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/BlockFrequencyInfo.h"
59 #include "llvm/Analysis/LoopInfo.h"
60 #include "llvm/Analysis/LoopIterator.h"
61 #include "llvm/Analysis/LoopPass.h"
62 #include "llvm/Analysis/ScalarEvolution.h"
63 #include "llvm/Analysis/ScalarEvolutionExpander.h"
64 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
65 #include "llvm/Analysis/TargetTransformInfo.h"
66 #include "llvm/Analysis/ValueTracking.h"
67 #include "llvm/IR/Constants.h"
68 #include "llvm/IR/DataLayout.h"
69 #include "llvm/IR/DebugInfo.h"
70 #include "llvm/IR/DerivedTypes.h"
71 #include "llvm/IR/DiagnosticInfo.h"
72 #include "llvm/IR/Dominators.h"
73 #include "llvm/IR/Function.h"
74 #include "llvm/IR/IRBuilder.h"
75 #include "llvm/IR/Instructions.h"
76 #include "llvm/IR/IntrinsicInst.h"
77 #include "llvm/IR/LLVMContext.h"
78 #include "llvm/IR/Module.h"
79 #include "llvm/IR/PatternMatch.h"
80 #include "llvm/IR/Type.h"
81 #include "llvm/IR/Value.h"
82 #include "llvm/IR/ValueHandle.h"
83 #include "llvm/IR/Verifier.h"
84 #include "llvm/Pass.h"
85 #include "llvm/Support/BranchProbability.h"
86 #include "llvm/Support/CommandLine.h"
87 #include "llvm/Support/Debug.h"
88 #include "llvm/Support/raw_ostream.h"
89 #include "llvm/Transforms/Scalar.h"
90 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
91 #include "llvm/Transforms/Utils/Local.h"
92 #include "llvm/Transforms/Utils/VectorUtils.h"
98 using namespace llvm::PatternMatch;
100 #define LV_NAME "loop-vectorize"
101 #define DEBUG_TYPE LV_NAME
103 STATISTIC(LoopsVectorized, "Number of loops vectorized");
104 STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization");
106 static cl::opt<unsigned>
107 VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden,
108 cl::desc("Sets the SIMD width. Zero is autoselect."));
110 static cl::opt<unsigned>
111 VectorizationInterleave("force-vector-interleave", cl::init(0), cl::Hidden,
112 cl::desc("Sets the vectorization interleave count. "
113 "Zero is autoselect."));
116 EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
117 cl::desc("Enable if-conversion during vectorization."));
119 /// We don't vectorize loops with a known constant trip count below this number.
120 static cl::opt<unsigned>
121 TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
123 cl::desc("Don't vectorize loops with a constant "
124 "trip count that is smaller than this "
127 /// This enables versioning on the strides of symbolically striding memory
128 /// accesses in code like the following.
129 /// for (i = 0; i < N; ++i)
130 /// A[i * Stride1] += B[i * Stride2] ...
132 /// Will be roughly translated to
133 /// if (Stride1 == 1 && Stride2 == 1) {
134 /// for (i = 0; i < N; i+=4)
138 static cl::opt<bool> EnableMemAccessVersioning(
139 "enable-mem-access-versioning", cl::init(true), cl::Hidden,
140 cl::desc("Enable symblic stride memory access versioning"));
142 /// We don't unroll loops with a known constant trip count below this number.
143 static const unsigned TinyTripCountUnrollThreshold = 128;
145 /// When performing memory disambiguation checks at runtime do not make more
146 /// than this number of comparisons.
147 static const unsigned RuntimeMemoryCheckThreshold = 8;
149 /// Maximum simd width.
150 static const unsigned MaxVectorWidth = 64;
152 static cl::opt<unsigned> ForceTargetNumScalarRegs(
153 "force-target-num-scalar-regs", cl::init(0), cl::Hidden,
154 cl::desc("A flag that overrides the target's number of scalar registers."));
156 static cl::opt<unsigned> ForceTargetNumVectorRegs(
157 "force-target-num-vector-regs", cl::init(0), cl::Hidden,
158 cl::desc("A flag that overrides the target's number of vector registers."));
160 /// Maximum vectorization interleave count.
161 static const unsigned MaxInterleaveFactor = 16;
163 static cl::opt<unsigned> ForceTargetMaxScalarInterleaveFactor(
164 "force-target-max-scalar-interleave", cl::init(0), cl::Hidden,
165 cl::desc("A flag that overrides the target's max interleave factor for "
168 static cl::opt<unsigned> ForceTargetMaxVectorInterleaveFactor(
169 "force-target-max-vector-interleave", cl::init(0), cl::Hidden,
170 cl::desc("A flag that overrides the target's max interleave factor for "
171 "vectorized loops."));
173 static cl::opt<unsigned> ForceTargetInstructionCost(
174 "force-target-instruction-cost", cl::init(0), cl::Hidden,
175 cl::desc("A flag that overrides the target's expected cost for "
176 "an instruction to a single constant value. Mostly "
177 "useful for getting consistent testing."));
179 static cl::opt<unsigned> SmallLoopCost(
180 "small-loop-cost", cl::init(20), cl::Hidden,
181 cl::desc("The cost of a loop that is considered 'small' by the unroller."));
183 static cl::opt<bool> LoopVectorizeWithBlockFrequency(
184 "loop-vectorize-with-block-frequency", cl::init(false), cl::Hidden,
185 cl::desc("Enable the use of the block frequency analysis to access PGO "
186 "heuristics minimizing code growth in cold regions and being more "
187 "aggressive in hot regions."));
189 // Runtime unroll loops for load/store throughput.
190 static cl::opt<bool> EnableLoadStoreRuntimeUnroll(
191 "enable-loadstore-runtime-unroll", cl::init(true), cl::Hidden,
192 cl::desc("Enable runtime unrolling until load/store ports are saturated"));
194 /// The number of stores in a loop that are allowed to need predication.
195 static cl::opt<unsigned> NumberOfStoresToPredicate(
196 "vectorize-num-stores-pred", cl::init(1), cl::Hidden,
197 cl::desc("Max number of stores to be predicated behind an if."));
199 static cl::opt<bool> EnableIndVarRegisterHeur(
200 "enable-ind-var-reg-heur", cl::init(true), cl::Hidden,
201 cl::desc("Count the induction variable only once when unrolling"));
203 static cl::opt<bool> EnableCondStoresVectorization(
204 "enable-cond-stores-vec", cl::init(false), cl::Hidden,
205 cl::desc("Enable if predication of stores during vectorization."));
207 static cl::opt<unsigned> MaxNestedScalarReductionUF(
208 "max-nested-scalar-reduction-unroll", cl::init(2), cl::Hidden,
209 cl::desc("The maximum unroll factor to use when unrolling a scalar "
210 "reduction in a nested loop."));
214 // Forward declarations.
215 class LoopVectorizationLegality;
216 class LoopVectorizationCostModel;
217 class LoopVectorizeHints;
219 /// Optimization analysis message produced during vectorization. Messages inform
220 /// the user why vectorization did not occur.
223 raw_string_ostream Out;
227 Report(Instruction *I = nullptr) : Out(Message), Instr(I) {
228 Out << "loop not vectorized: ";
231 template <typename A> Report &operator<<(const A &Value) {
236 Instruction *getInstr() { return Instr; }
238 std::string &str() { return Out.str(); }
239 operator Twine() { return Out.str(); }
242 /// InnerLoopVectorizer vectorizes loops which contain only one basic
243 /// block to a specified vectorization factor (VF).
244 /// This class performs the widening of scalars into vectors, or multiple
245 /// scalars. This class also implements the following features:
246 /// * It inserts an epilogue loop for handling loops that don't have iteration
247 /// counts that are known to be a multiple of the vectorization factor.
248 /// * It handles the code generation for reduction variables.
249 /// * Scalarization (implementation using scalars) of un-vectorizable
251 /// InnerLoopVectorizer does not perform any vectorization-legality
252 /// checks, and relies on the caller to check for the different legality
253 /// aspects. The InnerLoopVectorizer relies on the
254 /// LoopVectorizationLegality class to provide information about the induction
255 /// and reduction variables that were found to a given vectorization factor.
256 class InnerLoopVectorizer {
258 InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
259 DominatorTree *DT, const DataLayout *DL,
260 const TargetLibraryInfo *TLI, unsigned VecWidth,
261 unsigned UnrollFactor)
262 : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), TLI(TLI),
263 VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()),
264 Induction(nullptr), OldInduction(nullptr), WidenMap(UnrollFactor),
267 // Perform the actual loop widening (vectorization).
268 void vectorize(LoopVectorizationLegality *L) {
270 // Create a new empty loop. Unlink the old loop and connect the new one.
272 // Widen each instruction in the old loop to a new one in the new loop.
273 // Use the Legality module to find the induction and reduction variables.
275 // Register the new loop and update the analysis passes.
279 virtual ~InnerLoopVectorizer() {}
282 /// A small list of PHINodes.
283 typedef SmallVector<PHINode*, 4> PhiVector;
284 /// When we unroll loops we have multiple vector values for each scalar.
285 /// This data structure holds the unrolled and vectorized values that
286 /// originated from one scalar instruction.
287 typedef SmallVector<Value*, 2> VectorParts;
289 // When we if-convert we need create edge masks. We have to cache values so
290 // that we don't end up with exponential recursion/IR.
291 typedef DenseMap<std::pair<BasicBlock*, BasicBlock*>,
292 VectorParts> EdgeMaskCache;
294 /// \brief Add code that checks at runtime if the accessed arrays overlap.
296 /// Returns a pair of instructions where the first element is the first
297 /// instruction generated in possibly a sequence of instructions and the
298 /// second value is the final comparator value or NULL if no check is needed.
299 std::pair<Instruction *, Instruction *> addRuntimeCheck(Instruction *Loc);
301 /// \brief Add checks for strides that where assumed to be 1.
303 /// Returns the last check instruction and the first check instruction in the
304 /// pair as (first, last).
305 std::pair<Instruction *, Instruction *> addStrideCheck(Instruction *Loc);
307 /// Create an empty loop, based on the loop ranges of the old loop.
308 void createEmptyLoop();
309 /// Copy and widen the instructions from the old loop.
310 virtual void vectorizeLoop();
312 /// \brief The Loop exit block may have single value PHI nodes where the
313 /// incoming value is 'Undef'. While vectorizing we only handled real values
314 /// that were defined inside the loop. Here we fix the 'undef case'.
318 /// A helper function that computes the predicate of the block BB, assuming
319 /// that the header block of the loop is set to True. It returns the *entry*
320 /// mask for the block BB.
321 VectorParts createBlockInMask(BasicBlock *BB);
322 /// A helper function that computes the predicate of the edge between SRC
324 VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
326 /// A helper function to vectorize a single BB within the innermost loop.
327 void vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV);
329 /// Vectorize a single PHINode in a block. This method handles the induction
330 /// variable canonicalization. It supports both VF = 1 for unrolled loops and
331 /// arbitrary length vectors.
332 void widenPHIInstruction(Instruction *PN, VectorParts &Entry,
333 unsigned UF, unsigned VF, PhiVector *PV);
335 /// Insert the new loop to the loop hierarchy and pass manager
336 /// and update the analysis passes.
337 void updateAnalysis();
339 /// This instruction is un-vectorizable. Implement it as a sequence
340 /// of scalars. If \p IfPredicateStore is true we need to 'hide' each
341 /// scalarized instruction behind an if block predicated on the control
342 /// dependence of the instruction.
343 virtual void scalarizeInstruction(Instruction *Instr,
344 bool IfPredicateStore=false);
346 /// Vectorize Load and Store instructions,
347 virtual void vectorizeMemoryInstruction(Instruction *Instr);
349 /// Create a broadcast instruction. This method generates a broadcast
350 /// instruction (shuffle) for loop invariant values and for the induction
351 /// value. If this is the induction variable then we extend it to N, N+1, ...
352 /// this is needed because each iteration in the loop corresponds to a SIMD
354 virtual Value *getBroadcastInstrs(Value *V);
356 /// This function adds 0, 1, 2 ... to each vector element, starting at zero.
357 /// If Negate is set then negative numbers are added e.g. (0, -1, -2, ...).
358 /// The sequence starts at StartIndex.
359 virtual Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate);
361 /// When we go over instructions in the basic block we rely on previous
362 /// values within the current basic block or on loop invariant values.
363 /// When we widen (vectorize) values we place them in the map. If the values
364 /// are not within the map, they have to be loop invariant, so we simply
365 /// broadcast them into a vector.
366 VectorParts &getVectorValue(Value *V);
368 /// Generate a shuffle sequence that will reverse the vector Vec.
369 virtual Value *reverseVector(Value *Vec);
371 /// This is a helper class that holds the vectorizer state. It maps scalar
372 /// instructions to vector instructions. When the code is 'unrolled' then
373 /// then a single scalar value is mapped to multiple vector parts. The parts
374 /// are stored in the VectorPart type.
376 /// C'tor. UnrollFactor controls the number of vectors ('parts') that
378 ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
380 /// \return True if 'Key' is saved in the Value Map.
381 bool has(Value *Key) const { return MapStorage.count(Key); }
383 /// Initializes a new entry in the map. Sets all of the vector parts to the
384 /// save value in 'Val'.
385 /// \return A reference to a vector with splat values.
386 VectorParts &splat(Value *Key, Value *Val) {
387 VectorParts &Entry = MapStorage[Key];
388 Entry.assign(UF, Val);
392 ///\return A reference to the value that is stored at 'Key'.
393 VectorParts &get(Value *Key) {
394 VectorParts &Entry = MapStorage[Key];
397 assert(Entry.size() == UF);
402 /// The unroll factor. Each entry in the map stores this number of vector
406 /// Map storage. We use std::map and not DenseMap because insertions to a
407 /// dense map invalidates its iterators.
408 std::map<Value *, VectorParts> MapStorage;
411 /// The original loop.
413 /// Scev analysis to use.
422 const DataLayout *DL;
423 /// Target Library Info.
424 const TargetLibraryInfo *TLI;
426 /// The vectorization SIMD factor to use. Each vector will have this many
431 /// The vectorization unroll factor to use. Each scalar is vectorized to this
432 /// many different vector instructions.
435 /// The builder that we use
438 // --- Vectorization state ---
440 /// The vector-loop preheader.
441 BasicBlock *LoopVectorPreHeader;
442 /// The scalar-loop preheader.
443 BasicBlock *LoopScalarPreHeader;
444 /// Middle Block between the vector and the scalar.
445 BasicBlock *LoopMiddleBlock;
446 ///The ExitBlock of the scalar loop.
447 BasicBlock *LoopExitBlock;
448 ///The vector loop body.
449 SmallVector<BasicBlock *, 4> LoopVectorBody;
450 ///The scalar loop body.
451 BasicBlock *LoopScalarBody;
452 /// A list of all bypass blocks. The first block is the entry of the loop.
453 SmallVector<BasicBlock *, 4> LoopBypassBlocks;
455 /// The new Induction variable which was added to the new block.
457 /// The induction variable of the old basic block.
458 PHINode *OldInduction;
459 /// Holds the extended (to the widest induction type) start index.
461 /// Maps scalars to widened vectors.
463 EdgeMaskCache MaskCache;
465 LoopVectorizationLegality *Legal;
468 class InnerLoopUnroller : public InnerLoopVectorizer {
470 InnerLoopUnroller(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
471 DominatorTree *DT, const DataLayout *DL,
472 const TargetLibraryInfo *TLI, unsigned UnrollFactor) :
473 InnerLoopVectorizer(OrigLoop, SE, LI, DT, DL, TLI, 1, UnrollFactor) { }
476 void scalarizeInstruction(Instruction *Instr,
477 bool IfPredicateStore = false) override;
478 void vectorizeMemoryInstruction(Instruction *Instr) override;
479 Value *getBroadcastInstrs(Value *V) override;
480 Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate) override;
481 Value *reverseVector(Value *Vec) override;
484 /// \brief Look for a meaningful debug location on the instruction or it's
486 static Instruction *getDebugLocFromInstOrOperands(Instruction *I) {
491 if (I->getDebugLoc() != Empty)
494 for (User::op_iterator OI = I->op_begin(), OE = I->op_end(); OI != OE; ++OI) {
495 if (Instruction *OpInst = dyn_cast<Instruction>(*OI))
496 if (OpInst->getDebugLoc() != Empty)
503 /// \brief Set the debug location in the builder using the debug location in the
505 static void setDebugLocFromInst(IRBuilder<> &B, const Value *Ptr) {
506 if (const Instruction *Inst = dyn_cast_or_null<Instruction>(Ptr))
507 B.SetCurrentDebugLocation(Inst->getDebugLoc());
509 B.SetCurrentDebugLocation(DebugLoc());
513 /// \return string containing a file name and a line # for the given loop.
514 static std::string getDebugLocString(const Loop *L) {
517 raw_string_ostream OS(Result);
518 const DebugLoc LoopDbgLoc = L->getStartLoc();
519 if (!LoopDbgLoc.isUnknown())
520 LoopDbgLoc.print(L->getHeader()->getContext(), OS);
522 // Just print the module name.
523 OS << L->getHeader()->getParent()->getParent()->getModuleIdentifier();
530 /// \brief Propagate known metadata from one instruction to another.
531 static void propagateMetadata(Instruction *To, const Instruction *From) {
532 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
533 From->getAllMetadataOtherThanDebugLoc(Metadata);
535 for (auto M : Metadata) {
536 unsigned Kind = M.first;
538 // These are safe to transfer (this is safe for TBAA, even when we
539 // if-convert, because should that metadata have had a control dependency
540 // on the condition, and thus actually aliased with some other
541 // non-speculated memory access when the condition was false, this would be
542 // caught by the runtime overlap checks).
543 if (Kind != LLVMContext::MD_tbaa &&
544 Kind != LLVMContext::MD_alias_scope &&
545 Kind != LLVMContext::MD_noalias &&
546 Kind != LLVMContext::MD_fpmath)
549 To->setMetadata(Kind, M.second);
553 /// \brief Propagate known metadata from one instruction to a vector of others.
554 static void propagateMetadata(SmallVectorImpl<Value *> &To, const Instruction *From) {
556 if (Instruction *I = dyn_cast<Instruction>(V))
557 propagateMetadata(I, From);
560 /// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
561 /// to what vectorization factor.
562 /// This class does not look at the profitability of vectorization, only the
563 /// legality. This class has two main kinds of checks:
564 /// * Memory checks - The code in canVectorizeMemory checks if vectorization
565 /// will change the order of memory accesses in a way that will change the
566 /// correctness of the program.
567 /// * Scalars checks - The code in canVectorizeInstrs and canVectorizeMemory
568 /// checks for a number of different conditions, such as the availability of a
569 /// single induction variable, that all types are supported and vectorize-able,
570 /// etc. This code reflects the capabilities of InnerLoopVectorizer.
571 /// This class is also used by InnerLoopVectorizer for identifying
572 /// induction variable and the different reduction variables.
573 class LoopVectorizationLegality {
577 unsigned NumPredStores;
579 LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, const DataLayout *DL,
580 DominatorTree *DT, TargetLibraryInfo *TLI,
581 AliasAnalysis *AA, Function *F)
582 : NumLoads(0), NumStores(0), NumPredStores(0), TheLoop(L), SE(SE), DL(DL),
583 DT(DT), TLI(TLI), AA(AA), TheFunction(F), Induction(nullptr),
584 WidestIndTy(nullptr), HasFunNoNaNAttr(false), MaxSafeDepDistBytes(-1U) {
587 /// This enum represents the kinds of reductions that we support.
589 RK_NoReduction, ///< Not a reduction.
590 RK_IntegerAdd, ///< Sum of integers.
591 RK_IntegerMult, ///< Product of integers.
592 RK_IntegerOr, ///< Bitwise or logical OR of numbers.
593 RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
594 RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
595 RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
596 RK_FloatAdd, ///< Sum of floats.
597 RK_FloatMult, ///< Product of floats.
598 RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()).
601 /// This enum represents the kinds of inductions that we support.
603 IK_NoInduction, ///< Not an induction variable.
604 IK_IntInduction, ///< Integer induction variable. Step = 1.
605 IK_ReverseIntInduction, ///< Reverse int induction variable. Step = -1.
606 IK_PtrInduction, ///< Pointer induction var. Step = sizeof(elem).
607 IK_ReversePtrInduction ///< Reverse ptr indvar. Step = - sizeof(elem).
610 // This enum represents the kind of minmax reduction.
611 enum MinMaxReductionKind {
621 /// This struct holds information about reduction variables.
622 struct ReductionDescriptor {
623 ReductionDescriptor() : StartValue(nullptr), LoopExitInstr(nullptr),
624 Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
626 ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
627 MinMaxReductionKind MK)
628 : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
630 // The starting value of the reduction.
631 // It does not have to be zero!
632 TrackingVH<Value> StartValue;
633 // The instruction who's value is used outside the loop.
634 Instruction *LoopExitInstr;
635 // The kind of the reduction.
637 // If this a min/max reduction the kind of reduction.
638 MinMaxReductionKind MinMaxKind;
641 /// This POD struct holds information about a potential reduction operation.
642 struct ReductionInstDesc {
643 ReductionInstDesc(bool IsRedux, Instruction *I) :
644 IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
646 ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
647 IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
649 // Is this instruction a reduction candidate.
651 // The last instruction in a min/max pattern (select of the select(icmp())
652 // pattern), or the current reduction instruction otherwise.
653 Instruction *PatternLastInst;
654 // If this is a min/max pattern the comparison predicate.
655 MinMaxReductionKind MinMaxKind;
658 /// This struct holds information about the memory runtime legality
659 /// check that a group of pointers do not overlap.
660 struct RuntimePointerCheck {
661 RuntimePointerCheck() : Need(false) {}
663 /// Reset the state of the pointer runtime information.
670 DependencySetId.clear();
674 /// Insert a pointer and calculate the start and end SCEVs.
675 void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr,
676 unsigned DepSetId, unsigned ASId, ValueToValueMap &Strides);
678 /// This flag indicates if we need to add the runtime check.
680 /// Holds the pointers that we need to check.
681 SmallVector<TrackingVH<Value>, 2> Pointers;
682 /// Holds the pointer value at the beginning of the loop.
683 SmallVector<const SCEV*, 2> Starts;
684 /// Holds the pointer value at the end of the loop.
685 SmallVector<const SCEV*, 2> Ends;
686 /// Holds the information if this pointer is used for writing to memory.
687 SmallVector<bool, 2> IsWritePtr;
688 /// Holds the id of the set of pointers that could be dependent because of a
689 /// shared underlying object.
690 SmallVector<unsigned, 2> DependencySetId;
691 /// Holds the id of the disjoint alias set to which this pointer belongs.
692 SmallVector<unsigned, 2> AliasSetId;
695 /// A struct for saving information about induction variables.
696 struct InductionInfo {
697 InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {}
698 InductionInfo() : StartValue(nullptr), IK(IK_NoInduction) {}
700 TrackingVH<Value> StartValue;
705 /// ReductionList contains the reduction descriptors for all
706 /// of the reductions that were found in the loop.
707 typedef DenseMap<PHINode*, ReductionDescriptor> ReductionList;
709 /// InductionList saves induction variables and maps them to the
710 /// induction descriptor.
711 typedef MapVector<PHINode*, InductionInfo> InductionList;
713 /// Returns true if it is legal to vectorize this loop.
714 /// This does not mean that it is profitable to vectorize this
715 /// loop, only that it is legal to do so.
718 /// Returns the Induction variable.
719 PHINode *getInduction() { return Induction; }
721 /// Returns the reduction variables found in the loop.
722 ReductionList *getReductionVars() { return &Reductions; }
724 /// Returns the induction variables found in the loop.
725 InductionList *getInductionVars() { return &Inductions; }
727 /// Returns the widest induction type.
728 Type *getWidestInductionType() { return WidestIndTy; }
730 /// Returns True if V is an induction variable in this loop.
731 bool isInductionVariable(const Value *V);
733 /// Return true if the block BB needs to be predicated in order for the loop
734 /// to be vectorized.
735 bool blockNeedsPredication(BasicBlock *BB);
737 /// Check if this pointer is consecutive when vectorizing. This happens
738 /// when the last index of the GEP is the induction variable, or that the
739 /// pointer itself is an induction variable.
740 /// This check allows us to vectorize A[idx] into a wide load/store.
742 /// 0 - Stride is unknown or non-consecutive.
743 /// 1 - Address is consecutive.
744 /// -1 - Address is consecutive, and decreasing.
745 int isConsecutivePtr(Value *Ptr);
747 /// Returns true if the value V is uniform within the loop.
748 bool isUniform(Value *V);
750 /// Returns true if this instruction will remain scalar after vectorization.
751 bool isUniformAfterVectorization(Instruction* I) { return Uniforms.count(I); }
753 /// Returns the information that we collected about runtime memory check.
754 RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; }
756 /// This function returns the identity element (or neutral element) for
758 static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
760 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
762 bool hasStride(Value *V) { return StrideSet.count(V); }
763 bool mustCheckStrides() { return !StrideSet.empty(); }
764 SmallPtrSet<Value *, 8>::iterator strides_begin() {
765 return StrideSet.begin();
767 SmallPtrSet<Value *, 8>::iterator strides_end() { return StrideSet.end(); }
770 /// Check if a single basic block loop is vectorizable.
771 /// At this point we know that this is a loop with a constant trip count
772 /// and we only need to check individual instructions.
773 bool canVectorizeInstrs();
775 /// When we vectorize loops we may change the order in which
776 /// we read and write from memory. This method checks if it is
777 /// legal to vectorize the code, considering only memory constrains.
778 /// Returns true if the loop is vectorizable
779 bool canVectorizeMemory();
781 /// Return true if we can vectorize this loop using the IF-conversion
783 bool canVectorizeWithIfConvert();
785 /// Collect the variables that need to stay uniform after vectorization.
786 void collectLoopUniforms();
788 /// Return true if all of the instructions in the block can be speculatively
789 /// executed. \p SafePtrs is a list of addresses that are known to be legal
790 /// and we know that we can read from them without segfault.
791 bool blockCanBePredicated(BasicBlock *BB, SmallPtrSetImpl<Value *> &SafePtrs);
793 /// Returns True, if 'Phi' is the kind of reduction variable for type
794 /// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
795 bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
796 /// Returns a struct describing if the instruction 'I' can be a reduction
797 /// variable of type 'Kind'. If the reduction is a min/max pattern of
798 /// select(icmp()) this function advances the instruction pointer 'I' from the
799 /// compare instruction to the select instruction and stores this pointer in
800 /// 'PatternLastInst' member of the returned struct.
801 ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
802 ReductionInstDesc &Desc);
803 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
804 /// pattern corresponding to a min(X, Y) or max(X, Y).
805 static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
806 ReductionInstDesc &Prev);
807 /// Returns the induction kind of Phi. This function may return NoInduction
808 /// if the PHI is not an induction variable.
809 InductionKind isInductionVariable(PHINode *Phi);
811 /// \brief Collect memory access with loop invariant strides.
813 /// Looks for accesses like "a[i * StrideA]" where "StrideA" is loop
815 void collectStridedAcccess(Value *LoadOrStoreInst);
817 /// Report an analysis message to assist the user in diagnosing loops that are
819 void emitAnalysis(Report &Message) {
820 DebugLoc DL = TheLoop->getStartLoc();
821 if (Instruction *I = Message.getInstr())
822 DL = I->getDebugLoc();
823 emitOptimizationRemarkAnalysis(TheFunction->getContext(), DEBUG_TYPE,
824 *TheFunction, DL, Message.str());
827 /// The loop that we evaluate.
831 /// DataLayout analysis.
832 const DataLayout *DL;
835 /// Target Library Info.
836 TargetLibraryInfo *TLI;
840 Function *TheFunction;
842 // --- vectorization state --- //
844 /// Holds the integer induction variable. This is the counter of the
847 /// Holds the reduction variables.
848 ReductionList Reductions;
849 /// Holds all of the induction variables that we found in the loop.
850 /// Notice that inductions don't need to start at zero and that induction
851 /// variables can be pointers.
852 InductionList Inductions;
853 /// Holds the widest induction type encountered.
856 /// Allowed outside users. This holds the reduction
857 /// vars which can be accessed from outside the loop.
858 SmallPtrSet<Value*, 4> AllowedExit;
859 /// This set holds the variables which are known to be uniform after
861 SmallPtrSet<Instruction*, 4> Uniforms;
862 /// We need to check that all of the pointers in this list are disjoint
864 RuntimePointerCheck PtrRtCheck;
865 /// Can we assume the absence of NaNs.
866 bool HasFunNoNaNAttr;
868 unsigned MaxSafeDepDistBytes;
870 ValueToValueMap Strides;
871 SmallPtrSet<Value *, 8> StrideSet;
874 /// LoopVectorizationCostModel - estimates the expected speedups due to
876 /// In many cases vectorization is not profitable. This can happen because of
877 /// a number of reasons. In this class we mainly attempt to predict the
878 /// expected speedup/slowdowns due to the supported instruction set. We use the
879 /// TargetTransformInfo to query the different backends for the cost of
880 /// different operations.
881 class LoopVectorizationCostModel {
883 LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
884 LoopVectorizationLegality *Legal,
885 const TargetTransformInfo &TTI,
886 const DataLayout *DL, const TargetLibraryInfo *TLI,
887 const Function *F, const LoopVectorizeHints *Hints)
888 : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI), TheFunction(F), Hints(Hints) {}
890 /// Information about vectorization costs
891 struct VectorizationFactor {
892 unsigned Width; // Vector width with best cost
893 unsigned Cost; // Cost of the loop with that width
895 /// \return The most profitable vectorization factor and the cost of that VF.
896 /// This method checks every power of two up to VF. If UserVF is not ZERO
897 /// then this vectorization factor will be selected if vectorization is
899 VectorizationFactor selectVectorizationFactor(bool OptForSize);
901 /// \return The size (in bits) of the widest type in the code that
902 /// needs to be vectorized. We ignore values that remain scalar such as
903 /// 64 bit loop indices.
904 unsigned getWidestType();
906 /// \return The most profitable unroll factor.
907 /// If UserUF is non-zero then this method finds the best unroll-factor
908 /// based on register pressure and other parameters.
909 /// VF and LoopCost are the selected vectorization factor and the cost of the
911 unsigned selectUnrollFactor(bool OptForSize, unsigned VF, unsigned LoopCost);
913 /// \brief A struct that represents some properties of the register usage
915 struct RegisterUsage {
916 /// Holds the number of loop invariant values that are used in the loop.
917 unsigned LoopInvariantRegs;
918 /// Holds the maximum number of concurrent live intervals in the loop.
919 unsigned MaxLocalUsers;
920 /// Holds the number of instructions in the loop.
921 unsigned NumInstructions;
924 /// \return information about the register usage of the loop.
925 RegisterUsage calculateRegisterUsage();
928 /// Returns the expected execution cost. The unit of the cost does
929 /// not matter because we use the 'cost' units to compare different
930 /// vector widths. The cost that is returned is *not* normalized by
931 /// the factor width.
932 unsigned expectedCost(unsigned VF);
934 /// Returns the execution time cost of an instruction for a given vector
935 /// width. Vector width of one means scalar.
936 unsigned getInstructionCost(Instruction *I, unsigned VF);
938 /// A helper function for converting Scalar types to vector types.
939 /// If the incoming type is void, we return void. If the VF is 1, we return
941 static Type* ToVectorTy(Type *Scalar, unsigned VF);
943 /// Returns whether the instruction is a load or store and will be a emitted
944 /// as a vector operation.
945 bool isConsecutiveLoadOrStore(Instruction *I);
947 /// Report an analysis message to assist the user in diagnosing loops that are
949 void emitAnalysis(Report &Message) {
950 DebugLoc DL = TheLoop->getStartLoc();
951 if (Instruction *I = Message.getInstr())
952 DL = I->getDebugLoc();
953 emitOptimizationRemarkAnalysis(TheFunction->getContext(), DEBUG_TYPE,
954 *TheFunction, DL, Message.str());
957 /// The loop that we evaluate.
961 /// Loop Info analysis.
963 /// Vectorization legality.
964 LoopVectorizationLegality *Legal;
965 /// Vector target information.
966 const TargetTransformInfo &TTI;
967 /// Target data layout information.
968 const DataLayout *DL;
969 /// Target Library Info.
970 const TargetLibraryInfo *TLI;
971 const Function *TheFunction;
972 // Loop Vectorize Hint.
973 const LoopVectorizeHints *Hints;
976 /// Utility class for getting and setting loop vectorizer hints in the form
977 /// of loop metadata.
978 /// This class keeps a number of loop annotations locally (as member variables)
979 /// and can, upon request, write them back as metadata on the loop. It will
980 /// initially scan the loop for existing metadata, and will update the local
981 /// values based on information in the loop.
982 /// We cannot write all values to metadata, as the mere presence of some info,
983 /// for example 'force', means a decision has been made. So, we need to be
984 /// careful NOT to add them if the user hasn't specifically asked so.
985 class LoopVectorizeHints {
992 /// Hint - associates name and validation with the hint value.
995 unsigned Value; // This may have to change for non-numeric values.
998 Hint(const char * Name, unsigned Value, HintKind Kind)
999 : Name(Name), Value(Value), Kind(Kind) { }
1001 bool validate(unsigned Val) {
1004 return isPowerOf2_32(Val) && Val <= MaxVectorWidth;
1006 return isPowerOf2_32(Val) && Val <= MaxInterleaveFactor;
1014 /// Vectorization width.
1016 /// Vectorization interleave factor.
1018 /// Vectorization forced
1020 /// Array to help iterating through all hints.
1021 Hint *Hints[3]; // avoiding initialisation due to MSVC2012
1023 /// Return the loop metadata prefix.
1024 static StringRef Prefix() { return "llvm.loop."; }
1028 FK_Undefined = -1, ///< Not selected.
1029 FK_Disabled = 0, ///< Forcing disabled.
1030 FK_Enabled = 1, ///< Forcing enabled.
1033 LoopVectorizeHints(const Loop *L, bool DisableInterleaving)
1034 : Width("vectorize.width", VectorizationFactor, HK_WIDTH),
1035 Interleave("interleave.count", DisableInterleaving, HK_UNROLL),
1036 Force("vectorize.enable", FK_Undefined, HK_FORCE),
1038 // FIXME: Move this up initialisation when MSVC requirement is 2013+
1040 Hints[1] = &Interleave;
1043 // Populate values with existing loop metadata.
1044 getHintsFromMetadata();
1046 // force-vector-interleave overrides DisableInterleaving.
1047 if (VectorizationInterleave.getNumOccurrences() > 0)
1048 Interleave.Value = VectorizationInterleave;
1050 DEBUG(if (DisableInterleaving && Interleave.Value == 1) dbgs()
1051 << "LV: Interleaving disabled by the pass manager\n");
1054 /// Mark the loop L as already vectorized by setting the width to 1.
1055 void setAlreadyVectorized() {
1056 Width.Value = Interleave.Value = 1;
1057 // FIXME: Change all lines below for this when we can use MSVC 2013+
1058 //writeHintsToMetadata({ Width, Unroll });
1059 std::vector<Hint> hints;
1061 hints.emplace_back(Width);
1062 hints.emplace_back(Interleave);
1063 writeHintsToMetadata(std::move(hints));
1066 /// Dumps all the hint information.
1067 std::string emitRemark() const {
1069 if (Force.Value == LoopVectorizeHints::FK_Disabled)
1070 R << "vectorization is explicitly disabled";
1072 R << "use -Rpass-analysis=loop-vectorize for more info";
1073 if (Force.Value == LoopVectorizeHints::FK_Enabled) {
1074 R << " (Force=true";
1075 if (Width.Value != 0)
1076 R << ", Vector Width=" << Width.Value;
1077 if (Interleave.Value != 0)
1078 R << ", Interleave Count=" << Interleave.Value;
1086 unsigned getWidth() const { return Width.Value; }
1087 unsigned getInterleave() const { return Interleave.Value; }
1088 enum ForceKind getForce() const { return (ForceKind)Force.Value; }
1091 /// Find hints specified in the loop metadata and update local values.
1092 void getHintsFromMetadata() {
1093 MDNode *LoopID = TheLoop->getLoopID();
1097 // First operand should refer to the loop id itself.
1098 assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
1099 assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
1101 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1102 const MDString *S = nullptr;
1103 SmallVector<Value*, 4> Args;
1105 // The expected hint is either a MDString or a MDNode with the first
1106 // operand a MDString.
1107 if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
1108 if (!MD || MD->getNumOperands() == 0)
1110 S = dyn_cast<MDString>(MD->getOperand(0));
1111 for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
1112 Args.push_back(MD->getOperand(i));
1114 S = dyn_cast<MDString>(LoopID->getOperand(i));
1115 assert(Args.size() == 0 && "too many arguments for MDString");
1121 // Check if the hint starts with the loop metadata prefix.
1122 StringRef Name = S->getString();
1123 if (Args.size() == 1)
1124 setHint(Name, Args[0]);
1128 /// Checks string hint with one operand and set value if valid.
1129 void setHint(StringRef Name, Value *Arg) {
1130 if (!Name.startswith(Prefix()))
1132 Name = Name.substr(Prefix().size(), StringRef::npos);
1134 const ConstantInt *C = dyn_cast<ConstantInt>(Arg);
1136 unsigned Val = C->getZExtValue();
1138 for (auto H : Hints) {
1139 if (Name == H->Name) {
1140 if (H->validate(Val))
1143 DEBUG(dbgs() << "LV: ignoring invalid hint '" << Name << "'\n");
1149 /// Create a new hint from name / value pair.
1150 MDNode *createHintMetadata(StringRef Name, unsigned V) const {
1151 LLVMContext &Context = TheLoop->getHeader()->getContext();
1152 SmallVector<Value*, 2> Vals;
1153 Vals.push_back(MDString::get(Context, Name));
1154 Vals.push_back(ConstantInt::get(Type::getInt32Ty(Context), V));
1155 return MDNode::get(Context, Vals);
1158 /// Matches metadata with hint name.
1159 bool matchesHintMetadataName(MDNode *Node, std::vector<Hint> &HintTypes) {
1160 MDString* Name = dyn_cast<MDString>(Node->getOperand(0));
1164 for (auto H : HintTypes)
1165 if (Name->getName().endswith(H.Name))
1170 /// Sets current hints into loop metadata, keeping other values intact.
1171 void writeHintsToMetadata(std::vector<Hint> HintTypes) {
1172 if (HintTypes.size() == 0)
1175 // Reserve the first element to LoopID (see below).
1176 SmallVector<Value*, 4> Vals(1);
1177 // If the loop already has metadata, then ignore the existing operands.
1178 MDNode *LoopID = TheLoop->getLoopID();
1180 for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
1181 MDNode *Node = cast<MDNode>(LoopID->getOperand(i));
1182 // If node in update list, ignore old value.
1183 if (!matchesHintMetadataName(Node, HintTypes))
1184 Vals.push_back(Node);
1188 // Now, add the missing hints.
1189 for (auto H : HintTypes)
1191 createHintMetadata(Twine(Prefix(), H.Name).str(), H.Value));
1193 // Replace current metadata node with new one.
1194 LLVMContext &Context = TheLoop->getHeader()->getContext();
1195 MDNode *NewLoopID = MDNode::get(Context, Vals);
1196 // Set operand 0 to refer to the loop id itself.
1197 NewLoopID->replaceOperandWith(0, NewLoopID);
1199 TheLoop->setLoopID(NewLoopID);
1201 LoopID->replaceAllUsesWith(NewLoopID);
1205 /// The loop these hints belong to.
1206 const Loop *TheLoop;
1209 static void emitMissedWarning(Function *F, Loop *L,
1210 const LoopVectorizeHints &LH) {
1211 emitOptimizationRemarkMissed(F->getContext(), DEBUG_TYPE, *F,
1212 L->getStartLoc(), LH.emitRemark());
1214 if (LH.getForce() == LoopVectorizeHints::FK_Enabled) {
1215 if (LH.getWidth() != 1)
1216 emitLoopVectorizeWarning(
1217 F->getContext(), *F, L->getStartLoc(),
1218 "failed explicitly specified loop vectorization");
1219 else if (LH.getInterleave() != 1)
1220 emitLoopInterleaveWarning(
1221 F->getContext(), *F, L->getStartLoc(),
1222 "failed explicitly specified loop interleaving");
1226 static void addInnerLoop(Loop &L, SmallVectorImpl<Loop *> &V) {
1228 return V.push_back(&L);
1230 for (Loop *InnerL : L)
1231 addInnerLoop(*InnerL, V);
1234 /// The LoopVectorize Pass.
1235 struct LoopVectorize : public FunctionPass {
1236 /// Pass identification, replacement for typeid
1239 explicit LoopVectorize(bool NoUnrolling = false, bool AlwaysVectorize = true)
1241 DisableUnrolling(NoUnrolling),
1242 AlwaysVectorize(AlwaysVectorize) {
1243 initializeLoopVectorizePass(*PassRegistry::getPassRegistry());
1246 ScalarEvolution *SE;
1247 const DataLayout *DL;
1249 TargetTransformInfo *TTI;
1251 BlockFrequencyInfo *BFI;
1252 TargetLibraryInfo *TLI;
1254 bool DisableUnrolling;
1255 bool AlwaysVectorize;
1257 BlockFrequency ColdEntryFreq;
1259 bool runOnFunction(Function &F) override {
1260 SE = &getAnalysis<ScalarEvolution>();
1261 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1262 DL = DLP ? &DLP->getDataLayout() : nullptr;
1263 LI = &getAnalysis<LoopInfo>();
1264 TTI = &getAnalysis<TargetTransformInfo>();
1265 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1266 BFI = &getAnalysis<BlockFrequencyInfo>();
1267 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
1268 AA = &getAnalysis<AliasAnalysis>();
1270 // Compute some weights outside of the loop over the loops. Compute this
1271 // using a BranchProbability to re-use its scaling math.
1272 const BranchProbability ColdProb(1, 5); // 20%
1273 ColdEntryFreq = BlockFrequency(BFI->getEntryFreq()) * ColdProb;
1275 // If the target claims to have no vector registers don't attempt
1277 if (!TTI->getNumberOfRegisters(true))
1281 DEBUG(dbgs() << "\nLV: Not vectorizing " << F.getName()
1282 << ": Missing data layout\n");
1286 // Build up a worklist of inner-loops to vectorize. This is necessary as
1287 // the act of vectorizing or partially unrolling a loop creates new loops
1288 // and can invalidate iterators across the loops.
1289 SmallVector<Loop *, 8> Worklist;
1292 addInnerLoop(*L, Worklist);
1294 LoopsAnalyzed += Worklist.size();
1296 // Now walk the identified inner loops.
1297 bool Changed = false;
1298 while (!Worklist.empty())
1299 Changed |= processLoop(Worklist.pop_back_val());
1301 // Process each loop nest in the function.
1305 bool processLoop(Loop *L) {
1306 assert(L->empty() && "Only process inner loops.");
1309 const std::string DebugLocStr = getDebugLocString(L);
1312 DEBUG(dbgs() << "\nLV: Checking a loop in \""
1313 << L->getHeader()->getParent()->getName() << "\" from "
1314 << DebugLocStr << "\n");
1316 LoopVectorizeHints Hints(L, DisableUnrolling);
1318 DEBUG(dbgs() << "LV: Loop hints:"
1320 << (Hints.getForce() == LoopVectorizeHints::FK_Disabled
1322 : (Hints.getForce() == LoopVectorizeHints::FK_Enabled
1324 : "?")) << " width=" << Hints.getWidth()
1325 << " unroll=" << Hints.getInterleave() << "\n");
1327 // Function containing loop
1328 Function *F = L->getHeader()->getParent();
1330 // Looking at the diagnostic output is the only way to determine if a loop
1331 // was vectorized (other than looking at the IR or machine code), so it
1332 // is important to generate an optimization remark for each loop. Most of
1333 // these messages are generated by emitOptimizationRemarkAnalysis. Remarks
1334 // generated by emitOptimizationRemark and emitOptimizationRemarkMissed are
1335 // less verbose reporting vectorized loops and unvectorized loops that may
1336 // benefit from vectorization, respectively.
1338 if (Hints.getForce() == LoopVectorizeHints::FK_Disabled) {
1339 DEBUG(dbgs() << "LV: Not vectorizing: #pragma vectorize disable.\n");
1340 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1341 L->getStartLoc(), Hints.emitRemark());
1345 if (!AlwaysVectorize && Hints.getForce() != LoopVectorizeHints::FK_Enabled) {
1346 DEBUG(dbgs() << "LV: Not vectorizing: No #pragma vectorize enable.\n");
1347 emitOptimizationRemarkAnalysis(F->getContext(), DEBUG_TYPE, *F,
1348 L->getStartLoc(), Hints.emitRemark());
1352 if (Hints.getWidth() == 1 && Hints.getInterleave() == 1) {
1353 DEBUG(dbgs() << "LV: Not vectorizing: Disabled/already vectorized.\n");
1354 emitOptimizationRemarkAnalysis(
1355 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1356 "loop not vectorized: vector width and interleave count are "
1357 "explicitly set to 1");
1361 // Check the loop for a trip count threshold:
1362 // do not vectorize loops with a tiny trip count.
1363 const unsigned TC = SE->getSmallConstantTripCount(L);
1364 if (TC > 0u && TC < TinyTripCountVectorThreshold) {
1365 DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "
1366 << "This loop is not worth vectorizing.");
1367 if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)
1368 DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");
1370 DEBUG(dbgs() << "\n");
1371 emitOptimizationRemarkAnalysis(
1372 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1373 "vectorization is not beneficial and is not explicitly forced");
1378 // Check if it is legal to vectorize the loop.
1379 LoopVectorizationLegality LVL(L, SE, DL, DT, TLI, AA, F);
1380 if (!LVL.canVectorize()) {
1381 DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");
1382 emitMissedWarning(F, L, Hints);
1386 // Use the cost model.
1387 LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI, F, &Hints);
1389 // Check the function attributes to find out if this function should be
1390 // optimized for size.
1391 bool OptForSize = Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1392 F->hasFnAttribute(Attribute::OptimizeForSize);
1394 // Compute the weighted frequency of this loop being executed and see if it
1395 // is less than 20% of the function entry baseline frequency. Note that we
1396 // always have a canonical loop here because we think we *can* vectoriez.
1397 // FIXME: This is hidden behind a flag due to pervasive problems with
1398 // exactly what block frequency models.
1399 if (LoopVectorizeWithBlockFrequency) {
1400 BlockFrequency LoopEntryFreq = BFI->getBlockFreq(L->getLoopPreheader());
1401 if (Hints.getForce() != LoopVectorizeHints::FK_Enabled &&
1402 LoopEntryFreq < ColdEntryFreq)
1406 // Check the function attributes to see if implicit floats are allowed.a
1407 // FIXME: This check doesn't seem possibly correct -- what if the loop is
1408 // an integer loop and the vector instructions selected are purely integer
1409 // vector instructions?
1410 if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1411 DEBUG(dbgs() << "LV: Can't vectorize when the NoImplicitFloat"
1412 "attribute is used.\n");
1413 emitOptimizationRemarkAnalysis(
1414 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1415 "loop not vectorized due to NoImplicitFloat attribute");
1416 emitMissedWarning(F, L, Hints);
1420 // Select the optimal vectorization factor.
1421 const LoopVectorizationCostModel::VectorizationFactor VF =
1422 CM.selectVectorizationFactor(OptForSize);
1424 // Select the unroll factor.
1426 CM.selectUnrollFactor(OptForSize, VF.Width, VF.Cost);
1428 DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in "
1429 << DebugLocStr << '\n');
1430 DEBUG(dbgs() << "LV: Unroll Factor is " << UF << '\n');
1432 if (VF.Width == 1) {
1433 DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial\n");
1436 emitOptimizationRemarkAnalysis(
1437 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1438 "not beneficial to vectorize and user disabled interleaving");
1441 DEBUG(dbgs() << "LV: Trying to at least unroll the loops.\n");
1443 // Report the unrolling decision.
1444 emitOptimizationRemark(F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1445 Twine("unrolled with interleaving factor " +
1447 " (vectorization not beneficial)"));
1449 // We decided not to vectorize, but we may want to unroll.
1451 InnerLoopUnroller Unroller(L, SE, LI, DT, DL, TLI, UF);
1452 Unroller.vectorize(&LVL);
1454 // If we decided that it is *legal* to vectorize the loop then do it.
1455 InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
1459 // Report the vectorization decision.
1460 emitOptimizationRemark(
1461 F->getContext(), DEBUG_TYPE, *F, L->getStartLoc(),
1462 Twine("vectorized loop (vectorization factor: ") + Twine(VF.Width) +
1463 ", unrolling interleave factor: " + Twine(UF) + ")");
1466 // Mark the loop as already vectorized to avoid vectorizing again.
1467 Hints.setAlreadyVectorized();
1469 DEBUG(verifyFunction(*L->getHeader()->getParent()));
1473 void getAnalysisUsage(AnalysisUsage &AU) const override {
1474 AU.addRequiredID(LoopSimplifyID);
1475 AU.addRequiredID(LCSSAID);
1476 AU.addRequired<BlockFrequencyInfo>();
1477 AU.addRequired<DominatorTreeWrapperPass>();
1478 AU.addRequired<LoopInfo>();
1479 AU.addRequired<ScalarEvolution>();
1480 AU.addRequired<TargetTransformInfo>();
1481 AU.addRequired<AliasAnalysis>();
1482 AU.addPreserved<LoopInfo>();
1483 AU.addPreserved<DominatorTreeWrapperPass>();
1484 AU.addPreserved<AliasAnalysis>();
1489 } // end anonymous namespace
1491 //===----------------------------------------------------------------------===//
1492 // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and
1493 // LoopVectorizationCostModel.
1494 //===----------------------------------------------------------------------===//
1496 static Value *stripIntegerCast(Value *V) {
1497 if (CastInst *CI = dyn_cast<CastInst>(V))
1498 if (CI->getOperand(0)->getType()->isIntegerTy())
1499 return CI->getOperand(0);
1503 ///\brief Replaces the symbolic stride in a pointer SCEV expression by one.
1505 /// If \p OrigPtr is not null, use it to look up the stride value instead of
1507 static const SCEV *replaceSymbolicStrideSCEV(ScalarEvolution *SE,
1508 ValueToValueMap &PtrToStride,
1509 Value *Ptr, Value *OrigPtr = nullptr) {
1511 const SCEV *OrigSCEV = SE->getSCEV(Ptr);
1513 // If there is an entry in the map return the SCEV of the pointer with the
1514 // symbolic stride replaced by one.
1515 ValueToValueMap::iterator SI = PtrToStride.find(OrigPtr ? OrigPtr : Ptr);
1516 if (SI != PtrToStride.end()) {
1517 Value *StrideVal = SI->second;
1520 StrideVal = stripIntegerCast(StrideVal);
1522 // Replace symbolic stride by one.
1523 Value *One = ConstantInt::get(StrideVal->getType(), 1);
1524 ValueToValueMap RewriteMap;
1525 RewriteMap[StrideVal] = One;
1528 SCEVParameterRewriter::rewrite(OrigSCEV, *SE, RewriteMap, true);
1529 DEBUG(dbgs() << "LV: Replacing SCEV: " << *OrigSCEV << " by: " << *ByOne
1534 // Otherwise, just return the SCEV of the original pointer.
1535 return SE->getSCEV(Ptr);
1538 void LoopVectorizationLegality::RuntimePointerCheck::insert(
1539 ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr, unsigned DepSetId,
1540 unsigned ASId, ValueToValueMap &Strides) {
1541 // Get the stride replaced scev.
1542 const SCEV *Sc = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
1543 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
1544 assert(AR && "Invalid addrec expression");
1545 const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
1546 const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
1547 Pointers.push_back(Ptr);
1548 Starts.push_back(AR->getStart());
1549 Ends.push_back(ScEnd);
1550 IsWritePtr.push_back(WritePtr);
1551 DependencySetId.push_back(DepSetId);
1552 AliasSetId.push_back(ASId);
1555 Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
1556 // We need to place the broadcast of invariant variables outside the loop.
1557 Instruction *Instr = dyn_cast<Instruction>(V);
1559 (Instr && std::find(LoopVectorBody.begin(), LoopVectorBody.end(),
1560 Instr->getParent()) != LoopVectorBody.end());
1561 bool Invariant = OrigLoop->isLoopInvariant(V) && !NewInstr;
1563 // Place the code for broadcasting invariant variables in the new preheader.
1564 IRBuilder<>::InsertPointGuard Guard(Builder);
1566 Builder.SetInsertPoint(LoopVectorPreHeader->getTerminator());
1568 // Broadcast the scalar into all locations in the vector.
1569 Value *Shuf = Builder.CreateVectorSplat(VF, V, "broadcast");
1574 Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
1576 assert(Val->getType()->isVectorTy() && "Must be a vector");
1577 assert(Val->getType()->getScalarType()->isIntegerTy() &&
1578 "Elem must be an integer");
1579 // Create the types.
1580 Type *ITy = Val->getType()->getScalarType();
1581 VectorType *Ty = cast<VectorType>(Val->getType());
1582 int VLen = Ty->getNumElements();
1583 SmallVector<Constant*, 8> Indices;
1585 // Create a vector of consecutive numbers from zero to VF.
1586 for (int i = 0; i < VLen; ++i) {
1587 int64_t Idx = Negate ? (-i) : i;
1588 Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
1591 // Add the consecutive indices to the vector value.
1592 Constant *Cv = ConstantVector::get(Indices);
1593 assert(Cv->getType() == Val->getType() && "Invalid consecutive vec");
1594 return Builder.CreateAdd(Val, Cv, "induction");
1597 /// \brief Find the operand of the GEP that should be checked for consecutive
1598 /// stores. This ignores trailing indices that have no effect on the final
1600 static unsigned getGEPInductionOperand(const DataLayout *DL,
1601 const GetElementPtrInst *Gep) {
1602 unsigned LastOperand = Gep->getNumOperands() - 1;
1603 unsigned GEPAllocSize = DL->getTypeAllocSize(
1604 cast<PointerType>(Gep->getType()->getScalarType())->getElementType());
1606 // Walk backwards and try to peel off zeros.
1607 while (LastOperand > 1 && match(Gep->getOperand(LastOperand), m_Zero())) {
1608 // Find the type we're currently indexing into.
1609 gep_type_iterator GEPTI = gep_type_begin(Gep);
1610 std::advance(GEPTI, LastOperand - 1);
1612 // If it's a type with the same allocation size as the result of the GEP we
1613 // can peel off the zero index.
1614 if (DL->getTypeAllocSize(*GEPTI) != GEPAllocSize)
1622 int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
1623 assert(Ptr->getType()->isPointerTy() && "Unexpected non-ptr");
1624 // Make sure that the pointer does not point to structs.
1625 if (Ptr->getType()->getPointerElementType()->isAggregateType())
1628 // If this value is a pointer induction variable we know it is consecutive.
1629 PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
1630 if (Phi && Inductions.count(Phi)) {
1631 InductionInfo II = Inductions[Phi];
1632 if (IK_PtrInduction == II.IK)
1634 else if (IK_ReversePtrInduction == II.IK)
1638 GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
1642 unsigned NumOperands = Gep->getNumOperands();
1643 Value *GpPtr = Gep->getPointerOperand();
1644 // If this GEP value is a consecutive pointer induction variable and all of
1645 // the indices are constant then we know it is consecutive. We can
1646 Phi = dyn_cast<PHINode>(GpPtr);
1647 if (Phi && Inductions.count(Phi)) {
1649 // Make sure that the pointer does not point to structs.
1650 PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
1651 if (GepPtrType->getElementType()->isAggregateType())
1654 // Make sure that all of the index operands are loop invariant.
1655 for (unsigned i = 1; i < NumOperands; ++i)
1656 if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1659 InductionInfo II = Inductions[Phi];
1660 if (IK_PtrInduction == II.IK)
1662 else if (IK_ReversePtrInduction == II.IK)
1666 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1668 // Check that all of the gep indices are uniform except for our induction
1670 for (unsigned i = 0; i != NumOperands; ++i)
1671 if (i != InductionOperand &&
1672 !SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
1675 // We can emit wide load/stores only if the last non-zero index is the
1676 // induction variable.
1677 const SCEV *Last = nullptr;
1678 if (!Strides.count(Gep))
1679 Last = SE->getSCEV(Gep->getOperand(InductionOperand));
1681 // Because of the multiplication by a stride we can have a s/zext cast.
1682 // We are going to replace this stride by 1 so the cast is safe to ignore.
1684 // %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
1685 // %0 = trunc i64 %indvars.iv to i32
1686 // %mul = mul i32 %0, %Stride1
1687 // %idxprom = zext i32 %mul to i64 << Safe cast.
1688 // %arrayidx = getelementptr inbounds i32* %B, i64 %idxprom
1690 Last = replaceSymbolicStrideSCEV(SE, Strides,
1691 Gep->getOperand(InductionOperand), Gep);
1692 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(Last))
1694 (C->getSCEVType() == scSignExtend || C->getSCEVType() == scZeroExtend)
1698 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Last)) {
1699 const SCEV *Step = AR->getStepRecurrence(*SE);
1701 // The memory is consecutive because the last index is consecutive
1702 // and all other indices are loop invariant.
1705 if (Step->isAllOnesValue())
1712 bool LoopVectorizationLegality::isUniform(Value *V) {
1713 return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
1716 InnerLoopVectorizer::VectorParts&
1717 InnerLoopVectorizer::getVectorValue(Value *V) {
1718 assert(V != Induction && "The new induction variable should not be used.");
1719 assert(!V->getType()->isVectorTy() && "Can't widen a vector");
1721 // If we have a stride that is replaced by one, do it here.
1722 if (Legal->hasStride(V))
1723 V = ConstantInt::get(V->getType(), 1);
1725 // If we have this scalar in the map, return it.
1726 if (WidenMap.has(V))
1727 return WidenMap.get(V);
1729 // If this scalar is unknown, assume that it is a constant or that it is
1730 // loop invariant. Broadcast V and save the value for future uses.
1731 Value *B = getBroadcastInstrs(V);
1732 return WidenMap.splat(V, B);
1735 Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
1736 assert(Vec->getType()->isVectorTy() && "Invalid type");
1737 SmallVector<Constant*, 8> ShuffleMask;
1738 for (unsigned i = 0; i < VF; ++i)
1739 ShuffleMask.push_back(Builder.getInt32(VF - i - 1));
1741 return Builder.CreateShuffleVector(Vec, UndefValue::get(Vec->getType()),
1742 ConstantVector::get(ShuffleMask),
1746 void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
1747 // Attempt to issue a wide load.
1748 LoadInst *LI = dyn_cast<LoadInst>(Instr);
1749 StoreInst *SI = dyn_cast<StoreInst>(Instr);
1751 assert((LI || SI) && "Invalid Load/Store instruction");
1753 Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
1754 Type *DataTy = VectorType::get(ScalarDataTy, VF);
1755 Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
1756 unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
1757 // An alignment of 0 means target abi alignment. We need to use the scalar's
1758 // target abi alignment in such a case.
1760 Alignment = DL->getABITypeAlignment(ScalarDataTy);
1761 unsigned AddressSpace = Ptr->getType()->getPointerAddressSpace();
1762 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
1763 unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
1765 if (SI && Legal->blockNeedsPredication(SI->getParent()))
1766 return scalarizeInstruction(Instr, true);
1768 if (ScalarAllocatedSize != VectorElementSize)
1769 return scalarizeInstruction(Instr);
1771 // If the pointer is loop invariant or if it is non-consecutive,
1772 // scalarize the load.
1773 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
1774 bool Reverse = ConsecutiveStride < 0;
1775 bool UniformLoad = LI && Legal->isUniform(Ptr);
1776 if (!ConsecutiveStride || UniformLoad)
1777 return scalarizeInstruction(Instr);
1779 Constant *Zero = Builder.getInt32(0);
1780 VectorParts &Entry = WidenMap.get(Instr);
1782 // Handle consecutive loads/stores.
1783 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
1784 if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
1785 setDebugLocFromInst(Builder, Gep);
1786 Value *PtrOperand = Gep->getPointerOperand();
1787 Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
1788 FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
1790 // Create the new GEP with the new induction variable.
1791 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1792 Gep2->setOperand(0, FirstBasePtr);
1793 Gep2->setName("gep.indvar.base");
1794 Ptr = Builder.Insert(Gep2);
1796 setDebugLocFromInst(Builder, Gep);
1797 assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
1798 OrigLoop) && "Base ptr must be invariant");
1800 // The last index does not have to be the induction. It can be
1801 // consecutive and be a function of the index. For example A[I+1];
1802 unsigned NumOperands = Gep->getNumOperands();
1803 unsigned InductionOperand = getGEPInductionOperand(DL, Gep);
1804 // Create the new GEP with the new induction variable.
1805 GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
1807 for (unsigned i = 0; i < NumOperands; ++i) {
1808 Value *GepOperand = Gep->getOperand(i);
1809 Instruction *GepOperandInst = dyn_cast<Instruction>(GepOperand);
1811 // Update last index or loop invariant instruction anchored in loop.
1812 if (i == InductionOperand ||
1813 (GepOperandInst && OrigLoop->contains(GepOperandInst))) {
1814 assert((i == InductionOperand ||
1815 SE->isLoopInvariant(SE->getSCEV(GepOperandInst), OrigLoop)) &&
1816 "Must be last index or loop invariant");
1818 VectorParts &GEPParts = getVectorValue(GepOperand);
1819 Value *Index = GEPParts[0];
1820 Index = Builder.CreateExtractElement(Index, Zero);
1821 Gep2->setOperand(i, Index);
1822 Gep2->setName("gep.indvar.idx");
1825 Ptr = Builder.Insert(Gep2);
1827 // Use the induction element ptr.
1828 assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
1829 setDebugLocFromInst(Builder, Ptr);
1830 VectorParts &PtrVal = getVectorValue(Ptr);
1831 Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
1836 assert(!Legal->isUniform(SI->getPointerOperand()) &&
1837 "We do not allow storing to uniform addresses");
1838 setDebugLocFromInst(Builder, SI);
1839 // We don't want to update the value in the map as it might be used in
1840 // another expression. So don't use a reference type for "StoredVal".
1841 VectorParts StoredVal = getVectorValue(SI->getValueOperand());
1843 for (unsigned Part = 0; Part < UF; ++Part) {
1844 // Calculate the pointer for the specific unroll-part.
1845 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1848 // If we store to reverse consecutive memory locations then we need
1849 // to reverse the order of elements in the stored value.
1850 StoredVal[Part] = reverseVector(StoredVal[Part]);
1851 // If the address is consecutive but reversed, then the
1852 // wide store needs to start at the last vector element.
1853 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1854 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1857 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1858 DataTy->getPointerTo(AddressSpace));
1860 Builder.CreateAlignedStore(StoredVal[Part], VecPtr, Alignment);
1861 propagateMetadata(NewSI, SI);
1867 assert(LI && "Must have a load instruction");
1868 setDebugLocFromInst(Builder, LI);
1869 for (unsigned Part = 0; Part < UF; ++Part) {
1870 // Calculate the pointer for the specific unroll-part.
1871 Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
1874 // If the address is consecutive but reversed, then the
1875 // wide store needs to start at the last vector element.
1876 PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
1877 PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
1880 Value *VecPtr = Builder.CreateBitCast(PartPtr,
1881 DataTy->getPointerTo(AddressSpace));
1882 LoadInst *NewLI = Builder.CreateAlignedLoad(VecPtr, Alignment, "wide.load");
1883 propagateMetadata(NewLI, LI);
1884 Entry[Part] = Reverse ? reverseVector(NewLI) : NewLI;
1888 void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr, bool IfPredicateStore) {
1889 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
1890 // Holds vector parameters or scalars, in case of uniform vals.
1891 SmallVector<VectorParts, 4> Params;
1893 setDebugLocFromInst(Builder, Instr);
1895 // Find all of the vectorized parameters.
1896 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1897 Value *SrcOp = Instr->getOperand(op);
1899 // If we are accessing the old induction variable, use the new one.
1900 if (SrcOp == OldInduction) {
1901 Params.push_back(getVectorValue(SrcOp));
1905 // Try using previously calculated values.
1906 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
1908 // If the src is an instruction that appeared earlier in the basic block
1909 // then it should already be vectorized.
1910 if (SrcInst && OrigLoop->contains(SrcInst)) {
1911 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
1912 // The parameter is a vector value from earlier.
1913 Params.push_back(WidenMap.get(SrcInst));
1915 // The parameter is a scalar from outside the loop. Maybe even a constant.
1916 VectorParts Scalars;
1917 Scalars.append(UF, SrcOp);
1918 Params.push_back(Scalars);
1922 assert(Params.size() == Instr->getNumOperands() &&
1923 "Invalid number of operands");
1925 // Does this instruction return a value ?
1926 bool IsVoidRetTy = Instr->getType()->isVoidTy();
1928 Value *UndefVec = IsVoidRetTy ? nullptr :
1929 UndefValue::get(VectorType::get(Instr->getType(), VF));
1930 // Create a new entry in the WidenMap and initialize it to Undef or Null.
1931 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
1933 Instruction *InsertPt = Builder.GetInsertPoint();
1934 BasicBlock *IfBlock = Builder.GetInsertBlock();
1935 BasicBlock *CondBlock = nullptr;
1938 Loop *VectorLp = nullptr;
1939 if (IfPredicateStore) {
1940 assert(Instr->getParent()->getSinglePredecessor() &&
1941 "Only support single predecessor blocks");
1942 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
1943 Instr->getParent());
1944 VectorLp = LI->getLoopFor(IfBlock);
1945 assert(VectorLp && "Must have a loop for this block");
1948 // For each vector unroll 'part':
1949 for (unsigned Part = 0; Part < UF; ++Part) {
1950 // For each scalar that we create:
1951 for (unsigned Width = 0; Width < VF; ++Width) {
1954 Value *Cmp = nullptr;
1955 if (IfPredicateStore) {
1956 Cmp = Builder.CreateExtractElement(Cond[Part], Builder.getInt32(Width));
1957 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cmp, ConstantInt::get(Cmp->getType(), 1));
1958 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
1959 LoopVectorBody.push_back(CondBlock);
1960 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
1961 // Update Builder with newly created basic block.
1962 Builder.SetInsertPoint(InsertPt);
1965 Instruction *Cloned = Instr->clone();
1967 Cloned->setName(Instr->getName() + ".cloned");
1968 // Replace the operands of the cloned instructions with extracted scalars.
1969 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
1970 Value *Op = Params[op][Part];
1971 // Param is a vector. Need to extract the right lane.
1972 if (Op->getType()->isVectorTy())
1973 Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
1974 Cloned->setOperand(op, Op);
1977 // Place the cloned scalar in the new loop.
1978 Builder.Insert(Cloned);
1980 // If the original scalar returns a value we need to place it in a vector
1981 // so that future users will be able to use it.
1983 VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
1984 Builder.getInt32(Width));
1986 if (IfPredicateStore) {
1987 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
1988 LoopVectorBody.push_back(NewIfBlock);
1989 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
1990 Builder.SetInsertPoint(InsertPt);
1991 Instruction *OldBr = IfBlock->getTerminator();
1992 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
1993 OldBr->eraseFromParent();
1994 IfBlock = NewIfBlock;
2000 static Instruction *getFirstInst(Instruction *FirstInst, Value *V,
2004 if (Instruction *I = dyn_cast<Instruction>(V))
2005 return I->getParent() == Loc->getParent() ? I : nullptr;
2009 std::pair<Instruction *, Instruction *>
2010 InnerLoopVectorizer::addStrideCheck(Instruction *Loc) {
2011 Instruction *tnullptr = nullptr;
2012 if (!Legal->mustCheckStrides())
2013 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
2015 IRBuilder<> ChkBuilder(Loc);
2018 Value *Check = nullptr;
2019 Instruction *FirstInst = nullptr;
2020 for (SmallPtrSet<Value *, 8>::iterator SI = Legal->strides_begin(),
2021 SE = Legal->strides_end();
2023 Value *Ptr = stripIntegerCast(*SI);
2024 Value *C = ChkBuilder.CreateICmpNE(Ptr, ConstantInt::get(Ptr->getType(), 1),
2026 // Store the first instruction we create.
2027 FirstInst = getFirstInst(FirstInst, C, Loc);
2029 Check = ChkBuilder.CreateOr(Check, C);
2034 // We have to do this trickery because the IRBuilder might fold the check to a
2035 // constant expression in which case there is no Instruction anchored in a
2037 LLVMContext &Ctx = Loc->getContext();
2038 Instruction *TheCheck =
2039 BinaryOperator::CreateAnd(Check, ConstantInt::getTrue(Ctx));
2040 ChkBuilder.Insert(TheCheck, "stride.not.one");
2041 FirstInst = getFirstInst(FirstInst, TheCheck, Loc);
2043 return std::make_pair(FirstInst, TheCheck);
2046 std::pair<Instruction *, Instruction *>
2047 InnerLoopVectorizer::addRuntimeCheck(Instruction *Loc) {
2048 LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
2049 Legal->getRuntimePointerCheck();
2051 Instruction *tnullptr = nullptr;
2052 if (!PtrRtCheck->Need)
2053 return std::pair<Instruction *, Instruction *>(tnullptr, tnullptr);
2055 unsigned NumPointers = PtrRtCheck->Pointers.size();
2056 SmallVector<TrackingVH<Value> , 2> Starts;
2057 SmallVector<TrackingVH<Value> , 2> Ends;
2059 LLVMContext &Ctx = Loc->getContext();
2060 SCEVExpander Exp(*SE, "induction");
2061 Instruction *FirstInst = nullptr;
2063 for (unsigned i = 0; i < NumPointers; ++i) {
2064 Value *Ptr = PtrRtCheck->Pointers[i];
2065 const SCEV *Sc = SE->getSCEV(Ptr);
2067 if (SE->isLoopInvariant(Sc, OrigLoop)) {
2068 DEBUG(dbgs() << "LV: Adding RT check for a loop invariant ptr:" <<
2070 Starts.push_back(Ptr);
2071 Ends.push_back(Ptr);
2073 DEBUG(dbgs() << "LV: Adding RT check for range:" << *Ptr << '\n');
2074 unsigned AS = Ptr->getType()->getPointerAddressSpace();
2076 // Use this type for pointer arithmetic.
2077 Type *PtrArithTy = Type::getInt8PtrTy(Ctx, AS);
2079 Value *Start = Exp.expandCodeFor(PtrRtCheck->Starts[i], PtrArithTy, Loc);
2080 Value *End = Exp.expandCodeFor(PtrRtCheck->Ends[i], PtrArithTy, Loc);
2081 Starts.push_back(Start);
2082 Ends.push_back(End);
2086 IRBuilder<> ChkBuilder(Loc);
2087 // Our instructions might fold to a constant.
2088 Value *MemoryRuntimeCheck = nullptr;
2089 for (unsigned i = 0; i < NumPointers; ++i) {
2090 for (unsigned j = i+1; j < NumPointers; ++j) {
2091 // No need to check if two readonly pointers intersect.
2092 if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
2095 // Only need to check pointers between two different dependency sets.
2096 if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
2098 // Only need to check pointers in the same alias set.
2099 if (PtrRtCheck->AliasSetId[i] != PtrRtCheck->AliasSetId[j])
2102 unsigned AS0 = Starts[i]->getType()->getPointerAddressSpace();
2103 unsigned AS1 = Starts[j]->getType()->getPointerAddressSpace();
2105 assert((AS0 == Ends[j]->getType()->getPointerAddressSpace()) &&
2106 (AS1 == Ends[i]->getType()->getPointerAddressSpace()) &&
2107 "Trying to bounds check pointers with different address spaces");
2109 Type *PtrArithTy0 = Type::getInt8PtrTy(Ctx, AS0);
2110 Type *PtrArithTy1 = Type::getInt8PtrTy(Ctx, AS1);
2112 Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy0, "bc");
2113 Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy1, "bc");
2114 Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy1, "bc");
2115 Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy0, "bc");
2117 Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
2118 FirstInst = getFirstInst(FirstInst, Cmp0, Loc);
2119 Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
2120 FirstInst = getFirstInst(FirstInst, Cmp1, Loc);
2121 Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
2122 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
2123 if (MemoryRuntimeCheck) {
2124 IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
2126 FirstInst = getFirstInst(FirstInst, IsConflict, Loc);
2128 MemoryRuntimeCheck = IsConflict;
2132 // We have to do this trickery because the IRBuilder might fold the check to a
2133 // constant expression in which case there is no Instruction anchored in a
2135 Instruction *Check = BinaryOperator::CreateAnd(MemoryRuntimeCheck,
2136 ConstantInt::getTrue(Ctx));
2137 ChkBuilder.Insert(Check, "memcheck.conflict");
2138 FirstInst = getFirstInst(FirstInst, Check, Loc);
2139 return std::make_pair(FirstInst, Check);
2142 void InnerLoopVectorizer::createEmptyLoop() {
2144 In this function we generate a new loop. The new loop will contain
2145 the vectorized instructions while the old loop will continue to run the
2148 [ ] <-- Back-edge taken count overflow check.
2151 | [ ] <-- vector loop bypass (may consist of multiple blocks).
2154 || [ ] <-- vector pre header.
2158 || [ ]_| <-- vector loop.
2161 | >[ ] <--- middle-block.
2164 -|- >[ ] <--- new preheader.
2168 | [ ]_| <-- old scalar loop to handle remainder.
2171 >[ ] <-- exit block.
2175 BasicBlock *OldBasicBlock = OrigLoop->getHeader();
2176 BasicBlock *BypassBlock = OrigLoop->getLoopPreheader();
2177 BasicBlock *ExitBlock = OrigLoop->getExitBlock();
2178 assert(BypassBlock && "Invalid loop structure");
2179 assert(ExitBlock && "Must have an exit block");
2181 // Some loops have a single integer induction variable, while other loops
2182 // don't. One example is c++ iterators that often have multiple pointer
2183 // induction variables. In the code below we also support a case where we
2184 // don't have a single induction variable.
2185 OldInduction = Legal->getInduction();
2186 Type *IdxTy = Legal->getWidestInductionType();
2188 // Find the loop boundaries.
2189 const SCEV *ExitCount = SE->getBackedgeTakenCount(OrigLoop);
2190 assert(ExitCount != SE->getCouldNotCompute() && "Invalid loop count");
2192 // The exit count might have the type of i64 while the phi is i32. This can
2193 // happen if we have an induction variable that is sign extended before the
2194 // compare. The only way that we get a backedge taken count is that the
2195 // induction variable was signed and as such will not overflow. In such a case
2196 // truncation is legal.
2197 if (ExitCount->getType()->getPrimitiveSizeInBits() >
2198 IdxTy->getPrimitiveSizeInBits())
2199 ExitCount = SE->getTruncateOrNoop(ExitCount, IdxTy);
2201 const SCEV *BackedgeTakeCount = SE->getNoopOrZeroExtend(ExitCount, IdxTy);
2202 // Get the total trip count from the count by adding 1.
2203 ExitCount = SE->getAddExpr(BackedgeTakeCount,
2204 SE->getConstant(BackedgeTakeCount->getType(), 1));
2206 // Expand the trip count and place the new instructions in the preheader.
2207 // Notice that the pre-header does not change, only the loop body.
2208 SCEVExpander Exp(*SE, "induction");
2210 // We need to test whether the backedge-taken count is uint##_max. Adding one
2211 // to it will cause overflow and an incorrect loop trip count in the vector
2212 // body. In case of overflow we want to directly jump to the scalar remainder
2214 Value *BackedgeCount =
2215 Exp.expandCodeFor(BackedgeTakeCount, BackedgeTakeCount->getType(),
2216 BypassBlock->getTerminator());
2217 if (BackedgeCount->getType()->isPointerTy())
2218 BackedgeCount = CastInst::CreatePointerCast(BackedgeCount, IdxTy,
2219 "backedge.ptrcnt.to.int",
2220 BypassBlock->getTerminator());
2221 Instruction *CheckBCOverflow =
2222 CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, BackedgeCount,
2223 Constant::getAllOnesValue(BackedgeCount->getType()),
2224 "backedge.overflow", BypassBlock->getTerminator());
2226 // The loop index does not have to start at Zero. Find the original start
2227 // value from the induction PHI node. If we don't have an induction variable
2228 // then we know that it starts at zero.
2229 Builder.SetInsertPoint(BypassBlock->getTerminator());
2230 Value *StartIdx = ExtendedIdx = OldInduction ?
2231 Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
2233 ConstantInt::get(IdxTy, 0);
2235 // We need an instruction to anchor the overflow check on. StartIdx needs to
2236 // be defined before the overflow check branch. Because the scalar preheader
2237 // is going to merge the start index and so the overflow branch block needs to
2238 // contain a definition of the start index.
2239 Instruction *OverflowCheckAnchor = BinaryOperator::CreateAdd(
2240 StartIdx, ConstantInt::get(IdxTy, 0), "overflow.check.anchor",
2241 BypassBlock->getTerminator());
2243 // Count holds the overall loop count (N).
2244 Value *Count = Exp.expandCodeFor(ExitCount, ExitCount->getType(),
2245 BypassBlock->getTerminator());
2247 LoopBypassBlocks.push_back(BypassBlock);
2249 // Split the single block loop into the two loop structure described above.
2250 BasicBlock *VectorPH =
2251 BypassBlock->splitBasicBlock(BypassBlock->getTerminator(), "vector.ph");
2252 BasicBlock *VecBody =
2253 VectorPH->splitBasicBlock(VectorPH->getTerminator(), "vector.body");
2254 BasicBlock *MiddleBlock =
2255 VecBody->splitBasicBlock(VecBody->getTerminator(), "middle.block");
2256 BasicBlock *ScalarPH =
2257 MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
2259 // Create and register the new vector loop.
2260 Loop* Lp = new Loop();
2261 Loop *ParentLoop = OrigLoop->getParentLoop();
2263 // Insert the new loop into the loop nest and register the new basic blocks
2264 // before calling any utilities such as SCEV that require valid LoopInfo.
2266 ParentLoop->addChildLoop(Lp);
2267 ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
2268 ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
2269 ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
2271 LI->addTopLevelLoop(Lp);
2273 Lp->addBasicBlockToLoop(VecBody, LI->getBase());
2275 // Use this IR builder to create the loop instructions (Phi, Br, Cmp)
2277 Builder.SetInsertPoint(VecBody->getFirstNonPHI());
2279 // Generate the induction variable.
2280 setDebugLocFromInst(Builder, getDebugLocFromInstOrOperands(OldInduction));
2281 Induction = Builder.CreatePHI(IdxTy, 2, "index");
2282 // The loop step is equal to the vectorization factor (num of SIMD elements)
2283 // times the unroll factor (num of SIMD instructions).
2284 Constant *Step = ConstantInt::get(IdxTy, VF * UF);
2286 // This is the IR builder that we use to add all of the logic for bypassing
2287 // the new vector loop.
2288 IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
2289 setDebugLocFromInst(BypassBuilder,
2290 getDebugLocFromInstOrOperands(OldInduction));
2292 // We may need to extend the index in case there is a type mismatch.
2293 // We know that the count starts at zero and does not overflow.
2294 if (Count->getType() != IdxTy) {
2295 // The exit count can be of pointer type. Convert it to the correct
2297 if (ExitCount->getType()->isPointerTy())
2298 Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
2300 Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
2303 // Add the start index to the loop count to get the new end index.
2304 Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
2306 // Now we need to generate the expression for N - (N % VF), which is
2307 // the part that the vectorized body will execute.
2308 Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
2309 Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
2310 Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
2311 "end.idx.rnd.down");
2313 // Now, compare the new count to zero. If it is zero skip the vector loop and
2314 // jump to the scalar loop.
2316 BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx, "cmp.zero");
2318 BasicBlock *LastBypassBlock = BypassBlock;
2320 // Generate code to check that the loops trip count that we computed by adding
2321 // one to the backedge-taken count will not overflow.
2323 auto PastOverflowCheck =
2324 std::next(BasicBlock::iterator(OverflowCheckAnchor));
2325 BasicBlock *CheckBlock =
2326 LastBypassBlock->splitBasicBlock(PastOverflowCheck, "overflow.checked");
2328 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2329 LoopBypassBlocks.push_back(CheckBlock);
2330 Instruction *OldTerm = LastBypassBlock->getTerminator();
2331 BranchInst::Create(ScalarPH, CheckBlock, CheckBCOverflow, OldTerm);
2332 OldTerm->eraseFromParent();
2333 LastBypassBlock = CheckBlock;
2336 // Generate the code to check that the strides we assumed to be one are really
2337 // one. We want the new basic block to start at the first instruction in a
2338 // sequence of instructions that form a check.
2339 Instruction *StrideCheck;
2340 Instruction *FirstCheckInst;
2341 std::tie(FirstCheckInst, StrideCheck) =
2342 addStrideCheck(LastBypassBlock->getTerminator());
2344 // Create a new block containing the stride check.
2345 BasicBlock *CheckBlock =
2346 LastBypassBlock->splitBasicBlock(FirstCheckInst, "vector.stridecheck");
2348 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2349 LoopBypassBlocks.push_back(CheckBlock);
2351 // Replace the branch into the memory check block with a conditional branch
2352 // for the "few elements case".
2353 Instruction *OldTerm = LastBypassBlock->getTerminator();
2354 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2355 OldTerm->eraseFromParent();
2358 LastBypassBlock = CheckBlock;
2361 // Generate the code that checks in runtime if arrays overlap. We put the
2362 // checks into a separate block to make the more common case of few elements
2364 Instruction *MemRuntimeCheck;
2365 std::tie(FirstCheckInst, MemRuntimeCheck) =
2366 addRuntimeCheck(LastBypassBlock->getTerminator());
2367 if (MemRuntimeCheck) {
2368 // Create a new block containing the memory check.
2369 BasicBlock *CheckBlock =
2370 LastBypassBlock->splitBasicBlock(MemRuntimeCheck, "vector.memcheck");
2372 ParentLoop->addBasicBlockToLoop(CheckBlock, LI->getBase());
2373 LoopBypassBlocks.push_back(CheckBlock);
2375 // Replace the branch into the memory check block with a conditional branch
2376 // for the "few elements case".
2377 Instruction *OldTerm = LastBypassBlock->getTerminator();
2378 BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
2379 OldTerm->eraseFromParent();
2381 Cmp = MemRuntimeCheck;
2382 LastBypassBlock = CheckBlock;
2385 LastBypassBlock->getTerminator()->eraseFromParent();
2386 BranchInst::Create(MiddleBlock, VectorPH, Cmp,
2389 // We are going to resume the execution of the scalar loop.
2390 // Go over all of the induction variables that we found and fix the
2391 // PHIs that are left in the scalar version of the loop.
2392 // The starting values of PHI nodes depend on the counter of the last
2393 // iteration in the vectorized loop.
2394 // If we come from a bypass edge then we need to start from the original
2397 // This variable saves the new starting index for the scalar loop.
2398 PHINode *ResumeIndex = nullptr;
2399 LoopVectorizationLegality::InductionList::iterator I, E;
2400 LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
2401 // Set builder to point to last bypass block.
2402 BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
2403 for (I = List->begin(), E = List->end(); I != E; ++I) {
2404 PHINode *OrigPhi = I->first;
2405 LoopVectorizationLegality::InductionInfo II = I->second;
2407 Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
2408 PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
2409 MiddleBlock->getTerminator());
2410 // We might have extended the type of the induction variable but we need a
2411 // truncated version for the scalar loop.
2412 PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
2413 PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
2414 MiddleBlock->getTerminator()) : nullptr;
2416 // Create phi nodes to merge from the backedge-taken check block.
2417 PHINode *BCResumeVal = PHINode::Create(ResumeValTy, 3, "bc.resume.val",
2418 ScalarPH->getTerminator());
2419 BCResumeVal->addIncoming(ResumeVal, MiddleBlock);
2421 PHINode *BCTruncResumeVal = nullptr;
2422 if (OrigPhi == OldInduction) {
2424 PHINode::Create(OrigPhi->getType(), 2, "bc.trunc.resume.val",
2425 ScalarPH->getTerminator());
2426 BCTruncResumeVal->addIncoming(TruncResumeVal, MiddleBlock);
2429 Value *EndValue = nullptr;
2431 case LoopVectorizationLegality::IK_NoInduction:
2432 llvm_unreachable("Unknown induction");
2433 case LoopVectorizationLegality::IK_IntInduction: {
2434 // Handle the integer induction counter.
2435 assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
2437 // We have the canonical induction variable.
2438 if (OrigPhi == OldInduction) {
2439 // Create a truncated version of the resume value for the scalar loop,
2440 // we might have promoted the type to a larger width.
2442 BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
2443 // The new PHI merges the original incoming value, in case of a bypass,
2444 // or the value at the end of the vectorized loop.
2445 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2446 TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2447 TruncResumeVal->addIncoming(EndValue, VecBody);
2449 BCTruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2451 // We know what the end value is.
2452 EndValue = IdxEndRoundDown;
2453 // We also know which PHI node holds it.
2454 ResumeIndex = ResumeVal;
2458 // Not the canonical induction variable - add the vector loop count to the
2460 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2461 II.StartValue->getType(),
2463 EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end");
2466 case LoopVectorizationLegality::IK_ReverseIntInduction: {
2467 // Convert the CountRoundDown variable to the PHI size.
2468 Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
2469 II.StartValue->getType(),
2471 // Handle reverse integer induction counter.
2472 EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
2475 case LoopVectorizationLegality::IK_PtrInduction: {
2476 // For pointer induction variables, calculate the offset using
2478 EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
2482 case LoopVectorizationLegality::IK_ReversePtrInduction: {
2483 // The value at the end of the loop for the reverse pointer is calculated
2484 // by creating a GEP with a negative index starting from the start value.
2485 Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
2486 Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
2488 EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
2494 // The new PHI merges the original incoming value, in case of a bypass,
2495 // or the value at the end of the vectorized loop.
2496 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I) {
2497 if (OrigPhi == OldInduction)
2498 ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
2500 ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
2502 ResumeVal->addIncoming(EndValue, VecBody);
2504 // Fix the scalar body counter (PHI node).
2505 unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
2507 // The old induction's phi node in the scalar body needs the truncated
2509 if (OrigPhi == OldInduction) {
2510 BCResumeVal->addIncoming(StartIdx, LoopBypassBlocks[0]);
2511 OrigPhi->setIncomingValue(BlockIdx, BCTruncResumeVal);
2513 BCResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[0]);
2514 OrigPhi->setIncomingValue(BlockIdx, BCResumeVal);
2518 // If we are generating a new induction variable then we also need to
2519 // generate the code that calculates the exit value. This value is not
2520 // simply the end of the counter because we may skip the vectorized body
2521 // in case of a runtime check.
2523 assert(!ResumeIndex && "Unexpected resume value found");
2524 ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
2525 MiddleBlock->getTerminator());
2526 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2527 ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
2528 ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
2531 // Make sure that we found the index where scalar loop needs to continue.
2532 assert(ResumeIndex && ResumeIndex->getType()->isIntegerTy() &&
2533 "Invalid resume Index");
2535 // Add a check in the middle block to see if we have completed
2536 // all of the iterations in the first vector loop.
2537 // If (N - N%VF) == N, then we *don't* need to run the remainder.
2538 Value *CmpN = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ, IdxEnd,
2539 ResumeIndex, "cmp.n",
2540 MiddleBlock->getTerminator());
2542 BranchInst::Create(ExitBlock, ScalarPH, CmpN, MiddleBlock->getTerminator());
2543 // Remove the old terminator.
2544 MiddleBlock->getTerminator()->eraseFromParent();
2546 // Create i+1 and fill the PHINode.
2547 Value *NextIdx = Builder.CreateAdd(Induction, Step, "index.next");
2548 Induction->addIncoming(StartIdx, VectorPH);
2549 Induction->addIncoming(NextIdx, VecBody);
2550 // Create the compare.
2551 Value *ICmp = Builder.CreateICmpEQ(NextIdx, IdxEndRoundDown);
2552 Builder.CreateCondBr(ICmp, MiddleBlock, VecBody);
2554 // Now we have two terminators. Remove the old one from the block.
2555 VecBody->getTerminator()->eraseFromParent();
2557 // Get ready to start creating new instructions into the vectorized body.
2558 Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
2561 LoopVectorPreHeader = VectorPH;
2562 LoopScalarPreHeader = ScalarPH;
2563 LoopMiddleBlock = MiddleBlock;
2564 LoopExitBlock = ExitBlock;
2565 LoopVectorBody.push_back(VecBody);
2566 LoopScalarBody = OldBasicBlock;
2568 LoopVectorizeHints Hints(Lp, true);
2569 Hints.setAlreadyVectorized();
2572 /// This function returns the identity element (or neutral element) for
2573 /// the operation K.
2575 LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
2580 // Adding, Xoring, Oring zero to a number does not change it.
2581 return ConstantInt::get(Tp, 0);
2582 case RK_IntegerMult:
2583 // Multiplying a number by 1 does not change it.
2584 return ConstantInt::get(Tp, 1);
2586 // AND-ing a number with an all-1 value does not change it.
2587 return ConstantInt::get(Tp, -1, true);
2589 // Multiplying a number by 1 does not change it.
2590 return ConstantFP::get(Tp, 1.0L);
2592 // Adding zero to a number does not change it.
2593 return ConstantFP::get(Tp, 0.0L);
2595 llvm_unreachable("Unknown reduction kind");
2599 /// This function translates the reduction kind to an LLVM binary operator.
2601 getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
2603 case LoopVectorizationLegality::RK_IntegerAdd:
2604 return Instruction::Add;
2605 case LoopVectorizationLegality::RK_IntegerMult:
2606 return Instruction::Mul;
2607 case LoopVectorizationLegality::RK_IntegerOr:
2608 return Instruction::Or;
2609 case LoopVectorizationLegality::RK_IntegerAnd:
2610 return Instruction::And;
2611 case LoopVectorizationLegality::RK_IntegerXor:
2612 return Instruction::Xor;
2613 case LoopVectorizationLegality::RK_FloatMult:
2614 return Instruction::FMul;
2615 case LoopVectorizationLegality::RK_FloatAdd:
2616 return Instruction::FAdd;
2617 case LoopVectorizationLegality::RK_IntegerMinMax:
2618 return Instruction::ICmp;
2619 case LoopVectorizationLegality::RK_FloatMinMax:
2620 return Instruction::FCmp;
2622 llvm_unreachable("Unknown reduction operation");
2626 Value *createMinMaxOp(IRBuilder<> &Builder,
2627 LoopVectorizationLegality::MinMaxReductionKind RK,
2630 CmpInst::Predicate P = CmpInst::ICMP_NE;
2633 llvm_unreachable("Unknown min/max reduction kind");
2634 case LoopVectorizationLegality::MRK_UIntMin:
2635 P = CmpInst::ICMP_ULT;
2637 case LoopVectorizationLegality::MRK_UIntMax:
2638 P = CmpInst::ICMP_UGT;
2640 case LoopVectorizationLegality::MRK_SIntMin:
2641 P = CmpInst::ICMP_SLT;
2643 case LoopVectorizationLegality::MRK_SIntMax:
2644 P = CmpInst::ICMP_SGT;
2646 case LoopVectorizationLegality::MRK_FloatMin:
2647 P = CmpInst::FCMP_OLT;
2649 case LoopVectorizationLegality::MRK_FloatMax:
2650 P = CmpInst::FCMP_OGT;
2655 if (RK == LoopVectorizationLegality::MRK_FloatMin ||
2656 RK == LoopVectorizationLegality::MRK_FloatMax)
2657 Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
2659 Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
2661 Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
2666 struct CSEDenseMapInfo {
2667 static bool canHandle(Instruction *I) {
2668 return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||
2669 isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);
2671 static inline Instruction *getEmptyKey() {
2672 return DenseMapInfo<Instruction *>::getEmptyKey();
2674 static inline Instruction *getTombstoneKey() {
2675 return DenseMapInfo<Instruction *>::getTombstoneKey();
2677 static unsigned getHashValue(Instruction *I) {
2678 assert(canHandle(I) && "Unknown instruction!");
2679 return hash_combine(I->getOpcode(), hash_combine_range(I->value_op_begin(),
2680 I->value_op_end()));
2682 static bool isEqual(Instruction *LHS, Instruction *RHS) {
2683 if (LHS == getEmptyKey() || RHS == getEmptyKey() ||
2684 LHS == getTombstoneKey() || RHS == getTombstoneKey())
2686 return LHS->isIdenticalTo(RHS);
2691 /// \brief Check whether this block is a predicated block.
2692 /// Due to if predication of stores we might create a sequence of "if(pred) a[i]
2693 /// = ...; " blocks. We start with one vectorized basic block. For every
2694 /// conditional block we split this vectorized block. Therefore, every second
2695 /// block will be a predicated one.
2696 static bool isPredicatedBlock(unsigned BlockNum) {
2697 return BlockNum % 2;
2700 ///\brief Perform cse of induction variable instructions.
2701 static void cse(SmallVector<BasicBlock *, 4> &BBs) {
2702 // Perform simple cse.
2703 SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;
2704 for (unsigned i = 0, e = BBs.size(); i != e; ++i) {
2705 BasicBlock *BB = BBs[i];
2706 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
2707 Instruction *In = I++;
2709 if (!CSEDenseMapInfo::canHandle(In))
2712 // Check if we can replace this instruction with any of the
2713 // visited instructions.
2714 if (Instruction *V = CSEMap.lookup(In)) {
2715 In->replaceAllUsesWith(V);
2716 In->eraseFromParent();
2719 // Ignore instructions in conditional blocks. We create "if (pred) a[i] =
2720 // ...;" blocks for predicated stores. Every second block is a predicated
2722 if (isPredicatedBlock(i))
2730 /// \brief Adds a 'fast' flag to floating point operations.
2731 static Value *addFastMathFlag(Value *V) {
2732 if (isa<FPMathOperator>(V)){
2733 FastMathFlags Flags;
2734 Flags.setUnsafeAlgebra();
2735 cast<Instruction>(V)->setFastMathFlags(Flags);
2740 void InnerLoopVectorizer::vectorizeLoop() {
2741 //===------------------------------------------------===//
2743 // Notice: any optimization or new instruction that go
2744 // into the code below should be also be implemented in
2747 //===------------------------------------------------===//
2748 Constant *Zero = Builder.getInt32(0);
2750 // In order to support reduction variables we need to be able to vectorize
2751 // Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
2752 // stages. First, we create a new vector PHI node with no incoming edges.
2753 // We use this value when we vectorize all of the instructions that use the
2754 // PHI. Next, after all of the instructions in the block are complete we
2755 // add the new incoming edges to the PHI. At this point all of the
2756 // instructions in the basic block are vectorized, so we can use them to
2757 // construct the PHI.
2758 PhiVector RdxPHIsToFix;
2760 // Scan the loop in a topological order to ensure that defs are vectorized
2762 LoopBlocksDFS DFS(OrigLoop);
2765 // Vectorize all of the blocks in the original loop.
2766 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
2767 be = DFS.endRPO(); bb != be; ++bb)
2768 vectorizeBlockInLoop(*bb, &RdxPHIsToFix);
2770 // At this point every instruction in the original loop is widened to
2771 // a vector form. We are almost done. Now, we need to fix the PHI nodes
2772 // that we vectorized. The PHI nodes are currently empty because we did
2773 // not want to introduce cycles. Notice that the remaining PHI nodes
2774 // that we need to fix are reduction variables.
2776 // Create the 'reduced' values for each of the induction vars.
2777 // The reduced values are the vector values that we scalarize and combine
2778 // after the loop is finished.
2779 for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
2781 PHINode *RdxPhi = *it;
2782 assert(RdxPhi && "Unable to recover vectorized PHI");
2784 // Find the reduction variable descriptor.
2785 assert(Legal->getReductionVars()->count(RdxPhi) &&
2786 "Unable to find the reduction variable");
2787 LoopVectorizationLegality::ReductionDescriptor RdxDesc =
2788 (*Legal->getReductionVars())[RdxPhi];
2790 setDebugLocFromInst(Builder, RdxDesc.StartValue);
2792 // We need to generate a reduction vector from the incoming scalar.
2793 // To do so, we need to generate the 'identity' vector and override
2794 // one of the elements with the incoming scalar reduction. We need
2795 // to do it in the vector-loop preheader.
2796 Builder.SetInsertPoint(LoopBypassBlocks[1]->getTerminator());
2798 // This is the vector-clone of the value that leaves the loop.
2799 VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
2800 Type *VecTy = VectorExit[0]->getType();
2802 // Find the reduction identity variable. Zero for addition, or, xor,
2803 // one for multiplication, -1 for And.
2806 if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
2807 RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
2808 // MinMax reduction have the start value as their identify.
2810 VectorStart = Identity = RdxDesc.StartValue;
2812 VectorStart = Identity = Builder.CreateVectorSplat(VF,
2817 // Handle other reduction kinds:
2819 LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
2820 VecTy->getScalarType());
2823 // This vector is the Identity vector where the first element is the
2824 // incoming scalar reduction.
2825 VectorStart = RdxDesc.StartValue;
2827 Identity = ConstantVector::getSplat(VF, Iden);
2829 // This vector is the Identity vector where the first element is the
2830 // incoming scalar reduction.
2831 VectorStart = Builder.CreateInsertElement(Identity,
2832 RdxDesc.StartValue, Zero);
2836 // Fix the vector-loop phi.
2837 // We created the induction variable so we know that the
2838 // preheader is the first entry.
2839 BasicBlock *VecPreheader = Induction->getIncomingBlock(0);
2841 // Reductions do not have to start at zero. They can start with
2842 // any loop invariant values.
2843 VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
2844 BasicBlock *Latch = OrigLoop->getLoopLatch();
2845 Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
2846 VectorParts &Val = getVectorValue(LoopVal);
2847 for (unsigned part = 0; part < UF; ++part) {
2848 // Make sure to add the reduction stat value only to the
2849 // first unroll part.
2850 Value *StartVal = (part == 0) ? VectorStart : Identity;
2851 cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
2852 cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part],
2853 LoopVectorBody.back());
2856 // Before each round, move the insertion point right between
2857 // the PHIs and the values we are going to write.
2858 // This allows us to write both PHINodes and the extractelement
2860 Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
2862 VectorParts RdxParts;
2863 setDebugLocFromInst(Builder, RdxDesc.LoopExitInstr);
2864 for (unsigned part = 0; part < UF; ++part) {
2865 // This PHINode contains the vectorized reduction variable, or
2866 // the initial value vector, if we bypass the vector loop.
2867 VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
2868 PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
2869 Value *StartVal = (part == 0) ? VectorStart : Identity;
2870 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
2871 NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
2872 NewPhi->addIncoming(RdxExitVal[part],
2873 LoopVectorBody.back());
2874 RdxParts.push_back(NewPhi);
2877 // Reduce all of the unrolled parts into a single vector.
2878 Value *ReducedPartRdx = RdxParts[0];
2879 unsigned Op = getReductionBinOp(RdxDesc.Kind);
2880 setDebugLocFromInst(Builder, ReducedPartRdx);
2881 for (unsigned part = 1; part < UF; ++part) {
2882 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2883 // Floating point operations had to be 'fast' to enable the reduction.
2884 ReducedPartRdx = addFastMathFlag(
2885 Builder.CreateBinOp((Instruction::BinaryOps)Op, RdxParts[part],
2886 ReducedPartRdx, "bin.rdx"));
2888 ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
2889 ReducedPartRdx, RdxParts[part]);
2893 // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
2894 // and vector ops, reducing the set of values being computed by half each
2896 assert(isPowerOf2_32(VF) &&
2897 "Reduction emission only supported for pow2 vectors!");
2898 Value *TmpVec = ReducedPartRdx;
2899 SmallVector<Constant*, 32> ShuffleMask(VF, nullptr);
2900 for (unsigned i = VF; i != 1; i >>= 1) {
2901 // Move the upper half of the vector to the lower half.
2902 for (unsigned j = 0; j != i/2; ++j)
2903 ShuffleMask[j] = Builder.getInt32(i/2 + j);
2905 // Fill the rest of the mask with undef.
2906 std::fill(&ShuffleMask[i/2], ShuffleMask.end(),
2907 UndefValue::get(Builder.getInt32Ty()));
2910 Builder.CreateShuffleVector(TmpVec,
2911 UndefValue::get(TmpVec->getType()),
2912 ConstantVector::get(ShuffleMask),
2915 if (Op != Instruction::ICmp && Op != Instruction::FCmp)
2916 // Floating point operations had to be 'fast' to enable the reduction.
2917 TmpVec = addFastMathFlag(Builder.CreateBinOp(
2918 (Instruction::BinaryOps)Op, TmpVec, Shuf, "bin.rdx"));
2920 TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
2923 // The result is in the first element of the vector.
2924 ReducedPartRdx = Builder.CreateExtractElement(TmpVec,
2925 Builder.getInt32(0));
2928 // Create a phi node that merges control-flow from the backedge-taken check
2929 // block and the middle block.
2930 PHINode *BCBlockPhi = PHINode::Create(RdxPhi->getType(), 2, "bc.merge.rdx",
2931 LoopScalarPreHeader->getTerminator());
2932 BCBlockPhi->addIncoming(RdxDesc.StartValue, LoopBypassBlocks[0]);
2933 BCBlockPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2935 // Now, we need to fix the users of the reduction variable
2936 // inside and outside of the scalar remainder loop.
2937 // We know that the loop is in LCSSA form. We need to update the
2938 // PHI nodes in the exit blocks.
2939 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2940 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2941 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2942 if (!LCSSAPhi) break;
2944 // All PHINodes need to have a single entry edge, or two if
2945 // we already fixed them.
2946 assert(LCSSAPhi->getNumIncomingValues() < 3 && "Invalid LCSSA PHI");
2948 // We found our reduction value exit-PHI. Update it with the
2949 // incoming bypass edge.
2950 if (LCSSAPhi->getIncomingValue(0) == RdxDesc.LoopExitInstr) {
2951 // Add an edge coming from the bypass.
2952 LCSSAPhi->addIncoming(ReducedPartRdx, LoopMiddleBlock);
2955 }// end of the LCSSA phi scan.
2957 // Fix the scalar loop reduction variable with the incoming reduction sum
2958 // from the vector body and from the backedge value.
2959 int IncomingEdgeBlockIdx =
2960 (RdxPhi)->getBasicBlockIndex(OrigLoop->getLoopLatch());
2961 assert(IncomingEdgeBlockIdx >= 0 && "Invalid block index");
2962 // Pick the other block.
2963 int SelfEdgeBlockIdx = (IncomingEdgeBlockIdx ? 0 : 1);
2964 (RdxPhi)->setIncomingValue(SelfEdgeBlockIdx, BCBlockPhi);
2965 (RdxPhi)->setIncomingValue(IncomingEdgeBlockIdx, RdxDesc.LoopExitInstr);
2966 }// end of for each redux variable.
2970 // Remove redundant induction instructions.
2971 cse(LoopVectorBody);
2974 void InnerLoopVectorizer::fixLCSSAPHIs() {
2975 for (BasicBlock::iterator LEI = LoopExitBlock->begin(),
2976 LEE = LoopExitBlock->end(); LEI != LEE; ++LEI) {
2977 PHINode *LCSSAPhi = dyn_cast<PHINode>(LEI);
2978 if (!LCSSAPhi) break;
2979 if (LCSSAPhi->getNumIncomingValues() == 1)
2980 LCSSAPhi->addIncoming(UndefValue::get(LCSSAPhi->getType()),
2985 InnerLoopVectorizer::VectorParts
2986 InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
2987 assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
2990 // Look for cached value.
2991 std::pair<BasicBlock*, BasicBlock*> Edge(Src, Dst);
2992 EdgeMaskCache::iterator ECEntryIt = MaskCache.find(Edge);
2993 if (ECEntryIt != MaskCache.end())
2994 return ECEntryIt->second;
2996 VectorParts SrcMask = createBlockInMask(Src);
2998 // The terminator has to be a branch inst!
2999 BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
3000 assert(BI && "Unexpected terminator found");
3002 if (BI->isConditional()) {
3003 VectorParts EdgeMask = getVectorValue(BI->getCondition());
3005 if (BI->getSuccessor(0) != Dst)
3006 for (unsigned part = 0; part < UF; ++part)
3007 EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
3009 for (unsigned part = 0; part < UF; ++part)
3010 EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
3012 MaskCache[Edge] = EdgeMask;
3016 MaskCache[Edge] = SrcMask;
3020 InnerLoopVectorizer::VectorParts
3021 InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
3022 assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
3024 // Loop incoming mask is all-one.
3025 if (OrigLoop->getHeader() == BB) {
3026 Value *C = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 1);
3027 return getVectorValue(C);
3030 // This is the block mask. We OR all incoming edges, and with zero.
3031 Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
3032 VectorParts BlockMask = getVectorValue(Zero);
3035 for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
3036 VectorParts EM = createEdgeMask(*it, BB);
3037 for (unsigned part = 0; part < UF; ++part)
3038 BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
3044 void InnerLoopVectorizer::widenPHIInstruction(Instruction *PN,
3045 InnerLoopVectorizer::VectorParts &Entry,
3046 unsigned UF, unsigned VF, PhiVector *PV) {
3047 PHINode* P = cast<PHINode>(PN);
3048 // Handle reduction variables:
3049 if (Legal->getReductionVars()->count(P)) {
3050 for (unsigned part = 0; part < UF; ++part) {
3051 // This is phase one of vectorizing PHIs.
3052 Type *VecTy = (VF == 1) ? PN->getType() :
3053 VectorType::get(PN->getType(), VF);
3054 Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
3055 LoopVectorBody.back()-> getFirstInsertionPt());
3061 setDebugLocFromInst(Builder, P);
3062 // Check for PHI nodes that are lowered to vector selects.
3063 if (P->getParent() != OrigLoop->getHeader()) {
3064 // We know that all PHIs in non-header blocks are converted into
3065 // selects, so we don't have to worry about the insertion order and we
3066 // can just use the builder.
3067 // At this point we generate the predication tree. There may be
3068 // duplications since this is a simple recursive scan, but future
3069 // optimizations will clean it up.
3071 unsigned NumIncoming = P->getNumIncomingValues();
3073 // Generate a sequence of selects of the form:
3074 // SELECT(Mask3, In3,
3075 // SELECT(Mask2, In2,
3077 for (unsigned In = 0; In < NumIncoming; In++) {
3078 VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
3080 VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
3082 for (unsigned part = 0; part < UF; ++part) {
3083 // We might have single edge PHIs (blocks) - use an identity
3084 // 'select' for the first PHI operand.
3086 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3089 // Select between the current value and the previous incoming edge
3090 // based on the incoming mask.
3091 Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
3092 Entry[part], "predphi");
3098 // This PHINode must be an induction variable.
3099 // Make sure that we know about it.
3100 assert(Legal->getInductionVars()->count(P) &&
3101 "Not an induction variable");
3103 LoopVectorizationLegality::InductionInfo II =
3104 Legal->getInductionVars()->lookup(P);
3107 case LoopVectorizationLegality::IK_NoInduction:
3108 llvm_unreachable("Unknown induction");
3109 case LoopVectorizationLegality::IK_IntInduction: {
3110 assert(P->getType() == II.StartValue->getType() && "Types must match");
3111 Type *PhiTy = P->getType();
3113 if (P == OldInduction) {
3114 // Handle the canonical induction variable. We might have had to
3116 Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
3118 // Handle other induction variables that are now based on the
3120 Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
3122 NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
3123 Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx,
3126 Broadcasted = getBroadcastInstrs(Broadcasted);
3127 // After broadcasting the induction variable we need to make the vector
3128 // consecutive by adding 0, 1, 2, etc.
3129 for (unsigned part = 0; part < UF; ++part)
3130 Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
3133 case LoopVectorizationLegality::IK_ReverseIntInduction:
3134 case LoopVectorizationLegality::IK_PtrInduction:
3135 case LoopVectorizationLegality::IK_ReversePtrInduction:
3136 // Handle reverse integer and pointer inductions.
3137 Value *StartIdx = ExtendedIdx;
3138 // This is the normalized GEP that starts counting at zero.
3139 Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
3142 // Handle the reverse integer induction variable case.
3143 if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
3144 IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
3145 Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
3147 Value *ReverseInd = Builder.CreateSub(II.StartValue, CNI,
3150 // This is a new value so do not hoist it out.
3151 Value *Broadcasted = getBroadcastInstrs(ReverseInd);
3152 // After broadcasting the induction variable we need to make the
3153 // vector consecutive by adding ... -3, -2, -1, 0.
3154 for (unsigned part = 0; part < UF; ++part)
3155 Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
3160 // Handle the pointer induction variable case.
3161 assert(P->getType()->isPointerTy() && "Unexpected type.");
3163 // Is this a reverse induction ptr or a consecutive induction ptr.
3164 bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
3167 // This is the vector of results. Notice that we don't generate
3168 // vector geps because scalar geps result in better code.
3169 for (unsigned part = 0; part < UF; ++part) {
3171 int EltIndex = (part) * (Reverse ? -1 : 1);
3172 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3175 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
3177 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
3179 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
3181 Entry[part] = SclrGep;
3185 Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
3186 for (unsigned int i = 0; i < VF; ++i) {
3187 int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
3188 Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
3191 GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
3193 GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
3195 Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
3197 VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
3198 Builder.getInt32(i),
3201 Entry[part] = VecVal;
3207 void InnerLoopVectorizer::vectorizeBlockInLoop(BasicBlock *BB, PhiVector *PV) {
3208 // For each instruction in the old loop.
3209 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
3210 VectorParts &Entry = WidenMap.get(it);
3211 switch (it->getOpcode()) {
3212 case Instruction::Br:
3213 // Nothing to do for PHIs and BR, since we already took care of the
3214 // loop control flow instructions.
3216 case Instruction::PHI:{
3217 // Vectorize PHINodes.
3218 widenPHIInstruction(it, Entry, UF, VF, PV);
3222 case Instruction::Add:
3223 case Instruction::FAdd:
3224 case Instruction::Sub:
3225 case Instruction::FSub:
3226 case Instruction::Mul:
3227 case Instruction::FMul:
3228 case Instruction::UDiv:
3229 case Instruction::SDiv:
3230 case Instruction::FDiv:
3231 case Instruction::URem:
3232 case Instruction::SRem:
3233 case Instruction::FRem:
3234 case Instruction::Shl:
3235 case Instruction::LShr:
3236 case Instruction::AShr:
3237 case Instruction::And:
3238 case Instruction::Or:
3239 case Instruction::Xor: {
3240 // Just widen binops.
3241 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
3242 setDebugLocFromInst(Builder, BinOp);
3243 VectorParts &A = getVectorValue(it->getOperand(0));
3244 VectorParts &B = getVectorValue(it->getOperand(1));
3246 // Use this vector value for all users of the original instruction.
3247 for (unsigned Part = 0; Part < UF; ++Part) {
3248 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
3250 if (BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V))
3251 VecOp->copyIRFlags(BinOp);
3256 propagateMetadata(Entry, it);
3259 case Instruction::Select: {
3261 // If the selector is loop invariant we can create a select
3262 // instruction with a scalar condition. Otherwise, use vector-select.
3263 bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
3265 setDebugLocFromInst(Builder, it);
3267 // The condition can be loop invariant but still defined inside the
3268 // loop. This means that we can't just use the original 'cond' value.
3269 // We have to take the 'vectorized' value and pick the first lane.
3270 // Instcombine will make this a no-op.
3271 VectorParts &Cond = getVectorValue(it->getOperand(0));
3272 VectorParts &Op0 = getVectorValue(it->getOperand(1));
3273 VectorParts &Op1 = getVectorValue(it->getOperand(2));
3275 Value *ScalarCond = (VF == 1) ? Cond[0] :
3276 Builder.CreateExtractElement(Cond[0], Builder.getInt32(0));
3278 for (unsigned Part = 0; Part < UF; ++Part) {
3279 Entry[Part] = Builder.CreateSelect(
3280 InvariantCond ? ScalarCond : Cond[Part],
3285 propagateMetadata(Entry, it);
3289 case Instruction::ICmp:
3290 case Instruction::FCmp: {
3291 // Widen compares. Generate vector compares.
3292 bool FCmp = (it->getOpcode() == Instruction::FCmp);
3293 CmpInst *Cmp = dyn_cast<CmpInst>(it);
3294 setDebugLocFromInst(Builder, it);
3295 VectorParts &A = getVectorValue(it->getOperand(0));
3296 VectorParts &B = getVectorValue(it->getOperand(1));
3297 for (unsigned Part = 0; Part < UF; ++Part) {
3300 C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
3302 C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
3306 propagateMetadata(Entry, it);
3310 case Instruction::Store:
3311 case Instruction::Load:
3312 vectorizeMemoryInstruction(it);
3314 case Instruction::ZExt:
3315 case Instruction::SExt:
3316 case Instruction::FPToUI:
3317 case Instruction::FPToSI:
3318 case Instruction::FPExt:
3319 case Instruction::PtrToInt:
3320 case Instruction::IntToPtr:
3321 case Instruction::SIToFP:
3322 case Instruction::UIToFP:
3323 case Instruction::Trunc:
3324 case Instruction::FPTrunc:
3325 case Instruction::BitCast: {
3326 CastInst *CI = dyn_cast<CastInst>(it);
3327 setDebugLocFromInst(Builder, it);
3328 /// Optimize the special case where the source is the induction
3329 /// variable. Notice that we can only optimize the 'trunc' case
3330 /// because: a. FP conversions lose precision, b. sext/zext may wrap,
3331 /// c. other casts depend on pointer size.
3332 if (CI->getOperand(0) == OldInduction &&
3333 it->getOpcode() == Instruction::Trunc) {
3334 Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
3336 Value *Broadcasted = getBroadcastInstrs(ScalarCast);
3337 for (unsigned Part = 0; Part < UF; ++Part)
3338 Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
3339 propagateMetadata(Entry, it);
3342 /// Vectorize casts.
3343 Type *DestTy = (VF == 1) ? CI->getType() :
3344 VectorType::get(CI->getType(), VF);
3346 VectorParts &A = getVectorValue(it->getOperand(0));
3347 for (unsigned Part = 0; Part < UF; ++Part)
3348 Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
3349 propagateMetadata(Entry, it);
3353 case Instruction::Call: {
3354 // Ignore dbg intrinsics.
3355 if (isa<DbgInfoIntrinsic>(it))
3357 setDebugLocFromInst(Builder, it);
3359 Module *M = BB->getParent()->getParent();
3360 CallInst *CI = cast<CallInst>(it);
3361 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
3362 assert(ID && "Not an intrinsic call!");
3364 case Intrinsic::lifetime_end:
3365 case Intrinsic::lifetime_start:
3366 scalarizeInstruction(it);
3369 bool HasScalarOpd = hasVectorInstrinsicScalarOpd(ID, 1);
3370 for (unsigned Part = 0; Part < UF; ++Part) {
3371 SmallVector<Value *, 4> Args;
3372 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
3373 if (HasScalarOpd && i == 1) {
3374 Args.push_back(CI->getArgOperand(i));
3377 VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
3378 Args.push_back(Arg[Part]);
3380 Type *Tys[] = {CI->getType()};
3382 Tys[0] = VectorType::get(CI->getType()->getScalarType(), VF);
3384 Function *F = Intrinsic::getDeclaration(M, ID, Tys);
3385 Entry[Part] = Builder.CreateCall(F, Args);
3388 propagateMetadata(Entry, it);
3395 // All other instructions are unsupported. Scalarize them.
3396 scalarizeInstruction(it);
3399 }// end of for_each instr.
3402 void InnerLoopVectorizer::updateAnalysis() {
3403 // Forget the original basic block.
3404 SE->forgetLoop(OrigLoop);
3406 // Update the dominator tree information.
3407 assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
3408 "Entry does not dominate exit.");
3410 for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
3411 DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
3412 DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
3414 // Due to if predication of stores we might create a sequence of "if(pred)
3415 // a[i] = ...; " blocks.
3416 for (unsigned i = 0, e = LoopVectorBody.size(); i != e; ++i) {
3418 DT->addNewBlock(LoopVectorBody[0], LoopVectorPreHeader);
3419 else if (isPredicatedBlock(i)) {
3420 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-1]);
3422 DT->addNewBlock(LoopVectorBody[i], LoopVectorBody[i-2]);
3426 DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks[1]);
3427 DT->addNewBlock(LoopScalarPreHeader, LoopBypassBlocks[0]);
3428 DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
3429 DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
3431 DEBUG(DT->verifyDomTree());
3434 /// \brief Check whether it is safe to if-convert this phi node.
3436 /// Phi nodes with constant expressions that can trap are not safe to if
3438 static bool canIfConvertPHINodes(BasicBlock *BB) {
3439 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3440 PHINode *Phi = dyn_cast<PHINode>(I);
3443 for (unsigned p = 0, e = Phi->getNumIncomingValues(); p != e; ++p)
3444 if (Constant *C = dyn_cast<Constant>(Phi->getIncomingValue(p)))
3451 bool LoopVectorizationLegality::canVectorizeWithIfConvert() {
3452 if (!EnableIfConversion) {
3453 emitAnalysis(Report() << "if-conversion is disabled");
3457 assert(TheLoop->getNumBlocks() > 1 && "Single block loops are vectorizable");
3459 // A list of pointers that we can safely read and write to.
3460 SmallPtrSet<Value *, 8> SafePointes;
3462 // Collect safe addresses.
3463 for (Loop::block_iterator BI = TheLoop->block_begin(),
3464 BE = TheLoop->block_end(); BI != BE; ++BI) {
3465 BasicBlock *BB = *BI;
3467 if (blockNeedsPredication(BB))
3470 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
3471 if (LoadInst *LI = dyn_cast<LoadInst>(I))
3472 SafePointes.insert(LI->getPointerOperand());
3473 else if (StoreInst *SI = dyn_cast<StoreInst>(I))
3474 SafePointes.insert(SI->getPointerOperand());
3478 // Collect the blocks that need predication.
3479 BasicBlock *Header = TheLoop->getHeader();
3480 for (Loop::block_iterator BI = TheLoop->block_begin(),
3481 BE = TheLoop->block_end(); BI != BE; ++BI) {
3482 BasicBlock *BB = *BI;
3484 // We don't support switch statements inside loops.
3485 if (!isa<BranchInst>(BB->getTerminator())) {
3486 emitAnalysis(Report(BB->getTerminator())
3487 << "loop contains a switch statement");
3491 // We must be able to predicate all blocks that need to be predicated.
3492 if (blockNeedsPredication(BB)) {
3493 if (!blockCanBePredicated(BB, SafePointes)) {
3494 emitAnalysis(Report(BB->getTerminator())
3495 << "control flow cannot be substituted for a select");
3498 } else if (BB != Header && !canIfConvertPHINodes(BB)) {
3499 emitAnalysis(Report(BB->getTerminator())
3500 << "control flow cannot be substituted for a select");
3505 // We can if-convert this loop.
3509 bool LoopVectorizationLegality::canVectorize() {
3510 // We must have a loop in canonical form. Loops with indirectbr in them cannot
3511 // be canonicalized.
3512 if (!TheLoop->getLoopPreheader()) {
3514 Report() << "loop control flow is not understood by vectorizer");
3518 // We can only vectorize innermost loops.
3519 if (TheLoop->getSubLoopsVector().size()) {
3520 emitAnalysis(Report() << "loop is not the innermost loop");
3524 // We must have a single backedge.
3525 if (TheLoop->getNumBackEdges() != 1) {
3527 Report() << "loop control flow is not understood by vectorizer");
3531 // We must have a single exiting block.
3532 if (!TheLoop->getExitingBlock()) {
3534 Report() << "loop control flow is not understood by vectorizer");
3538 // We need to have a loop header.
3539 DEBUG(dbgs() << "LV: Found a loop: " <<
3540 TheLoop->getHeader()->getName() << '\n');
3542 // Check if we can if-convert non-single-bb loops.
3543 unsigned NumBlocks = TheLoop->getNumBlocks();
3544 if (NumBlocks != 1 && !canVectorizeWithIfConvert()) {
3545 DEBUG(dbgs() << "LV: Can't if-convert the loop.\n");
3549 // ScalarEvolution needs to be able to find the exit count.
3550 const SCEV *ExitCount = SE->getBackedgeTakenCount(TheLoop);
3551 if (ExitCount == SE->getCouldNotCompute()) {
3552 emitAnalysis(Report() << "could not determine number of loop iterations");
3553 DEBUG(dbgs() << "LV: SCEV could not compute the loop exit count.\n");
3557 // Check if we can vectorize the instructions and CFG in this loop.
3558 if (!canVectorizeInstrs()) {
3559 DEBUG(dbgs() << "LV: Can't vectorize the instructions or CFG\n");
3563 // Go over each instruction and look at memory deps.
3564 if (!canVectorizeMemory()) {
3565 DEBUG(dbgs() << "LV: Can't vectorize due to memory conflicts\n");
3569 // Collect all of the variables that remain uniform after vectorization.
3570 collectLoopUniforms();
3572 DEBUG(dbgs() << "LV: We can vectorize this loop" <<
3573 (PtrRtCheck.Need ? " (with a runtime bound check)" : "")
3576 // Okay! We can vectorize. At this point we don't have any other mem analysis
3577 // which may limit our maximum vectorization factor, so just return true with
3582 static Type *convertPointerToIntegerType(const DataLayout &DL, Type *Ty) {
3583 if (Ty->isPointerTy())
3584 return DL.getIntPtrType(Ty);
3586 // It is possible that char's or short's overflow when we ask for the loop's
3587 // trip count, work around this by changing the type size.
3588 if (Ty->getScalarSizeInBits() < 32)
3589 return Type::getInt32Ty(Ty->getContext());
3594 static Type* getWiderType(const DataLayout &DL, Type *Ty0, Type *Ty1) {
3595 Ty0 = convertPointerToIntegerType(DL, Ty0);
3596 Ty1 = convertPointerToIntegerType(DL, Ty1);
3597 if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
3602 /// \brief Check that the instruction has outside loop users and is not an
3603 /// identified reduction variable.
3604 static bool hasOutsideLoopUser(const Loop *TheLoop, Instruction *Inst,
3605 SmallPtrSetImpl<Value *> &Reductions) {
3606 // Reduction instructions are allowed to have exit users. All other
3607 // instructions must not have external users.
3608 if (!Reductions.count(Inst))
3609 //Check that all of the users of the loop are inside the BB.
3610 for (User *U : Inst->users()) {
3611 Instruction *UI = cast<Instruction>(U);
3612 // This user may be a reduction exit value.
3613 if (!TheLoop->contains(UI)) {
3614 DEBUG(dbgs() << "LV: Found an outside user for : " << *UI << '\n');
3621 bool LoopVectorizationLegality::canVectorizeInstrs() {
3622 BasicBlock *PreHeader = TheLoop->getLoopPreheader();
3623 BasicBlock *Header = TheLoop->getHeader();
3625 // Look for the attribute signaling the absence of NaNs.
3626 Function &F = *Header->getParent();
3627 if (F.hasFnAttribute("no-nans-fp-math"))
3628 HasFunNoNaNAttr = F.getAttributes().getAttribute(
3629 AttributeSet::FunctionIndex,
3630 "no-nans-fp-math").getValueAsString() == "true";
3632 // For each block in the loop.
3633 for (Loop::block_iterator bb = TheLoop->block_begin(),
3634 be = TheLoop->block_end(); bb != be; ++bb) {
3636 // Scan the instructions in the block and look for hazards.
3637 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
3640 if (PHINode *Phi = dyn_cast<PHINode>(it)) {
3641 Type *PhiTy = Phi->getType();
3642 // Check that this PHI type is allowed.
3643 if (!PhiTy->isIntegerTy() &&
3644 !PhiTy->isFloatingPointTy() &&
3645 !PhiTy->isPointerTy()) {
3646 emitAnalysis(Report(it)
3647 << "loop control flow is not understood by vectorizer");
3648 DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
3652 // If this PHINode is not in the header block, then we know that we
3653 // can convert it to select during if-conversion. No need to check if
3654 // the PHIs in this block are induction or reduction variables.
3655 if (*bb != Header) {
3656 // Check that this instruction has no outside users or is an
3657 // identified reduction value with an outside user.
3658 if (!hasOutsideLoopUser(TheLoop, it, AllowedExit))
3660 emitAnalysis(Report(it) << "value could not be identified as "
3661 "an induction or reduction variable");
3665 // We only allow if-converted PHIs with more than two incoming values.
3666 if (Phi->getNumIncomingValues() != 2) {
3667 emitAnalysis(Report(it)
3668 << "control flow not understood by vectorizer");
3669 DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
3673 // This is the value coming from the preheader.
3674 Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
3675 // Check if this is an induction variable.
3676 InductionKind IK = isInductionVariable(Phi);
3678 if (IK_NoInduction != IK) {
3679 // Get the widest type.
3681 WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
3683 WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
3685 // Int inductions are special because we only allow one IV.
3686 if (IK == IK_IntInduction) {
3687 // Use the phi node with the widest type as induction. Use the last
3688 // one if there are multiple (no good reason for doing this other
3689 // than it is expedient).
3690 if (!Induction || PhiTy == WidestIndTy)
3694 DEBUG(dbgs() << "LV: Found an induction variable.\n");
3695 Inductions[Phi] = InductionInfo(StartValue, IK);
3697 // Until we explicitly handle the case of an induction variable with
3698 // an outside loop user we have to give up vectorizing this loop.
3699 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3700 emitAnalysis(Report(it) << "use of induction value outside of the "
3701 "loop is not handled by vectorizer");
3708 if (AddReductionVar(Phi, RK_IntegerAdd)) {
3709 DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
3712 if (AddReductionVar(Phi, RK_IntegerMult)) {
3713 DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
3716 if (AddReductionVar(Phi, RK_IntegerOr)) {
3717 DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
3720 if (AddReductionVar(Phi, RK_IntegerAnd)) {
3721 DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
3724 if (AddReductionVar(Phi, RK_IntegerXor)) {
3725 DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
3728 if (AddReductionVar(Phi, RK_IntegerMinMax)) {
3729 DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
3732 if (AddReductionVar(Phi, RK_FloatMult)) {
3733 DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
3736 if (AddReductionVar(Phi, RK_FloatAdd)) {
3737 DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
3740 if (AddReductionVar(Phi, RK_FloatMinMax)) {
3741 DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<
3746 emitAnalysis(Report(it) << "value that could not be identified as "
3747 "reduction is used outside the loop");
3748 DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
3750 }// end of PHI handling
3752 // We still don't handle functions. However, we can ignore dbg intrinsic
3753 // calls and we do handle certain intrinsic and libm functions.
3754 CallInst *CI = dyn_cast<CallInst>(it);
3755 if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
3756 emitAnalysis(Report(it) << "call instruction cannot be vectorized");
3757 DEBUG(dbgs() << "LV: Found a call site.\n");
3761 // Intrinsics such as powi,cttz and ctlz are legal to vectorize if the
3762 // second argument is the same (i.e. loop invariant)
3764 hasVectorInstrinsicScalarOpd(getIntrinsicIDForCall(CI, TLI), 1)) {
3765 if (!SE->isLoopInvariant(SE->getSCEV(CI->getOperand(1)), TheLoop)) {
3766 emitAnalysis(Report(it)
3767 << "intrinsic instruction cannot be vectorized");
3768 DEBUG(dbgs() << "LV: Found unvectorizable intrinsic " << *CI << "\n");
3773 // Check that the instruction return type is vectorizable.
3774 // Also, we can't vectorize extractelement instructions.
3775 if ((!VectorType::isValidElementType(it->getType()) &&
3776 !it->getType()->isVoidTy()) || isa<ExtractElementInst>(it)) {
3777 emitAnalysis(Report(it)
3778 << "instruction return type cannot be vectorized");
3779 DEBUG(dbgs() << "LV: Found unvectorizable type.\n");
3783 // Check that the stored type is vectorizable.
3784 if (StoreInst *ST = dyn_cast<StoreInst>(it)) {
3785 Type *T = ST->getValueOperand()->getType();
3786 if (!VectorType::isValidElementType(T)) {
3787 emitAnalysis(Report(ST) << "store instruction cannot be vectorized");
3790 if (EnableMemAccessVersioning)
3791 collectStridedAcccess(ST);
3794 if (EnableMemAccessVersioning)
3795 if (LoadInst *LI = dyn_cast<LoadInst>(it))
3796 collectStridedAcccess(LI);
3798 // Reduction instructions are allowed to have exit users.
3799 // All other instructions must not have external users.
3800 if (hasOutsideLoopUser(TheLoop, it, AllowedExit)) {
3801 emitAnalysis(Report(it) << "value cannot be used outside the loop");
3810 DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
3811 if (Inductions.empty()) {
3812 emitAnalysis(Report()
3813 << "loop induction variable could not be identified");
3821 ///\brief Remove GEPs whose indices but the last one are loop invariant and
3822 /// return the induction operand of the gep pointer.
3823 static Value *stripGetElementPtr(Value *Ptr, ScalarEvolution *SE,
3824 const DataLayout *DL, Loop *Lp) {
3825 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
3829 unsigned InductionOperand = getGEPInductionOperand(DL, GEP);
3831 // Check that all of the gep indices are uniform except for our induction
3833 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i)
3834 if (i != InductionOperand &&
3835 !SE->isLoopInvariant(SE->getSCEV(GEP->getOperand(i)), Lp))
3837 return GEP->getOperand(InductionOperand);
3840 ///\brief Look for a cast use of the passed value.
3841 static Value *getUniqueCastUse(Value *Ptr, Loop *Lp, Type *Ty) {
3842 Value *UniqueCast = nullptr;
3843 for (User *U : Ptr->users()) {
3844 CastInst *CI = dyn_cast<CastInst>(U);
3845 if (CI && CI->getType() == Ty) {
3855 ///\brief Get the stride of a pointer access in a loop.
3856 /// Looks for symbolic strides "a[i*stride]". Returns the symbolic stride as a
3857 /// pointer to the Value, or null otherwise.
3858 static Value *getStrideFromPointer(Value *Ptr, ScalarEvolution *SE,
3859 const DataLayout *DL, Loop *Lp) {
3860 const PointerType *PtrTy = dyn_cast<PointerType>(Ptr->getType());
3861 if (!PtrTy || PtrTy->isAggregateType())
3864 // Try to remove a gep instruction to make the pointer (actually index at this
3865 // point) easier analyzable. If OrigPtr is equal to Ptr we are analzying the
3866 // pointer, otherwise, we are analyzing the index.
3867 Value *OrigPtr = Ptr;
3869 // The size of the pointer access.
3870 int64_t PtrAccessSize = 1;
3872 Ptr = stripGetElementPtr(Ptr, SE, DL, Lp);
3873 const SCEV *V = SE->getSCEV(Ptr);
3877 while (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V))
3878 V = C->getOperand();
3880 const SCEVAddRecExpr *S = dyn_cast<SCEVAddRecExpr>(V);
3884 V = S->getStepRecurrence(*SE);
3888 // Strip off the size of access multiplication if we are still analyzing the
3890 if (OrigPtr == Ptr) {
3891 DL->getTypeAllocSize(PtrTy->getElementType());
3892 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(V)) {
3893 if (M->getOperand(0)->getSCEVType() != scConstant)
3896 const APInt &APStepVal =
3897 cast<SCEVConstant>(M->getOperand(0))->getValue()->getValue();
3899 // Huge step value - give up.
3900 if (APStepVal.getBitWidth() > 64)
3903 int64_t StepVal = APStepVal.getSExtValue();
3904 if (PtrAccessSize != StepVal)
3906 V = M->getOperand(1);
3911 Type *StripedOffRecurrenceCast = nullptr;
3912 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(V)) {
3913 StripedOffRecurrenceCast = C->getType();
3914 V = C->getOperand();
3917 // Look for the loop invariant symbolic value.
3918 const SCEVUnknown *U = dyn_cast<SCEVUnknown>(V);
3922 Value *Stride = U->getValue();
3923 if (!Lp->isLoopInvariant(Stride))
3926 // If we have stripped off the recurrence cast we have to make sure that we
3927 // return the value that is used in this loop so that we can replace it later.
3928 if (StripedOffRecurrenceCast)
3929 Stride = getUniqueCastUse(Stride, Lp, StripedOffRecurrenceCast);
3934 void LoopVectorizationLegality::collectStridedAcccess(Value *MemAccess) {
3935 Value *Ptr = nullptr;
3936 if (LoadInst *LI = dyn_cast<LoadInst>(MemAccess))
3937 Ptr = LI->getPointerOperand();
3938 else if (StoreInst *SI = dyn_cast<StoreInst>(MemAccess))
3939 Ptr = SI->getPointerOperand();
3943 Value *Stride = getStrideFromPointer(Ptr, SE, DL, TheLoop);
3947 DEBUG(dbgs() << "LV: Found a strided access that we can version");
3948 DEBUG(dbgs() << " Ptr: " << *Ptr << " Stride: " << *Stride << "\n");
3949 Strides[Ptr] = Stride;
3950 StrideSet.insert(Stride);
3953 void LoopVectorizationLegality::collectLoopUniforms() {
3954 // We now know that the loop is vectorizable!
3955 // Collect variables that will remain uniform after vectorization.
3956 std::vector<Value*> Worklist;
3957 BasicBlock *Latch = TheLoop->getLoopLatch();
3959 // Start with the conditional branch and walk up the block.
3960 Worklist.push_back(Latch->getTerminator()->getOperand(0));
3962 // Also add all consecutive pointer values; these values will be uniform
3963 // after vectorization (and subsequent cleanup) and, until revectorization is
3964 // supported, all dependencies must also be uniform.
3965 for (Loop::block_iterator B = TheLoop->block_begin(),
3966 BE = TheLoop->block_end(); B != BE; ++B)
3967 for (BasicBlock::iterator I = (*B)->begin(), IE = (*B)->end();
3969 if (I->getType()->isPointerTy() && isConsecutivePtr(I))
3970 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3972 while (Worklist.size()) {
3973 Instruction *I = dyn_cast<Instruction>(Worklist.back());
3974 Worklist.pop_back();
3976 // Look at instructions inside this loop.
3977 // Stop when reaching PHI nodes.
3978 // TODO: we need to follow values all over the loop, not only in this block.
3979 if (!I || !TheLoop->contains(I) || isa<PHINode>(I))
3982 // This is a known uniform.
3985 // Insert all operands.
3986 Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
3991 /// \brief Analyses memory accesses in a loop.
3993 /// Checks whether run time pointer checks are needed and builds sets for data
3994 /// dependence checking.
3995 class AccessAnalysis {
3997 /// \brief Read or write access location.
3998 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
3999 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
4001 /// \brief Set of potential dependent memory accesses.
4002 typedef EquivalenceClasses<MemAccessInfo> DepCandidates;
4004 AccessAnalysis(const DataLayout *Dl, AliasAnalysis *AA, DepCandidates &DA) :
4005 DL(Dl), AST(*AA), DepCands(DA), IsRTCheckNeeded(false) {}
4007 /// \brief Register a load and whether it is only read from.
4008 void addLoad(AliasAnalysis::Location &Loc, bool IsReadOnly) {
4009 Value *Ptr = const_cast<Value*>(Loc.Ptr);
4010 AST.add(Ptr, AliasAnalysis::UnknownSize, Loc.AATags);
4011 Accesses.insert(MemAccessInfo(Ptr, false));
4013 ReadOnlyPtr.insert(Ptr);
4016 /// \brief Register a store.
4017 void addStore(AliasAnalysis::Location &Loc) {
4018 Value *Ptr = const_cast<Value*>(Loc.Ptr);
4019 AST.add(Ptr, AliasAnalysis::UnknownSize, Loc.AATags);
4020 Accesses.insert(MemAccessInfo(Ptr, true));
4023 /// \brief Check whether we can check the pointers at runtime for
4024 /// non-intersection.
4025 bool canCheckPtrAtRT(LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
4026 unsigned &NumComparisons, ScalarEvolution *SE,
4027 Loop *TheLoop, ValueToValueMap &Strides,
4028 bool ShouldCheckStride = false);
4030 /// \brief Goes over all memory accesses, checks whether a RT check is needed
4031 /// and builds sets of dependent accesses.
4032 void buildDependenceSets() {
4033 processMemAccesses();
4036 bool isRTCheckNeeded() { return IsRTCheckNeeded; }
4038 bool isDependencyCheckNeeded() { return !CheckDeps.empty(); }
4039 void resetDepChecks() { CheckDeps.clear(); }
4041 MemAccessInfoSet &getDependenciesToCheck() { return CheckDeps; }
4044 typedef SetVector<MemAccessInfo> PtrAccessSet;
4046 /// \brief Go over all memory access and check whether runtime pointer checks
4047 /// are needed /// and build sets of dependency check candidates.
4048 void processMemAccesses();
4050 /// Set of all accesses.
4051 PtrAccessSet Accesses;
4053 /// Set of accesses that need a further dependence check.
4054 MemAccessInfoSet CheckDeps;
4056 /// Set of pointers that are read only.
4057 SmallPtrSet<Value*, 16> ReadOnlyPtr;
4059 const DataLayout *DL;
4061 /// An alias set tracker to partition the access set by underlying object and
4062 //intrinsic property (such as TBAA metadata).
4063 AliasSetTracker AST;
4065 /// Sets of potentially dependent accesses - members of one set share an
4066 /// underlying pointer. The set "CheckDeps" identfies which sets really need a
4067 /// dependence check.
4068 DepCandidates &DepCands;
4070 bool IsRTCheckNeeded;
4073 } // end anonymous namespace
4075 /// \brief Check whether a pointer can participate in a runtime bounds check.
4076 static bool hasComputableBounds(ScalarEvolution *SE, ValueToValueMap &Strides,
4078 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, Strides, Ptr);
4079 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
4083 return AR->isAffine();
4086 /// \brief Check the stride of the pointer and ensure that it does not wrap in
4087 /// the address space.
4088 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
4089 const Loop *Lp, ValueToValueMap &StridesMap);
4091 bool AccessAnalysis::canCheckPtrAtRT(
4092 LoopVectorizationLegality::RuntimePointerCheck &RtCheck,
4093 unsigned &NumComparisons, ScalarEvolution *SE, Loop *TheLoop,
4094 ValueToValueMap &StridesMap, bool ShouldCheckStride) {
4095 // Find pointers with computable bounds. We are going to use this information
4096 // to place a runtime bound check.
4097 bool CanDoRT = true;
4099 bool IsDepCheckNeeded = isDependencyCheckNeeded();
4102 // We assign a consecutive id to access from different alias sets.
4103 // Accesses between different groups doesn't need to be checked.
4105 for (auto &AS : AST) {
4106 unsigned NumReadPtrChecks = 0;
4107 unsigned NumWritePtrChecks = 0;
4109 // We assign consecutive id to access from different dependence sets.
4110 // Accesses within the same set don't need a runtime check.
4111 unsigned RunningDepId = 1;
4112 DenseMap<Value *, unsigned> DepSetId;
4115 Value *Ptr = A.getValue();
4116 bool IsWrite = Accesses.count(MemAccessInfo(Ptr, true));
4117 MemAccessInfo Access(Ptr, IsWrite);
4120 ++NumWritePtrChecks;
4124 if (hasComputableBounds(SE, StridesMap, Ptr) &&
4125 // When we run after a failing dependency check we have to make sure we
4126 // don't have wrapping pointers.
4127 (!ShouldCheckStride ||
4128 isStridedPtr(SE, DL, Ptr, TheLoop, StridesMap) == 1)) {
4129 // The id of the dependence set.
4132 if (IsDepCheckNeeded) {
4133 Value *Leader = DepCands.getLeaderValue(Access).getPointer();
4134 unsigned &LeaderId = DepSetId[Leader];
4136 LeaderId = RunningDepId++;
4139 // Each access has its own dependence set.
4140 DepId = RunningDepId++;
4142 RtCheck.insert(SE, TheLoop, Ptr, IsWrite, DepId, ASId, StridesMap);
4144 DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *Ptr << '\n');
4150 if (IsDepCheckNeeded && CanDoRT && RunningDepId == 2)
4151 NumComparisons += 0; // Only one dependence set.
4153 NumComparisons += (NumWritePtrChecks * (NumReadPtrChecks +
4154 NumWritePtrChecks - 1));
4160 // If the pointers that we would use for the bounds comparison have different
4161 // address spaces, assume the values aren't directly comparable, so we can't
4162 // use them for the runtime check. We also have to assume they could
4163 // overlap. In the future there should be metadata for whether address spaces
4165 unsigned NumPointers = RtCheck.Pointers.size();
4166 for (unsigned i = 0; i < NumPointers; ++i) {
4167 for (unsigned j = i + 1; j < NumPointers; ++j) {
4168 // Only need to check pointers between two different dependency sets.
4169 if (RtCheck.DependencySetId[i] == RtCheck.DependencySetId[j])
4171 // Only need to check pointers in the same alias set.
4172 if (RtCheck.AliasSetId[i] != RtCheck.AliasSetId[j])
4175 Value *PtrI = RtCheck.Pointers[i];
4176 Value *PtrJ = RtCheck.Pointers[j];
4178 unsigned ASi = PtrI->getType()->getPointerAddressSpace();
4179 unsigned ASj = PtrJ->getType()->getPointerAddressSpace();
4181 DEBUG(dbgs() << "LV: Runtime check would require comparison between"
4182 " different address spaces\n");
4191 void AccessAnalysis::processMemAccesses() {
4192 // We process the set twice: first we process read-write pointers, last we
4193 // process read-only pointers. This allows us to skip dependence tests for
4194 // read-only pointers.
4196 DEBUG(dbgs() << "LV: Processing memory accesses...\n");
4197 DEBUG(dbgs() << " AST: "; AST.dump());
4198 DEBUG(dbgs() << "LV: Accesses:\n");
4200 for (auto A : Accesses)
4201 dbgs() << "\t" << *A.getPointer() << " (" <<
4202 (A.getInt() ? "write" : (ReadOnlyPtr.count(A.getPointer()) ?
4203 "read-only" : "read")) << ")\n";
4206 // The AliasSetTracker has nicely partitioned our pointers by metadata
4207 // compatibility and potential for underlying-object overlap. As a result, we
4208 // only need to check for potential pointer dependencies within each alias
4210 for (auto &AS : AST) {
4211 // Note that both the alias-set tracker and the alias sets themselves used
4212 // linked lists internally and so the iteration order here is deterministic
4213 // (matching the original instruction order within each set).
4215 bool SetHasWrite = false;
4217 // Map of pointers to last access encountered.
4218 typedef DenseMap<Value*, MemAccessInfo> UnderlyingObjToAccessMap;
4219 UnderlyingObjToAccessMap ObjToLastAccess;
4221 // Set of access to check after all writes have been processed.
4222 PtrAccessSet DeferredAccesses;
4224 // Iterate over each alias set twice, once to process read/write pointers,
4225 // and then to process read-only pointers.
4226 for (int SetIteration = 0; SetIteration < 2; ++SetIteration) {
4227 bool UseDeferred = SetIteration > 0;
4228 PtrAccessSet &S = UseDeferred ? DeferredAccesses : Accesses;
4231 Value *Ptr = A.getValue();
4232 bool IsWrite = S.count(MemAccessInfo(Ptr, true));
4234 // If we're using the deferred access set, then it contains only reads.
4235 bool IsReadOnlyPtr = ReadOnlyPtr.count(Ptr) && !IsWrite;
4236 if (UseDeferred && !IsReadOnlyPtr)
4238 // Otherwise, the pointer must be in the PtrAccessSet, either as a read
4240 assert(((IsReadOnlyPtr && UseDeferred) || IsWrite ||
4241 S.count(MemAccessInfo(Ptr, false))) &&
4242 "Alias-set pointer not in the access set?");
4244 MemAccessInfo Access(Ptr, IsWrite);
4245 DepCands.insert(Access);
4247 // Memorize read-only pointers for later processing and skip them in the
4248 // first round (they need to be checked after we have seen all write
4249 // pointers). Note: we also mark pointer that are not consecutive as
4250 // "read-only" pointers (so that we check "a[b[i]] +="). Hence, we need
4251 // the second check for "!IsWrite".
4252 if (!UseDeferred && IsReadOnlyPtr) {
4253 DeferredAccesses.insert(Access);
4257 // If this is a write - check other reads and writes for conflicts. If
4258 // this is a read only check other writes for conflicts (but only if
4259 // there is no other write to the ptr - this is an optimization to
4260 // catch "a[i] = a[i] + " without having to do a dependence check).
4261 if ((IsWrite || IsReadOnlyPtr) && SetHasWrite) {
4262 CheckDeps.insert(Access);
4263 IsRTCheckNeeded = true;
4269 // Create sets of pointers connected by a shared alias set and
4270 // underlying object.
4271 typedef SmallVector<Value*, 16> ValueVector;
4272 ValueVector TempObjects;
4273 GetUnderlyingObjects(Ptr, TempObjects, DL);
4274 for (Value *UnderlyingObj : TempObjects) {
4275 UnderlyingObjToAccessMap::iterator Prev =
4276 ObjToLastAccess.find(UnderlyingObj);
4277 if (Prev != ObjToLastAccess.end())
4278 DepCands.unionSets(Access, Prev->second);
4280 ObjToLastAccess[UnderlyingObj] = Access;
4288 /// \brief Checks memory dependences among accesses to the same underlying
4289 /// object to determine whether there vectorization is legal or not (and at
4290 /// which vectorization factor).
4292 /// This class works under the assumption that we already checked that memory
4293 /// locations with different underlying pointers are "must-not alias".
4294 /// We use the ScalarEvolution framework to symbolically evalutate access
4295 /// functions pairs. Since we currently don't restructure the loop we can rely
4296 /// on the program order of memory accesses to determine their safety.
4297 /// At the moment we will only deem accesses as safe for:
4298 /// * A negative constant distance assuming program order.
4300 /// Safe: tmp = a[i + 1]; OR a[i + 1] = x;
4301 /// a[i] = tmp; y = a[i];
4303 /// The latter case is safe because later checks guarantuee that there can't
4304 /// be a cycle through a phi node (that is, we check that "x" and "y" is not
4305 /// the same variable: a header phi can only be an induction or a reduction, a
4306 /// reduction can't have a memory sink, an induction can't have a memory
4307 /// source). This is important and must not be violated (or we have to
4308 /// resort to checking for cycles through memory).
4310 /// * A positive constant distance assuming program order that is bigger
4311 /// than the biggest memory access.
4313 /// tmp = a[i] OR b[i] = x
4314 /// a[i+2] = tmp y = b[i+2];
4316 /// Safe distance: 2 x sizeof(a[0]), and 2 x sizeof(b[0]), respectively.
4318 /// * Zero distances and all accesses have the same size.
4320 class MemoryDepChecker {
4322 typedef PointerIntPair<Value *, 1, bool> MemAccessInfo;
4323 typedef SmallPtrSet<MemAccessInfo, 8> MemAccessInfoSet;
4325 MemoryDepChecker(ScalarEvolution *Se, const DataLayout *Dl, const Loop *L)
4326 : SE(Se), DL(Dl), InnermostLoop(L), AccessIdx(0),
4327 ShouldRetryWithRuntimeCheck(false) {}
4329 /// \brief Register the location (instructions are given increasing numbers)
4330 /// of a write access.
4331 void addAccess(StoreInst *SI) {
4332 Value *Ptr = SI->getPointerOperand();
4333 Accesses[MemAccessInfo(Ptr, true)].push_back(AccessIdx);
4334 InstMap.push_back(SI);
4338 /// \brief Register the location (instructions are given increasing numbers)
4339 /// of a write access.
4340 void addAccess(LoadInst *LI) {
4341 Value *Ptr = LI->getPointerOperand();
4342 Accesses[MemAccessInfo(Ptr, false)].push_back(AccessIdx);
4343 InstMap.push_back(LI);
4347 /// \brief Check whether the dependencies between the accesses are safe.
4349 /// Only checks sets with elements in \p CheckDeps.
4350 bool areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4351 MemAccessInfoSet &CheckDeps, ValueToValueMap &Strides);
4353 /// \brief The maximum number of bytes of a vector register we can vectorize
4354 /// the accesses safely with.
4355 unsigned getMaxSafeDepDistBytes() { return MaxSafeDepDistBytes; }
4357 /// \brief In same cases when the dependency check fails we can still
4358 /// vectorize the loop with a dynamic array access check.
4359 bool shouldRetryWithRuntimeCheck() { return ShouldRetryWithRuntimeCheck; }
4362 ScalarEvolution *SE;
4363 const DataLayout *DL;
4364 const Loop *InnermostLoop;
4366 /// \brief Maps access locations (ptr, read/write) to program order.
4367 DenseMap<MemAccessInfo, std::vector<unsigned> > Accesses;
4369 /// \brief Memory access instructions in program order.
4370 SmallVector<Instruction *, 16> InstMap;
4372 /// \brief The program order index to be used for the next instruction.
4375 // We can access this many bytes in parallel safely.
4376 unsigned MaxSafeDepDistBytes;
4378 /// \brief If we see a non-constant dependence distance we can still try to
4379 /// vectorize this loop with runtime checks.
4380 bool ShouldRetryWithRuntimeCheck;
4382 /// \brief Check whether there is a plausible dependence between the two
4385 /// Access \p A must happen before \p B in program order. The two indices
4386 /// identify the index into the program order map.
4388 /// This function checks whether there is a plausible dependence (or the
4389 /// absence of such can't be proved) between the two accesses. If there is a
4390 /// plausible dependence but the dependence distance is bigger than one
4391 /// element access it records this distance in \p MaxSafeDepDistBytes (if this
4392 /// distance is smaller than any other distance encountered so far).
4393 /// Otherwise, this function returns true signaling a possible dependence.
4394 bool isDependent(const MemAccessInfo &A, unsigned AIdx,
4395 const MemAccessInfo &B, unsigned BIdx,
4396 ValueToValueMap &Strides);
4398 /// \brief Check whether the data dependence could prevent store-load
4400 bool couldPreventStoreLoadForward(unsigned Distance, unsigned TypeByteSize);
4403 } // end anonymous namespace
4405 static bool isInBoundsGep(Value *Ptr) {
4406 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
4407 return GEP->isInBounds();
4411 /// \brief Check whether the access through \p Ptr has a constant stride.
4412 static int isStridedPtr(ScalarEvolution *SE, const DataLayout *DL, Value *Ptr,
4413 const Loop *Lp, ValueToValueMap &StridesMap) {
4414 const Type *Ty = Ptr->getType();
4415 assert(Ty->isPointerTy() && "Unexpected non-ptr");
4417 // Make sure that the pointer does not point to aggregate types.
4418 const PointerType *PtrTy = cast<PointerType>(Ty);
4419 if (PtrTy->getElementType()->isAggregateType()) {
4420 DEBUG(dbgs() << "LV: Bad stride - Not a pointer to a scalar type" << *Ptr <<
4425 const SCEV *PtrScev = replaceSymbolicStrideSCEV(SE, StridesMap, Ptr);
4427 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PtrScev);
4429 DEBUG(dbgs() << "LV: Bad stride - Not an AddRecExpr pointer "
4430 << *Ptr << " SCEV: " << *PtrScev << "\n");
4434 // The accesss function must stride over the innermost loop.
4435 if (Lp != AR->getLoop()) {
4436 DEBUG(dbgs() << "LV: Bad stride - Not striding over innermost loop " <<
4437 *Ptr << " SCEV: " << *PtrScev << "\n");
4440 // The address calculation must not wrap. Otherwise, a dependence could be
4442 // An inbounds getelementptr that is a AddRec with a unit stride
4443 // cannot wrap per definition. The unit stride requirement is checked later.
4444 // An getelementptr without an inbounds attribute and unit stride would have
4445 // to access the pointer value "0" which is undefined behavior in address
4446 // space 0, therefore we can also vectorize this case.
4447 bool IsInBoundsGEP = isInBoundsGep(Ptr);
4448 bool IsNoWrapAddRec = AR->getNoWrapFlags(SCEV::NoWrapMask);
4449 bool IsInAddressSpaceZero = PtrTy->getAddressSpace() == 0;
4450 if (!IsNoWrapAddRec && !IsInBoundsGEP && !IsInAddressSpaceZero) {
4451 DEBUG(dbgs() << "LV: Bad stride - Pointer may wrap in the address space "
4452 << *Ptr << " SCEV: " << *PtrScev << "\n");
4456 // Check the step is constant.
4457 const SCEV *Step = AR->getStepRecurrence(*SE);
4459 // Calculate the pointer stride and check if it is consecutive.
4460 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
4462 DEBUG(dbgs() << "LV: Bad stride - Not a constant strided " << *Ptr <<
4463 " SCEV: " << *PtrScev << "\n");
4467 int64_t Size = DL->getTypeAllocSize(PtrTy->getElementType());
4468 const APInt &APStepVal = C->getValue()->getValue();
4470 // Huge step value - give up.
4471 if (APStepVal.getBitWidth() > 64)
4474 int64_t StepVal = APStepVal.getSExtValue();
4477 int64_t Stride = StepVal / Size;
4478 int64_t Rem = StepVal % Size;
4482 // If the SCEV could wrap but we have an inbounds gep with a unit stride we
4483 // know we can't "wrap around the address space". In case of address space
4484 // zero we know that this won't happen without triggering undefined behavior.
4485 if (!IsNoWrapAddRec && (IsInBoundsGEP || IsInAddressSpaceZero) &&
4486 Stride != 1 && Stride != -1)
4492 bool MemoryDepChecker::couldPreventStoreLoadForward(unsigned Distance,
4493 unsigned TypeByteSize) {
4494 // If loads occur at a distance that is not a multiple of a feasible vector
4495 // factor store-load forwarding does not take place.
4496 // Positive dependences might cause troubles because vectorizing them might
4497 // prevent store-load forwarding making vectorized code run a lot slower.
4498 // a[i] = a[i-3] ^ a[i-8];
4499 // The stores to a[i:i+1] don't align with the stores to a[i-3:i-2] and
4500 // hence on your typical architecture store-load forwarding does not take
4501 // place. Vectorizing in such cases does not make sense.
4502 // Store-load forwarding distance.
4503 const unsigned NumCyclesForStoreLoadThroughMemory = 8*TypeByteSize;
4504 // Maximum vector factor.
4505 unsigned MaxVFWithoutSLForwardIssues = MaxVectorWidth*TypeByteSize;
4506 if(MaxSafeDepDistBytes < MaxVFWithoutSLForwardIssues)
4507 MaxVFWithoutSLForwardIssues = MaxSafeDepDistBytes;
4509 for (unsigned vf = 2*TypeByteSize; vf <= MaxVFWithoutSLForwardIssues;
4511 if (Distance % vf && Distance / vf < NumCyclesForStoreLoadThroughMemory) {
4512 MaxVFWithoutSLForwardIssues = (vf >>=1);
4517 if (MaxVFWithoutSLForwardIssues< 2*TypeByteSize) {
4518 DEBUG(dbgs() << "LV: Distance " << Distance <<
4519 " that could cause a store-load forwarding conflict\n");
4523 if (MaxVFWithoutSLForwardIssues < MaxSafeDepDistBytes &&
4524 MaxVFWithoutSLForwardIssues != MaxVectorWidth*TypeByteSize)
4525 MaxSafeDepDistBytes = MaxVFWithoutSLForwardIssues;
4529 bool MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
4530 const MemAccessInfo &B, unsigned BIdx,
4531 ValueToValueMap &Strides) {
4532 assert (AIdx < BIdx && "Must pass arguments in program order");
4534 Value *APtr = A.getPointer();
4535 Value *BPtr = B.getPointer();
4536 bool AIsWrite = A.getInt();
4537 bool BIsWrite = B.getInt();
4539 // Two reads are independent.
4540 if (!AIsWrite && !BIsWrite)
4543 // We cannot check pointers in different address spaces.
4544 if (APtr->getType()->getPointerAddressSpace() !=
4545 BPtr->getType()->getPointerAddressSpace())
4548 const SCEV *AScev = replaceSymbolicStrideSCEV(SE, Strides, APtr);
4549 const SCEV *BScev = replaceSymbolicStrideSCEV(SE, Strides, BPtr);
4551 int StrideAPtr = isStridedPtr(SE, DL, APtr, InnermostLoop, Strides);
4552 int StrideBPtr = isStridedPtr(SE, DL, BPtr, InnermostLoop, Strides);
4554 const SCEV *Src = AScev;
4555 const SCEV *Sink = BScev;
4557 // If the induction step is negative we have to invert source and sink of the
4559 if (StrideAPtr < 0) {
4562 std::swap(APtr, BPtr);
4563 std::swap(Src, Sink);
4564 std::swap(AIsWrite, BIsWrite);
4565 std::swap(AIdx, BIdx);
4566 std::swap(StrideAPtr, StrideBPtr);
4569 const SCEV *Dist = SE->getMinusSCEV(Sink, Src);
4571 DEBUG(dbgs() << "LV: Src Scev: " << *Src << "Sink Scev: " << *Sink
4572 << "(Induction step: " << StrideAPtr << ")\n");
4573 DEBUG(dbgs() << "LV: Distance for " << *InstMap[AIdx] << " to "
4574 << *InstMap[BIdx] << ": " << *Dist << "\n");
4576 // Need consecutive accesses. We don't want to vectorize
4577 // "A[B[i]] += ..." and similar code or pointer arithmetic that could wrap in
4578 // the address space.
4579 if (!StrideAPtr || !StrideBPtr || StrideAPtr != StrideBPtr){
4580 DEBUG(dbgs() << "Non-consecutive pointer access\n");
4584 const SCEVConstant *C = dyn_cast<SCEVConstant>(Dist);
4586 DEBUG(dbgs() << "LV: Dependence because of non-constant distance\n");
4587 ShouldRetryWithRuntimeCheck = true;
4591 Type *ATy = APtr->getType()->getPointerElementType();
4592 Type *BTy = BPtr->getType()->getPointerElementType();
4593 unsigned TypeByteSize = DL->getTypeAllocSize(ATy);
4595 // Negative distances are not plausible dependencies.
4596 const APInt &Val = C->getValue()->getValue();
4597 if (Val.isNegative()) {
4598 bool IsTrueDataDependence = (AIsWrite && !BIsWrite);
4599 if (IsTrueDataDependence &&
4600 (couldPreventStoreLoadForward(Val.abs().getZExtValue(), TypeByteSize) ||
4604 DEBUG(dbgs() << "LV: Dependence is negative: NoDep\n");
4608 // Write to the same location with the same size.
4609 // Could be improved to assert type sizes are the same (i32 == float, etc).
4613 DEBUG(dbgs() << "LV: Zero dependence difference but different types\n");
4617 assert(Val.isStrictlyPositive() && "Expect a positive value");
4619 // Positive distance bigger than max vectorization factor.
4622 "LV: ReadWrite-Write positive dependency with different types\n");
4626 unsigned Distance = (unsigned) Val.getZExtValue();
4628 // Bail out early if passed-in parameters make vectorization not feasible.
4629 unsigned ForcedFactor = VectorizationFactor ? VectorizationFactor : 1;
4630 unsigned ForcedUnroll = VectorizationInterleave ? VectorizationInterleave : 1;
4632 // The distance must be bigger than the size needed for a vectorized version
4633 // of the operation and the size of the vectorized operation must not be
4634 // bigger than the currrent maximum size.
4635 if (Distance < 2*TypeByteSize ||
4636 2*TypeByteSize > MaxSafeDepDistBytes ||
4637 Distance < TypeByteSize * ForcedUnroll * ForcedFactor) {
4638 DEBUG(dbgs() << "LV: Failure because of Positive distance "
4639 << Val.getSExtValue() << '\n');
4643 MaxSafeDepDistBytes = Distance < MaxSafeDepDistBytes ?
4644 Distance : MaxSafeDepDistBytes;
4646 bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
4647 if (IsTrueDataDependence &&
4648 couldPreventStoreLoadForward(Distance, TypeByteSize))
4651 DEBUG(dbgs() << "LV: Positive distance " << Val.getSExtValue() <<
4652 " with max VF = " << MaxSafeDepDistBytes / TypeByteSize << '\n');
4657 bool MemoryDepChecker::areDepsSafe(AccessAnalysis::DepCandidates &AccessSets,
4658 MemAccessInfoSet &CheckDeps,
4659 ValueToValueMap &Strides) {
4661 MaxSafeDepDistBytes = -1U;
4662 while (!CheckDeps.empty()) {
4663 MemAccessInfo CurAccess = *CheckDeps.begin();
4665 // Get the relevant memory access set.
4666 EquivalenceClasses<MemAccessInfo>::iterator I =
4667 AccessSets.findValue(AccessSets.getLeaderValue(CurAccess));
4669 // Check accesses within this set.
4670 EquivalenceClasses<MemAccessInfo>::member_iterator AI, AE;
4671 AI = AccessSets.member_begin(I), AE = AccessSets.member_end();
4673 // Check every access pair.
4675 CheckDeps.erase(*AI);
4676 EquivalenceClasses<MemAccessInfo>::member_iterator OI = std::next(AI);
4678 // Check every accessing instruction pair in program order.
4679 for (std::vector<unsigned>::iterator I1 = Accesses[*AI].begin(),
4680 I1E = Accesses[*AI].end(); I1 != I1E; ++I1)
4681 for (std::vector<unsigned>::iterator I2 = Accesses[*OI].begin(),
4682 I2E = Accesses[*OI].end(); I2 != I2E; ++I2) {
4683 if (*I1 < *I2 && isDependent(*AI, *I1, *OI, *I2, Strides))
4685 if (*I2 < *I1 && isDependent(*OI, *I2, *AI, *I1, Strides))
4696 bool LoopVectorizationLegality::canVectorizeMemory() {
4698 typedef SmallVector<Value*, 16> ValueVector;
4699 typedef SmallPtrSet<Value*, 16> ValueSet;
4701 // Holds the Load and Store *instructions*.
4705 // Holds all the different accesses in the loop.
4706 unsigned NumReads = 0;
4707 unsigned NumReadWrites = 0;
4709 PtrRtCheck.Pointers.clear();
4710 PtrRtCheck.Need = false;
4712 const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
4713 MemoryDepChecker DepChecker(SE, DL, TheLoop);
4716 for (Loop::block_iterator bb = TheLoop->block_begin(),
4717 be = TheLoop->block_end(); bb != be; ++bb) {
4719 // Scan the BB and collect legal loads and stores.
4720 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
4723 // If this is a load, save it. If this instruction can read from memory
4724 // but is not a load, then we quit. Notice that we don't handle function
4725 // calls that read or write.
4726 if (it->mayReadFromMemory()) {
4727 // Many math library functions read the rounding mode. We will only
4728 // vectorize a loop if it contains known function calls that don't set
4729 // the flag. Therefore, it is safe to ignore this read from memory.
4730 CallInst *Call = dyn_cast<CallInst>(it);
4731 if (Call && getIntrinsicIDForCall(Call, TLI))
4734 LoadInst *Ld = dyn_cast<LoadInst>(it);
4735 if (!Ld || (!Ld->isSimple() && !IsAnnotatedParallel)) {
4736 emitAnalysis(Report(Ld)
4737 << "read with atomic ordering or volatile read");
4738 DEBUG(dbgs() << "LV: Found a non-simple load.\n");
4742 Loads.push_back(Ld);
4743 DepChecker.addAccess(Ld);
4747 // Save 'store' instructions. Abort if other instructions write to memory.
4748 if (it->mayWriteToMemory()) {
4749 StoreInst *St = dyn_cast<StoreInst>(it);
4751 emitAnalysis(Report(it) << "instruction cannot be vectorized");
4754 if (!St->isSimple() && !IsAnnotatedParallel) {
4755 emitAnalysis(Report(St)
4756 << "write with atomic ordering or volatile write");
4757 DEBUG(dbgs() << "LV: Found a non-simple store.\n");
4761 Stores.push_back(St);
4762 DepChecker.addAccess(St);
4767 // Now we have two lists that hold the loads and the stores.
4768 // Next, we find the pointers that they use.
4770 // Check if we see any stores. If there are no stores, then we don't
4771 // care if the pointers are *restrict*.
4772 if (!Stores.size()) {
4773 DEBUG(dbgs() << "LV: Found a read-only loop!\n");
4777 AccessAnalysis::DepCandidates DependentAccesses;
4778 AccessAnalysis Accesses(DL, AA, DependentAccesses);
4780 // Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
4781 // multiple times on the same object. If the ptr is accessed twice, once
4782 // for read and once for write, it will only appear once (on the write
4783 // list). This is okay, since we are going to check for conflicts between
4784 // writes and between reads and writes, but not between reads and reads.
4787 ValueVector::iterator I, IE;
4788 for (I = Stores.begin(), IE = Stores.end(); I != IE; ++I) {
4789 StoreInst *ST = cast<StoreInst>(*I);
4790 Value* Ptr = ST->getPointerOperand();
4792 if (isUniform(Ptr)) {
4795 << "write to a loop invariant address could not be vectorized");
4796 DEBUG(dbgs() << "LV: We don't allow storing to uniform addresses\n");
4800 // If we did *not* see this pointer before, insert it to the read-write
4801 // list. At this phase it is only a 'write' list.
4802 if (Seen.insert(Ptr)) {
4805 AliasAnalysis::Location Loc = AA->getLocation(ST);
4806 // The TBAA metadata could have a control dependency on the predication
4807 // condition, so we cannot rely on it when determining whether or not we
4808 // need runtime pointer checks.
4809 if (blockNeedsPredication(ST->getParent()))
4810 Loc.AATags.TBAA = nullptr;
4812 Accesses.addStore(Loc);
4816 if (IsAnnotatedParallel) {
4818 << "LV: A loop annotated parallel, ignore memory dependency "
4823 for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
4824 LoadInst *LD = cast<LoadInst>(*I);
4825 Value* Ptr = LD->getPointerOperand();
4826 // If we did *not* see this pointer before, insert it to the
4827 // read list. If we *did* see it before, then it is already in
4828 // the read-write list. This allows us to vectorize expressions
4829 // such as A[i] += x; Because the address of A[i] is a read-write
4830 // pointer. This only works if the index of A[i] is consecutive.
4831 // If the address of i is unknown (for example A[B[i]]) then we may
4832 // read a few words, modify, and write a few words, and some of the
4833 // words may be written to the same address.
4834 bool IsReadOnlyPtr = false;
4835 if (Seen.insert(Ptr) || !isStridedPtr(SE, DL, Ptr, TheLoop, Strides)) {
4837 IsReadOnlyPtr = true;
4840 AliasAnalysis::Location Loc = AA->getLocation(LD);
4841 // The TBAA metadata could have a control dependency on the predication
4842 // condition, so we cannot rely on it when determining whether or not we
4843 // need runtime pointer checks.
4844 if (blockNeedsPredication(LD->getParent()))
4845 Loc.AATags.TBAA = nullptr;
4847 Accesses.addLoad(Loc, IsReadOnlyPtr);
4850 // If we write (or read-write) to a single destination and there are no
4851 // other reads in this loop then is it safe to vectorize.
4852 if (NumReadWrites == 1 && NumReads == 0) {
4853 DEBUG(dbgs() << "LV: Found a write-only loop!\n");
4857 // Build dependence sets and check whether we need a runtime pointer bounds
4859 Accesses.buildDependenceSets();
4860 bool NeedRTCheck = Accesses.isRTCheckNeeded();
4862 // Find pointers with computable bounds. We are going to use this information
4863 // to place a runtime bound check.
4864 unsigned NumComparisons = 0;
4865 bool CanDoRT = false;
4867 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE, TheLoop,
4870 DEBUG(dbgs() << "LV: We need to do " << NumComparisons <<
4871 " pointer comparisons.\n");
4873 // If we only have one set of dependences to check pointers among we don't
4874 // need a runtime check.
4875 if (NumComparisons == 0 && NeedRTCheck)
4876 NeedRTCheck = false;
4878 // Check that we did not collect too many pointers or found an unsizeable
4880 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4886 DEBUG(dbgs() << "LV: We can perform a memory runtime check if needed.\n");
4889 if (NeedRTCheck && !CanDoRT) {
4890 emitAnalysis(Report() << "cannot identify array bounds");
4891 DEBUG(dbgs() << "LV: We can't vectorize because we can't find " <<
4892 "the array bounds.\n");
4897 PtrRtCheck.Need = NeedRTCheck;
4899 bool CanVecMem = true;
4900 if (Accesses.isDependencyCheckNeeded()) {
4901 DEBUG(dbgs() << "LV: Checking memory dependencies\n");
4902 CanVecMem = DepChecker.areDepsSafe(
4903 DependentAccesses, Accesses.getDependenciesToCheck(), Strides);
4904 MaxSafeDepDistBytes = DepChecker.getMaxSafeDepDistBytes();
4906 if (!CanVecMem && DepChecker.shouldRetryWithRuntimeCheck()) {
4907 DEBUG(dbgs() << "LV: Retrying with memory checks\n");
4910 // Clear the dependency checks. We assume they are not needed.
4911 Accesses.resetDepChecks();
4914 PtrRtCheck.Need = true;
4916 CanDoRT = Accesses.canCheckPtrAtRT(PtrRtCheck, NumComparisons, SE,
4917 TheLoop, Strides, true);
4918 // Check that we did not collect too many pointers or found an unsizeable
4920 if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
4921 if (!CanDoRT && NumComparisons > 0)
4922 emitAnalysis(Report()
4923 << "cannot check memory dependencies at runtime");
4925 emitAnalysis(Report()
4926 << NumComparisons << " exceeds limit of "
4927 << RuntimeMemoryCheckThreshold
4928 << " dependent memory operations checked at runtime");
4929 DEBUG(dbgs() << "LV: Can't vectorize with memory checks\n");
4939 emitAnalysis(Report() << "unsafe dependent memory operations in loop");
4941 DEBUG(dbgs() << "LV: We" << (NeedRTCheck ? "" : " don't") <<
4942 " need a runtime memory check.\n");
4947 static bool hasMultipleUsesOf(Instruction *I,
4948 SmallPtrSetImpl<Instruction *> &Insts) {
4949 unsigned NumUses = 0;
4950 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
4951 if (Insts.count(dyn_cast<Instruction>(*Use)))
4960 static bool areAllUsesIn(Instruction *I, SmallPtrSetImpl<Instruction *> &Set) {
4961 for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
4962 if (!Set.count(dyn_cast<Instruction>(*Use)))
4967 bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
4968 ReductionKind Kind) {
4969 if (Phi->getNumIncomingValues() != 2)
4972 // Reduction variables are only found in the loop header block.
4973 if (Phi->getParent() != TheLoop->getHeader())
4976 // Obtain the reduction start value from the value that comes from the loop
4978 Value *RdxStart = Phi->getIncomingValueForBlock(TheLoop->getLoopPreheader());
4980 // ExitInstruction is the single value which is used outside the loop.
4981 // We only allow for a single reduction value to be used outside the loop.
4982 // This includes users of the reduction, variables (which form a cycle
4983 // which ends in the phi node).
4984 Instruction *ExitInstruction = nullptr;
4985 // Indicates that we found a reduction operation in our scan.
4986 bool FoundReduxOp = false;
4988 // We start with the PHI node and scan for all of the users of this
4989 // instruction. All users must be instructions that can be used as reduction
4990 // variables (such as ADD). We must have a single out-of-block user. The cycle
4991 // must include the original PHI.
4992 bool FoundStartPHI = false;
4994 // To recognize min/max patterns formed by a icmp select sequence, we store
4995 // the number of instruction we saw from the recognized min/max pattern,
4996 // to make sure we only see exactly the two instructions.
4997 unsigned NumCmpSelectPatternInst = 0;
4998 ReductionInstDesc ReduxDesc(false, nullptr);
5000 SmallPtrSet<Instruction *, 8> VisitedInsts;
5001 SmallVector<Instruction *, 8> Worklist;
5002 Worklist.push_back(Phi);
5003 VisitedInsts.insert(Phi);
5005 // A value in the reduction can be used:
5006 // - By the reduction:
5007 // - Reduction operation:
5008 // - One use of reduction value (safe).
5009 // - Multiple use of reduction value (not safe).
5011 // - All uses of the PHI must be the reduction (safe).
5012 // - Otherwise, not safe.
5013 // - By one instruction outside of the loop (safe).
5014 // - By further instructions outside of the loop (not safe).
5015 // - By an instruction that is not part of the reduction (not safe).
5017 // * An instruction type other than PHI or the reduction operation.
5018 // * A PHI in the header other than the initial PHI.
5019 while (!Worklist.empty()) {
5020 Instruction *Cur = Worklist.back();
5021 Worklist.pop_back();
5024 // If the instruction has no users then this is a broken chain and can't be
5025 // a reduction variable.
5026 if (Cur->use_empty())
5029 bool IsAPhi = isa<PHINode>(Cur);
5031 // A header PHI use other than the original PHI.
5032 if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
5035 // Reductions of instructions such as Div, and Sub is only possible if the
5036 // LHS is the reduction variable.
5037 if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
5038 !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
5039 !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
5042 // Any reduction instruction must be of one of the allowed kinds.
5043 ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
5044 if (!ReduxDesc.IsReduction)
5047 // A reduction operation must only have one use of the reduction value.
5048 if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
5049 hasMultipleUsesOf(Cur, VisitedInsts))
5052 // All inputs to a PHI node must be a reduction value.
5053 if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
5056 if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
5057 isa<SelectInst>(Cur)))
5058 ++NumCmpSelectPatternInst;
5059 if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
5060 isa<SelectInst>(Cur)))
5061 ++NumCmpSelectPatternInst;
5063 // Check whether we found a reduction operator.
5064 FoundReduxOp |= !IsAPhi;
5066 // Process users of current instruction. Push non-PHI nodes after PHI nodes
5067 // onto the stack. This way we are going to have seen all inputs to PHI
5068 // nodes once we get to them.
5069 SmallVector<Instruction *, 8> NonPHIs;
5070 SmallVector<Instruction *, 8> PHIs;
5071 for (User *U : Cur->users()) {
5072 Instruction *UI = cast<Instruction>(U);
5074 // Check if we found the exit user.
5075 BasicBlock *Parent = UI->getParent();
5076 if (!TheLoop->contains(Parent)) {
5077 // Exit if you find multiple outside users or if the header phi node is
5078 // being used. In this case the user uses the value of the previous
5079 // iteration, in which case we would loose "VF-1" iterations of the
5080 // reduction operation if we vectorize.
5081 if (ExitInstruction != nullptr || Cur == Phi)
5084 // The instruction used by an outside user must be the last instruction
5085 // before we feed back to the reduction phi. Otherwise, we loose VF-1
5086 // operations on the value.
5087 if (std::find(Phi->op_begin(), Phi->op_end(), Cur) == Phi->op_end())
5090 ExitInstruction = Cur;
5094 // Process instructions only once (termination). Each reduction cycle
5095 // value must only be used once, except by phi nodes and min/max
5096 // reductions which are represented as a cmp followed by a select.
5097 ReductionInstDesc IgnoredVal(false, nullptr);
5098 if (VisitedInsts.insert(UI)) {
5099 if (isa<PHINode>(UI))
5102 NonPHIs.push_back(UI);
5103 } else if (!isa<PHINode>(UI) &&
5104 ((!isa<FCmpInst>(UI) &&
5105 !isa<ICmpInst>(UI) &&
5106 !isa<SelectInst>(UI)) ||
5107 !isMinMaxSelectCmpPattern(UI, IgnoredVal).IsReduction))
5110 // Remember that we completed the cycle.
5112 FoundStartPHI = true;
5114 Worklist.append(PHIs.begin(), PHIs.end());
5115 Worklist.append(NonPHIs.begin(), NonPHIs.end());
5118 // This means we have seen one but not the other instruction of the
5119 // pattern or more than just a select and cmp.
5120 if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
5121 NumCmpSelectPatternInst != 2)
5124 if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
5127 // We found a reduction var if we have reached the original phi node and we
5128 // only have a single instruction with out-of-loop users.
5130 // This instruction is allowed to have out-of-loop users.
5131 AllowedExit.insert(ExitInstruction);
5133 // Save the description of this reduction variable.
5134 ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
5135 ReduxDesc.MinMaxKind);
5136 Reductions[Phi] = RD;
5137 // We've ended the cycle. This is a reduction variable if we have an
5138 // outside user and it has a binary op.
5143 /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
5144 /// pattern corresponding to a min(X, Y) or max(X, Y).
5145 LoopVectorizationLegality::ReductionInstDesc
5146 LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
5147 ReductionInstDesc &Prev) {
5149 assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
5150 "Expect a select instruction");
5151 Instruction *Cmp = nullptr;
5152 SelectInst *Select = nullptr;
5154 // We must handle the select(cmp()) as a single instruction. Advance to the
5156 if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
5157 if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->user_begin())))
5158 return ReductionInstDesc(false, I);
5159 return ReductionInstDesc(Select, Prev.MinMaxKind);
5162 // Only handle single use cases for now.
5163 if (!(Select = dyn_cast<SelectInst>(I)))
5164 return ReductionInstDesc(false, I);
5165 if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
5166 !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
5167 return ReductionInstDesc(false, I);
5168 if (!Cmp->hasOneUse())
5169 return ReductionInstDesc(false, I);
5174 // Look for a min/max pattern.
5175 if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5176 return ReductionInstDesc(Select, MRK_UIntMin);
5177 else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5178 return ReductionInstDesc(Select, MRK_UIntMax);
5179 else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5180 return ReductionInstDesc(Select, MRK_SIntMax);
5181 else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5182 return ReductionInstDesc(Select, MRK_SIntMin);
5183 else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5184 return ReductionInstDesc(Select, MRK_FloatMin);
5185 else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5186 return ReductionInstDesc(Select, MRK_FloatMax);
5187 else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5188 return ReductionInstDesc(Select, MRK_FloatMin);
5189 else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
5190 return ReductionInstDesc(Select, MRK_FloatMax);
5192 return ReductionInstDesc(false, I);
5195 LoopVectorizationLegality::ReductionInstDesc
5196 LoopVectorizationLegality::isReductionInstr(Instruction *I,
5198 ReductionInstDesc &Prev) {
5199 bool FP = I->getType()->isFloatingPointTy();
5200 bool FastMath = FP && I->hasUnsafeAlgebra();
5201 switch (I->getOpcode()) {
5203 return ReductionInstDesc(false, I);
5204 case Instruction::PHI:
5205 if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
5206 Kind != RK_FloatMinMax))
5207 return ReductionInstDesc(false, I);
5208 return ReductionInstDesc(I, Prev.MinMaxKind);
5209 case Instruction::Sub:
5210 case Instruction::Add:
5211 return ReductionInstDesc(Kind == RK_IntegerAdd, I);
5212 case Instruction::Mul:
5213 return ReductionInstDesc(Kind == RK_IntegerMult, I);
5214 case Instruction::And:
5215 return ReductionInstDesc(Kind == RK_IntegerAnd, I);
5216 case Instruction::Or:
5217 return ReductionInstDesc(Kind == RK_IntegerOr, I);
5218 case Instruction::Xor:
5219 return ReductionInstDesc(Kind == RK_IntegerXor, I);
5220 case Instruction::FMul:
5221 return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
5222 case Instruction::FSub:
5223 case Instruction::FAdd:
5224 return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
5225 case Instruction::FCmp:
5226 case Instruction::ICmp:
5227 case Instruction::Select:
5228 if (Kind != RK_IntegerMinMax &&
5229 (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
5230 return ReductionInstDesc(false, I);
5231 return isMinMaxSelectCmpPattern(I, Prev);
5235 LoopVectorizationLegality::InductionKind
5236 LoopVectorizationLegality::isInductionVariable(PHINode *Phi) {
5237 Type *PhiTy = Phi->getType();
5238 // We only handle integer and pointer inductions variables.
5239 if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
5240 return IK_NoInduction;
5242 // Check that the PHI is consecutive.
5243 const SCEV *PhiScev = SE->getSCEV(Phi);
5244 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
5246 DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
5247 return IK_NoInduction;
5249 const SCEV *Step = AR->getStepRecurrence(*SE);
5251 // Integer inductions need to have a stride of one.
5252 if (PhiTy->isIntegerTy()) {
5254 return IK_IntInduction;
5255 if (Step->isAllOnesValue())
5256 return IK_ReverseIntInduction;
5257 return IK_NoInduction;
5260 // Calculate the pointer stride and check if it is consecutive.
5261 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5263 return IK_NoInduction;
5265 assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
5266 uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
5267 if (C->getValue()->equalsInt(Size))
5268 return IK_PtrInduction;
5269 else if (C->getValue()->equalsInt(0 - Size))
5270 return IK_ReversePtrInduction;
5272 return IK_NoInduction;
5275 bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
5276 Value *In0 = const_cast<Value*>(V);
5277 PHINode *PN = dyn_cast_or_null<PHINode>(In0);
5281 return Inductions.count(PN);
5284 bool LoopVectorizationLegality::blockNeedsPredication(BasicBlock *BB) {
5285 assert(TheLoop->contains(BB) && "Unknown block used");
5287 // Blocks that do not dominate the latch need predication.
5288 BasicBlock* Latch = TheLoop->getLoopLatch();
5289 return !DT->dominates(BB, Latch);
5292 bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB,
5293 SmallPtrSetImpl<Value *> &SafePtrs) {
5294 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5295 // We might be able to hoist the load.
5296 if (it->mayReadFromMemory()) {
5297 LoadInst *LI = dyn_cast<LoadInst>(it);
5298 if (!LI || !SafePtrs.count(LI->getPointerOperand()))
5302 // We don't predicate stores at the moment.
5303 if (it->mayWriteToMemory()) {
5304 StoreInst *SI = dyn_cast<StoreInst>(it);
5305 // We only support predication of stores in basic blocks with one
5307 if (!SI || ++NumPredStores > NumberOfStoresToPredicate ||
5308 !SafePtrs.count(SI->getPointerOperand()) ||
5309 !SI->getParent()->getSinglePredecessor())
5315 // Check that we don't have a constant expression that can trap as operand.
5316 for (Instruction::op_iterator OI = it->op_begin(), OE = it->op_end();
5318 if (Constant *C = dyn_cast<Constant>(*OI))
5323 // The instructions below can trap.
5324 switch (it->getOpcode()) {
5326 case Instruction::UDiv:
5327 case Instruction::SDiv:
5328 case Instruction::URem:
5329 case Instruction::SRem:
5337 LoopVectorizationCostModel::VectorizationFactor
5338 LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize) {
5339 // Width 1 means no vectorize
5340 VectorizationFactor Factor = { 1U, 0U };
5341 if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
5342 emitAnalysis(Report() << "runtime pointer checks needed. Enable vectorization of this loop with '#pragma clang loop vectorize(enable)' when compiling with -Os");
5343 DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
5347 if (!EnableCondStoresVectorization && Legal->NumPredStores) {
5348 emitAnalysis(Report() << "store that is conditionally executed prevents vectorization");
5349 DEBUG(dbgs() << "LV: No vectorization. There are conditional stores.\n");
5353 // Find the trip count.
5354 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
5355 DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');
5357 unsigned WidestType = getWidestType();
5358 unsigned WidestRegister = TTI.getRegisterBitWidth(true);
5359 unsigned MaxSafeDepDist = -1U;
5360 if (Legal->getMaxSafeDepDistBytes() != -1U)
5361 MaxSafeDepDist = Legal->getMaxSafeDepDistBytes() * 8;
5362 WidestRegister = ((WidestRegister < MaxSafeDepDist) ?
5363 WidestRegister : MaxSafeDepDist);
5364 unsigned MaxVectorSize = WidestRegister / WidestType;
5365 DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
5366 DEBUG(dbgs() << "LV: The Widest register is: "
5367 << WidestRegister << " bits.\n");
5369 if (MaxVectorSize == 0) {
5370 DEBUG(dbgs() << "LV: The target has no vector registers.\n");
5374 assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
5375 " into one vector!");
5377 unsigned VF = MaxVectorSize;
5379 // If we optimize the program for size, avoid creating the tail loop.
5381 // If we are unable to calculate the trip count then don't try to vectorize.
5383 emitAnalysis(Report() << "unable to calculate the loop count due to complex control flow");
5384 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
5388 // Find the maximum SIMD width that can fit within the trip count.
5389 VF = TC % MaxVectorSize;
5394 // If the trip count that we found modulo the vectorization factor is not
5395 // zero then we require a tail.
5397 emitAnalysis(Report() << "cannot optimize for size and vectorize at the same time. Enable vectorization of this loop with '#pragma clang loop vectorize(enable)' when compiling with -Os");
5398 DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
5403 int UserVF = Hints->getWidth();
5405 assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
5406 DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");
5408 Factor.Width = UserVF;
5412 float Cost = expectedCost(1);
5414 const float ScalarCost = Cost;
5417 DEBUG(dbgs() << "LV: Scalar loop costs: " << (int)ScalarCost << ".\n");
5419 bool ForceVectorization = Hints->getForce() == LoopVectorizeHints::FK_Enabled;
5420 // Ignore scalar width, because the user explicitly wants vectorization.
5421 if (ForceVectorization && VF > 1) {
5423 Cost = expectedCost(Width) / (float)Width;
5426 for (unsigned i=2; i <= VF; i*=2) {
5427 // Notice that the vector loop needs to be executed less times, so
5428 // we need to divide the cost of the vector loops by the width of
5429 // the vector elements.
5430 float VectorCost = expectedCost(i) / (float)i;
5431 DEBUG(dbgs() << "LV: Vector loop of width " << i << " costs: " <<
5432 (int)VectorCost << ".\n");
5433 if (VectorCost < Cost) {
5439 DEBUG(if (ForceVectorization && Width > 1 && Cost >= ScalarCost) dbgs()
5440 << "LV: Vectorization seems to be not beneficial, "
5441 << "but was forced by a user.\n");
5442 DEBUG(dbgs() << "LV: Selecting VF: "<< Width << ".\n");
5443 Factor.Width = Width;
5444 Factor.Cost = Width * Cost;
5448 unsigned LoopVectorizationCostModel::getWidestType() {
5449 unsigned MaxWidth = 8;
5452 for (Loop::block_iterator bb = TheLoop->block_begin(),
5453 be = TheLoop->block_end(); bb != be; ++bb) {
5454 BasicBlock *BB = *bb;
5456 // For each instruction in the loop.
5457 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5458 Type *T = it->getType();
5460 // Only examine Loads, Stores and PHINodes.
5461 if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
5464 // Examine PHI nodes that are reduction variables.
5465 if (PHINode *PN = dyn_cast<PHINode>(it))
5466 if (!Legal->getReductionVars()->count(PN))
5469 // Examine the stored values.
5470 if (StoreInst *ST = dyn_cast<StoreInst>(it))
5471 T = ST->getValueOperand()->getType();
5473 // Ignore loaded pointer types and stored pointer types that are not
5474 // consecutive. However, we do want to take consecutive stores/loads of
5475 // pointer vectors into account.
5476 if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
5479 MaxWidth = std::max(MaxWidth,
5480 (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
5488 LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
5490 unsigned LoopCost) {
5492 // -- The unroll heuristics --
5493 // We unroll the loop in order to expose ILP and reduce the loop overhead.
5494 // There are many micro-architectural considerations that we can't predict
5495 // at this level. For example, frontend pressure (on decode or fetch) due to
5496 // code size, or the number and capabilities of the execution ports.
5498 // We use the following heuristics to select the unroll factor:
5499 // 1. If the code has reductions, then we unroll in order to break the cross
5500 // iteration dependency.
5501 // 2. If the loop is really small, then we unroll in order to reduce the loop
5503 // 3. We don't unroll if we think that we will spill registers to memory due
5504 // to the increased register pressure.
5506 // Use the user preference, unless 'auto' is selected.
5507 int UserUF = Hints->getInterleave();
5511 // When we optimize for size, we don't unroll.
5515 // We used the distance for the unroll factor.
5516 if (Legal->getMaxSafeDepDistBytes() != -1U)
5519 // Do not unroll loops with a relatively small trip count.
5520 unsigned TC = SE->getSmallConstantTripCount(TheLoop);
5521 if (TC > 1 && TC < TinyTripCountUnrollThreshold)
5524 unsigned TargetNumRegisters = TTI.getNumberOfRegisters(VF > 1);
5525 DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters <<
5529 if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)
5530 TargetNumRegisters = ForceTargetNumScalarRegs;
5532 if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)
5533 TargetNumRegisters = ForceTargetNumVectorRegs;
5536 LoopVectorizationCostModel::RegisterUsage R = calculateRegisterUsage();
5537 // We divide by these constants so assume that we have at least one
5538 // instruction that uses at least one register.
5539 R.MaxLocalUsers = std::max(R.MaxLocalUsers, 1U);
5540 R.NumInstructions = std::max(R.NumInstructions, 1U);
5542 // We calculate the unroll factor using the following formula.
5543 // Subtract the number of loop invariants from the number of available
5544 // registers. These registers are used by all of the unrolled instances.
5545 // Next, divide the remaining registers by the number of registers that is
5546 // required by the loop, in order to estimate how many parallel instances
5547 // fit without causing spills. All of this is rounded down if necessary to be
5548 // a power of two. We want power of two unroll factors to simplify any
5549 // addressing operations or alignment considerations.
5550 unsigned UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs) /
5553 // Don't count the induction variable as unrolled.
5554 if (EnableIndVarRegisterHeur)
5555 UF = PowerOf2Floor((TargetNumRegisters - R.LoopInvariantRegs - 1) /
5556 std::max(1U, (R.MaxLocalUsers - 1)));
5558 // Clamp the unroll factor ranges to reasonable factors.
5559 unsigned MaxInterleaveSize = TTI.getMaxInterleaveFactor();
5561 // Check if the user has overridden the unroll max.
5563 if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)
5564 MaxInterleaveSize = ForceTargetMaxScalarInterleaveFactor;
5566 if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)
5567 MaxInterleaveSize = ForceTargetMaxVectorInterleaveFactor;
5570 // If we did not calculate the cost for VF (because the user selected the VF)
5571 // then we calculate the cost of VF here.
5573 LoopCost = expectedCost(VF);
5575 // Clamp the calculated UF to be between the 1 and the max unroll factor
5576 // that the target allows.
5577 if (UF > MaxInterleaveSize)
5578 UF = MaxInterleaveSize;
5582 // Unroll if we vectorized this loop and there is a reduction that could
5583 // benefit from unrolling.
5584 if (VF > 1 && Legal->getReductionVars()->size()) {
5585 DEBUG(dbgs() << "LV: Unrolling because of reductions.\n");
5589 // Note that if we've already vectorized the loop we will have done the
5590 // runtime check and so unrolling won't require further checks.
5591 bool UnrollingRequiresRuntimePointerCheck =
5592 (VF == 1 && Legal->getRuntimePointerCheck()->Need);
5594 // We want to unroll small loops in order to reduce the loop overhead and
5595 // potentially expose ILP opportunities.
5596 DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n');
5597 if (!UnrollingRequiresRuntimePointerCheck &&
5598 LoopCost < SmallLoopCost) {
5599 // We assume that the cost overhead is 1 and we use the cost model
5600 // to estimate the cost of the loop and unroll until the cost of the
5601 // loop overhead is about 5% of the cost of the loop.
5602 unsigned SmallUF = std::min(UF, (unsigned)PowerOf2Floor(SmallLoopCost / LoopCost));
5604 // Unroll until store/load ports (estimated by max unroll factor) are
5606 unsigned StoresUF = UF / (Legal->NumStores ? Legal->NumStores : 1);
5607 unsigned LoadsUF = UF / (Legal->NumLoads ? Legal->NumLoads : 1);
5609 // If we have a scalar reduction (vector reductions are already dealt with
5610 // by this point), we can increase the critical path length if the loop
5611 // we're unrolling is inside another loop. Limit, by default to 2, so the
5612 // critical path only gets increased by one reduction operation.
5613 if (Legal->getReductionVars()->size() &&
5614 TheLoop->getLoopDepth() > 1) {
5615 unsigned F = static_cast<unsigned>(MaxNestedScalarReductionUF);
5616 SmallUF = std::min(SmallUF, F);
5617 StoresUF = std::min(StoresUF, F);
5618 LoadsUF = std::min(LoadsUF, F);
5621 if (EnableLoadStoreRuntimeUnroll && std::max(StoresUF, LoadsUF) > SmallUF) {
5622 DEBUG(dbgs() << "LV: Unrolling to saturate store or load ports.\n");
5623 return std::max(StoresUF, LoadsUF);
5626 DEBUG(dbgs() << "LV: Unrolling to reduce branch cost.\n");
5630 DEBUG(dbgs() << "LV: Not Unrolling.\n");
5634 LoopVectorizationCostModel::RegisterUsage
5635 LoopVectorizationCostModel::calculateRegisterUsage() {
5636 // This function calculates the register usage by measuring the highest number
5637 // of values that are alive at a single location. Obviously, this is a very
5638 // rough estimation. We scan the loop in a topological order in order and
5639 // assign a number to each instruction. We use RPO to ensure that defs are
5640 // met before their users. We assume that each instruction that has in-loop
5641 // users starts an interval. We record every time that an in-loop value is
5642 // used, so we have a list of the first and last occurrences of each
5643 // instruction. Next, we transpose this data structure into a multi map that
5644 // holds the list of intervals that *end* at a specific location. This multi
5645 // map allows us to perform a linear search. We scan the instructions linearly
5646 // and record each time that a new interval starts, by placing it in a set.
5647 // If we find this value in the multi-map then we remove it from the set.
5648 // The max register usage is the maximum size of the set.
5649 // We also search for instructions that are defined outside the loop, but are
5650 // used inside the loop. We need this number separately from the max-interval
5651 // usage number because when we unroll, loop-invariant values do not take
5653 LoopBlocksDFS DFS(TheLoop);
5657 R.NumInstructions = 0;
5659 // Each 'key' in the map opens a new interval. The values
5660 // of the map are the index of the 'last seen' usage of the
5661 // instruction that is the key.
5662 typedef DenseMap<Instruction*, unsigned> IntervalMap;
5663 // Maps instruction to its index.
5664 DenseMap<unsigned, Instruction*> IdxToInstr;
5665 // Marks the end of each interval.
5666 IntervalMap EndPoint;
5667 // Saves the list of instruction indices that are used in the loop.
5668 SmallSet<Instruction*, 8> Ends;
5669 // Saves the list of values that are used in the loop but are
5670 // defined outside the loop, such as arguments and constants.
5671 SmallPtrSet<Value*, 8> LoopInvariants;
5674 for (LoopBlocksDFS::RPOIterator bb = DFS.beginRPO(),
5675 be = DFS.endRPO(); bb != be; ++bb) {
5676 R.NumInstructions += (*bb)->size();
5677 for (BasicBlock::iterator it = (*bb)->begin(), e = (*bb)->end(); it != e;
5679 Instruction *I = it;
5680 IdxToInstr[Index++] = I;
5682 // Save the end location of each USE.
5683 for (unsigned i = 0; i < I->getNumOperands(); ++i) {
5684 Value *U = I->getOperand(i);
5685 Instruction *Instr = dyn_cast<Instruction>(U);
5687 // Ignore non-instruction values such as arguments, constants, etc.
5688 if (!Instr) continue;
5690 // If this instruction is outside the loop then record it and continue.
5691 if (!TheLoop->contains(Instr)) {
5692 LoopInvariants.insert(Instr);
5696 // Overwrite previous end points.
5697 EndPoint[Instr] = Index;
5703 // Saves the list of intervals that end with the index in 'key'.
5704 typedef SmallVector<Instruction*, 2> InstrList;
5705 DenseMap<unsigned, InstrList> TransposeEnds;
5707 // Transpose the EndPoints to a list of values that end at each index.
5708 for (IntervalMap::iterator it = EndPoint.begin(), e = EndPoint.end();
5710 TransposeEnds[it->second].push_back(it->first);
5712 SmallSet<Instruction*, 8> OpenIntervals;
5713 unsigned MaxUsage = 0;
5716 DEBUG(dbgs() << "LV(REG): Calculating max register usage:\n");
5717 for (unsigned int i = 0; i < Index; ++i) {
5718 Instruction *I = IdxToInstr[i];
5719 // Ignore instructions that are never used within the loop.
5720 if (!Ends.count(I)) continue;
5722 // Remove all of the instructions that end at this location.
5723 InstrList &List = TransposeEnds[i];
5724 for (unsigned int j=0, e = List.size(); j < e; ++j)
5725 OpenIntervals.erase(List[j]);
5727 // Count the number of live interals.
5728 MaxUsage = std::max(MaxUsage, OpenIntervals.size());
5730 DEBUG(dbgs() << "LV(REG): At #" << i << " Interval # " <<
5731 OpenIntervals.size() << '\n');
5733 // Add the current instruction to the list of open intervals.
5734 OpenIntervals.insert(I);
5737 unsigned Invariant = LoopInvariants.size();
5738 DEBUG(dbgs() << "LV(REG): Found max usage: " << MaxUsage << '\n');
5739 DEBUG(dbgs() << "LV(REG): Found invariant usage: " << Invariant << '\n');
5740 DEBUG(dbgs() << "LV(REG): LoopSize: " << R.NumInstructions << '\n');
5742 R.LoopInvariantRegs = Invariant;
5743 R.MaxLocalUsers = MaxUsage;
5747 unsigned LoopVectorizationCostModel::expectedCost(unsigned VF) {
5751 for (Loop::block_iterator bb = TheLoop->block_begin(),
5752 be = TheLoop->block_end(); bb != be; ++bb) {
5753 unsigned BlockCost = 0;
5754 BasicBlock *BB = *bb;
5756 // For each instruction in the old loop.
5757 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
5758 // Skip dbg intrinsics.
5759 if (isa<DbgInfoIntrinsic>(it))
5762 unsigned C = getInstructionCost(it, VF);
5764 // Check if we should override the cost.
5765 if (ForceTargetInstructionCost.getNumOccurrences() > 0)
5766 C = ForceTargetInstructionCost;
5769 DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " <<
5770 VF << " For instruction: " << *it << '\n');
5773 // We assume that if-converted blocks have a 50% chance of being executed.
5774 // When the code is scalar then some of the blocks are avoided due to CF.
5775 // When the code is vectorized we execute all code paths.
5776 if (VF == 1 && Legal->blockNeedsPredication(*bb))
5785 /// \brief Check whether the address computation for a non-consecutive memory
5786 /// access looks like an unlikely candidate for being merged into the indexing
5789 /// We look for a GEP which has one index that is an induction variable and all
5790 /// other indices are loop invariant. If the stride of this access is also
5791 /// within a small bound we decide that this address computation can likely be
5792 /// merged into the addressing mode.
5793 /// In all other cases, we identify the address computation as complex.
5794 static bool isLikelyComplexAddressComputation(Value *Ptr,
5795 LoopVectorizationLegality *Legal,
5796 ScalarEvolution *SE,
5797 const Loop *TheLoop) {
5798 GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
5802 // We are looking for a gep with all loop invariant indices except for one
5803 // which should be an induction variable.
5804 unsigned NumOperands = Gep->getNumOperands();
5805 for (unsigned i = 1; i < NumOperands; ++i) {
5806 Value *Opd = Gep->getOperand(i);
5807 if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&
5808 !Legal->isInductionVariable(Opd))
5812 // Now we know we have a GEP ptr, %inv, %ind, %inv. Make sure that the step
5813 // can likely be merged into the address computation.
5814 unsigned MaxMergeDistance = 64;
5816 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Ptr));
5820 // Check the step is constant.
5821 const SCEV *Step = AddRec->getStepRecurrence(*SE);
5822 // Calculate the pointer stride and check if it is consecutive.
5823 const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
5827 const APInt &APStepVal = C->getValue()->getValue();
5829 // Huge step value - give up.
5830 if (APStepVal.getBitWidth() > 64)
5833 int64_t StepVal = APStepVal.getSExtValue();
5835 return StepVal > MaxMergeDistance;
5838 static bool isStrideMul(Instruction *I, LoopVectorizationLegality *Legal) {
5839 if (Legal->hasStride(I->getOperand(0)) || Legal->hasStride(I->getOperand(1)))
5845 LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
5846 // If we know that this instruction will remain uniform, check the cost of
5847 // the scalar version.
5848 if (Legal->isUniformAfterVectorization(I))
5851 Type *RetTy = I->getType();
5852 Type *VectorTy = ToVectorTy(RetTy, VF);
5854 // TODO: We need to estimate the cost of intrinsic calls.
5855 switch (I->getOpcode()) {
5856 case Instruction::GetElementPtr:
5857 // We mark this instruction as zero-cost because the cost of GEPs in
5858 // vectorized code depends on whether the corresponding memory instruction
5859 // is scalarized or not. Therefore, we handle GEPs with the memory
5860 // instruction cost.
5862 case Instruction::Br: {
5863 return TTI.getCFInstrCost(I->getOpcode());
5865 case Instruction::PHI:
5866 //TODO: IF-converted IFs become selects.
5868 case Instruction::Add:
5869 case Instruction::FAdd:
5870 case Instruction::Sub:
5871 case Instruction::FSub:
5872 case Instruction::Mul:
5873 case Instruction::FMul:
5874 case Instruction::UDiv:
5875 case Instruction::SDiv:
5876 case Instruction::FDiv:
5877 case Instruction::URem:
5878 case Instruction::SRem:
5879 case Instruction::FRem:
5880 case Instruction::Shl:
5881 case Instruction::LShr:
5882 case Instruction::AShr:
5883 case Instruction::And:
5884 case Instruction::Or:
5885 case Instruction::Xor: {
5886 // Since we will replace the stride by 1 the multiplication should go away.
5887 if (I->getOpcode() == Instruction::Mul && isStrideMul(I, Legal))
5889 // Certain instructions can be cheaper to vectorize if they have a constant
5890 // second vector operand. One example of this are shifts on x86.
5891 TargetTransformInfo::OperandValueKind Op1VK =
5892 TargetTransformInfo::OK_AnyValue;
5893 TargetTransformInfo::OperandValueKind Op2VK =
5894 TargetTransformInfo::OK_AnyValue;
5895 TargetTransformInfo::OperandValueProperties Op1VP =
5896 TargetTransformInfo::OP_None;
5897 TargetTransformInfo::OperandValueProperties Op2VP =
5898 TargetTransformInfo::OP_None;
5899 Value *Op2 = I->getOperand(1);
5901 // Check for a splat of a constant or for a non uniform vector of constants.
5902 if (isa<ConstantInt>(Op2)) {
5903 ConstantInt *CInt = cast<ConstantInt>(Op2);
5904 if (CInt && CInt->getValue().isPowerOf2())
5905 Op2VP = TargetTransformInfo::OP_PowerOf2;
5906 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5907 } else if (isa<ConstantVector>(Op2) || isa<ConstantDataVector>(Op2)) {
5908 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
5909 Constant *SplatValue = cast<Constant>(Op2)->getSplatValue();
5911 ConstantInt *CInt = dyn_cast<ConstantInt>(SplatValue);
5912 if (CInt && CInt->getValue().isPowerOf2())
5913 Op2VP = TargetTransformInfo::OP_PowerOf2;
5914 Op2VK = TargetTransformInfo::OK_UniformConstantValue;
5918 return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK,
5921 case Instruction::Select: {
5922 SelectInst *SI = cast<SelectInst>(I);
5923 const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
5924 bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
5925 Type *CondTy = SI->getCondition()->getType();
5927 CondTy = VectorType::get(CondTy, VF);
5929 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
5931 case Instruction::ICmp:
5932 case Instruction::FCmp: {
5933 Type *ValTy = I->getOperand(0)->getType();
5934 VectorTy = ToVectorTy(ValTy, VF);
5935 return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
5937 case Instruction::Store:
5938 case Instruction::Load: {
5939 StoreInst *SI = dyn_cast<StoreInst>(I);
5940 LoadInst *LI = dyn_cast<LoadInst>(I);
5941 Type *ValTy = (SI ? SI->getValueOperand()->getType() :
5943 VectorTy = ToVectorTy(ValTy, VF);
5945 unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
5946 unsigned AS = SI ? SI->getPointerAddressSpace() :
5947 LI->getPointerAddressSpace();
5948 Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
5949 // We add the cost of address computation here instead of with the gep
5950 // instruction because only here we know whether the operation is
5953 return TTI.getAddressComputationCost(VectorTy) +
5954 TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5956 // Scalarized loads/stores.
5957 int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
5958 bool Reverse = ConsecutiveStride < 0;
5959 unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
5960 unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
5961 if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
5962 bool IsComplexComputation =
5963 isLikelyComplexAddressComputation(Ptr, Legal, SE, TheLoop);
5965 // The cost of extracting from the value vector and pointer vector.
5966 Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
5967 for (unsigned i = 0; i < VF; ++i) {
5968 // The cost of extracting the pointer operand.
5969 Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
5970 // In case of STORE, the cost of ExtractElement from the vector.
5971 // In case of LOAD, the cost of InsertElement into the returned
5973 Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
5974 Instruction::InsertElement,
5978 // The cost of the scalar loads/stores.
5979 Cost += VF * TTI.getAddressComputationCost(PtrTy, IsComplexComputation);
5980 Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
5985 // Wide load/stores.
5986 unsigned Cost = TTI.getAddressComputationCost(VectorTy);
5987 Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
5990 Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
5994 case Instruction::ZExt:
5995 case Instruction::SExt:
5996 case Instruction::FPToUI:
5997 case Instruction::FPToSI:
5998 case Instruction::FPExt:
5999 case Instruction::PtrToInt:
6000 case Instruction::IntToPtr:
6001 case Instruction::SIToFP:
6002 case Instruction::UIToFP:
6003 case Instruction::Trunc:
6004 case Instruction::FPTrunc:
6005 case Instruction::BitCast: {
6006 // We optimize the truncation of induction variable.
6007 // The cost of these is the same as the scalar operation.
6008 if (I->getOpcode() == Instruction::Trunc &&
6009 Legal->isInductionVariable(I->getOperand(0)))
6010 return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
6011 I->getOperand(0)->getType());
6013 Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
6014 return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
6016 case Instruction::Call: {
6017 CallInst *CI = cast<CallInst>(I);
6018 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
6019 assert(ID && "Not an intrinsic call!");
6020 Type *RetTy = ToVectorTy(CI->getType(), VF);
6021 SmallVector<Type*, 4> Tys;
6022 for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
6023 Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
6024 return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
6027 // We are scalarizing the instruction. Return the cost of the scalar
6028 // instruction, plus the cost of insert and extract into vector
6029 // elements, times the vector width.
6032 if (!RetTy->isVoidTy() && VF != 1) {
6033 unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
6035 unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
6038 // The cost of inserting the results plus extracting each one of the
6040 Cost += VF * (InsCost + ExtCost * I->getNumOperands());
6043 // The cost of executing VF copies of the scalar instruction. This opcode
6044 // is unknown. Assume that it is the same as 'mul'.
6045 Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
6051 Type* LoopVectorizationCostModel::ToVectorTy(Type *Scalar, unsigned VF) {
6052 if (Scalar->isVoidTy() || VF == 1)
6054 return VectorType::get(Scalar, VF);
6057 char LoopVectorize::ID = 0;
6058 static const char lv_name[] = "Loop Vectorization";
6059 INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
6060 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
6061 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
6062 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfo)
6063 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
6064 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
6065 INITIALIZE_PASS_DEPENDENCY(LCSSA)
6066 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
6067 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
6068 INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
6071 Pass *createLoopVectorizePass(bool NoUnrolling, bool AlwaysVectorize) {
6072 return new LoopVectorize(NoUnrolling, AlwaysVectorize);
6076 bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
6077 // Check for a store.
6078 if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
6079 return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
6081 // Check for a load.
6082 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
6083 return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
6089 void InnerLoopUnroller::scalarizeInstruction(Instruction *Instr,
6090 bool IfPredicateStore) {
6091 assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
6092 // Holds vector parameters or scalars, in case of uniform vals.
6093 SmallVector<VectorParts, 4> Params;
6095 setDebugLocFromInst(Builder, Instr);
6097 // Find all of the vectorized parameters.
6098 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
6099 Value *SrcOp = Instr->getOperand(op);
6101 // If we are accessing the old induction variable, use the new one.
6102 if (SrcOp == OldInduction) {
6103 Params.push_back(getVectorValue(SrcOp));
6107 // Try using previously calculated values.
6108 Instruction *SrcInst = dyn_cast<Instruction>(SrcOp);
6110 // If the src is an instruction that appeared earlier in the basic block
6111 // then it should already be vectorized.
6112 if (SrcInst && OrigLoop->contains(SrcInst)) {
6113 assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
6114 // The parameter is a vector value from earlier.
6115 Params.push_back(WidenMap.get(SrcInst));
6117 // The parameter is a scalar from outside the loop. Maybe even a constant.
6118 VectorParts Scalars;
6119 Scalars.append(UF, SrcOp);
6120 Params.push_back(Scalars);
6124 assert(Params.size() == Instr->getNumOperands() &&
6125 "Invalid number of operands");
6127 // Does this instruction return a value ?
6128 bool IsVoidRetTy = Instr->getType()->isVoidTy();
6130 Value *UndefVec = IsVoidRetTy ? nullptr :
6131 UndefValue::get(Instr->getType());
6132 // Create a new entry in the WidenMap and initialize it to Undef or Null.
6133 VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
6135 Instruction *InsertPt = Builder.GetInsertPoint();
6136 BasicBlock *IfBlock = Builder.GetInsertBlock();
6137 BasicBlock *CondBlock = nullptr;
6140 Loop *VectorLp = nullptr;
6141 if (IfPredicateStore) {
6142 assert(Instr->getParent()->getSinglePredecessor() &&
6143 "Only support single predecessor blocks");
6144 Cond = createEdgeMask(Instr->getParent()->getSinglePredecessor(),
6145 Instr->getParent());
6146 VectorLp = LI->getLoopFor(IfBlock);
6147 assert(VectorLp && "Must have a loop for this block");
6150 // For each vector unroll 'part':
6151 for (unsigned Part = 0; Part < UF; ++Part) {
6152 // For each scalar that we create:
6154 // Start an "if (pred) a[i] = ..." block.
6155 Value *Cmp = nullptr;
6156 if (IfPredicateStore) {
6157 if (Cond[Part]->getType()->isVectorTy())
6159 Builder.CreateExtractElement(Cond[Part], Builder.getInt32(0));
6160 Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Cond[Part],
6161 ConstantInt::get(Cond[Part]->getType(), 1));
6162 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store");
6163 LoopVectorBody.push_back(CondBlock);
6164 VectorLp->addBasicBlockToLoop(CondBlock, LI->getBase());
6165 // Update Builder with newly created basic block.
6166 Builder.SetInsertPoint(InsertPt);
6169 Instruction *Cloned = Instr->clone();
6171 Cloned->setName(Instr->getName() + ".cloned");
6172 // Replace the operands of the cloned instructions with extracted scalars.
6173 for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
6174 Value *Op = Params[op][Part];
6175 Cloned->setOperand(op, Op);
6178 // Place the cloned scalar in the new loop.
6179 Builder.Insert(Cloned);
6181 // If the original scalar returns a value we need to place it in a vector
6182 // so that future users will be able to use it.
6184 VecResults[Part] = Cloned;
6187 if (IfPredicateStore) {
6188 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else");
6189 LoopVectorBody.push_back(NewIfBlock);
6190 VectorLp->addBasicBlockToLoop(NewIfBlock, LI->getBase());
6191 Builder.SetInsertPoint(InsertPt);
6192 Instruction *OldBr = IfBlock->getTerminator();
6193 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr);
6194 OldBr->eraseFromParent();
6195 IfBlock = NewIfBlock;
6200 void InnerLoopUnroller::vectorizeMemoryInstruction(Instruction *Instr) {
6201 StoreInst *SI = dyn_cast<StoreInst>(Instr);
6202 bool IfPredicateStore = (SI && Legal->blockNeedsPredication(SI->getParent()));
6204 return scalarizeInstruction(Instr, IfPredicateStore);
6207 Value *InnerLoopUnroller::reverseVector(Value *Vec) {
6211 Value *InnerLoopUnroller::getBroadcastInstrs(Value *V) {
6215 Value *InnerLoopUnroller::getConsecutiveVector(Value* Val, int StartIdx,
6217 // When unrolling and the VF is 1, we only need to add a simple scalar.
6218 Type *ITy = Val->getType();
6219 assert(!ITy->isVectorTy() && "Val must be a scalar");
6220 Constant *C = ConstantInt::get(ITy, StartIdx, Negate);
6221 return Builder.CreateAdd(Val, C, "induction");