasAnalysis>();
1508 } // end anonymous namespace
1510 //===----------------------------------------------------------------------===//
1511 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1512 // LoopVectorizationCostModel.
1513 //===----------------------------------------------------------------------===//
1515 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1516 // We need to place the broadcast of invariant variables outside the loop.
1517 Instruction *Instr = dyn_cast<Instruction>(V);
1519 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1520 Instr->getParent()) != LoopVectorBody.end());
1521 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1523 // Place the code for broadcasting invariant variables in the new preheader.
1524 IRBuilder<>::InsertPointGuard Guard(Builder);
1526 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1528 // Broadcast the scalar into all locations in the vector.
1529 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1534 Value *InnerLoopVectorizer::getStepVector(Value *Val, int StartIdx,
1536 assert(Val->getType()->isVectorTy() && "Must be a vector");
1537 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1538 "Elem must be an integer");
1539 assert(Step->getType() == Val->getType()->getScalarType() &&
1540 "Step has wrong type");
1541 // Create the types.
1542 Type *ITy = Val->getType()->getScalarType();
1543 VectorType *Ty = cast<VectorType>(Val->getType());
1544 int VLen = Ty->getNumElements();
1545 SmallVector<Constant*, 8> Indices;
1547 // Create a vector of consecutive numbers from zero to VF.
1548 for (int i = 0; i < VLen; ++i)
1549 Indices.push_back(ConstantInt::get(ITy, StartIdx + i));
1551 // Add the consecutive indices to the vector value.
1552 Constant *Cv = ConstantVector::get(Indices);
1553 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1554 Step = Builder.CreateVectorSplat(VLen, Step);
1555 assert(Step->getType() == Val->getType() && "Invalid step vec");
1556 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
1557 // which can be found from the original scalar operations.
1558 Step = Builder.CreateMul(Cv, Step);
1559 return Builder.CreateAdd(Val, Step, "induction");
1562 /// \brief Find the operand of the GEP that should be checked for consecutive
1563 /// stores. This ignores trailing indices that have no effect on the final
1565 static unsigned getGEPInductionOperand(const DataLayout *DL,
1566 const GetElementPtrInst *Gep) {
1567 unsigned LastOperand = Gep->getNumOperands() - 1;
1568 unsigned GEPAllocSize = DL->getTypeAllocSize(
1569 cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
1571 // Walk backwards and try to peel off zeros.
1572 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
1573 // Find the type we're currently indexing into.
1574 gep_type_iterator GEPTI = gep_type_begin(Gep);
1575 std::advance(GEPTI, LastOperand - 1);
1577 // If it's a type with the same allocation size as the result of the GEP we
1578 // can peel off the zero index.
1579 if (DL->getTypeAllocSize(*GEPTI) != GEPAllocSize)
1587 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1588 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1589 // Make sure that the pointer does not point to structs.
1590 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1593 // If this value is a pointer induction variable we know it is consecutive.
1594 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1595 if (Phi && Inductions.count(Phi)) {
1596 InductionInfo II = Inductions[Phi];
1597 return II.getConsecutiveDirection();
1600 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1604 unsigned NumOperands = Gep->getNumOperands();
1605 Value *GpPtr = Gep->getPointerOperand();
1606 // If this GEP value is a consecutive pointer induction variable and all of
1607 // the indices are constant then we know it is consecutive. We can
1608 Phi = dyn_cast<PHINode>(GpPtr);
1609 if (Phi && Inductions.count(Phi)) {
1611 // Make sure that the pointer does not point to structs.
1612 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1613 if (GepPtrType->getElementType()->isAggregateType())
1616 // Make sure that all of the index operands are loop invariant.
1617 for (unsigned i = 1; i < NumOperands; ++i)
1618 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1621 InductionInfo II = Inductions[Phi];
1622 return II.getConsecutiveDirection();
1625 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1627 // Check that all of the gep indices are uniform except for our induction
1629 for (unsigned i = 0; i != NumOperands; ++i)
1630 if (i != InductionOperand &&
1631 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1634 // We can emit wide load/stores only if the last non-zero index is the
1635 // induction variable.
1636 const SCEV *Last = nullptr;
1637 if (!Strides.count(Gep))
1638 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1640 // Because of the multiplication by a stride we can have a s/zext cast.
1641 // We are going to replace this stride by 1 so the cast is safe to ignore.
1643 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1644 // %0 = trunc i64 %indvars.iv to i32
1645 // %mul = mul i32 %0, %Stride1
1646 // %idxprom = zext i32 %mul to i64 << Safe cast.
1647 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1649 Last = replaceSymbolicStrideSCEV(SE, Strides,
1650 Gep->getOperand(InductionOperand), Gep);
1651 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1653 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1657 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1658 const SCEV *Step = AR->getStepRecurrence(*SE);
1660 // The memory is consecutive because the last index is consecutive
1661 // and all other indices are loop invariant.
1664 if (Step->isAllOnesValue())
1671 bool LoopVectorizationLegality::isUniform(Value *V) {
1672 return LAI->isUniform(V);
1675 InnerLoopVectorizer::VectorParts&
1676 InnerLoopVectorizer::getVectorValue(Value *V) {
1677 assert(V != Induction && "The new induction variable should not be used.");
1678 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1680 // If we have a stride that is replaced by one, do it here.
1681 if (Legal->hasStride(V))
1682 V = ConstantInt::get(V->getType(), 1);
1684 // If we have this scalar in the map, return it.
1685 if (WidenMap.has(V))
1686 return WidenMap.get(V);
1688 // If this scalar is unknown, assume that it is a constant or that it is
1689 // loop invariant. Broadcast V and save the value for future uses.
1690 Value *B = getBroadcastInstrs(V);
1691 return WidenMap.splat(V, B);
1694 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1695 assert(Vec->getType()->isVectorTy() && "Invalid type");
1696 SmallVector<Constant*, 8> ShuffleMask;
1697 for (unsigned i = 0; i < VF; ++i)
1698 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1700 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1701 ConstantVector::get(ShuffleMask),
1705 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
1706 // Attempt to issue a wide load.
1707 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1708 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1710 assert((LI || SI) && "Invalid Load/Store instruction");
1712 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1713 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1714 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1715 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1716 // An alignment of 0 means target abi alignment. We need to use the scalar's
1717 // target abi alignment in such a case.
1719 Alignment = DL->getABITypeAlignment(ScalarDataTy);
1720 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1721 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1722 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1724 if (SI && Legal->blockNeedsPredication(SI->getParent()) &&
1725 !Legal->isMaskRequired(SI))
1726 return scalarizeInstruction(Instr, true);
1728 if (ScalarAllocatedSize != VectorElementSize)
1729 return scalarizeInstruction(Instr);
1731 // If the pointer is loop invariant or if it is non-consecutive,
1732 // scalarize the load.
1733 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1734 bool Reverse = ConsecutiveStride < 0;
1735 bool UniformLoad = LI && Legal->isUniform(Ptr);
1736 if (!ConsecutiveStride || UniformLoad)
1737 return scalarizeInstruction(Instr);
1739 Constant *Zero = Builder.getInt32(0);
1740 VectorParts &Entry = WidenMap.get(Instr);
1742 // Handle consecutive loads/stores.
1743 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1744 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1745 setDebugLocFromInst(Builder, Gep);
1746 Value *PtrOperand = Gep->getPointerOperand();
1747 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1748 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1750 // Create the new GEP with the new induction variable.
1751 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1752 Gep2->setOperand(0, FirstBasePtr);
1753 Gep2->setName("gep.indvar.base");
1754 Ptr = Builder.Insert(Gep2);
1756 setDebugLocFromInst(Builder, Gep);
1757 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1758 OrigLoop) && "Base ptr must be invariant");
1760 // The last index does not have to be the induction. It can be
1761 // consecutive and be a function of the index. For example A[I+1];
1762 unsigned NumOperands = Gep->getNumOperands();
1763 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1764 // Create the new GEP with the new induction variable.
1765 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1767 for (unsigned i = 0; i < NumOperands; ++i) {
1768 Value *GepOperand = Gep->getOperand(i);
1769 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1771 // Update last index or loop invariant instruction anchored in loop.
1772 if (i == InductionOperand ||
1773 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1774 assert((i == InductionOperand ||
1775 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1776 "Must be last index or loop invariant");
1778 VectorParts &GEPParts = getVectorValue(GepOperand);
1779 Value *Index = GEPParts[0];
1780 Index = Builder.CreateExtractElement(Index, Zero);
1781 Gep2->setOperand(i, Index);
1782 Gep2->setName("gep.indvar.idx");
1785 Ptr = Builder.Insert(Gep2);
1787 // Use the induction element ptr.
1788 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1789 setDebugLocFromInst(Builder, Ptr);
1790 VectorParts &PtrVal = getVectorValue(Ptr);
1791 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1794 VectorParts Mask = createBlockInMask(Instr->getParent());
1797 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1798 "We do not allow storing to uniform addresses");
1799 setDebugLocFromInst(Builder, SI);
1800 // We don't want to update the value in the map as it might be used in
1801 // another expression. So don't use a reference type for "StoredVal".
1802 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1804 for (unsigned Part = 0; Part < UF; ++Part) {
1805 // Calculate the pointer for the specific unroll-part.
1806 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1809 // If we store to reverse consecutive memory locations then we need
1810 // to reverse the order of elements in the stored value.
1811 StoredVal[Part] = reverseVector(StoredVal[Part]);
1812 // If the address is consecutive but reversed, then the
1813 // wide store needs to start at the last vector element.
1814 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1815 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1816 Mask[Part] = reverseVector(Mask[Part]);
1819 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1820 DataTy->getPointerTo(AddressSpace));
1823 if (Legal->isMaskRequired(SI))
1824 NewSI = Builder.CreateMaskedStore(StoredVal[Part], VecPtr, Alignment,
1827 NewSI = Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
1828 propagateMetadata(NewSI, SI);
1834 assert(LI && "Must have a load instruction");
1835 setDebugLocFromInst(Builder, LI);
1836 for (unsigned Part = 0; Part < UF; ++Part) {
1837 // Calculate the pointer for the specific unroll-part.
1838 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1841 // If the address is consecutive but reversed, then the
1842 // wide load needs to start at the last vector element.
1843 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1844 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1845 Mask[Part] = reverseVector(Mask[Part]);
1849 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1850 DataTy->getPointerTo(AddressSpace));
1851 if (Legal->isMaskRequired(LI))
1852 NewLI = Builder.CreateMaskedLoad(VecPtr, Alignment, Mask[Part],
1853 UndefValue::get(DataTy),
1854 "wide.masked.load");
1856 NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
1857 propagateMetadata(NewLI, LI);
1858 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
1862 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
1863 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1864 // Holds vector parameters or scalars, in case of uniform vals.
1865 SmallVector<VectorParts, 4> Params;
1867 setDebugLocFromInst(Builder, Instr);
1869 // Find all of the vectorized parameters.
1870 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1871 Value *SrcOp = Instr->getOperand(op);
1873 // If we are accessing the old induction variable, use the new one.
1874 if (SrcOp == OldInduction) {
1875 Params.push_back(getVectorValue(SrcOp));
1879 // Try using previously calculated values.
1880 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1882 // If the src is an instruction that appeared earlier in the basic block
1883 // then it should already be vectorized.
1884 if (SrcInst && OrigLoop->contains(SrcInst)) {
1885 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1886 // The parameter is a vector value from earlier.
1887 Params.push_back(WidenMap.get(SrcInst));
1889 // The parameter is a scalar from outside the loop. Maybe even a constant.
1890 VectorParts Scalars;
1891 Scalars.append(UF, SrcOp);
1892 Params.push_back(Scalars);
1896 assert(Params.size() == Instr->getNumOperands() &&
1897 "Invalid number of operands");
1899 // Does this instruction return a value ?
1900 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1902 Value *UndefVec = IsVoidRetTy ? nullptr :
1903 UndefValue::get(VectorType::get(Instr->getType(), VF));
1904 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1905 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1907 Instruction *InsertPt = Builder.GetInsertPoint();
1908 BasicBlock *IfBlock = Builder.GetInsertBlock();
1909 BasicBlock *CondBlock = nullptr;
1912 Loop *VectorLp = nullptr;
1913 if (IfPredicateStore) {
1914 assert(Instr->getParent()->getSinglePredecessor() &&
1915 "Only support single predecessor blocks");
1916 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
1917 Instr->getParent());
1918 VectorLp = LI->getLoopFor(IfBlock);
1919 assert(VectorLp && "Must have a loop for this block");
1922 // For each vector unroll 'part':
1923 for (unsigned Part = 0; Part < UF; ++Part) {
1924 // For each scalar that we create:
1925 for (unsigned Width = 0; Width < VF; ++Width) {
1928 Value *Cmp = nullptr;
1929 if (IfPredicateStore) {
1930 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
1931 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
1932 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1933 LoopVectorBody.push_back(CondBlock);
1934 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
1935 // Update Builder with newly created basic block.
1936 Builder.SetInsertPoint(InsertPt);
1939 Instruction *Cloned = Instr->clone();
1941 Cloned->setName(Instr->getName() + ".cloned");
1942 // Replace the operands of the cloned instructions with extracted scalars.
1943 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1944 Value *Op = Params[op][Part];
1945 // Param is a vector. Need to extract the right lane.
1946 if (Op->getType()->isVectorTy())
1947 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1948 Cloned->setOperand(op, Op);
1951 // Place the cloned scalar in the new loop.
1952 Builder.Insert(Cloned);
1954 // If the original scalar returns a value we need to place it in a vector
1955 // so that future users will be able to use it.
1957 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1958 Builder.getInt32(Width));
1960 if (IfPredicateStore) {
1961 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1962 LoopVectorBody.push_back(NewIfBlock);
1963 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
1964 Builder.SetInsertPoint(InsertPt);
1965 Instruction *OldBr = IfBlock->getTerminator();
1966 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1967 OldBr->eraseFromParent();
1968 IfBlock = NewIfBlock;
1974 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
1978 if (Instruction *I = dyn_cast<Instruction>(V))
1979 return I->getParent() == Loc->getParent() ? I : nullptr;
1983 std::pair<Instruction *, Instruction *>
1984 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
1985 Instruction *tnullptr = nullptr;
1986 if (!Legal->mustCheckStrides())
1987 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
1989 IRBuilder<> ChkBuilder(Loc);
1992 Value *Check = nullptr;
1993 Instruction *FirstInst = nullptr;
1994 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
1995 SE = Legal->strides_end();
1997 Value *Ptr = stripIntegerCast(*SI);
1998 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
2000 // Store the first instruction we create.
2001 FirstInst = getFirstInst(FirstInst, C, Loc);
2003 Check = ChkBuilder.CreateOr(Check, C);
2008 // We have to do this trickery because the IRBuilder might fold the check to a
2009 // constant expression in which case there is no Instruction anchored in a
2011 LLVMContext &Ctx = Loc->getContext();
2012 Instruction *TheCheck =
2013 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
2014 ChkBuilder.Insert(TheCheck, "stride.not.one");
2015 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
2017 return std::make_pair(FirstInst, TheCheck);
2020 void InnerLoopVectorizer::createEmptyLoop() {
2022 In this function we generate a new loop. The new loop will contain
2023 the vectorized instructions while the old loop will continue to run the
2026 [ ] <-- Back-edge taken count overflow check.
2029 | [ ] <-- vector loop bypass (may consist of multiple blocks).
2032 || [ ] <-- vector pre header.
2036 || [ ]_| <-- vector loop.
2039 | >[ ] <--- middle-block.
2042 -|- >[ ] <--- new preheader.
2046 | [ ]_| <-- old scalar loop to handle remainder.
2049 >[ ] <-- exit block.
2053 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
2054 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
2055 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
2056 assert(BypassBlock && "Invalid loop structure");
2057 assert(ExitBlock && "Must have an exit block");
2059 // Some loops have a single integer induction variable, while other loops
2060 // don't. One example is c++ iterators that often have multiple pointer
2061 // induction variables. In the code below we also support a case where we
2062 // don't have a single induction variable.
2063 OldInduction = Legal->getInduction();
2064 Type *IdxTy = Legal->getWidestInductionType();
2066 // Find the loop boundaries.
2067 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
2068 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
2070 // The exit count might have the type of i64 while the phi is i32. This can
2071 // happen if we have an induction variable that is sign extended before the
2072 // compare. The only way that we get a backedge taken count is that the
2073 // induction variable was signed and as such will not overflow. In such a case
2074 // truncation is legal.
2075 if (ExitCount->getType()->getPrimitiveSizeInBits() >
2076 IdxTy->getPrimitiveSizeInBits())
2077 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
2079 const SCEV *BackedgeTakeCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
2080 // Get the total trip count from the count by adding 1.
2081 ExitCount = SE->getAddExpr(BackedgeTakeCount,
2082 SE->getConstant(BackedgeTakeCount->getType(), 1));
2084 // Expand the trip count and place the new instructions in the preheader.
2085 // Notice that the pre-header does not change, only the loop body.
2086 SCEVExpander Exp(*SE, "induction");
2088 // We need to test whether the backedge-taken count is uint##_max. Adding one
2089 // to it will cause overflow and an incorrect loop trip count in the vector
2090 // body. In case of overflow we want to directly jump to the scalar remainder
2092 Value *BackedgeCount =
2093 Exp.expandCodeFor(BackedgeTakeCount, BackedgeTakeCount->getType(),
2094 BypassBlock->getTerminator());
2095 if (BackedgeCount->getType()->isPointerTy())
2096 BackedgeCount = CastInst::CreatePointerCast(BackedgeCount, IdxTy,
2097 "backedge.ptrcnt.to.int",
2098 BypassBlock->getTerminator());
2099 Instruction *CheckBCOverflow =
2100 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, BackedgeCount,
2101 Constant::getAllOnesValue(BackedgeCount->getType()),
2102 "backedge.overflow", BypassBlock->getTerminator());
2104 // The loop index does not have to start at Zero. Find the original start
2105 // value from the induction PHI node. If we don't have an induction variable
2106 // then we know that it starts at zero.
2107 Builder.SetInsertPoint(BypassBlock->getTerminator());
2108 Value *StartIdx = ExtendedIdx = OldInduction ?
2109 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
2111 ConstantInt::get(IdxTy, 0);
2113 // We need an instruction to anchor the overflow check on. StartIdx needs to
2114 // be defined before the overflow check branch. Because the scalar preheader
2115 // is going to merge the start index and so the overflow branch block needs to
2116 // contain a definition of the start index.
2117 Instruction *OverflowCheckAnchor = BinaryOperator::CreateAdd(
2118 StartIdx, ConstantInt::get(IdxTy, 0), "overflow.check.anchor",
2119 BypassBlock->getTerminator());
2121 // Count holds the overall loop count (N).
2122 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2123 BypassBlock->getTerminator());
2125 LoopBypassBlocks.push_back(BypassBlock);
2127 // Split the single block loop into the two loop structure described above.
2128 BasicBlock *VectorPH =
2129 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
2130 BasicBlock *VecBody =
2131 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
2132 BasicBlock *MiddleBlock =
2133 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
2134 BasicBlock *ScalarPH =
2135 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
2137 // Create and register the new vector loop.
2138 Loop* Lp = new Loop();
2139 Loop *ParentLoop = OrigLoop->getParentLoop();
2141 // Insert the new loop into the loop nest and register the new basic blocks
2142 // before calling any utilities such as SCEV that require valid LoopInfo.
2144 ParentLoop->addChildLoop(Lp);
2145 ParentLoop->addBasicBlockToLoop(ScalarPH, *LI);
2146 ParentLoop->addBasicBlockToLoop(VectorPH, *LI);
2147 ParentLoop->addBasicBlockToLoop(MiddleBlock, *LI);
2149 LI->addTopLevelLoop(Lp);
2151 Lp->addBasicBlockToLoop(VecBody, *LI);
2153 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
2155 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
2157 // Generate the induction variable.
2158 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2159 Induction = Builder.CreatePHI(IdxTy, 2, "index");
2160 // The loop step is equal to the vectorization factor (num of SIMD elements)
2161 // times the unroll factor (num of SIMD instructions).
2162 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2164 // This is the IR builder that we use to add all of the logic for bypassing
2165 // the new vector loop.
2166 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
2167 setDebugLocFromInst(BypassBuilder,
2168 getDebugLocFromInstOrOperands(OldInduction));
2170 // We may need to extend the index in case there is a type mismatch.
2171 // We know that the count starts at zero and does not overflow.
2172 if (Count->getType() != IdxTy) {
2173 // The exit count can be of pointer type. Convert it to the correct
2175 if (ExitCount->getType()->isPointerTy())
2176 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
2178 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
2181 // Add the start index to the loop count to get the new end index.
2182 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
2184 // Now we need to generate the expression for N - (N % VF), which is
2185 // the part that the vectorized body will execute.
2186 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
2187 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
2188 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
2189 "end.idx.rnd.down");
2191 // Now, compare the new count to zero. If it is zero skip the vector loop and
2192 // jump to the scalar loop.
2194 BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, "cmp.zero");
2196 BasicBlock *LastBypassBlock = BypassBlock;
2198 // Generate code to check that the loops trip count that we computed by adding
2199 // one to the backedge-taken count will not overflow.
2201 auto PastOverflowCheck =
2202 std::next(BasicBlock::iterator(OverflowCheckAnchor));
2203 BasicBlock *CheckBlock =
2204 LastBypassBlock->splitBasicBlock(PastOverflowCheck, "overflow.checked");
2206 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2207 LoopBypassBlocks.push_back(CheckBlock);
2208 Instruction *OldTerm = LastBypassBlock->getTerminator();
2209 BranchInst::Create(ScalarPH, CheckBlock, CheckBCOverflow, OldTerm);
2210 OldTerm->eraseFromParent();
2211 LastBypassBlock = CheckBlock;
2214 // Generate the code to check that the strides we assumed to be one are really
2215 // one. We want the new basic block to start at the first instruction in a
2216 // sequence of instructions that form a check.
2217 Instruction *StrideCheck;
2218 Instruction *FirstCheckInst;
2219 std::tie(FirstCheckInst, StrideCheck) =
2220 addStrideCheck(LastBypassBlock->getTerminator());
2222 // Create a new block containing the stride check.
2223 BasicBlock *CheckBlock =
2224 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.stridecheck");
2226 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2227 LoopBypassBlocks.push_back(CheckBlock);
2229 // Replace the branch into the memory check block with a conditional branch
2230 // for the "few elements case".
2231 Instruction *OldTerm = LastBypassBlock->getTerminator();
2232 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2233 OldTerm->eraseFromParent();
2236 LastBypassBlock = CheckBlock;
2239 // Generate the code that checks in runtime if arrays overlap. We put the
2240 // checks into a separate block to make the more common case of few elements
2242 Instruction *MemRuntimeCheck;
2243 std::tie(FirstCheckInst, MemRuntimeCheck) =
2244 Legal->getLAI()->addRuntimeCheck(LastBypassBlock->getTerminator());
2245 if (MemRuntimeCheck) {
2246 // Create a new block containing the memory check.
2247 BasicBlock *CheckBlock =
2248 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.memcheck");
2250 ParentLoop->addBasicBlockToLoop(CheckBlock, *LI);
2251 LoopBypassBlocks.push_back(CheckBlock);
2253 // Replace the branch into the memory check block with a conditional branch
2254 // for the "few elements case".
2255 Instruction *OldTerm = LastBypassBlock->getTerminator();
2256 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2257 OldTerm->eraseFromParent();
2259 Cmp = MemRuntimeCheck;
2260 LastBypassBlock = CheckBlock;
2263 LastBypassBlock->getTerminator()->eraseFromParent();
2264 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
2267 // We are going to resume the execution of the scalar loop.
2268 // Go over all of the induction variables that we found and fix the
2269 // PHIs that are left in the scalar version of the loop.
2270 // The starting values of PHI nodes depend on the counter of the last
2271 // iteration in the vectorized loop.
2272 // If we come from a bypass edge then we need to start from the original
2275 // This variable saves the new starting index for the scalar loop.
2276 PHINode *ResumeIndex = nullptr;
2277 LoopVectorizationLegality::InductionList::iterator I, E;
2278 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2279 // Set builder to point to last bypass block.
2280 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2281 for (I = List->begin(), E = List->end(); I != E; ++I) {
2282 PHINode *OrigPhi = I->first;
2283 LoopVectorizationLegality::InductionInfo II = I->second;
2285 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
2286 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
2287 MiddleBlock->getTerminator());
2288 // We might have extended the type of the induction variable but we need a
2289 // truncated version for the scalar loop.
2290 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
2291 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
2292 MiddleBlock->getTerminator()) : nullptr;
2294 // Create phi nodes to merge from the backedge-taken check block.
2295 PHINode *BCResumeVal = PHINode::Create(ResumeValTy, 3, "bc.resume.val",
2296 ScalarPH->getTerminator());
2297 BCResumeVal->addIncoming(ResumeVal, MiddleBlock);
2299 PHINode *BCTruncResumeVal = nullptr;
2300 if (OrigPhi == OldInduction) {
2302 PHINode::Create(OrigPhi->getType(), 2, "bc.trunc.resume.val",
2303 ScalarPH->getTerminator());
2304 BCTruncResumeVal->addIncoming(TruncResumeVal, MiddleBlock);
2307 Value *EndValue = nullptr;
2309 case LoopVectorizationLegality::IK_NoInduction:
2310 llvm_unreachable("Unknown induction");
2311 case LoopVectorizationLegality::IK_IntInduction: {
2312 // Handle the integer induction counter.
2313 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
2315 // We have the canonical induction variable.
2316 if (OrigPhi == OldInduction) {
2317 // Create a truncated version of the resume value for the scalar loop,
2318 // we might have promoted the type to a larger width.
2320 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
2321 // The new PHI merges the original incoming value, in case of a bypass,
2322 // or the value at the end of the vectorized loop.
2323 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2324 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2325 TruncResumeVal->addIncoming(EndValue, VecBody);
2327 BCTruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2329 // We know what the end value is.
2330 EndValue = IdxEndRoundDown;
2331 // We also know which PHI node holds it.
2332 ResumeIndex = ResumeVal;
2336 // Not the canonical induction variable - add the vector loop count to the
2338 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2339 II.StartValue->getType(),
2341 EndValue = II.transform(BypassBuilder, CRD);
2342 EndValue->setName("ind.end");
2345 case LoopVectorizationLegality::IK_PtrInduction: {
2346 EndValue = II.transform(BypassBuilder, CountRoundDown);
2347 EndValue->setName("ptr.ind.end");
2352 // The new PHI merges the original incoming value, in case of a bypass,
2353 // or the value at the end of the vectorized loop.
2354 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) {
2355 if (OrigPhi == OldInduction)
2356 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
2358 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2360 ResumeVal->addIncoming(EndValue, VecBody);
2362 // Fix the scalar body counter (PHI node).
2363 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2365 // The old induction's phi node in the scalar body needs the truncated
2367 if (OrigPhi == OldInduction) {
2368 BCResumeVal->addIncoming(StartIdx, LoopBypassBlocks[0]);
2369 OrigPhi->setIncomingValue(BlockIdx, BCTruncResumeVal);
2371 BCResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2372 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
2376 // If we are generating a new induction variable then we also need to
2377 // generate the code that calculates the exit value. This value is not
2378 // simply the end of the counter because we may skip the vectorized body
2379 // in case of a runtime check.
2381 assert(!ResumeIndex && "Unexpected resume value found");
2382 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2383 MiddleBlock->getTerminator());
2384 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2385 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2386 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
2389 // Make sure that we found the index where scalar loop needs to continue.
2390 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
2391 "Invalid resume Index");
2393 // Add a check in the middle block to see if we have completed
2394 // all of the iterations in the first vector loop.
2395 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2396 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2397 ResumeIndex, "cmp.n",
2398 MiddleBlock->getTerminator());
2400 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
2401 // Remove the old terminator.
2402 MiddleBlock->getTerminator()->eraseFromParent();
2404 // Create i+1 and fill the PHINode.
2405 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
2406 Induction->addIncoming(StartIdx, VectorPH);
2407 Induction->addIncoming(NextIdx, VecBody);
2408 // Create the compare.
2409 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
2410 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
2412 // Now we have two terminators. Remove the old one from the block.
2413 VecBody->getTerminator()->eraseFromParent();
2415 // Get ready to start creating new instructions into the vectorized body.
2416 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2419 LoopVectorPreHeader = VectorPH;
2420 LoopScalarPreHeader = ScalarPH;
2421 LoopMiddleBlock = MiddleBlock;
2422 LoopExitBlock = ExitBlock;
2423 LoopVectorBody.push_back(VecBody);
2424 LoopScalarBody = OldBasicBlock;
2426 LoopVectorizeHints Hints(Lp, true);
2427 Hints.setAlreadyVectorized();
2430 /// This function returns the identity element (or neutral element) for
2431 /// the operation K.
2433 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
2438 // Adding, Xoring, Oring zero to a number does not change it.
2439 return ConstantInt::get(Tp, 0);
2440 case RK_IntegerMult:
2441 // Multiplying a number by 1 does not change it.
2442 return ConstantInt::get(Tp, 1);
2444 // AND-ing a number with an all-1 value does not change it.
2445 return ConstantInt::get(Tp, -1, true);
2447 // Multiplying a number by 1 does not change it.
2448 return ConstantFP::get(Tp, 1.0L);
2450 // Adding zero to a number does not change it.
2451 return ConstantFP::get(Tp, 0.0L);
2453 llvm_unreachable("Unknown reduction kind");
2457 /// This function translates the reduction kind to an LLVM binary operator.
2459 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
2461 case LoopVectorizationLegality::RK_IntegerAdd:
2462 return Instruction::Add;
2463 case LoopVectorizationLegality::RK_IntegerMult:
2464 return Instruction::Mul;
2465 case LoopVectorizationLegality::RK_IntegerOr:
2466 return Instruction::Or;
2467 case LoopVectorizationLegality::RK_IntegerAnd:
2468 return Instruction::And;
2469 case LoopVectorizationLegality::RK_IntegerXor:
2470 return Instruction::Xor;
2471 case LoopVectorizationLegality::RK_FloatMult:
2472 return Instruction::FMul;
2473 case LoopVectorizationLegality::RK_FloatAdd:
2474 return Instruction::FAdd;
2475 case LoopVectorizationLegality::RK_IntegerMinMax:
2476 return Instruction::ICmp;
2477 case LoopVectorizationLegality::RK_FloatMinMax:
2478 return Instruction::FCmp;
2480 llvm_unreachable("Unknown reduction operation");
2484 Value *createMinMaxOp(IRBuilder<> &Builder,
2485 LoopVectorizationLegality::MinMaxReductionKind RK,
2488 CmpInst::Predicate P = CmpInst::ICMP_NE;
2491 llvm_unreachable("Unknown min/max reduction kind");
2492 case LoopVectorizationLegality::MRK_UIntMin:
2493 P = CmpInst::ICMP_ULT;
2495 case LoopVectorizationLegality::MRK_UIntMax:
2496 P = CmpInst::ICMP_UGT;
2498 case LoopVectorizationLegality::MRK_SIntMin:
2499 P = CmpInst::ICMP_SLT;
2501 case LoopVectorizationLegality::MRK_SIntMax:
2502 P = CmpInst::ICMP_SGT;
2504 case LoopVectorizationLegality::MRK_FloatMin:
2505 P = CmpInst::FCMP_OLT;
2507 case LoopVectorizationLegality::MRK_FloatMax:
2508 P = CmpInst::FCMP_OGT;
2513 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
2514 RK == LoopVectorizationLegality::MRK_FloatMax)
2515 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2517 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2519 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2524 struct CSEDenseMapInfo {
2525 static bool canHandle(Instruction *I) {
2526 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2527 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2529 static inline Instruction *getEmptyKey() {
2530 return DenseMapInfo<Instruction *>::getEmptyKey();
2532 static inline Instruction *getTombstoneKey() {
2533 return DenseMapInfo<Instruction *>::getTombstoneKey();
2535 static unsigned getHashValue(Instruction *I) {
2536 assert(canHandle(I) && "Unknown instruction!");
2537 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2538 I->value_op_end()));
2540 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2541 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2542 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2544 return LHS->isIdenticalTo(RHS);
2549 /// \brief Check whether this block is a predicated block.
2550 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2551 /// = ...; " blocks. We start with one vectorized basic block. For every
2552 /// conditional block we split this vectorized block. Therefore, every second
2553 /// block will be a predicated one.
2554 static bool isPredicatedBlock(unsigned BlockNum) {
2555 return BlockNum % 2;
2558 ///\brief Perform cse of induction variable instructions.
2559 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
2560 // Perform simple cse.
2561 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2562 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
2563 BasicBlock *BB = BBs[i];
2564 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2565 Instruction *In = I++;
2567 if (!CSEDenseMapInfo::canHandle(In))
2570 // Check if we can replace this instruction with any of the
2571 // visited instructions.
2572 if (Instruction *V = CSEMap.lookup(In)) {
2573 In->replaceAllUsesWith(V);
2574 In->eraseFromParent();
2577 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
2578 // ...;" blocks for predicated stores. Every second block is a predicated
2580 if (isPredicatedBlock(i))
2588 /// \brief Adds a 'fast' flag to floating point operations.
2589 static Value *addFastMathFlag(Value *V) {
2590 if (isa<FPMathOperator>(V)){
2591 FastMathFlags Flags;
2592 Flags.setUnsafeAlgebra();
2593 cast<Instruction>(V)->setFastMathFlags(Flags);
2598 void InnerLoopVectorizer::vectorizeLoop() {
2599 //===------------------------------------------------===//
2601 // Notice: any optimization or new instruction that go
2602 // into the code below should be also be implemented in
2605 //===------------------------------------------------===//
2606 Constant *Zero = Builder.getInt32(0);
2608 // In order to support reduction variables we need to be able to vectorize
2609 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2610 // stages. First, we create a new vector PHI node with no incoming edges.
2611 // We use this value when we vectorize all of the instructions that use the
2612 // PHI. Next, after all of the instructions in the block are complete we
2613 // add the new incoming edges to the PHI. At this point all of the
2614 // instructions in the basic block are vectorized, so we can use them to
2615 // construct the PHI.
2616 PhiVector RdxPHIsToFix;
2618 // Scan the loop in a topological order to ensure that defs are vectorized
2620 LoopBlocksDFS DFS(OrigLoop);
2623 // Vectorize all of the blocks in the original loop.
2624 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2625 be = DFS.endRPO(); bb != be; ++bb)
2626 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
2628 // At this point every instruction in the original loop is widened to
2629 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2630 // that we vectorized. The PHI nodes are currently empty because we did
2631 // not want to introduce cycles. Notice that the remaining PHI nodes
2632 // that we need to fix are reduction variables.
2634 // Create the 'reduced' values for each of the induction vars.
2635 // The reduced values are the vector values that we scalarize and combine
2636 // after the loop is finished.
2637 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2639 PHINode *RdxPhi = *it;
2640 assert(RdxPhi && "Unable to recover vectorized PHI");
2642 // Find the reduction variable descriptor.
2643 assert(Legal->getReductionVars()->count(RdxPhi) &&
2644 "Unable to find the reduction variable");
2645 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2646 (*Legal->getReductionVars())[RdxPhi];
2648 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2650 // We need to generate a reduction vector from the incoming scalar.
2651 // To do so, we need to generate the 'identity' vector and override
2652 // one of the elements with the incoming scalar reduction. We need
2653 // to do it in the vector-loop preheader.
2654 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
2656 // This is the vector-clone of the value that leaves the loop.
2657 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2658 Type *VecTy = VectorExit[0]->getType();
2660 // Find the reduction identity variable. Zero for addition, or, xor,
2661 // one for multiplication, -1 for And.
2664 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2665 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2666 // MinMax reduction have the start value as their identify.
2668 VectorStart = Identity = RdxDesc.StartValue;
2670 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2675 // Handle other reduction kinds:
2677 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2678 VecTy->getScalarType());
2681 // This vector is the Identity vector where the first element is the
2682 // incoming scalar reduction.
2683 VectorStart = RdxDesc.StartValue;
2685 Identity = ConstantVector::getSplat(VF, Iden);
2687 // This vector is the Identity vector where the first element is the
2688 // incoming scalar reduction.
2689 VectorStart = Builder.CreateInsertElement(Identity,
2690 RdxDesc.StartValue, Zero);
2694 // Fix the vector-loop phi.
2696 // Reductions do not have to start at zero. They can start with
2697 // any loop invariant values.
2698 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2699 BasicBlock *Latch = OrigLoop->getLoopLatch();
2700 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2701 VectorParts &Val = getVectorValue(LoopVal);
2702 for (unsigned part = 0; part < UF; ++part) {
2703 // Make sure to add the reduction stat value only to the
2704 // first unroll part.
2705 Value *StartVal = (part == 0) ? VectorStart : Identity;
2706 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal,
2707 LoopVectorPreHeader);
2708 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
2709 LoopVectorBody.back());
2712 // Before each round, move the insertion point right between
2713 // the PHIs and the values we are going to write.
2714 // This allows us to write both PHINodes and the extractelement
2716 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2718 VectorParts RdxParts;
2719 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2720 for (unsigned part = 0; part < UF; ++part) {
2721 // This PHINode contains the vectorized reduction variable, or
2722 // the initial value vector, if we bypass the vector loop.
2723 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2724 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2725 Value *StartVal = (part == 0) ? VectorStart : Identity;
2726 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2727 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2728 NewPhi->addIncoming(RdxExitVal[part],
2729 LoopVectorBody.back());
2730 RdxParts.push_back(NewPhi);
2733 // Reduce all of the unrolled parts into a single vector.
2734 Value *ReducedPartRdx = RdxParts[0];
2735 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2736 setDebugLocFromInst(Builder, ReducedPartRdx);
2737 for (unsigned part = 1; part < UF; ++part) {
2738 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2739 // Floating point operations had to be 'fast' to enable the reduction.
2740 ReducedPartRdx = addFastMathFlag(
2741 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
2742 ReducedPartRdx, "bin.rdx"));
2744 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2745 ReducedPartRdx, RdxParts[part]);
2749 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2750 // and vector ops, reducing the set of values being computed by half each
2752 assert(isPowerOf2_32(VF) &&
2753 "Reduction emission only supported for pow2 vectors!");
2754 Value *TmpVec = ReducedPartRdx;
2755 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
2756 for (unsigned i = VF; i != 1; i >>= 1) {
2757 // Move the upper half of the vector to the lower half.
2758 for (unsigned j = 0; j != i/2; ++j)
2759 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2761 // Fill the rest of the mask with undef.
2762 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2763 UndefValue::get(Builder.getInt32Ty()));
2766 Builder.CreateShuffleVector(TmpVec,
2767 UndefValue::get(TmpVec->getType()),
2768 ConstantVector::get(ShuffleMask),
2771 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2772 // Floating point operations had to be 'fast' to enable the reduction.
2773 TmpVec = addFastMathFlag(Builder.CreateBinOp(
2774 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
2776 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2779 // The result is in the first element of the vector.
2780 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2781 Builder.getInt32(0));
2784 // Create a phi node that merges control-flow from the backedge-taken check
2785 // block and the middle block.
2786 PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
2787 LoopScalarPreHeader->getTerminator());
2788 BCBlockPhi->addIncoming(RdxDesc.StartValue, LoopBypassBlocks[0]);
2789 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2791 // Now, we need to fix the users of the reduction variable
2792 // inside and outside of the scalar remainder loop.
2793 // We know that the loop is in LCSSA form. We need to update the
2794 // PHI nodes in the exit blocks.
2795 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2796 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2797 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2798 if (!LCSSAPhi) break;
2800 // All PHINodes need to have a single entry edge, or two if
2801 // we already fixed them.
2802 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2804 // We found our reduction value exit-PHI. Update it with the
2805 // incoming bypass edge.
2806 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2807 // Add an edge coming from the bypass.
2808 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2811 }// end of the LCSSA phi scan.
2813 // Fix the scalar loop reduction variable with the incoming reduction sum
2814 // from the vector body and from the backedge value.
2815 int IncomingEdgeBlockIdx =
2816 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2817 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2818 // Pick the other block.
2819 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2820 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
2821 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2822 }// end of for each redux variable.
2826 // Remove redundant induction instructions.
2827 cse(LoopVectorBody);
2830 void InnerLoopVectorizer::fixLCSSAPHIs() {
2831 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2832 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2833 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2834 if (!LCSSAPhi) break;
2835 if (LCSSAPhi->getNumIncomingValues() == 1)
2836 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2841 InnerLoopVectorizer::VectorParts
2842 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2843 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2846 // Look for cached value.
2847 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
2848 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
2849 if (ECEntryIt != MaskCache.end())
2850 return ECEntryIt->second;
2852 VectorParts SrcMask = createBlockInMask(Src);
2854 // The terminator has to be a branch inst!
2855 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
2856 assert(BI && "Unexpected terminator found");
2858 if (BI->isConditional()) {
2859 VectorParts EdgeMask = getVectorValue(BI->getCondition());
2861 if (BI->getSuccessor(0) != Dst)
2862 for (unsigned part = 0; part < UF; ++part)
2863 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
2865 for (unsigned part = 0; part < UF; ++part)
2866 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
2868 MaskCache[Edge] = EdgeMask;
2872 MaskCache[Edge] = SrcMask;
2876 InnerLoopVectorizer::VectorParts
2877 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
2878 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
2880 // Loop incoming mask is all-one.
2881 if (OrigLoop->getHeader() == BB) {
2882 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
2883 return getVectorValue(C);
2886 // This is the block mask. We OR all incoming edges, and with zero.
2887 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
2888 VectorParts BlockMask = getVectorValue(Zero);
2891 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
2892 VectorParts EM = createEdgeMask(*it, BB);
2893 for (unsigned part = 0; part < UF; ++part)
2894 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
2900 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
2901 InnerLoopVectorizer::VectorParts &Entry,
2902 unsigned UF, unsigned VF, PhiVector *PV) {
2903 PHINode* P = cast<PHINode>(PN);
2904 // Handle reduction variables:
2905 if (Legal->getReductionVars()->count(P)) {
2906 for (unsigned part = 0; part < UF; ++part) {
2907 // This is phase one of vectorizing PHIs.
2908 Type *VecTy = (VF == 1) ? PN->getType() :
2909 VectorType::get(PN->getType(), VF);
2910 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
2911 LoopVectorBody.back()-> getFirstInsertionPt());
2917 setDebugLocFromInst(Builder, P);
2918 // Check for PHI nodes that are lowered to vector selects.
2919 if (P->getParent() != OrigLoop->getHeader()) {
2920 // We know that all PHIs in non-header blocks are converted into
2921 // selects, so we don't have to worry about the insertion order and we
2922 // can just use the builder.
2923 // At this point we generate the predication tree. There may be
2924 // duplications since this is a simple recursive scan, but future
2925 // optimizations will clean it up.
2927 unsigned NumIncoming = P->getNumIncomingValues();
2929 // Generate a sequence of selects of the form:
2930 // SELECT(Mask3, In3,
2931 // SELECT(Mask2, In2,
2933 for (unsigned In = 0; In < NumIncoming; In++) {
2934 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
2936 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
2938 for (unsigned part = 0; part < UF; ++part) {
2939 // We might have single edge PHIs (blocks) - use an identity
2940 // 'select' for the first PHI operand.
2942 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2945 // Select between the current value and the previous incoming edge
2946 // based on the incoming mask.
2947 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
2948 Entry[part], "predphi");
2954 // This PHINode must be an induction variable.
2955 // Make sure that we know about it.
2956 assert(Legal->getInductionVars()->count(P) &&
2957 "Not an induction variable");
2959 LoopVectorizationLegality::InductionInfo II =
2960 Legal->getInductionVars()->lookup(P);
2962 // FIXME: The newly created binary instructions should contain nsw/nuw flags,
2963 // which can be found from the original scalar operations.
2965 case LoopVectorizationLegality::IK_NoInduction:
2966 llvm_unreachable("Unknown induction");
2967 case LoopVectorizationLegality::IK_IntInduction: {
2968 assert(P->getType() == II.StartValue->getType() && "Types must match");
2969 Type *PhiTy = P->getType();
2971 if (P == OldInduction) {
2972 // Handle the canonical induction variable. We might have had to
2974 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
2976 // Handle other induction variables that are now based on the
2978 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
2980 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
2981 Broadcasted = II.transform(Builder, NormalizedIdx);
2982 Broadcasted->setName("offset.idx");
2984 Broadcasted = getBroadcastInstrs(Broadcasted);
2985 // After broadcasting the induction variable we need to make the vector
2986 // consecutive by adding 0, 1, 2, etc.
2987 for (unsigned part = 0; part < UF; ++part)
2988 Entry[part] = getStepVector(Broadcasted, VF * part, II.StepValue);
2991 case LoopVectorizationLegality::IK_PtrInduction:
2992 // Handle the pointer induction variable case.
2993 assert(P->getType()->isPointerTy() && "Unexpected type.");
2994 // This is the normalized GEP that starts counting at zero.
2995 Value *NormalizedIdx =
2996 Builder.CreateSub(Induction, ExtendedIdx, "normalized.idx");
2997 // This is the vector of results. Notice that we don't generate
2998 // vector geps because scalar geps result in better code.
2999 for (unsigned part = 0; part < UF; ++part) {
3001 int EltIndex = part;
3002 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3003 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx);
3004 Value *SclrGep = II.transform(Builder, GlobalIdx);
3005 SclrGep->setName("next.gep");
3006 Entry[part] = SclrGep;
3010 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
3011 for (unsigned int i = 0; i < VF; ++i) {
3012 int EltIndex = i + part * VF;
3013 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3014 Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx);
3015 Value *SclrGep = II.transform(Builder, GlobalIdx);
3016 SclrGep->setName("next.gep");
3017 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
3018 Builder.getInt32(i),
3021 Entry[part] = VecVal;
3027 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
3028 // For each instruction in the old loop.
3029 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3030 VectorParts &Entry = WidenMap.get(it);
3031 switch (it->getOpcode()) {
3032 case Instruction::Br:
3033 // Nothing to do for PHIs and BR, since we already took care of the
3034 // loop control flow instructions.
3036 case Instruction::PHI: {
3037 // Vectorize PHINodes.
3038 widenPHIInstruction(it, Entry, UF, VF, PV);
3042 case Instruction::Add:
3043 case Instruction::FAdd:
3044 case Instruction::Sub:
3045 case Instruction::FSub:
3046 case Instruction::Mul:
3047 case Instruction::FMul:
3048 case Instruction::UDiv:
3049 case Instruction::SDiv:
3050 case Instruction::FDiv:
3051 case Instruction::URem:
3052 case Instruction::SRem:
3053 case Instruction::FRem:
3054 case Instruction::Shl:
3055 case Instruction::LShr:
3056 case Instruction::AShr:
3057 case Instruction::And:
3058 case Instruction::Or:
3059 case Instruction::Xor: {
3060 // Just widen binops.
3061 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3062 setDebugLocFromInst(Builder, BinOp);
3063 VectorParts &A = getVectorValue(it->getOperand(0));
3064 VectorParts &B = getVectorValue(it->getOperand(1));
3066 // Use this vector value for all users of the original instruction.
3067 for (unsigned Part = 0; Part < UF; ++Part) {
3068 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3070 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
3071 VecOp->copyIRFlags(BinOp);
3076 propagateMetadata(Entry, it);
3079 case Instruction::Select: {
3081 // If the selector is loop invariant we can create a select
3082 // instruction with a scalar condition. Otherwise, use vector-select.
3083 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3085 setDebugLocFromInst(Builder, it);
3087 // The condition can be loop invariant but still defined inside the
3088 // loop. This means that we can't just use the original 'cond' value.
3089 // We have to take the 'vectorized' value and pick the first lane.
3090 // Instcombine will make this a no-op.
3091 VectorParts &Cond = getVectorValue(it->getOperand(0));
3092 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3093 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3095 Value *ScalarCond = (VF == 1) ? Cond[0] :
3096 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3098 for (unsigned Part = 0; Part < UF; ++Part) {
3099 Entry[Part] = Builder.CreateSelect(
3100 InvariantCond ? ScalarCond : Cond[Part],
3105 propagateMetadata(Entry, it);
3109 case Instruction::ICmp:
3110 case Instruction::FCmp: {
3111 // Widen compares. Generate vector compares.
3112 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3113 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3114 setDebugLocFromInst(Builder, it);
3115 VectorParts &A = getVectorValue(it->getOperand(0));
3116 VectorParts &B = getVectorValue(it->getOperand(1));
3117 for (unsigned Part = 0; Part < UF; ++Part) {
3120 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3122 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3126 propagateMetadata(Entry, it);
3130 case Instruction::Store:
3131 case Instruction::Load:
3132 vectorizeMemoryInstruction(it);
3134 case Instruction::ZExt:
3135 case Instruction::SExt:
3136 case Instruction::FPToUI:
3137 case Instruction::FPToSI:
3138 case Instruction::FPExt:
3139 case Instruction::PtrToInt:
3140 case Instruction::IntToPtr:
3141 case Instruction::SIToFP:
3142 case Instruction::UIToFP:
3143 case Instruction::Trunc:
3144 case Instruction::FPTrunc:
3145 case Instruction::BitCast: {
3146 CastInst *CI = dyn_cast<CastInst>(it);
3147 setDebugLocFromInst(Builder, it);
3148 /// Optimize the special case where the source is the induction
3149 /// variable. Notice that we can only optimize the 'trunc' case
3150 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3151 /// c. other casts depend on pointer size.
3152 if (CI->getOperand(0) == OldInduction &&
3153 it->getOpcode() == Instruction::Trunc) {
3154 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3156 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3157 LoopVectorizationLegality::InductionInfo II =
3158 Legal->getInductionVars()->lookup(OldInduction);
3160 ConstantInt::getSigned(CI->getType(), II.StepValue->getSExtValue());
3161 for (unsigned Part = 0; Part < UF; ++Part)
3162 Entry[Part] = getStepVector(Broadcasted, VF * Part, Step);
3163 propagateMetadata(Entry, it);
3166 /// Vectorize casts.
3167 Type *DestTy = (VF == 1) ? CI->getType() :
3168 VectorType::get(CI->getType(), VF);
3170 VectorParts &A = getVectorValue(it->getOperand(0));
3171 for (unsigned Part = 0; Part < UF; ++Part)
3172 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3173 propagateMetadata(Entry, it);
3177 case Instruction::Call: {
3178 // Ignore dbg intrinsics.
3179 if (isa<DbgInfoIntrinsic>(it))
3181 setDebugLocFromInst(Builder, it);
3183 Module *M = BB->getParent()->getParent();
3184 CallInst *CI = cast<CallInst>(it);
3185 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3186 assert(ID && "Not an intrinsic call!");
3188 case Intrinsic::assume:
3189 case Intrinsic::lifetime_end:
3190 case Intrinsic::lifetime_start:
3191 scalarizeInstruction(it);
3194 bool HasScalarOpd = hasVectorInstrinsicScalarOpd(ID, 1);
3195 for (unsigned Part = 0; Part < UF; ++Part) {
3196 SmallVector<Value *, 4> Args;
3197 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3198 if (HasScalarOpd && i == 1) {
3199 Args.push_back(CI->getArgOperand(i));
3202 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
3203 Args.push_back(Arg[Part]);
3205 Type *Tys[] = {CI->getType()};
3207 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3209 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
3210 Entry[Part] = Builder.CreateCall(F, Args);
3213 propagateMetadata(Entry, it);
3220 // All other instructions are unsupported. Scalarize them.
3221 scalarizeInstruction(it);
3224 }// end of for_each instr.
3227 void InnerLoopVectorizer::updateAnalysis() {
3228 // Forget the original basic block.
3229 SE->forgetLoop(OrigLoop);
3231 // Update the dominator tree information.
3232 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3233 "Entry does not dominate exit.");
3235 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3236 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3237 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3239 // Due to if predication of stores we might create a sequence of "if(pred)
3240 // a[i] = ...; " blocks.
3241 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3243 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3244 else if (isPredicatedBlock(i)) {
3245 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3247 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3251 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks[1]);
3252 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
3253 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3254 DT->changeImmediateDominator(LoopExitBlock, LoopBypassBlocks[0]);
3256 DEBUG(DT->verifyDomTree());
3259 /// \brief Check whether it is safe to if-convert this phi node.
3261 /// Phi nodes with constant expressions that can trap are not safe to if
3263 static bool canIfConvertPHINodes(BasicBlock *BB) {
3264 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3265 PHINode *Phi = dyn_cast<PHINode>(I);
3268 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3269 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3276 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3277 if (!EnableIfConversion) {
3278 emitAnalysis(VectorizationReport() << "if-conversion is disabled");
3282 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3284 // A list of pointers that we can safely read and write to.
3285 SmallPtrSet<Value *, 8> SafePointes;
3287 // Collect safe addresses.
3288 for (Loop::block_iterator BI = TheLoop->block_begin(),
3289 BE = TheLoop->block_end(); BI != BE; ++BI) {
3290 BasicBlock *BB = *BI;
3292 if (blockNeedsPredication(BB))
3295 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3296 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3297 SafePointes.insert(LI->getPointerOperand());
3298 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3299 SafePointes.insert(SI->getPointerOperand());
3303 // Collect the blocks that need predication.
3304 BasicBlock *Header = TheLoop->getHeader();
3305 for (Loop::block_iterator BI = TheLoop->block_begin(),
3306 BE = TheLoop->block_end(); BI != BE; ++BI) {
3307 BasicBlock *BB = *BI;
3309 // We don't support switch statements inside loops.
3310 if (!isa<BranchInst>(BB->getTerminator())) {
3311 emitAnalysis(VectorizationReport(BB->getTerminator())
3312 << "loop contains a switch statement");
3316 // We must be able to predicate all blocks that need to be predicated.
3317 if (blockNeedsPredication(BB)) {
3318 if (!blockCanBePredicated(BB, SafePointes)) {
3319 emitAnalysis(VectorizationReport(BB->getTerminator())
3320 << "control flow cannot be substituted for a select");
3323 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
3324 emitAnalysis(VectorizationReport(BB->getTerminator())
3325 << "control flow cannot be substituted for a select");
3330 // We can if-convert this loop.
3334 bool LoopVectorizationLegality::canVectorize() {
3335 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3336 // be canonicalized.
3337 if (!TheLoop->getLoopPreheader()) {
3339 VectorizationReport() <<
3340 "loop control flow is not understood by vectorizer");
3344 // We can only vectorize innermost loops.
3345 if (!TheLoop->getSubLoopsVector().empty()) {
3346 emitAnalysis(VectorizationReport() << "loop is not the innermost loop");
3350 // We must have a single backedge.
3351 if (TheLoop->getNumBackEdges() != 1) {
3353 VectorizationReport() <<
3354 "loop control flow is not understood by vectorizer");
3358 // We must have a single exiting block.
3359 if (!TheLoop->getExitingBlock()) {
3361 VectorizationReport() <<
3362 "loop control flow is not understood by vectorizer");
3366 // We only handle bottom-tested loops, i.e. loop in which the condition is
3367 // checked at the end of each iteration. With that we can assume that all
3368 // instructions in the loop are executed the same number of times.
3369 if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
3371 VectorizationReport() <<
3372 "loop control flow is not understood by vectorizer");
3376 // We need to have a loop header.
3377 DEBUG(dbgs() << "LV: Found a loop: " <<
3378 TheLoop->getHeader()->getName() << '\n');
3380 // Check if we can if-convert non-single-bb loops.
3381 unsigned NumBlocks = TheLoop->getNumBlocks();
3382 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3383 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3387 // ScalarEvolution needs to be able to find the exit count.
3388 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3389 if (ExitCount == SE->getCouldNotCompute()) {
3390 emitAnalysis(VectorizationReport() <<
3391 "could not determine number of loop iterations");
3392 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3396 // Check if we can vectorize the instructions and CFG in this loop.
3397 if (!canVectorizeInstrs()) {
3398 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3402 // Go over each instruction and look at memory deps.
3403 if (!canVectorizeMemory()) {
3404 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3408 // Collect all of the variables that remain uniform after vectorization.
3409 collectLoopUniforms();
3411 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
3412 (LAI->getRuntimePointerCheck()->Need ? " (with a runtime bound check)" :
3416 // Okay! We can vectorize. At this point we don't have any other mem analysis
3417 // which may limit our maximum vectorization factor, so just return true with
3422 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
3423 if (Ty->isPointerTy())
3424 return DL.getIntPtrType(Ty);
3426 // It is possible that char's or short's overflow when we ask for the loop's
3427 // trip count, work around this by changing the type size.
3428 if (Ty->getScalarSizeInBits() < 32)
3429 return Type::getInt32Ty(Ty->getContext());
3434 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
3435 Ty0 = convertPointerToIntegerType(DL, Ty0);
3436 Ty1 = convertPointerToIntegerType(DL, Ty1);
3437 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
3442 /// \brief Check that the instruction has outside loop users and is not an
3443 /// identified reduction variable.
3444 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
3445 SmallPtrSetImpl<Value *> &Reductions) {
3446 // Reduction instructions are allowed to have exit users. All other
3447 // instructions must not have external users.
3448 if (!Reductions.count(Inst))
3449 //Check that all of the users of the loop are inside the BB.
3450 for (User *U : Inst->users()) {
3451 Instruction *UI = cast<Instruction>(U);
3452 // This user may be a reduction exit value.
3453 if (!TheLoop->contains(UI)) {
3454 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
3461 bool LoopVectorizationLegality::canVectorizeInstrs() {
3462 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
3463 BasicBlock *Header = TheLoop->getHeader();
3465 // Look for the attribute signaling the absence of NaNs.
3466 Function &F = *Header->getParent();
3467 if (F.hasFnAttribute("no-nans-fp-math"))
3469 F.getFnAttribute("no-nans-fp-math").getValueAsString() == "true";
3471 // For each block in the loop.
3472 for (Loop::block_iterator bb = TheLoop->block_begin(),
3473 be = TheLoop->block_end(); bb != be; ++bb) {
3475 // Scan the instructions in the block and look for hazards.
3476 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3479 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
3480 Type *PhiTy = Phi->getType();
3481 // Check that this PHI type is allowed.
3482 if (!PhiTy->isIntegerTy() &&
3483 !PhiTy->isFloatingPointTy() &&
3484 !PhiTy->isPointerTy()) {
3485 emitAnalysis(VectorizationReport(it)
3486 << "loop control flow is not understood by vectorizer");
3487 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
3491 // If this PHINode is not in the header block, then we know that we
3492 // can convert it to select during if-conversion. No need to check if
3493 // the PHIs in this block are induction or reduction variables.
3494 if (*bb != Header) {
3495 // Check that this instruction has no outside users or is an
3496 // identified reduction value with an outside user.
3497 if (!hasOutsideLoopUser(TheLoop, it, AllowedExit))
3499 emitAnalysis(VectorizationReport(it) <<
3500 "value could not be identified as "
3501 "an induction or reduction variable");
3505 // We only allow if-converted PHIs with exactly two incoming values.
3506 if (Phi->getNumIncomingValues() != 2) {
3507 emitAnalysis(VectorizationReport(it)
3508 << "control flow not understood by vectorizer");
3509 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
3513 // This is the value coming from the preheader.
3514 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
3515 ConstantInt *StepValue = nullptr;
3516 // Check if this is an induction variable.
3517 InductionKind IK = isInductionVariable(Phi, StepValue);
3519 if (IK_NoInduction != IK) {
3520 // Get the widest type.
3522 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
3524 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
3526 // Int inductions are special because we only allow one IV.
3527 if (IK == IK_IntInduction && StepValue->isOne()) {
3528 // Use the phi node with the widest type as induction. Use the last
3529 // one if there are multiple (no good reason for doing this other
3530 // than it is expedient).
3531 if (!Induction || PhiTy == WidestIndTy)
3535 DEBUG(dbgs() << "LV: Found an induction variable.\n");
3536 Inductions[Phi] = InductionInfo(StartValue, IK, StepValue);
3538 // Until we explicitly handle the case of an induction variable with
3539 // an outside loop user we have to give up vectorizing this loop.
3540 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3541 emitAnalysis(VectorizationReport(it) <<
3542 "use of induction value outside of the "
3543 "loop is not handled by vectorizer");
3550 if (AddReductionVar(Phi, RK_IntegerAdd)) {
3551 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
3554 if (AddReductionVar(Phi, RK_IntegerMult)) {
3555 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
3558 if (AddReductionVar(Phi, RK_IntegerOr)) {
3559 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
3562 if (AddReductionVar(Phi, RK_IntegerAnd)) {
3563 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
3566 if (AddReductionVar(Phi, RK_IntegerXor)) {
3567 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
3570 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
3571 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
3574 if (AddReductionVar(Phi, RK_FloatMult)) {
3575 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
3578 if (AddReductionVar(Phi, RK_FloatAdd)) {
3579 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
3582 if (AddReductionVar(Phi, RK_FloatMinMax)) {
3583 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
3588 emitAnalysis(VectorizationReport(it) <<
3589 "value that could not be identified as "
3590 "reduction is used outside the loop");
3591 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
3593 }// end of PHI handling
3595 // We still don't handle functions. However, we can ignore dbg intrinsic
3596 // calls and we do handle certain intrinsic and libm functions.
3597 CallInst *CI = dyn_cast<CallInst>(it);
3598 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
3599 emitAnalysis(VectorizationReport(it) <<
3600 "call instruction cannot be vectorized");
3601 DEBUG(dbgs() << "LV: Found a call site.\n");
3605 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
3606 // second argument is the same (i.e. loop invariant)
3608 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
3609 if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
3610 emitAnalysis(VectorizationReport(it)
3611 << "intrinsic instruction cannot be vectorized");
3612 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
3617 // Check that the instruction return type is vectorizable.
3618 // Also, we can't vectorize extractelement instructions.
3619 if ((!VectorType::isValidElementType(it->getType()) &&
3620 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
3621 emitAnalysis(VectorizationReport(it)
3622 << "instruction return type cannot be vectorized");
3623 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
3627 // Check that the stored type is vectorizable.
3628 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
3629 Type *T = ST->getValueOperand()->getType();
3630 if (!VectorType::isValidElementType(T)) {
3631 emitAnalysis(VectorizationReport(ST) <<
3632 "store instruction cannot be vectorized");
3635 if (EnableMemAccessVersioning)
3636 collectStridedAccess(ST);
3639 if (EnableMemAccessVersioning)
3640 if (LoadInst *LI = dyn_cast<LoadInst>(it))
3641 collectStridedAccess(LI);
3643 // Reduction instructions are allowed to have exit users.
3644 // All other instructions must not have external users.
3645 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3646 emitAnalysis(VectorizationReport(it) <<
3647 "value cannot be used outside the loop");
3656 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
3657 if (Inductions.empty()) {
3658 emitAnalysis(VectorizationReport()
3659 << "loop induction variable could not be identified");
3667 ///\brief Remove GEPs whose indices but the last one are loop invariant and
3668 /// return the induction operand of the gep pointer.
3669 static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE,
3670 const DataLayout *DL, Loop *Lp) {
3671 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
3675 unsigned InductionOperand = getGEPInductionOperand(DL, GEP);
3677 // Check that all of the gep indices are uniform except for our induction
3679 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
3680 if (i != InductionOperand &&
3681 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
3683 return GEP->getOperand(InductionOperand);
3686 ///\brief Look for a cast use of the passed value.
3687 static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
3688 Value *UniqueCast = nullptr;
3689 for (User *U : Ptr->users()) {
3690 CastInst *CI = dyn_cast<CastInst>(U);
3691 if (CI && CI->getType() == Ty) {
3701 ///\brief Get the stride of a pointer access in a loop.
3702 /// Looks for symbolic strides "a[i*stride]". Returns the symbolic stride as a
3703 /// pointer to the Value, or null otherwise.
3704 static Value *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE,
3705 const DataLayout *DL, Loop *Lp) {
3706 const PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
3707 if (!PtrTy || PtrTy->isAggregateType())
3710 // Try to remove a gep instruction to make the pointer (actually index at this
3711 // point) easier analyzable. If OrigPtr is equal to Ptr we are analzying the
3712 // pointer, otherwise, we are analyzing the index.
3713 Value *OrigPtr = Ptr;
3715 // The size of the pointer access.
3716 int64_t PtrAccessSize = 1;
3718 Ptr = stripGetElementPtr(Ptr, SE, DL, Lp);
3719 const SCEV *V = SE->getSCEV(Ptr);
3723 while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V))
3724 V = C->getOperand();
3726 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
3730 V = S->getStepRecurrence(*SE);
3734 // Strip off the size of access multiplication if we are still analyzing the
3736 if (OrigPtr == Ptr) {
3737 DL->getTypeAllocSize(PtrTy->getElementType());
3738 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
3739 if (M->getOperand(0)->getSCEVType() != scConstant)
3742 const APInt &APStepVal =
3743 cast<SCEVConstant>(M->getOperand(0))->getValue()->getValue();
3745 // Huge step value - give up.
3746 if (APStepVal.getBitWidth() > 64)
3749 int64_t StepVal = APStepVal.getSExtValue();
3750 if (PtrAccessSize != StepVal)
3752 V = M->getOperand(1);
3757 Type *StripedOffRecurrenceCast = nullptr;
3758 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) {
3759 StripedOffRecurrenceCast = C->getType();
3760 V = C->getOperand();
3763 // Look for the loop invariant symbolic value.
3764 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
3768 Value *Stride = U->getValue();
3769 if (!Lp->isLoopInvariant(Stride))
3772 // If we have stripped off the recurrence cast we have to make sure that we
3773 // return the value that is used in this loop so that we can replace it later.
3774 if (StripedOffRecurrenceCast)
3775 Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
3780 void LoopVectorizationLegality::collectStridedAccess(Value *MemAccess) {
3781 Value *Ptr = nullptr;
3782 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
3783 Ptr = LI->getPointerOperand();
3784 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
3785 Ptr = SI->getPointerOperand();
3789 Value *Stride = getStrideFromPointer(Ptr, SE, DL, TheLoop);
3793 DEBUG(dbgs() << "LV: Found a strided access that we can version");
3794 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
3795 Strides[Ptr] = Stride;
3796 StrideSet.insert(Stride);
3799 void LoopVectorizationLegality::collectLoopUniforms() {
3800 // We now know that the loop is vectorizable!
3801 // Collect variables that will remain uniform after vectorization.
3802 std::vector<Value*> Worklist;
3803 BasicBlock *Latch = TheLoop->getLoopLatch();
3805 // Start with the conditional branch and walk up the block.
3806 Worklist.push_back(Latch->getTerminator()->getOperand(0));
3808 // Also add all consecutive pointer values; these values will be uniform
3809 // after vectorization (and subsequent cleanup) and, until revectorization is
3810 // supported, all dependencies must also be uniform.
3811 for (Loop::block_iterator B = TheLoop->block_begin(),
3812 BE = TheLoop->block_end(); B != BE; ++B)
3813 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
3815 if (I->getType()->isPointerTy() && isConsecutivePtr(I))
3816 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3818 while (!Worklist.empty()) {
3819 Instruction *I = dyn_cast<Instruction>(Worklist.back());
3820 Worklist.pop_back();
3822 // Look at instructions inside this loop.
3823 // Stop when reaching PHI nodes.
3824 // TODO: we need to follow values all over the loop, not only in this block.
3825 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
3828 // This is a known uniform.
3831 // Insert all operands.
3832 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3836 bool LoopVectorizationLegality::canVectorizeMemory() {
3837 LAI = &LAA->getInfo(TheLoop, Strides);
3838 auto &OptionalReport = LAI->getReport();
3840 emitAnalysis(*OptionalReport);
3841 return LAI->canVectorizeMemory();
3844 static bool hasMultipleUsesOf(Instruction *I,
3845 SmallPtrSetImpl<Instruction *> &Insts) {
3846 unsigned NumUses = 0;
3847 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
3848 if (Insts.count(dyn_cast<Instruction>(*Use)))
3857 static bool areAllUsesIn(Instruction *I, SmallPtrSetImpl<Instruction *> &Set) {
3858 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
3859 if (!Set.count(dyn_cast<Instruction>(*Use)))
3864 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
3865 ReductionKind Kind) {
3866 if (Phi->getNumIncomingValues() != 2)
3869 // Reduction variables are only found in the loop header block.
3870 if (Phi->getParent() != TheLoop->getHeader())
3873 // Obtain the reduction start value from the value that comes from the loop
3875 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
3877 // ExitInstruction is the single value which is used outside the loop.
3878 // We only allow for a single reduction value to be used outside the loop.
3879 // This includes users of the reduction, variables (which form a cycle
3880 // which ends in the phi node).
3881 Instruction *ExitInstruction = nullptr;
3882 // Indicates that we found a reduction operation in our scan.
3883 bool FoundReduxOp = false;
3885 // We start with the PHI node and scan for all of the users of this
3886 // instruction. All users must be instructions that can be used as reduction
3887 // variables (such as ADD). We must have a single out-of-block user. The cycle
3888 // must include the original PHI.
3889 bool FoundStartPHI = false;
3891 // To recognize min/max patterns formed by a icmp select sequence, we store
3892 // the number of instruction we saw from the recognized min/max pattern,
3893 // to make sure we only see exactly the two instructions.
3894 unsigned NumCmpSelectPatternInst = 0;
3895 ReductionInstDesc ReduxDesc(false, nullptr);
3897 SmallPtrSet<Instruction *, 8> VisitedInsts;
3898 SmallVector<Instruction *, 8> Worklist;
3899 Worklist.push_back(Phi);
3900 VisitedInsts.insert(Phi);
3902 // A value in the reduction can be used:
3903 // - By the reduction:
3904 // - Reduction operation:
3905 // - One use of reduction value (safe).
3906 // - Multiple use of reduction value (not safe).
3908 // - All uses of the PHI must be the reduction (safe).
3909 // - Otherwise, not safe.
3910 // - By one instruction outside of the loop (safe).
3911 // - By further instructions outside of the loop (not safe).
3912 // - By an instruction that is not part of the reduction (not safe).
3914 // * An instruction type other than PHI or the reduction operation.
3915 // * A PHI in the header other than the initial PHI.
3916 while (!Worklist.empty()) {
3917 Instruction *Cur = Worklist.back();
3918 Worklist.pop_back();
3921 // If the instruction has no users then this is a broken chain and can't be
3922 // a reduction variable.
3923 if (Cur->use_empty())
3926 bool IsAPhi = isa<PHINode>(Cur);
3928 // A header PHI use other than the original PHI.
3929 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
3932 // Reductions of instructions such as Div, and Sub is only possible if the
3933 // LHS is the reduction variable.
3934 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
3935 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
3936 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
3939 // Any reduction instruction must be of one of the allowed kinds.
3940 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
3941 if (!ReduxDesc.IsReduction)
3944 // A reduction operation must only have one use of the reduction value.
3945 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
3946 hasMultipleUsesOf(Cur, VisitedInsts))
3949 // All inputs to a PHI node must be a reduction value.
3950 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
3953 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
3954 isa<SelectInst>(Cur)))
3955 ++NumCmpSelectPatternInst;
3956 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
3957 isa<SelectInst>(Cur)))
3958 ++NumCmpSelectPatternInst;
3960 // Check whether we found a reduction operator.
3961 FoundReduxOp |= !IsAPhi;
3963 // Process users of current instruction. Push non-PHI nodes after PHI nodes
3964 // onto the stack. This way we are going to have seen all inputs to PHI
3965 // nodes once we get to them.
3966 SmallVector<Instruction *, 8> NonPHIs;
3967 SmallVector<Instruction *, 8> PHIs;
3968 for (User *U : Cur->users()) {
3969 Instruction *UI = cast<Instruction>(U);
3971 // Check if we found the exit user.
3972 BasicBlock *Parent = UI->getParent();
3973 if (!TheLoop->contains(Parent)) {
3974 // Exit if you find multiple outside users or if the header phi node is
3975 // being used. In this case the user uses the value of the previous
3976 // iteration, in which case we would loose "VF-1" iterations of the
3977 // reduction operation if we vectorize.
3978 if (ExitInstruction != nullptr || Cur == Phi)
3981 // The instruction used by an outside user must be the last instruction
3982 // before we feed back to the reduction phi. Otherwise, we loose VF-1
3983 // operations on the value.
3984 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
3987 ExitInstruction = Cur;
3991 // Process instructions only once (termination). Each reduction cycle
3992 // value must only be used once, except by phi nodes and min/max
3993 // reductions which are represented as a cmp followed by a select.
3994 ReductionInstDesc IgnoredVal(false, nullptr);
3995 if (VisitedInsts.insert(UI).second) {
3996 if (isa<PHINode>(UI))
3999 NonPHIs.push_back(UI);
4000 } else if (!isa<PHINode>(UI) &&
4001 ((!isa<FCmpInst>(UI) &&
4002 !isa<ICmpInst>(UI) &&
4003 !isa<SelectInst>(UI)) ||
4004 !isMinMaxSelectCmpPattern(UI, IgnoredVal).IsReduction))
4007 // Remember that we completed the cycle.
4009 FoundStartPHI = true;
4011 Worklist.append(PHIs.begin(), PHIs.end());
4012 Worklist.append(NonPHIs.begin(), NonPHIs.end());
4015 // This means we have seen one but not the other instruction of the
4016 // pattern or more than just a select and cmp.
4017 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
4018 NumCmpSelectPatternInst != 2)
4021 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
4024 // We found a reduction var if we have reached the original phi node and we
4025 // only have a single instruction with out-of-loop users.
4027 // This instruction is allowed to have out-of-loop users.
4028 AllowedExit.insert(ExitInstruction);
4030 // Save the description of this reduction variable.
4031 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
4032 ReduxDesc.MinMaxKind);
4033 Reductions[Phi] = RD;
4034 // We've ended the cycle. This is a reduction variable if we have an
4035 // outside user and it has a binary op.
4040 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
4041 /// pattern corresponding to a min(X, Y) or max(X, Y).
4042 LoopVectorizationLegality::ReductionInstDesc
4043 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
4044 ReductionInstDesc &Prev) {
4046 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
4047 "Expect a select instruction");
4048 Instruction *Cmp = nullptr;
4049 SelectInst *Select = nullptr;
4051 // We must handle the select(cmp()) as a single instruction. Advance to the
4053 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
4054 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->user_begin())))
4055 return ReductionInstDesc(false, I);
4056 return ReductionInstDesc(Select, Prev.MinMaxKind);
4059 // Only handle single use cases for now.
4060 if (!(Select = dyn_cast<SelectInst>(I)))
4061 return ReductionInstDesc(false, I);
4062 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
4063 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
4064 return ReductionInstDesc(false, I);
4065 if (!Cmp->hasOneUse())
4066 return ReductionInstDesc(false, I);
4071 // Look for a min/max pattern.
4072 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4073 return ReductionInstDesc(Select, MRK_UIntMin);
4074 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4075 return ReductionInstDesc(Select, MRK_UIntMax);
4076 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4077 return ReductionInstDesc(Select, MRK_SIntMax);
4078 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4079 return ReductionInstDesc(Select, MRK_SIntMin);
4080 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4081 return ReductionInstDesc(Select, MRK_FloatMin);
4082 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4083 return ReductionInstDesc(Select, MRK_FloatMax);
4084 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4085 return ReductionInstDesc(Select, MRK_FloatMin);
4086 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
4087 return ReductionInstDesc(Select, MRK_FloatMax);
4089 return ReductionInstDesc(false, I);
4092 LoopVectorizationLegality::ReductionInstDesc
4093 LoopVectorizationLegality::isReductionInstr(Instruction *I,
4095 ReductionInstDesc &Prev) {
4096 bool FP = I->getType()->isFloatingPointTy();
4097 bool FastMath = FP && I->hasUnsafeAlgebra();
4098 switch (I->getOpcode()) {
4100 return ReductionInstDesc(false, I);
4101 case Instruction::PHI:
4102 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
4103 Kind != RK_FloatMinMax))
4104 return ReductionInstDesc(false, I);
4105 return ReductionInstDesc(I, Prev.MinMaxKind);
4106 case Instruction::Sub:
4107 case Instruction::Add:
4108 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
4109 case Instruction::Mul:
4110 return ReductionInstDesc(Kind == RK_IntegerMult, I);
4111 case Instruction::And:
4112 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
4113 case Instruction::Or:
4114 return ReductionInstDesc(Kind == RK_IntegerOr, I);
4115 case Instruction::Xor:
4116 return ReductionInstDesc(Kind == RK_IntegerXor, I);
4117 case Instruction::FMul:
4118 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
4119 case Instruction::FSub:
4120 case Instruction::FAdd:
4121 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
4122 case Instruction::FCmp:
4123 case Instruction::ICmp:
4124 case Instruction::Select:
4125 if (Kind != RK_IntegerMinMax &&
4126 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
4127 return ReductionInstDesc(false, I);
4128 return isMinMaxSelectCmpPattern(I, Prev);
4132 LoopVectorizationLegality::InductionKind
4133 LoopVectorizationLegality::isInductionVariable(PHINode *Phi,
4134 ConstantInt *&StepValue) {
4135 Type *PhiTy = Phi->getType();
4136 // We only handle integer and pointer inductions variables.
4137 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
4138 return IK_NoInduction;
4140 // Check that the PHI is consecutive.
4141 const SCEV *PhiScev = SE->getSCEV(Phi);
4142 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
4144 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
4145 return IK_NoInduction;
4148 const SCEV *Step = AR->getStepRecurrence(*SE);
4149 // Calculate the pointer stride and check if it is consecutive.
4150 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4152 return IK_NoInduction;
4154 ConstantInt *CV = C->getValue();
4155 if (PhiTy->isIntegerTy()) {
4157 return IK_IntInduction;
4160 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
4161 Type *PointerElementType = PhiTy->getPointerElementType();
4162 // The pointer stride cannot be determined if the pointer element type is not
4164 if (!PointerElementType->isSized())
4165 return IK_NoInduction;
4167 int64_t Size = static_cast<int64_t>(DL->getTypeAllocSize(PointerElementType));
4168 int64_t CVSize = CV->getSExtValue();
4170 return IK_NoInduction;
4171 StepValue = ConstantInt::getSigned(CV->getType(), CVSize / Size);
4172 return IK_PtrInduction;
4175 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
4176 Value *In0 = const_cast<Value*>(V);
4177 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
4181 return Inductions.count(PN);
4184 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
4185 return LoopAccessInfo::blockNeedsPredication(BB, TheLoop, DT);
4188 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
4189 SmallPtrSetImpl<Value *> &SafePtrs) {
4191 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4192 // Check that we don't have a constant expression that can trap as operand.
4193 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
4195 if (Constant *C = dyn_cast<Constant>(*OI))
4199 // We might be able to hoist the load.
4200 if (it->mayReadFromMemory()) {
4201 LoadInst *LI = dyn_cast<LoadInst>(it);
4204 if (!SafePtrs.count(LI->getPointerOperand())) {
4205 if (isLegalMaskedLoad(LI->getType(), LI->getPointerOperand())) {
4206 MaskedOp.insert(LI);
4213 // We don't predicate stores at the moment.
4214 if (it->mayWriteToMemory()) {
4215 StoreInst *SI = dyn_cast<StoreInst>(it);
4216 // We only support predication of stores in basic blocks with one
4221 bool isSafePtr = (SafePtrs.count(SI->getPointerOperand()) != 0);
4222 bool isSinglePredecessor = SI->getParent()->getSinglePredecessor();
4224 if (++NumPredStores > NumberOfStoresToPredicate || !isSafePtr ||
4225 !isSinglePredecessor) {
4226 // Build a masked store if it is legal for the target, otherwise scalarize
4228 bool isLegalMaskedOp =
4229 isLegalMaskedStore(SI->getValueOperand()->getType(),
4230 SI->getPointerOperand());
4231 if (isLegalMaskedOp) {
4233 MaskedOp.insert(SI);
4242 // The instructions below can trap.
4243 switch (it->getOpcode()) {
4245 case Instruction::UDiv:
4246 case Instruction::SDiv:
4247 case Instruction::URem:
4248 case Instruction::SRem:
4256 LoopVectorizationCostModel::VectorizationFactor
4257 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) {
4258 // Width 1 means no vectorize
4259 VectorizationFactor Factor = { 1U, 0U };
4260 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
4261 emitAnalysis(VectorizationReport() <<
4262 "runtime pointer checks needed. Enable vectorization of this "
4263 "loop with '#pragma clang loop vectorize(enable)' when "
4264 "compiling with -Os");
4265 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
4269 if (!EnableCondStoresVectorization && Legal->getNumPredStores()) {
4270 emitAnalysis(VectorizationReport() <<
4271 "store that is conditionally executed prevents vectorization");
4272 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
4276 // Find the trip count.
4277 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4278 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
4280 unsigned WidestType = getWidestType();
4281 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
4282 unsigned MaxSafeDepDist = -1U;
4283 if (Legal->getMaxSafeDepDistBytes() != -1U)
4284 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
4285 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
4286 WidestRegister : MaxSafeDepDist);
4287 unsigned MaxVectorSize = WidestRegister / WidestType;
4288 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
4289 DEBUG(dbgs() << "LV: The Widest register is: "
4290 << WidestRegister << " bits.\n");
4292 if (MaxVectorSize == 0) {
4293 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
4297 assert(MaxVectorSize <= 64 && "Did not expect to pack so many elements"
4298 " into one vector!");
4300 unsigned VF = MaxVectorSize;
4302 // If we optimize the program for size, avoid creating the tail loop.
4304 // If we are unable to calculate the trip count then don't try to vectorize.
4307 (VectorizationReport() <<
4308 "unable to calculate the loop count due to complex control flow");
4309 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4313 // Find the maximum SIMD width that can fit within the trip count.
4314 VF = TC % MaxVectorSize;
4319 // If the trip count that we found modulo the vectorization factor is not
4320 // zero then we require a tail.
4322 emitAnalysis(VectorizationReport() <<
4323 "cannot optimize for size and vectorize at the "
4324 "same time. Enable vectorization of this loop "
4325 "with '#pragma clang loop vectorize(enable)' "
4326 "when compiling with -Os");
4327 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
4332 int UserVF = Hints->getWidth();
4334 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
4335 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
4337 Factor.Width = UserVF;
4341 float Cost = expectedCost(1);
4343 const float ScalarCost = Cost;
4346 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
4348 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
4349 // Ignore scalar width, because the user explicitly wants vectorization.
4350 if (ForceVectorization && VF > 1) {
4352 Cost = expectedCost(Width) / (float)Width;
4355 for (unsigned i=2; i <= VF; i*=2) {
4356 // Notice that the vector loop needs to be executed less times, so
4357 // we need to divide the cost of the vector loops by the width of
4358 // the vector elements.
4359 float VectorCost = expectedCost(i) / (float)i;
4360 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
4361 (int)VectorCost << ".\n");
4362 if (VectorCost < Cost) {
4368 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
4369 << "LV: Vectorization seems to be not beneficial, "
4370 << "but was forced by a user.\n");
4371 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
4372 Factor.Width = Width;
4373 Factor.Cost = Width * Cost;
4377 unsigned LoopVectorizationCostModel::getWidestType() {
4378 unsigned MaxWidth = 8;
4381 for (Loop::block_iterator bb = TheLoop->block_begin(),
4382 be = TheLoop->block_end(); bb != be; ++bb) {
4383 BasicBlock *BB = *bb;
4385 // For each instruction in the loop.
4386 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4387 Type *T = it->getType();
4389 // Ignore ephemeral values.
4390 if (EphValues.count(it))
4393 // Only examine Loads, Stores and PHINodes.
4394 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
4397 // Examine PHI nodes that are reduction variables.
4398 if (PHINode *PN = dyn_cast<PHINode>(it))
4399 if (!Legal->getReductionVars()->count(PN))
4402 // Examine the stored values.
4403 if (StoreInst *ST = dyn_cast<StoreInst>(it))
4404 T = ST->getValueOperand()->getType();
4406 // Ignore loaded pointer types and stored pointer types that are not
4407 // consecutive. However, we do want to take consecutive stores/loads of
4408 // pointer vectors into account.
4409 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
4412 MaxWidth = std::max(MaxWidth,
4413 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
4421 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
4423 unsigned LoopCost) {
4425 // -- The unroll heuristics --
4426 // We unroll the loop in order to expose ILP and reduce the loop overhead.
4427 // There are many micro-architectural considerations that we can't predict
4428 // at this level. For example, frontend pressure (on decode or fetch) due to
4429 // code size, or the number and capabilities of the execution ports.
4431 // We use the following heuristics to select the unroll factor:
4432 // 1. If the code has reductions, then we unroll in order to break the cross
4433 // iteration dependency.
4434 // 2. If the loop is really small, then we unroll in order to reduce the loop
4436 // 3. We don't unroll if we think that we will spill registers to memory due
4437 // to the increased register pressure.
4439 // Use the user preference, unless 'auto' is selected.
4440 int UserUF = Hints->getInterleave();
4444 // When we optimize for size, we don't unroll.
4448 // We used the distance for the unroll factor.
4449 if (Legal->getMaxSafeDepDistBytes() != -1U)
4452 // Do not unroll loops with a relatively small trip count.
4453 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
4454 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
4457 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
4458 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
4462 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
4463 TargetNumRegisters = ForceTargetNumScalarRegs;
4465 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
4466 TargetNumRegisters = ForceTargetNumVectorRegs;
4469 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
4470 // We divide by these constants so assume that we have at least one
4471 // instruction that uses at least one register.
4472 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
4473 R.NumInstructions = std::max(R.NumInstructions, 1U);
4475 // We calculate the unroll factor using the following formula.
4476 // Subtract the number of loop invariants from the number of available
4477 // registers. These registers are used by all of the unrolled instances.
4478 // Next, divide the remaining registers by the number of registers that is
4479 // required by the loop, in order to estimate how many parallel instances
4480 // fit without causing spills. All of this is rounded down if necessary to be
4481 // a power of two. We want power of two unroll factors to simplify any
4482 // addressing operations or alignment considerations.
4483 unsigned UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
4486 // Don't count the induction variable as unrolled.
4487 if (EnableIndVarRegisterHeur)
4488 UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
4489 std::max(1U, (R.MaxLocalUsers - 1)));
4491 // Clamp the unroll factor ranges to reasonable factors.
4492 unsigned MaxInterleaveSize = TTI.getMaxInterleaveFactor();
4494 // Check if the user has overridden the unroll max.
4496 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
4497 MaxInterleaveSize = ForceTargetMaxScalarInterleaveFactor;
4499 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
4500 MaxInterleaveSize = ForceTargetMaxVectorInterleaveFactor;
4503 // If we did not calculate the cost for VF (because the user selected the VF)
4504 // then we calculate the cost of VF here.
4506 LoopCost = expectedCost(VF);
4508 // Clamp the calculated UF to be between the 1 and the max unroll factor
4509 // that the target allows.
4510 if (UF > MaxInterleaveSize)
4511 UF = MaxInterleaveSize;
4515 // Unroll if we vectorized this loop and there is a reduction that could
4516 // benefit from unrolling.
4517 if (VF > 1 && Legal->getReductionVars()->size()) {
4518 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
4522 // Note that if we've already vectorized the loop we will have done the
4523 // runtime check and so unrolling won't require further checks.
4524 bool UnrollingRequiresRuntimePointerCheck =
4525 (VF == 1 && Legal->getRuntimePointerCheck()->Need);
4527 // We want to unroll small loops in order to reduce the loop overhead and
4528 // potentially expose ILP opportunities.
4529 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
4530 if (!UnrollingRequiresRuntimePointerCheck &&
4531 LoopCost < SmallLoopCost) {
4532 // We assume that the cost overhead is 1 and we use the cost model
4533 // to estimate the cost of the loop and unroll until the cost of the
4534 // loop overhead is about 5% of the cost of the loop.
4535 unsigned SmallUF = std::min(UF, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
4537 // Unroll until store/load ports (estimated by max unroll factor) are
4539 unsigned NumStores = Legal->getNumStores();
4540 unsigned NumLoads = Legal->getNumLoads();
4541 unsigned StoresUF = UF / (NumStores ? NumStores : 1);
4542 unsigned LoadsUF = UF / (NumLoads ? NumLoads : 1);
4544 // If we have a scalar reduction (vector reductions are already dealt with
4545 // by this point), we can increase the critical path length if the loop
4546 // we're unrolling is inside another loop. Limit, by default to 2, so the
4547 // critical path only gets increased by one reduction operation.
4548 if (Legal->getReductionVars()->size() &&
4549 TheLoop->getLoopDepth() > 1) {
4550 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionUF);
4551 SmallUF = std::min(SmallUF, F);
4552 StoresUF = std::min(StoresUF, F);
4553 LoadsUF = std::min(LoadsUF, F);
4556 if (EnableLoadStoreRuntimeUnroll && std::max(StoresUF, LoadsUF) > SmallUF) {
4557 DEBUG(dbgs() << "LV: Unrolling to saturate store or load ports.\n");
4558 return std::max(StoresUF, LoadsUF);
4561 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
4565 DEBUG(dbgs() << "LV: Not Unrolling.\n");
4569 LoopVectorizationCostModel::RegisterUsage
4570 LoopVectorizationCostModel::calculateRegisterUsage() {
4571 // This function calculates the register usage by measuring the highest number
4572 // of values that are alive at a single location. Obviously, this is a very
4573 // rough estimation. We scan the loop in a topological order in order and
4574 // assign a number to each instruction. We use RPO to ensure that defs are
4575 // met before their users. We assume that each instruction that has in-loop
4576 // users starts an interval. We record every time that an in-loop value is
4577 // used, so we have a list of the first and last occurrences of each
4578 // instruction. Next, we transpose this data structure into a multi map that
4579 // holds the list of intervals that *end* at a specific location. This multi
4580 // map allows us to perform a linear search. We scan the instructions linearly
4581 // and record each time that a new interval starts, by placing it in a set.
4582 // If we find this value in the multi-map then we remove it from the set.
4583 // The max register usage is the maximum size of the set.
4584 // We also search for instructions that are defined outside the loop, but are
4585 // used inside the loop. We need this number separately from the max-interval
4586 // usage number because when we unroll, loop-invariant values do not take
4588 LoopBlocksDFS DFS(TheLoop);
4592 R.NumInstructions = 0;
4594 // Each 'key' in the map opens a new interval. The values
4595 // of the map are the index of the 'last seen' usage of the
4596 // instruction that is the key.
4597 typedef DenseMap<Instruction*, unsigned> IntervalMap;
4598 // Maps instruction to its index.
4599 DenseMap<unsigned, Instruction*> IdxToInstr;
4600 // Marks the end of each interval.
4601 IntervalMap EndPoint;
4602 // Saves the list of instruction indices that are used in the loop.
4603 SmallSet<Instruction*, 8> Ends;
4604 // Saves the list of values that are used in the loop but are
4605 // defined outside the loop, such as arguments and constants.
4606 SmallPtrSet<Value*, 8> LoopInvariants;
4609 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
4610 be = DFS.endRPO(); bb != be; ++bb) {
4611 R.NumInstructions += (*bb)->size();
4612 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4614 Instruction *I = it;
4615 IdxToInstr[Index++] = I;
4617 // Save the end location of each USE.
4618 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
4619 Value *U = I->getOperand(i);
4620 Instruction *Instr = dyn_cast<Instruction>(U);
4622 // Ignore non-instruction values such as arguments, constants, etc.
4623 if (!Instr) continue;
4625 // If this instruction is outside the loop then record it and continue.
4626 if (!TheLoop->contains(Instr)) {
4627 LoopInvariants.insert(Instr);
4631 // Overwrite previous end points.
4632 EndPoint[Instr] = Index;
4638 // Saves the list of intervals that end with the index in 'key'.
4639 typedef SmallVector<Instruction*, 2> InstrList;
4640 DenseMap<unsigned, InstrList> TransposeEnds;
4642 // Transpose the EndPoints to a list of values that end at each index.
4643 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
4645 TransposeEnds[it->second].push_back(it->first);
4647 SmallSet<Instruction*, 8> OpenIntervals;
4648 unsigned MaxUsage = 0;
4651 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
4652 for (unsigned int i = 0; i < Index; ++i) {
4653 Instruction *I = IdxToInstr[i];
4654 // Ignore instructions that are never used within the loop.
4655 if (!Ends.count(I)) continue;
4657 // Ignore ephemeral values.
4658 if (EphValues.count(I))
4661 // Remove all of the instructions that end at this location.
4662 InstrList &List = TransposeEnds[i];
4663 for (unsigned int j=0, e = List.size(); j < e; ++j)
4664 OpenIntervals.erase(List[j]);
4666 // Count the number of live interals.
4667 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
4669 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
4670 OpenIntervals.size() << '\n');
4672 // Add the current instruction to the list of open intervals.
4673 OpenIntervals.insert(I);
4676 unsigned Invariant = LoopInvariants.size();
4677 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
4678 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
4679 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
4681 R.LoopInvariantRegs = Invariant;
4682 R.MaxLocalUsers = MaxUsage;
4686 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
4690 for (Loop::block_iterator bb = TheLoop->block_begin(),
4691 be = TheLoop->block_end(); bb != be; ++bb) {
4692 unsigned BlockCost = 0;
4693 BasicBlock *BB = *bb;
4695 // For each instruction in the old loop.
4696 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
4697 // Skip dbg intrinsics.
4698 if (isa<DbgInfoIntrinsic>(it))
4701 // Ignore ephemeral values.
4702 if (EphValues.count(it))
4705 unsigned C = getInstructionCost(it, VF);
4707 // Check if we should override the cost.
4708 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
4709 C = ForceTargetInstructionCost;
4712 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
4713 VF << " For instruction: " << *it << '\n');
4716 // We assume that if-converted blocks have a 50% chance of being executed.
4717 // When the code is scalar then some of the blocks are avoided due to CF.
4718 // When the code is vectorized we execute all code paths.
4719 if (VF == 1 && Legal->blockNeedsPredication(*bb))
4728 /// \brief Check whether the address computation for a non-consecutive memory
4729 /// access looks like an unlikely candidate for being merged into the indexing
4732 /// We look for a GEP which has one index that is an induction variable and all
4733 /// other indices are loop invariant. If the stride of this access is also
4734 /// within a small bound we decide that this address computation can likely be
4735 /// merged into the addressing mode.
4736 /// In all other cases, we identify the address computation as complex.
4737 static bool isLikelyComplexAddressComputation(Value *Ptr,
4738 LoopVectorizationLegality *Legal,
4739 ScalarEvolution *SE,
4740 const Loop *TheLoop) {
4741 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
4745 // We are looking for a gep with all loop invariant indices except for one
4746 // which should be an induction variable.
4747 unsigned NumOperands = Gep->getNumOperands();
4748 for (unsigned i = 1; i < NumOperands; ++i) {
4749 Value *Opd = Gep->getOperand(i);
4750 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
4751 !Legal->isInductionVariable(Opd))
4755 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
4756 // can likely be merged into the address computation.
4757 unsigned MaxMergeDistance = 64;
4759 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
4763 // Check the step is constant.
4764 const SCEV *Step = AddRec->getStepRecurrence(*SE);
4765 // Calculate the pointer stride and check if it is consecutive.
4766 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4770 const APInt &APStepVal = C->getValue()->getValue();
4772 // Huge step value - give up.
4773 if (APStepVal.getBitWidth() > 64)
4776 int64_t StepVal = APStepVal.getSExtValue();
4778 return StepVal > MaxMergeDistance;
4781 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
4782 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
4788 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
4789 // If we know that this instruction will remain uniform, check the cost of
4790 // the scalar version.
4791 if (Legal->isUniformAfterVectorization(I))
4794 Type *RetTy = I->getType();
4795 Type *VectorTy = ToVectorTy(RetTy, VF);
4797 // TODO: We need to estimate the cost of intrinsic calls.
4798 switch (I->getOpcode()) {
4799 case Instruction::GetElementPtr:
4800 // We mark this instruction as zero-cost because the cost of GEPs in
4801 // vectorized code depends on whether the corresponding memory instruction
4802 // is scalarized or not. Therefore, we handle GEPs with the memory
4803 // instruction cost.
4805 case Instruction::Br: {
4806 return TTI.getCFInstrCost(I->getOpcode());
4808 case Instruction::PHI:
4809 //TODO: IF-converted IFs become selects.
4811 case Instruction::Add:
4812 case Instruction::FAdd:
4813 case Instruction::Sub:
4814 case Instruction::FSub:
4815 case Instruction::Mul:
4816 case Instruction::FMul:
4817 case Instruction::UDiv:
4818 case Instruction::SDiv:
4819 case Instruction::FDiv:
4820 case Instruction::URem:
4821 case Instruction::SRem:
4822 case Instruction::FRem:
4823 case Instruction::Shl:
4824 case Instruction::LShr:
4825 case Instruction::AShr:
4826 case Instruction::And:
4827 case Instruction::Or:
4828 case Instruction::Xor: {
4829 // Since we will replace the stride by 1 the multiplication should go away.
4830 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
4832 // Certain instructions can be cheaper to vectorize if they have a constant
4833 // second vector operand. One example of this are shifts on x86.
4834 TargetTransformInfo::OperandValueKind Op1VK =
4835 TargetTransformInfo::OK_AnyValue;
4836 TargetTransformInfo::OperandValueKind Op2VK =
4837 TargetTransformInfo::OK_AnyValue;
4838 TargetTransformInfo::OperandValueProperties Op1VP =
4839 TargetTransformInfo::OP_None;
4840 TargetTransformInfo::OperandValueProperties Op2VP =
4841 TargetTransformInfo::OP_None;
4842 Value *Op2 = I->getOperand(1);
4844 // Check for a splat of a constant or for a non uniform vector of constants.
4845 if (isa<ConstantInt>(Op2)) {
4846 ConstantInt *CInt = cast<ConstantInt>(Op2);
4847 if (CInt && CInt->getValue().isPowerOf2())
4848 Op2VP = TargetTransformInfo::OP_PowerOf2;
4849 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
4850 } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
4851 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
4852 Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
4854 ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
4855 if (CInt && CInt->getValue().isPowerOf2())
4856 Op2VP = TargetTransformInfo::OP_PowerOf2;
4857 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
4861 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK,
4864 case Instruction::Select: {
4865 SelectInst *SI = cast<SelectInst>(I);
4866 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
4867 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
4868 Type *CondTy = SI->getCondition()->getType();
4870 CondTy = VectorType::get(CondTy, VF);
4872 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
4874 case Instruction::ICmp:
4875 case Instruction::FCmp: {
4876 Type *ValTy = I->getOperand(0)->getType();
4877 VectorTy = ToVectorTy(ValTy, VF);
4878 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
4880 case Instruction::Store:
4881 case Instruction::Load: {
4882 StoreInst *SI = dyn_cast<StoreInst>(I);
4883 LoadInst *LI = dyn_cast<LoadInst>(I);
4884 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
4886 VectorTy = ToVectorTy(ValTy, VF);
4888 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
4889 unsigned AS = SI ? SI->getPointerAddressSpace() :
4890 LI->getPointerAddressSpace();
4891 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
4892 // We add the cost of address computation here instead of with the gep
4893 // instruction because only here we know whether the operation is
4896 return TTI.getAddressComputationCost(VectorTy) +
4897 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4899 // Scalarized loads/stores.
4900 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
4901 bool Reverse = ConsecutiveStride < 0;
4902 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
4903 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
4904 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
4905 bool IsComplexComputation =
4906 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
4908 // The cost of extracting from the value vector and pointer vector.
4909 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
4910 for (unsigned i = 0; i < VF; ++i) {
4911 // The cost of extracting the pointer operand.
4912 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
4913 // In case of STORE, the cost of ExtractElement from the vector.
4914 // In case of LOAD, the cost of InsertElement into the returned
4916 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
4917 Instruction::InsertElement,
4921 // The cost of the scalar loads/stores.
4922 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
4923 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
4928 // Wide load/stores.
4929 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
4930 if (Legal->isMaskRequired(I))
4931 Cost += TTI.getMaskedMemoryOpCost(I->getOpcode(), VectorTy, Alignment,
4934 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
4937 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
4941 case Instruction::ZExt:
4942 case Instruction::SExt:
4943 case Instruction::FPToUI:
4944 case Instruction::FPToSI:
4945 case Instruction::FPExt:
4946 case Instruction::PtrToInt:
4947 case Instruction::IntToPtr:
4948 case Instruction::SIToFP:
4949 case Instruction::UIToFP:
4950 case Instruction::Trunc:
4951 case Instruction::FPTrunc:
4952 case Instruction::BitCast: {
4953 // We optimize the truncation of induction variable.
4954 // The cost of these is the same as the scalar operation.
4955 if (I->getOpcode() == Instruction::Trunc &&
4956 Legal->isInductionVariable(I->getOperand(0)))
4957 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
4958 I->getOperand(0)->getType());
4960 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
4961 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
4963 case Instruction::Call: {
4964 CallInst *CI = cast<CallInst>(I);
4965 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
4966 assert(ID && "Not an intrinsic call!");
4967 Type *RetTy = ToVectorTy(CI->getType(), VF);
4968 SmallVector<Type*, 4> Tys;
4969 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
4970 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
4971 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
4974 // We are scalarizing the instruction. Return the cost of the scalar
4975 // instruction, plus the cost of insert and extract into vector
4976 // elements, times the vector width.
4979 if (!RetTy->isVoidTy() && VF != 1) {
4980 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
4982 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
4985 // The cost of inserting the results plus extracting each one of the
4987 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
4990 // The cost of executing VF copies of the scalar instruction. This opcode
4991 // is unknown. Assume that it is the same as 'mul'.
4992 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
4998 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
4999 if (Scalar->isVoidTy() || VF == 1)
5001 return VectorType::get(Scalar, VF);
5004 char LoopVectorize::ID = 0;
5005 static const char lv_name[] = "Loop Vectorization";
5006 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
5007 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
5008 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
5009 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
5010 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfo)
5011 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
5012 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
5013 INITIALIZE_PASS_DEPENDENCY(LCSSA)
5014 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
5015 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
5016 INITIALIZE_PASS_DEPENDENCY(LoopAccessAnalysis)
5017 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
5020 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
5021 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
5025 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
5026 // Check for a store.
5027 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
5028 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
5030 // Check for a load.
5031 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
5032 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
5038 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
5039 bool IfPredicateStore) {
5040 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
5041 // Holds vector parameters or scalars, in case of uniform vals.
5042 SmallVector<VectorParts, 4> Params;
5044 setDebugLocFromInst(Builder, Instr);
5046 // Find all of the vectorized parameters.
5047 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5048 Value *SrcOp = Instr->getOperand(op);
5050 // If we are accessing the old induction variable, use the new one.
5051 if (SrcOp == OldInduction) {
5052 Params.push_back(getVectorValue(SrcOp));
5056 // Try using previously calculated values.
5057 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
5059 // If the src is an instruction that appeared earlier in the basic block
5060 // then it should already be vectorized.
5061 if (SrcInst && OrigLoop->contains(SrcInst)) {
5062 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
5063 // The parameter is a vector value from earlier.
5064 Params.push_back(WidenMap.get(SrcInst));
5066 // The parameter is a scalar from outside the loop. Maybe even a constant.
5067 VectorParts Scalars;
5068 Scalars.append(UF, SrcOp);
5069 Params.push_back(Scalars);
5073 assert(Params.size() == Instr->getNumOperands() &&
5074 "Invalid number of operands");
5076 // Does this instruction return a value ?
5077 bool IsVoidRetTy = Instr->getType()->isVoidTy();
5079 Value *UndefVec = IsVoidRetTy ? nullptr :
5080 UndefValue::get(Instr->getType());
5081 // Create a new entry in the WidenMap and initialize it to Undef or Null.
5082 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
5084 Instruction *InsertPt = Builder.GetInsertPoint();
5085 BasicBlock *IfBlock = Builder.GetInsertBlock();
5086 BasicBlock *CondBlock = nullptr;
5089 Loop *VectorLp = nullptr;
5090 if (IfPredicateStore) {
5091 assert(Instr->getParent()->getSinglePredecessor() &&
5092 "Only support single predecessor blocks");
5093 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
5094 Instr->getParent());
5095 VectorLp = LI->getLoopFor(IfBlock);
5096 assert(VectorLp && "Must have a loop for this block");
5099 // For each vector unroll 'part':
5100 for (unsigned Part = 0; Part < UF; ++Part) {
5101 // For each scalar that we create:
5103 // Start an "if (pred) a[i] = ..." block.
5104 Value *Cmp = nullptr;
5105 if (IfPredicateStore) {
5106 if (Cond[Part]->getType()->isVectorTy())
5108 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
5109 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
5110 ConstantInt::get(Cond[Part]->getType(), 1));
5111 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
5112 LoopVectorBody.push_back(CondBlock);
5113 VectorLp->addBasicBlockToLoop(CondBlock, *LI);
5114 // Update Builder with newly created basic block.
5115 Builder.SetInsertPoint(InsertPt);
5118 Instruction *Cloned = Instr->clone();
5120 Cloned->setName(Instr->getName() + ".cloned");
5121 // Replace the operands of the cloned instructions with extracted scalars.
5122 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
5123 Value *Op = Params[op][Part];
5124 Cloned->setOperand(op, Op);
5127 // Place the cloned scalar in the new loop.
5128 Builder.Insert(Cloned);
5130 // If the original scalar returns a value we need to place it in a vector
5131 // so that future users will be able to use it.
5133 VecResults[Part] = Cloned;
5136 if (IfPredicateStore) {
5137 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
5138 LoopVectorBody.push_back(NewIfBlock);
5139 VectorLp->addBasicBlockToLoop(NewIfBlock, *LI);
5140 Builder.SetInsertPoint(InsertPt);
5141 Instruction *OldBr = IfBlock->getTerminator();
5142 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
5143 OldBr->eraseFromParent();
5144 IfBlock = NewIfBlock;
5149 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
5150 StoreInst *SI = dyn_cast<StoreInst>(Instr);
5151 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
5153 return scalarizeInstruction(Instr, IfPredicateStore);
5156 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
5160 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
5164 Value *InnerLoopUnroller::getStepVector(Value *Val, int StartIdx, Value *Step) {
5165 // When unrolling and the VF is 1, we only need to add a simple scalar.
5166 Type *ITy = Val->getType();
5167 assert(!ITy->isVectorTy() && "Val must be a scalar");
5168 Constant *C = ConstantInt::get(ITy, StartIdx);
5169 return Builder.CreateAdd(Val, Builder.CreateMul(C, Step), "induction");